Vous êtes sur la page 1sur 422

1

COMPUTATIONAL FLUID DYNAMICS


OF BUBBLES, DROPS AND PARTICLES


E. Loth
University of Illinois at Urbana-Champaign

Draft for Cambridge University Press
April 6, 2009



ACKNOWLEDGEMENTS

This work is a result of my interactions with students at the University of Illinois and my
colleagues in the multiphase flow community. I have tried to collect and organize their
brilliant ideas and work in a simple engineering discussion. In particular, I want to express my
appreciation to Prof. G.M. Faeth who introduced me to this subject. I would also like to
acknowledge the various sponsors of our multiphase research including AEDC, AFOSR,
DARPA, NASA, NCSA, NSF, ONR, SAIC, as well as Cambridge University and their Press.
I am especially grateful to all those who provided figures and suggestions for the text (you
know who you are!). To honor all of those who have contributed, 100% of the author proceeds
will be directed to the International Red Cross.
Of course, this book would not have been possible without the support of my wonderful
parents, wife, and children, to whom this book is sincerely dedicated.


2
PREFACE

The goal of this book is to present a basic overview of the equations, techniques, and
guidelines concerning engineering-level numerical simulation of dispersed multiphase flow
motion. Unlike most single-phase CFD (Computational Fluid Dynamics), the problems for
multiphase CFD are often associated with first determining the appropriate set of equations to
solve as opposed to only concentrating on the appropriate numerical methods to solve the
equations. Thus, the text will concentrate on both determining the relevant governing
equations and models for a particular set of fluid physics of interest as well as the appropriate
methods for their numerical solution. Both experiments and theoretical analysis will be
discussed in order to help explain and support the general principles of the governing equations
and physics. It is hoped that this book allows the reader an introduction to various tools
(numerical methods) for a "job" (prediction of a particular set of dispersed-phase properties)
which are appropriate for various situations (flow conditions and numerical resources). The
choice of a tool is often based on an appropriate balance between performance (accuracy
with respect to specific predicted characteristics) and cost (required computational resources).
The key physical process considered herein is momentum transfer to and from the particles,
with particular attention to turbulent flow aspects.
The target audience for this book (graduate students, engineers, etc.) is not assumed to have
any particular exposure to multiphase flow but is assumed to be familiar with single-phase
fluid dynamics. To aid in notation and to review relevant concepts, a brief summary of the
single-phase equations and flow regimes are given in Appendix A, while pertinent aspects of
single-phase computational fluid dynamics are overviewed in Appendix B.
The first three chapters focus on the physics of multiphase flow. In Chapter 1, the
multiphase flow issues are introduced and equations of motion for a single spherical particle
are discussed. Chapter 2 characterizes size, shape and trajectories for both solid and fluid
particles in various flows. In Chapter 3, particle motion is considered with respect to the
dynamics of many particles. From this, the coupling regimes of multiphase flow are classified
in terms of effects of the surrounding fluid on the particle motion, the effects of the particles on
the surrounding fluid motion, and the way in which particles can affect each other.
The remaining five chapters focus on the methodologies for multiphase simulation.
Chapter 4 broadly discusses numerical approaches for treatment of both the continuous-phase
and the dispersed-phase. This includes guidelines for choosing numerical approaches for a
particular simulation objective, i.e. choosing "tools" based on performance and cost.
Appendix B reviews some of the computational approaches for single-phase flow for reference.
Chapter 5 considers multiphase approaches for particles which are so small that they are
approximately in equilibrium with the surrounding fluid. In contrast, Chapters 6 and 7 focus
on non-equilibrium formulations using the point-force approach to represent the surface-
integrated forces on a particle. In particular, Chapter 6 describes various theoretical and
empirical point-force force expressions for a particle, while Chapter 7 discusses numerical
approaches for the overall multiphase flow-fields which employ the point-force assumption.
Finally, Chapter 8 overviews resolved-surface numerical methods where the flow around the
particle surface is directly simulated in detail (no point-force approximation).
3
TABLE OF CONTENTS

1. Introduction to Multiphase Fluid Dynamics............................................................. 14
1.1. Scope of the Book .............................................................................................. 14
1.2. Examples of Multiphase Flow in Engineering Systems..................................... 14
1.2.1. Multiphase Flow in Energy Systems .......................................................... 15
1.2.2. Multiphase Flow in Processing Systems..................................................... 16
1.2.3. Multiphase Flow in Biological and Environmental Systems...................... 18
1.3. Basic Terminology and Assumptions ................................................................ 21
1.4. Description of Uncoupled Continuous-Phase Flow........................................... 25
1.5. Equation of Motion for an Isolated Spherical Particle....................................... 27
1.5.1. General Particle Equation of Motion.......................................................... 27
1.5.2. Quasi-Steady Drag Force............................................................................ 29
1.5.3. Surface Forces due to Accelerations........................................................... 39
1.5.4. Simplified Equations of Motion for an Isolated Particle ............................ 42
1.6. Mass and Heat Transfer for an Isolated Spherical Particle ................................ 44
1.7. Chapter 1 Problems ............................................................................................ 44
2. Characterizing Particle Sizes, Shapes and Trajectories ............................................ 49
2.1. Particle Size Distribution ................................................................................... 49
2.1.1. Definition of the Probability Distribution Function.................................... 49
2.1.2. Average and Effective Diameters ............................................................... 51
2.1.3. Analytical Probability Distribution Functions............................................ 55
2.2. Particle Shapes ................................................................................................... 58
2.2.1. Fixed-Shape Solid Particles ........................................................................ 59
2.2.2. Deformable Fluid Particles ......................................................................... 60
2.3. Dynamics of Particle Orientation, Size and Shape ............................................ 66
2.3.1. Orientation Dynamics in Free-Fall ............................................................. 67
2.3.2. Orientation Dynamics in Simple Shear Flow............................................. 71
2.3.3. Size Oscillations for Bubbles...................................................................... 72
2.3.4. Non-Spherical Shape Oscillations for Fluid Particles ................................ 77
2.4. Fluid Particle Break-up and Coalescence........................................................... 78
2.4.1. Break-up due to Rayleigh-Taylor Instabilties............................................. 79
2.4.2. Break-up due to Kelvin-Helmholtz Instabilities......................................... 82
2.4.3. Break-up due to Shear Deformation ........................................................... 82
2.4.4. Break-up due to Turbulence........................................................................ 85
2.4.5. Coalesence and break-up due to Collisions ................................................ 86
2.5. Chapter 2 Problems ............................................................................................ 44
3. Coupling Physics and Regimes for Multiphase Flow............................................... 93
3.1. Overview of Coupling Regimes......................................................................... 93
3.2. One-Way Coupling and Domain Stokes Number .............................................. 94
3.3. Turbulent Stokes Numbers............................................................................... 100
3.4. Turbulent Particle Dispersion........................................................................... 102
3.5. Turbulent Diffusion and Bias of Particles........................................................ 105
3.5.1. Diffusion of Mean Fluid Tracers .............................................................. 105
4
3.5.2. Diffusion of Finite-Inertia Particles.......................................................... 108
3.5.3. Diffusion of Finite Inertia Particles with Eddy-Crossing Effects............. 115
3.5.4. Lagrangian Average of Particle Relative Velocity................................... 121
3.5.5. Eulerian Average of Particle Relative Velocity........................................ 124
3.6. Brownian Diffusion.......................................................................................... 131
3.7. Two-Way Coupling Criteria............................................................................. 134
3.7.1. Concentration Properties and Interphase Force ........................................ 134
3.7.2. Momentum Coupling Parameter (Effect on Mean Velocity) ................... 136
3.7.3. Turbulence Coupling Parameter (Effect on Kinetic Energy) ................... 139
3.8. Three-Way and Four-Way Coupling Criteria .................................................. 147
3.8.1. Three-Way Coupling: Particle-Particle Fluid Dynamic Interactions........ 147
3.8.2. Four-Way Coupling: Particle-Particle Collisions ..................................... 150
3.9. Multiphase Flow-Coupling Summary.............................................................. 154
3.10. Chapter 3 Problems .......................................................................................... 156
4. Overview of Multiphase Flow Numerical Approaches .......................................... 158
4.1. Dispersed-Phase Flow Numerical Approaches................................................ 159
4.1.1. Lagrangian vs. Eulerian Approaches ........................................................ 159
4.1.2. Mixed-Fluid vs. Separated vs. Weakly-Separated Approaches................ 167
4.1.3. Point-Force vs. Distributed-Force Approaches......................................... 173
4.1.4. Resolved-Surface Approaches.................................................................. 176
4.2. Comparing Physics Capabilties of Multiphase Methods ................................. 178
4.2.1. Predicting Particle Dynamics................................................................... 179
4.2.2. Predicting Turbulent Dispersion, Diffusion and Biases ........................... 182
4.2.3. Predicting Two-, Three- and Four-Way Coupling.................................... 187
4.2.4. Physics-Based Choices of Multiphase Methodologies............................. 189
4.3. Comparing Computational Costs of Multiphase Methods............................... 189
4.3.1. Number of Cells for Continuous-Phase Methods ..................................... 190
4.3.2. Number of Cells or Locations for Dispersed-Phase Methods .................. 194
4.4. Chapter 4 Problems .......................................................................................... 195
5. Mixed-Fluid and Weakly-Separated-Fluid Approaches ......................................... 196
5.1. Mixed-Fluid Properties .................................................................................... 196
5.1.1. Mixed-Fluid Density and Mass Fractions................................................. 196
5.1.2. Mixed-Fluid Viscosity .............................................................................. 198
5.1.3. Mixed-Fluid Compressibility.................................................................... 202
5.2. Mixed-Fluid Methodologies............................................................................. 208
5.2.1. Mixed-Fluid Transport Equations................................................................ 208
5.2.2. Time-Averaged Transport Equations........................................................ 211
5.3. Weakly-Separated-Fluid Methods.................................................................... 219
5.3.1. Weakly-Separated-Fluid Transport Equations.......................................... 219
5.3.2. Time-Averaged Transport Equation and Particle Velocity ...................... 224
5.3.3. Capability of Approaches for Particle Velocity Effects ........................... 226
5.4. Chapter 5 Problems .......................................................................................... 227
6. Point-Force Components ........................................................................................ 229
6.1. General Decomposition of Point Forces .......................................................... 229
5
6.1.1. General Surface Point-Force Expression.................................................. 229
6.1.2. Theoretical Point-Force Momentum Equations........................................ 230
6.2. Drag Force........................................................................................................ 233
6.2.1. Spherical Solid Particles ........................................................................... 233
6.2.2. Influence of Continuous-Phase Compressibility and Rarefaction............ 236
6.2.3. Influence of Mass Flux, Turbulence and Particle Roughness................... 242
6.2.4. Non-Spherical Solid Particles................................................................... 247
6.2.5. Spherical Fluid Particles ........................................................................... 258
6.2.6. Deformed Fluid Particles .......................................................................... 263
6.2.7. Influence of Vorticity and Particle Rotation ............................................ 276
6.3. Lift Force and Torque ...................................................................................... 277
6.3.1. Vorticity-Induced Lift ............................................................................... 279
6.3.2. Spin-Induced Lift ...................................................................................... 282
6.3.3. Lift and Torque with both Particle and Fluid Rotation............................. 284
6.3.4. Lift and Torque for Non-Spherical Particles ............................................ 288
6.4. Virtual Mass Force........................................................................................... 291
6.5. Fluid-Stress Force ............................................................................................ 295
6.6. History Forces .................................................................................................. 296
6.7. Thermophoretic Forces .................................................................................... 300
6.8. Particle-Wall Fluid Dynamic Forces................................................................ 302
6.9. Particle-Particle Fluid Dynamic Forces ........................................................... 307
6.10. Particle-Wall and Particle-Particle Collision Effects....................................... 313
6.11. Chapter 6 Problems .......................................................................................... 307
7. Point-Force Numerical Techniques ........................................................................ 322
7.1. Eulerian Point-Force Formulations .................................................................. 322
7.1.1. Generalized Eulerian-Eulerian PDEs and Numerical Methods................ 323
7.1.2. Incompressible Particles in a Compressible Gas...................................... 325
7.1.3. Incompressible Particles in an Incompressible Fluid............................... 328
7.1.4. Bubbles in an Incompressible Liquid....................................................... 331
7.1.5. Methods for Polydispersion .................................................................... 335
7.1.6. Particle-Wall Interactions......................................................................... 335
7.1.7. Time-Averaged Turbulent Multiphase PDEs ........................................... 338
7.1.8. Eulerian Models for Turbulent Particle Diffusion.................................... 345
7.1.9. Eulerian Models for Continuous-Phase Turbulence Modulation ............. 351
7.2. Lagrangian Point-Force Formulations ............................................................. 353
7.2.1. Lagrangian ODEs for Particle Properties ................................................. 353
7.2.2. Interpolating Fluid Properties at the Particle Location............................. 357
7.2.3. Lagrangian Particle Advection Schemes .................................................. 361
7.2.4. Representing Polydisperse Distributions .................................................. 369
7.2.5. Lagrangian Models for Turbulent Diffusion............................................. 371
7.2.6. Lagrangian Models for Brownian Diffusion............................................. 383
7.2.7. Eulerian Cell-Averages of Lagrangian Particle Properties....................... 384
7.2.8. Time-Averages of Lagrangian Particle Properties.................................... 384
7.2.9. Particle-Wall Interaction Methods............................................................ 391
7.2.10. Particle-Particle Interaction Methods ....................................................... 391
7.3. Lagrangian Distributed-Force Formulations.................................................... 401
6
7.3.1. Spatially-Averaged Methods .................................................................... 406
7.3.2. Semi-Resolved Methods .................................................................................
7.4. Chapter 7 Problems ................................................................................................

8. Resolved-Surface Approach ................................................................................... 409
8.1. Surface-Fitted Grid Methods............................................................................ 409
8.1.1. Transport Equations, Grids, and Boundary Conditions............................ 410
8.1.2. Particle Equations of Motion .................................................................... 414
8.2. Continuous-Interface Methods for Solid Particles........................................... 416
8.3. Continuous-Interface Methods for Fluid Particles........................................... 416
8.3.1. Mixed-Fluid Transport Equations and Boundary Conditions.........................
8.3.2. Volume-of-Fluid Methods ...........................Error! Bookmark not defined.
8.3.3. Level-Set Methods.......................................Error! Bookmark not defined.
8.3.4. Front-Tracking Methods ..............................Error! Bookmark not defined.
8.4. Chapter 8 Problems .......................................................................................... 429
Appendix A: Single-Phase Flow Equations and Regimes...................................................
A.1. Continuous-Flow Governing Equations.................................................................
A.2. Inviscid Flow Equations and Mach Number Regimes...........................................
A.3. Incompressible Viscous Flow Equations ...............................................................
A.4. Reynolds Number Regimes for the Continuous-Phase..........................................
A.5. Average Properties and Equations for Turbulent Flow..........................................
A.5.1. Averages of Fluctuating Velocities.................................................................
A.5.2. Homogeneous Isotropic Stationary Turbulence..............................................
A.5.3. Integral-scales in Turbulent Flows..................................................................
A.5.4. Micro-Scales and Energy Cascade in Turbulent Flows..................................
A.5.5. Time-Averaged Navier-Stokes Equations ......................................................
A.5.6. Techniques for Closing the Turbulence Equations.........................................
A.5.7. Fully-Resolved and Partially-Resolved Techniques.......................................
Appendix B: Single-Phase Flow Discretization Techniques...............................................
B.1. Computational Grids and Local Shape-Functions .................................................
B.2. Weighted Residual Methods ..................................................................................
B.3. Finite Difference Methods .....................................................................................
B.3.1. Spatial Discretization......................................................................................
B.3.2. Temporal Discretization..................................................................................
B.3.3. Incompressible Flow Implicit Schemes..........................................................
B.3.4. Compressible Flow Flux-Based Schemes.......................................................
B.4. Finite Volume Method ...........................................................................................
B.5. Finite Element Method...........................................................................................
B.6. Spectral Method .....................................................................................................
B.7. Lattice Boltzman Method.......................................................................................
B.8. Direct Simulation Monte-Carlo Method ................................................................

7
Nomenclature

English Variables
a speed of sound (Eq. A.20)
A area or boundary surface (Eq. 1.1b)
A bubble oscillation amplitude (Eq. 2.49)
A* area normalized by the surface area of a sphere of same volume (Eq. 6.23)
b Richardson-Zaki exponent (Eq. 6.248)
B compressibility pressure constant (Eq. A.24)
B Spalding transfer number (Eq. 1.86)
Bo Bond number (Eq. 2.43)
c coefficient (Eq. 1.46)
const. an arbitrary constant of order unity
C force coefficient (Eq. 1.46)
C/S cross-section (Eq. 6.36)
c specific heat (Eq. 1.88)
C cumulative distribution function (Eq. 2.3)
Ca capillary number (Eq. 2.53)
d equivalent volumetric diameter of a particle (Eq. 1.1)
d* ratio of small to large particle diameter for a collision (Eq. 2.20)
d particle surface roughness (Eq. 6.113b)
d particle-path derivative (Eq. 1.13a)
D macroscopic length scale of physical domain (Fig. 1.20)
D fluid path derivative (Eq. 1.13b)
E particle aspect ratio (Eq. 2.36)
E Youngs modulus (Eq. 6.229)
e internal energy (Eq. 5.57)
e coefficient of restitution (Eqs. 6.223 and 6.232)
f Stokes drag correction factor (Eq. 1.55)
f frequency (e.g., Eq. 2.65) or a rate function (e.g., Eq. 7.13)
F general function (e.g., Eq. 6.258)
F force per unit mass (e.g., Eq. 7.127)
F force (Eq. 1.1)
f force per mixed-fluid volume
Fr Froude number (Eq. 3.16)
g gravity acceleration vector (Eq. 1.23)
G filter function (Eq. A.132)
G Carnahan-Starling equilibrium radial distribution function (Eq. 5.19)
h spread parameter for a size distribution (Eq. 2.23)
h enthalpy (Eq. 1.88)
H History force kernel (Eq. 6.193b)
H smoothed Heavy-side function (Eq. 8.??)
i unit vector (Eq. 1.16)
I impulse (Eq. 6.21)
8
I moment of inertia (Eq. 6.154)
I* normalized moment of inertia about broadside axis (Eq. 2.48)
j tensor index (Eq. 3.31)
J* normalized McLaughlin lift parameter (Eq. 6.133)
J Jacobian of the Eulerian-Lagrangian transformation (Eq. 7.195)
k turbulent kinetic energy of the surrounding fluid (Eq. A.71)
k thermal conductivity (Eq. 1.91)
K compressibility exponent (Eq. 2.63)
K
ij
viscous stress tensor (Eq. 8.3)
Kn Knudsen number (Eq. 1.19)
l wave-length (Eq. 2.76) or characteristic distance (Eq. A.81)

L Lagrange polynomial (Eq. 7.118)

L partial derivative operator (Eq. B.17)

m mass or shape oscillation mode (Eq. 2.74)
m viscosity function (Eq. 6.52)
m` mass flux (Eq. 1.84)
M Mach number (Eq. A.26)
Mo Morton number (Eq. 2.44)
MW Molecular weight of a gas (Eq. 1.86)
n wave-number (Eq. A.95)
n outward normal vector (Eq. 1.66)
n
p
number of particles per mixed-fluid volume (Eq. 3.141)
N number of realizations, particles, parcels, nodes, etc. (Eq. 3.29)
NSI no summation of indices (Eq. A.75)
N number of collisions per unit volume of mixture (Eq. 3.197)
O order of a term
Oh Ohnesorge number (Eq. 2.86)
p continuous-phase pressure without local flow around particle surface (Eq. 1.9e)
P pressure around or inside of particle (Eq. 1.6)
P probability distribution function (Eq. 2.1a)
q arbitrary variable (Eq. 2.26)
q conservative variable vector for governing equations (Eq. 7.6)
Q flux vector for governing equations (Eq. 7.6)
`
Q heat transfer rate (Eq. 1.90)
r radial distance from particle centroid (Fig. 1.25)
r
p
effective particle radius (Eq. 1.1)
R acceleration parameter (Eq. 1.81a)
Re Reynolds number (Eq. 1.26)
Rn random number with uniform probability between 0 and 1 (Eq. 7.169)
R gas constant (Eq. A.21)
s speed ratio (Eq. 6.104)
S vector of the transport source terms
S Shields number (Eq. 5.13)
Sc Schmidt number, i.e. ratio of momentum to scalar diffusivity (Eq.1.94)
Sh Sherwood number, i.e. ratio of mass to molecular diffusivity (Eq. 1.87)
9
St Stokes number (Eq. 3.12)

STP standard temperature and pressure defined as 20
o
C and 1 atmosphere (Table A.1)
t time (Eq. 1.16)
T temperature (Eq. 1.18)
T fluid stress vector (Eq. 1.18)
T torque (Eq. 6.152)
u continuous-phase velocity without local flow around particle surface (Eq. 1.9b)
U continuous-phase velocity including local flow around particle surface (Eq. 1.8a)
U uncertainty of a variable (Eq. 4.13)
v velocity of the particle centroid (Eq. 1.9a)
V internal particle velocity (Eq. 1.8b), or a general velocity with a subscript
w weighting function (Eq. B.18)
w relative velocity of the dispersed-phase (Eq. 1.9d)
W Faxen-corrected relative velocity (Eq. 6.3)
We Weber number (Eq 2.39)
W Wiener process (Eq. 7.183)
x streamwise direction for Cartesian coordinates (Eq. 1.16)
x position vector (Eq. 1.16)
x relative position vector (Eq. A.75)
x
i
a general Cartesian direction (Eq. 1.16)
X mass fraction (Eq. 3.145, Eq. 5.6)
y transverse or wall-normal direction for Cartesian coordinates (Eq. 1.16)
z spanwise direction for Cartesian coordinates (Eq. 1.16)
Y drag-power parameter (Eq. 2.18)
Z transfer function (Eq. 7.245 or Eq. 8.25)

Latin Variables
dispersed-phase volume fraction
collision impact parameter (Eq. 2.93)
Kronecker delta (Eq. A.10), boundary layer thickness (Eq. 4.76), or interface thickness
(Eq. 8.19)
discretization increment or sub-grid scale property
turbulent dissipation of continuous-fluid (Eq. A.73)
small perturbation (1)
azimuthal angle coordinate (Eq. A.57) or level-set function
velocity potential (Eq. 6.176) or discretization shape function (Eq. 4.42)
gas specific heat ratio (Eq. 6.93)

` shear rate (Eq. 5.53)
gamma function (Eq. 2.26)
particle mass loading (Eq. 3.144)
particle collision angle (Eq. 2.93, Fig. 2.47)
Boltzmann constant (Eq. 3.135)
Kolmogorov length scale of the turbulence (Eq. A.92) or viscosity function (Eq. 6.55)
integral-length scale of the turbulence (Eq. 3.75)
10
dynamic viscosity (Eq. 1.5)
*
p
viscosity ratio (Eq. 1.5)
momentum coupling parameter (Eq. 3.151)
kinematic viscosity (Eq. A.47)
density (Eq. 1.3)
*
p
density ratio (Eq. 1.4)


polar angle for particle-centered coordinates (Fig. 1.25 and Eq. A.57)
mass diffusivity (Eq. 1.87)

turbulent particle diffusivity normalized by that of a fluid tracer (Eq. 3.45)


surface tension (Eq. 1.6)
time-scale (Eq. 1.75)
correlation function (Eq. 3.35)
vapor mole fraction (Eq. 6.124)
volume (Eq. 1.1)
continuous-phase fluid vorticity (Eq. 1.10)
non-dimensional continuous-phase vorticity (Eq. 6.115)
angular rotation rate (of particle, unless otherwise specified) (Eq. 1.10)
non-dimensional particle rotation rate (Eq. 6.115)
decorrelation parameter (Eq. 7.165b)
stream function (Eq. 1.29)
velocity ratio for a particle-particle collision (Eq. 1.60)
Gaussian random number with mean of zero and variance of unity (Eq. 4.71)

English Subscripts
@p continuous-fluid property extrapolated to particle center of mass
b bubble
buoy buoyancy effect
Bass Basset history
Br Brownian motion
clean conditions where the interface is fully mobile
clus cluster of particle due to preferential concentration
cont contaminated conditions where the interface is immobilized
conv convective
coll particle-particle or particle-wall collisions
crit critical condition where flow transitions
cross eddy-crossing effect
diss dissipation
dom fluid domain of interest (entire computational domain)
dyn dynamic pressure
d volumetric diameter or dispersed phase
D drag or overall macroscopic domain scales
E Eulerian
E particle aspect ratio
e element
11
eff effective
eq equilibrium
exact exact solution
f continuous-phase fluid
fr wall friction
fm free-molecular
g gas
g gravity
gap gap between particle surface and wall or another particle
H history force effect
hy hybrid LES/RANS turbulence model
I interface between particle and surrounding fluid
ic particle internal circulation effect
inj injection
init initialization
int interphase interaction
inv inviscid
irreg irregularly-shaped characteristic
Kn Knudsen
KH Kelvin-Helmholtz instability
k due to gradients in the turbulent kinetic energy
L lift
L Lagrangian property
LN Log-Normal size distribution values
l liquid or wave-length
lam laminar
m mixed-fluid
m`
surface mass flux or blowing effect
m-m molecule-molecule interaction
mom mean continuous-phase momentum
M Mach number
mE moving Eulerian property
n number of particles
nat natural frequency
o reference state
osc related to oscillations
p particle-phase (dispersed-phase) or unhindered pressure
pP particles per parcel
p particles per computational cell
p-p particle-particle
p-w particle-wall
plastic onset of plastic deformation
prod turbulent production
proj projected
pseudo pseudo-turbulence
P parcel (computational group of particles) or local pressure
12
P parcels per computational cell
r radial direction
rel local difference between particle and surrounding fluid
ref reference condition
rep repulsion force
rms root of the mean of the squares, i.e. ( )
N
2
rms i
i 1
1
q q
N
=


rot rotational
RR Rosin-Rammler size distribution values
RT Rayleigh-Taylor instability
s sampling point
s constant entropy
S fluid stress
sep separated wake conditions
sf skin friction
shear linear velocity shear
side side force, perpendicular to lift and drag forces
slip molecular slip along a surface
sphere for an equivalent spherical particle
surf surface-averaged quantity, typically of particle
swim due to self-propulsion from shape oscillation
t tangential
term terminal characteristic of particle in quiescent fluid
tot total or stagnation conditions
trans transitional flow condition where flow becomes completely turbulent
turb turbulent
um un-mixed
vap vapor
w wall interaction effect
win window of history integral
yield yield stress value
Z distributed-force or velocity envelope

Latin Subscripts
finite volume fraction effects

added-mass effect or particle volume


cell resolution discretization or sub-grid scale property
integral-scale of turbulence
polar component with respect to particle-centered spherical coordinates
long-time or far-field property

Superscripts
in before interaction with wall
out after interaction with wall
n time-step index
13
+ normalized by wall shear-stress values
* non-dimensional, typically particle property normalized by continuous-fluid property

Functions of an Arbitrary Property q
q

Eulerian time-average, or spatial-filtered value for LES

q' instantaneous fluctuation from time-average ( q q) =
q Eulerian density-weighted (Favre-) average
q" instantaneous fluctuation from density-weighted average ( q q) =
q Lagrangian path-average
q instantaneous fluctuation from path-average ( q q) =
q ensemble-average for many particles
q

Concentration-weighted (Eq. 7.21)


q

Lagrangian transformation (Eq. 8.4)


[ ]
q convected quantity

Comparison Symbols
~ same order of magnitude
approximately equal

/
not approximately equal
equal by definition
<

less than or of the same order


>

more than or of the same order


at least an order of magnitude smaller
at least an order of magnitude larger (and thus will typically dominate)
normal component with respect a wall or to particle axis of symmetry
parallel component with respect a wall or to particle axis of symmetry
=f (q) variable is a function of parameter q
f (q) variable is not a function of parameter q
O(q) on the order of q, e.g., O(1) means the value is on the order of unity
q0 the limit where q can be neglected altogether

14
1. Introduction to Multiphase Fluid Dynamics
1.1. Scope of the Book
This book focuses on simulating engineering-level multiphase fluid dynamics. The term
multiphase flow was coined by the late Prof. Soo of the University of Illinois in 1965 and
includes fluid dynamics of multiple phases. Generally, the surrounding fluid can be a liquid or
a gas while the particles can be solid, liquid, or gas. The focus of this text is on the motion of
particles which are small relative to the overall domain and are dispersed, i.e. have
trajectories more influenced by their interaction with the surrounding fluid (e.g. drag force)
than by collisions with other particles.
Dispersed particle dynamics are common in many engineering systems. Because of this,
robust and accurate description of multiphase flow is important as knowledge of the
characteristics can lead to increases in performance, reduction in cost, and improved safety for
engineering systems. In recent years, Computational Fluid Dynamics (CFD) has become an
indispensable predictive tool for gathering information to be used for design and optimization
for fluid systems. Thus, the combination of CFD and multiphase flow has emerged as an
important research area, but one with unique characteristics and issues. For example,
multiphase flow approaches include a wide variety of approaches such as variations in
reference frame representation (Eulerian or Lagrangian), phase coupling (intra-phase and/or
interphase coupling), and particle/flow detail (e.g., high resolution around a single particle or
bulk description of thousands or millions of particles). These different approaches are
associated with large variations in computer usage and predictive fidelity. Thus, simulation of
dispersed multiphase flow requires careful consideration of both the flow regimes and the
relevant numerical approaches. As such, this text is designed to relate various multiphase fluid
physics to the appropriate governing equations and numerical schemes.
For additional information on such topics and on topics outside the scope of this book,
there are several excellent publications to which the reader is referred. Multiphase fluid
physics are discussed in the texts of Clift et al. 1978, Wallis 1969, Soo 1990, Crowe et al.
(1998), Kleinstreuer (2003), Brennen (2005) and Michaelides (2006). Reviews focusing on
computational multiphase flow include Elghobashi (1994), Faeth (1987), Shirolkar et al.
(1996), Shyy et al. (1997) and Tomiyama (1998), as well as Prosperetti & Tryggvason (2007).
Important studies which focus on the mathematical treatments of two-phase flow include Drew
& Passman (1998) and Prosperetti (1998) while heat and mass transfer aspects are discussed by
Williams (1965), Oran & Boris (1987), Kuo (1986), and Sirignano (1999).


1.2. Examples of Multiphase Flow in Engineering Systems
Dispersed particle-laden multiphase flow is important to many engineering applications
including aerospace, atmospheric, biological, chemical, civil, mechanical, and nuclear systems.
Table 1.1 lists some engineering applications for the four primary conditions of dispersed
multiphase flow: solid particles in a gas, liquid droplets in a gas, solid particles in a liquid, and
gas bubbles in a liquid. Examples from three types of systems which involve multiphase flow
are discussed in the following and include: energy systems (related to the generation of energy
or mechanical work), processing systems (related to the production and transport of fluids), and
environmental systems (related to natural and biological transport processes). As will de
discussed, these flow can be quite complex as they can include a wide range of flow and
15
particle length-scales and time-scale and also include issues of turbulence, convection, settling,
re-suspension, two-way coupling, collisions, coalescence, break-up, erosion, accretion, etc.
Numerical approaches of various systems must therefore be tailored to the relevant physics so
that the simulations can improve performance, alleviate problems, and/or predict overall
behavior.

1.2.1. Multiphase Flow in Energy Systems

Circulating Fluidized Beds

Combustion of coal and other solid fuels for power generation is often most efficiently
accomplished for power generation via a fluidized bed which involves many multiphase
processes. In such a system, the solid fuel is fed into the bed after crushing and the injected
upward-flowing gas flows causes the particles to be approximately suspended and exposed to
high convection velocities and heat transfer rates which, in turn, cause rapid combustion (Basu,
1999). At low flow rates, the fluidized bed is quite compact with so that particle-particle
collisions dominate the bed dynamics. However, the emphasis has recently been on higher-
efficiency low-polluting circulating fluidized beds (Fig. 1.1) where higher gas velocities and
dispersed flow conditions allow particles to burn as they move upward. To avoid particles
exiting the system, vertical channels can be employed to capture the particles and force them
back down into the combustion region for continued recirculation (Fig. 1.2a). Smaller particles
which exit the primary burner can have secondary recirculation in an economizer before they
are re-injected into the lower bed. For finer particles which are not re-circulated, downstream
stages employ other filtering devices such as the cyclone shown in Fig. 1.2b. Employing
various stages of filtration allows maximum particle removal while avoiding significant
pressure losses.


Air-Breathing and Rocket Engines

Air-breathing engines generally involve multiphase flows in three aspects of the process:
the presence of particles within the air intake, the injection and distribution of the fuel spray in
the combustion chamber, as well as the production and evolution soot particles (e.g. Fig. 1.3).
With regard to the air intake aspect, particle ingestion causes erosion of the turbo-machinery
components during near-ground operations in sandy or dusty regions. This problem can be
minimized by inlet flow curvature but the performance is sensitive to particle concentration,
size and trajectory. Water droplet ingestion is another issue at near freezing temperatures since
it can lead to ice formation on compressor blades which can degrade performance and/or
break-off in pieces leading to engine damage. The fuel spray is another important multiphase
aspect, and results in a complex cloud of droplets created from pressurized ejection. The size,
velocity, and spatial distribution of the spray are important in describing the performance of the
system and pollution emission, including soot concentration.
Rocket engines flows are a common aerospace application which involves multiphase flow.
A solid-propellant engine fuels often include aluminum fuel particles to increase thrust levels.
The particles break-off from the internal sidewall surfaces as the binder around them burns
16
away, and they are then transported along the core flow of the combustion chamber (Fig. 1.4).
The particles undergo complex changes, since they transition from solid to liquid and then
evaporate and combust. However, some of the particles do not vaporize before the nozzle
region which can lead to undesired build-up or ablation. There are also a variety of multiphase
issues associated with liquid-fueled rockets such as gasification transition as the fuel and
oxidizer is pumped from tanks and converted into gas.


Nuclear Reactors

Nuclear power plants rely heavily on heat removal by water, which will often boil due
to the high temperatures. Thus, vapor bubble flow convection has been widely studied due to
its importance in the safety and thermal efficiency of these systems. However, these flows can
be complex since they involve non-spherical bubble shapes with changing volumes.
Furthermore, the random passage of bubble wakes in the flow causes complex velocity
fluctuations beyond that of conventional turbulence. The phenomenon has been termed
pseudo-turbulence or burbulence (since it can occur even in laminar flow conditions and is
often most apparent for bubbly systems).

1.2.2. Multiphase Flow in Processing Systems

Chemical Reactors

Multiphase flow is important in many liquid chemical reactors since they often employ
bubble columns and air-lifts (Fig. 1.5) to cause high mass and/or chemical transport rates
among liquids, gasses, and solids. The reactor processes include fermentation, bio-
oxygenation, production of cell cultures and proteins, waste-stream treatment, etc. The transfer
rates are a function of the bubble size distribution, mean and fluctuating velocities and
trajectories, bubble coalescence and break-up, as well as particle-bubble interactions.
Advanced bubble columns employ impellers and non-linear configurations which further
complicate the fluid dynamics.
While most systems focus on chemical reactions outside of the bubbles, there is a related
process which conducts chemical reactions inside a bubble, called sono-chemistry. This
phenomenon is based on acoustically-driven micro-bubble oscillations which momentarily
create very high pressures and temperatures during the contraction phase. This is
accomplished by driving the bubbles with high-amplitude pressure fluctuations near their
natural frequency. As a result, the bubbles will undergo a rapid and violent collapse during the
compression portion such that the gas inside the bubble may be driven to temperatures of more
than 10,000
o
K and pressures of more than 10,000 atmospheres (Suslick & Flannigan, 2005).
As a result, the gas inside the bubble becomes plasma and emit light which can be seen with
the unaided eye. This emission is termed sono-luminescence and an example is shown in
Fig. 1.6a. These intense chemical reactions are valuable for a number of applications including
degradation of organic species in contaminated water and the production of protein micro-
spheres for drug-delivery (Fig. 1.6b), biomedical imaging, and blood substitutes.
17

Pump and Pipe Flows

Liquid pumping by propellers is used in many systems to achieve high throughput.
However, the associated high blade speeds can cause the local pressure to drop below the local
vapor pressure, giving rise to a multiphase flow phenomenon known as cavitation. If their
presence is extensive, the emergence of cavitating vapor bubbles can reduce the overall
pumping performance. In addition, cavitating bubbles may undergo violent condensation
collapse on the blade surfaces which can cause an intense localized pressure wave and cause
surface pitting.
Multiphase flow can also be a benefit since the use of carefully chosen polymers or micro-
bubbles can reduce the viscous losses in a channel flow. In particular, polymer chains added in
very small volume fractions (about 10
-4
) have been found to reduce the mean skin friction by
more than 80% by suppressing turbulent transport near the walls (Fig. 1.7). One of the most
famous examples of this is the successful use of drag-reducing polymer along the Alaskan oil
pipe-line (they have also been used for other flows including fire hose nozzles). Similarly, the
injection of micro-bubbles has been found to reduce the skin friction losses (also shown in Fig.
1.7), though the required volume fractions are typically larger (about 10
-2
or more). While not
as robust and understood, the bubbly flow drag reduction eliminates some environmental and
cost concerns associated with polymers. Therefore, this technique is also being actively
explored for implementation into commercial or military vessels as it could be highly
beneficial in terms of reduced fuel for transportation.
Multiphase flow is also important for analyzing systems which involve transporting
particles in pipes. Particle conveying systems are called slurries when they are liquid-driven
and pneumatic transport when they are air-driven. In both cases, the dynamics of particles
can involve regions where convection effects are important (though, much of the transport
involves dense multiphase regions). Understanding the trajectories of particles in pipe flows
(including sliding, bouncing and re-suspension dynamics) is important to minimize the wall
deposition, especially for turbulent flows. Even in laminar flows, wall deposition can occur
due to random Brownian motion resulting from the collisions of the surrounding fluid
molecules of the particle surface. This phenomenon can be employed for clean rooms
though proper design of the flow filtering.


Sloshing Dynamics

Multiphase flow is also important to the transport of liquids in containers where sloshing
may occur. Such transport is particularly important for Liquid Natural Gas (LNG) shipping as
oceanic vessels have steadily increased in size to respond to mounting international energy
demand. The sub-cooled and pressurized LNG tanks can suffer high impact sloshing loads
when large unsteady ship motions occur (Fig. 1.8a). Small-scale experiments (Fig. 1.8b) have
qualitatively shown that these impact pressures can be substantially cushioned by the presence
of dispersed bubbles due to the resulting compressibility of the mixed fluid. Cavitation and
condensation can also occur due to the rapid changes in pressure and including this phenomena
will further complicate the simulation methodology.

18

Sprayer Systems

Sprayer systems include applications such as surface coating/painting, powder production,
particle filtering, hazard mitigation, etc. For surface coating, it is desirable to apply a spray
with droplets of a particular size range since large droplets tend to cause surface anomalies and
small droplets may not deposit on the target zone and instead be carried away with air currents.
A particularly complex multiphase flow process is that of extreme-durability single-crystal
solid metallic coatings. To achieve such coatings, plasma sprays are employed (Fig. 1.9)
whereby a metal powder is injected into a supersonic high-temperature gas jet and the particles
become molten or semi-molten as they are entrained. The convected particles ideally impinge
on the target surface with an appropriate speed and size to liquefy (with little or no splashing)
and then crystallize. Careful choice of the particle material and sizes as well as the jet flow and
plasma temperatures can result in surface coatings with high toughness, such as used for gas
turbine blades to increase their wear resistance.
Spraying systems have also been used to produce powders, i.e. fine solid particles. In an
industrial spray dryer (Fig. 1.10a), this is typically accomplished by first emulsifying the
material into liquid form and then downward spraying the liquid mixture into a large chamber
in which dry-gas is injected. The droplets from the spray then solidify in flight and settle to
create a powder at the bottom of the chamber (Fig. 1.10b). Using multiple stages increases the
system thermal efficiency and particle size uniformity, and also prevents overheating of
powder particles along their trajectories. Achieving a desired powder size and consistency is
difficult but critical for the production of foods, detergents, pharmaceuticals, ceramics,
herbicides, pesticides, etc.

1.2.3. Multiphase Flow in Biological and Environmental Systems

Particle Motion in the Blood Stream and the Body

The bloodstream is the most important liquid flow within the human body and it may
contain particles by disease or from medical treatments. In terms of disease, cholesterol or
calcium particles can undergo wall deposition and thus can accrete on certain vessel surfaces
(perhaps due to low shear-stress and certain bio-chemical conditions). This is a complex
problem as the flow is pulsatile with three-dimensional geometries (especially near grafts) and
may also be transitional if not turbulent in the large vessels near the heart. The resulting
particle accretion is called plaque deposition and, if severe, can cause a blockage in the blood
vessel (called a stenosis) which can lead to poor circulation and break-off can even lead to
blockage occurs in the brain blood vessels resulting in an intracranial aneurysm, i.e., a stroke.
On the other hand, artificial introduction of particles can be used for purposes of drug
delivery or imaging diagnostics. For example, bubbles can also be injected to serve as a
contrast agent, i.e. a diagnostic marker within the body. In particular, the micro-bubbles
(about three microns in diameter and composed of helium, carbon dioxide or other gasses) can
be injected into the body. The small bubble size allows for low volume concentrations (e.g., a
0.0001%) that are harmless but high number concentrations (e.g. 20,000 bubbles per cc) that
provide a significant acoustic signal when excited by an applied sound field. These signals can
19
show blood-flow patterns in the heart to find any regions with poor circulation or to find
tumors in non-vascular regions (Fig. 1.11).


Particle Motion in Respiratory Passages

The introduction of particles to the human respiratory system is a related and important
dispersed multiphase problem. Oral spray-drug delivery by inhalers are based on the ability of
the aerosol droplets to avoid wall deposition so as to maximize air-borne delivery to the lung
cavity (Fig. 1.12). In the case of airborne pollutants, the opposite is desired, i.e., deposition of
particles in the mouth and nasal passages is favorable since it avoids particulate transport to the
sensitive tissue of the lungs. The probability of particle deposition in the respiratory system
varies significantly with particle size. For particles which are greater than one-micron in
diameter, inertia impact and curved passages dominate the deposition rates. For smaller
particles, say 0.1 to 1 micron in diameter, the velocity caused by gravitational settling becomes
dominant. The deposition of ultra-fine particles (0.1 microns in diameter) is dominated by
Brownian diffusion, i.e. diffusion by the random molecular collisions of the gas with individual
particles. If such particles do not deposit and reach the lungs, they can permanently settle and
do significant harm, particularly among children and the elderly.


Atmospheric Solid Particulates

Because of the potential impact to the respiratory system of humans and animals, there has
been a significant concern regarding the emissions of micron and sub-micron sized particles in
the atmosphere, commonly regarded as particulate pollution. As such, air quality is often
characterized in terms of the micro-particle concentration (Table 1.2). In 2003, over 25 urban
areas were found to have Air Quality Index (AQI) levels above the unhealthy threshold of 100
for more than 10 days. The modeling for these particles can extend to the entire planetary
atmospheric boundary layer as shown in Fig. 1.13a. Furthermore, interaction of such particles
with water droplets from clouds or rain may even affect weather patterns. Particles released in
urban environments have also been investigated as shown in Fig. 1.13b, where one must
consider the effect of the street canyons between buildings where particulate concentration
can rise due to re-circulating (trapped) air. Another motivation to understand such transport of
particles is to be able to assess the threat of terrorists placing a device which emits hazardous
particles. In particular, the consequences of the release of a highly lethal powder (e.g.,
anthrax) or the use of a dirty bomb (where an explosive device is used to scatter radioactive
materials) are now being carefully studied.
There are also natural events which can cause very high particulate concentrations,
including particles emitted from a large fire or by gigantic dust storms (Fig. 1.14). Knowing
the particle concentration and movement for such events is important to environmental
modeling and to determine regions suitable for aircraft operation and to assess impact on
surface water and agricultural areas. Another natural event involving particle injection into the
atmosphere is the case of volcanic eruptions for which ash particles can eject upwards at
speeds of hundreds of meters per second and disperse on a continental scale as a result. A
particularly deadly condition can occur if the particle-laden cloud has a somewhat lower
20
velocity but a very high particle mass content so that its effective much denser than the
surrounding unladen air. In this case, the particle cloud may fall and be driven along the
ground forming what geologists call a pyroclastic flow (Fig. 1.15). Such flows can have
temperatures up to 600
o
C with speeds up to 100 km/hr and can propagate up to 60 kilometers.
These particle-driven or gravity-current pyroclastic flows have the capacity to rapidly
incinerate and asphyxiate such that they have killed more people than traditional upward-
spewing volcanic eruptions (Chester, 1993).


Atmospheric Liquid Particles

The formation and coalescence of drops in clouds is another area in which multiphase CFD
is being applied to understand the controlling processes. The prediction and understanding of
rain cloud formation is important to weather prediction and climatology. However, the
mechanisms which cause clouds to rapidly transition from small droplets (about 1-20 microns)
into large rain drops (about 100-1000 m) during the formation of a thunderstorm are not well
understood.
For aircraft, droplets concentrations and size impact the visibility for pilots, but take on
additional significance at near-freezing conditions as they can rapidly accrete on wings and
other aircraft surfaces (Fig. 1.16a). The accretion, commonly known as aircraft icing, can result
in a substantial degradation to the aerodynamic lift on a control surface. The resulting loss of
maneuvering control has been responsible for several fatal aviation accidents. The accretion is
strongly influenced by particle inertia, turbulent dispersion and the surface roughness. To
understand this, ground-based experiments are conducted in icing wind tunnels whereby
droplets are injected upstream of a refrigerated wind tunnel test section (Fig. 1.16b).


Sediment and Bubble Transport in Bodies of Water

Sediment transport in natural bodies of water can play a significant role in the health and
maintenance of rivers and coastal areas. For example, silting can dramatically reduce the
capacity of rivers and has been cited as one of the reasons for some of the more devastating
floods in Asia and elsewhere. Coastal land erosion is another problem that has become more
severe in the last few decades since it can lead to substantial losses of animal habitats and
affect human population centers. Erosion can be driven by the lifting-up of particles (typically
by drag and lift forces) along with the subsequent convection and re-settling. However, these
aspects are complicated by the wave-induced and current-induced turbulent boundary layers.
Particles can also be introduced for environmental cleaning of waste streams by addition of
chemicals to form a precipitate of unwanted matter. To remove the precipitate, bubbles
(typically ranging from 30 to 300 microns) can be introduced to achieve air flotation since
they attach themselves to the suspended particles (Fig. 1.17a). The resulting bubble/particle
agglomerates rise to the top of the flotation tank where they are collected in a sludge blanket
that is later removed while the particle-free water passes below a baffle (Fig. 1.17b). The
sludge blanket is then removed to produce waste-free water.
Bubbles are also injected in plumes for reservoirs and wastewater treatments to increase
oxygen content and control pH levels. In these large-scale water bodies, the upward rise of the
21
bubbles entrains the surrounding liquid through complex turbulent mixing and interaction
between the phases. This flow mixture is complicated by presence of large bubbles which can
coalesce, deform due to inertia effects, and exhibit non-linear (zig-zag) trajectories.
Interestingly, bubble plumes has also been used by humpback whales and dolphins to catch
fish. For example, a whale (or several whales) will dive beneath a group of prey and then
slowly spiral upwards while blowing bubbles (Fig. 1.18). This creates a cylindrical wall of
bubbles, which traps the prey in a virtual net, allowing a whale to then swim and feed
through this region. Leighton (2004) suggested that the reason for the trapping is that the
bubbles create an acoustically-insulated environment.
Bubbles of specific sizes and concentrations are also released along the hull of military
ships for stealth. This process is called masker and is intended to modify or reduce the
underwater acoustic detectability of the vessel. The bubble trajectories are strongly affected by
the boundary layer along the ship and a downstream turbulent wake which affects their spatial
distribution. Their size distribution is also important because this determines the acoustical
insulation properties.


1.3. Basic Terminology and Assumptions

To describe multiphase flow regimes, a few essential definitions are needed. A "particle" is
defined herein as a relatively small unattached body immersed in a flow and can be solid
particle (such as dust) or a fluid particle (such as droplet or a bubble). Particles can be grouped
in natural and man-made categories as shown in Fig. 1.19 and can occur in a wide range of
sizes varying from less than a micron to greater than a centimeter. To quantify the particle size,
the volumetric diameter (d) is often used. Denoting an individual particle volume as
p
this
diameter is equal to that of a sphere which has the same net volume (Fig. 1.20a):

( )
1/3
p p
d 2r 6 / 1.1
The above equation also defines the particle volumetric radius (r
p
) as that for a spherical
particle of the same volume. . If the particle is a sphere, its surface area is simply:

2
p sphere p
A A 4 r = =
1.2
For a non-spherical particle,
p sphere
A A > and this difference can important for surface-averaged
phenomena such as mass and heat transfer. The particle mass can be used along with the
particle volume to define the particle density:

p p p
m /
1.3
In this text, we will assume the particle density is uniform so that the center of mass is equal to
the centroid (x
p
) for spherical particles. The particle density ratio can be defined as

*
p p f
/
1.4
We may refer to very-heavy particles as having
*
p
1 and very-buoyant particles as having
*
p
1. Examples of very-heavy particles include solid particles and droplets in a gas flow
22
(which fall in the direction of gravity). Examples of very-buoyant particles include bubbles
and hollow particles in a liquid (which rise opposite the direction of gravity). Examples of
particles with density ratios of the order of unity include drops in immiscible liquids. Herein,
immiscible indicates that there is no molecular mixing at the interface so that the drop
properties (mass, volume, etc.) remain distinct from those of the surrounding liquid.
The macroscopic length scale of the continuous-flow domain (such as a channel width or a
jet diameter) is defined as D. In this text, the particle size will be assumed to be small in
comparison, i.e. dD (Fig. 1.20b). Since each particle is assumed to be surrounded by a fluid
which otherwise fills the domain, the particles in general will be referred to as the dispersed-
phase and the surrounding fluid will be referred to as the continuous-phase. Most of the
analyses herein will assume a single dispersed-phase and a single continuous-phase, but they
are generally extendable to multiple phases.
Dispersed flow, the focus of this book, indicates that particle motion is generally dominated
by the interaction with the surrounding flow as opposed to the presence of nearly particles. At
low particle concentrations, there is one-way coupling whereby the dispersed-phase motion
is affected by the continuous-phase but not vice-versa. The dispersed-phase motion is typically
controlled by particle inertia, drag and gravitational forces. At higher concentrations, the
dispersed-phase can also affect the continuous-phase through by changing the momentum or
effective density and this is referred to as two-way coupling. At even higher concentrations,
three-way coupling may occur whereby particle-particle fluid dynamic interactions can occur
when particles in close proximity, but not touching. In particular, three-way coupling occurs
when the flow around a particle will influence the flow around a neighboring particle, e.g.,
though wakes. Finally, four-way coupling refers to particle-particle collisions when actual
contact is made. As noted in 1.1, dispersed-flow conditions arise when particle motion is
dominated by the continuous-phase flow interactions. On the other hand, dense flow occurs
when the particle motion is dominated by particle-particle interactions, i.e. three-way or four-
way coupling effects are stronger than one-way coupling effects. Details and criteria of these
different coupling regimes will be discussed in Chapter 2. In the remainder of this chapter, we
will consider a single isolated particle whose influence on the surrounding flow will be
negligible (consistent with one-way coupling).
The continuous-phase surrounding the particle will always be a fluid (gas or liquid) and it
will be associated with the subscript "f". It will generally be assumed to behave as a
continuum such that a dynamic fluid viscosity (
f
) and density (
f
) of this phase can be defined.
The subscript "p" will be used when referring to particle properties, e.g., the volume-averaged
particle density is
p
. For a fluid particle, the viscosity of the particle (
p
) becomes finite and
may play a role in the particle dynamics. However, if the particle is a solid, the viscosity of the
particle can be considered infinite. A fluid associated with either phase can also be assumed to
have either gas or liquid properties (denoted by subscripts g and l).
For fluid particles, the ratio of the viscosities can be important and is defined as

*
p p f
/
1.5
The viscosity ratio will tend to exhibit the same tendencies as that for the density ratio, i.e.,
liquid drops in a gas correspond to
*
p
1 and
*
p
1 while gas bubbles in a liquid correspond to
*
p
1 and
*
p
1. Similarly, drops in an immiscible liquid correspond to
*
p
~1 and
*
p
~1.
23
Density, viscosity and other fluid properties are listed in Table A.1 for two gasses and two
liquids at STP (standard temperature and pressure) conditions of 20
o
C and 1 atm.
The shape curvature of liquid particles is controlled by the interface stress. For a particle at
rest, this stress is based on the Young-Laplace equation which states that the difference
between the surrounding fluid pressure (P
f
) and the particle pressure (P
p
) is based on surface
tension () and the local interface curvature radii in the two orthogonal directions (r
I,1
and r
I,2
):

p f
I,1 I,2
1 1
P P
r r

= + +



1.6
The I subscript denotes and interface property and for conciseness, P
f
will be generally
expressed as P in this text. The pressure relationship for a sphere becomes:

p p
P P 2 / r = +
1.7
Note that surface tension is a property of an interface between two fluids and values for a
liquid (as in Table A.1) are typically referenced to air.
To determine the fluid dynamics forces, the various velocities of the multiphase flow
should be defined. A velocity or a force will be represented as a vector quantity when given in
bold-face (e.g., U) and as a tensor quantity when given with an index (e.g., U
i
). The scalar
magnitude of a vector will be represented by a regular type-face (e.g., U equals |U|). There are
two sets of velocities that one may employ: surface-based and centroid-based. The surface-
based velocities are shown in Fig. 1.21a and include the continuous-fluid velocity (U) is
defined only outside of the particle surface and takes into account the particle presence by
following streamlines around the body and by taking into account the particle wake. Similarly,
the internal particle velocity (V) is defined only within
p
and may account for rotation in the
case of a solid particle or recirculation in the case of a fluid particle. These two velocities are
thus defined as

velocity outside a particle including disturbance around the surface
velocity inside a particle including any rotation or recirculation

U
V

1.8a

1.8b
The centroid-based velocities are shown in Fig. 1.21b and include the particle translational
velocity (v) at the particle centroid (x
p
) and the unhindered continuous-fluid velocity which
neglects local particle disturbances so that it is exists in all portions of the domain, including
with the particle. The unhindered velocity at the particle center of mass represents the fluid
velocity which would occur if the particle was not present and is defined as u
@p
. This can be
used with the particle velocity to defined the particle relative velocity (w) and these various
velocities are thus defined with the unhindered continuous-phase pressure (p) as:

( ) ( )
( ) ( ) ( )
p p
@p p
@p
t 0
( t t) ( t)
translation velocity of particle centroid
t
continuous-phase velocity neglecting local particle disturbances
t , t
t t t
p continuous-phase pressure neglecting local

+

x x
v
u
u u x
w v u
particle disturbances


1.9a

1.9b

1.9c

1.9d

1.9e
24
Note that V=v for a solid non-rotating spherical particle, but is otherwise different. Note also
that U=u and P=p far way from the particle or if there is no particle, but are otherwise different.
To demonstrate the difference between u and U, consider a single-particle falling in a large
tank with quiescent conditions (stagnant fluid). The flow disturbances near the particle surface
result in U0 locally; however, since the fluid is stagnant in the absence of the particle, u=0
throughout. However, if there are other particles which as a collection cause the overall fluid to
move, then u
@p
(and u in general) can include changes due to the neighboring particles but not
due to the single particle at whose location it is defined. It is important to note that u
@p
is un-
physical (the continuous-phase does not exist inside the particle) and only will be used for
mathematical and computational convenience.
The unhindered continuous-phase vorticity () similarly neglects local disturbances caused
by an individual particle and is defined as twice the fluid angular rotation:

f
2
f
1.10
The fluid vorticity hypothetically extrapolated to x
p
is then
@p
which can be used with the
particle angular velocity about the particle center of mass (
p
of Fig. 1.21b) to define the
relative angular velocity as

p,rel
(t)
p
(t)

-
f@p
(t) =
p
-
f@p
1.11
Often, the subscript f will be implied for , while the subscript p will be implied for .
Forces, velocities, and particle positions can generally vary in space (x) and time (t).
Temporal variation can be considered in terms of various reference frames including a fixed
Eulerian reference frame, a particle-Lagrangian reference frame, and a fluid-Lagrangian
reference frame (Fig. 1.22). For example, the Eulerian derivative of an arbitrary parameter
q is based on temporal changes at a fixed point in the domain, x (though, the domain itself
may be moving at some velocity) and thus is given as



t 0
q q( t t) q( , t)
t t

+


x, x

1.12
In comparison, the particle-path and fluid-path Lagrangian time derivatives are defined along a
path specified by v and u respectively, i.e.




t 0
q q( t t t) q( , t) q
q
t t t

+ +
= +

x v , x
v
d
d

t 0
q q( t t t) q( , t) q
q
t t t

+ +
= +

x u , x
u
D
D

1.13a


1.13b
This assumes that the variable q can be considered as a continuum throughout the domain (this
may not be possible with certain properties) so that these two derivatives can be related for a
scalar field as:


q q
q
t t
= + w
d D
d D
1.14
For a vector field q, the Lagrangian derivatives are related in vector form and tensor form as:
25


( ) ( )
t t t

+ +

q q q
= v q = w q
d D
d D

i i i i i
j j
j j
q q q q q
v w
t t x t x

+ +

= =
d D
d D

1.15a

1.15b
In these equations, the LHS is the Lagrangian time derivative along the particle path while the
RHS is Lagrangian time derivative along the unhindered fluid path.
Vector quantities will be often represented in Cartesian coordinates for convenience. For
example, the three Cartesian components of the particle velocity and position vector are based
on the unit vectors in the x, y, and z directions as

v(t) v
x
(t)

i
x
+ v
y
(t)

i
y
+ v
z
(t)

i
z
x
p
(t) x
p
(t)

i
x
+ y
p
(t)

i
y
+ z
p
(t)

i
z
1.16a

1.16b
However, sometimes the velocities, positions and forces will be given in spherical coordinates
fixed to the particle to examine flow inside or in the immediate vicinity of a particle. For the
declaration of several other variables and symbols used in the text, the reader is further referred
to the nomenclature. However, new variables will generally be defined as they are introduced.


1.4. Description of Uncoupled Continuous-Phase Flow

In this section, we consider the condition where the continuous-flow is uncoupled, i.e.,
there is a negligible impact of the particles on the surrounding flow evolution given by u. This
neglects disturbances in the immediate vicinity of the particle surface as well as macroscopic
changes due to the collective action of the particles. This uncoupled condition is consistent
with a small number of particles in the domain, such that the continuous-phase behaves as in a
single-phase flow and whose conservation equations are discussed in Appendix A. This
condition also serves as a baseline to which we can add the influence of particles on the
surrounding flow, i.e., the coupled conservation equations to be discussed in Chapters 5, 7 &
8. Often, a continuum flow assumption is often made as discussed below.


Continuum Approximation

To establish the conventional derivative-based conservation equations for the continuous-
phase which will be introduced in the next sub-section, it is important that the surrounding
fluid can be considered as a continuum. The continuum approximation is reasonable if the
relevant length-scales for the flow-field (e.g., D) are much larger than the average inter-
molecular spacing of the continuous-phase (l
m-m
). In this case, a large number of molecular
interactions occur over the relevant length-scale, such that only the averaged effect is important
and the individual discrete (random) interactions are not significant. The breakdown of
continuum occurs quite differently in a liquid than in a gas. For liquids, the inter-molecular
spacing is on the order of the molecular length-scale (less than a nanometer), whereas gasses
typically have an inter-molecular spacing which may be on the order of a thousand times the
molecular diameter. Therefore, the continuum approximation is more likely to be violated with
26
gas flows than with liquid flows for a given particle size.
To determine the continuum criteria for a gas, the average mean-free path must be obtained.
This is defined as the average distance traveled by gas molecules between collisions with each
other. The mean free-path is inversely proportional to the product of the number of molecules
per unit volume and their effective cross-sectional area (which is based on the molecular
diameter). However, the molecular diameter is not a directly measurable quantity, so a more
practical means of assessing the mean-free path is to relate it to the average rms (root-mean-
square) speed of the molecules (v
m-m
) and the measured gas viscosity and density:


m-m
m-m
2
v

g
g
l
1.17
For a perfect gas, the average speed of the molecules is related to the gas temperature and gas
constant (defined in Appendix A.1), such that


8
m-m
v T

=
g g
R 1.18
and is therefore roughly one-third greater than the gas speed of sound (a
g
). Based on this, the
molecular mean-free path for air at STP is about 0.07 m and thus the continuum assumption
is generally reasonable except at very small scales. Since gasses in general have similar
dynamic viscosities which are only a moderate function of absolute temperature
(approximately square root dependence), substantial increases in the mean-free path primarily
occur at low gas densities such as those found in vacuum conditions (where the mean free path
is inversely proportional to pressure) or very high altitudes (e.g., the mean free path at 130 km
in the atmosphere is about one meter).
To characterize when effects of molecular collisions can be neglected and a continuum can
be assumed, it is helpful to parameterize the competing effects. This is accomplished with the
Knudsen number (Kn), which is defined as the ratio of the mean-free path to the relevant flow
length-scale. In multiphase flow, one can define both a macroscopic (continuous-phase)
Knudsen number and a particle Knudsen number for the continuous-phase flow (U) as follows


m-m
D
m-m
p
Kn
D
Kn
d

l
l

1.19a

1.19b
In general, very small values of the Knudsen number (e.g., less than 10
-2
) are consistent with
the continuum approximation and a no-slip boundary condition on a solid surface. This is
because a very large number of molecular collisions occur over the relevant length-scale, such
that the effects of their discrete interactions are negligible. Boundary conditions for small but
finite Knudsen number (about 10
-1
) as illustrated in Fig. 1.23 can be incorporated by assuming
a finite slip velocity at solid surfaces. For Knudsen numbers of order unity, the flow can not be
considered as a continuum with respect to the relevant length-scales and non-continuum
aspects should be incorporated directly. As the Knudsen number becomes much greater than
unity, a free-molecular flow regime is established, which is dominated by individual molecular
collisions.
Based on this definition, continuum approximation criteria for the macroscopic fluid
dynamics and the local fluid dynamics over the particle can be defined as
Kn
D
0 macroscopic flow acts as a continuum 1.20a
27



Kn
p
0 flow around the particle acts as a continuum 1.20b
For air at STP, the Kn
p
criterion suggests that particles which are on the order of 5-10 microns
or larger can be reasonably considered to obey the continuum approximation. While this
assumption will generally be used in this text, the non-continuum effects for micron-sized or
sub-micron particles will be discussed in Chapters 6 and 8. The assumption of a continuum for
the macroscopic continuous-phase flow will be used throughout the text.


General transport equations

Assuming a continuum fluid with respect to the uncoupled flow, the transport equations
can be written by assuming no mass, momentum, or energy transfer between the fluid system
and the particles or between the fluid system and the surroundings. The general (three-
dimensional, unsteady, compressible, viscous) conservation equations of mass, momentum,
and energy are given by Eqs. A.5, A.6 and A.13. If the flow is compressible, these partial
differential equations (PDEs) are supported by an equation of state to relate pressure changes
to density changes discussed in Section A.1. The Mach number is often used to assess the
flow compressibility, whereby higher Mach numbers indicate larger fluid dynamic changes in
pressure and density. Various inviscid formulations are discussed in A.2, while viscous
formulations are discussed in A.3 and A.4 in terms of the three primary viscous regimes:
laminar, transitional, and turbulent flow.



1.5. Equation of Motion for an Isolated Spherical Particle

This section will discuss general and specific equations of motion for an isolated, spherical
particle subjected to drag and gravitational forces and simple flows. The resulting basic
equations introduced in this section will be useful to illustrate the most common aspects of
multiphase fluid dynamics discussed in Chapter 2-4.

1.5.1. General Particle Equation of Motion

The overall particle translational equation of motion simply specifies that the rate of change
of the particles linear momentum is equal to the net sum of the forces acting on the particle.
This yields a general Lagrangian equation of momentum given by
( )
p body surf coll
m t = + + v / F F F d d
1.21
The right hand side (RHS) includes the forces associated with these temporal changes. Body
forces (F
body
) are those proportional to the particle mass, surface forces (F
surf
) are those
proportional to the particle surface area and related to the surrounding fluid stress, and collision
forces (F
coll
) includes the effects of other particles or walls which may come in contact with
28
the particle. These forces control the particle velocity, which is needed for the companion
Lagrangian equation of particle position

p
t

x
v vv

d
d
1.22
Integration of this equation yields the particle trajectory as a function of time.
In the following, the RHS of Eq. 1.21 will be constructed in terms of particle and
continuous-phase characteristics under specific assumptions. In this text, the body forces are
assumed to be represented solely by the gravitational force (F
g
) which acts in the direction of
the gravity acceleration vector (g), such that
F
body
F
g
= m
p
g 1.23
This assumes that other body forces (such as electromagnetic forces) are negligible. In this
chapter, collision forces will be neglected since we have assumed an isolated particle.
The surface force for a spherical particle can be written in terms of the pressure and viscous
stresses (Eq. A.60) acting over a differential surface area (Eq. 1.2):


( )
surf r rr r r r p
P K K K dA

= + + +

F i i i i
1.24
In this equation, i
r
is the unit normal in the outward radial direction while i

and i

are in the
polar () and azimuthal () directions. Note that this results is general in that no assumptions
are made of the relative velocity, vorticity or temperature and thus requires an exact solution of
the flow to implement. Another approach to describe the surface forces is to decompose it into
a linear sum of various fluid dynamic forces related to certain flow properties:


F
surf
= F
D
+ F
L
+ F

+ F
H
+ F
S
+ F
Br
+ F
T

1.25
These individual components include forces due to:
drag (F
D
) which resists the relative velocity,
lift (F
L
) which arise due to particle spin or fluid shear,
virtual-mass (F

) which is related to the surrounding fluid that accelerates with the particle,
history (F
H
) which takes into account unsteady stress over the particle,
fluid-stress (F
S
) which stems from the fluid dynamic stresses in the absence of the particle,
Brownian motion (F
Br
) random motion from discrete molecular interactions, and
thermophoresis (F
T
) force due to molecular interactions along a temperature gradient.
Note also that we have not included two effects associated with continuous-phase flow
rotation: centrifugal and Coriolis forces. The centrifugal force is associated with particles
being thrown out of rotating flow regions. It is not a surface force but instead represents the
inertia associated with a particle which tends to translational motion and thus resists rotational
motion. Similarly, the Coriolis force is not a surface force but is instead a relative acceleration
to a continuous-phase system which is uniformly rotating. As such, both effects are already
implicitly incorporated into Eq. 1.21 so long as a Cartesian coordinate system is employed, but
would need to be added if one chooses a rotating coordinate system.
The above decomposition is the one of several assumptions that will employed in this
chapter to obtain a closed-form solution for the surface force. Others include negligible
compressibility effects for the surrounding fluid flow as well as continuum conditions
(Kn
p
0) so that the above-mentioned thermophoresis and Brownian forces are negligible.
29
Next, we assume that the surrounding flow-field velocity gradients and particle spin are
negligible so that particle lift need not be considered (F
L
=0). Finally, the particle is assumed
to be spherical with constant diameter and density. Using the above assumptions, a simplified
surface force expression can be obtained as discussed in the following two sections.
While the surface forces discussed in this Chapter are limited to the above conditions,
Chapter 6 will consider more general conditions so that these assumptions can be relaxed. In
particular, it will address issues associated with various shapes and particle surface conditions,
flow compressibility, non-continuum effects, lift forces, particle-particle interaction, particle-
wall interactions, etc.

1.5.2. Quasi-Steady Drag Force

The quasi-steady drag force arises from pressure and viscous stresses applied to the particle
surface. The resulting force resists the relative velocity so that it is defined to act in the
direction opposite of the particle relative velocity (w). In this section, the force will be
considered for steady continuum conditions. Furthermore, the particle and the far-field
velocities are both assumed to be steady, i.e. w is constant, in order to obtain the quasi-steady
drag force. If there are no spatial gradients in the unhindered continuous-phase flow, i.e.
u=0, then the u
@p
is the same as that in the far-field, i.e. w=v-u

=v-U

.
The dependence of the drag force on the magnitude of the relative velocity is primarily
dictated by the particle Reynolds number (Re
p
), defined as

f
p
f
Re

w d

1.26
The particle Reynolds number is the non-dimensional ratio of fluid inertial forces to viscous
forces with respect to the fluid dynamics in the vicinity of the particle. The Reynolds number
can be used to categorize the flowfield regime seen by the particle (Fig. 1.24). The fluid
dynamics around the particle, in turn, determine the quantitative relationship between the
relative velocity and the drag force exerted on the particle. In the following, the particle Mach
number (M
p
=w/a) is assumed to be small so that the flow can be considered incompressible
(see A.2).


Creeping Flow: Stokes Solution

When the inertial effects are negligible, the viscous and pressure fields dominate the flow-
field solution resulting in a fully attached laminar flow over the particle (e.g., Fig. 1.24a). The
dominance of viscosity also means that the shear stresses are felt far from the particle surface.
As a result, this low Reynolds number condition is often called creeping flow. For this
condition, Stokes (1851) derived the drag force in this limit of negligible convection and by
assuming incompressible continuum flow of constant density and viscosity around the particle.
Simplifying the incompressible Navier-Stokes momentum equation based on these
assumptions and additionally neglecting gravitational and unsteady terms, yields a balance
30
between the pressure gradient and the viscous stresses (Eq. A.63). Considering this equation
for the case of flow around a stationary particle (v=0) yields the Stokes equation:


2
f
P = U for Re
p
0 1.27
The neglect of gravity eliminates the hydrostatic pressure gradient associated with buoyancy
(Eq. A.56), but this effect will be re-introduced in the next section. For a particle, application
of this PDE and the proper surface boundary conditions can describe the local velocity field of
the continuous-phase fluid around the particle. As such, the resulting velocity field is affected
by the particle displacement and thus is different from the unhindered velocity near the particle
surface (i.e. Uu).
The PDE of Eq. 1.27 can be expressed in terms of axisymmetric flow in spherical
coordinates (see A.3) with a control volume is consistent with the particle center as shown in
Fig. 1.25, and moves at the speed of the particle. In this case, the radial and tangential velocity
components based on polar coordinates (U
r
and U

) are defined throughout the surrounding


flow (rr
p
) and the swirl velocity is assumed to be negligible, i.e. no azimuthal velocity (U

=0).
From this, Eq. 1.27 can be decomposed into radial and tangential momentum equations as
discussed in A.3:


2 2
r r r r r
2 2 2 2 2 2 2
f
U 2U cot U U 2U U U 1 P 2 1 cot 2
r r r r r r r r r


= + + +


2 2
r
2 2 2 2 2 2 2
f
U U U U U U 1 P 2 1 cot 2
r r r r r sin r r r


= + + + +


1.28a

1.28b
By introducing the Stokes spherical stream function which satisfies continuity, the radial and
tangential velocity components of the continuous-phase fluid can be expressed as


U
r
=
2
1
r sin

= -
1
r sin r



1.29a


1.29b
For convenience, we can introduce a specialized spherical differential operator
2
defined as


2
2
2 2
q sin 1 q
q
r r sin

= +



1.30
This special operator is similar to, but different than, the true Laplacian operator of Eq. A.59.
Substituting these velocity components of Eq. 1.29 into Eq. 1.28 and applying this special
operator notation, yields the stream-function based momentum equations


( )
2 f
2
P
r r sin

=


( )
2 f
1 P
r r sin r

=


1.31a


1.31b
The pressure term can be eliminated by cross-differentiating Eq. 1.31a with respect to and Eq.
1.31b with respect to r to yield a single fourth-order PDE:
31


2 2
( ) 0 = 1.32
Boundary conditions are needed to allow solution of this differential equation.
The boundary conditions for a solid non-rotating sphere are given by a no-slip condition on
the particle surface, i.e. U
r
=U

=0 at r=r
p
. Based on Eq. 1.29, these boundary conditions can be
written in terms of the stream function


0
r

= =

as r = r
p
1.33
Far from the surface, the boundary conditions correspond to uniform flow in the horizontal
direction (x-direction) so that the radial and tangential far-field components can be written as:


r
U U cos

= as r
U U sin

= as r
1.34a
1.34b
Note that this coordinate system translates the continuous-phase velocities to ones relative to
the velocity of the particle centroid. Based on Eq. 1.29, this yields a boundary condition on the
stream function given by


2 2
1
2
U r sin const.

= + as r 1.35
The constant in the above equation is arbitrary and so can be set to zero. The solution may be
obtained by assuming separation of variables so that /sin
2
() is a function of radius that
satisfies the governing equation (Eq. 1.32) and the above boundary conditions (Eq. 1.35). In
this manner, the stream function can be obtained as (White, 1991):


3
2
p p 2
3
3r r
U r
1 sin
2 2r 2r


= +




1.36
The first term of the RHS expression corresponds to the far-field condition (Eq. 1.35), the
second term is called the Stokeslet (a viscous correction), and the third term is called the
doublet (which is an inviscid correction). The viscous correction decays much more slowly
with increasing distance from the particle surface than the inviscid correction because of the
low Reynolds number of the flow. From Eq. 1.29, the radial and tangential velocities can be
obtained (with free-stream, doublet and Stokeslet contributions) as


3
p p
r 3
3
p p
3
3r r
U U cos 1
2r 2r
3r r
U U sin 1
4r 4r



= +




=




1.37a


1.37b
The resulting streamlines are symmetric both front-to-back and top-to-bottom, which is
consistent with the flow seen in Fig 1.24a and the condition of reversibility associated with Eq.
A.63. This indicates that the particle influence is felt as far upstream as it is felt downstream.
To obtain the pressure field, the pressure gradients can first be expressed in terms of radial
and tangential pressure gradients
32


P P
dP dr d
r

= +

1.38
The derivatives of the pressure on the RHS of this equation are given by Eq. 1.29 in terms of
the spherical differential operator which, in turn, can be obtained from Eqs. 1.30 and 1.36 as


2
p 2
3U r
sin
2 r

= 1.39
The pressure can then be obtained by substituting the RHS expression into Eqs. 1.31 and 1.38
and integrating to yield


p f
2
3U r
P P cos
2r

= 1.40
This can be evaluated on the sphere surface by setting r=r
p
. The result indicates that the
pressure is symmetric top-to-bottom but is anti-symmetric in the streamwise direction. Since
w<0 (Fig. 1.25), the pressure is higher on the front of the particle (=0) than on the back (=),
giving rise to a drag in the positive x-direction.
The surface force can be considered in terms of the pressure and viscous stresses from Eq.
1.24. The pressure component of this force can be written as


pressure r p
P dA =

F i
1.41
Based on symmetry of the flow about x-axis, the net drag will be in the x-direction so that this
integral can be solved by taking the streamwise component of pressure stress (Pcos) over the
upper surface (=0 to ) and introducing a factor of 2 for the lower surface. The differential
area can be based on a spherical element (dA
p
=
2
p
r sind) to obtain:


2
pressure p x p f
0
2 r Pcos sin d 2 r

= =

F i U 1.42
This is known as the pressure drag or form drag.
To determine the shear component of drag, the viscous stresses of Eq. A.60 must be
considered. For a particle with no swirl velocity, the terms associated with U

and gradients in
can be neglected. Furthermore, the boundary conditions of Eq. 1.34, conditions of symmetry
and the assumption of a constant particle radius eliminate result in only one non-zero viscous
stress component on the surface:


r
r f
U U 1
K r
r r r


+





1.43
Given the fluid velocity from Eq. 1.37 and this term, the shear force contribution of Eq. 1.24
can be integrated on the particle surface to yield:


F
shear
= 4 r
p

f
U

1.44
This is called the friction drag and turns out to be twice the pressure drag. The total drag
force is simply the sum of the form and friction drags. We can generalize the above analysis to
the case of a particle moving at a steady speed relative to the flow so that the approaching
33
velocity for Fig. 1.25 is U

-v. If the unhindered flow is uniform, u


@p
=U

then the approach


velocity is -w (Eq. 1.9d) so that the total drag in the particle reference frame is:

D pressure shear p f f
6 r 3 d = + = = F F F w w
1.45
Therefore, the creeping flow conditions yields a drag force linearly proportional to relative
velocity. This is often referred to as the Stokes drag, owing to his derivation and is in good
agreement with the experimental results for small but finite Reynolds numbers (e.g., within 2%
accuracy for Re
p
<0.1).
To non-dimensionalize the drag force, one may define a drag coefficient (C
D
) based on the
total drag force normalized by projected cross-sectional area and the far-field dynamic
pressure (
f
w
2
based on Eq. A.41). The definition is shown below along with value for
creeping flow based on Eqs. 1.26 and 1.45:

( )
D D
D 1 2 2 1 2
f
8 proj f
2
f
D p 1 2 2
p f
8
F F
C
d w
A w
3 d w 24
C Re 0
Re d w
=


= =

for


1.46a

1.46b
Therefore, the creeping flow drag coefficient is inversely proportional to Reynolds number.
While some theoretical results can be employed to extend the Stokes theory to small but finite
Re
p
values (as will be discussed in 6.2.1), unfortunately there is no closed-form analytical
solution for the viscous flowfield once the particle Reynolds number is greater than unity.


High Reynolds Number Flow: Newton-Based Drag

At the other extreme of very high Reynolds numbers (Re
p
1), some theoretical analysis is
also possible if we assume the viscous effects are confined to a small regions near the surface
so that much of the external flow can be analyzed with inviscid potential flow theory. For
incompressible flow, the equations of motion for mass and momentum in spherical coordinates
are given by Eqs. A.57 and A.58. Considering only the flow outside of the boundary layer, an
inviscid boundary condition can be approximately applied at the particle surface, i.e. u
r
=0,
while the far-field boundary conditions are the same as Eq. 1.34. The stream function solution
turns out to be a linear combination of a uniform stream and a point-doublet (White, 1991),
which can be given in terms of radial and tangential velocity components by Eq. 1.29. If we
consider the flow solution in terms of the relative velocity of a moving particle, the result is:


( )
( )
( )
2 2
p p 1 1 2 2 2 2
p p 2 2 2 2
p p
3 3
p p
r 3 3
3 3
p p
3 3
r r
r r
U v r sin wr sin
r r r r
r r
U U v cos 1 wcos 1
r r
r r
U U v sin 1 wsin 1
2r 2r



= =




= =




= + = +




1.47a


1.47b


1.47c
For the velocities, the first term in the parentheses is the far-field solution while the second
term is the inviscid doublet. In contrast to the viscous solution, the tangential surface velocity
34
is finite and given by Eq. 1.47c as
3
2
U wsin

= . Using this velocity and the Bernoulli


relation (Eq. A.41), the inviscid surface pressure over a sphere is given as:


2 2
f
9
1
P P w 1 sin
2
4


= +



1.48
The pressure is thus symmetric about the fore and aft surface, with the stagnation pressure
recovered at both =0 and . Such a pressure distribution can be expressed by defining a
pressure coefficient which is the difference between the static and the far-field pressures
normalized by the far-field relative dynamic pressure:


P 2
1
2 f
P P
C
w


1.49
The potential flow C
P
based on Eq. 1.48 can be compared to that for the Stokes surface
pressure distribution based on Eq. 1.40:


( )
2
P
C 1 9 / 4 sin = for potential flow
P p
C 6cos / Re = for creeping flow
1.50a
1.50b
The C
P
for potential flow thus tends to unity at the stagnation points and a minimum at =/2,
while the creeping flow solution yields a much higher C
P
at the stagnation points since the
pressure rise is based on viscous effects and not a dynamic pressure recovery.
Integrating this symmetric inviscid pressure distribution over the particle surface leads to
the solution F
D
=0. This is known as dAlemberts paradox, i.e., that the net force acting on a
body moving though an inviscid fluid is zero. This is, of course, inconsistent with the drag
commonly observed on bodies at high Reynolds numbers. While much of the flow around the
particle may be consistent with the above theoretical analysis, The reason for this difference
lies in the viscous effects which are dominant in the thin region near the particle surface. This
region is defined as the boundary layer and is important because it allows for non-zero
viscous stress to act on the body owing to the no-slip condition. In addition, the boundary
layer determines if and where the flow will separate from the particle surface. At high particle
Reynolds numbers, the boundary layer always separates and yields a turbulent wake. Once the
flow separates, the concept of a thin boundary layer for the downstream portion is lost and the
outer potential flow solution is also unreasonable. Prior to separation, the attached boundary
layer can generally have two states: laminar or turbulent.
The condition where the attached boundary layer is thin and laminar (and the wake is fully
turbulent) is called the sub-critical regime (Fig. 1.24c). For a solid sphere in uniform flow,
this regime has an approximate Reynolds number range given by 2,000<Re
p
<300,000 and
includes laminar separation point just before the mid-plane (~80
o
). In the upstream attached
region, the pressure distribution similar to the inviscid solution given by Eq. 1.48 as shown in
Fig. 1.26. However, in the downstream separated region, the pressure is nearly constant (since
there is little dynamic pressure in this region) and is controlled by the value near separation.
Since the downstream pressure is lower than that of the front stagnation region, this yields a net
pressure drag opposing the relative velocity. The portion of drag associated with friction is
related to the shear stress acting on the surface. Before the separation point, the velocity field
just above the boundary layer closely follows the inviscid solution (Eq. 1.47). The no-slip
condition within the attached boundary layer results in a shear stress component of drag
opposing the relative velocity. After the separation point, the wall shear stress is nearly
negligible (due to lower velocities and lower gradients in the detached flow zone) so that there
35
is little viscous drag in this region. As such, the upstream portion of viscous drag dominates.
Thus, friction and form drag are both important in this Reynolds number range and are
controlled by the separation mechanisms, which is related to viscous effects close to the
particle surface. This indicates that drag can be substantial even at high Reynolds numbers
which explains dAlemberts paradox.
Since both the pressure and shear stresses are related to the local dynamic pressure and
since the separation point is approximately constant for the sub-critical regime, it can be
expected that the resulting drag force will be simply proportional to the free-stream dynamic
pressure and the particle frontal area. Based on Eq. 1.46, this suggests that the drag coefficient
be approximately constant in the regime, which is consistent with measurements in this Re
p

range (Fig. 1.27). The occurrence of a nearly constant drag coefficient for large particles was
first recognized by Sir Isaac Newton when he analyzed drag resistance for projectiles in 1710.
As such, this value has been called the Newton drag coefficient and the critical drag
coefficient. The minimum and maximum Reynolds numbers for which the drag is
approximately constant can be defined as Re
p,sub
and Re
p,crit
. Thus, the critical drag coefficient
may be defined for the general case and the solid sphere case as:

D,crit
C const. for Re
p,sub
<Re
p
<Re
p,crit

D,crit
C 0.40 0.45 for 2,000<Re
p
<300,000 for a solid sphere

1.51a
1.51b
As will be discussed in 6.2, non-spherical particles may have Newton regimes but with
different critical drag coefficient and with different Reynolds number ranges because of
changes in boundary layer transition and separation physics.
For a solid sphere at Re
p
>300,000, the attached boundary layer becomes turbulent and
causes the flow separation point to be shifted to a further downstream position (~120
o
). The
reduced separation region associated with a wire tripping is shown in Fig. 1.24d, which is
significantly smaller than that associated with the sub-critical flowfield shown in Fig. 1.24c.
This allows the pressure to follow the potential flow solution over a larger portion of the
surface and thus substantially increases the pressure in the aft region as compared to the sub-
critical case (Fig. 1.26). As a result, the drag coefficient in this super-critical regime drops
dramatically, which creates a phenomenon termed the drag crisis (Fig. 1.27). This reduction
in drag was first observed by G. Eiffel, who was both an avid fluid dynamicist and the famous
French architect. In the super-critical regime, the drag again becomes sensitive to the
Reynolds number since the turbulent boundary layer separation point varies with Re
p
. The
maximum Reynolds number for which the drag is approximately constant is termed the
critical Reynolds number, Re
p,crit
. For a solid sphere, this is generally defined when the drag
coefficient first drops below 0.3 and is about 300,000.
Because most multiphase flows have particle with Re
p
<Re
p,crit
, the super-critical region is
usually not applicable unless the particles are very large and moving at very high speeds. For
example, a 75 mm (3) diameter baseball moving at 36 m/s (80 mph) in air at STP has an Re
p

of about 1.8x10
5
consistent with sub-critical conditions for a smooth sphere. However, the
stitching on the surface prematurely trips the attached boundary layer to become turbulent so
that the drag crisis occurs with a reduced resistance and thus faster speeds. A similar drag
reduction occurs with the addition of dimples on a golf ball, thus allowing longer trajectories.


Drag at Intermediate Reynolds numbers
36

The intermediate Reynolds number is that between creeping flow conditions (with fully
attached flow) and that below the sub-critical Newton-regime (with fully turbulent wakes and
nearly constant drag coefficient), i.e. 0.1<Re
p
<2,000. Based on flow visualizations such as Fig.
1.24, the intermediate Reynolds number flowfield in the rear of the particle undergoes several
transitions: from an attached laminar wake for Re
p
<20 (Fig. 1.24a), to a separated but still
laminar wake (Fig. 1.24b), to an unsteady transitional wake. As the Reynolds number
increases from 25 to 100, the point of separation point moves upstream from the rear
stagnation point with 180
o
at Re
p
=25 to 108
o
at Re
p
=100, and modest changes for higher
Reynolds numbers. In all these cases, the front portion of the particle surface is characterized
by an attached laminar boundary layer. As seen in Fig. 1.26, the C
P
upstream distribution also
moves from the creeping flow distribution towards the potential flow distribution (given by Eq.
1.50) as the Reynolds number increases. The downstream C
P
distribution contains regions of
nearly constant value associated with flow separation. Larger regions of flow separation are
associated with decreased recovery at the rear stagnation point, and this commensurately lower
pressure at the rear of the particle is associated with an increase in drag.
Based on measurements of the drag coefficient for intermediate Reynolds numbers (Fig.
1.27), the drag coefficient transitions from the creeping flow value given by Eq. 1.46 to the
nearly constant critical drag coefficient of Eq. 1.51b. While there is no general analytical
solution for intermediate Reynolds numbers, empirical drag coefficient expressions can be
used to describe the overall variation, e.g., the expression by White (1991):
D
p p
24 6
C 0.4
Re 1 Re
= + +
+
for Re
p
<2x10
5

1.52
The first term on the RHS is the Stokes limit, the last term on the RHS is the Newton limit, and
the intermediate term is simply an empirical correction. The agreement with experimental data
for this expression is good for a large range of conditions as shown in Fig. 1.27. Also shown
in this figure is the Putnam (1961) fit, which is another empirical expression which is suitable
for low and intermediate values of Reynolds numbers:
D 1/3
p p
24 4
C
Re Re
= + for Re
p
<1000
1.53
A similar fit (which would be difficult to distinguish from the Putnam fit in Fig. 1.27) for small
to moderate Reynolds numbers is the Schiller-Naumann (1933) expression, which is


( )
0.687
D p
p
24
C 1 0.15Re
Re
= + for Re
p
< 1000
1.54
This is the most commonly used drag coefficient in multiphase flows since it is quite accurate
and since many particles are constrained to Re
p
values in this range. A more detailed
comparison of these drag coefficient expressions is given in 6.2. Unlike the analytical
creeping flow drag, such empirical drag coefficients are non-unique and only valid for
conditions for which they have been calibrated. There are, unfortunately, many other empirical
expressions which can arise in computational multiphase fluid dynamics.
One may quantify the departure of the drag from the Stokes solution by normalizing the
drag force by the creeping flow solution for a sphere in an incompressible uniform continuum,
37

D
f
F
f
3 d w



1.55
This ratio is called the Stokes correction factor so that drag force can be written as

1 2
D f f D
8
3 d f d C = F w = - w w
1.56
If only Reynolds number effects are considered (i.e. a solid sphere in incompressible flow), the
Stokes correction factor correction factor can be denoted as f
Re
, which can be related to the
drag coefficient and expressed using Whites fit as:

D p p p
D
Re
D p p p
F (Re ) Re /4 Re
C
f 1
F (Re 0) (24/ Re ) 60 1 Re
= = + +
+

1.57
To account for conditions that may include continuous-fluid compressibility, non-spherical
particle shapes, etc., a set of detailed models for C
D
and f will be discussed in 6.2. However,
White (1991) notes that the drag coefficients for non-spherical particles tend to be in the ball
park of those for a sphere with the same volume. This is also generally true with respect to
differences between solid and fluid particles (discussed next) so that the expressions developed
in this section are generally reasonable first-order approximations.


Drag for Spherical Fluid Particles

Compared to the solid sphere discussed above, a fluid sphere (drop or bubble) can
allows internal recirculation. Because of this, the flow past the particle need not come to rest
on the surface as there will be a finite amount of slip (U

0 at r=r
p
) which will yield a reduced
drag. This condition for Re
p
0 permits an analytical solution if the interface between the
phases can be considered clean (i.e. fully mobile) and that the surface tension is uniform over
the particle (reasonable if the effect of contaminants is negligible). The free-stream boundary
conditions of Eq. 1.34 for U and the stream function PDE of Eq. 1.32 for

for a rigid sphere
in creeping flow still apply. However, the mobile interface leads to modified surface boundary
conditions. In particular, a spherical fluid particle leads to the following tangential and normal
stress balances at the interface (denoted by subscript I) based on K
r
and K
rr
of Eq. A.60:


r r
p f p
I
I p
r r
p f
p I I
U V 2
P P 2 2
r r r
V U V U 1 1 1
r r
r r r r r r r


+ =




+ + =




1.58a


1.58b
The first equation (Eq. 1.58a) relates the interface pressure jump to the deviatoric (normal)
viscous stresses and the surface tension. This simplifies to the spherical version of Eq. 1.7 in
quiescent conditions. The second equation (Eq. 1.58b) relates the tangential shear stress to the
gradients in surface tension and replaces the noslip boundary condition used for solid particles.
If one assumes a constant surface tension (reasonable for uncontaminated conditions with no
temperature gradients), negligible particle viscosity (reasonable for bubbles) and negligible
radial velocity (appropriate for constant diameter), then Eq. 1.58b simplifies to:
38


I
U
0
r r



1.59
This equation is often termed the stress-free condition since it is consistent with zero
tangential stress of the liquid on the bubble surface.
If one defines an internal stream function
p
based on the internal particle velocity V,
the boundary conditions at r=r
p
for steady flow can be expressed as:


p
p
p
f p 2 2
p
f p p 2 2
p
r r
1 1
r r r r r r
1 1
P 2 2 P 2
r r sin r r r sin
=

=


=





+ =





1.60a

1.60b

1.60c


1.60d
The first equation (Eq. 1.60a) indicates that the interface bounds the two velocity domains and
that the surface of the sphere is a streamline (no mass transfer across the particle surface). The
second equation (Eq. 1.60b) indicates that the tangential velocity is continuous across the
interface. The third equation (Eq. 1.60c) stems from Eq. 1.58b if one assumes constant
surface tension (consistent with isothermal conditions) and that U
r
=V
r
=0 at the interface
(consistent with a constant particle diameter). The fourth equation (Eq. 1.60d) stem from the
normal stress balances of Eq. 1.58a.
The analytical solution for the continuous-phase and particle-phase streamlines for the
above boundary conditions is generally called the Hadamard-Rybczynski solution as it was
independently derived by both in 1911. Details of the stream function derivation are given by
Batchelor (1967) and with the solution:


* 3 *
2
p p p p 2
* 3 *
p p
3r (2 3 ) 3r
wr
1 sin
2 2r(3 3 ) 2r (3 3 )
+
= +

+ +



1.61
The RHS expression reverts to the solid sphere expression of Eq. 1.36 as
*
p
. The
theoretical internal motion associated with
p
is called the Hills vortex and compares
reasonably well with experimental streamlines as shown in Fig. 1.28. The external pressure
distribution just outside of the particle can be obtained using the relationships of Eqs. 1.31 and
1.61 to yield:


*
p f p
2 *
p
3wr (2 3 )
P P cos
2r (3 3 )

+
=
+

1.62
The internal pressure within the fluid particle (P
p
) is associated with a discontinuity at the
interface due to the surface tension and shear stress by Eq. 1.60d and is thus given by:
39


p p
p 2 *
p p
5wr
2
P P cos
r 2r (1 )

= +
+

1.63
In the limit of no relative velocity, the internal pressure is simply specified by the surface
tension jump, i.e. Eq. 1.7. Thus, a spherical shape is reasonable for creeping flow conditions
consistent with the analysis of Saito (1913).
The external pressure (of Eq. 1.62) and shear stress (both the tangential and the
deviatoric normal components) can be integrated over the particle surface to determine the drag
in steady flow conditions yielding

F
D
= - 6 r
p

f
w


*
p
*
p
2 3
3 3
+
+

1.64
If we define the Stokes correction for internal circulation (f
*
) as the ratio of the drag force
acting on fully mobile fluid sphere to the drag force on a solid sphere at Re
p
0, we have

* *
D p p p
* * *
D p p p
F ( , Re 0) 2 3
f
F ( , Re 0) 3 3

+
=
+

1.65
Therefore, the drag of a fluid sphere at this condition with
*
p
0 (e.g., a gas bubble in a
liquid) is 2/3 of that for a solid sphere. However, contaminants in the system often cause the
interface to become immobile (or partially immobile) such that this drag reduction is not
realized (or only partially realized). Furthermore, spherical liquid droplets falling in a gas will
tend to be characterized by
*
p
so that the correction of Eq. 1.65 will be negligible. The
impact of contamination, deformation and impact of finite density and viscosity ratios on drag
at finite Reynolds numbers will be discussed in 6.2.5 and 6.2.6. Furthermore, low density
ratio fluid particles tend to be especially influenced by the acceleration-based forces in addition
to the velocity-based drag force.

1.5.3. Surface Forces due to Accelerations

Assuming again a solid spherical particle in a continuum incompressible flow, the surface
forces associated with accelerations of the far-field continuous-phase velocity and the particle
velocity are considered in this section. This will give rise to three types of acceleration forces:
the fluid-stress force, the added mass force, and the history force.
The first acceleration force is the fluid-stress force (F
S
) which arises due to the continuous-
phase pressure and shear stresses that will exist in the absence of the particle. For example,
consider the deformable fluid element in the steady diverging nozzle flow of Fig. 1.29. As the
element moves downstream, it will decelerate according to the area increase. This deceleration
along the fluid element Lagrangian path (x
f
) is simply controlled by the fluid stresses (pressure
and shear) integrated over the fluid element surface. The surface-area integration (over A
f
) can
be converted to a volume integration (over
f
) by applying the divergence theorem. For
example, the pressure stress which acts opposite the outward unit normal vector n of the fluid
element surface yields a fluid-stress force based on pressure divergence within the volume:
40
( ) ( )
f f f
p dA p d p = =

n
1.66
The RHS expression assumes that the pressure gradients are approximately uniform in the
vicinity
f
(reasonable for small length-scales compared to changes in the fluid properties). A
similar conversion can be applied to the viscous stresses if the differential volume is assumed
small, so that the net fluid stress over the volume of a fluid element is given by:
( )
S,f f ij
= p K + F
1.67
Therefore, the unhindered stresses (which neglect the influence of the particle) give rise to a
fluid-stress force for either a fluid element or a solid spherical particle which inhabits the same
space (Maxey and Riley, 1983). Within the fluid volume, one may also relate the pressure and
viscous stresses by the single-phase momentum equation (Eq. A.6b):
( )
f f f f
p K
t

+ =


u
g
D
D

1.68
These unhindered fluid-stress forces (which neglect the influence of the particle within
f
) will
act on a fluid element, e.g., to decelerate it along its path as shown in Fig. 1.29.
Now consider a spherical particle which occupies this volume (
p
=
f
). As will be
discussed in 6.5 and by Maxey (1993), a particle is subjected to the same fluid-stress forces
per unit volume that a fluid element would experience from LHS of Eq. 1.68. Thus, a particle
which occupies a volume
p
will have a fluid-stress force given by:
( )
@p
S p ij f p
p K =
t

= +


u
F g

D
D

1.69
This shows that the fluid-stress force can be considered in terms of the fluid and gravitational
accelerations and is proportional to the displaced continuous-phase mass

f @p f p
m
1.70
If the continuous-phase fluid acceleration is negligible, then the fluid-stress force reduces to
the buoyancy force

S buoy f p
- F F g

for steady flow
1.71
Thus the buoyancy force is equal to the hydro-static pressure gradient of Eq. A.56 integrated
over the particle volume and acts in the opposite direction of the gravity. Since the fluid stress
is proportional to the continuous-phase density, it will tend to be more important for particles
with low density ratios, i.e.
*
p
1, e.g. it is the cause of bubble rise.
The second acceleration force is the added-mass or virtual mass force (F

) which arises
when the particle acceleration is different from that of the continuous-phase field, i.e. when
there is a non-zero relative acceleration
( )
t 0 w / d d . This relative acceleration gives rise to an
additional force because a portion of the surrounding fluid (in the vicinity of the particle) is
carried along with the particle with the same acceleration. This is due to the particle boundary
conditions which prevents mass flux through the surface. The surrounding fluid mass which
moves the particle is referred to as the added-mass or the virtual mass (m

). It is
schematically described in Fig. 1.30 as an integrated effect that is strongest near the particle
surface and weakens further away. For a solid sphere, the added-mass can be related to the
displaced fluid mass using by defining c

as the virtual mass coefficient:
41

f@p
m = c m


1.72
Analytically, the force require to accelerate the added-mass associated with a sphere at the
relative particle acceleration is defined as the added-mass force, i.e.

@p
f p f p
= c = c
t t t






u
w v
F

d
d d
d d d


1.73
The added-mass coefficient, c

, is typically taken as since this corresponds to both the value


for both the Stokes regime and inviscid flow (Maxey, 1993), and moreover it has been found
to be reasonable for a wide range of Reynolds numbers for spheres (see 6.4).
The third acceleration force arises when the viscous stresses on the particle surface are
unsteady and is called the history force (F
H
) since it will depend on the temporal development
of this flow. This force can be significant if the accelerations are large and the viscous effects
are strong but is zero if inviscid flow is considered (unlike the added mass force which is
present for both viscous and inviscid flow conditions). The history force arises due to the
unsteady development of the viscous region due to the particle no-slip condition. It is
generally treated separately from the quasi-steady drag force F
D
, but is, in fact, the unsteady
portion of the drag force so that the two forces together are the total drag.
To aid in the understanding of the history force, one may consider a flat plate which is at
rest in an initially quiescent flow (u=0 throughout) but then is impulsively accelerated to a
fixed velocity. The velocity field is governed by the one-dimensional unsteady flow equation:


2
x x
f f 2
u U P
t x x

= +

1.74
The creeping flow solution is Stokes First Problem (Crowe et al. 1998) and indicates that the
boundary layer above the plate grows in time after the impulsive acceleration (i.e. when the
plate is moving at a constant velocity) which causes the shear force to decay since the velocity
gradient is distributed an increasing distance. Eventually, the flow will be steady with a fixed
shear force on the plate. In the same way, a particle initially at rest and then accelerated to a
constant speed will have a viscous drag that decays in time until it finally reached the steady-
state solution given by Eq. 1.45.
Based on these same principles, the derivation for the history force on a spherical particle
was first formulated by Basset (1888) for creeping flow conditions (Re
p
0). Assuming the
particle is initially at a quasi-steady state and the continuous-flow is spatially uniform away
from the particle, the theoretical Basset history force is given by:


t
2 t 0
H f f
0
3 1
d
2 t t
=

= +



w w
F
d
d
d

1.75
Thus, the influence of previous accelerations on the current force decays as a function of t
-1/2
,
a result which is consistent with that of the flat plate. As with quasi-steady drag, there is no
analytical solution for the history forces for Re
p
>1. However, as discussed in 6.6, this force
tends to be negligible when Re
p
1, when the particle relative acceleration are weak compared
to those on the RHS of Eq. 1.69, or when the particle density ratio is high (
*
p
1).

42
1.5.4. Simplified Equations of Motion for an Isolated Particle

Translational Equations of Motion

Depending on the assumptions used, a variety of equations of motion can be put forth.
Based on the above surface and body force expressions at creeping flow (Re
p
0), the particle
equation of motion for the translational acceleration of an isolated, rigid, non-rotating,
spherical particle in a continuous-phase velocity field is:

(
p
+ c


f
)
p
t
d
d
v
= -3 d
f
w

+
p
(
p
-
f
) g +

t
@p @p 2 t 0
f p f f
0
3 1
c d
t t 2 t t
=


+ +


u
w w

d
d
d
d d
D
D
u

1.76
This is the classic Basset-Boussinesq-Oseen (BBO) equation. It is worth noting that this result
is non-empirical, i.e. it can be obtained directly from the incompressible continuum flow
equations (Eq. A.54) applied to the particles exterior surface.
If one wishes to employ a similar equation of motion at larger Reynolds numbers, the
creeping drag force is generally replaced by the empirical drag force (Eq. 1.56). The added
mass and fluid stress expressions are unchanged since they are also theoretically correct in the
potential flow limit, and have been found to be generally reasonable for a wide range of Re
p

conditions. However, the history and lift forces are generally neglected for non-creeping
conditions since they are somewhat controversial and often weaker for higher Reynolds
numbers (although significant advances have been made for some conditions as discussed in
6.3 and 6.6). Thus, a corresponding non-creeping equation of motion can be given as
( ) ( )
@p @p
p f p f p f p f p
c 3 d f c
t t t


+ = + + +


u
w g


d
d
D
d
d D
u
v

1.77
This equation of motion is used often by many multiphase researchers in computations for
general Re
p
conditions with the above assumptions. For a more general equation of motion,
Chapter 6 will discuss incorporation of non-spherical and compressible flow effects, lift due to
vorticity and/or particle rotation, history force, thermophoresis, etc.
The particle acceleration terms have been placed on the LHS (left-hand-side) of Eq. 1.77
indicating that an effective mass of the particle can be defined as the sum of the particle mass
and the virtual mass:
m
eff
m
p
+ m

= (
p
+ c

f
)
p
1.78
One can consider two limits of this effective mass expression. If the particle density is much
greater than that of the surrounding fluid (
*
p
1), as is typically the case of a liquid or solid
particle in a gas, then the effective mass is approximately equal to m
p
. However, if the particle
density is much less than that of the surrounding fluid (
*
p
1), as is typically the case of a gas
bubble in a liquid, then the effective mass is approximately equal to m

.
The terms associated with gravity in Eq. 1.77 can be similarly grouped together to form an
effective gravitational force which is proportional to the difference between the particle mass
and the displaced mass (the Archimedes principle):
43
( ) ( )
g,eff p p f p f @p
m m = = F g g
1.79
If the particle density is much greater than that of the surrounding fluid (
*
p
1), then the
effective gravitational force can be approximated as the body force of the particle (F
g,eff
F
g
).
However, if the particle density is much less than that of the surrounding fluid (
*
p
1), then the
effective gravitational force can be approximated as the buoyancy force acting on the particle
(F
g,eff
F
buoy
).
With the above notations and the assumption of a spatially-uniform unhindered continuous-
phase velocity ( 0 = u ), the isolated spherical particle equation of motion becomes:

( )
@p
eff D g,eff f p
m + 1 c
t t

= + +

u
v
F F



d
d
1.80
This equation may written in terms of particle acceleration by defining an acceleration
parameter (R) as follows:

( )
( )
f f
*
p f p
@p
f
2 *
p
c 1 c
R
c c
18 f
- + 1 R R
t t d c

+ +
=
+ +

= +
+
u
w v
g



d
d

1.81a


1.81b
The acceleration parameter relates the importance of the fluid acceleration to the particle
acceleration. Assuming an added-mass coefficient of , R approaches 3 for very-buoyant
particles (
*
p
1) indicating the fluid acceleration effect can be substantial. However, R tends
to 0 for very dense particles (
*
p
1) indicating the fluid acceleration effect (last term on RHS
of Eq. 1.81b) will be small. Even in steady flows, the acceleration parameter can be important
since it helps relate the particle acceleration to effective gravitational acceleration. For
example, very-heavy particles (
*
p
1) initially accelerate at g but very-buoyant particles
(
*
p
1) initially accelerate at -2g, due to the added mass effect.


Terminal Velocity

If only drag and effective gravitational forces are important, the relative particle
velocity will eventually an relative equilibrium speed which is called the terminal velocity
(w
term
). This velocity is independent of the initial conditions and is equal to the steady-state
velocity achieved in quiescent conditions and so can be obtained based on the balance between
drag and effective gravitational forces of Eq. 1.81b. It will be in the direction of g for
*
p
>1 or
g for
*
p
<1 and can be expressed by either the Stokes correction factor or the drag coefficient:
44

( )
2
p f
term
f term
p f
term
f D,term
d
18 f
d
4
w
3 C

=

g
w
g

1.82a


1.82b
The magnitude of the terminal velocity increases for increasing particle diameter or density
difference increases, but this sensitivity is stronger in the creeping flow regime where f is
approximately constant. Examples of typical terminal velocities (using Table A.1 and Eq.
1.83) are 3.8 m/s (Re
p
250) for a 1 mm water drop an 0.11 m/s (Re
p
110) for a 1 mm
contaminated bubble rising in water. A linear (Stokes) drag generally requires Re
p,term
< 0.1
which corresponds to diameters less than 60 m for a water drop in air or less than 40 m for a
bubble in water.
For creeping flow or Newton-drag, w
term
can be computed explicitly since either f or C
D

is a constant in these two regimes. In the intermediate Reynolds number regime, these
coefficients are a function of Re
p
so that an iterative solution is required to determine w
term
, e.g.,
using Eqs. 1.26, 1.57, and 1.82a). To avoid this, Ferry & Balachandar (2001) developed an
approximate explicit correlation for the Stokes correction:

( )
3 2
p,Stokes f p f f
Re gd 18 =
( )
0.18
term p,Stokes p,Stokes
f 1 0.76Re 1 0.025Re

+ +

for Re
p
<1000
1.83a

1.83b
From this f
term
expression, w
term
and Re
p,term
may be computed using Eq. 1.82a and 1.26. This
non-iterative approach yields a drag coefficient accurate to within 2% when compared to
experimental data of Fig. 1.27.


1.6. Mass and Heat Transfer for an Isolated Spherical Particle

While fluid dynamics are the focus of this text, simplified equations for heat and mass
transfer are introduced herein to provide the reader with the basic relationships. For more
detailed discussion on heat and mass transfer, the reader is referred to texts of Clift et al.
(1978), Soo (1990) Crowe et al. (1998), Sirignano (1999), Brennen (2005) and Michaelides
(2006). These equations presented herein are idealized for the simple case of an isolated
spherical particle with uniform internal temperature and density. The continuous-phase
surrounding the particle is assumed to have uniform temperature and radiation and chemical
reactions are neglected. Furthermore, the kinetic energy associated with the fluid velocity
emanating from the particle surface due to any mass transfer will be neglected compared to
changes in the internal energy (temperature) of the particle. As such, the below relations are
reasonable for most non-combusting particles.
The particle mass transfer rate is defined as the change in particle mass per unit time and
can be related to the change in particle diameter:

( )
2
p p
p
d m
d
m
t 2 t

= `
d d
d d

1.84
45
The rate at which mass is transferred from the surrounding fluid to the particle is thus defined
as positive, e.g.
p
m` >0 for a condensing drop and
p
m` <0 for an evaporating drop. The mass
transfer will be governed by two effects: 1) how close the surrounding fluid is to saturation for
the particle species and 2) how fast the particle mass species can move (diffuse) through the
surrounding fluid.
The first effect is related to the particle species mass fraction within the surrounding
fluid defined as X
p@f
. For example, this can represent the vapor fraction of the particle matter
if the surrounding fluid is a gas or the liquid fraction of the particle matter if the surrounding
fluid is a liquid. The normalized difference between the species mass fraction on the surface,
which is assumed to be saturated (X
p@f,sat
) and that in the far-field (X
p@f,
) can be quantified by
the Spalding transfer number, B, as:

p@f,sat p@f,
p@f,sat
X X
B
1 X


1.85
For a gas, the mass fraction is equal to the ratio of the particle species vapor pressure (P
vap
) to
the local unhindered gas pressure (p) times the ratio of the particle species vapor molecular
weight (MW
vap
) to the overall gas molecular weight (MW
g
). If the particle has a uniform
internal temperature, then the vapor pressure at the surface (P
vap,surf
) is equal to the saturated
vapor pressure at the droplet temperature (T
p
). In contrast, the unhindered vapor fraction
(p
vap,
) will be based on the gas temperature (T
g
) and need not be saturated so that Eq. 1.85
becomes:

( )
vap,surf vap
vap
vap vap,surf
P p
B
p MW / MW P

g

1.86
If the surrounding gas has the same vapor pressure as the saturated value found on the particle
surface, the Spalding number (and mass transfer) will be zero indicating equilibrium. However,
high particle temperatures or low vapor concentrations in the far-field will lead to evaporation
with B
vap
>0 and evaporation, i.e.
p
m` <0. In contrast, low drop temperatures or high vapor
concentrations in the far-field will yield to condensation with B
vap
<0 and
p
m` >0.
The second effect is the diffusivity of the particle species while in the surrounding fluid,
which is denoted herein by
p@f
. For a droplet evaporating or condensing in a gas,
p@f
is
given by the molecular diffusion of the droplet vapor within the surrounding fluid, which
indicates its net diffusivity (Eq. A.48). For a bubble condensing in a liquid or a particle
melting in a liquid,
p@f
becomes the molecular diffusion of the particles species in a liquid
based on conditions just outside of the particle surface. Higher diffusivity allows the species to
move faster toward the particle (positive mass transfer) or faster away from the particle surface
(negative mass transfer). One may then define a normalized mass transfer as the Sherwood
number (based on the surrounding fluid density and the particle diameter):

p
f p@f
m
Sh
d B


`

1.87
This allows the mass transfer relationship of Eq. 1.84 can be written as:

f p@f p
2
p p
2 B Sh 2m
d
t d d

= =

`
d
d

1.88
46
For creeping flow (Re
p
0) and saturated conditions (B0), an exact solution for the mass
transfer rate can be obtained using the Stokes solution of velocity field coupled with solution of
the concentration diffusion equation for negligible convection. The resulting analytical
solution corresponds to Sh=2 (Clift et al. 1978). If the Spalding number and the diffusivity are
also constant (consistent with weak changes in temperature) and the particle density is constant,
the particle surface area changes linearly in time. This is often referred to the d
2
law which
can be expressed in terms of a mass-transfer time-scale (
m

`
) as:

f p@f 2 2 2
t o t o
p m
2
p t o
m
f p@f
4 B Sh
t B
d d t d 1
B
d
4 B Sh
= =
=

= =

`
`

1.89a


1.89b
The mass-transfer time-scale indicates the time it will take for the particle mass to disappear
for B>0, but indicates the time for the particle surface area to double for B<0.
The particle heat transfer from the continuous-phase to a single particle is defined as
p
`
Q but particle energy can also be modified by phase change so that the Lagrangian ODE is
given by:

( )
p,p p
p phase p p
T
m m
t
= +
`
`
d c
h Q
d

1.90
In this equation, c
p,p
is the specific heat (at constant pressure) of the particle and h
phase
is the
enthalpy required for particle matter to change from a phase consistent with the particle to a
phase consistent with the surrounding fluid phase. For a droplet in a gas, h
phase
is equal to the
heat of vaporization which is a positive quantity. An evaporating drop will experience a
positive heat transfer (
p
`
Q >0), a negative mass transfer (
p
m` <0) and a T
p
given by the wet-bulb
temperature. The temperature condition indicates that the LHS of Eq. 1.90 is approximately
zero so that the two RHS terms must be equal and opposite. The heat transfer rate from the
continuous-phase is generally proportional to the particle diameter, the thermal conductivity of
the continuous phase (k
f
), the temperature difference between the far-field fluid and the particle
surface (T
f
-T
p
), and the relative velocity of the particle. For static conditions, this
proportionality can be used to define the non-dimensional heat transfer as the Nusselt number:

( )
p
f f p
Nu
d T T


`
Q
k

1.91
A more formal and general definition of the Nusselt number is given by Sirignano (1999)
which allows internal temperature variations, whereas the form herein assumes that the particle
internal temperature is uniform and equal to T
p
. For creeping flow conditions (with no
convective heat transfer), the heat transfer around the particle is similar to that for mass
transfer except governed by thermal diffusivity instead of mass diffusivity. As a result, the
exact solution of the temperature field PDEs similarly yields a constant non-dimensional
transfer rate which corresponds to Nu=2.
Assuming the specific heat of the particle is constant, the temperature evolution given
by Eq. 1.90 can be written in terms of the temperature difference and the change in particle
diameter:
47
( )
p phase
f
f p 2
p p,p p,p
T 3
6Nu d
T T
t d d t
= +
d h
k d
d c c d

1.92
If the surrounding fluid temperature is approximately constant and mass transfer is neglected,
the ODE can be integrated to show that the relative temperature difference will decay
exponentially based on a thermal response time (
T
):

p f
f
2
po f p p,p T
2
p p,p
T
f
T (t) T
12 t t
exp exp
T T d
d
12

= =







k
c
c
k

1.93a


1.93b
The particle temperature can be considered constant for t
T
and can be considered to be in
equilibrium with the surrounding fluid for t
T
. Thus, the thermal response time is an
indication of the time-scale to reach equilibrium if there is an initial temperature difference.
For droplets in a gas or solid particle in a liquid, the mass transfer and heat transfer time-scales
of Eqs. 1.89b and 1.93b are similar and these phenomena can be coupled through Eq. 1.90. For
bubbles, the particle thermal inertia is so small that its temperature ir primarily governed by
that of the surrounding liquid unless there are very rapid changes in liquid temperature or
bubble volume (Owis & Nayfeh, 2003).
For finite Reynolds and Spalding numbers of drops, the heat and mass transfer rates
increase beyond their creeping values due to convection effects. These effects can be modeled
with the empirical Ranz-Marshall (1952) relationships which assume that the mass transfer is
proportional to the heat transfer and which employ the Schmidt and the Prandtl numbers of the
surrounding fluid (Eqs. A.51 and A.52) as follows:

( )
( )
1/2 1/3
p f
ln 1 B
Sh 2 0.6 Re Sc
B
+
= +
( )
( )
1/2 1/3
p f
ln 1 B
Nu 2 0.6 Re Pr
B
+
= +
1.94a


1.94b
While these relationships are widely employed, more detailed expressions can take into
account variable Spalding numbers, high mass transfer rates, effects of particle spacing, etc.
(Sirignano, 1999).


1.7. Chapter 1 Problems

For all problems in all chapters, show all steps and discuss all assumptions.

1.1) Describe a research area or engineering application that involves multiphase flow.
Discuss some of the key particle interaction physics and issues of current research.
1.2) Express the particle Knudsen number (Kn
p
) in terms of the particle Reynolds number
(Re
p
) and relative Mach number (M
p
=w/a). Determine the minimum diameter particle that is
considered to be in continuum for air at a standard altitude of 10 km where p=26,400 Pa,
T=50
o
C and
f
=1.447x10
-5
kg/m-sec.
48
1.3) Starting from Eqs. A.6a and A.8, derive the Stokes flow PDE of Eq. 1.27 for flow
about a particle in the creeping flow limit (Re
p
0) by non-dimensionalizing terms with
particle diameter and relative velocity as well as fluid density and viscosity.
1.4) Starting from the velocity field of Eq. 1.37, obtain the friction drag given by Eq. 1.44.
1.5) For inviscid flow over a sphere, show that the stream-function of Eq. 1.47a is
consistent with the velocity field of inviscid velocity field of Eqs. 1.47b and 1.47c using the
relationships of Eq. 1.29. Further show that Eqs. 1.47b and 1.47c satisfy the far-field and
surface boundary conditions as well as continuity equation given by Eq. A.57.
1.6) Starting from the pressure field solution of Eq. 1.62, obtain the form drag for an
uncontaminated spherical gas bubble which has negligible viscosity.
1.7) Starting from Eq. 1.77, obtain Eqs. 1.81 and 1.82 and note all assumptions. For a 200
micron diameter rain drops at STP obtain w
term
and Re
p
to within 1% by iteration with Whites
drag coefficient and Table A.1 properties. Compare this result to the non-iterative technique of
Eq. 1.83.
1.8) Starting from Eq. 1.90, obtain the ODE for the particle temperature, and then apply
the necessary assumptions to obtain the solution given by Eq. 1.93.
49

2. Characterizing Particle Sizes, Shapes and Trajectories
2.1. Particle Size Distribution

There are a wide variety of mechanisms which can cause size distributions in
multiphase flows such as solid particles break-up (e.g., from crushing), drop break-up (e.g.,
from spray turbulence and shear), or bubble break-up (from fragmentation of air jets). In
addition, particle-particle coalescing can lead to polydisperse size distributions of raindrops or
soot particles. In many multiphase flows, it is important to understand the size distribution
which can be expected. For solid particles, this can be measured by using a series of finer and
finer screens (which filter particles with sizes greater than a certain diameter). The size
distribution can be quantified by noting the mass fraction (by weight) or number fraction (by
count) which remains after each screen. A particle group whose diameters are all nearly
uniform is said to have a mono-disperse size distribution. Conversely, a particle group
whose diameters vary significantly is said to have a poly-disperse size distribution. Size
variations within a particle cloud will result in a variety of particle responses and thus a range
of particle paths and relative velocities, etc. For example, fluidized beds typically involve a
complex mixture of solids with gases for which the particle size (and relative velocity)
determines such aspects as bed expansion, fluidization velocity, and various mass and heat
transfer properties. Similarly, droplet sprays have a size distribution which is due to complex
break-up mechanisms near the nozzle. In both cases, the distribution of particle sizes is critical
to the performance of the combustion system. Therefore, it is important to characterize this
aspect in a particle mixture. The following section focuses on describing these distributions
through statistical integration.

2.1.1. Definition of the Probability Distribution Function

Measurements of the size distribution are often obtained by separating the distribution into
a series of bins, which are groups of particles of approximately the same diameter. The
resulting size distribution is commonly referred to as a size probability distribution function
(PDF). The number-based probability distribution function, P
N
(d), is the probability per bin
width (d) that an individual particle diameter will be in the range of d-d to d+d. Thus,
P
N
(d) has units of 1/length and in a discrete sense is the number of particles in a bin (N
pd
)
normalized by both the total number of particles (N
p
) and the bin width (d) and the sum of the
probabilities times the bin width is unity:

( )
N,i i p d,i p
(d ) N / N d

P
d
N
N,i i
i 1
(d ) d = 1

P

2.1a

2.1b
In the latter equation, the total number of bins is N
d
. An example probability size distribution
for a spray is shown in Fig. 2.1a, where the rectangular sky-line function describes the
measured bin distribution with a discrete representation. If the sampling error for the number
50
of particles counted in a bin is Gaussian, the uncertainty of a bin probability is proportional to
(N

)
-1/2
, e.g., 100 particles are needed for 10% uncertainty and 10,000 for 1% uncertainty.
As in Fig. 2.1a, the measured data can also be fit with a continuous analytical probability
distribution function (example functions will be discussed in the next section). The integral of
the continuous representation of the PDF over all possible particle diameters is also unity:

N
0
(d) d = 1

P d 2.2
The discrete and continuous descriptions will be equivalent as the bin sizes tend to zero if a
statistically large number of particles (with respect to the desired accuracy level) is considered.
The integral of the PDF to a certain diameter d (normalized by the integral over all sizes)
represents the fraction of particles which have a diameter of d or less. This fraction is defined
as the cumulative distribution function (CDF) based on the number of particles, C
N
(d):

d
N d
0
N N
0
N
0
(d) d
(d) = (d) d
(d) d

P d
C P d
P d
2.3
As such, C
N
(d) equals unity (100%) for the largest possible particle diameter. As an example,
the discrete and interpolated CDF corresponding to Fig. 2.1a are given in Fig. 2.1b.
In addition to the number-based distributions discussed above, one may also define
volume-based distributions. The volume based probability distribution (P

) is based on the
volume of particles within a bin normalized by the bin width and the total volume of particles,
while the cumulative volume-based distribution function (C

) is based on the integral value:



p d p
N N
p,i p,i
i 1 i 1
d d
3
N d
0 0
3 0
N
0 0
1
(d) /
d
(d) d (d) d d
(d) = (d) d =
(d) d (d) d d

= =


P
P d P d
C P d
P d P d

2.4a




2.4b

The cumulative volume-based distribution function is equal to the mass-fraction distribution
function, assuming the particle density is uniform. Knowledge of C

can be used to define the
mass-fraction diameter (d
q%
) where q% represents the volume fraction associated with
particles of diameter d or smaller, i.e.

q% q%
d d

C @

2.5
Thus, half the particle mass is contained in the range dd
50%
. The mass fraction diameter can
also be used to define broadness of a size distribution using the ratio of the diameter with
C

=90% to the diameter defined by C

=10%, i.e. d
90%
/d
10%
. If the particle distribution is very
narrow, this ratio will tend toward unity. For example, Gauthier et al. (1999) suggested that a
51
particle mixture can be considered approximately monodisperse for d
90%
/d
10%
<1.2. However,
as the particle size distribution gets broader, d
90%
/d
10%
will increase indicating a more
polydisperse distribution.
There are two common uses for number and volume distributions in numerical multiphase
flow applications. The first is to use the size-distribution to obtain effective characteristics for
a particle-cloud defined as many particles moving as a group, i.e. with similar velocity and
location. The average diameter within the cloud is often described as the effective particle
diameter so that only a single diameter needs be considered for the equation of motion. The
second is to use the size-distribution to generate a statistically large number of individual
particles (each with individual diameters and equations of motion) such that the net distribution
is properly represented. This is often done by approximating the distributions with analytical
functions. The effective diameter concept is discussed in the next section, while common
analytical distribution functions used for the second aspect are treated in 2.1.3.

2.1.2. Average and Effective Diameters

When there are a large number of particles, it may not be possible to measure or simulate
the motion of each particle. Therefore, average particle characteristics can be useful in
describing a group of particles. For multiphase flows with a large variation in particle diameter,
an average diameter can be employed to represent the dynamics or coupling of the particles.
This can be especially important for Eulerian representations of particles (discussed in 4.1)
since the PDEs for the particle concentration and momentum distribution assume an average
particle diameter evaluated over a computational control volume, and representing multiple
particle diameters often requires multiple sets of Eulerian PDEs. Thus, it is computationally
convenient and expedient to identify an average particle diameter to represent the polydisperse
distribution. For a polydisperse diameter distribution, one can define several average
diameters based on a measured or interpolated distribution function as discussed below. In
addition, effective diameters for certain fluid dynamic phenomenon can be determined based
on such averages and identification of the controlling physics.


Average Weighted Diameters

There are many average diameters that can be used to describe a size distribution
depending on how the probability distribution function is weighted. A general expression for a
particle diameter average (d
ij
) with various diameter weightings can be based on the ratio of
two differently weighted PDFs:

1
i j
i j
i j N N
0 0
d d d d d d d




( )
( ) ( ) P d P d
2.6
Three common averages are the number-averaged diameter (d
10
), the area-averaged diameter
(d
20
) and the volume-averaged diameter (d
30
) which are defined as
52

10 N
0
1
2
2
20 N
0
d d d d
d d d d

( )
( )
P d
P d

1
3
3
30 N
0
d d d d

( ) P d
2.7a


2.7b


2.7c
These equations can employ either the discrete or continuous expressions for P
N
(d). For a
given size distribution, increasing the weighting power causes an increase in the average
diameter. For example, the number-average diameter will be less than the area-averaged
diameter which is less than the volume-averaged diameter, i.e., d
10
<d
20
<d
30
, etc.
The three diameters of Eq. 2.7 can help characterize the effective diameter for processes
that are proportional to particle number, surface area or particle volume. For example, the
average drag for a particle in a Newton-based regime is proportional to the particle area and
would be best represented by d
20
. Alternatively, the representation of the net gravitational
force for a particle cloud is proportional to the particle volume and would be best represented
by d
30
. Many physical processes involve a balance between surface-averaged and volume-
averaged physics. Some examples include drag force vs. gravitational force, droplet
vaporization rate vs. particle thermal inertia, surface tension energy vs. particle kinetic energy,
etc. Because of the ubiquity of this balance, two common average diameters used in
engineering systems are the Sauter mean diameter (SMD) and the volume-width mean
diameter (VWMD) which are defined as

3
3 2 30
32 N N 2
20 0 0
3
3 30
31 N N
10 0 0
d
d (d)d d (d)d d
d
d
d (d)d d (d)d d
d


=
=


P d P d
P d P d

2.8a

2.8b
As shown below, these mean diameter are effective diameters for a multiphase system
governed by drag and gravitational forces under certain conditions.


Drag-Based Effective Diameters

To obtain the average diameter which best represents the motion of a cloud of particles, it
is important to determine the forces that dominate the particle motion. If the particle
accelerations are weak, then Eq. 1.80 indicates that these forces include the effective
gravitational force and the drag force. The particle cloud will then move at some single
effective relative velocity, w
eff
, which is characterized by a drag effective diameter, d
eff
. To
obtain this effective diameter, one may apply the terminal velocity force and assume that the
particle mixture is governed by a uniform particle density as well as a uniform fluid density
and viscosity. The effective diameter and relative velocity of the particle cloud can be
obtained from the integrated force distributions based on the size PDF.
The net effective gravitational force magnitude can be obtained by integrating over all
particle sizes based on Eq. 1.79 as:
53

p p
N N
p f p f
3
g,eff p,i N
i 1 i 1
p p o
g g
1
F (d)d d
N N 6

= =

= =

P d 2.9
This indicates that this net force is proportional to d
30
. For the net drag force for a polydisperse
mixture moving at the effective relative velocity, Eq. 1.46 applied to a size distribution yields:

p
N
2 2
D f eff N D
i 1
p o
1
F w (d)C (d)d d
N 8

P d
2.10
The integrand includes the drag coefficient, which is a function of particle Reynolds number.
By assuming the surrounding fluid density and viscosity are constant, we can consider the two
limits of Newton-based drag and Stokes drag to evaluate the effective diameter and velocity.
Particles in the inertia-dominated drag regime (about Re
p
>2000) have a drag coefficient
which is independent of Reynolds number (C
D,crit
) such that the drag force is proportional to
the area-averaged diameter d
20
. Note that this result is independent of the exact value of
C
D,crit
(e.g., those caused by particle deformation, flow compressibility, circulation within a
fluid particle, three-way interaction, etc.) as long as the drag coefficient is not a function of
particle Reynolds number. The balance of the net gravitational force and the net drag force is
given by F
g,eff
=F
D
. For the inertia-dominated drag regime (Rep > 2000), Eqs. 2.9 and 2.10
can be combined to yield

2 2 * 3
eff D,crit N p N
o o
3
w C (d)d d g 1 (d)d d
4

=

P d P d
2.11
The ratio of the two integrals is equal to the Sauter mean diameter (Eq. 2.8a). Therefore, the
group relative velocity (equal to the group terminal velocity) becomes:
( )
*
eff p 32 D,crit
w 4 / 3 1 d g / C = 2.12
Thus, the Sauter mean diameter is the drag-effective diameter (d
eff
=d
32
) for particles governed
by Newton-based drag, regardless of the particle-size PDF (i.e., it can be monodisperse, bi-
modal or polydisperse). This result is consistent with the common usage of the SMD to
characterize a polydisperse fluidization (Wen & Yu, 1966; Wu & Baeyens, 1998).
Now let us consider the other extreme of creeping flow where C
D
=24/Re
p
. Substituting this
drag coefficient (which is inversely proportional to the particle diameter) into Eq. 2.10 and
then equating this with Eq. 2.9 yields a balance of drag and gravitational forces:

* 3
eff f N p N
o o
18w (d)d d g 1 (d)d d

=

P d P d
2.13
The effective velocity is then proportional to the square of the volume-width diameter
(d
eff
=d
31
) of Eq. 2.8b:
( )
2 *
eff 31 p f
d 1 18 = w g


2.14
54
The inclusion of other effects (due to shape, flow compressibility, slip-surface, three-way
interaction, etc.) will not significantly affect this result so long as C
D
is proportional to 1/Re
p
,
which is often be true for creeping flow (6.2).
The above analysis can be extended to the intermediate Reynolds number regime to
determine the effective diameter and velocity. For this, it is important to determine the drag
coefficient dependence on particle diameter for Eq. 2.10, which in turn is equivalent to
determining the C
D
dependence on Re
p
if
p
,
f
, and
f
are fixed. If one assumes a power
relationship for small changes in Re
p
, this can be expressed as

( )
Y 2
D Y p
C c Re

2.15
In this relationship, the local slope in a log-log version C
D
vs. Re
p
is the power exponent, Y-2,
and c
Y
is a coefficient. The exponent term Y is a function of Reynolds number whereby Y=1
and c
Y
=24 for the creeping flow limit but Y=2 and c
Y
=C
D,crit
for the inertia-dominated limit.
To obtain an analytical expression for Y for intermediate Re
p
, one may take the natural log of
both sides of Eq. 2.15:
( )
D p Y
ln(C ) Y 2 ln(Re ) ln(c ) = +
2.16
Considering only small changes in Re
p
so that the last term on the RHS is approximately a
constant, and then differentiating this equation with respect to Re
p
yields:

p
D D
p D p
Re
[ln(C )] C
Y 2
[ln(Re )] C Re

=


2.17
A closed-form expression for the LHS can be obtained from the White empirical drag
coefficient expression (Eq. 1.52), so that the drag-power parameter for a solid particle is:

( )
p
2
p p p
p
3 Re
24 24 6
Y 2 0.4
Re Re 1 Re
1 Re



= + + +

+

+




2.18
The variation of this parameter with particle Reynolds number is given in Fig. 2.2a. However,
the drag coefficient for a pure (uncontaminated) spherical bubble is proportional to 1/Re
p
for a
wide range of Re
p
conditions (as will be discussed in 6.2.5), in which case Y1.
The effective diameter for Eq. 1.82 is then d
3Y
with a corresponding effective velocity:

2
(3 Y)
3 2 * *
N 3Y p p
0
eff
Y
f Re f Re
N
0
(d)d d d 1 g 1 g
w
18 f 18 f
(d)d d




= =



P d
P d

2.19
To determine the importance of the Y parameter, one may compare the effective terminal
velocity based on the Sauter mean diameter (d
32
) to one based on the correct diameter (d
3Y
).
This velocity ratio should approach unity if the SMD is reasonable to use. To illustrate this
importance consider a simple bi-disperse mixture composed of two particle sizes (d
small
and
d
large
). The size ratio for this mixture is defined as:

*
small small large
d d / d
2.20
55
The influence for a range of diameter ratios is plotted in Fig. 2.2.b where it can be seen that the
velocity ratio approaches unity for high Re
p
indicating that the SMD is appropriate. However,
use of the d
3Y
becomes more important for smaller Re
p
values and as the size distribution
broadness increases, e.g. the creeping flow effective velocity over-prediction by 20% for
*
small
d 0.33 = . For a more general polydisperse distribution, the effective diameter and particle-
cloud velocity based on Eq. 2.19 have been found to be reasonable for a fluidized beds when
the size distribution is not too broad (Loth et al. 2004)

eff
w(d) w if d
90%
/d
10%
< 3
2.21
Broader size distributions exhibited substantial vertical segregation of particle (e.g., larger
particles found at lower heights) indicating a single effective velocity is no longer reasonable.
The above analysis can also be modified for the case where the particle density varies
according to particle diameter,
p
(d), by defining an effective particle density as
p,eff
. Wu &
Baeyens (1998) simply define this effective density as the sum of the total particle mass
normalized by the sum of the total particle volume:

3 3
p,eff N p N
0 0
(d) (d)d d (d)d d



P d P d
2.22
In this case, the particle density ratio for w
eff
becomes
p,eff
/
f
, where the same effective
diameter dependencies (with Y) still arise.
This general approach can be used for other multiphase characteristics (density, viscosity,
velocity, temperature, etc.) if the key physics are identified and the PDF is known. If discrete
experimental distributions are used, they must be integrated for different Re
p
values, etc. As
such, it is often convenient to use analytical functions to describe such PDFs since they can
lead to corresponding analytic expressions for mean diameters, etc. Such analytic PDFs are the
topic of the next section.

2.1.3. Analytical Probability Distribution Functions

Despite the wide variety of mechanisms which can cause particle size distributions
(crushing, break-up, coalescence, etc.), most size distribution functions are qualitatively similar.
Because of this, several analytic functions have been proposed to described the PDFs using a
mean diameter, a broadness parameter, and a normalized distribution function. Such
representations are useful in computational multiphase methods to allow rapid estimation of
average diameters or to allow random numerical sampling for a size distribution (7.2.2). In
the following, the two most common analytical PDFs of particle size are discussed: the Rosin-
Rammler distribution and the Log-Normal distribution.


Rosin-Rammler Distribution

The Rosin-Rammler distribution was originally developed to characterize pulverized coal,
but was found to be quite robust in representing the cumulative mass PDF (C

) for a wide
variety of powdered materials (Rosin & Rammler, 1933). As a result, this is one of the most
56
common size distribution functions in existence today and is used in a wide variety of
applications. The PDF form is relatively simple and includes two empirical constants:
( ) RR
1
h
RR
(d) 1 exp d / d


=


C

2.23
In this equation, d
RR
and h
RR
are respectively the Rosin-Rammler reference diameter and
spread parameter, which describe a mean size and the broadness of the distribution. The mean
size corresponds to d
RR
d
63%
, which can be seen by substituting d=d
RR
into Eq. 2.23. Size
distributions for two different spread parameters are shown in Fig. 2.3a where it can be seen
that an increasing h
RR
corresponds to an increasing difference between d
10%
and d
90%
. The
spread parameter is zero for a monodisperse (uniform) size distribution and is of order unity for
a very polydisperse (broad) size distribution. Historically, experimental data have been plotted
on a log-log scale as in Fig. 2.3b to obtain the spread parameter (equal to the inverse of the
slope on such a plot) and the reference diameter, d
RR
(equal to diameter corresponding to
C

0.63). Such reasonable agreement for mass-based CDF is typical such that the Rosin-
Rammler distribution is the most common representation of mass distribution for pulverized
materials.
The Rosin-Rammler spread-parameter can be used to quantify the broadness of a size
distribution in terms of the ratio of the 90% diameter to the 10% diameter (Eq. 2.5):
( )
RR 90% 10%
h 0.324ln d / d
2.24
For example, the h
RR
=0.2 of Fig. 2.3a corresponds to a diameter ratio of 1.85 and larger
spreads will be consistent with larger diameter ratios. This spread parameter can also be used
to determine whether a particle size distribution is approximately monodisperse. The
definition for monodisperse given by Gauthier et al. (1999) is d
90%
/d
10%
<1.2, which
corresponds to h
RR
<0.06, while that given by Crowe et al. (1998) is (d
rms
-d
50%
)/d
50%
<0.1, which
corresponds to h
RR
<0.05.
In order to compute various diameter averages, Rosin & Rammler (1933) derived the
number-based PDF in terms of the spread parameter and the reference diameter as:

( )
( )
( )
RR
RR
RR
RR
RR
1 4h
1
h
h
N
RR RR RR
d
d
d /
(d) exp d /
h d 1 3h


=



P
2.25
This solution includes the mathematical Gamma function, defined for an arbitrary variable q by

q 1 b
0
2
(q) b e b

d
2.26
The Gamma function is plotted in Fig. 2.4a and has the properties (q+1)= q(q), ()=
1/2
,
and (1)= (2)=1, but is undefined at q=0. Based on the properties of the Gamma function,
the general relationship for an average diameter-weighted probability can be found from
integrating Eq. 2.25:

[ ]
( )
( )
i
RR i
N RR
RR 0
1 (i 3)h
(d)d d d
1 3h

+
=

P d
2.27
57
Based on this equation and Eq. 2.6, the Sauter and volume-width diameters become:

( )
( )
32 RR
RR
31 RR
RR
1
d d
1 h
1
d d
1 2h
=

=


2.28a


2.28b
Note that these mean diameter expressions indicate the limits of applicability for the Rosin-
Rammler distribution function. In particular, it is required that h
RR
<1 in order to compute d
32

and required that h
RR
< to compute d
31
(else the diameters can be negative or undefined due to
properties of the Gamma function). This unfortunately prevents computation of effective
diameters for size distributions which are very broad. Furthermore, d
10
values based on the
Rosin-Rammler distribution function are sometimes in poor agreement with the experimentally
measured d
10
values, a problem discussed by Mugele & Evans (1951). If accurate descriptions
of number-based distribution instead of mass-based distribution are desired, Kuo (1986)
suggests using other distribution functions, which are discussed below.


