NUMERICAL SIMULATIONS OF UNSTEADY

FLOWS IN A PULSE DETONATION ENGINE

BY THE SPACE-TIME CONSERVATION ELEMENT
AND SOLUTION ELEMENT METHOD


DISSERTATION


Presented in Partial Fulfillment of the Requirements
for the Degree Doctor of Philosophy
in the Graduate School of The Ohio State University

By

Hao He, M.S.

*****


The Ohio State University

2006



Dissertation Committee:

Prof. S.-T. (John) Yu, Advisor
Prof. Walter R. Lempert
Prof. Mohammad Samimy
Prof. Ahmet Selamet
Approved by


_________________________
Advisor

Graduate Program in Mechanical Engineering






Copyright by
Hao He
2006

ii
ABSTRACT
This dissertation is focused on the development of a numerical framework for
time-accurate solutions of high-speed unsteady flows with specific applications to the
development of the Pulse Detonation Engine (PDE) concept. The numerical method
employed is the space-time Conservation Element and Solution Element (CESE) method,
which is a novel numerical method for time-accurate solutions of nonlinear hyperbolic
equations. As a part of the outcome of the present work, a general-purposed two- and
three-dimensional CESE code has been developed. The code can use both structured and
unstructured meshes composed of tetrahedrons or hexahedrons for three-dimensional
flows and quadrilaterals and triangles for two-dimensional flows. The code is fully
parallelized for large-scaled calculations. Parallel computing is based on domain
decomposition. Message Passing Interface (MPI) is used for data communication
between computer nodes. Extensive pre and post processing codes have been developed
to streamline the numerical calculations, so that the core solver could be directly
connected with commercial grid generators and solution animation software. The code
has been applied to analyze various flow fields related to the PDE concept. First,
numerical results of one-, two- and three-dimensional detonation waves are reported. The
chemical reactions were modeled by a one-step, finite-rate, irreversible global reaction.
The classical Zeldovich, von Neumann, and Doering (ZND) analytical solution was used

iii
to set up the initial conditions as well as for code validation. In the three-dimensional
calculations, detonations in square, round, and annular tubes at different sizes were
successfully simulated. Salient features of detonation waves were crisply resolved,
including triple points and numerical soot traces on walls. Second, as a promising
detonation initiation means, implosion with shock focusing was investigated. Polygonal
and circular converging shock fronts were simulated. In two-dimensional calculations, we
found a double-implosion mechanism in a successful detonation initiation process. Third,
the plume dynamics of a PDE fueled by propane/air mixtures were studied to support the
prototype development at NASA Glenn Research Center (GRC). Numerical results show
that in each PDE cycle the engine is actively producing thrust forces only in about 6% of
one cycle time period. The rest of the time is occupied by the blow-down and refueling
processes. Since the PDE tube is always open, the processes depend on the flow
conditions outside the PDE tube. In the near-field plume, complex shock/shock and
shock/vortex interactions were found. In the far field, a spherical expansion wave is the
dominant flow feature. This dissertation work is synergy of a very accurate and efficient
CFD method, i.e., the CESE method, and the modern parallel computing technology. The
resultant software is a state-of-the-art numerical framework, ready to be applied to obtain
time-accurate solutions of hyperbolic equations. This approach of applying the CESE
method and parallel computing could point to a new direction for high-fidelity
simulations of complex flow fields of advanced propulsion systems.


iv






Dedicated to my parents

v
ACKNOWLEDGMENTS
I would like to express my sincere gratitude to my advisor, Professor Sheng-Tao
John Yu, for his patience, guidance and encouragement. My thanks go to Professor
Walter R. Lempert, Professor Mohammad Samimy, and Professor Ahmet Selamet for
serving as the members of my dissertation committee.
I would like to express my appreciation to Dr. Zeng-Chan Zhang and Dr. Moujin
Zhang for the instructive discussions regarding numerical methods and algorithm. I also
wish to thank Bao Wang and Minghao Cai for their help on maintaining the computer
facilities.

vi
VITA
November, 1973Born in China
July, 1995..B.E., Engineering Mechanics, Tsinghua
University, Beijing, China
May, 2003..M.S., Mechanical Engineering, Wayne State
University, Detroit, USA
1995-1999..Staff Member, Department of Engineering
Mechanics, Tsinghua University, Beijing, China.
1999-2003.Graduate Research Assistant, Department of
Mechanical Engineering, Wayne State University,
Detroit, USA
2003-present..Graduate Research Assistant, Department of
Mechanical Engineering, The Ohio State
University
PUBLICATIONS
Bao Wang, Hao He, and S.-T. John Yu, 2005, Direct Calculation of Wave Implosion for
Detonation Initiation in Pulsed Detonation Engines, AIAA Journal, Vol. 43, No. 10, pp.
2157-2169.
FIELDS OF STUDY
Major Field: Mechanical Engineering