Log-Normal Distribution

The Log-Normal size distribution assumes a Gaussian distribution of particle size about a
log-based normalized diameter for the number-based PDF

( )
2
LN
N
LN LN
ln d / d
1 1
(d) exp
2 h 2 h d



=



P
2.29
In this equation, h
LN
represents the Log-Normal spread parameter and d
LN
represents the Log-
Normal reference diameter. This distribution can be integrated using Eq. 2.3 to give the
cumulative number distribution

( )
LN
N
LN
ln d / d
1
(d) 1 erf
2 h 2



= +



C
2.30
In this equation, erf is the Error function, which is defined for a variable q by

2 q
b
0
2
erf (q) e b

d
2.31
The Error function is plotted in Fig. 2.4b and has the following properties: erf(-q)=-erf(q),
erf()=1, and erf(0)=0. Substituting d=d
LN
into Eq. 2.30, shows that the Log-Normal mean
diameter corresponds to a C
N
=50%, i.e. the diameter which equally divides the number-based
particle population. The mass-based cumulative function can be similarly obtained:

( )
2
LN LN
LN
ln d / d 3h
1
(d) 1 erf
2 h 2




= +



C
2.32
Thus, the diameter which equally divides the particle mass is given by d
50%
=d
LN
exp(3h
LN
2
).
58
Example fits of the Log-Normal function for number-based and volume-based cumulative
distributions are given for spray data in Fig. 2.5a. In general, the fit is quite reasonable and it
can be seen that C

<C

for a given diameter, which is consistent with d


30
>d
10
. Note that the
spread parameter from Fig. 2.5a corresponds to a relatively broad distribution, which is beyond
the capability of the Rosin-Rammler function for characterizing the volume-width diameter
(see Table 2.1). The Log-Normal distribution generally provides a reasonable representation
of the number-based size distribution for droplets and bubbles as shown in Fig. 2.5b. Based on
Eqs. 2.5 and 2.32, the Log-Normal spread parameter can be related to the size ratio as:
( )
LN 90% 10%
h 0.39ln d / d
2.33
A monodisperse distribution defined by d
90%
/d
10%
<1.2 thus corresponds to h
LN
<0.07.
The general equation for a mean diameter of the Log-Normal distribution can be obtained
from Eqs. 2.6 and 2.29 as:

1 2
ij LN LN
2
d d exp (i j)h = +


2.34
This can be used to relate the Sauter mean diameter to the volumetric-width mean diameter and
to relate the effective diameter (for an arbitrary Re
p
) to the log-normal mean diameter:

d
32
=d
31
exp(h
LN
2
/2)
d
eff
= d
3Y
= d
LN
exp [h
LN
2
(Y-3)/4]
2.35a

2.35b
Thus, the Log Normal distribution allows any mean diameter to be defined for any spread
parameter, which is not true for the Rosin-Rammler PDF. This is shown in Table 2.1 along
with the ratio of the SMD to VWMD indicating that the distinctions become significant for
braoder distributions, e.g. d
90%
/d
10%
>2. In addition, it should be noted that effective velocities
are not appropriate for h
RR
>0.36 or h
LN
>0.43 based on Eq. 2.21.
It should be noted that the predictive performance for the PDF described by the Log-
Normal distribution is not always good. For example, comparison with the Houghton droplet
data (Fig. 2.6) indicates that the Rosin-Rammler approach gives a better characterization. An
even better result is obtained using the upper-limit function (Kuo, 1986), which is a variant of
the Rosin-Rammler method that incorporates a maximum droplet size to prevent excessive tails
at large diameters. There are several other size distribution functions including the Nukiyama-
Tanasaw function (Kuo, 1986), the log-hyperbolic function (Crowe et al. 1998), and
distributions based particularly on bubble break-up (Lasheras, 1998). As such, there is no
single robust and universal distribution function.



2.2. Particle Shapes

While the most commonly used particle description is that of a sphere, non-spherical
shapes frequently occur due to natural or manufacturing processes. In this section, we will
consider particle shapes which are fixed or have reached an equilibrium at the terminal velocity.
Often the shapes are mild deviations from the spherical case and can be classified as
ellipsoidal-particles, i.e. the projected area is an ellipse when viewed from any of the three axes.
An ellipsoid has three different cross-sectional diameters and has a volume based on these
59
three diameters, i.e. d
3
=d
1
d
2
d
3
.
A special types of ellipsoid is the spheroid where one of the three cross-sectional shapes is
a circle and the other two are equal ellipse. This is equivalent to having two equal cross-
sectional diameters. This leads to an axis of symmetry which is perpendicular to the circular
cross-section. The axis lengths along the axis of symmetry (d
&
) and about the axis of
symmetry (d

) are shown in Fig. 2.7a, and their ratio defines the particle aspect ratio:

E d / d

=
|

2.36
Spheroids can be generally classified as oblate (E<1) or prolate (E>1) and quantified by their
aspect ratio. For oblate spheroids, the shape becomes a circular disk in the limit E0 , while
prolate spheroids approach a needle shape as E. The volume of an ellipsoid is proportional
to the product of the three major axes so that for both prolate and oblate spheroids:

2
p
6
= d d

|

2.37
Therefore, the volume-based diameter (Eq. 1.1) is given by

1/ 3 2/ 3
d=d E =d E

|

2.38
Of course, E=1 reverts the geometry to a sphere.
Such rounded shapes often occur for fluid particles because surface tension minimizes local
curvature. In contrast, solid particles can contain edges and points if they are formed by
crystallization which can lead to simple geometric shapes that are cylindrical, conic or cubic as
shown in Fig. 2.7b. However, particles often take on more complex and irregular shapes if
they are formed by crushing or fragmentation. Some discussion on solid and fluid particles is
given in the next two sections.

2.2.1. Fixed-Shape Solid Particles

Non-spherical solid bodies can be classified as regularly-shaped particles (ellipsoids, cones,
disks, etc.) or irregularly-shaped particles (non-symmetric rough surfaces). Regularly-shaped
solid particles are based on simple cross-sections, such as those in Fig. 2.7. On the other hand,
irregular particles, such as the soot particle of Fig. 2.8, do not possess symmetry which
prevents analytic representations of the surface. In either case, it is convenient to use particle
mass and density to obtain the volume-based diameter (Eq. 1.1).
Regularly-shaped particles can be divided into those without edges and those with edges.
Regular solid shapes without edges commonly include spheroids and ellipsoids. Regular solid
shapes with edges include cones, cylinders and cuboids (also called rectangular parallel-
pipeds or rectangular boxes). Cuboids are made up of connecting orthogonal rectangular
faces and characterized by three perpendicular lengths, e.g., a cuboid with a length ratio of
1:1:1 is simply a cube. Regularly-shaped particles with edges are often associated with
crystallization, which can create natural directions for growth. There are also a wide variety of
water crystal shapes that naturally form in the atmosphere. The well-known dendrites (snow-
flakes) have shapes which are six-armed, nearly-symmetric, highly-intricate and flat (Fig. 2.9a).
These dendrites form at temperatures somewhat below freezing due to a process whereby
liquid vapor condenses on ice crystals as they grow. This accounts for the many frilly surfaces
60
that grow on the crystal. However, there are dozens of other ice crystal shapes which can form
and most are generally simpler in geometry and without porosity. For example, needles and
hexagon disks are common shapes for diameters of 1 mm or less (Fig. 2.9b and 2.9c). Man-
made crystals such as carbon nano-tubes and protein crystals tend to form in disks and needle
shapes. Very large solids such as hail and graupel (several mm in diameter) form by
agglomeration (instead of crystaline growth) and resemble oblate spheroids.
Irregularly shaped particles can stem from random coagulation (Fig. 2.8) but can also result
from break-up of larger solid objects or particles. In particular, grinding and crushing
processes (such as used to produce coal particles and sand grit) create complex shapes that
cannot be defined by a simple geometric surface. Examples of such particles at micron scales
are shown in Fig. 2.10, which can be seen to be qualitatively similar to crushed particles at
much larger scales, e.g. gravel. The degree of deviation from a sphere of such irregular
particle can be characterized by the ratio of surface area to that of a sphere with equivalent
volume as will be discussed in 6.2.4.

2.2.2. Deformable Fluid Particles

While non-spherical solid particles generally owe their shape to formation or break-up as
discussed above, dispersed fluid particles (bubbles and drops away from walls) can deform
based on the interplay of surface tension and the fluid-dynamic stresses on the particle surface.
Thus, the shape of a fluid particle is not permanent and deforms according to the stresses
imparted by the surrounding environment. The surface tension forces will always drive a free
particle toward a spherical shape, whereas initial conditions and/or fluid-dynamic forces are the
primary sources of non-sphericity. Thus zero-gravity (or micro-gravity) conditions with steady
uniform flow will result in a spherical shape once the particle relative velocity becomes
negligible. However, initial conditions, acceleration or gravity can cause significant
deformation owing to the relative velocity and accompanying fluid stresses. This section
focuses on the steady shapes expected for fluid-particles associated with a fixed relative
velocity. The dynamics of particle size and shape caused by acceleration or instabilities will be
discussed later in 2.3.


Weber Number and Terminal Velocity Parameters

The importance of the fluid surface stresses on particle deformation tends to scale with the
dynamic pressure seen by the particle since it is the surface pressure distribution and surface
tension which controls the deformation. This pressure distribution, in turn, can be primarily
characterized the dynamic pressure and the Reynolds number (which controls importance of
viscosity), and secondarily by the viscosity ratio (which controls recirculation). However, the
deformation is also related to the ratio of the continuous-fluid stresses (which cause
deformation and are proportional to the dynamic pressure) to the surface tension stresses
(which resist deformation). This ratio is defined as the Weber number by:


2
f
wd
We

2.39
61
where is the surface tension between the particle and the surrounding fluids (see Table A.1).
The influence of Reynolds number, Weber number and viscosity ratio on both shape and
drag has been investigated theoretically by many researchers with perhaps the most important
results are given by Saito (1913), Taylor & Acrivos (1964), and Moore (1965). Saito
examined the creeping flow limit of Re
p
0 (where inertial terms are neglected), and showed
that a fluid particle of any viscosity ratio and Weber number will remain spherical, consistent
with analysis of Eq. 1.63. This criterion can be qualitatively expressed as
Re
p
1 particles rapidly tend to spherical geometry 2.40
To quantify this relationship, one may suggest that fluid particles are spherical for Re
p
<0.1.
Experimentally observed shapes at such Reynolds numbers are at least qualitatively consistent
with this result. For example, the data of Pan & Acrivos (1968), Grace (1973), Grace et al.
(1976), and Bhaga & Weber (1981) suggest approximate criteria of Re
p
<0.2 for clean bubbles
and Re
p
<0.5 for contaminated bubbles and drops. Taylor & Acrivos extended this condition to
small but finite Re
p
values (see 6.2.6 for details) and found that the deformation is principally
controlled by Weber number, and that the onset of ellipsoidal condition (often defined as
E0.9) occurs at We0.64 for clean gas bubbles.
At very high Reynolds numbers, Moore analyzed gas bubbles and found that onset of
ellipsoidal shapes occurs at We=0.79. These results are also consistent with experiments,
which showed that very large Weber numbers coincide with highly deformed bubbles and
drops. In summary, the degree of deformation at finite Re
p
can be qualitatively expressed as

We

1 particles rapidly tend to spherical geometry
We ~ 1 moderate deviations from a sphere can occur
We 1 large deviations from a sphere can occur
2.41
These trends are consistent with measurements and simulations for a wide range of Reynolds
numbers (from 1 to 10,000). In particular, such observations indicate approximately spherical
conditions are roughly given by

0.9 E 1.0 when Re
p
< 0.2 or We

< 0.8
2.42
A quantitative description of aspect ratio and drag dependence on Re
p
, We, density ratio and
viscosity ratio will be discussed in 6.2.6.
While the instantaneous shape and drag will vary with the instantaneous Weber number
and Reynolds number, it is convenient to consider first the steady-state shapes at terminal
velocity since this condition is well documented. However, the terminal velocity requires
knowledge of the drag coefficient which may not be known since it can be a complex function
of the shape. Therefore, it is convenient to consider dimensionless parameters which do not
require the drag or terminal velocity to be known a priori. To determine these, we note that
the terminal shape and terminal velocity (and thus Re
p,term
and We
term
) are uniquely specified
by the particle volumetric diameter, interface surface tension, and gravity along with dispersed-
phase and continuous-phase fluid densities and viscosities. This allows identification of two
non-dimensional parameters which do not involve velocity: Bond number (Bo) and Morton
number (Mo). The Bond number (also called the Etvs number) is proportional to particle
volume and is given by

2
p f
gd
Bo


2.43
62
This parameter represents the ratio of effective gravitational forces to surface tension forces,
where the former is proportional to the density difference between the phases and the
gravitational acceleration. Note that the effective gravitational force itself does not cause
deformation, e.g., a bubble which is initially released at zero velocity will have little
deformation even if the Bond number is high. However, once such a bubble reaches terminal
velocity, the effective gravitational forces are exactly balanced by drag and the associated
dynamic pressure determines the terminal deformation. In this sense, high Bond numbers are
qualitatively similar to high Weber numbers in terms of Eq. 2.41. The Morton number is also a
function of the density difference, but includes an additional dependence on the continuous-
phase viscosity:

4
f p f
2 3
f
g
Mo


2.44
This parameter has no direct physical interpretation but is conveniently constant for a given set
of continuous-phase and the dispersed-phase fluids.
The above two parameters are convenient in terms of test conditions. For example, it is
straightforward to conduct measurements with Mo fixed while varying Bo (by simply changing
particle size) or to keep Bo fixed and change Mo (by simply changing fluid viscosity). In
comparison, it is more difficult to hold Re
p
fixed while varying We (and vice-versa) since
changing either viscosity or particle velocity will change w. As such the static parameters of
Mo and Bo can be more convenient to characterize terminal steady-state conditions while the
dynamic parameters Re
p
and We are best for describing non-equilibrium (ww
term
) ellipsoidal
conditions. Furthermore, the gravitational effects associated with Bo and Mo help specify the
hydrostatic pressure gradients at terminal conditions which influence shape deformations. It
should be noted that these parameters can be related only at the terminal conditions:

2
term
4
p,term
We Mo
Re Bo
=
2.45
Note that the terminal Weber and Bond numbers can also be related by the terminal drag
coefficient through Eqs. 1.82b, 2.39, and 2.43:
term
D,term
4Bo
We
3C
=
2.46
Therefore, fluid particles in the Newton-based drag regime (constant C
D
) will have a linear
relationship between We
term
and Bo. In the following sub-sections, we will review the well-
documented terminal shapes for drops in gasses (
*
p
1) and for gas bubbles or liquid drops in
liquids (
*
p
1 or
*
p
~1).


Terminal Shapes for Droplets in a Gas

Fluid particles in a gas with a low terminal Weber number are expected to be nearly
spherical. As an example, a water drop in air is spherical (aspect ratios within 2% of unity) for
diameters of 0.5 mm or less, such as the shape shown in Fig. 2.11a. As the drop diameter and
63
corresponding Weber number increase, the drop tends towards an oblate spheroid (the tear-
shaped rain drop is a myth). This spheroidal shape is due to the fact that the internal pressure
is approximately constant but the external pressure around a sphere is highest at the incoming
stagnation point (bottom of the drop) and the lowest along the sides. Oblate shapes (E<1) are
the most common at terminal conditions, though some special conditions can lead to prolate
shapes (Helenbrook & Edwards, 2002). One reason for the preference in oblate shapes has to
do with trajectory stability for these conditions, as will be discussed in the next section. For
water droplets in air, E0.9 (often defined as the onset of significant oblate deformation)
corresponds to a We2 and a diameter of about 1 mm.
At higher Weber numbers, the lower (fore) surface becomes substantially flatter than the
top (aft) surface due to a lower pressure in the rear as compared to the front stagnation point
and hydro-static pressure gradients. As an example, a 5.8 mm water drop with a Weber
number of order ten is shown in Fig. 2.11b. While not obvious from the photograph of Fig.
2.11b, drops undergoing significant deformation can develop a dimple on the upstream (i.e.
lower) surface. This dimple is shown more clearly by the cross-section schematic in Fig. 2.12
for the 6 mm drop. The oblate shape causes the drag coefficient to be greater than that for a
spherical droplet of equivalent volume due to increased frontal area. This increased drag for
deformed particle results in a lower terminal velocity than that expected for a spherical particle
(quantitative relationships will be given in 6.2).
For We1, the dynamic pressure at the leading stagnation point (the center of the bottom
surface) increases further relative to surface tension which can lead to severe exaggeration of
the dimple. In particular, a large drop falling in a gas often forms a bag shape whereby the
middle of the drop is hollowed out yielding a parachute type of effect (see Figs. 2.11c and
2.12). However, this condition is typically short-lived as it is highly unstable and will lead to
break-up of the droplet. Therefore, raindrops are rarely larger than 10 mm since they will
break up if they exceed this diameter. This break-up condition corresponds to a critical Bond
number discussed with more detail in 2.4.1.


Terminal Shapes for Bubbles and Drops in a Liquid

Gas bubbles and immiscible liquid particles in a liquid have very similar shapes at terminal
conditions and thus are often analyzed in a similar fashion. For the common range of Morton
numbers for bubbles and drops (<10
-6
), the evolution of these shapes as We increases is shown
in Fig. 2.13, with specific examples of bubbles shown in Figs. 2.14. At low Weber numbers,
both types of fluid particles will be nearly spherical at low Weber numbers. An example of
this is shown in Fig. 2.14a for small bubbles effervescing in a glass of champagne (naturally
obtained by French researchers). The aspect ratio of these bubbles which are 1 mm or less in
diameter is typically within a few percent of unity, depending on the water contamination.
Interestingly, this is roughly the same size criterion for deformation of water drops in a gas at
terminal velocity since
f
2
term
w is similar for both cases.
At Weber numbers of order of unity, fluid particles in a liquid typically take the shape of an
oblate ellipsoid. However, the aft portion (bottom) of a bubble will tend to be flatter as shown
in Fig. 2.14b. This deviation from an ellipsoidal shape is related to a combination of buoyancy
and flow recirculation effects which yield a separation region with nearly uniform pressure that
causes the rear interface to become nearly horizontal when surface tension is weak.
64
A further increase in the Weber number leads to different shapes depending on the
Reynolds number. For Re
p,term
>100 leads to a spherical-cap shape (Fig. 2.14c). This
geometry is defined by a hemi-spherical shape on the top and a flat surface on the bottom,
which will generally undulate due to unsteadiness in the wake. As shown in Fig. 2.15a, the
wake is generally turbulent and open in the sense that a readily identified recirculation
pattern is no longer applicable. At higher Morton numbers, moderate Reynolds numbers
(about 40<Re
p
<100), the wake can produce an approximately steady recirculation pattern (Fig.
2.15b). In both cases, the approximately flat rear portion indicates that the surface tension is
very weak so that the continuous-phase hydrostatic pressure is simply equal to the internal
bubble pressure on this surface.
For We
term
>18 and Re
p,term
>100, effects of viscosity and surface tension become negligible
and a spherical-cap bubble shape is experimentally observed. The aspect ratio of height to
diameter for this bubble shape is approximately E1/4 independent of further increases in
Weber number or Reynolds number. The upper surface of this shape can be approximated as
spherical with a radius of curvature given by r
I
9d/8. Assuming inviscid flow, negligible
surface tension, and a balance between the continuous-phase and bubble-phase pressure
distribution over this portion of the surface, the drag coefficient and an associated relationship
between Weber and Bond number (using Eq. 2.46) are given by:

D,term
term
C 8 / 3
Bo 2We

for spherical-cap bubbles


2.47a

2.47b
Thus, a spherical-cap shape is consistent with Bo greater than 36. Similarly, the Re
p
>100
criterion is consistent with Mo les than about 10
-4
based on Eq. 2.45. This combination
(Bo>36 and Mo<10
-4
) can be conveniently specified a priori without knowledge of the
terminal velocity to predict the occurrence of a spherical cap bubble based only on bubble size.
However, such bubbles can be difficult to initiate and will tend to break up in unsteady flows.
As the Morton numbers become larger, effects of fluid viscosity become more significant.
Thin disk-like shapes are possible depending on the viscosity ratio and initial conditions and at
very high Weber numbers (>100), the bubble can form a skirt which extends downward from
the spherical cap portion partially enclosing the wake (Fig. 2.16). At Re
p
<40, an oblate
ellipsoidal cap can form. However, it should be noted that the transition Weber number for
various bubble shapes can be affected by contamination.


Influence of Contaminants

As noted in 1.5.2, contaminants can affect the internal flow-field of a fluid particle. These
surface-active contaminants are called surfactants and are typically organic molecules or
sub-micron particles which can adhere to the fluid-particle interface. As an example of the
influence of surfactants, detergent mixed into water can create bubbles that are highly stable.
The presence of surfactants can reduce surface tension and, in high concentrations, eventually
renders the interface immobile. As such, surfactants can increase drag from a condition
governed by recirculation (Eq. 1.64) to one governed by a no-slip interface (Eq. 1.33). This, in
turn, can affect the particle terminal velocity, wake and shape.
Surfactants can delay the transition from spherical to ellipsoidal bubbles for a given
fluid particle diameter because they result in a decreased terminal velocity and thus reduced
65
Weber numbers. For example, this transition occurs approximately at diameters of 1.1 mm in
clean (uncontaminated) water but occurs at diameters of 1.7 mm for tap water (Clift et al. 1978
and Wu & Gharib, 2002). In addition, the reduction in surface tension over the surface due to
contamination can lead to ellipsoidal shapes that are comparatively more flattened. Since the
velocity change is not as significant with liquid drops in a gas (since
p

f
), the importance of
surfactants on spherical drops is generally negligible. Surfactants are also not important for
highly deformed particles, since surface tension and flow recirculation plays a weaker role in
their shapes (Clift et al. 1978).
Predicting the conditions which will cause a bubble to be contaminated is difficult. If the
fluid particle is initially clean, the surfactants generally attach themselves to the interface and
are swept to the aft region where their concentration can build up (Bel Fdhila & Duineveld
1996; Cuenot et al. 1997). Therefore, the effect of contaminants can be path- and time-
dependent since the fluid particle may be initially clean. The gradients of surfactant
concentration on a fluid particle surface can lead to significant gradients in the interfacial
surface tension. The resulting surface tension gradients lead to tangential stresses (known as
Marangoni stresses) that oppose surface

convection so that recirculation is inhibited. In order
to account for this phenomenon, the stagnant cap model was developed by Griffith (1962),
which assumes that the absorbed surfactant molecules are dragged towards the rear of the fluid
particle by the adjacent liquid (Griffith, 1962). In the model, the interface is separated into two
parts, a mobile portion which represents the clean surface, and an immobile (no-slip) interface
that represent the contaminated surface.
The level of the surfactant concentration on the fluid particle interface can be broadly
classified into three regimes: clean, partially-contaminated, and contaminated. The clean
condition allows a fully mobile interface with full internal recirculation as in Fig. 1.28 and
corresponds to
clean
=180 in Fig. 2.17. As the concentration on the surface increases, parts of
the fluid particle surface become immobile and the recirculation is limited to a fraction of the
surface (0<
clean
<180). This partially-contaminated condition is generally associated with
an increased drag (e.g., reduced terminal velocity). The contaminated or fully-
contaminated condition corresponds to
clean
=0. In this case, the particle surface is saturated
with surfactants so that it is immobile, i.e. the surface effectively reverts to a no-slip condition
(and thus, terminal velocity) equivalent to that of a solid particle.
These three regimes can be related to the concentration of surfactants in the liquid. It is
especially difficult to remove such impurities from water. Furthermore gas-water interfaces
are especially susceptible to the influence of surfactants particularly for small surface-to-
volume ratios. As such, it is difficult to achieve clean conditions for water for bubble
diameters less than 1 mm. As an example, Fig. 2.18 identifies the partially-contaminated
regime for various surfactants and multiphase systems. Roughly, concentration levels below
10
-4
g/L can be considered clean (or sometimes called pure) for most fluid particles, while
concentrations above 10
-2
g/L can be considered fully-contaminated. However, the boundaries
between the regimes are a function of bubble diameter and surfactant type. The characterizing
properties of surfactants include solubility, surface pressure, cation concentration, organic
concentration, molecular weight, etc. As a result, generalizations based only on overall mass
concentration are difficult since different surfactants can have different levels of effectiveness.
For example, consider a 1.6 mm bubble in water with a surfactant concentration of 10
-3
g/L of
either sodium dodecyl sulfate or the more effective Triton X-100 as shown in Fig. 2.18. The
former yields a clean bubble whereas the latter yields a fully-contaminated bubble.
66
Water itself is often classified as tap water, distilled water or hyper-clean water, each of
which can be assigned an approximate surfactant concentration level as shown in Table 2.2.
Distilled water is obtained by heating to form steam which is condensed elsewhere, leaving
behind impurities which have higher boiling points than water. However, many organic
substances are easily dissolved in water and have similar or lower boiling point then water,
making them notoriously difficult to remove. In order to further purify water, various filtration
systems (deionization, reverse osmosis, etc.) and UV light are employed to remove surfactants
and produce hyper-clean water. Even with these techniques, a clean condition in terms of
bubble surface mobilization can only be expected for bubbles of diameters greater than 0.7 mm
for hyper-clean water. As a result, few, if any, researchers have obtained clean conditions for
bubbles with diameters less than 0.5 mm in water solutions.


Comparing Deformation in Liquids vs. Gases

It is interesting to compare the shape regimes for
*
p
1 and
*
p
1. The theoretical and
experimental qualitative boundaries for the different shape categories are shown in Fig. 2.19.
Here it can be seen that the approximate criterion for onset of ellipsoidal shape is We>2 for
water drops in air and We>0.8 for uncontaminated bubbles in water. Similarly, comparing Fig.
2.11b with 2.14b indicates that although the drop has a larger Weber number and Bond number,
the bubble has a larger deformation. This indicates that bubbles are more readily deformed for
a given dynamic pressure. This difference can be attributed to the weak internal circulation for
a drop in a gas compared to that for a clean bubble. As a result, the bubble has higher
pressures in the aft portion and lower pressures at the sides for a given dynamic pressure as
compared to that for a drop at the same Re
p
. This effect can be explained by looking at Fig.
1.26. A clean bubble (which does not have flow separation) yields a pressure distribution close
to that of inviscid potential flow while the pressure distribution for a drop is close to that for a
sub-critical solid sphere. The more exaggerated pressure variation found for the bubble
translates into higher deformations as compared to those for a drop at the same Weber number.
This also explains why contaminated bubbles (which can have flow separation) exhibit less
deformation than clean bubbles (Fig. 2.19).
The differences in the surface pressure distribution coupled with hydrostatic gradients also
result in different shape asymmetry with respect to the upstream (fore-portion) and
downstream (aft-portion) of the fluid particle. In particular, a drop in a gas yields a flatter
portion on the upstream portion (bottom surface) while a bubble in a liquid yields a flatter
portion near the aft-portion (though this happens to also be on the bottom surface). This
difference will be discussed in 2.4.1 to explain why bubble shapes are more stable than drop
shapes. In particular, droplets will tend to break up at Weber numbers on the order of 10 while
the spherical cap bubble shape can be stable at Weber numbers on the order of 100 or more.


2.3. Dynamics of Particle Orientation, Size and Shape

Various phenomena can lead to time-dependent changes in particle orientation, size and
shape. Even in the simple case of a particle accelerating from rest to its terminal velocity, there
can be significant changes in the particle orientation and shape. For example, a solid particle
67
may experience changes in its orientation if it is non-spherical and a fluid particle may deform
as its relative velocity increases. Continuous-phase flow gradients (especially turbulence) can
also cause orientation dynamics, deformations, and even particle break-up. Continuous-phase
pressure fluctuations can also cause a gas bubble to undergo significant size oscillations, which
can lead to non-spherical shapes and eventually cause break-up at high amplitudes. In the
following, the controlling physics of these phenomena are overviewed. Techniques to describe
these interactions for point-force particles will be discussed in Chapter 7, whereas Chapter 8
will discuss resolved-surface methods which automatically capture such effects for individual
particles (without the need for analytical or empirical models).

2.3.1. Orientation Dynamics in Free-Fall

When a particle is spherical, its trajectory in free-fall is generally rectilinear and
parallel to the direction of gravity. However, non-spherical particles can undergo various types
of secondary motion with respect to their shape orientation (e.g., the particle axis of
symmetry may oscillate about the axis of gravity) and/or with respect to their path orientation
(e.g., the particle trajectory may oscillate about a vertical line). The capability for such
motions is related to a particles wake, which is determined by its shape and Reynolds number
for solid particles. In addition, fluid particles can have flow recirculation and dynamic shapes
which lead to further complexity. The orientation dynamics of solid and fluid particles are
discussed in the following sections for the simple condition of free-fall in a quiescent
continuous-flow. It should be noted that three second-order tensors can be used to describe
the orientation dynamics associated with surface forces and torque for a general particle shape
(Clift et al. 1978). However, we will forgo such complexity until Chapter 6, and instead
discuss here the dynamics for some simple shapes and conditions.


Solid Particles

The most commonly studied non-spherical particle shape is the spheroid. Many other non-
spherical particles can be roughly approximated as spheroids, so that this shape is an important
baseline condition. One can consider the dynamics of these particles for the Stokes flow limit,
intermediate Reynolds numbers, and the Newton-based drag limit. At creeping flow
conditions (Re
p
0), a constant-density spheroidal particle with a moderate aspect ratio will
experience negligible torque if the axis of symmetry is located either exactly parallel or
perpendicular to the direction of gravity. However, its preferred orientation of falling will be
with the longest dimension(s) perpendicular to the gravity direction, i.e. falling broadside,
due to stability considerations. To see this, consider a needle particle falling in the length-
wise or vertical direction where the axis of symmetry parallel to gravity. A slight angular
perturbation from this length-wise orientation will yield a torque which tends to exaggerate this
perturbation (similar to an airfoil at an angle of attack) so that the vertical orientation is
dynamically unstable (Galdi et al. 2002). In fact, this instability is the reason why an arrow,
rocket or bullet (which can be approximated as prolate spheroids) must be stabilized with fins
or a spin velocity in order to retain a streamwise orientation with lower frontal area to allow
less drag and higher speeds. Now consider the same needle falling broadside. An angular
68
perturbation from this orientation will yield a torque which tends to counter this perturbation,
thus returning the particle to its original orientation. This preferred orientation is consistent
with that found for oblate ellipsoidal fluid particles in free fall (Figs. 2.11b and 2.14b), even
though it has higher drag.
In some exceptions, a particle at low Reynolds numbers will not necessarily prefer the
broadside orientation. This includes the presence of neighboring particles (or nearby-walls)
which oppose a change in orientation through three-way coupling effects such that the vertical
orientation can be preferred. Also, non-Newtonian viscous properties for the continuous-phase
can cause the pointed-down orientation to be stable (Galdi et al. 2002). In addition, very small
particles for which Kn
p
is no longer small can have their orientations perturbed by random
molecular collisions, a process referred to as Brownian motion (3.6). Finally, irregularly-
shaped particles generally do not possess a torque-free equilibrium orientation. Instead, they
are typically unstable for all orientations and thus rotate and tumble in free-fall.
At moderate Reynolds numbers (about Re
p
<100), the preferred orientations and
dynamics of a solid particle in free-fall are similar to those for creeping flow conditions. For
example, ellipsoids, cylinders, and disks will tend to free-fall with a broadside orientation (e.g.,
Fig. 2.20) and irregularly-shaped particles will tumble but have a path-averaged orientation
which is approximately broadside (Gogus et al. 2001).
At higher Reynolds numbers (about Re
p
>100), non-spherical particles will typically fall
in a broadside condition but with significant secondary motion due to the presence of strong
unsteadiness in their wakes. For spheroids with significant eccentricity (E<0.5 or E>2),
terminal trajectories may exhibit substantial periodic movement in which the minor axis rotates
to trace out a cone or tumbles end-over-end motion. A well-studied particle shape for this
secondary motion is that of a circular disk (a very low aspect ratio oblate ellipsoid). In free-fall
at high Reynolds numbers, the disk motion can be complex and depends on the particles
moment of inertia about the broadside axis. In particular, Willmarth (1964) observed four
types of motion for disks (Fig. 2.21): a) steady-motion falling in the broadside orientation, b)
periodic-motion with back and forth oscillations, c) chaotic motion which continually
transitions between periodic and tumbling motion, and d) tumbling-motion with end-over-
end flips. For the periodic-motion, any initial disturbance results in a continuous back and
forth motion whereby the particle slides at an angle towards one direction, flattens out, and
then repeats this motion in the other direction (Fig. 2.21b). For the intermediate unsteady
condition known as chaotic-motion, the periodic motion is enhanced to larger amplitudes until
tumbling sets in, but this tumbling is not preserved and instead it abruptly transitions back to
the periodic mode after some time, whereby the process can be repeated (Fig. 2.21c). Thus,
this motion regime will also have a mean horizontal glide in between that of periodic-motion
and tumble-motion. The tumble-motion results in trajectories where the slide motion becomes
so exaggerated that it points the outward edge upwards and the inner edge downward,
reversing the process (Fig. 2.21d). The horizontal drift associated with the chaotic and
tumbling motion can be quite substantial and because of this Stringham et al. (1969) classified
the regime as glide-tumble. Many of these motions can be observed in nature through the
fall of leaves and seeds to the ground.
In order to characterize the different regimes of the disk trajectories, a non-dimensional
moment of inertia can be defined (Willmarth, 1964) based on disk geometry and density ratio
69

*
p p *
5
f
d
d 64d

|
I
I 2.48
In this relationship, d

is the disk diameter and d



is the disk thickness based on notation of Fig.
2.7a. Based on experiments, a regime map was constructed as a function of I* and Re
p
as
shown in Fig. 2.22. This map shows that steady-motion is found for all particles with
Reynolds numbers of 100 or less. This is consistent with the stable trajectory expected for
broadside ellipsoids at low Reynolds numbers. At higher Reynolds numbers, the unsteadiness
in the wake prevents stable trajectories and instead tumbling motion is observed for particles
with Re
p
>100 and I*>0.04. For high I*, the tumbling motion is more periodic with little
horizontal drift. Intermediate inertias at these Reynolds numbers tend to cause chaotic motion,
while lower inertias give rise to periodic motion without the disk flipping over.
In comparisons to disks, other non-spherical shapes tend to have less secondary motion.
In particular, Stringham et al. (1969) noted that oblate spheroids with E>0.5 are less prone to
orientation dynamics as compared to disks (consistent with having lower non-dimensional
inertias than disks) and will only have minor periodic oscillations for Re
p
<10
4
. Cylinders and
prolate spheroids tend to fall broadside at Re
p
>100 but with secondary motions which are a
function of d / d
|
and
*
p
(Clift et al. 1978).


Fluid Particles

As discussed in 2.2, droplets free-falling in a gas will tend to have spherical shapes at
low Bond numbers and Reynolds numbers. This spherical symmetry leads to torque-free
conditions, such that rectilinear motion is obtained for terminal velocity conditions. As
deformation increases yielding oblate spheroids or more distorted shapes, the trajectory will
generally retain a steady rectilinear condition for Re
p
<200. However, higher Reynolds
numbers will yield an unsteady wake accompanied by lateral periodic trajectory oscillations
(deVries et al. 2001). These motions are qualitatively consistent with those for ellipsoidal
solid particles, i.e., they generally fall in to the rocking regime (equivalent to the periodic
motion defined for disks) for moderate aspect ratios. However, the tumbling regime noted for
solid particles with low aspect ratios (about 0.3 or less) is not generally observed for highly
deformed fluid particles and instead nearly rectilinear motion is more common.
For bubbles and drops in a liquid, spherical shapes again give rise to a rectilinear path in
quiescent flow. The appearance of oblate ellipsoidal shapes is often followed a secondary
motion phenomena with respect to particle shape orientation and pathlines, e.g. the rocking
motion shown in Fig. 2.23. The minimum diameter for which secondary motion occurs is
sometime referred to as the oscillation diameter (d
osc
) and is on the order of 1.3-1.7 mm for air
in water. The onset of secondary motion is typically categorized with zig-zag trajectories
(sinusoidal motion within a plane) or spiral trajectories (3-D helicoidal motion). The zig-zag
motion is closely contained within a single vertical plane but the spiral motion yields a
trajectory which approximately circumscribes a vertical elliptical cylinder surface. Examples
of trajectories are shown for a bubble in water in Fig. 2.24 where the lateral amplitude
deviation is denoted by A
osc
. In tap water, the motion is typically helicoidal for 1.3 mm d
2.0 mm, zig-zag followed by helicoidal for 2.0 mm d 3.6 mm, and just zig-zag for 3.6 mm
70
d 6 mm (Clift et al. 1978; Ellingsen and Risso, 2001). Larger bubbles between 7 mm and
15 mm in diameter result in approximately rectilinear motion with rocking (Aybers & Tapucu
1969). Finally, spherical cap bubbles (d > 15 mm) yield a return to nearly rectilinear motion,
though weak rocking can still occur for turbulent wake conditions (Fan & Tsuchiya,

1993).
Measurements of perturbation amplitude from the mean path (A
osc
) for a variety of bubble
diameters are shown in Fig. 2.25 for air in water. While there is significant variation, the
amplitude in quiescent conditions is generalized herein as


osc osc osc
3.5 mm d d 6.5 mm 0 mm < < A for else A
2.49
Thus the lateral zig-zag motion can be approximated using a frequency (f
osc
) as:


( )
osc osc osc o
y sin 2 t t

A f
2.50
The lateral motion can be extended to approximate helicoidal motion by superimposing a
similar x
osc
expression which is /2 out of phase with y
osc
. For air bubbles between 1 and 6
mm in diameter, the zig-zag frequency (f
osc
) in water is about 7 Hz, while the spiraling
frequency is about 5 Hz (Mercier & Forslund, 1973).
While the above may serve as general guidelines, deVries et al. (2001) and Wu & Gharib
(2002) noted that water purity and release mechanism can affect the secondary motion in terms
of A
oscl
, f
osc
and especially d
osc
. In particular, the oscillation diameter for secondary motion
may vary substantially as shown in Table 2.3 depending of fluid type, contamination level and
temperature. In general, cleaner conditions and lower Morton numbers (e.g., for warmer
conditions) allow for larger oscillation diameters, but there are additional influences of
viscosity ratio, Weber number and Reynolds number which influence this transition. Although
not included in the above equations, the trajectory oscillations can result in a net horizontal
drift away from the vertical axis, but this is generally limited to a small fraction of A
osc
over a
single oscillation (Mercier & Forslund, 1973).
Non-rectilinear motion is generally a result of the wake dynamics but it is not necessarily
caused by unsteady vortex shedding (as is the case for solid ellipsoidal particles). This was
confirmed with experiments using shadowgraphs of the bubble trajectory during spiraling
motion as shown in Fig. 2.26a where the flow is clearly attached with a laminar wake. It can
also be seen that the wake is double-threaded since it consists of two counter-rotating vortex
filaments due to the ellipsoidal shape. The curvature of these filaments leads to a wake-
induced lift on the bubble, which causes an instantaneous sideways force. This movement, in
turn, leads to asymmetric production of vorticity in the filament during its lateral motion such
that the side force causes a change in direction, resulting in zig-zag or spiral behavior (deVries,
et al. 2001). For Re
p
greater than about 1000 (or less at higher surfactant levels), a
conventional unsteady wake shedding phenomenon will occur and the trajectory oscillation
amplitudes tend to increase, though some evidence of the two counter-rotating filaments can
still be seen near the bubble surface (Fig. 2.26b). In most of these cases, the bubble shape does
not change (or changes very little), whereas the orientation changes significantly.
The oscillation frequency and amplitude of bubble trajectories in quiescent conditions will
similarly occur when the surrounding liquid is moving so long as the relative velocity is
considered. The pervasiveness of this motion even in turbulent flow conditions was observed
in the experiments of Ford & Loth

(1998) where the motion relative to the instantaneous
continuous-fluid flow exhibited similar amplitude and oscillation frequencies as seen for
71
quiescent conditions (Fig. 2.27). This indicates that the correlation between the bubble wake
and its dynamics is approximately preserved in tap water even when subjected to a
substantially unsteady and non-uniform surrounding.

2.3.2. Orientation Dynamics in Simple Shear Flow

In the following, solid and fluid particles are considered with respect to their motion and
orientation in a simple shear flow defined by Eq. A.34.

Solid Particles

Solid particles in a linear shear flow (Fig. A.3a) can undergo complex dynamics with
respect to shape orientation. For creeping flow conditions, Jeffrey (1922) showed analytically
that this condition causes the axis of symmetry of a spheroid to rotate in a periodic orbit for a
given aspect ratio (Fig. 2.28). The time-scale for this motion is given by

shear
2 1
E
E


= +



2.51
This behavior has been confirmed experimentally for ellipsoidal particles by Gauthier et al.
(1998). These orbits are also a function of the initial condition such that the trajectory never
theoretically forgets its original orientation and maintains a specific orbit. However, random
forcing due to Brownian motion or turbulence can result in transition between various orbits.
Similar behavior has also been observed for approximately ellipsoidal particles such as
cylinders and disks, though with somewhat modified rotation rates. If a gravitational force is
also important to the motion (i.e. terminal velocities on the order of tip rotation velocities),
then the dynamics can be either stabilized or enhanced depending on the relationship between
the shear forces and body forces stemming from the particles orientation to gravity (Broday et
al. 1998). In addition, particles without fore-and-aft symmetry (such as Figs. 2.8 & 2.10) will
generally have a chaotic tumbling motion while undergoing shear since the axis of drag will
not necessarily be along the center of mass, resulting in a net torque.


Fluid Particles

Under continuous-fluid shear with negligible gravity (especially for immiscible drops in a
liquid where
*
p
~1), the fluid particle will stretch. The strength of the shear (Eq. A.34) can be
compared to the viscous forces by defining a shear-based Reynolds number

2
f shear
f
d
Re


2.52
For linear shear applied to an initially spherical particle stretching will produce an ellipsoidal
shape with the long-dimension oriented in the direction of shear (Fig. 2.29). Cox (1969) found
that the particle shape for Re

0 is an ellipsoid where the three diameters are give by the


ratios E
1/2
:1:E
-1/2
and where the longest and shortest diameters are in the shear plane
72
(perpendicular to the vorticity vector). He also noted that deformation increases proportional
to the shear Capillary number (Ca

) which can be related to a deformation time-scale (

) as:

f
shear shear
d
Ca


=



2.53
The shear Capillary number is an indicator of the ratio of shear stresses to surface tension
stresses. For long times (t

) and small deformations (Ca

1), the equilibrium orientation


angle for the fluid particle is theoretically given by:

*
p 1
19 Ca
1
tan
4 2 40

= +




2.54
This angle approaches 45
o
for weak shear and approaches 90
o
for high shear. An example of a
stretched fluid particle is shown in Fig. 2.29 where the theoretical angle is reasonable.
However, higher deformations can lead to break-up (to be discussed later).

2.3.3. Size Oscillations for Bubbles

Fluid particles can also deform radially with changes to their overall volume if the particle
material is significantly compressible. In particular, bubbles can undergo significant size
oscillations. To describe this phenomenon, one must relate the change in the bubble pressure
to change in the bubble volume. This involves the equation of state and the dynamics of the
surrounding liquid. For weak volumetric oscillations, the bubble remains nearly spherical and
the volume change can be simply related to the change in diameter. The analysis for a
spherical compressible bubble surrounded by an incompressible liquid was first obtained by
Lord Raleigh in 1917. He assumed a uniform internal pressure and neglected surface tension,
gravity, and viscosity effects. The effects of surface tension and viscosity were added by
Plesset in 1949 resulting in the well-known Raleigh-Plesset equation which is overviewed
below and further details are given by Brennen (1995).
Consider a gas bubble of radius r
p
, of density
g
, and pressure P
g
that is surrounded by an
incompressible liquid of density
l
and pressure P. Assume that all the motion is radial with a
spherical coordinate system emanating from the bubble center (r=0) as shown in Fig. 2.30.
The liquid velocity is only radial U
r
and is equal to the interface velocity at r=r
p
:
( ) ( )
p r p r p
r U r , t V r , t = `
2.55
Considering any spherical surface in the surrounding liquid region (r>r
p
), conservation of mass
necessitates that the liquid flux though any such surface be conserved if the liquid is
incompressible and no mass transfer occurs, i.e.

2 2
r p p
r U r r const. = = `
2.56
The conservation of momentum from the incompressible Navier-Stokes equation with only
radial variations (Eq. A.58a and A.59) is

2 r r r
r r 2
U U U 1 P 1
u r 2U
t r r r r r

+ = +



l
l l

2.57
73
Substituting, the expression for U
r
from Eq. 2.56 into Eq. 2.57 and solving for the radial
pressure gradient outside of the bubble (Batchelor, 1967) yields
( ) ( )
2
2 2
p p p p 2 5
1 P 1 2
r r r r
r r t r

=

` `
l

2.58
Note that the viscous terms have conveniently dropped out of this expression and that the RHS
consists of only time-derivative terms. Integrating this momentum equation to the far-field
(r) relates the surface liquid pressure (P
surf
) to the far-field liquid pressure (P

) as

( )
( )
2
2
p p
2 surf
p p 4
r r
P P 1
r r
r t 2r


=

`
`
l

2.59
To solve this ODE one must obtain a state relationship between the liquid pressure and the
bubble radius. This can be obtained by examining the force balance across the bubble interface
assuming that the bubble pressure is uniform throughout and defined as P
g
. The liquid pressure
on the outside surface of the bubble can be defined as the interface pressure which is uniform if
the bubble is stationary (w=0). The pressure difference across the interface can be obtained
from Eq. 1.58a assuming negligible internal viscosity and then related to
p
r` by Eq. 2.56 as:

surf
r
p
surf p p p
P P
4 U 2 2
2 r
r r r r


= = +



`
g
l l
l l l l l

2.60
Combining this result with Eq. 2.59 relates bubble pressure to the interface acceleration (
p
r `` ) as:
( )
2
p
p p p
p p
P P 4 r
3 2
r r r
2 r r

= + + +

`
`` `
g l
l l

2.61
Thus, surface tension dictates the pressure difference for the steady-state equilibrium condition.
Further assuming that the bubble pressure is based on an ideal gas (Eq. A.21) yields:

,eq ,eq ,eq surf p,eq
P T P 2 / r = = +
g g g g
R
2.62
In this equation, r
p,eq
is the equilibrium bubble radius based on
p
r 0 = ` and R
g
is the gas constant
for the bubble fluid.
In order to close the ODE, we must relate the bubble radius to the bubble pressure by an
equation of state for the gas (based on conservation of energy) and a density-volume
relationship (based on conservation of mass). The corresponding relationship for pressure,
density, and radius ratios can be given in a generalized form as:

p,eq
,eq ,eq p
3
P r
P r

= =


g g
g g
g g
K K
2.63
In this equation, K
g
is the gas compressibility exponent where K
g
= for isentropic conditions
(Eq. A.27) and K
g
=1 for isothermal conditions (T
g
=T
l
). The isentropic limit is reasonable for
fast dynamics and negligible heat transfer, i.e. rapidly oscillating large bubbles with a high
volume to surface ratio. The isothermal condition is reasonable when heat transfer dominates,
i.e. slow dynamics of small bubbles so that the temperature of the surrounding liquid dictates
the bubble temperature.
Substituting Eq. 2.63 into Eq. 2.61 yields the generalized Raleigh-Plesset ODE:
74

( )
3
2
,eq p,eq p
p p p
p p p
P r 4 r
P 2 3
r r r
r r 2 r

= + +




`
`` `
g
K
g l
l l l
2.64
This equation can be solved numerically for the general case once the initial conditions and
boundary conditions are specified. An important characteristic of this equation is the natural
frequency of the system, f
nat
. This can be obtained by linearizing the above ODE and
considering small perturbations about the equilibrium condition as discussed by Brennen
(1995). The resulting natural frequency is given by:

2
nat 2 2
p,eq p,eq
1
3 P 2(3 1)
4 r r

= +




g g
l
f K K
2.65
For bubble diameters on the order of 100 m or more, surface tension can generally be
neglected. Furthermore, such bubble diameters will have natural frequencies much greater
than that of the bubble thermal response (Eq. 1.93b) so that an isentropic (vs. isothermal)
description is appropriate. These conditions yield the Minnaert (1933) frequency:

( )
2 2
nat eq
3 P d

=
l
f 2.66
For example, a 3 mm diameter bubbles at small depths will yield a Minnaert frequency of
about 2 kHz. For smaller bubbles, viscosity, surface tension and even vapor pressure effects
become important so that frequencies for bubbles of radius 1-100 m are limited to about 4
kHz. Note that these frequencies are much higher than the 5-7 Hz associated with secondary
motion of bubble trajectories (2.3.1), so that the two mechanisms are uncorrelated.
Consider a bubble subjected to a sinusoidal pressure oscillation of low frequency (ff
nat
)
such that it has an isothermal response (K
g
=1). If the amplitudes are also weak (1), the
imposed pressure can be expressed as:
[ ]
eq
P P 1 sin(2 t) = + f
2.67
In this case, the liquid radial momentum is negligible so that the RHS of Eq. 2.64 is neglected
and internal bubble pressure and radius are given by the quasi-steady values:

[ ]
( )
,eq
1/3
p p,eq
P P 1 sin(2 t)
r r 1 sin 2 t

= +
= +

g g
f
f


2.68a

2.68b
Such a result is shown in Fig. 2.31 for f = f
nat
/20 and 1/ 20 = , whereby the internal pressure is
at a maximum when the radius is at a minimum, and the latter has non-dimensional oscillation
amplitudes which are smaller than that for the pressure. One may also consider a bubble which
is initially set in motion based on Eq. 2.68 within a finite liquid volume. If the surrounding
fluid is incompressible and inviscid, there will be no damping or radiation loss and the radial
motion will be self-sustaining, i.e. the dynamics may continue indefinitely. Incorporation of a
finite viscosity and/or finite liquid compressibility will convert the kinetic energy of the system
into unrecoverable thermal energy and thus damp out the behavior while incorporation of a far-
field pressure will serve as a sink for the radiated energy.
75
An interesting and different result occurs when faster bubble oscillations occur with
frequencies that are on the order of f
nat
. In this case, the dynamics governed by Eq. 2.64 must
be solved numerically if the amplitudes are no longer asymptotically small. For a continuous
external forcing given by finite-amplitude sinusoidal far-field pressure oscillation, the
corresponding radial dynamics are quite different than that given by the quasi-steady case.
For example, Fig. 2.32 shows pressure and radial oscillations for inviscid conditions (generally
reasonable for air-water bubbles greater than 30 m in diameter) for bubbles excited by
frequencies on the order of the natural frequency by an ambient pressure oscillation with an
amplitude of 0.35 = . In these cases, one may observe a dramatically larger and more
complex bubble radius sensitivity to pressure fluctuations as compared to the quasi-steady case.
Indeed, sono-luminescence and use of micro-bubble as a contrast-agent (1.2.2 and 1.2.3)
take advantage of this sensitivity. The condition with f =0.8 f
nat
yields very sharp decreases in
the bubble radius during some of the contractions. This is related to the large inertia of the
surrounding liquid which causes the surrounding fluid motion to continue contracting even
though the bubble pressure has started to rise above the equilibrium condition. As a result, the
bubble continues to contract significantly below the minimum radius given by the quasi-steady
oscillation and this contraction builds up a higher gas pressure until the inward liquid motion is
eliminated. After this point, there is a rapid outward expansion. Near the maximum bubble
diameter, there is also a tendency for the liquid to continue moving outward despite the local
bubble pressure dropping below the far-field condition. The condition with f=1.8f
nat
yields
radius variations with less amplitude but which are more non-linear.
Since forcing near the natural frequency causes much larger amplitude oscillations than
that of the non-resonant conditions, accurate analysis of this condition generally requires
inclusion of both liquid compressibility and more detailed internal gas dynamics. To
incorporate liquid compressibility with a finite speed of sound in the liquid (a
l
) , Keller &
Kolodner (1956) used a spherical unsteady counterpart of Eq. A.45 for constant speed of sound
to obtain a linearized wave-equation for the liquid velocity potential outside of the bubble:

2
2
2 2
1
a t

=

l

2.69
If the velocity potential is known from this ODE, the liquid pressure distribution can be
determined with the spherical Bernoulli equation given the initial pressure (P
init
) at t=0:
( ) ( )
2
1
init
2
P P / t / r

= +

l

2.70
This equation can be applied at the interface with Eq. 2.60 to relate the liquid velocity potential
to the bubble radius. If viscosity and surface tension are neglected (reasonable if the bubble is
not small), the resulting boundary conditions are given by

surf p
p p
P P r=r
r / r r=r
=
= `
g
at
at

2.71a

2.71b
These boundary conditions, coupled with the bubble state relation (Eq. 2.63) and the governing
ODE (Eq. 2.69) can be used to solve the bubble dynamics given initial (equilibrium) conditions.
However, numerical methods are needed as this system is non-linear. Assuming that the radius
oscillations are small, Keller & Kolodner linearized the differential equation to give an
analytical solution:
76

init init init p
1 2 2 2
p,eq p,eq p,eq
3 P 3 P 3 P r
1 c exp cos 1 t c
r 2 a r r a




= + +







g g g
l l l l l
K K K

2.72
In this equation, c
1
and c
2
are constants based on the initial amplitude and phase angle of the
bubble oscillations. This result is an exponentially damped sinusoidal oscillation which shows
that the oscillations will reduce in magnitude due to finite speed of sound. Note that as a
l

(tending to an incompressible surrounding liquid), the damping effect is eliminated and an
impulsively started oscillation will continue unabated at the natural frequency given by Eq.
2.65.
A condition of a bubble oscillation that particularly requires compressibility condition is
that of a step-function forcing with high amplitude. An underwater explosion can be
characterized by this condition and will cause strong expansion of the combustion gas products,
which can then lead to oscillating bubble dynamics that decay due to water compressibility.
Figure 2.33 shows experimental data of the gas cloud radius from a 100 meter deep explosion
of 1 kg of tetryl. In addition, the analytical prediction based on the linearized compressibility
correction (given by Eq. 2.72) is shown along with numerical solution for non-linear
compressibility and dynamics (given by solution of Eqs. 2.64, 2.69, 2.70 and 2.71). In each
case, an equilibrium radius and initial amplitude were selected to fit the experimental data.
The dynamics and decay resulting from acoustic energy dissipated outside of the bubble are
accurately predicted at short times (first few cycles) using non-linear theory, but the long-time
dynamics are reasonably linearized compressibility. Note that the equilibrium radius
physically changed during the oscillations due to continued product release so that a larger r
p,eq

was used for the long-time behavior than was needed to describe the short-time behavior. This
result indicates that the linearized solution is reasonable for small to moderate amplitude
variations, but that complex behavior associated with very high amplitudes and the spikes in
bubble collapse can only be captured by the non-linear analysis.
To extend the liquid compressibility and finite relative velocity, Prosperetti (1987) included
a general perturbation analysis to the Rayleigh-Plesset equation, which he combined with the
pressure jump across the interface (Eq. 2.60) and the unhindered pressure (Eq. 1.9e) to yield:


( )
2
1 2 3 2
p p p p p p p p p p
4
p p
4 3 1 1 2
r r r r r 6r r r 2r P P r w
2 a r r


+ + + =



`` ` ``` ` `` ` `
l
g l l
l l

2.73
The viscous effect is the same as that appearing in Eq. 2.64, while the relative velocity effect is
incorporated through a dynamic pressure (
l
w
2
) and is significant when the relative velocity
is on the order of the interface speeds (
p
r` ). Liquid compressibility (which changes this to a 3
rd

order ODE ) will also be important if the interface speeds are no longer negligible to the liquid
speed of sound. In such cases, the internal Mach number is likely to be significant such that
there may be internal radial pressure waves (e.g. Eq. 2.63 may not be sufficient to describe gas
pressure throughout). For strong oscillations, heat transfer and non-spherical motion effects
can become important (Feng & Leal, 1997). This is particularly true for sono-luminescence.
For sono-luminescence (Fig. 1.6a), the bubble diameter is very small (a few microns) and
subjected to ultrasonic excitation near the natural frequency. In this case, the assumption of
uniform properties inside the bubble breaks down because of the internal shock wave.
Therefore, this condition requires a description of the radial variation of the gas flow within the
bubble using the Euler equations of motion, e.g., numerical solution of Eq. A.44 written in
77
spherical coordinates (Kameda & Matsumoto, 1996). Secondly, assumptions of an ideal gas
and constant molecular weight can break down such that the use of variable gas constants with
species equations may be needed to provide good comparison with experiment (Feng & Leal,
1997). Moreover, if the interface velocities are no longer small compared to the liquid sound
speed, a more physical description of liquid compressibility, such as the Tait equation given by
Eq. A.24, may be needed. Finally, the surrounding fluid bulk viscosity and thermal
conductivity can also be significant in damping strong oscillations for small bubbles. For
example, viscous and thermal effects were found to be critical in determining the dynamics of
a 3 micron bubble driven by pressure oscillations with 0.3 = (Moshaii et al. 2004).

2.3.4. Non-Spherical Shape Oscillations for Fluid Particles

This section considers two types of shape oscillations: 1) harmonic modes in response to
external excitations, and 2) tangential oscillations which are self-generated. The first type of
shape oscillation modes are important as they can relate to bubble break-up mechanisms, while
the second topic is associated with locomotion (swimming) of micro-organisms, which can be
thought of as self-propelling particles.


Shape Oscillations due to External Perturbations

When a fluid particle is exposed to pressure fluctuations or unsteady conditions, it can
undergo shape oscillations. The simplest condition is the first mode (m=1) which corresponds
to volumetric oscillations for a spherical shape as discussed in the previous section. Higher
modes are called harmonics and differ in that there are asymmetric (non-spherical) shape
oscillations with no volumetric change. The second mode (m=2) corresponds to oscillations
between an oblate ellipsoid on one axis to a sphere to a prolate ellipsoid on a perpendicular
axis and then back again (Fig. 2.34a). Higher modes occur with more complex shapes, e.g., the
m=3 and m=5 conditions will lead to triangular and star-like configurations that oscillate
approximately within a plane (Figs. 2.34b and 2.34c). Turbulence can also excite such shape
oscillations as shown in Fig. 2.35, although with less symmetry.
Lamb (1945) derived the natural frequencies for the simple geometric modes assuming
small oscillations about a spherical shape for a fluid particle immersed in an infinite fluid. To
obtain these frequencies, the velocity potential was employed for sinusoidal oscillations of the
instantaneous particle surface about the equilibrium sphere radius with an internal pressure
based on surface tension and the (linearized) local surface curvature. The natural frequency of
a harmonic mode (m>1) was then found to be

2
nat ,m
3
p f p
m(m 1)(m 1)(m 2)
(m 1) m r
+ +
=
+ +

f
2.74
Despite the stringent assumptions of this theory, experiments have shown the predicted
frequencies to be in good agreement for bubbles (Feng & Leal, 1997) and droplets (Azuma &
Yoshihara, 1999). The second mode (Fig. 2.34a) has the lowest frequency given by:
78

2
nat ,2 3
p f p
24
(3 2 )r

=
+
f
2.75
This yields f
nat,2
of 760 Hz for a 1 mm diameter water drop in air, and 950 Hz for a 1 mm gas
bubble in water. The second mode is normally the most important as it is the one most likely
to be initiated. Higher modes may be manifested in controlled conditions with m as high as 10.
However, it is often difficult to singularly excite only one mode since weak asymmetries (e.g.,
due to walls or neighboring particles) often lead to interactions between various shape modes
resulting in complex shape dynamics (Azuma & Yoshihara, 1999).
In general, the non-spherical frequencies are lower than that predicted for the mode 1
spherical oscillation given by Eq. 2.66. However, these frequencies become comparable for
very small bubbles (about 100 m or less). This can yield significant coupling between the
shape and size oscillations (Feng & Leal, 1997). Such coupling can also lead to bubble break-
up during sono-luminescence.


Self-Generated Shape Oscillations for Swimming Organisms

Many researchers have examined the dynamics of micro-organisms with self-locomotion
using analyses of fluid particles, albeit with complex interfaces and shapes. Single- and
multiple-cell bodies contain fluid like interiors enclosed in a semi-deformable membrane (often
with finite porosity). The drag on these particles is similar to that of contaminated
immiscible drops in a liquid of similar density. However, the important difference is that
micro-organisms often use shape oscillations for self propulsion for their survival, e.g. to swim
in the direction of higher food or oxygen content within a liquid. The swimming is often
accomplished by the motion of small (relative to their body size) eukaryotic flagella or cilia on
the organisms surface (Fig. 2.36a) which move in a coordinated fashion to evoke surface
waves. The surface waves include both tangential and radial motions, and generally map out an
elliptical path (Fig. 2.36b). The radial motion is critical; otherwise, there would not be a net
propulsive force.
The surface traveling waves move from the tip (=0) to the base (=), with wavelengths
on the order of 10-20% of the body length and radial amplitudes of about 5% of the body
length. These metachronal waves have frequencies on the order of 10
2
-10
3
Hz for typical
bacterium dimensions. This yields swimming speeds of about 5d/sec, e.g., a 30 m diameter
body will move roughly 150 m/sec relative to the surrounding liquid (Brennen & Winet,
1977). The resulting body Reynolds numbers are very small, about 10
-6
to 10
-3
so that the
creeping flow theory may be used. Theoretical analysis indicates that locomotion efficiency
will be maximized in creeping flow if the vorticity production is minimized (Stone & Samuel,
1996). Therefore, small cilia compared to the body length are best for efficiency, while the
larger flagella are best for maximizing speed.


2.4. Fluid Particle Break-up and Coalescence

Under significant local stresses, a fluid particle may not only deform but also break-up into
two or more smaller particles. The initial particle is sometimes called the parent particle and
79
the particles resulting from break-up are sometimes called the children particles. There are
thermodynamic mechanisms which can cause fluid particle break-up, such as bubble break-up
by cavitation collapse or drop break-up by internal boiling. However, the focus here is on fluid
dynamic break-up mechanisms due to: Rayleigh-Taylor instabilities (due to density
differences), Kelvin-Helmholtz instabilities (due to velocity differences), shear deformation,
turbulent interaction and particle-particle collisions. The latter two mechanisms can also lead
to coalescence, especially for fluid particles of similar size.

2.4.1. Break-up due to Rayleigh-Taylor Instabilties

If a fluid particle is subjected to high accelerations and has a density different from that of
the continuous-phase, the interface may become unstable. In particular, instabilities are likely
when the deformations due to dynamic pressure become severe such that the surface tension
may be insufficient to maintain the particles surface integrity. One may expect such break-up
physics to scale with the Weber number, since it controls the degree of deformation. High
Weber numbers can result from high terminal velocities, a high injection velocity (used to help
break up droplets in the case of internal combustion engine sprays), and surrounding fluid
unsteadiness (e.g., turbulence or shock waves). In most cases, the time-dependent instability
which causes break-up can be attributed to the acceleration of the particles interface and the
density difference between the phases, as discussed in the following section.


Resulting from Gravitational Acceleration

As noted in Fig. 2.11c, a droplet with a high density ratio (
*
p
1) at terminal conditions will
typically form an indentation in the upstream portion of the droplet which leads to a bag drop.
The wall of the bag becomes thinner with time until it shatters which also causes break-up of
the toroid rim (Fig. 2.37). The break-up can be traced to the instability which causes the
dimple at the leading surface to grow and eventually form a bag (Fig. 2.12). This mechanism
for this growth is the Rayleigh-Taylor instability.
For the purposes of both physical explanation and theoretical analysis, it is convenient to
examine the Rayleigh-Taylor (RT) instability in terms of two immiscible fluids of different
densities and separated by an initially flat interface (Fig. 2.38a). If the lighter fluid is on the
top (e.g., oil on top of water), the interface is both statically and dynamically stable. If the
denser fluid is on top (e.g., water on top of oil), the interface is only statically stable and any
curvature perturbation can lead to a dynamic instability due to hydrostatic forces. If surface
tension is negligible, a Rayleigh-Taylor can occur when an increase in density occurs in the
direction opposite to gravity. However, the onset of this instability is also sensitive to the
wavelengths since it may be counteracted by capillary forces which scale with surface tension
and the interface diameter (/D). For example, a thin tube such as a straw can indefinitely
support a heavier fluid above a lighter fluid. However, a large container can not similarly
support a heavier fluid above a lighter fluid, e.g., hydrostatic force will destabilize a water
layer over a bath of oil if there is a interface is subjected to a perturbations. Thus, larger
wavelengths are more unstable.
80
The Rayleigh-Taylor instability can be analyzed by considering an interface which is
initially steady and nominally horizontal. As a result, both fluids are subjected to gravitational
acceleration normal to the interface, but are otherwise un-accelerated. To determine whether a
disturbance will be damped or excited, one may examine the dynamics of an interface
perturbation. If one assumes that the fluids are incompressible, irrotational and initially
stagnant, this velocity potential satisfies the Laplace equation (Eq. A.37). The potential for a
sinusoidal perturbation of magnitude
o
, wave-length l, and angular frequency f is given by:

2 i(x / t )
o
e

=
l f
2.76
In this equation, x is the distance along the interface (Fig. 2.38b). The induced perturbation
velocity field for the above velocity potential can then be used to determine the spatially
varying and unsteady static pressure on either side of the interface with Eq. A.41. The pressure
difference across the interface is simply given by surface tension and surface curvature (Eq.
1.6) for inviscid flow. Assuming that the amplitude of the perturbation is small, one may
derive the dispersion relation for exponential growth rate (Lamb, 1945) from which the
Rayleigh-Taylor instability frequency is identified as:

( )
3
upper lower
RT
upper lower upper lower
1 2 g 2
2


=


+ +


f
l l
2.77
A disturbance will be: unstable if this frequency is real, neutrally stable if this frequency is zero,
and dynamically stable if this frequency is imaginary. Thus, an unstable condition requires
that the first term in square brackets is positive (i.e.
upper
>
upper
) and that the second term is
smaller (low enough surface tension or large enough wavelength). The critical wavelength
associated with a neutrally stable condition (f
RT
=0) is given by:

( )
crit ,RT
upper lower
2
g

=

l
2.78
Thus, l<l
crit,RT
indicates a stable condition while l>l
crit,RT
will lead to instability. The most
unstable wavelength can be obtained by taking the derivative of the f
RT
with respect to the
wavelength and setting this to zero. This yields 3 l
crit,RT
, which is the wavelength where the
disturbances grow the fastest. Viscosity effects will reduce the growth rates of the disturbance
but will not change the critical frequency and wavelengths given by Eqs. 2.77 and 2.78.
The next step is to relate the instability characteristics to the interface of a fluid particle.
The RT instabilities will be most profound at the interface location where the density increase
is in the opposite direction of gravity, i.e. where
p f
( )n g is largest. For a drops in a gas,
this will occur on the lower surface (Fig. 2.39a) while for a bubble it will occur on the lower
surface (Fig. 2.39b). The assumption of an initially flat interface is reasonable in these area if
the Weber number is significantly greater than unity, especially for the case of a drop in a gas
(Fig. 2.11b). The maximum wavelength that can be supported on the particle interface may be
estimated as half the drop circumference. In this case, the critical wavelength (and associated
instability) can be related to the droplet diameter by:

1
2 crit ,RT crit ,RT
d l
2.79
Thus, droplets with d<d
crit,RT
are expected to be stable whereas larger drops (which allow larger
wavelengths) are expected to develop an RT instability.
81
The above analysis may be used to determine the maximum droplet size in terminal
velocity conditions. In particular, Eqs. 2.78 and 2.79 show that the critical break-up diameter
corresponds to a critical value of the Bond number (defined in Eq. 2.43) as

2
crit ,RT p f
crit
gd
Bo ~16

=


2.80
Thus droplets in terminal velocity free-fall with smaller Bond numbers are expected to be
stable, while larger Bond numbers can be expected to break-up. Based on Eq. 2.80, the
predicted maximum terminal size for water drops in air with this criterion is 11 mm which
reasonably matches the experimental observation of about 10 mm. The simple flat-interface
analysis is also consistent with experimental evidence for several fluid particles for which

p
>
f
, i.e. drops in a gas and some immiscible drops in a liquid (Clift et al. 1978). Since large
distorted drops at high Re
p
values have a drag coefficient of approximately 8/3 (6.2.6), this
Bond number can be expressed in terms of a terminal Weber number based on Eq. 2.47b. Thus,
the critical Weber number is given by:

crit ,RT,term
We ~ 8
2.81
Note that this criterion is based on steady-state terminal velocity conditions.
Next we consider why the Rayleigh-Taylor instability has a much weaker impact on a
bubble despite having the same density difference and same surface tension. The difference
can be traced to the local interface shape where the instabilities are generated. For the bubble,
the top interface is expected to be the location where instabilities grow fastest (Fig. 2.39b).
However, unlike the droplet at high Weber numbers, this region of a bubble has significant
convex curvature as shown in Fig. 2.14b (and as discussed in 2.2.2). As such, the bubble will
be more resistant to Rayleigh-Taylor instabilities because: 1) the leading-edge curvature yields
an initial surface tension force which will resist convex (dimple) instabilities, and 2) the
fraction of the interface that is nearly flat is less than that for a drop, and 3) the higher
recirculation rates in bubbles (compared to that in drops) allows instabilities to convect away
from the leading edge before they can cause break-up. As a result, bubbles are much less
sensitive to gravitational RT instabilities than droplets of similar Weber numbers. Droplets in
an immiscible liquid are even less sensitive to Rayleigh-Taylor instabilities since the density
difference is far smaller than for drops in a gas.


Resulting from Particle Acceleration

In the case where the relative particle acceleration is much higher than that dictated by
gravitational conditions, the Bond number is no longer an applicable parameter. To investigate
such conditions, Joseph et al. (1999) conducted experiments with shock-induced instability and
break-up of drops. They noted that bag-type break-up could occur at very high Weber
numbers (about 5,000) as shown in Fig. 2.40a. A close-up image of the drop where the surface
instabilities begin is shown in Fig. 2.40b. This portion is consistent with the region most
susceptible to Rayleigh-Taylor instabilities (Fig. 2.39a) and may thus be analyzed accordingly.
However, one must employ the interface acceleration due to particle acceleration since this is
very high (more than 10
6
m/s
2
) compared to that of gravitational acceleration. If the local
particle acceleration is used in the Rayleigh-Taylor instability criteria of Eqs. 2.78 and 2.79,
the critical wavelength and diameter are given by:
82

crit ,RT
p f
crit ,RT crit ,RT
2
w / t
d 2 /

=


l
d d
l

2.82a

2.82b
The experimental results showed that instabilities of this wavelength appeared on the droplet
surface (Fig. 2.40b), even at very high Weber numbers (about 5,000). The results suggested
that high accelerations will produce break-up for d>d
crit,RT
given by Eq. 2.82, though further
studies are needed to confirm this result.


Resulting from Interface Radial Acceleration

When a fluid particle is exposed to pressure fluctuations, the resulting size oscillations
(discussed in 2.3.3) can also lead to break-up due to asymmetric instabilities. The break-up
due to size oscillations for liquid droplets in a gas is relatively rare because the relative
incompressibility of the internal liquid yields low accelerations of the radius. However, the
break-up of bubbles due to size oscillation instabilities is widely known to occur because of the
strong sensitivity of the bubble volume to external pressure perturbations. As an example, the
collapse of an acoustically excited bubble can result in several children bubbles (Fig. 2.41).
These interface oscillations can be associated with a critical break-up diameter by substituting
p
r `` for particle relative velocity in Eq. 2.82. The frequencies for such oscillations are typically
much higher than that associated with Rayleigh-Taylor instabilities, so that these two processes
are effectively decoupled. For example, a 6 mm diameter bubble will have a natural volume
oscillation frequency of about 8 kHz (Eq. 2.66) whereas the most unstable Rayleigh-Taylor
frequency can be found by Eq. 2.77 to be about 6 Hz.

2.4.2. Break-up due to Kelvin-Helmholtz Instabilities

The Kelvin-Helmholtz (KH) instability is associated with flows which have tangential
velocities that vary in magnitude. It is caused by hydrodynamic amplification of perturbations
and can arise on a interface which has a velocity discontinuity as shown in Fig. 2.42. For this
idealized condition, a sinusoidal surface wave (Fig. 2.38b) will tend to cause a local increase in
the upper velocity on its crest based on potential flow (similar to the velocity increase seen
around the sides of an inviscid sphere). This, in turn, will cause a lower local pressure on the
crest which serves to pull the crest higher. The opposite will occur in the wave valley, i.e. the
local velocity will decrease causing the pressure to increase further which serves to push the
indention downward. The amplification of the surface wave perturbations results in interface
instabilities. A classic example of this is wind-induced wave instabilities on a body of water.
In this case, perturbations are amplified by the above KH mechanism but are stabilized by
gravitational forces which tend to make the interface flat. This balance effectively determines
the wave height. Another example is that of a shear layer as shown by Fig. A.5, where a
velocity difference between two parallel flowing streams results in eddy and braid structures.
In both fluids are of the same density and are miscible with low viscosity, there are no
stabilizing gravitational or surface tension forces to deter the instabilities so that they lead to
turbulent flow for any wavelength (Fig. A.5c). However, if the two immiscible fluids are
83
separated by an interface, the KH instabilities are limited to a critical wavelength based surface
tension. To determine this wavelength, one may again use inviscid potential flow theory
(Drazin, & Reid, 1981). The resulting dispersion relation for a perturbed planar interface
yields a KH instability frequency given by:

2 flow
KH flow flow
flow stag flow
1 2
U U




=



+



f
l l

2.83
An unstable condition corresponds to a real root and this can be used to determine the critical
length scale as:

flow stag
crit ,KH 2
flow stag flow
2
U
+

=




l
2.84
Therefore, a given wavelength is more likely to be unstable as the velocity increases or as the
surface tension reduces.
To translate the above instability analysis for an initial flat interface to that of a fluid
particle, one may assume that the fluid inside the particle is approximately stagnant (so that

stag
=
p
) while the continuous-phase velocity over the surface drives the instabilities as shown
in Fig. 2.42 (so that
flow
=
f
). Furthermore, the fluid velocity over the surface (U
flow
) can be
assumed to be proportional to the particle relative velocity (w) and the critical wavelength
(l
crit,KH
) will be proportional to the critical particle diameter for break-up (d
crit,KH
). This may be
used to define the critical Weber number and relate this to the fluid densities, while
experiments may be used to determine the constants of proportionality:

( )
2
p f
f crit ,KH
crit ,KH
p
w d
We 12
+



2.85
For the condition of
*
p
1 (droplets in a gas at high Re
p
), this corresponds to We
crit
12 which
is the well-known dynamic criterion of Taylor (1949). Note that the constant of proportionality
is difficult to predict theoretically since the edges of a fluid particle will have large curvature
and will be highly convection. Nevertheless, some researchers have obtained a critical Weber
number assuming V
1
w and assuming the most unstable wavelength (instead of the critical
wavelength) should be used, which yields the same RHS as Eq. 2.85 except with 4=12.57 as
the constant, so that the difference is minor. For droplets moving at high velocities (e.g., due
to spray injection), We
crit
~12 has been found to be a reasonably accurate predictor for the
break-up which typically occurs with the bag-type mode of Fig. 2.37 whereas higher Weber
numbers (e.g. We>50) lead to aerodynamic sheet-stripping and explosive break-up modes
(Theofanous et al. 2004).
Several researchers have also noted that viscous effects can stabilize a drop and allow
increased critical Weber numbers. To characterize this influence, it is convenient to define the
Ohnesorge numbers based on the respective viscosities and densities of either the surrounding
fluid or of the particle fluid:
84

1/4
f
f
p f
*
p p
p f
*
p
p
We Mo
Oh
Re Bo d
Oh Oh
d

= =



=


2.86a


2.86b
These two dimensionless parameters respectively relate external and internal viscous stresses
to surface tension stresses. Often, Oh
p
is typical larger than Oh
f
for drops in a gas so that one
may expect Oh
p
(i.e. internal stresses) to dominate the viscous influence. Based on
experiments for drops with Oh
p
values as high as 5, Pilch & Erdman (1987) proposed the
following empirical dynamic correlation for drop break-up:
( )
1.6
crit ,KH p
We 12 1 1.07Oh = +
2.87
However, this viscous correction is approximate since it is based on particular set of test
conditions and since RT instabilities due to rapid acceleration or deceleration can also be
important (Faeth, 1987). In fact, current detailed models for drop break-up are based on a
combination of RT and KH instabilities.
It interesting to compare the inviscid (high Re
p
) Rayleigh-Taylor and Kelvin-Helmholtz
instability criteria of Eqs. 2.81 and 2.85 for different density ratios. For
p f
> (droplets in a
gas), the Kelvin-Helmholtz prediction of We
crit
12 is less restrictive than that the Rayleigh-
Taylor criterion of We
crit,term
8. However, this R-T value is more appropriate for terminal
velocity conditions while the K-H is more appropriate for unsteady interactions at high relative
velocity. For
p f
< (bubbles), the KH instability based on Eq. 2.85 negligible and the RT
instability was discussed previously to be weak, and this combination is consistent with
spherical-cap bubbles at very large Weber numbers. For
p f
(droplets in immiscible
liquids), the Kelvin-Helmholtz instability will typically dominate (e.g., We
crit
24) since the
Rayleigh-Taylor instability (based on densitydifference) will be weak in comparison.

2.4.3. Break-up due to Shear Deformation

When exposed to a continuous-phase linear shear flow (Fig. A.3a), a fluid particle will
deform and rotate (2.3.2 and Fig. 2.29). As the shear rate increases, deformation will elongate
the particle until instabilities arise and the drop or bubble interface will rupture. This
deformation is influenced by the shear Reynolds number (Re

of Eq. 2.52) and by Capillary


number (Ca

of Eq. 2.53). These two parameters can be combined to form a shear Weber
number given by:

3 2
f shear
d
We Re Ca


= =

2.88
Many of the drop break-up experiments have been conducted in immiscible liquids with
*
p
~1
where the relative velocities are sufficiently small to allow detailed observation of the shear-
induced deformation. Two such examples are shown in Fig. 2.43 where both cases have
*
p
1 but their Ca

(defined in Eq. 2.53) values differ by a factor of four. The low shear (small
85
Ca

) case of Fig. 2.43a exhibits a break-up at an aspect ratio (E) of about one-third. The
resulting children particles include two primary drops and a few small satellite drops in the
region of rupture. The high shear (high Ca

) case of Fig. 2.43b, the deformation is more


influenced by viscosity and can survive greater aspect ratio before rupture (Fig. 2.43b).
The critical Capillary number for break-up of fluid particles is denoted Ca
,crit
. While
robust criteria are not generally available for this parameter, guidelines are available from
numerical simulations. For example, the
p f
results of Fig. 2.44 indicate strong sensitivity
of viscosity ratio at creeping flow conditions (Re

1) with break-up conditions limited to


viscosity ratios of 0.005<

<3. This limited region of finite Ca


,crit
values is attributed to a
strong viscous coupling between the particle and continuous-phase fields (Khismatullin et al.
2003). As Re

increases, the deformation increases due to increased inertia and the critical
capillary number (Ca
,crit
) reduces. At Re

=50, the break-up criterion becomes independent of


the viscosity ratio and roughly characterized by We
,crit
2.5. This criterion is consistent with
experiments of break-up at high Reynolds numbers (Clift et al. 1978). However, the role of
surfactants and density ratio numbers complicate matters further such that the understanding of
shear break-up criteria is incomplete.

2.4.4. Break-up due to Turbulence

Many engineering and natural multiphase systems include turbulent flow, and often drops
or bubbles will break-up when exposed to sufficiently strong levels of turbulence. Since many
generations of particle splitting may occur, this yields a spectrum of particle sizes. In general,
the mechanisms of deformation and break-up can include all those discussed in 2.4.1 and
2.4.2: indentation due to high relative velocity, stretching due to high velocity shear,
acceleration instability from Rayleigh-Taylor (near the leading edge), or shear instability from
Kelvin-Helmholtz modes (near the sides). The characteristics of turbulence are discussed in
A.5 which includes a wide range of non-linear turbulent structures (vortices or other features
such as seen in Fig. A.5c) which can drive fluid particle deformations. As such, turbulent
break-up is quite complex and primarily predicted with empirical models.
In some cases, droplets are first stretched into in ligaments as in Fig. 2.45. This can be
attributed to shear deformation (as in Fig. 2.43) or shape oscillations (e.g. m=2 of Fig. 2.34a)
Such ligaments typically have a dumb-bell type shape with a thinner filament in between the
larger masses. This filament will then break up due to surface instabilities along its length,
yielding two primary drops along with several smaller satellite drops which arise from the
break-up of the connecting filament. However, many other highly deformed shapes have been
found to lead to break-up. Thus, the process is best considered from a probabilistic point of
view owing to the wide range of mechanisms, shapes, and outcomes that are possible.
Similarly, bubbles subjected to turbulence can undergo a variety of break-up scenarios
including shape oscillations and shear deformation as shown in Fig. 2.46. These results are
similar to those observed by other researchers for gravitational conditions (Martinez-Bazan et
al. 1999a; Qiand (2003) who also noted that higher instantaneous Weber numbers tended to
yield a higher number of daughter bubbles.
Despite the complexity of turbulent deformation and break-up physics for drops and
bubbles, some simple scaling techniques can be used to qualitatively determine the maximum
86
particle size which can exist. In particular, Hinze (1975) postulated that the local shear stresses
deform a drop and will break it apart if these stresses exceed those resisting deformation
(particle surface tension and particle viscosity). Hinze suggested that the local turbulent
kinetic energy was important (whether the break-up was due to shear deformation, shape
oscillations, Rayleigh-Taylor instabilities, Kelvin-Helmholtz instabilities, etc.). Further, Hinze
proposed that the turbulent structures whose wavelength (l) is similar to the drop diameter (d)
are primarily responsible for break-up. This is because very large structures will be seen as
nearly uniform by the drop while very small structures will be weaker and will not cause
whole-body deformations to the drop. To determine the intensity of kinetic energy at a scale of
the particle diameter, the corresponding turbulent wavelength was assumed to occur in the
inertial sub-range and thus proportional to (d)
2/3
based on Eq. A.96b. Assuming further that
the effect of the viscosity (and thus the Ohnesorge number) is negligible in the deformation and
break-up, Hinze proposed a constant of proportionality from experimental studies for droplet
breakup so that the critical turbulent Weber number is given by:

2/3 5/3
f crit
,crit
2 d
We 1.2

2.89
Any fluid particle with d>d
crit
is thus likely to be broken-up in turbulence. This critical Weber
number has been found to reasonably predict size distributions of drops (Clift et al. 1978) and
bubbles (Martinez-Bazan et al. 1999b; Millies & Mewes, 1999).
Since turbulent dissipation is not always easily measured, Risso and Fabre (1998)
proposed a modified critical Weber number expression based on turbulent kinetic energy and
integral length-scale and determined a critical value based on micro-gravity experiments:

5/3
f crit
k,crit 2/3
0.7 kd
We 3

2.90
A similar Weber number was defined by Qiand (2003) but with a critical value given by
We
k,crit
2 based on the results of Sevik & Park (1973). Still other proposed formulations
include effects of viscosity through the Ohnesorge number, which can be especially important
for low Re
p
conditions (about 10 or less). Furthermore, Martinez-Bazan et al. (1999b) noted
that the distribution of the turbulence spectra can also be important. The variety of proposed
forms for the critical Weber number reflects the complexity and empiricism associated with
turbulence modeling.
Turbulent flow can also lead to particle-particle interactions which can result in break-
up and coalescence of fluid particles. The collision frequencies as a function of drop size and
turbulent flow properties will be discussed in 3.8.2. The type of outcome which arise from
these collisions is discussed in the following section.

2.4.5. Coalesence and Break-up due to Collisions

Fluid particles can coalesce, rebound or break-up if they collide with other fluid particles.
This can be important to overall size distributions. For example, growth of cloud drops to rain
drops is primarily attributed to particle-particle collisions. Similarly, bubble size distributions
are primarily governed by a balance between turbulent break-up and bubble coalescence.
Break-up may also occur on wall surfaces due to droplet splatter or bubble splitting, but such
87
effects are typically secondary in terms of determining size distributions. The following
discusses the potential outcomes for drops in a gas, the most commonly studied case, followed
by some discussion for bubbles and drops in liquids.


Droplet Collisions in a Surrounding Gas

Droplet-droplet collisions in a gas (
*
p
1) occur when two droplets move towards each
other in flight and come into direct contact. The most common interaction is a binary collision
whereby two particles collide (trinary and more complex collisions are generally rare). A
binary collision typically results in significant fluid particle deformation, particularly if the
particles are of similar size. The drop size difference can be characterized by normalizing the
small diameter by the large diameter (Eq. 2.20). The relative velocity of the particle centroids
as they move towards each other (i.e. the incoming velocity) is given as

p p 1 2 large small
= v v v v v
2.91
The collision kinetic energy from this velocity can be compared to that of surface tension by
defining a collisional Weber number:

2
p p p small
p p
v d
We

2.92
This equation uses the diameter of the smaller droplet to allow a finite value even in the
limiting case of a drop interacting with an unbounded liquid pool. If the collisions are not
head-on, the skewness of the collision can be characterized by the incoming impact angle
in
,
which is defined as the angle between the particle-particle contact plane and the relative
incoming velocity vector (Fig. 2.47). This angle can be used to define an impact parameter:
( )
in
cos
2.93
Thus, a head-on collision corresponds to
in
=90
o
and =0 while a near grazing condition
corresponds to
in
=0
o
and =1.
The dynamics which occurs depends primarily upon We
p-p
and and can be classified by
the five types of interactions shown in Fig. 2.48 and listed below (in order of increasing We
p-p

and thus impact speeds):
SC: slow coalescence where the droplets move slowly enough that there is little
deformation and there is sufficient time for the interfaces to merge by diffusion;
B: bounce where the impact speeds allow for significant deformation but the interaction
time is too short for any coalescence;
FC: fast coalescence whereby large deformation dynamics occur after impact causing the
interfaces to break and merge due to hydrodynamic instabilities, after which the dynamics
subside and a single combined drop is formed;
HS: head-on separation where the droplets impact at roughly head-on angles and
temporarily coalesce due to the high impact speeds, but the high inertias of the moving
mass regions cause a shape reflection (as if reflecting off an inviscid wall) which leads to
separation of the original primary drops but with a residual filament (which can form a
satellite drops); and
88
GS grazing separation or off-center separation where the droplets impact at roughly
grazing conditions and temporary coalesce but then separate due to shape reflections (often
more complex than a simple mode 2 oscillation), which leads to two primary droplets and a
residual filament which can lead to many satellite drops.
Brenn et al. (2001) showed that increasing the collisional Weber number for the GS regime
increases the number of small satellite drops due to increased instability of the filament.
The range of interactions observed for droplets as a function of Weber number and
collision angle is shown with a qualitative map in Fig. 2.49. This figure shows that slow
coalescence (SC) is observed at very low Weber numbers, while bouncing (B) dominates at
higher Weber numbers, particularly for off-axis collisions. For on-axis collisions, an increase
in Weber number leads to fast coalescence (FC) and eventually head-on separation (HS). For
nearly grazing conditions, the bounce outcome is followed at higher Weber numbers by
grazing separation (GS). The regime boundaries in Fig. 2.49 can be characterized by the three
intercepts (We
SC
, We
B
, and We
FC
) as well as the Weber number for =0.5 for grazing
separation (We
GS
). Predictions of these four characteristic Weber numbers and of the collision
angle dependence are also shown in this figure. The predictions are based on models which
incorporate influence of relative diameter (based on the diameter ratio of Eq. 2.20), viscosity
effects (through the Ohnesorge numbers of Eq. 2.86), and interface dynamics.
The first intercept (We
SC
) occurs at the boundary between slow coalescence and head-
on bouncing. This boundary occurs when the time required for a slow coalescence merger to
occur (
SC
) equals the contact time between two drops during a head-on bouncing collision
(
cont
). Thus, the Weber number associated with
SC
=
cont
defines We
SC
. The contact time-
scale has been extensively studied and shown by Foote (1974) to vary inversely and
quadratically to the impact velocity (and thus the impact Weber number) as:

( )
4 3
cont p p p p
p p
d v d / We

= 2.94
In the following, a model for
SC
is put forth.
The merger in the slow coalescence process can be regarded as a two-step process:
condensation of a liquid bridge between the drops based and drainage of a significant portion
of the droplet fluid into this bridge to ensure interface merger. The two steps can be expressed
in terms of time-scales for bridging and draining (
bridge
and
drain
) along with two empirical
coefficients (c
bridge
and c
drain
):

SC bridge bridge drain drain
c c = +
2.95
The bridge time-scale is controlled by molecular diffusion of particle species vapor in the gas
when the two drop interfaces are in very close proximity, about 100 angstroms. If the
surrounding gas is initially dry, the bridge time will be proportional to the droplet area (d
2
)
and inversely proportional to the diffusion rate of the droplet species in the surrounding gas
(
p@f
of Eq. 1.87). However, if the surrounding gas is already saturated with particle species
vapor, then the bridge time will be zero. As such, the bridge time-scale can be expressed in
terms of the surface vapor pressure (at T
p
and saturated conditions) and the far-field vapor
pressure (at T
g
but not necessarily saturated conditions) as:
89

3/2
2
vapor,surf vapor
ref
bridge
p@f ,ref ref vapor,surf
P p
T d p
p T P








2.96
In this equation,
p@f ,ref
is the binary diffusivity at a reference pressure and temperature (p
ref

and T
ref
) while the term in the square brackets accounts for finite vapor pressure in the far-field.
The drainage time-scale can be related to particle viscosity and a bridge length-scale as:

3
drain p p p
d Oh d / = = 2.97
The RHS expression indicates that drainage time is linearly proportional to the particle
Ohnesorge number and a square root term which is often defined as the capillary time-scale.
Using
SC
=
cont
, Eqs. 2.95-2.97 and the empirical values for the constants (c
bridge
=0.15 and
c
drain
=5.0 based on by Loth et al. 2008), the Weber number defining the head-on SC boundary
is given by:

( )
1
3
SC bridge p p
We 0.15 d 5Oh


+


2.98
This model indicates a reduced SC probability as gas pressure, gas density or droplet viscosity
increase, which is consistent with experimental results, e.g., Figs. 2.49 & 2.50. Such figures
also show that the collision angle dependence on this boundary is approximately linear so that
the overall boundary can be estimated as:
( )
SC/B SC
We =We 1-
2.99
As such, slow coalescence is not expected to occur for grazing conditions.
For the B/FC boundary (between bouncing and fast coalescence), the associated
deformation dynamics (Fig. 2.48) indicate that the drops are flattened up against each other and
separated by a thin parallel gas layer. The separating gas layer will remain intact if the
interface oscillations are damped which will then lead to a bounce outcome (We
p-p
<We
B
).
Otherwise, wave instabilities will form which will lead to interface merging and fast
coalescence (We
p-p
>We
B
). Therefore, the boundary is determined by the dynamics of the
interfaces and the cushioning gas for which gas viscosity, gas pressure, droplet viscosity and
surface tension are all expected to play a role. Their role is not understood theoretically but
Loth et al. (2008) collapsed the experimental data with the following dependence:

1/2
f f
p p B
Oh Oh 1
0.46 0.105
Oh Pd Oh Pd We




+ 2.100
This includes two empirical coefficients and a non-monotonic variation of this parameter to
predict the head-on boundary. The influence of collision angle is better understood and is
reasonably predicted by a model of Estrade et al. (1999), which can be approximated as:


B
B/ FC 1.5
We (1 )
We
(1 )
+
=

2.101
This model is based on the assumption that fast coalescence will occur if the initial kinetic
energy exceeds that needed to create a minimum deformation of the droplets.
The boundary between fast coalescence and head-on separation was treated
90
theoretically by Ashgriz & Poo (1989) for inviscid conditions which incorporates collision
angle dependence and correlates well with experiments at low droplet viscosities. Their set of
four equations can be approximated with a single equation (for computational convenience) as:



( ) ( )
p
FC
FC/HS * *
small small
582Oh +12
We
We =
1-3 d 1-3 d
=

2.102
For the head-on boundary Weber number, Qian & Law (1997) noted an influence at high liquid
viscosities which can be incorporated as:

FC p
We =12+582Oh
2.103
Similarly, the boundary between fast coalescence and grazing separation (FC/GS) was
successfully obtained analytically by Brazier-Smith et al. (1972) for the inviscid limit. Their
result describes the boundary dependence in terms of collision angle and size ratio as:

( ) ( )
( )
( )
( )
2/3
2 3 11/3
* * * * 3
small small small small
GS
FC/GS
2 2
* * 5
small small
d +d - 1 d 1+d
We
We
5.24
1+d d

+


=

2.104
Again, the head-on boundary Weber number requires a correction for highly viscous drops,
which was proposed by Loth et al. (2008) as:


( )
GS p
We =21+630 Oh
2.105
It is important to note that We
FC/HS
and We
FC/GS
are independent of the gas properties (density,
pressure, etc.), unlike that for We
SC
and We
B
.
In general, predictions based on these four boundaries yield at least qualitative results
(e.g., Figs. 2.49 & 2.50). The lack of prediction accuracy can be partially attributed to the
empirical character of the models but also to experimental uncertainty and other factors. One
such factor is a difference in electrostatic charge which can promote coalescence even when
the charges are of the same sign (Czys & Ochs, 1998). Another factor is that of non-spherical
droplet shapes, which can greatly complicate collision outcomes and lead to a large number of
satellite drops (Willis & Orme, 2000).


Drop and Bubble Collisions in a Liquid

Bubbles (
*
p
1) and immiscible droplets (
*
p
~1) in a liquid can have similar outcomes
as those seen for drops in a gas. However, the collision Weber numbers are typically limited to
values on the order of unity due to the reduced inertia of the particles and the higher viscosity
of the surrounding fluid. As a result, collision outcomes tend to be confined to slow
coalescence or bouncing.
For head-on clean bubble-bubble interactions in gravitational conditions, the slow
coalescence regime is typified by We
SC
0.2, while bouncing tends to occur for 0.2<We
p-p
<3
(Duineveld, 1995). In addition, Duineveld noted that We
SC
depends on whether there is vortex
shedding or contamination. This latter is important because surfactants can prevent the
drainage needed to rupture the interfaces so that there are fewer opportunities for coalescence
(Walter & Blanch, 1986). For example, bubbles in saltwater tend to have longer lifetimes and
larger sizes than bubbles in freshwater due to the stability provided by the contaminant. As
91
another example, detergents include surfactants which can prevent coalescence and yield
foams (or emulsions in the case of immiscible droplets in liquids). Because of the complex
nano-scale physics involved in the drainage, models for coalescence rates can be quite
empirical, especially in contaminated systems (Tsouris & Tavlarides, 1994).


2.5. Chapter 2 Problems

2.1) Consider a particle mixture which has a triangularly-shaped number probability
distribution function given in the following figure.

Determine: a) d
10
, b) d
30
, c) d
31
, and d) d
32
. Obtain and plot the number-based cumulative
distribution function.
2.2) a) Derive an analytical expression for Y beginning with Eq. 2.15 but using the Putnum
fit (Eq. 1.53) instead of the White fit. b) Is the behavior of the Putnam Y in the Newton drag
regime physically appropriate? Explain why or why not. c) Consider a bi-disperse mixture
with two particle sizes (d
large
=2d
small
) which each constitute 50% of the total particles by
number. If the particles move as a single cloud with a mean Re
p
consistent with Y=1.5,
obtaining d
3J
/d
32
and discuss whether it is reasonable to describe this mixture with a single
effective cloud velocity and the Sauter Mean Diameter.
2.3) a) Obtain the Eq. 2.24 relation between the Rosin-Rammler spread parameter and the
diameter ratio (d
90%
/d
10%
) starting with the C

distribution function of Eq. 2.23. b) Obtain the


similar relation given by Eq. 2.33 starting with the Log-Normal C

distribution function of Eq.


2.32. c) Show that d
10
<d
20
<d
30
etc. for a polydisperse log-normal distribution.
2.4) Based on the definitions of Re
p
, We, Mo and Bo, confirm the relationships given in
Eqs. 2.45, 2.46 and 2.86.
2.5) Consider a spherical cap bubble rising at terminal velocity at a high Bond number
such that surface tension effects are weak. The top portion of the bubble shape can be
approximated with a constant radius of curvature (r
cap
=9d/8) while the bottom is a flat
horizontal interface.





a) If there is no surface tension, the liquid dynamic pressure distribution over the top surface of
the bubble is given by the inviscid Bernoulli relation (Eq. 1.48) and must be balanced by the
hydrostatic pressure change. Write the pressure balance and then apply small angle
d
20 m
P
N

10 m 30 m
r
cap


92
approximations (1) for the trigonometric functions to obtain the corresponding drag
coefficient of Eq. 2.47a ; b) Show that this drag coefficient yields the relationship given by Eq.
2.47b; c) Discuss the qualitative differences between drop deformation in a gas and bubble
deformation in a liquid for both moderate and high Weber numbers.
2.6) a) Starting from Eqs. 2.55-2.57, derive the isothermal version of Rayleigh-Plesset
ODE (given by Eq. 2.64). b) Obtain an expression for the corresponding natural frequency of
this ODE by assuming a constant far-field pressure, negligible viscosity and weak bubble
oscillations of the form ( )
p p,eq
r r 1 sin 2 t = +

f with 1. Hint: neglect higher-order in and
consider the remaining sin(2ft) terms to solve for f
nat
.
2.7) For a 2 mm diameter bubble in water at isothermal conditions with a far-field liquid
pressure of one atmosphere, obtain: a) the natural frequencies with m=1,2,3; b) the
characteristic frequency (inverse of characteristic time-scale) for shear deformation, and c) the
most unstable frequency for Rayleigh-Taylor instability (if it exists), and d) the level of
turbulent dissipation for which bubble break-up is likely to occur.
2.8) Consider monodisperse ethanol drops of 20 microns in diameter at STP conditions
injected into dray air (similar to that seen for a pre-compression fuel spray). Compute the
maximum relative velocity where: a) slow coalescence occurs, b) where head-on bounce
occurs, c) and where fast coalescence will occur at any angle (i.e. when We
FC/HS
=We
FC/GS
).





93
3. Coupling Physics and Regimes for Multiphase Flow

3.1. Overview of Coupling Regimes

Multiphase flows can have more than one type of dispersed-phase (e.g., particles and
bubbles in a liquid) and/or more than one type of continuous-phase (e.g., particles in two
immiscible liquids). But the most common occurrence is for one dispersed-phase and one
continuous-phase. There are a variety of interactions which can occur between the two phases
(particle-fluid interactions) and among the dispersed-phase (particle-particle interactions).
Particle-fluid interactions encompass both the effects of the continuous-phase on the particle
trajectories and the effect of the particles on the continuous-phase flow distribution. Particle-
particle interactions can refer to two separate mechanisms: fluid-dynamic interactions and
collisions. Particle-particle fluid-dynamic interactions occur when the proximity of one
particle directly affects the fluid dynamic forces of a neighboring particle, such as by wake
drafting. In contrast, particle-particle collisions occur when particles come into direct contact,
e.g., the particles can rebound, shatter, or coalesce. Similarly particle-wall collisions can result
in particles which reflect, break-up or adhere.
There are several regimes which describe the degree and type of coupling which occurs
between particle motion and that of its surroundings. These regimes are illustrated in Fig. 3.1
and can be used to classify the character of the multiphase flow, and thus to help determine
appropriate numerical techniques. The broadest division is between dispersed and dense flow.
Dispersed flow (which is the focus of this text) is defined herein as the conditions when the
particles can be considered to approximately move as individual entities. In dense flows, the
interactions between particles dominate the particle motion. An increase in particle
concentration by volume decreases the average inter-particle spacing. Such a change makes
fluid dynamic interactions and collisions among particles are more likely, and thus can cause a
change from dispersed to dense flow.
Additional divisions can be made for dispersed flow based on the degree of coupling,
which tends to increase as the number of particles increases. By order of increasing mass and
volume fractions (e.g., starting from an isolated particle), dispersed-flow includes four different
coupling regimes:
a) one-way: particle-phase motion affected by the continuous-phase but not vice-versa,
b) two-way: particle-phase also affects the continuous-phase through fluid coupling,
c) three-way: flow disturbances caused by particles affect the motion of nearby particles,
d) four-way: contact collisions also influence the overall particle motion.
The first three regimes include only fluid dynamic interactions, whereas the fourth regime
includes particle collisions. These regimes occur in many practical multiphase flows. An
example effect of one-way coupling is the drag of a particle, which is caused by the surface
fluid dynamic forces. For two-way coupling, and an example effect of is the upward
momentum of a liquid induced by a bubble rising in a plume. An example effect of three-way
coupling would be a particle drafting behind another particle. Small rain drops colliding and
coalescing to form larger rain drops is an example of four-way coupling. So long as fluid
dynamic forces dominate the particle movement, these are dispersed flow regimes.

94
Dense flow, where particle-particle (three-way or four-way) interactions dominate, can be
divided into two categories: collision-dominated and contact-dominated. For collision-
dominated conditions, the particles tend to move at speeds on the order of the terminal velocity
and therefore the interactions yield high rebound velocities. In these conditions, four-way
interactions dominate over one-way effects so that the importance of the surrounding fluid is
reduced. As the particle concentration increases further, the time of contact tends to increase
as compared to the time of free-flight motion. This will eventually lead to contact-dominated
flows, where the particles tend to spend most of the time in contact with other particles, e.g.,
particles roll and rub against each other. Often, the effects of the continuous fluid on the
particles are weak and can be neglected, e.g., granular flows typically are modeled by
neglecting the presence of the surrounding-fluid. Dense flow and dispersed flow can also co-
exist in some multiphase flows. For example, sediment transport of a particle bed (e.g., sand,
dust or snow) by boundary layer flows (wind or water) have dense-flow regimes near the
particle bed and dispersed flow in the outer region.
The rest of this chapter will examine the physic and develop approximate criteria for the
various coupling regimes which characterize a multiphase flow. It should be noted that
defining the boundaries between these various regimes can be subjective, but qualitative
relationships help determine the expected type of coupling. This, in turn, helps determine the
appropriate equations of motion (for both phases) and the appropriate numerical strategy. This
is important in order to capture the relevant physics at a reasonable computational cost.


3.2. One-Way Coupling and Domain Stokes Number

To begin, we will consider the one-way coupling condition where the surrounding fluid
is not influenced by the dispersed-phase (i.e., the flow motion is independent of the particle
presence), but the particle dynamics directly respond to the fluid velocity. This can range from
simple surrounding-fluid velocity variations which are approximately one-dimensional (such as
inviscid flow though a nozzle), to the more complex diffusion dynamics associated with
turbulent or molecular fluctuations. The one-way coupling assumption is reasonable in many
multiphase flow conditions and can be particularly advantageous in terms of devising an
expedient numerical strategy. There are two reasons for this. Firstly, the surrounding flow-
field (especially, if steady) can be determined beforehand without regard to the particle
conditions, i.e., the solution does not require the position, velocity and trajectories of the
particles. Secondly, several particle conditions can be applied to this same flow-field to allow
a parametric variation (e.g., vary injection location, velocity or particle size). Thus, the one-
way coupling assumption can represent a significant reduction in required computational
resources and/or scheme development, and therefore this simplification should be used when it
is appropriate. The conditions associated with two-way coupling will be discussed in 3.7.


Particle Response Time

In order to assess one-way coupling, it is helpful to consider the particle response to a
simple continuous-phase flow field. In particular, the temporal response to the restoring drag
force for a given effective mass can be used to define a particle response time as
95

2 *
p f p eff
p
D f D
( c )d 4( c )d m
F 18 f 3C

+ +
= =

w
w

3.1
Thus, particle response time is effectively proportional to
p
for
*
p
1, but to
f
for
*
p
1. For
the linear drag regime where f
Re
is a constant (Eq. 1.57 with Re
p
1) so that
p
is proportional
to d
2
and independent of the local relative and so can be known a priori. This is quite
convenient since many two-phase interactions are characterized to the particle response time.
For the Newton drag regime where C
D
is approximately constant (Eq. 1.51 with Re
p
1),
p
is
proportional to d but is inversely proportional to |w|. As such,
p
can vary in time and space if
there are local relative velocity changes and this variation can similarly occur for intermediate
Re
p
values due to non-linear drag effects (Eq. 1.57). For convenience, an a priori estimate of
the response time can be made for intermediate and high Reynolds numbers by assuming that
the relative velocity can be approximated by the terminal velocity and by combining Eqs. 1.81a,
1.82a and 3.1:

term
p p,term
w
g 1 R
=

for near-equilibrium conditions


3.2
Therefore, the terminal particle response time for a very-heavy particle (R0) tends to w
term
/g
while that for a very-buoyant particle (R3) tends to w
term
/2g. The response time for a
neutrally-buoyant particle (R=1 & w
term
=0) can instead be obtained via Eq. 3.1 as d
2
/12f
f
.
If heat transfer also occurs, it is interesting to compare this momentum response time to the
thermal response time (Eq. 1.93b). The ratio of these two response times for creeping flow is
( )
p T f f p,p
2 3 = k c
3.3
As an example, this ratio is about 0.2 for water droplets in air (Table A.1) such that the thermal
changes will be slower than (but on the order of) the momentum changes.


Solutions to Linear Equations of Motion

Based on Eq. 1.77 for an isolated solid spherical particle (which neglects Basset history, lift
and particle-particle interactions), the particle equation of motion can be given as
( ) ( ) ( )
@p @p
p f p f f 2
f
c 1 c
t (d /18 f ) t

+ = + + +

u v u
v
g



D
d
d D

3.4
This assumes that fluid accelerations along the particle and fluid path are approximately equal:

@p @p
t t

u u

D d
D d
3.5
This relationship is exact in the creeping flow regime (Maxey, 1993) and is otherwise
reasonable if the fluid stress is a secondary force or if the fluid spatial gradients are weak. The
result of Eq. 3.4 can then be rearranged using the particle response time (
p
of Eq. 3.1) to yield
an equation for the particle acceleration equivalent to Eq. 1.81b:
96

@p @p
p
(1 R) R
t t
= +

v - u u
v
+ g



d
d
d d

3.6
For the linear drag regime,
p
is a constant so that this equation can be decomposed into
independent equations of motion for each direction. In contrast, a non-linear drag regime
associated with finite Re
p
values can not be linearly decomposed and is not as convenient to
solve analytically.
The above particle equation of motion can be integrated for the linear drag regime if u
@p
is
constant along the particle path (e.g. the flow is uniform). If the initial particle velocity is
denoted v
o
, the particle velocity can be written as:

( )
( )
( )
p p p
t / t / t /
0 @p p
(t) e 1 e 1 R 1 e

= + + v v u g
3.7
This solution can be used to formulate an exponential numerical scheme for the particle
equation of motion with the following notation: the change in time is denoted by t, the value
of a property q at time t
n
is denoted by q
n
(known value), and the value at time t+t=t
n+1
is
denoted by q
n+1
(unknown value to be solved). In this case, the updated velocity and position
values can be expressed as:

( )
( )
( )
( )
( )
n n
p p
n
p
t/ t/
n+1 n n n
@p p
t/
n+1 n n n n n n
p p p p @p p
e 1 e 1 R
1 R 1 e 1 R t

= + +

= + +

v v u g
x x + w g u g

3.8a

3.8b
This one-step exponential scheme is popular is exact for Stokesian particles in a uniform
steady flow (for which particle response time and fluid velocity are constant). It can also be
employed at moderate Reynolds number conditions and weakly varying flows by assuming
that u
@p
and
p
vary only weakly through the time-step and can be estimated by the value at t
n
.
However, the time-step should be chosen to be sufficiently small so that the truncated terms
have little influence. Since the trajectory dynamics are dictated by the particle response time,
this infers:


p
t <
3.9
Advanced integration schemes to handle highly non-linear drag, strong flow acceleration along
the particle path, lift, collisions, history force, etc. are discussed in 7.2.3. However, Eq. 3.8 is
reasonable and useful for simple flows where drag and gravitational forces dominate.
The gravitational effects can be shown more clearly by incorporating the terminal velocity
using Eq. 3.2 and again assuming linear drag with uniform steady flow:
( )( )
p p
t / t /
0 @p term
e 1 e

= + + v v u w
3.10
Thus, the particle relative velocity tends to the terminal velocity for long times, i.e. t
p
. Thus,
particles will take longer to respond to continuous-phase velocity (and any of its changes) as
p

increases. Also, particles with very small
p
will adapt more quickly to their equilibrium state
(e.g., will tend to always move at a relative velocity equal to their terminal velocity).
If we neglect gravitational effects (g=0), then the solution for the relative velocity is simply
related to the initial conditions and the particle response time:
97
( ) ( )
term o term p
w w w w exp t / =
3.11
At long-times, the particle will reach a terminal velocity with respect to the fluid ww
term
. If
we neglect gravitational effects (g=0), this reverts to w=w
o
exp(-t/
p
) indicating that the relative
velocity will decrease exponentially at a time scale
p
, such vu for long-times, e.g., the
relative velocity will reduce by about 95% at t=3
p
. These trends are illustrated in Fig. 3.2,
indicating that the particle response time-scale dictates the time it takes for a particle to adapt
to the surrounding flow based on drag.


Particle Stokes Number

If the continuous-phase flow field varies along the particle path, the relative magnitude of
the particle response time can be normalized by the time-scale associated with the continuous-
fluid velocity changes (
f
). This ratio is defined as the particle Stokes number

2
p p f
f f f
( c )d
St
18 f

+
=

3.12
The Stokes number is an important measure of the degree of correlation which may be
expected between the particle velocity and the fluid velocity field for a given fluid time-scale.
For a very small Stokes number (St0), the time for the particle velocity to react to the
changes in continuous-phase velocity is negligible such that v(t)u
@p
(t)+w
term
(based on Eqs.
3.10 and 3.11). As the Stokes number increases, the particle will take longer to accommodate
to changes in the surrounding fluid velocity. At a Stokes number of order of unity (St~1), one
would expect v to be both substantially influenced by u but also lag in phase with u. For a
very large Stokes number (St1), the particle inertia is too great to be significantly affected by
the continuous-fluid unsteadiness. In general, the continuous-fluid time-scale may contain
both overall macroscopic time-scales as well as smaller local time-scales (i.e., several
frequencies). In such cases, one Stokes number is not sufficient to describe all the particle
dynamics. In the following, the time-scale associated with the overall continuous-phase flow
domain is first considered, while more local time scales associated with turbulence will be
considered in the next section.


Domain Stokes Number

For an unsteady flow, the domain time-scale (
D
) can be associated with the overall period
of the flow oscillations. For example, flows can be dominated by fluid-structure interactions
(as in oscillations of a piston in an internal combustion engine, or aeroelastic flutter), pressure
disturbances (as in acoustic or shock-wave oscillations), or pulsatile flow (as in blood flow or
wave-induced motion). For a steady flow in the Eulerian reference frame,
D
can be
characterized by a macroscopic length-scale (D) and velocity-scale (u
D
):

D D
D/ u
3.13
98
The macroscopic scales for the continuous-phase are based on the relevant flow conditions
over which flow changes may occur. For example, D for an external flow may be set as the
object length or width and u
D
set as the free-stream speed. For a channel flow (Fig. 3.3a) D
may be set as a divergence length or channel diameter and u
D
set as the mean velocity. While
the choice of D and u
D
is somewhat arbitrary, use of consistent characterization will allow
comparisons between two different flow or two different particles.
Once the macroscopic time-scale is defined, the macroscopic Stokes number is simply

D p D
St /
3.14
This Stokes number is important in determining whether (and to what degree) particles will
generally follow the continuous-phase streamlines. In particular, the degree of this tracking is
based on whether the St
D
is small, intermediate, or large. To illustrate this dependence, Fig.
3.3b shows the influence of St
D
on particle velocity in the diverging channel of Fig. 3.3a
assuming gravitational and initial condition effects are negligible (w
term
/u
D
0 and w
o
/u
D
0).
For St
D
1, the particle velocities will be similar to that of the continuous-fluid since the
particles adjust very quickly to the local fluid conditions and follow their streamlines. An
example of this is a very small water (fog or cloud) droplet which would follow the flow
streamlines closely. For St
D
1, a particle would not significantly adjust to an evolving channel
velocity due to its large response time (
p
). For example, a large hail particle or a bullet with
very high inertia would only be weakly affected by variations in the surrounding flow. In
between these two extremes, a particle with St
D
~1 will be significantly affected by the
continuous-fluid but will have substantial differences between v and u. An example of this
could be medium-sized water drops which will have significant inertia but will be affected by
the surrounding flow. The response of the particle dynamics to integral-scale turbulence is
thus primarily a function of Stokes number and may be summarized as:

St
D
1 particle motion weakly affected by the macroscopic flow
St
D
~1 particle motion substantially modified by the macroscopic flow
St
D
1 particle motion closely follows the macroscopic flow
3.15
Quantitative values for these criteria depend on the flow geometry and the specific definition
used for the domain scales, e.g., whether D is based on channel length or channel width, etc.
While the time-scale ratio given by the Stokes number is the most important parameter for
particle dynamics, other parameters are important and can dominate under certain conditions.
Two such parameters include the velocity-scale ratio (w
term
/u
D
) and the acceleration parameter
(R). These three parameters are consistent with a non-dimensionalization of Eq. 3.7, and they
can be related to the macroscopic Froude number (Eq. A.62a) by:

( )
2
D D
D
term D
u St
Fr | 1 R |
gD w / u
=
3.16
While the Froude number is independent of the particle properties, but the RHS expression
indicates it control which particle parameters are likely to be more important. For example,
high-speed flows with Fr
D
1 are more likely to have particles with significant Stokes numbers,
while low-speed flows with Fr
D
1 are more likely to have particles with significant velocity
ratios. For linear drag, the particle response time is independent of Re
p
so that macroscopic
99
particle dynamics will then be uniquely determined by the initial particle conditions and three
of the following four non-dimensional parameters (R, Fr
D
, St
D
, u
D
/w
term
).
If the particle Reynolds number is not small, then the Stokes correction factor (f) is
dependent upon Re
p
as are the instantaneous
D
and St
D
. For an overall characterization, one
may define a path-averaged quantity ( q ) using an initial time (t
0
) and a time shift ()as:

0
0
t
t
1
q q t
+

d 3.17
Using a time period based on the domain time-scale, the path-averaged Stokes number can be
expressed in terms of the Stokes correction by Eqs. 3.1 and 3.12:

( )

0 D 0 D
0 0
2
t t
p f
D
D p 2
t t
D D f D
( c )d
1 1
St St t t 1/ f
18
+ +

+
= = =





d d
3.18
If the average particle relative velocity is similar to its terminal conditions, then:

term
D
D,term
D
w
St St
g 1 R
=


3.19
This form is convenient for a given terminal velocity.
If the average relative velocity differs substantially from its terminal velocity, Eq. 3.19 is
no longer valid then an average of the inverse of the correction factor should be employed. In
this case, it may be convenient to express the path-average in terms of the Reynolds number
variations. If the initial conditions is given by Re
p,o
and the final state is given by
Re
p,ref
=w
ref
d/
f
where w
ref
is the final relative velocity, then the path-average on the RHS of Eq.
3.18 can be expressed as:

( )

( ) ( )
p,ref
p
p,o
p
p,ref p,o
Re
Re
Re Re
Re Re
1
1/ f 1/ f

d
3.20
If Re
p
<1,000 so that the Putnam fit (Eq. 1.53) is reasonable and Re
p,ref
Re
p,o
, an analytical
solution can be obtained:

1/ 3
p,ref 1 1
2/ 3
p,ref p,ref
Re
1 6
f 18 tan
Re Re 6








3.21
This resulting

D St can then be used to characterize the response of particles to a given


continuous-phase flow field.
An application of this result is that of cloud droplets which may collect on a wing moving
at speed u

as shown in Fig. 3.4a. The small drops with a rapid response time will tend to
move around the airfoil with the air-flow streamlines with no impact. In contrast, large drops
will impinge directly and intermediate drops will have some trajectory deflection but some of
these will impact. It is important to know which drops will impact since they can then freeze
and adhere to the airfoil, which can form a safety hazard (Fig. 1.16). The probably of impact
can be quantified in terms of an impact efficiency which can be related to the particle Stokes
number. For example, particles with St
D
1 will tend to an impact efficiency of zero while
100
those with St
D
1 will tend to impact efficiency of unity. The macroscopic time-scale (
D
) for
the continuous-fluid velocity variations can be simply defined as the ratio of airfoil chord
length to flight speed. However, the particle response time (
p
) for non-linear drag will vary in
throughout the trajectory so a path-average is needed. For this, one may assume that that the
droplet is initially at a terminal velocity (Re
p
=Re
p,term
) but transitions to a high relative velocity
nearly equal to the airfoil velocity at the final state (Re
p,ref
Re
p,
=du

/
f
). Further assuming
Re
p,
Re
p,term
, one may employ Eqs. 3.18 and 3.21 to describe the path-averaged Stokes
number and use this to assess the likelihood of impact for a given airfoil shape and angle of
attack (Langmuir, 1946). As shown in Fig. 3.4b,

D St correlates well with the impact


efficiency (which is the fraction of droplets per frontal area which will impact) independent of
the reference Reynolds number. This indicates that the path-averaged Stokes number
reasonably scales the particle dynamics.



3.3. Turbulent Stokes Numbers

Since turbulence contains a wide range of time scales, there are a correspondingly wide
range of Stokes numbers (A.5). At the micro-scales where most of the dissipation occurs,
homogeneous turbulence away from the walls is characterized by the Kolmogorov scales (Eq.
A.93) while non-homogeneous turbulence close to walls is characterized by the inner scales
(Eq. A.100). The corresponding micro-scale Stokes numbers can be defined as:

p
2
p p fr
fr f
St
u
St


=


3.22a

3.22b
Since these represent sensitivity of particle dynamics to the fastest time-scales of the
turbulence, particles with small micro-scale Stokes numbers effectively can be considered as
fluid tracers with respect to turbulent dispersion. If the micro-scale Stokes number is very
large, then the particles are unaffected by the small-scale turbulence and may instead be
affected by the large-scale turbulence.
The largest turbulent time-scales are the integral-scales (Eq. A.78) which contain most of
the turbulent energy. The corresponding integral-scale Stokes number is defined as:

p
St

L

3.23
Note that the denominator is the integral time-scale along the Lagrangian fluid path (Eq.
A.78a). This is most appropriate for small low-inertia particles since these tend to follow the
instantaneous fluid path (Fig. 3.5). For large high inertia particle which tend to move with the
mean flow, the moving-Eulerian time-scale would be more appropriate but
L
and
mE
are
typically similar (Eq. A.79) so that this distinction is not critical so that the above definition is
reasonable.
101
The integral Stokes number is perhaps the most important to describe turbulent diffusion
since the large turbulent structures yield the greatest velocity and position fluctuations. To
demonstrate its influence, one may consider particle dynamics in a free-shear layer since this
flow has well-defined large-scale structures (Fig. A.5c). A schematic of the various particle
trajectories as a function of St

are shown in Fig. 3.6a. A particle with very small inertia


(St

1) will be strongly affected by the turbulence so that its trajectory will closely follow the
large-scale structures of the fluid. In the other extreme, a particle with very large inertia
(St

1) will continue on its trajectory with only a small influence from any turbulent structures.
In between these extremes, intermediate response time particles (St

~1) will have paths


substantially influenced by the vortices but with significant deviations from the fluid paths.
Thus, the response of the particle dynamics to integral-scale turbulence may be summarized as:

St

1 particle motion weakly affected by the turbulent structures


St

~1 particle motion substantially modified by integral-scale structures


St

1 particle motion closely follows integral-scale structures


3.24
This is similar to the criteria used for the macroscopic features (Eq. 3.15).
The particle density ratio also has an effect on the particle trajectories, particularly for
intermediate Stokes numbers. Very-heavy particles ( * 1) with St

~1 will tend to be
centrifuged to the outer edges of the vortices and will also tend to collect in the high shear
regions. This is seen in Fig. 3.6b which includes an experimental visualization of solid
particles dispersing in a gas shear flow. In contrast, very-buoyant particles (
*
p
1) will tend to
move to the cores of vortices (which have low pressure) as demonstrated in Fig. 3.6c. This is
due to the fluid stress forces which cause particles to move in the opposite direction of a
pressure gradient (Eq. 1.69) and becomes important when this effect is stronger than that due to
hydrostatic pressure gradients. To characterize this non-dimensionally, one may relate the
pressure gradient of a turbulent eddy pressure to the turbulent velocities with Eq. A.43 and use
this relationship to define an integral-scale Froude number (analogous to the domain Froude
number of Eq. 3.16):

( ) ( )
2
f
2
p u
u u
Fr
g g

=


L

3.25a

3.25b
We note that the approximation for the RHS of Eq. 3.25b employs the integral time-scale and
assumes that c

1 (as defined by Eq. A.90). It also shows that the turbulent Froude number
can be expressed as a ratio of hydro-dynamic acceleration (u

) to the hydro-static
acceleration (g). As such, the behavior seen in Fig. 3.6c is consistent with a substantial Fr

. In
contrast, a bubble rising in very weak turbulence (Fr

0) will be dominated by the balance of


drag forces and hydrostatic effects and thus ignore movement to vortex cores and instead move
along vertical lines at the terminal velocity.
It is also instructive to note that the Stokes number and the Froude number can be related
(by Eqs. 3.2, 3.23 and 3.25) to the non-dimensional terminal velocity defined as:

,term * term
term
St 1 R
w
w
u Fr

=
3.26
102
The value
*
term
w is often called the drift parameter since it characterizes the particle deviation
from the continuous-fluid fluctuations. Since the Froude number only a function of the
continuous-phase properties, this indicates that the Stokes number and the terminal velocities
ratios have the same dependence on particle size, density, etc. for a given flow field. As noted
for the macroscopic dimensionless values (Eqs. 3.14 and 3.16), the above relationships indicate
that the particle response to integral-scale turbulence is primarily a function of four non-
dimensional parameters (Fr

, St

,
*
term
w , R), of which three are independent.
Turning now to the smallest scales of the turbulence, we consider the impact of
Kolmogorov Stokes numbers on particle dynamics. Particles with St

1 can be considered as
nearly immiscible fluid tracers since their terminal velocity (and relative velocities) will be
small with respect to the velocity fluctuations of the even smallest eddies. Thus, these micro-
particles will act as fluid tracers with vu so long as molecular motion is ignored. Particles
with St

~1 respond only approximately to the small-scale vortex features, resulting in a


significant relative velocity fluctuations. These scales correspond to high vorticity and fluid
stress fluctuations so that St

~1 can lead to a turbulent bias from the terminal velocity under


certain conditions. Particles with St

1, will not be significantly affected by the smallest


energy dissipating scales.
Similar arguments can be applied to wall-shear flows in terms of St
+
(corresponding to the
smallest-scale features). For example, Fig. 3.7 shows a flow with St
+
~1 (and St

1) whereby
the particles faithfully follow the large-scale turbulent features (such as boundary layer
ejections) far away from the wall (Fig. 3.7a), but show preference to concentrate in low-speed
streaks for thin turbulent features close to the wall (Fig. 3.7b). The results shown in this figure
are based on Direct Numerical Simulation (DNS) which resolves all the length scales and
frequencies of turbulent flow for the continuous-phase fluid. The DNS technique is an
effective supplement to experimental measurements to help understand multiphase turbulent
flows (though it is limited to modest Reynolds numbers will be discussed in 4.3.1).
Note that the coupling criteria used for the integral-scale turbulence given by St

values in
Eq. 3.24 can be similarly applied to the Kolmogorov-scale turbulence based on by St

or the
inner-scale turbulence based on by St
+
. These different regimes can be further used to
understand and model the particle concentration patterns and spread rates caused by the
continuous-flow turbulence, i.e. turbulent dispersion. This subject is explored in more detail
in the following two sections. Afterwards, 3.6 will discuss the related topic of Brownian
diffusion where very small particle can diffuse owing to the random arrival of fluid molecules.


3.4. Turbulent Particle Dispersion

Turbulent dispersion for two-phase flows refers to the evolution of particle, droplet, or
bubble concentrations due to underlying turbulent features of the continuous phase. This
process is very important for many engineering flows since rapid distribution of injected
particles throughout a domain is often most effectively accomplished by turbulent dispersion.
The power of turbulence is thus employed in many energy systems to rapidly disperse particles.
For example, piston and turbine engines often introduce higher turbulence levels so that
injected fuel droplets will be distributed throughout the combustion chamber to ensure more
uniform distribution of the fuel vapor to improve efficiency and reduce pollution. Similarly,
103
bubble columns or plumes (Fig. 1.5) are typically turbulent to promote high mass and chemical
transfer rates between the two phases. Plasma sprays also use turbulence to ensure uniform
coating (Fig. 1.9). Many of the other particle distribution phenomenon discussed in 1.2 are
also a result of turbulent dispersion. Therefore, it is important to understand and predict the
spread of particles and droplets due to turbulence.
In many ways, particle mixing is similar to the continuous-phase species mixing discussed
in A.5.5. In particular, the instantaneous spread of particles is related to the underlying
turbulent structures and the instantaneous gradients of particle concentration. However, it can
also be considered in the mean with respect to average turbulence intensity and mean particle
concentration gradients. Because of this, particle distribution due to turbulent dispersion is
often considered in terms of two aspects:
1. preferential concentration, which refers to the spatio-temporal details of the particle
concentration field generated by instantaneous and local features (such as vortex
structures) of the turbulence; and
2. mean diffusion, which refers only to the overall mean (time-averaged) spread rate of
particles caused by the turbulence.
The difference between these two aspects is illustrated in Fig. 3.8 for the case of particles
injected into a turbulent boundary layer. The instantaneous particle positions associated with
preferential concentration (also called structural dispersion) is depicted in Fig. 3.8a. Here
it can be seen that local variations in the particle distribution give rise to complex patterns, as
also seen in Fig. 3.7. In contrast, Fig. 3.8b depicts the corresponding time-averaged particle
concentration associated with mean diffusion where the mean particle cloud progressively
spreads out laterally as it convects in the streamwise direction. The differences between
preferential concentration and mean diffusion are similar to those between local mixing and
overall diffusion of a fluid tracer in a turbulent flow. However, it should be kept in mid that
the particle response time (via the effective Stokes number) will lead to additional complexities.
The remainder of this section discusses some aspects and implications of preferential
concentration along with basic definitions for mean particle diffusion, while 3.5 discusses the
latter in substantial detail.


Preferential Concentration

The description preferential concentration was termed by Eaton & Fessler (1994) to explain
the phenomenon of particles tending towards specific types of turbulent features. As discussed
in the above boundary layer example, the spatio-temporal variations in the particle distribution
associated with preferential concentration can be substantial. The spatial variations can be
especially noted in Figs. 3.6-3.7 where pockets of high and low particle concentration are
found to occur after some time. As such the instantaneous particle concentration can differ
substantially from the mean particle concentration profile. These differences are a function of
the turbulence intensity, turbulence structure and the particle inertia (via Stokes numbers). In
particular, preferential concentration is correlated with the particle Stokes number based on the
smallest turbulence scales, e.g., based on St
+
in the case of near-wall regions (Fig. 3.7) and
based on St

in the case of free-shear flows or regions away from the wall. For example, Fig.
3.9 shows a snapshot of the instantaneous distribution particle concentration within a plane for
St

~1 and approximately HIST flow (A.5.2). This snapshot demonstrates a strong correlation
104
between regions of low particle concentration and regions with high continuous-phase vorticity,
consistent with centrifugal dispersion of particles. Other experiments at St

~1 but with
anisotropic non-homogenous turbulence also yielded similar results (Fig. 3.10a). For particles
with longer response times (St

~10 and St

~1), experiments in a stirred-box configuration in


nearly zero-gravity show that particle concentration tendencies are still present but are not as
distinct (Fig. 3.10b). Particles with still longer response times (St

~100 and St

~10) tend to be
homogeneously distributed with respect to the turbulent integral-scales indicating that their
inertia is too high to have significant correlations with even the largest turbulent structures.
The effects of preferential concentration can impact the macroscopic characteristics of the
flow and may be critical to the efficiency (and simulation) of energy systems. For example,
preferential concentration in an internal combustion engine chamber can cause droplets to
cluster in certain regions instead of being distributed uniformly. Droplet evaporations in
such clustered regions will lead to local fuel-rich pockets which can dramatically change the
local chemical kinetics and overall efficiency. To avoid this, Sankaran and Menon (2002)
showed that preferential concentration of fuel droplets can be reduced by introducing swirl-
type combustors. The clustering phenomenon can also lead to increased particle-particle
collision which may explain how rain-clouds are quickly formed. In particular, the rapid
coalescing of small cloud droplets of about 10-100 microns into rain droplets of about 200-
1000 microns has been related to turbulence and preferential concentration (Sundaram &
Collins, 1997).
Preferential concentration can also lead to changes in the mean relative velocity of particles,
by at least two mechanisms: preferential bias and clustering bias. Preferential bias is a one-
way coupling mechanism owing to particle trajectories favoring certain paths and flow
structures. Clustering bias is a two-way coupling mechanism whereby high particle
concentrations in certain regions lead to clusters which act approximately as a single larger
entity. For heavy particles, this concentration can impart a local downwash on the flow so that
the accompanying particles will fall faster than that of an isolated particle (while buoyant
particles may rise faster due to an imparted upwash). These bias mechanisms will be discussed
in 3.5.5.


Mean Particle Diffusion

A common characterization of mean turbulent diffusion is the spatial spreading of particles
(both longitudinally and laterally) from an injection point. With respect to the mean spreading,
it is helpful to define a particle cloud as a group of particles which originated from a single
point in space but were released at different times (Fig. 3.11). Turbulent events will lead to a
unique particle trajectory for each release time. Each path-averaged velocity can be defined as:

0
t
p p p
t
0
1
( , t) t
+

d v v x
3.27
In this equation, is the time period since the initial release time (t
0
). The resulting particle
position is also a function of the duration time and the release point x
0


p p o
( ) = + x v x
3.28
105
Next, we define the notation, q as the ensemble average over a statistically large number of
particles (N
p
), i.e.

p
p
N
1
k
N
i k
q q




=

3.29
One may then define the ensemble-averaged particle position, the particle position deviation
from the mean, and the ensemble-averaged variation about the mean as follows:

( ) ( )
( ) ( ) ( )
( )
p
p
p
p
N
1
p p,k
N
k 1
p p p
N
1
p p pk pk
N
k 1




=




=


x x
x x x
x x x x

3.30a

3.30b

3.30c
For convergence, this ensemble-average should consider releases distributed over a long time
period compared to

. As illustrated in Fig. 3.11, the ensemble-averaged particle position


characterizes the mean trajectory while the ensemble-averaged variation characterizes the
mean spread of the particle about the mean trajectory at a given time. To incorporate
anisotropic diffusion, the mean spread can be given in tensor form as
pi pj
x x and the
corresponding spread rate of the particle cloud is given as:

pi pj pi pj
1 1
p,ij
2 2
x x x x
t

=

d d
d d



3.31
The magnitude of
p,ij
is also known as the particle diffusivity and can be a result of
turbulent diffusivity (3.5) or Brownian diffusion (3.6). In order to model the turbulent
diffusivity, it is helpful to first consider the turbulent diffusion of fluid tracers (particles with
zero relative velocity and zero inertia), as discussed in the next section. This will be followed
by the additional consideration of inertial and gravitational effects on mean turbulent diffusion
in a following section.


3.5. Turbulent Diffusion and Bias of Particles
3.5.1. Diffusion of Mean Fluid Tracers

In order to eventually characterize the mean turbulent diffusion of particles (which can be
quite complex), it is helpful to first consider the mean diffusion of a fluid tracer which has zero
inertia. The fluid tracer can be thought of as an infinitely small immiscible particle, i.e., it has
a Stokes number of zero and is not subjected to molecular diffusion. The spread of a single
fluid-tracer can be represented in terms of instantaneous velocity and position (similar to those
defined for a general particle by Eqs. 3.27 and 3.28) while the mean diffusion of many such
tracers can be related to the means square of perturbations from the ensemble-averaged path:
106

( )( )
0
i fi
St 0
t
fi fi fi,0 fi fi,0
t
0
fi fi fi fi fi fj
1 1
turb,ij
2 2
v u
x ( ) u dt x u x
x ( ) x ( ) x ( ) x ( ) x x
t t

+ = +

=

D D
D D

3.32a


3.32b


3.32c
Taylors theory for diffusion (Reeks, 1977) assumes homogeneous turbulence (i.e. the mean
fluid statistics are the same at all positions in space) and relates the above spread rate to the
integral of the fluctuating velocities with a time shift:

0 0
t t
turb,ij i 0 j i j ,ij
t t
0 0
( ) u (t )u (t) dt u u ( )dt
+ +


L


3.33
The RHS expression was obtained by employing the Lagrangian correlation of velocity (Eq.
A.77a) and assumes that the ensemble-averaged turbulent velocity fluctuations and correlations
among are the fluid-tracers are given by the average values, which is consistent with the
assumption of homogeneous turbulence (A.5.2).
If one further assumes that the turbulence is isotropic and thus has negligible cross-
correlations (Eq. A.74), then the fluid-tracer diffusion also has the same properties, i.e.
( )
turb t ,11 t ,22 t ,33 turb,ij
i j
>


3.34
A similar statement can be made about the Lagrangian correlation function (Eq. A.76), so that
the resulting diffusion can then be written in terms of the turbulence strength (Eq. A.72) as:

0
0
t
2
turb
t
( ) u (t) dt
+

L


3.35
Based on this equation, one may readily identify two asymptotic values for the fluid-tracer
diffusivity: a short-time limit defined by

<
L
and a long-time limit defined by

>
L
.
For short-times,
L
is approximately unity since it is near the autocorrelation value (see Fig.
A.12) and the diffusivity is then:

2
turb
( ) u



for
L
3.36
This indicates that the short-time diffusivity will increase linearly with time. For long-times,
the integral in Eq. 3.35 will simply become the Lagrangian integral time-scale of the velocity
fluctuations as defined in Eq. A.78a. Therefore, the long-time diffusion rate (denoted by
turb,
) is linearly proportional to the turbulent kinetic energy and the integral time-scale:

2
turb,
u


L

for
L

3.37
This result also indicates that the long-time diffusion of fluid tracers becomes independent of
time (i.e. the particles spread out at a constant rate) which is a result of diffusion being
influenced by the entire spectrum of turbulent scales. This limit is often the most relevant
since mean characteristics are typically realized at time scales much greater than the turbulent
107
integral-scale. For intermediate times, the particle spread rate can be obtained by assuming
that the correlation function has an exponential dependence (Eq. A.80a) so that
( )
2
turb
/
( ) u 1 e



=
L
L




3.38
This relationship is shown in Fig. 3.12a and recovers Eqs. 3.36 and 3.37 in the short- and long-
time limits. It also shows that turbulent diffusivity is proportional to the turbulent kinetic
energy and reaches a long-time equilibrium.
To obtain the positional rms of fluid-tracers as a function of time, the above equation can
be integrated for each coordinate direction as:
( )
2
f f f f f f
/
x x y y z z 2u 1 e




= = =

L
L L


3.39
Note that the diffusion for a fluid tracer in HIST is isotropic so that fluid-tracer mean cloud
volume will grow like a sphere about the trajectory dictated by the mean velocity. The RHS
expression indicates that the short-time growth rate is quadratic (proportional to
2
) while the
long-time rate is linear (proportional to ). The theoretical behavior is shown in Fig. 3.12b and
compares quite favorably with measurements and DNS data for which the HIST approximation
is reasonable. For example, a two-dimensional flow with mean velocity only in the x-
direction ( )
x
u u = will result in
f f
y z 0 = = . If the strengths of the velocity fluctuations
are anisotropic but the flow is homogenous and the integral time-scale is a constant, the spread
is a tensor given by:
( )
fi fj i j
/
x x 2u u 1 e




=

L
L L


3.40
As such, the diffusion in each direction will be proportional to the strength of the fluctuations
in that direction. For example, the anisotropy in the velocity fluctuations for the inner portion
of a turbulent boundary layer shown in Fig. A.10 will lead to anisotropic diffusion
with
f f f f f f
x x z z y y > > , i.e. a particle cloud would grow like an ellipsoid.
The presence of non-zero cross-correlations (i.e.
i j
u u 0 for i j ) can also lead to a lateral
drift with respect to the mean velocity path. To quantify this drift, Hinze (1975) derived a
general equation for the ensemble-averaged mean velocity by assuming long-times and weak
turbulence. A similar simplified equation can be obtained for 2-D flow with mean velocity
only in the x-direction:

f,x x x x x
f, y y y y y
u u u u u
u u u u u
= = +
= = +


3.41a

3.41b
Now consider the path-average lateral movement of a fluid tracer in this flow

0 0
t t
y y y x y x y f
f f x
2 2
t t
f x x x x x x 0 0
u u u u u u u y y y u 1 1
dt 1 dt
x x x u u u u u u
+ +
+
=

+



3.42
This equation first approximates the ensemble-average with the path-average (reasonable for
many particles for long-times), and then approximates the integrand using the approximations
of Eq. 3.41 which are equivalent to assuming that the product of the mean velocities is much
108
smaller than the cross-correlation, i.e. (
x y x y
u u u u < ). As an example, consider the jet flow
shown in Fig. 3.13a for which tracers are released below the centerline so that the cross-
correlation is non-zero and the above assumptions are reasonable. The measured concentration
distribution in the lateral direction is shown in Fig. 3.13b and indicates an approximately
Gaussian distribution whose width is consistent with Eq. 3.40 and whose shift is consistent
with Eq. 3.42. As shown in the figure, this net migration is generally small compared with the
overall diffusion spread rate. It should be noted that gradients in the turbulence intensity can
also lead to a net migration, and this aspect will be discussed in 3.5.5.

3.5.2. Diffusion of Finite-Inertia Particles

Correlations of Fluid and Particle Velocities along the Particle Path

To assess particle diffusivity due to the surrounding turbulence, it is important to define the
correlation and integral time-scale of the continuous-phase fluctuations along the particle path:

0
i @p, j p
f @p,ij
i j
@p,ij f @p,ij
t
u ( , t)u ( ( ) , t )
()
u ( , t)u ( , t)
()

+ +


x v x
x x
d
NSI

3.43a


3.43b
In Eq. 3.43a, the correlation may be anisotropic and there is no summation of indices (NSI) for
the RHS. Similar to the fluid-tracer correlation function (Eq. A.75) and diffusion definition
(Eq. 3.33), the Lagrangian particle correlation function and associated turbulent diffusivity can
be written in terms of the velocity fluctuations, i.e.

0
i j j p
p,ij
i j j
t
p,ij i 0 j
t
0
v ( , t)v ( , t )
()
v ( , t)v ( , t)
( ) v (t )v (t) dt
+
+ +


=

x x
x x
x
NSI


3.44a


3.44b
In the fluid-tracer limit, the above correlation functions and the integral time-scale revert to
that of Eqs. A.77a and A.78a. As such, the particle diffusivity reverts to the fluid-tracer
diffusivity of Eq. 3.33. To quantify the non-tracer effects, it is helpful to first define a
normalized particle diffusivity ratio,

, as the ratio of the particle turbulent diffusivity to the


fluid-tracer turbulent diffusivity:

*
ij p,ij turb,ij
/
NSI



3.45
In the most complex case of NAsT (Non-homogeneous Anisotropic Turbulence), the diffusion
ratio tensor can be a function of particle inertia, relative velocity, and location. In the simplest
case of HIST with particles of negligible inertia and relative velocity, the diffusion ratio will
simply be unity, i.e.,

1 as St

0. In the following, particles in HIST will be considered


for arbitrary inertias but with small relative velocity compared to that of turbulence, e.g., a
109
negligible terminal or injection velocity compared to u

. Later sub-sections will account for


relative velocity effects.


Zero Eddy-Crossing Particle Diffusivity

A theoretically important limit for examining turbulent diffusion is that with finite particle
inertia (and finite Stokes number) but with negligible mean relative velocity compare to the
turbulence velocity. In this condition, the particles do not cut-through the turbulence in any
preferred direction and this is called the zero eddy-crossing limit. This limit is reasonable if
terminal velocities and injection relative velocities are small compared to the turbulent velocity
and there are no significant spatial variations in the mean velocity and turbulence. For such a
condition, the ensemble-averaged particle motion, i.e. 0 w . An arbitrary quantity (q) that
satisfies these conditions will be denoted as follows:
( )
term rms inj rms w 0
q q w / u 0, w / u 0






3.46
This limit is reasonable at long-times for small particles in strong turbulence or for particles in
micro-gravity conditions. For this limit, one may define a normalized continuous-phase
turbulence intensity based on the particle mean relative velocity (which is approximately the
inverse of the normalized terminal velocity of Eq. 3.26):

*
*
term term
u u 1
u
w w w

=
3.47
Thus, the limit defined by Eq. 3.46 is consistent with
*
u

or
*
term
w 0 .
For the zero eddy-crossing limit, the particle velocity correlation function is approximately
a scalar and can be related to the particle velocity fluctuations based on Eq. 3.44 as

0
0
t
p p,ii i i p
t
( ) ( ) v ( , t)v ( , t ) t
+
= + +

x x d x


3.48
Based on this equation and the analysis used for Eq. 3.36, the short-time particle diffusion is
consistent with the particle velocity correlation function of approximately unity, so that:
( ) ( )
2 2
*
p turb rms rms
/ v / u

for t
L
3.49
Thus, the short-time diffusion ratio is proportional to the ratio of particle velocity fluctuations
to those of the fluid.
To determine the diffusion ratio at intermediate- and long-times, one must integrate the
particle equation of motion. To make this problem analytically tractable, a linear solid particle
drag is assumed (f=1) and the lift and history forces are assumed to be negligible. Note that
this latter assumption may not be appropriate for small density ratios (i.e. bubbly flows),
especially with respect to the history force effects but this will be addressed in a later sub-
section. Based on Eqs. 1.77 and 3.1 and the above assumptions, the particle equation of
motion is:
110

p p
R
t t
+ = +




d
d
d d
@p @p
u u
v v

3.50
For c

= , R becomes 3/(1+2
*
p
) based on Eq. 1.81a. For the zero eddy-crossing mean limit
of Eq. 3.46, the ensemble-averaged particle trajectory will be the same as a fluid tracer,
i.e.
p o
= +
@p
x u x . Hinze (1975) approximated the continuous-phase velocity fluctuations as
a Fourier integral of the turbulent energy spectrum and then applied the above particle equation
of motion to relate
f @p
to
p
for HIST. For convenience, it is helpful to define an effective
Stokes number in terms of the fluid time-scale along the particle path (Eq. 3.43b):

p 2
@p
@p
St





3.51
Based on this definition, the above assumptions, and Eq. 3.50, Hinze obtained the particle
diffusion (after some calculation) for zero eddy-crossing as:

( ) ( ) ( ) ( )
2 2 2 2
@p @p @p p p, w 0
2 2
@p @p
( ) R St exp / St 1 R exp /
1
u 1 St



=


3.52
This result can be normalized by Eq. 3.38 to give the turbulent particle diffusion ratio:

( ) ( ) ( ) ( )
( )( ) ( )
2 2 2 2 2
@p @p @p @p p
*
2
@p @p
1 St St R exp / St 1 R exp /
/ 1 St 1 exp /


+ +
=


L L
3.53
As such, the particle Stokes number influences the diffusivity ratio which is often described as
the inertia effect.
For long-times, the terms in the square brackets of Eq. 3.53 approach unity. As such, the
long-time particle diffusion ratio with zero eddy-crossing is given by :

*
, w 0
@p, w 0

L

3.54
The influence of particle inertia (Stokes number) on
@p, w 0
determines the long-time
diffusion. For small inertia particles, the particle path is very close to the fluid path so that:

*
, w 0
1


for
St 1

<
L

3.55
In the limit of very high inertias, the particle path will tend to the moving Eulerian path (shown
in Fig. 3.5) so that so that
@p

mE
(Reeks, 1977; Stock, 1996). Thus

*
, w 0

mE
L
for
St

L

3.56
Based on Eq. A.79,


L mE
for HIST, so that long-time diffusion reverts to approximately
unity for both Stokes number limits. As proposed by Csanady (1963), this further suggests:
111

@p, w 0

L mE

3.57
From this, it is reasonable to assume that the influence of Stokes number on long-time
diffusion is generally weak for HIST with negligible gravity. However, Wang & Stock (1993)
noted that significant differences between
L
and
mE
can occur. To account for this, they
developed a model to describe the transition of
@p
between these two time scales as a
function of particle Stokes number. Thus, Eq. 3.57 may only be qualitatively correct in some
cases, though it is commonly used.
Assuming Eq. 3.57 is appropriate, Eq. 3.53 simplifies to:

( ) ( ) ( )
( ) ( )
2 2
p
*
w 0
2
1 R exp / St exp /
1
1 St 1 exp /



+

= +


L
L
L
L

3.58
The dependence of this diffusivity ratio as a function of path duration for high density ratio
particles (
*
p
1 > ) is plotted in Fig. 3.14a for various integral-scale Stokes numbers. In general,
the diffusion ratio increases with both path duration and inertia. At very long-times compared
to the particle inertia (
p
/ St , i.e.

> >
L L
), the diffusion ratio approaches unity. As such,
the diffusion for very small particles or for very long-times is consistent with the fluid tracer
limit (i.e. it becomes independent of the particle inertia). In the limit of very large particle
inertia ( St 1

>
L
) at short-times, the diffusion ratio tends to zero for short-times consistent
with the inability of the particle to respond to the turbulent fluctuations. If the Stokes number
is taken as the path-averaged value (

St
L
) which includes the Stokes correction factor, these
trends are consistent with experimental results for non-linear drag particle diffusion of Crowe
et al. (1988) and Wang & Stock (1993).
It is interesting to consider the impact of density ratio for a Stokes number of unity, for
which Eq. 3.58 can be evaluated using the rule of LHopital. The result is shown in Fig. 3.14b
and indicates the influence of
*
p
at short-times is substantial. In particular, heavy particles
will spread less for short-times, neutrally-buoyant particles spread at the same rate as fluid-
tracers for all times, and low density particles will actually spread faster than fluid tracers (up
to eight times faster). However, history and buoyancy forces (neglected in these predictions)
will both oppose such large diffusion increases for low density particles. As such, these
predictions and those of Eqs. 3.52, 3.53 and 3.58 may be only qualitatively correct for bubbles
or other low-density particles.
Perhaps the most important result of Eq. 3.58 is that the diffusion ratio for long-times and
zero gravity approaches that of the fluid-tracer independent of the density ratio and Stokes
number:

*
, w 0
1

=


for all
St
L

and R 3.59
This can also be inferred from application of Eq. 3.57 when considering the limits of Eq. 3.55
and 3.56. This result occurs because the long-time integration and ensemble-averaging
eventually dominate over any lag associated with particle inertia. Thus, the spread rate will
eventually be the same for all particles. However, the time it takes to achieve the long-time
112
limit increases with the Stokes number so that the net spread for such particles will always be
less at a given time, as discussed below.


Particle Positional Variance

To see the effect that particle inertia has on the instantaneous spread of a particle cloud, the
particle diffusion rate can be integrated in time:

0
1 1 1 2 2 2
p p p p
2 2 2
t w 0 w 0 w 0
w 0
t x y z


= =

d
3.60
To determine the RHS terms, one may integrate Eq. 3.52 (which neglects the history force).
The result can be normalized by the spread for tracer particles (Eq. 3.39) to obtain a
dimensionless particle spread in the y-direction as:

( )
p
2
2 2 2
p p
w 0
3
2 2
y y
1 St R 1 R
e 1 St e 1
1 St 1 St 2 u


+




=






L
L L
L
L
L L

3.61
As with the particle diffusion rate, this result may not be quantitatively correct for finite R
values, e.g., for bubbles. The theoretical dependency of this particle positional variance is
shown in Fig. 3.15a for various Stokes numbers with R0, e.g., heavy particles in a gas. As
expected, the diffusion rate (slope of the curve) is shown to eventually approach unity for large
times (
p
> ) but is much less than the fluid-tracer diffusion rate for short and intermediate
times for large Stokes numbers. The initially lower slope indicates that the particle spread at
high inertias will always lag behind the diffusion of fluid-tracers.
These predictions are compared to experimental HIST particles dispersion in Fig. 3.15b for
a wake flow studied by Snyder & Lumley (1971) with particles of various sizes and densities.
In the experiments, the integral time-scale was assumed to be approximately constant and
based on a value at mid-point downstream (
L
=0.25k/). For the lightest particles (hollow
glass beads), the theoretical predictions for zero eddy-crossing compare reasonably well with
the experimental data when matching the particle Stokes number. This can be reasonably
expected since the terminal velocity was a small fraction of the turbulent velocity fluctuations
(i.e.
*
term
w 1 < ) such that the assumption of negligible gravity effects on diffusion was
reasonable. However, as the particle terminal velocity increased (for the corn pollen and solid
glass particles), the measured particle spread rate was found to decrease substantially as
compared to the zero eddy-crossing diffusion predictions. This discrepancy is a result of an
eddy-crossing effect associated with finite relative velocity, discussed in the next section.


Particle Velocity Fluctuations in Laminar Flow with a Single Frequency

Before addressing the fluctuations in turbulent flow, the simpler condition of particles
subjected to laminar flow which is excited by a single frequency is discussed. Consider a
spatially-uniform oscillatory flow based on a time period of oscillations given by
D
. Consider
a particle equation of motion given by Eq. 1.76 which assumes creeping flow (Re
p
0) but
113
incorporates the Basset history (Eq. 1.75). If one assumes the influence of the initial
conditions is negligible (dw/dt=0 at t=0), the theoretical velocity amplitude ratio is analytically
tractable and yields (LEsperance et al. 2006):

( )
( )
( ) ( )
D D D
rms
2 2
rms
D
D D D
St R 1 St 3RSt / 2
v
1
u
1 3RSt / 2 St 3RSt / 2
+

+


+ + +

for laminar flow

3.62
The history force effect is represented by the square root terms. As shown in Fig. 3.16, this
single-frequency prediction compares well with laminar flow experiments for a variety of
density ratios at low Re
p
values. The above result for very-heavy particles (R0) yields a
negligible influence of the history force so that:

( )
2
rms
2
2
rms
D
v 1
u
1 St


+

for laminar flow and
*
p
1 >
3.63
The result for particles whose densities are on the order of the fluid or much lower indicates a
significant influence of the history force when St
D
is finite.


Particle Velocity Fluctuations

One may define the turbulent particle kinetic energy in a fashion similar to that for the fluid
kinetic energy (Eq. A.71)

1
p i i
2
k v v 3.64
Using the same assumptions used for particle diffusion, Hinze (1975) represented the fluid and
particle velocities as Fourier integrals for all the frequencies (f) of the turbulent spectrum:

1 2
0
3 4
0
u (c cos t c sin t) d
v (c cos t c sin t) d

= +
= +

f f f
f f f

3.65a


3.65b
These equations include four length-scales constants (c
1
, c
2
, c
3
, and c
4
) which can be obtained
by substituting the RHS expressions into the particle equation of motion given by Eq. 3.50.
Setting the sum of the sine and cosine terms on both sides of this equation to be equal allows
the following relations to be obtained:

2 2 3
p p
3 1 2 2 2 2 2
p p
3 2 2
p p
4 1 2 2 2 2 2
p p
(R 1) (R 1)
c 1 c c
1 1
(R 1) (R 1)
c c 1 c
1 1

= + +

+ +



= + +

+ +


f f
f f
f f
f f

3.66a


3.66b
114
Taking the square of Eqs. 3.65a and 3.65b applying a long time averaging (as in Eq. A.70)
yields:

2 2
2 1 2
0
2 2
2 3 4
0
(c +c )
u d
2
(c +c )
v d
2

f
f

3.67a

3.67b
To integrate the fluid velocity fluctuations, one may assume that the Lagrangian correlation
coefficient for the fluid motion may be represented by an exponential function (Eq. A.80a).
Hinze (1975) obtains

2 2 2 2 2
2 1 2

1 (c +c )
u '
4
+
=


L
L
f

3.68
Substituting this results and Eqs. 3.66a and 3.66b into Eq. 3.67b yields

( )( )
2 2
2 2
p p

2 2 2
2
0 p
k 1 R
2 1 St R v
d
k 1 St 1 1
u

+ +
= = =
+ + +

2
L L
2
L L
f
f
f f

3.69
Thus, the normalized particle kinetic energy is a function of the Stokes number and the density
ratio and tends to unity for the tracer particle limit.
The appropriateness of the above result can be considered for both very-heavy particles
(R0) and bubbles (R3). For the heavy particle limit, the particle fluctuations reduce as
Stokes number increases:

2

2
v 1
1 St
u

L
for turbulent flow and
*
p
1 > 3.70
Compared to the oscillating laminar flow (Eq. 3.63), this Stokes number dependence has
similarities (e.g., no history force effect) but also substantial differences. As such, the response
of a particle due to a turbulent spectrum is fundamentally different than that due to a single
frequency. As shown in Fig. 3.17, the turbulent result of Eq. 3.70 compares reasonably with
simulation data of turbulent flows for heavy particles with negligible mean relative velocities,
even when effects of non-linear drag are included (Sommerfeld, 2001; Squires, 2007). It is
interesting that the kinetic energy ratio reduces with Stokes number independent of whether
short-times or long-times are considered. This is because the reduction due to Stokes number
in Eq. 3.70 is caused by an inertial decorrelation as particles velocities become less dependent
on fluid fluctuations. In contrast, the diffusion rate varies substantially between short-times
(Eq. 3.49) and long-times (Eq. 3.54) since it eventually incorporates all the low-frequency
contributions so that it becomes independent of Stokes number.
For bubbles (R3), Eq. 3.69 yields k
p
/k=9 for St

1, i.e. the bubble velocity fluctuations


much larger than the fluid fluctuations for significant inertia. However, experiments and DNS
results (as discussed in the next section) typically indicate that bubble kinetic energies are
typically limited to that of the continuous-phase, i.e. k
p
/k1, contrary to this prediction. The
problem of Eq. 3.69 may be firstly attributed to its neglect of history forces, which can
115
dramatically limit particle velocity fluctuations for the laminar flow case as noted in Eq. 3.62
and Fig. 3.16. This conclusion is also supported by analysis from Reeks & McKee (1984)
which indicated that Basset history coupled with initial relative velocities can affect the
turbulent flow diffusion even at long-times. Unfortunately, it is difficult to incorporate finite
Reynolds number history force effects into theoretical predictions of particle kinetic energy in
turbulent flows. A second problem associated with Eq. 3.69 is that bubbles and low density
particles are more likely to have significant mean relative velocities. This is because they tend
to be associated with lower speed flows and thus lower Froude numbers. This leads to higher
terminal velocity ratios for a given Stokes number based on Eq. 3.25. Such significant relative
velocities would violate the assumption of Eq. 3.46, on which Eq. 3.69 is based. Before
addressing this issue (in 3.5.3), the particle relative velocity fluctuations are investigated.
Using the velocity fluctuations of Eq. 3.65 and again neglecting the mean relative velocity,
the relative particle velocity fluctuations can also be represented as a Fourier integral
( ) ( )
3 1 4 2
0
w v u c - c cos t c - c sin t d

= +

f f f
3.71
Based on Eq. 3.66, these parameter differences in the integral can be expressed as

2 2 3
p p
3 1 1 2 2 2 2
p p
3 2 2
p p
4 2 1 2 2 2 2 2
p p
(R 1) (R 1)
c - c c c
1 1
(R 1) (R 1)
c - c c c
1 1

= +
+ +

= +
+ +
2
f f
f f
f f
f f

3.72a


3.72b
The mean-square of Eq. 3.71 can then be obtained as

( ) ( )
2 2
3 1 4 2
2
0
c - c c - c
w d
2T

+

=

f
3.73
This equation can be integrated using Eq. 3.68 to yield

( )( )
2 2 2 2

2 2 2
2
0 p
2 St (R 1) w (R 1)
d
1 St 1 1
u


= =
+ + +

L L
2
L L
f
f
f f

3.74
Thus, the relative particle velocity fluctuations are a function of the Stokes number and the
density ratio, but tend to zero for the tracer particle limit. As with Eq. 3.69, this result must be
modified for conditions with significant eddy-crossing effects and/or low particle density ratios,
which is the following subject.

3.5.3. Diffusion of Finite Inertia Particles with Eddy-Crossing Effects

Eddy-Crossing and Eddy-Interaction Time-Scales

The above discussion assumed that mean particle relative velocity was small compared
to the continuous-phase turbulence. If this assumption is not reasonable, then w may
116
substantially veer from the mean fluid path. Conveniently, this does not necessarily lead to any
changes in the magnitude of the continuous-phase fluctuations if the turbulence is homogenous.
For example, a stationary (v=0) particle in HIST will experience the same
rms
u as a particle
moving with the mean flow (v=u ). However, these two particles will see a different frequency
of the continuous-phase fluctuations, since the stationary particle will have fluctuations wash
over it at a faster rate than the particle which moves with the flow. Thus the time-scale for
turbulence along the particle path may no longer be given by
L
if u v . In general, a
particle with substantial relative velocity will experience a shorter time-scale since the particle
path will cut through an eddy, rather than remain with it during its entire Lagrangian lifetime.
This is referred to as the eddy-crossing effect since it is based on the time it takes for the
particle to cross a typical turbulent structure (of size ) and is qualitatively shown in Fig. 3.18.
The eddy-crossing effect may be quantified by considering particles which move at high
relative speeds to the turbulence so that they cut through the turbulence structures in times
which are much shorter than the integral time-scale. This limit is equivalent to Taylors
frozen turbulence hypothesis noted in A.5.3 whereby the turbulence structures do not have
time to evolve along the particle path during their interaction with the particle. As a result, the
change of fluid velocity fluctuations seen by the particle is simply a result of a spatial
decorrelation (and not temporal correlation) given by the spatial shift w. Application of this
frozen turbulence assumption to Eq. 3.43 yields the integral-time of turbulence along the
particle path for the eddy-crossing effect. This time-scale is denoted
cross
, and the integral of
the associated correlation function happens to out to be equal to the integral-length scale,
defined by Eqs. A.81 for streamwise fluctuations and by A.84 for transverse fluctuations:
( ) ( )
0 0
cross
1
, 0 d , 0 dx
w w

= = = =

w w x x


3.75
Recall from Eq. A.87 that the integral length-scale for streamwise velocity fluctuations is
different from that for lateral fluctuations because of the continuity effect so that the eddy-
crossing time-scale is anisotropic (even for HIST):

cross
cross
/ w
/ w



| |



3.76a

3.76b
For isotropic turbulence, the ratio of the streamwise to lateral length-scales is 2 (Eq. A.89) and
thus so is the ratio of the streamwise to lateral eddy-crossing time-scales. However, typical
anisotropic free-shear and boundary-layer flows tend to have a length-scale ratio that is
somewhat larger, about 2-3 (Hinze, 1975).
Next, we consider the decorrelation time along the particle path for intermediate conditions
(e.g.,
*
u

~1) where the eddy-crossing effect is significant but does not dominate. A general
relationship was rigorously derived by Wang & Stock (1993) for anisotropic particle-fluid
correlations and diffusivities. However, Csanady (1963) obtained a simple relationship for the
particle-path decorrelation time-scale in HIST flow based on a quadratic combination of the
integral-scale and eddy-crossing frequencies:
117
( ) ( )
1/ 2
2 2
@p cross
0
, d



= +

w
L
x
3.77
This equation was found to reasonably characterize particle dispersion in approximately HIST
flow (Csanady, 1963) and appropriately recovers the frozen-flow and zero-crossing limits for
high and low relative velocities. To show this, consider particles whose mean relative speed is
equal to the terminal velocity. In this case,
cross
is proportional to /w
term
by Eq. 3.75, while

L
is proportional to /u

by Eq. A.90. As such,


*
term
w 1 > leads to
cross

L
which is
consistent with the frozen turbulence limit for which the eddy-crossing effect will dominate so
that
@p

cross
. In contrast,
*
term
w 1 < leads to
cross

which is consistent with the zero


eddy-crossing limit of 3.5.2 so that
@p

L
.
As shown in Fig. 3.19a, the decorrelation time of the fluid turbulence observed by the
particle given by Eq. 3.77 will be shorter as
*
term
w increases due to the eddy-crossing effect.
Based on the above, the time-scale for fluid velocity decorrelation be given as

2 2 2
@p,i term i
/ 1 w /

= +
L L
3.78
This assumes that
L
is isotropic (same for both streamwise and lateral velocity fluctuations)
but allows for anisotropy in the integral length-scales consistent with the discussion of A.5.3
(Sawford, 1991). One may then define an effective Stokes number which incorporates the
eddy-crossing effect and integral length-scale anisotropy as:

( )
2
@p,i term i
St St 1 w /

= +
L L



3.79
This time-scale ratio is important for turbulent dispersion since it compares the particle
response time to the effective turbulence time-scale seen by the particle, and thus controls the
overall response.


Diffusion Ratio for Eddy-Crossing Effects

Noting that the integral time-scale seen by the particle will be modified by the eddy-
crossing effect by Eq. 3.78, the diffusivity of particles given by Eq. 3.53 can be used to obtain
the corresponding long-time diffusion ratio as:
( )
1/ 2
2
@p,i *
,i term i
1 w /


= = +

L
L

3.80
This is similar to the result of Eq. 3.54 which is independent of both particle inertia and density
ratio, but differs in that the eddy-crossing effect leads to diffusion which is reduced in
magnitude. The diffusion can also be anisotropic even for HIST flows due to directional
variations associated with the eddy-crossing time (Eq. 3.76). The differences in streamwise
and lateral length scales can be related to the turbulence structure parameter defined in Eq.
A.90:
118

( ) ( )
1
2
c
u u



|
L L

3.81
This constant can be used with Eqs. 3.26 and 3.76 to give the long-time diffusion in the
longitudinal and lateral directions for particles primarily driven by drag and gravitational
forces:

( )
( )
1/ 2
2
* *
term
1/ 2
2
* *
term
1+ w / c
1+ w / 2c



=






|

3.82a

3.82b
These results are plotted in Fig. 3.19b for HIST flow where the long-time diffusion ratios
without the gravitational effect tend toward unity (consistent with Eq. 3.59). As the particle
relative velocity increases, the influence of the eddy-crossing effect reduces the diffusion ratios
anisotropically (i.e.
* *


|
), so that the streamwise diffusion will be twice that of the
lateral diffusion when the eddy-crossing effect dominates, i.e.
*
term
w 1. >
The predicted lateral diffusion using c

=1.5 is shown in Fig. 3.20a compared against a


wide range of Stokes numbers and density ratios in free shear flows. This includes
measurements and simulations of heavy particles in grid-generated turbulence respectively by
Wells & Stock (1983) and Elghobashi & Truesdell (1984) as well as measurements of bubbles
in water by Lasheras (1998) and Poorte & Biesheuvel (2002). As expected, very small
particles with high turbulence and negligible terminal velocity will have a long-time particle
diffusivity which is approximately equal to the scalar diffusivity. On the other hand, particles
with a large terminal velocity will have a particle diffusivity which is much lower. This trend
applies to both bubbles in a liquid and particle in a gas, which is consistent with result of Eq.
3.80 which is independent of
*
p
. It should also be noted that while the experiments include
particle Reynolds numbers much greater than unity with non-linear drag, the theoretical
diffusion model, which assumes Stokes drag, behaves reasonably well.
However, it can be seen that the data spread is significant indicating that the value of c


may not be unique, and indeed this scatter can be accommodated by ranging this coefficient
from 0.5 to 5. Similar trends of long-time diffusion ratio with
*
term
w are shown in Fig. 3.20b
for fully-resolved DNS data for heavy particles in a stirred box and bubbles in an outer portion
of a boundary layer (where history forces have been included). The trends again demonstrate
the importance of the eddy-crossing effect in controlling long-time turbulent diffusion for
particles. However, the heavy particle results are better correlated with the Csanady-based
theory that uses c

=0.5, indicating further that this constant may have variations from flow to
flow. Indeed, Wang & Stock (1993) noted that the structure parameter may vary as much as
0.1 to 10 depending on the individual flow details. Furthermore, c

values for NAsT flows


can vary as a function of location and even direction (Bocksell & Loth, 2006).
Finally, we return to the Snyder & Lumley data which we were unable to predict when the
terminal velocity was significant (Fig. 3.15b), and evaluate the lateral particle spread again but
with the eddy-crossing and continuity effects included (based on Eqs. 3.60, 3.79, and 3.97).
119
The normalized theoretical result for heavy particles in the y-direction (which is perpendicular
to the mean relative velocity) is given by

( )
2
@p
rms @p
p p
2
@p @p p
y y
2 2
exp exp
1 St
u



+


=






3.83
Note that this equation can be modified for streamwise diffusion by substituting the lateral
integral time-scale and Stokes number with the values
@p @p
and St

| |
.
If we specify c

=1.5, Fig. 3.21 shows that the particle spread rates are reasonably
represented by the theoretical diffusion ratios, despite the assumption that the integral time-
scale is constant (it actually increased nearly three-fold over the measuring range). The heavier
particles (copper and solid glass) were associated with higher relative velocities and, as
expected, the resulted in reduced experimental diffusion rates according to the eddy-crossing
effect. The lightest particles (hollow glass) were dominated by the eddy life time with respect
to their diffusion and thus yielded the largest diffusion rates. It can also be seen that the initial
diffusion spread is approximately quadratic like that of a fluid tracer (Eq. 3.36) and that the
eddy-crossing effect is important for intermediate (and not just long) times.
In summary, the four main characteristics associated with diffusion of particles in
homogeneous isotropic stationary turbulence are:
a) inertia effect which reduces

as St
L
increases for short and intermediate times,
a) long-time limit which yields
*

1 for zero-crossing conditions


c) eddy-crossing effect which reduces
*

as
*
term
w increases
d) continuity effect which causes
*

|
to be twice
*


for large
*
term
w
In addition, the particle turbulent energy ratio (k
p
/k) is found to reduce for heavy particles as
St

increases, but is approximately unity for bubbles. These characteristics can be used to
develop numerical models for turbulence diffusion of particles (discussed in 7.17 and 7.2.5)
and for particle velocity bias (discussed in the next section). In the following, the effects of
relative velocity are incorporated into the models for fluctuations of particle velocity and of
particle relative velocity.


Particle Kinetic Energy for Eddy-crossing Effects

Since the eddy-crossing effect reduces the integral time-scale of turbulence seen the
particle, Eq. 3.69 can be modified using 3.80 to obtain the kinetic energy ratio (neglecting
history force effects):

( )
( )
2
2
2
term i
i,rms
2
i,rms
term i
1 St R 1 w /
v
u
1 St 1 w /


+ +

=


+ +
L L
L L

NSI



3.84
This equation indicates that variations in
i
(Eq. 3.81) can lead to anisotropy in the particle
velocity fluctuations when the eddy-crossing is significant. For very-heavy particles (R0) in
HIST, the overall particle kinetic energy is the sum of all three components:
120

( ) ( )
p
2 2
term term
k
2/ 3 1/ 3
k
1 St 1 w / 1 St 1 w /

= +
+ + + +
| L L L L



3.85
If the terms in the parentheses are limited to values of 4 or less, this can be approximated as:

( )
1
2
p
term
k
1 St 1 w /
k



+ +


L L


3.86
Employing the structure parameter of Eq. 3.81, the term in the square brackets can be defined
as an effective Stokes number for isotropic diffusion:

( )
2
1 *
@p term
2
St St 1 w / c

+
L
3.87
The turbulent kinetic energy is therefore simplified as

( )
1
p @p
k / k 1 St

+
L


for
*
p
1 > 3.88
This result has been found to be reasonable for a wide range of particle Reynolds numbers and
flow conditions (Fig. 3.22).
For bubbles (R3), the kinetic energy ratio has been used for moderate Stokes numbers by
Mudde & Simonin (1999). However, at high Stokes numbers, this equation yields bubble
kinetic energies which are an order of magnitude greater than that of the liquid turbulence.
This substantial increase is in contrast to experimental evidence for bubbles which indicate that
the turbulence ratio is generally of order unity (Sun & Faeth, 1986). This difference can be
attributed to the neglect of history force effects, which are known to reduce the velocity
fluctuations in laminar conditions (e.g., Eq. 3.62). Furthermore, lift forces and eddy-crossing
effects will reduce the degree of correlation of the particle and fluid velocity fluctuations. As a
result, the fluid-stress force effects (which are responsible for predictions of k
P
>k in Eq. 3.84 at
high Stokes numbers) will be reduced. Based on the data of Fig. 3.20b, one may propose an
empirical model which suggests that the increases due to fluid-stress force effects are
approximately canceled out by the history force, lift and eddy-crossing effects so that:

p
k k for
*
p
1 < and
p
Re 1 >
3.89
Of course, more research is needed to develop a quantitative description of bubble turbulence
intensity and for other low-density particle with finite R values.
The relative velocity fluctuations obtained for linear drag in HIST Eq. 3.74 for
*
p
1 > can
be extended to include eddy-crossing effects by using the transformation of Eq. 3.87 to modify
the effective Stokes number seen by the particle:

2
2
@p
rms
2
rms @p
St
w w
u 1 St
u

=

+


3.90
It is instructive to consider the limits of very small and very large Stokes numbers for the RHS
expression. Particles with very small Stokes numbers and linear drag will rapidly respond to
all the turbulent frequencies so that they are generally in a static equilibrium, i.e.
121

rms rms
w / u 0 for
@p
St 1 <
3.91
In contrast, very large Stokes number particles do not respond significantly to the turbulence
and their relative velocity fluctuations will simply equal to the continuous-phase velocity
fluctuations.

RMS RMS
w u for
@p
St 1 >
3.92
As shown in Fig. 3.23, the Stokes number variations for the relative velocity fluctuations are
reasonably represented with available data. Further, as long as the turbulence is weak,
homogenous and isotropic, the initial conditions are negligible, and the drag force is linear, the
time-averaged relative velocity will tend to the terminal velocity for the above conditions:

term
w w for
@p
St 1 <
3.93
In addition, the mean particle Reynolds number will tend to Re
p,term
. However, as will be
discussed in the next section, these mean quantities can differ from the terminal velocity values
if some of the above restrictions are relaxed.

3.5.4. Lagrangian Average of Particle Relative Velocity

Substantial turbulence levels can cause changes in mean particle relative velocity as well as
the corresponding particle Reynolds number. The following sub-section deals with several
mechanisms which can cause such changes.


Reynolds Number Changes due to Relative Velocity Fluctuations

For a particle falling in steady state in non-turbulent conditions, the particle Reynolds
number is simply equal to the terminal Reynolds number. However, this is not true if there are
significant relative velocity fluctuations, e.g., arising from turbulence. To show this, consider
the path-averaged Reynolds number of a particle:


f f
p
f f
d w d w
Re

=

3.94
To obtain the path-averaged relative velocity magnitude, one may decompose the relative
velocity vector along the particle path in terms of a trajectory-averaged mean and a fluctuation
about this mean using the notation:

= + w w w
3.95
Assuming the mean continuous-phase flow is uniform and the fluctuations are isotropic and
homogeneous, this particle average velocity after a long-time (t
p
) will come to an
equilibrium value. Taking a path-average of the equation of motion (Eq. 1.80) in these
conditions, the average drag force must balance the effective gravitational force:
122


D g,eff

= F F 3.96
Assuming constant particle diameter, shape and viscosity, this indicates

term term
f f = w w 3.97
Further assuming that the drag is linear so that the path-averaged Stokes correction is a
constant (f=1 for a solid sphere), leads to a steady-state result consistent with Eq. 3.93:

term
= w w
3.98
Thus, the path-averaged velocity vector for linear drag is simply the terminal velocity since
lateral fluctuations of w (perpendicular to g) will cancel-out for long times.
However, the magnitude of the fluctuations in both lateral directions will always be
positive at any instant, so that a time-average of this magnitude will be non-zero. For weak
fluctuations (
RMS term
w w < ), this effect will be negligible so that

term
w w w . For strong
fluctuations (
rms
w w > ), all three components increase the mean magnitude of the relative
velocity so that

rms
w 3w . From these two limits, one may construct an approximation
for general conditions as:

( )

3
2
i i rms
i 1
w w w w 3w
=
= + +


3.99
Using the path-averaged relative velocity of Eq. 3.95 and assuming homogenous turbulence
and isotropic particle statistics, Eq. 3.90 can be rewritten as

@p
rms
rms @p
St
w
u 1 St

=
+




3.100

Based on Eqs. 3.94, 3.99 and 3.100, the path-averaged particle Reynolds number becomes:


2 2 2
@p x y z
p
rms rms

p,term term term term
@p
3St w w w Re
w u
1 3 1
Re w w w
1 St

+ +

= + +
+
3.101
This indicates that the mean particle Reynolds number will increase with the turbulence
intensity. While Eq. 3.100 assumes Re
p
1, Fig. 3.24 also shows that this equation is
reasonable for non-linear drag regimes. This figure also shows that

p
Re can greatly exceed
Re
p,term
for high Stokes number and turbulence intensities. This can be important because
increases in the particle Reynolds number beyond the creeping flow limit can significantly
enhance heat and mass transfer rates (Eq. 1.94). The relative velocity fluctuations can also
affect the mean relative velocity if the drag is non-linear, as discussed in the following.


Non-linear Drag Bias

123
For a particle with linear drag, the path-averaged relative velocity will equal the terminal
velocity (Eq. 3.98). However, a particle with non-linear drag with relative velocity fluctuations
can yield a modified relative velocity. To show this, consider a heavy particle in the Newton
drag regime (2,000<Re
p
<300,000) where C
D
is approximately constant and where drag is the
only surface force applied to the particle. Whether at terminal conditions or in turbulent flow,
the long-time averaged drag force magnitude is then equal to the gravitational force based on
Eq. 3.96. Equating these drag forces under these two conditions yields:

1 1 2 2 2
f term D term g,eff D,z f z D
8 8
d w C F F F d w wC = = = = 3.102
The terminal velocity can then be related to the time-averaged velocity as:

( ) ( ) ( ) ( )

2
2 2
2 2
term z z z z z x y z z z
w w w w w w w w w w 3w w = = + + + + +


3.103
The approximation on the RHS assumes that the fluctuating terms dominate and are isotropic.
These terms can be approximated using Eq. 3.100 (which assumes weak turbulence) as


1

2
@p
2 2
term @p term
St
w 3u u
1
w 1 St u u w




+

+ +


for C
D
=const. 3.104
The result indicates that the mean relative velocity will be generally less than the terminal
velocity, e.g., heavy particles will have a lower fall rate once in turbulence. Since the above
results indicates ( )

( )
D D
F w F w , care must be taken when computing the average non-linear
drag force in turbulence.
A similar bias also exists for intermediate particle Reynolds numbers. The difference
between the mean drag and the force computed with only the mean relative velocity can be
expressed in terms of the Stokes correction factor and the vertical velocity (w
z
) as:


( )

( )

( )

z
D
z z z z
D
z z z z
f f w
F w fw f w f w fw
1 1
F w f w f w f w f w

+
= = = + = +
`
`
` ` ` `

3.105
The Schiller-Naumann expression of Eq. 1.54 can be used for the mean Stokes correction:

0.687
p

f 1 0.15Re +
for Re
p
<1000

3.106
The instantaneous correction for small fluctuations (
i,rms
w w < ) is then:

( )
0.687
0.687
0.687
p
p z
f 1 0.15Re 1 0.15Re w/ w = + = + 3.107
If one assumes
i
w w < , then the term in parentheses can be approximated to first-order as:


0.687
2 2 2 0.687 0.687
y
z x z z z
z z z z z z
w
w w w w w
1 1 2 1 1 0.687
w w w w w w


+ + + + + +


3.108
Combining Eqs. 3.107 and 3.108, the RHS correlation of Eq. 3.105 can be expressed as:
124


( )

( )

( )
0.687 2
p
D z
z
0.687 2
D z
z p
F w w
fw 0.103Re
1 1
F w w f w
1 0.15Re

+ +
+
`
3.109
Incorporating Eq. 3.100 then yields:


( )

( )

0.687 2
p @p D
RMS
0.687
D term @p
p
St F w
u 0.103Re
1
F w w 1 St
1 0.15Re

+

+
+

for Re
p
<1000
3.110
Therefore, increasing Stokes number, turbulence intensities and particle Reynolds number all
increase the LHS ratio beyond unity. As shown in Fig. 3.25, this approximation is reasonable
for a fixed particle (infinite Stokes number) exposed to turbulent intensities as high as 25%.
As with the Newton drag regime, the mean relative velocity for the intermediate drag
regime will also decrease in magnitude compared to the terminal velocity. However, Eqs.
3.106 and Eq. 3.110 yield implicit relationships for this change which must be solved
iteratively. An implicit model for the non-linear drag effect was also developed by Mei (1994)
based on an intermediate Re
p
form. An approximate explicit expression which gives similar
results using the Y function of Eq. 2.18 is given by:


( )
( )
1/ 2

@p term
Re z
2
*
term term @p
term
St 3 Y 1 w w
1 1 1
w w 1 St
1 w


+

+
+


3.111
This equation is constructed to approach the Newton-drag form (Eq. 3.104) for high turbulence
and high Re
p
and the creeping-flow form (Eq. 3.98) for low Re
p
. As shown in Fig. 3.26, this
HIST-based approximation is quantitatively consistent with simulations of Stout et al. (1995)
for particles subjected to stochastic and uncorrelated fluid velocity fluctuations for which only
drag and gravitational forces were present. This equation is also qualitatively consistent with
experiments of particles in an oscillating fluid (Tunstall & Houghton, 1968), though the neglect
of history force, fluid stress fluctuations and the container boundaries may be responsible for
the differences. In general, the results of Fig. 3.26 show that non-linear drag will reduce the
magnitude of trajectory-averaged relative velocity as particle Reynolds number, relative
turbulence levels (
*
u

) and/or Stokes number increase. In many multiphase flows, the non-


linear drag changes are moderate because the latter two effects tend to counteract. For
example, large particles will tend to have high Stokes numbers but large terminal velocities, i.e.
small
*
u

levels, while small particles tend to be characterized by small Stokes numbers but
with large
*
u

levels.

3.5.5. Eulerian Average of Particle Relative Velocity

In addition to the non-linear drag bias mentioned above, turbulence can also introduce
biases that impacts the Eulerian mean relative velocity. These biases generally do not affect
the relative velocity along the Lagrangian path, but instead arise because the particle
trajectories are biased towards certain regions so that the mean flow that the particle sees is
125
different from that seen by fluid tracers (3.4). If one assumes that these effects act linearly,
they can be identified using an Eulerian time-average of Eq. 3.95:

term Re pref clus k


= + = + = + + + + + w w w w w w w w w w w 3.112
The RHS includes the Lagrangian-averaged relative velocity, followed by preferential bias
(due to particles tendency to be located in up-washes or down-washes), cluster bias (due to the
change in local effective density for a cluster of particles), turbo-phoresis (due to turbulent
gradient bias) and concentration-diffusivity (due to concentration gradient bias).
Before discussing these biases, it is useful to establish the baseline condition where they do
not have an effect. Consider a single particle which rises or falls in a continuous-phase flow
with homogenous isotropic turbulence identified by
RMS
u and u ( u )

= . If we assume that the


particle has very weak or very strong inertia compared to the Kolmogorov scales (i.e.
St 1 or St 1

< > ), the continuous-phase velocity fluctuations along the particle path will be
independent of the particle dynamics since the trajectory will either be near that of a fluid
tracer or be insensitive to the turbulence structures. In this case, the fluctuations seen by the
particles will have zero mean, i.e.


@p @p
u u u 0 =

3.113
Similarly, there will be no bias in fluid accelerations or vorticity. If the trajectory-averaged
relative velocity and that measured in an Eulerian frame (based on Eq. 3.112) are equal, then



term Re
+ w w = w w


3.114
Note that Eqs. 3.112 and 3.114 both allow for non-linear drag bias effects to be included in the
trajectory-averaged relative velocity, but this velocity is approximately the terminal velocity
for low Re
p
. If Eulerian biases are significant, then particles will see continuous-phase
velocities or surrounding particle concentrations which are biased, so that the terms on the
RHS of Eq. 3.112 may be important.


Preferential Bias and Clustering Bias

The trajectories associated with preferential concentration can also affect the Eulerian-
averaged terminal velocity of a particle, since particles can be attracted to certain features of an
unsteady flow. For example, heavy particles tend to move toward the outside of eddy due to
centrifugal forces (an inertia effect), whereas bubbles are often attracted to the center of eddy
cores where the pressure gradient is lowest due to fluid-stress forces. This can lead to a
significant increase in the mean relative velocity for very-heavy particles and a decrease for
very-buoyant particles. As a result, there are substantial changes in the local fluid velocity
seen by the particles, i.e. preferential bias as shown schematically in Fig. 3.27. Thus
particles with * 1 will tend to move towards the outside edges of vortices and thus catch
down-drafts so that their overall settling velocity is enhanced, which is confirmed by the
experiments of Yang & Shy (2003). Very-buoyant particles (with * 1) will tend to move
towards the center of the vortices and thus will also experience downward continuous-fluid
velocity perturbations but this will lead to a decrease in the rise velocity, an effect which is
most extreme when St

is on the order of unity (Wang & Maxey, 1993). A visualization of this
126
phenomenon is shown in Fig. 3.28 where it can be seen that heavy particles tend to be located
in downdrafts. Thus, even if the particle velocity relative to the local fluid region remains
fixed, the tendency for the particle to ride in local regions which have a bias from the mean
fluid velocity will allow the particle to have a bias when considered from an Eulerian (full
domain) point of view.
To describe the impact of particle dynamics for a heavy particle with finite Stokes number,
consider the difference between the continuous-phase velocity averaged along the particle path
(Lagrangian) and the velocity averaged at a fixed point (Eulerian):


u u u


3.115
The Eulerian-averaged particle velocity is then


= + = + v = w+u w w +u w u +u


3.116
If many particles are considered and the turbulence is homogeneous, the path-averaged
properties can be approximated as Eulerian ensemble-averaged properties. Furthermore, the
ensemble-averaged properties (which are number-weighted) can be approximated as the
particle concentration-weighted properties. These approximations applied to the path-averaged
continuous-phase velocity fluctuations yields an Eulerian correlation:


u u u u
u u u u u u
+
= = =



3.117
This bias can be interpreted as a correlation between particle concentration and continuous-
phase turbulence, owing to the attraction of the particle to certain flow structures. For
preferential bias, heavy particles are more likely to be located in downdraft regions and
bubbles are more likely to be located to updraft regions (though there may be exception for
nearly neutrally-buoyant particles as will be discussed). This preference can be expressed as

0 > u g . This bias is likely to be significant when the particle trajectories have sufficient
inertia to allow trajectories separate from a fluid-tracer but small enough inertia such that they
are strongly influenced by the continuous-phase turbulent structures. There have been both
theoretical and empirical attempts to model this bias as discussed below.
The preferential bias for an isolated particle (=0) in a homogenous isotropic turbulent
flow was analyzed by Maxey (1987) assuming weak turbulence (
*
u 1

< ), weak inertia


( St 1

< ) and linear drag as the only surface force. The correlation of particles with turbulent
fluctuations was found to yield a bias in the fluid velocity seen by the particles, i.e. a non-zero

0 =
u . This bias can be expressed in terms of the particle response time, the turbulence
intensity and the intermediate length scale (Eq. A.97) as:

2
p
pref 0
term int er
u
3
8 w g


g
w u
l
for St 1

< &
*
p
1 >
3.118
This was extended to bubbles by Spelt & Biesheuvel (1997) by including fluid-stress force to
account for bias of continuous-phase velocity acceleration along the particle path, i.e. a non-
zero

0
t
=
u / D D . This preferential bias given by Eq. 3.118 can be approximately generalized
for bubbles and arbitrary density ratio (various R values) as:
127


( )
2
p
pref
term int er
u 3 1 R
8 w g



g
w
l
for St 1

<
3.119
To relate the RHS expression to particle Stokes number, note that the intermediate length scale
will generally be bounded by the integral length-scale and the Taylor length-scale. If one
assumes l
inter
l
Taylor
and applies Eq. A.98, this theoretical bias (normalized by the integral
velocity scale) can be approximated in terms of the Kolmogorov Stokes number (Eq. 3.22a) as:


( )
( )
pref
3 1 R St
0.3St 1 R
u g g 8 15



w
g g
for St 1

< 3.120
Consistent with experimental and DNS data, this result indicates that the bias phenomena
occurs when the particle dynamics interacts with the micro-scales (St

of order unity) while its


magnitude is dictated by integral velocity scale (u

). This first dependence is attributed to the


ability of a particle to be preferentially located while the second dependence is associated with
the strength of the vortices (Yang & Lei, 1998). Spelt & Biesheuvel also showed a second-
order effect for very-buoyant particles associated whereby lateral motion induces a lift force
downward, which can exaggerate the above preferential bias.
The preferential bias effect has been studied with one-way coupled DNS results for heavy
particles (R0), where an increase in the fall velocity occurs at small Stokes numbers. This is
consistent with the above theory based on St 1

< when the terminal velocity is equal to the


Kolmogorov velocity (Fig. 3.29a). As the particle inertia becomes larger, the theoretical result
no longer applies and over predicts the preferential bias effect, which instead peaks at St

of
order unity, after which it decays. This indicates that high inertias cause the particle to be
unresponsive to preferential location, which is consistent with experimental and computational
results (Fig. 3.9a and 3.10a). Using the Taylor length-scale defined by Eq. A.98, Yang & Lei
(1998) found that this trend is quantitatively consistent for Taylor Reynolds numbers (defined
as u

l
Taylor/

f
) ranging from 20 to 150.
Additionally, DNS and experimental results have indicated that the preferential bias effect
has a dependence on turbulence intensity with a peak at
* *
term
w ( 1/ u ) 0.4

= (Fig. 3.29b). This


dependency and peak is not predicted from Eq. 3.120 but can be qualitatively explained based
on expected particle dynamics. In particular, high terminal velocities (
*
term
w 1 > ) will cause
the particles to have strong eddy-crossing effects can not be preferentially located for
significant times. In contrast, low terminal velocities (
*
term
w 1 < ) will cause the particles to
closely follow the fluid reducing the probability of the relative movement needed for
preferential location. The DNS results of Fig. 3.29 can be used to develop an empirical model
for preferential bias of heavy Stokesian particles with linear drag in HIST as

( )
*
pref
0 term
3 2
*
term
St w
1.15
u u 1 2.2St g
1 4.4 w
=





+
+

w
u g
for
*
p
1
3.121
128
As shown in Fig. 3.29, this empirical approximation is approximately consistent with the Spelt
& Biesheuvel theory (Eq. 3.120) for low inertia particles ( St 1

< ), while also being consistent


with the experiments and DNS for other conditions.
Since the introduction of non-linear drag is found to increase the overall drag for heavy
particles, this effect can counteract some of the increases caused by the preferential bias
(Nielsen, 2007). This is shown in Fig. 3.30a where non-linear drag effects tend to reduce the
impact of preferential bias and in some cases can actually lead to a net reduction in the settling
velocity. To predict the combination of these two effects, the empirical models may be
combined linearly (assuming gravity is in the z direction) as



term Re pref
+ + w w w w

3.122
This linear combination is generally reasonable as shown in Fig. 3.30a, though tends to
exaggerate the net effects at larger Stokes numbers where the agreement is only qualitative.
This indicates the difficulty in quantitatively predicting the bias effect for general conditions.
Another difficulty in predicting the change in settling velocity is that the above theories and
empirical relations are limited to sparse (one-way coupling) conditions. However, finite mass
loading (, to be defined in 3.7.1) coupled with preferential concentration (Figs. 3.9 and 3.10)
can cause particles to collect locally in clusters. These clusters can act as single mixed-fluid
entities whose effective density is higher than that of the surrounding flow which has little to
no particle concentration. As a result, cloud clusters of heavy particles will induce a local
downward velocity of the mixture which exaggerates the settling velocity compared to the
terminal velocity. This increase in the fall (or rise) velocity can be referred to as a clustering
bias. As shown in Fig. 3.30b, heavy particles with mass loadings as small as 1% can cause
substantial increases in the Eulerian-frame fall velocity. Also shown is the model of Eq. 3.122
which underestimates the net change in settling velocity when significant mass loadings are
present. A model for this cluster velocity was developed by Aliseda et al. (2002) based on a
balance of the effective weight and drag of the cluster cloud. In general, the cluster velocity
(V
clus
) is proportional to the local concentration and the (d
clus
/d)
2
, where d
clus
is the cluster
cloud diameter. This cluster bias increases the mean relative velocity of the particle observed in
an Eulerian reference frame. It can be effectively considered as a three-way coupled bias to the
velocity-field so that Eq. 3.122 becomes


pref clus pref clus


+ = + u w w w w V

3.123
Using measurements of the local concentration and cluster diameter, this yielded reasonable
predictions of the increase in settling velocity at finite volume fractions (Aliseda et al. 2002).
However, it is difficult to correlate the mean volume fraction and particle size to the cluster
cloud concentration and diameter. Indeed, it can be seen in Fig. 3.30b that the cluster velocity
is not generally linearly proportional to the mean volume fraction.
For bubbles and low density particles, the preferential bias effect will be exaggerated due to
fluid-stress terms and will furthermore cause a net decrease in rise velocity as compared to the
quiescent terminal velocity. The theoretical result of Eq. 3.120 for weak turbulence and small
inertias is generally consistent with stochastic numerical simulations, DNS results and the
available experimental data of Poorte & Biesheuvel (2002), so long as linear drag is
appropriate. Note that this theoretical result does not include the lift or history force effects,
which are initially second-order but are expected to be more significant as the inertia and
turbulence levels are stronger. At higher Stokes numbers or turbulence levels, the theoretical
129
result tends to over-predict the impact. An improvement in prediction for these conditions can
be obtained by extending the empirical model of Eq. 3.121 to include density ratio effects:


( )
( )
( )
pref ,i
3 2
2
*
term
term
1 R St
1.15
w g
1 2.2 1 R St
1 4.4 w




+
+


w
g

3.124
This result is at least qualitatively reasonable when compared to the experimental bubble and
DNS data in Fig. 3.31a, but there is a fair amount of uncertainty due to the neglect of history
force and/or lift effects in the DNS results as well as substantial differences with respect to the
artificial spectra results. Furthermore, all the results were for dilute conditions and two-way
coupling effects (such as clustering bias) may give additional modifications to the Eulerian rise
velocity.
For sparsely distributed particles of intermediate density ratio, one may expect a decrease
in terminal velocity for buoyant particles and an increase for heavy particles based on Eq.
3.124. However, there is little experimental or DNS data which has investigated a range of
density ratios, though there is one experimental study which considered buoyant drops with a
specific density ratio of 0.85 in turbulent water flows. The measurements of the mean Eulerian
velocity are shown in Fig. 3.31b and this yields the interesting result that both decreases and
increases were observed with respect to the terminal velocity. The increases are surprising
since preferential bias and non-linear drag effects are expected to only yield velocity reductions,
and even in those cases the predictions are not consistent with the experimental trends. The
only clear trend is that the measurements for
*
u 1

> lead to a velocity increase regardless of


the Stokes number. It is possible that clustering bias may explain these increased rise rates,
especially if local cluster concentration levels and/or diameters increase with relative
turbulence intensity. However, there is insufficient data on local cluster concentration levels
and diameters in the buoyant drop studies for this to be verified. Freidman & Katz (2002)
developed a trapping model which indicated that the change in terminal velocity scales linearly
with turbulence intensity for
*
u 1

> . This model can be empirically represented as




( )
2
* *
term
term term
1
w g
w 50 w

+
w g
for
*
p
<

1
3.125
In summary, the variety of trend observed in Figs. 3.29-3.31 indicate that much additional
work is needed to understand the impact of non-linear drag, preferential bias and clustering
bias, especially for density ratios of order unity.


Turbulence Gradient Bias (Turbophoresis)

If particles are subjected to turbulence which is significantly non-homogenous (i.e. includes
substantial spatial gradients in turbulent kinetic energy), a phenomenon known as "turbo-
phoresis occurs which leads to a mean drift of particles from regions of high turbulence into
regions of low turbulence. This effect is often negligible in the streamwise direction, but can
be significant in the transverse direction where the mean particle velocity may be otherwise
small. This phenomenon is somewhat similar to the preferential bias in that particles ride in
130
regions which (due to the particle inertia) are biased towards lower turbulence intensity regions.
This turbophoretic velocity can be conceptualized in terms of two regions where one has a high
turbulence and the other has a low turbulence level (Fig. 3.31Xa). If we consider an interface
between these two regions, the probability that a particle crosses the interface is proportional to
the local turbulence level. Therefore, the mean flux is towards the lower turbulence region, i.e.
in the direction of k .
To model the effect of turbo-phoresis, consider the equation of motion for very-heavy
particles (
*
p
1) so that the drag is the only surface force. Assuming linear drag (Re
p
1), the
time-averaged equation of motion from Eqs. 3.4-3.6 becomes

i i
p
v w
t
=

d
d

3.126
If one further assumes small particle response times, the particle accelerations approach that of
the continuous-phase so that the LHS can be approximated as:

j j i
i i i i i i i i
j j j i
j j j j j
u u u
v u u u u u
u u u u
t t t x x x x x

= + = +

d d D
d d D
3.127
The first and second terms on the RHS can be neglected if the mean velocity gradients are
weak and since the velocity fluctuations are divergence-free (Eq. A.104b). Assuming isotropic
turbulence (Eq. A.74b) and substituting this result into the force balance of Eq. 3.126 yields a
turbophoretic relative velocity bias for small response times as:

2
k,i p
3
i
k
w
x

for St 1

<
L
&
*
p
1 >
3.128
The above result indicates that turbulence tends to move particles from regions of high
turbulence intensity to regions of low turbulence intensity as discussed above. An important
example of turbo-phoresis occurs with particles in a turbulent boundary layer. Since the
transverse gradients in the y-direction are the largest (Fig. A.10), this effect will give induce a
normal particle velocity away from the portion of the boundary layer with maximum
turbulence intensity. In the near wall region, this is consistent with measurements of particle
deposition fluxes (Fig. 3.32). A derivation of turbo-phoresis for general density ratios but
small Stokes numbers is given in 5.3.2 while a non-algebraic version for any Stokes number
is noted in 7.1.8.


Concentration Gradient Bias

Another Eulerian gradient-based drift which can arise is that due to particle concentration.
This bias arises because turbulence will generally diffuse a mean concentration gradient, e.g.
Fig. A.17 for fluid tracer. For particles, this can be explained by considering a control surface
for a homogenous isotropic turbulent flowfield. If the mean particle concentration on both
sides of the surface are equal then the mean flux through this surface will be simply be zero.
However, if there is a particle concentration gradient perpendicular to this surface with higher
concentration region on the LHS of the surface, then more particles are apt to pass from the
131
LHS than from the RHS (Fig. 3.31Xb). This yields a number-weighted mass flux of particles
to the RHS since these particles are correlated with a right-ward fluid velocity fluctuation.
This concentration-gradient bias can thus be considered in terms of Eq. 3.117 as:


p
N
i
,i i i,k
i k
p
u 1
w u u
N



= =


3.129
The RHS correlation between concentration and velocity fluctuations can be expected to
become stronger as the concentration gradient or turbulence intensity increases and will be in
the direction of - . As such, it can be modeled with the diffusion gradient hypothesis of Eq.
A.121 as:


2
i
,i
turb i
c k
u 1
w
Sc x


=

3.130
Additions to this model have been proposed by Elghobashi & Abou-Arab (1983) which may be
important in on-homogenous turbulence.


3.6. Brownian Diffusion

As mentioned earlier, the two primary phenomena that give rise to particle diffusivity are
turbulent and Brownian diffusion. While turbulent diffusion is created by the semi-organized
and semi-random interactions of particles with turbulent structures, Brownian diffusion is
created by the random collisions of the continuous-phase fluid molecules on the particle
surface (Fig. 3.33a). The resulting wandering trajectory was first studied in 1827 with a
microscope by a botanist named Robert Brown, who observed motion of organic and inorganic
particles which were quite small (about 1 m diameter) and within in a liquid. This motion
can be important for the diffusion of small particles in micro-fluidic liquid delivery devices,
which are important for many bioengineering processes. For gas flows, Brownian motion is
typically important for small particles in flows with low speeds and small distances, e.g.,
aerosol particles in respiratory tracts (Fig. 1.12) and near-wall particle deposition in pipe flows
(Fig. 3.32). Brownian diffusion is analogous to turbulent mixing in that the two phenomena
can be considered as stochastic processes in the mean. However, unlike turbulent dispersion,
Brownian motion does not have any spatial or temporal structures since the molecular
collisions are only governed by random statistical probabilities. As such, only mean diffusion
is relevant for the description of Brownian motion.
This phenomenon was not fully understood or accepted for some time until Einsteins 1905
quantitative analysis on the statistical molecular theory of heat in liquids outlined the laws
governing particle movements due to molecular interactions. Einsteins explanation was so
compelling that it played a major role in transforming the concept of molecules from a
hypothetical (and controversial) postulate to a practical and accepted fact. The key to
Einsteins analysis is that the fluid-dynamic force exerted on the particle can be decomposed
into a continuum-based friction force based on the no-slip condition along with a random force
resulting from the molecular collisions (F
Br
). This decomposition is appropriate because the
frequency of molecular collisions is much higher than the response time of the particle. One
132
may define u as the continuum fluid velocity, i.e. defined by the ensemble average speed of the
fluid molecules at a given point. In contrast, the particle velocity v is assumed to incorporate
non-continuum effects. Based on the definition of relative velocity (Eq. 1.9d), the particle
dynamic equation in steady flows for general Knudsen numbers is then given by

p f Kn Br
m 3 d f (t)
t
= +
v
w F
d
d
3.131
If we consider the flow around the particle to be nearly in a continuum (Kn
p
1), then the
overall motion of the particle can be approximated with a Stokesian drag response (f
Kn
=1).
However, increased Knudsen numbers can cause a reduction in drag, i.e. f
Kn
<1 (6.2.2). In the
following, we will retain f
Kn
but consider it to be a constant. Since Brownian motion is most
commonly studies in gas flows, we will further assume that the ratio of particle density to fluid
density is very high so that the particle history force and added mass effects are negligible.
While the added mass effect can be incorporated by replacing m
p
with m
eff
of Eq. 1.78 in the
above equation, it will be shown that the long-time Brownian diffusion is independent of this
effect and a similar comment can be made for the history force (Mainardi & Pironi, 1996).
In the following, the basic principles of Brownian diffusion will be analyzed for quiescent
conditions (u=0) since the corresponding movement and diffusion effect can be linearly super-
posed to particle trajectories in flows with continuum velocity (Fig. 3.33b). Let us further
focus on the motion associated with just one Cartesian direction with
x p
v x / t d d . and define
x
p
as the movement of a particle from an initial location, i.e., define x
p
(t=0)=0. Since the
molecules have no preferred direction, the ensemble-averaged force function will have a zero
mean, i.e.
Br
F 0 = . If initial particle position and velocity is given by x
p
(t=0)=v
x
(t=0)=0, the
ensemble-averaged particle location and velocity will be stationary:
p
x 0 = and
x
v 0 = .
The instantaneous 1-D dynamic equation of Eq. 3.131 with these assumptions can then be
expressed in terms of the particle position as:

p p
p f Kn Br
x x
m 3 d f F (t)
t t t

= +


d d
d
d d d

3.132
If we multiply this equation with the particle position, use the chain rule to expand the LHS
derivative, and then take the ensemble average over many interactions for all terms, this yields:

2
p p p
p p p f Kn p p Br
x x x
m x m 3 d f x x F (t)
t t t t

= +


d d d
d
d d d d

3.133
Since the forcing function is random with no bias, the last term on the RHS can be eliminated.
The second term on the LHS can be re-written as
2
p x
m v , which is proportional to the random
translational kinetic energy of the particle velocity fluctuations in one direction. This can be
related to the overall kinetic energy of the fluid as follows.
Over a time period much greater than the particle response time, an equilibrium condition
will exist such that the random particle energy in all directions will be eventually equal to the
random kinetic energy of the surrounding fluid molecules. This is referred to as the equi-
partition of kinetic energy between the molecules and the particle. The random kinetic
energy associated with a fluid is isotropic and equal to 3/2 of the product of the absolute
133
temperature of the surrounding fluid and the Boltzmann constant (). Equating this fluid
kinetic energy with all three components of the particles kinetic energy gives

( )
3 1 2 2 2
f p x y z
2 2
T m v v v = + +
3.134
Note that the Boltzmann constant can be related to the universal gas constant of Eq. A.21
normalized by Avogardos number
( ) ( )
23 2 2 23
univ
1.38x10 m kg / Ks / 6.022x10 / mol

= =R
3.135
Since the particle random energy driven by Brownian motion must also be isotropic, our 1-D
dynamic equation can be written as:

p x
p f f Kn p x
x v
m T 3 d f x v
t
=
d
d

3.136
If this ordinary differential equation is integrated twice and evaluated for long-time diffusion,
the particle positional variation becomes independent of particle mass and is given by:

2 f
p
Br
f Kn
2 T t
x
3 d f

=


3.137
This equilibrium spread rate thus describes the particle positional variance resulting from many
molecular collisions. For a spherical coordinate system, the radial spread from a point source
is given as

2 2 f
p p
Br Br
f Kn
2 T t
r 3 x
d f

= =


3.138
Note that the Brownian spread is independent of particle density and the frequency of the
molecular collisions (thus the exact form of F
Br
is not critical to this result).
The translational Brownian diffusion rate in a Cartesian direction can be defined based on
the spread rate which gives a constant value for long times:

p p
1
Br f
Br
2
f Kn
x x
T
t 3 d f

=

d
d
3.139
The diffusion rate increases linearly with temperature, just as observed for molecular diffusion
of a gas (and many liquids), which is consistent with the increased energy of the random
molecular motion. The inverse dependence on particle size in this equation explains why
Brownian diffusion is only observed in small particles. In addition, the diffusion rate decreases
with increased dynamic viscosity, which is consistent with an increased hindering of the
particle motion. For example, 1 m diameter particles will have a radial diffusivity in room
temperature water which is on the order of 1 m
2
/s, so that release from a single point will
yield a movement on the order of the particle diameter over a 1 second period. By comparison,
this same particle in air in 1 second will diffuse about 8 particle diameters assuming Stokes
drag. However, small particles in gasses are likely to have a finite Knudsen number. In this
case, a non-continuum Stokes correction (to be discussed in 6.2.3) should be included since
the no-slip condition is no longer strictly enforced on the particle surface. As the particle
134
becomes smaller and tends toward the mass and size of a continuous-fluid molecule, the inertia
time response will become negligible compared to the time-scale of the molecular interactions.
This will render the particle diffusivity to be effectively a molecular diffusivity, i.e.
Br f
,
where the latter is discussed in A.3.
The presence of Brownian motion can cause particles to spread from regions of high
concentration to low concentration in the same manner as that for turbulence (Eq. 3.130). This
leads to Brownian concentration-gradient bias for the relative particle motion:


f
Br,i
f Kn i
T
w
3 d f x

=


3.140
In general, this bias will be negligible compared to that of turbulent diffusion, but can be
significant for very small particles acting over short distances. An additional molecular
interaction effect can arise if there are strong gradients in temperature, which causes particles
to move in the direction of lower temperatures. This phenomenon is called thermophoresis
and is generally negligible when particles observe a continuum, i.e. the particle Knudsen
number is very low. However, for finite Kn
p
values and high temperature gradients, this effect
can be important and will be discussed in 6.7.



3.7. Two-Way Coupling Criteria

Two-way coupling in two-phase flows generally denotes the interphase transfer of mass,
momentum or energy and is important to the fluid dynamic description of both phases. In this
text, only the momentum-based two-way coupling will be considered whereby the presence of
the particles affects the velocity field of the continuous-phase flow. This type of coupling has
three primary modes: 1) changes in the net momentum of the continuous-phase, 2) changes in
continuous-phase turbulence, and 3) changes in the effective fluid properties including
viscosity, density, specific heat, etc. All three types of coupling are proportional to the number
of particles in the domain and can additionally depend on the particle mass and/or interphase
force. In the following, we first discuss the key parameters that drive interphase coupling and
then consider the limits at which two-way coupling should be considered. Crowe et al. (1998)
gives similar criteria for the limits associated with mass and energy transfer.

3.7.1. Concentration Properties and Interphase Force

The particle concentration, particle volume fraction, particle loading, and particle
interphase force are key parameters used to determine whether two-way momentum coupling
is important for a multiphase flow. The particle concentration is often expressed in terms of
the volume of the mixed-fluid (
m
), which is the space taken up by both the continuous-phase
fluid and the particles. To characterize particle concentration as a continuum variable
throughout a domain, the mixed-fluid volume should be large enough to encompass many
particles. Defining
p,
N

as the total number of particles in a mixture volume, the continuum


135
assumption is then reasonable for
p,
N 1

> . The corresponding particle number density (n


p
) is
then defined as the number of particles per unit volume of the mixed-fluid, i.e.

p,
p, p,
p N
m
f p,k
i k
N N
n


=
=


3.141
For example, an aerosol spray region may have an n
p
value of 10
3
particles/mm
3
. If the mixed-
fluid volume is taken to be the domain volume (
D
), one may define a macroscopic number-
density (n
p,D
) as

p
pD
D
N
n


3.142
In this equation, N
p
represents the total number of particles in the domain. The macroscopic
number-density will give a reasonable estimate of the local number density for an
approximately uniform particle concentration. However, significant differences can occur
between the local and domain number-densities if particles are only located in specific regions
(e.g., injection locations) or if particle trajectories have preferred directions (e.g.,
gravitationally settle). Flowfields can also cause particle concentrations to be localized, e.g.,
preferential concentration associated with turbulence can cause peak concentrations in low
vorticity regions (Fig. 3.9).
The volume of dispersed-phase per unit volume of mixed fluid can be defined as the
volume fraction (), which depends on the ensemble-averaged particle volume (Eq. 1.1) and
number density:

p
N
3 p,k
k 1
p p p
m
volume of particles
n d n
volume of mixture 6

= = =

3.143
For very-heavy particles (
p

f
), the dispersed-phase may only occupy only a small fraction of
the mixed-fluid volume but can constitute a substantial portion of the mixed-fluid mass.
Therefore, it is often convenient in this case to define a mass loading () based on the ratio
of net particle mass to net continuous-fluid mass in the mixed-fluid volume

p,
N
*
p,k m p
i k
f
m /
mass of particles
mass of fluid (1 ) (1 )

=

= =

3.144
One may also define the mass fraction as:

p,
N
p,k m p
i k
p
p f p f
m /
mass of particles
X
mass of mixture (1 ) (1 )

=

= =
+ +

3.145
For particles in a gas with a mass loading of order unity, the volume fraction is quite small. In
such cases, the mass loading can be approximated as X
p
=
*
p
. If the mass loading is also
small, then it is approximately equal to the mass fraction.
Another key parameter for two-way coupling is the magnitude of the force that a particle
exerts on the fluid (-F
int
), which is equal and opposite to the force that the fluid imparts on a
136
particle (F
int
). This force represents the interphase momentum transfer between the continuous
fluid and a single particle. This force is equal to the sum of the surface forces that act on a
particle (F
surf
) except the fluid-stress force (F
S
). The latter force is excluded since it acts on
the particle surface in the same way that it acts on a fluid element with an equivalent surface
(Drew, 1983). Thus, the fluid-stress force is already accounted for in the continuous-phase
equations of motion (e.g., Eq. A.6a). As such, the interphase force is:
F
int
F
surf
F
S
3.146
Based on Eq. 1.25, the interphase force of a particle on the flow can be written as

F
int
= F
D
+ F
L
+ F

+ F
H
+ F
Br
+ F
T

3.147
This force can be simplified under various circumstances as follows.
If the density ratio is very high (
*
p
1) as is the case of particles in a gas, it is common to
neglect added mass, fluid stress and history forces but retain drag, lift and thermophoretic
effects (e.g., the simulations of Fig. 3.32) so that
F
int
F
D
+ F
L
+ F
T
3.148
If the density ratio is very low (
*
p
1) as is the case of bubbles in a liquid, the body force is
negligible compared to the hydrostatic surface forces and the particle inertia is negligible
compared to the added mass force. As a result, the net surface force is approximately zero so
that the interaction force equals the fluid-stress force

F
surf
0
F
int
- F
S

3.149a

3.149b
In this limit, the interphase coupling only involves the continuous-fluid velocity properties,
though the net two-way coupling will also be a function of the bubble concentration field
(which is dependent on the particle velocity field via one-way coupling effects).
Finally, consider the case for which the particle moves at a constant velocity (ww
term
)
which is given by the balance of drag and effective gravitational forces:
( )
int g,eff p f p
= F F g
3.150
This is the terminal velocity approximation and is appropriate once the initial transients are
diminished and if the flow variations in space and time are mild in comparison, i.e. the terminal
velocity is much greater than any fluid velocity variations. As such, this assumption will not
be generally reasonable for micro-gravity conditions, neutrally-buoyant particles, high particle
injection velocities, highly unsteady (such as turbulent) flows, etc.

3.7.2. Momentum Coupling Parameter (Effect on Mean Velocity)

The presence of particles can modify the mean flowfield of the surrounding fluid. In an
internal flow, this can appear in the form of modified mean flow speed and/or mean pressure
gradient. To determine the conditions where such an effect is possible, a momentum coupling
137
parameter can be defined in terms of the net change in momentum caused by two-way coupling
averaged over the domain, i.e.

( ) ( )
( )
f f
0
D
f
0
D
u u
momentum imparted by the particles
unladen momentum of continunous phase u
=
=

=




3.151
In general, if the mean-flow coupling parameter is small (
D
1), one may reasonably neglect
the influence of the particles of the continuous-fluid bulk momentum, and therefore assume
one-way coupling. In order to assess this parameter for general conditions, it is helpful to first
consider two limiting conditions: the mixed-fluid limit (St
D
1) and the fully-separated
limit where the particle velocity (St
D
1). The momentum coupling parameters for these two
limits are next discussed, followed by a combined expression for intermediate conditions.


Mean-Flow Coupling for Mixed-Fluid Conditions

One way that the dispersed-phase can affect the continuous-flow is through modification of
the effective or mixed-fluid properties. This mechanism is generally realized when the
dispersed-phase is small enough that it moves with the same velocity of the continuous-phase.
As noted in Eq. 3.15, this is consistent with the small domain Stokes number (St
D
1). The
result for two-way coupling is a single mixed-fluid velocity field for both phases:

m St 0 @p
v v u
3.152
In this case, two-way coupling effects are primarily limited to changes in the mixed-fluid fluid
density in regions of high particle concentration.
The mixed-fluid density (also called the effective density or modified density) is
simply the mass of all phases per unit volume and can be related to the volume fraction as


m p f
(1 ) +
3.153
In the limit of very-heavy particles and low volume fractions, this equation combined with Eq.
3.144 yields a simplified mixed-fluid density and volume fraction relationship:


m f
f p
(1 )
/
+

for
*
p
1 and 1 <
3.154a
3.154b
As such, the effective density and volume fraction increase linearly as mass loading increases
and the assumptions used in Eq. 3.154 are consistent with mass loadings of order unity or less.
An example of a flow for which the mass loading effects are strong is the pyroclastic flow of
Fig. 1.12, where higher effective density causes particle-laden flow to spread along the ground.
For very-buoyant particles, the mixed-fluid density is:


m f
(1 ) for
*
p
1
3.155
In this case, the voids created by bubbles serve to reduce the effective density as particle
concentration increases. An example of two-way coupling due to this effect is a buoyant
plume caused by micro-bubbles (often used in aerating large water bodies), where an upward
flow is induced by the reduced effective density.
138
Since the change in fluid momentum is proportional to the density change, the mean-flow
momentum coupling parameter for mixed-fluid conditions approximated over the domain is:

*
D,m m,D f f D p
/ 1
3.156
Modifications of other fluid properties can also occur, e.g., mixed-fluid viscosity, specific heat
ratio, speed of sound, etc. (5.1.1). However, these effects are often secondary and similarly
depend on mass loading at high density ratios and particle volume fraction at low density ratios,
so that the above coupling parameter still qualitatively applies. Thus, mixed-fluid coupling is
only expected to be negligible when
D,m
1 < .


Mean Flow Coupling for Unmixed Conditions

The conditions where continuous-phase and particle-phase velocities are distinct can be
considered separated or unmixed conditions. In this case, the coupling is related to the
interphase force (F
int
) instead of an effective density change. For example a vertical pipe with
an upward gas flow subjected to the addition of falling particles will tend to have a decreased
gas velocity (Fig. 3.34a). The impact of the force of the particles on the continuous-fluid
momentum based on Eq. 3.151 can be related to the time-scale over which the change occurs
(
mom
) for a domain which contains N
p
particles as:

mom p int,un
D mom int
D,um 3
f D f D p
N F
F
u D u





3.157
The RHS expression is based on an average particle volume fraction (
D
) and the momentum
time-scale can be based on either the particle response time (
p
) or the domain time-scale (
D
),
depending on the type of interaction. A simple approximation is to take the maximum of these
two-scales:
( )
mom p D
max ,
3.158
If one assumes the interphase force is dominated by drag, the above results can be combined
with the definitions of Eqs. 3.1 and 3.13 to yield:
( ) ( )
* 1
D,um D p D D
c w / u max 1,St

= +
3.159
Therefore, the momentum coupling increases as the particle concentration, relative density and
relative velocity increase. The heavy particle limit for small Stokes numbers is similar to the
momentum-coupling parameter defined by Crowe et al. (1998).
If the particle response time is significant and dominates the domain time-scale, the mean
coupling parameter has the following density limits:

( )
D,um D D
w / u for
p

f
and
D
St 1 >
( )
D,um D D
c w / u

for
p

f
and
D
St 1 >
3.160a

3.160b
The combined limits are similar to the mixed condition (Eq. 3.156) in that that they scale
linearly with the particle loading ( for very-heavy particles and for very-buoyant particles).
139
However, the unmixed coupling also depends on the relative velocity. As such, the impact on
momentum will be consistent with the relative velocity direction. For example, heavy particles
falling at near terminal velocity will incur a downward component on the continuous-phase
flow. However, relative velocities can be dictated by high injection velocities such as droplet
sprays in engines (Gosman & Ioannides, 1981) or rapid variations with respect to time or space
such as shock waves (Sivier et al. 1994). In other cases, boundary reflections and high particle
inertias may primarily determine the relative velocity, e.g., the horizontal pipe flow of Kussin
& Sommerfeld (2001) where the particle velocities were dictated by wall collisions (Fig.
3.34b).


Mean Flow Coupling for General Particle Velocity

The coupling parameter based on effective density (Eq. 3.156) can be empirically
combined with that based on interphase force (Eq. 3.157) as

D D,m D,um
+
3.161
It can be expected that higher values of
D
will yield significant changes in the continuous-
phase flowfield due to the presence of particles. Figure 3.35 shows two-way coupling effects
for several multiphase channel flows (including the data shown in Fig. 3.34) in terms of the
momentum coupling parameter. In Fig. 3.35a, the center-line velocity change becomes
significant for
D
>0.3. In these studies, the particles generally lagged the continuous-phase
mean flow velocity, which is consistent with a reduction in the continuous-flow velocity as
two-way coupling increased. Changes in the axial pressure gradients for channel flows are
shown in Fig. 3.35b where two-way coupling effects became more pronounced as
D

increased. These studies included cases where particle velocity lagged or led the continuous-
phase velocity, which corresponded to increases or decreases in the pressure gradient needed to
establish the mean flow. The effect of
D
on the pressure gradient effect is greater than that
on the centerline velocity since the pressure loss tends to scale with the velocity squared.
Based on such results, a reasonable criterion for the influence of particle forces on the
mean continuous-phase velocity or pressure can be approximately given by

D
>


0.1 for significant two-way momentum coupling 3.162
Note that this mean velocity coupling parameter only indicates the potential for two-way
coupling to occur since it does not contain information as to whether or how the particles will
accelerate and/or decelerate the flow (which will be determined by the detailed particle and
continuous velocity-field interactions).

3.7.3. Turbulence Coupling Parameter (Effect on Kinetic Energy)

If the flow is turbulent, the coupling associated with the continuous-phase turbulent kinetic
energy is often referred to as turbulence modulation. This phenomenon is quite complex as
it can yield both increases and decreases in turbulence. These relationships are diagrammed in
Fig. 3.36 along with the primary continuous-phase phenomena that modify turbulence
(production at large scales and dissipation at small scales). It is generally difficult to predict
140
whether particles will increase or decrease a flows turbulence, let alone predict the magnitude
of the change. As such, turbulence modulation has been the subject of much research and
debate (Eaton, 1999). However, the type of changes which can occur can be illustrated by
considering particles which are stationary (fixed in space) so that they act as effectively as a
wire mesh as the flow moves past them. If the incoming flow is laminar or of low-turbulence,
a mesh with large wire diameters can yield local turbulent wakes will produce significant
overall turbulence in the flow (Snyder & Lumley, 1971). However, if the incoming flow is
already substantially turbulent, then a mesh with small wire diameters can reduce the
turbulence levels by breaking up the large-eddies and cause increased dissipation.
Correspondingly, wind tunnels commonly incorporate fine meshes to reduce the overall
turbulence levels in their test section. With particles that move and react to the turbulence, the
interactions become more complicated and such generalizations based on only size may not be
appropriate.
To determine whether two-way coupling on turbulence is likely, a kinetic energy coupling
parameter may be defined based on the change expected due to particles (similar to Eq. 3.151)
and this parameter can be considered in terms of both turbulence production and dissipation:

0
k
0
k k
k
=
=


3.163
To estimate this parameter, we first consider the turbulence for the condition of mixed-fluid
(negligible relative velocity), followed by the condition of heavy particles with significant
density and relative velocity. Lastly, we will consider the bubbly flow condition ( * 1) which
can often include the additional complexity of pseudo-turbulence.


Turbulence Coupling for Mixed-Fluid Conditions

Similar to mixed-fluid conditions for mean velocity, particles which have a small relative
velocity compared to that of the fluid turbulence intensity can be considered to yield mixed-
fluid conditions (Eq. 3.152). This criterion w u

< can be expressed as


*
term
w 1 < using the
parameter defined in Eq. 3.26. Druzhinin & Elghobashi (1999) and Ferrante & Elghobashi
(2003) investigated such conditions with numerical simulations and found that the change in
momentum associated with increased mixed-fluid density led to longer-lived particle velocity
fluctuations is a production mechanism for continuous-phase turbulence. However, the
particle-fluid coupling also increased the continuous-phase velocity gradients which is a
damping mechanism, so that generally decreases in turbulence were observed. As with the
momentum coupling parameter for mixed-fluid conditions (Eq. 3.156), these turbulence
coupling effects are primarily related to changes in the effective density so that the turbulence
coupling parameter can be approximated as:

*
k,m p
1
3.164
This overall reduction in turbulence is similar to that observed for liquid flows with the
addition of polymers or micro-bubbles (Fig. 1.7) or small glass particles (Zisselmar & Molerus,
1979).

141

Turbulence Dissipation Coupling

If the particles have a substantial relative velocity (e.g.,
*
term
w of order unity or greater),
then the effects associated with a modified density (given by Eq. 3.153) will be reduced since
the phases will not move in unison. The resulting associated motion and forces between the
phases can lead to additional mechanisms of production and/or dissipation beyond that of
single-phase turbulence (Eq. A.124). The kinetic energy coupling parameter can be considered
as the sum of these multiphase generation and loss mechanisms:

k pr od diss
=
3.165
An example multiphase flow with two-way turbulence dissipation is a pipe flow with small
particles. This case was experimentally investigated by Kulick et al. (1994) for downward
flow with 50 m glass and 70 m copper particles. As shown in Fig. 3.37, the turbulence
distribution near the pipe center (y=0) was reduced as the mass loading increased, especially
for the copper particles which had higher terminal velocities. These trends indicate turbulence
dissipation tends to occur for small particle sizes and Reynolds numbers (about 5 or less for
these flows). In the following, the dissipation coupling parameter is considered.
To estimate the loss of continuous-phase turbulent kinetic energy due to the particle force,
it is generally reasonable to neglect unsteady particle wake effects as well as particle-particle
and particle-wall interactions. In this case the loss of turbulent kinetic energy of the
continuous-phase includes two components: 1) the transfer of fluctuating fluid energy to
thermal energy (heat) through viscous dissipation and 2) the transfer of fluctuating fluid energy
to particle kinetic energy. The first loss component was illustrated above as analogous to
turbulence moving past a screen, representing particles fixed in space (v=0). In this case,
viscous losses will convert some of the turbulent kinetic energy into heat, representing a net
loss. Recalling that work rate is the dot-product of the force and velocity vectors, the loss in
kinetic energy is equal to the ensemble-average of the interphase force and the fluid velocity
fluctuations along the particle path, i.e.
int @p
F u . The second loss component considers how
the particle movement varies due to the turbulence. In particular, it represents the work done
by the fluid fluctuations to impart particle velocity fluctuations. This work rate can be
expressed as the path-averaged product of fluid force applied to the particles and the change in
particle velocity, i.e.
int
F v . The two kinetic energy loss mechanisms can thus be linearly
combined as
int
F w .
To model this combined energy loss, it is convenient to assume that the interphase force
can be approximated as the drag force (Eq. 3.148) and that the drag is linear, i.e. constant
particle response time. In this case, the average turbulent kinetic energy consumed per particle
can be written as (Kenning & Crowe, 1997):
( )
2 2 2 2 eff eff eff
int,i i i i i 1 2 3 rms
p p p
m m m
F w w w w w w w 3 w + = + + =

3.166
Parthasarathy & Faeth (1990) noted that the RHS term represents the turbulence dissipation
directly caused by relative velocity fluctuations of a particle. Based on Eqs. 3.1 and 3.143, the
dissipated energy per unit volume of mixture can then be expressed as:
142
( )
p f p f 2
p int,i i rms i i i i i
p p
c c
n F w 3 w 3 w v w u

+ +
=


3.167
This indicates the relative velocity fluctuations are primarily responsible for turbulence
dissipation. Note that these approximations assume a linear drag for the particles, but can be
reasonable for non-linear drag so long as the mean relative velocity is much greater than the
relative fluctuations so that
p
does not vary considerably along the path. Also note that the
RHS term shows the dissipation loss decomposed into two separate parts. The portion related
to
i i
w u was identified by Eaton (1999) as the work done by the turbulence against the
particle drag force and is consistent with the first loss mechanism. The portion related to
i i i
w v was identified by Kenning & Crowe (1997) as the work done by the carrier fluid to
oscillate the dispersed-phase and is the second loss mechanism.
Based on the above, the particle-induced dissipation of turbulent kinetic energy per unit
mass of the continuous-phase can be defined as

p
k,diss int,i i
diss f
n
k
F w
t

=



S
3.168
This result, in the limit of heavy particles, has been derived by Sundaram & Collins (1996).
Based on Eqs. 3.100 and 3.167, the particle-induced dissipation for HIST flow can be
approximated as

( ) ( )
p f p f @p 2
rms
diss f p f p @p
3 c 2 c k St
k
w
t 1 St

+ +



+

3.169
This can be compared the unladen (single-phase) dissipation rate which can be approximated
by Eqs. A.91 and A.124 for homogenous isotropic turbulence as:
( )
0
0
k / t c k /
=
=
=
L

3.170
The dissipation coupling parameter (
diss
) can then be defined as particle-induced dissipation
rate normalized by the single-phase dissipation rate:

( )
( )
( )
p f @p
diss
diss
f @p
0
2 c k / t St
k / t c St 1 St

=
+

+
3.171
This indicates that particles which interact with the flow but which have significant inertias
(St

~1) will yield the most impact on turbulence dissipation. Changes in streamwise velocity
fluctuations for several channel flows are shown in Fig. 3.38, which indicates significant
reductions as the coupling parameter increases. Significant changes can occur for
diss
as low
as 0.02. However, this correlation is qualitative and may be sensitive to differences in channel
geometry (rectangular vs. circular), flow direction (upwards vs. downwards), degree of wall
reflections, as well as the complexities associated with the non-homogeneous anisotropic flow-
field.


143
Turbulence Production Coupling

As particle size increases significantly, the result of two-way coupling can also lead to
turbulence increases. For example, consider the multiphase upward pipe flows shown in Fig.
3.39. These conditions yield a transition to increased turbulence production in portions of the
flow as the particle size and mass loading increase. Production dominates for all locations for
Fig. 3.39a whose larger particles are characterized by Re
p
~2200 (such that fully-turbulent
wakes are expected). For the moderate size particles of Fig. 3.39b, the turbulence is reduced
near the walls but is enhanced near the centerline. The near-wall reduction is consistent with
locally higher turbulence levels of the single-phase flow so that particle-based damping is more
likely to have an effect. In contrast, the near-centerline unladen turbulence has little single-
phase turbulence so that particle-based turbulent production is more likely to be seen.
Therefore, two-way turbulence coupling can lead to both enhancement and dissipation of the
turbulence depending on the local single-phase flow characteristics. In the following, the
mechanisms for turbulence production are considered.
In order to understand more clearly the turbulence production by the particles, it is helpful
to consider a condition with zero mean shear rate. A simple example of this is the continuous
settling of particles in an otherwise quiescent bath so that that any turbulence generated is due
solely to the particles. This condition was studied experimentally by Parthasarathy and Faeth
(1990) for a range of terminal Reynolds numbers (38 to 545) and a range of volume fractions
(10
-5
to 10
-4
). The measured turbulence levels were found to be anisotropic with longitudinal
fluctuations approximately twice that of lateral fluctuations, i.e.
,RMS ,RMS
u 2u


|
. The
measured velocity correlations indicated that these fluctuations arose from fully-developed
turbulence and not the simple random arrival of particles, whose wakes are nominally laminar
for most of the Re
p
values investigated (Chen et al. 2000). This turbulence was formed
through the non-linear amalgamation of the particle wakes and led to cascade of kinetic energy
with a -5/3 slope as observed in single-phase turbulence (Fig. A.16). The measurements also
indicated that the rate of kinetic energy production is a fraction of the mean work applied by
the particles per unit volume and can be represented as
pr od p int f
c n F w/ . The coefficient c
prod

can be estimated from experiments by assuming turbulence production is balanced by the
unladen turbulent dissipation and by particle-induced dissipation:

prod p int,i i f 0 p int,i i
c n F w n F w
=
= +
3.172
For
*
term
w 1 > and St 1

< , the particle-induced dissipation term can be neglected and the


conventional dissipation can be obtained from measured integral-scale properties of the
turbulence (Eq. A.91). For the LHS, the particle velocity and interphase force can be assumed
to be consistent with terminal velocity conditions (as used in Eq. 3.167) so that the turbulent
production rate due to particles can be defined and approximated as:

prod p prod p f 2
k,prod int,i i term
prod f f p
c n c ( c )
k
F w w
t

+

=



S
3.173
Based on the above, the high-loading production coefficient can be obtained as shown in Fig.
3.40, which suggests c
prod
~0.025 for a wide range of conditions (38<Re
p
<550). However, this
value is qualitative since there is significant scatter in the data and potentially a weak
144
dependence on flow or particle Reynolds number. Furthermore, Yarin & Hetsroni (1994)
suggested that turbulence production fraction was proportional to C
D
4/3
and roughly doubles for
pipe flow as compared to HIST flow. Assuming the above c
prod
value as a constant, the
turbulence production coupling parameter can be estimated from Eq. 1.61 as:

( )
( )
( )
( )
*
2
Pr od p Pr od p int,i i prod *
pr od term 2
f
0
k / t c c c n F w
w
k / t u / St


=
+
=

L
3.174
This equation uses the same normalization as for dissipation coupling (Eq. 3.171) except for a
sing change so that the coupling parameter is a positive quantity.
The production and dissipation parameters (which are both defined as positive values) can
be combined to give an overall turbulence coupling parameter for the continuous-phase
streamwise fluctuations as

( )
( )
( )
*
2
p
@p *
pr od diss pr od term
@p
c
2St
c w
St c 1 St


+

+


3.175
Thus, the turbulent energy dissipation can involve productions from the average relative
velocity as well as dissipation from the relative velocity fluctuations (Al Taweel & Landau,
1977). This result indicates that particles with higher terminal velocity ratios are more likely
to lead to a net increase in the turbulence. Given that c

is on the order of 0.2-0.3 and typical


particle Stokes numbers in many multiphase flows are on the order of 0.1 to 1.0, the condition
for which the particle-induced production just cancels out the particle-induced dissipation (no
net increase) corresponds to
*
term
w values on the order of 5-12. This is consistent with
experimental results (e.g., Tsuji et al. 1984, Maeda et al. 1980, Kulick et al. 1994 and others)
which indicate

*
term
w 8 >

for net increase in turbulence by particles (0.1<St

<1) 3.176
This criterion explains why flows with smaller particles tend to be associated with turbulence
reduction (Fig. 3.37) while flows with larger particles tend to be associated with turbulence
enhancement (Fig. 3.39a). Rewriting the criterion for a net increase as
term
u w / 8

<

also
explains why a single particle type can lead to turbulence reduction in low u

regions but
enhancement in high u

regions. For example, the channel flow of Fig. 3.39b reveals particle-
induced production in the center where single-phase turbulence is low (recall Fig. 3.37a) and
particle-induced dissipation nearer the wall. This criterion may also be extended to conditions
where the particle mean relative velocity is not dictated by terminal velocity. For example, the
transition from dissipation to production takes place at w/u

~7 for the channel flow of Kussin


& Sommerfeld (2001).
Other criteria for turbulence enhancement have been proposed based on particle size or
particle Reynolds number. In particular, Gore & Crowe (1989) suggested that turbulence
production results d/>0.1 while Hetsroni (1989) suggested that particles with highly unsteady
wakes (e.g., Re
p
>400) will tend to increase turbulence, while smaller particles with laminar
steady wakes (e.g., Re
p
<100) will tend to decrease turbulence by increasing dissipation. These
criteria are often reasonable but are not always robust. For example, Parthasarathy & Faeth
(1990) and Chen et al. (2000) measured turbulence increases for small particles at low
145
Reynolds numbers (d/0.03 and Re
p
<110). Furthermore, Tsuji et al (1984) measured an
overall turbulence decrease for much larger particles at higher Reynolds numbers (d/0.2 and
Re
p
>1000). As such, the criterion for heavy particles is not completely understood. Further
complications arise for buoyant particles as discussed below.


Pseudo-Turbulence

For particles that have densities much less than that of the continuous-phase, another
phenomenon that can affect Eulerian-measured values of
rms
u is pseudo-turbulence. This is
caused by Eulerian-averaged fluid displacements associated with the added mass of passing
particles as illustrated in Fig. 3.41. Therefore, pseudo-turbulence is not considered to be true-
turbulence since it is simply the random arrival of added mass regions at a measuring volume
and can occur even in laminar flow. This effect was investigated by Theofanous & Sullivan
(1982) and Lance & Bataille (1991a, 1991b) who made Eulerian measurements of the
continuous-phase velocity fluctuations in pipe flows and in nearly HIST flows with ellipsoidal
bubbles. The associated pseudo-turbulence caused the velocity fluctuations to be significantly
enhanced as the volume fraction and relative terminal velocity increased (Fig. 3.42). Because
the added mass causes this phenomenon, pseudo-turbulence is especially important for bubbles,
and is sometimes called burbulence.
Pseudo-turbulence may be theoretically modeled for a spherical bubble by noting that the
added mass is traveling at a speed of w
term
with respect to the continuous phase mean-flow.
The associated kinetic energy of the added mass for this case is
2
1
2 p f term
c w

as discussed in
6.4. If three-way coupling effects are negligible, then the effects of all the added masses on
the fluid may be linearly superimposed and defined as the added-mass continuous-phase
kinetic energy (k

). This kinetic energy for a single particle can be related to the local
continuous-phase disturbances due to added mass. Assuming that the added mass coefficient is
isotropic, the pseudo-turbulence kinetic energy per unit mass of fluid is given by (van
Wijndgaarden, 1998):
( )
1 1 2
i i term
2 2
k u u c w

3.177
This result is consistent with the sum of the individual components energies derived from
potential flow theory (Biesheuvel & van Wijndgaarden, 1984):
( ) ( )
1 2
i i term ii term,i term,i
10
u u c 3w w w

= + 3.178
Thus, 40% of the kinetic energy is in the direction of the relative velocity while the lateral
directions each have 30%.
If the bubbles are ellipsoidal as in Fig. 2.11b, then the added mass will become anisotropic
with both parallel and perpendicular components. The added mass coefficients in the two
directions can be related through potential flow theory (Loewenberg, 1993) as:
( )
c 1/ 1 2c

= +
|

3.179
For ellipsoidal bubbles of air in water rising broadside, typical values for c
|
(related to
motion in the direction of the relative velocity) range from 0.5-3 as deformation increases
146
(Bataille et al. 1991). Measurements of the continuous-phase fluctuations in the longitudinal
direction made by Lance & Bataille (1991a) can be used to estimate the pseudo-turbulence as
the increase associated with finite volume fraction (Fig. 3.42) and, as noted in Fig. 3.43, the
induced fluctuations are approximately given by:
( ) ( )
2
term
=0
u u u u u u c w


| | | | | | |

3.180
At the larger volume fractions, the experimental data tend towardc
|
=2 which is consistent
with increased deformation for larger diameters (van Wijndgaarden, 1998 and 6.4).
It is important to note that the result of Eq. 3.180 is about two-fold greater than expected by
rectilinear theory (Eq. 3.177). The measured increase in fluctuations levels may be attributed
to the non-rectilinear path of the bubbles (Figs. 2.23-2.27), which lead to potential flow
fluctuations in the continuous-phase. Also note that the energy associated with the lateral
fluctuations in quiescent flow is expected to be substantially less, e.g., reduced by ten-fold for
c
|
=2 based on Eq. 3.179. As such, one may expect that most of the kinetic energy stems
from the streamwise fluctuations, i.e.

1 2
term
2
k c w


|

3.181
This empirical relationship is similar in form to the theoretical outcome for spherical bubbles
(Eq. 3.177). The effective Eulerian kinetic energy (k
eff
) which may be measured can be
modeled as the sum of the true turbulence (k) and the pseudo-turbulence

eff
k k k

= +
3.182
Strong interactions between the two types of turbulence (when both are significant) can also
increase isotropy of the overall turbulence (Lance & Bataille, 1991b). However, it should be
kept in mind that particle dispersion will only be influenced by the shear-induced turbulence.
As such, the analysis in 3.3-3.5 and 3.7 does not require modification for pseudo-turbulence.
To include pseudo-turbulence, the overall turbulence coupling parameter of Eq. 3.165 can
be extended to include the kinetic energy of Eq. 3.181 as:

( )
2
*
k,eff pr od diss term
c w

+
|
3.183
The effect of
k,eff
on measured turbulence averaged over pipe cross-sections for particles with
*
p
1 is shown in Fig. 3.42a. The figure shows that turbulence production for solid particles
and drops is associated with
k,eff
>0 while attenuation correlates with
k,eff
<0. The transition
between the lower left quadrant and the upper right quadrant is consistent with
*
term
w 8 >

, as
expected from Eq. 3.176. This transition is also at least qualitatively consistent with
turbulence modulation data reviewed by Gore & Crowe (1989). If bubbles are considered, the
pseudo-turbulence and terminal velocity effects dominate so that generally only turbulence
enhancement has been observed (
k,eff
>0). Representative data are shown in Fig. 3.42b (a log-
log version of the upper quadrant of Fig. 3.44a) for which velocity fluctuation increases as the
turbulence coupling parameter rises. Based on these two figures, an approximate criterion for
the impact of particles on the continuous-fluid kinetic energy can be given as:
|
k,eff
| >

0.02 for significant two-way turbulence coupling 3.184


147
Of course, the turbulence coupling parameter only indicates the capacity for the particles to
affect the kinetic energy of the continuous-phase whereas the quantitative changes are a
function of the flow details and the interphase exchange mechanisms.


3.8. Three-Way and Four-Way Coupling Criteria

Three-way and four-way coupling refers to the addition of two separate mechanisms:
particle-particle fluid dynamic interactions and particle-particle contact interactions (Fig. 3.1).
The first mechanism, particle-particle fluid dynamic interaction, can be defined to occur when
two particles do not come into direct contact with each other, but their respective virtual or
viscous volumes intersect so that they interact in a fluid dynamical sense. The second
mechanism, particle-particle collision, refers to the actual surface-to-surface contact
interactions when physical volumes of particle intersect.
In some cases, both types of interactions may be important. For example, consider the
condition where solid particles are falling at nearly their terminal velocity in a quiescent fluid.
If a particle happens to arrive at a location which is just downstream of another particle it will
see the upstream particles wake which will represent a lower relative velocity as compared to
the conditions if the particle was isolated. This condition is shown in Fig. 3.45a and is termed
drafting and results in lower drag for the downstream particle such that it travels faster and
eventually catches up to the upstream particle. As it approaches the surface of the upstream
particle, there will be some fluid dynamic repulsion observed at short separation distances due
to the fluid between the particles which may hinder contact. If this hindering is not significant
because of low fluid density or viscosity, the particles may collide with a significant velocity
causing the trailing particle to bounce upwards with respect to its relative motion. However, if
the particle-particle approach velocity is small and the fluid dynamic repulsion is significant,
the trailing particle will slow down substantially before contact and may gently touch on the
particle below it. This is referred to as kissing and may result in both particles traveling
together for a short period (Fig. 3.45b). However, this geometry is generally unstable to small
velocity perturbations (similar to the instability of prolate particle falling vertically), such that
the particles will then tend to separate and to rotate about each other. This is referred to as
tumbling and can lead to a nearly stable condition of parallel motion of both particles (Fig.
3.45c). This condition tends to create particles grouped roughly along horizontal planes rather
than being randomly dispersed (Zenit et al. 2001).
In this section, we seek to qualitatively determine the conditions for which these two
effects become significant to the overall particle motion (i.e., where the multiphase flow is no
longer dilute). It is helpful to know this since particle break-up and/or coalescing are more
likely to occur as particle collisions increase. It is also important to establish such criteria since
inclusion of the three-way and four-way coupling mechanisms will substantially complicate the
numerical treatment and increase the required computational resources (as will be discussed in
Chapter 4).

3.8.1. Three-Way Coupling: Particle-Particle Fluid Dynamic Interactions

148
Particle-particle fluid dynamics effects are often characterized in terms of the inter-particle
spacing (l
p-p
). This is defined as the distance between centroids of two particles, and for
spherical particles is related to the gap between two particle surfaces (l
gap
) and their radii as:

p-p p p p,1 p,2
= x x l l
( ) ( )
1
gap p-p p,1 p,2 p-p 1 2
2
r r d d + = + l l l
3.185a

3.185b
Assuming a uniform lattice distribution and a uniform particle size, the mean particle spacing
(using the notation of Eq. 3.29 and employing the nearest neighbor) can be approximated by
the volume fraction as
( )
1/3
p-p gap
d / 6 d + l l 3.186
Thus, increased volume fractions will lead to reduced particle spacing, which will lead to
increased fluid dynamic interactions between particles. For reference, a volume fraction of 1%
corresponds to an inter-particle gap of 2.7 diameters.


Significant Three-way Coupling

Three-way coupling arises when a particles movement is affected by the presence of
surrounding particles. One way to assess such fluid dynamic interaction is to compare the drag
of an isolated single particle as compared to that average drag of particles in a finite
concentration mixture. At very low Reynolds numbers (Re
p
1), the particle drag becomes
especially sensitive to particle spacing since the viscous effects are felt far from the particle.
Batchelor (1972) analyzed this condition and determined the effective mixed-fluid viscosity
(
m
) for a dilute matrix of particles in a linear shear flow (Fig. A.3a). This was accomplished
by considering the net increase in overall fluid shear stresses due to the presence of particle
boundary conditions, yielding:

f p
m
f f p
5
2
1
+

= +


+


3.187
This result has been confirmed by several experiments for volume fractions of up to 5% (Zuber,
1964; Davis & Acrivos, 1985). Thus for solid particles (
p
), a 1% volume fraction is
associated with a 2.5% increase in the effective viscosity. For linear drag, this will yield a
commensurate reduction in the particle terminal velocity:

term, 0
term m
term term, 0 f
w
f
w f
=
=

= =

for
p
Re 1 <
3.188
This is roughly consistent with particle-laden flows at higher Reynolds numbers (6.9) which
demonstrate three-way coupling effects for volume fractions on the order of 1% or more. For
example, experiments and simulations have shown that particles with a lateral surface gap of
three diameters yield a 2% increase in drag for Re
p
values ranging from 50 to 150 (Kim et al.
1993, Chen & Wu, 1999) while similar changes are found for streamwise gaps of four to five
149
diameters (Sirignano, 1993; Yuan & Prosperetti, 1994). Based on such findings, a criterion for
the onset of particle-particle fluid dynamic interactions can be given approximately as:
>

0.01 for significant three-way coupling in mixed-conditions 3.189


However, this criterion should be considered as only a rough guideline since and often criteria
are put forth that are even more restrictive.
Turbulent flow can be especially sensitive to three-way coupling because variations in
particle concentration will cause local reductions in particle spacing. Because of this,
Elghobashi (1994) recommends a criterion of >0.1% for significant three-way coupling at
low Reynolds numbers. Preferential concentration in turbulence may also lead to clustering
bias as discussed in 3.5.5. This phenomenon can substantially affect Eulerian-observed fall
velocities for values as low as 10
-5
for particles in a gas, which correspond to mass loadings
on the order of 1% or more. Note that the clustering bias begins with preferential
concentration (a one-way coupling mechanism) which leads to local downdrafts or updrafts
imparted by heavy or buoyant particles (a two-way coupling mechanism) which then results in
the accompanying particles have increased Eulerian velocities (a three-way coupling
mechanism!). To take this into account, Eq. 3.189 can be modified where turbulent
preferential concentration is strong to establish a criterion for three-way coupling or
*
p
1 as:
>

0.01 for significant three-way coupling in clustering-conditions 3.190


For buoyant particles, the two-way coupling mechanism for clustering bias will be based on
volume fraction instead of mass loading. As such, the equivalent criterion to Eq. 3.190 for
these flows would be >

0.01 which happens to correspond to Eq. 3.189.




Dense Flow

Once the inter-particle gap is of the order of the particle radius, the virtual volumes can
overlap significantly such that particles tend to move as a close-knit group instead of as
individual entities. It is difficult to determine a precise volume fraction where the flow
transitions from dispersed to dense conditions in terms of particle-particle fluid dynamic
interaction, but several researchers have proposed some typical values. For example, Zuber
(1964) suggested that volume fractions below 5% are consistent with the assumption that each
particle makes its own independent contribution to the local resistance of its own motion such
that only minor corrections are needed. However, the drag and added mass effects corrections
for three-way coupling are generally reasonable up to volume fractions of about 10%-20%
(Sangani et al. 1991) while higher volume fractions are best considered as a porous medium
(Gidaspow, 1994; Zhang & Reese, 2003). A transition regime spanning 5%-20% roughly
corresponds to an inter-particle gaps (l
gap
) ranging from one diameter (=6.5%) to one radius
(15.5%). Assuming that the transition roughly occurs at an intermediate volume fraction of
about 10% is consistent with many researchers, and this is the criterion proposed herein:

<

0.1 for dispersed-phase flow


>

0.1 for dense flow based on three-way coupling


3.191a

3.191b
150
Again, as with the other coupling parameters put forth, these are rather approximate criteria
and should be viewed as only qualitative.

3.8.2. Four-Way Coupling: Particle-Particle Collisions

For particle-particle collisions, any direct contact is likely to significantly affect the particle
trajectories. The contact arises because of inter-particle velocities (v
p-p
) created by differences
in the relative velocities between neighboring trajectories. They can occur for a variety of
reasons. For gravitationally dominated inter-particle velocities, collisions can arise due to
differences in terminal velocity which are a result of fluid-dynamic interaction (Fig. 3.45) or
variations in particle size or shape (Fig. 3.46a). In the latter case, the inter-particle velocity
will be proportional to the differences in drag, e.g., due to w
term
. When the inter-particle
velocities are controlled by changes in the mean flow (e.g., turning flows or shock flows),
particles of different inertias will respond with different paths leading to potential collisions
(Fig. 3.46b), and this effect can be related to St
D
. Turbulence and unsteady flows can also
cause particles to interact and collide (Fig. 3.46c). This behavior will be more noticeable
between particles with different inertias, but will be still substantial for monodisperse particles
(i.e. no size variations are needed for this phenomenon to occur). In such cases, the inter-
particle velocity will increase with turbulence intensity and will be a function of the relevant
particle inertia scales (St

or St

).
Particle-particle collisions become significant to the overall particle trajectories when the
frequency of collisions for a particle along its path (f
coll
) becomes high. The particle collision
time-scale can then be defined as the inverse of this frequency, i.e.

coll coll
1/ f
3.192
To compare the importance of collisions to drag forces with regard to the particle trajectory,
one may then compare the particles fluid dynamic response time (
p
) to the collision time-
scale (
coll
). If the particle responds quickly to the continuous-fluid velocities and only
occasionally undergoes collisions (
p

coll
), its trajectory will be dominated by the fluid
dynamic interactions. In this case, four-way coupling will not dominate (Fig. 3.47a). In the
other extreme for which there are many collisions during the particles fluid dynamic response
time (
coll

p
), the particle trajectory will be primarily determined by the particle-particle
contact interactions (Fig. 3.47b). This comparison is quantified by defining a collision Stokes
number as

coll p coll p coll
St / = f
3.193
Thus a high collision Stokes number

means that the fluid dynamic response is comparatively
slow so that collisions will dominate the particle motion. Conversely a low collision Stokes
number indicates that the particle response time is comparatively fast. Based on the coupling
regime definitions of 3.1, these regimes can therefore be summarized as follows

St
coll
1 for negligible particle-particle collisions
St
coll
1 for dense flow based on four-way coupling
3.194a
3.194b
151
Approximate thresholds of St
coll
<0.1 and St
coll
>10 can be used for these two limits. As per Fig.
3.1, four-way coupling dense flows can be sub-divided into collision-dominated flows (where
collisions between particles dominate their motion) and contact-dominated flows, such as
granular flows (where particle contact for finite time periods determines the particle motion).
These two regimes can be identified by considering the ratio of average contact time to the
time scale between collisions, whereby large ratios would suggest contact-dominated flows.
In order to establish the conditions when particle-particle collisions are dominant,
secondary or negligible, an estimate of the particle collision frequency along the trajectory of a
single particle is needed. The frequency of collisions per particle can be determined from the
inter-particle velocity (v
p-p
), the particles per unit volume (n
p
of Eq. 3.141), and the swept area
(A
p-p
). This area is shown in Fig. 3.48 and is based on a diameter equal to the sum of the strike
and target particle (d
1
+d
2
). The collision frequency is therefore:
( )
2
1
coll p p p p p 1 2 p p p
4
A v n d d v n

= = + f

3.195
Combining Eqs. 3.1, 3.143, 3.193, and 3.195 for a monodisperse particle distribution yields a
collision Stokes number similar to that defined by Crowe et al. (1998):

*
p p p
coll
f
v ( c )d
St
3 f

+
=

3.196
Thus, the importance of collisions scales linearly with the particle volume fraction and inter-
particle velocity. To determine the latter, we will consider the two primary mechanisms which
lend themselves to theoretical analysis: collisions due to terminal velocity variations and
collisions due to turbulence.


Collisions due to Terminal Velocity Variations

For conditions where the inter-particle velocity can be determined, the collision frequency
per unit volume can be defined as N. In a monodisperse mixture, the number of collisions per
unit volume in the mixture is proportional to the number concentration (n
p
) and the collision
frequency, i.e.

2 2
p coll p p p
n d n v

= N f
3.197
For a polydisperse mixture, the number of collisions of particles of diameter d
i
with particles of
diameter d
j
, will be proportional to the product of the concentrations and the inter-particle
velocity (Eq. 3.195):

( )
2
1
ij i j p,i p, j p p
4
d d n n v

= + N 3.198
If the inter-particle velocity is associated with a difference in the terminal velocity between the
two particle sizes then v
p-p
=|w
term,i
-w
term,j
|. If we consider, the particle to have a linear drag law
(small Re
p
), then the terminal velocity is proportional to the particle diameter squared (Eq.
1.82a). For the non-linear drag case, we should include the Stokes correction factor such that
152
( ) ( )( )
2
2 2
ij,term i j p,i p, j p f i term,i j term, j
f
g
d d n n d / f d / f
72

= +

N
3.199
This result can be simplified if the difference between the two particle diameters is small, i.e.
( )
j i
d d 1 where 1 = + < . Such an approximation may be reasonable for collisions among
particle distributions which are approximately monodisperse as discussed in 2.1.3. For
example, conditions with d
90%
/d
10%
=1.2 can be characterized with 0.1 . Employing the
weak size variation approximation into Eq. 3.199 and ignoring higher-order terms, yields

4
ij,term p,i p, j p f i
f
g
n n d
9 f

N
3.200
Consider a bi-disperse particle distribution which is equally divided by number concentration,
i.e.
p,i p,j p
1
2
n n n . Assuming that the volume fraction is small ( 1 < ) and that only binary
collisions occur, the collisional Stokes number based on Eqs. 3.1, 3.143, 3.197 and 3.200 is:

( )
3
p f p f
coll,term 2 2
f
d g c
St
108 f

+
=

3.201
For a given volumetric concentration, collision dominated flow for terminal velocity is thus
more likely as the size disparity, particle diameter and density difference increase.


Collisions due to Turbulence

For turbulent flow, there are tow fundamental models for the limits of very high particle
inertia and very low particle inertia: kinetic-based collisions and diffusion-based collisions.
These two limits will be considered in the following along with a simple approximation to
represent particles with intermediate inertias.
For kinetic-based collisions, particle trajectories are assumed to intersect based on random
fluctuations of the particles in turbulent flow. To model this condition, Abrahamson (1975)
assumed that the fluid velocity correlations between two colliding particle trajectories could be
neglected, i.e. the trajectories arise from separate eddies. This is sometimes referred to as an
uncorrelated collision model and is reasonable when the particle do not closely follow the
surrounding fluid velocity correlations (e.g., St
L
1). In this case, the collision frequency can
be related simply to the strength of the particle velocity fluctuations

2
ij,kinetic p,rms p,i p, j
4 d v n n = N

3.202
The particle velocity fluctuations can be related to the continuous-phase velocity fluctuations
for heavy particles (Eqs. 3.70 and 3.63). Assuming a monodisperse size distribution and using
the integral time-scale of Eq. A.90 yields a kinetic-based collisional Stokes number
coll,kinetic
St 24
St
d
1 St

+
L
L



3.203
153
Thus, the importance of collisions scales linearly with the particle volume fraction (), but the
relationship to particle size is more complex since there is a dependency on integral-scale
Stokes number. Note that variations in the particle size distribution or a mean shear flow as
well as eddy-crossing effects (via Eq. 3.84) can increase the collision frequency further.
In the other limit of high path correlation, Saffman and Turner (1956) investigated a
diffusion-based mechanism for collisions of equal sized particles in HIST. The Saffman-
Turner theory assumed d and St

1 so that only weak variations from the fluid path are


considered. For a mono-disperse size distribution, the collision frequency stems from local
velocity gradients associated with turbulent dissipation and can be expressed in terms of the
Kolmogorov time-scale (Eq. A.93) as

i
ij,diff p, j coll,diff ,i p, j
16
n n
5

= =

N f 3.204
Thus, the small particle collisional Stokes number (based on diffusion theory) is given by

coll,diff
St St


3.205
As can be seen in Fig. 3.49, the actual collision frequency is bounded by the above two
limiting cases and tends towards the diffusion theory (Eq. 3.205) for small Stokes numbers
while tending to the kinetic theory (Eq. 3.203) for larger Stokes numbers. This, collisions in
turbulence, St
coll,turb
, will generally bounded by these two limits

coll,diff coll,turb coll,kinetic
St St St
3.206
For intermediate Stokes numbers, the finite but incomplete correlation of fluid velocity
fluctuations seen by the particles yields a collision frequency to be more than that given by the
Saffman-Turner diffusion theory but less than that of the Abrahamson kinetic theory.
Unfortunately, this intermediate regime does not lend itself to theoretical analysis as the
effects of particle inertia, preferential concentration, and integral-scale turbulence
characteristics become important (Sundaram & Collins, 1997; Ferry & Balachandar, 2002).
Therefore, several authors have given empirical corrections which combine these two regimes,
e.g., Williams and Crane (1983), Sundaram & Collins (1997), and Kruis & Kusters (1997).
Herein we put forth a simple correlation based on a path-averaged St
L
(similar to that used for
domain Stokes number in Eq. 3.161)

2
coll,diff coll,kinetic
coll,turb
St St St
St 24
St St
1 St 1 St d
1 St


+


= +
+ +
+

L
L
L L
L

3.207
This empirical combination will be bounded and properly tend towards the two theoretical
limits for very small and very large values of St
L
. Based on Fig. 3.49, this combination
provides qualitative agreement with detailed numerical simulations. This relationship also
indicates that the collisional Stokes number is linearly proportional to the volume fraction.
Comparison with Eq. 3.201 indicates that the turbulent collision rate will generally dominate as
compared to that associated with terminal velocity variations if the size distribution is narrow,
about d
90%
/d
10%
<1.2.


154
Collisions due to Brownian Motion

Finally, a collision rate due to Brownian motion was obtained by Smoluchowski (1916) in
terms of the temperature and the Boltzmann constant (, given in Table A.1) as

( )
2
i j
ij,Br p,i p, j
f i j
d d
2 T
n n
3 d d
+

N
3.208
This overall frequency can be converted into a collisional Stokes number for a monodisperse
distribution (d
j
=d
j
) based on Eqs. 3.1, 3.143, 3.193 and 3.197 and assuming Stokesian drag:
( )
coll,Br p f 2
f
8 T
St c
9 d


= +


3.209
This collisional Stokes number decreases with particle diameter while that for turbulent
diffusion (Eq. 3.205) showed an increase with particle inertia and diameter. As such, the
Brownian collision rate will be negligible in comparison except when particles are quite small.



3.9. Multiphase Flow-Coupling Summary

As discussed in 3.3-3.5, the particle Stokes number is the primary indicator of the
particles ability to follow the flow. In particular, the one-way coupling regimes can be
summarized in order of decreasing particle inertia as follows:

St
D
1 particle motion negligibly affected by the macroscopic fluid motion
St

1 particle motion negligibly affected by the turbulent structures


St

~1 particle motion substantially modified by integral-scale structures


St

1 particle motion closely follows integral-scale structures


St

~1 particle motion substantially modified by Kolmogorov structures


St

1 particle motion closely follows all turbulent structures


3.210
Similarly the two-way, three-way, and four-way multiphase-coupling criteria discussed in
3.7-3.8 can be summarized as:

D
1 for negligible two-way momentum coupling by particles
|
k
|1 for negligible two-way turbulence coupling by particles
1 for negligible three-way coupling in mixed-conditions


1 for negligible three-way coupling with heavy particle clustering
>

0.1 for dominate three-way coupling: dense flow


St
coll
1 for negligible four-way coupling on particles
St
coll
1 for dominate four-way coupling: dense flow
3.211
It should be noted that these criteria are qualitative since they do not account for differences
associated with: domain flow geometry, domain unsteadiness, non-linear drag, eddy-crossing
effects, particle contamination, size distribution, anisotropic or non-homogenous turbulence
features, etc. However, it is helpful to understand when specific coupling mechanisms are
155
important as this can dictate the appropriate numerical method for analysis (a subject to be
discussed in the next chapter).
To illustrate the conditions associated with different coupling regimes, the above criteria
can be used to construct regime maps as shown in Figs. 3.50 and 3.51 as a function of particle
size and concentration. These particular regime maps provide approximate boundaries for the
various coupling conditions for two common conditions with substantially different density
ratios: solid particles droplets in air (
*
p
~1000) and air bubbles in water (
*
p
~0.001). The latter
case is also approximately representative of liquid or solid particles in a liquid (
*
p
~1). For the
particle phase, the maps consider a range of particle sizes from 10 microns (below which non-
continuum effects may become significant) to 5 mm (above which bubbles become highly
ellipsoidal). The drag coefficients used in these regime maps are primarily based on the
expressions given in 1.5.2, but some corrections are made for ellipsoidal bubble shapes (based
on expressions in 6.2.6). The demarcations of the various coupling regimes are based on Eqs.
3.210 and 3.211 assuming monodisperse particles moving approximately at their terminal
velocity. For the continuous-phase, domain sizes are set as 100 mm which is a typical scale for
many laboratory flows, and ensures a large D/d ratio (ranging from 20 to 10,000). To allow
similar domain Reynolds numbers for both maps, the air flow speed was set to be an order of
magnitude larger than the water flow speed (5 m/s vs. 0.5 m/s). The flow is considered to be
homogeneous isotropic stationary turbulent flow with u

0.05u
D
and

0.05

0.05
D

(typical of the middle portion of fully-developed turbulent channel flows at these domain
Reynolds numbers).
The case of particles in air is shown in Fig. 3.50 in terms of particle diameter vs. mass
loading (). For very low mass loadings (), only one-way coupling is present. This coupling
at the smallest diameters is consistent with St

~1 whereby preferential concentration effects


can be significant, e.g., Fig. 3.26. At larger diameters, particle inertia will lead to particle
concentration features on the integral-scale (Fig. 3.6). In this regime the overall diffusivities of
the particle to the flow has begun to reduce compare to that for scalars. At very large Stokes
numbers, the particles tend to cut through the flowfield with only minor deviations from
terminal velocity trajectories.
As mass loading increases, the first coupling regime encountered is typically an affect on
the continuous turbulence levels, and it is important to note that impact of the particles on the
spectra at higher frequencies can come at loadings which are an order of magnitude smaller
(Druzhinin & Elghobashi, 1999). For large values of
*
term
w , pseudo-turbulence is more likely
which results in the turbulence becoming more sensitive to the particle presence and generally
enhanced as compared to the single-phase conditions. At moderately larger values of mass
loading, mean flow coupling becomes important especially for small particles when mixed-
fluid conditions apply. Particle-particle collisions associated with turbulent mixing then
become important, particularly for intermediate-sized particles since the collision frequencies
become large while the particle response times are still modest. Note that significant size
variation will result in particle collision effects becoming important at much lower volume
fractions. At still higher mass loadings, three-way coupling effects become important (=1%),
though these can be important at even lower void fractions if preferential concentration causes
clustering. The end of the dispersed flow regime and the onset of dense flow correspond to
conditions where three-way coupling dominates (especially for smaller particles) or collision
frequency dominates (especially for larger particles).
156
For the bubbly flow considered in Fig. 3.51, we consider the increased particle
concentration in terms of volume fraction (similar results can be expected for particles in liquid
with. In comparing with Fig. 3.50, we note that most of the above trends are again evident, but
there are some significant exceptions. For example, bubbles in water have lower terminal
velocities (compared to particle in a gas for the same diameter) due to increased continuous-
phase viscosity. The corresponding reduced Stokes numbers result is more likely to lead to
tracer particle behavior for the one-way coupling effects. Furthermore, the importance of
volume loading (compared to mass loading) for bubbly flows leads to increased sensitivity to
three-way coupling effects as compared to that for two-way or four-way coupling effects. In
particular, particle-particle collisions due to turbulence become important for volume fraction
of only a few percent and can even be more significant for polydisperse size distributions. For
example, a distribution given by d
90%
/d
10%
~10 would cause collision effects to become
important for volume fractions as little as 0.1%. The two-way coupling to the continuous-
phase turbulence tends to be dominated by production and pseudo-turbulence mechanisms so
that enhancement is expected for most conditions. Finally, dense flow occurs once the volume
fraction of particle becomes substantial, e.g., reaches 10%. It should be emphasized that the
above regime maps indicate the qualitative capacity of a multiphase flow to incur coupling, but
do not dictate the magnitude or direction of the modification.


3.10. Problems

3.1) Derive the exponential explicit scheme of Eq. 3.8a for the particle velocity starting
from Eq. 3.4 assuming a constant fluid velocity throughout the domain and Stokesian drag
(f=1).
3.2) Use the scheme of Problem 1 to predict the one-dimensional velocity of a very-heavy
particle (R0) for zero gravity (g=0) a total time integration of 3
p
. Assume that the particle
is initially at rest (v=0 at t=0) but is subjected to a constant fluid velocity of constant
magnitude u
o
. Use a reasonably small time-step, and plot together the predicted solution for
this condition in terms of v/u
o
as a function of t/
p
. Discuss the trend in particle velocity
evolution. Also show that this scheme is exact (not a function of time-step) by comparing it
with the analytical solution. Under what conditions will the exponential scheme no longer be
exact?
3.7) Consider a HIST wake flow of air at STP with the following turbulent properties: a
Lagrangian fluid time-scale of 0.1 s, a kinetic energy of 0.06m
2
/s
2
, c

=0.2 and c

=1.5
a) compute St

and St

for an isolated 100 micron water drop and comment on the type of one-
way coupling expected for these Stokes numbers;
b) compute the long time transverse diffusion rate and the particle turbulent kinetic energy
c) explain qualitatively why some turbulent bias effects can be neglected with respect to wfor
this particle and why significant ones would give an increase or decrease compared to w
term
;
then compute wbased on the significant bias effects
3.8) Starting from Eq. 3.35, derive Eq. 3.38 and also show that this result yields the limits
given by Eqs. 3.36 and 3.37 for short and long times, respectively, and yields the fluid-tracer
spread of 3.38.
3.9) Derive the general particle spread rms as a function of time and the long-time of Eq.
3.137 starting from Eq. 3.132.
157
3.10) For the conditions of problem 7 and a mass loading of 10%:
a) determine
k
and comment on whether you would expect a change in the continuous-phase
turbulence due to two-way coupling and whether any such change would likely be an increase
or a decrease;
b) determine the turbulent and Brownian collisional Stokes numbers (St
coll, turb
and St
coll, Br
)
and comment on whether collisional effects will be significant or dominate particle motion in
either case.


158

4. Overview of Multiphase Flow Numerical Approaches

In this chapter, our attention is turned to numerical formulations for treating multiphase
flow. The numerical methodology is strongly related to the flow physics of interest and the
governing equations for both phases. The numerical resources required are also related to the
minimum spatial and temporal discrimination needed and on the mathematical character of the
solution. The emphasis will be on defining the general characteristics of the various available
approaches, discussing their appropriateness for specific simulation objectives (predicting
mean diffusion, determining a bubble shape evolution, etc.) and quantifying the associated
computational resources (programming, CPU time, CPU memory, etc.).
The first section of this chapter deals with the various approaches categorized by their
treatment of the reference frames, velocity fields, and interphase surface forces. The second
section gives some guidelines on comparing the various methodologies in terms of the
predicting various multiphase fluid physics such as one-way coupling (including turbulent and
Brownian diffusion), two-way coupling (which modifies the equations and techniques), as well
as three-way and four-way coupling (which describe particle-particle interactions). The third
section discusses the computational requirements primarily in terms of number of continuous-
phase and particle phase nodes required. Note that numerical methodologies for the uncoupled
continuous-phase flow are described in Appendix B. This uncoupled flow is important as it
drives the particle dynamics in dispersed flow conditions, and also serves as a base-line
description for the coupled continuous-flow equations.
In general, this chapter is intended to give a brief survey of the tools available, some
guidelines on the jobs for which these tools are best suited, and the costs for their
implementation. This should allow one to select a numerical technique which is optimal for
the desired multiphase analysis. The following chapters (5-8) will provide more depth and
details of the mechanics and performance of the various approaches.






159

4.1. Dispersed-Phase Flow Numerical Approaches

There are a large variety of numerical methods for multiphase flow, and so it is helpful to
consider the various ways to categorize the approaches. The most common categorization is
by the reference frame for the particles: Lagrangian vs. Eulerian approaches. Other major
categorizations include: Mixed-Fluid vs. Separated-Fluid approaches and Point-Force vs.
Resolved-Surface approaches. These three classifications are overviewed in Table 4.1 in terms
of the momentum PDEs assuming no chemistry or non-continuum effects. These conservation
equations will be put forth assuming simplified volumetric averaging principles for dispersed
particles in a continuum. However, a more formal derivation of these equations requires
careful consideration of the spatial and temporal averaging limits and the assumption of small
but finite particle size. As such, the reader is particularly referred to details given by Drew
(1983), Zhang & Prosperetti (1994), Crowe et al. (1998), and Prosperetti (2007). There are
subtle differences among these derivations in these sources which are particular to different
assumptions.

4.1.1. Lagrangian vs. Eulerian Approaches

One of the key divisions between the various multiphase numerical methods is that of the
reference frame for treating the dispersed-phase. While continuous-flow CFD formulations are
generally considered in an Eulerian reference, the dispersed-phase characteristics (velocity,
concentration, diameter, etc.) are commonly treated with either Lagrangian or Eulerian
representations. If we consider the case where the particles are small in size compared to the
grid resolution of the continuous-phase (larger particles will be considered later), a natural
approach is to use particle trajectories, i.e. the Lagrangian dispersed-phase representation. In
this case, the particle characteristics (velocity, position, etc.) are declared and updated along
the particle path-lines, which are based on the center of mass of the particle (or along the center
of mass for a cloud of particles) as shown in Fig. 4.1a. In contrast, the Eulerian dispersed-
phase representation generally declares the particle characteristics (velocity, concentration,
etc.) at nodes coincident with those of the continuous-phase grid (Fig. 4.1b). Hybrid
approaches are generally rare, though this may be an important area of research in the future.
The two different reference approaches (Lagrangian vs. Eulerian) for the particle
characteristics lead to different formulations in the dispersed-phase equations of motion, as
discussed in the following.


Lagrangian Dispersed-Phase Equation of Motion

In the Lagrangian methodology, the particle path-lines are determined by the particle
equation of motion. Generally, this yields an ordinary differential equation (ODE) for each
particle (as in 1.5) using particle path derivatives (with temporal discretization as discussed in
2.2). Thus, the equation of motion for the Lagrangian dispersed-phase coupled with the
definitions of particle velocity (1.3) and mass transfer rate (1.6) in tensor form are given by
160


( ) ( ) ( )
i
p body surf coll body int coll p ij
i i
v
m F F F F F F p K
t
= + + = + + + +
d
d

pi i
x / t v d d
p p
m / t m ` d d
4.1a

4.1b

4.1c
The RHS forces of Eq. 4.1a are considered along the path of the particle, where the surface
force is separated into interphase forces and fluid-stress forces as noted in Eq. 3.146, and
where the fluid-stress forces are given by Eq. 1.67. Note that these ODEs are formally coupled,
e.g., the particle velocity is needed for integrating the particle position, but the particle velocity
and mass transfer are also a function of the local flow conditions.
The particle energy equation can be neglected altogether if the continuous-phase flow is
approximately isothermal (e.g., incompressible adiabatic flow) and the relative temperature
differences with the particles are small. However, these conditions are not satisfied in many
conditions, particularly if there are significant heat and mass transfer effects. An energy ODE
along the particle path can be expressed in terms of the particle specific energy (e
p
), the particle
specific heat at constant pressure (c
p,p
) and the absolute temperature (e.g., Kelvin):


p p,p p
e T c
4.2
Based on this definition, the Lagrangian transport can be generalized to include gradients in the
surrounding fluid temperature but limited to some specific assumptions (discussed below) as


( )
p
p p p phase p eff
e
m m T
t
= + +
`
`
d
Q h k
d
4.3
As in Eq. 1.90, the RHS includes the specific enthalpy required for phase change of the particle
material (h
phase
) and the heat transfer rate per particle (
p
`
Q ). The last RHS term of Eq. 4.3 is the
energy transfer due to temperature gradients in the surrounding fluid with an effective
conductivity,
eff
k . This can be approximated by the mixed-fluid conductivity (k
m
) which is is
based on the volume fraction contributions (Crowe et al., 1998) so that:


eff m p f
(1 ) + k k k k
4.4
This mixed-fluid conductivity definition assumes that heating of the particles and continuous-
phase occurs simultaneously and is independent of whether the mixture is miscible or
immiscible so long as the thermal equilibrium approximation applies. This assumption for the
effective conductivity is reasonable if the particles have a small thermal response time (
T
of
Eq. 1.93b) compared to the time-scale of temperature changes in the fluid. If the particle
thermal response time is very large, the thermal diffusivity will tend to that conducted only
through the continuous-phase, i.e.
eff f
(1 ) k k . Since the differences in thermal
conductivities for most substances does not vary widely (e.g., Table A.1), the difference
between these two limits is generally not significant for dispersed flows (<10%). As such, Eq.
4.4 is both common and generally reasonable (Crowe et al. 1998, Elghobashi, 2006).
For the above energy equation, the internal temperature of the particle is assumed to be
unaffected by the four following mechanisms: particle compressibility, chemical reactions,
internal temperature gradients, or the normal surface velocity. There are specific conditions for
161
which these assumptions are not reasonable. Firstly, a compressible fluid particle with rapid
volumetric changes (e.g., subjected to strong acoustic waves) will undergo changes in internal
temperature, which can include reversible pressure work as well as irreversible viscous
dissipation (Crowe et al. 1998). This is particularly true for bubbles in liquids, in which case a
Rayleigh-Plesset momentum equation (2.3.3) can be supplemented with a similar energy
equation (Brennen, 2005). Secondly, the particle or surrounding fluid chemical composition
may change in time and/or space which can cause changes in the heat and mass transfer rates
as well as the fluid properties. This is especially common in the combustion of multi-
component drops in gasses, and methods for these effects are discussed by Sirignano (1999).
Thirdly, the particle internal temperature distribution may not be uniform and affected by
finite-rate internal conduction (as well as internal convection for fluid particles) which
indicates that the particle surface-averaged temperature is different than its volume-averaged
temperature. Such non-uniformity can be important for particles immersed in flows with rapid
temperature changes or particle with large thermal inertias, and requires additional RHS terms
and/or internal temperature modeling (Sirignano, 1999, Michaelides, 2006). Lastly, neglecting
the kinetic energy associated with the normal surface velocity assumes that the efflux velocity
on the particle surface due to mass transfer (
r,m
u
`
) is small compared to the particle
translational energy. Crowe et al. (1998) derived this additional energy term but not that it is
nearly always negligible compared to the heat of phase change.


Particle Concentration with the Lagrangian Approach

Since the dispersed-phase is treated in terms of individual particle paths (or group of
particles) for the Lagrangian method, no continuum approximation is required. As such, this
approach is also sometimes referred to as the discrete approach. When coupled with an
Eulerian representation of the continuous-phase, this combined treatment is termed the
Eulerian-Lagrangian approach. Thus the Lagrangian particle trajectories are generally not
coincident with the Eulerian continuous-phase mesh (Fig. 4.1a). To translate particle
trajectories to Eulerian concentration fields, cell-averaged concentrations are typically used.
For the concept of a local concentration to be well-posed, the computational cell should be
large enough to encompass at least a single particle, i.e.:

p
<

for a well-posed cell-volume for concentration


4.5
If this is not satisfied (e.g., the cell-volume is wholly located within a particle), the concept of a
number of particles per cell volume is not reasonable.
The cell-averaged number density or volume fraction (Eq. 3.143) can be obtained by
determining all the particles whose centroids (x
p
) is within the discrete cell volume (

) and
then summing up their respective volumes (
p, j
):

p
p p p, j
j 1
N
1
n

= =

4.6
Note that

is the cell volume that includes both the continuous-phase and the dispersed-
phase and thus is sometimes referred to as the mixed-fluid cell volume. Also, note that a
162
particle volume that overlaps a cell boundary will only partially contribute, i.e.
p, j
only
includes the portion of the particle volume within the discrete cell. There are also other high-
order approximations based on relative distance of the particle to the cell center (as will be
discussed in 7.2.7), but these are still based on a summation of individual properties. The
Eulerian conversion of Eq. 4.6 is reasonable provided there are many particles per cell (as will
be discussed in the next sub-section). Similar equations can convert Lagrangian particle
velocities and temperatures to cell-averaged Eulerian particle velocities and temperatures.
In multiphase conditions where the number of physical particles is very high (e.g.,
hundreds or thousands of particles per computation cell), it may be impossible or inconvenient
to track all of the particles with Lagrangian ODEs. An alternative approach is to use
representative Lagrangian particles called parcels. These representative particle groups are
essentially a small cloud of particles that are all modeled with the same Lagrangian trajectory.
These parcels are also sometimes called computational particles or super-particles. If we
denote N
pP
as the number of physical particles in a parcel, corresponding ODEs for the parcel
trajectory and momentum (where v
P
as the velocity of the parcel) can be given as


( )
( ) ( )
pP
pP P,i pP P,i
pP p P,i pP body surf coll
i
N particles / parcel
N x / t N v
N m v / t N F F F

=
= + +
d d
d d

4.7a

4.7b

4.7c
If there is no mass transfer, break-up or coalescence, the number of particles per parcel will be
constant along its trajectory. In this case, the set of equations for a single parcel are effectively
the same as those for a single particle so that the parcel approach is much more
computationally efficient. For example, N
pP
=100 allows the number of computed trajectories
to be 100-fold less for a fixed number of particles in the domain.
To express the volume fraction of particles in a computational cell, a simple method is to
sum over the number of Lagrangian parcels within an Eulerian cell (defined as N
P
):

P
p p pP, j p, j
j 1
N
1
n N

= =


4.8
Details of this and other averaging techniques will be discussed in 7.2.7. To initialize parcels
in a computational cell for a region with a monodisperse particle mixture, the initial number of
parcels for each cell (N
P,init
) is often set uniformly throughout the domain. In this case, the
initial number of particles per parcel (N
pP,init
) can be expressed as:


( ) ( )
pP,init p P
init
init
N / N

= for monodisperse parcel initialization
4.9a
A similar initialization of parcels can be used for polydisperse size distributions (as will be
discussed in 7.2.4). The parcel volume can be defined and expressed as:


( ) ( )
P
P,init p pP P
init
init
volume of mixed-fluid occupied by parcel
N / / N


= =

4.9b

4.9c
The scale of this volume determines the spatial accuracy of the point-force approximation
because the continuous-phase properties for F
surf
of Eq. 4.7c are obtained at the parcel center
163
(x
P
) but are meant to represent the properties throughout
P
. If the cloud is large, there can be
a substantial variation of continuous-phase flow variables across its length. This problem can
be minimized by limiting the parcel cloud to a length-scale on the order of the minimum
continuous-phase length-scale which, in turn, can be estimated as the continuous-phase grid
resolution (given by Eqs. B.14a, B.14b and B.14c):

P
<

for well-posed parcel volume


4.10
This criterion specifies that the Lagrangian parcel cloud volume be bounded by the Eulerian
computational volume of the continuous-phase flow. This limitation is also consistent with the
summation of Eq. 4.8, which effectively assumes that all the particles associated with a parcel
contribute to the computational cell in which x
P
is located. As such, Eq. 4.10 places an upper
bound N
pP
for the parcel method.