vii
TABLE OF CONTENTS
Page
Abstract ....................................................................................................................... iii
Dedication................................................................................................................... iiv
Acknowledgments..........................................................................................................v
Vita....................................................................................................................... vi
List of Tables ............................................................................................................... ix
List of Figures................................................................................................................x
Nomenclature............................................................................................................ xvii
Chapters:
1. Introduction................................................................................................................1
1.1 Literature Review.............................................................................................1
1.2 Motivation and Objectives...............................................................................4
1.3 Numerical Methods..........................................................................................6
1.4 Organization.....................................................................................................8
2. The Model Equations.................................................................................................9
2.1 Three-Dimensional Euler Equations................................................................9
2.2 Euler Equations for Reacting Flows ..............................................................11
2.3 The ZND Solution..........................................................................................14
3. The Space-Time CESE Method...............................................................................27
3.1 Conventional Finite Volume Method ............................................................26
3.2 The CESE Method.........................................................................................32
3.3 The Modified CESE Method .........................................................................40
3.4 Two-Dimensional Euler Solver .....................................................................47
3.4.1 Conservation Elements and Solution Elements ..................................48
3.4.2 Approximations within a Solution Element........................................51
3.4.3 Evaluation of u
m
..................................................................................53
3.4.4 Evaluation of u
mx
and u
my
....................................................................56

viii
3.5 Three-Dimensional Euler Solver ...................................................................59
3.5.1 Conservation Elements and Solution Elements ..................................60
3.5.2 Approximations within a Solution Element........................................64
3.5.3 Evaluation of u
m
..................................................................................66
3.5.4 Evaluation of u
mx
, u
my
and u
mz
.............................................................69
4. Parallel Computation ...............................................................................................72
4.1 Introduction....................................................................................................72
4.2 Hardware System...........................................................................................78
4.3 Software and Programming ...........................................................................84
4.4 System Specifications ....................................................................................88
4.5 Useful Links...................................................................................................99
5. Detonations ............................................................................................................100
5.1 Introduction..................................................................................................100
5.2 One-Dimensional Detonations.....................................................................103
5.3 Two-Dimensional Detonation Waves..........................................................110
5.4 Three-Dimensional Detonation Waves in a Duct ........................................111
5.5 Detonations in Circular and Annular Tubes ................................................130
6. Implosion and Explosion .......................................................................................146
6.1 Introduction..................................................................................................145
6.2 Two-Dimensional Implosion and Explosion ...............................................150
6.3 Three-Dimensional Implosion and Explosion .............................................163
7. Pulse Detonation Engine Aeroacoustics ................................................................172
7.1 Introduction..................................................................................................172
7.2 Initial Condition of Detonation Wave..........................................................178
7.3 Near Field Calculations................................................................................181
7.4 Far Field Calculations..................................................................................193
7.5 Concluding Remarks....................................................................................202
8. Conclusions............................................................................................................203
8.1 Achievements and Findings.........................................................................203
8.2 Recommendations for Future Work.............................................................207
Bibliography ..............................................................................................................211

ix
LIST OF TABLES
Table Page
4.1 Parallel computing performance...................................................................86
4.2 Specifications of the1
st
generation cluster in 1999 .......................................90
4.3 Specifications of the 2
nd
generation cluster in 2000 .....................................92
4.4 Specifications of the 3
rd
generation cluster in 2002......................................95
4.5 Specifications of the 4
th
generation cluster in 2003......................................97

x
LIST OF FIGURES
Figure Page
2.1 A schematic of the ZND detonation wave....................................................15
2.2 Diagram of the Rayleigh lines and Hugoniot curves with different ..........21
2.3 Flow variable profiles of a one-dimensional ZND detonation wave from
the analytical solution: (a) mass fraction of reactant; (b) pressure; (c)
density, and (d) velocity. Parameters: = 1.2, q
0
= 50, E
+
= 50, and f =
1.6..................................................................................................................26
3.1 A schematic of the space-time integral .........................................................29
3.2 Space-time meshes by the conventional finite volume methods in (a) one
and (b) two spatial dimensions .....................................................................30
3.3 Space-time geometry of the conventional finite volume methods in E
2
.......31
3.4 Schematics of the CESE method in one spatial dimension. (a)
Zigzagging SEs. (b) Integration over CE to solve u and u
x
at the new
time level.......................................................................................................34
3.5 CE
+
and CE
-
for solving u and u
x
at the new time level in the a- scheme ..38
3.6 Grid points in the x-y plane for the original CESE method..........................39
3.7 Schematics of the modified CESE method in one spatial dimension: (a)
the staggered space-time mesh, (b) SE (j, n), shown as the yellow part,
and CE (j, n)..................................................................................................45
3.8 The space-time mesh in two spatial dimensions: (a) grid points in the x-y
plane, (b) SE and CE for the two-dimensional scheme ................................49
3.9 Spatial translation of the quadrilateral
* * * *
1 2 3 4
A A A A .......................................57
3.10 A schematic of the 3D spatial mesh in the x-y-z space.................................60