Eulerian Dispersed-Phase Continuum Approximation

The Eulerian description applied to the dispersed-phase (Fig. 4.1b) generally requires that
the particle properties (such as concentration, velocity or temperature) can be approximated as
a continuum. This is generally known as the particle-phase continuum assumption (Drew &
Passman, 1999). This continuum treatment assumes that the particle concentration field is
well-posed, is continuously differentiable and is gradient-resolved throughout the
domain. The well-posed criterion is established by Eq. 4.5 and the other two criteria are
discussed in the following.
The continuously-differentiable assumption is related to the number of particles per
computational cell (N
p
). To illustrate this, consider the larger and smaller computational
volume (with respect to particle spacing) given in Fig. 4.2. The larger volume contains many
particles so that the number density (n
p
of Eq. 3.141) does not depend significantly on the exact
volume position. In contrast, the smaller volume has a length-scale that is on the order of the
average inter-particle spacing (defined in Eq. 3.186). In this case, the local concentration can
vary dramatically depending on the specific location of the volume (e.g., it can be finite or zero
depending on whether it happens to capture a particle). These differences are important if one
wishes to represent concentration gradients. For example, consider two neighboring control
volumes with a particle concentration defined as n
p,i
and n
p,i+1
(Fig. 4.3a). For cell volumes
with many particles, n
p
will tend to vary smoothly between cells. Thus the concentration at an
intermediate location can be obtained by applying a Taylor series expansion for n
p,i
:

1/ 2
2 3 2 3
p p p
p,i p,i 2 3
i
i i
n n n
x 1 x 1 x
n n ...
2 x 2! 2 x 3! 2 x
+

= + + + +








4.11
Combining this with a similar expansion for n
pi+1
but with a backwards shift (x) yields:

( )
1/ 2
2
p,i 1 p,i
p,i
n n
n O x
2
+
+
+
= + 4.12
By dropping the higher order terms of Eq. 4.12, we obtain a finite difference of n
p,i+1/2
based
on the cell volume discretization. The linear variation in n
p
approximated in Eq. 4.12 is also
consistent with linear shape functions used for Finite-Volume or Finite-Element approaches
164
(B.1). The concentration given by Eq. 4.12 can be compared to the concentration defined by
the particles actually within the i+1/2 control volume, and the difference is the discretization
error. While some discretization error is always expected whenever the inter-particle spacing is
non-uniform it will generally be small for large N
p
.
A much different result arises if N
p
is no longer large. For example, Fig. 4.3b shows a
grid resolution which is on the order of the local inter-particle spacing (x~l
p-p
). For these
particular volume locations, n
pi+1
happens to be one while n
p,i
is zero. This discontinuous
change in the number density distribution between the cells indicates that a finite difference
mean between two adjoining cells will give substantially different results depending on the
selected position of the numerical control volumes. For example, the n
p,i+1/2
based on Eq. 4.12
will be about half that obtained by the dashed-line control volumes. Thus, finite difference
expressions of the gradients which assume a continuously-differentiable concentration (Eq.
4.11) are no longer reasonable to employ since the results are highly sensitive to control
volume location.
The above discretization error for the cell-averaged concentration is related to the using
statistical analysis of uncorrelated events. In this case the statistical uncertainty of the number
density (
p
n U ) is proportional to the square root of the sampling frequency:

1
p p p
n ~ n N

U 4.13
Based on this, a frequency of 100 particles per cell translates into an uncertainty of 10%, while
10
4
particles would be needed to ensure an uncertainty of 1%. Of course, if N
p
is order unity,
then the error in n
p
will be on the order of 100% as shown in Fig. 4.3b. Thus, a large number
of particles per cell is needed for accurate predictions of the gradients, i.e. an Eulerian
representation requires:

p
N 1

> for a continuously-differentiable concentration


4.14
In a sense, the continuum assumption for the multi-phase flow is similar to that for a single-
phase flow as discussed in 1.4. For example, the latter assumes a large number of molecules
over for any fluid control volume so that fluid properties such as density, pressure and velocity
can be taken as continuum values, i.e. Kn
f
0 so that individual molecular interactions can be
ignored.
The third assumption required is that the continuum Eulerian particle concentration field is
gradient resolved. This requires that the mixed-fluid volume is small enough to capture the
continuum particle concentration gradient (n
p
). This will be satisfied when the length-scale
associated with the continuum concentration gradients is much larger than that associated with
the control volume:

1/3
p p
x n / n

= < for a gradient-resolved particle concentration


4.15
Thus, the deterministic concentration gradient is grid independent if this criterion is satisfied.
It should be kept in mind that local variations in particle spacing associated with a random
distribution should not be considered (else this constraint can not be satisfied).
Since Eq. 4.14 automatically satisfies Eq. 4.5 and since Eq. 3.186 relates concentration to
spacing, all three of the Eulerian continuum assumptions may be satisfied by:
165

p p p p
x n / n

< < l for a particle concentration continuum


4.16
Thus, an Eulerian treatment requires a cell resolution length-scale which is both larger than the
mean inter-particle spacing and smaller than the deterministic concentration gradient length-
scale. This indicates that the cell discretization can have problems if it is too big (e.g., using
the entire domain for the cell volume would not capture any concentration gradients) or if it is
too small (e.g., movement of a single particle from the cell volume can cause the large changes
in the concentration).
Note that the instantaneous continuum approximation constraint (Eq. 4.16) may be relaxed
if just the mean properties are of interest. For example, the time-averaged particle number
density (Drew & Passman, 1999) is given by:

p p p
0
1
n n (t)dt

= =

as
4.17
If the total number of particles considered over the time integral is large, then
p
n can be
considered as a continuum even if the number of particles in a cell at a given time instant is
small or even zero. For example, a computational cell which has an average of 0.1 particles
based on Eq. 4.17 can be considered equivalent to a 10% probability that a particle will be
located in the cell volume at any given time. As such, the uncertainty of Eq. 4.13 would only
arise if one wishes to convert this probability into an instantaneous realization. Using such a
time-average allows the constraints of Eqs. 4.14 and 4.15 to be relaxed, but the continuum
approximation still requires that the cell-volume be well-posed (Eq. 4.5) so that:

p p
d x n / n < < for a mean particle concentration continuum
4.18
This can allow a much broader range of computational resolutions as compared to that of Eq.
4.16. In addition, use of a mean concentration gradient eliminates the issue of random particle
spacing. If an unsteady flow is considered, the averaging of Eq. 4.17 can instead be considered
over a time-step t. In this case, the requirements for temporal discretization of unsteady
concentration fields are analogous to those for spatial discretization of concentration gradient
fields, as shown in Fig. 4.4.


Eulerian Dispersed-Phase Equation of Motion

In the Eulerian approach, the particle concentration and momentum are solved with
transport equations in a manner similar to the density and momentum for the continuous-phase
(i.e., analogous to equations in Appendix A). To allow this, we assume that the particle
characteristics (such as n
p
or ) can be described as a continuum as discussed above. The
conservation equations for the Eulerian dispersed-phase mass, momentum and energy can be
derived in the same manner as that for the continuous-phase (A.1). For continuity, the rate of
mass change in a control volume is equal to the mass flux through its boundaries and any mass
source term within the control volume. Written in PDE form and ignoring the effects of
Brownian diffusion and particle collisions (at least until 4.2.2), this governing equation per
unit volume of the mixed-volume fluid in conservative form becomes:
166


( ) ( )
p p p p j
p p
j
m n m n v
n m
t x

+ =

` 4.19
Note that the above RHS Eulerian source term is also the RHS Lagrangian source term of Eq.
4.1c since both LHS terms are consistent. The particle momentum is obtained by equating the
rate of momentum change with the forces based on Eq. 4.1a:


( ) ( )
( ) ( )
p p i p p i j
p body int p i p ij
i
j
m n v m n v v
n F F m v p K
t x


+ = + + + +


` 4.20
Note that a RHS mass transfer term is a result of using a conservative form for the LHS.
The energy equation (assuming a continuum in the particle concentration) equates the rate
of particle energy change with convection and energy transfer (Crowe et al. 1998). This can be
obtained in a manner similar to Eq. A.15 assuming a thermal diffusion due to temperature
gradients and a mixed conductivity and assuming that the particles due no work (since there is
no particle pressure):


( ) ( )
( ) ( )
p p p p p p j
p p m p p p phase p p
j
m n e m n e v
n T n m + m e
t x

+ = + +

`
` ` k Q h 4.21
Note that the last term on the RHS also account for the heat transfer from the continuous-phase
and the energy required for phase change. In addition, a mass flux term as in Eq. 4.20 also
appears due to the conservative form.
The above transport equations can be considered in terms of steady-state properties if
the dispersed-phase field is ergodically stationary. In this case, a time-average of Eqs. 4.19-
4.21 assuming constant particle mass yields


( )
( )
( ) ( )
( )
( )
p j
j
p i j
p p body p int p coll p p ij
i
j p
p p j
p p
p p m
j p
n v
0
x
n v v
1
n F n F n F n p n K
x
n e v
n
n T
x

= + + + +

= +

`
Q
k

4.22a


4.22b

4.22c
Representations of the time-averaged terms will require turbulence models (A.5.5-A.5.6) and
will be discussed in Chapter 7.
Once the Eulerian conservation equations are discretized, each (cell-averaged) node in the
computational domain will include cell-averaged particle-phase variables but require the
constraints of Eq. 4.18. When coupled with an Eulerian representation of the continuous-phase,
this overall multiphase treatment is termed the Eulerian-Eulerian or two-fluid approach.
Importantly, this approach allows the dispersed-phase PDEs to be treated with the same
discretization techniques as used for the continuous-phase PDEs. It also conveniently allows
the same grid for both phases so that the grid resolution is equivalent (x
f
=x
p
). Because of
this the average interphase coupling terms are treated in the same manner for both phases so
167
that Eulerian-Eulerian approaches are especially efficient and accurate when two-way coupling
effects are present (to be discussed later).
To summarize, the Eulerian approach for the dispersed-phase: 1) requires a continuum
assumption for the particle characteristics, and 2) leads to PDEs for the particle characteristics
on a discrete grid similar to that for the continuous-phase characteristics. When coupled with
transport equations for the continuous-phase, this is often called the two-fluid method or the
Eulerian-Eulerian multiphase formulation. In contrast, the Lagrangian approach leads to
ODEs along the particle path. In either case, the particle relative velocity must be obtained and
the possible methods for this are discussed in the following section.

4.1.2. Mixed-Fluid vs. Separated vs. Weakly-Separated Approaches

Depending on how the particle relative velocity is treated, a second type of classification is
mixed-fluid vs. separated-fluid. For the separated-fluid approach, the continuous-phase
PDEs employ an Eulerian field of PDEs while the particle field is treated with a separate set of
PDEs for an Eulerian representation or ODEs for a Lagrangian representation. These are
coupled through a relative velocity as discussed above. For the mixed-fluid approach, the
particle relative velocity is considered negligible so that the surrounding fluid and particles are
considered numerically as one fluid with a single PDE for the mixture. In addition, there is a
third (hybrid) method called the weakly-separated-fluid approach, which generally assumes
that the relative velocity accelerations are small. These three methods are discussed below.


Mixed-Fluid Treatment

The mixed-fluid approximation assumes that the dispersed-phase and the continuous-phase
are in local kinetic and thermal equilibrium, i.e., the relative velocities and temperatures
between the two phases are negligible in comparison to variations of the overall flow-field

p f f
v u u
T T T

<
<
for mixed-fluid treatment
4.23a

4.23b
Neglecting these relative components at all points (x) in the domain yields the mixed fluid
values for velocity and temperature


m @p
m p
( ,t) ( ,t) ( ,t)
T ( ,t) T( ,t) T ( ,t)


v x u x v x
x x x

4.24a

4.24b
For flows with weak accelerations, the kinetic equilibrium is equivalent to the terminal velocity
being much less than any relevant continuous-phase velocity scales, e.g., w
term
u
D
for steady
macroscopic flows and w
term
u

for turbulent flows. From a dynamics point of view, this is


equivalent to having infinite interphase drag between the particles and the fluid, i.e. negligible
inertia and zero response time. Therefore, the mixed-fluid approach is consistent with particle
sizes so small that all the relevant Stokes numbers are much less than unity, e.g., St
D
1 if
macroscopic effects are important or St

1 if all turbulent features are important. If there are


variations in flow temperature, the temperature differences between the two phases are
similarly assumed to be negligible which coincides with infinite heat transfer rates. Therefore,
168
the mixed-fluid approach is only reasonable for very small particles with respect to relative
velocity (equivalent to an immiscible mixture distributed throughout a domain). Since the
velocities and temperatures of both phases are now assumed to be represented by single values,
this mixed-fluid approach has also been termed the Locally Homogeneous Flow approach
(Faeth, 1987), the Single-Fluid Scalar Transport approach, and the Modified-Density
approach.
A key aspect for the mixed-fluid method is the use of the mixed-fluid density as defined in
Eq. 3.153. The mixed-fluid density from an Eulerian reference frame is simply based on a cell-
average of the phase masses within its volume (Eq. 3.153):

m
(x,t) = (x,t)
p
(x,t) + [1-(x,t)]
f
(x,t) 4.25
This is illustrated for a one-dimensional discretization in Fig. 4.4 where the mixed-fluid density
can be assumed to vary in x or t so long as the number of particles within a space or time
increment is sufficiently large, i.e. the continuum approximation is ensured. If the particle and
surrounding-fluid densities are constant, the volume fraction alone is sufficient to characterize
the presence of the dispersed-phase throughout the domain. Otherwise, local pressure
variations can cause changes in the particle density (such as bubble volume changes with
hydrostatic pressure) and/or the continuous-fluid density (such as compressible flow
conditions). In this case, equations of state are needed to relate pressure to the phase densities.
Use of the mixed-fluid approximation results in a single set of momentum conservation
equations for the flow mixture (as opposed to one set for the continuous-phase and one set of
the dispersed-phase). Because of this, the mixed-fluid treatment allows significant numerical
simplicity for all dispersed conditions, and thus should be used when appropriate. Furthermore,
collision effects are generally negligible for a mixed-fluid since v=u everywhere. The resulting
mixture equations can thus be obtained for transport of mass, momentum and energy in
conservative form in a manner similar to that used for Eqs. A.5a, A.6a and A.15 yielding:


( )
( ) ( )
( ) ( )
( ) ( )
m m, j
m
j
m m,i m, j m m,i m,ij
m i
j i j
m m m, j
m,i m m
m,ij ij m m
j j
v
0
t x
v v v K
p
g
t x x x
e v
v e
K p T
t x x

+ =

+ = +



+ =

+ k

4.26a



4.26b


4.26c
This equation introduces a mixed-fluid viscous stress tensor (K
m,ij
) which can be written in
terms of a mixed-fluid viscosity


mj
mi
m,i j m ij m
j i
V
V 2
K
x x 3

= +




v
4.27
Recall that the mixed fluid viscosity for Stokesian conditions and small volume fractions is
given by Eq. 3.187.
Some key aspects of these equations are worth noting. First, the mixed-fluid set of
equations is inherently a two-way coupled system since all phases act in concert. Second, the
particle size and shape are directly employed (and effects not incorporated) since the relative
velocities are assumed insignificant. As a result, the particle characteristics to be solved for are
169
generally limited to the cell-averaged and
p
distributions (where the latter is simply constant
if the particle is incompressible). Chapter 5 will discuss details of the mixed-fluid state
equations and other characteristics.


Separated-Fluid Treatment

The separated-fluid treatment generally assumes the continuous-phase and dispersed-
phase have distinct transport equations. These can be in the form of either an Eulerian-
Eulerian or Eulerian-Lagrangian approach, whereby the interphase processes are related by the
relative differences. If we assume that the particle velocity is defined in an Eulerian sense such
that Eq. 4.14 is satisfied, then one may define an Eulerian relative velocity (similar to Eq. 1.9d)
as an average over many particles in a control volume centered at x so that:


( ,t) ( ,t) ( ,t) w x = v x u x
4.28
The mass, momentum and energy conservation equations of the dispersed phase were given by
Eq. 4.1 for the Lagrangian approach and by Eqs. 4.19-4.21 for the Eulerian approach (which
neglects particle collisions and Brownian diffusion). The latter can be re-written in terms of
particle volume fraction and the interphase force (Eq. 3.146):


( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
p p j
p
j p
p i p i j
int,i body,i p i ij
j p
p p p p j
p p phase p eff
j p
v
m
t x
v v v
F F m v p K
t x
e e v
m +e T
t x

+ =


+ = + + + +



+ = + +


`
`
`
` Q h k

4.29a


4.29b


4.29c
If the body force is only gravitational, then F
body,i
is given by Eq. 1.23, and if the continuous-
phase flow is incompressible and weakly coupled, then F
S,i
(last term on RHS of Eq. 4.29b)
can be expressed by Eqs. 1.69 and 4.6. These separated-fluid Eulerian-Eulerian multiphase
formulation will be discussed further in 7.1.1 while a similar discussion and presentation is
given in 7.2.5 for the Eulerian-Lagrangian multiphase formulation.
For two-way and three-way coupling, the continuous-phase mass, momentum and energy
must also take into account interphase transfer due to
p int p
m , F , and
`
` Q . However, these
conservation equations must also take into account that the continuous-phase mass only
occupies part of the mixed-fluid volume. For example, the continuous-phase fluid mass per
mixed-fluid volume is given by
f
(1 ) . Application of the Reynolds Transport Theorem of
Eq. A.4 to this quantity yields the continuous-phase continuity equation, where the interphase
mass transfer is included as a sink term:


[ ] f j
f
p
j p
(1 ) u
(1 )
m
t x


+ =

` 4.30
The RHS is thus equal and opposite to the source term used for the particle-phase Eulerian
transport of Eq. 4.29a.
170
For the momentum equation, the Reynolds Transport Theorem (Eq. A.1) and the Gauss
Divergence Theorem (Eq. A.3) can be used to describe the change in continuous-phase
momentum per unit volume, i.e.
f i
(1 ) u . This change can be related by Newtons second
law to the body, fluid-stress, and interphase coupling forces applied to the continuous-phase as:


[ ]
( )
f i j
f i
f i
j
i j ij
int,i p i
j p
(1 ) u u
(1 ) u
(1 ) g
t x
(K p )
(1 ) F m v
x



+ =

+ +




`

4.31
The first term on the RHS is simply the body force applied to the continuous-phase mass per
unit volume. The second term on the RHS can be obtained by subtracting the fluid-stress force
applied to the particle (Eqs. 4.6 and 1.69) from the total fluid-stress force to all phases:

( ) ( )( ) ( )( )
p
N
S,f m ij p,k ij m ij
k 1
p K p K 1 p K

=
= + + = +

F 4.32
This result is consistent with Eq. 1.67 and indicates that the portion applied to the continuous-
phase is the complement of the portion applied to the particles in Eq. 4.29b. A rigorous
derivation and discussion of this term is given by Prosperetti (2007). The third term is the
momentum transferred to the particle phase, which is equal to and opposite of the particle
momentum source terms of Eq. 4.29b. The interphase force term can be related to the fluid
dynamic particle surface force by Eq. 3.146:


( ) ( )
i j ij
int,i surf ,i
p p j
(K p )
F F
x


=





4.33
Substituting the RHS expression into Eq. 4.31 yields a common form on the continuous-phase
momentum transport:


[ ] ( )
surf ,i p i f i j i j ij f i
f i
j j p
F m v (1 ) u u (K p ) (1 ) u
(1 ) g
t x x
+

+ = +

`
4.34
Further assuming constant viscosity and density yields the momentum equation in Table 4.1.
The coupled energy equation for the continuous-phase can be similarly obtained in
conservative form by employing the same assumptions used for particle energy transport (Eq.
4.3) and neglecting the following effects: energy interactions associated with particle volume
change or shape changes, kinetic energy of the particle surface efflux due to mass transfer, and
energy release associated with chemical reactions. The resulting equation is:


[ ]
[ ]
( ) ( ) ( ) ( )
( )
f tot j
f tot
j
i j ij i
eff
f j j
j j
i
p p phase p p i i int,i i i j ij
p j
(1 ) e u
(1 ) e
t x
(1 )(K p )u
T
(1 ) g u (1 )
x x
v
m +e m v v F v K p
x



+ =




+ +


+ + + +


`
` `
k
Q h

4.35
A detailed derivation and accounting of these terms is given by Crowe et al. (1998) and an
additional shaft work term is identified by Brennen (2005). Herein, the physical interpretation
171
of these terms is outlined. The LHS terms (first line) include the temporal and spatial changes
in the total fluid energy within the surrounding fluid volume, and the first three RHS terms
(second line) are similar to those for the uncoupled single-phase energy equation given by Eq.
A.13 except that they are limited to the surrounding fluid volume fraction. These include the
work associated with moving the fluid against the body force (1
st
term) and against the fluid-
stress forces (2
nd
term), as well as the thermal diffusion within the continuous-phase volume
(3
rd
term). The remaining RHS terms (third line) are associated with the interphase energy
coupling. The term in the square brackets includes the heat and mass transfer terms which
appeared as sinks in the particle energy equation (Eq. 4.29c), the kinetic energy sink term
associated due to generation of particle mass, and the work required to move the particles
based on interphase force. The last term represents the work required to move the particles
based on fluid stress. These two work terms can be rearranged using Eq. 4.33 as follows:


( )
( )
( )
i j ij i
i
int,i i i j ij surf ,i i
p j p j
(K p )v
v
F v (K p ) F v
x x




+ = +

4.36
The first term on the RHS is the work associated the net surface force while the second term
combined with the second term on the RHS of Eq. 4.35 describes the work associated with
relative particle velocity against the fluid-stress forces.
If the mass loading and volume fractions are also small (1 and 1), then only one-way
coupling occurs and the continuous-phase governing equations (Eqs. 4.30-4.36) revert to the
single-phase equations in Appendix A. If these parameters are not small but the equilibrium
assumption of Eq. 4.24 is valid, then the continuous-phase Eulerian conservation equations can
be combined with the Eulerian particle-phase equations (Eq. 4.29) to yield the mixed-fluid
conservation equations (Eq. 4.26). These continuous-phase conservation equations neglect the
flowfield deflections around an individual particle surfaces but retain the viscous drag forces
on the particles themselves. Note that the continuous-phase flow equations can be treated as
inviscid in the limit of no particle influence but still include two-way coupling owing to drag of
the particles stemming from an Re
p
based on
f
. For example, one may simulate attenuation of
a shock-wave due to particles by neglecting K
ij
in Eqs. 4.31-4.35 while retaining F
int,i
or F
surf,i
.
The differences between the mixed-fluid and separated-fluid approaches are illustrated in
Fig. 4.5 and Table 4.1. Note that the mixed-fluid approach generally uses only the Eulerian
reference frame for both phases assuming negligible velocity and temperature differences
between the phases. In contrast, the separated-fluid approach allows for strong and unsteady
relative velocity differences, requires more differential equations, and includes both Eulerian-
Eulerian and Eulerian-Lagrangian formulations. The following section discusses a hybrid
approach for conditions which are in between that of separated flow and mixed-fluid flow.


Weakly-Separated-Fluid Treatment

An intermediate method between the mixed-fluid method and the separated-fluid method is
the weakly-separated-fluid or partially-mixed method. The simplest and most common
version of the weakly-separated-fluid method assumes that the particle relative acceleration is
negligible in comparison to the fluid acceleration.
172

t t

v u (

d D
d D
i.e.
t t
<
w u (

d D
d D
for weakly-separated fluid treatment 4.37
This assumption is consistent with negligible particle inertia and indicates that the
gravitational force alone determines the relative velocity, i.e.


@p term
( ,t) ( ,t) + = v x u x w
4.38
This is also called the terminal velocity approach. The particle is also considered to be in
thermal equilibrium: T
p
=T. As a result, no momentum or energy PDEs is needed for the
dispersed phase (Table 4.1). Thus, the continuous-phase equations only require the addition of
a dispersed-phase transport equation for the particle concentration, and the Eulerian and
Lagrangian versions of this transport are simply:


( )
( )
p j@p term, j p
p
j p
pi
i @p term,i
(u w )
m
t x
x
u w
t
+


+ =

= +
`
d
d

4.39a


4.39b
Since particle inertia is neglected, this approach is reasonably accurate when St1.
A slightly modified version of this approach can take into account small but finite particle
acceleration. As discussed in 5.3.1, this yields a perturbation correction to Eq. 4.38 which is
appropriate for small particle response time as:
( )
@p 2
@p term p p
+ (1 R) O
t




u
v = u w - +

D
D

4.40
This correction has been found to be reasonable for macroscopic flow with St
D
<1 or for
turbulent flows resolved by DNS with St

<1. This Stokes number restriction is less stringent


than that for the mixed-fluid approach but more severe than that for the separated-fluid
approach. Note that the partially-mixed or weakly-separated-fluid approach places no
restrictions on the terminal velocity magnitude, as is the case with the separated-fluid approach.
Generally, the weakly-separated-fluid approach employs one-way coupling so that Eq. 4.39
can be solved after the continuous-phase flow solution for u
j
is determined. This is convenient
since a single solution of the continuous-phase PDEs can be used with Eq. 4.40 to obtain
particle velocity fields for conditions with different terminal velocities, response times, and
density ratios. Because of this, Ferry & Balachandar (2001) named the Eulerian-Eulerian
version of this treatment (e.g., Eq. 4.39a) as the fast-Eulerian approach. Similarly, the
Eulerian-Lagrangian weakly-separated-fluid version (e.g., Eq. 4.39b) can be called the fast-
Lagrangian approach). If there is two-way coupling, these particle continuity equations must
be solved simultaneously with continuous-phase PDEs (Eqs. 4.30-4.35).
As with other approaches, the weakly-separated-fluid treatment can be applied in a time-
averaged sense for turbulent flow by decomposing the particle velocity into time-averaged and
fluctuating components. If the turbulence is homogenous and isotropic, the particle drag is
linear, and the particle concentration is uniform, the time-averaged particle velocity is given by


term
( ,t) ( ,t) + = v x u x w
4.41
173
Since the time-averaged continuous-phase is based only on mean integral-scale turbulence, the
time-averaged weakly-separated-fluid approach only requires St

<0.5 instead of St

<0.5 or
St
+
<0.5 (depending on whether particle is near wall or not) for fully-resolved turbulence.
Additional corrections to Eq. 4.41 can be made to include local gradients of the continuous-
phase flow field as well as preferential bias and non-linear drag (5.3.2).

4.1.3. Point-Force vs. Distributed-Force Approaches

As noted in the two previous sections, the separated-fluid and weakly-separated-fluid
treatments require knowledge of the surface forces acting on the particle (F
surf
) to determine its
movement. These forces are also needed if two-way and three-way coupling effects are to be
included for the continuous-phase fluid solution. In determining F
surf
, there are three primary
approaches: the "point-force" representation for particles smaller than the grid-scale (d<x),
the "distributed-force" representation for particles on the order of the grid-scale (d~x), and the
resolved-surface representation for particles much larger than the grid-scale (dx). The
first two are discussed in this section while the third is discussed in the following section.


Point-Force Treatment

The point-force (also referred to as the point-mass or point-volume) treatment is
consistent with the discussions in the previous chapters and the mixed-fluid and separated-fluid
treatments described above. The point-force technique has the advantage of eliminating the
need to simulate the detailed flow around the particle. For example, if Lagrangian particles are
considered with shape functions (discussed in B.1), then the relative velocity is given by


N
p p @p p p j j p
j 1
( , t) ( , t) ( , t) ( , t) (t) ( )

=
= =

w x v x u x v x u x
4.42
In this equation, N

is all the number of computational cells associated with a particle location.


Use of the shape function allows the velocity of all surrounding nodes to be included in the
evaluation of u
@p
. This is shown schematically in Fig. 4.6a for a two-dimensional grid. For an
Eulerian treatment of the dispersed-phase averaged over several particles, Eq. 4.28 can be used
to express the relative velocity in terms of velocities at the grid nodes:


i i i
( , t) ( , t) ( , t) = w x v x u x
4.43
As discussed in 1.3, the continuous-phase characteristics in either case do not include the
fluid disturbances associated with individual particles. This indicates that the particle should
be smaller than the length-scales associated with continuous-phase gradients. Since this is also
true for the discretization of the continuous-phase, it yields a size criterion based on
computational resolution:
d<x for point-force treatment 4.44
Thus, the particle size should be small compared to the continuous-phase discretization so that
interpolating to a particle location is reasonable and occurs within a single control volume. A
more rigorous criterion would be dx such that the particle will see only very weak
continuous-fluid gradients over the length scale associated with its diameter. As such, the
174
computational resolution for the continuous-phase flow must be as large or larger than the
particle diameter, while still being sufficiently small enough to provide grid-independent
results for the continuous-phase solution. Since the computational resolution is generally set
by the latter constraint, the point-force technique is generally limited to particles which are
smaller than any resolved spatial features within the computational domain. For example, a
fully-resolved description of a turbulent flow would require that the particle diameter be
smaller than the smallest eddy scales, i.e. d< away from walls or d
+
<1 near walls which is
consistent with Eq. 4.44. As such, the point-force technique is formally limited to particles
smaller than the micro-scale for resolved turbulence simulations.
Once the unhindered flow characteristics are known, a point-force equation (as discussed in
1.5) can be used describe the surface forces based on a linear combination of drag, lift, added-
mass, history, fluid-stress, etc.


F
surf
= F
D
+ F
L
+ F

+ F
H
+ F
S
+ etc.



4.45
The fluid dynamic forces on the particle can be obtained with empirical or theoretical
treatments (e.g., F
surf
=-3df
f
w if only drag is to be included). As will be discussed in
Chapter 6, there are a variety of point-force expressions to take into account aspects such as: a)
particle shape, size, and orientation; b) effects of Reynolds and Mach numbers; c) slip and no-
slip surface conditions, d) turbulence levels, etc. The surface force representations are tailored
to specific situations based on the physics of interest, test conditions, etc. As such, there is a
large variety and number (dozens) of point-force expressions which have appeared in the
multiphase literature. This non-uniqueness indicates there is no single standard expression
which should be applied to all conditions and one must instead choose the appropriate point-
force equations just as carefully as one chooses the appropriate numerical solution techniques.
If two-way coupling is to be included, the effects of the particles should be distributed back
in terms of cell-averaged properties for the Eulerian treatment of the continuous-phase given
by Eqs. 4.30, 4.31 and 4.35. If Lagrangian particles are employed, then the coupling
contribution to a continuous-phase node can be computed from the associated global shape
function and the computational volumes of the surrounding elements defined in B.3. For
example, the interphase force coupling terms of Eq. 4.31 at node j can be expressed as:


e
p
e
p
N
N
j p,k D,e,k
k 1
e 1 ,e
N
N
int int,k D,e,k
k 1
e 1
p ,e
j
1
1

=
=
=
=

=






F F

4.46a


4.46b
In these equations, N
e
is the number of elements associated with node j,
,e
is the volume of
each of those elements, N
p
is the number of particles associated with an element e,
p,k
is the
volume of each of those particles, F
int,k
is the interphase force for each of those particles, and
De,k
is the shape function for particle k associated with node j and element e. Use of shape
functions allows the particles contributions to vary linearly as it moves between two nodes.
An illustration of particle force mapping from the Lagrangian dispersed-phase nodes to the
continuous-phase nodes is shown schematically in Fig. 4.6b for a two-dimensional grid. This
can be extended to parcels by replacing N
p
with N
P
for the RHS summations and multiplying
the individual terms with the number of particles per parcel (N
pP
).
175

Note that Eq. 4.46 assumes that an evenly spaced distribution of Lagrangian particles will
lead to a uniform volume fraction and interphase force distribution on the Eulerian mesh. This
can be approximately ensured if computational cells contain one or more particles, i.e.

>
p
/.
This limits particle spacing to be less than the grid scale, resulting in an exception to Eq. 4.44:
d<x/
1/3
for Lagrangian point-force treatment with two-way coupling 4.47
The similar limitation for parcels is satisfied by Eq. 4.10.
For an Eulerian treatment of the dispersed-phase, the interphase coupling terms can be
obtained directly via the cell-averaged particle concentration. For example, the interphase
force per unit volume evaluated at the fluid node j is simply


( ) ( )
j int j
int
p p, j
j
, t , t

=




x F x
F
4.48
As compared to Eq. 4.46b, this Eulerian-Eulerian formulation is numerically convenient
(especially if there are many particles per computational cell) since communication of data
between the phases occurs at coincident node locations so that there is no interpolation error.
Furthermore, the Eulerian approach does not include the added restriction of Eq. 4.47. The
characteristics of the point-force method are summarized in Fig. 4.7a, and indicate that this
technique is limited to small particle sizes but is highly efficient for many particles.


Distributed-Force Treatment

The distributed-force or semi-resolved technique is employed when the particle diameter
is on the order of the spatial resolution of the continuous-phase, i.e.
d~x for distributed-force treatment 4.49
In this case, the variations in the flow properties in the region of the particle should be
considered since the fluid interactions between the two phases occurs over distances extending
several particle diameters (1.5.2). Furthermore, the feedback of the particle on the
surrounding fluid will be As such, the coupling should extend over a region larger than a single
grid cell (Fig. 4.7b). To include the surface force on the particle, there are two distributed
force methods: spatially-averaged and semi-resolved. The spatially-averaged approach
obtains unhindered conditions by surface and volume averages in the vicinity of the particle as
opposed to using a single interpolated point (u
@p
). As discussed in 7.3.1, this is approach is
generally limited to d<3x without two-way coupling, but d<x with two-way coupling. In
contrast, the semi-resolved approach directly integrates two-way coupling effects and thus
allows for larger particles (or finer resolutions) based on 5x<d<10x as discussed in 7.3.2.
Note that both techniques share a similar approach for distributing the interphase force on the
fluid.
For two-way coupling (impact of the particle on the continuous-phase flow), there are
many variants of the distributed-force technique. To impart F
int
over a region, a Lagrangian-
Eulerian force transfer function (Z
F
) can be used about the vicinity of the particle located at x
p
.
This is illustrated in Fig. 4.8b where the Lagrangian particle force is mapped onto the Eulerian
grid. Similar distributions can be applied for mass and heat transfer. The distributed-force
176
technique is thus as a hybrid between the point-force approach and the resolved-surface
approach (discussed below).

4.1.4. Resolved-Surface Approaches

For very large particles, the resolved-surface approach can be used if the detailed local flow
over the particle surface is numerically discretized and computed. This approach resolves the
fluid surface stresses so that they may be computationally integrated to compute the resulting
fluid dynamic forces on the particle. This method is also called the full DNS approach and
requires the computational resolution to sufficiently describe the detailed stress distribution
over the particle surface, e.g.
dx for resolved-surface treatment 4.50
This criterion requires many fluid nodes per individual particle, e.g., hundreds or thousands
depending on the particle geometry, Reynolds number, etc. As such, this technique is only
reasonable when there are few particles in the computational domain. However, this technique
has the advantage of allowing complex particle geometries and incorporating detailed non-
linear flowfields around the particles without resulting to theoretical or empirical surface force
descriptions, which may be inappropriate for complex particle or local fluid behavior (Fig. 4.8).
Since the surface-force is obtained from the surrounding flow, it does not require any force
decomposition in terms of lift, drag, added-mass, history, fluid-stress, etc. since all these
effects are directly incorporated by the discrete surface integration. In particular, no
assumptions of particle shape, particle Reynolds number, particle or flow acceleration, surface
conditions, flow gradients, etc. are required for use of this formulation. Thus, there is no need
to determine u
@p
for a particle relative velocity and no empirical or analytical force
expressions are required. Furthermore, in the case of a fluid-particle (e.g., droplet or bubble),
the interior fluid dynamics may also be directly simulated using internal discretization (this can
be important since the recirculation can affect the surface stresses). Also note that two-way,
three-way, and four-way coupling are automatically included in the resolved-surface approach
since the interstitial fluid and contact dynamics can be fully resolved as well. As such, the
resolved-surface technique is the most desirable in terms of accuracy as it allows the most
physically realistic surface force method, but is also the most computationally intensive
approach.
There are two primary treatments for the resolved-surface method which depend on how
the interface between the particle and the continuous phase is treated: the Surface-Fitted Grid
Method (SFGM) and the Continuous-Interface Method (CIM). These two approaches are
illustrated in Fig. 4.9 (where the outward normal of the interface is n) and are overviewed in
the following and discussed in more detail in Chapter 8.


Surface-Fitted Grid Method

The surface-fitted grid approach is typically used when the particle shape is simple (e.g., a
sphere or an ellipsoid) so that the grid resolution over the surface is straightforward. In this
case, the interface is discontinuous, i.e. infinitely thin (Fig. 4.9b). For a solid particle, there is
177
only flow field outside of the particle and the PDEs for U with
f
can be based on transport
equations discussed in Appendix A. For example an incompressible viscous flow momentum
equation is given in Table 4.1 for SFGM. The boundary condition on the particle surface can
be either slip if an inviscid approximation is reasonable or no-slip if viscosity is important.
The viscous boundary condition for a fluid particle interface which is contaminated (2.2.2)
will be the same as that for a solid no-slip surface. Numerical discretization and solution for
this flow can be based on the methods outlined in Appendix B. An example of a resolved-
surface SFGM vector field solution for a spherical solid particle with a viscous boundary
condition is shown in Fig. 4.10a, where the particle causes a wake flow.
For a fluid particle with significant recirculation, the particle interior should also be
discretized. In this case, a separate set of conservation equations with V and
p
is needed for
the internal particle fluid dynamics. Furthermore, the viscous boundary condition for a clean
interface is replaced with a balance of the tangential stresses between the continuous-phase to
that of the particle-phase at the surface:


( )
ij j ij j
p
K n K n =
4.51
In this equation, n
j
is the local surface normal unit vector and for a spherical particle this leads
to Eq. 1.58b. If the particle is also allowed to deform, then the normal stress conditions must
also be considered. If the viscous stresses associated with compressive strain are ignored
(reasonable for incompressible flow), the normal conditions for a deformed surface are given
by Eq. 1.6.
Once the flow field around the particle surface is known, the surface force can be
determined by numerically integrating the pressure and the continuous-phase shear stress
(based on U) over the discretized surface elements:


( )
surf ,i i ij j p
F Pn K n A = +


4.52
In this equation, A
p
is a discrete particle surface area element. This can be combined with Eq.
4.1a to move the particle to the next position with a new velocity and acceleration.


Continuous-Interface Method

There are several types of continuous-interface techniques and they can be broadly sub-
divided into two classes: those that treat fixed-shape particles with no internal flow description
(e.g., solid particles) and those that treat deformable fluid particles with an internally computed
flow (e.g., drops and bubbles). The former class is often used with complex solid particle
shapes where the surrounding flow field (U) is obtained at fluid nodes in the vicinity of the
interface and beyond. Since no internal particle velocity is computed, the centroid velocity (v)
is obtained explicitly by numerical integration of the particle surface stresses at the interface.
The latter class is the more common resolved-surface approach for fluid particles since it takes
advantage of the continuous-interface methods ability to handle complex interface dynamics
(8.2). Since it treats the interface as a fluid mixture with variable properties, it is sometimes
termed the Eulerian mixed-fluid interface approach. As shown in Table 4.1, this approach
uses a mixed-fluid velocity (V
m
) throughout the domain as a local volume fraction () to
indicate whether a cell contains no portion of the particle (=0), some portion of the particle
178
(0<<1), or is within the particle (=1). Solution of V
m
can be used to infer the surrounding
fluid velocity (U) where =0 and the internal particle velocity (V) where =1. The net particle
trajectory (v) is obtained implicitly based on the average movement of the interface surface
(which typically defined at = ). This class is illustrated in Fig. 4.9c. The main benefit of
this approach is that V
m
can be solved on a simple Cartesian grid (Fig. 4.9c) which generally
allows the fastest computation for a given number of grid points. This is particularly
advantageous if the particle shape is complex and/or changes due to fluid dynamic
deformations as it avoids the complex surface gridding needed for the SFGM approach. An
example of a bubble simulated with the CIM treatment is shown in Fig. 4.10b where it can be
seen that a complex shape results from the balance of fluid dynamic and surface tension forces.
For the mixed-fluid interface approach, an individual Eulerian cell which happens to be
located within the interface or in the interior of the particle will only contain a small fraction of
the total particle volume (

p
). This limitation is consistent with Eq. 4.50, i.e. the interface
should generally be smaller than the overall particle diameter to ensure that the interface
thickness does not significantly corrupt the solution. The finite interface thickness is non-
physical and is a result of the need to allow a continuum of the volume fraction so that an
Eulerian conservation equation for the mixture fraction and local density can be employed
across the interface. Thus, the interface is constructed to be as thin as possible while still
allowing a stable continuum variation of the mixed-fluid density across the grid, which
typically results in a thickness of 3-4 computational cells and special techniques must be used
to ensure that the interface does not numerically diffuse beyond a certain threshold.
If one neglects the distributed surface tension force, the resolved-surface PDEs (for xd)
are similar to the mixed-fluid PDEs (for dx) as shown in Table 4.1. However, there are key
differences in the volume fraction range and the fluid properties. For example, the mixed-fluid
approach employ an immiscible viscosity such as Eq. 3.187 (limited to small ), whereas the
CIM will employ a miscible viscosity based linearly on volume fraction (where 01):
( )
m f p
1 = +
4.53
Similar relationships can be employed for thermal conductivity, specific heats, etc. Thus, the
mixed-fluid approach assumes an immiscible mixture while the Continuous-interface Method
fluid properties and the equation of state are for a miscible mixture (this difference is illustrated
in Fig. 4.11). A key aspect of the CIM approach is that the surface force of Eq. 4.52 and the
stress boundary conditions of Eqs. 1.6 and 4.51 are implicitly and automatically satisfied. As
such, no separate ODEs are needed to update the particle velocity and rotation and F
surf
need
not be computed.


4.2. Comparing Physics Capabilties of Multiphase Methods

As discussed above, there are three main classifications for multiphase numerical
treatments: Lagrangian vs. Eulerian, Mixed-Fluid vs. Separated-Fluid, and Point-Force vs.
Resolved-Surface. As shown in Table 4.1, these various techniques have different equations
which must be numerically solved. Each of these descriptions has been used extensively for
multiphase flows, and the choice of which numerical approach to employ depends significantly
on which particle and fluid physics are relevant and of interest. This section contains an
overview of the relationships between the various numerical treatments and particle-fluid
179
phenomena. In particular, the computational approaches outlined in 4.1 will be reviewed
according to their capabilities (and limitations) in terms of describing the multiphase physics
outlined in Chapters 1-3.

4.2.1. Predicting Particle Dynamics

Predicting Particle Surface Forces and Relative Velocities

The most accurate approach is to use a resolved-surface technique to predict the particle
dynamics (Fig. 4.8). The particle surface stresses are based on the unsteady three-dimensional
flow field around the particle (and inside if the particle is a fluid). The resolved-surface
methods thus can capture directly the surface force and torque applied to the particle for a wide
variety of conditions. No assumption of linear decomposition of surface forces is needed since
the total stress on the surface automatically includes any linkages between drag, lift, added
mass effects, history forces, etc. Thus, the resolved-surface avoids any empirical point-force
corrections which may not be appropriate for non-spherical particles, e.g., the deforming
bubble shown in Fig. 4.10b.
Even when the particle is spherical, the resolved-surface method generally yields a more
accurate description of the surface force (and thus of the particle dynamics). For example,
Bagchi & Balachandar (2003) noted that a resolved-surface technique was needed to accurately
predict the insanities surface forces for d>10, i.e. the particle length-scale greatly exceeded
the Kolmogorov length-scale. Since more than 100 points were employed around the particle
circumference, this is consistent with the Eq. 4.50. The point-force approach was found to be
quite reasonable for reasonable for d<, and the distributed-force approach may be used for
intermediate particle sizes, e.g. d<3. Since the spectrally-resolved grid-spacing away from
the particle was about x~3, these limits are consistent with Eqs. 4.44 and 4.49.
In this case, it is helpful to consider the relationship between the micro-scale length ratios
(which determine the computational approach) and the micro-scale Stokes numbers (which
determine the relative inertia of the particles). Assuming Stokesian drag (f=1), the
Kolmogorov ratios can be related by combining Eqs. 3.1, 3.22a, A.92a


*
p
St d
18
c

=
+

4.54
For a particle density ratio of 900 (approximately that of water drops in air) where one wishes
to accurately predict the instantaneous surface forces, this suggests:

d>10 St

>5000 resolved-surface approach


d~3 St

~500 distributed-force with separated-fluid approach


d~ St

~50 point-force w/ separated-fluid approach


d~0.1 St

~0.5 point-force w/ weakly-separated-fluid approach


d<0.03 St

<0.05 mixed-fluid approach


4.55
However, the guidelines become more restrictive in terms of Stokes number at very low
particle density ratios. For example, gas bubbles in a liquid with c

= suggest:

d>10 St

>3 resolved-surface approach


d~3 St

~0.3 distributed-force approach


d< St

<0.03 mixed-fluid approach


4.56
180
These criteria similarly apply in wall-bounded flows (e.g. a resolved-surface approach is
needed for d
+
>10) since the near-wall micro-scale ratios can be related by combining Eqs. 3.1,
3.22b and A.101 as:


*
fr p
d St
d 18
y c
+
+

=
+

4.57
Therefore, caution should be used when applying point-force technique to low density particles
in that the instantaneous force may not be properly resolved. However, the low-frequency and
mean surface force effects can still be captured with a distributed-force technique for d or
d
+
1. As such, integral-scale and mean flow particle dynamics are generally captured with a
point-force or distributed-force technique so long as d, etc.


Predicting Particle-Wall Interactions

Capturing the wall interaction physics will depend on whether a resolved-surface, mixed-
fluid or point-force approach is employed. With a resolved-surface approach, the fluid
dynamic effects with wall interactions can be incorporated directly if gap region is sufficiently
discretized. For wall collisions, the reflection dynamics for solid particles can be captured by
modeling the internal solid mechanics but are more commonly described with coefficients of
restitution as a function of impact speed and particle properties (6.10). For fluid particles,
reflection can be captured with a deformable interface by including surface tension effects. In
the extreme of a mixed-fluid approach, the particles move in unison with the continuous-phase
so that wall boundary conditions are placed only on v
m
.
For a point-force separated-fluid approach, the capability to handle the particle-wall
interactions depends on the type of boundary conditions. If the particles are near the wall, the
fluid dynamic interaction can often be modeled within the point-force approximation for
simple conditions, e.g., spheres at creeping flow conditions, as discussed in 6.8. Wall-contact
interactions can often be handled with a boundary condition. For deposition or stick conditions,
the boundary condition is free so that v is unconstrained at the surface and particles simply
flux into the wall (Fig. 4.12a). The flux rate per unit wall area is based on the particle
concentration and the particle velocity normal to the wall surface (v

). There is no major
problem with either the Eulerian or Lagrangian methods in representing this behavior.
However, if the particles reflect, slide or roll along the surface, then the reference frame can
have a significant effect in terms of capturing the associated physics.
A Lagrangian approach can explicitly determine when a particles surface will intersect
with the wall surface. The rebounding particle velocity can be determined based on models for
the coefficients of restitution and friction. For example, a reflecting boundary condition will
re-direct the particle velocity once within a particle radius of the surface (Fig. 4.12b). If
parallel and normal coefficients of restitution (e

and e

) are known, then the outgoing


velocities are related to the incoming velocities by


out in
out in
v v
v v

=
=
| | |
e
e

4.58a
4.58b
181
This boundary condition ensures mass conservation (zero net flux through the wall). Similar
relationships can be used to obtain a Lagrangian sliding condition (e

=0) and a change or


preservation of the particle rotation (6.10).
In contrast, the Eulerian representation of a particle velocity field is limited to a single
particle velocity for a given control volume. Since mass conservation dictates a zero-flux
condition to the wall, the cell-averaged velocity must be parallel to the wall, i.e.


v 0

=
4.59
Thus the Eulerian continuum representation of v can not be multi-valued to represent both the
particles moving towards and away from the wall. As such, it incorrectly describes the wall-
normal momentum variations. The Eulerian scheme also has difficulties in terms of
representing the tangential velocity reductions associated with e
&
since these are associated
with discrete collisions. Because of this, the tangential effects are simply ignored with this
approach.


Predicting Size Distributions

To simulate particles with a size distribution, generally the distribution is specified at some
initial time or location and one must predict the ensuing particle velocities. When the particles
within the computational domain are not all of a single size or shape, the Lagrangian technique
can be simply combined with statistical sampling to represent the particle size distribution. For
example, Log-Normal or Rosin-Rammler functions (2.1.3) can be used with a random number
generator with a statistically large number of particle trajectories to reasonably recover the size
distribution PDF. If parcels are employed for clouds of monodisperse particles, then a similar
number of trajectories are needed to ensure recovery of the PDF.
To incorporate polydispersion with the Eulerian treatment, one may employ an effective
velocity and diameter for the particles within the control volume (see 1.3). However, the
particle velocity will be different from the predicted velocity if the size distribution is large or
particle Reynolds number is small. This can be seen by comparing the actual relative particle
velocity for a particular diameter with the computed (effective) particle velocity and diameter
by Eq. 2.19 that

( )
3 J
J
eff 3J
d / d

= w w


4.60
For a Newton-based drag (with constant drag coefficient and Y=2), a 10% diameter deviation
from mean diameter (d
3Y
) would correspond to a relative velocity error of 3%. However, a
Stokesian particle (Re
p
0) will have Y=1 such that a diameter deviation of 10% within a
control volume will lead to a relative velocity error of about 20%. Similar problems can arise
due to initial conditions, flow-field gradients, and turbulent diffusion and these errors can
accumulate over time. Thus, an Eulerian particle field computed with effective diameters and
velocities can only be used to represent only a modest variation of size diameters.
If the particle size distribution is broad and an Eulerian approach is still preferred (due to
two-way coupling and/or large numbers of particles), then multiple Eulerian particle fields can
be employed. In this case, each field represents a bin of particles with a particular diameter
range and mean diameter. The number of particles per unit volume then becomes a vector
182
(n
pk
) where k indicates the bin number (Crowe et al. 1998). This approach is sometimes called
the multi-group, multi-fluid or multi-bin approach. Each bin can be specified with a
specific particle diameter range and represented by a bin diameter d
k
and velocity v
k
(other
characteristics can be similarly discretized, e.g., temperature). The sum of all the volume
fractions from these bins represents the total volume fraction of the dispersed-phase


K
k
k 1

=
=


4.61
In this equation, K represents the total number of bins. As such, K transport equations are
needed for each Eulerian particle group based on a bins effective diameter and velocity:


( )
( ) ( )
( )
k i,k
k
j
k p i,k j,k k i,k
k
k p i surf coll
i,k
j p,k
v
0
t x
v v v
g F F
t x

+ =

+ = + +


4.62a


4.62b
Thus, the general dispersed-phase velocity vector is now represented as a tensor which is a
function of both direction and bin number. This decomposition of the size PDF into several
fields can also accommodate mass exchanges in between bins.
As may be expected, the above Eulerian bin approach can be computationally intensive
since the number of dispersed-phase variables at each node will increase linearly with the
number of bin (Eulerian fields) than are used. To minimize computational memory and CPU,
one typically wishes to minimize the number of bins by making each bin as wide (in terms of
particle diameter range) as reasonable. The bin width can be specified according to the desired
accuracy, e.g., Eq. 4.60 can be used to estimate velocity variations if gravity dominates. If the
particles are of sufficiently small Stokes number, the bin momentum equation can be used (e.g.,
Eq. 4.62b). Another way to address the polydisperse issue for an Eulerian treatment is assume
a specific PDFs form and then numerically solve for the mean diameter and broadness at each
location, i.e. develop transport PDEs of the PDF constants (e.g., d
RR
and h
RR
of Eq. 2.23) and
then numerically solve these throughout the discretized domain (see 7.1.1). This tends to be
more efficient than bin methods for simple flowfields, but can be difficult to incorporate
complex geometries and wall boundary conditions, etc.


4.2.2. Predicting Turbulent Dispersion, Diffusion and Biases

Predicting the impact of continuous-phase turbulence on the particle trajectories and
concentration fields is highly dependent on the numerical modeling used to describe the
continuous-phase turbulence. In general, the choice of continuous-phase treatment should
mirror the particle-fluid physics desired, particularly with respect to the length and time scales.
As discussed in A.5.5-A.5.7, there is a wide variety of continuous-phase techniques in terms
of ability to resolve turbulent structures including: DNS (resolves all the eddies), LES (resolves
the eddies into the inertial sub-range), POD (resolves just the most energetic eddies), and
RANS (no eddies resolved, just integral-scales identified). Based on the concept of spatially-
183
filtered equations using a filter G, each technique can then be characterized by a filter length-
scale,
G
, below which eddies structures will be directly simulated. For the techniques
mentioned above, the filter length-scale can be related to various turbulent length-scales:
~ for DNS, <
G
< for LES,
G
~ for POD, and
G
~ for RANS.


Filter Stokes Number

To describe the impact on particle trajectories with specific response times, it is convenient
to convert these filter length-scales to filter time-scales. To do this, one may decompose the
physical velocity field associated with the various approaches into a resolved portion (

u )
which is computed directly, and an unresolved portion (

u ) whose effects are modeled:




( , t) ( , t) ( , t)

= + u x u x u x
4.63
The unresolved (spatially-filtered) portion of the flow-field requires modeling and can be
characterized by the spatially-filtered kinetic energy (k

) discussed in A.5.7. For isotropic


turbulence, the sub-grid time-scale can be approximated in a manner similar to Eq. A.91 using
Eqs. A.133 and A.137:


1/ 2
c k / c / k

4.64
From this and the particle response time (Eq. 3.1), one can define the filter Stokes number


p
St /


4.65
This parameter can then be used to determine which particles have a large enough response
time to see a sufficiently resolved flow-field to accurately capture their trajectories:


St 1

> all particle dynamics are captured


4.66
This condition thus allows direct capture of mean diffusion (including anisotropic and non-
homogeneous turbulence effects), non-linear drag and preferential concentration effects such as
preferential bias and clustering bias.
In the other extreme, particles with St

1 (e.g., RANS flows where St

=0) will have too


short of a frequency response to capture particle inertia effects and will effectively behave as
passive tracers to the resolved flow structures. Therefore, the physics of non-linear drag bias
preferential bias, clustering bias, and turbo-phoresis can not be directly captured. However,
some empirical models can be used to mimic these effects as will be discussed in Chapters 5
and 7 and are briefly outlined below.


Filtered Diffusion Techniques

Mean particle diffusion can be modeled using the modeled turbulent kinetic energy (k

)
and particle characteristics based on the analysis presented in 3.5. In terms of accuracy, this
presents two problems. First, such models depend on the accuracy of the continuous-phase
RANS predictions for mean velocities, turbulent kinetic energy, dissipation, etc. which have
been found to be problematic for complex geometries or complex NAsT flows (Piomelli, 1997).
184
Second, the particle diffusion models are similarly based on empirical arguments and
coefficients (e.g., c

) so that they are also not strongly robust in complex geometries or flows.
For a point-force separated-fluid approach there are two basic mean diffusion techniques: 1)
deterministic diffusion for the Eulerian particle approach and 2) stochastic diffusion for the
Lagrangian particle approach. These two techniques are overviewed in the following, and
contrasted according to their ability to capture the diffusion physics for HIST flows with
constant particle diameter and density.
The deterministic diffusion approach is best suited for Eulerian treatments of the particle
and employs the diffusion particle rate (
p
, defined in 3.5.2) and relates it to the particle
velocity fluctuations via the gradient transport model (Crowe et al. 1998)
i p
i
v
x


4.67
This approach is equivalent to the Fickian diffusion applied in the single-phase flow (Eq.
A.105). From this, the conservation of particle concentration of Eq. 4.39a can be written as

( )
i
p
i i i
v
x x x

=




4.68
For this approach, the time of release can not be readily incorporated. Therefore, a long-time
diffusion rate (
p
) is generally used in this equation. This diffusion for HIST can be obtained
from 3.5 but is more commonly modified to employ the local relative velocity and filtered
turbulent energy via Eq. A.72 as

1/ 2
2
p
2 w
1+
c 2k / 3






=




4.69
The fluid turbulent diffusion rate can in turn be related to properties of the RANS solution, e.g.,

2
t t
c k
Sc Sc

4.70
The turbulent Schmidt number is typically a constant (e.g., Sc
turb
=0.7). Thus a uniform
flowfield will result in a uniform spread rate based on long-time diffusion as shown in Fig.
4.13b (whereas short-time and intermediate-time physics are difficult to include).
Incorporating additional complexities in terms of the flow and geometries is discussed in
7.1.5.
For a Lagrangian approach, stochastic techniques are typically used to model the mean
turbulent diffusion. The simplest of these is the discontinuous random-walk (DRW) approach
where the continuous-phase velocity seen by the particle is the sum of the mean velocity field
and a fluctuation component which can be derived from the rms and a random number
generator:

i p i p i,rms p i
u ( , t) u ( ) u ( ) (t) = + x x x
4.71
185
In this equation, is a Gaussian random number with a zero mean and a unity variance which
recovers the mean and rms values of the LHS for a large number of samples. For isotropic
turbulence, the rms values are based on the filtered kinetic energy:

rms
u 2k / 3

= 4.72
In order to ensure the correct diffusion time-scale for these perturbation, a clock-time (
clock
) is
initiated once a random number has been sampled in that direction. The velocity perturbation
is kept fixed until this clock-time exceeds the eddy-interaction time of the particle after which
it is re-sampled and the time-clock is restarted (details of this technique will be given in 7.2.5).
The resulting velocity fluctuations are qualitatively illustrated in Fig. 4.14, which gives the
trajectories like those drawn in Fig. 4.13c. As such, this provides a net diffusion if a
statistically large number of particle paths is considered. Lagrangian stochastic approaches
allow a temporal development from short-time to long-time diffusion and have been successful
for mean particle diffusion in NAsT flow. However, as with the Eulerian approach, its
performance depends on accurate RANS predictions of the continuous-phase turbulence. In
addition, it should be considered in terms of turbulent biases, as discussed below.


Capturing or Modeling Turbulent Biases

Depending on the numerical method employed, the different types of turbulent bias
mentioned in 3.5.4 and 3.5.5 can be captured directly, modeled approximately or neglected
altogether. These dependencies are summarized in Table 4.2 for various approaches, in order
of increasing complexity required to compute the particle velocity field. The mixed-fluid
approach allows the simplest description of the multiphase field, but cannot account for any of
the relative velocity effects. The weakly-separated-fluid approach avoids this limitation and
allows most of the relative velocity effects to be modeled without need of ODEs or PDEs for
the particle velocity field, so long as the relevant Stokes number is on the order of unity or less.
However, it assumes a quasi-equilibrium condition for all cells so that initial conditions effects
must be neglected (this would be a problem for a spray for which droplets are far from terminal
velocity until far downstream). The separated-fluid approach can capture directly effects of
initial conditions, non-linear drag bias, turbo-phoresis, preferential bias and clustering bias.
However, incorporating clustering bias requires that the three-way coupling among groups of
particle be simulated which requires: a) fully resolving the particle surfaces (i.e. resolved-
surface approach), b) using theoretical interactions available only for limited conditions (e.g.,
creeping flow or inviscid interactions at large distances), or c) models of three-way coupling
(which are largely empirical).
If a time-averaged approach is employed, then the effects of turbo-phoresis, preferential
bias, and clustering bias cannot be incorporated directly and instead require
theoretical/empirical modeling. An advantage of the Lagrangian description of the particles for
a time-averaged approach is that the non-linear drag effects can be taken into account directly
since relative velocity fluctuations are incorporated. In contrast, non-linear drag effects must
be modeled with an Eulerian particle description.
If a Large Eddy Simulation approach is employed and the particles have substantial inertia
compared to the sub-grid time-scales, then the particle dynamics will be independent of the
sub-grid flowfield (Eq. 4.66). In this case, simply using the resolved continuous-phase flow
186
field for the particle trajectories will allow bias effects to be captured as found for DNS
simulations. If the LES temporal resolution is coarsely resolved or the particle dynamics are of
high frequency such that the opposite limit is reached (St

1), then the particle velocity effects


are only captured in a manner similar to that of a RANS approach, i.e. any bias effects must be
modeled. For an intermediate condition of St

<1, the weakly-separated-fluid approach can be


used to simulate the sub-grid particle dynamics. On the other hand, if LES filter and particle
conditions allow the resolved flow time-scales to be sufficiently small such that St

>1, then
unsteady effects associated with turbo-phoresis, non-linear drag, and preferential bias can be
approximately incorporated without modeling.


Predicting Particle Brownian Motion

To capture Brownian motion physics, the random molecular interactions with an
individual particle can be represented numerically by two extremes: full resolution of
individual molecule-particle collisions (as in Fig. 1.23) and mean diffusion based on
empirical/theoretical models for the net effect of many such interactions as discussed in 3.6.
Because a fully-resolved approach is generally impractical due to the extremely large number
of molecular interactions associated with even 1 m particles, mean diffusion techniques are
generally used when non-continuum particle effects are weak (Kn
p
1). Typically, this involves
a mean diffusion technique coupled with a point-force model. Basic methods for both Eulerian
and Lagrangian particle reference frames are outlined below.
For an Eulerian particle representation in laminar flow, theoretical Brownian diffusion
rates (e.g., Eq. 3.139) can be applied to the particle mass transport equation of Eq. 4.29a in a
manner similar to that for turbulent diffusion (Eq. 4.68) and molecular diffusion (Eq. A.50a).
The resulting transport equation is


( )
( ) ( )
p
p p Br p
p
m
t


+ = +


v ` 4.73
An additional term can be added to the RHS with respect to diffusion due to particle-particle
collisions (though such a term is often non-linear). This approach captures the long-time
diffusion associated with monodisperse particles. If there are a large variety of particle sizes,
then one may need to consider several Eulerian-phase fields (e.g., using Eq. 4.62) or turn to a
Lagrangian approach. In the limit as the particle size approaches the molecule size, the
diffusion rate tends to
f
so that the transport mechanism reverts to that for laminar species
diffusion (Eq. A.50). For turbulent flow,
Br pt
< so that Brownian diffusion will be
generally negligible in a computational approach that includes time-averaged turbulence.
Lagrangian approaches for mean Brownian diffusion are advantageous in that they may
handle a large size distribution, intermediate diffusion times, and wall reflections in a straight-
forward manner. In this approach, the unsteady Brownian force of Eq. 3.131 can be simulated
in a stochastic fashion to compute the discrete particle paths, which are then averaged over
many realizations. The surrounding velocity seen at the particle (u
i@p
) can be specified in
terms of a continuum component and a stochastic (Brownian) component. The latter can be
modeled with an approach is similar to Eq. 3.122 by including a Brownian Force based on a
random number generator. Further details of this approach will be given in 7.2.6.
187


4.2.3. Predicting Two-, Three- and Four-Way Coupling

As with one-way coupling physics discussed above, resolved-surface approaches can also
directly handle two-way, three-way, and four-way coupling aspects. In particular, resolved-
surface approaches determine the detailed flow around the particle so that two-way coupling is
captured directly. Also, the fluid between particles is numerically discretized so that the
particle-particle fluid dynamic interactions (e.g., drafting which yields reduced drag for a
particle trailing in the wake of another particle) can be captured (three-way coupling). By
considering the contact mechanics between the interfaces of two or more particles, the
resolved-surface method can also handle four-way coupling directly including such aspects as
kissing and tumbling (Fig. 3.45) which can continuously rearrange particle distributions.
Furthermore, proper interface treatment for splitting and amalgamation of particles can
simulate individual break-up and coalescence dynamics.
When point-force approaches are used instead, many of the above aspects can only be
empirically or analytically modeled for simple conditions (e.g., spherical solid particles in
nearly uniform continuous-phase flow). This is discussed below in terms of Eulerian and
Lagrangian reference frames for two-way, three-way and four-way coupling.


Point-Force Two-Way Coupling

For either Lagrangian or Eulerian point-force treatments, inclusion of two-way coupling
effects requires description of the particle interphase forces on the continuous-phase transport
equations so that the phases are fully coupled. The interphase coupling force (F
int
as defined
in Eq. 3.146) is equal in magnitude and opposite in direction to the particle force acting on the
continuous-phase (-F
int
). Thus, any physical shortcoming associated with the point-force
predictions of the particle interphase force or particle positions (as discussed in 4.2.1) will
necessarily reduce the accuracy of the two-way coupling aspects.
Additional issues can arise with respect to the particle reference frame. The use of an
Eulerian frame for the dispersed-phase (4.1.3) allows u and v to be discretized in the same
fashion so that w is everywhere defined with a consistent control volume. This Eulerian-
Eulerian approach allows both phases to be treated with similar discretization and numerical
techniques. When Lagrangian approaches are employed, the sum of interphase particle force
per mixed-fluid volume (n
p
F
int
) can be distributed to the continuous-phase nodes by shape
functions as in Fig. 4.6b. However, numerical errors can occur if the number of parcels per
cell (N
P
) is not sufficiently high:
N
P
>1 for point-force treatment with two-way coupling 4.74
Often four or more parcels per cell are employed for accuracy, but this leads to significant
computational overhead communicating the parcel information to the Eulerian reference frame,
the Lagrangian approach is generally less computationally efficient. For example, Druzhinin
& Elghobashi (1998, 1999) examined two-way coupling in a turbulent flow with DNS and
188
found the Eulerian-Eulerian treatment to be almost an order of magnitude more efficient than
the Eulerian-Lagrangian treatment for similar accuracy levels.


Point-Force Three-Way Coupling

In a Lagrangian point-force approach, it is possible to track each particle and its neighbors
individually so that three-way coupling effects can be modeled based on particle-particle fluid
dynamic distances. However, this approach is computationally impractical for large numbers
of particles. Furthermore, corrections to the isolated drag force based on separation distance
(l
p-p
) are not well understood for finite Re
p
conditions since the interactions are anisotropic and
often involve several particles (not just binary interactions). As such, it is simpler and much
more common to employ effective viscosity treatments based on mean particle concentrations
(Eq. 3.187). More detailed treatment will be discussed in 6.9. Such ensemble-averaged
approaches can be used for Lagrangian particle and parcels as well as the Eulerian dispersed-
phase treatment by modifying F
D
accordingly. However, such ensemble-averaged approaches
cannot directly incorporate individual dynamics of particle-fluid-particle events, which cause
attraction or repulsion (e.g., drafting, kissing, and tumbling).


Point-Force Four-Way Coupling

To incorporate four-way coupling with a point-force approach, one may employ an
approach that models each particle-particle interaction or simply tries to capture the mean
effects of many interactions. Capturing every particle interaction in a Lagrangian reference is
the most accurate technique for the same reasons associated with particle-wall reflections
(4.2.1). To determine when such collisions occur, each particle path must be considered in
terms of the neighboring particle positions for each time-step to determine potential collision
partners. Once a collision is identified, the collision dynamics can be obtained by models of
the restitution and friction (Crowe et al. 1988). The simplest particle-particle collision model
is conservation of linear momentum, which assumes solid elastic spheres with no mass flux
interaction (no coalescence or break-up). However, finite energy losses can be straightforward
to incorporate via coefficients of restitution (for linear particle momentum) and friction (for
angular particle momentum) such as that used in the particle-wall interaction model.
Incorporation of non-spherical shapes substantially complicates particle collision detection and
momentum interaction such that empirical probabilistic models are generally employed. The
same is true for break-up and coalescence physics.
If the number of particles in the system is large (e.g., 10
5
or more) and/or the collision
frequency becomes high (St
coll
1), then it may be more computational efficient to capture only
the mean collisional effects. The mean collisional approach is most easily handled with an
Eulerian description of the particles based on collision rates given in 3.8.2. However, when
the flow approaches a dense condition in terms of particle-particle collisions (St
coll
1), the
ensemble-averaged particle collisions can be modeled as an effective stresses on the particle
concentration. In particular, collisions tend to increase the particle velocity fluctuations and
tend to disperse the particles to regions of lower concentration. These stresses can be
incorporated into the particle transport equations using kinetic theory concepts such as
189
granular-temperature (which stems from gas kinetic models which relate random molecular
motion to gas pressure). This approach is not described further since this text is focused on
dispersed flow (Eq. 3.189), but additional details for dense flow treatment are given by
Gidaspow (1994) and Crowe et al. (1998).


4.2.4. Physics-Based Choices of Multiphase Methodologies

It is instructive to consider the links between the various dispersed-phase numerical
approaches and the continuous-phase numerical approaches based on the above discussions.
The primary treatments discussed above for the continuous-phase and dispersed-phase
formulations are presented in Fig. 4.15 in terms of the relevant physics. On the top of this
figure, the various continuous-phase methods are determined by the type of flow and the
phenomena of interest. The chosen approach dictates the smallest relevant continuous-phase
length-scale to be resolved (l
min
). For example, this length-scale for a mean flow description
could be a shear layer or boundary layer thickness (). The l
min
, in turn, determines the
minimum continuous-phase resolution needed (x). The lower portion of this figure then
compares this resolution to the particle size (d) to determine the appropriate dispersed-phase
approach (e.g., point-force, distributed-force or resolved-surface). The appropriate reference
frame (Eulerian or Lagrangian) for the particle approach is then based on which multiphase
physics are most critical.


4.3. Comparing Computational Costs of Multiphase Methods

Ideally one would like to employ the most accurate techniques available for a given two-
phase flow problem. For example, if the flow is turbulent, a direct numerical simulation
(DNS) description for the continuous-phase would be preferred. If the particle Reynolds
number is not small or the particle has complex shape or surface condition, then a resolved-
surface simulation (RSS) is preferred. These approaches are desirable from a physics
standpoint as they do not rely on any empirical approximations, as discussed above.
Unfortunately, the computational resources required for these approaches are often too
intensive for most engineering calculations. For example, a high flow Reynolds number may
prohibit DNS while a high number of particles may prohibit RSS (and trying to accomplish
DNS and RSS simultaneously is even more challenging). Because the practical issue of
computational resources must also be weighed in the decision of computational strategy, most
multiphase flows are simulated with some level of modeling. Therefore, the previous
discussion on accuracy of numerical approaches should be complemented with discussions
on computational cost, which is the subject of this section.
To give a guide regarding the cost of various techniques, estimates of the required
computational resources for various approaches are put forth. The resources are most closely
tied to numerical resolution and so will be characterized in this manner. The general principle
for numerical resolution is to employ a sufficiently high spatial and temporal resolution such
that the results are discretization-convergent (further decreases in grid or time-step size do not
lead to significant differences in the predicted results). Assuming an Eulerian description of
190
the continuous-phase, the number of unknowns can be estimated as proportional to the number
of continuous-phase grid cells (N
f
). As such, one may expect CPU time and memory to also
scale with N
f
. Similarly, the computational resources required for the dispersed-phase can be
expected to scale with the number of dispersed-phase cells or particle locations (N
d
), where
particle properties (e.g., velocity, position, etc.) must be determined. If this information is
known for a specific numerical approach, one may roughly gauge the order of CPU
requirements. This can be correlated with the previous discussion on the fluid-particle physics
that can be captured for a particular approach, so that one may estimate the physics one can
afford to resolve within a given budget of computational resources.
It should be noted that the following estimates for N
f
and N
d
are rough guides only, since
the actual computational resources can vary widely due to variations in dimensionality,
flowfield complexity or technique aspects. For example, a compressible viscous flow will
require more CPU and unknowns per nodal location than that for a two-dimensional inviscid
irrotational incompressible flow. Furthermore, the cost of computing unknowns depends
markedly on the type of time-stepping (explicit vs. implicit, single-step vs. multi-step, etc.) and
the computer processing speed and architecture (e.g., massively parallel computers are better
suited for explicit techniques while serial machines are better suited for implicit techniques).
In addition, the below approximations for the number of continuous-phase or dispersed-phase
cells do not take into account substantial differences which may occur due to variations in
boundary conditions (e.g., a complex domain will often require more computational cells

than a
similarly-sized spherical or cubic domain) or due to spatial discretization accuracy (e.g.,
spectral methods in simple domains may allow higher spatial accuracy for a given number of
unknowns). Note also that inclusion of two-way coupling does not substantially increase the
number of unknowns but the numerical solutions may require longer times for convergence or
may need to be re-run for each new particle condition.

4.3.1. Number of Cells for Continuous-Phase Methods

Resolution Length Scales

The primary factors that determine the number of continuous-phase cells in the domain (N
f
)
are the dimensionality (e.g., 2-D vs. 3-D) and the range of numerical length-scales. With
respect to the length-scales, the smallest will be based on the minimum numerical spatial
resolution while the largest will be based on the domain scales. For the latter, an average
domain length-scale (D) can be based on the computational domain volume, i.e. D
3
~D
x
D
y
D
z
.
On the other hand, the minimum length scale is often subjective or not known until grid
dependency studies are completed. Some guidelines are discussed below for uncoupled
continuous-phase flows and discussed later for coupled continuous-phase conditions.
In general, the grid resolution is based on l
min
as noted in Fig. 4.15. For many flows, this
length-scale tends to be based on geometric features. In such a case, we typically require about
5-20 grid points to resolve the smallest feature, e.g.,

min min
x /10 l 4.75
Note that l
min
is not necessarily the smallest possible feature but instead the smallest feature
which the user deems as physically relevant. For example, resolution of roughness elements
191
on a wall can be important since the elements can significantly affect particle deposition (Fig.
1.16a). However, the effect of roughness elements can also be treated in an average sense by
imposing an appropriate boundary condition at the surface for rough walls or may even be
neglected altogether if the elements are small.
In addition, the grid resolution can be based on flow features. For inviscid flows, shock-
waves or other contact discontinuities arise in the system and these may require high resolution
in their vicinity to ensure that they are resolved to reasonable thicknesses. For laminar
boundary layer flows, a mesh with about 20 points across the boundary layer thickness () is
generally sufficient. Assuming that is in the direction of the y-axis, this gives

min
y / 20 4.76
Typically, the gradients in the span-wise and streamwise directions are larger for attached
flows so that resolution in these directions can be anisotropic. However, flow separation can
lead to isotropic flow features and thus require a more uniform resolution.
For turbulent flows, the spatial resolution will depend primarily on whether any spatial-
filtering is employed as discussed in A.5.7. A time-averaged RANS approach is by far the
least computationally intensive, especially for flows which are 2-D for both domain geometry
and boundary conditions. If the flow includes only shear layers (of thickness ) then the above
resolutions used for laminar flow will generally suffice for RANS flows. For spatially-filtered
approaches like LES, resolution down to the sub-grid cut-off scale will be required:

min min min G
y x z
4.77
For DNS, all scales are resolved down to the Kolmogorov scales, i.e.

min min min
y x z 4.78
For wall-bounded flows, additional near-wall resolution is needed because of the viscous sub-
layer so that a minimum resolution (based on inner units of Eq. A.101) is generally given for
RANS, LES and DNS by

min
y 1
+
4.79
For RANS flows, the streamwise and spanwise resolutions will be similar to those given above
for laminar boundary layers. For spatially-filtered approaches like LES, the normal resolution
will be similar to that given by Eq. 4.77 while for DNS the streamwise and normal resolutions
along the wall are on the order of the viscous sub-layer thickness, e.g.,
min min
x z ~ 5
+ +
.
In the special case of the resolved-surface treatment for particles, the geometric-based l
min

will be the particle size (d) since the flow over and around the particle must be discretized and
accurately simulated. Moreover, a thin boundary layer of thickness may also form over the
particle so that the criteria of Eq. 4.76 or 4.79 should be applied for the normal direction.


Number of Continuous-Phase Fluid Cells
Once the relevant minimum spatial resolution and domain length scales are determined,
their ratio can be used to estimate the total computational resolution in terms of N
f
. If uniform
192
resolution is used throughout a simple Cartesian domain (i.e. x=x
min
, etc.), then the number
of nodes can be approximated as the number of cells which scale with the domain lengths:

x y z
f fx fy fz
D D D
N N N N
x y z


4.80
A similar equation for two-dimensional flows yields a quadratic dependence. However, grid
stretching if often used to reduce the number of grid points (and thus the required
computational resources) so that Eq. 4.80 can be considered an upper bound. If this stretching
is uniform then the grid size for an individual element can be described (Eq. B.30) as
( )
i 1
i o
x x 1

= + 4.81
Overall grid stretching is typically limited to be less than 20% from cell to cell (-0.2<

<0.2)
since larger changes tend to degrade spatial accuracy (Eq. B.31). This magnitude is reasonable
for both continuous mesh variations used in structured grids or layered levels of h-refinement
or p-refinement used in unstructured meshes (B.1). The total number of cells in the x-
direction (N
f,x
) can then be related to the macroscopic domain length in the x-direction:

( )
( )
fx
fx
N
N
i 1
x
i 1
o
1 1
D
1
x

+
= + =

4.82
This estimate can be extrapolated to a 3-D domain for the total number of cells with reference
to the minimum grid resolution if, and only if, stretching is used in all three direction as:

( ) ( ) ( )
( ) ( ) ( )
x x min y y min z z min
f
x y z
ln 1 D / x ln 1 D / y ln 1 D / z
N
ln 1 ln 1 ln 1


+ + +

+ + +



4.83
Such a uniform grid stretching assumes that there is one region which requires the minimum
resolution and the grid is stretched in all the outward directions from this region, which is often
a good approximation for the case of a flow around a single particle treated via the resolved-
surface approach. Roughly, Eq. 4.83 is a lower bound on the total number of fluid cells
required, whereas Eq. 4.80 is an upper bound.
For many continuous-phase flow grids, the above estimations are reasonable (Loth, 2000).
However, spectral methods are much more efficient at capturing spatial scales since the
flowfield is solved in wave-space instead of physical-space. Therefore, Eq. 4.83 will
substantially over-estimate the number of fluid nodes for spectral methods, though these
methods are only appropriate for simple domain configurations where there are no sharp flow
gradients, e.g., incompressible turbulence in simple channel.
As noted above, determining the number of nodes for turbulent flows is generally more
complex because it is a function of the flow Reynolds number and turbulent scales. However,
the resources for internal flows are roughly related to the macroscopic Reynolds number (Re
D
).
As shown in Fig. 4.16, the number of grid cells increases slowly with Re
D
and can be roughly
approximated a priori as

0.2
f D
N 1000Re 2-D RANS wall-bounded requirements 4.84
193
This correlation is similar to that based on skin-friction variation which yields a dependency of
Re
D
1/4
(Loth, 2000). For external flows, the resolution is typically reduced since the boundary
layer thickness normalized by streamwise length (/D) generally decreases with Re
D
. Thus an
approximation of N
f,D
for external flows is more difficult since flow separation and geometry
can greatly affect the boundary layer thickness, but Eq. 4.84 is a reasonable upper bound.
For LES, the nodal resolution will be a function of the spatial-filter length-scale
G
defined
in Eq. A.132. This filter is generally related to the grid resolution, which can vary depending
on the range of macro to micro-scales of the flow and the available computational resources.
Piomelli

(1997) estimated the nodal resolution in each direction as Re
D
2/5
for free-shear flows,
yielding a 3-D dependency of Re
D
6/5
. A wall-bounded flow with a viscous sub-layer requires
additional resolution but this proportionality is expected to be consistent and indeed a rough
approximation for internal flows based on a survey of LES studies (Fig. 4.16) is:

1.2
f D
N 0.2Re LES wall-bounded requirements 4.85
Again, external turbulent flows are more difficult to estimate but Eq. 4.85 can be used as a
conservative guideline, i.e. nodal requirements may be expected to vary significantly for an
equivalent Re
D
(Dorgan & Loth, 2005). An approach which is intermediate to LES and RANS
(in terms of both resolution and computational resources) is the Hybrid RANS/LES approach.
However, the number of nodes for a hybrid approach is difficult to estimate in advance, but
some rough estimates have been put forth by Loth (2000).
For DNS in free-shear regions, /~Re

3/4
as noted in B.3. If we assume that D/u
D
is
linearly related to /u

, then D/~Re
D
3/4
such that the total number of cells or nodes required
will scale with Re
D
9/4
. For example, consider a 1 cm diameter water jet flowing at 1 m/s (a
relatively modest flow field in terms of size and speed). This gives rise to a turbulent flow
with Re
D
of 10
4
, which corresponds to D/ of 10
3
. The total number of nodes in 3-D will be
on the order of 10
9
. While grid stretching and higher-order elements can reduce the number of
nodes required, computational resources limit DNS to only modest flow Reynolds numbers.
DNS for wall-bounded flows (which resolve the viscous sub-layer) is even more restrictive,
e.g., Re
D
7/2
(Piomelli, 1997). As shown in Fig. 4.16, this proportionality is reasonable for
internal flow simulations with simple geometries so that a rough estimate for turbulent
conditions can be given by:
( )
3.5
f D
N 0.002Re DNS wall-bounded requirements 4.86
Since Eq. 4.86 assumes orthogonal shape functions, it will substantially underestimate the
number of nodes if there are complexities in domain geometry. In summary, the choice of
RANS, LES or DNS should be a balance between available computational resources (and time)
with the level of robustness desired in the continuous-phase flow solution (for a given flow
Reynolds number).
Generally, the inclusion of two-way coupling for point-force or distributed-force
approaches will not significantly increase the total number of computational points required.
However, if a resolved-surface method is used, then fluid cells are needed to resolve the flow
around the particle surface. If the particle boundary layer is thin, roughly 10,000 fluid cells
may be needed around an individual solid particle so that
194

4
f f , 0 p
N N 10 N
=
+ Lagrangian resolved-surface requirements
4.87
This is a first-order estimate as the actual number of needed nodes or cells may increase for
complex particle shapes, high particle Reynolds numbers, and whether any interior nodes are
needed for a fluid particle.

4.3.2. Number of Cells or Locations for Dispersed-Phase Methods

Next, the dispersed-phase computational requirements are considered in terms of the
number of locations where particle characteristics (velocity, etc) are to be simulated. The most
common Eulerian and Lagrangian approaches are generally based on the number of particles in
the simulation, as shown in Fig. 4.17. A more detailed classification of approaches is given in
Fig. 4.18 based on Stokes number, particle size, and relative number of fluid nodes to the
number of particles (or parcels).
The point-force techniques have particle diameters less than the grid length scale and can
be distinguished by Eulerian or Lagrangian approaches to describe the particle motion. For the
Eulerian approach, the dispersed-phase unknowns are generally computed with the same
resolution as that of the continuous-phase, i.e.
N
d
= N
f
Eulerian point-force requirements 4.88
Note that this result is independent of the total number of particles (N
p
) in the domain. As such,
Eulerian methods are generally applied for conditions when N
p
N
f
as this affords significant
computational efficiency per particle.
If the physics suggest a Lagrangian approach (e.g., to capture particle-wall effects or
turbulent diffusion) and there are a large number of particles (N
p
N
f
), then a Lagrangian parcel
approach is appropriate. For example, a typical combustion spray may include 10
12
drops but
the combustion chamber may be modeled with only 10
6
fluid cells. If two-way coupling is
important, then one must ensure sufficient parcel resolution (Eq. 4.74). If the particle sizes are
uniform, generally four or more parcels per Eulerian cell are reasonable so that
N
d
= N
P
~ 4N
f
Lagrangian parcel point-force requirements 4.89
The number of parcels that are needed will increase if polydisperse size distributions and/or
unstructured grids are to be considered. If the total number of particles in the domain
decreases such that it is on the order of the number of fluid nodes or less, a conventional
Lagrangian point-force approach for individual particles should be used. In this case, the
number of particle nodes (parcels) is simply equal to the number of particle trajectories at each
time-step within the domain
N
d
= N
p


Lagrangian point-force requirements 4.90
If the particle diameters become on the order of the continuous-phase grid size, then a
Lagrangian distributed-force approach is needed. In this case the number of particle nodes is
again given by Eq. 4.90 but each particle then interacts with a number of fluid nodes. If the
number of particles is reduced down to a relatively small number (about 1-100), then a
resolved-surface simulation may be appropriate. For a Lagrangian interface approach, each
particle trajectory can be tracked based on surface integrated forces so that
195
N
d
= N
p


Lagrangian resolved-surface requirements 4.91
Since there are only a few particles, this is not a significant computational cost. However, the
number of fluid cells will substantially increase beyond the uncoupled resolution (Eq. 4.87).
In summary, computational techniques for particles, drops, and bubbles become more
computationally intensive as the level of physics to be resolved increases. Similarly the
required computational resources increase as the number and/or size of the particles increase.
Thus, individual access to computing memory and time, turnaround constraints, and the desired
prediction characteristics should all be carefully considered. Use of the concepts illustrated by
Figs. 4.14 and 4.17 allows engineers to select numerical approaches in their most reasonably
reduced form (to minimize computational effort) while ensuring a sensible solution (providing
sufficient accuracy of physics). Effectively, this is the CFD equivalent of Einsteins comment:
Make everything as simple as possible, but not simpler. Fortunately, the continuing increase
of computing power will allow increased resolution and consequently improved representation
of the fluid physics and/or the capability to handle more complex systems.
In the next four chapters, different multiphase numerical techniques will be discussed in
more detail, starting from approaches appropriate to the smallest particles and ending with
approaches appropriate to the largest particles (in line with the sequence given in Fig. 4.17).


4.4. Problems

4.1) Discuss the primary reasons for choosing a particular technique among a specific set of
techniques for the following sets:
a) Lagrangian particle vs. Lagrangian parcel approaches,
b) Lagrangian parcel vs. Eulerian bin approaches for polydisperse particles,
c) mixed-fluid vs. weakly-separated approaches vs. separated-fluid, and
d) point-force vs. distributed force vs. resolved-surface approaches.
4.2) Describe the two-way coupled continuous-phase momentum transport of Eq. 4.31
starting from the Reynolds Transport Theorem of Eq. A.1.
4.3) a) Describe Eq. 4.57. b) Develop a guide like Eqs. 4.55 or 4.56 for sand particles in
water with
p
=2.5
f
.
4.4) Starting from the separated-fluid Eulerian dispersed-phase and continuous-phase
momentum equations (Eqs. 4.29b and 4.31), derive the mixture momentum equation (Eq.
4.26b) and explain any assumptions needed.
4.5) a) Select a dispersed multiphase flow problem and discuss the flow aspects which you
would like to predict. b) Identify a typical flow Reynolds numbers and the relevant
particle Stokes number. c) Describe the numerical approaches and transport equations
for each phase that would be appropriate and explain your reasoning. c) Estimate the
number of continuous-fluid nodes and particle nodes which may be needed.
196
5. Mixed-Fluid and Weakly-Separated-Fluid Approaches

The mixed-fluid and weakly-separated-fluid techniques assume that the particle response
time is relatively small compared to the resolved fluid time scales. To their advantage, these
two techniques are computationally efficient in that they require only a modest increase in
computational resources beyond that of an uncoupled single-phase flow simulation. Both
techniques typically take advantage of mixed-fluid properties in terms of density, viscosity,
speed of sound, etc. which is described in the first part of this chapter. The rest of the chapter
then deals with mixed-fluid approximation for the transport equations, whereby the relative
velocity is assumed to be negligible. This is followed by a discussion on the weakly-separated-
fluid approach whereby a finite relative velocity is included based on terminal velocity and
local conditions. For both the mixed-fluid and weakly-separated-fluid approach, unsteady and
time-averaged formulations are considered and brief comments are made regarding the
associated numerical aspects.


5.1. Mixed-Fluid Properties
5.1.1. Mixed-Fluid Density and Mass Fractions

The mixed-fluid approach assumes that the particles and the continuous-phase fluid are in
local kinetic and thermal equilibrium, i.e., the relative velocities and temperatures are not
significant (Eq. 4.24) and only cell-averaged quantities of density, momentum and energy are
transported (Eq. 4.26). These assumptions are often also used to define a colloidal
suspension which refers to suspended small particles (with respect to the time-scale of the
overall fluid dynamics). The mixed-fluid formulation allows substantial computational
simplicity since neither particle size nor terminal velocity plays a role in the computations (and
do not need to be specified). The method also automatically allows for two-way coupling and
does not require the flow be dilute.
With the mixed-fluid approach, the volume fraction () can be used to describe the
concentration of the dispersed-phase within a grid cell. The multiphase mixture is assumed to
be represented by a single mixed-fluid density (
m
), velocity (v
m
) and temperature (T
m
). This
is reasonable if the path-averaged magnitude of the particle relative velocity is small compared
to the relevant path-averaged convective velocity. A rigorous constraint for the mixed-fluid
approach can be based on both the mean and fluctuating components of the relative velocity:


rms
w u < for mixed-fluid approximation
5.1
If this constraint is satisfied, the particles will effectively follow the continuous-phase path-
lines as noted by Eq. 4.24a, creating an immiscible mixture.
The above criterion generally requires that each particle terminal velocity and response
time is small compared to the key continuous-phase velocity and time-scales. If we assume
that the fluid processes are governed by a general wave-length (l), the mixed-fluid criteria are:
197


term
w
1 St 1
u
< < and
l
l
for mixed-fluid approximation
5.2
In turbulent flow, the integral-scale properties can be important to mean particle diffusion or
overall turbulence modulation. However, wavelengths down to the micro-scales will become
important if aspects such as near-wall deposition or preferential bias are important. Thus,
implementation of the mixed-fluid approximation criteria depends on the wave-length of
interest, e.g.


term D D
w u St 1 < < and for mixed-fluid at macroscopic scales
term
w u St 1

< < and for mixed-fluid at turbulent integral-scales
term
w u St 1

< < and for mixed-fluid at all turbulent scales
5.3a

5.3b

5.3c
As an example, consider the Snyder & Lumley (1971) turbulent wake flow of Fig. 3.21 which
includes a continuous-phase (air) mean velocity of 6.55 m/s with u

~ 0.13 m/s,

~ 0.06 sec,
u

~ 0.03 m/s,

~ 0.003 sec (A.5.3 and A.5.4). For terminal velocity conditions, the
Kolmogorov-scale criteria for spherical particles with
p
=1000 kg/m
3
can be satisfied with
particle diameters as large 10 m (
term
w /u St 0.1

) while the integral-scale criteria would
allow diameters as large as 20 m (
term
w /u 0.1 and St 0.02

). In the latter case, the
velocity ratio is more restrictive which is consistent with Fr

<1 for this flow (Eq. 3.25).


The most important mixed-fluid property is density. The cell-averaged mixed-fluid density
(Eq. 4.25) is simply a function of the volume fraction and density ratio of the individual phase
components


( )
m p f
1 = +
5.4
If there are multiple particle fields and associated volume fractions, then the density can be
defined as


( )
n n
m p,i i f i
i 1 i 1
1
= =

+




5.5
The mixture density can also be used to define the mass-fractions (Eq. 3.145). If there is only
one type of particle, the mass fractions for each phase are given by


p p
p
p f m
f f
f
p f m
X
(1 )
(1 ) (1 )
X
(1 )

=
+

=
+

5.6a


5.6b
Thus, the mass factions are everywhere conserved so that Eq. 5.6b can be replaced by


p f
X X 1 + =
5.7
Similar to Eq. 5.5, this can also be extended to include multiphase particle phase groups.
If heat transfer effects are included (e.g., over a boundary surface), a mixed-fluid thermal
conductivity can be employed for the mixed-fluid volume. Since the particles and continuous-
phase are heated simultaneously (Eq. 4.24b) and this transfer is volume-based, a first-order
198
approximation for mixed-fluid thermal conductivity can be given in terms of the phase volume
fractions (Crowe et al. 1998) :


m p f
(1 ) = + k k k
5.8
If the particle conductivity is greater than that of the surrounding fluid, the mixed-fluid has an
increase net conductivity. The specific heat of the mixture is based on the mass-fractions and
can be obtained by considering an energy balance with constant temperature (T
m
):

p,m m f p, p f p,
(1 ) ( ) = + +
g s g s
c c c c c
5.9
In the following, the viscosity and compressibility properties are discussed for various
multiphase mixtures based on the mixed-fluid assumption.

5.1.2. Mixed-Fluid Viscosity

The mixed-fluid viscosity (or effective viscosity) is the ratio of the bulk stress to the bulk
rate of strain for the mixture averaged over a volume that includes many particles. Often the
mixed-fluid viscosity is normalized by the continuous-phase viscosity to define an effective
dimensionless viscosity:


* m
m
f


5.10
In the following, three-way coupling effects on the mixed-fluid viscosity is considered for both
solid and fluid particles, first without particle-particle interactions and then with Brownian-
induced particle-particle interaction.


Solid Particles in a Fluid without Particle-Particle Interactions

The Newtonian mixed-fluid viscosity for solid spherical particles was derived by Einstein
in 1906 for the dilute flow limit (no inter-particle effects) with creeping flow conditions. This
bulk viscosity for immiscible mixtures can be obtained by considering an isolated particle in a
straining field and assuming the effects can be summed linearly for many particles. It is
interesting that the initial derivation by Einstein contained an error in which he later found and
corrected in 1911 when comparison with experimental data showed a quantitative deviation.
Details of a similar derivation of the effective viscosity are given by Batchelor (1972) whereas
a simple overview is given below.
For a single particle, a pure strain is imposed in the far-field and the particle is located such
that extrapolated velocity to the particle center of mass (e.g., centroid for uniform density
particle) is zero. The strain is assumed to be very weak such that velocity gradients non-
dimensionalized by density, viscosity and particle diameter are small. Further assuming a
creeping flow approximation allows the inertial terms in the Navier-Stokes equations to be
neglected (Eq. A.63). Finally, a no-slip boundary condition is assumed (Eq. 1.27). From the
governing equation and the boundary conditions, the continuous-phase velocity can be
obtained analytically in the region extending from the particle surface to the far-field. From
this, the stress tensor in the continuous-phase (Eq. A.9) can be determined at any point. This
199
can then be used to determine the additional energy dissipation owing to the distortion of the
velocity field to satisfy the boundary conditions on the particle surface. By linearly summing
this increased dissipation for several particles within a dispersed-phase continuum for small ,
the Einstein viscosity correction is expressed to first-order as:


( )
* 2
m
1 2.5 = + + O
5.11
While this result assumes that the particles are small (consistent with negligible relative
velocity of the particles in the mixed-fluid approach), it does not assume that they are
monodisperse. Thus, a volume fraction of 5% for a polydisperse distribution (which has an
inter-particle gap of about 1.2 diameters by Eq. 3.186) corresponds to a 12.5% increase in
viscosity. This result neglects particle-particle interactions and so is only expected to be
reasonable for 1. Experiments of sedimenting spheres at low Reynolds numbers have found
that this viscosity ratio to be reasonably for small volume concentrations, e.g., of 10% or less
as shown in Fig. 5.1a. Increased sensitivity has been found for non-spherical particle. For
example, the mixed-viscosity for oblate ellipsoids with aspect ratio E (<1) can be approximated
(van der Kooij et al. 2001) as


( )
* 2
m
2 1
1 2.5 E 2
3 E


+ + +



+ O
5.12
Similar results are found for long slender ellipsoids (E>1) by Harding et al. (1995). Thus, this
approximation is reasonable for aspect ratios (E) ranging 0.25 to 4, and indicates maximum
increases in viscosity of about 20% for =5%.
Rheology measurements have been conducted between parallel plates with a small gap and
have generally confirmed the above results for modest volume fractions, about <10% as
shown in Fig. 5.1a. However, gravitational settling can cause concentration non-uniformities,
e.g., heavier particles tend to the lower surface. Since the settling can be counteracted with
viscous re-suspension at high shear rates, the particle concentration field and thus the mixed-
fluid viscosity can be a function of the shear rate, i.e. the mixed-fluid viscosity is effectively
non-Newtonian (Drew & Passman, 1999). Acrivos et al. (1993) noted that this effect is
primarily a function of the ratio of viscous to buoyancy forces which can be quantified in terms
of the Shields number of the mixture (S), which for flow perpendicular to gravity and a channel
height of D is given by:


9
f shear p f
2
/ Dg = S
5.13
If the viscous re-suspension effects are dominant (S1), the concentration gradient effect will
be negligible. In this case, a mixed-fluid viscosity is reasonable and is independent of shear
rate. However, at Shields number of order unity or less, the gravitational effect will be
substantial, indicating that the mixed-fluid approach may not be reasonable.


Fluid Particles in a Fluid without Particle-Particle Interactions

One may also apply a similar creeping-flow analysis to fluid particles (Batchelor, 1970) by
assuming an uncontaminated surface such that the particle surface boundary conditions are
200
based on a tangential stress balance which allows finite tangential slip (Eq. 1.58b). In a
manner similar to that derived for the drag on fluid particles in creeping flow (1.5.2), the first-
order mixed-fluid viscosity (for nearly dilute flow) can be obtained as


*
p *
m *
p
2.5 1
1
1
+
= +


+


5.14
This is the result used in Eq. 3.187 to estimate three-way coupling effects. As the particle
viscosity approaches infinity (or the surface becomes fully-contaminated), the viscosity ratio
reverts to the solid-particle case of Eq. 5.11. However, when the particle viscosity is negligible
compared to the continuous-phase (as in the case of uncontaminated gas bubbles in a liquid)
then the mixed-fluid viscosity ratio (for nearly dilute flow) simply becomes


*
m
1 = + 5.15
Thus a volume fraction of 5% corresponds to only a 5% increase in viscosity for such bubbly
mixtures. This reduced impact of volume fraction is consistent with the reduced stress and
dissipation which occurs for mobile interface conditions as compared to no-slip conditions.
It is important to note that the above discussion on viscosity assumes that the two-phases
are immiscible so that individual particle volumes and shapes may be defined at any time. This
is in contrast to the miscible condition where the phases are fluids mixed at the molecular scale.
In the miscible case of two fluids, the viscosity is based on the volume fraction of its
components, i.e. ( )
* *
m p
1 + . For highly deformable liquid particles in a liquid
mixture, there may be a tendency towards this limit but generally smaller particles have very
high relative surface tension so that Eq. 5.15 is more appropriate.


Solid Particles in a Fluid with Particle-Particle Interactions

Particle-particle interactions can occur at higher volume fractions and influence the
effective viscosity of the mixed-fluid. These interactions may include chemical and electrical
mechanisms at dense conditions, but primarily include Brownian motion and particle-particle
collision dynamics as discussed herein (van der Werff and deKruif, 1989). To assess the
influence of Brownian motion, one may define a Brownian frequency (f
Br
) based on the
isolated Brownian diffusivity (given by Eq. 3.139) normalized by the particle diameter:


Br f
Br 2 3
f
4 4kT
d 3 d

=

f
5.16
The mixed-fluid fluid shear rate can be defined in a manner similar to Eq. A.34 as:


m,x
m,shear
v
y


5.17
This may be normalized by Eq. 5.16 to define a shear frequency ratio:


3
f m,shear * shear
shear
Br f
3 d
4kT

=
f
f
f
5.18
201
This ratio is proportional to the particle Peclet number and compares the importance of viscous
diffusion to Brownian diffusion for a shear flow. The two extremes of this ratio are defined as
the low-shear or low-frequency limit (
*
shear
1 < f ) for which the associated viscosity is defined
as
*
m,0
, and the high-shear or high-frequency limit (
*
shear
1 > f ) for which the associated
viscosity is defined as
*
m,
. Data from these two limits are shown in Fig. 5.1 as a function of
particle volume fraction. The low-shear limit (
*
m,0
) represents a long-time equilibrium
conditions where Brownian diffusion dominates while the high-shear limit (
*
m,
) represents a
short-time dynamic conditions where Brownian motion is negligible (Verberg et al. 1997).
As shown in Fig. 5.1, both limits are reasonably represented by the Einstein viscosity (Eq.
5.7) for small volume fractions, even though this theoretical result was obtained for the high-
shear limit. In fact, the high-shear limit is reasonably represented by this viscosity for volume
fractions as much as 20% as shown in Fig. 5.1a. At higher volume fractions, the particle-
particle surface interactions become important which can be related to the fraction of particle
pairs which are in contact (G), i.e. the fraction of particles that are colliding, exchanging
momentum and thus contributing to the dissipation. Assuming isotropy and hard spheres, this
fraction is equal to the equilibrium radial distribution function which can be represented by the
Carnahan-Starling approximation given as:


( )
3
1
2
1

G = 5.19
The increase in mixed-fluid viscosity is expected to depend on this fraction and, in fact, this
fraction can be used to reasonably represent the high-shear limit (Verberg et al. 1997), i.e.


*
m,
G
5.20
As shown in Fig 5.1a, the Carnahan-Starling approximation is quite reasonable for volume
fractions as high as 50%.
Brady (1993) derived an expression for the low-shear viscosity which includes a higher-
order correction associated with the increase in viscosity due to Brownian stresses


( )
2
*
m,o
12
5

+
G
G
5.21
A similar expression obtained by Verberg et al. (1997) gives nearly identical results for
<40%. There are also several empirical expressions for the low-shear limit of viscosity such
as the Eiler fit used by Leighton & Acrivos (1987):


2
*
m
max
1.5
1
1 /

+



5.22
In this equation,
max
is the maximum volume fraction, also referred to as the packing limit, for
which values ranging from 0.58 to 0.68 are reported in the literature depending on shear rate,
packing organization, and particle charge (van der Werff & de Kruif, 1989; Stickel and Powell,
2005). Variations on the Eiler fit are given by Chapman & Leighton (1991) with a numerator
202
of 1.59 (vs. 1.5), and by Stickel and Powell (2005) with a numerator of 1.25(which allow the
expression to correctly approach the Einstein viscosity in the dilute limit).
The above results indicate that the equilibrium mixed-fluid viscosity is always less than the
dynamic mixed-fluid viscosity (i.e.
* *
m,o m,
). This can be attributed to the increased
mixing (mobilization of the particle) caused by Brownian motion at long time-scales and is
consistent with the experimental differences between Figs. 5.1a and 5.1b. As a consequence,
subjecting a multiphase suspension to increasing shear-rates in time can result in a viscosity
consistently reducing until it reaches
*
m,
, a phenomenon often described as shear-thinning.
Models to determine the effective viscosity for intermediate frequencies with substantial
volume fractions are given by Verberg et al. (1997) and Brady (1993). However, this
complication can generally be avoided for dispersed conditions (<10%) since the Einstein
expression (which is independent of the frequency) is quite reasonable (Figs. 5.1a and 5.1b).

5.1.3. Mixed-Fluid Compressibility

Relationships between pressure and density may be used to account for mixed-fluid
compressibility aspects. These relationships depend on whether the continuous phase is a gas
or a liquid, and each of these cases is discussed in the following two sub-sections, along with
the effective speed of sound of the mixture. The latter is important to determine whether the
flow speed can induce compressibility (analogous to the single-phase behavior of Fig. A.4).


Solid or Liquid Particles in a Gas

The mixed-fluid compressibility relationships for solid or liquid particles in a gas ( * 1)
typically make use of the dusty gas assumption (Marble, 1970). This assumption indicates
that the net particle volume is negligible (1) but the particle mass can be significant
compared to that of the gas. As such, bulk properties associated with flow compressibility are
conventionally described as a function of the mass loading (), which represents the ratio of
dispersed-phase mass to continuous-fluid mass within a mixed-fluid volume (Eq. 3.144). Thus,
the mixed-fluid density can be approximated as
( )
m f
1 +
5.23
Assuming that the mixed-fluid obeys an ideal equation of state, the unhindered pressure is

m m m m
p p T = = R dusty-gas equation of state
5.24
Based on the mixed-fluid assumption, we note that thermal equilibrium stipulates T
m
=T
g
(=T
p
).
Based on the negligible particle volume fraction assumption, the mixed-fluid pressure is
defined as that external to the particles and thus equal to the gas pressure, i.e. p
m
=p
g
(Wallis,
1969). In Eq. 5.24, the mass loading affects the density but not the pressure or temperature so
that an ideal equation of state requires:


( )
m
/ 1 = +
g
R R
5.25
Therefore, the particles decrease the gas constant, which is equivalent to an increased
203
molecular weight for a dusty gas based on Eq. A.21.
It is often of interest to determine the speed of sound for a particular flow to note the
degree of compressibility one might expect, since this may affect the manner in which
boundary conditions are employed (A.2) and numerical methods are chosen (B.3.2). To
determine the sound speed for the dusty-gas assumption, we must first obtain the specific heats
of the mixture. If the dispersed-phase is a solid, the specific heat is approximately independent
of pressure, i.e.
p

s s, s,
c c c . A similar approximation can be made if the particle is instead a
liquid. If the particle specific heat is normalized by the gas specific heat as
*
s
c , it can be used to
define the mixed-fluid specific heat based on Eq. 5.9:

( )
*
p,
*
f p, p f p,
p,m p,
m f
/
(1 ) ( )
1
1 1

+ + +
= =

+ +

s s g
g s g s
s
g
c c c
c c c c
c
c c

5.26a

5.26b
The mixed-fluid specific heat at constant volume
,m
( )

c can be similarly obtained. In either


case, the mixed-fluid specific heat will increase or decrease depending on whether the ratio of
Eq. 5.26a is greater or less than unity. For most solid and liquid particles in a gas,
*
s
c >1 is
greater than one so that
p,m
c increases with particle loading. However,
*
s
c can be less than
unity if the particles are composed of a metal or if the gas has a low molecular-weight, e.g.,
hydrogen or helium.
The particles can also affect the specific heat ratio. Assuming that the overall mixture is
calorically perfect (see A.1), the "mixed" specific heat ratio can be obtained from Eq. 5.9 as
(Rudinger, 1964)


*
p, p,m
m *
,m ,
1
1

+
+
= =

+ +

g s
s
g
g s g s
c c c
c
c c c c

5.27
For finite particle loadings, this yields 1
m

g
such that the dusty gas approximation results
in a reduced compressibility exponent. This result is expected since the particle-phase is
incompressible (i.e. has an intrinsic compressibility exponent of unity). From Eqs. 5.25 and
5.27, the acoustic speed can then be obtained as:


( )( )
* 2
m m m
2
*
1 a
a
1 1
+
= =

+ +
s
g g g
g s
c R
R
c

5.28
As such, the mixed-fluid sound speed will be reduced as or
*
s
c increases (Fig. 5.2a).
The mixed-fluid approximation can break down at high-speeds based on the criterion of
Eq. 5.1 since particle dynamics (and thus relative velocities) can be important. For example,
consider air at STP flowing though a 0.1 meter diameter nozzle at nearly sonic conditions so
that the macroscopic fluid time-scale (
D
) is about 0.3 msec. For a particle to move
approximately in equilibrium with this flow, it should have a much smaller
p
(e.g., less than
0.1 msec), which for a particle density of 1000 kg/m
3
corresponds to a diameter of less than 6
microns. However, if the fluid domain time-scales are based on much larger scales such as the
volcanic ash expulsion flow (Fig. 1.15), then particles on the order of 100-200 microns in
diameter may be reasonably considered to be in equilibrium.

204

Bubbles in a Liquid

The compressibility of a bubbly mixture in a liquid can be studied analytically using the
mixed-fluid assumptions. The relationship between density and pressure which determines the
mixture equation of state is affected by the compressibility of the particle phase but also by
frequency of the flow (f
l
) relative to the natural frequency of the bubble size oscillation (f
nat

from Eq. 2.66). For a low-frequency formulation (f
l
f
nat
), the bubble size and volume can be
considered to be in quasi-steady equilibrium with the local pressure of the surrounding fluid
and the surface tension. This allows us to relate the internal bubble pressure (P
g
) to the
continuous-phase unhindered liquid pressure (p) from Eq. 1.7 or Eq. 2.62, i.e.:


b
P p 2 / r = +
g

5.29
If one assumes that the particle density is much less than that of the liquid, the mixed-fluid
density can be approximated as:
( ) ( )
m p f
1 1 = +
l

5.30
Next, one may define a reference mixed-fluid volume is
mo
, where the subscript o refers to
a reference condition. As the pressure changes, the volume will be modified but the mass will
be preserved such that

m m mo mo
=
5.31
If the liquid is incompressible, its associated volume will be constant and equal to the reference
value, i.e.

f m o mo
(1 ) (1 ) = =
5.32
The bubble compressibility can be given by the polytropic exponent K
g
(Eq. 2.63) so that the
particle volumes within the mixed-fluid (
p
) can be related to the local pressure. The mixed-
fluid volume can then be related to the reference conditions:


1/
o p,o
m f p o mo o mo
p
p 2 / r
(1 )
p 2 / r
+
+ = +


+

g
K
5.33
If the bubble radius is large enough (e.g., greater than 10 m for bubbles in water), then the
pressure difference due to surface tension can be ignored and a simple relationship between
mixed-fluid density and pressure can be obtained (Peregrine & Thais, 1996) by dividing Eq.
5.33 by Eq. 5.31:


1/
o o o
m mo mo
1 p 1
p

= +



g
K
5.34
Solving for the pressure yields


( )
m o
o mo m o
p
p 1


=




g
K
bubbly-flow equation of state 5.35
Therefore the pressure can be related directly to the reference conditions and the mixed-fluid
density, which is convenient for numerical formulations as an equation of state. This change in
pressure can also be expressed in terms of the change of volume fraction as:
205


( ) ( )
( )
o o
o
o o
1
p
p 1
+

=




g
K
5.36
The RHS approximation assumes isothermal conditions and small volume fractions. The result
indicates that a pressure increase leads to a commensurate linear decrease in the volume
fraction as the gas-phase portion is compressed. If the liquid is also compressible, then its
pressure-density relationship (e.g., Eq. A.24) must be incorporated in the above analysis to
obtain the mixed-fluid equation of state.
The speed of sound of a multiphase mixture can also be related to the speeds of sound for
each phase by considering perturbations in the volume fractions (Brennen, 2005). Defining a
volume perturbation using the notation
o
+ and normalizing this by the reference
mixed-fluid volume yields the following volume ratios for continuous-phase, particle-phase
and mixed-fluid:


( )
( )
fo f fo o f f
f
mo mo mo fo f
fo
p po p p po o
p
mo mo mo po p
po p
p
f p
mo m m
mo mo mo
(1 ) m
p
p
m
p
p
1
+
= = =

+
+

+
= = =

+
+

+
+
= = +


5.37a



5.37b


5.37c
Combining these equations and neglecting second-order terms, e.g., 1/(1+)1- , one can
obtain the normalized change in the mixed-fluid volume in terms of the surrounding fluid
pressure change (p) as:


o o m
mo o o
P
1
p p
p p P

=





g g g
g g g

5.38
The reciprocal of
p
p / p for a bubble can be obtained in terms of surface tension (Eq. 5.29):


( )
1/3
1/3
p p p
p 1 4 2 1
1 2 1 2 1
P P r 3m P 3r P


= = =



g
g
g g g g g
5.39
The LHS of Eq. 5.38 normalized by the pressure change can be written for constant mass as:


mo m m m m
mo 2
mo m mo
(1/ ) 1 1
p p p p

= =




5.40
If one assumes adiabatic isentropic processes, Eq. A.20 indicates that the two terms in
parentheses on the RHS of Eq. 5.38 are related to the speed of sound for the liquid and the gas
while the term in parentheses on RHS of Eq. 5.40 is related to the speed of sound of the
mixture. Therefore, combining the above equations and dropping the o subscript yields:


m m m
m 2 2 2 2
m m b b b
1 1
a p p 1 2 / 3 a r a a


= = +





g g l l
s s

5.41
206
This result was extended by Brennen (2005) to include effects of mass transfer in the form of
evaporation and condensation.
The above equation can be simplified by neglecting surface tension, which is reasonable if
2
p
a r <
g g
and consistent with d>10 m. This assumption leads to the well-known Minnaert
(1933) equation:


2 2 2
m m
1 1
a a a

= +

g g l l
adiabatic bubbly-flow speed of sound
5.42
This result is equivalent to stating that the mixture acoustic impedance (LHS) is equal to the
volumetric average of the acoustic impedance for the particles and the surrounding fluid.
[Interestingly, the Minnaert relationship combined with the dusty-gas assumptions of low
volume fraction and very-heavy particles leads to ( )
2 2
m
a a / 1 = +
g
, which is the same result
given by Eq. 5.28 if
*
0
s
c .]
The above assumption of an isentropic process with no heat transfer allows the pressure-
density derivative for the gas-phase to be expressed as
2
P / a T = =
g g g g g
R . However, this
may be unreasonable if liquid temperature controls the bubble temperature so that it is
approximately isothermal while undergoing volume changes. This is especially true if there is
high heat transfer rates caused by significant water liquid vapor within the bubble or bubble
surface to volume ratios are high (smaller bubbles). In this case, the pressure-density
derivative should be expressed as P / T =
g g g g
R , so that one should substitute
2 2
a / for a
g g
in
Eqs. 5.41 and 5.42 to obtain isothermal relationships. Figure 5.3a shows a comparison of
experimental data of gas bubbles in water with the speed of sound predictions from adiabatic
and isothermal assumptions at zero frequency. In general, the isothermal approximation is
reasonable over a wide range of volume fractions, consistent with a mixed-fluid with infinite
heat transfer rates. It is interesting to note that intermediate volume fractions for adiabatic or
isothermal conditions lead to a sonic velocity which is substantially less than either of the
single-phase velocities! Because of this, choked flow of bubbly systems can occur at quite
modest speeds. As such, compressibility corrections should be considered whenever the
mixture Mach number is no longer small, e.g., if M
m
>0.2 (which can correspond to speeds as
low as 4 m/s in air/water mixtures).
If one further assumes liquid incompressibility to simply the process, the isothermal
mixture speed of sound of Eq. 5.42 becomes:


2
m
p
a
(1 )
=
+

l g

5.43
This reduced isothermal prediction yields results which are quite similar to the fully
compressible isothermal predictions as shown in Fig. 5.3a with differences of only about 1%
for >10
-3
. However, the reduced isothermal prediction is undefined (tends to infinity) for
0. This reduced prediction can also be linearized for small but finite void fraction to give
an even simpler expression:


2
m
a p /
l
5.44
This linearized isothermal prediction is not valid in either the very small or very large
volume fraction limits, but is reasonable for 0.001<<0.1 (Fig. 5.3a).
207
Now, let us consider bubbly flow subjected to frequencies that are no longer small
compared to the natural frequency of the bubble (f
nat
, Eq. 2.66). In this case, we need to include
the bubble size dynamics via the Rayleigh-Plesset equation of Eq. 2.64 instead of the quasi-
steady pressure balance given by Eq. 5.29. Brennen (2005) discusses a linearized approach to
determine the coherent portion of the wave speed as a function of the signal frequency (f).
This is based on the same assumptions used for Eq. 5.35 (bubbles with negligible mass in an
incompressible liquid) and yields a wave speed as a dispersion relation


( )
2 2 2
2 2
p,eq nat 2
m diss 2 2
nat nat eq eq
4 r
a 1 c
3 1

= +



f
-f f
f f

5.45
In this equation,
eq
and r
p,eq
refer to an equilibrium volume fraction and bubble radius about
which the cycle oscillations will occur, where the latter can be obtained by Eq. 5.29. Also,
included is a dissipation constant (c
diss
) which represents the effective damping due to viscous
effects, acoustic radiation, etc. If the only damping is due to viscosity (caused by the radial
motion of bubble interface), then the damping coefficient can be given as


diss 2
nat p,eq
2
c
r

l
f

5.46
Brennen (2005) discusses inclusion of other effects such as mass transfer.
The differences between Eq. 5.44 and Eq. 5.45 indicate significantly different physics
between the low-frequency regime and the high-frequency regime. First of all, there is no
damping for low-frequencies whereas the damping can be very high for the high-frequency
regime, especially as volume fraction increases. Because of this, most experimental studies on
attenuation have been limited to volume fraction below 1%. Secondly, the high-frequency
relation of Eq. 5.45 indicates that the wave propagation is dispersive which is caused by
scattering from the bubbles as they oscillate. This leads to an incoherent portion of the wave
speed if the bubble distribution is randomly distributed. A third issue is that the wave speed is
sensitive to the bubble diameter. This is indicated in Fig. 5.3b where the prediction of wave
speed as a function of frequency is much improved if the polydisperse distribution is taken into
account (as compared to a monodisperse prediction based on a mean bubble size). This figure
(for a small volume fraction of 0.02%) also indicates that while the speed of sound is reduced
compared to the single-phase speed for low frequencies, it can actually be increased for high-
frequencies. It should also be noted that c
diss
was empirically set to be 0.5 to match the
experimental results (which was much greater than the value given by Eq. 5.46) and the
attenuation is found to increase dramatically (by several orders of magnitude) for frequencies
close to the natural frequency (Drew and Passman, 1999; Brennen 2005). This suggests that
some form of the Rayleigh-Plesset equation with liquid compressibility (e.g. Eq. 2.73) should
be integrated for these conditions. If the wave speed is on the order of the gas speed of sound
(typical for shock fronts in dilute bubbly mixtures) then the internal gas pressure should be
obtained by solution of the compressible flow equations (Kameda & Matsumoto, 1996). For
an Eulerian treatment of mixed-fluid, the ODEs for the bubble dynamics and internal gas
dynamics can be applied on average over a cell volume.


Solid or Liquid Particles in a Compressible Liquid

208
The compressibility of a mixture of particles in a liquid can predicted by using Minnearts
impedance relationship for the speed of sound of a mixture (Eq. 5.42) by assuming both phases
are represented with a single velocity and a single pressure field and four-way coupling and
surface tension effects are negligible. Experimental results for water containing small solid
particles (about 1 micron in diameter) are shown in Fig. 5.4 where the agreement with
Minnearts impedance relationship can be seen to be quite reasonable (Brennen, 2005). If both
phases are incompressible (e.g., solid particles in a low-speed liquid), then no equation of state
is needed for the mixture.



5.2. Mixed-Fluid Methodologies
5.2.1. Mixed-Fluid Transport Equations

The Partial Differential Equations (PDEs) which govern mixed-fluid flow are based on
transport equations similar to that of single-phase flow, i.e. continuity, momentum and energy.
However, as additional transport equation for dispersed-phase mass transport is needed. These
PDEs are discussed below in an Eulerian framework, which is consistent with typical
numerical implementation.


Particle Mass and Mixed-Fluid Mass Transport

It should be noted that two conservation equations are generally needed for the mixed-fluid
methodology: one for mass for the dispersed fluid (which allows determination of the
distribution) and one for mass for the mixed-fluid (which allows determination of the
m

distribution). Note that each of these transport equations must be supplemented with proper
boundary and initial conditions for solution as well as the equation of state for the mixture
(discussed in the previous section). For the particle concentration, the transport equation of Eq.
4.73 can be obtained in the mixed-fluid limit as


( )
( ) ( )
p
p m p Br p
p
m
t


+ = +


v ` 5.47
This includes the rate of mass transfer per particle from the continuous-phase to the dispersed-
phase
p f
(m m ) = ` ` which may occur because of evaporation, condensation, solidification, etc.
This can be complimented with the continuous-phase equation given by Eq. 4.30. However, a
mixed-fluid mass transport (based on the linear combination of the particle-phase and
continuous-phase equations) is instead more commonly used:


( ) ( )
m
m m Br p
t


+ =

v 5.48
The RHS term can be neglected if the Brownian diffusion speeds of Eq. 3.140 are relatively
small, i.e.
Br,i m,i
w v < .
209
For dusty-gas flows with negligible Brownian motion, the transport equation for the mixed-
fluid density is simply


( )
m
m m
0
t

+ =

v 5.49
However, the transport equation is often written in terms of the mass loading since the volume
fraction is generally negligible. Assuming no mass transfer, the mixed-fluid version of the
continuous-phase flow (Eq. 4.30) becomes


( )
f
f m
0
t

+ =

v 5.50
Combining this PDE with the dusty-gas version of Eq. 5.47 and the approximation of Eq.
3.154b yields a transport equation for the mass loading (Collins et al. 1994):


f f m
0
t

+ =

v 5.51
Therefore, the mass conservation transport PDEs for dusty-gas are given by Eqs. 5.49 and 5.51.
If both phases are incompressible (
f
and
p
are constant), the transport equations for
mixture density and the volume fraction (Eqs. 5.47 and 5.48) will be simplified. If one further
assumes negligible effects of mass transfer and Brownian motion, then
m
/ t 0 D D so that the
mass conservation equations become:


m
m
0
t
0

+ =

=
v
v

5.52a

5.52b
This combination of a passive transport PDE for the volumetric fraction coupled with a
divergence-free mixture velocity field is commonly used for bubbly flows. The latter
incompressibility condition eliminates the need for an equation of state so that the pressure can
be obtained in a manner similar to the case of incompressible single-phase flow solvers, e.g.,
via a Poisson equation, as is be discussed at the end of this section.
For very low-speed flows with significant particle volume fraction, it may be important to
retain the Brownian effects with an incompressible version of Eq. 5.47 and further include
mass diffusion associated with fluid-shear effects. For the latter, Leighton & Acrivos (1987)
modeled the diffusion associated with spatial variations in mixed-fluid viscosity ratio (Eq.
5.10) and mixed-fluid shear rate (Eq. 5.17):


( ) ( )
2 2 2 *
m Br m,shear m m,shear m
c d c d ln
t


+ = + +

v 5.53
Compared to Eq. 5.47, there are two new terms on the RHS due to three-way coupling in
addition to that associated with Brownian diffusion. One of these is associated with gradients
in the shear rate and represents effects of hydrodynamic interactions between particles. The
other term is associated with gradients in viscosity which arise from particle concentration
gradients. These two terms include two empirical coefficients c

and c
m
which Subia et al.
(1998) respectively set as 0.1 and 0.15 based on comparisons with experiment at low shear
Reynolds numbers (Re

1). Anisotropic versions of these terms were found to give improved


agreement with experiments as discussed by Fang et al. (2002).

210

Mixed-Fluid Momentum and Energy Transport

It is helpful to define a mixed-fluid viscous shear stress similar to that used in Eq. A.9:


m, j m,i m,k
m,ij m ij
j i k
v v v
2
K
x x 3 x

+





5.54
Noting that the mixed-fluid density (Eq. 5.4) and the mixed-fluid viscosity (5.1.2) are both
functions of the volume fraction, the mixed-fluid momentum transport of Eqs. 4.26b and 4.27
can be written as


( )
( )
m m
m m m m m,ij
p K
t

+ = +

v
v v g 5.55
This approaches the single-phase momentum equation of Eq. A.6a and A.9 for small volume
fraction and is equivalent to the mixed-fluid PDE of Table 4.1 if flow divergence effects are
neglected (Eq. 5.52a). This form assumes negligible collisions with respect to the pressure so
that continuous-phase pressure is sufficient to describe the pressure stresses. As discussed in
3.8.2, neglecting particle-particle collisions is reasonable for small Stokes numbers, which is
consistent with the mixed-fluid assumption (Eq. 5.1). If combined with the continuity equation
of Eq. 5.49 (which neglects Brownian motion), the RHS can be written as
m m
/ t v D D ,
approaches the single-phase form given by Eq. A.6b for small volume fraction:
The conservation of mixed-fluid energy is generally formulated based on the characteristics
of the continuous-phase. If the continuous-phase is a compressible gas and both the potential
energy and phase change enthalpy are neglected, then it is numerically convenient to express
the mixed-fluid energy transport in conservative form. Using the total energy (as in Eq. A.13)
and a mixed-fluid velocity yields:


( )
( ) ( )
m m,tot
m m,tot m m m m,ij m m m
e
e K p ( T )
t


+ = +

v g v + v k
5.56
In the above equation, k
m
is the mixed-fluid thermal conductivity coefficient (Eq. 5.30), K
m,ij
is
the mixed-fluid viscous stress (Eq. 5.54), and e
m,tot
is the total energy of the mixture which
includes both internal energy and kinetic energy. The internal energy can be written in terms
of the mixed-fluid temperature using the mixed-fluid specific heat of Eq. 5.9:


1 2 2
1
2 m,tot m m p,m m m
2
e e v c T v + = +
5.57
If one employs the dusty gas approximation, the continuous-phase is assumed to be a
calorically perfect gas and the dispersed phase is a solid or liquid with negligible volume
fraction but much higher density, i.e. 1 < and
p f
> . In this case, use of Eq. 5.26b
allows the mixed-fluid total energy to be represented as


( ) ( )
1 * 2
m,tot p, m m
2
e 1 T / 1 v = + + +
g s
c c
5.58
On the other hand, if the initial and boundary conditions include no temperature variations (as
in many low-speed flows and bubbly flows), this yields isothermal conditions so that the
energy equation can be dropped entirely (Ahuja et al. 2000).


Numerical Issues
211

Because of their similarity, standard single-phase techniques can be used for the mixed-
fluid equations if their mathematical and resolution properties are similar. For example,
steady-state mixed-fluid PDEs are typically solved with the same convergence acceleration
techniques used for single-phase flow, e.g., the local time-stepping and multi-grid schemes of
B.3.2. However, the multi-phase PDEs include and additional transport equation for the
dispersed-phase mass transport. While this PDE can sometimes be treated like a variable
density mass continuity PDE, conditions with high spatial gradients in particle concentration
(due to concentration fronts) may require special techniques. In particular, non-linear and/or
higher-order schemes (mentioned in B.3.2) are commonly used to ensure that numerical
diffusion does not become problematic.
The compressibility of the phases will also be critical to the application of the numerical
techniques. For example, a dusty-gas variable formulation (Eqs. 5.49, 5.51, 5.55, 5.56 and
5.58) can be expressed in the conservative flux-based forms (Eqs. B.75 and B.76) and can be
solved with the explicit compressible schemes (B.3.4). This can be used in combination with
a dusty-gas equation of state (Eq. 5.24). On the other hand, an incompressible bubbly flow
formulation (Eqs. 5.52 and 5.55) can be solved with primitive variables using pressure-based
techniques (B.3.3). If an implicit three-step MAC technique (Eqs. B.70, B.71b and B.71c) is
used, Eqs. B.54b and 5.55 are transformed as:

( )
( )
( )
n
1 pred n n n 1
m,i m,i pred,i pred,i
2
m i
pred 2
m,i
m
2
i i
n 1 pred
m,i m,i
f i
1 p
v v t 3
x
v p
x t x
p
t
v v
x

= +

=

- f f

5.59a


5.59b


5.59c
Note that Eq. 5.59b is implicit and must be solved iteratively for the pressure, which may
include a hydrostatic pressure gradient (Eq. A.56). The assumption of a constant gas density
required for this formulation is reasonable if the pressure variations are weak. For example, a
water flow with vertical changes of 0.1m or less and velocity changes of 1 m/s or less will lead
to pressure changes of 1% or less (based on Eq. A.41). For larger pressure changes, one may
need to take into account compressibility of the bubbles using a conservative-variable
formulation and a bubbly flow equation of state (e.g., Eq. 5.29 or Eq. 5.35).

5.2.2. Time-Averaged Transport Equations

The above mixed-fluid approach can be directly used in turbulent flows by resolving all-
scales with an unsteady three-dimensional approach and application of the full unsteady
Navier-Stokes equations. However, time-averaged formulations are often employed to reduce
the numerical resources required for simulation. These formulations represent the turbulence
with mean integral-scale values (to model mean turbulent diffusion, etc.), so that only the
criteria of Eq. 5.3b are needed to employ a mixed-fluid approach. Such formulations below
212
can be also considered in an LES framework assuming the particle dynamics are weak, i.e.
St

1 (discussed in 4.2.2).
Since the mixed-fluid can have a variable density, the time-averaged equations are best
obtained with the density-weighting, i.e. using Favre-averaging (discussed in A.5.5). For
example, the Favre-averaging and decomposition for a mixed-fluid property are given by:


m m
m
m
q
q

5.60a


m m m
q q q = +
5.60b
The double prime superscript denotes the density-weighted perturbation. The averaging
properties of Eqs. A.111 and A.112 similarly apply. For the mixed-fluid density, one may also
combine Eqs. 5.4 and 5.60a as


( ) ( )
m m p m f m m m p f m
q q 1 q q 1 q

= + = = +


5.61
Therefore, Favre-averaging also applies to the volume-averaged phase densities, e.g.,


p m p m
q q = 5.62
In the following, the mixed-fluid properties are discussed followed by formulation of the
Favre-averaged mass, momentum and energy mixed-fluid equations in tensor notation.


Time-averaged Mixture Properties

The mean mixture density can be obtained from Eq. 5.4 as:


( )
m p f p f
1 = + + 5.63
The RHS is commonly simplified by neglecting the fluctuation terms. This is obviously
appropriate if the particle and fluid phase density are constant (e.g., solid particles in a liquid).
Even when turbulence causes significant phase density fluctuations, these fluctuation terms can
typically be neglected. For the example of a dusty-gas, the particle density is typically constant
(so
p
0 = ) while the turbulence-induced variations of the surrounding gas density are
negligibly correlated with those of particle volume fraction (i.e.
f f
< ). In the other
extreme of a bubbly flow, the surrounding liquid density can be considered constant (so
f
0 = ) while any turbulence-induced changes in the bubble density will be negligible
compared because
p f
(1 ) < . Thus, the mean mixed-fluid density is approximately:


( )
m p f
1 +
5.64
As a result, the instantaneous variation of the mixed-fluid density is given as:


( )
m m m p f

5.65
Thus, volume fraction fluctuations affect the instantaneous mixed-fluid density but not the
mean mixed-fluid density.
213
Using similar arguments, the mixed-fluid viscosity can be similarly expressed in terms of
mean properties for dispersed-phase conditions. For example, using the Einstein viscosity
correction of Eq. 5.11 yields:


( )
m m m m m m f
1 2.5 = = + 5.66
Other property relations (e.g., c
p
, k
m
, a
m
, e
p
, e
f
, etc.) can be correspondingly expressed as
functions of only mean volume fraction, density, pressure etc.


Mixed-Fluid Equation of State

If either of the phases are themselves significantly compressible, a mixed-fluid equation of
state is needed to relate
m
and p. For the dusty-gas equation of state (Eqs. 5.24 and 5.25),
the time-average can be approximated as:


( )
m m m m
p 1 T T + =
g g
R R
m m m m
m
T T
p T
1 1 p / p

=
+ + +
g g
g g
R R
R
5.67a

5.67b
The approximation used in Eq. 5.67a is the conventional single-phase assumption (Speziale et
al. 1991) that the correlations between the gas density and temperature turbulent fluctuations
do not make a significant contribution to the mean pressure. This is generally reasonable
except for conditions of combustion or hypersonic flow. The approximation used in Eq. 5.67b
is that the pressure fluctuations of the gas are not strongly related to the mass loading
fluctuations, i.e. p p < . This is similar to the argument used to obtain Eq. 5.64.
For a bubbly flow with turbulent time-scales larger than the thermal response time-scale
(Eq. 1.93b), an isothermal behavior is expected and the instantaneous state relationship of Eq.
5.35 can be rewritten in terms of the local volume fraction (instead of
o
) as:



( )
m mo
m
o mo mo
1
p
1
p

= = +


5.68
Taking a time-average of this equation and assuming p p < for low-speeds yields a
mean equation of state given by:



m
o mo
p
1
p

= +


5.69
This form is convenient in that it does not include any perturbation correlations.
To examine conditions where the criterion for low-speeds is reasonable, note that typical
pressure fluctuations in turbulence are related to the velocity fluctuations, e.g.,

rms
p 0.6 u u
l
(Hinze, 1975). If the velocity fluctuations are limited by
rms
u 0.1u < and the
mean velocity is limited by 13 m/s, then
rms
p 0.01p < . Furthermore, Eq. 5.35 indicates that the
volume fraction and pressure are approximately inversely proportional, i.e. / p / p so
that



4
p / p 10

< for low-speeds with no mass transfer 5.70


214
This is the assumption is used often (e.g., Chahed et al. 2003) and a similar relationship can be
put forth regarding density,
p p
/ 1 < , and temperature. However, the above inequality
also assumes that the pressure frequencies are small compared to the bubble natural frequency
so that the bubble is near equilibrium and non-linear volume oscillation effects are negligible.


Transport Equations for Mixed-Fluid Density and Volume Fraction

The time-average of the mixed-fluid continuity equation (Eq. 5.48) gives


( )
m m,i
m
i
v
0
t x

+ =


5.71
An unsteady term is retained as a numerical convenience to provide a pseudo-steady
formulation (introduced with Eq. A.114) which allows a time-marching schemes to converge
the solution to a steady-state. Using Eq. 5.60a, a Favre-weighted average yields:


( )
m m,i
m
i
v
0
t x

+ =


5.72
This averaging eliminates the fluctuating correlations which would have occurred with a
Reynolds-average (e.g., compare Eq. A.113a to Eq. A.113b). Similarly, the time-average of
the mixed-fraction PDE (Eq. 5.47) with a pseudo-steady formulation is given by:


( ) ( )
( )
p p m,i
p Br p
i p
v
m
t x


+ = +


` 5.73
Based on Eq. 5.64, the volume-averaged phase densities can be used for Favre-averaging (Eq.
5.61) so that the transport equation becomes:


( ) ( )
( )
p p m,i
p Br p
i p
v
m
t x


+ = +

` 5.74
Again, it is seen that Favre-averaging can eliminate turbulent fluctuation correlations. If the
particles are incompressible and one neglects Brownian motion and mass transfer and does not
retain the artificial unsteady term, then the mixture-fraction PDE is greatly simplified:


( )
m,i
i
v
0
x

=

5.75
This result indicates that the volume-fraction weighted velocity is divergence-free. This
property may be exploited numerically with a pressure-based solver, similar to the use of Eqs.
5.59b and 5.59c for Eq. 5.52b For chemical reactions and other special cases, it may be useful
to employ an additional transport equation for , which can then be used to prescribe a
probability distribution function for the volume fraction.
For later transport equations, the difference between the Reynolds-averaged velocity and
the Favre-averaged velocity is needed and this can be obtained with Eqs. 5.59 and A.112:

m,i m,i m m,i m m,i m,i


v v v / v v = = = 5.76
The term on the RHS is sometimes expressed as the drift velocity (Viollet & Simonin, 1994)
and represents the flux associated with turbulent mixing, i.e. it causes mean mixed-fluid
215
density to diffuse. This term requires turbulence modeling as will be discussed in a later sub-
section.


Mixed-Fluid Momentum Transport

For the mixed-fluid momentum, a conservative form is often employed since the density
variations due to the multiphase properties can be substantial (B.3). Applying a time-average
with a pseudo-steady formulation to Eq. 5.55 yields:


( ) ( )
m m,i m m,i m, j
m, j m,i m,k
m i m
j i i j i k
v v v
v v v
p 2
g
t x x x x x 3 x





+ = + +






5.77
By employing Favre-averaging (Eq. 5.60), the triple-correlation on the LHS can be simplified.
The velocity gradient terms on the RHS can also be simplified and then approximated as

m, j m, j m, j m, j
m m m m m m m
i i i i
v v v v
x x x x

= =

5.78
The last approximation assumes that the variations in the effective viscosity are not
significantly correlated to the fluctuations in the velocity gradient, e.g.,

m m, j
v 0 . This is
quite common and is consistent with using a quasi-steady viscosity expressions (5.1.2). The
momentum equation can then be re-written as


( ) ( )

( )
m m,i m, j m m,i
m i
j i
m m,i m, j
m, j m,i m,k
m m
i j i k j
v v v
p
g
t x x
v v
v v v
2

x x x 3 x x

+ =


+ +







5.79
This is similar to Eq. A.115b for a single-phase stratified-density field. To close the equations,
the velocity fluctuation tensor correlation (

m,i m, j
v v ) requires turbulence modeling, which will
be discussed in a later sub-section.


Mixed-Fluid Energy Transport

The average energy equation can also be obtained using enthalpy, internal energy or total
energy as the transport variable. There are many forms of the average energy equation
resulting from a variety of assumptions and choices of enthalpy or energy for the transport
variable (e.g., Bilger, 1976; Kuo, 1986, Shyy et al. 1997; Chung 2002). If total energy is used
for the transport variable, time-averaging Eq. 5.56 yields:


( ) ( )
m m,tot m m,tot m,i
m
m i m,i m,i m,ij m, j m
i i i
e e v
T
g v pV K v
t x x x



+ = + +



+ k 5.80
Application of Favre-averaged with a simplification of the heat transfer term yields:
216


( ) ( )

( )
m m,tot m m,tot m,i
m
m i m,i m,i m m, j m,ij
i i i
m m,tot m,i m,i m,ij m,i m,ij
i
e e v
T
g v pv v K
t x x x
e v v K v K
x


+ = + +


+ +

+
+

k

5.81
The heat transfer term was simplified using two assumptions: 1) the thermal conductivity
coefficients of each phase given in Eq. 5.8 are constant so that mixed-fluid conductivity is only
a function of , and 2) the temperature fluctuations are weakly correlated to the density
fluctuations so that
m m
T T

. Both of these assumptions are generally reasonable for low-


speed flows (Bilger, 1976). The total internal energy on the LHS includes internal and kinetic
energy (Eq. 5.57) and can also be written in terms of Favre-averaged temperature by assuming
the specific heats of each phase (Eq. 5.9) are constant:



( )
1 1
m m,tot m p,p m,tot m,i m,i m p,p m,tot m,i m,i m,i m,i
2 2
m p,p p p,p f p,f
e T v v T v v v v
(1 )
= + + +
+

c c
c c c

5.82a

5.82b
The RHS mixed-fluid fluctuations of total energy can then be expressed as:



( ) m m,i m p,m m m,i m m,k m,k m,i m,k m,k m,i
e v T v v v v v v v = + + c
5.83
Additional correlations appear with respect to the viscous shear stress, e.g.,


m, j m,i m,k
m,ij m ij
j i k
v v v
2
K
x x 3 x

= +




5.84
The RHS neglects fluctuations between the viscosity and the velocity and utilizes the drift
velocity of Eq. 5.76. As such, the above transport equations require relations for terms such as
m m,i
v ,

m,i m, j
v v and

m m,i
T v , which are (finally) discussed in the following.


Turbulence Modeling

The above transport equations involve correlations quite similar to the single-phase counter
parts discussed in A.5.5. For example,
m m,i
v is associated with diffusion of the mixed-fluid
density from regions of high concentration to regions of low concentrations. To show this
qualitatively, consider the turbulent mixing shown in Fig. A.18 and assume that the upper
white region has no particles while the lower black region represents fluid mixed with many
particles such that it has a higher mixed-fluid density, i.e.
m
/ y 0 < . In this case, an
upward velocity fluctuation (
m,y
v 0 > ) will tend to carry high density fluid from below to upper
regions introducing a local increase in the mixed-fluid density (
m
0 > ). Similarly, a
downward velocity would tend to carry lower density fluid producing a local decrease in the
mixed-fluid density. In both cases, this leads to
m m,i
v >0 which is an upward diffusion and is
in the opposite direction of the mixed-fluid density gradient. Furthermore, increased mean
gradients will lead to increases turbulent transport of the density perturbations. Furthermore,
high spatial gradients in mixed-fluid density and high turbulent viscosity levels can both be
expected to increase the diffusion magnitude. Based on the above phenomenology, one may
217
expect that
m m,i
v will be related to
m
and to
turb,m
. Indeed, the density-velocity
correlations can be modeled via Fickian diffusion as is used for the single-phase transport (Eq.
A.121) as :


turb,m
m
m m,i m m,i
turb i
v v
Sc x

= =


5.85
The negative sign is consistent with diffusion in the opposite direction of the gradient and
indicates that the drift velocity identified in Eq. 5.76 serves to reduce the mixed-fluid density
gradient. As with the continuous-phase version, Sc
t
is generally taken as 0.7 based on several
experiments which compared turbulent mass diffusion to turbulent momentum diffusion (Faeth,
1987).
A similar argument can be made regarding the Favre-averaged temperature and velocity
fluctuations so that one may expect one may expect that

m m,i
T v will be related to
m
T

and
to
turb,m
. Again applying a Fickian diffusion model yields:

turb,m
m
m m m,i
turb i
T
T v
Pr x


5.86
This form uses a turbulent Prandtl number which is the ratio of momentum diffusion to thermal
diffusion, which is analogous to the laminar version given by Eq. A.52 and is generally taken
as 0.9 (Shyy et al. 1997).
Perhaps the most important correlation is

m,i m, j
v v which is needed to close the momentum
equation. This term is associated with diffusion of the mixed-fluid momentum from regions of
high velocity to regions of low velocity. As with the temperature and density, this term is
expected to be proportional to the density gradients and turbulent viscosity. As with the single-
phase equivalent given by Eqs. A.115b and A.116b, it is referred to as a turbulent-stress on the
fluid and is typically modeled by assuming a Boussinesq gradient diffusion process:



m, j m,i m,k
m,ij m m,i m, j m,turb ij m m
j i k
v v v
2
K v v k
x x 3 x

+ +





5.87
This closure model mimics the laminar viscosity stress which appears on the RHS of Eq. 5.79.
It also includes the turbulent kinetic energy to satisfy the condition when there are no spatial
gradients of the mean mixed-fluid velocity:

1
m m,i m,i
2
k v v

5.88
The mixed-fluid turbulent viscosity is also expected to scale with the strength of the turbulence.
As with the stratified flow relations of Eqs. A.117, A.122 and A.123, a common approach is to
model this viscosity in terms of the turbulent kinetic energy and dissipation:


2
m,t m m,turb m m,turb m m m
c k /

= =

5.89
In the RHS expression, c

is the empirical constant widely taken as 0.09 (Eq. A.120) and


allows an effective viscosity given as.


( )
m,eff m m m,turb
= +
5.90
As such, closure of the above equations lastly requires equations for of the mixed-fluid
turbulent kinetic energy and dissipation in terms of mean flow properties.
218
The most common approach for these two terms is to use a two-equation turbulence model.
For free-shear flows, a popular version of this is the k
m
-
m
formulation (Faeth, 1987). In
particular, Favre-averaged transport PDEs can be written for both the mixed-fluid kinetic
energy and dissipation to incorporate the physics associated with the convection, generation,
diffusion and dissipation using the methodology used for those in Eq. A.126:


( ) ( )

( ) ( )

m m m m m,i
m, j
m
m,ij m,i m,eff m m
i i i i i
m m m,i m, j m m
m
m,ij m,i
i i i m
k k v
v
k p
K v
t x x x x x
v v
1.44 p
K v
t x x x k




+ = +




+ =




2
m m
m,eff m
i i m
1.9
x x k

+




5.91a




5.91b

As with Eqs. A.125b and A.126a, each LHS represents the mean convection with a pseudo-
steady formulation. The first two terms on each RHS are the production terms associated
with the mean velocity and pressure gradients. Each of the third terms includes the effective
viscosity (Eq. 5.90) and represents the turbulent diffusion of the conserved quantity. The last
term represents dissipation and is therefore negative. The reader is referred to A.5.5-A.5.6 for
additional discussion on the assumptions and modeling issues associated with these equations.
There are several important aspects of these turbulence model equations. Firstly, they
employ empirical coefficients (0.09, 1.3, 1.44, 1.9, etc.) which have been obtained by
validation with simple shear flows, but which are not universal. Secondly, they arise from
simplifying assumptions that fluctuation correlations can be related to gradients of the mean
velocity field. Thirdly, several terms have been neglected in expressing these two equations.
Some of these are discussed by Shyy et al. (1997) and in fact there are many different versions
of the PDEs depending on which terms are included and how they are modeled, etc. Generally
these terms are secondary, but in some cases are helpful to improve calibration fidelity. Lastly,
these transport equations can be replaced by those of other turbulence models such as the k-
model (which tends to give better treatment for all boundary layers) and stress-tensor models
which allow for the prediction of anisotropic turbulence intensities. Thus employment of this
approach for complex conditions, especially with NAsT flows can be expected to lead to
significant error.


Axisymmetric Shear-Layer Examples

For axisymmetric free-shear flows with the boundarylayer approximation, the above
transport equations can be compactly summarized (Faeth, 1987) as:


( ) ( )
m m,x m m,r m,t
q
q
v q v q
q
r
x r r r r c r


+ =





S
5.92
The source term and coefficients are given in Table 5.1 where the various qs represent
transport quantities of mixed-fluid mass, dispersed-phase mass, mixed-fluid momentum,
turbulent kinetic energy, and turbulent dissipation. This formulation assumes low Mach
number isothermal free-shear flows so that wall effects associated with
m,t
and many of the
compressibility corrections are negligible. Faeth (1987) discussed the solution of these
219
equations for several multiphase flowfields whereby the coupled PDEs were solved by adding
time-derivative terms to the transport equations, i.e. the pseudo-unsteady formulation
discussed in A.5.5.
Sample predictions of this RANS mixed-fluid approach are shown in Fig. 5.5 based on
particle-laden circular jet data of McComb & Salih (1977, 1978). The particles for these flows
have small terminal velocities (
*
term
w 1 < ), so that the Stokes number criterion of Eq. 5.3b is
the most important constraint for use of the mixed-fluid approximation. The titanium dioxide
particles (d=2.3 m and
p
=4,260 kg/m
3
) were consistent with a mixed-fluid condition since
the particle response time (based on the local centerline conditions) was small, about St

~0.01.
As a result, they were reasonably modeled with the mixed-fluid approximation. The tungsten
particles (d=5.7 m and
p
=19,300 kg/m
3
) were placed in a smaller jet and yielded a larger
response time (about St

~2). Consequently these higher inertia particles were not modeled


well with the mixed-fluid approximation. In particular, these particles yielded reduced lateral
diffusion consistent with increased Stokes numbers (Fig. 3.21). Similar trends were found for
bubbly flows (Sun & Faeth, 1986; Loth & Faeth, 1989).



5.3. Weakly-Separated-Fluid Methods

For many particle flows, the mixed-fluid approach is often unacceptable since interphase
differences can have a significant effect on the mean convection of the individual phases. For
example, errors as large as 50% have been found with respect to experimental particle
concentrations and particle velocities once particle response times become significant (e.g., Fig.
5.5b). Furthermore, the physics of preferential concentration and turbo-phoresis can not be
captured with the mixed-fluid approach, which assumes w0. Therefore, separated-fluid
representations of the particle and continuous-fluid fields which directly take these differences
into account are more common. However, they are also more computationally demanding. An
intermediate option which retains the computational convenience of the mixed-fluid method
but allows relative velocity physics is the weakly-separated-fluid approach. This approach
assumes that the relative velocity is a function only of the instantaneous continuous-phase
conditions and the particle response time. In its simplest form with w=w
term
, the weakly-
separated-fluid approach has been called the terminal velocity or laminar sedimentation
approach (Hubert, 2001). The more general form has been called this the fast Eulerian
method by Ferry & Balachandar (2001) call the weakly-separated-fluid method the fast
Eulerian method because of its computational efficiency while Rani & Balachandar (2003)
call it the equilibrium Eulerian method because it neglects the particle velocity history. In
general, this method is often reasonable for particles with finite but modest Stokes numbers
(e.g., less than unity) and is discussed below for both instantaneous and averaged continuous-
phase flow fields.

5.3.1. Weakly-Separated-Fluid Transport Equations

The basic concept of the weakly-separated-fluid approach is to obtain the particle velocity
field solely from the instantaneous particle characteristics (e.g., size and density) and the
220
instantaneous continuous-phase velocity and acceleration field (using the point-force
approximation). This will require the particle response time influence to be small compared to
the continuous-phase time-scales so that the particle will not be significantly affected by
previous particle conditions, e.g., injection conditions are assumed to be unimportant.


Weakly-separated Eulerian Particle Field

To obtain the weakly-separated-fluid flow equations, the Lagrangian particle momentum
equation can be transformed into an Eulerian equation for particle velocity as a function of the
continuous-phase properties (and independent of particle history). To do this, let us initially
assume that wu and St1 so that w may be considered in terms of a small perturbation
(whereas mixed-fluid assumes w0 and St0). For the weakly-separated condition, Eq. 3.5
is therefore appropriate since the difference between du
@p
/dt and Du
@p
/Dt (based on Eq. 1.15)
will be similarly small, i.e.
( ) ( )
@p @p @p
w u u u <
5.93
In fact, this approximation becomes exact Re
p
1 (Maxey, 1987). Based on Eq. 3.6 for a solid
spherical particle (which neglects Basset history, lift and particle-particle interactions), the
particle equation of motion can be rearranged as:

@p
@p p
(1 R) R
t t

+


u
v
v = u + g



d
d
d d

5.94
The particle acceleration term on the RHS must be treated in order to allow the weakly-
separated flow approach. Taking the Lagrangian time derivative of all the terms in Eq. 5.94
yields (Ferry & Balachandar, 2001):
( )
2
2
@p @p @p
p p 2 2
R
t t t t t

= + +



u u u
v v
( ( (
( (

d d d
d d
O
d d d d d

5.95
The first approximation indicates that the accelerations differ by a term of order
p
while the
second approximation is consistent with Eq. 3.5. Substituting this result on the RHS of Eq.
5.95 into the particle acceleration term on the RHS of Eq. 5.94 and applying Eq. 3.5 yields:

@p 2
@p p p
(1 R) O( )
t





D
D
u
v = u + g - +
5.96
Substituting expressions for the terminal velocity (Eq. 1.82a) and the Eulerian time derivative
(Eq. 1.15) and neglecting the second-order terms on the RHS (equivalent to assuming small
p
),
the first-order weakly-separated-fluid particle velocity is given by
( )
@p
@p term p @p @p
(R 1)
t

= +


u
v u + w + u u


5.97
This is the equation obtained by Druzhinin (1995) and indicates that this equilibrium particle
velocity is simply a function of particle response time and the continuous-phase properties.
221
The RHS of this equation can be obtained for each time-step to solve for the Eulerian
particle velocity field (which is independent to that at the previous time-step), e.g.
( ) ( )
n 1
@p n 1 n 1 n 1 n 1 n 1
@p term p @p @p
R 1
t
+
+ + + + +

= +



u
v u + w + u u


5.98
There are no stability constraints on this algebraic solution and the RHS time-derivative term
can be obtained by backward differencing with the previous time-step or central differencing
with the n and n+2 time-step values of u
@p
for higher-accuracy. Ferry & Balachandar (2001)
extended the above derivation to include the second-order terms and non-linear response times
as well as the effects of lift and Basset-history forces. However, Eq. 5.98 generally is
reasonable for Stokes numbers less than unity and Re
p
less than 10.
If the fluid accelerations and gradients are weak ( u / t g and u g / u < < ) or the particle
response time is negligible compared to fluid time-scales, then Eq. 5.98 simply reverts to:

n 1 n 1
@p term
+ +
= v u + w
5.99
This is the terminal velocity approach whereby the particle velocity field is simply linearly
shifted from the continuous-phase velocity field (both have the same acceleration field). In
either case, the above equations remove the need for integrating a particle momentum PDE.
The resulting velocity field from Eq. 5.98 or Eq. 5.99 can then be used in the particle
concentration PDE of Eq. 4.29a (assuming constant particle density) as:


p
j
j p p
m
v
t x


+ =

`

5.100
For the terminal velocity approach, this equation simplifies to a linearized advection equation.
The numerical solution of the concentration transport equation is generally chosen to be
consistent with the methods used for the continuous-fluid (App. B), i.e. same spatial
discretization and time-step integration. Note that for one-way coupled conditions, the
continuous-phase velocity is independent of the particle velocity. This is especially convenient
for handling polydisperse size distributions, since one may then construct several bins to
describe the different sets of particle diameters (as discussed in 4.2.1). For a single solution
of u, the particle velocity fields for all the bins can be based on bin versions of Eq. 5.98. Each
of the particle velocity fields can then be used to determine each of the concentration fields
using bin versions of Eq. 5.100.
The weakly-separated-fluid approach can also be extended to include some two-way
coupling by noting that the fluid dynamic surface force based on the above assumptions and
neglecting particle collisions can be written based on Eqs. 1.21 and 1.23 under the assumption
of constant particle mass. The surface force per unit particle volume can then be
approximately written in terms of the fluid acceleration by Eq. 5.95


@p
surf
p p
p
t t


=



u
F v
g g
(

D
d
d D
5.101
The Lagrangian fluid acceleration term on the RHS can then be converted to a temporal and
spatial and derivatives of conservative quantities (e.g., Eq. A.6b). The modified RHS can then
be substituted into the continuous-phase momentum equation (Eq. 4.31) and written using the
222
modified density of Eq. 5.4 as:

( )
( ) ( )
m
m m i j ij
K p
t

+ =

u
uu g + 5.102
This two-way coupled weakly-separated-fluid equation for u is identical to that for the mixed-
fluid approach for V
m
given by Eq. 5.44. Thus, the continuous-phase momentum in either
approach is only affected by the modified density (Ferry & Balachandar, 2001). Similarly, the
effect of the particles in terms of three-way coupling and mixture compressibility for the
weakly-separated-fluid approach can also be included by using the mixed-fluid properties
(5.1.2 and 5.1.3).


Weakly-Separated Lagrangian Particle Field

An approach for a weakly-separated Lagrangian particle field using an Eulerian
continuous-phase flow field was developed by Healy & Young (2005) based on an integration
of the linear-drag particle dynamics equation of Eq. 3.6. In particular, one may use a central
difference in time to estimate the fluid acceleration along the particle path (based on Eq. B.49)
to incorporate its effect in terms of the include fluid-stress force as:
( ) ( )
p p
t / t /
n 1 n 1 n n n 1 n
@p @p p @p @p
e (1 R) R / t 1 e

+ + +

= + +

v u v - u g + u u
5.103
This is similar to the integration scheme of Eq. 3.8a except that it employs the continuous-fluid
velocities at both time-steps. This requires that the continuous-phase flow field be solved first
but ensures numerical stability even when the particle response times is small, i.e.
p
t < .
Once the particle velocities at the old and new time-steps are known from Eq. 5.103, the
particle position may be similarly updated by modifying Eq. 3.8b to give:

( )
( ) ( )
( )
p
p
t/
n+1 n n+1/2
p p p
t/ p n+1/2 n 1 n
@p p @p @p p
+ 1 e
R
1 R t 1 e
t

+
=


+ + +





x x v
u g u u

5.104
For non-linear drag, the particle response time in Eqs. 5.103-5.104 can be replaced with
n
p
,
which can be computed using w
n
via Eq. 3.1. Note that this approach uses v
n
to help compute
v
n+1
, and as such can account for initial conditions. This is in contrast to the weakly-separated
Eulerian approach which does not use v
n
at all, and thus can not allow for effects of initial
conditions as indicated Table 4.2. While the weakly-separated Lagrangian approach is not
necessarily more efficient than the fully-separated Lagrangian approach in terms of computing
the particle velocity, it is more computationally efficient in terms of computing particle
concentration fields as discussed in the following.
For a known Lagrangian particle velocity field, the local instantaneous particle
concentration may be obtain based on the number of particles within each cell (Eq. 4.6), but
such methods require many particles per cell to achieve high-accuracy, especially if two-way
coupling is important (Eq. 4.26). However, the weakly-separated assumptions allow another
option which is much more computationally efficient. This method can be derived starting
from an Eulerian particle concentration PDE. For example, Eq. 5.100 can be expanded as
223


j j p
j
j j j p p
v v m
1 1
v
t x x t x


+ + = + =






`

d
d

5.105
The LHS can be rearranged to yield a Lagrangian ODE for concentration change:


( )
p
p
m ln
t m

= + v
`

d
d

5.106
This technique has been called the full Lagrangian method by Healy & Young (2005) since it
allows particle concentration to be carried as an additional quantity along trajectories. A key
for this method is to obtain the divergence of the particle velocity field, which is not normally
available with a Lagrangian method. However, the weakly-separated conditions allows this
spatial derivative to be described using Eq. 5.97
( )
@p
@p p @p @p
(R 1)
t

= +


u
v u + u u


5.107
Substituting this equation into Eq. 5.106 and discretizing with a central-differencing scheme
yields the concentration along a representative particle path defined by Eq. 5.103

( ) ( )
n 1/ 2
n 1
@p p
@p p @p @p n
p
m
ln t R 1
t m
+
+


= + +




u
u u u
`

5.108
This can then be interpreted as an ODE for the parcel concentration. If the continuous-phase
and particle momentum equations are first integrated to t
n+1
, then the RHS can be estimated
using Eqs. B.49 and B.50. The volume fraction along the path can then be integrated, though
care must be taken with the time-step so that that the RHS is small compared to unity to ensure
accuracy.
This method has several advantages. Firstly, note that all the terms on the RHS are
straightforward to obtain for an Eulerian discretization of continuous-phase velocity field, e.g.,
the spatial derivatives at time-steps n and n+1 can be averaged while the temporal derivative
for n+1/2 can be obtained with central differencing between time-steps n and n+1. This
technique requires that each particle be initialized with the local particle concentration field
and then Eq. 5.108 can be used to find its temporal evolution. The particle concentration
within a computational cell is then simply given by the average within that cell, i.e.

j
j 1
N
1
N


5.109
Other methods to convert the Lagrangian concentrations to Eulerian concentrations will be
discussed in 7.2.7. The major benefit of this method is that it is much more computationally
efficient than the conventional counting method of Eq. 4.6. For example, 2-D particle
flowfield simulations by Healy & Young (2005) required only 200 pathlines with the full
Lagrangian method to obtain a statistically converged particle concentration, while about
20,000 conventional pathlines were required to produce the same level of accuracy.


224
Weakly-Separated Advantages

There are key numerical efficiency advantages to using the weakly-separated-fluid
approach as compared to the conventional Eulerian separated-fluid methods. The major
advantage of the weakly-separated-fluid approach is that the dispersed-phase momentum PDEs
(three in 3-D flow) are eliminated and replaced by the algebraic expression such as given in Eq.
5.98. This is particularly convenient if one wishes to describe several bins of the particle phase
to allow a polydisperse distribution. In addition, the separated-flow PDEs often limit the
particle velocity updates to time-steps based on some fraction of the particle response time
(e.g., t=
p
/8). This can be overly restrictive if
p

f
. While, there are some approximate
methods which can avoid this restriction (Ling et al. 1998), use of the weakly-separated
approach allows t to be determined instead by
f
(as well as the PDE character and time
integration scheme as discussed in B.3.2)
The major advantage of the weakly-separated-fluid over the mixed-fluid approach is an
improved description of particle velocity that can approach the separated (exact) solution. This
is demonstrated in Fig. 5.6a where the changes of particle velocity in a wall-bounded turbulent
flow are reasonably reproduced up to a Stokes number of unity. Inclusion of the fluid
acceleration terms for the weakly-separated-fluid approach (Eq. 5.97) tends to be somewhat
more accurate than that of the terminal velocity approach (Eq. 5.99). Similarly, Fig. 5.6b
shows that the relative velocity fluctuations (related to preferential bias) are reasonably
modeled with the weakly-separated approach for St
+
<1. Note that the ability to incorporate the
particle dynamics associated at these conditions through Eq. 5.97, also allows the weakly-
separated approach to capture the physics of non-linear drag bias and turbo-phoresis directly,
whereas the terminal velocity or mixed-fluid approaches can not capture preferential bias or
clustering bias.
However, it should be noted that the weakly-separated-fluid approach can not easily
incorporate particle collisions in an Eulerian treatment. Because of this, non-physically large
particle concentrations may occur due to streamline crossings if the instantaneous Stokes
numbers get too large (Ferry & Balachandar, 2001). Such errors in the particle transport (Eq.
5.100) can lead to sharp concentration gradients and numerical instabilities for higher-order
schemes. To prevent non-physical concentration gradients and numerical instabilities, artificial
diffusion can be added (Rani & Balachandar, 2003). However, the inclusion of physical
diffusion (from Brownian motion or turbulence) in the transport equation can also eliminate
this problem. Incorporation of such diffusion is discussed in the next section.

5.3.2. Time-Averaged Transport Equation and Particle Velocity

As with the mixed-fluid equations, time-averaged weakly-separated-fluid formulations for
turbulent flow may also be developed. Such time averaging assumes that the particle response
times do not exceed the integral time-scales (i.e. St

<1) while for LES the particle response


times should not exceed the sub-grid time-scales (i.e. St

<1). A time-average of the mixture


fraction transport of Eq. 4.73 (which includes Brownian transport) can be used to obtain a PDE
for . If the continuous-phase is incompressible and does not include two-way coupling
effects, it is likely to be given in a Reynolds-averaged formulation (e.g., Eqs. A.104a and
A.115a). In this case, it is helpful to use Reynolds-averaging (not a Favre-averaging) for the
225
mixture fraction PDE. If one also retains the unsteady term for a pseudo-unsteady formulation,
the result becomes:


( )
i
i
Br
i i i i
v
v
t x x x x

+ + =




5.110
For small particle inertias, the second term on the LHS can be modeled in the same manner as
Eq. 5.85 as a turbulent diffusion component (Young & Leeming, 1997):


( )
t ,m i
Br
i i t ,m i
v
t x x Sc x



+ = +





5.111
In order to apply this particle concentration transport equation, we need the time-averaged
particle velocity field. This can be obtained directly from the continuous-phase velocity field
through the weakly-separated-fluid approach outlined below.


Averaged Particle Velocity

Using the perturbation assumption of Eq. 5.97, the time-averaged particle velocity can be
written in terms of the terminal velocity and the mass-weighted continuous-phase velocity as:

i@p i@p
i i@p term,i p j@p j@p
j j
u u
v u w (R 1) u u
x x

= + +



+
5.112
As before, a correction is included based on particle response time. The last correlation term
on the RHS of Eq. 5.112 is related to non-homogenous turbulence which will be discussed in
the next section. However, if the turbulence is approximately homogeneous and the spatial
gradients of the turbulence and the mean flow are weak, then the velocity gradient terms
become negligible and we recover the terminal velocity approach of Eq. 5.99

i i@p term,i
v u w = +
5.113
Experiments in gravity-dominated ( Fr 1

> ) for have shown This has been found to be a


reasonable approximation for particles in homogeneous weak turbulence if the flow is dilute
and gravity dominated and the particle drag is linear (Crowe et al. 1996 and Brennen, 2005).
Note that variations due to two-way, three-way and four-way coupling can be incorporated
by including mixed-fluid effects on the density, viscosity and collisional pressure. However,
there are three types of coupling conditions where the average relative velocity will be
significantly different than that given by the terminal velocity. These conditions include
unsteady effects associated with: 1) turbo-phoresis, 2) non-linear drag bias, 3) preferential bias,
and 4) clustering bias. These are discussed in the following three sub-sections.


Turbulent Bias Effects

226
Let us consider flows with substantial spatial variations of the time-averaged turbulence
intensity based on Eq. 5.112. In this case, a model for the correlation term on the far RHS
must be obtained. This term can be rewritten (temporarily dropping the @p notation) as

j j
i i
j i
j j j
u u u
u
u u
x x x

=

5.114
For incompressible flow, the second term on the RHS can be neglected since
j j
u / x =0 (refer
to Eq. 5.69). The first term can then be modeled as in Eq. 5.87 for anisotropic turbulence, or as
1
3 i
k/ x for isotropic turbulence (Bocksell & Loth, 2006). Neglecting the mean velocity field
gradients, Eqs. 5.112 and 5.114 can be combined to give the change in relative velocity due to
turbulence gradients:

i j
i term,i k,i p
j
u u
w w w (R 1)
x


= =

for St 1

<
L
5.115
The RHS is the turbophoretic velocity seen by particles of small inertias. This reverts to that
derived for very-heavy particles in Eq. 3.128, whereby an increase of the turbulent kinetic
energy in one direction leads to a turbophoretic velocity in the opposite direction for a high
density particle (R>1) or in the same direction for a low density particle (R<1). As shown in
Fig. 3.25, the turbophoretic velocity (along with lift) is important to predict particle deposition
physics for St
+
of 0.1 to about 20 (corresponding to St

of about 0.01 to 1). This figure also


indicates that Brownian diffusion (RHS of Eq. 5.110) becomes more important at smaller
particle inertias, while turbulent diffusion (2
nd
term on LHS of Eq. 5.111) becomes more
important at higher inertias.
An assumption in the weakly-separated-fluid approach is that the drag is linear so that the
mean relative velocity is simply equal to the terminal velocity. However, a non-linear drag
behavior (variable
p
) coupled with finite fluctuations in the relative velocity will result in
w <w
term
as discussed in 3.5.4. Furthermore, preferential concentration effects can be
incorporated for isotropic turbulence based on Eq. 3.122 so that Eq. 5.112 becomes:


i@p i@p j@p
i i@p Re,i pref,i p j@p
j j
u u u
v u w w (R 1) u
x x

= + + +



+
5.116
This could also be extended to include cluster bias effects, but these are not well understood as
discussed in 3.5.5 and noted in Table 4.2.

5.3.3. Capability of Approaches for Particle Velocity Effects

To conclude this chapter, it is worthwhile to review the capabilities of the mixed-fluid
techniques. The mixed-fluid approach (and the other methods) can incorporate two-way and
three-way coupling effects based on modified density, viscosity, etc. It can also be employed
with a time-averaged (or LES) description of the continuous-phase flow. The main attribute of
mixed-fluid approach is that it allows the simplest computational description of the multiphase
field, but can not account for any of the relative velocity or particle inertia effects (Table 4.2).
227
The weakly-separated approach allows for any magnitude of terminal velocity to be
included and can also allow for weak inertial effects, e.g., St
D
<1 for nonturbulent flows and
St

<1 for turbulent flows. For these conditions, the unsteady weakly-separated-fluid approach
allows all the relative velocity effects to be included without need of ODEs or PDEs for the
particle velocity field. Favre-averaged continuous-phase fields can also be employed with the
weakly-separated-fluid approach to include terminal velocity effects in a time-averaged sense.
However, these approaches can suffer limitations in terms of predicting preferential
concentration and associated bias effects, since these phenomena tend to be strongest for
Stokes numbers of order unity (3.5.5).
Though not listed or discussed above, it should be noted that Large Eddy Simulations can
similarly incorporate the weakly-separated-fluid approach to allow algebraic representations of
the particle velocity field if the particle dynamics are weak (i.e. St

<1). In this case, the


effects of turbo-phoresis can be modeled in a manner similar to that for a steady time-averaged
approach, i.e. Eq. 5.116 can be interpreted as an LES transport equation if
i
u represents the
unfiltered velocity field and
i
u represents the filtered velocity field (B.4). The separated-
fluid approach is more robust as it can handle all particle Stokes numbers) but is
computationally intensive. Various formulations for the point-force representations needed for
the separated-fluid approach are discussed in the next chapter.



5.4. Problems

5.1) Based on Eqs. 5.29 and 5.30, derive the bubbly-flow speed of sound relations given by
Eqs. 5.41 and 5.42.
5.2) Consider a 1-D mixed-fluid multiphase flow moving at a constant velocity of 5 m/s
which contains water and very small neutrally-buoyant constant-diameter spherical
particles so that a mixed-fluid approach is appropriate.
a) Derive appropriate discretized equation for the Eulerian particle volume fraction.
b) Use these equations to solve for the unsteady volume fraction distribution in a domain
of x=0,L where L=100 m with the initial conditions: =0 for x0.4L, =0.1 for
x0.5L and varies linearly for 0.4Lx0.5L. Use a grid resolution of 100 cells and
a sufficiently small temporal resolution to ensure time-step independence (and stability).
c) On one graph, plot the volume fraction distribution for t = 0, 1, 2 and 4 seconds and
discuss the trends and what effects, in any, are associated with your numerical scheme.
5.3) a) Starting from Eq. 5.54, derive the following mixed-fluid momentum equation (and
state all assumptions requited):

( )
( )
m mi mj
mj
mi
m i m m,t m
j i i j i
v v
v
v p 2
g k
x x x x x 3


= + + +







b) Based on Eq. 5.87, qualitatively explain the fluid physics which relate the mean
velocity gradients to the velocity fluctuations and the sign of this first term.
Quantitatively show that the second term must be included in Eq. 5.87 by considering i=j.
5.4) a) Starting with Eq. 1.77, derive Eq. 5.98 and show that is second-order accurate with
respect to
p
. b) Continue with a time-average of this expression to obtain Eq. 3.128.
228
5.4) Starting with Eq. 5.105, derive Eq. 5.108.
5.5) Consider a potential flow solution of flow past a cylinder whose stream function is given
by
( )
2 2 2
o
u y 1 r / x y


= +

. For a very heavy particle (R0) with linear drag, a
Stokes number defined as
D p o
St u / r

= , and a terminal velocity given by w


term,y
=-
0.1u

i
y
:
a) Use the Eulerian weakly-separated approach of Eq. 5.98 to analytically obtain v
y
/u

at
x/r
o
=-1.1 for y/r
o
ranging from -3 to +3 for St
D
=0.2 and 2.
b) Use the terminal-velocity approach of Eq. 5.99 to similarly obtain v
y
/u

.
c) Use the separated-fluid Lagrangian approach of Eq. 3.8 to obtain v
y
/u

for particles
when they cross the plane at x/r
o
=-1.1 but are initially released with a velocity equal to
that of the fluid

and at positions with y/r
o
of -2, -1, 0, +1, +2, +3 and x/r
o
=-2. Make sure
your time-step is small enough that your results are time-accurate.
d) For St
D
=0.2, plot v
y
/u

as a function of y/r
o
for the three different techniques (use solid
and dashed lines for the two Eulerian approaches and symbols for the discrete values of
the separated-fluid approach). Make a similar plot for St
D
=1. Discuss the differences
among the three techniques for these two conditions.


229

6. Point-Force Components

This chapter will discuss the individual components (drag, lift, etc.) for the surface-
averaged fluid stresses based on the point-force assumption for an individual particle. The
intention of this section is to account for different multiphase conditions with the aim to
reproduce the integrated surface-force with reasonable fidelity. The components forces are
generally formulated in terms of the surrounding-fluid which is unhindered by the presence of
the particle under consideration. As such, these components may be used to integrate the
particle equation of motion. However, collision effects associated with a nearby individual
particle or a wall are considered in terms of instantaneous and discontinuous changes in the
particle momentum.
Some of the surface-force expressions can be based on direct theoretical results while other
formulations are given as empirical descriptions of analytical functions, experimental data sets,
or resolved-surface simulations (RSS). However, the various formulations will often be only
applicable (or validated) for a specific range of particle-fluid conditions. In addition, there are
still several complex flow regimes where no force formulations can be recommended with
confidence, e.g., no accurate models are available for instantaneous drag and lift for irregular
particle shapes. As such, it is not possible to identify a single point-force expression of motion
with surface-averaged expressions that is robust for all conditions!
Within this chapter, a baseline point-force expression is first considered for relatively
simple conditions to serve as a reference for more complex conditions. This is followed by
separate discussions for the effects of fluid dynamic surface forces and individual contact
forces on a particle equation of motion. In particular, drag force, lift force and torque,
unsteady forces, particle-wall and paired particle-particle interactions, etc. are considered, each
for a variety of conditions but for an isolated particle. The final section of this chapter then
considers effects associated with flow fields of finite particle volume fractions. It should be
noted that only fluid dynamic and contact effects will be discussed, whereas issues associated
with chemical reaction, ionization, magnetic fields, electric fields, photonic fields, van der
Waals forces, etc. will not be considered in this text and the reader is referred to Clift et al.
(1978), Soo (1990), and Crowe et al. (1998).


6.1. General Decomposition of Point Forces
6.1.1. General Surface Point-Force Expression

The fluid dynamic surface force (F
surf
) for a single particle is generally represented as a
linear sum of several individual and independent forces. As noted in Eq. 1.25, this includes
drag (F
D
), lift (F
L
), virtual mass force due to relative acceleration (F

), far-field fluid-stresses
acting on the particle surface (F
S
), history force (F
H
), Brownian force (F
Br
) and the surface
force due to thermal gradients (F
T
). If we examine the deterministic portion of this equation,
the general expression for F
surf
and F
int
(Eq. 3.146) becomes
230


F
surf
= F
D
+ F
L
+

F
S
+

F

+

F
H
+

F
T
= F
int
+

F
S
6.1
The stochastic effect of Brownian motion which will not be discussed in Chapter 6 since it is
convenient to separate this force out for many numerical techniques. In particular, it will be
incorporated through concentration diffusion in 7.1 and as a random force in 7.2. As such,
the thermophoretic force discussed in Chapter 6 is considered to only incorporate the mean
effect of molecular interactions. The RHS of Eq. 6.1 has three important caveats (similar
statements can be made with respect to the fluid dynamic torque acting on a particle):
1) Linear decomposition is not guaranteed except for some simple conditions which
permit theoretical treatment (two well-known examples are treated in the next section).
Thus, the linear decomposition is not necessarily robust as there can be non-linear
interactions between the various forces. Such interactions are not well understood but
are often small enough to be neglected for many conditions.
2) Many force expressions are empirical. This arises because theoretical treatments are
often limited to small particle Reynolds numbers. Similarly, complex fluid dynamics
(e.g., turbulence) or complex geometries (e.g., irregular particles) will generally
necessitate using experimental or resolved-surface simulation results to validate or even
determine reasonable point-force expressions.
3) The point-force treatment assumes that continuous-fluid gradients (over the particle
length scale) are small or negligible. The quasi-steady drag force generally neglects
spatial gradients while lift, added-mass, history forces, fluid-stress and thermophoresis
forces are based on weak spatial gradients. Since the continuous-fluid properties are
based on extrapolated to the particle centroid, the occurrence of high gradients or non-
linear flow around the particle may not be reasonable with a point-force approximation,
such that distributedforce or resolved-surface techniques may instead be needed (see
4.1.3).
To numerically address conditions where these caveats become problematic, resolved-surface
simulation (RSS) techniques are requires and are discussed in Chapter 8.

6.1.2. Theoretical Point-Force Momentum Equations

There are many analytical expression of the surface force, starting from the seminal work
of Stokes in 1851, which produced Eq. 1.46. The conditions which permit a full theoretical
solution typically include spheroidal shapes at creeping flow conditions (Re
p
0) or inviscid
flow conditions (
f
=0). In particular, there are two important particle equations of motion for a
solid sphere of uniform mass: the Maxey-Riley equation for creeping irrotational flow and the
Auton-Hunt-Prudhomme equation for inviscid rotational flow. These two equations often
serve as baseline equations of motion to which additional effects are incorporated.
The Maxey-Riley equation (1983) assumes a non-rotating rigid particle in an unsteady,
incompressible flow. The relative particle velocity is allowed to be finite but small enough in
terms of Reynolds number that non-linear convective terms can be neglected (similar to Eq.
A..64). The flow is also allowed to be non-uniform but with weak spatial gradients so that they
can be considered small perturbation to the overall flowfield. These latter assumptions can be
summarized in terms of the relative velocity to the continuous-phase and its gradients:
231


p
f
@p 2
@p 2
@p 2
Re 1
u
d
u
u
d

<
<
<

6.2a

6.2b

6.2c
Recall that the continuous-fluid properties at the particle locations ignore local disturbances
caused by the particle and are imaginary since the surrounding fluid does not physically exist
inside the particle. As such, u
@p
is a helpful mathematical representation but is not a
physically measurable quantity.
The Maxey-Riley was important in that it extended the so-called BBO equation to non-
uniform flows and eliminated some inconsistencies of previous equations. As discussed in
1.5.4, the BBO equation was based on earlier work for uniform, but unsteady, flow past a
particle and stemmed from contributions by Basset (1888), Boussinesq (1903) and Oseen
(1927). The Maxey-Riley derivation is more complex than the Stokes drag derivation of 1.5.2
and the final result using the decomposition of Eq. 6.1 is given by:


D f
@p
S f p
2
2
@p
f p
t
3 2
H f f
2
0
2
2
@p
3 d
t
d
40
c
t
(0)
d
t t
d
24

=

=




=


= +





D
D
d
d
d d
d
F W
u
F g
w - u
F
W/ W
F
W w - u

6.3a

6.3b


6.3c


6.3d

6.3e
Note that the added mass coefficient ( c

) for Eq. 6.3c is based on the spherical shape and


that the finite W(0) in Eq. 6.3d is based on later work by Maxey (1993) and was not in the
original Maxey-Riley derivation (since this was assumed to be zero). At creeping conditions,
the effect of linear flow non-uniformity was found to be negligible (i.e. no lift force) but
second-order gradients did make a contribution. The corrections associated with these
gradients (represented by
2
u
@p
) are often called Faxen terms, since Faxen (1922) was first to
obtain such contributions. For example, the above quasi-steady drag (F
D
) is similar to the
Stokes drag of Eq. 1.45 except that it includes a flow non-uniformity term, which is
incorporated by defining a Faxen-corrected relative velocity W (Eq. 6.3e). The fluid-stress
term (F
S
) is the same as that given by Eq. 1.69 (since this stems directly from Eq. 1.68), but the
added-mass force (F

) and the history force (F


H
) differ from Eqs. 1.73 and 1.75 due to Faxen
corrections.
For uniform flow and no initial relative velocity, the above equation reverts to the BBO
equation given in Eq. 1.76. An important point about the Maxey-Riley equation is that the
232
above combination of forces was obtained within a single theoretical framework. Thus, the
above decomposition is exact and complete within the bounds of the above stated conditions.
Auton (1987) and Auton et al. (1988) considered a much different limit: a sphere moving
relative to an inviscid but rotational fluid. These studies assumed a fixed particle subjected to
a far-field linear shear flow (Fig. A.3a) and employed slip conditions on the particle surface
(Eq. 4.54). Auton assumed the spatial velocity gradient to be weak compared to change in
relative velocity across the particle. To extend this to unsteady and straining flows, Auton et al.
additionally assumed that the temporal velocity gradient was weak. These two assumptions
can be written as


shear
2
w
d
w
t d

<
<
w


6.4a


6.4b
As was shown with the simpler case of uniform steady flow of Eq. 1.50a the inviscid
assumption leads to zero net drag on the body (F
D
=F
H
=0). Including the hydrostatic effect, the
surface force has the form F
surf
=F
S
+F

+F
L
where these first two components are given
(neglecting higher-order terms) as:


@p
S f p
@p
f p
t
c
t t


=



=


u
F g
u
v
F
D
D
D
d
D d

6.5a


6.5b
These authors noted that the lift contain both an inertial force (related to the added-mass effect)
and a rotational force. For a sphere in inviscid flow, the combination was found to be


( )
1
L f p
2
= F w
6.6
The sum of forces given by Eq. 6.5 and 6.6 can be called the AHP particle dynamic equation
since the authors were Auton, Hunt and Prudhomme. As with the Maxey-Riley equation, it is
a complete and consistent result for the stated assumptions.
It is interesting to compare the AHP equation to the Maxey-Riley (or the BBO) equation.
We note that the fluid stress is the same in both cases, so that one may expect Eq. 6.3b to be
generally robust for spherical particles. The added mass terms are slightly different between
Eq. 6.3c and 6.5b, though Maxey (1993) noted that the differences are negligible in the
creeping flow limit if
2
u
@p
is negligible (i.e. weak gradients). For the condition of small but
finite Reynolds numbers, Coimbra and Kobayashi (2002) and others obtain the fluid-stress
force of Eq. 6.5b, so that form may be assumed as reasonably robustness. The remaining
forces components are quite different in that drag and history effects are included in the
creeping flow limit whereas only lift is included for the inviscid limit.
In order to address conditions which are inconsistent with both the Maxey-Riley or AHP
assumptions (e.g., finite Reynolds numbers, non-spherical shapes, deformability,
compressibility, etc.), the individual force terms are generally modified. In such circumstances,
the drag force (F
D
) is defined as acting in the direction of -w, while the lift force (F
L
) for zero-
233
rotation is defined as acting in the direction
shear
w . As will be discussed in the rest of this
chapter, there are a wide variety of analytical and empirically derived force terms which have
been developed to address such conditions. The theoretical expressions are often limited to
small-perturbations and simple conditions while the empirical expressions allow for more
complex conditions but are subject to the experimental uncertainties and bias on which they are
based. In either case, the point-force expressions are typically limited to specific regimes and
are generally obtained independently (based on the assumptions of Eqs. 6.1 and 6.2) so that
force interactions are neglected. Thus, the three caveats mentioned in the previous section
should be kept in mind, i.e. one must always be aware of the assumptions upon which an
individual surface-force expression is obtained to determine the extent of its validity.


6.2. Drag Force

Since the drag force is typically the most important surface-force, it will be discussed first.
As discussed in 1.5, the general expression for drag force can be written in terms of a
correction to the Stokesian drag (f) or in terms of a drag coefficient (C
D
)


( )
2
D f f D
3 d f / 8 d C = F w = w w
6.7
Recall that f is more convenient for low particle Reynolds numbers while C
D
is more
convenient for high particle Reynolds numbers, and both are simply related by


D
p
24f
C
Re
=
6.8
As such, f=1 corresponds to spherical solid particles in a uniform incompressible continuum
creeping flow based on Eq. 1.46, while f1 represents a correction to the Stokes drag. For
example, f
Re
denotes a correction due to finite particle Reynolds number, i.e.

D p
D
Re
D p f
F (Re )
F
f
F (Re 0) 3 d w
=


6.9
Similarly, f
shape
denotes a correction due to non-sphericity via the normalized area ratio, f
*

denotes a correction due to finite viscosity ratio, f
Kn
denotes a correction due to non-
continuum effects via the Knudsen number, f
M
denotes a correction due to compressibility via
the Mach number, etc. While some of these corrections can be combined as a linear product,
e.g., f(Re,Kn)=f
Re
f
Kn
under some conditions, this may not be appropriate in many cases, e.g.,
f(E,)f
E
f

. Expressions for point-force drag for a variety of flow conditions will be


discussed in 6.2, beginning with Reynolds number effects.


6.2.1. Spherical Solid Particles

Flow over a Solid Sphere

234
The most straightforward particle condition to consider is that of a smooth isolated solid
sphere with no rotation surrounded by an incompressible fluid. To consider the rationale for
employing various drag expressions, it is helpful to consider the corresponding flow physics.
The flow about the particle can vary from fully attached to massively-separated conditions as
shown in Fig. 6.1. The separated flow is initially laminar, and then becomes transitional and
finally turbulent. For the transitional and turbulent conditions, the separation point is
controlled by downstream vortices due to instabilities of the shear layer, which is often referred
to as vortex shedding. For transitional flow, the shedding frequency increases (normalized by
the free stream velocity and particle diameter) and moves closer to the surface as Re
p
increases
(Clift et al. 1978). The shedding reaches the surface itself at about Re
p
of 6000, where the
shedding frequency (normalized by w/d) is maximized. For 6000<Re
p
<2x10
5
, the separation
occurs at a point which rotates around the sphere at the shedding frequency yielding coherent
structures. If one considers the broad range of Reynolds number regimes for such a particle,
the progression can be summarized as follows:
1) Re
p
1: creeping flow where convective terms are weak or negligible (Fig. 1.24a)
2) Re
p
<22: laminar steady attached flow but significant convective terms (Fig. 6.1a-b)
3) 22<Re
p
<130: laminar separation with laminar wake (Figs. 1.24b, 6.1c-f)
4) 130<Re
p
<1500: laminar boundary-layer/separation with unsteady (transitional) wake
5) 1500<Re
p
<250,000: laminar boundary-layer/separation with turbulent wake (Fig. 1.24c)
6) Re
p
>250,000: turbulent boundary-layer/separation with turbulent wake (Fig. 1.24d)
Flows over particle which are nearly spherical and nearly solid will generally progress through
these same regimes, though the Reynolds number associated with each regime may differ.


Stokes Correction for a Solid Sphere

The variation in particle drag coefficient for these regimes is shown in Fig. 6.2. In the
creeping flow limit f=1 so that the drag coefficient varies inversely with particle Reynolds
number (Eq. 6.8). To consider the drag at small but finite particle Reynolds numbers, Oseen
(1910) included the effects of the convective terms in a linearized fashion (rather than
neglecting them altogether as in the creeping flow condition). In particular, the surface force
was derived with the linearized momentum equation (Eq. A.64). The justification for this
linearization is that the convective terms for small Re
p
become significant (compared to the
viscous terms) primarily at large distances from the body (where U approaches U

). The
analytical solution is discussed by Clift et al. (1978) and the resulting surface pressure
coefficient (defined by Eq. 1.49) is given by:


2
p
p
6cos 1 3cos
C
Re 2 2

= +
6.10
This can be compared to the Stokes solution for surface pressure distribution (Eq. 1.39):


p p
C 6cos / Re =
6.11
The additional contribution of the Oseen result is symmetric about the particle equator so that
the form drag for the Oseen solution is still given by the creeping flow result (Eq. 1.42). On
the other hand, the convective terms lead to an increased surface shear stress so that a higher
friction drag occurs. As a result, the Stokes correction for the Oseen drag is:
235


f
Re
= 1 + (3/16) Re
p
for Re
p
1 6.12
While derived assuming small Re
p
values, this result turns out to be reasonable for Reynolds
numbers of order unity. For example, the Stokes formulation will underestimate the drag by
more than 7% at Re
p
=0.5, whereas the Oseen drag overestimates it by only 2%. Proudman
(1969) obtained a higher-order approximate solution which is even reasonable for Re
p
values
as large as 4.
At higher Reynolds numbers, convection effects become stronger and empirical
expressions are needed because the flow-field is not analytically tractable. For example, Clift
et al. gives a ten-part empirical curve set for sphere drag up to Re
p
of 10
6
, i.e. the conditions
include creeping flow and up to and past the drag crisis. However, single equations are often
preferred for numerical applications. Using experimental drag measurements for solid spheres
up to flow separation, Clift et al. (1978) suggest the following expression


( ) 10 p
0.82 0.05log Re
p Re
f 1 0.1315Re

= + for Re
p
<20 6.13
There are a large number of curve-fits which have been proposed for the sphere correction data
at higher Re
p
values for which the flow is separated. A common and reasonably accurate for
the drag of particles with laminar and transitional wakes is given by the Schiller-Naumann drag
expression (Eq. 1.54) whose Stokes correction is given by:


0.687
Re p
f 1 0.15Re = + for Re
p
< 1000 6.14
The combination of a laminar boundary layer and turbulent wakes (sub-critical conditions)
reduces the sensitivity of the flow and the drag coefficient with respect to Reynolds number.
Consistent with Newtons drag law, the drag coefficient in this regime nearly levels off to a
constant value which is termed C
Dcrit
and has a value of about 0.4-0.45. The transition from
creeping flow all the way to sub-critical flow can be modeled by the fit of Dallavalle (1948):


2
p
Re
0.4Re
f 1
24

= +


for Re
p
< 2x10
5
6.15
However, the White fit of Eq. 1.52 tends to give a more accurate representation for this same
range of Reynolds numbers. Upon careful examination, there is actually a modest increase in
the drag coefficient for 6,000 < Re
p
< 2x10
5
which is caused by mild forward increases in the
separation point. The White and Dallavalle do not capture the slight rise in C
D
, but this is
captured by the more accurate sub-critical expression is given by Clift and Gauvin (1970):


( )
1.16
0.687
D p
p
p
24 0.42
C 1 0.15Re
42,500
Re
1+
Re


= + +


for Re
p
< 2x10
5

6.16
This expression is accurate to within 6% error of experimental data (Fig. 6.2) and the term in
square brackets demonstrates that this is an extension of the Schiller-Naumann drag correction
of Eq. 6.14. Note that all these expression assumes that there is no mass flux at the particle
surface. If mass flux does occur, e.g. due to combustion, this can serve to reduce or increase
236
the drag depending on the flux direction. This effect is called surface blowing and is
discussed in the next section.
At still higher Reynolds number, the boundary layer becomes turbulent, rendering it less
prone to separation. This phenomenon is associated with a more advanced separation point
(compare Figs. 1.24c and 1.24d), and the resulting improved pressure recovery on the aft
surface causes the drag coefficient to rapidly decrease (below 0.1). This is known as the drag
crisis which occurs at Re
pcrit
. The critical Reynolds number for a smooth sphere in uniform
flow is on the order of 300,000 but even small roughness elements or surrounding-fluid
turbulence levels can cause a significant reduction. Both of these effects are discussed in the
next section.

6.2.2. Influence of Continuous-Phase Compressibility and Rarefaction

The level of compressibility of the fluid can become important at high relative particle
velocities, while the level of rarefaction can become important for small particles with respect
to the molecular mean-free path. In both cases, the effects are generally seen for fluid and
solid particles in gasses (

1), so this section will focus on that condition. The two key
parameters to characterize compressibility are the Mach number and Knudsen number,
respectively. For a gas, the Mach number is based on the surrounding speed of sound


p
g g g
w w
M
a T
=
R

6.93
This assumes a thermally perfect gas with a fixed specific heat ratio () and gas constant (R
g
) as
discussed in A.1. Compressibility will generally be significant when the relative Mach
number is no longer small, e.g., 0.4 or more (i.e. w>140m/s for air at STP). Such conditions
can not be expected based on terminal velocity. Instead, significant relative Mach numbers are
generally due to particles interacting with high-speed flows. Examples include particles in a
plasma spray or a rocket engine with a supersonic exit flow, particles ablating form a surface in
a hypersonic boundary layer, and particles subjected to a shock wave. In these cases,
significant relative Mach numbers can occur for durations on the order of the particle response
time (Sivier et al. 1994). Thus, compressibility effects can be important to provide an
understanding of the particle dynamics in such conditions.
To assess the validity of a continuum surrounding flow, the particle Knudsen number is
defined as the ratio mean-free-path length of the surrounding molecules to the particle diameter
(Eq. 1.19b). This can be written in terms of the particle Mach and Reynolds numbers for an
ideal gas (Schaaf & Chambr, 1958) as


p
m-m
p
p
M
Kn
d 2 Re

=



l

6.94
The magnitude of the Knudsen number leads to a range of conditions as noted in Table 6.5.
Small values of the Knudsen number (<10
-3
) are consistent with a continuum no-slip
approximation whereby a very large number of molecular collisions occur at the particle
surface. The no-slip criterion can expressed as
237


f
p
f f
Kn 1 or d
a

< > for continuum flow around the particle


6.95
Since gas viscosity and speed of sound typically have a modest variation with temperature, the
continuum approximation is generally reasonable for particles of a significant size, e.g.,
diameters greater than 10 microns in atmospheric conditions. However, larger sizes are
required in low-density gases, e.g., near vacuum conditions or at very high altitudes. When
Kn
p
is no longer small, the flow around the particle can no longer be considered as a
continuum. In this case, the molecules do not have a sufficiently high collision rate with the
particle and each other. As a result, non-continuum effects lead to a difference between the
mean molecular velocity and the mean particle surface velocity, i.e. a partial-slip condition.
This phenomenon is referred to as accommodation. For slip flow, the non-continuum effect
is weak and can be considered a small departure from the no-slip condition (allowing some
theoretical analysis). Knudsen numbers much greater than unity are generally considered
free-molecular flow conditions where the molecules impact and reflect from the particle with
negligible molecular-molecular interactions (theoretical analysis is again possible in this
regime). The transition regime corresponds to Kn
p
of order unity (e.g., Fig. 6.36) and is the
most complex to analyze. Before addressing the drag coefficient associated with the different
Mach and Knudsen number regimes, it is useful to note whether compressibility or rarefaction
dominate the local flow physics.
Based on Eq. 6.94, one may generally consider the drag coefficient to be a function of two
independent parameters (among M
p
, Re
p
and Kn
p
). At very high Reynolds numbers (Re
p
1),
the physics tends to be dominated by compressibility (M
p
) since Kn
p
will be generally small in
these conditions. At very low Reynolds numbers, the physics will tend to be dominated by
rarefaction (Kn
p
), since M
p
will be generally small in these conditions. These two regimes are
illustrated in Fig. 6.37 in terms of an empirically determined drag coefficient for a sphere in
thermal equilibrium, i.e.
p g
T T

= developed by Loth (2008d). These curves (supported by


experimental and numerical results as discussed below) indicate a nexus of the compressibility
and rarefaction effects at a single Reynolds number, i.e.


D p p p
C 1.63 Re =45 M Kn at for all and
6.96
The independence of compression or rarefaction at this condition (for thermal equilibrium) was
first noted by Zarin (1970). This condition corresponds to a particle with a steady separated
wake in the incompressible continuum limit. Unfortunately, the corresponding flow is not
analytically tractable. However, the nexus condition can be exploited to provide a robust
prediction of the drag by considering separate models for Re
p
<45 (where rarefaction effects are
know to dominate since typical M
p
values are small) and for Re
p
>45 (where compressibility
effects are known to dominate since typical Kn
p
values are small). In the following, the
rarefaction-dominated regime is first considered followed by the compression-dominated
regime.


Rarefaction-Dominated Regime

The effect of rarefaction becomes important for finite Knudsen numbers and low Reynolds
numbers as discussed above. When the flow is in continuum (Eq. 6.95 satisfied), the drag
force for a spherical particle given by the Schiller-Naumann expression of Eq. 6.14 is
238
reasonable (for Re
p
<45). For finite Knudsen numbers but small Reynolds numbers, a Stokes
correction can be defined as


D p p
Kn
f
F (Kn , Re 0)
f
3 d


6.97
This is often referred to as the Cunningham (1910) correction factor and depends on the
surface properties. For a smooth particle, the angles of incidence and reflection at the
molecular length scales will be equal (specular reflection) and the tangential momentum will
be conserved. For a rough particle, the normal velocity component is assumed zero while the
molecular reflection will be random (diffuse reflection) and some portion of the tangential
momentum will be lost. The accommodation coefficient is defined as the fraction of molecules
which undergo diffuse reflection upon impacting the particle (whereas the remaining fraction
undergoes specular reflection).
For small but finite Knudsen numbers, the normal boundary condition is simply given by
U
r
=0 at r=r
p
if there is no mass transfer. However, the tangential condition can be expressed in
terms of the velocity gradients, the surface temperature gradient, and a tangential momentum
coefficient (c

) as given by Talbot et al. (1980):




p
p
r
m m
r r
r r
T
U U 1
U c r c
r r r T


=
=


+








g g
g g
l
R

( )
accom accom
c 2 c / c


6.98a

6.98b
The first term on the RHS of Eq. 6.98a is proportional to the surface shear stress (in square
brackets) but becomes negligible if the mean-free path of the molecules is relatively small
(continuum flow). For finite Kn
p
values this term can be significant and is proportional to c

,
and thus the accommodation coefficient. The latter is constrained by c
accom
1 and has a
typical value of 0.9 on most surfaces, which corresponds to c

1.22, though this coefficient


can range from 1.0 to 1.5. Talbot et al. (1980) also pointed out that the c

value used for the


velocity term may be different from that used for temperature term. However, the difference is
typically quite small (about 3%) so that a single value of c

, as used herein, is generally


reasonable.
Assuming weak temperature gradients so that the second term of Eq. 6.98a can be
neglected, Basset (1888) theoretically obtained a Knudsen correction factor for slip flow:


2
Kn p p p
f 1 2c Kn + (Kn ) Kn 1

= < for O
6.99
In the other limit of free molecular flow (
p
Kn 1 > ), Epstein (1929) obtained a theoretical result
by assuming the velocity distribution of the molecules striking the sphere is based on a
Maxwell distribution (and unaffected by the presence of the sphere) yielding


Kn p
p
8 2 /(c 1)
f Kn 1
36Kn

+ +
> for
6.100
Combining this correction factor (which is inversely proportional to Kn
p
) with its definition
(Eq. 6.97) and Eq. 6.94, indicates that the drag force is independent of viscosity in the free-
molecular flow limit. This is because the molecular velocity distribution is no longer
Maxwellian, since binary collisions are no longer the dominant process for exchange of energy
between the gas molecules. Note that some of the early studies defined Knudsen number based
in particle radius, but generally it is based on particle diameter, as is the case for this text.
239
These two definitions led to a factor of two typos in the above equation (and some of the below
equations) in the equations given by some references.
To bridge the gap from very low to very high Knudsen numbers, Phillips (1975) proposed
an approximate theoretical solution. This assumed low Re
p
and M
p
but allowed for
intermediate Knudsen number values and obtained


2 2 2
p p
Kn 2 2 2 2 3
p p p
15c 6c Kn (16 16c 4 )(c 2)Kn
f
15c 24c Kn 36c (c 1)Kn 72(c 2)(c 1)Kn


+ + + +
=
+ + + + + +

6.101
A reasonably accurate empirical representation which is consistent with c

1.22 and the


above equation is that discussed by Clift et al. (1978) :


Kn
p p
1
f
1 Kn 2.514 0.8exp( 0.55/ Kn )
=
+ +


6.102
In the limit of particle Reynolds and Mach numbers much less than unity, both the theoretical
and empirical expressions (Eqs. 6.101 and 6.102) show that the slip effect results in decreasing
equivalent drag as the Knudsen number increases. As shown in Fig. 6.38, both expressions
also compare favorably with the experimental data of Milliken (1911, 1923) for which the
charge of an electron was first measured! Also shown in this figure is the Direct Simulation
Monte-Carlo (B.8) data of Benson et al. (2004) which follows the trends as well. To account
for finite Reynolds numbers, the above Knudsen number effect can be extended by including a
Schiller-Naumann correction as


( )
p
0.687
D,Kn,Re Kn p
p
24
C f 1 0.15Re
Re
= +
6.103
This expression, which will be shown to be quite robust, can use either a theoretical or an
empirical Knudsen correction (e.g., Eq. 6.101 or Eq. 6.102), but assumes the particle Mach
number is small. Note that use of the drag coefficient form, instead of the drag correction form,
is more appropriate in free-molecular flow for which the drag is no longer proportional to
viscosity.
As the particle Mach number becomes significant, e.g., of order unity or more, a free
molecular creeping flow limit can be obtained in terms of the molecular speed ratio:


p
s M / 2 6.104
Assuming equal tangential and normal accommodation coefficients, Stadler & Zurick (1951)
and Schaaf & Chambr (1958) derived the free-molecular compressibility limit for s1 for


( ) ( ) ( ) ( )
2 2 4 2
p
D,fm 4 3
1 2s exp s 4s 4s 1 erf s
T
2
C
2s 3s T s

+ +
= + +

6.105
The first two terms on the RHS are associated with diffuse reflection effects while the third
term is associated with specular reflection and includes the effect of the particle temperature
ratio. This ratio is based on T

which is the far-field gas temperature as seen by the particle, i.e.


the uncoupled gas temperature hypothetically extrapolated to the particle centroid (T
f@p
). This
is normally equal to the particle temperature for thermal equilibrium conditions, but T
p
may be
different than the ambient temperature if the particle is undergoing rapid heating or cooling
(e.g., due to injection conditions, shock passage, etc.). Hersch et al. (1969) noted that the form
of Eq. 6.105 is equal to the Epstein (1929) model if one assumes that the thermal
240
accommodation coefficient is equal to the momentum accommodation coefficient. The free
molecular drag coefficient for equilibrium particle temperature is defined as


( ) ( ) ( ) ( )
p
2 2 4 2
D,fm D,fm,T T 4 3
1 2s exp s 4s 4s 1 e rf s
C C
2s s

=
+ +
= = +


6.106
In the limit of very high speed ratios, this gives
D,s 1
C
>
=2, which is often taken as the lower
limit for creeping flow.
The free molecular limit can be empirically corrected for finite Reynolds numbers by
ensuring that the nexus given by Eq. 6.96 is satisfied (regardless of particle temperature) with
an drag coefficient expression proposed by Loth (2008d):


D,fm
D,fm,Re
p D,fm
C
C
Re C
1 1
1.63 45
=


+



6.107
In this case, the limit of very high speed ratios and equilibrium particle temperature but finite
Reynolds numbers becomes
( )
1
D,s 1,Re p
C 0.5 0.0169 Re

= +
>
, which is often taken as the lower
limit for low Reynolds number drag. This limit and the creeping free-molecular limits of Eq.
6.106 are shown in Fig. 6.39, where it can be seen that there is a substantial difference between
the free-molecular values and the continuum Clift-Gauvin drag coefficient, especially at small
particle Reynolds numbers.
While the free-molecular compressibility limit is appropriate for s1 (including hypersonic
conditions), it becomes undefined (infinite) for the zero Mach number limit. To additionally
blend the finite Mach number limit with the zero Mach number limit, the following empirical
combination may be used (Loth, 2008d):


4
p D,fm,Re D,Kn,Re
D,fm,M,Re p 4 4
p p
M C C
C Re 45
1 M 1 M
= +
+ +
for
6.108
This expression represents the overall fit to the rarefaction regime and simultaneously allows
for Reynolds, Mach, Knudsen, and temperature ratio effects for the rarefaction dominated
regime. For the low Mach and Reynolds number conditions of Fig. 6.39, this reverts to the
theoretical creeping limits of Eqs. 6.99 and 6.100 for very small or very large Knudsen
numbers. For the finite Mach number conditions, this model correctly asymptotes to the free
molecular limit (shown by the dotted lines in Fig. 6.39) including the limit of s1 where
D
C 2 . This model also properly represents Reynolds number effects as shown by
comparison with experimental data and numerical data in Fig. 6.39. The latter is based on
Direct Simulation Monte-Carlo (DSMC) whereby representative molecular interactions are
employed to make statistical predictions (B.8). Because of these attributes, the present model
consistently outperforms those of Carlson & Hoglund (1964), Crowe (1967), Henderson (1976),
and Hermsen (1979) in the rarefaction-dominated regime, e.g., as shown in Fig. 6.40.
Temperature ratio effects are also included in the present model via the specular term of Eq.
6.105. As shown in Fig. 6.41a, this leads to increases in drag as the particle temperature
increases due to an additional viscous contribution. The effect is highest at the creeping
compressible flow limit (where the model reverts to Eq. 6.105) and reduces as the Re
p

increases. For all these conditions, the present drag model gives the best prediction of the
temperature influence and a typical comparison with previous models is shown in Fig. 6.41b.
241
Beyond the nexus point (in the compression dominated region), little data is available to
support whether the temperature increases or decreases the drag but in general the effect of
p
T / T

is small (Aerosty, 1962; Bailey & Hiatt, 1972). The next section discusses the
influence of Mach and Reynolds numbers in this regime.

.
Compression-Dominated Regime

Various compressible flow features will determine the flowfield over a sphere depending
on the Mach number (Fig. 6.43). The compression-dominated regime generally yields
increased drag as the Mach number increases (recall Fig. 6.37), which can be quantified in
terms of a drag ratio defined as:

D,crit
M
D,crit ,M 0
C
C
C
=

6.109
As with the drag ratio for shape effects, the baseline (incompressible) critical drag coefficient
is taken as 0.42.
The increase in drag is initially due to the influence of compressibility on boundary layer
separation for M
p
>0.2 (Schlichting, 1979). The compressibility is generally most significant at
the point with the highest velocity, which is typically just above the boundary layer at =90
o

according to potential flow theory (Eq. 1.47c). As the Mach numbers increases the maximum
flow velocity at this point will eventually reach a sonic speed. The free-stream Mach number
associated with this condition is defined as the critical Mach number (M
crit
), about 0.6 for a
sphere (Hoerner, 1965). Beyond this speed, gas dynamic waves appear. For example, flow at
a subsonic but super-critical Mach number is shown in Fig. 6.42a, where one can see weak
expansion waves as the flow becomes locally supersonic, followed by a lambda-shock pattern
at the top of the sphere. This boundary layer shock interaction promotes earlier flow
separation than that seen for the incompressible condition (Fig. 1.24d). As a result, the drag
significantly increases for M
p
>M
crit
(Fig. 6.43).
As the Mach number becomes supersonic, a bow-shock forms in front of the sphere (Fig.
6.42b) but the point of flow separation moves further back since there is no longer a lambda-
shock present. The corresponding drag initially increases until it is maximized at about M
p
of
1.5 (Fig. 6.43). At higher Mach numbers, the flow retains a bow-shock but the stand-off
distance is reduced and boundary layer separation is further delayed due to supersonic
expansion over the rear portion (Fig. 6.42c). As a result, the drag ratio is somewhat reduced
and eventually approaches a constant at hypersonic Mach numbers,
M
C 2

(Fig. 6.43).
Since the rear flow separation is primarily dictated by the gas dynamics, the flow and the drag
coefficient are not sensitive to Re
p
for M
p
>1.5, i.e. there is no drag-crisis when the boundary
transitions from laminar to turbulent conditions (Fig. 6.2). As a result, Crowe (1967) termed
the supersonic C
D
as the inviscid drag coefficient because boundary layer properties have
little influence. There is no theoretical correction for Mach number effects at these conditions
so the drag ratio must be described empirically, and can be modeled as a two-part equation:
242


( )
( )
{ }
M p p
2
M p p
5 2
C tanh 3ln M 0.1 M 1.45
3 3
C 2.044 0.2exp 1.8 ln M /1.5 M 1.45

= + +


= +

for
for

6.110a

6.110b
This is similar to the fit given by Crowe et al. (1972) and follows the collected data in Fig. 6.43.
In order to represent the influence of Mach number between the nexus point (where the
effect is negligible) and the high Re
p
regime (where the effect is given by Eq. 6.110), it is
convenient to employ a modified Clift-Gauvin drag expression in the following form


0.687 M
D p M
M
p
1.16
p
0.42C 24
C 1 0.15Re H
42, 500G
Re
1+
Re

= + +

for Re
p
>45
6.111
This expression involves the critical drag coefficient ratio defined in Eq. 6.109 but also
introduces two empirical functions of the particle Mach number given by Loth (2008d) as:


( )
4
M p p
M p p
M
M
M
G 1 1.525M M 0.89
G 0.0002 0.0008tanh 12.77 M 2.02 M 0.89
0.258C
H 1
1 514G
= <

= + >

=
+
for
for
6.112a

6.112b

6.112c
The form of G
M
is constructed to ensure that the incompressible Clift-Gauvin limit is recovered
as M
p
approaches zero while H
M
is constructed to ensure that the nexus of Eq. 6.96 is satisfied.
As shown in Fig. 6.44a, this model is more robust than previous models in terms of predictive
performance, which was true for a wide range of M
p
. In general, the variation in drag
coefficient becomes weaker as the Mach number increases (Figs. 6.37 and 6.44b). Finally, it
should be noted that all of the above compressibility and rarefaction effects assume spherical
shapes, since little is known about the orientation dynamics and drag of non-spherical objects
in free trajectories.


6.2.3. Influence of Mass Flux, Turbulence and Particle Roughness

Effect of Surface Mass Flux

An aspect which can affect the particle drag is a finite mass flux at the surface. It is
generally most significant for particles in a gas where it has been termed blowing since rapid
evaporation or sublimination can lead to significant radial gas velocities. For example, small
combusting kerosene droplets can have radial out-gassing velocities on the order of the relative
velocity. The surface fluid velocity (based on the mass flux) and the particle relative velocity
can be used to define a dimensionless mass flux velocity which can be simplified as an velocity
integral for constant density as:
243

( )
r,m surf
p *
m
2
surf f
u dA
m
u
wA w d
=

`
`
`

6.119
The flow solution for flow around a sphere with Re
p
0 was obtained by Dukowicz (1982) for
surface mass flux at isothermal conditions with uniform viscosity as a function of the mass-
flux Reynolds number, defined as follows:

*
p m p
f
m
vapor f
Re u m
Re
2 2 d

=

`
`
`

6.120
For arbitrary mass-flux Reynolds number with either blowing (
m
Re
`
>0) or suction (
m
Re
`
<0),
the corresponding creeping-flow drag correction for isothermal mass flux becomes:

( )
( ) ( )
m
m
Re
m
m m 2
1 Re
m m
2
1 1 Re e
2
f Re
3
1 Re e 1 Re


+
=
+ +

`
`
`
` `
` `

6.121
For small mass fluxes, this tends to
m m
f 1 7Re / 24
` `
, consistent with a drag force of
p
m w ` .
A similar low Re
p
result was obtained by Eisenklam et al. (1967) but based on an Oseen-type
drag and a prescribed vapor concentration distribution and tends to
m m
f 1 13Re / 24
` `
for small
mass fluxes. However, experiments also obtained by Eisenklam et al. for both evaporating and
bringing drops with Re
p
in the range of 0.1 to 1 yielded trends closer to
m m
f 1 Re
` `
assuming
a viscosity for Re
p
based on an average of far-field gas and local vapor viscosities; a result
consistent with later measurements and predictions (Sirignano, 1999). All these forms show
that outward (blowing) surface velocity will reduce the effective drag while inward (suction)
velocity will cause a drag increase. This is because blowing tends to remove vorticity from the
surface thus reducing the net friction drag on the particle.
At Re
p
>1, the mass flux from the particle surface tends to be directed downstream and thus
can be interpreted as a thrust once a far-field control volume is employed, again giving rise to
reduced drag for blowing conditions. For isothermal constant-viscosity conditions, resolved-
surface simulations with blowing by Niazmand & Renksizbulut (2003) and Cliffe & Lever
(1985) for Re
p
up to 500 were reasonably represented by the following empirical relationships:

( )
Re,m Re m
D
D,m 0 Re Re
f f f 1
C
C f f
=
+

` `
`
for Re
p
10 and
m
Re 5
`

( )
-0.2
Re,m * D
m
D,m 0 Re
f
C
1 20u
C f
=
= +
`
`
`
for 10Re
p
500 and
m
Re 5
`

6.122a


6.122b
However, mass transfer tends to occur when there is temperature difference in which cases, the
above relationships may no longer be reasonable and heat transfer becomes important to
describe the detailed physics. There have been several empirical relations proposed based on
resolved-surface simulations as discussed by Sirignano (1999), but often these results do not
tend correctly to the solid particle drag expression for zero blowing. However, the
experimental drag data of Eisenklam et al. (1967) and Yuen and Chen (1976) for both
evaporating and condensing drops in gasses are reasonably correlated up to Re
p
of 500 by
244
using the empirical 1/3-rule. This rule employs an evaporating gas temperature (
m
T
`
)
which is 1/3 of the unhindered gas temperature (which does not take into account the
droplets presences) and 2/3 of the droplet surface temperature.

m f p,surf
1 2
T T T
3 3
+
`
6.123
Similarly, an evaporating gas mixture is defined based on an evaporating vapor mole fraction
(
m

`
) equal to 1/3 of the far-field vapor-fraction and 2/3 of the drop surface vapor-fraction.

m f p,surf
1 2
3 3
+
`
6.124
These results can then be used to compute an evaporating viscosity,
m

`
, equal to the volume-
weighted viscosities of the species of the evaporating gas mixture taken at the evaporating
temperature (e.g., via relations such as Eq. A.16). This evaporating viscosity is then used to
define an evaporating Reynolds number:

( )
( )
f f f
f
p, m p
m m m m
T ,
wd
Re Re
T ,

=

`
` ` ` `

6.125
This Reynolds number can then be used for the particle drag (e.g., via Eq. 6.16 or 6.14), i.e.

( )
( )
D p, m
Re,m
D
Re D,m 0 D p
C Re
f
C
f C C Re
=
=
`
`
`

6.126
Thus, evaporation tends to reduce the effective Reynolds number which therefore reduces drag,
while condensation would yield the opposite trend. Dukowicz (1984) employed resolved-
surface simulations and found that this result yields generally reasonable predictive
performance when extended to conditions with moderate Re
p
values. The predictions with this
simple adjustment are shown in Fig. 6.51 and are reasonable, especially for Re
p
<10. As such,
Eq. 6.126 is a commonly used and reasonably robust empirical relationship for drag
modification for droplets in gasses with significant mass flux. The correlation can be improved
somewhat (Fig. 6.51) by defining a corrected evaporating Reynolds number for Eq. 6.126 as
( )
p,m,corr p,m p, m
Re Re 1 0.007Re +
` ` `

6.127
However, this corrected 1/3-rule expression is generally limited to Re
p
<1000, beyond which
droplet deformation tends to occur which will dominate drag modifications (Fig. 6.28).
As far as combining the above relationships, Niazmand & Renksizbulut (2003) used
resolved-surface simulations to show that the effects of spin and blowing can be approximately
combined as follows:

Re, ,m Re, Re,m
D
D, 0,m 0 Re Re Re
f f f
C
C f f f

= =




` `
`
for Re
p
500,
m
Re 5
`
and
*
p
<1
6.128
However, experimental drag data for rotation combined with blowing or suction is lacking so
that this relationship should be viewed with some caution.
245
Finally, we mention the impact of surrounding flow planar strain. Bagchi & Balachandar
(2002c) noted that the increase in drag for 10<Re
p
<300 with RSS results is at least
qualitatively consistent with the increase associated with potential flow theory, and even
quantitatively in the absence of flow separation (Re
p
<20). While the trends are complex at
higher Re
p
values, the net changes in drag were generally small (on the order of a few %)
indicating that the neglect of strain or other aspects of Mechanism 5 is generally reasonable.


Effect of Turbulence on Drag

The effect of turbulence magnitude and frequency on the drag of particles is a matter of
considerable debate and complexity. There are at least seven mechanisms by which turbulence
may affect particle mean relative velocity with respect to the average fluid velocity, i.e. w. In
particular, turbulence can introduce:
1) non-linear drag bias due to finite rms of the relative velocity (decreases w)
2) preferential bias for heavy particles (increases w) or buoyant particles (decreases w)
3) turbo-phoresis (bias w towards regions of lower kinetic energy)
4) clustering bias which can induce three-way coupling (increases w)
5) local fluid vorticity and/or particle rotation to increase drag (decreases w)
6) non-linear/unsteady flow gradients about the particle (modifies w)
7) reduced drag-crisis critical Reynolds number, especially with roughness (modifies w)
Mechanisms 1-7 are taken into account with full numerical resolution of all flow scales for
both the particle and the flow, i.e. use of RSS and DNS (Table 4.2). However, this combined
approach can be extremely intensive in terms of computational resources and so is only
practical for a few particles in a simple flow at a modest Reynolds number.
For the point-force approach, the focus of this chapter, the ability to capture the first five
mechanisms for is related to the numerical method employed (Table 4.2) and is summarized
below. If one considers the point-force approach in conjunction with DNS, then Mechanisms
1-3 can be taken into account directly since no modeling is required for the flow field details,
e.g., the interactions between the unsteady turbulent structures and the particle trajectories are
fully captured. This can also be true for Mechanism 4 with respect to one-way coupling, but
the increase in volume fraction may also lead to three-way and four-way coupling effects
which requires modeling (3.8, 6.7). The impact of vorticity or particle rotation on drag
(Mechanism 5), can be empirically modeled at finite Re
p
as noted in the following section
(6.2.7).
If the turbulent structures are unresolved because of the use of RANS (or unresolved down
to the particle response time when using LES), then models are needed to account for
Mechanisms 2-3 (3.5.5). In addition models are needed for Mechanism 1 for an Eulerian
treatment of the particles (3.5.4) and an additional correction is needed for Lagrangian
treatment of particles in non-homogeneous turbulence (7.2.5). Mechanisms 4-5 are difficult
to model within an averaged approach.
The remainder of this section focuses on point-force models for Mechanisms 6 and 7. For
Mechanism 6, consider a particle which is held stationary with respect to a relative mean
velocity but is subjected to velocity fluctuations and turbulent structures of the incoming flow.
Such a scenario is consistent with wind tunnel measurements past fixed objects. An example
of this is shown in Fig. 6.45, where it can be seen that an increase in turbulence (normalized by
246
the relative velocity) resulted in a more unsteady and complex wake (Wu & Faeth, 1994). If
one employs the instantaneous relative velocity given by DNS coupled with a point-force drag
with d< (based on Eq. 4.44), then the impact of this mechanism is small. This is
demonstrated in Fig. 4.14b which compares RSS and point-force descriptions of the surface
force for such a case. However, as the particle becomes larger (d) a distributed-force model
is needed and random fluctuations about the predicted surface force arise as shown in Fig.
4.14a, though the average drag is reasonably represented. A similar result is shown in Fig.
6.46 for a bubble with d, whereby the drag model is reasonable on average, but the
fluctuations are only approximately captured. No robust model has been proposed to account
for the instantaneous discrepancies for a point- or distributed-force description with d.
However, several studies have been conducted to model the effect of Mechanism 6 on the
average drag when a RANS description of the flow is employed. Published data for solid and
liquid particles reviewed by Crowe et al. (1998) indicate a variety of results for the mean effect
of turbulence on a fixed particle, including both increases and decreases in drag. Unfortunately,
Crowe et al. concluded that there is insufficient evidence to propose a quantitative robust
variation of the non-turbulent drag behavior for sub-critical Reynolds numbers, so that
assuming the mean drag (based on the mean relative velocity) is not significantly affected by
turbulence levels for Re
p
<100 is not an unreasonable assumption. In fact, detailed experiments
by Warnica et al. (1995) lend further support to this assumption.

Effect of Turbulence and Roughness on the Drag Crisis


In contrast to many of the above mechanisms, the reduction of Re
p,crit
by turbulence
(Mechanism 7) has received much more attention. This effect arises because increased free-
stream fluctuations introduce instabilities which cause the particles boundary layer to
transition sooner to turbulent flow. A similar effect is caused by roughness on the surface of a
particle. As a result, the drag crisis occurs at lower Reynolds numbers in the presence of
increased turbulence or particle roughness. To quantify the impact on Re
p,crit
(defined when C
D

drops below 0.3), the particle surface variations (
RMS
d ) are normalized by its diameter and the
free-stream turbulence levels
RMS
(u of Eq. A.72) are normalized by the relative velocity:


* RMS
* RMS
RMS
u
u
w
d
d
d



6.113a

6.113b
The roughness effect is shown in Fig. 6.47a for a relatively low turbulence level for
*
RMS
d up to
1.5%, which leads to a four-fold reduction in Re
p,crit
. Conditions with very high surface
irregularities (
*
RMS
d on the order of 15% or higher) have been studied in the context of irregular
particles as discussed in 6.2.4. While very high Re
p
studies for such conditions are scarce,
there is no evidence of any drag coefficients below 0.3 for such particles. Furthermore, such
particles are not likely to have a drag-crisis at all since the highly irregular surfaces will cause
the particles to have turbulent flow over the particle at modest Re
p
values for which the
measured C
D
is approximately constant (e.g., non-circular cross section particles in Fig. 6.8).
247
Unfortunately, there is not sufficient quantitative data at intermediate surface roughness levels
(0.015<
*
RMS
d <0.15) to determine the behavior of the transition.
The effect of turbulence is qualitatively equivalent to that of small roughness levels on the
particle surface. This effect is shown in Fig. 6.47b for particles that are effectively smooth
(
*
RMS
d <10
-4
) for which Re
p,crit
reduces to values as low as 20,000 at turbulence levels of 20%.
On the other hand, turbulence levels less than 1% have little effect on critical Reynolds number.
The combined effect of roughness and turbulence intensity is also shown in Fig. 6.47b, where
it can seen that turbulence has a more profound effect and that the combination gives even
larger reductions in Re
p,crit
. Correspondingly, roughness becomes less important at high
turbulence intensities. An empirical expression which approximately correlates the combined
effects of the data in Fig. 6.47a and Fig. 6.47b (and is shown on both figures) is given by


( ) ( ) ( )
-0.03 0.55 -0.07
* * *
10 p,crit RMS RMS RMS
log Re 4.4 d +0.002 -1.9 u d +0.002 = 6.114
However, this relationship belies a complex influence of turbulence for
*
RMS
u >0.1. In
particular, high turbulence gives rise to a trans-critical regime occurs for which the drag
coefficient twice drops below 0.3 as Re
p
increases. This regime is shown in Fig. 6.48 whereby
the drag coefficient undergoes an initial drop, followed by a rise, followed by a second drop.
While the mechanisms for these rapid changes are not fully understood, Clamen & Gauvin
(1969) suggest the following. At high turbulence but moderate Re
p
values, the laminar
boundary layer separates sooner but then undergoes reattachment creating a separation bubble
as has been noted on some airfoils. The resulting wake turns out to be significantly smaller
which causes the first C
D
drop below 0.3. At a somewhat larger Re
p
value, this laminar
separation bubble fails to reattach leading to a massively separated condition and very high C
D

values. Finally, the boundary layer transition to turbulent before separation and so stays
attached longer, creating a smaller wake and dropping the C
D
below 0.3 again. As such, drag
can be quite sensitive at high turbulence levels, e.g., at turbulence levels of 35% C
D
can change
by an order of magnitude for only a factor of three change in Re
p
. Furthermore, Neve &
Shansonga (1989) reported that the drag coefficient sensitivity to turbulence is also function of
the non-dimensional turbulent length-scales seen by the particle (/d), though it disappears
altogether for d. As such, significant caution is needed when employing the standard sub-
critical drag expressions of 6.2.1 (which assume approximately constant C
D
from
1000<Re
p
<Re
p,crit
) when the turbulence levels are beyond a few %. With respect to large-scale
roughness below Re
p,crit
, there are various theories.

6.2.4. Non-Spherical Solid Particles

The drag of non-spherical solid particles will depend on the degree of non-sphericity as
well as their orientation to the flow (since the drag will generally be anisotropic with respect to
direction). For non-spherical solid bodies, particles may be classified as regularly-shaped
(ellipsoids, cones, disks, etc. as shown in Fig. 2.7b) or irregularly-shaped (non-symmetric
rough surfaces as shown in Fig. 2.8-2.9). The drag associated with these two types of particle
shapes is discussed below, beginning with spheroids and ellipsoids.


248
Spheroids and Ellipsoids in Creeping Flow (Re
p
1)

As discussed in 2.2, ellipsoids are the most commonly characterized regularly-shaped
particle. In particular, the spheroid is defined by an aspect ratio (E) based on the ratio of the
diameters along and about the axis of symmetry (Eq. 2.36). Oblate spheroids will have E<1
with the limiting shape E=0 corresponding to a disk; while prolate spheroids will have E>1
with the limiting shape E= corresponding to a needle (Fig. 2.7a). In both cases E=1
corresponds to a sphere. We may define the Stokes correction for ellipsoid of aspect ratio (f

)
as the ratio of creeping solid spheroid drag force to creeping solid sphere drag force (where
both have the same volume equivalent diameter):

D p
E
f
F (E, Re 0)
f
3 d w



6.17
Thus f
E
approaches unity as E approaches unity. Exact, limiting and approximate solutions for
the drag on spheroids at creeping flow conditions (Re
p
0) were derived by Oberbeck (1876).
In particular, he obtained corrections for flow parallel to the axis of symmetry (
E
f
|
) and flow
perpendicular to the axis of symmetry (
E
f

). These are given in Table 6.1 along with


approximations for small departures from the spherical shape, i.e. |E-1|1, as well as small
departures from the limiting solutions (disk or needle geometries).
The trends for the parallel drag correction factor as a function of aspect ratio are shown in
Fig. 6.3 where the approximate and limiting conditions are shown to be reasonable over a wide
range of aspect ratios in comparison to the exact values. It can also be seen that variations in E
from 0.01 to 100 result in only moderate changes in the overall drag correction (less than four-
fold). This arises because increases in friction drag for longer bodies are approximately
balanced by decreases in pressure drag, while the reverse is true for shorter bodies. It is
interesting that the minimum drag shape and orientation is a prolate spheroid with aspect ratio
E=1.955 moving parallel to the axis of symmetry, where
E
f
|
=0.9555 (Fig. 6.3). However, this
condition is not generally stable and the more commonly occurring broadside orientation has a
drag correction which is greater than unity
Generally, spheroids traveling at an initial orientation with no rotation will continue to fall
at the same orientation, though a broadside orientation tends to be more stable (2.3.1). If the
initial orientation is parallel or perpendicular to the relative velocity, no torque exists since the
drag acts in the opposite direction to the gravitational force. However, it is possible for a
spheroid to be at an orientation which is neither parallel nor perpendicular. In this case, the
drag force may be obtained based on a linear combination of
E
f
|
and
E
f

because of the linear


drag associated with creeping flow, i.e.


( )
D f E E
3 f f

= + F w w
| |

6.18
In this equation, and

w w
|
are the components of relative velocity which are parallel and
perpendicular to the particles axis of symmetry (and whose vector sum is w). If both velocity
components are non-zero, this leads to an interesting result as F
D
will not be parallel to w
(compare Eqs. 6.7 and 6.18). As a result, the terminal velocity direction will differ from the
direction of gravity, i.e. the particle will have a glide angle with respect to g causing a lateral
249
velocity component (White, 1991). One may recast the force of Eq. 6.14 into a drag
component in the relative velocity direction and then define the remaining fluid dynamic force
as lift. However, the force of Eq. 6.14 is often denoted as drag while lift is defined to be the
result of non-zero continuous-phase shear or non-zero relative rotation of the particle (6.3).
Similarly, the drag force for ellipsoids with aspects ratios E1, E2 and E3 (and other
orthotropic particles) will be based on the Stokes correction factor in each direction and the
component of relative velocity associated with each axis direction


( )
D f 1 E1 2 E2 3 E3
3 f f f = + + F w w w
6.19
In Brownian motion, all orientations can be assumed to be equally possible for many
realizations (or long times) since molecular interaction is random. In this case, one may
consider an average correction factor by integrating over all possible orientations. For particles
which are regularly shaped about three axes, this integration yields a correction which is based
on the average of the inverse correction factors (Clift et al. 1978), i.e.:


E E1 E2 E3
3 1 1 1
f f f f
= + +
6.20
Since spheroids have two axes which are equal in length and perpendicular to the axis of
symmetry, this yields
E1 E E2 E3 E
f f and f f f

= = =
|
. If one assumes that Stokesian drag is
appropriate which is reasonable for most small particles in a liquid (whereas small particles in
air undergoing Brownian motion will generally include non-continuum effects as discussed in
6.2.5), then the net correction for oblate and prolate shapes based on Eq. 6.20 and Table 6.1
becomes:


1/ 3 2
E
1
1/ 3 2
E
2
E 1 E
f for E<1
cos E
E E 1
f for E>1
ln E E 1

=

+



6.21a



6.21b
This average drag correction is also reasonable to within a few percent (Davis, 1991) for the
wide range of orientations caused by Jeffrey orbits, which result from shear flow (2.3.2). As
may be expected, the minimum drag shape for a given particle volume when averaging over all
orientations is a sphere (for which f
E
=1).


Other Regularly-Shaped Particles in Creeping Flow (Re
p
1)

Regularly-shaped non-spheroidal particles such as cones, cubes, etc. (Fig. 2.7b) do not
typically have analytical solution for the drag even in the creeping flow limit (Re
p
1).
Therefore, the shape correction values are typically obtained by terminal velocity experiments
comparing a non-spherical particle to a spherical particle with the same volume and thus the
same volumetric diameter (Eq. 1.1). Since the drag is linearly proportional to the relative
velocity for creeping conditions, use of Eq. 1.82a yields:
250

( )
p
D,shape term,shape term,sphere
shape
f term,sphere term,shape
Re 1 & const.vol.
F f w
f
3 d w f 1 w
= =
=
<

6.22
As a first approximation, the shapes and corresponding drag corrections of many particles may
be approximated as ellipsoids by determining an effective aspect ratio (E). As shown in Fig.
6.3, this approach works well for cylinders since their shape is quite similar to that of a
spheroid. In particular, the drag increases for cylinders correspond reasonably well to the
limiting cases of spheroidal disks and needles. However, the drag for a cylinder with a length
to diameter ratio of unity is somewhat larger (by 14.4%) than that predicted for a sphere,
primarily due to the increased surface area for the cylinder. This difference becomes more
profound if the particles have additional edges or indentations as this causes extra surface area
(A
surf
) for the same volume. One may also from the spheroidal drag relations that the projected
cross-sectional area (A
proj
) with respect to the direction of the particles relative velocity would
be important for non-spherical drag effects.
As such, to estimate the shape factor for non-spheroidal regularly-shaped particles, it is
common to consider two dimensionless area parameters: the surface area ratio and the
projected area ratio. Each of these areas can be normalized by the surface area of a sphere
which has the same volume, i.e.

proj * * surf
surf proj 1 2 2
4
A
A
A , A
d d



6.23
The surface area ratio will always be greater than one since the spherical geometry has the
minimum surface area for a given volume, i.e.,
*
surf
A 1. The inverse of the surface area ratio is
more commonly defined as the sphericity ratio (Wadell, 1934), the sphericity (Clift et al.
1978), or the shape factor (Crowe et al. 1998). For a cylinder with a length to diameter
ratio defined by E
cyl
, geometric relations can provide the surface area ratio and equivalent
volume diameter:

( )
1/ 3
cyl cyl cyl *
cyl surf cyl 1/ 3
2
cyl
cyl
L 2E 1 3E
E , A , d=d
d 2
18E
+
=



6.24
The projected area ratio will depend on the orientation of the particle as well as its shape. For
example, a long cylinder will have
*
proj
A >1 if it falls broadside, whereas it will have
*
proj
A <1 if
it falls vertically along its axis. Equations for
*
surf
A and
*
proj
A for double-cones, cuboids, etc.
are given as a function of orientation in Clift et al. (1978) for non-isometric conditions. There
are also other measures of particle non-sphericity discussed by Clift et al. (1978) but these are
the two most commonly used parameters for particles with symmetry and turn out to be the
most effective in correlating the drag (Ganser, 1993; Madhav & Chhabra, 1995; Xie & Zhang,
2001).
In general, one would expect that larger values of
*
proj
A or
*
surf
A would correspond to larger
drag values, and indeed this is the case. Leith (1987) suggested the following correlation of
these two area ratios for the Stokes shape correction factor
251

* *
shape proj surf
1 2
f A A
3 3
= + 6.25
This relationship was based on the fact that one-third of the drag of a sphere is form drag
(related to the projected area) while two-thirds is friction drag (related to the surface area), and
that the form and skin friction drags are proportional to the particle diameter. This ratio holds
reasonably well for many non-spherical particles, especially those with small deviations from a
sphere. In particular, isometric particles with equal sides (cubes, octahedron, tetrahedron, etc.)
have a projected area which is approximately the same for all orientations so that the projected
areas for these shapes can be estimated as unity. Table 6.2 lists some regular particle shapes
along with measured and predicted Stokes correction factors. The relations are generally
reasonable for mild to moderate non-sphericity, but differences become significant for large
number of convolutions or holes in the shape (e.g., the four-sphere cluster).
If the particle can be reasonably approximated as having a surface area close to that of a
spheroid, one can use the corresponding surface area ratio relationships (Clift et al. (978):

( )
2/ 3 4/ 3 2
*
surf
2 2
1/ 3
* 1 2
surf 2/ 3
2
E E 1 1 E
A ln E 1
2
4 1 E 1 1 E
1 E
A sin 1 E E 1
2E
2 1 E


+
= +



= +

for
for

6.26
Unlike the surface area, the projected area depends on orientation. For a spheroid falling at a
broadside orientation, the projected area ratio can be given from Eqs. 2.37 and 2.38 as

* 1/ 3
proj 2
2
* -2/3
proj 2
d d
A E E 1
d
d
A =E E 1
d

= =
=
|
for
for

6.27
Since creeping drag is not highly sensitive to details in particle shape, the spheroidal
approximation is reasonable for many particles, e.g., cylinders as shown in Fig. 6.3.
From Table 6.2, it can be noted that increased deviations from a spheroid result in
decreased accuracy of Eq. 6.25. Highly flattened or elongated particle shapes (E1 or E1)
tend to be over-predicted as discussed by Ganser (1993). For example, application of Eq. 6.25
to disks yields a drag which is about 14% too high for both orientations when compared to the
exact solution. More complex correlations have also been offered but in general empirical
measurements for a specific shape are recommended when the deviation from spherical
condition is extreme.


Irregularly-Shaped Particles in Creeping Flow (Re
p
1)

Many naturally occurring particles are irregular in shape as shown in Fig. 2.10. For those
that closely resemble regularly-shaped particles (such as Fig. 2.10c), the above creeping drag
corrections (Table 6.1 or Eq. 6.25) are generally adequate. However, many particles are too
irregular (e.g., Figs. 2.8, 2.10a and 2.10b) and pose a significant challenge for a number of
252
reasons. Firstly, the surface area or projected area of such particles may be difficult or
impossible to measure so that approximations based on the above area ratios are of little use.
Secondly, experiments for irregular particles at creeping flow conditions are not as plentiful as
that for regular particles. Third, the shape characterization of the particles is often not
documented (e.g., a study may mention simply that sand particle of an effective diameter was
used). Finally, within any mean shape characterization for such particles, irregularity generally
leads to random protuberances orientations so that the statistical variation of drag is often large.
Given these challenges, it can be expected that only approximate and probabilistic
predictions are possible for highly irregular particles. A wide variety of shape-characterizing
parameters have been put forth for irregular particles, but the most common and most
successful is the Corey shape function (CSF). This function is given by
max max med
d / d d which
employs the three lengths of a particle in mutually-perpendicular directions: the particles
longest dimension (d
max
), the shortest dimension (d
min
), and the intermediate or medium
direction (d
med
). A different representation of the CSF is the max-med-min area:

* max med
mmm 2
min
d d
A
d

6.28
This form is chosen for its consistency with Eq. 6.23 so that increasing area ratios correspond
to increasing drag corrections. Since particles will generally lie on a plate on their broadside,
the maximum and medium dimensions can be obtained through visualization from above,
while the minimum dimension (thickness) can be obtained from visualization from a side. This
must then be repeated for a statistically large number of samples to obtain an average max-
med-min area. While cumbersome, this shape characterization is much more convenient than
the sphericity because the surface area of irregular particles is often impractical to measure,
and the same is true for other measures like not-roundedness, anisometry and bulkiness ratio.
Furthermore, the max-med-min area has been found to be generally better at correlating drag
modifications (Clift et al. 1978; Dressel, 1985; Jimenez & Madsen 2003, Smith & Chueng,
2003). Note that a projection area is difficult to characterize for an irregular particle since its
asymmetry generally leads to torque imbalances and thus complex tumbling. However, the
free-fall orientation for irregular particles is not uniformly random over all possible angles as
they tend to fall mostly in the so-called broadside orientation so that d
min
is roughly parallel to
the average fall direction (2.3.1).
Some approximate values for
*
mmm
A for irregular non-spherical particle shapes are given by
Jimenez & Madsen (2003), e.g., 1.2 for smooth well-rounded shapes, 2 for shapes with some
edges but naturally rounded, and 6 for crushed sediment with sharp edges. Experimental
correlation of this area ratio with the Stoke drag correction for particles in free-fall with Re
p
1
is given in Fig. 6.4. As expected for irregular shapes, there is significant scatter but a strong
trend is observed with increasing area ratio correlated with increasing drag. A general curve fit
(averaged over many orientations and particle samples) was given by Loth (2008c) as

( )
0.09
*
shape mmm
f A = 6.29
This is nearly identical to the fit proposed by Dressel (1985) except that a 0.1 exponent was
used. Other shape factors (roundedness, perimeter, circularity, shape entropy,
polygonal harmonics, etc.) have also been put forth to characterize the shape irregularity, but
253
the impact of the max-med-min area tends to be strongest since creeping flow is not sensitive
to sharp edges (Clift et. al. 1978; Tran-Cong et al. 2004). However, it should be kept in mind,
that Eq. 6.29 is only a rough estimate, as evidenced by the large experimental data spread
shown in Fig. 6.4. For particles, for which even
*
mmm
A is not readily available, it is useful to
note that a theorem of Hill and Power (1956) which states that the drag coefficient in creeping
flow is at most the drag of a circumscribing sphere and at least the drag of an inscribing sphere
(Fig. 6.5).


Non-Spherical Particles in the Newton-Drag Regime (Re
p,sub
<Re
p
<Re
p,crit
)

Consistent with how we treated sphere drag in 1.5.2, it is helpful to consider the drag at
high Reynolds numbers before proceeding to intermediate values. As with spherical particles,
non-spherical solid particles with Reynolds numbers in the range of 10
3
-10
5
tend to have drag
coefficients which are approximately constant for a given shape. Since this is evidence of a
Newton-drag regime, these drag coefficients can be defined as a critical drag coefficients
specific to a given shape: C
D,crit,shape
. This drag coefficient can be normalized by that of a
sphere with the same volume (similar to the definition of f
shape
) and the Reynolds number
regime can use the same notation as that for Eq. 1.51b:

D,shape,crit
shape
D,sphere,crit
d=const.
C
C const.
C
for a given shape
6.30
Thus, C
shape
can be thought of as a Newton drag correction; though Thompson & Clark (1991)
referred to this value as scruple. Consistent with the Clift-Gauvin expression, this correction
will be based herein on
D,sphere,crit
C 0.42 , which is the approximate average for a sphere over
10
4
<Re
p
<10
5
(Fig. 6.2). A simple way to obtain this correction is compare the terminal
velocities of a non-spherical particle with a spherical particle of the same volume (and density)
and employ Eq. 1.82b:

( )
2
shape term,sphere term,shape
C w / w = 6.31
Measured values of C
shape
based on free-fall trajectories of different particle geometries are
given in Table 6.3 and further geometries are given by Lasso & Weidmann (1986). Since
particle trajectories at Re
p
>200 tend to have some minor oscillations due to unsteady separated
wake effects, measurements from time-averaged terminal velocities will formally yield path-
averaged Newton drag corrections, i.e.
shape
C .
While corrections for a few specific particle types are known, it is helpful to develop
correlations for generally predicting the Newton drag correction. Similar to the creeping flow
conditions, area ratios are expected to dominate the drag corrections. While drag at high
Reynolds numbers is normally defined by the projected area, it is difficult to determine
*
proj
A
for some particles since the trajectories will generally include secondary motion so that they
are not always falling in a broadside orientation. Therefore, several authors (Clift et al. 1978;
Thompson & Clark, 1991; Ganser, 1993) recommend only using sphericity, i.e. relating C
shape

254
to
*
surf
A . The surface area ratios for cylinders and spheroids are given in Eqs. 6.23 and 6.24,
while Table 6.3 gives area ratios for other shapes. This table also shows that the Newton drag
corrections tend to increase with
*
surf
A with greater sensitivity than at creeping flow conditions,
i.e. C
shape
values of Table 6.3 are larger than f
shape
values of Table 6.2 for the same shape. This
may be attributed to the increased sensitivity of particle boundary layer separation to increased
shape convolutions at high Re
p
conditions. This sensitivity influences the extent of separated
wake flow and the pressure drag component (in contrast, creeping conditions lead to attached
flow over the particle regardless of any shape convolutions.
The drag corrections for shapes in Table 6.3 which have aspect ratios of order unity or less
as well as and data for disks (0.012Ecyl0.515) and a spheroid (E=0.5) free-falling in liquids
are plotted as a function of the surface area ratio in Fig. 6.6. Despite the wide variety of
particle shapes and surface area ratios, a reasonable correlation was obtained by (Loth, 2008c):

( ) ( )
1/ 2
* *
shape surf surf
C 1 1.5 A 1 6.7 A 1 = + + E 1 for and
*
p
~1 6.32
However, this fit does not take into account inertia effects which can be significant at low
aspect ratios and high density ratios. In particular, disks at high Reynolds numbers can
undergo significant pitching and/or tumbling depending on the non-dimensional moment of
inertia (Eq. 2.48), which in turn is proportional to density ratio (Fig. 2.22). Measured drag
coefficients for disks falling in air (Willmarth et al. 1964) indicate a somewhat reduced drag
for a given area ratio. However, data of Stringham et al. (1969) indicate that inertia effects are
further modified once the Reynolds number based on the disk diameter becomes larger than
2000. As such, the above drag correction expression is only approximate.

For cylinders and prolate spheroids (E>1), secondary motion has also been found to be
significant only at extreme aspect ratios. For example, Jayaweera & Mason (1965) found that
unsteady motion such as fluttering was present in liquids at high Reynolds numbers up to
E
cyl
<25, whereas Gorial & OCallaghan (1990) found that cylinders in air were generally stable
for E
cyl
>4. However, both studies (and those they reviewed) indicated that for cylinders with
E
cyl
>10 the drag coefficient defined using broadside projected area was approximately unity,
i.e.
* *
D D,proj proj proj
C C A A = , corresponds to a shape factor of 2.4
*
surf
A . Data collected and
reviewed by Christiansen & Barker (1965) for cylinders of E
cyl
=1.75 and E
cyl
=2.5 indicate
secondary motion is sensitive to * and that affects the mean drag coefficients (up to 20%),
but in general cylinders and prolate ellipsoids can be approximately represented for a wide
variety of density ratios by the following relationship (Loth, 2008c):

( )
* *
shape surf surf
C 1 0.7 A 1 2.4 A 1 + + E 1 > for 6.33
Interestingly, the result for long cylinders (E
cyl
1) is approximately consistent with that based
on a fixed finite-length cylinder in the broadside orientation, i.e.
shape shape
C C

. However,
this is not generally true for some other non-spherical particle shapes because secondary
motion tends to yield higher drag than that of a fixed broadside motion. As such, averaged
free-fall data is important to obtain reasonable results of the mean drag.
To account for secondary motion, a potentially more accurate but complex technique is to
determine the time-evolving particle orientation during its trajectory based on applied torque
255
and particle dynamics, and then apply a C
shape
based on this instantaneous local orientation.
While more demanding numerically, this has the advantage of employing the more detailed
drag data associated with various orientations (which can be obtained from fixed wind or water
tunnel studies) and furthermore allows inclusion of the lift and side forces to simulate lateral
motion. An example of this was conducted by List et al. (1973) for an ellipsoid of E=0.5,
whereby the oscillatory motion was investigated and at least qualitative agreement was
obtained with respect to modeling secondary motion.
The case of irregular particles is further complicated by the randomness of the shapes and
dynamics. However, a review of experimental studies for mean drag of irregularly-shaped
particles suggests that it is again reasonable to use the max-med-min area ratio (as was used for
the Stokes correction factor). A qualitative dependency of the mean Newton shape factor for
available data is shown in Fig. 6.7 which can be reasonably correlated (Loth, 2008c) as:

*
shape mmm
C 6A 1 = 6.34
This relationship, unlike Eq. 6.29 leads to a drag correction greater than unity as
*
mmm
A approaches unity. This trend may be attributed to the increased drag associated with
tumbling and flow separation for surface roughness at high Re
p
(whereas creeping flow drag is
relatively insensitive to roughness). However, as with the Stokes correction for irregular
particles, there is substantial uncertainty associated with Eq. 6.34 such that it is only a rough
guide to the expected drag correction sampled over many particles.


Non-Spherical Particles at Intermediate Reynolds Numbers

Based on the above, non-spherical particles at intermediate Re
p
can also be expected to
produce a significant increase in drag coefficient as compared to that for spheres. Such a result
is shown in the inset of Fig. 6.8 for the regularly-shaped data of Haider & Levenspiel (1989).
Determining a robust dependency at intermediate Re
p
for non-spherical particles has
challenged many researchers and there have been more than a dozen forms of correlations (see
review by Chhabra et al. 1999). The most successful approaches are those by Ganser (1993)
and Cheng (1997) which use a combination of the Stokes drag correction and the Newton drag
correction. These approaches assume that the dependence which governs the variation from
Re
p
0 to Re
p,crit
is functionally similar for all particle shapes and that the differences are
simply the corrections at the two extremes (given by f
shape
and C
shape
). Ganser argued this
dependency based on dimensional analysis, and it can be expressed via
* *
D p
C Re ) = f( by
normalizing the drag coefficient and Reynolds number as:


* D
D
shape
shape p
*
p
shape
C
C
C
C Re
Re
f


6.35a


6.35b
Evidence of this dependency can be seen in Fig. 6.8 where a free-fall data for a large variety of
non-spherical particles from Jayaweera & Mason (1965), Haider & Levenspiel (1989), Madhav
256
& Chhabra (1995), Tran-Cong et al. (2004) are found to approximately collapse for these
normalized parameters. In this conversion, a broadside orientation was employed for the
Stokes correction, i.e. oblate particles are related to
E
f
|
while prolate particle are related to
E
f

.
Upon close inspection of Fig. 6.8, one notes that there tends to be two different curves
along which the data from different shapes collapse. These two curves are related to the
cross-section of the particle shape relative to the flow or C/S for short. In particular,
particles which can have a circular C/S such as spheres and needles (which tend to fall
broadside) are fitted by the solid line. In contrast, particles which will have a non-circular
C/S such as cubes, cones, disks (which tend to fall broadside) are fitted by the dashed lines.
Further data for agricultural particles are given by Gorial & OCallaghan (1990) and Xie &
Zhang (2001) and generally follow the trends given above. This segregation may be related to
the fluid dynamics over the particle surface. In particular, particles which present a circular
C/S to the flow will have a surface separation point that will vary substantially with Re
p

(6.2.1) while particles which present a non-circular C/S to the flow tend to have a separation
point at the surface edges which initiates at a low Rep (about 10 or less) and stays fixed
regardless of Re
p
changes beyond that level.
Another example of the normalized drag curve is shown in Fig. 6.9 for particle shapes that
correspond to various forms of atmospheric precipitation including: spherical drops
(corresponding to clouds and small rain drops), various forms of snow in the form of dendrites,
needles and disks (recall examples in Fig. 2.9), as well as hail shapes in the forms of a cone-
hemisphere and a spheroid with E=0.5. The limiting shape corrections for the spheroid and the
needles are well represented by Table 6.1 and Eqs. 6.32 and 6.33. The data again falls along
the two curves seen in Fig. 6.8 corresponding to whether the particle cross-section is circular or
non-circular. One would expect all prolate spheroids to yield a drag-curve with a circular
cross-section but an interesting issue is whether this is also true for some oblate spheroids.
The spheroid data in Fig. 6.9 and the E=0.5 resolved-surface simulations of (Comer &
Kleinstreuer, 1995) indicate that such a drag-curve holds for E0.5, while the disk data (with
E
cyl
0.05) indicate non-circular cross-section characteristics. Other simulations by Comer &
Kleinstreuer for Re
p
ranging from 40-120 indicate that spheroids with E=0.2 lie approximately
in the middle of the circular and non-circular cross-section curves.
Use of a dimensionless Clift-Gauvin expression based on Eq. 6.16 yields


( )
( )
*
*
*
0.687
*
D p
p
1.16
p
24 0.42
C 1 0.15 Re
42,500
Re
1+
Re

= + +


for circular C/S

6.36
As shown in Figs. 6.8 and 6.9, this gives good correlation for particles for a wide range of
Reynolds numbers whose relative cross-section (C/S) is approximately circular, e.g., spheres,
cylinders and even oblate spheroids with E>0.5. As such, Eq. 6.36 can be referred to as the
circular C/S Clift-Gauvin fit. For moderate particle Reynolds numbers, a normalized
Schiller-Naumann expression may be similarly defined by modifying Eq. 6.14, i.e.


* 0.687
shape p
f f [1 0.15(Re ) ] = + for
*
p
Re < 1000 with circular C/S 6.37
Other expressions for spherical particles could also be normalized in this fashion.
257
For the second group of particles whose relative cross-section is non-circular, a modified
version of the Clift-Gauvin formula similar to that proposed by Ganser can be obtained based
on the best fit to the data as


( )
( )
*
*
*
0.74
*
D p
-0.5
p
p
24 0.42
C 1 0.035 Re
Re
1+ 33 Re

= + +


for non-circular C/S
6.38
This correlation is referred to herein as the non-circular Ganser fit and, as shown in Fig. 6.8,
gives reasonable agreement for a wide variety of shapes. One may also consider the form
proposed by Cheng for sand particles, which is similar to normalizing the Dallavalle drag
expression (Eq. 6.15) except that Cheng used an exponent of 1.5 (vs. 2). However, this was
based on limited irregular particle data (in terms of both shape and Reynolds number range)
and neither of these exponent values gives good agreement with the more comprehensive non-
spherical data presented in Fig. 6.8 and 6.9. A successful variant given by Loth (2008c) simply
employs an exponent of unity (i.e.
D shape p shape
C 24f / Re 0.4C = + ) so that:

*
D
*
p
24
C 0.4
Re
= + for non-circular C/S
6.39
This is referred to as the non-circular C/S Cheng fit and is found to give a remarkably good
correlation as shown in Fig. 6.8 and 6.9, though it does not allow for potential dips which may
occur at about
*
p
Re of about 1000 (prior to reaching the critical drag coefficient).
The maximum
*
p
Re used for the normalized drag correlations is not well known for most
shapes. Particles that are nearly spherical but with small-scale rough surfaces (about 1% or
less) or spheroids with moderate aspect ratios (e.g., for <E<2) can yield a decreased Re
p,crit

but generally display a nearly constant Newton-like drag coefficient for
*
p
Re up to 40,000
(Pettyjohn & Christiansen, 1948). However, the Newton regime for disks tends to be limited
to
*
p
Re >1000 since inertia (I*) effects can become important thereafter. In contrast, the
maximum Re
p
for the Newton regime of irregularly-shaped particles is generally high or
unknown since they exhibit little or no drag crisis (Albertson, 1952; Alger & Simmons, 1968;
and Gogus et al. 2001). This is because such particles tend to have a consistent bluff-body
separation and tumbling motion for a large range of Re
p
. With respect to irregular particles,
the results are generally is consistent with Eq. 6.39 but data are not available over the entire
range of Reynolds numbers (from Stokes to Newton behavior) for a single shape set in order to
establish this correlation as definitive.
In summary, the key to the drag coefficient for non-spherical particles primarily is to
obtaining the Stokesian and Newton corrections for the particle and then employ a normalized
expression for the drag coefficient. These values can be obtained from the preceding figures
and tables for C
shape
and broadside values of f
shape
or upon correlations given in Eqs. 6.25, 6.29,
6.32-6.34. However, it should be kept in mind that the correlations are based on mean free-fall
characteristics and that the area ratios give only a gross generalization of the shape
characteristics, e.g., regularly-shaped particles of different shapes but the same sphericity ratio
can have significantly different Newton-based drag coefficients (that are not predicted by the
258
above functions). Thus, it is always best to obtain C
D
based on experiments with that same
particle shape and Reynolds number, especially for shapes that are substantially non-spherical.

6.2.5. Spherical Fluid Particles

Now let us consider the situation which arises when the particle is a fluid. If the
instantaneous Weber number is sufficiently low (2.2.2), such particles are generally spherical
in shape (e.g., Figs. 2.11a and 2.14a). The regime for spherical fluid particles as a function of
Bond number and terminal Reynolds number is illustrated in Fig. 2.19, and generally any
particles in the creeping flow regime (Re
p
0) can be considered spherical. An approximately
spherical geometry persists for terminal velocity conditions with Re
p
<300 for water drops in air
and with Re
p
<100 for gas bubbles in water. This section will focus on spherical fluid particles,
while the following section will focus on deformed fluid particles.
A spherical fluid particle can have different drag than that of a solid spherical particle due
to the influence of the viscosity ratio and surfactants. If the particle is spherical and is fully
contaminated its surface can be approximated with a no-slip condition. In this case, the
expressions for a solid spherical particle (6.2.1) are appropriate. However, if the particle is
clean or partially-contaminated, then there can be a slip flow along the particle interface which
is driven by internal circulation and the relative velocity. This tends to reduce the impact of
viscous drag, but can also change the form drag. In terms of the viscosity ratio, fluid particles
can be generally divided into two groups:

1 (liquid drops in gas) and

of order unity or
less (immiscible liquid drops or gas bubbles in a liquid). The two groups can behave much
differently at finite Reynolds numbers. However, for creeping flow (discussed in the below
sub-section), a unified theoretical treatment can be applied for fluid particles of all viscosity
ratios.


Fluid Spheres at Creeping Flow Limit

If the fluid sphere has no contaminants, then an exact analytical solution can be obtained
for the creeping flow regime as discussed in 1.5.2. Based on the stream function (Eq. 1.61),
the tangential velocity solution outside of the fluid sphere is given by



* 3 *
p p p p
* 3 *
p p
r 2 3 r
U wsin 1
4r 1 4r 1

+
=

+ +



6.40
The limit as the viscosity ratio approaches zero (e.g., a gas bubble in a liquid) gives a
tangential velocity on the particle surface (r=r
p
) given by


1
2
U wsin

= for
*
p
0
6.41
This result is 1/3 of that for an inviscid free-slip surface (Eq. 1.47c), and indicates that the flow
over the particle surface includes a stagnation point at =0 and but peaks at a magnitude of
w at the top and bottom locations which drives the internal circulation (Fig. 1.28). The
difference is attributed to the stress-free boundary condition (Eq. 1.59).
259
The corresponding drag force can be obtained from the pressure, shear-stress, and
deviatoric normal stress contributions for the fluid particle so that the Stokesian correction
associated with internal circulation follows the Hadamard-Rybczynski result


*
* *
D p p p
*
f
p
F ( , Re 0) 2 3
f
3 d w
3 3

+
=

+


6.42
Recall that
*
p
is the particle viscosity normalized by the external fluid viscosity (Eq. 1.5). It is
instructive to evaluate three limits of this expression:
* * *
p p p
, 1, and 0. The high
viscosity ratio condition is generally satisfied when the fluid is liquid and the external fluid is a
gas, e.g., a rain-drop. In this limit, the tangential velocity reverts to the Stokes flow solution of
a solid particle, yielding a negligible drag correction:


* f 1

= for drops in a gas
6.43
However, it should be noted that the internal fluid motion can be important to the overall heat
transfer between a drop and the surrounding gas since the temperature field depends on internal
convection. The second condition of intermediate viscosity ratio is approximately satisfied for
the case where the internal fluid and the external fluid are immiscible liquids of similar
viscosities with no surface-active contaminants. The resulting flow yields a drag which
happens to correspond to the average of their two drags


*
1
f 5 / 6

= for clean immiscible drops in a liquid
6.44
The third condition of negligible viscosity ratio is theoretically satisfied for the case where the
internal fluid is air and the external fluid is water, such as a gas bubble, where the lower drag
corresponds to


*
0
f 2 / 3

= for clean bubbles in a liquid
6.45
This yields a 3/2 increase in the terminal velocity compared to that for a solid sphere (with the
same density difference and diameter). However, such changes are not observed in
experiments with water due to insufficient purity as discussed below.
The purity assumption used above for the Hadamard-Rybczynski solution indicates that the
interface is free of any surface-active contaminants (surfactants). In view of this, some careful
measurements were conducted in fluids for which contaminants can be nearly completely
removed. This includes experiments of air bubbles in aqueous dextrose solutions and mercury
drops in pure glycerin, for which the predicted terminal velocities agreed well with the
experimental results (Clift et al. 1978). However, interfaces with high surface tensions such
air/water and air/liquid-metals are extremely susceptible to contaminants, especially at small
diameters, so that it is generally not possible to eliminate their effect even with careful
distillation and filtering (as discussed in 2.2.2). As a result, the presence of surfactants in
even modest amounts (i.e., amounts so small as to have a negligible effect on viscosity or
density of the fluid) can partially or fully immobilize the surface flow. If the surface is fully
immobile, then internal circulation is effectively nullified and a solid-sphere drag will result.
Such contamination is especially observed for small particles due to their increased surface-to-
volume ratio and increased sensitivity to surfactants of a given size. In particular, bubble
diameters of 100 m or less (consistent with Re
p
<1) in water or other easily contaminated
260
liquids are generally assumed to have contaminated surfaces yielding a drag consistent with
solid particle behavior:


* f 1

for contaminated conditions (e.g., small gas bubbles in water)


6.46
Examples of fluid particles which tend to the limits of Eqs. 6.43, 6.45 and 6.46 are shown in
Fig. 6.10.
As the Stokes solid-particle drag solution was extended to include linearized inertial terms
by the Oseen expression, Brenner & Cox (1963) extended the Hadamard-Rybczynski
expression to include inertial terms for a spherical fluid particle. This includes a small
correction for small but finite Re
p
values:


( )
*
2
* *
p p
2
p p p
* *
p p
2 3 2 3
3
f Re Re ln Re
16
3 3 3 3


+ +

= + +

+ +

O
for clean conditions

6.47
The corresponding drag coefficient for
*
p
(e.g., drop in a gas) approaches the Oseen
expression (Eq. 6.12) but approaches C
D
=2+16/Re
p
for
*
p
0 (e.g., clean gas bubble in a
liquid). These linearized expressions are reasonable up until about Re
p
of unity (as was found
for the Oseen correction). However, it should be recalled that small bubbles in water are
unlikely to be clean due to the presence of surfactants.
Partially-contaminated fluid spheres can be expected to have drag coefficients in between
that given by the Hadamard-Rybczynski and Stokes solutions. Various approaches to account
for this effect can be related to the angle of the clean portion (
clean
) as defined in Fig. 2.17.
Sadhal & Johnson (1983) obtained an exact solution for the variable viscosity case at the
creeping flow limit


( )
*
clean
1
*
clean clean clean clean
p
3
,
* *
p
p
2 2 sin sin 2 sin3
2 3
f
3 3
2 3 3

+ +
+
= +
+
+
6.48
The above references note that the determination of
clean
is related to the gradient of surface
tension over the surface caused by the surfactant concentration and is based on a large host of
physio-chemical processes including chemistry, ionization, solubility, absorption, surface
tension, surfactant dynamics, etc. Furthermore, the contamination process is also time-
dependent in that injection of a bubble may initially be characterized as clean (with a lower
Stokesian correction) but over a distance associated with several bubble diameters, the
collection of surfactants and dynamic sweeping of these to the rear of the bubble will tend to
increase the stagnant region, until a steady-state condition has been reached. In fact, the
precise method of injection and detachment dynamics can affect the initial concentration levels
so as to change the particle shape for a given bubble diameter (Wu & Gharib, 2002). As such,

clean
may depend on particle sizes, transit times, release mechanisms, contaminant
concentration levels and chemical make-up.


Fluid Spheres at Intermediate and Large Reynolds Numbers

261
When the Reynolds number becomes much greater than unity, the above expressions are no
longer applicable. As with the solid-sphere case, the fluid physics change substantially for
large Reynolds numbers, but are also sensitive to
*
p
. For example, Fig. 6.11 shows
streamlines at Re
p
=100 for fluid particles of different viscosity ratios. The spherical droplet in
a gas at
*
p
=55 (about a 600 micron water drop free-falling in air), yields steady flow
separation and an external flowfield (Fig. 6.11a) which emulates that of the solid particle (Fig.
6.1). This is because the high viscosity of the fluid particle reduces the internal velocities to a
small fraction of the overall relative velocity, so that the surface condition is quite similar to a
no-slip condition. In contrast, a spherical gas bubble in a liquid at
*
p
=0 with an
uncontaminated interface (about an 800 micron gas bubble free-rising in clean liquid) will have
a much different external flowfield (Fig. 6.11b). In particular, the internal recirculation yields
a surface velocity in the direction of the relative flow thereby lessening the local shear stress
and eliminating any wake separation of the external liquid. In fact, a clean spherical gas
bubble in a liquid will have no wake separation for all Reynolds numbers. The streamlines of a
fully-contaminated spherical gas bubble have negligible internal recirculation and so yield a
steady flow separation (Fig. 6.11c), effectively identical to that of a solid particle or a droplet
in a low viscosity gas.
The surface pressure distributions for different viscosity ratios at Re
p
=100 (Fig. 6.12) are
consistent with the above flow physics. The solid-sphere and liquid-sphere in a gas give nearly
identical results with a similar separation point of about 130
o
. However, the clean gas sphere
in a liquid closely follows the inviscid potential flow solution up until the aft region (at about
130
o
), where the pressure recovery is not as high. The reduced recovery results in bubble
pressure drag, though a viscous component also exists and actually dominates the drag for
Re
p
<200 (Clift et al. 1978).
As expected from the above, the particle drag for droplets in a gas or for contaminated
bubbles closely mimics that for a solid-sphere, e.g., to within 1-2% as shown in Fig. 6.10.
Therefore, one may simply invoke the empirical Schiller-Naumann expression given by Eq.
6.14 for fluid particles up to Re
p
of 1000 so long as there is negligible deformation. However,
once the effective viscosity ratio is reduced, the drag on fluid spheres will be less than that of
solid spheres due to flow recirculation (as was found for the creeping flow condition).
In the extreme of a clean bubble with a very low viscosity ratio (
*
p
1), the lack of
separation fortunately allows closed-form theoretical solutions. In particular, Levich (1949,
1962) analyzed the flow by assuming a thin boundary layer on the surface and in the wake and
approximating the flow outside of these viscous regions with the potential flow solution (recall
Fig. 6.12). By assuming high Reynolds numbers and a stress-free surface boundary
condition on the bubble surface (Eq. 1.59), he then related the viscous dissipation in the
boundary layer to work done by drag. The result yields a linear drag which can be expressed in
terms of a Stokes correction as


( )
1/ 2
Re 1, * 0 p
f 2 Re


= +
>
O for clean bubbles with Re
p
1
6.49
The corresponding drag coefficient is given by C
D
=48/Re
p
. Moore (1963) extended the
accuracy of this theory to include an additional term appropriate for intermediate Reynolds
numbers. The resulting Stokes correction can be expressed as:
262


( )
11/ 6
Re 1, * 0 p
p
2.21
f 2 1 Re
Re




= +


>
O for clean bubbles with Re
p
1
6.50
The term in the square brackets represents the correction to the Levich drag. Moores
expression is found to be reasonable for clean bubbles with Re
p
values as low as 100. To
bridge the gap between the Moore and Hadamard-Rybczynski theoretical results, Mei et al.
(1994) proposed the following empirical blended function which tends properly to each limit


1
Re, * 0
p p
2 8 1 3.315
f 1 1+
3 Re 2 Re






= + +





for clean bubbles
6.51
For a wide range of Reynolds numbers, this expression was found to show good comparison
with both clean bubble experimental data (Fig. 6.10) and resolvedsurface simulations (Takagi
& Matsumoto, 1996; Magnaudet & Eames, 2000).
Since flow separation may occur for intermediate values of
*
p
, analytical solutions are not
available at high Reynolds numbers. Therefore, some guidance must be taken from
experiments and resolved-surface simulations (Fig. 6.13). As with the Hadamard-Rybczynski
result, particle drag was found to monotonically decrease as the viscosity ratio is reduced. The
increment in drag for a clean fluid particle beyond that of a clean bubble can be normalized by
the difference between the solid-sphere and clean bubble results. This ratio is defined as f*
which varies from zero to unity as the particle viscosity ratio increases. This ratio was
modeled empirically by Loth (2008a) using a viscosity ratio function (m):


( )
2 3
p
Re, * Re, * 0
Re, * Re, * 0 p
0.01Re 0.4 0.8 1.4
f f
f *
f f 1 0.01Re


+ +

=
+
m m m m

( )
p p f
+ m
6.52a


6.52b
As shown in Fig. 6.12, this model combined with Eqs. 6.14 & 6.51 gives reasonable
predictions at least until a Re
p
of about 1000, beyond which fluid particles tend to deform.
Similar correlations for the viscosity ratio effect are given by Clift et al. (1978) and Feng &
Michaelides (2001), though these include equation broken down for different Re
p
ranges. At
higher Reynolds numbers, there is no quantitative data available for spherical fluid particles
with intermediate viscosities (e.g., immiscible drops in a liquid) since they are generally
deformed at this condition. Before addressing the non-spherical fluid particle drag, we
address the effects of contamination at intermediate Reynolds numbers for spherical particles.
The contaminant influence is a function of the surfactant properties and their concentration
as well as the chemical composition of both the continuous-phase and dispersed-phase liquid
properties. For contaminated conditions, the fluid particle will have a negligible value of
clean

and so will follow the rigid sphere drag coefficient curve due to immobilization of the interface.
Similarly, clean conditions will yield drag coefficients corresponding (at least qualitatively) to
the formulation of Eq. 6.52. For partially-contaminated conditions, no analytical solutions are
available. However, Magnaudet & Eames (2000) noted that the theoretical variation presented
in Eq. 6.48 can be at least qualitatively extended to higher Reynolds numbers (based on
263
resolved-surface simulations) by assuming the same functional dependence on the clean bubble
angle but using the intermediate Reynolds number corrections for the clean and contaminated
limits. If we write this relationship for the more general variable viscosity case it becomes:


* *
p clean p
* *
p p
1
clean clean clean clean Re , , Re ,
3
Re , Re ,
f f 2 2 sin sin 2 sin3

f f 2



+ +
=

6.53
The limits on the LHS * *
p p
Re , Re ,
(f , f )

can be obtained from Eqs. 6.14 and 6.52 so that
knowledge of
clean
would allow computation of the drag correction. While it is generally
difficult to relate
clean
to surfactant concentration, particle and continuous phase properties,
and time of residence, the models of Bel Fdhila & Duineveld (1996) show reasonable
agreement with their experimental results. In addition, Table 2.2 and Fig. 2.18 indicate general
relationships between water treatments, surfactant concentration and level of contamination for
various bubbles sizes. However, the impact of particle deformation becomes important as the
bubble size becomes larger, an issue discussed in the next section.

6.2.6. Deformed Fluid Particles

Deformation due to Relative Velocity

Deformation of fluid particles can occur when the surface tension no longer dominates the
particle geometry. The most commonly addressed deformation condition is that due to relative
velocity. Common shape changes which can occur for falling drops in air and rising bubbles in
water are shown in Fig. 6.14 in terms of the terminal velocity. The larger particle sizes yield
higher dynamic pressures and lower surface tension forces which cause the particle to go from
spherical shapes to oblate ellipsoids. In the case of water droplets in air, this deformation
eventually leads to the bag break-up (Fig. 2.11c). For bubbles in water, the ellipsoidal regime
is followed by the spherical cap regime for quiescent flows (Fig. 2.14c). In general, the bubbles
in tap water have lower terminal velocities due to surface contamination effects for millimetric
sizes. Small bubbles in distilled water (about d<0.5 mm) often behave as tap water bubbles
because it is difficult to eliminate surfactant effects at these scales, whereas large bubbles (d>5
mm) are not sensitive to contamination since the drag will be largely governed by the bluff
body separation of the spherical cap.
As discussed in 2.2.2, the non-dimensional parameters which govern fluid particle
deformation include the Weber number (ratio of relative hydrodynamic stress to surface
tension stress), Bond number (ratio of relative hydrodynamic force to the effective
gravitational force), Morton number (dimensionless group based on external and internal fluid
properties). It was noted that the Mo and Bo parameters are more convenient to use for an
experimentalist examining terminal velocity conditions. As such, there are a wide variety
correlations for shape and terminal drag coefficient based on static parameters of Bond number
and/or Morton number, e.g., Harmathy (1960), Wallis (1974), Beard (1976), Clift et al. (1978),
Ishii & Chawla (1979), Tomiyama (1998), Raymond & Rosant (2000), and Tomiyama et al.
(2002a). However, computational fluid dynamics generally requires the shape and drag of
fluid particles at both equilibrium and non-equilibrium conditions. For example, a small
264
droplet injected into a flow at a high relative speeds (as in many fuel injection systems) can
have high relative velocities and deformations, even though its shape at terminal conditions
would be spherical. In such unsteady cases, the static or terminal parameters given by Mo and
Bo are expected to be inappropriate and dynamic parameters (which do not involve gravity) are
needed.
Therefore, analysis of unsteady conditions requires the use of dynamic parameters (We and
Re
p
) to describe the instantaneous relevant physics. A third dynamic parameter for relative
velocity is given by the particle capillary number (ratio of viscous to surface tension stresses):


f
p
w We
Ca
Re


6.54
The parameter can be independently used with We or Re
p
to specify the dynamic conditions
and is expected to better describe any shapes in creeping flow (Table 6.4). However,
deformation is generally negligible at this condition (Eq. 2.40) so that We and Re
p
are the more
common set of dynamic parameters used to the deformation conditions. Another parameter is
the Tadaki number which is defined as Re
p
Mo
0.23
. The Tadaki number been shown to be quite
reasonable in correlating bubble and drop shapes at terminal velocities (Fan & Tsuchiya, 1990).
Unlike Bo and Mo, this parameter includes the velocity (though Rep). However, it is only
quasi-dynamic since it incorporates a density difference and the gravitational acceleration. As
such, it is not wholly appropriate at unsteady (non-equilibrium) conditions.
The dynamic parameters of We and Re
p
, along with
*
p
and
*
p
, will be used in this section
to provide a quantitative description of aspect ratio (E) and quasi-steady drag coefficient (C
D
).
In particular, the deformation and drag for clean bubbles for small Re
p
is considered, followed
by that for successively higher Reynolds number. Next, deformable drops in a gas are
considered followed by contaminated drops or bubbles in liquids, both of which effectively
correspond to
*
p
1. Finally, intermediate viscosity ratio conditions are considered, i.e. clean
drops in a liquid.


Aspect Ratios of Deformed Fluid Particles at Low Reynolds Numbers

In general, deformation in free-falling quiescent conditions at creeping flow conditions
(Re
p
0) is negligible for the large majority of multiphase flows based on Eq. 2.40.
Correspondingly, the influence of deformation on drag is generally small. However, there are
some multiphase systems which consist of immiscible drops or bubbles in a high viscosity
liquid for which the deformation can be significant, e.g., drops with Mo>1 can have terminal
ellipsoidal shapes for Re
p
values less than unity (Fig. 2.19). For convenience, such shapes can
be characterized with the aspect ratio (E) of Eq. 2.36. However, it should be noted that an
oblate spheroid shape is only reasonably for modest deformations (about E>0.8), whereas
stronger deformations can correspond to more complex shapes (Fig. 2.13).
To analyze uncontaminated fluid particles, Taylor & Acrivos (1964) considered the
geometry of deformation under the assumptions of clean conditions, linearized inertial terms,
and small deformations. The latter assumptions are consistent with We1 and Re
p
1. Taylor
& Acrivos obtained a first-order solution for particle shape and showed that the initial
265
deformation is consistent with that of a spheroid. The resulting aspect ratio can be expressed in
terms of the particle Weber number (We) and a viscosity ratio function (
*
) as


( )
( )( )
*
*
*
* *
p p
*3 *2 *
* p p p 3
*
p
1 2 We
E 1 3 We
1 We
1 12
1 81 57 103 3
80 20 40 4 12
16 1


=
+

+
= + + +

+


6.55a


6.55b
The viscosity function does not vary significantly, i.e.
*
ranges from 0.052 for a bubble in a
liquid (
*
p
&
*
p
0) to 0.063 at for a high viscosity drop in a gas (
*
p
&
*
p
). A correction
to Eq. 6.55a was also obtained by Taylor & Acrivos by including additional terms of
order
2
p
We / Re . The result indicated that larger deformations will give rise to non-symmetric
shapes tending toward a spherical cap. The overall aspect ratio for this higher-order theory
(assuming small Weber number) becomes:


( )
( )
*
p
*
*
p p
2 11 10
We
E 1 3 We 1
Re 7 10 10


+
+
+



6.56

The term in the square brackets indicates the higher-order correction. Comparison of Taylor-
Acrivos shapes with resolved-surface simulations (RSS) at Re
p
=0.5 are shown in Fig. 6.16
indicating that the above theory is reasonable for predicting shapes with We<0.5.
Quantitative comparisons of the aspect ratio at low Reynolds numbers (Re
p
2) against data
of clean bubbles (
*
p
0) from experiments and resolved-surface simulations are shown in Fig.
6.17. The aspect ratio predicted by the Taylor-Acrivos theory (Eq. 6.55a) is reasonable for all
these Reynolds numbers up to We<0.5, beyond which it tends to over-predict the deformation
(consistent with Fig. 6.16). The higher-order Taylor-Acrivos theory (Eq. 6.56) is also shown
for Re
p
=1 and is reasonable within its theoretical limits of
p
Re 1 < and
2
p
We / Re 1 < . At
higher Weber numbers, the aspect ratio tends to reach a constant minimum value
(E
min
=
We
E

). However, the value of this is difficult to determine since there is scatter in the
measurements (since instabilities can arise) and the simulations (attributed to the computational
complexity of handling large deformations). Since there is no theory, an empirical aspect ratio
for uncontaminated bubbles at intermediate Re
p
can be constructed as:


( )
min E
E 1 (1 E ) tanh c We =
6.57
Based on data for clean bubble systems over a range of Reynolds numbers (from 0.2 to 5000)
with shapes which progress from spheres to ellipsoids to spherical caps, E
min
and c
E
can be
modeled (Loth, 2008a) as


( )
( )
min p
E p
E 0.25 0.55exp 0.09Re
c 0.165 0.55exp 0.3Re
= +
= +

6.58a

6.58b
Comparison with Re
p
=1 is shown in Fig. 6.17 where the fit approaches that of the Taylor-
Acrivos theory at low Weber numbers. However, this empirical description will generally be
266
unreasonable for highly sheared or unsteady flows or extreme viscosity systems which yield
more complex deformations, e.g., the skirted or disk shapes of Figs. 2.13-2.19 (though these
shapes only occur for We>100). It should also be noted that the fore-aft (non-ellipsoidal)
asymmetries associated with high deformations in terminal conditions is related to hydrostatic
pressure (Eq. A.56). However, Eqs. 6.57 & 6.58 take this into account since We
term
and Re
p,term

specify the Bo and Mo though Eqs. 2.45 & 2.46; while for non-equilibrium conditions, break-
up will generally occur before the E
min
condition is reached. It should also be noted that the
above models assume that the shape is a function of the local We and Rep and is not
undergoing unsteady deformation. For most flows, this is a reasonable assumption since the
natural frequency of the bubble shape oscillations is on the order of several kHz (Eq. 2.66) so
that the shapes adapt almost immediately to the local flow conditions.