xi
4.1 A schematic of a Beowulf cluster .................................................................73
4.2 Domain decomposition examples: (a) mesh for a simple three-
dimensional domain of a box, (b) tow-dimensional mesh for a
flow-over-cylinder problem, (c) a three-dimensional mesh of an annulus
tube; different colors represent different sub-domains .................................76
4.3 Network performance of two gigabit Ethernet systems with Intels
traditional x86 architecture and CSA, respectively, and a fast Ethernet
system with single, double and triple fast Ethernet network interface
cards (NICs) installed on each computer node, where the channel
bonding technology is used for the multiple card cases. Benchmarked by
netperf 2.1 .....................................................................................................81
4.4 A schematic of channel bonding technology applied on a Fast Ethernet
system with dual network cards used on each node .....................................82
4.5 A schematic of the speedup ..........................................................................86
4.6 The 1
st
generation cluster in 1999.................................................................89
4.7 The 2
nd
generation cluster in 2000................................................................91
4.8 The 3
rd
generation cluster in 2002 ................................................................93
4.9 The 4
th
generation cluster in 2003 in OSU ...................................................96
5.1 A schematic of the piston problem.............................................................104
5.2 Dimensionless shock front pressure history in the piston problem for
2 . 1 = , 50
0
= q , 50
+
= E , and (a) f = 2.0, (b) f = 1.6..............................105
5.3 Dimensionless time history of the shock front pressure in the instability
problems for 2 . 1 = , 50
0
= q , 50 =
+
E , and (a) f = 1.8, (b) f = 1.6,
with the mesh resolution 50 pts/L
1/2
............................................................107
5.4 The dimensionless p-v diagram of the detonation for parameters 2 . 1 = ,
50
0
= q , 50 =
+
E and f = 1.8....................................................................108
5.5 Variation of the dimensionless peak pressure with mesh resolution for
various schemes. The relative mesh spacing is defined as 5/n with n as
the number of mesh nodes for the half-reaction zone. Flow parameters:
2 . 1 = , 50
0
= q , 50 =
+
E , and f = 1.6.....................................................109

xii
5.6 Two-dimensional detonation waves: (a) mass fraction of reactant, (b)
pressure, (c) vorticity and (d) temperature. Flow parameters: f = 1.6, =
1.2, q
0
= 50, E
+
= 50....................................................................................111
5.7 Snapshots of pressure contour at different time for a detonation wave
propagating in a square duct: (a) t = 12.0, (b) t = 13.5. The cross section
size is 88. The unit length is the half reaction length. Flow parameters:
f = 1.6, = 1.2, q
0
= 50, E
+
= 50 .................................................................113
5.8 Snapshots of different contour for a detonation wave propagating in a
square duct: (a) temperature, (b) concentration of the reactant. t = 13.5.
Cross section size: 88. Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
=
50.................................................................................................................114
5.9 Snapshots of pressure contour on a cross section near the shock front of
detonation waves propagating in a square duct: (a),(b),(c) rectangular
mode; (d),(e),(f) diagonal mode. Cross section size: 88. Flow
parameters: f = 1.6, = 1.2, q
0
= 50, E
+
= 50..............................................117
5.10 Numerical soot trace on the side wall: (a) rectangular mode, (b) under-
expanded diagonal mode, (c) fully developed diagonal mode with four
heads. Cross section size: 88 for cases (a) and (b), 2020 for case (c).
Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
= 50.....................................119
5.11 Snapshots at different time: (a), (b), (c) pressure iso-surfaces; (d), (e), (f)
pressure contour on a cross section near the wave front. Cross section
size: 4040. Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
= 50 ...............120
5.12 Numerical soot trace on one side wall. Cross section size: 4040. Flow
parameters: f = 1, = 1.2, E
+
= 35, (a) q
0
= 30; (b) q
0
= 40; (c) q
0
= 60.....122
5.13 Numerical soot trace on one side wall. Cross section size: 4040. Flow
parameters: f = 1, = 1.2, q
0
= 50, (a) E
+
= 20; (b) E
+
= 40; (c) E
+
= 45;
(d) E
+
= 50...................................................................................................123
5.14 Numerical soot trace on one side wall. Cross section size: 4040. The
reactants are H
2
-Air mixtures at: (a) = 0.7; (b) = 0.8; (c) = 1.0; (d)
= 1.4 ........................................................................................................123
5.15 Cell width versus equivalence ratio for detonations in the H
2
-Air
mixtures. Cell width is normalized based on the minimal
numerical/experimental cell width obtained at the stoichiometric
condition ( = 1.0) .....................................................................................127

xiii
5.16 Numerical soot trace of the two-dimensional calculation of a detonation
propagating in a H
2
-Air mixture at = 1.4................................................127
5.17 Time history of the pressure peaks at the shock front for two- and
three-dimensional calculations of a CJ detonation wave in a H
2
-Air
mixture at = 1.4. The peak pressure and time are both dimensionless...129
5.18 The TDW combustor geometry and the set-up...........................................132
5.19 A schematic of the mesh used for the three-dimensional calculations on
spinning detonation in a circular tube. The tube radius is 2. The unit
length is the half reaction length.................................................................133
5.20 Calculated soot trace on the tube wall of a single head spinning detonation
wave. The tube radius is 2. The unit length is the half reaction length.
Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50 ....................................135
5.21 A schematic of a single-head detonation wave and comparison between
the experimental picture and the numerical result. Tube radius: 2. Flow
parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50..............................................135
5.22 A schematic of the single-head spinning detonation in a rotating frame....136
5.23 Pressure iso-surfaces of a spinning detonation at different time. Tube
radius: 2. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50 ....................138
5.24 Snapshots of pressure contour on the sliced x-z and x-y planes at
different time. Tube radius: 2. Flow parameters: f = 1.2, = 1.2, q
0
= 50,
E
+
= 50. .......................................................................................................139
5.25 Snapshots of pressure contour on the tube wall of a spinning detonation
at different time. Tube radius: 2. Flow parameters: f = 1.2, = 1.2, q
0
=
50, E
+
= 50..................................................................................................140
5.26 Calculated multi-headed detonation wave in a round tube: (a) pressure
iso-surfaces; (b) pressure contour on the sliced y-z plane; (c) pressure
contour on the sliced x-y plane. Tube radius: 6. Flow parameters: f = 1.2,
= 1.2, q
0
= 50, E
+
= 50..............................................................................142
5.27 Calculated soot trace on the tube wall of a multi-headed detonation wave.
Tube radius: 6. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50 ...........143

xiv
5.28 A schematic of the mesh for the 3D calculations on detonations in an
annular tube. The external and internal radius is 6 and 3, respectively. The
unit length is the half reaction length..........................................................143
5.29 Pressure iso-surfaces of an annular detonation at different time. External
and internal radius: 6 and 3. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+

= 50 .............................................................................................................144
5.30 Calculated pressure contour on the y-z and x-z planes of the multi-
headed detonation wave in an annular chamber. External and internal
radius: 6 and 3. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50 ...........145
5.31 Numerical soot trace on the outer and inner tube walls of the multi-headed
detonation wave in an annular chamber. External and internal radius: 6
and 3. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50..........................145
6.1 A schematic of the mesh for 2D implosion/explosion calculations............151
6.2 Snapshots of pressure contour of a square front implosion at time = 0,
0.1, 0.2, 0.25, 0.3, 0.4, 0.5, 0.65, and 0.8. Domain radius: 1.
Circumradius of the square front: 0.707, i.e., edge length 1.0. P
out
/P
in
=
20.
out
/
in
= 20............................................................................................153
6.3 Snapshots of pressure contour of an octagonal front implosion at time =
0, 0.09, 0.18, 0.27, 0.315, 0.36, 0.495, 0.63, and 0.72. Domain radius: 1.
Circumradius of the octagonal front: 0.7. P
out
/P
in
= 20.
out
/
in
= 20 ..........154
6.4 Snapshots of pressure contour of a circular front implosion at time = 0,
0.09, 0.18, 0.27, 0.36, 0.45, 0.585, 0.72, 0.90, respectively. Domain
radius: 1. Radius of the circular front: 0.667. P
out
/P
in
= 20.
out
/
in
= 20.....155
6.5 Snapshots of pressure contour of a circular front implosion with
chemical reaction at time = 0, 0.09, 0.18, 0.27, 0.315, 0.36, 0.405, 0.54,
0.90, respectively. Domain radius: 1. Radius of the circular front: 0.667.
P
out
/P
in
= 20.
out
/
in
= 20. Detonation parameters: f = 1, = 1.4, q
0
= 50,
E
+
= 20 ........................................................................................................156
6.6 The evolution of the polygonal shock. In each polygonal section, the
flow field is a planar shock wave entering a converging channel ..............157
6.7 Dimensionless time histories of (a) temperature, and (b) pressure, at the
focal point for the three non-reacting cases. P
out
/P
in
= 20.
out
/
in
= 20. .....161