Aspect Ratio of Clean Bubbles at Moderate to High Reynolds Numbers

For moderate Reynolds numbers (about 1<Re
p
<100) where there is neither a thin boundary
layer nor a weak convection condition, there is no theoretical solution to predict the aspect
ratio and instead empirical expressions must be employed. For these conditions, experiments
and resolved-surface simulations of deforming bubbles in high-viscosity uncontaminated
liquids have shown that bubble shape is primarily a function of Re
p
and We (Bhaga & Weber,
1981; Ryskin & Leal, 1984; Kojima et al. 1968; Raymond & Rosant, 2000). Results from
these studies are shown in Fig. 6.18 along with the fit of Eq. 6.57. While there is some scatter,
the data for a given Reynolds number tend to collapse as a function of Weber number with an
initial trend similar to that of the Taylor-Acrivos theory followed by a leveling off at E
min
(all
of which tends to follow the empirical fit). As in Fig. 6.17, the value of E
min
decreases as the
Reynolds number increases. This is attributed to the increased importance of inertia terms
which cause further deformation by creating a high pressure on the front and rear and a low
pressure on the sides. However, the aspect ratio only describes a gross feature of the bubble
shape since the shape is no longer spheroidal at high deformations. For example, the bubble
trailing-edge forms a dimple at Re
p
of about 3 which peaks in magnitude at Re
p
of about 40
and then disappears at Re
p
of about 100 (Bhaga & Weber, 1981; Ryskin & Leal, 1984).
Furthermore, the data in Fig. 6.18 exclude the skirted bubble condition (Fig. 2.16 and 2.19)
which occur at We>100 even in this intermediate Re
p
range (Clift et al. 1978).
At high Reynolds numbers, Moore (1965) obtained theoretical deformations for low Weber
numbers. In particular, he assumed that the viscous flow over the bubble was confined to a
thin boundary layer which remains attached. This attached-wake condition is only possible if
the bubble remains clean (since a no-slip condition of a contaminated particle will lead to
strong separation as noted in Fig. 6.11). Moore noted that this attached flow can persist for
aspect ratio as low as 0.6. Blanco & Magnaudet (1995) conducted several simulations and
showed that the boundary between attached and separated wakes for deforming clean bubbles
is a function of Re
p
and E as shown in Fig. 6.19. This figure also includes two sample
streamline simulations on either side of this boundary, whereby the larger We condition yields
a wake vortex. However, data of the terminal velocity conditions for air bubbles in hyper-
clean water of Duineveld (1995) shown that they may be generally attached.
267
Moores low order theory for deformation is based on a pressure distribution arising from
irrotational flow over a stress-free spherical surface. Assuming a small Weber number, this
yields an oblate spheroid with an aspect ratio:


1 9
1 We
E 64
= +

6.59

Moore also obtained a higher-order approximation by employing the potential flow over an
oblate ellipsoid (Lamb, 1945) which resulted in an implicit relationship for aspect ratio as a
function of Weber number


( )
( )
( )
1/ 3 2 3
2
2 1
3
2
4E 1 E 2E
We E 1 E cos E
1 E

+
=


6.60
For moderate deformations, this high-order relationship can be approximated explicitly (Loth,
2008a) as


2 3
1 9
1 We 0.0089We 0.0287We
E 64
+ + for E>0.5 6.61
The average difference between this result and Eq. 6.60 is 0.3% for E>0.5, so that Eq. 6.62 is
generally reasonable for the entire attached flow regime (Fig. 6.19). It should also be noted
that the hydrostatic pressure gradient does not affect the aspect ratio analysis if fore-aft
symmetry is assumed (i.e. only spheroids are concerned). However, this effect can be
important at high deformations as it leads to significant fore-aft asymmetry (2.2.2).
Predictions of aspect ratio as well as results from experiments and RSS for Re
p
100 are
shown in Fig. 6.20. The results indicate that both the low-order theory (Eq. 6.59) and the high-
order theory (Eq. 6.60 or 6.61) are quite reasonable for small We, consistent with Moores
assumptions. It is interesting that the experimental and RSS results tend to be bounded by the
two theories at moderate Weber numbers, but then tends to a minimum value independent of
Re
p
. Also shown in Fig. 6.20 is the empirical relationship for the aspect ratio in this regime
based on Eq. 6.57. This corresponds well with Moore theory at low We values but the aspect
ratio of bubbles in water tends to be somewhat over-predicted by the empirical result. This
may be due to the difficulty of obtaining a completely uncontaminated surface for water as
discussed in 2.2.2.
Note that the theories are no longer suitable for E<0.5 since they do not account for the
open-wake which develops and since they predict E0 as the surface tension becomes weak.
In practice, bubbles with E<0.5 differ substantially from an ellipsoid (as the fore-aft symmetry
is broken) and at maximum deformation tends toward an (approximate) universal minimum
given by Eq. 6.57, i.e.


We
E 0.25

for Re
p
>100
6.62
This limit is consistent with oblate spherical-cap bubbles (Figs. 2.14c and 2.15a) and has been
pointed out by several researchers for high Re
p
conditions (Clift et al. 1978). The shape can be
regarded as lenticular, i.e. a spherical segment for the upper portion and a flat portion for the
rear. While there has not been success for a direct quantitative theoretical prediction of Eq.
6.62, the interface geometry is qualitatively consistent with neglecting surface tension and
268
imposing a nearly inviscid pressure distribution around a spherical segment portion followed
by an approximately constant pressure in the wake and along the base of the bubble.
Since the above models are functions of the instantaneous Weber number, they would be
expected to be appropriate at velocity conditions which are not necessarily the mean rise
velocity. To test this, predictions of aspect ratio were compared to experimental data of
bubbles rising in clean water whereby the trajectory oscillations lead to relative velocity
variations. Sample comparisons are shown in Fig. 6.21 where the agreement is reasonable
though the variation amplitude tends to be somewhat under-predicted.


Drag of Deformable Clean Gas Bubbles

There are three limiting conditions where drag can be theoretically obtained for clean fluid
particles: a) linearized inertia and small deformation
p
( Re 1, We 1) < < , b) bubbles with thin
boundary layers, attached wake and small deformations
p
( Re 1, We 1) > < , and c) spherical-
cap bubbles with a separated wake
p
( Re 1, We 1) > > . Each is discussed in the following, and
then later are used for a generalized bubble (and droplet) drag for arbitrary
*
p
, Re
p
and We.
Including only linearized inertial terms, Taylor and Acrivos (1964) used the low Re
p

theoretical deformation of Eq. 6.55a to obtain the drag for an oblate spheroid along the axis of
symmetry. They extended the spherical stream function work of Proudman & Pearson (1957)
by assuming only small departures from a spherical shape. Their result is approximately
second-order accurate in Reynolds number and can be expressed in terms of the Stokes
correction factor as:


( )
( )
( )
2
*2 *
* *
p p
p p 2
p * p p 2 * *
*
p p
p
8 3
2 3 2 3
3 6
f Re We Re lnRe
3 3 16 3 3 5
3 3

+ + +
= + + +


+ +
+

|
O
6.63
The first two terms on the RHS are the corrections obtained for spherical fluid particles in
linearized flow (Eq. 6.47) while the last term is due to deformation. At low Reynolds numbers
(where the second term is neglected), the Stokes correction approaches f=2/3+0.055We for
bubbles (
*
p
1) and f=1+0.025We for drops in a gas (
*
p
1), indicating that bubbles are more
sensitive to deformation changes than drops. In either case, significant changes occur for
Weber numbers greater than 0.1.
Drag for the second condition (an attached thin boundary layer over the entire bubble
surface with weak deformation) was obtained by Moore (1965) by assuming an oblate ellipsoid
with the aspect ratio given by Eq. 6.60. The (outer) potential flow solution was obtained
(assuming inviscid flow) to find the dissipation associated with a thin boundary layer
consistent with the stress-free condition. From this, the un-separated flow drag can be obtained
to order
1/ 2
p
Re as


E
D,Moore E
p p
2.21H 48
C G 1
Re Re

=




6.64
The parameters G
E
and H
E
are functions of the aspect ratio. The first can be given directly as
269


( )
( )
( )
2 2 1
7/3
3/ 2
2
E 2
2 1
E 1 E 1 2E cos E
E
G 1 E
3
E 1 E cos E

+
=


6.65
This function can also be approximated for E>0.5 (attached flow conditions) as


E 2
0.4256 0.4466
G 0.1287
E E
+ + 6.66
The second function must be numerically obtained and a table of values is given by Moore
(1965); though a more convenient approximation for E>0.5 (Loth, 2008a) is


E 2
0.5693 0.4563
H 0.8886
E E
+ 6.67
The drag of Eq. 6.64 reverts to the spherical form (Eq. 6.50) as E approaches unity, i.e. G
E
1
and H
E
1. Moore also points out that the contribution of H
E
becomes negligible at very high
Re
p
values, though this requires Mo on the order of 10
-14
or less for terminal conditions. Note
that these results are unreasonable for conditions of flow separation (Fig. 6.19), an issue which
will be addressed below.
The third theoretical result (Davies & Taylor, 1950) is for the spherical-cap bubble rising at
terminal velocity for which surface tension boundary layer effects can be ignored and the
bubble Reynolds number is high. The inviscid drag coefficient is given by Eq. 2.47a:


D,We
C 8/ 3

= for Re
p

6.70
This well-known result has been found to reasonably represent drag on spherical-cap bubbles
(Clift et al. 1978). Joseph (2006) this result to include viscous effects as:


D,We
p
8 14.24
C
3 Re

= + for Re
p
1
6.71
However, the additional term is generally less than 1% for most spherical cap bubbles as they
occur at high Re
p
values. As will be discussed later, there is experimental evidence
that
D
C 8/ 3 is also reasonable for disintegrating drops as We (Simpkins & Bales, 1972)
and thus may be reasonable for a large range of viscosity and density ratios.
For variable viscosity and a wide range of Reynolds numbers, Darton & Harrison (1974)
and Clift et al. (1978) proposed an empirical expression for C
D
at maximum deformation by
linearly combining the creeping flow drag of Eq. 6.42 with the Re
p
0 drag of Eq. 6.70:


*
p
D,We D,max *
p p
2 3
8 24
C C
3 Re 3 3

+
= +


+


6.72
This can be further extended to include the shape correction of Table 6.1 and Eq. 6.58a as


*
p 1/ 3 min
D,We min *
p p
2 3
E 8 24 4
C E
3 Re 3 3 5 5

+

+ +



+



6.73
270
Both of these are linear combinations of the low and high Re
p
expressions, and as such are
consistent with the form successfully employed for non-spherical solid particles with non-
circular cross-sections (Eq. 6.39). In particular, Eq. 6.71 tends to be more accurate for bubbles
at high Re
p
while 6.75 tends to be more accurate for fluid particles at low Re
p
. However, Eq.
6.72 is straightforward and remarkably robust for many conditions, as will be seen.
Furthermore, it closely approximates the other two expressions in their respective regimes so is
quite suitable for general use.
The above theoretical results (and the empirical result of Eq. 6.72) are evaluated with
experiments and direct simulations of bubble drag coefficient as shown in Fig. 6.22 for various
Re
p
values as a function of Weber number. As may be expected, the Taylor-Acrivos theory is
reasonable for Re
p
<1 and We<1, but tends to overestimate the drag for higher Re
p
or We. As
also expected, the Moore theory is reasonable for We<4 and Mo=4.45x10
-10
(which
corresponds to Re
p
>100 and E>0.5). However, it underestimates the drag at higher Weber
numbers, which can be attributed primarily to the onset of a separated wake. At We1 for this
condition, the drag coefficient tends to approach an equilibrium value consistent with the
spherical-cap theory of Eq. 6.70. For Re
p
values as low as 2, the expressions of Eqs. 6.72 and
6.73 give good prediction of the maximum-distortion drag coefficient.
To model the drag at intermediate Re
p
and We values (where the flow is separated but
maximum deformation has not yet been reached), a normalized drag coefficient increment can
be defined (similar to that used in Eq. 6.52a) as


D D,We 0 *
D
D,We D,We 0
C C
C
C C


6.74
For clean bubbles,
D,We
C

is given by the spherical-cap case of Eq. 6.72 whereas


D,We 0
C

can be based on the spherical expression of Eq. 6.51, i.e.




1
D,We 0
p p p
16 8 1 3.315
C 1 1+
Re Re 2 Re





= + +







6.75
When plotted as a function of We, data from experiments and RSS can be primarily grouped in
terms of Re
p
>100 and Re
p
<100.
Considering first the category of Re
p
>100, one can employ Moores theory (Eq. 6.64 and
6.57) to obtain the drag increment at moderate Weber numbers. This gives good predictions
for the attached flow region for liquids of various Morton numbers (Fig. 6.23). For the
separated regime, Mendelson (1967) noted that the terminal velocity was approximately
independent of Re
p
and instead dominated by surface tension and gravitational effects. He then
considered the bubble as an interfacial surface wave disturbance and equated the wavelength to
the bubble perimeter to obtain the terminal velocity as
term f
w 2 / d gd/ 2 = + . This result can
be expressed in terms of the terminal Weber number for a bubble as


term
D,sep
term
We 2 8
C
3
We

=



6.76
271
Clift et al. (1978) and Fan & Tsuchiya (1990) considered minor variations of this expression
(e.g., the value of 2 was replaced by 3 or other numbers). However an improved empirical
correlation as shown in Fig. 6.23 is given by (Loth, 2008a):


( )
D,separated D,We 0
*
D,sep
D,We D,We 0
C C
C 2.5tanh 0.2We 1.5
C C


6.77
A shown in Fig. 6.23, the separated drag is appropriate for We>5 while the Moore drag is
appropriate for We<3. For intermediate conditions (3<We<5), a simple approximation is to use
the maximum value of the Moore prediction and the empirical separated-flow prediction:


( ) D D,Moore D,sep
C max C , C for Re
p
>100
6.78
This combination gives good predictions as shown in Fig. 6.23. Note that the demarcation
between the attached and separated regimes for bubbles in distilled water is responsible for a
peak in terminal velocity (Fig. 6.14), as well as a rapid rise in drag coefficient (Fig. 6.15).
Measurements and RSS results for the intermediate Re
p
condition of 5<Re
p
<100 are shown
in Fig. 6.24 which generally collapse as a single function given by Loth (2008a):


( )
1.6
D
C * tanh 0.021We for 5<Re
p
<100
6.79
For Re
p
<5, the Taylor-Acrivos theory initially applies at small We but soon after the drag
coefficient saturates at the maximum-deformation value which is typically only a few % higher
than the spherical value (see Fig. 6.24). It is difficult to suggest an appropriate C
D
for Re
p
<5
because: a) the C
D
variations are small, b) shapes may be sensitive to initial conditions and c)
the experimental data are scant (Fig. 6.22). However, Eq. 6.79 is at least qualitatively correct.
To overview the performance of the above methods, the measured drag coefficients for a
variety of liquid are shown in Fig. 6.25 based on:


4
p,term
D,term 3
term
Re
4 Bo 4
C Mo
3 We 3 We
= = 6.80
This figure shows the drag coefficient limits for We (Eq. 6.73) and We0 (Eq. 6.75) and
as well as the intermediate deformation predictions based on these two limits with Eq. 6.78 for
Re
p
>100 and Eq. 6.79 for Re
p
<100 and the definition of Eq. 6.74, i.e.:


( )
*
D D,We 0 D D,We D,We 0
C C C C C

= +
6.81
These predictions are generally accurate and robust indicating that drag of clean bubbles which
are deformed by relative velocity can be reasonably characterized by dynamic parameters such
as We and Re
p
.


Aspect Ratio and Drag of Drops in Gases

Now let us focus on droplets in a gas for which
*
p
1 and
*
p
1. The shapes generally
proceed from spheres to oblate ellipsoids and then dimple in the front stagnation point
272
eventually leading to bag break-up (Figs. 2.11 and 2.12). For unsteady flows, dynamic Kelvin-
Helmholtz break-up corresponds to a critical Weber number (Eq. 2.85),


crit,KH
We 12
6.82
However, the break-up criterion for a drop at terminal velocity in quiescent conditions is given
by the Rayleigh-Taylor instability


crit,RT
We 8
6.83
For a rain droplet at terminal velocity, this corresponds to a diameter of about 6 mm at STP
consistent with atmospheric observations.
For steady-state or quasi-steady conditions, one can expect the shape of a drop in a gas to
be dominated by the instantaneous Weber number. Unfortunately, there is no quantitative
theory for deformation (or drag) for drops in a gas. This is because the flow over the surface is
generally separated since the high viscosity ratio results in nearly a no-slip boundary condition
and since deformation typically occurs at Re
p
>200 (Fig. 6.14). As such, one must rely upon
empirical correlations. Experiments for drops of different liquids in different gas pressures
were carefully studied by Reinhart (1964). The resulting aspect ratios correlate well with
Weber number as shown in Fig. 6.26 which can be empirically characterized (Loth, 2008a):


( ) E 1 0.75tanh 0.07We = for We<We
crit

6.84
Note that the maximum Weber numbers recorded in these drop-tube experiments are generally
bounded by Eqs. 6.82 and 6.83. At higher Weber numbers, one may observe a variety of
break-up modes, e.g., bag, vibrational, shear, stripping, and shattering (Theofanous et al. 2004).
While Eq. 6.84 is not intended for We>We
crit
, photographs of drops which are disintegrating
due to shock waves at very high Weber numbers (We>1000) indicate an instantaneous
lenticular shape with surface stripping with a main body aspect ratio of approximately ,
consistent with the limit of this correlation.
Once the shape is known, it is possible to apply the drag corrections for non-spherical solid
shapes (6.2.4) since the internal circulation effects will be small. Indeed, this gives
reasonable results for modest deformations (about We<3) when the shape can be approximated
as an ellipsoid. However, higher Weber numbers yield more complex shapes and so a more
straightforward and accurate approach is needed to directly relate the drag to the Reynolds and
Weber numbers. The normalized drag increment (defined by Eq. 6.74) for a drop in a gas
corresponds well with the dimensionless group
0.2
p
We Re as shown in Fig. 6.27. Note that the
0.2 exponent was empirically determined to give best data collapse and that this group can be
rewritten as
1.1 0.2
f
We Oh

or in terms of We, Oh
p
,
*
p
and
*
p
(Eq. 2.86). The data trends can also
be approximated with the following functional dependence (Loth, 2008a):


( ) ( ) ( )
0.2 0.2 0.2
p p p
2 3
* 3 5 7
D
We Re We Re We Re C 3.8x10 3x10 9x10

+ + 6.85
This model only applies to We<We
crit
(since it becomes unbounded for We), while the
drag of drops disintegrating by shock waves is better modeled with Eq. 6.70 (Fig. 6.28). As
such, there may be an instantaneous drag increment for intermediate We as suggested by the
dashed line of Fig. 6.27, but there is insufficient data for quantitative determination. To
273
summarize the drag coefficient prediction, Fig. 6.28 presents data and predictions for falling
drops (based on Eq. 6.85) and disintegrating drops (based on Eq. 6.73), where it is seen that the
expressions are reasonably robust for a wide range of conditions (400<Re
p
<7000).


Aspect Ratio and Drag of Contaminated Bubbles and Drops in Liquids

Like drops in a gas, contaminated bubbles and drops in liquids yield a separated wake for
even modest Re
p
values so that no theoretical solutions for the deformation and drag are
available, and again we must rely upon experimental evidence and empirical correlations. For
the aspect ratio, there is good correlation with Weber number for a wide variety of drop and
bubble conditions as indicated in Fig. 6.29, whereby the results can be approximated (Loth,
2008a) as


( ) E 1 0.75tanh 0.11We =
6.86
This Weber number dependence is qualitatively similar to the results for other fluid particle
cases in that the minimum aspect ratio tends to 0.25. However, there are some significant
differences which can be gleaned by comparing Fig. 6.29 to Figs. 6.20 and 6.26. The first
difference is that deformations for contaminated fluid particles are less sensitive to Weber
number as compared to clean bubbles in a liquid. This effect can be partially attributed to the
increased velocity and therefore lower pressure (Fig. 6.12) at the sides of bubbles for stress-
free (vs. no-slip) boundary condition. A second difference is that deformation for drops in
liquids is more sensitive to Weber number than that for drops in a gas. This effect may be
related to differences in
*
p
whereby drops in liquids are not as influenced by hydrostatic
pressure gradients, a trend which is similar to that predicted for low Re
p
conditions (Eq. 6.55a).
A third difference is that clean and contaminated bubbles can achieve steady-state
spherical-cap shapes at high We values, whereas drops in liquids and drops in gas tend to break
up by a We
crit
of about 12). This effect is attributed to differences in
*
p
since a droplet
undergoing asymmetric shape oscillations has substantial inertia moving in different directions
which can more easily overcome the restoring surface tension. In contrast, the interior fluid of
bubbles has negligible inertia so that the interface shape dynamics are instead governed by
internal pressure and the surrounding fluid. However, bubbles can be unstable at high We
conditions. For example, Davies & Taylor (1950) noted significant difficulty in creating the
initial conditions for large spherical cap bubbles, lest they break up into smaller bubbles. In
fact, contaminated bubbles tend to be constrained to We<10 (Fan & Tsuchiya, 1990).
Furthermore, E
min
is not universal since toroidal bubbles can form at high We depending on the
initial conditions (Bonometti & Magnaudet, 2006) while skirted bubbles can form at We>100
and skirted drops at We>10 at intermediate Reynolds numbers (Clift et al. 1978).
The incremental drag for contaminated bubbles (
*
p
1) can be constructed in a manner
similar to that for drops in a gas (using Eqs. 6.16, 6.72, and 6.74). The experimental trends
indicating influence of Weber number for this condition are shown in Fig. 6.30 with an
empirical fit given by Loth (2008a) as:
274


( )
1.6
* 0.2
D p
C tanh 0.0038 We Re





6.87
Note that there is some scatter in the data so that this fit is effectively qualitative for
intermediate deformations. For contaminated drops (
*
p
~1), similar trends are observed though
the dependence on Weber number is more complex in form (Fig. 6.31). Hu & Kitner (1955)
proposed that this complexity is due to the onset of trajectory oscillations and irregular (e.g.,
asymmetric) shapes. The recommended an empirical fit for terminal velocity conditions as a
function of We and Mo, but no empirical fit other than Eq. 6.87 is proposed herein for dynamic
conditions because of the trend complexity. To see the reasonableness of this fit, predictions
for contaminated bubbles and drops are shown in terms of terminal Reynolds numbers in Fig.
6.32. This shows that the empirical representations are reasonably robust for small Reynolds
numbers but are more qualitative for larger Reynolds numbers with intermediate deformation.


Aspect Ratio and Drag of Clean Drops in Liquids

The final fluid particle condition examined for isolated drag behavior is that of clean drops
in liquids. Clift et al. (1978) suggested that the aspect ratio for such particles may be related to
*
p
and Bo. However, when data is considered in terms of Weber number as in Fig. 6.33, no
substantial dependence on viscosity ratio is observed for 0.8<
*
p
<2.4. When compared to other
fluid particle cases, the aspect ratio trends for clean drops in a liquid (
*
p
~1) are similar to
those for contaminated conditions (
*
p
1) at low We but tend to those for clean bubble
conditions (
*
p
1) at high We. This result may be due to increased sensitivity of even small
amounts of surfactants which result for very small particles (e.g., Fig. 6.14) in conjunction with
the reduced amount of recirculation possible when
*
p
is order unity. In contrast, large drops
with increased Weber number are much less susceptible to contamination so that they can tend
to clean bubble deformations. This issue is further complicated by the fact that significant
secondary motion can occur for immiscible drops at these conditions. As such, the behavior
for
*
p
~1 is not well understood and only a crude approximation to aspect ratio can be put forth
for We<4 (Loth, 2008a):


( )
4
E 1 0.75tanh 0.07We 0.001We +
6.88
The normalized drag increment for clean drops in liquids is shown as a function of We in Fig.
6.34, where the lower-bound C
D
is that for a sphere of variable viscosity (Eq. 6.52) and the
upper bound is that for a maximum deformation fluid particle (Eq. 6.73). The results indicate
considerable scatter, again without a strong correlation to
*
p
. The normalized drag increment
for clean drops in a liquid (
*
p
~1) shows significant scatter but approximately follows the fit
given by Loth (2008a):


( ) { }
D
C * 0.5 1 tanh We 5.5 +
6.89
275
For partially-contaminated conditions, there is insufficient data to establish quantitative
behavior.
In summary, fluid particles which deform due to relative velocities exhibit shapes and
drags which depend strongly on Reynolds number and Weber number. Experimental data and
resolved-surface simulations agree well with available theoretical predictions (Taylor &
Acrivos, Moore, and Davies & Taylor) in their respective applicable conditions, while
empirical relations can be constructed for fluid particles of intermediate Weber and Reynolds
number. However, only qualitative drag increments are available for clean or contaminated
drops in liquids. Other flow mechanisms beside relative velocity, can cause shape deformation
including proximity to a wall and shape oscillations (2.3.3). However, these shapes tend to be
highly unsteady and little is known about the drag force is affected. As a result, spherical
shape drag coefficients are typically used (though the unsteady component of the drag can be
important as will be discussed in 6.4-6.6). Another deformation mechanism which happens to
be reasonably well-studied (and probably more common) is that associated with as imposed
external shear, as discussed in the next section.


Deformation and Drag due to Shear

As shown in Figs. 2.29, 2.40 and 2.46, shear regions can cause significant deformations
in drops and bubbles. The case of moderate shearing in creeping flow (Re

0, where Re

is
defined by Eq. 2.52) leads to ellipsoidal conditions with three principal axes as discussed in
2.3.2. In particular, Cox (1969) found that the analytical shape is given an ellipsoid with the
three diameters give by the ratios E
1/2
:1:E
-1/2
where the longest and shortest dimensions are in
the shear plane (perpendicular to the vorticity vector). For small deformations and weak shear
(Re

1), the aspect ratio is a function of the viscosity ratio by the relation

( ) ( )
*
p
2
2
* *
p p
5(19 16)
1 E
1 E
4( 1) 19 40/ Ca

=
+
+ +

6.90
In this result, Ca

is the shear Capillary number defined in Eq. 2.53. This theoretical E


evaluated at
*
p
1 < is generally reasonable for bubbles at these high relative shear conditions as
shown in Fig. 6.35. Note that these bubbles were contaminated, which indicates the shear
deformation (unlike relative velocity deformation) is not strongly affected by surfactants or
internal recirculation. Also shown in Fig. 6.35 is the empirical fit of Kariyasaki (1987):

0.6
1 E
0.43Ca
1 E

=
+
6.91
Interestingly, the Cox theory reasonably predicts the aspect ratio for Re
p
values as large as 5
indicating viscous effects remain critical even at these conditions. However, it should be kept
in mind that this theory assumes that the deformation is dominated by fluid shear and not
influenced by relative velocity, i.e.
*
shear
1 > where

* shear
shear
d
w

6.92
276
One may expect that the convective terms would become more important at high Re

, such that
deformation will be determined by the shear Weber number of Eq. 2.88. The qualitative
parameters for influence are shown in Table 6.4.
The effect of shear deformation on drag for creeping flows is generally small but can be
estimated by determining the aspect ratio (Eq. 6.89) and combining this with the relative
orientation to compute the spheroidal correction (Table 6.1 and Eq. 6.18). However,
Kariyasaki found that
*
shear
values up to unity had little impact on the measured C
D
for
contaminated bubbles with 0.005<Re
p
<5. Similarly, Tomiyama et al. (2002b) conducted
investigations with uncontaminated bubbles at 5<Re
p
<80, and found similar drag insensitivity
for
*
shear
values up to 0.2. As such, the effects of shear on the drag of deformable particle are
often ignored. For solid particles, the effects tend to be more significant as discussed in the
next section.

6.2.7. Influence of Vorticity and Particle Rotation

The impact of vorticity and rotation have on particle drag can be characterized by non-
dimensional continuous-phase vorticity for linear shear (Eq. A.35 and Fig. A.3a) and for a
simple vortex (Eq. A.34 and (Fig. A.3b) as well as the non-dimensional particle rotation:

*
shear f ,shear
*
vortex f ,vortex
*
p p
d / w
d / w
d / w




6.115a

6.115b

6.115c
As usual, analytical solutions are available at small Reynolds numbers while experiments and
RSS results are available at larger Re
p
values.
For a particle in a solid-body rotation vortex at the creeping flow limit, Heron et al. (1975)
determined that the drag increases as

( )
* *
vortex vortex p vortex p
5
f 1 Re / 2 Re
14
= + + O 6.116
This theoretical correction assumes small vorticity levels, so it is not clear whether a significant
correction occurs at large rotation rates or larger Reynolds numbers. However, experiments by
Sridhar & Katz (1995) for contaminated bubbles in a vortex flow at 20<Re
p
<80 and
0.04<
*
vortex
<0.1 indicated no influence of rotational vorticity on the drag coefficient
(f
vortex
0), suggesting that the effect on drag is confined to smaller Re
p
values.
For the case of linear shear, Saffman (1965) found that the drag, to leading order, is simply
equal to the Stokes drag, i.e. f
shear
=1. This lack of influence of vorticity on drag approximately
extends to Reynolds numbers as high as 10 based on resolved-surface simulations of solid
particles (Dandy & Dwyer, 1990; Bagchi & Balachandar, 2002a) and for bubbles (Legendre &
Magnaudet, 1998). At higher Reynolds numbers, resolved-surface simulations yield drag
increases for particles and bubbles in shear as shown in Fig. 6.49 which can be approximated
for Re
p
500 and
*
shear
0.8 (Loth, 2008b) as:
277

( )
( )
0.7
* * D
p shear
D, 0
1.2
0.8 * * D
p shear
D, 0
C
1 0.00018Re for 1
C
C
1 0.0035Re for 1
C
=
=
= +
= +
>
<

6.117a


6.117b
However, caution should be used when employing these empirical corrections since they are
based on a limited set of conditions with no experimental verification.
For particle rotation at the small Reynolds limit with linearized inertia terms, Rubinow &
Keller (1961) found that there is no effect of spin on particle drag, i.e. the Oseen correction is
recovered (Eq. 6.12). For particle spin at finite Re
p
values, there have been several
experimental and computational studies as shown in Fig. 6.50. The experimental and
numerical trends indicates that for Re
p
<20 (for which the flow is attached over the particle),
there is little change in drag, i.e. f
Re,
f
Re
. However, significant rotation rates for Re
p
=250
yielded enhanced wake asymmetry and unsteadiness and drag increases of as much as 20%
were found at Re
p
=250 (Fig. 6.50). To describe these trends, Niazmand & Renksizbulut
(2003) developed an empirical correlation for the combined effect of particle spin and
Reynolds number on drag for Re
p
<300 and
*
p
<1, which is given by

p
Re /1000
*
p Re,
D
D, 0 Re
f
C
1
C f 2

=

= +



6.118
At yet higher Reynolds numbers where the wake is fully turbulent (e.g., Re
p
>3,000), both drag
decreases (for
*
p
<2 or
*
p
>15) and increases (for 4<
*
p
<12) have been observed with
variations of as much 20%. However, the complex trends and the scatter in the data (Fig.
6.50) prevent a robust quantitative correction. As such, Crowe et al. (1998) simply suggest
ignoring the effect of spinning on drag at these conditions, i.e.
D D, 0
C C
=
for Re
p
>3,000. In
the next section, the lift force will be discussed which is much more sensitive to spin and shear
rates.




6.3. Lift Force and Torque

While the drag and gravitational forces are typically the most important to the particle
motion, the lift force can often be significant and can even approach the magnitude of the other
two forces in some circumstances (Leal, 1980; Drew, 1983). As such, there are a number of
multiphase systems where consideration of lift is vital. The lift has been noted to be important
for: lateral migration in tubes (Saffman, 1965), effectiveness of micro-centrifuges (Heron et al.
1975), particle deposition in boundary layers (Young & Leeming, 1997), bubble void
distribution in pipes (Kariyasaki, 1987), and tissue manufacturing in rotating bioreactors
(Ramirez et al. 2003). In most cases, the lift is quasi-steady so that it can be described in
terms of the instantaneous flowfield and particle conditions. However, a history-based
unsteady lift can be important in highly unsteady conditions, as will be discussed in 6.6. For
278
non-continuum conditions, Wang (1972) investigated the effect of particle spin in free
molecular flow and obtained a lift force which acted in a direction opposite to that of drag in
continuum creeping flow conditions. Volkov (2007) investigated the transition between the
continuum and the free molecular regimes for spinning particles in order to obtain the critical
Knudsen number where particle spin did not yield lift.
The two primary mechanisms for steady lift on a particle in a continuum include vorticity
in the continuous-fluid (e.g., Fig. 6.52a) and rotation of the particle (Fig. 6.52b). One may also
consider a special combination case of free-rotation (Fig. 6.52c) where there is no torque on
the particle (T=0) so that the particle attains an equilibrium (steady-state) spin related to the
imposed shear. The direction of the lift is defined to be perpendicular to the relative velocity
(and drag). A positive lift and a positive C
L
is defined in the direction of
f
w for vorticity
effects and in the direction of
p
w for particle spin effects, as demonstrated in Figs. 6.53a and
6.53b. The lift components for vorticity and spin can then be based on these directions and the
associated lift coefficients can be based on the same force normalization used for the drag
coefficient (Eq. 1.46):

( )
2 2
L L f
8
C F / w d


2 2
L L L f L L
8
w d C C




= + = +




w w
F F F
w w


6.129a


6.129b
In many cases, the magnitude of the vorticity or the spin is typically expressed in terms of *
and
*
p
(Eq. 6.115). These parameters are thus proportional to a velocity difference across the
particle (due to vorticity or spin) normalized by the particles relative velocity.
An interesting demonstration of shear-induced lift is the stable levitation of a ping-pong
ball with a steady laminar jet. The ball will remain in the jet center since deviation to either
side will expose it to a shear flow which will provide a restoring force. An example of spin-
induced lift is that of a tennis-ball hit with top-spin. This creates a down-ward force which
allows fast shots to better arc down into play. Positive lift in both cases is generally caused by
higher pressure and lower velocities on the bottom-side of the particle as compared to the top-
side. This can inferred from the flow past a spinning baseball in Fig. 6.53 by noting that the
streamlines are closer together at the top of the ball than they are on the bottom of the ball
indicating higher velocities and lower pressures on the top (and a positive C
L
). However,
certain high Re
p
conditions can produce a negative C
L
due to complexities with flow separation,
as will be discussed in the following sections.
The different types of lift are often associated with the founding theories for each. For
continuous-phase vorticity induced lift (Fig. 6.52a), there are three types: Saffman lift which
is based on particles at low Re
p
subjected to a linear-shear flow (Saffman, 1961), Heron lift
which is based on particles at low Re
p
subjected to a vortex with solid-body rotation (Heron et
al. 1975), and Auton lift which is based on particles in inviscid flow subjected to vorticity
(Auton, 1987). The lift associated with particles which rotate (Fig. 6.52b) is termed herein the
Robins-Magnus lift after experimentally-observed lift of spheres by Robins in 1742 and
rotating cylinders by Magnus in 1853, though the concept of sphere lift was first introduced by
Newton in 1672 (Barkla & Auchterlonie, 1971). In all of these cases, the formulation for lift
force can be influenced by the factors discussed in 6.2, such as effect of Reynolds number,
non-sphericity, deformation, internal circulation, etc. However, it should be kept in mind that
279
the current understanding of such effects on lift is not as complete as in the case of drag,
especially for non-creeping conditions and especially when the effects of continuous-fluid
vorticity and particle rotation are combined. Note that only vorticity or spin which is
perpendicular to the relative velocity is considered herein since this is the component which is
related to the traditional lift and since other vorticity or spin components are do not, at least
theoretically, lead to other side forces.

6.3.1. Vorticity-Induced Lift

For viscous conditions, there are two fundamental types of flows with vorticity which
yield a theoretical lift coefficient: linear shear flow (Eq. A.34 and Fig. A.3a) for which
C
L
=C
L,shear
and simple vortex flow (Eq. A.35 and Fig. A.3b) for which C
L
=C
L,vortex
. These
two conditions will be considered separately in the following sub-sections.


Linear-Shear (Saffman) Lift for Solid Spheres

The linear-shear vorticity is defined as having a uniform gradient of the velocity in one
direction, e.g., Eq. A.34. Saffman (1965, 1968) noted that there is no lift due to linear shear in
the limit of creeping flow with no inertial terms (Re
p
0 via Eq. A.63). However, he derived a
finite lift for linearized inertial terms using a matched expansion of an inner and outer solution
and the assumptions that Re
p
Re

1, where the shear Reynolds number is defined by Eq.


2.52 and can be related to the non-dimensional shear of Eq. 6.115a as:

2 *
f f f p shear
Re d / Re

=
6.130
His solution for lift without particle spin can be expressed as

( )
2
L,Saff f f shear shear
1.615 d / 1 0.334 Re

= F w
6.131
Saffman noted that the lift is due to a transverse flow which extends far from the particle but is
induced by the particles surface boundary conditions. For the assumed conditions of weak
shear, a reasonable approximation is to neglect the second term in the parentheses. Using this
approximation and normalization based on Eqs. 6.129a and 6.115a, a Saffman lift coefficient
for the leading-order shear-induced lift coefficient can be defined as


*
shear f shear
L,Saff 2
p
12.92 12.92
C
w Re

=

for Re
p
1 6.132
Since the lift-to-drag ratio can be conveniently expressed as C
L
/C
D
based on Eqs. 1.46 and
6.129a, this equation can be used to assess the importance of lift at finite Reynolds numbers by
comparing with Eq. 6.12. For example, the lift-to-drag ratio is 5.4% at Re
p
=0.1 and
*
shear
=1
which may be significant with respect to lateral diffusion of particles in high shear flows, e.g.,
near the wall of a boundary layer.
McLaughlin (1991) extended Saffmans result using Fourier analysis to eliminate the
restriction of Re
p
Re

, though again assuming small but finite values of Re


p
and Re

. The
solution yields an integral which can be numerically evaluated and the result can be described
280
in terms of ratio of this lift to that given by Saffman. This ratio can be defined as J* and was
empirically represented by Mei (1992) as:


* *
shear shear
10
p p
*
L,McL L,Saff
*
log
Re Re
0.3 1 tanh 0.191 tanh 6 1.92
J C C
5 2
J
2 3

+ + +








6.133a


6.133b
Theory and data for the McLaughlin lift normalized by Saffman lift as well as the above are
shown in Fig. 6.54, where it can be seen that the latter is quite reasonable. It can be seen that
the lift ratio tends to unity (consistent with Saffmans solution) at large values of
*
shear p
/ Re
where the lift force would be the largest. At small values of
*
shear p
/ Re , the predicted J*
reduces from unity to nearly zero values and this trend is consistent with experiments and
resolved-surface simulations. This reduction in effective is due to convection effects which
diminish the far-field influence of the surface conditions. At very low vorticity values, J* even
becomes slightly negative and the Mei fit is not as accurate in this range. However, the total
lift will tend to be a very small fraction of the drag force for this condition, so that the Mei
approximation of Eq. 6.133b is quite reasonable for most conditions.
For finite Re
p
conditions, much of the recent understanding has come from resolved-
surface simulations (RSS) since experiments are quite difficult. As shown in Fig. 6.56a, the
McLaughlin lift force (derived for Re
p
1) turns out to be reasonable even for Re
p
<50.


*
L,shear L,McL L,Saff
C C J C = for Re
p
50
6.134
However, higher Reynolds numbers can result in a significant flow negative lift coefficient.
Kurose & Komori (1999) were the first to fully recognize this numerically and experimentally.
They attributed the result to flow separation which causes both pressure and viscous
contributions to reverse the direction of lift. In particular, two separation regions were
observed on the rear side of the sphere so that the top-to-bottom rear pressure distribution was
comparable and the downward pressure on the upper fore-surface caused the change in lift
direction. This yielded a small negative lift coefficient at high Re
p
values (Fig. 6.55b). The
numerical results associated with Reynolds number for which significant separation occurs can
be fairly well correlated with the following empirical fit (Loth, 2008b):


( )
1/ 3
p *
L,shear shear 10
Re
C 0.0525 0.0575tanh 5log
120




+



for Re
p
> 50
6.135
While these expressions yield fair agreement for the lift of non-rotating solid spherical particles
in an unbounded sheared domain based on simulation data, they should be used with some
caution since no quantitative experimental data is available at finite Re
p
.


Linear-Shear (Saffman) and Vorticity (Auton) Lift for Drops and Bubbles

Similar to the above, shear-flow lift has also been studied for bubbles, with theoretical
results available for uncontaminated conditions. For finite but small Re
p
, the equivalent
version of the McLaughlin solid particle lift was obtained for spherical fluid particles of
arbitrary viscosity ratio by Legendre & Magnaudet (1997) as
281


2 2
* *
p p *
L,L&M L,McL L,Saff * *
p p
2 3 2 3
C C C J
3 3 3 3
+ +
= =


+ +

for Re
p
1 6.136
This form of C
L,shear
recovers the Saffman and McLaughlin expressions in the limit of
*
p
,
while approaching a value that is 4/9 of these expressions for a bubble of negligible viscosity.
Their theory also showed that the leading-order drag term is simply that of the Hadamard-
Rybczynski result (Eq. 6.42). As such, the lift to drag ratio for a clean bubble at Re
p
=0.1 and
*=1 is about 3.6%. Fully-contaminated bubbles or drops have a high effective viscosity ratio
so that their lift force would be expected to follow that of the solid particles.
Clean bubbles in shear at finite Reynolds numbers have also been considered, primarily via
resolved-surface simulations by Legendre & Magnaudet (1998). At high Re
p
values (about
500) and positive velocity gradients (about 0.4 in the y-direction), the flow over the bubble
remained attached with the tangential velocity tending to the potential flow solution, while the
thin wake was found to deflect downwards. Since there is no separation, one may expect that
such conditions would tend toward the inviscid lift solution derived by Auton (Eq. 6.6) which
gives rise to an upward lift in this case. Since the Auton lift is linearly proportional to the
vorticity and to the relative velocity, many researchers define a second type of lift coefficient
based on vorticity and particle volume (which can be related to the previously defined lift
coefficient of Eq. 6.129a) by:

* L L L
L * 3
f p f
6
F F 3C
C
w 4 d w

= =


6.137
With this form, the inviscid rotational Auton lift (Eq. 6.6) can be expressed in expressed in
terms of the two types of coefficient as:

*
L ,Auton
C 1/ 2

=
6.138a

L ,Auton
3 w
C
8 d



6.138b
The first of these two equations is the form commonly cited for the inviscid lift coefficient. It
is important to note that the inviscid lift force is the same in both a linear shear flow and in a
pure vortex flow (which is not true for the low Reynolds number limit, as discussed later).
The RSS results for lift approach the inviscid limit for various shear levels as shown in Fig.
6.56. At low Re
p
values, the RSS results tend to the theoretical limit
*
L,McL
C . To represent the
trends of their results, Legendre & Magnaudet developed an empirical expression which
approaches the theoretical limits of Eq. 6.136 at low Re
p
and Eq. 6.138a at high Re
p
:


( )
( )
( )
1/ 2
2
2
*
p
L,McL *
L,L&M 3/ 2
*
p
p shear
1 16/ Re
C
1 4
C
2 9 1 29/ Re
1 0.2Re /



+


= +


+

+



6.139
Other available resolved-surface simulation results exhibit similar trends. However,
experimental bubble lift data tends to be somewhat lower, which may be attributed to partial
contamination which would tend the lift to the solid sphere trends (Fig. 6.55). In addition, the
282
experimental data may have included shape deformation which can also reduce the lift (as will
be discussed in 6.3.4).


Vortex (Heron) Lift for Solid Spheres

If the continuous-phase vorticity is defined by a vortex with solid body rotation (Eq. A.35
and Fig. A.3b), the lift is different as compared to that for linear shear. In the limit of small but
finite Re
p
, Heron et al. (1975) showed that the theoretical lift coefficient for a particle passing
through the center of the vortex (neglecting time derivatives) becomes:

*
L,vortex vortex p
C 5.091 Re = for Re
p
1 6.140
This result is about 24% larger than that for the Saffman lift (Eq. 6.132). As shown in Fig.
6.57, the vortex-induced lift gives somewhat reasonable results for Reynolds numbers as high
as 100. The available data are primarily for contaminated bubbles which tend to behave like
solid particles. In general, the vortex-based lift can be as much as two order of magnitude
greater than shear-induced lift for Re
p
values of 25 to 100 (Bagchi & Balachandar, 2002b). At
higher Reynolds number, it is not clear whether the vortex-based lift will: a) tend to the
inviscid Auton limit as is the case for uncontaminated bubbles in shear, or b) depart from the
Auton limit as is the case for solid particles. Furthermore, some of the measured lift values in
this figure can be associated with the spin of the bubbles, which complicates their
interpretation.

6.3.2. Spin-Induced Lift

As previously noted, the lift associated with particles which rotate is termed Robins-
Magnus lift. Unfortunately, there is no closed form potential flow solution for a spinning
sphere. However, the mechanism for this lift can be illustrated by examining the potential flow
solution around a cylinder with circulation as shown in Fig. 6.58. These streamlines are
qualitatively representative of the flow observed for a spinning particle if no boundary layers,
flow separation, or other viscous effects are included. At low spin rates, the stagnation points
tend to move downward and the higher velocities on the top lead to lower pressures and
therefore an upward lift force (in the direction of
p
w ). As the circulation tends towards
unity, the stagnation points approach each other on the bottom of the surface leading to even
larger velocity and pressure differences between the top and the bottom. However, higher
rotation rates yields a closed streamline region surrounding the particle such that differences do
not substantially increase further. For small but finite Re
p
(linearized Navier-Stokes equation),
Rubinow & Keller (1961) obtained the analytical solution assuming small spin rates (
*
p
1 < )
and a velocity field given by u

=
p
r
p
(r
p
/r)
3
. This field is a linear combination of the Oseen
solution and that due to a rotating sphere in a fluid at rest at infinity. The resulting particle-
spin lift force (F
L
) obtained by Rubinow & Keller is given as


( )
3
L,R&K f p
8
d

= F w
6.141
283
Based on Eqs. 6.115c and 6.129a, the associated lift coefficient becomes


*
L,R&K p
C = for
*
p p
Re 1 1 < < and
6.142
It is interesting to note that the rotation-based lift is independent of viscosity (unlike the shear-
based Saffman lift). This is based on a no-slip condition which is appropriate for solid particle
and contaminated fluid particles. The stress-free condition associated with a clean bubble
(Eq. 1.59) can be expected to produce negligible lift in comparison.
As mentioned above, there is no closed-form inviscid solution for a sphere that satisfies the
spinning surface boundary condition. As such, only the creeping flow Rubinow & Keller
solution is available for this flow and we define a normalized spin lift coefficient accordingly:


* *
L L, L,R&K L, p
C C / C C /

=
6.143
Note that this coefficient is different from that defined by Eq. 6.137 and that the Rubinow &
Keller conditions of Eq. 6.142 yield
*
L,R&K
C 1 = . To consider the lift beyond these conditions,
one must rely on measurements and resolved-surface simulations. Such results are shown in
Fig. 6.59 for Re
p
<2000, where it can be seen that increasing Re
p
tends to reduce the lift
coefficient. Furthermore, the normalized lift tends to reduce as
*
p
increases beyond 1
consistent with the flow differences noted in Fig. 6.58. This trend is reflected by the empirical
model of Tanaka et al. (1990) for high Re
p
:


* *
L ,Tanaka p
C min 0.25, 0.5/

=


6.144
Other proposed models, not shown, assume the normalized lift coefficient is simply a constant,
e.g.,
*
L
C

=0.4 (Tri et al, 1990) or


*
L
C

=0.55 (Bagchi & Balachandar, 2002a), which are


reasonable estimates for Re
p
values of 20-120 at low and high spin rates, respectively. To
predict over a large range of conditions, the present fit shown in Fig. 6.59 was developed by
Loth (2008b) and is given by:


( )
( ) { }
* * 1/ 2
L p p
C 1 0.675 0.15 1 tanh 0.28 2 tanh 0.18Re


= + +



6.145
This provides a more robust prediction in terms of both Re
p
(up to 2000) and
*
p
(up to 20)
while also approaching the theoretical limit for small Reynolds numbers (Eq. 6.142). As with
the drag coefficient, these lift correlations for Re
p
>200 represent the time-averaged component,
since force unsteadiness results from the unsteady separated wake region.
At much higher Reynolds numbers (>50,000), the results can be grouped into two
categories as shown in Fig. 6.60: sub-critical (where the surface is smooth and laminar
separation occurs) and supercritical (where the surface is rough, tripped or at a high enough
Re
p
to ensure turbulent transition before separation). Regarding the sub-critical conditions, the
high
*
p
trends are reasonably described by Eq. 6.144 and 6.145. However, there is a curious
reversal of lift direction at low spin rates, leading to negative lift coefficients. This was first
observed by Maccoll (1928) and later discussed by Davies (1949), who related it to differences
in flow separation between the top and the bottom, though this effect is still not clearly
understood. A reasonable fit for the overall trend at sub-critical conditions is given by Loth
(2008b) as:
284


( )
* * *
L ,sub p p
C min 0.25tanh 0.5 0.1, 0.5/


=


6.146
For supercritical conditions, low spin rate lift coefficients tend toward the Rubinow & Keller
limit indicating high lift-generation efficiency owing to substantial wake deflection (Fig. 6.53)
while high spin rates tend toward the sub-critical limit. The data includes substantial scatter,
but a reasonable fit is that proposed by Sawicki et al. (2003) for baseballs:


* *
L ,sup p
C min 0.75, 0.3 0.09/

= +


6.147
The appearance of negative lift for high but sub-critical Re
p
values may be related to the rapid
increase in wake volume and the onset of three-dimensional unsteady rotation of the upstream
separation point, which both begin at Re
p
~1300. This separation behavior persists until Re
crit
,
after which the boundary layer transitions to turbulence and the wake volume drastically
reduces since the separation angle moves aft (from about 80
o
to about 120
o
). This increased
reattachment may be the reason for the return to positive lift.


6.3.3. Lift and Torque with both Particle and Fluid Rotation

In this section, the consequence of combined spin and shear will be considered in terms of
particle lift and the resulting particle torque for non-equilibrium conditions. This is followed
by a discussion of torque-free equilibrium spin rates and the resulting lift.


Combined Fluid Vorticity and Particle Spin

Saffman (1965, 1968) showed that the theoretical particle spin lift of Rubinow & Keller
can be linearly combined with the shear-induced lift of Eq. 6.131. Thus, the first-order lift
coefficient can be obtained by combining Eqs. 6.132 and 6.142. Assuming spin and shear are
both perpendicular to the relative velocity and yield a positive lift (which is reasonable if the
particle spin is driven by the shear), the resulting lift coefficient is given by:


*
* shear
L p
p
12.92
C
Re

= +

6.148
This form assumes that lift due to continuous-phase vorticity stems from shear-based lift and
not vortex-based lift, an assumption which was found reasonable for turbulent flows (Loth &
Dorgan, 2009). This analytical solution also assumes that Re

1, Re
p
Re

,
*
p
1, and Re
p
1.
This last assumption also indicates that shear-induced lift will dominate spin-induced lift if
*
p

and
*
shear
are of the same order of magnitude.
Salem & Osterl (1998) showed that this result can be empirically extended to Re
p
<50 by
incorporating the shear correction of Eq. 6.133 and spin correction defined by Eq. 6.143:
285


*
* * * shear
L p L
p
12.92
C J C
Re

= +

6.149
Thus, the lift coefficient when vorticity and rotation are in the same direction is given by:


( ) ( ) ( )
L shear p L,shear shear L p
C , C C

+
6.150
Bagchi & Balachandar (2002a) showed that this linear combination can apply even when the
non-dimensional shear and spin rates that are not small and are not in the same direction:


( ) ( ) ( )
L p L,shear shear p L shear p
, 0, 0 0, 0

= + = F F F
6.151
Ass shown in Fig. 6.6, this combination is typically reasonable for
*
shear
and
*
p
values as
high as 0.4 and for Re
p
values as high as 100.


Torque and Angular Momentum

To employ the spin-induced lift for a free particle, one must keep track of the instantaneous
spin rate. To do this, a point-force approximation to the angular momentum of the particle can
be constructed to relate particle rotation to torque (T) and angular moment of inertia about the
centroid (I
p
) along the particle path. The torque can stem from both fluid dynamic and
collision stresses. The surface torque (T
surf
) is the sum of moments associated with the
surrounding fluid stress on the particle surface. The collisional torque (T
coll
) includes the
effects of other particles or walls coming into contact with the particle and will be discussed in
6.10. The rotational equation of motion specifies that the rate of change of angular
momentum is equal to the sum of torques acting on the particle:

( )
p p
surf coll
t
= +
d I
d

T T
6.152
For a spherical particle of constant radius, the torque only includes the viscous stresses of Eq.
A.60 which contribute to the fluid stress vector (Eq. A.11) while the moment of inertia
appearing on the LHS is independent of orientation:


( ) ( )
surf p p p r r r p
dA r K K dA

= = +

r T i i i T
6.153a

2 5
m d d
= = I
p p
p
10 60
6.153b
If the particle is ellipsoid with uniform density, the torque should be modified to include
pressure components and a surface normal while the moment of inertia about an individual axis
is based on the diameters of the other two axes:

3
2
j
j 1
j i
m
d
=

I
p
p,i
20

6.154
286
The moment of inertia about the axis of symmetry for a thin oblate spheroid (E1) is thus
4
d d / 60

| p
, which approaches that of a disk (Eq. 2.48). There are various theoretical and
empirical expressions for this torque as discussed below.
For a sphere in quiescent flow (
f
=0), the unsteady angular momentum equation was
obtained by Basset (1888) for creeping flow as


( ) ( )
t 4
p 3 f
surf f p
0 f
t
p f f
f
2 2
0
d
d
6 t
4 t 4 t
4 d
exp e rf
3 d d

=




/ d d
- d
d
+ d
d

T

6.155
Similar to the resistance of translational momentum (Eq. 6.3a), the angular momentum is
resisted by the fluid viscosity and includes an unsteady history torque component. Note that
hydrostatic and pressure distributions do not contribute to torque as they operate perpendicular
to the particle surface. Furthermore, the surface torque for a sphere does not have components
associated with gravity or added mass.
A similar equation was derived by Happel & Brenner (1973) for creeping flow which
assumed weak angular acceleration so that the history torques (unsteady terms on the RHS of
Eq. 6.155) could be neglected but which allowed for a finite fluid vorticity so that a relative
angular rotation (Eq. 1.11) could be considered. The resulting surface torque is given by:


T
surf
= -
f
d
3
(
p
-
f
/2) = -
f
d
3

p.rel
6.156
To extend the particle torque to finite Reynolds numbers, one may individually modify the two
RHS components of Eq. 6.156 with separate Reynolds number corrections, i.e.


T
surf
= -
f
d
3
(f

p
- f

f
) 6.157
This form introduces a spin correction (f

) and a vorticity correction (f

), both of which are


simply unity for Re
p
0. Similar to the relationship between Stokes correction and drag
coefficient (Eq. 1.57), these correction factors are often defined in terms of a torque Reynolds
number and a torque coefficient:

2
f p
f
d
Re


surf
1 2 5
f p
2
64 f
C
Re r

=

T

6.158a


6.158b
Results from experiments and several RSS studies with Re
p
= Re

=0 have shown that this form


is reasonable for Re

as high as 30. As the rotational Reynolds number increases further (up to


2,000), the torque coefficient increases beyond the creeping flow solution as shown in Fig.
6.62. This trend can be described with a fit by Loth (2008b):

0.7
f 1 0.012Re

= + 6.159
While this data and the empirical fits are for Re
p
=0, the predicted torque correction at zero
shear was found reasonable for Re
p
of at least 20 based on Salem & Osterl (1998). An
287
expression for the vorticity correction, at least at equilibrium conditions (f
,eq
) can be estimated
by considering the torque-free condition as discussed below.


Torque Response Time and Torque-Free Conditions

To consider the angular acceleration dynamics for a sphere at small rotation rates, one may
combine the ODE of Eq. 6.152 with the moment of inertia of Eq. 6.153 and the surface torque
of Eq. 6.157 (which neglects history torques). If one also neglects effects of collisions
(T
coll
=0), this yields a linear ODE characterized by an angular momentum response time
particle (

):

1
p p
2
f f
t

f

d
d



2
p,rel
p
surf f
d
60


6.160a


6.160b
These equations are analogous to the linear momentum ODE (Eq. 3.6) and the particle
response time (Eq. 3.1). As such, it is interesting to compare the angular and translational
response time for Re
p
0:


p
p p f
0.3
c

=
+
6.161
Thus, high-density particles (
p

f
) will have an angular response which is only 30% of the
translational response indicating that such particles approach spin equilibrium faster than
translational equilibrium. Thus, such particles with a small Stokes number of such particles
will generally be in spin-equilibrium (
p,eq
). In the case of low density particles (
p

f
), the
angular response time becomes very small because there is no added mass effect associated
with the torque. As such, a bubble (contaminated or clean) can be reasonably assumed to be in
spin equilibrium for almost any Stokes number.
If there are no external torques applied to a spherical particle and changes in the vorticity
field are long compared to the angular response time, the particle will come to a spin-
equilibrium (T
surf
=0). As mentioned above, this is especially likely for bubbles. The resulting
equilibrium spin-rate can be non-dimensionalized by Eq. 6.115c and related for this condition
to the non-dimensional shear rate (Eq. 6.115a) as:

*
p,eq p,eq
d / w


6.162
For creeping flow, the equilibrium particle rotation equals the fluid rotation (Eq. 6.156) so that

1 * *
p,eq f
2
=

for Re 0


6.163
For finite Re
p
and Re

(defined by Eq. 6.130), the normalized equilibrium particle spin can be


expressed in terms of the creeping flow torque corrections via Eqs. 6.162 and 6.157:
( )
1 * *
p,eq f ,eq ,eq
2
f / f

=
6.164
288
An Oseen-like first-order correction for the equilibrium spin rate was obtained by Lin et al.
(1970) by assuming small shear Reynolds numbers and no relative velocity:
( )
1 * * 3/ 2
p,eq f
2
1 0.0385Re

= for Re 1

< and for


p
Re 0 =
6.165
Note that RHS Reynolds number has a subscript error in the version given by Loth (2008b).
Figure 6.63 compares this result to measurements by Poe & Acrivos (1975) for Re

of order
unity or greater (and Re
p
=0) and this comparison indicates that the reduction in spin rate is less
severe than predicted by Lin et al. theory. A similar qualitative conclusion is reached when
comparing the theory to results of resolved surface simulations (RSS) for spheres at finite Re
p

values. These RSS and experimental results were used to construct an empirical model for the
equilibrium spin rate (Loth, 2008b) given as:
( )( )
1 * * 0.5
p,eq f p p
2
1 0.0075Re 1 0.062Re 0.001Re

=
6.166
Again, the RHS Reynolds number has a subscript error in the version given by Loth (2008b).
This model, as well as the theory and data sets all tend toward the creeping flow solution of Eq.
6.163 in the limit of negligible Reynolds numbers. One may use this result to estimate the
torque correction under shear at finite Reynolds numbers by employing Es. 6.164 and recalling
f

1 for Re
p
20:
( )( )
0.5
,eq p p
f 1 0.0075Re 1 0.062Re 0.001Re


6.167
A similar expression was proposed by Salem & Osterl (1998) based on RSS results, though it
did not tend to the theoretical limit for small Reynolds numbers. As such, more study is
needed to understand the torque behavior at higher Re
p
and non-equilibrium conditions.
Importantly, the equilibrium spin rate can be used to obtain the corresponding lift which
includes a shear-based lift and a spind-based lift, where the latter is determined by the
equilibrium spin-rate due to the surrounding vorticity. The result can be expressed using Eqs.
6.129, 6.132 and 6.149 as:


( )
2 2
L,eq f L,eq
8
w d C

= F w w
* * * *
L,eq L, L, ,eq shear p p,eq L
C C C 4.113 J Re C

= + +
6.168a

6.168b
This form can be combined with Eqs. 6.133b, 6.145, and 6.166 to include the finite Reynolds
number effects for Re
p
20. For solid particles at low Re
p
, the shear-based lift will dominate
based on Eq. 6.148. However, the shear-based lift becomes weak at higher Re
p
, such that the
spin-based lift will be substantial (if not dominant) in comparison. Finally, it is interesting to
compare the lift and drag forces at spin-equilibrium. The simulation results for Re
p
<40
indicate a trend roughly approximated by Loth (2008b) as

*
L D p,eq
C / C / 4
6.169
Based on Eq. 6.166, this indicates that the lift can be generally neglected when
*
f
1 < .

6.3.4. Lift and Torque for Non-Spherical Particles

Lift and torque for non-spherical particles has received much less attention than the
spherical case. This is attributed to two reasons: 1) the equations of motion become more
complex in that three angles of orientation (along with three components of torque) must be
289
formally integrated in time, and 2) only limited theoretical and experimental information is
available for the lift and torque of non-spherical particles. The review by Leal (1980)
discusses dynamics and theoretical results for non-spherical particles and deformable particles,
and some of these aspects are also overviewed below.


Lift and Torque for Solid Particles

One condition for which exact solutions are available is that for solid spheroids and other
orthotropic particles at creeping flow conditions. If the surrounding flow is uniform, the total
hydrodynamic force is simply based on a linear combination of the Stokes corrections and
velocity components associated with each axis of shape symmetry (Eq. 6.14 and 6.19). If two
or more of these axes contain non-zero components of relative velocity and different Stokes
shape corrections, then the total hydrodynamic force will not be parallel to the total relative
velocity vector. As such, the portion of force which is perpendicular to w may be interpreted
as a lift component while the remainder can be interpreted as the drag force. While this defines
the lift for such particles, there is no net torque for creeping flow symmetric particles (unless
Brownian motion is significant) so that the orientation is steady.
The addition of linear shear at small but finite Re
p
values for non-spherical particles is
generally associated with three components of lift force and torque (Jeffrey, 1922), where the
moments of inertia are given by Eq. 6.153. Gavze & Shapiro (1997) summarize the lift and
torque equations, which revert to the versions for a sphere as the aspect ratio approaches unity.
For ellipsoidal particles with zero net relative velocity (e.g., neutrally-buoyant particles), the
mechanism of shear-based lift is removed, but the torque remains finite. If the fluid shear is
constant in space and time, the torque equations for spheroidal particles can be explicitly
integrated to yield the classical Jeffrey cyclic motion in linear shear flow (2.3.2).
For larger particle Reynolds number, the lift or torque data is scarce. One data set using
resolved-surface simulations is given by Dwyer & Dandy (1990) who investigated lift for
spheroids at 45
o
inclination and 10<Re
p
<65 for various aspect ratios. At much higher
velocities, List et al. (1973) measured lift and moment coefficients for spheroids at various
orientations and aspect ratios at 40,000<Re
p
<400,000. However, such data is not sufficiently
broad to propose quantitative robust relations for non-spherical solid particles for Re
p
>1.


Shear Lift for Deformable Fluid Particles

In conditions of significant shear and significant Weber number, fluid particles will deform
and along the direction of shear (Fig. 2.43a). At high Re
p
and We conditions, the bubble
shapes tend to hydrofoil cross-section due to asymmetric pressure loads. The shape evolution
of a spherical bubble subjected to shear is demonstrated in Fig. 6.64a. The result of this
deformation can have a profound effect on lift. This is demonstrated in Fig. 6.64b which
contrasts the shape and trajectory of bubbles in a shear for different Weber numbers. The
approximately spherical bubble at low We moves towards the right, which is consistent with
the direction based on theory and experiments for solid spherical particle lift in a shear flow, i.e.
i
w
. However, the high We bubble undergoes high deformation yielding a cross-sectional
shape of a hydrofoil which moves in the opposite direction. This movement to the right is
290
consistent with the lift on a hydrofoil at a positive angle of attack. Thus, the deformed case
gives a lift in the direction of - i
w
, i.e. a negative coefficient, a result which has been
observed in experiments of drops and bubbles for a wide range or Reynolds numbers, about
0.01<Re
p
<1600 (Kariyasaki, 1987; Tomiyama et al. 2002b; Ford & Loth, 1998).
Empirical expressions for the lift coefficient were developed by Kariyasaki (1987) for
Re
p
<4 and by Tomiyama

(1998) for Re
p
>4. Both of these were based on the concept that a
highly deformed bubble acts as a hydrofoil, whereby the lift is proportional to deformation
(which provides an effective angle of attack) and vorticity (which provides an effective
circulation). The Tomiyama model assumes that the lift is a linear sum of that due to sphere-
based shear and an additional component owing to hydrofoil deformation, i.e.

* * *
L,shear,We L,shear L,We
C C C = +
6.170
The deformation component was suggested to be a function of Bond number, becoming
significant for Bo>4 and approaching a constant of about -0.6 for Bo>10. To generalize this
result to dynamic conditions, the Tomiyama fit can be expressed in terms of Weber number
and drag coefficient (Eq. 2.81) for clean conditions as
( )
*
L,We D
C 0.25 1 tanh 0.7C We 5 = +


6.171
This gives reasonable agreement for linear shear conditions with clean bubbles over a
significant range of Morton numbers with moderate Re
p
as shown in Fig. 6.65, while a similar
but more exaggerated trend is observed for contaminated air bubbles at high Re
p
. However, lift
measurements by Bataille et al. (1991) for bubbles in distilled water in a spinning cylinder
were more consistent with the inviscid lift expression (Eq. 6.140) indicating little change due
to deformation, i.e.
L,vortex,We L,vortex
C C .


Oscillating Lift for Deformable Fluid Particles

As noted in 2.3.1, ellipsoidal fluid particles at high Re
p
in quiescent liquids have been
found to yield trajectory oscillations which include helicoidal paths and sinusoidal paths. This
unsteady lift due to oscillating side-forces is result of the double-thread wake instabilities (Fig.
2.26), and not necessarily vortex shedding. The effect can included accurately with a fully
resolved approach as in 8, but is often neglected for a point-force approximation, since the lift
has typically a near-zero mean when averaged over several oscillations. However, the
oscillatory side-force is generally instantaneously larger than the shear-based side force and so
can be important to local bubble-bubble and bubble-wall interactions. To capture this effect
approximately with a Lagrangian point-force method, one may assign a periodic or stochastic
side force of zero-mean to the bubble dynamic equation. For example, Tomiyama (1998)
proposed a technique based on stochastic sampling of a Gaussian distribution of velocity
fluctuations which provided qualitative agreement for pipe flows.
One may also extract a deterministic side force model by considering a sinusoidal path in
quiescent flow given by Eqs. 2.49 and 2.50 so that the lateral velocity (perpendicular to
gravity) and associated acceleration are given by:
291


( )
( ) ( )
osc osc osc osc o
osc osc osc osc o
v 2 cos 2 t t
v / t 2 sin 2 t t
=

=

2
A
d d A

6.172a

6.172b
In this equation, t
o
is the initial time of release. For a bubble rising in a quiescent liquid at
nearly terminal velocity with weak path oscillations so that
osc osc term
2 w < A , the lateral
equation of motion can be approximated as:


( ) ( )
3 2 2
f osc f term L,osc term osc D,term
6 8
d c v / t d w C w v C

d d
6.173
This equation assumes an isotropic added mass and can be rearranged as


osc osc
L,osc D,term 2
term term
v v 4dc
C C
w 3w t

+
d
d

6.174
Substituting the velocity and acceleration of Eqs. 6.172a and 6.172b and employing the
instantaneous drag coefficient and relative velocity yields a time-dependent lift coefficient
which can be written in terms of an oscillation coefficient (c
osc
)


( ) ( )
osc osc osc
2
L,osc osc D osc o osc osc o
osc
c 2 / w
4d
C c C cos 2 t t c c sin 2 t t
3







A
A

6.175a

6.175b
This lift coefficient can then be employed as part of an overall point-force approximation for
non-quiescent conditions. Such a linear superposition of forces is often reasonable since the
quiescent frequency and amplitude of the oscillations are approximately maintained along the
relative path, even in turbulent flow (e.g., Fig. 2.27). To consider a symmetric helicoidal
(spiral) motion, a second orthogonal component can be added which is /2 out of phase, but
with the same amplitude (Ellingsen & Risso, 2001). The tumble of disks and solid object
could also be modeled in this vein. Note that the volume-averaged or time-averaged oscillating
lift coefficient is approximated as zero for this model which is reasonable since the angle
between the mean path and the gravity vector is on the order of or (Mercier et al. 1973).
Thus, the oscillatory lift is typically neglected for Eulerian or parcel-based methods but can be
useful for Lagrangian point-force method applied to individual bubble or particle paths. In
reality, the oscillations are caused by instabilities in the wake and the dynamics are controlled
by particle and fluid acceleration effects. In the next three sections, forces caused by
acceleration effects will be discussed in detail including virtual-mass force (6.4), fluid-stress
force (6.5) and history force (6.6).



6.4. Virtual Mass Force

The virtual mass (also called the added-mass or apparent mass) represents the portion
of the surrounding continuous-fluid mass which will be accelerated along with the particle.
This phenomenon occurs whether the surrounding fluid is assumed viscous or inviscid, so long
as it is perturbed by the particles presence. The added mass effect is due to the rigidity of the
292
particle surface, i.e. a fluid particle whose shape simply accommodates to the surrounding flow
(zero surface tension and zero buoyancy) will have no added mass. This influence of the
particle surface is generally reduced at further distances away from the particle (Fig. 1.30), so
that the virtual mass does not relate to a specific region of the surrounding-fluid. Instead, the
virtual mass reflects the integrated mass of continuous fluid which moves with a particle and
its proportionality to the fluid density (
f
) and the particle volume (
p
) is given by the added
mass coefficient (c

). The virtual mass force is related to the force required to accelerate the
surrounding fluid during the particle acceleration. Its effect is included in the particle surface
forces (Eq. 6.1) since it is formally associated with fluid stress acting on the particle surface. It
can be quite important for particle with low density, e.g., for bubbles the virtual mass effects
dominates the particle inertia effect and leads to an increased effective gravitational
acceleration of -2g, as noted in 1.5.4. The added mass effect occurs for all ranges of particle,
e.g., its effect is critical to describe the dynamics of a submarine undergoing a maneuver.


Spherical Particles

The added mass force can be related to a work principle. In particular, the product of the
added mass force and the particle relative velocity must balance the work rate required to
change the kinetic energy of the surrounding fluid caused by the particles movement. For an
isolated spherical particle in quiescent flow, the virtual mass force in creeping conditions (Eq.
6.3c) has the same form as in inviscid conditions (Eq. 6.5b). The latter result was obtained by
Lamb (1945) using the velocity potential . The basic steps for this solution for a sphere are
given by Crowe et al (1998) and outlined below.
For irrotational flow, the velocity potential around a body can be used to describe the
flowfield disturbances caused by the particle movement. Such a flowfield is accurate in the
inviscid limit but is also a reasonable approximation in the limit of thin viscous boundary
layers and wakes (no flow separation), since most of the fluid mass affected by the particle
movement is outside these regions. Using the definition of the velocity potential ( U as
per Eq. A.36) the kinetic energy of the fluid throughout the domain is given by:


( )
1 1 2
f D f D
2 2
K.E. U d d = =


6.176
For incompressible flow with constant density, the continuity (Eq. A.37) allows the volume
integral to be expressed as a divergence, so that it can be converted into an area integral:


( )
1 1
f D f p p
2 2
K.E. d dA = =

n
6.177
The RHS result employs the divergence theorem where the particle surface identifies the inner
boundary of the fluid domain and the surface normal of this area and the particle is n
p
.
For a spherical particle, the surface normal is the normalized radius vector:


p p p
/ = n r r
6.178
Next, it will be assumed that the unhindered fluid is at rest (u
@p
=0) so that U represents the
disturbances caused by a particle moving at velocity v. In this case, the potential flow solution
on the surface becomes based on Eq. 1.47 is given by:
293


1
3 2
p
2
r vcos / r = 6.179
Noting that
p
vcos = n and substituting these results into the RHS of Eq. 6.177 yields


( )
1 1 1 3 2 2
f p f p
3 2 2
K.E. r v v = =
6.180
The term in the parentheses on the RHS can be thought of as the mass associated with the
surrounding fluid kinetic energy, which will be in addition to the particle mass kinetic energy
( m
p
v
2
). As noted previously, the product of the added mass force and the particle velocity
must balance the change in the surrounding fluid kinetic energy. Note that the added mass
force should be in the opposite direction of the particle velocity since the work required should
resist particle inertia increases. With the assumption, the balance may be written as


( ) ( ) ( )
1 1 1 2
,p f p f p
2 2 2
K.E. / t v / t / t

= = = F v v v d d d d d d
6.181
The term in parenthesis on the RHS is thus the added mass force which acts in the direction
opposite of the particle velocity. For the case of a sphere whose radius is changing (e.g., an
oscillating bubble volume), the added mass force can be rewritten as


( )
1
,p f p
2
/ t

= F v d d
6.182
This reverts to Eq. 6.181 for constant particle volume.
Now, consider the condition where the particle is stationary (v=0) but the surrounding fluid
is accelerating. The product of this component of added mass force and the far-field velocity
must again balance the change in the disturbed surrounding fluid kinetic energy. In this case,
the added mass force applied to the fluid should be equal in magnitude and opposite in
direction of that applied to the added mass force applied to the particle, so that


( ) ( )
1
,f @p f p @p @p
2
K.E. / t / t

= = F u u u D D D D
6.183
Since velocity potentials for different fields can be linearly combined to describe the combined
field (A.2), the net added mass force for the case of acceleration of both the particle and
unhindered fluid is given by


@p 1
,f ,p f p
2
t t


+ =


u
v
F F F
D
d
D d

6.184
This is consistent with the derivation of Auton et al. (1988) given in Eq. 6.5b and indicates a
force if the particle accelerates relative as the surrounding fluid with an added mass coefficient
of , which Lamb referred to as the inertia-coefficient.
When the ambient fluid is also accelerating, the creeping flow and inviscid added-mass
expressions are not exactly consistent. In particular, the Maxey-Riley (MR) equation includes
du/dt while the Auton-Hunt-Prudhomme (AHP) equation and that given by Drew & Lahey
(1987) includes Du/Dt. However, Maxey (1993) noted that the differences between the
creeping and inviscid flow limits are generally negligible aside from Faxen effects. For finite
Re
p
conditions with spherical particles, there have been several experimental and numerical
studies (Bataille et al. 1991; Mei et al. 1991, Mei & Klausner, 1992, Mei & Klausner, 1994,
294
Legendre & Magnaudet, 1998; Kim et al. 1998, and Wakaba & Balachandar, 2005) which
have investigated the added mass force. These studies all showed that that the AHP equation
for the added mass force is remarkably reasonable for a wide variety of Re
p
values for both
solid and fluid surface conditions. They also demonstrated that the empirical (and often used)
form proposed by Odar & Hamilton (1964) is incorrect, which related to their misinterpretation
of the history forces. As such, Eq. 6.5b may be conveniently employed for isolated spherical
particles. However, as discussed in 3.7.3, effects of other spherical particles can be linearly
combined. The effect of walls will be addressed in 6.8.


Solid Non-Spherical Particles

Non-spherical particle geometries will result in deviations from the MR and AHP equations.
For a general particle shape, an added mass coefficient tensor (c
ij
) can be used to relate force
in the i direction to motion in the j direction (Yih, 1969)


( )
,i ,ij f p j,@p j
F c u / t u / t

= D D d d
6.185
Analytic expressions for the virtual mass coefficient tensor have been obtained for some simple
shapes in the limits of creeping flow and potential inviscid flow. For a spheroid, the added
mass coefficient simplifies to a vector (c
i
). For this case, Lamb (1945) and Lai & Mockros
(1972) obtained the added mass using the potential flow solution about a spheroid. For prolate
or oblate spheroid accelerating parallel to its axis of symmetry, the added mass can
respectively be given as a function of the aspect ratio:


2 2
2 2 2
1 2
2 2 1
Eln E E 1 E 1
c E>1
E E 1 Eln E E 1
Ecos E 1 E
c E<1
E 1 E Ecos E


+


=

+



=

|
|
for
for

6.186a



6.186b
Interestingly, these c

values are equal to the ratio of form drag to friction drag (Clift et al.
1978). Loewenberg (1993) notes that the parallel added mass for a disk approaches
f
(d

)
3
/3,
for a cylinder approaches 4.12
f
(d

)
3
, and that the perpendicular added mass coefficients are
related to the parallel counterparts by Eq. 3.179. In the case of a rotating spheroid, similar
added mass components arise which are proportional to the
`
(Broday et al. 1998). These
should be added to the torque equation (in contrast, there are no added mass terms for a
spherical particle). As with the sphere, these added masses are independent of viscosity and
equal to the inviscid potential flow values, i.e. are invalid in the case of flow separation
(Benjamin, 1987). As such, the above results may be expected to be reasonable for bodies in
creeping flow or at high Reynolds numbers with thin attached boundary layers (e.g.,
uncontaminated bubbles which are nearly spherical).
However, objects with separated wakes will tend to have increased c

values that are
difficult to predict (Brennen, 2005). One analytically tractable separated condition is the
spherical cap shape (typical of bubbles at We1 and Re
p
1) if one assumes a closed wake
295
potential flow region based on a Hills vortex. In this case, the added mass can be obtained by
linearly combining the effects associated with the spherical fore-section and the recirculating
closed-wake aft-section (Kendoush, 2003). The geometry of a spherical cap can be defined by

cap
as the angle from the axis of symmetry at which a flat rear is located (
cap
=180
o

corresponds to a sphere and
cap
=0
o
corresponds to a circular disk). Defining a cap-angle
parameter as c
cap
=cos(
cap
), the added mass coefficient is:


( ) ( ) ( ) ( )
( ) ( )
cap cap cap cap
cap cap
3 8 10
3
1144 940 c 410 c 135 c 81 c
c
560 840 c 280 c

+ +
=
+
|

6.187
The cap angle parameter may be computed from the bubble aspect ratio as follows


cap
c 1 2E = for E 0.5
cap
2
2
1 4E
c
1 4E

=
+
for E 0.5

6.188a

6.188b
The theoretical added mass coefficients for spheroidal and spherical cap shapes are shown as a
function of aspect ratio in Fig. 6.66, along with measurements and resolved-surface simulations
of deformed rising bubbles. In general, the experiments and simulations follow the trend of
both theories for modest deformations, though the open wake spherical-cap bubbles have
higher added mass values than given by the closed-wake theory.



6.5. Fluid-Stress Force

The fluid-stress force represents the force required to accelerate the fluid which would
occupy the particle volume if the particle was not present (Clift et al. 1978). As such, it is
sometimes referred to as the fluid acceleration force since it arises when the continuous-fluid is
accelerated (which can be present in steady flows such as flow through a contraction section).
To derive this force, the fluid-stresses in the absence of the particles presence (i.e.
unhindered fluid-stresses) are posed as an area integral over the particle surface and then this
is equated to a volume integral by the Gauss theorem applied to a tensor (Eq. A.2):

( ) ( )
ij ij j p ij ij p
j
p K n dA p K d
x

+ = +



6.189
If the kernel of the RHS integral is approximately constant within the vicinity of the particle
(Maxey & Riley, 1983), the fluid stress can be approximated as

ij ij
S,i p p
i j i j
K K
p p
F d
x x x x


= + +




6.190
If flow is incompressible, the gradient terms can be expressed in terms of the uncoupled single-
phase Lagrangian fluid acceleration and hydrostatic force by Eq. A.6b. The resulting force is
then (as given by Eq. 1.69) a combination the fluid acceleration and the buoyancy force:
296

@p,i
S,i f p i
u
F = g
t




D
D

6.191
The RHS approximates the Lagrangian acceleration averaged over the particle volume as the
Lagrangian acceleration extrapolated to the particle centroid. However, it does not require any
assumptions of particle shape or a specific form of the viscous stress. As discussed in 1.5.3,
last term on the RHS is the related to the hydro-static pressure of the surrounding flow, i.e. the
buoyancy force (Eq. 1.71). Since this derivation makes no assumptions as to whether the
surrounding fluid is viscous, it is no surprise that the results is the same for the creeping flow
conditions of Maxey or the inviscid, irrotational limit of Auton et al. (1988). In fact, this result
is also true for a particle of any arbitrary shape or condition so long as the point-force
approximation is reasonable.
To demonstrate the importance of fluid stress and added mass, Batchelor (1967) considered
the case of a particle and the surrounding-fluid initially at rest and then subjected to passage of
an acoustic wave for inviscid conditions. Neglecting gravity, the resulting particle equation of
motion is given by v=3u/(2
*
p
+1), such that particles with very-heavy ratios would tend to be
negligibly influenced by the fluid acceleration, whereas particles with very-buoyant ratios
would tend to have particle velocities and displacements three times that of the continuous-
phase fluid (a result confirmed with experiments).



6.6. History Forces

The history force (F
H
) arises due to the temporal development of the particle viscous
region (around the surface and in the wake caused by to the particle no-slip condition) and is
typically treated separately from the quasi-steady drag force (F
D
). The history force generally
acts in the direction opposite to the relative velocity, and therefore can be considered the
unsteady portion of the drag. However, it can also act laterally to contribute to an unsteady
portion of the lift force. In the following, the drag-based history force is discussed for solid
spherical particle (for small and then larger Reynolds numbers), followed by spherical fluid
particles, and then non-spherical solid particles. Finally, the lift-based history force is
discussed.


Drag-Based History Force for Solid Spheres

The history force is related to the time-scales for diffusion and convection of momentum
away from the particle

2
diff f f
conv diff p
d /
d / w / Re

=

6.192a
6.192b
The drag-based history force for solid spheres in the creeping flow regime (Re
p
0,
diff

conv
)
was first formulated by Basset (1888) and, as such, is commonly known as the Basset force.
The basic principles of the history force derivation are discussed by Crowe et al. (1988) for the
case of an accelerated flat plate, and by Clift et al. (1978) and Maxey (1993) for a sphere,
297
where the latter result is given by Eq. 6.3d. For spatially uniform flow and a particle starting
from rest, the history force and the associated integrand (kernel H) are simplified as:


( )
t
H f
-
3 d H t

w
F
d
d
d

( )
2
diff f
Basset
f
d 1
H
4 t 4 t

= =




6.193a



6.193b
The unsteady drag component is thus based on the time integral of the particle acceleration for
the entire trajectory. In comparison to the steady-state drag of Eq. 1.45, the history force can
be neglected if the time-scales of the relative velocity changes are slow. From a path-averaged
point of view, this is approximately equivalent to the following criterion for creeping flow:

diff
w t w < d /d for negligible Basset history force
6.194
For very heavy particles (

1), the diffusion time-scale is generally small due to the low


density of the surrounding fluid. As such, it is often negligible except for the case of very
rapid acceleration, such as small particles in the vicinity of a shock wave. In contrast,
particles in high density liquids are more likely to have a significant history force effect
(Armenio & Fiorotto, 2001).
The Basset kernel has a dependence on t
-1/2
which indicates that the decay of influence is
slow. However, this is no longer true at finite, though small, Reynolds numbers (Sano, 1981)
where faster decays are possible. For Re
p
greater than unity, Mei & Adrian (1992) used an
asymptotic analysis to investigate the history force at finite convective conditions for small
amplitude (10%) fluctuations of the mean flow. They showed that the finite Re
p
empirical
expression of F
H
by Odar & Hamilton (1964) was unsuitable since it employed an incorrect
added mass expression. Furthermore, Mei & Adrian determined that the decay rate for finite
Re
p
is faster than that given by Eq. 6.193a when t
diff
. In particular, the long-time decay rate
was found to be proportional to t
-2
(instead of t
-1/2
of Eq. 6.193b). Examining conditions with
weak perturbations, i.e.
rms
w w < , they established the following long-time kernel:


3
3
2 2
diff conv diff
long time H2 H2
diff p
1 1 3
H c c
t 4 t 4Re




= + = +





6.195
The first term in the square brackets represents the theoretical result for small but finite Re
p
at
long times, while the second term includes an empirical coefficient (c
H2
) for the high Re
p
limit
(where
conv
is no longer much slower than
diff
). Lawrence & Mei (1995) showed that both the
short-time decay of t
-1/2
and long-time decay of t
-2
are also valid for non-parallel change of
direction so long as deceleration is weak.
The transition between the short-time decay (Eq. 6.193a) and the long-time decay (Eq.
6.195) is shown in Fig. 6.67. Mei & Adrian (1992) suggested a history force kernel which
combines the short and long-time limits to also allow description for arbitrary times by
including an empirical power exponent (c
H1
) as follows:
298


( )
( )
H1
H1
H1
c
1/ c
1/ c
Re Basset long-time
H H H


= +


6.196
This expression tends to the limits of Eq. 6.193a for short-times and to Eq. 6.195 for long-
times. Based on resolved-surface simulations conducted by Mei et al. (1991) for Re
p
values up
to 100, they proposed empirical coefficients of c
H1
=2 and c
H2
=0.105. Figure 6.68 shows the
relative importance of the history force for a particle falling from rest in a quiescent liquid.
The speed at which the particle tends to the terminal velocity is reduced once history forces are
included. It can be seen that the experimental data at finite Re
p
initially follows the creeping
flow prediction but then shows a reduced influence of history force for longer times. This
trend is consistent with the transitions described by Eq. 6.196 and Fig. 6.67.
Kim et al. (1998) investigated the coefficients the form suggested by Mei & Adrian and
proposed a more detailed formulation (including six constants determined by calibration with
their own resolved-surface simulations) to increase robustness. However, their model for weak
accelerations simply reverts to the form of Eq. 6.196, except that their values of the constants
are c
H1
=2.5 and c
H2
=0.126. Dorgan & Loth (2006) conducted extensive comparisons of these
models with a wide variety of falling particle experimental data, including Re
p,term
ranging from
9 to 853 and
p
/
f
ranging from 1.17 to 9.32. They found that both models were reasonable but
slightly modified constants of c
H1
=2.5 and c
H2
=0.2 allowed for the best overall fit, e.g., Fig.
6.68. An efficient but approximate prediction for the history force is also shown in this figure
as the window-based approach, which will be discussed later in 7.2.3.
To characterize the transition from t
-1/2
to t
-2
dependence, one may define a history
transition time-scale,
H
, as shown in Fig. 6.67. In particular, this time-scale can be defined by
equating a Basset kernel integrated over a period which extends
H
in the past with the finite
Re
p
history kernel integrated over all times:


H
t t
Basset Re
t
H H


d d
6.197
The LHS integral can be evaluated analytically using Eq. 6.193b while the RHS can be
numerically integrated using Eqs. 6.195-6.196. For c
H1
=2.5 and c
H2
=0.2 (which most closely
match experimental data), the transition time can be approximated as (Dorgan & Loth, 2007):


2
H diff
p
0.502
0.123
Re

+


6.198
This relationship indicates that
H
for creeping flow,
H

diff
for Re
p
=0.5, and

H
0.015
diff
for high Re
p
. Using the history time definition, the criterion of Eq. 6.194 can be
extended to arbitrary Reynolds numbers as

H
w t w < d /d for negligible history force
6.199
This indicates that particles with high Reynolds numbers are not strongly influenced by history
forces unless the unsteadiness is very strong.
The above experiments and predictions considered only particles which accelerated in one
direction. If a particle reverses direction, though, it can ingest its own wake which further
complicates the history force. If the mild decelerations are weak the behavior does not vary
299
significantly from the weak acceleration condition discussed above. However, if the particle
ingests its own wake (due to an abrupt stop, rapid deceleration, or direction reversal), the long-
time decay rate of t
-2
(shown in Eq. 6.195) is instead given by a t
-1
decay rate as discussed by
Lawrence & Mei (1995) and Lovalenti & Brady (1995). For oscillating particles in a quiescent
mixture with a motion v
o
cos(2t/
osc
), one would thus expect that the Mei & Adrian form is
reasonable for short-times, i.e.
osc

H
. Even at intermediate times (
osc
~
H
), the Mei & Adrian
form still gives reasonable results compared to the data and better than that of the Bassett
model. An example of this is shown in Fig. 6.69 in terms of the hydrodynamic surface force
for an oscillating particle in a still fluid. The representation will no longer be reasonable for
very slow oscillations (
osc

H
), however in this case the unsteady force is rather weak. For
example, the history force at
osc
0.37
H
at Re
p
=20 only represents about 2% of the total drag
(Mei et al. 1991), so that it can be reasonably neglected altogether.


Drag-Based History Force for Fluid Spheres

For fluid particles of variable viscosity, the force can be obtained in the frequency domain.
However, a closed-form solution is only available for the inviscid bubble limit (Lovalenti &
Brady, 1995):


( ) ( )
t
2
H, * 0 f f f
-
d
4 d exp 36 t / d erfc 6 t / d d
d
=

w
F 6.200
This result yields a history force which is generally smaller for a clean bubble (than that for a
contaminated or solid surface) due to the existence of finite slip on the particle boundary. For
finite Reynolds numbers up to 300, Mei et al. (1994) proposed a detailed empirical correction
which included short-time and long-time components with coefficients based on RSS results
and found reasonable agreement for bubble trajectories in uncontaminated liquids (Park et al.
1995). However, the force is generally weak for clean bubbles with Re
p
>50 (Magnaudet &
Eames, 2000). In fact, for impulsively started or oscillating bubbles where the boundary layer
is thin compared to the quasi-steady result, Magnaudet & Eames (2000) noted that the unsteady
drag is reasonably represented simply by Moores quasi-steady result (C
D
=48/Re
p
).


Drag-Based History Force for Non-Spherical Solid Particles

For spheroids, Lai & Mockros (1972) obtained the history force for movement parallel to
the axis of symmetry in Basset (creeping flow) conditions. This change in the unsteady drag
from sphere to spheroid was found to be related to the change noted for the quasi-steady drag
(f
E
given by Table 6.1 and Fig. 6.3). For motion parallel to the spheroid axis of symmetry:


( ) ( )
t
2
H,E f E Basset
-
d
3 d f H t d
d

| |
w
F 6.201
Thus, the history force becomes more important compared to quasi-steady forces as the particle
becomes highly oblate or prolate. The above proportionality was generalized to some solid
300
particles of arbitrary geometry in creeping flow by Lawrence & Mei (1995), i.e. the deviation
in unsteady-drag resistance (from a sphere) was the square of the deviation in the steady-drag
resistance.


Lift-Based History Force for Solid Spheres

The unsteady component for the lift force has been derived by Coimbra & Kobiyashi
(2002) for the limits imposed by Saffman (linearized inertial terms with Re
p
1 and a weak
shear based on Re

1). The combination of steady lift (Eq. 6.131) and unsteady lift can be
written as:
( )
( )
t 3
2 f f f
L
0
/
d
1.615d 1.615
2 t



= +




w
F w

d d
d
6.202
The second term represents the unsteady contribution which can be important when the
frequency associated with vorticity and/or relative velocity becomes significant compared to
the diffusive frequency (
2
f
/ d ). Asmolov and McLaughlin (1999) also derived an expression
for an unsteady Saffman lift force in frequency space whereby lift can be represented by the
steady Saffman lift multiplied by a function of the relative frequency normalized by the fluid
shear rate. The functional dependence was obtained numerically and also approximated with
an analytical functional, whereby the real part approaches unity for low frequencies (i.e. tends
to the steady-state solution) but reduces significantly at higher frequencies. Simulations of
particle dynamics in rotating flows by Lim et al. (2005) indicate that the history portion of the
lift is often negligible, but can be significant for high frequencies and small values of
p
/
f
.
For finite Re
p
values, there is no analytical solution so one must rely primarily on
experiments and simulations. For the unsteady lift component, Wakaba & Balachandar (2005)
examined the unsteady component of the lift for a solid particle in a uniform shear but with a
constant acceleration in the relative velocity for a finite duration with Re
p
ranging from 5-125.
They found that the unsteady lift kernel at finite Re
p
was substantially smaller than the
creeping flow predictions, a result consistent with the unsteady drag behavior at finite Re
p
.
They also proposed a model for the unsteady lift kernel overall reduction which yields an
effectively inviscid lift history force with a decay that scales with
conv
(instead of
diff
).
However, they were not able to establish a quantitative model for an oscillating force that also
appeared, nor did they examine conditions of deceleration, unsteady shear or particle rotation.
As such, significant work is needed to understand the unsteady lift component for generalized
fluid vorticity, spin and Reynolds number conditions.


6.7. Thermophoretic Forces

While many of the above surface forces are based on relative velocity or acceleration
effects, the thermophoretic effect is caused by temperature gradients in the continuous fluid
surrounding the particle. It arises due to the gradient of the kinetic energy of the surrounding
molecules and can be significant (compared to drag) for non-continuum conditions, e.g., in
small particles in gasses. In particular, the hot side of a particle experiences collisions from
molecules with higher velocities as compared to that of the cool side according to kinetic gas
301
theory (Fig. 6.80). This yields a force in the direction opposite to the temperature gradient.
Thermo-phoresis can be used to drive particles away from a hot wall or toward a cold wall, the
sooting of kerosene lamps is an example of the latter. The fabrication of micro-processors
depends on thermo-phoresis because the repulsion and or deposition of impurities on the wafer
is controlled during elevated temperature conditions. In addition, this phenomenon is used in
thermal precipitators to filter sub-micron-sized particles from gas flows. Thermophoretic
forces can also play an important role in plasma flows, combustors and heat exchangers.
The thermophoretic force is generally written as a function of the particle Knudsen number
(Kn
p
) as defined in Eq. 6.94. As with drag in rarefied flows (6.2.2), thermophoresis has three
primary regimes: near-continuum (Kn
p
1), transition (Kn
p
~1), and free-molecular (Kn
p
1). As
with drag, there are theoretical approaches for the extreme cases, whereas the transition regime
is generally modeled with empirical relations based on experiments and resolved-surface
simulations. A second influential parameter is the ratio of the gas to particle thermal
conductivities:


p p
15
=
g g g *
k R
k
k 4k

6.240
If the particle thermal conductivity is small compared to the gas (k*1), then it will have an
interior temperature gradient which approaches that of the surrounding flow for near-
continuum conditions. However, most gas/particle combinations are best described by k*1
for which the particles internal temperature field is more uniform, resulting in increased
temperature differences between the particle and the gas at the interface. Various models for
the force are considered below for steady-state conditions.
For the near-continuum regime, one of the earliest approaches for the thermophoretic force
was presented by Epstein (1929). However, the first analysis to formally allow for small k*
conditions was completed by Brock (1962) using the surface boundary condition with
tangential accommodation (Eq. 6.98a). Brock assumed a uniform temperature gradient in the
far-field and imposed a heat flux boundary condition on the sphere surface based k*. The gas
and the particle temperature fields were assumed to satisfy Laplaces equation with a
temperature-jump boundary condition at r=r
p
given by
g p T m m g
T T c ( T / r)

= + l , where c
T
is a
temperature accommodation coefficient, about 2.18 (Loyalka, 1992). Using the Stokes stream
function, the Brocks force due to thermophoresis is:


( )
( )
p
p T
T,Kn 1
@p
p p T
2 c
6 d 2Kn c
T c
T
2 c
1 6Kn 1 2 4Kn c
c


+



=




+ + +



<
F
*
g g
g
g
*
k
k

6.241
The negative sign indicates a force opposing the unhindered gas temperature gradient, i.e. the
gradient which would exist in the absence of the particle. Talbot et al. (1980) noted that this
theory is reasonable for near-continuum (Kn
p
1), indicating that thermophoresis (like non-
continuum drag effects) becomes more important as particles become smaller or the gas
pressure decreases.
302
The large Knudsen number (l
m-m
d) limit can be obtained assuming no collisions between
molecules, as was employed for free-molecular drag theory in 6.2.2. The result by Waldmann
(1961) for monatomic gases is


p
T,Kn 1
p
@p
T
d
2 Kn T

=


>
F
g
g g
g

6.242
It turn out that Brocks theory (Eq. 6.241) approaches the free-molecular limit (Eq. 6.242) in
the limit of Kn
p
1, but unfortunately is not accurate for intermediate Kn
p
conditions, as will be
discussed below. However, this indicates that the intermediate regime may (and indeed does)
have a similar linear dependence on the gas viscosity, particle diameter and gas thermal
gradient. Thus, it is appropriate to consider a normalized thermophoretic force based on the
free-molecular limit, which is only be a function of Kn
p
and k*:


p
* T
T
T,Kn 1
F
F
F

>

6.243
To observe the detailed dependence on Kn
p
for the intermediate regime, one may consider data
from resolved-surface simulations based on the Boltzmann-Krook-Welander (BKW) and
Bhatnagar-Gross-Krook (BGK) models of the Boltzmann equation (Loyalka, 1992). Data from
such simulations are shown in Fig. 6.81 for the non-dimensionalized force. The results
indicate that the force is a strong function of the Knudsen number but only a weak function of
the conductivity ratio. Generally, the Brock theory is appropriate for Kn
p
<0.01, while the free
molecular theory is appropriate for Kn
p
>20.
As discussed by Zheng (2002), there have been several models of various levels of
complexity which have proposed to describe the intermediate Kn
p
functional dependence in
terms of the secondary parameters: k*, c

and/or c
T
. However, a simple model which neglects
these secondary parameters and gives reasonable agreement with the resolved-surface
simulations is given by


1.7
p *
T 1.7
p
Kn
F
1.15 Kn

+

6.244
In Fig. 6.82, this model is compared to experimental data in various gasses (air, argon and
helium) with conductivity ratios (k*~0.1) and low conductivity ratios ( 1 <
*
k ). Though there
is considerable spread in the data and perhaps some evidence of a k* dependence, the simple
empirical fit yields good agreement for Kn
p
greater than 0.01 for all the conductivity ratios
studied. In contrast, the Brock theory tends to overestimate the force for Kn
p
greater than 0.01
(by about four-fold) and unfortunately there is little experimental data at lower Kn
p
values,
where it is expected that the k* dependence will be more significant. However, the force is
generally negligible in this regime so this issue may not be of substantial practical importance.



6.8. Particle-Wall Fluid Dynamic Forces

303
Often, conditions for pipe flows and boundary layer flows can produce high concentrations
of particles very near the wall, e.g. due to thermophoresis. In particular, several studies have
shown such high concentrations at transverse positions on the order of the particle diameter for
solid particles (Young and Hanratty, 1991; Kaftori et al. 1995; Young & Leeming, 1997) as
well as for gas bubbles (Zun et al. 1992; Marie et al. 1997; Felton & Loth, 2001). In these
cases, the consideration of particle wall interactions can be crucial. In general, one may
consider two types of forces that arise as a particle approaches a wall, i.e., fluid dynamic
surface forces (a function of the continuous-fluid resistance due to the particles proximity to
the wall) and collision forces (which arise only from surface contact with the wall). Only the
fluid dynamic surface forces will be considered in this section (primarily in terms of drag and
lift), while contact forces will be considered in the following section.


Fluid Dynamic Corrections for Solid Particles

In general, wall-correction forces are particularly important for particles at low Reynolds
numbers (Re
p
0) since creeping flow conditions give rise to a more extended region of fluid
dynamic interaction. These corrections become especially important for small distances to the
wall. Since herein it is assumed that the particle is small compared to the fluid domain (Fig.
1.20), the analytical solutions and measurements for a particle traveling exactly along the
centerline of a cylindrical pipe are not substantially important (Clift et al. 1978). It is instead
more relevant for general computational multiphase tools to focus on the fluid dynamic
interactions with a nearby planar surface. For such a geometry, the distance is defined as
length between the particle centroid to the nearest wall surface point (Fig. 6.70), i.e.,


w p w
r min x x
6.203
In the following, we will consider the near-wall drag-force corrections for the Stokesian and
inviscid limits for spherical particles. This will be followed by a discussion on the corrections
for intermediate Reynolds number particles and then non-spherical shapes.
For Stokesian flow, the change in drag force was analyzed by Faxen and a third-order
approximation for tangential and normal drag corrections of a spherical particle without
rotation moving in a quiescent fluid was later obtained by Happel & Brenner (1973) for
p p w
Re r / r 1 < < as


2 3
p p p
w ,Faxen 2 3
w w w
2 3
p p p
w ,Faxen 2 3
w w w
r r r
9 81 217
f 1
16 r 256 r 4096 r
r r r
9 81 473
f 1
8 r 64 r 512 r

= + + +
= + + +
|

6.204a


6.204b
The increase in drag force is due to confinement which causes additional stresses imposed on
the surface of the particle and wall. The commonly cited first-order approximation (just the
first correction term on the RHS) also applies for a general particle shape in that overall drag is
based on product of the correction (f=f
E
f
w
). The first-order term dominates for particles far
from the wall but is relatively small for wall distances of more than 20 diameters.
When particles are very close to the wall the flowfield between them will be primarily a
function of the gap between the two surfaces
304


gap w p
r r r
6.205
For the condition of small gaps (
gap p
r r < ), a theoretical result can be derived with Stokesian
flow and lubrication theory. In particular, Goldman et al. (1967) obtained the correction:


gap
w ,Goldman
p
r
8
f ln 0.9588
15 r

= +



|

6.206
The constant (last term on the RHS) is sometimes neglected but was included by Goldman et al.
to allow an improved asymptotic behavior far from the wall (
gap p
r r > ). At the other extreme
of
gap
r 0 , the predicted drag becomes infinite as the particle comes in contact with the wall.
This suggests that a small particle can never impact the wall. However, surface roughness and
particle inertia can allow collisions to take place, as will be discussed in 6.10. For
intermediate distances, ONeil (1964) obtained an exact solution in integral form for which
tabled results were obtained via numerical computation (i.e. it is a simplified form of a
resolved-surface simulation). As shown on Fig. 6.71, this analytical result bridges the gap
between the lubrication and Faxen limits and is well matched by the experimental data
available at Re
p
1. An empirical description of the tabled solutions given by ONeil is given by


2
3/2
p p gap
w ,O' Neil 3/2 3/2
gap p gap p gap p
r r r
9 9 16
f 1 ln
16 r r 16 r r 45 r r

+


+ + +

|
6.207
This expression tends to the Goldman theoretical limit for small gaps and the Faxen theoretical
limit for large gaps.
As there is an increase in Re
p
, the flow transitions from a Stokes layer wall interaction
(described above) to an Oseen wall interaction for which
p w p
r / r Re 1 < < . The wall
correction for the latter can be obtained by using matched asymptotic expansions (Vasseur &
Cox, 1977). This leads to a correction as a ratio of wall distance to a viscous length scale in
the form


*
w w f p w p
r r w/ Re r / 2r =
6.208
Based on an experimental study of Takemura (2004), the Oseen and Stokes regimes can be
blended together for an improvement of the Faxen force at finite Re
p
conditions. For
*
w
r <10,
this Takemura approximated the parallel drag correction as:


2 3
p p p
w * *2 *3 2 3
w w w w w w
r r r
1 9 81 217
f 1
1 0.6956r 0.03425r 0.2319r 16 r 256 r 4096 r

+ + +

+ + +

|

6.209
The terms associated with
*
w
r correspond to the Oseen corrections which dominate for
*
w
r >1.5.
By employing an overall correction of f=f
Re
f
w
&
, this result shows good predictions at Re
p
=1
(Fig. 6.71) and can be reasonably extended to Re
p
of about 3. Unfortunately, results for
smaller gap distances (to determine when transition to lubrication theory may be appropriate),
perpendicular force components, or higher Re
p
values are scare so that no corrections are put
forth herein for those conditions.
For fluid particles, the presence of a wall also induces a transverse normal force (i.e. lift)
away from the wall. The low Reynolds number solution was obtained by Vasseur & Cox
305
(1977) in integral form for solid particles. The theoretical result was approximated by
Takemura and Magnaudet (2003) with the following function:


( ) ( )
4.58
6 * *
Lw w w
9
C 5.78x10 r exp 0.292r
8


+


for
*
w
r <10
( )
2.09
*
Lw w
C 8.94 r

for 10<
*
w
r <300
6.210a

6.210b
This compared well for their experimental data at Re
p
<1, and was empirically corrected to
account for Re
p
<92 (to within 20%) based on


( ) ( )
( ) p
2 2tanh 0.01Re
0.5 0.08
Lw,Re Lw p p w p
C C 1 0.6Re 0.55Re r / 3r

+ 6.211
Effects of particle rotation and a shear flow in the vicinity of the wall will also cause changes
in the lift and torque that may be predicted (Goldman et al. 1967, Cherukat et al. 1994, and
Magnaudet, 2003).
To include the wall effects on the added mass forces, Soo (1990) assumed a potential flow
interaction, which is identical to the case of two spheres moving on opposite sides of an
inviscid wall. By examining on the kinetic energy of the fluid caused by the motion of both
spheres, the effective added mass force associated with a wall can be obtained as


3
p 1
,w f p f p
2
w w
3
2
2
p 1
,w f p f p
2
w w w
r v v
v 3 9
F v
t r 32 t 32 r
r v
v v v 3 9 9
F
t r 16 t 32 r 64 r




= +



= + +



| |
| |
|
d d
d d
d d
d d

6.212a


6.212b
In these RHS expressions, the first acceleration term is the added mass force for a particle in an
infinite fluid while the remaining terms are the third-order correction due to wall effects and
include both acceleration and velocity terms.


Fluid Dynamic Corrections for Clean Spherical Bubbles

For clean bubbles with weak wall effects (r
w
r
p
), the Stokesian drag correction (Magnaudet,
2003) is given by


2 3
p p p
w
w w w
2 3
p p p
w
w w w
r r r
3 9 27
f 1
8 r 64 r 512 r
r r r
3 9 27
f 1
4 r 16 r 64 r


= + + +



= + + +


|

6.213a


6.213b
These corrections are similar in form to the Faxen correction for a solid particle (Eq. 6.204),
though not quite a strong. The Oseen theory for a clean spherical bubble was obtained in
integral form by Takemura et al. (2002), who also obtained experimental drag corrections that
varied between these two limits depending on
*
w
r . At high Re
p
values, a thin attached
boundary layer occurs so that a viscous potential flow solution can be employed for the clean
bubble drag (6.2.5). Based on this approach, Kok (1993) extended the first-order drag
(C
D
=48/Re
p
) assuming r
w
r
p
to obtain a wall correction:
306


p
p
3
p
w ,Re 1
w
3
p
w ,Re 1
w
r
1
f 1
4 r
r
1
f 1
8 r


= +



= +


| >
>

6.214a


6.214b
Measurements of dynamic bubble trajectories (Tsao & Koch, 1997; de Vries et al. 2001)
indicate that this correction when combined with the added mass force of Eq. 6.211 is at least
qualitatively reasonable for Re
p
>200.
For the lift force of a clean bubble, the lift of Eq. 6.210 can be modified for arbitrary
particle viscosity ratio using the same dependence of Eq. 6.136, i.e.


*
2
*
* Lw
Lw,
2 3
C C
3 3

+
=

+

for
p
Re 1 < 6.215
Experiments of clean bubbles (Fig. 6.72) show that Eq. 6.214a, while derived for low Re
p

values is reasonable for Re
p
<27. The creeping flow wall impact on shear-induced and rotation-
induced lift is also discussed by Magnaudet (2003).
To consider high Re
p
conditions, the potential flow solution can be used to obtain the
inviscid lift for flow about a clean bubble moving parallel to the wall in uniform flow yielding:


4 3
p p
Lw
w w
r r
3 1
C 1
8 r 8 r


= +





6.216
This solution indicates an attraction force which is due to an increased velocity (and thus lower
pressure) as the inviscid flow passes between the particle and the wall. This is opposite in
direction to the Stokes solution lift, which acts as a repulsive force. As such, the negative lift
associated for Re
p
>80 far from the wall is consistent with the potential flow solution as shown
in Figs. 6.73 and 6.74. Closer to the wall or for intermediate Re
p
values, the lift coefficient
rapidly approaches zero. For intermediate Re
p
values, Takemura & Magnaudet (2003)
developed an empirical combination of the above inviscid result (Eq. 6.216) and the low Re
p

solution (6.165) which helps bridge the attractive and repulsive regimes.


Fluid Dynamic Corrections for Non-Spherical Particles

Gavze & Shapiro (1996) examined flows for various solid particle shapes. They noted that
the wall interactions can be complex and coupled, i.e., can give rise to a modified resistance
tensor in which acceleration in one direction can affect forces in all three directions. This can
give rise to rotational translational coupling. As such, the particles can have an oscillatory
motion close to the wall since the wall force will be modified as the shape orientation to the
wall changes. Fortuitously, Gavze & Shapiro found that the effect of the wall decreased as the
particles non-sphericity increased. As such, the spherical wall corrections can be considered as
an upper-bound to the corrections for non-spherical particles. Further, if the spherical
corrections are negligible, it is likely that the non-spherical corrections can also be ignored.



307
6.9. Particle-Particle Fluid Dynamic Forces

The final section of this chapter focuses on fluid dynamic interactions between particles.
Unlike, the case for particle-wall interactions which typically involve two solid objects, the
interstitial fluid dynamic forces among particles typically involve many interactions for each
particle. As such, they are generally modeled in an ensemble fashion (and not individually) in
terms of the effect of finite volume fraction. Typically, finite volume fraction effects occur at
low speeds and low Knudsen numbers so that the surrounding fluid can be assumed
incompressible and in continuum. Following the decomposition Eq. 6.1, the interphase forces
which account for finite volume fraction (as denoted by subscript ) can be given as:


F
int,
= F
D,
+

F
H,
+ F
,
+ F
L,


6.245
Each of these forces is discussed in the following under the assumption of weak volume
fraction gradients, with particular emphasis on the drag force.


Effect on Quasi-Steady Drag for Solid and Fluid Spheres

For the purposes of computing drag for a given condition, it is numerically most convenient
to define the correction associated with volume fraction for a fixed relative velocity and
diameter, i.e. a fixed Re
p
. If one assumes Stokesian conditions, this correction is defined as the
drag force at finite volume fraction (F
D,
) normalized by that at zero volume fraction (F
D
):


p
D, D,
D D
w Re
F C
f
F C


=



6.246
However, observations of the effect of particle concentration are most commonly observed in
terms of the terminal velocity ratio defined as


term,
term
w
w


6.247
In this definition, w
term
is the terminal velocity of an isolated particle (0) whereas w
term,
is
the path-averaged and ensemble-average terminal velocity in a quiescent fluid whose particle
volume fraction is given by but otherwise has the same particle density and diameter and
fluid density and viscosity. Since the particles are generally randomly arranged in space, path-
averaging is typically needed even at low Re
p
conditions due to variable wake and virtual mass
interactions (three-way fluid dynamic coupling) and particle contact interactions (four-way
particle-particle coupling).
For the limit of 1 and Re
p
0, some theoretical results have put forth. In particular,
Burger (1942) proposed the relation 1/(1 6.875 ) = while Batchelor (1972) similarly
obtained 1 6.55 = . As shown in Fig. 6.83, these two expressions give similar
dependencies for small volume fractions (about 1% or less) but then diverge significantly.
Furthermore, neither is quantitatively accurate when compared to experimentally obtained
velocity ratios at significant volume fractions. This is consistent with other theoretical results
(Zuber, 1964). Such differences are consistent with the departure of linearized mixed-fluid
308
viscosity theory from experimental results at finite volume fractions (Fig. 5.1). However, the
departure occurs sooner for Fig. 6.83 since the flowfield perturbations for an individual particle
become stronger at finite relative and thus becomes non-linear at even modest particle spacing.
For significant volume fractions and arbitrary Reynolds numbers, the relative velocity ratio
is generally best described with an empirical relationship. In particular, Richardson & Zaki
(1954a) expressed the velocity ratio as a function of the fluid volume fraction using the
following power form:


( )
b
1 = 6.248
This is the most commonly employed form to described finite volume fraction effects and b
has been termed the Richardson-Zaki exponent. In general, volume fraction effects
represented by this form from experiments will primarily include hydrodynamic interactions,
but also collision interactions at high particle concentrations. For Stokesian drag of solid
particles, this exponent has been observed to be approximately constant (b
o
4.5) for a wide
variety of particle density ratios as shown in Fig. 6.83. Thus, a reasonable empirical
approximation for this regime is given by:


( ) ( )
o
b 4.5
1 1 = for Re
p
0 6.249
To relate this result to the drag correction of Eq. 6.246 requires some care since the particle
Reynolds number is not held constant in obtaining the terminal velocity ratio. The following
employs the method of di Felice (1994) to obtain f

.
Consider a one-dimensional force balance in the gravitational direction (z) for the particle
momentum equation at terminal conditions in a quiescent fluid:


p p p int,
p
0 g F
z

= +

6.250
This indicates that the interphase force is balanced by the body and fluid-stress forces. The
latter is only based on the hydrostatic pressure gradient (Eq. A.56) since the surrounding fluid
is at rest. Assuming a homogeneous mixture of particles, this gradient is given as


( )
f p
p
1 g
z

= +

6.251
Since the interphase force only includes drag for steady conditions in a quiescent fluid, the
above equations can be combined to yield the drag force at finite volume fraction


( ) ( )
1 1 3 2 2
D,term, p f f term, D,term,
6 8
F d g 1 d w C

= =
6.252
This may be compared to the drag force on an isolated particle


( )
1 1 3 2 2
D,term p f f term D,term
6 8
F d g d w C = =
6.253
From these two equations, the two forces are simply related


( )
D,term, D,term
F F 1

=
6.254
From this, the drag correction can be evaluated using the two different terminal velocities
309


( ) ( )
term,
D, D,term D,term,
2
D D D
w
F F 1 C 1
f
F F C


= = =



6.255
For Stokesian conditions, the drag coefficient is proportional 1/Re
p
, and using Eq. 6.249 yields


( ) ( )
( )
o
1 b
p,term,
2
p,term
Re 1 1
f 1
Re


= = =


6.256
This indicates the Richardson-Zaki form of the terminal velocity ratio (Eq. 6.248) for creeping
flow yields a drag with separable dependencies on particle Reynolds number and volume
fraction effects, i.e. C
D
=(24/Re
p
)f

. Similarly, shape and viscosity ratio corrections can be


included so long as the Richardson-Zaki form is preserved. Note that the theoretical Batchelor
correction for small void fraction can be similarly obtained as


( )
( )
( )
1
1
f 1 7.55
1 6.55

=


6.257
Note that this result is analytically consistent for small volume fractions, while Eq. 6.256 is
empirically consistent for larger volume fractions, and thus is more commonly employed.
For finite Re
p
conditions, several investigators (e.g., Maude & Whitmore, 1958; Wen & Yu,
1966; Wallis, 1969) assumed that the drag force also has separable dependency, i.e. can be
written as a product of Reynolds number and volume fraction functions:

( ) ( ) ( ) ( )
o
Y
1 b
2
D, f f 1 p 2 p
F / Re Re 1

= F F 6.258
The LHS is proportional to
2
D p
C Re while for the RHS assumes the behavior can be described
by two function for the two primary variables. In particular, F
1
accounts Reynolds number
effects for an isolated particle based on Eq. 2.15 while F
2
accounts for finite volume fraction as
noted in Eq. 6.256. Based on Eqs. 6.255 and 6.258, the velocity ratio dependence becomes:


( )
o
Y 2
b
p,term D,term 2 2 Y
D, 0 p,term,
Re C
1
C Re

=

= = =



6.259
From this, the terminal velocity ratio at any Re
p
yields a power relation of the form:


( )
0
b / Y
1 = 6.260
This result is demonstrated in Fig. 6.84 for solid or contaminated spherical particles with
various density ratios based on the Y given by Eq. 2.18, which assumes Whites form of the
drag coefficient for solid spheres.
An overall comparison for a large range of Reynolds numbers is given in Fig. 6.85. In
general, the separable relationship is reasonable and allows the general drag coefficient to be
given using only the empirical coefficient b
o
as:


( )
( )
( )
( )
( )
o
1 b / Y D, p D, p
D p D p
F Re , C Re ,
1
F Re C Re


= =
6.261
310
Based on Eq. 6.255, this result yields the same drag correction for both linear and non-linear
drag for spherical particles of variable concentration, which is


( ) ( )
o
1 b 3.5
p
f 1 1 Re

= for all 6.262


This universal correction is obviously convenient. However, some investigators have proposed
empirical relations for b as a function of Re
p
by fitting the data, e.g., the form proposed by
Rowe (1987) shown in Fig. 6.85 and given by:


( )
3/ 4
p
3/ 4
p
4.7 2.35 0.17Re
b
1 0.17Re
+
=
+
6.263
Similar empirical expressions of Syamlal et al. (1993) and di Felice (1994), and such fitted
correlations tend to improve on the simple power relation expression (Fig. 6.85). The drag
correction for forms such as Eq. 6.263 can be approximated with Whites drag fit (Eq. 1.52) as:


( )
( )
( )
( )
p, 1 2b 1 2b D p, p,
D p
p p
24 6
0.4
Re C Re 1 Re
f 1 1
24 6
C Re
0.4
Re 1 Re

+ +
+
=
+ +
+

6.264
This expression employs a corrected Reynolds number given by ( )
b
p, p
Re Re 1

.
For clean spherical bubbles, it is possible to obtain a theoretical relationship at high
Reynolds numbers because the flow is attached. Such an analysis was put forth by Sangani et
al. (1991) by assuming isolated particle behaves as C
D
=48/Re
p
with a potential flow interaction
with other particles. The velocity ratio result can be expressed as


( )
1
1 3.91

= + 6.265
This result compares well with measured velocities for nearly spherical bubbles at high Re
p
in
distilled water as shown in Fig. 6.86. The corresponding theoretical drag correction is then


( )( ) f 1 1 3.91

= +
6.266
Since a clean spherical bubble at high Re
p
has a linear drag law (Eq. 6.49), the appropriate
power relation corrections is given by Eq. 6.256; while the contaminated case with a non-linear
drag is given by Eq. 6.263. The experimental data tends to obey the clean bubble power
relation for small volume fractions and then towards the solid-sphere relation for larger volume
fraction, which is a result consistent with the onset of flow separation around individual
bubbles as wake disturbances grow. However, these two power relation results give a
reasonable upper and lower bound for the nearly clean experimental data as shown in Fig. 6.86
(in terms of terminal velocity ratio) and in Fig. 6.87 (for the Richardson-Zaki exponents). This
is fortunate, since it does not matter which is used because their common separable form (Eq.
6.258) yields the same drag correction. Thus, one may reasonably employ Eq. 6.262 for clean,
contaminated and (thus) partly contaminated conditions. However, the purely theoretical
Sangani et al. theory is recommended for the particular case of clean conditions with volume
fraction of 10% or less.

311

Effect on Quasi-Steady Drag for Non-Spherical Solid and Fluid Particles

Data for non-spherical solid particles is primarily available for the Stokes (Re
p
0) and
Newton (Re
p
1) drag regimes. For small Reynolds numbers, experimental studies have
focused on cylindrical shapes and have shown that the particle orientation can be quite
sensitive to the volume fraction and nearby wall interactions, especially true at high
eccentricities. As a result substantial semi-random changes in the terminal velocity can occur
for volume fractions as little as 0.1% for which the effects are generally insignificant for
equivalent spherical particles. Moreover, the particles may tend to form lateral groups which
yield a stronger reduction in w
term
as increases. As shown in Fig. 6.88, this yields to
considerable variation in the terminal velocity ratio. Herzhaft & Guazzelli (1999) suggested
that b
o
9 is a more appropriate power exponent for substantially non-spherical particles, and
indeed this correlates the data better than the spherical value (for which b
o
4.5) as also shown
in the Fig. 6.88. However, the influence of non-sphericity seems to be significantly reduced at
higher Reynolds numbers. For example, terminal velocity measurements of cylinders,
hexagonal prisms, cubes and plates for Re
term
>500 and A
surf
* up to 5.7 (Richardson & Zaki,
1954b) yielded exponent values ranging from 2.01-2.55 which are similar to the spherical
value (b=2.25). Such results suggest a drag correction used with spherical particles at high Re
p

conditions, i.e. given by ( )
3.5
f 1

, is reasonable for other shapes. This is in contrast to


the increase in b found for creeping flow conditions. Unfortunately, there is insufficient data at
intermediate Reynolds numbers to establish b as a function of Re
p
and A
surf
*.
For bubbles with finite Weber numbers, the non-sphericity becomes even more
complicated since the degree of particle deformation is sensitive to both relative velocity and
volume fraction. Bubbles with various diameters (and thus deformation levels) are shown in
Fig. 6.89. The moderate deformation bubbles shown in Fig. 6.89a (2<We
p
<6) have terminal
velocity ratios which are quite similar to those observed at low deformation and Weber
numbers (e.g., Fig. 6.86), i.e. the data tends to obey the clean bubble power relation for small
volume fractions and then towards the solid-sphere relation for larger volume fraction. This
indicates that the drag corrections given by Eqs. 6.262 or 6.266 can also be extended to
moderately deformed bubbles for significant volume fractions. The high deformation bubbles
shown Fig. 6.89b (We
p
>6) behave similarly at the low volume fractions consistent with
dispersed flow (<10%). However, >10% lead to substantially different behavior with
increases in the terminal velocity. This phenomenon was also observed in resolved-surface
simulations by Sankaranarayanan et al. (2003), and is primarily attributed to two effects:
change in shape and drafting. Regarding the change in shape, both experiments and
simulations have shown that increasing causes high Weber number bubbles to transition
from an oblate shape to a spherical shape and eventually a prolate shape and this change is
consistent with a reduced drag. Furthermore, highly deformed bubbles (unlike non-spherical
solid particles) tend to line themselves up vertically and the increased top and bottom
proximity of neighboring bubbles results in drafting further reducing drag. This can lead to
>1 which is consistent with a negative Richardson-Zaki exponent (b<0). Such a trend is has
not been predicted analytically, though Sankaranarayanan et al. suggest an approximate model.
To summarize, the measured and predicted Richardson-Zaki exponents for several types of
bubbles are shown in Fig. 6.87 and are generally bounded by the clean bubble and solid sphere
312
curves. Thus, a single separable drag correction (Eq. 6.262) may be reasonable for bubbles
with low to high Reynolds numbers, clean to contaminated conditions, and low to moderate
Weber numbers (We
p
<6), i.e.


( )
3.5
f 1

6.267
Furthermore, this result is also reasonable for higher Weber numbers for the dispersed regime
(<10%).


Effect of Volume Fraction on Other Fluid Dynamic Forces

The effect of volume fraction on the added mass force for solid spheres was proposed by
Zuber (1964) in the creeping flow limit as


( )
( )
( )
1 2
1 1
c 1 3
2 1 2

+
= +


6.268
This result was obtained by simply considering a fluid mass to be accelerated as that inscribed
by the neighboring particles. Subsequently, several researchers have conducted more detailed
analysis based on various assumptions of the velocity probability distributions associated with
the neighboring bubbles. These approaches generally yield slightly different results compared
to Zubers result. For example, the 3 term in the above equation is replaced by 2.76 in the
expression given by van Wijndgaarden (1976) but by 3.32 in the expression given by
Biesheuvel & Spoelstra (1989). However, Zhang & Prosperetti (1994) showed that this result
is only formally consistent when the mixture velocity (instead of the continuous-phase
velocity) is employed in the added mass force expression (Eq. 6.184). This difference in the
relative acceleration usage can be significant for large volume fractions and large accelerations.
Using the conventional definition of the added-mass force of Eq. 6.184 for particles and fluid
accelerating independently, the added mass coefficient of Zhang & Prosperetti (1994) to first-
order is:


( ) ( )
,
1
c 1 c 1
2

= = 6.269
The result is also reasonable for high Reynolds number conditions though it can be sensitive to
oscillations frequency (ten Cate & Sundaresan, 2006). Similarly, the inviscid lift is
proportional to the effective mass of the available fluid (Prosperetti, 2007) as is the history
force (ten Cate & Sundaresan, 2006):


( )
L, L
F F 1

=
( )
H, H
F F 1

=
6.270a

6.270b
These relationships may only expected to be reasonable for solid, spheres at nearly dilute
concentrations and little is known about the appropriate forms at other conditions.
The final fluid dynamic force on the particle to consider a finite volume fraction is the fluid
stress. For a single particle in an infinite incompressible medium, the fluid-stress force is given
by Eq. 6.190. For both inviscid flow (Drew, 1983) and viscous flow (Crowe et al. 1998), the
313
fluid stress acting on a particle surrounded by a flow with finite volume fraction is simply
equal to that for zero volume fraction (Eq. 6.191), i.e.
( )
S, f p ij
= p K

+ F
6.271
This results from the assumption that the fluid stresses (which, by definition, neglect the
presence of the particle upon which they are being applied) effectively act throughout the
domain, e.g., for bubbles in a liquid, the contributions due to surface tension contributions
cancel on average if the control volume is much larger than the particle volumes. This
implicitly assumes that the stresses are independent of volume fraction, which is an important
but subtle result of proper averaging within a mixed-fluid cell (Prosperetti, 2007). Other forms
are possible depending how pressure in the multi-phase mixture is defined (Zhang &
Prosperetti, 1994), but the above formulation is consistent with a conservative Eulerian PDE
for the particle momentum, which can be shown by physical reasoning. For example, consider
a fluid and a particle at rest with no gravitational forces but with a finite particle concentration
gradient. In this case, the fluid stress should be zero, i.e., a concentration gradient should not,
by itself, introduce non-physical momentum source.
In summary, one may write the fluid dynamic force for solid monodisperse spheres in
dispersed conditions in a manner similar to that of a single particle (Eq. 6.1) as
( )( )
surf int S D L H S
f 1

= + = + + + + F F F F F F F F
6.272
Here, a sub-script is implied for the surface and interphase forces, i.e. they are assumed to
include finite volume fraction effects. It may also be reasonable to employ this form in an
approximate manner for various Reynolds numbers, size distributions, shapes, particle surface
conditions, etc. but there is insufficient data and theory to ensure it is generally reasonable. If
one neglects Brownian and thermophoretic forces and assumes spin-equilibrium lift, this form
of the surface force can be combined with Eqs. 1.23, 6.5 and 6.168a to describe the particle
acceleration

( )( )
( )
( )
( )
( )
( )
D H L,eq p f coll f f @p
p p f p f p f
f 1 c 1
t t c c c


+ + +

= + +
+ + +
F F F g +F u
v
D
d
d D
6.273
The three groups of terms on the RHS include: 1) drag and lift forces adjusted for particle
concentration, 2) gravitational and collision terms, and 3) fluid stress and added mass terms
associated with continuous-phase acceleration. The one remaining force of Eq. 6.273 which
has not yet discussed in detail is that of the collision force. This force is discussed in the next
two sections, first with respect to particle-wall interactions and then with respect to particle-
particle interactions.


6.10. Particle-Wall and Particle-Particle Collision Effects

Particles colliding with walls or other particles are important in many multiphase flows,
especially for systems which involve coating, deposition, filtering, erosion, and four-way
coupling as discussed in 1.2. The collision point-force models discussed in this section
involve only two bodies, which are consistent with the dispersed flow assumption of 3.1
(whereas dense flows must generally account for simultaneous interactions between three or
more bodies). The impacting particle characteristics will be designated with a subscript 1,
314
while the target particle characteristics will be designated with a subscript 2. Particle-wall
collisions and particle-particle collisions can be treated in a unified manner since the limit of
the target particle mass tending to infinity (m
2
) is equivalent to a collision with a wall.
This section discusses a framework to determine the post-collision conditions based on the
pre-collision conditions and the particle, wall and fluid properties. These discontinuous jump
relations for the net interaction include fluid dynamic and contact force effects based only on
the net changes. Such a method is commonly known as the hard-sphere model since the
particle remains spherical shape for the analysis. One may alternatively employ a continuous
treatment for the unsteady particle deformation throughout the collision interaction using small
time-steps and discretized differential equations for the dynamics and the interior particle
stresses. This is known as the soft-sphere model which is often only practical for a modest
number of collisions. In a sense, the hard-sphere model is consistent with a point-force
approximation while the soft-sphere model is consistent with a resolved-surface approach with
time-dependent deformation (Crowe et al. 1998). Herein we will focus on the hard-sphere
model approach consistent with the intent of this chapter. Furthermore, issues associated with
van der Waals forces (due to molecular interactions between solids), chemical interactions, and
electrostatic charges will not be included but are discussed by Crowe et al. (1998). This model
involves both contact forces and fluid dynamic forces expressed in terms of the pre- and post-
collision characteristics as described below.


Collision Jump Characteristics

The case of a particles colliding with a wall in the hard-sphere model can be described as
per Fig. 6.74, where the incoming translational and angular velocities are defined as v
in
and
in
.
The pre-collision state is also defined to be before any fluid dynamic interaction with the wall
have occurred, e.g., a diameter or more away from the surface. The case for two particles
colliding is similarly defined but also includes non-zero target particle velocities (Fig. 2.47).
Once the interaction is completed (in terms of both collision and fluid dynamic forces), the
translational and angular velocities are defined as v
out
and
out
. The jump change between
these two states can be written as


out in
out in


v v v


6.217a

6.217b
Thus the changes in the strike particle velocity translational and angular velocities are v
1
and

1
and similarly defined for the target particle. If the particle mass remains constant, the
translational collision impulse that the target particle (of mass m
p2
) exerts on the impacting
particle (of mass m
p1
) is based on the particles change in velocity. Similarly, the angular
collision impulse is based on the moments of inertia (Eq. 6.154) and the particles change in
angular momentum. These two impulses can be written as


coll p1 1 p2 2
coll coll p1 1 p1 p2 2 p2
m m
/ r / r
=
=
I v v
n I I I

6.218a

6.218b
315
The angular collision impulse includes the collision vector (n
coll
) which is normal to the
contacting surfaces. For spherical particles with centroids given by x
p1
and x
p2
, this vector is
defined as:


p2 p1
coll
p2 p1

x x
n
x x

6.219
If the target body is a wall then n
coll
is simply perpendicular and outward to the wall surface
(n
w
). The collision vector allows us to decompose a centroid-based particle velocity into its
normal and tangential components


( )
coll coll



|
v v n n
v v v

6.220a

6.220b
These components are illustrated in Fig. 6.74 for a particle impinging on a wall.
For a collision of two bodies which may be spinning, the incoming centroid relative
velocity of Eq. 2.91 can be extended to describe the relative velocity of the surfaces, i.e. the
relative interface velocity between the two bodies just before collision. This relative velocity
includes both translational (i.e. centroid-based) and angular components of velocity:


( ) ( )
coll 1 2 p1 1 p2 2 coll
r r + v v v n
6.221
For the case of a single spinning particle impacting a wall,
in in in
coll p wall
r = v v n
. For two
particles impacting without spin,
in
coll p-p
= v v
(based on Eq. 2.91). The collision velocity can be
broken down into surface normal and tangential components:


( ) ( )
coll coll coll coll 1 2 coll coll 1 2
coll coll coll

= =


v v n n v v n n v v
v v v
|

6.222a

6.222b
For example,
in
coll p-p
= v v
for two particles without spin (as shown in Fig. 2.47) and
in in
,coll coll
= v v
for a head-on collision with a wall. In order to close this system of equation for the hard-
sphere model, the change in the normal and tangential momentum must be modeled. This will
be discussed in the following for: a) normal collisions of spherical particles, b) oblique and
spin interactions for spherical particles, and c) collisions of non-spherical particles or irregular
surfaces.


Normal Collisions for Spheres

A normal-wall collision is defined as a particle approaching a wall with a translational
velocity direction that is perpendicular to the contact surface (
in
=90
o
) and without spin (
in

=0). In this case, the normal momentum is reversed in direction and reduced by an amount
equal to the normal coefficient of restitution (e

):


out in

= v v e for a hard wall collision 6.223
316
Thus, a coefficient of restitution of unity for a particle impacting a wall simply indicates that
the normal and tangential velocities change sign but keep the same magnitude, which is a
perfectly elastic collision. However, any energy losses will result in values less than unity.
For a binary collision, the coefficient of restitution can be combined with the impulse
equations to obtain the normal velocity changes for each particle as:


( )
in
p2 ,2 p1 ,1 coll coll
m m 1 m

= = + v v v e
6.224
This equation employs the collision mass (m
coll
) defined as:


p1 p2
coll
p1 p2
m m
m
m m

+

6.225
For a collision of particles of equal size, m
coll
is m
p1
; while for a particle impacting a rigid
wall, m
coll
=m
p1
so Eq. 6.223 is recovered.
As noted above, a perfectly elastic interaction is defined by e

=1, while inelastic collisions


will yield e

<1. The two most common inelastic regimes are associated inviscid rebound and
viscous rebound. Both these conditions are visco-elastic with a dissipation of energy but no
permanent damage to the structure of the particle (or wall). However, a third regime arises for
solid particles impacting with high relative velocities which is associated with plastic
deformation. These three regimes are discussed below.
For the inviscid regime, the particle coefficient of restitution is not affected by the
surrounding fluid and is instead related only to energy dissipation within the bodies (within the
particles and/or wall). This coefficient of restitution is denoted e

and is often referred to as
the dry limit since it occurs in the absence of liquids, i.e. in the presence of gasses for which
the viscosity is not significant to the collision. It is approximately independent of the impact
velocities and only a function of the material properties, but can also depend weakly on
roughness. Typical empirical values (based on experiments) are listed in Table 6.6 for both
solid and fluid particles. Note that steel is one of the most efficient materials in this regard,
which is the reason it makes such a good demonstration of momentum conservation in the
Newtons cradle toy based on an array of suspended steel balls.
For the viscous regime, additional energy is dissipated due to the continuous-phase fluid
viscosity between the contacting surfaces so that e

<e

. In general, liquids are used as the


surrounding fluid in order to observe significant effects of viscosity, which become most
important in the short time when the gap between the surfaces becomes very small. Davis et al.
(1986) used lubrication theory to determine the pressure distribution and losses in the thin film
which drains between the two surface just before and after the contact time. A key parameter
which determines the influence of viscous effects was found to be the impact Stokes number
(St

), which compares the time it takes to slow the particle down (which can be approximated
as the Stokesian response time) to the drainage time-scale (which is the time required to push-
out the fluid as the gap between the particle and the surface goes to zero). The latter can be
estimated with the collision time scale (v/d) so that:


( )
in
p f
f
+c v d
St
9

6.226
317
Defined in this way, the added mass effects are included (Loth, 2000; Legendre et al. 2005).
The influence of the impact Stokes number on the normalized restitution coefficient of various
solid particles is shown in Fig. 6.75 along with the theoretical prediction of Davis et al. (1986)
for steel and glass. Here it can be seen that the limit of high impact Stokes numbers is
consistent with the inviscid limit and e

. At lower St

values, viscous effects become more


important and the particle inertia is more likely to be halted. This yields reduced rebound
velocity ratios until finally there is no rebound at all for St

of order unity or less.
If two particles are considered (instead of interaction with a wall), the response time is
controlled by that of the smallest particle whereas the drainage can be expressed in terms of a
collision diameter (d
coll
) as


( )
1 2
coll
1 2
in 2
p f ,coll small
f coll
d d
d
d d
+c v d
St
9 d


6.227a


6.227b
Thus, a particle impacting another particle of equal diameter will have an effective collision
diameter which is half the particle diameter. The commensurate doubling of the impact Stokes
number is consistent with a doubling of the fluid space between the surfaces so that drainage
and a rebound velocity is more likely than that for a solid wall impact with this same diameter.
For fluid particles, Legendre et al. (2005) modeled the bouncing deformation and velocity
as a simple mass-spring dissipative system. The resulting restitution coefficient is given by:


ref
St St
exp exp
St 35




=



e
e

6.228
Note that St
ref
35 is an empirical constant set to correlate the available solid and fluid
particle data (Fig. 6.75 and 6.76). However, St
ref
is generally dependent on several parameters
including particle roughness, density, shape and elasticity (Davis et al. 1986). For fluid
particles, there may be an influence of capillary number (Legendre et al. 2005) and one must
also be aware that the conditions allow for a Type II rebound as discussed in 2.4.5, i.e. the
impact Weber number is high enough to avoid coalescence but is low enough to avoid break-
up (e.g., 0.2<We
p-p
<3 for bubbles in water). Therefore, Eq. 6.228 is reasonable but only
approximate.
The final normal collision regime for rebound is that associated with plastic deformation.
Such deformation becomes important when the impact velocity exceeds the yield velocity
(v
yield
). The theoretically yield velocity for spheres can be obtained by equating the initial
kinetic energy with the elastic work done up to the point of material yield (Labous et al. 1997)


( )
2
5
yield yield p
v / 3 K / 10 = E 6.229
This result includes dependence on the particle material in terms of Youngs modulus of
elasticity (E) and the yield stress (K
yield
). The yield velocity can be related to the plastic
velocity, which is the velocity beyond which the restitution changes from elastic to plastic and
can be approximated (Johnson, 1985) as
318


( )
4
plastic yield
v 1.94v

e 6.230
As such, plastic deformation is a function of the material properties but not particle size.
Plastic velocities for nylon particles are on the order of 10 m/s, compared to only 0.1-1.0 m/s
for steel spheres (indicating a practical upper limit of velocity for the Newtons cradle).
In the limit of no particle breakage or cracking, the above relations can be used to
prescribe the restitution of particle and/or surfaces with no viscous effects:



e e for vv
plastic

( )
1/ 4
in
plastic
v / v


e e for v>v
plastic
6.231a

6.231b
This empirical correlation gives reasonable predictions as shown in Fig. 6.77 for the visco-
elastic and plastic regimes over a wide range of impact speeds for steel particles. Similar
results have been obtained for other materials and for both particle-particle and particle-wall
collisions in a gas (Walton, 1993, Labous et al. 1997; McNamara and Falcon, 2005).