xv
6.8 Dimensional time histories of (a) temperature, and (b) pressure, at the
focal point for the imploding reacting flow. P
out
/P
in
= 20.
out
/
in
= 20.
Detonation parameters: f = 1, = 1.4, q
0
= 50, E
+
= 20 ..............................162
6.9 A schematic of the mesh for 3D implosion/explosion problems................164
6.10 Snapshots of pressure contour on the sliced planes x = 2 and y = 0 of an
octagonal front implosion at time = 0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35,
0.5, 0.75, 1.25. Tube radius: 1. Circumradius of the octagonal front: 0.75.
P
out
/P
in
= 20.
out
/
in
= 20.............................................................................167
6.11 Snapshots of pressure contour on the sliced planes x = 4 and y = 0 of an
octagonal front implosion at time = 0, 0.24, 0.4, 0.56, 0.8, 1.8, 2.8, 4.0.
Tube radius: 2. Circumradius of the octagonal front: 0.3. P
out
/P
in
= 20.

out
/
in
= 20..................................................................................................169
6.12 Dimensionless time histories of (a) temperature, and (b) pressure, at the
focal point for the three cases with different initial condition and mesh
size. Tube radius: 2. P
out
/P
in
= 20.
out
/
in
= 20............................................171
7.1 A typical PDE cycle....................................................................................173
7.2 A schematic of setting up the initial condition ...........................................179
7.3 A schematic of computational domain for the near-field calculations .......181
7.4 A schematic of the sampling point distribution in the computational
domain for near field calculations. Domain size: 12.5 ft in length, 4.5 ft
in radius.......................................................................................................183
7.5 Time history of pressure at the closed end of the PDE tube.......................184
7.6 Three snapshots of pressure contour of a PDE plume at different time of
the initial stage of expansion. Domain size: 12.5 ft in length, 4.5 ft in
radius. t = 0.642, 1.124, and 1.927 ms, respectively...................................186
7.7 Vortex-shock interaction in the near field of a PDE plume........................188
7.8 Time histories of pressure at two checking points located at (a) (5.66,
5.875) and (b) (22.64, 5.875).................................................................189
7.9 Time histories of pressure and density at the location (5.66 in, 0.49 ft)
for a three-cycle calculation........................................................................190

xvi
7.10 Calculated wave characteristics at sampling points along L-1, L-2, L-3
and L-4: (a) peak pressure, and (b) the wave speed. The experimental
data is measured on L-1 only. .....................................................................192
7.11 A schematic of the computational domain and the distribution of
sampling points for far field calculations. Domain size: 57 ft in length,
30 ft in radius ..............................................................................................194
7.12 Three snapshots of pressure contour of a PDE plume at different time in
the far field. Domain size: 57 ft in length, 30 ft in radius...........................196
7.13 Time histories of pressure at various locations on the circles at (a) 10 ft
and (b) 50 ft away from the PDE tube exit .................................................198
7.14 Amplitude of pressure fluctuations along the central line at different
distance from the PDE tube exit .................................................................200
7.15 Sound strength (in dB) at different distance from the PDE tube exit .........201

xvii
NOMENCLATURE
Variables and Constants:
C
p
specific heat at constant pressure
C
v
specific heat at constant volume
D detonation wave velocity
E total energy
E
+
activation energy
e specific internal energy
e
i
sensible internal energy
e
c
chemical potential
f overdriven factor
f, f
m
flux component in x direction
g
m
flux component in y direction
h space-time density flux
h specific enthalpy
h
f
heat of formation
h
m
space-time density flux component
K pre-exponential factor of the Arrhenius kinetics
K
f
forward reaction rate
L
1/2
half reaction length
M Mach number
M
w
molecular weight
n normal vector of space-time surface
P, p pressure
q
0
heat release due to chemical reaction
q
m
flux component in z direction
R gas constant
R
u
universal gas constant
R, r radius

xviii
S surface of space-time volume
s surface element of space-time volume
T temperature
t time
u x-component of velocity
u, u
m
conserved flow variables
V space-time volume
v y-component of velocity
v specific volume (1/)
Z mass fraction of the reactant
specific heat ratio
mass fraction of the product

m
source term
density
area of space-time surface
equivalence ratio
source term due to chemical reactions
Subscripts:
0 unburned state of gas mixture
in inside value
j temporal index of space-time mesh grid
n spatial index of space-time mesh grid
out outside value
ref reference value
CJ Chapman-Jouguet variable
Superscripts:
* dimensionless variable


1
CHAPTER 1
INTRODUCTION
1.1 Literature Review
The Pulsed Detonation Engine (PDE) is a novel propulsion concept based on the
use of the high-pressure gases generated by repetitive detonation waves for thrust forces
(Kailasanath, 2001; Kailasanath et al., 1999; Li et al., 2000; Ebrahimi et al., 2000;
Schauer et al., 2001). For a fly regime of Mach 2 to 4, a PDE could improve vehicle
performance and cost effectiveness as compared with conventional air-breathing ramjet
engines. Though not yet implemented on any operational aircraft, there has been much
work done on the PDE related concepts during the past decades.
Among many technical issues about the PDE concept, it was recognized that
successful development of a PDE lies in the development of robust and repeatable
detonation initiation processes with minimum use of sensitized gases, e.g., oxygen, and
energy deposition in the forms of plasma and/or electric arcs. In the past, Nicholls et al.
(1957), Krzycki (1962), and Wortman et al. (1992) had difficulties in achieving reliable

2
detonation initiation by using relatively insensitive fuel/air mixtures. Helman et al. (1986),
however, successfully achieved detonation in an ethylene/air mixture by first detonating
an ethylene/oxygen mixture inside a smaller tube, i.e., a pre-detonator or initiator. The
blast wave from the initiator would propagate into the main combustor chamber to
initiate the detonation wave for propulsion. Aided by a similar pre-detonator, Hinkey et al.
(1995) developed a hydrogen-fueled PDE. Nevertheless, reliable detonation initiation
process has been one of major stumbling blocks in the further development of the PDE
concept.
In addition to the detonation initiation process, it is also important to study the
aero acoustics of the PDE plume. It has been envisioned that a PDE would generate thrust
force by releasing a series of detonation waves from the detonation chamber. The
propagating blast wave at a frequency of 30 to 100 Hz would generate an intensive
acoustic field around the PDE. In the near field, theses blast waves could impose
significant vibrations to the fuselage of the vehicle. In the far field, the acoustic signature
of the vehicle could be of concern. Moreover, the developing blast waves in the near field
could also impact the fluid flow inside the detonation chamber. As a result, the PDE
cycle, i.e., purging, fueling, detonation initiation, and blow down processes, could be
influenced.

3
As another subject, researchers are now investigating the possibility of developing
a continuous detonation engine by using a spinning detonation wave around an annular
chamber. Since it is a continuous detonation wave, no repetitive detonation initiation
processes, as that in a typical PDE, is needed. A group of Russian scientists at the
Lavrentev Institute of Hydrodynamics, Siberian Division, Russian Academy of Sciences,
Novosibirsk, have been developing a novel detonation-based annular combustor since
late 50s (Mikhailov and Topchiyan, 1965; Nikolaev and Topchiyan, 1977;
Bykovskii and Mitrofanov, 1980, 2000; Bykovskii and Vedernikov, 1996). They
considered combustion of acetylene-oxygen mixtures in annular chambers with
constriction of the exit cross section. Experimental observation indicated that
conventional turbulent subsonic combustion of the mixture prevailed in a large part of the
annular combustor. However, further investigation showed distinct wave structure
occurred at the leading edge of the reaction zone and the combustion mode was actually
related to a spinning detonation wave along the circumferences of the annulus. The
reaction zone was apparently continuously initiated by the rotating detonation wave with
velocities close to the speed of the sound of the combustion products, i.e., the CJ velocity
of the detonation for the given fuel/air mixture. In the past decades, they reported various
flow geometries and operational conditions of the above-mentioned transverse detonation
waves (TDW) in annular cylindrical chambers. Combustion of both gaseous and
two-phase propellants was considered in their works.

4
The most remarkable feature of these combustors is that the axial flow was
subsonic while the detonation waves traveled tangentially in the azimuthal
direction in the annular chamber. It is imperative to understand why the detonation
wave would be confined to spin at a near CJ velocity only in the azimuthal
direction without progressing upstream in the axial direction.
1.2 Motivation and Objectives
To fly higher and faster and even enter the earth orbit, the advanced propulsion
devices are required. Detonation wave could be used in the further development of
advanced propulsion systems. Computational Fluid Dynamics (CFD) has been playing an
important role in aerospace engineering for years. This Ph.D. research is a continuation of
these efforts in applying the state-of-the-art numerical method to modeling detonating
flows in the development of the envisioned PDE propulsion systems. The goals of the
present dissertation are twofold:
(i) By combining a high-fidelity CFD method and the modern parallel
computing technology, we plan to develop a general numerical
framework to provide high-fidelity numerical solutions of unsteady
flows.
(ii) We plan to use this numerical tool to conduct detailed simulation of
detonation waves and detonation initiation processes.