Oblique and/or Spin Collisions for Spheres

For an oblique or a spin collision, the normal momentum change will still be governed by
the above relations. However, surface tangential momentum will change its magnitude. For a
wall collision, this change can be characterized by the tangential coefficient of restitution:


out in
,coll ,coll
= v v
| | |
e for a hard wall collision
6.232
This jump relation for momentum change can be combined with the impulse equations to
obtain the surface tangential velocity changes for a binary collision with different masses.
Using Eqs. 6.221-6.225, these changes can then be decomposed into centroid and angular
velocity changes for two particles as:


( )
( )
in
p2 ,2 p1 ,1 coll coll
in
p2 2 2 p1 1 1 coll coll coll
2
m m 1 m
7
10
m d m d 1 m
7
= = +
= = +
v v v
n v
| | | |
| |
e
e

6.233a

6.233b
Therefore, all outgoing particle velocity components are known once the restitution
coefficients are specified.
The tangential coefficient of restitution can be related to the collision velocity ratios
defined below as the tangential velocity normalized by the incoming normal velocity as:


( )
in
1
coll in in
in
@ 0
in
coll
out
coll out in
in
coll
v
tan
v
v
v


=

=
|
|
|
e

6.234a


6.234b
319
Note that the incoming collision velocity ratio (Eq. 6.234a) can be related to the impact angle

in
(defined in Fig. 6.75) for the condition of no incoming spin (
in
0 = ), but will have an
outgoing spin if colliding obliquely due to the torque applied by the wall. For an oblique
collision with the hard-sphere model, two different regimes can occur depending on the impact
angle: roll and slide. The roll regime generally occurs at near-normal collisions (
in
~90
o
)
while slide generally occurs at grazing collisions (
in
~0
o
). Note that this refers to the relative
velocity at the surfaces since a near grazing condition will actually result in a substantial
angular velocity about the particle centroid, but this leads to substantial local slip between the
surfaces. In contrast a particle which impact with a nearly-normal condition may only produce
a small amount of rotation about the particle centroid after the collision, but there will be little
slippage between the surfaces. The transition between the roll and slide regimes is defined by
in
crit
. For
in
0 = , this is consistent with a critical impact angle defined by
in
crit
.
The restitution coefficient for the roll regime is denoted as
roll |
e so that


out in
roll
=
|
e for
in in
crit

6.235
This coefficient is approximately independent of the collision angle and the surrounding fluid
and is instead determined by the material properties and surface conditions of the collision
bodies. This coefficient is bounded by
roll
1 0
|
e , with a typical value of 0.35. Since this
coefficient is always positive, it is associated with a reversal of direction for the tangential
surface velocity (Eq. 6.232). The roll regime is shown in Fig. 6.78 for both particle-particle
and particle-wall interactions for which it can be seen that the collision velocity ratios are
approximately linearly related in this range, consistent with a constant value for
roll |
e . Also
shown is the soft-sphere model developed by Maw et al. (1976), which employs a temporal
integration of the deformation based on Hertzian theory (though the soft-sphere model still
requires specification of c

). The soft-sphere model tends to be more accurate and captures the


initial micro-slide region (which can occur at nearly normal collisions, i.e.
in
1 < ) as well as
the more gradual slope change behavior near transition. However, the hard-sphere model is
reasonably representative and much more widely used due to its simplicity and computational
efficiency.
The slide regime has a significantly different behavior as the tangential velocity is
dominated by the sliding coefficient of friction. In particular, the magnitude of tangential
impulse is equal (by Coulombs law) to the product of a coefficient of friction (c

) and the
normal impulse. As a result, the collision velocity ratios for sliding are related by


( ) ( )
out in
7 / 2 c 1

= +e for
in in
crit

6.236
Foe shallow angle collisions, the velocity ratio is approximately preserved, i.e.
out in
such that 1
|
e . This result is shown in Fig. 6.78 and gives good predictions
above the critical velocity ratio. The coefficient of friction is primarily a function of the
surface (and not material) properties of the particles in dry conditions, where it is denoted by
c

, but is sensitive to viscosity for collisions in liquids. In particular, the friction coefficient
for smooth particles can be significantly reduced compared to the dry value. For example,
steel particles with a roughness of 10
-6
diameters have wet c

values on the order of 0.02-0.001


depending on
f
, much less than their c

value (about 0.11). While Joseph and Hunt (2004)


320
discuss a semi-empirical relationship to qualitatively describe this reduction in terms of the
local shear strain, more work is needed to allow quantitative predictions. As such, there is not
a robust model for c

in wet conditions indicating empirical evidence must often be employed.


However, if the particle has a surface roughness which is on the order of theoretical minimum
liquid thickness (based on elasticity and Stokes number), then c

can be used as a reasonable
estimate for c

in wet conditions. For example, this is a generally good assumption for glass
particles.
To determine the transition point between these two regimes, one may define the critical
velocity ratio as the point where the roll and collision regimes intersect, i.e.
in
crit
occurs
when
in in
roll slide
= . This can be determined by equating Eqs. 6.234b and 6.236:


( )
( )
in
crit
roll
7c 1
2 1

+
=
+
|
e
e

6.237
Thus, the restitution needed is given as


( )
roll in
7c 1
min , 1
2

+
=


| |
e
e e
6.238
This result is consistent with the data of Figure 6.78 for which
roll |
e ~0.35 is valid up to
in
crit

given by c

~0.22, beyond which the slope changes from -


roll |
e to +1 for the slide regime. If one
considers the collision of a single non-spinning particle with a wall, the outgoing particle
tangential velocity and spin are given by Eqs. 6.233 and 6.225 as


( )( )
( )( )( )
out in
out in
1 2 / 7 1
10 / 7 v / d 1

= +

= +
v v
| | |
| |
e
e
for a hard wall collision
6.239a

6.239b
As such, an oblique collision will yield positive spin (as defined by Fig. 6.75) for both the roll
regime (
,roll
0.35
|
e ) and the slide regime (
,slide
1 0.35 < <
|

e ). In the limit of very shallow


angles (
in
1 > ), the incoming tangential velocity becomes very high but a reduced tangential
restitution coefficient due to sliding limits the amount of outgoing particle spin.
To demonstrate the robustness of the hard-sphere model, several particle materials and
different fluid viscosities are considered in Fig. 6.79, where the velocity ratios are normalized
by the critical value for each set of conditions. For the rolling regime,
roll |
e 0.35 is assumed
for all cases since this reasonably collapsed the data for a wide variety of materials in both dry
conditions (surrounding fluid is a gas) and wet conditions (surrounding fluid is a liquid). The
inviscid restitution and friction coefficients (

e and c

) are primarily a function of the


particle material (the values listed in Table 6.6 were used in the figure) while

e is reasonably
described in wet conditions as per Eq. 6.228. The wet c

values were based on experimental


results, except for the glass particles which were based on the dry friction coefficient
depending on
f
. In general, the overall prediction of the collision velocity ratios (which
determine the post-collision quantities) is generally accurate to within experimental scatter. If
the walls are rough and/or the particles are non-spherical, then the local collision angle must be
321
obtained based on the detailed geometry to capture these effects. The specific orientation for
an individual particle is often difficult to track so that such effects are often ignored or are
modeled numerically using a stochastic approach, as discussed in 7.2.8.

322

7. Point-Force Numerical Techniques

For the separated fluid method using a point-force representation, separate transport
equations are employed for the continuous-phase and the dispersed-phase, such as the
incompressible momentum equations in Table 4.1. For the continuous-phase, various forms of
the continuous-phase PDE and state equations without particle coupling are discussed in
Appendix A, while a two-way coupled PDE set is given by Eqs. 4.30-4.36. These continuous-
phase equations are typically solved with an Eulerian discretization as discussed in Appendix B
based on the flow condition, e.g., explicit conservative formulations for compressible
continuous-phase flows and implicit formulations with pressure-solvers for incompressible
continuous-phase flows.
The dispersed-phase equations are generally solved in either Eulerian or Lagrangian format.
Some guidance of the format to choose was provided in Chapter 4 based on physics (Fig. 4.16)
and numerics (Fig. 4.18). In general, Lagrangian approaches are typically more physically
robust if the following flow properties are of interest: particle break-up, particle reflection from
surfaces, counter-flowing particles, high particle concentration gradients, significant
polydispersivity, particle-particle interactions, history and turbulent diffusion. In addition, the
Lagrangian method is uniquely capable of capturing path-dependent processes such as particle-
particle and particle-wall collisions, as well as particle history force effects. However, the
Eulerian approach can allow a consistent numerical scheme and grid to allow increased
accuracy for interphase and intra-phase coupling and is generally more efficient for large
numbers of particles (so long as there is not strong polydispersivity of particle size within a
computational volume). Note that there has been some work towards a mixed Eulerian-
Lagrangian description for the particles (e.g., Nihei & Nadaoka, 1988; Sivier et al. 1996) but
mixed approaches are still uncommon as they do not yet combine the attributes of both
Eulerian and Lagrangian methods.
The techniques in this chapter are roughly arranged in terms of increasing particle size as
per Figs. 4.16 and 4.19. The first part of this chapter deals with Eulerian techniques for the
dispersed-phase while the second part deals with Lagrangian techniques. In each of these two
sections, it is assumed that a point-force approximation and a prescribed equation of motion are
employed (as discussed in Chapter 6). Each of these two sections will initially discuss one-
way coupled descriptions of the particle phase followed by approaches for handling two-way
coupling as well as effects particle-wall and particle-particle collisions. The third section of
this chapter will discuss distributed-force methods (reasonable for particles whose sizes are on
the order of the spatial discretization of the continuous-phase). Distributed-force techniques
are used to modify the Lagrangian point-force technique to allow interaction over a finite
volume of the continuous-phase region for each particle. It should be noted, that all three
approaches require analytical and empirical expressions of the Chapter 6 point-force couplings.


7.1. Eulerian Point-Force Formulations

In the following, Eulerian treatments are considered where the particles are treated as
point-forces and averaged within a cell. This allows a continuum approach (in terms of mass
323
or volume fraction) so that the particle concentrations are described with transport PDEs. The
first section reviews the basic mathematical description of the Eulerian-Eulerian formulation in
the context of point-force particle for the transport PDEs and the equation of state. This is
followed by forms of the equations that are best suited to the flow-field conditions, e.g.,
specialized for various levels of compressibility for the continuous-phase and the dispersed-
phase.

7.1.1. Generalized Eulerian-Eulerian PDEs and Numerical Methods

There are many forms of the dispersed-phase and continuous-phase PDEs that have been
used for the Eulerian-Eulerian formulation. To allow for compressibility of both the
continuous and dispersed phase, the Eulerian mass, momentum and energy transport equations
are most generally given in terms of the continuous-phase density () and the particle volume
fraction (). The transport PDEs can be fundamentally derived from using a differential
control volumes which is large enough to contain many particles so that one may consider to
be a continuum throughout the domain and yet is small enough such that spatial gradient in are
approximately independent (Eq. 4.16).
For Eulerian-Eulerian treatments, the mass conservation equations for the dispersed and
continuous-phases with Brownian diffusion can be obtained from Eqs. 4.29a, 4.30 and 4.73 as:


( ) ( )
( )
[ ]
p p j
p
Br p
j p
f j
p f
j p
v
m
t x
(1 ) u
m (1 )
t x


+ = +





+ =

`
`

7.1a


7.1b
Recall per Eq. 4.28 that u
j
includes two-way coupling effects averaged over a cell volume but
does not include local influence of an individual particle (as discussed in 1.3).
A diffusion term due to particle collisions can also be added to the RHS of Eq. 7.1a, with a
form similar to that for the Brownian diffusion except with
coll
(Gidaspow, 1994). However,
the ensemble-averaged Eulerian collision diffusivity must be related to inter-particle velocity
fluctuations (e.g., due to turbulence or continuous-phase velocity gradients) so that collision
effects are more readily included in the Eulerian particle momentum equation (Crowe et al.
1998). However, these terms are neglected herein for the Eulerian-Eulerian treatment since our
focus is on dispersed-phase flows where collision effects will not dominate. On the other hand,
the weak collision effects are conveniently included in the Eulerian-Lagrangian treatment, as
will be discussed in 7.2.
The momentum equations can be given in conservative form with ensemble-averaged stress
(RHS) terms. For the dispersed-phase momentum, the fluid-stress has the form of Eq. 6.271
and this is generally separated from the other fluid dynamics terms so that the conservation
equations of Eqs. 4.29b and 4.31 are generally given in terms of the interphase coupling force.


( ) ( )
p i p i j
i j p i int,i
p i
j i j p p
v v v
K m v F
p
g
t x x x

+ = + + +




`


7.2a

324


[ ]
( )
f i j i j f i
f i int,i p i
j i j p
(1 ) u u K (1 ) u
p
(1 ) g F m v
t x x x



= + +




+ `

7.2b

The continuous-phase momentum transport includes a fluid stress term as the conservative
counterpart of the dispersed-phase fluid stress. The RHS of both equations include the
momentum added to the particles due to mass transfer. The energy transport equations also
include effects of thermal gradients and heat flux due to phase change (Eqs. 4.29c and 4.35):


( ) ( )
( ) ( )
p p p p j
eff f p p phase p
j p
e e v
T m +e
t x


+ = + +


`
` k Q h 7.3a



[ ]
( )
( ) ( )
( )
f tot j i j ij i
f tot
f j j
j j
i eff f
i j ij p p phase p i i int,i i
j j p
(1 ) e u (1 )(K p )u
(1 ) e
(1 ) g u
t x x
v T
K p (1 ) m +e + v v F v
x x



+ = + +


+ + +


`
`
k
Q h


7.3b


Based on Eqs. 1.1, 1.88 and 1.91, the mass and heat transfer terms can be written as


( )
f p@f p
2
p p p
p
f
p f 2
p p p
6 Sh m
B
d
6Nu
T T
d

`
`
Q
k

7.4a


7.4b
The mass transfer is driven by the Spalding number (B of Eq. 1.85) which accounts for the
difference between surface and far-field vapor pressures saturated conditions and by the
Sherwood number which accounts for finite Re
p
effects (Eq. 1.94a). The heat transfer is driven
by the relative temperature between drop surface and the far-field, the Nusselt number which
again accounts for finite Re
p
effects (Eq. 1.94b), and the thermal conductivity which is
typically based on the volume fraction (Eq. 4.4):


eff p f
(1 ) = + k k k
7.5
Finally, these equations must be supplemented by an equation of state which relates the
individual phase densities to the overall pressure (e.g., 5.1.3).
The dispersed-phase transport equations can be solved with the conventional finite-
difference, finite-volume finite-element or spectral methods. Typically, the Eulerian dispersed-
phase PDEs are solved with the same computational grid, discretization and numerical scheme
as used for the continuous-phase PDEs. For one-way coupling conditions, it should be kept in
mind that the dispersed-phase PDEs are hyperbolic because particle characteristics can only
propagate downstream in the direction of particle velocity (similar to the case of supersonic
flow for the continuous-phase). From a numerical standpoint, this indicates that explicit
upwind methods are appropriate for the dispersed-phase (Murai & Matsumoto, 1999).
An advantage of one-way coupling for steady flows is that the dispersed-phase PDEs may
be numerically obtained after (a posteriori) the continuous-phase flow solution is obtained. As
such, several dispersed-phase solutions can be obtained for a single continuous-phase
(converged) solution. In contrast, unsteady continuous-phase flows generally require that both
325
phases be simultaneously integrated in time (since storing all temporal realizations of the
velocity field is typically impractical), generally with the same time-step intervals. For two-
way point-force coupling, a consistent numerical methodology (including spatial discretization
and temporal integration) should be used for both phases to ensure interphase transfer is
consistent, i.e. equal and opposite in a discrete sense.
As with single-phase CFD, another key-issue is the compressibility of the phases. The
three most common conditions include: incompressible particles in a compressible gas,
incompressible particles in an incompressible fluid, and compressible bubbles in an isothermal
incompressible liquid (though some comments will also be made for the condition when the
liquid is weakly compressible). In the following, these three regimes will be discussed for
along with the appropriate numerical methods. This is followed by comments on the numerical
treatment for wall collisions and turbulent dispersion, which can be difficult to simulate with
Eulerian-Eulerian approaches.

7.1.2. Incompressible Particles in a Compressible Gas

In many applications the continuous-phase is a gas while the dispersed-phase is an
incompressible solid particle or droplet (constant
p
). In such conditions, the particle density
ratio is large ( * 1) so that the volume fraction of the particles is typically very small or
negligible (1). Using these assumptions, one may write the multiphase conservative
transport equations of Eqs. 7.1-7.3 in a flux-based form (similar to Eq. B.75):


y
x z
t x y z


+ + + =

Q
Q Q q
S 7.6
The conservative transport variables and fluxes can then be given by


( ) ( )
f y
f x
f
2
f x y
f x f x
2
f y
f y f x y
f y z
f z f x z
y f tot f tot x f tot
x y
y x
2
x
x
y
x y
z
x z
p
p x
u
u
u u
u p u
u p
u u u
u u
u u u
u e p e u e p
, ,
v v
v
v
v
v v
v
v v
e
e v




+


+






+ +


= = =




q Q Q
( )
f z
f x z
f y z
2
f z
z f tot
z
z
x z x y
2
y z
y
2
z
y z
p z
p y
u
u u
u u
u p
u e p
,
v
v v v v
v v
v
v
v v
e v
e v










+


+


=







Q

7.7
In this form, the continuous-phase and dispersed-phase equations can be treated with explicit
conservative approaches. For the source terms, particle Brownian motion and gas-phase body
forces are typically neglected because of the speeds associated with compressible gas flow:
326


( )
( )
( )
( ) ( ) ( )
p p
xy
xx xz
x int,x p x
p
xy yy yz
y int,y p y
p
yz
xz zz
z int,z p z
p
i j i eff f p p phase p i i int,i i
j p
m /
K
K K
g F m v
x y z
K K K
g F m v
x y z
K
K K
g F m v
x y z
K u T m +e v v F v
x


+ + +

+ + +


+ + +

=

+ + + +


S
`
`
`
`
`
` k Q h
( )
( )
( )
( )
p p
x int,x p x p
y int,y p y p
z int,z p z p
p p phase p p
m / m
g F m v / m
g F m v / m
g F m v / m
m +e / m
















+ +


+ +


+ +



+


`
`
`
`
`
` Q h

7.8

The equations must be complemented by an equation of state, e.g., p=
g
R
g
T

from Eq. A.21.
The interphase-force for particles in a gas with * 1 is often given as:


int,i D,i L,i T,i Br,i
F F F F F

= + + +
7.9
This form neglects the added-mass, fluid stress and history force effects since these effects tend
to only be significant when the particle density is on the order of or less than that of the
surrounding fluid. While some formulations include the fluid stress and added-mass effects for
conditions with high particle accelerations, the history force is almost always neglected since it
requires long-time integral information along the particle path which is not readily available in
an Eulerian formulation (and is best included by a Lagrangian formulation for the particle-
phase as noted in 7.2). Furthermore, the interphase force neglects the three-way coupling
effects since volume fraction effects are assumed to be small (1).
If wall boundary layers are not important, the continuous-phase shear stress and thermal
conduction terms are often neglected (equivalent to setting K
ij
=k
eff
=0). In this case, the source
term will only include viscous effects associated with two-way coupling from particle drag and
thermal conductivity effects only associated with heat transfer to the particles (Sommerfeld,
1987). In this limit, some compressible multiphase flow formulations use a total particle
energy (e
p,tot
) which combines internal energy and kinetic energy as


1 1
p,tot p i i p,p p i i
2 2
e e v v T v v + = + c
7.10
so that the energy transport equation and its source term can be written as:


( ) ( )
( )
p,tot p,tot j
p p phase p i i int,i i
j p
e e v
m e v v F v
t x m


+ = + + + +


`
` Q h 7.11
327
This form integrates kinetic energy effects similarly to the continuous-phase energy transport
equation given by Eq. A.13.
For the Eulerian-Eulerian multiphase formulations, spatial discretizations and schemes are
similar to those applied for single-phase compressible flows. Generally, the homogenous
PDEs (neglecting the source/sink term) are solved first. This is then followed by an ODE
integration of the source term contributions. For example, in a predictor-corrector scheme (e.g.,
Eq. B.85), the convective fluxes for both phases are computed for the predictor step sing a
forward-difference scheme to n+1/2, after which the viscous and interphase source/sink terms
are added, generally with a centered high-order scheme. The corrector step (e.g. Eq. B.77)
then employs the half-time-step fluxes to obtain the n+1 solution to the homogenous problem,
after which the viscous and interphase source/sink terms are added. For the gas-phase
equations, the fluxes can be obtained using the conventional single-phase approaches (Chung,
2002; Knight, 2006).
However, as mentioned in the previous section, the particle equations often require special
treatment due of the lack of a pressure term. As a result, the Jacobian of the flux vector has
characteristic velocities equal to the particle velocity (whereas the gas-phase Euler equation
can have information which propagates both upstream and downstream). Formally, this
indicates that the particle transport equations should only involve upwind information, e.g., the
MacCormack method applied in the v direction (Wang, et al. 2001) or flux-vector splitting
using only the flux components in the v direction (Chang et al. 1996). Similarly, the Riemann
solver must be fully upwind, which simplifies the construction so that the interface fluxes can
be determined by considering the four local analytical solutions associated with the different
signs and magnitudes of the left and right velocities of an interface. As discussed by Saurel et
al. (1994), conditions with like velocity signs are assumed to have interface properties equal to
the upstream values, conditions with opposite signs and diverging velocities are assumed to
leave a void, and conditions with opposite signs and converging velocities are assumed to
combine both fluxes. Thus, the particle fluxes in the x-direction are:
( )
( )
( )
( ) ( )
x,p left right
i 1/ 2,left
x,p left right
i 1/ 2,right
x,p
i 1/ 2
x,p x,p left right
i 1/ 2,left i 1/ 2,right
v , v 0
v , v 0
v 0, v 0
+
+
+
+ +
>
<
=
> <
Q
Q
Q
Q Q
if
if
+ if
left right
0 v 0, v 0

< >

if

7.12
Similar relations can be used for the y-direction and z-direction particle fluxes. To obtain the
left and right states needed for the particle fluxes, one may use a forward-differencing scheme
based on the flux Jacobian with either the non-conservative variables (Collins et al. 1994) or
the conservative variables (Saurel et al. 1994; Chang et al. 1996).
For flows with steep gradients, 2
nd
-order accurate schemes can yield numerical oscillations
so that Total Variation Diminishing (TVD) and other limiting schemes discussed in B.3.4 are
often used (Hosangadi et al. 1993). Such non-linear schemes are preferred to 1
st
-order accurate
schemes as they can reduce the artificial (numerical) diffusion of particle concentration,
momentum and energy in regions of sharp spatial density or particle velocity gradients.
Application of the Flux Corrected Transport method (FCT) of Eq. B.83 can be accomplished
by using a 2
nd
-order accurate scheme (e.g., Eq. B.73) for the high-order solution and a 1
st
-order
accurate scheme (e.g., Eq. B.38) for the low-order solution. This can also be extended to
328
finite-element formulations and unstructured grids. For example, Sivier et al. (1994) employed
a 2
nd
-order Taylor-Galerkin finite-element approach to first solve the homogeneous uncoupled
(LHS) parts of the gas and particle equations given by Eq. 7.7, after which the interphase
coupling terms (RHS terms) were added, and finally FCT was employed as a limiter. Such
limiting is usually performed on the gas conservative variables (
f
,
f
u, and
f
e
tot
) but on the
dispersed-phase primitive variables (, v, and e
p
) since this improves stability with respect to
on interphase coupling parameters (Collins et al. 1994; Chang et al. 1996). Furthermore, a
minimum particle concentration is generally specified throughout the domain so as to avoid
any negative values resulting from numerical errors. This value is generally set to be very
small (e.g., 10
-10
), but should be large enough to avoid the influence of round-off errors. A
sample simulation result with the Eulerian-Eulerian approach is shown in Fig. 7.1 for the
attenuation (by two-way coupling) of shock-wave speed through a dusty gas. Another example
simulation for particle concentration transport for shock-induced lofting is shown in Fig. 7.2.


7.1.3. Incompressible Particles in an Incompressible Fluid

If both phases are incompressible, this allows for simplification of the governing equations
provided that proper attention to their numerical solution is paid as the resulting system of
equations becomes more stiff. In the following, the continuous-phase equations are first
considered followed by the dispersed-phase equations. In both cases, the temperature field is
assumed to be constant or at least decoupled from the velocity field so that only the continuity
and momentum equations of each phase need to be considered.
The particle continuity equation in conservative form is given by:


( )
( ) ( ) ( )
j
p
Br conv j non conv Br p
j p
v
m
, v , , m
t x m

= + + = +

`
` f f 7.13
The RHS includes two rate functions f
conv
and f
non-conv
and, similar to Eq. B.66, temporal
integration of this PDE may be obtained by using the Adams-Bashforth scheme for the
convective term (Eq. B.65) and the Crank-Nicolson scheme for the diffusion and sink terms
(Eq. B.51):


( ) ( )
n 1 n
n n 1 n 1 n
conv conv non conv non conv
1 1
3
t 2 2
+
+


= + +

f f f f 7.14
An upwind spatial discretization is used for the convective terms (e.g., Eqs. 7.11 and 7.12)
while a central-differencing can be used for the diffusion and sink term, e.g., application of Eq.
B.28 for constant
Br
for two-dimensional flow yields:


( )
i+1,j i, j i 1, j i,j+1 i, j i, j 1 2 non conv
2 2
i,j
Br
2 2
=
( x) ( y)

+ +
= +

f

7.15
The application of this discretization for the n+1 implicit terms of Eq. 7.14 for 1-D conditions
results in a simple tri-diagonal equation that can be solved directly by the Thomas algorithm
(Eq. B.54). For 2-D and 3-D conditions, one may apply a 1-D Laplacian operator for each
coordinate direction with the ADI technique (Eq. B.56).
329
The particle momentum equation is often written in non-conservative form for
consistency with that of the continuous-phase. Such a form can be obtained by combining the
LHS of Eqs. 7.1a and 7.2a and then assuming constant particle density and neglecting
Brownian diffusion. From this, the momentum equation in vector notation becomes


( )
2 int
f
p p
1
t

+ = + + +




F v
v v g p u
7.16
Note that the pressure is defined as a single Eulerian field throughout the domain but not inside
a particle (e.g., the pressure within a bubble would be locally higher based on Eq. 2.62). The
interphase force in this case can be represented by Eq. 6.272 to include three-way coupling
effects, such that the RHS will be equivalent to the RHS of Eq. 6.273. However, the form in
Eq. 7.16 is often used since it is analogous to the incompressible continuous-phase momentum
form of Eq. B.59a and thus can be solved similarly with pressure-based techniques of B.3.3.
For example, application of the Marker-and-Cell approach which separates the convective and
non-convective force functions and neglects the pressure gradient effects yields the following
particle equation of motion


( ) ( )
pred n
conv j non conv f int
v ,
t

+ +

v v
F g = F F

7.17

The convection and non-convection force contribution can be computed with explicit and
implicit treatments. The velocity can then be corrected to account for pressure as in Eq. B.69:
( )
n 1 pred n 1
p
t / p
+ +
= v v 7.18
Note that these two steps can be combined in the case of one-way coupling since the pressure
field is known in advance. If the volume fraction is significant (>1%) and there are strong
volume fraction gradients in the flow, mass conservation errors can become significant for the
above form. To help avoid such problems, one may employ the particle momentum equation
in the conservative form of Eqs. 7.7. and 7.8. Such an approach ensures conservation of the
dispersed-phase mass and momentum, which can be important if there are substantial gradients
for either within the flowfield. In either case, these equations must be coupled to the
continuous-phase pressure and velocity fields.
In many applications, the continuous-phase flow (liquid or gas) travels at low Mach
numbers such that the flow may be considered incompressible. As with single-phase flows of
this type, a staggered grid is generally employed with the fluid velocities and pressure variables
(and now volume fractions) each defined on a unique stencil (Fig. B.6b). Similarly, the
continuous-phase transport equations are often given in non-conservative form assuming the
velocity field is divergence free (Eq. A.29). To assess the validity of this assumption in
multiphase flows, the divergence may be obtained by manipulation of Eqs. 7.1a and 7.1b for
finite volume fraction as:


p
p f
m
1
1 t 1


= =




w v
u
`
D
D

7.19
This is referred to as the velocity divergence effect (Ferrante & Elghobashi, 2004) and can
be important for high volume fractions and relative velocities.
330
For small relative velocities and small volume fractions, Eq. 7.19 indicates that the
continuous-phase velocity field is approximately divergence-free. In this case, the continuous-
phase momentum equation can be expressed in non-conservative form as


( )
p 2 int
f
f p p
m
1
t


+ + +




w
F u
u u g p u
`

7.20
This continuous-flow momentum equation is similar to that of Eq. B.59a except that interphase
sink terms are included on the RHS. The pressure field equation, which replaces the continuity
equation for the continuous-phase, requires no multiphase modification due to the assumption
of a single field pressure noted above. As such, the single-phase pressure techniques discussed
in B.3.3 can be used, e.g., the PISO approach of Eq. B.59b. Similarly, if a Marker-and-Cell
(MAC) technique is used, the pressure equation is given by Eq. B.68. The solution of the
pressure field then allows the correction to the predicted continuous-phase velocity by Eq. B.69
(and that of the particle velocity). As discussed in B.3.3, the SIMPLE and PISO methods are
effective for intermediate and low Reynolds number flows (e.g., Re
D
~1 or Re
D
1) while the
MAC method is quite popular for moderate to high Reynolds number flows (e.g., Re
D
1).
For significant relative velocities or volume fractions, Eq. 7.19 indicates that the
continuous-phase velocity field is not divergence-free. In this case, the above techniques
should be modified. One option is to separate out the terms which are not divergence-free and
place them on the RHS as sink terms (as was done for Eq. 7.16). Another option is to define
concentration-weighted pressure and velocity fields as


( )
( )
1
p 1 p


u u


7.21a

7.21b
Using these variables and assuming no mass transfer, the momentum equation becomes


( ) ( )
( )
( )
2 int
f
f p
1 p
1 p 1
t 1


+ = + +




F u
uu g u




7.22

This equation can be discretized similarly to Eq. B.61 by separating the pressure gradient and
identifying convection and surface-force contributions to the remainder of the RHS:


( ) ( ) ( )
pred n
conv non conv f int
f
1
, p, , p 1
t

+ +

u u
uu F g


= F F

7.23

The value of at any time-step and location can be obtained from the dispersed-phase
continuity equation (discussed above). The influence of the momentum function defined in Eq.
7.17 can be included with a combination of an Adams-Bashforth and Crank-Nicolson schemes
as in Eq. 7.14 to yield:

( ) ( ) ( )
pred n
1 1 n n 1 n 1 n
conv conv non conv non conv
2 2
3 1
t
+

= + + +

u u
g

F F F F 7.24
This pressure and the correction to the predicted velocity can then be obtained as:
331


( )
( )
2 pred
f
n 1 pred
n 1
f
p / t
t p
1
+
+
=

=



u
u u



7.25a

7.25b
This requires an iterative approach which can be accelerated with fast Fourier transforms and
allows determination of the fluid velocity once the volume fraction is known (Ferrante &
Elghobashi, 2005).

7.1.4. Bubbles in an Incompressible Liquid

While the above techniques may be used for bubbly flow, special treatments are required if
the volume fractions are no longer small (e.g. more than 1%), bubble density changes, and the
relative velocities are significant. In such cases, the velocities of both phases will be quite
sensitive to pressure gradients and will lead to high stiffness for the above Poisson pressure
equations. To avoid this, the flow transport equations can instead be discretized in
conservative form and complemented with an equation of state. The latter can be based on
quasi-equilibrium conditions if the pressure oscillations are slow compared to the bubble kHz
frequencies associated with volumetric oscillations (Eq. 2.66). Assuming that the bubble is also
thermal equilibrium with the surrounding liquid, the bubble density can be given by Eq. 2.62:

3
p
p
T p 3m
4 2

=



g g g
R
7.26
In this expression, p represents the unhindered liquid pressure. For a large bubble, the surface
tension effect can be neglected yielding:

p
p / T
g g
R
7.27
This simple explicit relationship is reasonable to within 1% accuracy for bubble diameters
which are 300 m or more in water.
The PDEs of Eq. 7.1 can be used for the bubble and liquid mass transport equations by
assuming an incompressible liquid and neglecting Brownian motion:


( ) ( )
p p j
p
j p
v
m
t x

+ =

`

( )
j
j p
j j f p
u
u m
t x x

+ =

`

7.28a


7.28b
The second equation will be used later to provide a continuity equation that supports a
pressure-field solution. The first equation can be discretized in time-using a 1st-order accurate
backward difference for the unsteady terms. There are several options for the non-linear
convective terms. If one applies Picard linearization whereby
( ) ( )
n 1/ 2 n
n 1
p j p j
v v
+
+

(Prosperetti et al. 2007), the particle continuity equation becomes:
332


( ) ( ) ( ) ( )
n 1 n n n 1
n+1 n
p p p j p p
j
t v t m /
x
+ +


= +

`
7.29
Solution of this equation and Eq. 7.26 once the updated pressure and velocity (p
n+1
and v
n+1
),
are obtained from the momentum equations.
The bubble and liquid mass momentum equations can use similar assumptions for the
mass transport to yield the following momentum equations from Eq. 7.2:


i j int,i
i j p
K F
p
0
x x

= + +



7.30a



[ ] f i j i j int,i f i
f i
j i j p
(1 ) u u K F (1 ) u
p
(1 ) g
t x x x


= +




+

7.30b

The interphase force on the RHS of both equations can be expressed using Eq. 6.272 but
simplified using bubbly flow assumptions assuming spherical or nearly spherical shapes. For
the drag force, the particle response time can neglect bubble mass and be corrected for finite-
volume effects as
p,
based on Eq. 6.267, which is reasonable for spherical and moderately
distorted bubbles (e.g., We<5). For the lift force, the bubble angular momentum can be
neglected so that a spin-equilibrium is assumed (e.g., Eq. 6.151 and 6.166 for spheres). For the
added mass, the potential flow theory of Eq. 6.184 can be used but with particle time derivative
approximated by that along the fluid path. Based on these assumptions, the interphase force
can be expressed as:


( )( ) ( )( )
( ) ( )
( ) ( )
int,i D,i L,i ,i f p i p, L,i ,i
3.5 2
2
p f
f
p,
f f
L,i f p L, L, ,eq p
i
i
i i i i
,i f p j j
j j
3
4
F f F 1 F F w / 1 F F
c d
c d 1
18 f 18
w
F C C
d
u u v v
F c u v
t x t x



= + + = + +
+




=



= +




- -
w + w

7.31a

7.31b

7.31c


7.31d
This equation neglects the history force term which is typical for Eulerian treatment and
consistent with an assumption of weak relative accelerations or high bubble Reynolds numbers
(Eq. 6.199). If such assumptions are not appropriate, the Lagrangian method of 7.2 is better
suited to deal with the history force integral along the bubble trajectory.
The interphase force can be substituted into the equations of motion as follows:


i j L,i
i i i i f i
f j j
j j i j p, p
K F
v u v u w 1 p
c v u
t t x x 1 x x



+ = + +





-

7.32
This form is convenient in that the LHS contains temporal derivatives of the bubble and
continuous-phase velocity and a similar form can be applied to the continuous-phase
momentum equation.
333
For discretization, these transport equations can be evaluated at t
n+1
with second-order
central-difference for the time-derivatives, use of particle response time at t
n
(as in Eq. 3.8a),
and similar use of use of the lift-force at t
n
. For numerical solution with an incompressible
continuous-phase liquid, the pressure gradient can be handled via one of the predictor-corrector
methods of B.3.3.
The predictor typically employs the pressure gradient at t
n
to allow an explicit step and uses
partial elimination to achieve separable equations. In particular, the equations for the phase
velocities can be decoupled at the predictor level by replacing the relative velocity with the
difference between the particle and fluid velocities. If one assumes
p
t < , the predictor
form of the velocities can be expressed as:


( )
( )
n
pred n n n
i f ,i p,i
f i
n
n
n
p,
p, pred n n n n
i p,i f ,i p,i n n
p, p, f i
1 c t
p
u
c x
t c c t
c
t p
v
c t c t c x


+
= +




+ +
= + +

+ +




F F F
F F F F

7.33a


7.33b
These equations use the following functions:


( )
( ) ( )
n
n
n
n n n
n n
ij p, L,i n n n n n i i
p,i i i j j n
j j f j f p p,
n n n
n n
ij L,i n n n n n n n n n i i
f ,i i i i j j i
j j f j f p
c
1
K c 1 F
u v 1
v u t u v
x x c x c c t
K F
v u
u u v t v 1 u g
x x x



+ + +

+




+ + + + +

-
F
F
F F F F


7.34a


7.34b


7.34c
The resulting explicit relationships directly account for the coupling between the phases and
help to eliminate instability problems (Wang et al. 1994). The convective spatial derivatives
can be obtained with upwind differencing while the other spatial derivatives can be obtained
with central-differencing. To reduced numerical diffusion, Tomiyama & Shimada (2001) and
others suggest a limiter, such as TVD, on the volume fraction equation.
For the correction step using a MAC method (B.3.3), one may define an pressure
increment as p and use this to update the above predictor-step velocities as:


n 1 n
p p p
+

( )
( )
n
p,
n 1 pred
i i n
f p, i
n 1 pred
i i
f i
t c c t
t p
v v
c c t x
1 c t
p
u u
c x

+

+


+ +

=
+


+

=


7.35a

7.35b


7.35c

A pressure-based equation for p can be obtained by weighting Eqs. 7.35b and 7.35c by and
1- and take the divergence of the combination:
334


( ) ( )
( )( )
( )
n 1 pred
j j j j j j
j j
n n n
p,
n
n
f j p, i
v u u v u u
x x
t c c t
t p
1 c 1
c x c t x
+


+ +
=


+ +

+

+




7.36
To satisfy continuity for both phases, one may combine Eqs. 7.28a and 7.28b for a net
continuity equation to be satisfied at the new time-level (yielding a similar LHS):


( )
n 1
n pred
n n
j j j
p p j p j
n n
j p p p p j j
v u u
m v v
p p
x t x p t p x
+
+




`

7.37
The RHS employs the linearized equation of state given by Eq. 7.27 to replace the density
derivatives with pressure derivatives. Combining Eq. 7.36 and 7.37 to eliminate the n+1
velocity values and placing the pressure gradient terms on the LHS yields:


( )( )
( )
( )
n n n
p,
n
n
f j p, i
pred
n pred
n n
j j j
j p
n n
j j p p
t c c t
t p
1 c 1
c x c t x
v u u
v m
p p

x p t p x




+ +

+

+



+


= + +

`

7.38
This is similar to the equation used by Tomiyama & Shimada (2001) which neglected mass
transfer and particle compressibility effects (last three terms on the RHS). It can be solved for
p using implicit iterative techniques similar to those used for Eq. B.71b, e.g. coordinate
direction sweeps of SOR (Eq. B.58).
A similar approach is discussed by Wang et al. (1994) but with a SIMPLE-scheme and
other implicit methods are discussed by Prosperetti et al. (2007). In general, the overall
numerical methodology can be summarized as follows:
a) the existing pressure and volume fraction (p
n
&
n
) are used to compute the predicted
velocities with Eqs. 7.33a and 7.33b
b) the predicted velocities are used for Eq. 7.38, which is then iteratively solved for the
pressure increment (p) to find the new pressure based on Eq. 7.35a;
c) the velocity fields can be corrected with Eqs. 7.35a and 7.35b
d) the updated velocity fields and bubble density (from Eq. 7.27) can be used to obtain the
updated volume fraction with Eq. 7.29
For steady-flows, one may simply iterate on steps a)-d) until convergence. For complex flows
where higher-order temporal accuracy is needed, the volume fractions in steps b)-d) can be
improved by use of an explicit Adams-Bashforth model to predict
n+1/2
or the volume-fraction
equation can be converted into a two-step predictor-corrector equation by introducing the
predictor step after step a).
For flows at high-speeds (about 20 m/s or more), compressibility of the overall mixture can
become important. In this case, the above predictor-corrector iterative schemes should be
replaced by compressible flow explicit schemes more consistent with B.3.4. To use the Roe
335
upwind differencing scheme for the convective terms (a common approach), the Eigenvalues
and Eigenvectors of the multiphase system must be obtained. For a compressible bubbly flow
system, these are given by Venkateswaran et al. (2002) and Owis & Nayfeh (2003) and are
based on an equation of state of the mixture (5.1.3). These studies also included a pre-
conditioning technique to avoid numerical stiffness for regions of the flow where the velocity
is much less than the acoustic speed (Eq. 5.42). This pre-conditioning effectively adjusts the
local speed of sound so that the local Mach number does not drop below a small but finite
value (about 0.02 or so), i.e. a weak artificial compressibility is introduced to avoid excessively
small time-steps based on Eq. B.47b.


7.1.5. Methods for Polydispersion

There are three primary techniques to model multiphase flows with a variation in particle
sizes: effective diameter approach, bin approach, and distribution function approach. The
effective diameter approach simply uses an average diameter that reasonably represents the net
response of the particles within a mixed fluid volume. For flows where gravitational and drag
forces are the primary drivers for the relative velocity, the effective diameter is d
32
(Sauter
mean diameter) for Re
p
>500, d
31
(volume-width diameter) for Re
p
<1, and d
3Y
for intermediate
Reynolds numbers (Eqs. 2.6-2.8). An effective diameter is generally reasonable for moderate
levels of broadness, e.g., when 80% of the particle mass fraction is contained between a
maximum diameter and a minimum diameter whose ratio is less than 3:1 (Loth et al. 2004).
However, flows with fragmented particles, sprays or bubbles will tend to have a broader
distribution so that a single effective diameter is no longer reasonable.
The second approach is to use a multi-bin distribution (Eq. 4.62) which can also handle
multiple phases (e.g., bubbles and particles in a liquid). The generalized set of PDEs for
particle mass transport can be written as:


( ) ( )
( )
k p,k k p,k kj
k p,k
Br,k p,k k
j p,k
v
m
t x


+ = +


`
7.39
In this equation, k represents an individual phase or bin, K represents the total of all dispersed
phase bins,
k
is the volume fraction of phase k per overall mixed-fluid volume, v
k
is the
phase/bin particle velocity, and
p,k
m` is the rate of mass transferred to phase k per particle.
This rate can include both interphase transfer from the fluid and intra-phase transfer from other
particle bins. The latter may be obtained based on models for collision frequencies (3.8.2)
coupled with models for collision outcomes (2.4.5). Similar to the net volume fraction of Eq.
4.61, the mass transfer to the continuous phase needed for Eq. 7.1b is given by:


K
k p,k p
k 1
p,k p
m m
=

=

` `

7.40
Since this accounts for mass transfer between phases, it is independent of break-up and
coalescence processes which only affect the intra-phase fluxes.
The particle momentum equations (Eq. 7.2a) may be written in conservative form for
individual bins as:
336


( ) ( )
( )
k p,k k,i k p,k k,i k, j
i j
k p,k i
j i j
k
int,i,k p,k i,k
p,k
v v v
K
p
g
t x x x
F m v

+ = +



+ +

`


7.41

The last term in parentheses on the RHS is the intra-phase momentum transfer which
represents the interaction of a single bin/phase. Therefore, the continuous-phase momentum
equation (Eq. 7.2b) will be based on the sum of the bin contributions:


( )
( )
K
int,i p i
k
int,i,k p,k i,k
k 1
p p,k
F m v
F m v
=
+

= +



`
` 7.42
The multi-bin approach for particle energy transport based on Eq. 7.3a can be expressed as


( ) ( )
( ) ( )
k p,k p,k k p,k p,k j,k p,k p,k
k phase,k p,k eff
j p,k p,k
e e v
m
+e T
t x

+ = + +




`
` Q
h k
7.43
The first two terms on the RHS are the heat and mass transfer terms for each bin which can be
summed over the K bins to obtain the net heat and mass transfer terms. These are used along
with those involving F
int
v and v as the interphase coupling terms to the continuous-phase
energy transport (RHS terms of Eq. 7.3b).
To use these approaches, the bin diameters (d
k
) must be computed a priori and are used
throughout the domain as effective diameters for particles within that bin. Figure 7.3 shows
experimental results of a plume with bubbles (diameter range of roughly 1.5-6 mm) as well as
predictions with the effective diameter and multi-bin approach. It can be seen that the bubble
distributions with three Eulerian bins is more consistent with the experiments because it takes
into account the polydispersion of bubble sizes. Note that the diameter ranges for each bin
must be selected a priori based on expected overall ranges of the particle sizes.
A more complex but generally more robust Eulerian approach for a polydisperse system is
to consider a transport equation for the size distribution function (Williams, 1965). This
distribution function must then be computed at all nodes, but can represent a continuum of bin
distributions. A computationally-efficient variant of this approach is the Direct Quadrature
Method of Moments (DQMOM) whereby the size distribution function is characterized by a
finite number of moments (Fan et al. 2004). For a single moment, this requires an additional
transport equation for each bin which is the moment of the particle diameter and the continuity
equation (e.g., Eq. 7.39). For example, if particle density is constant and Brownian motion is
neglected, the single mass transport for each bin is replaced by two equations:


( )
( ) ( )
k kj
k
k
j
k k kj
k k
k k k k
j
n v
n
n
t x
d n v
d n
d n d n
t x

+ =

+ = +

`
`
`

7.44a


7.44b
The bin diameters are thus unknowns to be solved locally along with the number densities
(rather than being fixed globally and a priori as in the conventional bin method). As such, the
337
RHS terms include changes in bin k due to exchanges between other bins (e.g., break-up,
agglomeration) and the continuous-phase (e.g., evaporation, condensation). A key advantage
of this DQMOM technique is that the bin resolution is effectively optimized for each location,
e.g., regions of the flow with only very small particles will be resolved locally with small bin
diameters and other regions with broader size distributions will be resolved with a broader
range of bin diameters.


7.1.6. Particle-Wall Interactions

For the case of particle-wall interaction, there are three main outcomes as shown in Fig.
7.4: absorption, accommodation and reflection. The absorption condition is easiest to model
with an Eulerian approach. It effectively requires no boundary conditions so that the particle
velocity is not affected by the presence of the wall and the mass flux to a discrete surface area
may be simply computed from
p surf
v A

. The accommodation outcome indicates that the


particle loses all normal momentum upon wall impingement but remains within the
computational domain. As described in 6.10, this outcome corresponds to
out
=0 which may
occur for
in
0 (grazing conditions) or
|
< e e (weak restitution, e.g., St 10

< ). For an
Eulerian approach, an accommodation outcome may be approximated by applying the
boundary condition v 0

= (slip conditions), though this neglects any changes in tangential


momentum, i.e. assumes 1 =
|
e . Furthermore, imposition of such a boundary condition on the
Eulerian particle field can be felt in upstream nodes (as an artificial repelling force) prior to the
particle-wall impact unless the particle momentum equation is numerically treated with pure
upwind differencing. Even if pure up-winding is used to alleviate this non-physical streamwise
diffusion for upstream nodes, artificial transverse diffusion away from the wall is difficult to
avoid (Durst et al. 1984).
The condition of reflection cannot be generally described with a conventional Eulerian
approach. For example, consider a steady flow of particles towards a wall as shown in Fig. 7.5.
In the physical case (Fig. 7.5a), the particle velocities near the wall are actually multi-valued
within a finite volume as they consists of both incoming and reflecting trajectories.
Application of the accommodation boundary condition used for Fig. 7.4b can be considered a
crude approximation as it gives an average velocity field between that of the reflected and un-
reflected conditions (Fig. 7.5b). However, the net particle kinetic momentum is unphysically
reduced and the particle-wall collision rate can not be obtained.
However, the reflection outcome can be handled accurately with an Eulerian treatment if
one employs the multi-bin approach. In particular, incoming particles are treated as one
field/bin and once-reflected particles as a second field/bin (Slater & Young, 1988). This
approach is shown in Fig. 7.5c and can also be used for conditions where particle trajectories
may cross. However, it requires solution of an additional set of continuity and momentum
equations throughout the domain each time an individual trajectory undergoes a reflection or a
trajectory-crossing. If there are multiple reflection for a single particle, this would be typically
impractical to simulate with an Eulerian approach. Such a problem can arise for particles or
bubbles bouncing along a wall in turbulent flow (Kaftori et al. 1995, Felton & Loth, 2001) or
338
for particles in a turbulent flow where trajectory-crossing can be common within an Eulerian
mixed-fluid grid volume once the particle Stokes numbers are of order unity or greater.
Another Eulerian approach to handle multi-valued velocity fields has been put forward by
Desjardins et al. (2007) based on the DQMOM approach discussed in the previous sub-section.
In particular, an additional transport equation for the velocity-based moments of the particle
momentum. This increases the number of PDEs but allows a distribution function of the
particle velocity field at a given point in space. Predictions of particle motion in frozen
turbulence are shown in Fig. 7.6 with various approaches. The two-node Quadrature approach
yields a much improved representation of the concentration field, i.e. one that approaches that
of a full Lagrangian simulation. The Eulerian DQMOM technique may be extended to allow
for collision effects, but generally Lagrangian approaches (to be discussed in 7.2) are
preferred when particle-particle and wall-reflection effects are important.

7.1.7. Time-Averaged Turbulent Multiphase PDEs

Turbulence is a condition for which flow unsteadiness must be taken into account.
However, if a fully-resolved DNS description of the continuous-phase turbulence is not
practical, then RANS approaches (A.5.5-A.5.6) or LES approaches (A.5.7) can be used. In
either case, temporal or spatial averaging of the multi-phase equations is needed.
The RANS point-force equations have been rigorously derived by Elghobashi and Abou-
Arab (1983). The time-averaged equations specialized for incompressible particles in a
compressible gas are discussed by Faeth (1987) and Shyy et al. (1997), while those for a
compressible bubble are discussed by Ishii (1975) and Besnard et al. (1991).The Eulerian-
Eulerian transport equations are discussed below based on a two-fluid system while Tomiyama
& Shimada (2001) discuss a multi-bin dispersed-phase approach. In the following, the multi-
phase RANS equations will be overviewed using the pseudo-unsteady formulation of A.5.5.
This is a common formulation and allows extension to LES and hybrid RANS/LES techniques.
Specific discussion of turbulent particle diffusion and two-way coupling in the following two
sections.


Phase-based Favre-averaging and Compressibility

Time-averaging of the Eulerian particle PDEs with variable density was discussed in 5.2.2
using the conventional Favre-average definitions (e.g., Eq. 5.60a). This formulation allowed
significant simplification of the resulting PDEs by eliminating several fluctuation correlations
and also allowed a conservative mass transport equation for the mean variables in the absence
of interphase mass transfer (Wang et al. 1994). A similar approach can be extended to the
point-force equations of this chapter by using a phase-based Favre-average. In this case, the
respective spatial density (mass per mixed-fluid volume) of the dispersed-phase and
continuous-phase are used to create a particle-based or a fluid-based Favre-average:
339


p p
p
p
f f
f
f
q
q
(1 ) q
q
(1 )


7.45a


7.45b
As such, Eq. 7.45a is applicable for particle variables such as
int,i p p p
v, w, F m , e , and
`
` Q , while
Eq. 7.45b is applicable for continuous-fluid variables such as
f ij f
u, p, T , K , and e . The sub-
scripts of Eqs. 7.45a and 7.45b are often assumed to be understood and thus not shown, e.g.


p
p
f
f
v
v
(1 ) u
u
(1 )


7.46a

7.46b
These Favre-averaged velocities are numerically convenient since they allow direct
specification of the mass flux at inflow planes or determination of the mass flux at outflow or
wall absorption planes.
Similar to Eq. A.110, an arbitrary property may be decomposed in terms of mean and
fluctuating components using density-weighted average notation:


p p p
f f f
q q q
q q q
= +
= +


7.47a
7.47b
If the particle and continuous-phase densities are constant, the above averaging reduces to a
volume-fraction weighted averaging:


p p
p p
f f f f
f f
q q
q q
(1 )q q q q
q q
1 1 1 1

= +


= =


7.48a

7.48b
For both equations, the RHS expressions show the difference between time-weighting and
volume-fraction weighting. Generally, the fluctuations are not important to the equation of
state of a gas for subsonic flow, so that they can be simply based on either time-averaged or
densityweighted quantities as is convenient. However, averaging the transport equations must
be considered carefully and is discussed in the following.



Mean Mass Transport Equations

Using the above assumptions of uniform densities of both phases coupled with volume-
averaging applied to the continuity equations given in Eq. 7.1 yields:


( )

j
p
j p
v
m
t x m

+ =

`
7.49a
340

j
p
j p
(1 )u
m
(1 )
t x m



+ =

`
7.49b
As with the continuous-phase time-averaged turbulence equations (A.5.5), the time-derivative
terms in the RANS limit are non-physical but are retained to allow convergence using time-
marching schemes. It is also instructive to retain these for the LES limit to incorporate the
physical unsteadiness of the resolved velocity field.
Note that Brownian diffusion effects are not included in the weighted particle continuity
equations and will instead appear as a momentum force, which will be discussed in the next
sub-section. As a result, volume-fraction weighting eliminates all correlation terms in the
continuity equations. Thus, only mean variables are included in these equations. Furthermore,
if mass transfer can be neglected, the mass conservation equations can be written without
source terms.


Mean Momentum Transport Equations

When volume-fraction weighting is applied (constant densities for each phase), the time-
average of the particle and continuous-phase momentum transport PDEs (Eq. 7.2) becomes:


( ) ( )

( )

( ) i j j
i j
i j i
i int,i p i
j p j i p j
v v
v v
K v
1 p
g F m v
t x x x m x



+ = + + +



` 7.50a


[ ]

( )
int,i
p i i j
i j i
f i
j f i j f p
F m v (1 )u u
K (1 )u
1 p
g
t x x x

+



= +




+
`

7.50b
Note that the volume-fraction weighting avoids triple correlations on RHS as compared to a
Reynolds-averaging. It can be numerically helpful to write the stress term of Eq. 7.50a in
terms of the continuous-phase velocity by applying Eqs. 6.190 and 6.191. Unfortunately, the
fluid stress term of Eq. 7.50a is weighted with so that this will produce additional terms if
one employs fluid properties weighted with 1- as specified by Eq. 7.45b. This can be shown
by an example expansion of a similar but simpler term:


( ) K K K K K K / 1 = + = +

7.51
However, if one assumes that the viscous stress fluctuations are negligible compared to the
turbulent stresses,

p i j j
K v v < , and assumes that the pressure correlations are negligibly
correlated with the volume fraction, i.e. p p < , the fluid stress term of Eq. 7.50a can be
approximated as:

i j
i i
f j i
j i j
K
u u p
u g
x x t x


+




7.52
Such assumptions are certainly reasonable for flows with heavy particles but are also
commonly used for bubbly flows. In particular, the first assumption is reasonable except very
341
close to the wall (see discussion following Eq. A.118) while the second is reasonable for low-
speeds with weak mass transfer (Eq. 5.70). Note that the primary effect of mass transfer on
momentum is still retained in Eq. 7.50a, despite this approximation.
Substituting Eq. 7.52 into Eq. 7.50a yields a particle momentum equation in terms of
particle and fluid-phase velocities. Furthermore, the mass transfer term can be eliminated by
subtracting the time-average of the product of the continuity equation multiplied by the particle
velocity, yielding:




int,i
i i f i i
j i j i
j p j p
F
v v u u
v g u g
t x t x m


+ = + + +






7.53
A similar combination of the continuity and momentum equation for the continuous-phase
momentum (Eq. 7.50b) yields:


( )
( ) ( )
int,i
p i
i j p i
i i
j i
j f i j f p p
F m v
K m u
u u 1 p
u g
t x x x 1 1 m
+


= + +




+
`
`

7.54
Both transport equations thus include the Favre-averaged interphase force, which can next be
written in terms of the transport properties.
The interphase-force can then be broken into the various point-force components based on
Eq. 6.1 and extended to include Brownian diffusion as a force:



int,i D,i L,i ,i T,i Br,i
F F F F F F

= + + + + 7.55
Note that the history force is neglected since it requires long-time integral information along
the particle path which is not available in an Eulerian formulation. This assumption can be
reasonable for
*
p
1 (see 6.6) or when the velocity accelerations are due primarily to turbulent
fluctuations (so that their mean effects are negligible). Otherwise, the history force is best
included by a Lagrangian formulation for the particle-phase (7.2). The remaining forces are
individually discussed in the following.
The Favre-averaged drag force can be written in terms of mean and fluctuating velocities:



( ) D,i f i f i i
F 3 d fw 3 d fw f w = = +


7.56
The first term on the RHS can be expressed in terms of a Favre-averaged particle response time
(Eq. 3.1) defined as:


( )
2
p f
p
f
c d
18 f

7.57
The second term on the RHS of Eq. 7.56 is negligible for the linear-drag regime (e.g., Re
p
0
or clean bubbles with Re
p
1) but will be positive for a non-linear drag. For Re
p
<1000, this can
be modeled with Eq. 3.110 by equating trajectory-averaging to volume-weighted averaging:

0.687
p @p
k k
i i 0.687
j j @p p
St
u u 0.103Re
f w fw
w w 1 St
1 0.15Re





+

+



7.58
342
For particles in the Newton regime (Re
p
>1000), a volume-weighted average of Eq. 3.104 can
instead be used.
Since lift is linear with respect to relative velocity for both Re
p
1 (Eq. 6.131) and inviscid
conditions (Eq. 6.6), the unsteadiness along the Lagrangian trajectory is generally neglected
when considering the Lagrangian mean. Thus, the average lift force and the equilibrium lift
coefficient vector can be modeled using Eq. 6.129b as


2 2 2
L f L, L,
8
d w C w C




= +




w w
F
w w


7.59
If one neglects second-order correlations such as

w 0 = (reasonable since lift is already


typically small compared to drag) and assumes that spin-equilibrium (reasonable when
averaged over many particles), Eqs. 6.168 and 7.59 can be combined to yield:

( )
2 2 i
L,i f L,eq
8
F w d C

w
w


7.60
The virtual mass force of Eq. 6.5b can be similarly averaged as:


i i i i i i
,i f p f p j j
j j
u v u u v v
F c c u v
t t t x t x



= = +






D d
D d

7.61
The thermophoretic force can be modeled using the Eulerian-averaged temperature gradients
since again this force is typically secondary to drag. Based on Eqs. 6.242-6.244 this yields:

0.7
p
T,i 1.7
p i
d T Kn
F
2T 1.15 Kn x





+

g g g
g

7.62
The Brownian diffusion can be included as a force based on the diffusion velocity via Eq.
3.140 (Slater et al. 2003) as:


f
Br,i
i
T
F
x




7.63
This choice is often computationally convenient compared to inclusion in the mass continuity
equation as in Eq. 7.13, but either approach is appropriate.
The particle correlations appearing in Eqs. 7.53 and 7.61 can be approximated by taking the
time-average of the product of the Eq. 7.49a with
i
v :




( ) i j
i i i i
j j j j
j j j j j
v v
v v v v 1
v = v v v
x x x x x


+ +


7.64
This equation assumes

i p
v m 0 ` , which is satisfied for creeping flow since the mass transfer is
approximately independent of relative velocity for Re
p
1 (Eq. 1.87 and 1.94a). However, this
assumption is also reasonable for higher Reynolds numbers (Sirignano, 1999). Similarly, the
continuous-phase velocity correlations of Eqs. 7.54 and 7.61 can be approximated as
343




i j
i i i i
j j j j
j j j j j
(1 )u u
u u u u
u = u u u
x x x x (1 ) x



+ +


7.65
If thermophoretic and Brownian forces are small compared to turbulent diffusion (reasonable
for particles in a continuum) and the continuous-phase fluctuations are divergence-free (cf. Eq.
7.23), the interphase force becomes



( )

( )
2 2
int,i f i i f L,eq,i
8
i j i j
i i i i
f p j j
j j j j
F 3 d fw f w w d C
v v (1 )u u
u u v v
+ c u v
t x (1 ) x t x x

= + +





+ +





7.66
The terms in the square brackets can be combined with the LHS of Eq. 7.53 and the
acceleration parameter (Eq. 1.81a) to yield a particle momentum equation as follows:


( )


( )

( )
i i i i i
j i j L,eq,i
j j p
i j i j
i
j j p
v v u u w R 3w
v 1 R g R u C
t x t x 1 c 4d
v v (1 )u u
f w
R
1 x x f



+ = + + + +


+






+ +








7.67


The lift and fluid stress terms can be typically neglected for particles in a gas (since R0) while
the thermophoretic and Brownian forces can be typically neglected for bubbly flow (since
Kn
p
0). The three terms on the RHS in curled brackets are due to turbulent fluctuations. The
first is the non-linear drag effect which can be approximated with Eq. 7.58, while the other two
include effects associated with turbophoresis and concentration gradient as discussed in 3.5.5.
To close these terms, models will be given in 7.1.8 for

i j
u u and in 7.1.7 for

i j
v v .
The continuous-phase momentum of Eq. 7.54 can be similarly manipulated to yield:


( )
( )
i j f p i p p i
i i
j i
j f i j f p p
i i i j
L,eq,i *
p
j p
K m u m v
c u u 1 p
1 u g
(1 ) t x x x (1 )
w f w / f (1 )u u
3wC
c
1 c
(1 ) (1 ) x (1 ) 4d



+ = + +






+



+ + +




+

` `

( ) i j
i i
j
j j
v v
c v v 1
+ v
(1 ) t x x





+ +




7.68


Collecting the mean continuous-phase velocities on the LHS yields:
344

*
i j p
i i i
j f i
j f i j p
*
p i p p i
L,eq,i
i i
j
p j
K ( c )
u u w 1 1 p
u g
t x 1 c x x 1 c
m u m v
3wC
c v v
+ v
(1 c )m 1 c t x 4dc


+
= + +

+ +







+


+ +


+

` `


( )
*
i j i j
p
i
j j p
v v (1 )u u
( c )
f w c

1 c (1 ) x 1 c x f



+

+ +

+ +





7.69


For low-speed flow, the resulting forms of Eqs. 7.67 and 7.69 are often numerically solved by
first ignoring the pressure gradient effects and using explicit schemes for the mean velocity
terms of both phases, followed by while an iterative solution for the pressure field and a
subsequent velocity correction (7.1.4). For the case of particles in a gas (
*
p
1) with finite
mass loading (
*
p
) but negligible void fraction (1), Eq. 7.69 may be neglected can be
simplified as:



i j
i j p i * i i i i
j i p
j f i j p p j p
(1 )u u
K m v
u u w f w 1 p
u g
t x x x m (1 ) x f


= + + +





+

`



7.70

The last three terms of Eq. 7.69 and the last two terms of Eq. 7.70 include the effects of the
fluctuation quantities, which will be modeled in the next section.


Mean Energy Transport Equations

The particle internal energy transport equation of Eq. 7.3a can be time-averaged to give:


`
( )

( )

( )
p j p
p p p p phase eff f
j p p
e v e
m e m h T
t x m



+ = + + +


`
` ` Q k 7.71
If one includes the mass and heat transfer relationship of Eqs. 7.4 and 7.5, neglects correlations
between volume fraction and temperature fluctuations (Eq. 5.70), and uses the continuity
equation to allow a non-conservative formulation, this equation may be rewritten as:


( )

( )

( )

( )

f p f f f ,p p phase
p p
j 2
j p
p f f p j
p j
6 Nu T T ShB e h
e e
v
t x d
(1 ) T e v

x

+ +



+ =

+

+

k
k k

7.72
In this equation, the gradients of the volumefraction correlations have been neglected since
these effects will be small compared to the turbulent diffusion of particle energy by the terms
(similar to the assumption used for Eq. 7.50). Further simplifications are possible if the heat
and mass transfer coefficients are assumed to be have weak fluctuation correlations (which is
345
especially appropriate for Re
p
0 since this limit yields constant Sherwood and Nusselt
numbers by Eq. 1.94).
The continuous-phase energy equation (Eq. 7.3b) can be similarly time-averaged as:

( ) ( ) ( )

( )

i j ij i eff i j ij
i
tot tot
j j j
j f j f j f j
p p phase p i i int,i i j tot
f p j
(1 )(K p )u T K p
v
e e
u g u
t x (1 ) x (1 ) x x
m h +e + v v F v
(1 )u e

(1 ) (1 ) x





+ = + + +


+ +




`
`
k
Q


7.73


This result employs the previous assumptions of this section and reduces to the single-phase
time-averaged energy equation in the limit of 0 .
The mean transport equations for multiphase mass, momentum and energy (e.g., Eqs. 7.49,
7.67, 7.69, 7.72 and 7.73) yields several correlation terms which are essential to predict particle
distribution and properties with RANS approaches. If one is able to employ LES or Hybrid
simulations, than the most of the turbulent diffusion can be resolved directly and modeling is
only needed for the second-order sub-grid diffusion. Various turbulent diffusion models are
discussed in the next section.

7.1.8. Eulerian Models for Turbulent Particle Diffusion

Closure of the mean mass, momentum and energy PDEs of the previous section requires
modeling the fluctuation correlations in terms of mean transport variables. As with single-
phase flow turbulence modeling (discussed in A.5.5-A.5.6), there are two primary approaches
to model the multiphase fluctuation correlations: algebraic correlations and PDE correlations.
Generally, algebraic turbulence modeling is used to relate particle-phase diffusivity and bias to
the continuous-phase turbulence properties (such as
t
, k, , etc.) which are modeled with one-
or two- equation turbulence models. Sometimes PDE transport equations are used for both
phases, e.g., four-equation model for k, , k
p
and
p
(Kashiwa & Gore, 1987; Viollet &
Simonin, 1994). However, such approaches are not commonplace due to their complexity and
increased computational requirements associated with the additional equations, which
introduce many new terms to model. The simpler algebraic particle diffusivity correlations
will be discussed below, first for RANS and then for PDF approaches. This is followed by a
discussion on LES/Hybrid models.


RANS Correlations for Particle Diffusion and Bias

A wide variety of turbulence models have been used to correlate particle Reynolds stress
with the continuous-phase Reynolds stress with varying degrees of success (Soo, 1990). If one
assumes that the particle fluctuations are in local equilibrium with the continuous-phase
turbulence, then an algebraic relationship between the quantities can be employed. Note that
this does not require that the continuous-phase turbulence be in equilibrium with its mean
quantities, i.e. a one-equation or two-equation transport model can still be used for
t
. The
primary particle velocity correlation appearing in the momentum equations is given by the last
terms in Eqs. 7.67 and 7.69 which can be expanded using the chain-rule as:
346

( )

i j
i j i j
j j j
v v
v v v v
x x x

= +

7.74
The two terms on the RHS include turbulent transport of particle momentum by mean
concentration gradients and by turbulence gradients, which were discussed in 3.5.5.
To obtain the particle velocity fluctuations for heavy particles (R=0), the terminal velocity
in Eq. 3.84 can be replaced by the Favre-averaged mean particle velocity to yield:


1
2
p
i i
i i i i
i @p,i
w u u
v v u u 1 1+
1 St





= +

+


L
L
NSI 7.75
For very-buoyant particles (R=3), an anisotropic version Eq. 3.89 is recommended:



i i i i
v v u u NSI 7.76
The anisotropic length-scale can be approximated by using the component of the Favre-
averaged particle velocity as:


( )
i i
1 w / w

+
7.77
For fluctuations in the transverse and streamwise direction along the particle path, this equation
respectively yields
i
= and
i
2

=
|
, consistent with most HIST flows (Eq. A.89) and
even some NAsT flows (A.5.5). Similar equations have been used for both bubbly flows and
heavy particle flows (Mudde & Simonin, 1999; Slater et al. 2003). If k and are given, only

and

are needed to complete the above formulation. However, these values take on
different forms for free-shear flows compared to boundary-layer flows.
For Eulerian-based diffusion models free-shear flows (jets, wakes, shear layers, separated
regions, etc.), the time and length scales can be related to the fluid Favre-average of kinetic
energy and dissipation:

( )
3/ 2
c k /
c u 2/ 3 c c k /



=


L
L
for free-shear flows
7.78a

7.78b
These equations include integral-scale coefficients based on Eqs. A.90 and A.91 with typical
values used for Eulerian-based diffusion models given by:


c 0.21
c 1.3

=
=

7.79a

7.79b
These values are also approximately reasonable near the outer region for pipe and channel
flows since this region is approximately HIST. However, there is significant variation in the
values used by different researchers. For example, Chen & Pereira (1995) and Mudde &
Simonin (1999) suggest c

=0.75, Chahed et al. (2003) suggests c

=1.85, and Lightstone &


Hodgson (2004) suggest c

=2. These differences may be partially attributed to variations in


turbulent structures for different types of free shear flows as noted by Wang & Stock (1993)
347
and in Fig. 3.20. However, variations in these coefficients by different researchers can be
traced to the uncertainty and empiricism associated with turbulent diffusion models.
For turbulent boundary layers, Eqs. 7.78 & 7.79 are not appropriate due to strong changes
in the turbulent kinetic energy, especially near the wall. The outer region will have time-scales
that are proportional to about 20% of the boundary layer thickness () and friction velocity (u
fr
).
For the near-wall region, an empirical expression for this time-scale by Kallio & Reeks (1989)
compares well with DNS statistics (Fig. 7.7):


( ) ( )
2
wall wall
7.1 0.57 y 0.0013 y
+ + +

= +
L
for
wall
y
+
<100 7.80
A generalized form which approximated this result close to the wall and relaxes to the outer
region value far from the wall can be constructed as:


( ) ( )
2
f wall fr fr
min 10 0.4y / u , 0.2 / u
+


+

L
for boundary-layer flows
7.89
This compares reasonably with experiments (Hinze, 1975) and simulations (Fig. 7.7). The
turbulent length-scales in a boundary layer are also different from that of free-shear flows. As
shown in Fig. A.13b, the lateral length-scale in the outer region (y/>0.15) is on the order of
25% of the boundary layer thickness while the inner region decreases towards the wall. Using
Eq. A.129X, The overall integral length-scale can be approximated as:
( )
3
1
wall
2
0.25 1 exp y / 25
+





|

7.82
The integral time and length scales from Eqs. 7.81 and 7.82 can also be used for pipe and
channel flows by replacing the boundary layer thickness with the hydraulic radius for the outer
region values (Iliopoulos & Hanratty, 1999).
If turbulent bias effects are important, additional modeling may be important to express the
Favre-averaged relative velocity since the Favre-averaging definition differs between that for
particle properties and that for fluid properties (Eq. 7.48). This difference can be related to the
bias mechanisms associated with turbulent structures discussed in 3.5.5 and modeled as:


( )
i i i
i i i i i pref ,i clus,i
v u u
w v u v u w w
1

= = + +


7.83
The RHS expression takes into account the differences between Eqs. 7.48a and 7.48b and
includes the preferential and cluster bias mechanisms employed in Eq. 3.123. Note that the
other bias mechanisms identified in Eq. 3.112 are not needed for Eq. 7.83 as they are already
included within the formulation of v though the last three terms in Eq. 7.67. In particular, the
non-linear drag bias is modeled by Eq. 7.58 while the turbo-phoresis and concentration
gradient bias effects are modeled using Eq. 7.74 and description of the continuous-phase
turbulence field (to be discussed in 7.1.8).
The preferential bias can be approximately modeled by Eq. 3.121 for heavy particles as:


( )
i
pref ,i 3 2
St g w
w 1.15
1 2.2St g
1 4.4 w/ u





+
+


7.84
348
The Kolmogorov scales can be related to the prediction of the continuous-phase dissipation and
the laminar viscosity by Eqs. A.92b and A.93:

( )
( )
1/ 2
f
1/ 4
f
/
u

=
=

7.85a

7.85b
For particles which densities on the order of the surrounding fluid or much lower than the
surrounding fluid, the preferential bias is more difficult to predict as discussed in 3.5.5, and
thus frequently is not modeled. Similarly, the clustering bias is difficult to model and so is
neglected.
To convert the Favre-averaged velocity fields to a conventional time-averaged fields,
proper translation is required. The continuous-phase velocity transformation from Eq. 7.48b
can be approximated (Wang et al. 1994) by Eq. 3.130 as:


( )
2
i
i i i
t i
c k
u
u u u
1 1 Sc x


= +

7.86
The Eulerian averaged particle velocity can be transformed by Eq. 7.48b and the bias
approximated by analogy of Eq. 3.130 but with particle kinetic energy (Eq. 3.64) as:


2
p
i
i i i
t i
c k
v
v v v
Sc x


= +

7.87
Other fluid and particle characteristics can be similarly transformed.
If the particle energy also varies, the energy transport equation is needed which leads to
additional correlations which must be modeled or neglected. The primary turbulent diffusion
of the energy is given by the last term in Eq. 7.72 which expands as:

( )

p j
p j p j
j j j
e v
e v e v
x x x

= +

7.88
The first term on the RHS represents the diffusion due to particle concentration gradients while
the second represents diffusion due to gradients in the particle energy fluctuations. The
correlation associated with these terms can be modeled with gradient diffusion hypothesis (Eq.
A.121). Applying this to the continuous-phase energy with isotropic turbulent diffusion yields

2
tot tot
tot j t
j t j
c k
e e
e u
x Sc x

7.89
For the correlation of Eq. 7.88, one may instead use the isotropic particle diffusivity based on
Eqs. 3.37, 3.45, 3.80 and 7.75 (Chen & Pereira, 1995; Chahed et al. 2003) which yields:

1/ 2
2
2
p p
p j p,t , j
j t i j
e e c k
w
e v 1+
x Sc x

L
NSI 7.90
349
This allows for anisotropic diffusion due to eddy-crossing effects based on Eq. 7.77. There are
also triple (and higher) correlations which appear in the energy equations but these are often
ignored since they are weak or difficult to predict (unfortunately, it is often the latter).
Numerical techniques applied to these equations are similar to those for the steady-state
viscous particle-phase equations as described in 7.1.2-7.1.4, e.g., the time derivative terms are
employed to march the system to convergence with local time-stepping (B.3.2). Since fine-
scale structures do not need to be captured, many of the schemes are second-order accurate in
space and use a first-order temporal discretization (Crowe, 1982; Deshpande et al. 1993; Slater
et al. 2003). For explicit time-marching, upwinding is often used for the particle transport to
avoid numerical instabilities (Chen & Pereira, 1994). However, this often requires local time-
stepping to be limited on the order of the particle response time (Deshpande et al. 1993) and
the addition of limiter or other forms of artificial diffusion are often needed to ensure stability
(Tomiyama & Shimida, 2001). Sample RANS predictions for one-way gas-particle flows are
shown in Fig. 7.8 and also Fig. 5.5b. Another example of RANS-based particle predictions is
given in Fig. 3.32 for Brownian diffusion, turbophoresis and turbulent diffusion. For two-way
coupling, the equations become numerically stiff. This may be treated by making the time-step
proportional to
p
/(1+) for small mass loading. However, large mass or volume-fraction
loading are best treated with the Partial Factorization technique discussed in 7.1.4 (Wang et al.
1994).


PDF Transport Methods

Another option to treat time-averaged particle dispersion is to use an Eulerian model
constructed from a kinetic-based PDF formulation. This approach solves transport equations
for the probability distributions directly. The governing equations are based on the statistical
properties of the underlying turbulence and particle characteristics which are then integrated
over all probabilities to yield a deterministic set of PDF equations. The PDF can describe the
probability of a particle having a certain particle velocity and diameter which can vary as a
function of position, and time, P(v,d,x,t). In such an approach, a path-averaged PDF is defined
as P for which a modified Fokker-Planck equation is written (Reeks, 1993) Once this mean
PDF is known, all the statistics particle characteristics can be obtained through integration, e.g.,
mean concentration, velocity, diffusion, etc (Elghobashi, 1994). The transport equation for
PDF can be written as:


( )
p diff
1 1
t


+ =

v
v
x v
P
P P
7.91
where
diff
is a time-scale for diffusion in phase space, which is a function of dispersion tensors
related to particle trajectory integrals. Reeks noted that this time-scale description must be
developed properly to insure Galilean invariance on the mean. A detailed review of PDF
modeling is given by Mashayek & Pandya (2003) for a variety of approaches. This review also
shows good agreement between the PDF approach and DNS data for diffusion and mass
transfer in homogeneous isotropic turbulence and homogenous turbulent shear flows
The PDF method is attractive from a number of viewpoints: it allows a more fundamental
description of diffusion in probability space (including higher-order moments), it can allow for
350
non-Gaussian distributions, and it can include wall boundary conditions more accurately since
it involves wall velocity information. Because of this, the PDF approach is a useful tool to
investigate and apply closure models within simple (e.g., 1-D) flowfields. However, it can be
somewhat complex to include complex geometries and two-way coupling by such a probability
space. Furthermore, such formulations still require descriptions of the continuous-phase mean
flow and turbulence properties throughout the domain, e.g., by a RANS approach if not known
in advance, which is often a limitation for flows with separation or strong unsteadiness.


LES/Hybrid Correlations for Particle Diffusion and Bias

Both of the above time-averaged approaches to multi-phase turbulent flows are directly tied
to the continuous-phase mean-flow and turbulence properties, i.e. predictions of particle
velocity and concentration distributions are limited by the performance capability of RANS.
This can be particularly worrisome for complex flows for which RANS may not be accurate or
appropriate (A.5.5). Furthermore, the neglect of some particle correlation terms (e.g., the
high-order correlations of the energy equations) or the difficulty in representing other
correlations (e.g., the preferential bias model of Eq. 7.84) introduces further uncertainties to
this overall approach. As such, LES and Hybrid methods (A.5.7) represent an alternative
which reduces the problems and reliance of the continuous-phase turbulence modeling and the
particle correlations. Resolving the large-scale turbulent structures with a partially-resolved
approach (such as LES) allows one to robustly capture non-homogenous anisotropic turbulent
(NAsT) dispersion. Furthermore, if the particle response time is longer than the sub-grid time-
scale (St

>1), such an approach can capture non-linear drag bias and preferential concentration
effects (3.4).
The particle correlations discussed above can be used with an LES approach with some
modifications applied as sub-grid correlations. If the box filter of Eq. A.132 is assumed for the
spatial filtering, the mean transport equations in 7.1.7 can still be employed whereby the
averaged quantities (e.g., , v, u, etc. ) now represent the unsteady three-dimensional resolved
field quantities and the fluctuation correlations now represent the sub-grid quantities. In
particular, one may replace k with k

and

with

. Note that does not require replacement


since the turbulence dissipation is assumed to occur in the smaller scales. Similarly, the
corrections for a hybrid description can be accomplished by replacing k with k
hybrid
,

with

hybrid
. For example, the Nichols-Nelson model (Eqs. A.141-A.144) can be employed to
provide the following:


hybrid hybrid hybrid
c k /

=
7.92a


2
hybrid hybrid
c k /

=
7.92b
Thus, and again, much of the empiricism of the particle coupling is negligible if the particle
have a response time longer than that of the unresolved time-scales, i.e.
p

hybrid
(4.2.2).
From a numerical standpoint, the LES region must be three-dimensional with a grid resolution
that is fine enough to resolve the large majority of the turbulence energy. The same is true
with the temporal accuracy with global time-stepping to ensure that the sub-grid turbulent
kinetic energy is a small fraction of the resolved turbulent kinetic energy. Not that any
351
unresolved turbulence (sub-grid or RANS) may require modification if turbulence modulation
is important, as discussed in the next section.


7.1.9. Eulerian Models for Continuous-Phase Turbulence Modulation

To close the system of equations for the above formulations, the Favre-averaged Reynolds
stress of the continuous-phase may be modeled similar to that of Eqs. A.116b and A.120:

2
j
i k
i j ij ij
j i k
u
u u k 2 2
u u c k
x x 3 x 3



= +




7.93
Generally, two-equation approaches are employed to describe the continuous-phase turbulence
since this provides independent time and length scales of the turbulence (e.g., Eq. 7.78). In
turn, these can be used to model inertia and eddy-crossing effects for particle diffusion (e.g.,
Eq. 7.75). If one-way coupling conditions are considered (0 and 0), the RHS
continuous-phase kinetic energy and dissipation can be given from the single-phase k-
equations (Eq. A.125 or Eq. A.126). Another option is to use a Reynolds stress approach (Eq.
A.130) in which case the LHS of Eq. 7.93 is replaced by Eq. A.130.
For two-way coupling, the turbulence properties are modified by the particles due to
mechanisms discussed in 3.7.3. For the k- approach, a volume-fraction weighting for the
transport variables and turbulent stress tensor is again convenient and can be obtained via Eqs.
A.71, A.73, A.116b, and A.117 and A.122. The multiphase k- equations using these
definitions are similar to the variable single-phase counterparts (Eq. A.126) but they also must
take into account the effects of interphase coupling on turbulence, i.e. turbulence modulation.
This is generally incorporated through source/sink terms for both equations. In particular,
changes in kinetic energy dissipation due to particle fluctuations are given by S
k,diss
(defined by
Eq. 3.168) and changes due to production from particle wakes are given by S
k,prod
(defined by
Eq. 3.173), while particle-induced changes in the turbulent dissipation are given by S
,diss
. The
resulting multiphase k- PDEs can then be written as

( ) ( )

( )
( )
( ) ( )

f f i
j
ij i f f t
i i i i i
f f k,diss k,prod
f f i
ij
i
k u k
u
p k
K u
t x x x x x

u
1.44 K
t x k



+ = + +



+ +

+ =


S S
j
t
i f f
i i i i
2
f f ,diss
u
p
u
x x x 1.3 x
1.92
k


+ +



S

7.94a




7.94b
A rigorous discussion of the principles and issues associated with these multi-phase averaged
transport equations is given by Elghobashi & Abou-Arab (1983) while models for the various
source terms are discussed in the following.


352
Dissipation of Turbulence by Particles

As noted in Fig. 3.36 and 3.7.3, the presence of particles can both decrease and increase
turbulence kinetic energy. The decrease is due to dissipation caused by the particle
fluctuations and typically dominated shear-induced turbulence. The dissipation coupling term
(S
k,diss
) represents the kinetic energy consumed by particle relative velocity fluctuations and can
be modeled based on Eq. 3.168 as

*
@p p f p
k,diss
p f @p @p p
St c c
2 k
2 k
1 St

+ +
=



+ +

S
7.95
Thus, the dissipation increases with higher values of the particle concentration, density ratio
and kinetic energy. If the added mass term is neglected, the RHS expression is consistent with
the gas-particle model of Lightstone & Hodgson (2004). If the eddy-crossing effects are also
neglected, this expression reverts to that of Ahmadi & Ma (1990) and is quite similar to that of
Chen & Wood (1985). The particle-induced sink in dissipation is typically assumed to have a
rate based on the kinetic energy sink and the dissipation time-scale, i.e.

*
p k,diss
,diss
@p p
c
2
k

+

+

S
S
7.96
If added mass effects are neglected, the RHS expression reverts to that of Chen & Pereira
(1995) for high Stokes numbers and to Lightstone & Hodgson (2004). For bubbly flow, where
the particle mass is negligible, the above model is consistent with Lahey (2005).


Production of Turbulence and Pseudo-Turbulence by Particles

The production term of turbulence due to particle wakes is often neglected as the
dissipation process generally dominates. However, it can be important when the particle
relative velocities are high, about
*
term
w 8 > as per Eq. 3.176. Unfortunately, there is not a
strong consensus a model for this effect. One approach is to assume a form based on Eq. 3.173
which allows for general density ratios

2
prod p f
k,prod
p f
c w c




S 7.97
This assumes a constant production coefficient (about c
prod
=0.025) based on high-loading
results of Fig. 3.40. A similar form proposed by Yarin & Hetsroni (1994) models the
production coefficient in terms of particle wake length, particle spacing and integral length-
scale of the turbulence.
To additionally include pseudo-turbulence, which can be important where the particle
density ratio is on the order of unity or less, such as bubbly flow, it is common to use a model
the volume-averaged equilibrium value given by Eq. 3.177 (and Eq. 3.181) as:

1 1 2 2
,eq
2 2
k c w c w

7.98
353
An isotropic version of this net kinetic energy can be obtained from Eq. 3.178:
( )
2
1 2
i i ii i
10
u u c 3 w w


= +


7.99
The simplest approach is to assume that the pseudo-turbulence is everywhere in equilibrium
and can be summed with the conventional turbulence of Eq. 7.94a to provide an effective net
turbulence, i.e. Eq. 3.82 can be replaced by

eff ,eq
k k k

+
7.100
However, prediction fidelity can be improved by employing a non-equilibrium transport
equation for the pseudo-turbulence (Lopez de Bertodano et al. 1994) as

( )
,eq
i t ,eff
i i p
k k
k k
u k
t x x


= +



7.101
This can be combined with Eq. 3.82 and the conventional transport equations for k and as
well as an effective turbulent viscosity given by:

2
turb,eff turb turb
c k / 0.6d w

= + = +
7.102
This approach was found to give reasonable predictions for distribution of volume fraction and
the anisotropic turbulent fluctuations (Fig. 7.9). A more sophisticated approach is to write a
full transport (anisotropic) PDE for the pseudo-turbulence stress tensor as employed by Chahed
et al (2003). However, all such approaches are limited to the predictive ability of the RANS
techniques, e.g., poor prediction can be expected for separated flow regions.



7.2. Lagrangian Point-Force Formulations

For the Lagrangian approach, the characteristics of the particles (or parcels) are identified
with some initial conditions (including location, particles/parcel, physical attributes, etc.) upon
initialization or injection. The trajectory ODEs of position and velocity are then integrated in
time and space along their respective particle paths along with any changes in the particle
characteristics (temperature, mass, size, etc.) associated with the motion. The following
sections include a list of the ODEs, numerical aspects to integrate these ODEs, represent
polydisperse size distributions, model turbulent diffusion, model Brownian diffusion, include
two-way coupling, and treat for particle-wall/particle-particle interactions.

7.2.1. ODEs and Introduction of Particle Properties

ODEs for Particle Trajectories and Characteristics

The particle equations for the point-force approach are written in the particle path
derivative formulation, forming a set of ordinary differential equations. The equations for
particle position, mass, linear momentum, angular momentum, and energy are given by Eqs.
354
1.84, 1.87, 1.90, 1.92, 4.1 and 6.152 as:


p
t
=
x
v
d
d

p body coll surf
m
t
= + +
v
F F F
d
d

( )
p p
surf coll
t
= + T T
d I
d


7.103a


7.103b


7.103c


( )
p
p f f ,p
2
d
2m 2 B Sh
t d d

= =

` d
d

( )
( )
p,p p phase p
f
f p 2
p p
T 3h
6Nu d
T T
t d t m d
= =
`
d c Q
k d
d d

7.103d


7.103e
Note that the momentum equation (Eq. 7.103b) can accommodate three-way coupling effects
in the surface force (from Eq. 6.272).
The particle momentum equations in cylindrical coordinates must be modified if one
considers the motion relative to a rotating reference frame of angular velocity
fr
and radius r
fr
.
This results in additional terms known as centrifugal and Coriolis terms for both the fluid
momentum and particle momentum equations. The latter is given (Kochevsky, 2005) as


( )
p body surf coll p p f fr fr p p fr
m 2
t
= + +
v
F F F r v
d
d
7.104
It should be kept in mind that these additional two terms on the RHS are simply a consequence
of observing the accelerations in a relative rotating frame (which disappear in a non-rotating
reference frame). As such, these two terms are sometimes called pseudo-forces.
If the flow-field is steady (or time-averaged as in RANS) and only one-way coupling is
assumed, then the particle trajectories can be integrated simultaneously and independently of
the flow solution. If the flow-field is unsteady, then the particle trajectories should be solved
simultaneously with the continuous-phase flow-field evolution. If the time-variations in the
pressure are slow compared to the natural frequency, then the bubble volume can be considered
to be in local equilibrium with the surrounding fluid pressure (Eq. 7.26 or Eq. 7.27). Otherwise,
a Rayleigh-Plesset ODE can be integrated in time to capture the oscillation dynamics.
If liquid compressibility can be ignored, the ODE of Eq. 2.73 can be discretized with a
second-order central differencing in time for the interface acceleration (similar to that used in
Eq. B.28) and second-order backward differencing for the interface velocity (similar to that
used in Eq. B.27b) to yield:


( )
( )
( )
( )
( )
( )
2
n 1 n 1 n 1 n n 1 n 2 n n
p p p p p p
n
p 2 2 n
p
n n 1 n 2
2
p p p
1 n
n 4
p
r 2r r 3 3r 4r r
P p
2
r
r
t 2 2 t
4 3r 4r r
w
2 r t
+

+ +

=


+
+

g
l l
l
l

7.105
355
The gas pressure (p
g
) can be approximated using Eq. 2.63 if no internal variations are
considered. However, the local liquid pressure (p
l
) may be a function of the local volume
fraction and so will require an implicit solution for this equation, e.g. coupling with the
iterative Poisson-like equations for the unsteady pressure as will be discussed later in this
section. Integration of this equation will generally require temporal resolution on the order of
the natural frequency (Eq. 2.66). If this is much smaller than the bubble response time (given
by Eq. 3.1), then sub-stepping may be used for the shape oscillations coupled with larger time-
steps for the bubble motion. As noted in 4.1.1, the above equations are equally applicable if a
parcel approach is applied.


Injection or Initialization of Particles and Parcels

Introduction of particle trajectories can be computed with individual particles or with
parcels. Furthermore, particles can be injected at a point or at a plane are initialized
throughout a volume. These four different cases are discussed below starting with
initialization and injection of individual particles.
For initialization with individual particles, the particle-laden region is identified and the
particles are distributed according to their number density (n
p
of Eq. 3.141). For a 3-D region
which has a uniform distribution of monodisperse particles, their centroids can be placed on a
simple lattice with equal spacing in the x, y and z directions (and related to the volume fraction
of Eq. 3.143) as:

( ) ( ) ( )
1/3 1/3
p p p p p
init init init
x y z / n

= = = = 7.106
The number-averaged particle volume will simply be uniform for a monodisperse size
distribution or can be randomly sampled for a polydisperse size distribution as discussed in
7.2.4. For 2-D, this same resolution can be used for x and y assuming that the domain has a
unit width of z
p
. If the number concentration varies in any direction, then the spacing
between particles in that direction can vary accordingly. The particles are typically numbered
and stored on a list with the other initial particle characteristics (velocity, spin, temperature,
etc.). For polydisperse size distributions (where diameter in an additional characteristic), the
sizes also require initialization (to be discussed in 7.2.4).
For injection with individual particles, trajectories are typically released over a fixed area
(A
inj
) based on a net injection mass-flow rate (
inj
m` ) and the normal component of the
prescribed particle injection velocity (v
inj
). This can be related to the injection volume fraction
though the flux transport of Eq. A.1. If q is taken as the mass of particles per mixed-fluid
volume (
p
), then the first term on the RHS represents the particle mass flux through a
control surface. Applying this relation to the injection cross-sectional area and averaging over
time yields:


( )( )
inj
inj p inj inj p inj inj
A
m A v A = =

v n `


7.107
If the time-step is fixed between particle injections (e.g., set equal to the continuous-phase
time-step) and injection is in the x-direction, the particle spacing for a rectangular y-z
356
initialization plane is given by the particle mass flux and can be related to the injection volume
fraction by the above equation:
( )
p p p 2 2
p p
inj
inj inj
A
y z
m t v t

= = =



`

7.108
Note that the resulting streamwise particle spacing (x) will initially be equal to (tv)
inj
in this
case. To allow for the advantages of an isotropic particle spacing (uniform in all three
directions), this streamwise spacing can be substituted into the RHS of Eq. 7.108. The result is
then consistent with a spacing given by Eq. 7.106 if the time-step between new particle
injections is given by (x
p
/v)
inj
. As new particles are injected, they can be added to the particle
list. In some, multiphase flows there may be a very large number of particles which can have
an impact on memory and CPU requirements.
A parcel approach can be important if the number of physical particles to be modeled at a
given time is computationally impractical. For example, a distribution of 25 micron diameter
water droplets distributed in a 1 liter volume with a 10% mean mass fraction corresponds to
more than 10
7
droplets. To avoid tracking such a large number of particles, a parcel may be
employed, which is also referred to as a packet, a computational particle and a super-
particle. Each parcel represents a cloud of N
pP
(particles per parcel) particles which share a
prescribed set of physical attributes (location, velocity, temperature, diameter, density, shape,
etc.) and follow the same ODEs (Eq. 4.7). As such, a parcel retains the advantage of a
Lagrangian method while significantly reducing the computational cost since the trajectories to
be tracked is reduced by a factor of N
pP
.
For initialization with monodisperse parcels, the number of particles per parcel for a
given N
P
can be computed with Eq. 4.9a. For a three-dimensional domain with isotropic
Eulerian mesh volumes (Eq. B.13), the initial parcel locations can be placed on a simple lattice
given by:
( ) ( ) ( ) ( )
1/3 1/3
1/3
P P P P P p pP
init init
init init
x y z / N N /

= = = = = 7.109
For injection with monodisperse parcels, an equation similar to Eq. 7.108 can be applied:
( )
pP p p pP p 2 2
P P
inj
P P P
inj inj inj
N A N
y z
m t v t N v t


= = = =



`

7.110
Similar to individual particle injection, the time-step between new parcel injections can be set
as (y
P
/v)
inj
, which recovers a distribution like Eq. 7.109 just after injection. For steady
continuous-phase flow, one may simply specify N
P
to compute the appropriate time-step
between injections. For unsteady continuous-phase flows for which t is specified for both
phases (e.g. by CFL criterion of Eq. B.43), N
P
can be adjusted so that t
inj
equals or is some
multiple of t.
Typically, N
pP
is set to be a uniform value for the initialization region based on a targeted
value of N
P
. For one-way coupling and monodisperse particle sizes, the parcel volume should
be limited by the Eulerian grid size (Eq. 4.10) to ensure reasonable interpolation (to be
discussed in 7.2.2), i.e.
357

P
N 1

for one-way coupling


7.111
Polydisperse size distributions (to be discussed in 7.2.4) may require even greater resolution.
For two-way coupling with Lagrangian parcels (to be discussed in 7.2.7), it is helpful to have
several parcels per cell so that the feedback to the Eulerian mesh is approximately uniform, and
a typical target value is:

P
N 4

for two-way coupling


7.112
The parcel positions can also be based on consistent values of shape functions for a given N
P

(Sivier et al. 1996) which can be convenient for unstructured grids. If initialized particles say
within regions of similar grid resolution during their trajectories, these criteria can be used with
Eq. 7.109.
If the Eulerian mesh is significantly non-uniform in size, then parcels initialized in a region
with coarse grid resolution and the above criteria may end up in a region with much finer
resolutions so that the above criteria will be violated locally. To prevent this, one may
conservatively set N
pP
based on the minimum grid resolution anywhere in the domain (rather
than the grid resolution in the region of initialization). If the leads to a large number of parcels,
one may employ dynamic parcel adaptivity to locally adjust N
P
to an optimum value,
though this technique can be diffusive (Sivier et al. 1996).

7.2.2. Interpolating Fluid Properties at the Particle Location

Employing a Lagrangian trajectory scheme requires a particle tracking technique which can
operate within discrete framework on the continuous-fluid solution. Generally, three steps are
needed: 1) determination of the next Eulerian host cell of the particle centroid, 2) interpolation
of the continuous-fluid properties to the particle centroid, and 3) integrate the particle to the
next time-step. The first two steps are discussed below, while the advection step is addressed
in 7.2.3. When a parcel is employed (vs. an individual particle), the following techniques can
be used to obtain the parcel cloud centroid.


Host Cell Determination Schemes

Identification of the host cell (i.e. the computational cell within which the particle centroid
is located) is typically straightforward for structured grids. For example, a grid in a Cartesian
cuboid domain (which can be described with three pairs of orthogonal rectangles) will typically
include a mesh with the total number of cells given by Eq. 4.80, i.e. the total number of cells is
the product of the cell resolution in each direction given by N
fx
N
fy
N
fz
. If the particle location
(x
p
, y
p
, z
p
) is known then the corresponding host cell indices (i
p
, j
p
, k
p
) are obtained by simple
linear interpolation. For example, the x discretization and the x host cell index are given by:

( )
( )
x fx max min fx
p p min
x D / N x x / N
i 1 integer x x / x
=

= +


7.113a

7.113b
358
The latter equation uses the computational operator integer(q), where q is a real number and
the result is an integer value which rounds off any fraction below unity, e.g., integer(5.73)=5.
Similar equations can be used for the y and z discretizations and the y and z host cell indices.
If, and only if, a uniform stretching factor is used (Eq. 4.81), the host cell index is given by
( ) ( )
{ } p x p min o x
i 1 integer ln 1 x x / x / ln 1


= + + +


7.114
Similar equation can be written for the stretched y and z host cell indices. For a curvilinear
coordinate structured grid with hexahedron (six-sided) elements, the particle location is first
mapped into the transformed coordinate space using the mapping Jacobian (Anderson et al.
1997) after which the host cell may be found based on any stretching within the transformed
coordinates. Optimum search methods for multiple block meshes (e.g., Chimera grids) with
curvilinear coordinates are given by Chen (1997). However, the computational time associated
with determining the host cell for structured grids is generally small compared to overall
computational time per iteration.
For unstructured grids (e.g., the triangular grid regions in Fig. B.1), determination of the
host cell can be more complex since the cells are generally not numbered according to an
ordered increase in location, i.e. an explicit equation such as Eq. 7.114 is not available. A
brute-force global search over all the cells to determine i
p
is generally impractical due to the
large computational cost required. Therefore, efficient host cell searching techniques are
important whenever particles are injected or whenever they move to a new cell. In general,
such techniques for unstructured grids first determine whether the current host cell is simply
the old host cell (this is the cell where the particle was previously located or a first guess in the
case of initial injection). If so, the search is complete as is the case for particle P1 in Fig. 7.10
where it remains in element E1. This result will occur often when the particle movement is
small compared to the length-scale of the cell. If the old host cell is not the current host cell,
the most probable adjoining cell is interrogated to determine if it is the host cell, e.g., element
E2 is searched to find particle P2. If this predicted host cell is still not the host cell (e.g., for
particle P3), an adjoining cell to the predicted cell is next considered (e.g., E3 for particle P3).
If this is still not the host cell, this process is repeated until convergence. Thus, the search
moves along a path of adjoining cells until the host cell is located. This overall approach
includes two key aspects: interrogation of a cell to see if it is the host cell and determination of
the most probable adjoining cell when this is not the case.
To interrogate a cell with planar facets (e.g., any of the cell shapes shown in Fig. B.2),
there are two basic approaches: the shape function approach and the dot product approach.
The shape function approach is well suited for triangular or tetrahedral cells. This technique is
detailed by Lhner and Ambrosiano (1990) and uses the linear shape functions at each node
with respect to the particle position. For the example of the a triangular element of Fig. 7.11,
the local shape function associated with the node 3 can be expressed using Eqs. B.7 and B.8 in
terms of the three nodal coordinates and the particle position as:


( )
( )
2 p p 2 1 p p 1 1 2 2 1
3,p
3,p
2 3 3 2 1 3 3 1 1 2 2 1
x y x y x y x y x y x y
A
= =
A x y x y x y x y x y x y

+ +

+ +
7.115
Using this equation a particle on a node will yield a shape function () of unity while a particle
placed farther away within the element will have a linearly reduced shape function will reach a
value of zero at the opposite edge. A particle lying outside of this element edge will yield a
359
negative value for the shape function. Thus, shape functions can be used to test whether a
particle is within an element. If all the shape functions are positive (e.g., particle P
1
of Fig.
7.11), the host cell has been found. However, if any of the shape functions are negative (e.g.,

1
with respect to particle P
2
), this is not the host cell. Of the adjoining elements, the most
probable to be the host cell is that which shares the edge associated with a negative shape
function. This method is very efficient and can be readily extended to 4-noded tetrahedral
elements in 3-D. However, it is based on elements for which each node has a unique opposing
face and so will not work for certain elements (e.g., five-side pyramids, six-sided hexahedrons,
and five-sided prisms). Patankar and Karki (1999)

employed a method for 2-D rectangular
elements (e.g., a square) and 3-D hexahedral elements (e.g., a cube) by using a local coordinate
system along the two or three primary directions. This is also an efficient method, but such a
technique does not work directly for tetrahedral elements and prisms.
To handle a variety of cell shapes (e.g., triangular, rectangular, tetrahedral and hexahedral),
the dot-product method is preferred. In this method, two vectors associated with a given cell
face and a particle location are examined. The first is the face-normal vector (n
j
) which is
defined as positive in the outward direction for element face j. This vector is dotted with a
face-particle vector (d
j,i
) which points from the centroid of element face j to the location of
particle i. If a negative dot product is encountered between these two vectors for all the cell
faces (e.g., particle P
1
in Fig. 7.11), the host cell has been found and the search is terminated.
However, a positive dot product for any of the faces indicates the particle lies beyond the cell
face and a new host cell must be determined (e.g., particle P
2
in Fig. 7.11). In this case, one
must determine which cell to interrogate next. If only one face-particle vector is negative, the
adjoining element which shares this face is assumed to be the most probable host cell.
It is also possible for a particle position to have two positive dot products, especially if the
particle has traversed more than one element during a time-step or if there is no previous host
cell (e.g., the particle has just been injected) so that the first guess is far from the actual host
cell. For the case where multiple faces yield negative dot products, Martin et al. (2009)
investigated two possible choices: 1) choose the maximum positive dot product (MPDP) after
looping through all the faces to select the next host cell, and 2) choose the first positive dot
product (FPDP) when looping through the faces to select the next host cell. While the first
option will generally provide a more probable host cell and thus fewer cells to interrogate, the
second option can reduce the number of faces to be searched since often only one dot product
is positive. However, the differences are minor (about 20% in each case) and both methods are
robust and quite fast.


Interpolation Schemes

The interpolation step is important since it obtains flowfield properties at the particle
centroid (e.g., the local continuous-fluid density, velocity, and temperature) from which the
interphase phenomena (e.g., drag, heat and mass transfer, etc.) can be computed to predict the
new particle characteristics (e.g., mass, location, velocity, and temperature). This interpolation
depends substantially on the continuous-phase representation since the interpolation scheme is
generally chosen to be consistent with the numerical representation of the continuous-fluid
variables in space and time. In most numerical methods, the variables are assumed to vary
linearly within a local cell. This condition will be discussed first followed by that for high-
360
order variations (e.g., quadratic or cubic are also possible) and then spectral methods (for
which global basis functions are used).
For continuous-phase properties which vary linearly within a cell, the local shape functions
can be used to interpolate the nodal values (q
f
) to the particle location (Eq. B.1):


( )
N
@p i i i p
i 1
q = q ,

x x 7.116
For example, the interpolation of the fluid velocity based on linear shape functions in a 2-D
Cartesian grid cell (Fig. 7.12a) is simply bi-linear:


p 1 p 1
@p 1 2 1 3 1
2 1 3 1
y -y x -x
[ ] [ ]
y -y x -x

= + +



u u u - u u - u
7.117
The x
1
,y
1
coordinate is the lower left grid point and can be obtained from the host cell node
indices i
p
,j
p
, e.g. based on Eq. 7.114. For 3-D, a similar tri-linear interpolation can be written.
For triangular or tetrahedral grids with linear interpolation functions, one may use the shape
functions of Eqs. B.7 or B.11.
Higher-order elements similarly employ a summation of the non-linear shape functions at
the particle position coupled with the fluid properties at the nodal positions. However,
computational speed can be increased by employing trajectory reconstruction within the cell
based on a local coordinate system (Nelson et al. 2004). This expedites the interpolation of
flow properties and the identification of elemental boundary crossing. This is accomplished by
using a polynomial representation of the trajectory segment (constructed after each particle
movement) which is checked for intersections with element boundaries. Once a cell face is
crossed, the technique is terminated and restarted at the element boundary. Coppola et al
(2001) discuss the two basic strategies for tracking particles in higher-order elements. The first
requires a transformation Jacobian to allow for accurate interpolation of the flow properties
using local coordinates. The second performs the entire integration in the local coordinate
system. The first method involves more computational steps but the latter requires iterations to
determine the boundary intersection location due to the discontinuous behavior of the
transformation Jacobians. Coppola et al (2001) also developed a hybrid strategy using a
guided search in order to eliminate the non-linear iterations and yield a two- or three-fold
increase in speed.
For spectral representations of the continuous-phase, care must be taken to interpolate the
variables. One option is to use Eq. 7.116 with the global basis functions for each particle.
While this would be accurate and consistent, it would also be computationally impractical for a
large number of particles since N

in this case refers to the entire spectrum of functions for the


domain and is generally quite large. Another option is to use linear interpolation based on the
surrounding nodal values, but this truncates the higher-order accuracy of the flow solution.
Therefore, a common compromise is to use higher-order Lagrange polynomial interpolation.
This approach uses Lagrange polynomials for a given particle location (L
i
) which are
distributed over N
L
nodes in each direction. For a 1D grid in the x-direction, this
interpolation can be written as:


L
N
@p i i
i 1
q = q L
=


7.118a



361
L
N
p j
i
j 1, j i
i j
x x
L =
x x
=


7.118b
For multiple dimensions with an orthogonal grid, the interpolation becomes:


L L L
N N N
@p i,j,k i, j,k
i 1 j 1 k 1
q = q L
= = =


L L L
N N N
p a p b p c
i, j,k
a 1,a i b 1,b j c 1,c k
i a j b k c
x x y y z z
L =
x x y y z z
= = =









7.119a


7.119b
For linear interpolation, this reduces to second-order accurate linear interpolation over four
nodes given by Eq. 7.117 (N
L
=2, Fig. 7.12a). However, turbulent flows are generally
interpolated with polynomial sets that are quadratic (N
L
=3) or cubic (N
L
=4, Fig. 7.12b), to
ensure sufficient accuracy for Lagrangian statistics (Squires & Eaton, 1990). Other options
include Hermitian interpolations which additionally use derivatives at the nodal locations to
allow a more compact form (Balachandar & Maxey, 1989; Squires, 2007).

7.2.3. Lagrangian Particle Advection Schemes

The numerical technique for Lagrangian particle advection is typically implemented with a
finite difference formulation to update the particle unknowns. Two key issues should be kept
in mind: accuracy and stability. The issue of stability is generally related to explicit schemes
since most implicit schemes are unconditionally stable. To demonstrate this, consider a simple
particle equation of motion (Eq. 3.6) that just includes a linear drag (constant
p
)

p
/ t / = v v d d
7.120
The exact solution to this ODE is simply
( )
o p
exp t / = v v
7.121
For a numerical solution, the Euler explicit scheme of Eq. B.36 can be used which is first-order
accurate with respect to the particle acceleration and second-order accurate for the particle
velocity. Using the discrete notation introduced in Eq. 3.8, the particle velocity scheme is:
( )
n 1 n 2
p
1 t / ( t )
+
= + v v O
7.122
After n+1 timesteps, this particle velocity can be expressed in terms of particle time-step:

( )
n
n 1
o p
p p
1 CFL
CFL t /
+
=

v v

7.123a

7.123b
As shown in Fig. 7.13a, this scheme is numerically unstable for
p
CFL 1 > , a result typical of
explicit particle schemes and similar to that for first-order explicit continuous-phase methods.
On the other hand, consider the backward implicit scheme which employs a first-order
accurate representation of the particle acceleration at the new time-step:

( )
1
n 1 n 2
p
1 CFL ( t )

+
= + + v v O 7.124
362
This scheme is unconditionally stable as illustrated in Fig. 7.13b, which is typical of implicit
schemes. However, both of the above schemes can suffer from inaccuracies when the time-
step is on the order of the particle response time. Significant improvements can be made by
employing the Crank-Nicolson scheme based on Eq. B.51 which evaluates Eq. 7.120 at the half
time-step and can be made into an explicit scheme (for this linear ODE) by using second-order
averaging for the RHS:

n 1/ 2
n 1/ 2 n 1 n
2
p p
( t )
t 2
+
+ +
+
= = +



v v v v d
O
d
7.125
The LHS term is then replaced with second-order central differencing to yield:

p n 1 n 3
p
2 CFL
( t )
2 CFL
+

= +


+

v v O
7.126
This yields a significantly improved accuracy as shown in Fig. 7.13b, an increases the range of
numerically stability to
p
CFL 2 < .
The principles associated with these basic explicit and implicit schemes are applied, in the
following, to increasingly complex equations of motion. If one neglects collisions (to be
discussed in 7.2.8 and 7.2.9), the particle equation of motion based on Eqs. 6.273 and 7.103b
can be expressed as:

( )
( )
( )
( )
( )
( )
( )
( )
( )
p f
L
D
p p f p f p p f
f f @p H
D g L FA H
p f p p f
1
f
t c c c
c 1 1

t c c

= + +
+ + +
+

+ + = + + + +
+ +
g
F
F v
u F
d
d
D
D
F F F F F


7.127

The RHS has been broken down into four terms which have units of acceleration and are
related to: drag (F
D
), gravity and buoyancy (F
g
), lift (F
L
), fluid acceleration (F
FA
) and relative
particle acceleration associated with the history force (F
H
). The latter two terms require special
care if they are to be retained, and they may be important for unsteady motion of particles and
bubbles in liquids. Thermophoresis effects can be included in the same manner as that for lift.
The above equation is supplemented by the particle position equation (Eq. 7.103a). The next
sub-sections consider conditions with increasingly complex versions of these terms for:
a) approximately linear drag, nearly constant lift and no acceleration effects (F
FA
=F
H
=0)
b) approximately Newton-based drag, nearly constant lift and no acceleration effects
c) generalized surface forces and no acceleration effects
d) inclusion of fluid acceleration and/or three-way coupling effects, and
e) inclusion of history force effects.
This is followed by two sub-sections on global and local particle time-stepping.


Approximately Linear Drag with No Accelerating Effects

If one neglects three-way coupling (reasonable for a sparse distribution of heavy particles
in a gas), assumes that all non-drag forces are approximately steady over the times-step, the
particle acceleration for linear drag can be expressed as:
363

( )
( )
f @p
g L
2 *
p
18f
t d c

=
+
v - u
v

d
+ +
d
F F
7.128
If lift and u
@p
are approximately constant throughout the timestep, the particle equations may
be integrated to create an explicit exponential schemes (as in Eq. 3.8 and as in Eqs. 5.104-
5.105) given by:

( )
( )
( )
( )
( )
n n
p p
n
p
t/ t/
n+1 n n n n
@p g L p
t/
n+1 n n n n n n n n
p p p g L p @p g L p
e 1 e
1 e t


= + + +


= + + + +

v v u
x x + w u
F F
F F F F

7.129a

7.129b
This one-step exponential scheme is popular as it is inherently stable and exact if
p
, F
L
and
u
@p
are indeed constant.
If
p
, F
L
or u
@p
vary within a timestep, this will reduce the accuracy of the scheme. For
example, consider the limit of small particle response time such that the relative velocity exact
solution is simply w
term
at all times. In this case, the new particle velocity is given by
n+1 n+1
@p term
= + v u w , but the above expression yields
n+1 n
@p term
= + v u w . This problem can be
approximately overcome by using a predictor-corrector scheme. For example, the result of Eq.
7.129 can be defined as a predictor step
n+1
pred
v and
n+1
pred
x , which can then be used to estimate the
fluid velocity and particle response time at t
n+1/2
by central-averaging (Eq. B.50):

( )
( )
1 n+1/ 2 n+1 n 2
@p @p,pred @p
2
2
p f
n+1/ 2 2
p n+1 n
f pred
( t )
c d
1 1
( t )
36 f f

= +
+
= +


u u u +O
+O

7.130a

7.130b
These values can be used for a corrector step generalized for quasi-steady force effects as:

( )
( )
( )
( )
( )
n+1/2 n+1/2
p p
n+1/2
p
t/ t/
n+1 n n+1/ 2 n+1/ 2 3
@p g L
t/
n+1 n n+1/ 2 n n+1/ 2 n+1/ 2 n+1/ 2
p p p @p g L p
n+1/ 2 n+1/ 2 n+1/ 2
@p g L p
e 1 e ( t )
1 e


= + + +


= +


+ + +

v v u
x x + v u
u
+O F F
F F
F F
( )
3
t t +O

7.131a


7.131b
Further improvements in accuracy can be obtained by including the velocity information at t-t
(Barton, 1996). While based on the assumption of linear drag, the above scheme is generally
reasonable for particles at intermediate Reynolds numbers, e.g. Re
p
values of 40 or more. At
higher Reynolds numbers, different schemes may be developed based on the Newton-drag
regime as discussed below.


Approximately Newton-Based Drag with No Accelerating Effects

For the Newton drag regime, C
D
is approximately constant. A Newton-based scheme can
be developed by evaluating the particle ODE of Eq. 7.126 at t
n+1/2
under this assumption. A
364
second-order central-difference discretization (Eq. B.49) can be used for the LHS particle
acceleration. The RHS can again be written in terms of a drag force and quasi-steady forces.
The non-linear drag term can be treated with quadratic linearization (based on a Taylor series
expansion similar to that used for Eq. B.50) to give second-order accurate expression (for 1-D
with no sign change):

( )
n+1/ 2
n+1 n 2
w w w w ( t ) + O 7.132
Again assuming negligible three-way coupling effects (1) and that F
L
and u
@p
are
approximately constant over the time-step, the discretized ODE becomes:

( ) ( ) ( )
n+1 n n n+1 * n 3
D p g L
3 C t 4d c t ( t )


= + + +

v v w w + +O F F
7.133
This can be rearranged as a one-step scheme for the velocity and while the particle position of
Eq. 7.103a can be integrated based on second-order central differencing in time (Eqs. B.49 and
B.50). This schemes are then given as:
( )
( )
( )
( )
-1
n n n n n
D @p D
n+1 n n 3
g L
* *
p p
3 tC 3 t C
t 1+ t
4d c 4d c



= + + +
+ +

w u w
v v

+O F F
( ) ( )
1 n+1 n n+1 n 3
p p
2
t t = + + x x v v +O

7.134a

7.134b
These one-step schemes have forms similar to those in Eq. 7.126 which makes them
convenient and efficient. However, these Newton-based schemes are not exact even when
the drag coefficient is constant (unlike the exponential scheme which is exact for linear drag).
Thus, the time-step should be chosen to be sufficiently small so that the truncated terms have
little influence per Eq. 3.9. Typical values of the time-step are given by t
p
/8 (Bocksell &
Loth, 2001).
As with the exponential schemes, the Newton-based schemes can also be applied in the
intermediate Reynolds number regime through linearization. In particular, the drag coefficient
may be evaluated based on the most recent condition, i.e.
( )
n n
D D p
C C Re = . Such scheme are
widely used for spray simulations (Amsden et al. 1989). The choice of exponential or Newton-
based schemes for particles which are in between the linear and Newton regime can be based
on accuracy considerations. For example, errors in particle velocity as a function of particle
Reynolds number are shown by Fig. 7.14 for the two schemes. The results indicate that the
exponential scheme is preferred for lower Re
p
values while the Newton-based scheme is
preferred at higher Re
p
values. To minimize error, one may choose between the two schemes
depending on the local conditions, e.g., the exponential scheme if Re
p
<65 and otherwise the
Newton-based scheme (Lee & Loth, 2007). If C
D
, F
L
or u
@p
significantly vary within the
timestep, a predictor-corrector method can again be used. However, it is often more efficient
and robust to treat the surface forces in a generalized fashion as follows, especially if three-
way coupling is important.


Generalized Surface Forces with No Accelerating Effects

365
A popular option for generalized surface forces with three-way coupling is the forward-
differencing Adams-Bashforth scheme (Eq. B.65). This can be applied to Eq. 7.127 with
thermophoresis effects added in by defining a generalized function (F
G
):

G D L g T
+ + + F F F F F
( ) ( ) ( )
n n 1
1 n+1 n 3
G @p G @p
2
t 3 , , t


= +


v v v u v u +O F F
7.135a

7.135b
This one-step explicit scheme employs two previous time-steps and is second-order for particle
acceleration and third-order for particle velocity. It also treats all the surface and body forces
in a consistent fashion so that drag need not dominate other surface forces. The particle
position can again be integrated based on Eq. 7.134b. This scheme is generally more accurate
that the above Crank-Nicolson schemes but requires storage of the velocity at two previous
time-levels. However, it is not self-starting, i.e. the first time-step must be initialized with
another scheme.
An alternative method which is self-starting and requires only one previous time-level of
data to be stored is a predictor-corrector method based on the 2
nd
order Runge-Kutta method
which can be coupled with particle position integration of Eq. 7.134b as:


( )
( ) ( )
( ) ( )
( ) ( )
1 n+1/ 2 n n n
G @p
2
1 n+1/ 2 n n+1/ 2 n 3
p p
2
n+1 n n 1/ 2 n 1/ 2 3
G @p
1 n+1 n n+1 n 3
p p
2
= , t
t t
, t t
t t
+ +

= + +
+
= + +
v v + v u
x x v v
v = v + v u
x x v v
+O
O
+O
F
F

7.136a
7.136b
7.136c
7.136d
The first particle position estimation (Eq. 7.136b) is needed to interpolate the continuous-phase
velocity at t
n+1/2
, and thus requires that u
n+1/2
is known (or can be estimated from u
n
and u
n+1
) if
the continuous-phase flow is updated ahead of time.
A more common explicit method for particle integration is the 4
th
order Runge-Kutta
scheme which can be expressed (without showing the intermediate x
p
integration steps) as:


( )
( )
1 1 n+1/ 2,pred n n n n n
G @p G
2 2
1 n+1/ 2 n n 1/ 2,pred
G
2
n+1,pred n n 1/ 2
G
1 n+1 n n n 1/ 2,pred n 1/ 2 n 1,pred 5
G G G G
6
t , t
t
t
t 2 2 t
+
+
+ + +
= + = +
= +
= +
= + + + + +

v v v u v
v v
v v
v v O
F F
F
F
F F F F

7.137a
7.137b
7.137c
7.137d
This is more accurate than all the above schemes but requires two more computations for
n 1
p
+
x
(and any associated host cell searches and fluid interpolations). Similar integrations can be
used for particle position, mass, torque and temperature with the ODEs of Eq. 7.103.


Schemes for Significant Fluid Acceleration Effects

366
Inclusion of fluid-acceleration effects can lead to instabilities with the above explicit
techniques. Furthermore, two-way coupling effects require that the continuous-phase flow
field be updated in concert. Therefore, implicit schemes are generally preferred for these
conditions, particularly for bubbly flow. A popular such method to discretize Eq. 7.127
without the history force (F
H
=0) is the Adams-Moulton scheme which employs central-
differencing in time and iterations until convergence. If applied, the resulting scheme can be
expressed with an iteration index k as:

( ) ( ) ( ) ( )
k 1 k k
n+1 n n 1 n n 1/ 2 3
G G FA
t
t
2
+
+ +


= + +


v v + + O F F F 7.138
The fluid acceleration term can be obtained by applying Eq. 1.13b to the Eulerian continuous-
phase velocity field solutions at t
n
and t
n+1
:
( )
n 1/ 2
n 1 n n n n
p @p p
@p
/ t ( t t t) ( , t) t
+
+
= +

u u x +u , u x D D
7.139
Typically two to three iterations are sufficient for convergence. For example, to integrate
bubble trajectories, Dorgan & Loth (2009) employed an Adams-Bashforth scheme as the
predictor followed by two iterative Adams-Moulton schemes (which use values at n-1, n and
n+1) for each time-step integration.
Like the Adams-Bashforth scheme, the Adams-Moulton is second-order accurate for
particle acceleration but its truncation term is five times less and it is also unconditionally
stable for linear drag dynamics (though the time-steps should be less than
p
for sake of
accuracy). If the
p
is much smaller than the t dictated by the continuous-phase solution, then
the weakly-separated Lagrangian approach of 5.3.1 can be used to remove this restriction.


Schemes for the History Force

To include the general history force of Eq. 7.141, a numerical integration over all previous
time-steps in required based on the forms discussed in 6.6. This integration requires
significant storage and close attention to temporal resolution. In general, the time-step should
also be a small fraction of the particle response time and the time-scales associated with
momentum change (Eq. 6.192) for sake of accuracy:


( )
diff p conv
t min , , <
7.140
Due to this restriction, explicit schemes are typically used to incorporate the history force
effect, i.e. it is incorporated directly as
n
H
F . For creeping flow, one may employ a rectangular-
rule integration based on mid-point evaluations of the acceleration and the kernel (Reeks &
McKee, 1984). However, the time-step required for proper convergence with this method can
be quite small for unsteady flows, e.g., t=0.001
diff
(Alexander, 2004). A higher-order
formulation for the history force which allows larger time-steps and can be used for a
generalized kernel (H
Re
of Eq. 6.196) is given by Kim et al. (1998) as:
367


( ) ( ) ( )
( )
( )
( ) ( )
( ) ( )
( )
( ) ( ) ( ) ( )
n t
n
n n H
f
0
n 1 n 1 n n
n 1
i 1 i 1 i i
i 1
0.05 t
H t 8 2 H 0.05 t 4 H 0.1 t
3 d 3
0.15 t H t 2 H 0.55 t H 0.1 t
t
H n i+1 t 2 H n i+1/2 t H n i t
6


= =


+ + + +


+ + + +

F
w w w
w w w w
w w w w
` ` `
` ` ` `
` ` ` `
d

7.141
This scheme uses a combination of Simpson and Gauss Quadrature integration to allow
accuracy of approximately O(t
5
) and allows temporal convergence with t=0.01
conv
.
With respect to the memory storage, the history force requires temporal integration over all
times so that the associated data files of the past particle velocities can be cumbersome for
Lagrangian integration. For example, storing the relative velocity vectos for 100,000 particles
over 1000 time-steps using 4-byte real values would require 1.2GB of memory, far in excess of
that required if the history force is neglected. For this reason, many multiphase simulations
have simply ignored the history force out of practical necessity. One method to make the
history force at finite Re
p
more efficient (in terms of both CPU time and memory) is to use a
window-based technique. This approach takes advantage of the transition time definition of
Eq. 6.197 to limit the total integration period to
H
and assumes moderate changes in relative
velocity. For a particle trajectory with zero initial acceleration at t=0, this approximation can
be expressed using a window time (
win
) as

win
win H
t t
H,Re
Re Basset
f
0 t
t max(0, t )
H H
3 d










F
w w d d
d d
d d

7.142a

7.142b

The RHS approximation assumes that the particle acceleration is nearly constant over
transition time interval and thus can be simply approximated as the change in relative velocity:

( ) ( )
win win
(t) t t t

w w w d / d / 7.143
The RHS integral is then straightforward to evaluate as it can be obtained analytically from Eq.
6.193b. This result combined with Eq. 7.143 yields the window-based history force as:


( ) ( )
2
H,Re win f f win
3d (t) t t t

F w w / 7.144
Finally, the transition time used for Eq. 7.142a can be obtained by a path-average of Eq. 6.198:


( )

2
2
H f f p win
(t) d 0.502 Re (t, t ) 0.123

= +

7.145
The path-averaged Re
p
can be efficiently computed with a running sum if the time-steps are
uniform. The numerical efficiency advantage of using the window-based model is evident
when one considers savings in computational time, e.g., CPU cost per time-step can be orders
of magnitude less than the full history force integration as shown in Fig. 7.15. This figure also
shows that the Basset and Mei& Adrian simulations take only half as much CPU as that of the
Kime et al and Dorgan & Loth simulations. This can be traced to the kernel given by Eq.
368
6.196 from which it can be seen that the former two simulations use a square and a square root
operation (since c
H1
=2) while the latter two simulations use two general exponent operations
(since c
H1
=2.5) which are much slower at the hardware level of most computers.
In terms of predictive performance, the window-based model gives good results when the
changes in the particle acceleration are modest relative to the overall changes. Since the
derivative of the acceleration is sometimes called the jerk, this is equivalent to noting that the
particle jerk within the window period is small compared to the overall jerk along the particle
path. For the falling particle case, this is equivalent to specifying
H

p
which is generally
satisfied for a large range of terminal Reynolds numbers and density ratios. In practice,
reasonable predictions are obtained for even faster particle response times, about
H
~
p
. An
example of this is the prediction shown in Fig. 6.68, but the window-model is no longer
quantitatively reasonable for
H
>1.2
p
(Dorgan & Loth, 2007). For particles subjected to a
mean relative velocity with small fluctuations about the mean, the modest jerk criterion implies

osc
but again reasonable results are expected for
H
as large as
osc
. This condition is
probably closest to that of a particle falling through turbulence. However, turbulence leads to
complications associated with multiple flow time-scales, non-rectilinear motion and spatial
non-uniformities such that it is not clear whether any finite Re
p
model will be reasonable.
Another problematic case for the window-model is that of wake re-ingestion because it
leads to a modification of the long-time decay, as discussed above. For the oscillating particle
in a quiescent flow, the modest jerk criterion still applies, i.e. the window-model is appropriate
of the slow oscillations are slow, i.e.
H

osc
. In practice, reasonable agreement is found for
values as large as
H

osc
but predictive ability deteriorates at larger
H
values. An example
of the latter condition where the model breaks down is shown in Fig. 6.69. Therefore, the
window-model should be used with caution if wake re-ingestion can occur.


Global Particle Time-steps

If the continuous-phase flow is unsteady, the particle trajectories should be computed
simultaneously. A convenient and common choice is to use the same time-step for both the
particle and the continuous-phase, where the latter is generally determined by the associated
scheme stability and accuracy (B.3.2). For example, Eq. B.48 can be extended as:
( )
p f i, j,k
t t min t =
7.146
Another option is to base the particle time-step on accuracy and stability of the particle
equation of motion. For example, the Crank-Nicolson scheme of Eq. 7.126 was found to be
quite accurate for CFL
p
values less than unity (Fig. 7.13b). If history force is not included
(F
H
=0), a CFL
p
of 1/8 was found to ensure high-accuracy even in highly unsteady conditions,
i.e. within 1% error in particle velocity (Lee & Loth, 2007). Thus, the global particle time-step
may be based on:
( ) ( )
1
min p p p
8
t min CFL min =
7.147
If this time-step is much greater than the continuous-phase time-step, then one may update the
particle trajectories less frequently by sub-cycling the fluid time-step for the continuous-phase
Eulerian field before applying a Lagrangian integration (Kersey et al. 2008). However, this
benefit is not significant if the continuous-phase flow solution is more computationally
intensive than that for the particle trajectories or if there are strong two-way coupling aspects.
369
If the particle time-step time is constrained to be much less than the fluid-time-scale due to
small particle response times, sub-cycling of the particle trajectory can be performed between
two known fluid time-steps. This may be accomplished by assuming a linear (or higher-order)
variation of the continuous-phase fluid properties between time-steps. However, this too is
somewhat uncommon, since such conditions are consistent with small particle Stokes numbers
for which the mixed-fluid or weakly-separated techniques in Chapter 5 may instead be
appropriate. On exception is when the particle time-step is constrained by collision times
(four-way coupling) which requires special time integration schemes as will be discussed in
7.2.9.


Local Particle Time-steps

If the surrounding flow is steady and only final trajectory positions are needed, then the
particles trajectories can be independently integrated with their own time-step. One may use
Eq. 7.147 as the lower bound time-step for a particle j

1
p, j p, j
8
t =
7.148
As such, the difference between Eq. 7.147 and 7.148 is analogous to the difference between
global time-stepping (Eq. B.48) and local time-stepping (B.47c). However, larger time-steps
may be possible for portions of a particle trajectory where the dynamics are weak (e.g., at
terminal velocity with F
FA
=F
H
=0). As such, one may establish a target time-step based on the
quasi-steady function of Eq. 7.135a since
QS
t v F . A local advection time-step can be
defined using an advection coefficient (c
adv
) as:

adv adv RHS
t c / v F
7.149
This coefficient can then be set to limit the change in particle velocity. As shown in Fig. 7.16,
c
adv
=0.01 is approximately consistent with a particle velocity error of about 1%. It is also
desirable to have a maximum (upper bound) for the particle time-step consistent with the
movement limited by discretization of the continuous-phase flow. This accounts for any
spatial gradients in fluid properties such as shock waves or boundary layer regions. Using the
host cell computational length-scale (Eqs. B.14a, B.14b and B.14c), a maximum time-step for a
fixed Eulerian grid can then be defined for the velocity at t
n
as:

n n
max
t c / v / v

=
7.150
This essentially limits the particle crossing to a single computational cell per time-step.
Typical values for c

are unity velocity (Lee & Loth, 2007; Reichardt & Sommerfeld, 2007).
The minimum, advection and maximum time-steps can then be combined as
( )
p, j min, j adv, j max, j
t min max t , t , t

=


7.151
This time-step is fully adaptive, i.e. it can be used for each particle for each integration cycle
and can lead to order of magnitude less time-steps as compared to only using Eq. 7.147. This
approach is similar to that of Reichardt & Sommerfeld (2007).

7.2.4. Representing Polydisperse Distributions

370
In many multiphase flows, the particle characteristics are represented by a probability
distribution function, e.g., size distribution discussed in 2.1. If the distribution is to be
computationally rendered with Lagrangian technique, then two techniques are possible. The
first, and most rigorous, approach is to associate each parcel with a Probability Distribution
Function (PDF) which defines the range of attributes. This technique is sometimes referred to
as the Lagrangian PDF approach and requires integration of the PDF transport equation
directly. This technique can be computationally efficient but is also complex and can suffer
from parcel cloud volume deformation and growth due to the variety of particle trajectories
associated with different particle inertias (Fox, 2003). The second, and more common,
approach is to employ many parcels which are individually monodisperse, i.e. the diameter is
assumed constant within a given parcel. This approach allows the above methodologies for
advection to be used, and so is typically favored due to its programming simplicity. The
diameter for each parcel in this approach can be specified at particle injection regions with
either a deterministic bin distribution or a stochastic statistically-sampled distribution. The
latter is also useful for initializing individual polydisperse particles.
For a deterministic bin approach, the set of bins may be separated by equal increments of
particle diameter. Thus, each bin my have a mid-point bin diameter (d
i
) based on the total
number of bins (N
d
) and the bin width (d):


( )
( )
1
i min
2
max min d
d d i d
d d d / N

= +
=
for diameter-based bin width parcels
7.152a

7.152b
To ensure that the initial size distribution incorporates this resolution, N
d
parcels can be co-
located in each a cell with each of the parcel volumes set equal to the cell volume. As such,
several parcels occupy the same volume and the same centroid (x
P
). If four parcels of each bin
size are initialized in a cell to ensure resolution of two-way coupling and particle size
distribution, the total number of parcels per cell then becomes N
P
=4N
d
. However, this
requirement can often be relaxed if two-way coupling is not significant or at least not strongly
dependent on polydispersion effects.
To initialize the particles per parcel for a particular bin diameter (d
i
), it is helpful to first
compute the total number of particles in a given cell (N
p
) based on the average particle
volume, to secondly compute the number of these particles which correspond to the bin
diameter (N
p,i
) based on the number-based probability distribution (Eq. 2.1a) whose discrete
sum should be enforced as unity (Eq. 2.1b), and to thirdly compute the number of particles in
each parcel for the bin diameter (N
pP,i
):


( )
( ) ( )
( ) ( )
3
p p 30
init init init
p ,i p N,i
init init init
pP,i p ,i d P
init init
N / 6 / d
N N d
N N N / N



= =

=

=
P
7.153a

7.153b

7.153c
The initial locations of these groups of parcels can be spaced uniformly in a manner similar to
that for monodisperse parcels (Eq. 7.109):
( ) ( ) ( )
1/3
1/3
P P P P d P
init
init
x y z N / N

= = = = 7.154
371
One may similarly construct injection locations based on Eq. 7.110.
Computational bins may be also be constructed so that each has the same mass of particles,
which may be more computationally efficient if many bins are needed. This is equivalent to
separating by equal increments of the volume-based cumulative distribution function (

C of Eq.
2.4b). If there are N
C
bins, the mass of particles per bin normalized by the mass of particles
for all bins is equal to 1/N
C
. Since the distribution function is generally specified as a function
of particle diameter, C(d), it must be inverted to obtain diameter as function of the distribution
function, i.e. d(C), as shown in Fig. 7.17. If the distribution function is analytical, this
inversion can be straightforward. For example, the bin diameters for a Rosin-Rammler
distribution (Eq. 2.23) can be obtained as:

( )
( )
RR
1
,i
2
h
i RR ,i
i / N
d d ln 1

=

=

C
C
C

for volume-based bin width parcels

7.155a

7.155b
The initial number of particles per parcel for a given bin diameter can be set by Eq. 7.153 but
with N
C
(instead of N
d
) and by using P
N
from Eq. 2.25 and d
30
from Eq. 2.27. Another option
is to construct the bins so that each has the same number of particles:
( )
1
N,i
2
i / N

=
C
C

for number-based bin width parcels

7.156
For a log-normal distribution, the bin diameter can be obtained in a manner similar to Eq.
7.155b but with Eq. 2.30. In this case, the number of particles per parcel is uniform for all bins
and simply equal to N
p
/N
P
. In either case, the parcels for different bin diameters can again
be superposed with the same centroid and the parcel groups spaced according to Eq. 7.154.
For a large value of N
C
, the size distribution is well resolved locally but this unfortunately can
result in a large number of parcels for the domain.
One option to avoid a large number of parcels per cell is to use a statistically-sampled
approach whereby parcels of different diameters are no longer superposed but instead are
initialized randomly in space (and in time if injected). The sample diameters can be obtained
stochastically by using C
N
and a series of random numbers (Rn) which are uncorrelated and
have an equal probability of occurrence with a range between 0 and 1 (Fig. 7.17b):

N,i i
Rn = C

for number-based random-diameter parcels

7.157
This approach again allows the number of particles per parcel to be constant and equal to
N
p
/N
P
. However, the number of parcels per cell may now be a modest number similar to that
used for monodisperse parcels (e.g. N
P
=4) and the initialization or injection of these parcels
can be based on Eq. 7.109 or Eq. 7.110. The effects of polydispersion can be obtained in an
average-sense so long as the predicted properties are based on a statistically large number of
parcel samples, e.g., time-averaged particle concentrations or fluxes. For a steady uncoupled
continuous-phase flow, the parcels may be released simultaneously or serially (as is
computationally convenient) until statistical convergence is achieved.

7.2.5. Lagrangian Models for Turbulent Diffusion

372
To directly capture all effects of turbulence on particle movement (e.g., mean diffusion,
non-linear drag, preferential bias, clustering bias, etc.), the continuous-phase field seen by the
particle must be fully resolved and integrated in time simultaneously with the particle
trajectories. However, the computational requirement for full resolution of the turbulence are
often impractical (or even impossible) such that some time-averaging (RANS) or spatial
averaging (LES, DES, etc.) is used for the continuous-fluid flow solution. For RANS and box-
filtered LES, the continuous-fluid velocity is decomposed into resolved and unresolved
components as


@p,i i p i p
u (t) = u ( , t) +u ( , t) x x

7.158
The first term on the RHS represents the mean velocity for RANS and the resolved velocity for
LES, while the second term represents the instantaneous velocity perturbation for RANS or the
unresolved perturbation for LES. The former term is given directly from the continuous-phase
flow solution, whereas the latter term is only known through some mean statistical properties,
e.g., k and for a RANS simulation (A.5.6). To move the particle from t to t+t with an
explicit solver (such as Eq. 7.129), the value of u
@p,i
(t) is used and thus the terms on the RHS
of Eq. 7.158 are needed. The mean velocity (
i
u ) can be simply interpolated from the particle
position from the continuous-phase RANS solution by the techniques discussed in 7.2.2. The
key challenge is to represent the velocity perturbation (
i
u ) along a Lagrangian particle path
based on known continuous-phase statistics (k, , etc.).
The primary numerical method to represent the unresolved velocity fluctuations is to model
them stochastically in each component direction. From this, the instantaneous continuous-fluid
velocity can be modeled via Eq. 7.158 and used to compute the particle trajectories with the
techniques discussed in 7.2.3. If a sufficiently long period and sufficiently large number of
paths are employed, this should provide overall convergence for mean diffusion and will
further allow natural incorporation of non-linear drag bias since the particle relative velocities
are treated dynamically. However, this approach includes many of the assumptions and
limitations associated with the Eulerian particle diffusion models of 7.1.8.
The simplest and oldest of the stochastic Lagrangian approaches is the Discontinuous
Random Walk (DRW) for HIST (Homogenous Isotropic Stationary Turbulence), whereby a
single velocity perturbation is used throughout a particle-eddy interaction within a RANS
framework. Additional flexibility and accuracy may be obtained with the Continuous Random
Walk (CRW) approach, particularly with respect to NAsT (Non-homogenous Anisotropic
Turbulence) and LES. The following introduce DRW and CRW approaches for RANS-based
HIST flows, followed by considerations for NAsT flows and LES flows.


DRW Models for HIST

The random walk techniques for a RANS-modeled HIST flow assume that the unresolved
spectrum of turbulent eddies is characterized by an eddy strength (
rms
u ). These parameters can
be combined together with a series of Gaussian random numbers in each direction (
i
) to
simulate velocity perturbations along the particle path (Eq. 4.71):

i p i p i,rms p i
u ( , t) u ( ) u ( ) (t) = + x x x
7.159
373
The key to this technique is to determine how often the stochastic sampling takes place, as this
should be related to the time-scale of interaction between the particle and the underlying
turbulence, consistent with the Eulerian diffusion models of 7.17.
The first Lagrangian stochastic technique to obtain
i
u was proposed by Gosman &
Ioannides (1981) and is known as the Discontinuous Random Walk (DRW) approach. The
approach centers on particle-eddy interaction periods which are a combination of the integral
time-scale,

, and the eddy crossing time-scale which is the ratio of the integral length-scale
and the relative velocity, /w (Figs. 3.18 and 3.19). This approach is based on ample
experimental evidence that the large-scale structures primarily control turbulent diffusion for
both heavy and buoyant particles (3.3). The DRW approach is typically based on: a) an
interaction time based on the minimum of the two important time-scales, b) an integral time-
scale based on the kinetic energy and dissipation, c) directional variations in length-scale based
on the anisotropic continuity effect of Eq. 7.77 & 7.78, and d) an transverse integral length-
scale based on the turbulent mixing length and a length-scale constant (Eqs. A.128 & A.129):


( )
( )
@p,i i
i i
3/ 4 3/ 2
min , / w
c k /
1 w / w
c c k /


=
=
= +
=

7.160a

7.160b

7.160c

7.160d
Typical DRW values for c

and c

are given in Table 7.1 but these are only representative due
to the empiricism and uncertainty associated with turbulence modeling. This employs the
instantaneous relative velocity, but one may also employ w
term
for gravity-dominated flows
with long-times (Graham, 1998). Another approach for the eddy-crossing effect modeled by
Eq. 7.160c is to use a separate auto-correlation function in the transverse direction to include
the negative loop of the autocorrelation shown in Fig. A.14b (Launay et al. 1998). For all
these approaches, the turbulence strength is used to create a stochastic value of
i
u along the
particle path.
In the DRW method, a set of Gaussian random numbers (
i
) are generated for each
continuous fluid velocity component and have mean of zero with a variance of unity, i.e.


i
rms,i
= 0
1

=

7.161a
7.161b
Gaussian random numbers are often obtained using uniform-probability random numbers (Rn).
This second type of random number has an equal probability between 0 and 1 (and thus an
average value of ) and is commonly available in many computational libraries. A popular
method to obtain values from Rn values is to use the Box-Muller (1958) transformation.
This converts a pair of uniform probability random numbers into a pair of Gaussian random
numbers:


( ) ( )
( ) ( )
1 1 2
2 1 2
2ln Rn cos 2 Rn
2ln Rn sin 2 Rn
=
=

7.162
Gaussian numbers can be generated efficiently for most processors by using the polar version
of the Box-Muller transformation, which uses the equivalent algebraic equations. Given a
374
Gaussian random number, a velocity perturbation can then be obtained using Eq. 4.71 & 4.72
based on the local turbulent kinetic energy.
The frequency that these random velocities are changed is based on the particle-eddy
interaction time-scale. In particular, once a velocity fluctuation in a direction is obtained
through this process the time-lapse through the interaction is initialized (
clock,i
=0). The
velocity perturbation is held fixed until the eddy-interaction time is exceeded (
clock,i
>
@p,i
),
after which a new Gaussian random number is obtained and the process is repeated, as shown
in Fig. 4.15. The algorithm for each direction can be summarized as follows:

set
clock,i
=0
sample a random number value for each direction considered:
i
(t)
specify the velocity fluctuation in each direction as u
rms

i

integrate x
pi
and v
i
based on t and increment
clock,i
by t
if
clock,i
<
@p,i
, repeat previous two steps
if
clock,i

@p,i
, go to first step
7.163
If there is a crossing trajectory effects, the interaction time-scales and therefore clock times
will be different for each velocity component indicating that all random number may be
sampled at different times. Generally 5-10 time-steps are enough to integrate the particle
equation of motion during the eddy-interaction time, i.e. small enough time-steps to ensure
convergence. An example prediction with this model is shown in Fig. 7.19 for particles
dispersing in a grid-generated wake. Since this flow is approximately homogenous and
isotropic, the DRW model gives quite reasonable results, and similar to the theoretical result
(Eq. 3.79 and Fig. 3.21).
A particular advantage of the stochastic Lagrangian method for mean turbulent PDEs is
that the Eulerian diffusion models of 7.1.8 are no longer needed. Furthermore, the complex
and empirical Eulerian source term models for two-way coupling on mean flow (Eq. 7.66) can
be replaced by time-averaging the unsteady interphase forces within an Eulerian cell.
Similarly, the turbulence coupling equations with Eulerian source term models of Eqs. 7.95
and 7.97 can be replaced by Eqs. 3.168 and 3.173. Methods for such cell averaging will be
discussed in 7.2.7
There have been several proposals to advance this basic DRW model. More sophisticated
random number distributions can be prescribed to satisfy third-order and fourth-order moments
of the turbulence (kurtosis and flatness) but these are rarely employed as the RANS models
generally do not provide such statistical information. The eddy lifetime can also be
stochastically sampled using the value of Eq. 7.75 as the mean to help mimic the variations of
time-scales expected turbulence as shown by Graham (1998) and Bocksell & Loth (1999), but
the improvements in performance are not substantial. To allow a more formal prescription of
the velocity-correlation between the continuous-phase path and the particle path, one may also
employ an exponential-cosine decorrelation parameter (noted in Fig. A.12) as well as apply
correlations with the values at previous time-steps (Berelemont et al. 1990).
Due to its simplicity and computational efficiency, DRW model has been widely used (in
various forms) for free-shear flows and has shown good performance when the distributions of
k and are well predicted by the RANS solution (e.g., Faeth, 1987; Amsden et al. 1989; and
DeAngelis et al. 1997). However, if the RANS predictions of k and are inaccurate (which is
often the case for complex geometries and/or flow separation because of the empiricism of
turbulence models), then random-walk models will show corresponding deficits in
375
performance (e.g., under-predicted turbulence levels will lead to under-prediction of particle
diffusion). In addition, the ratio c

/c

is assumed to be constant in the random-walk method.
However, this ratio is proportional to the turbulence structure parameter that has been found to
physically vary for non-equilibrium turbulent flows (Stock, 1996; Graham, 1998) and within
boundary layers (Kallio & Reeks, 1989; Ushijima & Perkins, 1999). While some empirical
adjustments can be made to address these limitations, this shows that the robustness of the
particle mean diffusion model depends on the robustness of the continuous-phase turbulence
models.
There are two additional shortcomings of the DRW method that can be resolved by using
the CRW method. Firstly, changes in the local turbulence during the eddy-interaction time are
not well accommodated. This neglect can lead to substantial inaccuracies for non-
homogeneous flows (MacInnes & Bracco, 1992). Secondly, the step-function type
perturbations employed by the DRW model yield continuous-fluid accelerations which are
unsuitable (either zero or infinite along the particle path). This is problematic if stress-gradient
and/or history forces are to be included (which can be important for buoyant particles) since
these forces depend on continuous-fluid accelerations.


DRW Models for HIST

The CRW model allows for a continuous variation of the simulated velocity fluctuations by
correlating the turbulence statistics in time based on a Langevin ODE


i 1 i 2 i
u / t c u (t) c (t) = + d d
7.164
The two coefficients c
1
and c
2
can be related to the time-scale and intensity of the fluctuations.
The discrete version of this ODE is a Markov chain which is used to obtain the instantaneous
velocity for HIST (Legg & Raupach, 1982) based on that at the previous time-step, i.e. u (t-t),
the velocity fluctuations (
RMS
u ), and a decorrelation parameter (
j
):


2
i p ij i p j p ij rms p i j p
j p @p, j p
u ( , t) u ( , t t) ( , t) u ( ) (t) 1- ( , t)
( , t) exp t / ( , t)

= +

=

x x x x x
x x

7.165a

7.165b
This perturbation can be added to the mean velocity as shown in Eq. 7.158. In the first
equation,
ij
is the Kronecker delta of Eq. A.10 (=1 for i=j, else =0) and the rms of the velocity
stems from Eq. A.74b. The decorrelation parameter of the second equation employs an
exponential decay consistent with Eq. A.80a and Fig. A.12. It also employs an eddy-
interaction time-scale consistent with Eq. 7.160a, except with different values for c

and c

are
used (Table 7.1). The difference in constants stems from the fact that the DRW technique
effectively includes a linear (instead of exponential) decorrelation.
For trajectory accuracy, the time-step should be small compared to the eddy-interaction
time and statistical independence for mean particle diffusion is generally achieved (within 1%)
for t<
@p
/8 (Bocksell & Loth, 2001). This temporal resolution is consistent with an energy
spectrum which decreases by about one order of magnitude from the energy levels contained in
the largest scales. An example of the resulting perturbation time history of a CRW model is
shown in Fig. 7.18 where it can be seen that the CRW approach produces a range of fluctuation
376
frequencies more consistent with actual turbulent signals (Fig. A.9). The expressions given in
Eq. 7.165 also ensures that the most recent values of the turbulence intensity and eddy-
interaction time are used at each new time-step and that the large (non-physical) step-function
changes found in the DRW model have been reduced to smaller changes between time-steps.
For heavy particles in HIST flow, the CRW model yields similar but somewhat better
performance (Fig. 7.19). Thus, both random walk models can capture all four of the primary
HIST diffusion aspects summarized in 3.5.3: appropriate long-time limit behavior, an inertia
effect at short and intermediate times, the crossing trajectory effect, and the continuity effect.
For stratified or compressible flows with a Favre-averaged continuous-phase solution, the
value produced by the Markov chain can be set as
i
u . Thus, Eq. 7.165a can be replaced by
i i i
u u u = + (based on Eq. 5.60) to obtain the instantaneous velocity.
As previously mentioned, a potential advantage of the CRW model (over the DRW model)
is the finite acceleration of the fluid fluctuations along the particle path. These fluid
accelerations (which can be estimated by the discrete changes along the particle path) may be
used to compute the fluctuating components of the fluid-stress effect or the Basset history force.
Similarly, vorticity perturbations from the time-averaged vorticity vector ( ) may be modeled
to allow incorporation of lift fluctuations which may lead to additional particle diffusion. For
example, a Markov-chain model for each component of the vorticity can be specified as


i,@p p i p i p
( , t) = ( , t) + ( , t) x x x

2
i p ij i p j p ij rms p i j p
( , t) ( , t t) ( , t) ( ) (t) 1- ( , t)

= + x x x x x
j p p, j p
( , t) exp t / ( , t)



x x
7.166a

7.166b

7.166c
For the vorticity fluctuations, one may employ '
rms
= (/
f
)
1/2
by assuming isotropic
dissipation and where the
i
values are distinct (uncorrelated) from those used for velocity.
However, there are unresolved issues associated with incorporation of vorticity, e.g., what is
the appropriate decorrelation time for vorticity fluctuations since they are generally different
from velocity fluctuations for fluid tracer paths (Loth & Stedl, 1999) but not well studied along
particle paths. Similar issues can arise for the fluid acceleration (needed for Basset history and
fluid-stress forces) which may require some modification of the Markov chain. Fortunately,
fluid stress, history force and lift fluctuations are not generally needed since these effects do
not play a strong role in long-time diffusion (Fig. 3.20).


CRW Models for Non-Homogenous Isotropic Turbulence

Non-linear spatial variations in the mean velocity in a turbulent flow-field generally give
rise to non-homogenous turbulence. The magnitude of the turbulence gradients can be
approximately assessed by comparing with decorrelation length, i.e.


k
k
c 1
k


< for nearly-homogenous turbulence 7.167
If the turbulence gradients are weak (e.g., a grid-generated wake flow), their effect on the
CRW (and DRW) approach does not require modification. However, many free-shear flows
377
(such as jets and shear layers) and especially wall-bounded flows do not satisfy the above
inequality. As such, special treatment is required for the CRW (or DRW) approach to take into
account the effects of turbulence gradients. In particular, a drift correction must be added to
the above formulations to when Eulerian statistical properties, e.g., k(x), are used to
characterize fluctuations along a Lagrangian path.
To see why a drift correction is needed, consider the change in the continuous-fluid
velocity based on the definition of the fluid Lagrangian derivative


( )
i i i i i i
i j j j j
j j j j
u u u u u u
u u u u u
t t t x x x x

= + + + + +

D
D

7.168
If we consider, a steady flow such that
i
u is independent of time (e.g., RANS solution), then
the change in the Lagrangian change in the mean velocity is only a function of the mean
velocity gradients


( )
i
i j
j
u
u u
t x

D
D

7.169
This indicates that sampling the mean velocity at each new time when applying Eq. 7.164 is
satisfactory for mean gradients in velocity. However, the Lagrangian change of the velocity
fluctuations is given by:


( )
i i i i
i j j j
j j j
u u u u
u u u u
t t x x x

= + + +

D
D

7.170
Taking the time-average of this equation and noting
i
u =0 yields the net change along the path,
which can be used to define the net drift correction based on central-difference scheme for the
time derivative


( ) ( )
i i i
i i i j
j
u (t) u (t t) u
u u u u
t t t x

= =

d D
d D
for St

0
7.171
This drift correction represents the mean change in the velocity perturbations along a fluid
Lagrangian reference frame for non-homogenous turbulence. For a fluid-tracer in HIST flow
(Eq. A.74), this drift correction can be incorporated within the Markov chain of Eq. 7.165b as:


[ ]
2
i ij i j ij rms i j i
j i 1
i
i j ij
2
j j j
u (t) u (t t) (t) u (t) (t) 1- (t) u (t)
u u
u 1 k
u u St 0
x x 3 x

= + +

=

for

7.172a

7.172b
The drift correction term is thus
i
u which accounts for the mean change in the fluctuation
quantities along the Lagrangian, and is similar to that developed by Legg & Raupach (1982)
The drift correction is equivalent to a force in the direction of the turbulence gradient and is
consistent with the pressure gradient which arises due to inhomogeneous turbulence as noted
by Legg & Raupach and as can be derived from Eq. A.119.
To show the importance of this drift correction, a continuous-phase flow was prescribed
with uniform flow in the x-direction but a transverse peak in kinetic energy at y/h=0.5 that
tapers to lower values at y=0 and y=h. This kinetic energy distribution yielded a significant
378
transverse gradient with a c
k
, on the order of 0.3. Computational fluid-tracer particles (St

=0)
were injected uniformly upstream of this flow and allowed to convect downstream in the x-
direction. Since the flowfield is incompressible, the tracer concentration distribution should
remain uniform at all downstream locations to satisfy continuity. This exact solution is shown
in Fig. 7.20a as a solid line and can be compared with the tracers advected using the CRW
approach with and without the drift correction. It can be seen that this correction is needed to
avoid nonphysical drift of the fluid-tracers towards regions of low-turbulence (channel edges).
Similar behavior was found for a turbulent jet flow with particle as shown in Fig. 7.20b, i.e.,
lack of a drift correction caused fluid-tracer particles to be non-physically over-populated in
the edges of the jet where the turbulence was low causing the centerline to be artificially
depleted. However, the drift corrected over compensated near the jet exit where the particle
Stokes number was highest. Therefore, the fluid-tracer drift-correction of Eq. 7.172b is not
robust when particle inertia is significant.
To incorporate finite-inertia effects into the drift correction, the particle Lagrangian
derivative should instead be used for Eq. 7.168, i.e.


( )
i i i i i i
i j j j j
j j j j
u u u u u u
u v v v v
t t t x x x x

= + + + + +

d
d

7.173
This leads to a mean drift correction given by


( )
i
i j
j
u
u v
t x

=

d
d

7.174
Bocksell & Loth (2006) showed that for heavy particles with Stokesian drag and long-times
(t

), this drift can be related to the fluid-tracer drift in terms of the particle Stokes number:


( )
i
i i j
j p j
u t t k
u t u u
t 1 St x 3 x




=



+ +


d
d

7.175
The RHS expression assumes isotropic turbulence as was assumed for the RHS of Eq. 7.172b.
This result indicates that the drift correction is unaltered for small particles (since they closely
follow the fluid streamlines) but should be reduced as the particle inertia increases. For St

1,
the drift correction tends to zero since such particles are no longer correlated with the fluid
turbulence, i.e. v u 0 .
The above drift correction is appropriate for low to moderate turbulence gradients.
However, some flows have very high gradients. For example the boundary layer turbulence of
Fig. A.10 has large turbulence gradients near the wall, which correspond to c
k
(defined by Eq.
7.167) as large as 3. In this case, the drift correction of Eq. 7.172b is insufficient based on a
fluid-tracer test as shown in Fig. 7.21. In particular, numerically computed fluid-tracers (which
should faithfully follow the flow) in the boundary layer were found to collide with the wall.
This non-physical behavior is associated with the rapid changes in turbulence levels along the
Lagrangian path for each time-step, such Eq. 7.175 is not capable of avoiding this problem.
For example, a fluid-tracer located at y
+
10 will be exposed to turbulent velocities on the order
of

y so that a randomly generated a wall-ward velocity may cause a unphysical wall collision
in one time-step using a standard DRW or CRW model, even when the drift correction is
379
employed (Fig. 7.21). To avoid this problem, Iliopoulos & Hanratty (1999) recommend a
modified Langevin equation based on the normalized continuous-fluid velocity fluctuations:


( )
i i,rms 1 i i,rms 2 i
u / u t c u (t) / u (t) c (t) = +

d / d
7.176
This approach more directly accounts for changes in the turbulence intensities along the
particle path. The corresponding Markov chain for isotropic turbulence to describe each
component of instantaneous fluctuating velocity (
i
u ) at the new particle location compared to
that at the previous particle location is then expressed in terms of the rms values (
rms
u ) at both
locations. This results in a normalized version of the random walk model given by Eq. 7.172a
and the drift correction given by Eq. 7.175 as:


ij i 2 i i
j ij i j j
rms rms j rms
u (t t)
u (t) u t
(t) (t) 1- (t) u
u (t) u (t t) 1 St x u


= + +

+

7.177
This isotropic approach can be referred to as the Normalized Continuous Random Walk
(NCRW) approach. The NCRW model is compared to the standard CRW method in Fig. 7.21
for tracer particles, where it can be seen that the former technique substantially improves
accuracy near the wall where the turbulence gradients are largest.


CRW Models for Anisotropic Turbulence

In the anisotropic case, the streamwise turbulent fluctuations are not necessarily equal to
the transverse or spanwise fluctuations and the off-diagonal terms can be non-zero, e.g.


2
1 1 2 2 3 3
3
1 2 2 3 1 3
u u u u u u k
u u u u u u 0



7.178a

7.178b
Since the stress tensor is symmetric, there include six independent correlations. If the
turbulence modeling for the continuous-phase is based on a Reynolds-stress formulation
(A.5.6) or some other anisotropic RANS formulation, then the CRW model can be
constructed to take advantage of this additional information to improve its fidelity.
The simplest anisotropic model employs only diagonal terms of the stress tensor and
neglects the off-diagonal terms, i.e.
i j
u u 0 for ij. This can be referred to as the diagonal
stress-tensor approach and the resulting CRW model can be written using a anisotropic version
of Eq. 7.177:


[ ]
i,rms n 2
i ij i j ij rms i j i
i,rms
u (t)
u (t) u (t) u (t) (t) 1- (t) u
u (t t)

= + +




NSI
7.179
The decorrelation parameter (
j
) of Eq. 7.165b and the eddy-interaction time of Eq. 7.160 may
still be used since integral time scales are generally isotropic even when the velocity
fluctuations are significantly anisotropic.
If all six components of the stress tensor are known, the new continuous-phase velocity
fluctuation in a given direction can be statistically correlated with the previous fluctuations for
all three directions. This is sometimes can be referred to as the full stress-tensor approach.
380
The set of CRW equations needed to describe this approach are quite lengthy (Bocksell & Loth,
2006) but essentially relate
i
u (t) to
j
u (t) and assuming finite values of
i j
u u
.
Prediction for
the full and diagonal stress-tensor approaches are shown in Fig. 7.22 for near-wall particle
injection in a turbulent boundary layer and compared to the DNS predictions where the
turbulence is fully resolved. The exact DNS results show that particles with very small
inertia diffuse rapidly away from the near-wall injection region like fluid-tracers, whereas the
high inertia particles tend to remain close to the wall. For the low inertia particles, the
isotropic CRW predictions underestimate near-wall concentration and overestimate the
diffusion away from the wall since the actual lateral fluctuations are weaker due to anisotropy
(Fig. A.10). Including this effect with the diagonal or full stress-tensor CRW approach
yields substantially improved predictions. Note that the improvement of the full stress
approach is minor so that its complexity is not critical and can generally be avoided. It is also
shown the inertial drift correction (Eq. 7.175) considerably outperforms the tracer drift
corrections (Eq. 7.172b) for large particle Stokes numbers (Fig. 7.22b).
As noted in Figs. 7.19-7.22, the CRW approach is capable of simulating a wide range of
conditions, especially if used in a normalized form with a diagonal or full stress-tensor and
an inertia-based drift correction. Furthermore, turbulent bias associated with non-linear drag,
concentration-gradients and turbo-phoresis need not be modeled as they are already directly
incorporated into stochastic Lagrangian methods. However, the lack of turbulent structure
precludes capture of the preferential bias effect. This effect may be captured by adding the
modeled bias of Eq. 7.84 to the RHS of Eq. 7.158. Additionally and more importantly, it
should be noted that results in Figs. 7.19-7.22 are based on exact continuous-fluid time-
averaged turbulence from experiments or DNS. As such application of a random-walk model
with a RANS flowfield introduces modeling errors in k, , etc. which degrades the predictions.
This modeling problem can be avoided to some degree if an LES or hybrid RANS/LES
approach is employed, as discussed below.


CRW with LES or Hybrid RANS/LES Flows

The random walk technique can be extended to partially-resolved continuous-fluid
approaches such as the LES or DES techniques discussed in A.5.7. In this case, the resolved
unsteady continuous-phase velocity, ( , t) u x , represents the deterministic component to which a
stochastic sub-grid fluctuation , ( , t) u x , is added. The strength of the stochastic velocity
fluctuations is made to be consistent with the unresolved/sub-grid turbulent kinetic energy,
k ( , t)

x . The advantage of this approach, as compared to a RANS-CRW technique, is that the


bulk of the kinetic energy as well as the bulk of the anisotropic and non-homogenous
turbulence effects can be captured directly by the large-scale structures associated with u . As
a result, particle dispersion associated with these effects need not be modeled empirically. For
example, Yang & Lei (1998) found that LES with a resolved length-scale of 1.6 gave nearly
identical results in terms of preferential bias as that from DNS with resolved wavelength of
0.6. This is important because, the partially-resolved LES required an order of magnitude
less CPU time than the fully-resolved DNS.
It should be noted that sub-grid particle diffusion may be ignored completely (no stochastic
model needed) if St

1 based on Eqs. 4.64-4.66. Otherwise, sub-grid diffusion can be


381
important for particle diffusion and preferential concentration and there have been recent
models constructed for this purpose. For example, Rybalko et al. (2008) developed an
isotropic CRW model for hybrid RANS/LES flow solutions (as typified in Fig. A.20) in order
to predict particle diffusion. This hybrid CRW model reverts to a RANS CRW model in the
attached boundary layers, where the mean particle diffusion driven by stochastic fluctuations
dominates. This model also reverts to an LES sub-grid CRW model in the free-shear and
highly-separated flow regions, where the large scale structures are resolved and dominate
particle diffusion. In particular, the Nichols-Nelson hybrid model (Eqs. A.141-A.144) uses a
hybrid Markov chain similar to that for RANS flow (Eqs. 7.165 & 7.172). To accommodate
the hybrid aspects, this chain uses the hybrid turbulent kinetic energy and the resolved velocity
field:


i ihy,i ihy,i
2 n 2
ihy,i ij ihy,i j ij i hy j i
3
i p hy@p,i
u (t) u (t) u (t)
u (t) u (t t) (t) (t) k 1- (t) u
( , t) exp t /
= +
= + +

=

x

7.180a

7.180b

7.180c
The hybrid turbulent kinetic energy (k
hy
) is given by Eq. A.143 and reverts to the sub-grid
kinetic energy (k

) in the LES portions of the flow.


The decorrelation time-scale of the velocity fluctuations along the particle path can be
expressed in terms of the local instantaneous hybrid time and length scales in a manner similar
to that given in Eq. 7.160:


( )( )
hy@p,i hy hy,i
hy hy RANS
3/ 4 3/ 2
hy,i hy RANS i
min , / w
c k /
max c k c / , 1 w / w


=

=
= +

7.181a


7.181b

7.181c
The hybrid length-scale tends to the sub-grid length scale as the model approaches LES
conditions. However, the dissipation is controlled by the smallest-scale structures so is
independent of whether a RANS or LES region is considered, i.e.
hy RANS
= = . As with
the drift correction for the RANS version, a similar correction for gradients in the unresolved
kinetic energy should be included in the hybrid CRW model, e.g., Eq. 7.185 can be modified
as:


( )
hy hy hy
i hy,i
,hy i hy p i
k k
1 t t
u t u =
t 3 1 St x 3 x



=


+ +

d
d

7.182
This reverts to the drift correction of Eq. 7.175 in RANS regions with non-homogeneous
turbulence.
An example LES flow field is shown in Fig. 7.23a where particles are tracked both with
and without the CRW model. It can be seen that the addition of the stochastic fluctuations
yields some diffusion of the particles but most of the particle dispersion is governed by the
turbulent kinetic energy of the large-scale resolved structures. It is interesting to note that the
stochastic diffusion is not symmetric about the deterministic particle paths. This is due to the
gradients in the sub-grid kinetic energy which bias the particle velocity according to Eq. 7.182.
This LES flow solution was deemed reasonable in that the combination of resolved and sub-
382
grid kinetic energy approximately matched that of the fully-resolved DNS kinetic energy (Fig.
A.21). The quantitative diffusion of particles in this LES flow is shown in Fig. 7.23b and
compared to the deterministic diffusion of particles in a fully-resolved DNS flow-field. It can
be seen that the CRW model only accounts for a minority of the net diffusion, but that its
inclusion avoids over-prediction or under-prediction of particle spread in certain regions.


SDE for RANS and LES Flows

Another option for treating Lagrangian diffusion is to sample the continuous-fluid velocity
from a Stochastic Differential Equation (SDE) as proposed by Haworth & Pope (1987). This
may be based on the RANS equation of motion for a fluid tracer in differential form:


1 3
i i i i
turb turb
2 4
f i turb
u u u 1 p
c c
t x t

= + +




D DW
D D

7.183
The first term on the RHS is similar to that of Eq. A.118 since it includes the acceleration of a
fluid tracer due to mean pressure gradients. However, it neglects viscous stress effects
(assuming high Reynolds numbers) and hydrostatic effects (assuming uniform density), though
these effects can be included if needed (Shortoban et al. 2004). The second and third terms of
Eq. 7.183 include a coefficient c
turb
, which is related to the turbulent acceleration. The second
term forces the instantaneous velocity towards the mean velocity field over a turbulent time-
scale (
turb
), which is proportional to the integral time-sale of the continuous-phase turbulence
seen by the particle (
@p
). If
turb
=k/, then c
turb
=2.1 ensures that the mean Reynolds-stress for
this stochastic process is consistent with Eq. A.130 (MacInnes & Bracco, 1992). If an SDE is
used to represent the continuous-phase velocity fluctuations along the path of a particle with
finite relative velocity, an adjustment must be made to
turb
in order to take into account the
reduced eddy-interaction time due to eddy-crossing effects. This can be accomplished by
employing an eddy-crossing effect as noted by Reynolds & Lo Iacono (2004). A form for the
turbulent time-scale based on Pozorski & Minier (1998) and Eq. 7.77 assuming isotropic
turbulence is:


( )
1/ 2
2
turb,i 2
i
k 3w
1+
2k 1 w / w



=

+

7.184
This relationship can be used with Eq. 7.183 and c
turb
=2.1.
The third term includes the vector W
i
which represents the isotropic Wiener process, i.e.
has a mean of zero and variance equal to the differential time. It can be discretized with a
Gaussian random number (introduced in Eqs. 7.161a and 7.161b) for each direction


1/ 2 i
j ij
t
t


DW
D
7.185
Combining this equation with a central-difference temporal discretization for the LHS of Eq.
7.183 yields:
383


n-1 n-1 n-1
n n-1 i i
i i j ij
f i turb
p u u 1
u u t 2.075 t 2.1 t
x

= +




7.186
This type of discretization is called the Euler-Maruyamma method. For integration, the
particle time-step should be significantly smaller than the turbulent time-step to ensure
statistical convergence. It should be noted that this simple one-step integration scheme is quite
common since higher-order schemes do not necessarily preserve the proper stochastic
properties (Colucci et al. 1998).
Sample SDE results for a RANS-type turbulent wake flow are shown in Fig 7.24, which
indicates reasonable predictions. While these predictions are not as accurate as those obtained
with DRW and CRW models for this particular HIST flow (Fig. 7.19), the SDE approach does
not require drift corrections (as in Eq. 7.185) for non-homogeneous turbulence. This
robustness can be traced to its use of the pressure gradient in the fluid velocity updating
scheme which effectively incorporates the effects of gradients in turbulence (recall discussion
associated with Eq. A.119). The SDE model can be extended to LES and RANS/LES flows.
For example, Shortoban et al. (2004) used the above approach to incorporate stochastic sub-
grid fluctuations to and LES flowfield with
3/ 2
k

= (Eq. A.137). As shown in Fig. 7.25, the


LES predictions of the particle spread are improved with the inclusion of the sub-grid diffusion
provided by the SDE model. There is much active research on SDE models and further
developments are expected.


7.2.6. Lagrangian Models for Brownian Diffusion

There are several possible approaches to incorporate Brownian diffusion into a Lagrangian
point-force approach. One technique is to employ analytical expressions for velocity and
displacement of particles with correlated Gaussian functions for each time-step increment
(OBrien, 1990). However, such a theoretical approach is not applicable in turbulent flows or
for non-linear drag caused by Reynolds number, Knudsen number or history force effects
(Ounis et al. 1991; Mainardi & Pironi, 1996). The most common approach is to model this
diffusion using a stochastic Brownian force and a statistically large number of particles. For
example, consider the particle acceleration based on Eq. 3.131:
[ ]
f Kn Br p
/ t 3 d f (t) / m = + v w F d d
7.187
Since time-steps for integration will be generally small to include Brownian motion, a first-
order stable implicit scheme is often used (Ounis et al. 1991):

n+1 n n n+1 n
Br
p p
( t)
t m

= +

v v u v F

+O
7.188
This scheme shares the stability attributes of Eq. 7.124. The Brownian force per particle mass
can be modeled using a discretized Wiener process which satisfies the long-time diffusion of
Eq. 3.139:
384

p,Br n+1 n n
i i i j ij
p p
2
1 t
v v u
1 t / t

= + +

+




7.189
Predictions in Fig. 7.26 indicate good agreement with theory for conditions where Brownian
diffusion dominates.

7.2.7. Eulerian Cell-Averages of Lagrangian Particle Properties

For two-way coupling with Lagrangian particles, it is important to transfer the particle
mass, momentum and energy to Eulerian cell-averaged values consistent with the discretization
of the Eulerian continuous-phase. This first requires computation of the interphase coupling
terms at each particle location. Recall that the techniques in 7.2.2 can be used to interpolate
the Eulerian continuous-phase properties to the particle location (e.g., u
@p
and T
f@p
). These are
then combined with the Lagrangian particle properties (e.g., v and T
p
) to obtain the interphase
coupling term for an individual particle along its path. For example, Lagrangian values of
p
m`
can be obtained from Eq. 1.87,
p
`
Q from Eq. 1.91, and
int,i
F from Eq. 6.245. In order for these
values to be applied to the coupled continuous-phase equations (such as Eqs. 7.1b, 7.2b and
7.3b), they must be cell-averaged on the Eulerian grid. Furthermore, these coupled equations
also require a cell-averaged particle volumetric (or mass) concentration such as . Even for
one-way coupling conditions, it is often important to predict the Eulerian distribution patterns
or flux rates from the Lagrangian trajectories. Methods to obtain the Eulerian cell-averaged
properties from the individual particle properties are discussed in the following.


Instantaneous Properties

The most common approach for cell-averages is use a simple-summation of the particles
within the control volume (
,i
) that surrounds an Eulerian node (x
i
) as shown in Fig. 7.27.
Defining N
p,j
as the number of particles within this volume, the averaged particle
concentration is given by Eq. 4.6, i.e.:

p ,i
N
i p, j
i
j 1
,i
1

= =


7.190
Similarly, the cell-averaged mass, momentum and heat transfer per volume of mixed-fluid are:
385

p ,i
p ,i
p ,i
N
p
pp, j
j 1
p ,i
i
N
p
p, j
j 1
p ,i
i
N
int
int, j
j 1
p ,i
i
m
1
m
1
1


=





=




F
F
`
`
`
`
Q
Q
7.191a


7.191b


7.191c
For a parcel treatment, the cell-averaged properties are similarly based on the number of
parcels within a cell (N
P
) as well as the number of particles within a parcel (N
pP
). The cell-
averaged value for the particle volume concentration using parcels is thus given by Eq. 4.8:
( ) ( )
p ,i p ,i
N N
i p, j pP, j P, j
j 1 j 1
,i ,i
1 1
N

= =

= =



7.192
Note that the RHS term uses the parcel volume instead of the particle volume. Similar parcel-
based cell-averaging can be applied to the mass, momentum and heat transfer properties of Eq.
7.191. Such summation averages are commonly used for coupled equations with a large
number of parcels per cell (Dukowicz, 1980).
For the simple-summation approach,
i
is based on the cell-volume (includes only one
parcel in Fig. 7.27). Such averaging can lead to discontinuous cell-averaged quantities as
particles or parcels enter and leave the volume. For example, the parcel P
j
in Fig. 7.27 does not
yet contribute but will yield a two-fold increase in
i
once it crosses into the volume. Similar
discontinuities in the two-way coupling terms of Eq. 7.191 can lead to numerical instabilities
when there are only a few particles or parcels in a computational volume, i.e. Eq. 7.112 is not
satisfied. This especially problematic for unstructured grids with a wide range of grid sizes
(Fig. 7.28). This problem can be lessened by employing discrete shape functions (e.g., Eq.
B.16) and obtaining a volume fraction which is node-based (instead of cell-based for the
simple-summation technique). For the shape function approach, the Eulerian-based volume
fraction and interphase force transfer at a node i for parcels becomes:


( ) ( )
p ,i
N
i D P, j i p P, j pP, j
j 1
,i
1
x , x x N


( ) ( )
p ,i
N
int
D P, j i int, j P, j pP, j
j 1
p ,i
i
1
x , x x N

F
F
7.193a



7.193b
For the shape-function approach, N
p
is based on the surrounding cells of a node, e.g., the three
parcels in the dashed-box of Fig. 7.27. However, the cell volume (
,i
) in these equations for a
2-D domain is still based on the solid-line box of Fig. 7.27. This cell-volume should be
reduced by half if the node happens to lie on a planar wall boundary since parcels are on one
side. Similar equations can be used for momentum and energy transfer of Eq. 7.191b and
7.191c.
386
The shape-function approach is equivalent to using a weighting based on proximity to a
node and its benefits compared to the simple-summation technique can be significant. For
example, consider a single particle located in between two Eulerian nodes x
1
and x
2
on a one-
dimensional grid (Fig. B.3). For the summation technique of Eq. 7.192, this particle movement
yields a volume fraction at node x
2
which is zero for x* and then discontinuously changes to
a constant for x*>. As such, the simple-summation method often requires many parcels per
cell (N
P
1) so that the discontinuity effects between cells are not large. In contrast, the same
particle movement for the shape function approach will yield linear changes in volume fraction.
For example, the volume fraction at node x
2
based on Eq. 7.193a and the linear shape function
of Eq. B.4 is given by:


p p p 1 p *
2 p,2
2 1
x x
x
x x


= = =




7.194
For parcel representations, this smoothly varying characteristic is helpful in that it only
requires satisfaction of the Eq. 4.10 criterion (equivalent to N
P
1). As such, the shape
function approach initialized with the same number of parcels per cell is more robust than the
conventional summation approach.
In some cases, parcels may be initialized such that Eq. 4.10 is satisfied but movement of
the parcels into regions with finer grid resolution can subsequently violate this criterion. This
can be caused by non-uniform distribution of the Eulerian grid size and/or unstructured grids
(Fig. B.1), but is particularly problematic for regions where adaptive grid refinement is
employed. For example, Fig. 7.28 shows a shock attenuated by a dusty gas where there are
several levels of grid refinement near the shock. For a parcel initialization with N
P
=4, this
refinement results in a low average number of parcels per cell (about in this case) near the
shock and leads to large two-way coupling errors (Fig. 7.28a). In contrast, a higher initial
parcel resolution (Fig. 7.28b) led to a more physical and accurate description of the flow. This
can be addressed by significantly increasing the number of parcels in the coarse regions
upstream. Another option is to use parcel adaptivity whenever the parcel resolution becomes
locally too small (Sivier et al. 1996). The adaptivity is accomplished by breaking up parcels
into groups of smaller parcels when there are only a few within a cell. This allows for a
reduced number of total parcels in the domain but also introduces numerical diffusion of the
particle concentration.
A much different approach to reduce the number of parcels required is to use the full
Lagrangian method introduced by Osiptsov (1988) and developed in detail by Healy & Young
(2005). In this approach, particle concentration is carried along as an additional characteristic
(along with velocity, temperature, position, etc.) along the particle trajectory. This method
becomes straightforward for weakly-separated flows as discussed in 5.3.1, but requires
additional ODEs for significant Stokes numbers because one must include the divergence of
the particle velocity field on the RHS of Eq. 5.106. For fully-separated conditions, this
divergence can be expressed in terms of a path derivative of the Jacobian tensor (J
ij
) of the
Eulerian-Lagrangian transformation. The determinant of this Jacobian gives the change in
volume of a differential parcel volume along the parcel path over a time shift of . If there is
no mass transfer, this change is inversely proportional to the change in the volume fraction
which allows a discretized change in the parcel volume fraction for a given time-step:
387


( )
( )
( )
( )
n
P P
t t o P
n 1
P P P
t o t t
= =
+
= =

= =

J
7.195
To compute the LHS determinant, all the components of the Jacobian must be determined. If
the particle surface force consists only of a linear drag force, Healy & Young (2005) noted that
these components can be obtained from a set of coupled ODEs:


ij
ij
ij
i
kj ij
p k
t
u 1
t x
=

=



`
`
`


dJ
J
d
dJ
J J
d

7.196a


7.196b
Since the Jacobian is not generally symmetric, this pair of equations yields eight ODEs in 2-D
flows and eighteen ODEs in 3-D flows. These simultaneous equations can be integrated along
the particle path using a predictor-corrector technique, from which the updated particle
concentration can be obtained through Eq. 7.195 and the simple-summation of Eq. 5.109. For
monodisperse particles, the volume fraction in a cell can then be averaged using either a cell-
based summation or a node-based shape function approach:

( )
p ,i
N
i P, j
j 1
P ,i
1
x
N


( ) ( ) ( )
p ,i p ,i
N N
i D P, j i P, j D P, j i
j 1 j 1
x , x x x , x

= =
=


7.197a


7.197b
Note that a reasonable volume fraction can be obtained with just one particle per cell for the
summation approach (Eq. 7.197a) and with just one particle for the surrounding cells for the
shape-function approach (Eq. 7.197b). Therefore, the full Lagrangian approach is often
worthwhile since the overall number of particle trajectories will be much less than that with
traditional Lagrangian concentration methods. For example, Healy & Young (2005) found that
the total number of parcels in the domain can be reduced by 20-50 fold in 2-D flows
(indicating substantial improvement in overall efficiency) and that the savings are even more
dramatic for 3-D flows.
This approach can also be extended to polydisperse particle distributions by using a bin
method. In this case, one parcel from each bin can be used at injection. The full Lagrangian
technique can then be used at any downstream location to get the Eulerian bin volume fraction
in each cell (e.g., with at least one parcel per bin per cell for Eq. 7.197a). The net volume
fraction can then be summed from the bin contributions according to Eq. 4.61. Similarly, the
Eulerian mass, momentum and heat transfer of Eq. 7.191 can be computed for each bin and
then converted to an overall Eulerian transfer for all bins by summation. If there is particle
mass transfer, the change in net particle volume should be included, i.e. Eq. 7.195 becomes:


( )
n 1/ 2
n 1 n
P P p p
/ t m /
+
+
+ ` J 7.198
The term in parentheses on the RHS is the rate of increase in a particle volume, which is
assumed to apply uniformly to all particles within the parcel. Care should be taken if the
388
particle evaporates or dissolves completely or if the particle undergoes break-up or coalescence.
A similar approach can be used for compressible particles to accommodate changes in particle
volume due to pressure or size dynamics.

7.2.8. Time-Averages of Lagrangian Particle Properties

In many cases, the time-averaged properties are of interest. If the continuous-phase flow is
uncoupled to the particle trajectories, time-averaging can be accomplished with local particle
time-stepping and flux-averaging after the continuous-phase solution is obtained. If the
continuous-phase flow is unsteady or there is two-way or three-way coupling, time-averaging
requires simultaneous evaluation of both the Lagrangian particle field and the Eulerian
continuous-phase flow field. These two approaches are outlined below.


Flux-Averaging for Flows with One-Way Coupling

Flux-based averaging employs a face of a computational cell through which particle
trajectories are expected to cross, e.g. faces which are perpendicular to the streamwise
direction of the particles. As such, this method is typically used for particles which are
injected at a steady state in one location ore region and move in approximately one direction.
For a given overall particle mass flow rate, the total number of particles (N
p
) or parcels (N
P
)
injected in the domain can be set by a physical time period of injection (
inj
). Flux averaging is
popular for continuous-phase flows which are not coupled to the particles, since this approach
conveniently avoids direct cell-averaging of particle properties. The integration and time-
averaging technique depends on whether the continuous-phase flow is steady or unsteady.
If the continuous-phase flow is steady, one may accelerate the computation of the particle
trajectories by using local particle time-steps (t
p
), e.g., based on Eq. 7.151. A fixed number
of particle or parcels can be selected and tracked until they exit the domain or disappear due to
mass transfer (e.g., evaporate or dissolve). Since their trajectories are independent, they may
be computed sequentially or in parallel depending on which process is more efficient for the
computer hardware. If the continuous-phase flow is unsteady, the particles trajectories must be
integrated simultaneously with an equal timestep (Eq. 7.146).
In either case, the flux across an individual cell can be related to a time-average by
employing the principle of Eq. A.1. For example, the average particle mass flux based on the
total number of trajectory crossings (N
p
) can be related to that based on the concentration of
particles per mixed-fluid volume () evaluated over a discrete cell face cross-sectional area
(A

) as:


( )( )
p
N
p, j p
j 1
inj A
1
m A

=
=

v n


7.199
Assuming that the particle density (
p
) is independent of velocity or concentration fluctuations
(as in 5.2.2), the mean volume fraction at an Eulerian node can be expressed as a sum over the
number of particles that cross its face (N
p,i
) if individual particles are considered or number of
parcels that cross its face (N
P,i
) if parcels are employed:
389


p ,i
N
p, j
i
j 1
p ,i inj j i
m
1
A

v n
for individual particles
P ,i
N
pP, j p, j
i
j 1
p ,i inj j i
N m
1
A

v n
for parcels
7.200
One may set
inj
=t and compute the spanwise and transverse spacings based on Eq. 7.108 for
individual particles and Eq. 7.110 for parcels. The number of trajectories (and thus timestep)
should be large enough to ensure the averaged statistics in any cell face are sufficiently
converged. The total number of trajectories required for such convergence will increase if the
cell face area decreases or if polydisperse size distributions (7.2.4) or stochastic turbulent
interactions (7.2.5) are included. The number of trajectories can be increased by using a
larger injection time or using fewer particles per parcel.
To reduce the total number of trajectories required, one may use a shape-function approach
for additional accuracy. For example, the mean volume fraction based on parcels can be
expressed in terms of the shape function based on the location where the particle crosses the
cell face surface :


P ,i P ,i
N N
D,i, j pP, j p, j
i D,i, j
j 1 j 1 j i p ,i inj
N m
1
A

= =


=






v n

7.201
In these equations, the parcel volume is assumed to have a length-scale less than or equal to the
cell flux area. If the latter criterion is not satisfied by the grid resolution, then the initial parcel
volume (Eq. 4.9) should be reduced or the data collection grid should be coarsened.
One may also obtain Favre-averaged quantities using a trajectory weighting. For example,
the particle velocity defined by Eq. 7.45a can be obtained for parcels as:


( )
P ,i P ,i
N N
i i pP, j p, j j i pP, j p, j
j 1 j 1
N m N m

= =

=



v n v n 7.202
The same mass-weighted averaging of individual particle trajectories can be used for the
interphase mass, momentum and heat transfer terms of Eq. 7.191.


Time-Averaging with Two-Way Coupling

If the continuous-phase Eulerian flow includes two-way coupling, one must integrate the
unsteady Lagrangian trajectories and the continuous-phase flow field simultaneously.
Different techniques are employed depending on whether the continuous-phase is steady or
unsteady.
If the continuous-phase is unsteady, a single global time-step is typically used for both the
continuous-phase and particle equations of motion. In this case, two-way coupling is based on
instantaneous particle and fluid properties. Thus the particle trajectories can be based on the
techniques discussed in as discussed in 7.2.3 and the continuous-phase can employ shape-
function cell-averaging for the instantaneous interphase source terms as discussed in 7.2.7.
390
Both of these fields may then be averaged starting from the initial conditions with a simple
running sum. For example, volume fraction can be averaged based on the ratio of two sums:


( ) ( )
N N
n n n
i i
n 1 n 1
t t

= =
=

7.203
In this equation, N

is the total number of time-steps for integration. Similar averages can be


used for the other particle-phase and continuous-phase properties.
If the continuous-phase flow is steady, one must converge this to a steady solution based on
a time-average of all the Lagrangian particle trajectories. A common approach for this is to
iterate between the two fields with the following steps (Yan et al. 2007):
1) compute the uncoupled Eulerian continuous-phase steady flow by integrating the
pseudo-unsteady equations with local time-stepping until convergence,
2) use the flow of 1) to compute a large number of Lagrangian particle trajectories to obtain
initial time-averaged source terms for each Eulerian cell,
3) converge the coupled continuous-phase equations with the source terms from 2)
4) re-integrate the Lagrangian particle equations and the pseudo-unsteady continuous-
phase equations simultaneously until both fields are converged.
For the fourth step, the particle equations are integrated in the flow-field defined by the most
recent continuous-phase solution for centroid interpolation. In contrast, the continuous-phase
flow equations employ interphase source terms based on a particle field which is both cell-
averaged and time-averaged, e.g., with a running sum. Assuming constant time-steps, the
time-average can also be applied to LHS of Eq. 7.191 as:


( )

( ) ( )
N
n n
p p i p,i
i i
n 1
1
m m m
N

= =

` ` `
7.204
Similar averages can be applied to obtain
p int,i
and F
`
Q . Note that the RHS involves cell-
averaged quantities which may be obtained from either simple summation or a shape-function
approach (7.2.7). Since the number or particles or parcels used for such averaging is typically
large, the more efficient simple summation is sufficient and typically employed (Sommerfeld,
1987).
For turbulent flow, the source coupling terms to the continuous-phase can be based on a
time-averaged of the particle trajectories. This allows direct use of Eqs. 3.141, 3.168, 3.173,
7.96, 7.191c, and 7.203 to construct the following source terms:

p p p
N N N
k,diss int, j j int, j j
j 1 j 1 k 1
f p
1 1
F w F w
N

= = =


=





S
,diss k,diss
/ k



S S
( )
p p
N N
prod
k,prod int, j j
j 1 j 1
f p
c
F w
N

= =

=


S
7.205a


7.205b


7.205c
As such, these formulations directly incorporate the stochastic fluctuations of the relative
velocity and interphase force.

391
7.2.9. Particle-Wall Interaction Methods

Particles collisions and fluid dynamic interactions can be directly taken into account
with Lagrangian methods. The interaction and particle properties can be used to determine
whether an individual collision will lead to a reflection, coalescence, or a break-up. In the
following, the case of wall collisions is first considered, followed by a more complex particle-
particle case, and both cases first consider numerical issues associated with individual particles
and then those associated with parcels (groups of particles)


Spherical Particle-Wall Interactions

For particle interactions with walls, a computational method should first identify when
a particle is within the vicinity of the wall. The particle-wall distance (r
w
) is illustrated in Fig.
6.70 and can be obtained numerically by examining the computational cell in which the
particle is located. If one side of the cell includes a wall surface that is planar and defined by a
boundary node (x
w
) and a unit normal away from the wall (n
w
), this can be combined with the
particle position (x
p
) to quantify its proximity:


( )
w p w w
r = x x n
7.206
If there is more than one wall in a computational cell, one should employ the minimum value.
The wall-normal and wall-tangential velocities relative to the wall may be obtained as:


( ) ( )
( )
w
p w w
p w
v
v


=

v n
v v n
v v v v
|

7.207a

7.207b

7.207c
In these equations, a positive wall-normal velocity is defined as positive in the direction of the
wall, consistent with Fig. 6.70, while the tangential velocity is the remaining component of the
relative velocity with the wall.
Fluid dynamic interactions are discussed in 6.8 and assume that the particles are close
but are not in contact. The drag effects are the most sensitive to wall proximity especially for
solid particles with Re
p
<1 and r
w
<10r
p
as shown in Fig. 6.71. In general, the Stokes wall
correction (f
w
) can be simply incorporated into the drag force so that the particle velocity and
position can be integrated from t
n
to t
n+1
as discussed in 7.2.3. Wall effects on added mass,
history force and lift force can be similarly included but tend to be significant when
*
p
1 <

,
r
w
<3r
p
and
p
Re 1 <

. As such, these effects are generally ignored for particles in gas flows and
particles not in close proximity to the wall.
For wall collision effects, it is important to determine if and when a particle may collide
with the wall between time t
n
and t
n+1
. Generally, the particle ODE is first integrated with an
explicit scheme between these two times assuming that collision does not occur. If contact
occurs, the collision time (t
coll
) can be estimated using the particle radius, wall-normal distance,
and the average wall-normal velocities:
392


n 1
w p
r r
+
for a wall-collision
( ) ( )
n n n n 1
coll w p
t t 2 r r / v v
+

= + +
7.208a

7.208b
The particles can then be reintegrated to the contact point (from t
n
to t
coll
) so that the incoming
impact velocities (and rotation rates) relative to the wall as identified in Fig. 6.74 can be
obtained. If a stick boundary condition is uniformly applied (Fig. 7.4a), then the particle flux
can be recorded and the particle removed from a list used for further integration. This is often
reasonable if particle deposition can be ensured due to rough walls or high attraction forces
(electro-static, magnetic, van der Walls, etc.). In the more general case where reflection or
slide may be possible, the kinetic energy of the particle should be considered along with the
wall surface properties and even continuous-phase fluid properties. In general, the hard-sphere
model is used (e.g.,
in
v

is based on v

at t
coll
) and the interaction is assumed to yield a step
function change in the particle normal and tangential momentum (6.10).
For a solid particle, the dry restitution and friction coefficients of Table 6.6 can be
used to respectively compute the normal restitution coefficient (with either Eq. 6.228 or 6.231)
and the tangential restitution coefficient (with Eq. 6.238). The normal restitution coefficient
can then be employed with Eq. 6.223 to find the outgoing wall-normal velocity. Similarly, the
tangential restitution coefficient can be used with Eq. 6.239a&b if there is no initial particle
angular velocity to find the outgoing wall-tangential velocity and angular velocity. The
outgoing values can then be used to integrate the particle motion from t
coll
to t
n+1
. Note that the
post-collision particle acceleration is generally assumed to be zero for this integration so that
added mass and history effects can be ignored for this particular sub-interval. However,
Gondret et al. (2002) investigated collision trajectories in liquids and showed that the incoming
particle acceleration is at least partially retained after the collision. They modified the finite
Reynolds number history force to account for the velocity reversal using an empirical
correction factor based on their experiments.
For fluid particles, surface tension effects need to be additionally considered for a
surface collision. For liquid drops impacting on a liquid surface or liquid film, one may use the
criteria of 2.4.5 to determine the type of outcome. The relative Weber number (We
p-p
of Eq.
2.92) can be computed based on the wall-normal velocity and combined with the incoming
impact angle (Fig. 6.74) to determine whether coalescence, bounce or break-up is expected.
Coalescence yields a stick type of condition whereas bounce can be modeled similar to a solid
particle (discussed above) using near unity dry restitution coefficients as noted in Table 6.6.
Bubbles interacting with gas interfaces may be treated in a similar fashion. However, the
dynamics of these interactions is not well understood so that the outcome criterion is often
simplified to either bounce or coalescence, where the latter occurs for We
p-p
<0.2 (see
discussion in 2.4.5). When interacting with a solid surface, the outcome is typically assumed
to be bouncing for a bubble. In contrast, a drop may undergo bouncing, spreading or splashing.
However, the criteria for these three regimes are complicated by effects of surface roughness,
relative surface tension of the material and fluid viscosity (Rein, 1993, Range and Feuillebois,
1998, and ikalo et al. 2005).
For the above models, it is assumed that individual particles are tracked. However, a
similar approach can be used for parcels by replacing the particle centroid and velocities with
the parcel centroid and velocities for the interactions. However, the particle proximity to the
wall is still based on the relationship of r
w
to r
p
so that the average collision time based on all
393
particle is estimated. Thus, some of the particles in the parcel may overlap the wall (Fig. 7.29).
This is only reasonable if there are a large number of trajectories used to compute a flux. This
can be verified using a parcel-resolution test, whereby N
pP
(of Eq. 4.7a) is reduced until
convergence is achieved. For parcel reflection, one may note that a spherical particle volume
is preserved in size and shape if the incoming and reflected angles are the same (Fig. 7.29), i.e.
the normal and tangential restitution coefficients are equal.


Collisions of Non-Spherical Particles or Rough Surfaces

The conditions of non-spherical solid particles or non-planar surfaces can be treated either
deterministically or stochastically. The deterministic approach requires detailed knowledge of
the particle surface in relation to the local wall angle, as well as the coefficients of restitution
and friction. Such information is generally not available, so that a stochastic approach is
instead employed. The stochastic approach commonly considers surface angle deviations
referenced to spherical particles and/or flat walls. Techniques for sampling and employing
it in a collision model are discussed in the following.
For a particle impacting a rough wall, the wall surface can be defined to have two outward
unit vectors as shown in Fig. 7.30:
w
n (which describes the average wall normal), and
w
n (which describes the local wall normal). One may also define the two average tangential
unit vectors:
t ,1
n (which is along the average surface shown in Fig. 7.30) and
t ,2
n (which is
outward from this plane and normal to
w
n and
t ,1
n ). Based on these two tangential surface
vectors and the local wall normal, one may define the two perturbation angles from
w
n as
1

and
2
, such that:


( )( ) ( )( ) ( )( )
w 1 2 w 1 2 t ,1 2 1 t ,2
cos cos sin cos sin cos = + + n n n n
7.209
Given a set of perturbation angles, the local wall normal can be used to compute the incoming
particle trajectory angles and velocities and their outgoing values. Since the perturbation
angles are generally random, they can be represented by a standard deviation (
rms
) based on
both the roughness angle distribution and the particle size (Sommerfeld & Huber, 1999).
Based on this variation, stochastic sampling of the perturbation angles may be obtained using a
Gaussian random variable ( of Eq. 7.161) as
1 rms
= .
If particle reflections are sampled over many realizations with the above perturbation angle
sampling and Eq. 7.209, there will be diffuse reflection about the mean outgoing angle,
out
(Fig. 7.30). For particles that impact the mean wall at a non-normal angle, the roughness
can also introduce a bias effect. This occurs because an incoming trajectory angle from the
above technique may yield
in
<0
o
, which is unphysical. If this occurs, this sample is not used
and new perturbation angles are sampled until a physical (positive) angle occurs. This
resulting bias has been termed the shadow effect, and represents the fact that deep
indentation regions of the wall can not be reached by very shallow trajectories. Example
probability distributions of surface perturbation angles based on this approach are shown in Fig.
7.31 for three different incoming angles. The wall normal condition yields a Gaussian
394
distribution function with zero-mean. However, as the impact angle is reduced, the shadow
effect leads to a positive bias in the mean reflected angle.
In the case of a non-spherical particle impacting a smooth wall, a diffuse reflection pattern
can also occur. To model this,
in
can be defined as the impact angle which would occur if the
particle was spherical, i.e. if the particle outward normal is taken as radially from the particle
centroid. Since a non-spherical particle can include a non-radial component in its surface
normal (Fig. 7.32), a particle surface angle perturbation can be sampled from this variation
using the same concept used for the wall angle perturbation. Sommerfeld (2002) examined
non-spherical particles with no spin and found that they were approximately equal to the
surface area ratio (
*
surf
A of Eq. 6.23):


( )
1/ 3
*
rms surf
A 7.210
However,
rms
reduces as spin rate magnitude increases for large surface area ratios, since
orientations with high r
p
values become more probable at contact.
Collisions between non-spherical particles or between non-spherical particles and a rough
wall can be modeled stochastically using stochastic sampling for both the strike and contact
surfaces. However, experiments indicate that non-spherical restitution and friction coefficients
may increase at low impact angles (about <25
o
). If this variation is included empirically,
stochastic simulations give reasonable representations of mean collision properties as shown in
Fig. 7.33. Note that at shallow impact angles, the mean restitution becomes greater than unity
because the instantaneous impact angles will tend to be larger than the mean impact angle (due
to shadow effect bias). Similar results were observed in measurements by Hamed et al. (1995).

7.2.10. Particle-Particle Interaction Methods

Particle-Particle Fluid Dynamic Interaction

As discussed in 4.2.3, point-force Lagrangian methods are not well suited to handle
individual fluid dynamic interaction of particles with each other, i.e. three-way coupling effects.
Part of this unsuitability lies in the complexity of describing a generalized force associated
with multiple surrounding particles at various relative angles and locations. Another part lies
in the computational expense of determining the proximity of all relevant particles (this issue is
simplified to some degree for binary collisions, as will be discussed later). Because of this, the
more common approach is to use a simple-summation approach based on Eulerian
representations of the particle concentration field. As such, the methods to predict from the
Lagrangian trajectories (7.2.7-7.2.8) can be combined with a particle equation of motion that
incorporates three-way coupling effects (Eq. 7.133b).


Identifying Particle-Particle Collisions

For particle-particle collisions, computation of the locations and times of the collisions
depend on whether individual particles or parcels are being tracked. For individual particles,
the process is similar to that for wall collisions (Eq. 7.208) and generally includes a hard-
395
sphere collision model as will be discussed in the next sub-section. However, particle-particle
collisions are typically more difficult to identify because both bodies are small, non-planar,
moving at velocities which must be predicted with ODEs, and since collision pairs may come
from different computational cells. To find all the possible collision pairs, one may use a brute
force search of all particles with respect to all other particles but this would result in a
2
p
N
operation, where N
p
represents the total number of particles in the domain. Such a search is
computationally prohibitive for a large number of particles, e.g., 1000 or more. As such,
efficient search techniques are required with respect to identifying collision pairs. There are
two main techniques for this identification: the proactive method and the retroactive
method (Sundaram & Collins, 1996). Both assume that the particle velocities are deterministic
but the proactive method examines whether trajectories crossed during the timestep while the
retroactive method examines whether particles overlap at the end of the timestep.
The proactive method is similar to that used for the particle-wall interaction. It assumes
that a particle will generally undergo only one collision within a given time-step, i.e. between
t
n
and t
n+1
. This can generally be satisfied if the time-step is much smaller than the average
time between collisions. This criterion can be expressed in terms of the collision frequency of
Eq. 3.192 as


coll
t 1 < f to resolve particle collisions with proactive method
7.211
Generally, an approximate collision frequency is computed in advanced and assumed constant
for the entire domain and for all time, though it can be adapted. A specific choice for f
coll
t
depends on the degree of variation expected, but values of 10% to 20% are often reasonable
and indicate that most particles will not collide during the time-step. During a given time-step,
each interrogated particle considers all the neighboring particles to see if any of them are a
collision partner.
To identify the neighboring particles in the proactive method, the computational domain
can be divided into collision cells. These computational cells have a volume
coll
which is
smaller than the whole domain but large enough to ensure that collision partners arise from a
common cell or adjoining cells. Thus, the length-scale of these cells should be less than the
maximum distance a particle can travel in a time-step:


( )
1/3
coll max
v t < for collision cells and individual particles 7.212
Often, the collision cells are set equal to the Eulerian computational cells for the continuous-
phase (
coll
= ) since the RHS of Eq. 7.212 is similar to the CFL criterion for a fluid time-
step (Eq. B.43) or the discretization criterion for a particle time-step (Eq. 7.150). Since
potential collision partners are located no more than one cell from the interrogated particle, a 3-
D simulation results in a search of twenty-seven cells (the host cell and the 26 adjoining cells).
Searching for collision partners using the proactive method can be accomplished by a test
for intersecting paths. The first step of the proactive method is to use an explicit advection
scheme to compute all the new particle velocities throughout the domain while neglecting
collisions. Next, the relative particle velocity (defined by Eq. 2.91) between particles i and j is
computed during the time-step based on Eq. 3.21b:


( ) ( ) ( ) ( )
1 1 n 1/ 2 n 1 n n 1 n n 1 n n 1 n
p p i i j j p,i p,i p, j p, j
2 2
/ t
+ + + + +


+ =

v v v v + v x x x x
7.213
396
Use of this constant relative velocity throughout the timestep yields linear change in particle
spacing (l
p-p
of Eq. 3.185a) with time. The initial particle spacing and that at contact (if it
occurs) are defined as


n n n
p p p,i p, j
x x l
coll
p p p,i p, j
r r

+ l
7.214a

7.214b
The latter equation is consistent with l
gap
(of Eq. 3.185b) equal to zero, i.e. when the particle
spacing equals the sum of the particle radii. The above equations can then be combined as a
function of the collision time (t
coll
=t
n
+t
coll
):


coll n n n 1/ 2
p p p p coll p p coll p p
(t t ) t
+

+ = v l l l at collision
7.215
The collision criterion can be expressed as a quadratic equation for the collision time as:


( ) ( ) ( ) ( )
2 2 2
n 1/ 2 n 1/ 2 n n
p p coll p p p p coll p p p,i p, j
v t 2 t r r 0
+ +

+ + + = v l l 7.216
The two roots of this equation can be evaluated and the expected collision time is given by the
minimum positive real root. No collision is expected if the roots are negative or imaginary.
For each particle, collision times need only be considered for j>i (to avoid identifying the same
collision twice) and only the minimum real collision time is saved for all particle neighbors.
These two steps corresponds to Box 1 in Fig. 7.34 and are completed for all possible collision
partners. These steps are the most computationally intensive but can be accomplished in
parallel since collision dynamics have not yet been implemented. The next step is the first
decision diamond in Fig. 7.34 which compares each t
coll
to the timestep to determine whether
a collision may occur before t
n+1
:


coll
0 t t < collision expected within the time-step
7.217
Once this is complete, collision times are computed (using Eq. 7.213, 7.214 and 7.216 by
considering each particles neighbors based on the collision search cells. Since collision
dynamics tend to require small time-steps and many evaluations, single-step explicit advection
schemes are commonly used, e.g., the Adams-Bashforth scheme (Sundaram & Collins, 1996).
The next step in the proactive approach is to sort collisions in order of expected
occurrence within a given time-step, i.e. from smallest to largest t
coll
. This step can be
partially parallelized by considering particles that can not interact, e.g., in non-adjacent cells.
From this list, the earliest collision in the domain is first considered by moving the two
particles involved up to the point of their collision and then implementing a hard-sphere model
(discussed in the next sub-section). This yields the after-collision particle velocities (v
out
) and
new trajectories which completes the steps in Box 2. Since these two particles now have new
trajectories, the collision list must be updated (Box 3) since the remaining expected collisions
may have changed. In particular, new collision pairs may need to be added and others
removed. This continues until all collisions within a time-step have been addressed, i.e.
returning to Boxes 2 and 3 as needed. Finally, all the particles are moved to the end of the
time-step (Box 4) whereby the non-colliding particles are moved using the velocities computed
in Box 1 while the colliding particles use the outgoing velocities computed in Box 2.
397
While this general approach is common, there are many variations that have been proposed.
Sigurgeirsson et al. (2001) investigated such proactive methods to optimize the collision search
cell size and the time-step and to also consider the more complex two-way coupling condition
where the particle affects the surrounding fluid. Sundaram & Collins (1996) also considered
retroactive methods whereby the particles are first moved (ignoring collisions) after which
the particles whose volumes overlap at t
n+1
are assumed to have collided. This method is
sometimes called the overlap method. The overlapping trajectories are then re-integrated
from the initial time only up to the collision time. After this, a hard-sphere collision model is
used to compute the new velocities to integrate the particles to the end of the timestep. This
process continues until no overlaps are present. This method is about two times faster for a
given time-step but can miss some collisions if the overlap is no longer present at the end of a
time-step. To avoid this problems, the time-step should be reduced so that a particle only
moves a small fraction of its diameter towards another particle:


p p
t d / v

< to resolve collisions with retroactive method


7.218
Since this is generally more restrictive than Eq. 7.211 for dispersed flows with point-force
particles, the retroactive method is primarily practical for dense conditions, e.g., >10%.
However, this criterion can be relaxed to allow substantially larger time-steps by implementing
a second grazing detection scheme with the retroactive method (Wunsch et al. 2008).


Particle-Particle Collision Outcomes

To describe the outcome of a particle-particle collision as illustrated in Fig. 2.47, many of
the particle-wall collision techniques still apply but with a generalization of the variables as
noted in 6.10. For example, the wall-normal (n
w
) is replaced by the more general collision
normal (n
coll
), the particle mass (m
p
) is replaced by the more general collision mass (m
coll
), etc.
If the particles are solid and each retain their mass (no coalescence or break-up), their
reflection velocities depend on the normal and tangential coefficients of restitution. The
normal coefficient of restitution (e

) can be used to compute the change in the normal velocity


component with Eq. 6.224. For example, the outgoing normal velocity for particle 1 due to a
collision with particle 2 is given by:


( )( ) ( )
out in in in
,1 ,1 ,1 ,2 p1 p2
1 / 1 m / m

= + + v v v v e
7.219
The change in tangential velocity depends on whether the interaction is in the sliding or rolling
regime, which can be determined by Eq. 6.237. If there is no initial spin velocity for either
particle, this criterion can be based on the difference in tangential velocity:


( )( ) ( ) ( )
in in in in
,1 ,2 ,1 ,2 roll
v v 7 / 2 v v c 1 / 1

< + +
| | |
e e for rolling regime
7.220
Empirical values for c

, e

, and e
roll
are discussed in 6.10, Table 6.6, and Figs. 6.75-6.79.
Depending on the outcome of the Eq. 7.220 criterion, the outgoing tangential velocity is given
by Eqs. 6.234 and 6.239 as:
398


( )( )( ) ( )
( )( ) ( )
out in in in
,1 ,1 ,roll ,1 ,2 p1 p2
out in in in
,1 ,1 ,1 ,2 p1 p2
2 / 7 1 / 1 m / m
c 1 / 1 m / m

= + +
= + +
v v v v
v v v v
| | | | |
| |
e
e
when rolling
when sliding

7.221a

7.221b
The outgoing particle velocities given by Eqs. 7.219 and 7.221 can then be used to move all the
particles which have collided to the end of the time-step.
If the particles are drops or bubbles, one may also need to consider whether coalescence or
break-up will occur. This can be accomplished by computing the impact Weber number (Eq.
2.92) and impact parameter (Eq. 2.93). These can then be compared to the collisional boundary
Weber numbers to determine the phenomenological outcome. For example, if We
SC/B
<We
p-
p
We
B/FC
, the particles will bounce and the above equations can be used. However, if We
p-
p
We
SC/B
, the particles will coalesce. In this case, the total mass and momentum can be
combined and the resulting mass and velocity assigned to the first particle and the second
particle can be eliminated:


( )
( )
out in in
p1 p1 p2
out in in in in out
1 p1 1 p2 2 p1
out in in out out
1 p1 1 1 p2 2 2 p1 1
m m m
m m / m
m d m d / m d
= +
= +
= +
v v v

for coalescence
7.222a

7.222b

7.222c
For We
p-p
>We
B/FC
, the particles may coalesce or separate depending on the combination of
We
p-p
and (Fig. 2.49).
Note that the above approach for collision time and outcome assumes that the particle
velocities are based on deterministic continuous-phase velocities. As such, care must be taken
when employing stochastic approaches which model unresolved turbulent scales for each
particle trajectory by sampling uncorrelated random numbers (7.2.5). Such approaches can
not represent the spatial correlations of turbulence whereby nearby particles see similar
continuous-phase velocities and thus are less likely to collide. This can lead to an over-
prediction of collision frequency when using uncorrelated stochastic approaches (Eq. 3.206).
This is especially problematic at small Stokes numbers when the velocity correlation effect is
highest. In this case, a stochastic parcel approach can be used instead of tracking individual
particles, as will be discussed in the next sub-section. A parcel approach is also useful when
the number of particles in the domain is much larger than the number of Eulerian nodes in the
domain (Fig. 4.19).


Parcel-Parcel Collision Models

For a parcel-parcel collision, the above deterministic approach to predict collisions and
outcomes is replaced by a statistical approach similar to the Direct Simulation Monte-Carlo
(DSMC) method used to study molecular collisions. In this statistical approach, all parcels are
first advected with a collision-less ODE for the full time-step t (typically with a single-step
explicit scheme). Next, each of the individual parcels is considered for possible collision by a
random selection of a fictitious particle from another parcel within the collision cell to
provide a relative velocity. The determination of whether a parcel collision occurs is based
only on the mean probability using the collision cell concentration. It is not based on relative
positions of the parcels, so neither a search process or collision time computation are needed.
399
This is much different than the deterministic method for individual particles which requires a
search of adjoining cells (Fig. 7.34) and a calculation of the time of contact (Eq. 7.216).
The collision frequency (f
coll,j
) for particles in strike parcel i to collide with a fictitious
particle from target parcel j can be based on Eq. 3.195 as:


( ) ( )( )
2
1
coll, j P P P P p, j i j P,i P, j pP, j coll
4
A v n d d v v N /

= = + f 7.223
This collision frequency bases the number density of the target particles on the collision cell
volume. The above frequency can then be related to the probability of collision (P
coll
) within a
given time-step by considering for an individual target parcel and then summing over all
possible target parcels


( )
coll, j coll, j coll, j
1 exp t t = P f f
Pcoll
N
coll coll, j
j 1 =
=

P P
7.224a

7.224b
The RHS approximation of Eq. 7.224a is often used since the probability within a given time-
step is generally assumed to be much less than unity. Assuming that the strike parcel does not
collide with itself, the number of target parcels in the cell is N
Pcoll
=N
P
-1, where N
P
is the
number of parcels in a continuous-phase cell. Since collisions within a parcel are neglected,
N
Pcoll
should be sufficiently large on average to obtain statistical convergence of four-way
coupling effects. This is illustrated in Fig. 7.35 where the many physical particles in a cell are
modeled by a three parcels (N
P
=3), any two of which can be used as target parcels to
determine the relative probability of collision.
This collision probability is then used to determine whether the entire parcel i undergoes a
collision by summing the probabilities over all possible target parcels. For example, consider a
10% probability for a strike particle to collide. This is consistent physically with 10% of the
particles in the strike parcel undergoing a collision within a given time-step. If this net
probability is 10% for ten time-steps, it is assumed that the above process is statistically
equivalent to all the particles from a strike parcel colliding in one of the ten time-steps and
none of the particles to collide in the other nine time-steps.
For computational efficiency, a second statistical assumption is made to determine whether
a parcel will undergo collision: only one of the other parcels are used to consider a fictitious
particle per time-step and this is chose by the relative probability that this particle will lead to a
collision. This can be accomplished with the modified Nanbu method (Iller & Neunzert, 1987)
which selects a target parcel (j) for parcel (i) by a weighted stochastic process:


( )
Pcoll i
j integer N Rn 1 = +
7.225
This expression used the integer function of Eq. 7.113b and the uniform random number of Eq.
7.169. The same random number is used to determine whether a collision occurs for parcel i:


( )
i Pcoll coll, j
Rn j / N > P parcel i collides with a fictitious particle j
7.226
For this scheme, it is important to note that the collision dynamics are used only to update
parcel i velocities while the parcel j velocities are not modified. However, by sampling
random numbers for each of the parcels separately, each will have a chance to undergo a
400
collision independent of whether any other parcel collides. This independence is
computationally advantageous since the advection, evaluation for collision, and reflection for
each parcel can be done in parallel and without sorting lists.
The sampled probability for collision of parcel i for a single time-step with Eqs. 7.225-
7.226 is N
Pcoll
P
coll,j
. This converges to the probability of Eq. 7.224 for large N
Pcoll
after many
time-steps so long as the sampled probability never exceeds 100%. The latter aspect thus
places a time-step requirement for the modified Nanbu method:


( ) ( )
Pcoll coll, j
N t max 100% < f
7.227
This time-step constraint is in addition to Eq. 7.211 but turns out to be similar if the collision
frequency is approximately uniform between parcels. This uniformity can be ensured to some
degree by using parcel initialization to control the number of particles in each parcel (N
pP
). For
monodisperse particle distributions, Eq. 7.223 suggest that an initialization ideally creates
parcels with a uniform number of particles per parcel (Eq. 4.9a). For polydisperse distributions,
the collision frequency is proportional to N
p
d
2
v
rel
. If one assumes the relative velocity is
approximately linearly proportional to particle diameter (reasonable for intermediate Re
p

conditions based on Eq. 1.82), then the collision frequency is proportional to N
p
d
3
which
suggests parcels should be initialized with equal volume fractions (Eq. 7.155a).
If the criterion of Eq. 7.227 is satisfied, the relative velocities between i and j can be used
to compute the collision velocities of Eq. 6.222 to determine the outcome. However, the
outcome depends on the collision normal of the fictitious particle, which has no definite
position. Therefore, a set of random numbers is sampled to find the angles of collision
(Sommerfeld, 2001 and Berelemont et al. 2001). For solid particles, these angles can be used
to compute the outgoing velocity of parcel i based on Eqs. 7.219-7.221. For fluid particles, the
collision Weber number and impact parameter is first obtained to determine the type of
outcome as discussed in the previous sub-section. For bounce, the outgoing velocity is again
based on Eqs. 7.219-7.221. For coalescence, the single particle models of Eq. 7.222 may be
used but with mass conservation within parcel i (parcel j is not modified):


( )
out in in
p,i p,i p, j
out in in out
pP,i pP,i p,i p,i
out in in in in out
i p,i i p, j j p,i
m m m
N N m / m
m m / m
= +
=
= + v v v
for parcel coalescence
7.228a

7.228b

7.228c
Similar equations are used for temperature conservation, and the new diameter of parcel i is
based on Eq. 7.228a. A similar approach may be used for break-up. These approaches assume
a deterministic particle velocity.
For turbulent flows with random and independently sampled continuous-phase
velocities (7.2.5), the stochastic parcel approach may be used with some modification. In
particular, it is important to account for correlation of the continuous-phase velocities between
nearby particles if the particles do not have a high Stokes number. This is because nearby
particle of moderate Stokes numbers will tend to see and be influenced by similar
continuous-phase structures. The resulting correlation of u
@p
between the two particles
indicates that they are less likely to collide with each other. The collision frequency predicted
by Eq. 7.223 does not account for this and will result in an over-prediction. This can be
corrected by taking into account the ratio of correlated collision frequency to uncorrelated
collision frequency (from Eqs. 3.203 and 3.207):
401


coll,turb
coll,turb, j coll, j coll, j
coll,kinetic
St
dSt St
St 1 St
24St 1 St


= +


+
+


L
L
L L
f f f
7.229
Stochastic versions of this have also been developed for LES flows (Sommerfeld, 2001;
Berelemont et al. 2001) using fictitious particle velocities and sizes sampled from long-time
cell averages. Such approaches are useful for small monodisperse particles in flows which are
steady in the mean.


7.3. Lagrangian Distributed-Force Formulations

As discussed in 4.1.3 and 4.1.4, the point-force particle and the resolved-surface
particle represent two extremes for computational representation. The first extreme is a
rigorous point-force treatment where the particle scales must be small enough that the
continuous-phase velocity gradients are negligible compared to the relative velocity. This
generally implies that the diameter is much smaller than the continuous-fluid discretization
scales (dx). For one-way coupling, this criterion is often relaxed to be d<x. The other
extreme is the resolved-surface treatment whereby the continuous-fluid discretization scales
should be small enough to ensure that the detailed flow over the particle surface is resolved
(xd and td/v). Between these two extremes, one may have particles of intermediate size
(too large to be treated by the point-force method but too small to be treated with a resolved-
surface treatment).
There are two primary methods to incorporate a distributed-force: the spatially-
averaged approach and the semi-resolved approach. In the spatially-averaged method,
the continuous-phase properties seen by the particle (velocity, vorticity, etc.) are based on
spatial averages over the particle surface or particle volume to compute the components of the
surface force (drag, lift, added mass, etc.). The spatially-averaged technique is thus an
extension of the Lagrangian point-force method of 7.2 and is efficient for one-way coupling.
In the semi-resolved method, a single Eulerian velocity field is employed which integrate the
disturbance velocity field caused by the particle so that it represents the velocities of both
phases. This technique is a lower resolution version of the resolved-surface method of 8.2,
and allows self-consistent two-way and three-way coupling though is limited to small Re
p

values. A rule-of-thumb for the discretization limits of these approaches is given in terms of
increasing particle size as:

1/3
d x < point-force approach with two-way coupling
d<x point-force approach with one-way coupling
d<3x spatially-averaged approach with one-way coupling
d>10x semi-resolved approach
d>30x resolved-surface approach
7.230
The point-force limits are consistent with Eqs. 4.44 and 4.4.6X and the resolved-surface limit
is consistent with Eqs. 4.50 and 4.75. The criteria for the spatially-averaged and semi-resolved
approaches will be explained in the following discussion of these two methods.

402
7.3.1. Spatially-Averaged Methods

The spatially-averaged method is based on an Eulerian representation of the continuous-
phase flow and a Lagrangian description of the particle trajectories and velocities. Previous
versions of this approach have employed a simple volume-average for the continuous-phase
velocity within a filter-radius. In particular, an Eulerian-Lagrangian velocity transfer function
(Z
u
) can be specified as the Heaviside function, i.e. unity when the distance from the particle
centroid (r) is within a filter-width given as a
u
r
p
:


( )
( )
( )
Z
Z
u p
u j p
N
j u j p
j 1
avg p N
u j p
j 1
1 r a r
Z ,
0
( )Z ,
( )
Z ,
=
=

=

x x
u x x x
u x
x x
for
else


7.231a


7.231b

For equal spacing and many nodes within the filter radius, the summation normalization yields
a simple volume average. The relative particle velocity for drag and lift forces can then be
obtained as w
avg
=v-u
avg
. Bagchi & Balachandar (2003) investigated this transfer function with
a
u
values of 1.2 and 10 for a fixed solid particle in a turbulent flowfield (Fig. 7.36). The
spatially-averaged force results were compared to both resolved-surface force predictions and
point-force predictions. In general, the spatially-averaged method compared reasonably with
the resolved-surface predictions for d=1.50.5x with mean Re
p
of 260 (Fig. 7.37a). For
larger particles with d=9.63x with a mean Re
p
of 600, the spatially-averaged method
yielded poor correlations but at least gave the correct level of fluctuations, whereas the point-
force gave large non-physical oscillations (Fig. 7.37b). A spatially-averaged method consistent
with known theoretical limits for a single spherical solid particle surface force was put forth by
Loth & Dorgan (2009). It employs a surface-average (Z
u
=1 only at r=r
p
) of the continuous-
phase properties for the quasi-steady drag, history and lift forces but a volume-average (Z
u
=1
only at rr
p
) for the fluid-stress and the added-mass forces. In the following, the justification
for these averages is discussed for both low Reynolds number and inviscid theoretical limits,
followed by a generalized semi-empirical expression based on these averages as well as
discrete representations for computational implementation.


Surface Force Expression in Non-Uniform Flow

The quasi-steady drag force for general particle shapes was considered for a non-uniform
continuous-phase velocity field by Brenner (1964) using the reciprocal theorem. The drag
force at small Reynolds numbers was found to be related to the average continuous-phase
velocity on the particle surface (u
surf
), e.g. for a sphere this becomes:


( ) ( )
f
D p f surf
3
dA 3 d
d

= =

F v u v u 7.232
In this equation, A
p
is the particle surface area (Eq. 1.2). The particle torque can be similarly
related to the surface averaged-vorticity at small Reynolds numbers as:


( )
( )
1 3 f
surf p f p f ,surf
2
6
dA d
d

= =

r v u T 7.233
403
For a sphere, a Taylor series can be used to replace the surface average of a quantity q with the
q and the even-numbered derivatives evaluated at the center (Hobson, 1931) as


( ) ( )
2 4
2 4
sphere sphere @r 0
r 0 r 0
d d
qdA A q q q ...
24 1920
=
= =

= + + +


7.234
For Stokes flow,
4
0 = u everywhere so that the drag force can then be expressed with the
conventional Faxen corrections of Maxey-Riley (Eqs. 6.3a and 6.3e) as:


( )
( )
2
2
D f surf f @p
@p
d
3 d 3 d
24

= =



F v u v u u
7.235
The history force of Maxey-Riley includes a similar Faxen correction (Eq. 6.3d), also
consistent with u
surf
. Similarly, the interphase heat transfer on the RHS of Eq. 7.103e for small
Re
p
is based on a surface-averaged temperature difference: T
p
-T
f
,
surf
(Michaelides, 2003). For
Stokes flow,
2
0 = so that
f
,
surf
=
f@p
indicating that the torque of Eq. 7.233, simply reverts
to creeping flow result (Eq. 6.156). Thus, the low Reynolds number drag, heat transfer and
torque for a solid particle in a non-uniform flow can be based on surface-average values of the
continuous-phase properties. However, this is not true for fluid particles. In particular,
Lovalenti & Brady (1993) showed that the Faxen correction is not consistent with a surface-
average and, in fact, tends to zero for clean bubble. As such, the drag force of a clean bubble
should employ a centroid-based fluid velocity (Eqs. 1.64 and 6.45) and not a surface-average,
which is qualitatively consistent with the results of Merle et al. (2005).
For small but finite Reynolds numbers, Lovalenti & Brady (1993) also showed that the
fluid stress and added mass force for a solid sphere can be written in terms of volume-based
averages of the fluid accelerations:


( )
S f p f p
vol
f p f p f p
vol vol
d
t t
c d c c
t t t t t



= =






= =



u u
F g g
v u
v u v u
F
D D
D D
d
d d d D
d d d d D

7.236a


7.236b
The RHS approximation of the fluid acceleration is appropriate in the limit of small Reynolds
numbers (Maxey, 1993). As such, the fluid stress and added mass forces for a solid particle in
a non-uniform flow arise directly from a volume average. Since the fluid stress force is
independent of Reynolds number (6.5) and this is generally true for the added mass force
(6.4), the above volume-averages is expected to be appropriate for any solid particle in
incompressible flow (Loth & Dorgan, 2009). Note that the above volume-average is not
appropriate for fluid particles, at least for low Re
p
conditions (Lovalenti & Brady, 1993).
By applying two additional assumptions, the above volume-averaged result for added mass
can be simplified to yield the conventional Faxen correction. The first of these assumptions is
to exchange the time derivatives with the spatial-averaging. This is often a reasonable
approximation because the fluid-acceleration gradients are typically weak since they are
higher-order than the fluid-velocity gradients. Next, a Taylor series can be used to replace the
spherical volume average of a quantity q in terms of properties at the center:


( ) ( )
2 4
2 4
sphere sphere r 0
r 0 r 0
d d
qd q q q ...
40 4480
=
= =

= + + +


7.237
Making the second assumption of creeping flow for which
4
0 = u yields Faxen corrections for
the fluid stress and added mass forces:
404


( )
( )
2
2
S f p @p f p
@p
2
2
f p @p
@p
d
t 40
d
c
t 40


= +




=



F u u g
F v u u
D
D
d
d

7.238a


7.238b
This added-mass correction is equal to that of MR (Eq. 6.3c) but the correction for the fluid-
stress force was not included by MR (Eq. 6.3b) since they assumed that the fluid stress was
approximately constant in the vicinity of the sphere.
Unlike the case for drag, there is no general analytical solution for the lift force of a sphere
in non-uniform vorticity since this problem is generally intractable (Saffman, 1965).
However, the Saffman lift force derivation is based on a surface integral of the square root of
the vorticity, so that Eq. 6.131 can be expressed in terms of the surface-averaged vorticity and
continuous-phase velocity to give a lift force which is at least third-order accurate in particle
diameter (Loth & Dorgan, 2009). Similarly, the inviscid lift force is theoretically problematic
to extend to flows with non-uniform vorticity (Auton et al. 1988). But since the Auton
derivation stems from a surface integral of weak vorticity, the lift force of Eq. 6.6 for flows
with pure rotation (no strain) can also be based on surface averages of the vorticity and
continuous-phase velocity.
Based on the above and Re
p
corrections to the quasi-steady drag, history force and lift force
discussed in Chapter 6, a semi-empirical expression for the surface force for a solid sphere at
finite Reynolds number and non-uniform unsteady flows can be given as:


( ) ( )
( )
( )
( )
surf surf f surf f p f p
vol
t
surf
f p,surf L surf surf
0
3 df 1 c c
t t
3 d H Re ,
t



= +




+

u v
F v - u - g +
v - u
F u
D d
D d
d
d
d


7.239
Recall from Eq. 3.146, the interphase force (F
int
) can be obtained by simply subtracting out the
fluid stress force of Eq. 7.236a. It should be noted that this result has been derived in the limit
of small Reynolds number so that the three caveats of 6.1.1 should be kept in mind.


Discrete Representation of Surface Force

For the discrete form of the surface- and volume-averages, it is straightforward to use
sampling at particular locations. For the surface-average, gradients in all three directions can
be captured with six sampling points (x
s
) on the surface at left/right (x
p
r,y
p
,z
p
), top/bottom
(x
p
,y
p
r,z
p
) and front/back (x
p
,y
p
,z
p
r) positions as shown in Fig. 7.38. The surface averaged
velocity is then


6
surf p s,i surf
p
i 1
1 1
dA
A 6

u u u u
7.240
The discrete surface averaging can be expressed in terms of an Eulerian-Lagrangian transfer
function described by local shape functions for each computational cell as:
405


( )
( )
( )
( )
Z
Z
j s j s
u, surf j p
N
j u, surf j p
j 1
u, surf p N
u, surf j p
j 1
,
Z ,
0
( )Z ,
( )
Z ,

x x x x x
x x
u x x x
u x
x x
for
else


7.241a


7.241b
This transfer function and surface-averaged velocity should be used instead of Eq. 7.231 for
drag. A similar six-point surface-averaged vorticity can be used for lift. For a fifth-order
polynomial variation in velocity, the discrete expression becomes:


( ) ( )
2 4
2 4
surf @p
@p @p
d d
24 1152

= + + u u u u 7.242
This indicates that a conventional point-force, which only uses the first term on the RHS, is
only second-order accurate with respect to particle diameter. The discrete distributed-force
expression (Eq. 7.242) has the same second-order terms as the exact expression (Eq. 7.234)
and so is fourth-order accurate in diameter. For a Stokes flow,
4
0 = u so that the discrete
distributed-force method becomes exact and also equal to the Faxen correction (Eq. 7.235).
For the volume-average of fluid acceleration, it is convenient and computationally efficient
to use the same six surface sampling points along with an additional sampling point at the
particle centroid (x
@p
). This combination can be optimized using a fourth-order accurate
Simpsons rule for spherical integration to approximate the volume average:


6
p
p vol s,i @p vol,
i 1
1 3 1 2
d
t t 5 6 t 5 t t

=


+

u u u u u D D D D D
D D D D D
7.243
Application of this discrete approximation yields:


2 4
2 4
vol, @p
@p @p
d d
t t 40 t 1920 t



= + +



u u u u D D D D
D D D D

7.244
Again, the discrete sampling method and the Faxen correction (Eq. 7.238) are exact for Stokes
flow conditions and are fourth-order accurate with respect to particle diameter for non-
Stokesian flows (based on Eq. 7.237). The error for the point-force and the discrete sampling
method are shown in Fig. 7.39 for a sinusoidal shear layer with zero mean, i.e. u =0. The
trends indicate that the spatially-averaged force may be an order of magnitude more accurate
than the point-force. For a finite mean flow, this error will reduce proportional to
rms rms
u / (u u ) + but further studies are needed for finite Reynolds numbers.


Two-Way, Three-Way, and Four-Way Coupling

Unfortunately, two-way coupling with the spatially-averaged method suffers from the same
particle spacing limits of the point-force method (Eq. 4.47) in evenly-spaced particles are
ensured to result in a uniform force coupling to the continuous-phase equations. As such, it
does not provide a clear advantage and has not been used in this manner. Fortunately, particle-
wall and particle-particle interactions (four-way coupling) with the spatially-averaged method
can be incorporated with the collision techniques discussed in 7.2.10. In particular, the
deterministic proactive approach can be used since Eq. 7.211 will not be overly restrictive for
large particle sizes.
406

7.3.2. Semi-Resolved Methods

When particle sizes are too large for the spatially-averaged one-way coupling approach or
too large for the point-force two-way coupling approach, a semi-resolved technique can be
employed while still avoiding the computational resolution needed for a resolved-surface
approach. A common semi-resolved method is the force-coupling method (FCM) of Maxey
& Patel (2001). The FCM velocity field is obtained by solving an Eulerian PDE similar to that
for unhindered continuous-phase velocity except that the interphase force for an individual
particle is included. For example, for heavy particles with no mass transfer, Eq. 7.20 can be
written as a PDE for the combined velocity field (V
m
) as:


( )
int m
m m i j
f f p
1
p K
t

+ = +

F V
V V g -
7.245
For an approximately steady particle velocity, the interphase force can be set equal to the
gravitational force monopole:


( )
int grav p p f
= = F F g
7.246
This force can be distributed as shown in Fig. 4.7b using an assumed spatial variation. For
creeping or inviscid flow, the interaction of the particle with the continuous-phase can be
described by an approximately isotropic radial decay (1.5.2). Thus, a reasonable choice for
the Lagrangian-Eulerian force transfer function is a clipped Gaussian decay based on the radial
distance (r) from the particle-centroid to a continuous-phase node j:


( )
( )
j j p
2
1
j F F F
2
F j p
F F
r
exp r / r r a r
Z ,
0 r a r

>

x x
x x
for
for

7.247a


7.247b
The second equation includes a force radius (r
F
) which determines how rapidly Z
F
tapers off
radially from unity the particle centroid. It also uses a filter width (a
F
r
p
) chosen to clip once the
transfer function is small, e.g. a
F
=3 so that Z
F
=1.1% at the edge (beyond which it is zero). The
inner region where Z
F
is non-zero is called the force envelope.
To determine an appropriate envelope radius, consider a single Stokesian particle falling at
a steady rate in a quiescent bath (so that V
m
0 far from the particle). Imposition of F
int
on Eq.
7.245 will induce a relative velocity as shown in Fig. 7.40a. It can be shown that setting
r
F
=r
p
/(3)
1/2
for the coupling force of Eq. 7.246 will result in velocity field whose average value
at r=r
p
(i.e. on the points where the physical particle surface would lie) will equal to the
particle terminal velocity, i.e. V
m,surf
=v
term
=w
term
(Maxey & Patel, 2001). This can be seen by a
plot of the vertical velocity on a horizontal plane in Fig. 7.40b. Typically 5-10 grid points
across the particle diameter ensures suitable grid resolution (Maxey & Patel, 2001), a result
consistent with the criterion noted in Eq. 7.230. Thus, application of this force envelope
results in a single combined velocity field (V
m
) which approximates the continuous-phase flow
(U) away from the particle but approximates the particle velocity by:
407


m,surf
= v V
7.248
This surface-average approach in robust in that it automatically incorporates the Faxen drag
correction and is generally reasonable for Re
p
<10 (Climent & Maxey, 2003). It is also
convenient in that an ODE to determine the particle velocity (Eq. 7.103b) is no longer needed
and Eq. 7.248 can be used directly for the particle position ODE (Eq. 7.103a).
Another attribute of the FCM approach is that it directly incorporates two-way and three-
way coupling effects if multiple particles are considered, e.g. settling velocity changes with
volume fraction were found to be consistent with the empirical Richardson-Zaki formulation of
Eq. 6.249 (Maxey & Patel, 2001) and trajectory interactions of Fig. 3.45 were also reproduced
(Lomholt et al. 2003). Furthermore, particle spin and torque may be included by adding a
force dipole (Lomholt et al. 2003).
One drawback of the FCM interphase force of Eq. 7.246 is that it assumes that the particle
acceleration is weak compared to gravity, i.e. inertia effects are assumed to be small (similar to
the weakly-separated method of 5.3). However, this limitation can be removed by including
particle inertia and acceleration (Reichardt & Sommerfeld, 2007):


( ) ( )
t t t
int p p p f p
/ t
+
= + F v v g
7.249
Note that this approach may require iteration between Eqs. 7.245, 7.248 and 7.249 since the
new particle velocity is not known until V
m
is computed. A second drawback is that the FCM
method can not employ the collision techniques discussed in 7.2.10 since the particle velocity
is a derived quantity (although, collision dynamics may be incorporated with a repulsion force
to be discussed in 8.1.2). A third drawback is that the FCM approach assumes attached flow
over a simple particle geometry. As such, it is not readily extended to higher Reynolds number
and more complex particle geometries (effects more accurately handled with the techniques of
Chapter 8).


7.4. Chapter 7 Problems

7.1) Consider dilute spherical solid particles with density of 2000 kg/m
3
and specific
heat of 2000 J/kg-K which pass through a normal shock wave of air with =1.4 (whose
thermal conductivity and viscosity can be approximated by the values in Table A.1). The
flow moves from left to right and has a Mach number of 1.4 and STP conditions just
upstream of the shock and the particles are at the same speed and temperature of the air flow
upstream of the shock wave. Neglect gravitational effects and assume one-way coupling for
the gas-phase with a surface force based on drag.
a) Write the Eulerian particle transport equations for mass, momentum and energy and
choose appropriate expressions for particle drag and heat transfer based on Reynolds number
b) Determine an appropriate numerical scheme and write the discrete version of the particle
transport equations.
c) Write a program to obtain the spatial variation of particle properties for a given particle
diameter.
d) Plot the particle concentration, velocity and temperature spatial variations for three
different particle diameters: 10 m, 25 m and 100 m.and discuss the results.
408
7.2) Consider a downward flow of water and sand between two vertical plates which are
50 cm long and 1 cm apart. The water flow is not coupled to the particles and is laminar with
a parabolic velocity profile that peaks at 0.1 m/s while the sand concentration at injection is
uniform with =0.01. Assume that the sand particles are spherical with a specific gravity of
two and a diameter of 100 m. and are subjected to gravity, drag and lift forces, and are in
spin-equilibrium based on Eqs.???. For the particle velocity assume slip conditions at the
side walls and an initial velocity equal to the liquid velocity.
a) Write the one-way Eulerian particle mass and momentum transport equations..
b) Determine an appropriate numerical scheme and write the discrete version of these
equations.
c) Write a program to determine the spatial distribution of particle velocity and concentration.
d) Plot the downstream particle velocity and concentration distribution as a function of
horizontal distance between the plates at the exit plane (50 cm below injection) with and
without including the lift force and discuss the result.
7.3) Derive Eq. 7.30 starting from Eq. 7.29.
7.4) Derive Eq. 7.34 starting from Eq. 7.2.

3.3) Derive the exponential explicit scheme similar to Eq. 7.135?? but which assumes
Stokesian drag (f=1), zero gravity (g=0) and constant fluid acceleration (not velocity)
throughout the domain. Approximate the acceleration with a central-difference scheme, i.e.

( )
n+1/2 n+1 n
@p @p @p
u / t u u / t - .
3.4) Write a computer program using the exponential scheme of Problem 3) to predict the
velocity evolution of a particle with
f
=
p
(i.e. St
D
=1) in response to a fluid velocity given by
u(t)=u
o
sin(2t/
f
) with v=0 at t=0. Using a reasonably small time-step, plot the continuous-
phase velocity evolution (u/u
o
) as a function of time (t/
p
) along with non-dimensional particle
velocities (v/u
o
) for R=0 and R=3. Explain any differences between the velocities and discuss
which surface force effects dominate each particles motion.
3.5) Obtain the Newton-drag explicit one-step scheme of Eq. 3.21a for the particle velocity
starting from Eq. 3.4 assuming constant fluid velocity. Show all steps and the discrete
temporal accuracy of the scheme assuming that C
D
is constant.
3.6) Consider a particle is initially at rest (v=0 at t=0) but is subjected to a constant fluid
velocity of constant magnitude u
o
and a non-linear drag coefficient given by Eq. 1.52. The
initial particle Reynolds number of Re
p,o
=u
o
d/
f
can be used to compute the initial drag
coefficient (C
D,o
) and the initial particle response time (
p,o
). Rewrite the Newton-drag
scheme of Problem 5 for a one-dimensional flow with zero gravity (g=0) but variable C
D
in
terms of v/u
0
, t/
p,o
and C
D
/C
D,o
. Predict v/u
0
for a very-heavy particle (R0) from t/
p,o

ranging from 0-3 for Re
po
=0.1, 10 and 1000. On one plot, show the numerical solutions as
well as the Stokes analytical solution (Re
p
1) and discuss effects of non-linear drag.



409

8. Resolved-Surface Approach

The resolved-surface method is an accurate approach to predict particle surface forces,
especially for complex and deforming shapes and for flows with spatial gradient on the order
of the particle length-scale. This is in contrast to point-force approximations which employ
analytical or empirical force components that are appropriate for simple flows and geometries.
A resolved-surface simulation (RSS) requires sufficient spatial discretization of the particle
surface and the local flow disturbance it causes. This allows prediction of the detailed surface
stress distribution without empiricism or analytical limitations. It should be kept in mind that
the numerical resources for RSS are significant for a single particle and this often restricts such
simulations to a modest amount of particles as discussed in 4.3.
In general, resolved-surface simulations can be divided into two primary classes based
upon how the particle interface is described with respect to the computational grid of the
domain. In particular, RSS can employ surface-fitted grid methods (SFGM) or continuous-
interface methods (CIM) as illustrated in Fig. 4.9. The key difference between these two
methods is the treatment of the computational grid shown in Fig. 8.1. SFGM divides the
physical domain into two computational domains: one for the surrounding continuous-phase
(
f
) and one for the particle volume (
p
). In contrast, the CIM treats the entire domain as a
single computational zone (
m
) and employs a numerically-imposed finite-thickness interface
(
I
) to determine whether the local fluid belong to the particle-phase (=1), the continuous-
phase (=0) or the (non-physical) mixture within this interface (0<<1). Figure 8.2 further
distinguishes the treatment of the density and PDEs for these two methods.
In general, the surface-fitted grid methods are well suited for particles and domains which
are simple in shape and geometry and are discussed first in 8.1. The continuous-interface
methods can be used for particles of complex shape while allowing for a simple structured grid
domains. The techniques for solid particle are first discussed in 8.2 followed by a larger
variety of techniques for fluid particles in 8.3. Finally, a related methodology for fluid
particles called Sharp-Interface Methods (SIM) are briefly discussed in 8.4. This technique is
more computationally demanding but limits the numerical interface to a single grid cell.


8.1. Surface-Fitted Grid Methods

The discontinuous-interface description is the simplest conceptually as it involves a
straightforward division of the particle domain and the surrounding domain by an explicit
interface boundary (A
p
), across which is a step jump in the density and viscosity. Such a
discontinuity is appropriate because the physical interface thickness between the particle and
the fluid is generally very small (thickness of a number of molecules) compared to that of the
particle length-scale. Since the surface-fitted method requires placing a mesh that conforms to
this surface, these methods are often used for solid or non-deforming fluid particles, such that
the interface tracking and mesh creation is straightforward. Placing the grid along the particle
surface allows high-accuracy of surface shear stress, which is critical to the surface force.
If the particle is solid, then the surface-fitted grid method is simplified to one set of PDEs
since the particle interior does not need to be treated as a flow-field. If the particle is a fluid
410
(drop or bubble), then its internal flow-field may be described with the fluid dynamic PDEs
appropriate to the physics of concern. Note that these PDEs need not be the same as those for
the continuous-phase, e.g., for high-speed flow past a drop, one may reasonably use
compressible Eulerian equations for
f
while using incompressible Navier-Stokes equations
for
p
. In contrast, for low-speed flow past an oscillating bubble, one may reasonably use
incompressible Navier-Stokes for
f
while using compressible inviscid equations for
p
. This
segregation allows improved efficiency and accuracy if the fluid physics of the two domains
are significantly different. Similarly, the separation of the physical domain into two
computational domains also allows different numerical techniques to be used for each phase-
field. For example, an explicit scheme time-integration scheme is more appropriate for a
compressible flow domain whereas an implicit time-integration scheme is often more
appropriate for incompressible flow domain. Thus the physical and numerical approaches
(PDEs, time-integration schemes, grid structure, spatial-discretization, grid-scale, etc.) can be
idealized for each phase field in order to optimize the overall accuracy and efficiency. In the
following, the surface-fitted grid methodology is discussed first in terms of the continuous-
phase including equations of motion, computational grids, and surface boundary conditions.
Next, the particle centroid equation of motion is considered in terms of the trajectory ODE and
implementation of fluid dynamic and collision forces.

8.1.1. Transport Equations, Grids and Boundary Conditions

PDEs for Flow Surrounding a Particle

For the flow outside of the particle (in
f
), the governing equations for the resolved-surface
approach can use various assumptions (e.g., incompressible, isothermal, inviscid, creeping,
steady, etc.) and the choice is generally based on physics (capturing certain phenomena) and
practicality (ensuring that the required computational resources are reasonable). If the flow is
generally a continuum (Kn
p
~0), the basic forms of the governing equations are discussed in
Appendix A (and associated numerical approaches are discussed in B.1-B.7). If non-
continuum perturbation are important (Kn
p
1), one may modify the continuum equations by
simply replacing the no-slip tangential boundary condition on the particle surface with that
given by Eq. 6.48a. Flows for which the Knudsen number is no longer small generally require
statistical techniques such as the Direct Simulation Monte-Carlo Method of B.8.
In the following, the flow around the particle is assumed to be viscous, incompressible,
non-stratified and isothermal. As such, the Navier-Stokes transport equations of Eq. A.55 can
be written as


0 U =
2
f f f f
P
t

+ = + +

U
U U g U
8.1a

8.1b
Note that the variable p corresponds to the continuous fluid pressure which includes variations
induced by the particle surface. This momentum equation is shown in Table 4.1 for the SFGM
and can be simplified for steady flows by neglecting the transient term (1
st
term on the LHS)
and/or for creeping flow by neglecting the convective term (2
nd
term on the LHS). The
411
transport equations can also be specified in 2-D (axisymmetric) or 3-D (spherical or Cartesian)
coordinates, depending on whether the flow field is axisymmetric and whether the particle
remains spherical. For axisymmetric flow, the flow-field is sometimes solved in terms of
stream functions and vorticity (Eq. A.30), e.g., Ryskin and Leal (1984) for solid particles and
Mei & Klausner (1994) for bubbles. However, more generally, these equations are
numerically solved using a momentum equation for velocity and a Poisson equation for
pressure correction as discussed in B.3.3.
For compressible flow conditions, the fluid density must be included as a variable along
with a fluid equation of state. The respective PDEs for mass, momentum and energy are
generally written in terms of the conservative variables:


( )
f
f
0
t

+ =

U
( )
( )
f
f f i j
P K
t

+ = +

U
UU g -
( )
( ) ( ) ( )
f tot
f tot f i j
e
e P K
t

+ = +

U g U- U U

8.2a

8.2b

8.2c

In these equations, K
ij
represents the viscous stress tensor (Eq. A.9) written as a function of U
and
tot
e includes the internal and kinetic energy (Eq. A.13). These equations are generally
solved with flux-based schemes as discussed in B.3.4.
The above equations of motion assume that the computational domain is fixed in space or
is moving at a constant velocity, e.g., at a speed of U

. As such, they can be for a single


spherical particle moving at a steady centroid velocity (v) which is conveniently solved with a
stationary computational grid around the particle surface and a far-field velocity (U

) that is
equal to the relative flow velocity to the particle. However, if the particle moves relative to
geometric features of the domain of if the particles relative velocity is unsteady then special
numerical schemes must be employed. If the particle surface moves within the domain, then an
Arbitrary Lagrangian Eulerian (ALE) formulation is generally used.
For the case where the particle relative velocity is unsteady, the domain coordinate system
and velocity field require a Lagrangian transformation:


p
(t) (t)
( , t) ( , t) (t)
=
=
x x x
U x U x v



8.3a

8.3b
Any static flow variables like density, static pressure, static temperature remain the same in
this transformation. Similarly, any properties associated with the particle motion such as drag
force or particle acceleration are unaffected, e.g.,
D D
F (t) F ( t ) =

.
For the transport equations, the spatial coordinate system is translated but not stretched so
that = x x

. As such, all spatial derivatives are unmodified, e.g. =

and
ij ij
K K =

.
However, the time derivative of a property q varying in physical time can be expressed as
partial derivatives of the transformed time moving and transformed space as follows


i
i
x q q t q q q
. q . q
t t t x t t t t

= + = + =

x
v


8.4
This makes use of the relations / t = x v

and t t =

. Applying these transformations to


the compressible transport equations of Eq. 8.3 leads to additional source term for momentum
and energy PDEs:
412


( )
f
f
0
t

+ =


f
f f i j f
P K
t t

+ = +

U v
UU g -



( )
( ) ( ) ( )
f tot
f tot f i j f f
e
e P K .
t t

+ = + +

v
U g U- U U g v U



8.5a

8.5b

8.5c

Note that the total specific energy must be written in terms of transformed velocity field
2
tot
e e + U / 2

and all terms involving v are zero since the particle centroid velocity is not
a field property.


Computational Grids

A structured surface-fitted grid is well-suited for a single particle with a simple shape. In
the simplest case of a spherical particle in an unbounded fluid, a static spherical grid can be
employed which is centered at the particle centroid as shown in Figs. 4.10a (close-up), 4.17c
(conceptual), and 8.3 (entire domain). Finer radial resolution is generally required near the
particle surface especially. For example, r
o
/r
p
=2.5% is reasonable for solid particles up to Re
p

of 20 (after which flow separation may occur) while 1% is reasonable for a Re
p
of 500
(Legendre & Magnaudet, 1998). As noted by Comer & Kleinstreuer (1995), similar resolution
may be used for ellipsoidal particles (with an ellipsoidal grid). If the boundary layer on the
particle surface becomes turbulent, then radial resolution can be based on Eq. 4.79, i.e. r
o
+
~ 1.
In general, the far-field edge for particles in unbounded flow will be many particle diameters
away as shown in Fig. 8.4, especially for lift at low Reynolds numbers as shown in Fig. 8.4.
Similarly, domain sizes are needed to properly capture history forces at low Reynolds numbers
(Chang & Maxey, 1994). Based on the domain size, r
o
, and a 10-20% grid stretching from
the surface, Eqs. 4.81-4.82 can then be used to determine the total number of radial nodes in
the domain.
For multiple particles or particles moving with respect to confined domains, surface-fitted
grids are more commonly achieved with Chimera grids (Fig. 8.5) or unstructured grids (Fig.
8.6). Chimera grids, also referred to as oversetting grids, superpose surface-fitted grids onto
a global grid as shown in Fig. 8.5. As such, simple computational grid can be combined and
still allow computationally-efficient structured grid solvers to be used. However, interpolation
must be conducted at each time-step between all overlapping active cells. This requires
protocols to ensure that the nearest overlapping cell neighbor can be defined rapidly, and the
degree of overset will increase the overhead associated with this technique.
Unstructured-grids can allow for particles in close proximity and/or whose shapes are non-
spherical and typically employs tetrahedral elements e.g., Fig. B.2b but can be combined with
structured grid is used near the particle surface through use of a hybrid grid (Fig. B.1). As the
particles change positions change relative to each other or to the computational domain
geometry, the grid can be deform after each time-step while preserving the overall connectivity.
However, re-meshing is often needed after significant movement to avoid highly skewed (high-
aspect ratio) elements which can lead to substantial errors in the solution calculation. Thus,
one should keep in mind the additional computational overhead associated with deformation
and re-meshing with unstructured grids. Another important consequence with the use of
413
unstructured grids is that finite-difference and spectral schemes are difficult to employ so that
finite-volume and finite-element schemes are commonly used. As such, the numerical
discretization is typically limited to second-order accurate schemes, which means that
additional node densities are required as compared to grids used in conjunction with higher-
order schemes. A third consequence of unstructured grids is that each element is related
individually to other elements through an inter-connectivity matrix rather than through planes
and index increments in each dimension). Because of this, implicit solvers require point-
implicit approaches which are less efficient than rather than line- or plane-implicit approaches.


Surface Boundary Conditions and Internal Flowfields

The numerical solution of U requires proper boundary conditions on the particle surface.
For a non-spinning spherical solid particle in viscous flow with a centroid-based grid, this
simply can be given as U=0 at r=r
p
. This boundary condition may also be employed for fluid
particle which has negligible recirculation, e.g., a highly contaminated drop or bubble. In the
more general non-spherical particle case, the boundary conditions should be related to the local
surface velocity on the particle interface, V
I
. If the particle is solid but not rotating nor
deforming, then this particle interface velocity is simply equal to the particle translational
velocity (v). A solid particle that rotates at about its centroid (x
p
) will have a velocity at the
interface (x
I
) given in terms of the interface radius (r
I
) as shown in Fig. 8.7:


( ) ( )
I I p
I I p p I

= +
r x x
V x v x r

8.6a

8.6b
Based on the assumption of zero mass transfer through the interface, the velocity boundary
conditions on the surface can be expressed as inviscid or viscous. The viscous boundary
conditions on a solid particle yield a no-slip surface such that the particle and continuous-phase
velocities must be equal at the interface:


I I
= U V for viscous surface boundary conditions
8.7
For inviscid boundary conditions, only the normal component of the surface velocity must be
imposed:


I I
= U n V n for inviscid surface boundary conditions
8.8
However, such a boundary condition is only reasonable for special conditions, e.g. a clean
bubble for which drag is completely ignored, and so is rarely used.
For a fluid particle with significant recirculation, the internal particle domain should be
also discretized, e.g., with the same grid resolution used just above the particle surface. In this
case, a separate set of conservation equations is needed for V and
p
. Assuming viscous
incompressible flow for the particle interior fluid dynamics, the PDE form is the same as that
of Eq. 8.2:


0 V =
2
p p p p p
P
t

+ = + +

V
V V g V
8.9a

8.9b
414
In this equation P
p
,
p
and
p
correspond to the fluid characteristics inside the particle. The
continuity-based boundary conditions for the normal component of V
I
at x
I
are consistent with
Eq. 8.8, implying no flux through the interface.
The momentum-based boundary conditions based on the balance of the stresses at the
interface relate the tangential velocity and pressure differences. If the particle is spherical and
the interface is uncontaminated, the tangential and normal stress balances are given by Eq. 1.58.
The tangential balance of Eq. 1.58b includes a RHS term proportional to surface tension
gradients which can arise if there are gradients in surfactant concentration and/or temperature.
To simulate the former, transport equations for the concentration distribution on the surface
can be written in terms of convection and diffusion (Jan & Tryggvason, 1991). The normal
stress balance of Eq. 1.58a can be combined with a gas equation of state for spherical bubble
oscillation to yield a Rayleigh-Plesset equation (Eq. 2.64), which may integrated in time to
give the surface radial velocity and position.
For deformable particles which are non-spherical but axisymmetric and steady, one may
iterate between the transformed stress balance boundary conditions and the interface curvature
to converge to steady-state conditions (Ryskin & Leal, 1984; Ellenboro & Edwards, 2002). If
the fluid particle loses spatial symmetry and is three-dimensional, then one may employ a
Cartesian form of Eq. 1.58. However, this leads to a complex implicit relationship between the
surface shape and the surrounding pressure variations, especially for unsteady flow. As such,
particles with unsteady deformable shapes are most commonly treated with the methods
described in 8.2.

8.1.2 Particle Equations of Motion

Trajectory ODE

The surface-fitted grid particle equations of motion based on Eq. 7.103a-c are:


p
t = x / v d d
( )
p body surf coll
m t = + + v / F F F d d
( )
p p
surf coll
t
= + T T
d I
d


8.10a

8.10b

8.10c
If the collision effects are not included (discussed in the next sub-section), typical numerical
schemes to solve the ODEs are discussed in 7.2.3. For example, heavy particle with linear
drag can be integrated with exponential schemes based on weak fluid acceleration exponential
schemes for linear drag explicit second-order accurate schemes such as the Adams-Bashforth
and Runge-Kutta schemes . The particle Generally, the particle time-step is the same as the
fluid time-step. However, if particle movement requires computationally expensive grid
deformation or re-generation, sub-cycling may be employed so long as CFL
p
(defined in Eq.
7.123b) is sufficiently small for temporal accuracy. As with the point-force method, the body
force is based on gravitational forces (Eq. 1.23). However, the fluid dynamic force and torque
are based on direct numerical integration of the surface stresses while the collision force and
torque are based on interactions with neighboring particles or walls.
415
The surface force can be obtained based on the flow around the particle by summing the
continuous-phase pressure and the shear stresses (based on U and Eq. A.8) for incremental
surface areas (A
p
, based on Eq. 1.2). In Cartesian coordinates, this force can be written as:


( )
surf ij j p p
P K n A A = + =

F n T
8.11
In this equation, T is the discrete fluid stress vector defined by Eq. A.11 as force per area for a
surface with an outward normal n. For a spherical particle, the surface force can be written in
terms of the viscous and pressure stresses in spherical coordinates given by Eq. 1.24.
The fluid dynamic torque about the particle center is based on the surface stresses
perpendicular to the interface radius, r
I
(defined in Eq. 8.8a and shown in Fig. 8.3). For a non-
spherical particle, both the pressure and viscous stresses can produce a moment and the torque
can be written in Cartesian form as


( )
surf I p
A =

r T T
8.12
For a sphere, the pressure does not contribute and the torque simplifies to Eq. 6.153a.


Collision Forces

To incorporate the effects of drop, bubble or particle collisions in liquids within the
resolved surface framework, a repulsion force is often used. This avoids the velocity
discontinuities associated with the techniques of 7.2.10, which can be difficult to integrate
with respect to the local flowfield around the particle. Furthermore, the point of contact can
lead to singularities in the flow solution if lubrication forces (which are generally
computationally expensive to include) are not resolved. A simple, robust and quite common
strategy of which is that proposed by Glowinski et al. (2001) is to employ a repulsion force
when the gap between particles (l
gap
) is small. As shown in Fig. 8.8, the gap length defined for
spherical particles in Eq. 3.185a can be supplemented with a gap normal defined from strike
particle j in the direction of target particle k as:


( ) ( )
gap, j k j k p, j p,k p,k p, j p, j p,k
r r r r

= + = + x x l l
( )
gap, j k p,k p, j p,k p, j
/

= n x x x x
8.13a

8.13b
The repulsion force is turned off when the gap is more than a repulsion length-scale (l
rep
) but is
otherwise constructed to be along the gap normal (so that tangential forces are neglected):


( )
gap, j k rep
2
coll, j k
2
rep g,eff , j gap, j k gap, j k rep rep gap, j k rep

c F /


>

0
F
n
if
if
l l
l l l l l

8.14
This repulsion force at zero-gap is generally scaled with the effective gravitational force (Eq.
1.79) using a coefficient (c
rep
) of order unity and decreases quadratically as the gap approaches
the repulsion length-scale. Typically, this length-scale is a small fraction of the target particle
diameter or of the mesh size normal to the surface (x). For example, Ferrante & Elghobashi
(2007) set l
rep
=2x and c
rep
=5 in order to primarily minimize the number of over-lapping
particles and secondarily minimize the number of particles with non-zero repulsion forces.
416
The above formulation assumes that the fluid dynamic forces are on the order of the
gravitational force. If this is not the case, then a typical drag force magnitude may be
employed in Eq. 8.14.
The repulsion force concept can also be used to handle collisions with walls by employing
an equal diameter image particle which is an equal distance below the wall. Thus, if the closest
point to particle k on the wall is x
p,j
, the effective gap may be set as


gap, j k p,k p, j p, j
2r

= x x l if j=wall
8.15
Typically, particles will only have binary collisions for dispersed-conditions but the repulsion
force on particle k can be generally summed over all possible partners (particles or walls) as:


coll
N
coll,k coll, j k
j 1, j i

=
=

F F
8.16
Note that equal and opposite repulsion forces are applied each of two particles in a binary
collision. For each particle, the net collision force can be used along with the surface force of
Eq. 8.11 to complete the ODE of Eq. 8.10b. Note that the integration time-step must be
sufficiently resolved to apply the repulsion force to avoid large overlaps, e.g.


rep p p
t / v

<l to resolve repulsion force


8.17
For example, the LHS can be set as 1/4
th
of the RHS. This is often not overly restrictive for
l
rep
=2x since it is similar to the CFL criterion for the continuous-phase integration scheme (Eq.
B.43). Hu (2007) discusses other collision schemes based on fluid dynamic forces, including
those with lubrication forces.
A much different approach is the soft-sphere model discussed by Crowe et al. (1998),
which is more appropriate for particles in gasses where fluid dynamic effects are weak and
repulsion is instead dominated by the particle stiffness. This method employs a repulsion force
which is finite only after the particles have overlapped as:


( )
gap, j k
coll, j k
stiff gap, j k damp p, j p,k gap, j k gap, j k gap, j k
0
c c 0


>


0
F
v v n n
if
if
l
l l

8.18
This equation employs stiffness and damping coefficients (c
stiff
and c
damp
). For a solid particle,
these can be based on Hertzian contact theory using Youngs modulus and the coefficient of
restitution. A similar equation can be used for the tangential contact force needed for the
collision torque of Eq. 8.10c. This approach may also be conceptually applied to spherical
droplets if the internal fluid dynamics are neglected. Soft-sphere models assume small
physical deformations that do not require deformation of the numerical grid around the particle.
In the case of fluid particles with substantial deformations, it is often more appropriate to
employ immersed surface methods as discussed below.


8.2. Continuous-Interface Methods for Solid Particles

Employing a continuous-interface allows for a single set of transport equations, and for the
flow outside and around the particle to be treated within a single computational domain with
417
simple (and fast) structured uniform meshes. Thus it can be a powerful technique when
particles have complex shapes, are large in number, and/or move relative to each other and to
domain surfaces. There are many approaches in this category but most are related to the
Immersed Boundary Method (IBM) developed to simulate the flow in heart by Peskin (1977).
When applied to fluids, these variants include Immersed-Interface Methods, Mirroring
Immersed Methods, Smooth Profile Methods, and Volume of Solid Methods. Herein, we
generalize these techniques as Continuous-Interface Methods whereby the interface has some
thickness and can reside arbitrarily within a grid cell, i.e. is not constrained to be along a cell
face.
A particularly efficient and fast IBM method to handle solid surfaces not aligned along
Eulerian grid points was developed by Fadlun et al. (2000 which uses the continuous-phase
PDEs (Eq. 8.1) except that the momentum equation includes an interface force per unit mass
(F
I
) within the vicinity of the particle surface


2 f
I RHS I
f f
P
t

= + + + = +

U
U U g U F F F
8.19
For computational convenience, the RHS includes a function (F
RHS
) with units of force per
mass (i.e. acceleration) that accounts for all the RHS terms except F. The purpose of the
interface force is to allow the Eulerian mixed-fluid velocity near the interface to approximate
the particle surface velocity given by Eq. 8.6b, i.e. to satisfy the surface boundary condition.
To show this, consider the Eulerian nodes and the Lagrangian nodes in Fig. 8.9a. If one
ignores the surface boundary condition (F
I
=0) for Eq. 8.19, and employs a second-order central
temporal discretization, the unconstrained or free velocity prediction at t
n+1
of the nearest
Eulerian node is given by:


free n n 1/ 2
i i RHS
t
+
= + U U F for i=Eulerian mesh node index 8.20
On the other hand, if one assumes a linear variation in U between the Lagrangian particle
surface velocity (V
I
) and the second nearest Eulerian node (U
i+1
), the desired velocity of the
neighboring Eulerian node (
des
i
U ) to enforce this boundary condition can be interpolated as:


( )( ) ( )
des
i I i 1 I i I i 1 I
/
+ +
= + U V U V x x x x
8.21
If one applies the temporal discretization of Eq. 8.20 to Eq. 8.19, the updated velocity is:


( )
n 1 n n 1/ 2 n 1/ 2
i i RHS I
t
+ + +
= + + U U F F
8.22
To achieve the desired velocity of Eq. 8.21 (which implicitly satisfies the surface boundary
conditions), Eqs. 8.20 and 8.22 can be combined to obtain the necessary interphase force:
( )
n 1/ 2 des free
I,i i i
/ t
+
= U U F
8.23
Including this force in the solution of Eq. 8.20 throughout the domain is the heart of direct
forcing method. It effectively incorporates the particle surface boundary condition (to
second-order accuracy) without including a mesh point exactly on the surface, i.e. by using
only the Eulerian nodes. Prosperetti & Zhang (2005) improved the accuracy of this boundary
condition by employing linearized Stokes flow solution in the vicinity of the surface. The
direct forcing method is similar to earlier feed-back forcing methods but the latter are
constrained to smaller time-steps and thus are not as commonly used (Fadlun et al. 2000).
The direct forcing method was improved to incorporate moving particles of arbitrary
geometry by Uhlmann (2005) by including N
s
Lagrangian points attached to the surface of
418
each particle at positions x
s
. The desired velocity on these points is simply the surface velocity
(V
I
) given by the boundary condition of Eq. 8.6b. This can be used to compute the interphase
force which acts on each Lagrangian surface node assuming the free fluid velocity is also
computed is also computed on these nodes:


( )
n 1/ 2 free
I,s I s
/ t
+
= V U F for Lagrangian interphase force
8.24
Since the free velocity is computed on the Eulerian nodes and since the interphase force is
needed on the Eulerian mesh nodes, Uhlmann specified a numerical interface of finite
thickness
I
as shown in Fig. 8.6b. Over this interface, a smooth (continuously-differentiable)
transfer function for both Lagrangian-Eulerian transfer and Eulerian-Lagrangian transfer was
defined by the product of the transfer functions over each dimension:


( ) ( ) ( ) ( )
I i s x i s y i s z i s
Z , Z x , x Z y , y Z z , z = x x
8.25
For example, the transfer function along the x-coordinate is given below as a function of the
distance from the grid interface normalized by the local grid resolution:


2
i s i s
i i
x
2
i s i s i s
i i i
x i s i
x x x x
1
1 1 3 0.5
3 x x
Z
x x x x x x
1
5 3 1 3 1 0.5 1.5
6 x x x
Z 0 x x / x 1.5



=
for
for
else for

8.26
This function thus acts like a delta function over three grid nodes, i.e.
I
=3x.
Assuming uniform Eulerian grid spacing in all three directions, Z of Eq. 8.25 can be used
to convert free velocity and interphase force between the particle surface and the numerically
thick interface as:


( ) ( )
f f
N N
free free
s i I i s I i s
i 1 i 1
Z , Z ,
= =

=



U U x x x x for EulerianLagrangian
( ) ( )
s s
N N
n 1/ 2 n 1/ 2
I,i I,s I i s I i s
s 1 s 1
Z , Z ,
+ +
= =

=



x x x x F F for LagrangianEulerian
8.27a


8.27b
The first sum is over the N
f
Eulerian nodes near the particle while the second is over the N
s

Lagrangian nodes on the particle surface. A particularly convenient aspect of this
methodology is that the surface force (and surface torque) needed to update the Lagrangian
particle is simply based on the sum of the interphase forces (and the respective moment arms)
at the nearby Eulerian nodes:


( )
( ) ( )
f
f
N
n 1/ 2 n 1/ 2
surf f I,i ,i
i 1
N
n 1/ 2 n 1/ 2
surf p i f I,i ,i
i 1
+ +

=
+ +

=
=
=

F
x x
F
T F

8.28a

8.28b
This ensures that the interphase force acting on the fluid is exactly consistent with that acting
on the particle. Uhlmann (2007) found that this technique predicted the terminal drag to within
2% for Re
p
as high as 360 using d/x of about 13 and N
s
of about 500 (Fig. 8.10).
419
In summary, the algorithm for this method includes the following nine steps for each time-
step assuming there is a single particle:
1) compute the free velocity at the Eulerian nodes by solving Eq. 8.19 with F
I
=0
2) compute the free velocity at the Lagrangian nodes by applying Eq. 8.27a
3) compute the interphase force at the Lagrangian nodes by Eq. 8.24 and 8.6b
4) compute the interphase force at the Eulerian nodes by Eq. 8.27b
6) compute the Eulerian velocity field by solving Eq. 8.19 with F
I

7) compute the particle surface force and torque based on Eq. 8.28
8) compute the particle collision force based on Eqs. 8.13-8.16
9) compute the new particle Lagrangian velocity and position by solving Eq. 8.10
Steps 1) and 6) can be completed numerically using the conventional incompressible flow
solver techniques of B.3.3, e.g. using Eqs. B.66. B.68 and B.69. Another option is to use the
explicit Lattice-Boltzmann techniques of B.7, e.g. Hlzer & Sommerfeld (2006) and Becker
et al. (2007). Note that steps 1, 4), 6) and 7) need not be computed inside of the particle where

I
=0 to save on total operations. Also note that, steps 2)-5) and 7)-9) must be repeated if there
are multiple particles.
A similar approach is the Smoothed-Profile Method (SPM) which also applies a force per
unit volume over a finite thickness interface but also within the particle surface. The SPM
employs the same PDE of Eq. 7.247 (used by the Force-Coupling Method of 7.3.2) for a
combined velocity field (V
m
) based on the net interphase force (F
int
). If the particle
acceleration is weak and there are no collisions, this is simply the effective gravitational force
of Eq. 7.249. To distribute this force in the flowfield, smoothed volume fraction profile can be
defined based on an interface thickness and the tanh function. For a sphere, this is given by:


( ) ( )
{ }
1
2 i p p i p I
x , x tanh 4 r / 1

= +

x x
8.29
This distribution yields: 0 when x
i
is more than
I
/2 outside of the particle surface, 1
when x
i
is more than
I
/2 inside of the particle surface, and 0<<1 inside of the interface. The
interphase force at an Eulerian node can then be computed (similar to Eq. 8.24) as


( )
n 1/ 2 free
I,i i I m,i
/ t
+
= V V F NSI
8.30
The PDEs can then be computed using a semi-implicit penalty method with a specific timestep
to minimize error given by
f I
t 0.18 = . An advantage of this technique is that the particle
velocity may be computed directly from the concentration field as:


f f
N N
p i m,i i
i 1 i 1
/
= =

=



v V
8.31
However, a disadvantage of this technique is that incorporation of particle inertia effects
requires that the interaction force be computed by Eq. 7.258 which is coupled to Eq. 8.31. This
coupling may lead to time-lag errors or require an implicit force-field which may require
iteration on the overall method for each timestep. Including collision effects would further
complicate the Smoothed Profile Method.


8.3. Continuous-Interface Methods for Fluid Particles

420
As with solid particles, the Continuous-Interface Method for fluid particles allows the
entire domain to be treated with a single set of transport equations with simple (and fast)
structured uniform meshes. In the case of fluid particles, the velocity within the particle
(including any recirculation) and particle deformation is also obtained by numerically
accounting for surface tension with a finite thickness interface. The thickness of this mixture
region is generally kept as small as numerically possible, e.g., a few grid cells wide.
Employing a continuous-interface for fluid particles allows incorporation of particle
recirculation and interface deformation. A general principle of the method is finite thickness
interface on which the fluid particle surface boundary conditions are effectively enforced. The
interface thickness perpendicular to the particle surface normal (n) and is scaled by the
Eulerian grid resolution (x). A second principle is that a single fluid velocity (V
m
) is defined
throughout the computational domain and equals the continuous-fluid velocity (U) outside of
the interface and the particle velocity (V) inside of the interface. If one defines x
I
as the local
surface coordinate in the direction of n and assumes that the interface thickness is centered on
the particle surface (Fig. 8.9a), the velocity field based on these two principles can be
expressed as


1
2 I I
m
1
2 I I
x
x
>
=

<

U
V
V
if
if

8.32
Similarly, P
m
will tend to P
f
outside the interface and P
p
inside the interface. A third principle
is that a volume of fluid () is used to quantify the volume of a particle portion that is
captured within a computational cell normalized by the volume of that same cell (Fig. 4.12b).
This quantity varies smoothly within the interface and has the following limits on either side:


1
2 I I
1
2 I I
0 x
1 x
>
=

<

if
if

8.33
As such, effectively serves as a smoothed heavy-side function. In this context, it has been
referred to as a marker function or a color function and can be used to describe transition
of fluid density and viscosity through the numerical interface:


( )
( )
m p f
m p f
1
1
= +
= +

8.34a

8.34b
A similar expression may written for thermal conductivity if heat transfer effects are important.
It is important to note that the mixed-fluid region in the vicinity of the interface is non-
physical since the actual interface is best represented as a discontinuity. The thickness is
purely employed for numerical convenience so that the fluid properties (density and viscosity)
can vary continuously so that the flow-field can be described by a single velocity field. The
thickness of this interface (
I
) is chosen to be as small as possible (to be close to the physical
situation) without giving rise to numerical instabilities (to allow differential numerical
schemes). Generally, this balance leads to an interface-thickness of about 2-4 grid cells. In the
next section, the transport equations are identified while the following three sections discuss
the primary numerical techniques to solve these equations.

8.3.1 Mixed-Fluid Transport Equations

Even though a variable density is used in the mixed region, if
f
and
p
are themselves
421
constant and there is no mass transfer then the density of a fluid mixture remains constant
along its fluid path. Thus, the continuity equation from Eq. 1.13 is given by:


m m
m m
0
t t

= + =

V
D
D
8.35
Based on the above assumption, one may employ an incompressible formulation for the V
m

transport equations of mass, momentum and energy. If one further assumes isothermal
conditions with no heat transfer, the energy PDE is not needed and the remaining PDEs are
simply:


m
0 = V
( )
m
m m m,i j
m
1
P K
t

= + +

V
V V g - + F
8.36a

8.36b
The second equation describes momentum transport of the mixed-fluid where K
m,ij
represents
the viscous stress tensor (Eq. A.9). The last term on the RHS represents the surface tension
force per unit mass of the mixed fluid where r
I
is the interface radius of curvature and n is the
outward surface normal. As such, the interface boundary conditions of Eq. 1.58 are implicitly
incorporated within the formulation. Some of the most common techniques for solving this
system are discussed in the next three sections and include: volume-of-fluid , level-set and
front-tracking methods.

8.3.2 Volume-of-Fluid Methods

The volume-of-fluid (VOF) method is perhaps the most common of the continuous-
interface methods for immiscible fluids. It employs a marker function to identify the fluid in
the computational cell as defined by Eq. 8.33. In the VOF technique, Eq. 8.35 can be restated
as an advection equation for this function as


m
0
t

+ =

V 8.37
Employing this equation allows the interface to move based on the mixed-fluid velocity
(instead of requiring a Lagrangian ODE as in Eq. 8.10) and allows computation of the local
density and viscosity (through Eq. 8.34). Furthermore, it can be used to determine the surface
normal and surface curvature based on the continuum surface method (Brackbill et al. 1992):


I
1
r

=
n
n

8.38a

8.38b
These can be used to determine the distributed surface tension force on the RHS of Eq. 8.36.


I m
r

n
F
8.39
To solve this system of equations, one may use the procedure described by Nichols et al.
(1978) for each time-step starting from the initial conditions:
1) compute the mixed-fluid velocity using an incompressible iterative technique
2) compute the VOF function () using its advection equation (Eq. 8.37)
3) update the distributions of density, viscosity, surface normal curvature
422
4) repeat steps 1)-3) until finial time (or convergence for steady flows)
The first and second steps require special care as discussed below.
For the first step, the strong spatial variations in density and viscosity make solution of the
incompressible momentum equation more challenging than those of 7.1.3. A common
approach to handle such variable-density flows using a variation of the conventional Marker-
and-Cell (MAC) approach discussed in B.3.3. If one adapts the form of Eqs. B.61-B.66 to Eq.
8.36b, the predicted velocity may be based on convective and non-convective updates as:


( ) ( )
( )
1 1 pred n n n 1 pred n
m m conv conv non conv non conv
2 2
conv m m
n
non conv m,i j
m
t 3 t
1
K



= + + +
=
= +

V V
V V
g -
F F F F
F
F F

8.39a

8.39b

8.39c
Similarly, the updated pressure and corrected velocity based on Eqs. B.667-68 may be
expressed as


pred n 1
m
n
m
n 1
n 1 pred
m m n
m
P
t
P
t
+
+
+

=


V
V V

8.40a

8.40b
As in B.3.3, the first of these equations requires an iteration but is substantially complicated
by the variable density effect so that conventional solvers are no longer appropriate. Instead,
non-linear elliptical solvers are employed (Wesseling, 2004).
For the second step, a low diffusion monotonic technique is needed to ensure that the
interface thickness remains reasonably bounded but that no undershoots or overshoots sin a
occur (which would be non-physical and may lead quickly to numerical instability). One may
use the non-linear schemes of B.3.4 which work well for shock discontinuities but such
schemes take advantage of inherent steepening of compressive waves into shock waves,
whereas no such physics are inherent in Eq. 8.37. Furthermore, the penalties for undershoots
and overshoots are much more severe for the interface treatment since the density ratios tend to
be much more exaggerated. Fortunately, the simplicity of Eq. 8.37 allows for a reconstruction
approach that allows for reasonably sharp monotonic interfaces. This is discussed in detail by
Tryggvason et al. (2007) and is briefly outlined below for a one-dimensional advection of .

Vous aimerez peut-être aussi