5
Specific contribution by this research work include
(i) The extension of the space-time Conservation Element and Solution
Element (CESE) method for two- and three-dimensional chemically
reacting flows.
(ii) Quadrilateral cells and hexahedral cells will be used for two- and
three-dimensional calculations, respectively.
(iii) The solver supports parallel computing for fine-mesh resolution in
large spatial domains in both two and three spatial dimensions.
(iv) Thorough code validation by comparing the CFD results with the
classical ZND solutions, and in-depth studies of the aerodynamic
structure of the three-dimensional detonation waves.
(v) Detailed simulation of the PDE plume dynamics for acoustic
distribution, which is pertinent to a prototype PDE being tested at the
NASA Glenn Research Center (GRC);
(vi) Study of the detonation initiation process by using the
implosion/explosion mechanism, which will play a key role in an
operational PDE in the future.


6
1.3 Numerical Methods
The CESE method is a novel numerical scheme for nonlinear hyperbolic
conservation laws. Developed by Chang (1995) and coworkers (Chang et al., 1999; Yu
and Chang, 1997; Wang and Chang, 1999; Zhang et al., 1999, 2002), the tenet of the
CESE method is a unified treatment of space and time in calculating flux conservation.
Contrast to modern upwind schemes, the method does not use a Riemann solver or a
reconstruction procedure as the building blocks. Nevertheless, the shock capturing
capabilities of the CESE method is superb. In particular, the CESE method is capable of
capturing shocks and acoustic waves simultaneously, while the magnitude of the pressure
jump across the shock wave is several orders in magnitude greater than the averaged
amplitude of pressure fluctuations in the acoustic waves.
In the setting of the CESE method, the space-time domain is divided into
non-overlapping Conservation Elements (CEs), in which the space-time flux conservation
is imposed by integrating the conservation laws in a novel space-time integral form.
Inside a CE, the flow solution could be discontinuous. Since the summation of the
non-overlapping CEs is equal to the whole space-time domain, flux conservation is
enforced locally and globally. To facilitate the integration over CEs, non-overlapping
Solution Elements (SEs) are also defined in the space-time domain. Inside each SE, flow
variables and fluxes are assumed to be continuous and are discretized by a predetermined

7
function. No flow discontinuity is allowed inside a SE. The flow discontinuity only
occurs across the boundary of neighboring SEs. In general, SEs do not coincide with CEs.
In this work, the flow solution in each SE is discretized by the first-order Taylor series
expansion. Thus, a linear distribution of the flow variables is assumed in the SEs, and the
scheme is second-order accurate in both space and time.
To date, numerous highly accurate solutions have been obtained by the CESE
method, including cavitations (Qin et al., 2001), aero acoustics of supersonic jets (Loh et
al., 1995), turbulent flows with dense sprays (Im et al., 2004), MHD flows (Zhang et al.,
2003), and flows with complex shock systems (Wang et al., 1998). Previously,
applications of the CESE method to detonations were carried out by Park et al. (1999).
Important detonation wave features were reported, including the values of the peak
pressures and the oscillation frequencies of certain one- and two-dimensional detonation
waves. Further investigation of two- and three-dimensional detonations by the CESE
method was reported by Zhang et al. (2001) and He et al. (2001). Recently, Im et al.
(2002) and Wang et al. (2004) applied the CESE method to calculate certain detonation
initiation processes with detailed finite-rate chemistry models. They reported detonation
initiation results by shock reflection in conventional shock tubes, and spherical and
cylindrical implosion/explosion processes. Chemical reactions of H
2
/O
2
/Ar and C
3
H
8
/air
gas mixtures were simulated by finite-rate models.

8
In the present numerical framework, parallel computing is an important part,
because large-sized fine meshes require huge computer resources, which may exceed the
power of a single PC or a workstation. In this research work, parallel computing has been
preformed by using a Beowulf computer cluster. In the numerical method, parallel
computing is realized by domain decomposition. The mesh is first partitioned either
manually or by using software, e.g., METIS. In the calculations, the data transmission
between the computer nodes was done by using Message Passing Interface (MPI).
1.4 Organization
This dissertation is organized as follows: Chapter 2 presents the theoretical model
equations for detonations. Chapter 3 illustrates the CESE method. Chapter 4 introduces
the parallel computing technology and its implementation for the present work. Chapter 5
addresses the studies on detonation waves in one, two, and three spatial dimensions.
Chapter 6 discusses the two- and three-dimensional implosion/explosion processes used
for detonation initiation and transition. Chapter 7 presents the results for PDE plume aero
acoustics simulation. At the end, the effectiveness and limitation of the present research
is discussed and the conclusions are provided.


9
CHAPTER 2
THE MODEL EQUATIONS
2.1 Three-Dimensional Euler Equations
The three-dimensional Euler equations for can be expressed in the following vector
form:
0
m m m m
u f g q
t x y z

+ + + =

, (2.1)
where m =1, 2, 3, 4, 5, indicating the continuity equation, the three momentum equations
along the three Cartesian coordinates, and the energy equation, respectively. u
m
is the
flow variables, f
m
, g
m
, q
m
are the components of the flux vectors:
[ ]
m
u
u v
w
E

=




,
2
[ ]
( )
m
u
u p
f uv
uw
E p u

+

=



+

,
2
[ ]
( )
m
v
uv
g v p
vw
E p v

= +



+

,
2
[ ]
( )
m
w
wu
q wv
w p
E p w

=

+


+

,
(2.2)
where is the density, u, v, w are the three velocity components, and p is the pressure.

10
The specific total energy E is defined as
2 2 2 2 2 2
1 1
( ) ( )
2 ( 1) 2
p
E e u v w u v w
-
= + + + = + + +
, (2.3)
where e is the specific internal energy, and = C
p
/C
v
is the specific heat ratio.
To proceed, the above equations are non-dimensionalized based on a reference
state, denoted by a subscript 0. Usually the standard state or the freestream conditions
are taken. A reference velocity is defined as
0
RT , which resembles the speed of the
sound in the reference state. Furthermore, a reference length x
0
is needed. It is usually
defined as one of the geometric characteristic length, e.g., the diameter of the hole for a
free jet problem. The specific total energy E and the specific internal energy e are made
dimensionless by RT
0
.
With the above definitions of the reference state and the space and time scales, the
dimensionless variables (denoted by a bar) are defined as follows:
0

=
,
0
p
p
p =
,
0
0
0
p
RT
T
T
T

= =
,
0
RT
u
u =
,
0
RT
v
v =
,
0
RT
w
w =
,
0
x
x
x =
,
0 0
/ RT x
t
t =
,
0
RT
E
E =
. (2.4)
By substituting these dimensionless variables into the governing equations, one obtains a
set of dimensionless Euler equations, which remain unchanged in the form as compared
with Eqs. (2.1)-(2.3).

11
2.2 Euler Equations for Reacting Flows
Compared with the original form of the Euler equations, the Euler equations for
reacting flows have an additional equation for chemical reactions. The three-dimensional
version can be expressed in the following vector form:
m
m m m m
z
q
y
g
x
f
t
u
=

, (2.5)
where m =1, 2, 3, 4, 5 and 6, indicating the continuity equation, the three momentum
equations along the three Cartesian coordinates, the energy equation, and the species
equation, respectively. There is one more term shown in u
m
, f
m
, g
m
, q
m
, and
m
:

=
Z
E
w
v
u
u
m

] [ ,

+
+
=
uZ
u p E
uw
uv
p u
u
f
m

) (
] [
2
,

+
+
=
vZ
v p E
vw
p v
uv
v
g
m

) (
] [
2
,

+
+
=
wZ
w p E
p w
wv
wu
w
q
m

) (
] [
2
,

0
0
0
0
0
] [
m
, (2.6)
where Z is the mass fraction of the reactant. The specific total energy E is defined as
2 2 2 2 2 2
0 0
1 1
( ) ( )
2 ( 1) 2
p
E e Zq u v w Zq u v w
-
= + + + + = + + + +
, (2.7)

12
where q
0
is the heat release due to chemical reactions. To close the equation set, the gas
mixture is assumed polytropic, i.e., (i) the mole number of the reacting gas mixture is a
constant; (ii) the molecular weight of the reacting gas mixture is a constant; (iii) the gas
mixture is ideal, and (iv) the specific heats C
p
and C
v
of the gas mixture are constants. For
an ideal gas, p = RT, where T is the temperature and R = R
u
/M
w
is the gas constant with
R
u
as the universal gas constant and M
w
the molecular weight of the gas mixture. To
model the chemical reactions, a one-step irreversible reaction is employed. The source
term in the species equation due to chemical reaction in Eq. (2.6) can be written as
Z T R E K
u
) / exp(
+
=
, (2.8)
where K is the Arrhenius coefficient and E
+
is the activation energy. The species equation
and the continuity equation will enforce mass conservation of the reactant, the product,
and the reacting gas mixture.
In the present research, the above equations and their two-dimensional version are
used to simulate two- and three-dimensional detonations and implosion/explosion
processes.
To model PDE plumes, the axis-symmetric version of the above equations is
employed:
m
m m m
y
g
x
f
t
u
=

, (2.9)

13
where the flow variables u
m
, the flux f
m
, g
m
, and the source term
m
are

=
Z
E
v
u
u
m

] [ ,

+
+
=
uZ
u p E
uv
p u
u
f
m

) (
] [
2
,

+
+ =
vZ
v p E
p v
uv
v
g
m

) (
] [
2
,

+
+

y vZ
y v p E
y v
y uv
y v
m
/
/ ) (
/
/
/
] [
2
. (2.10)
To proceed, the above equations are made dimensionless and the state of the
unburned gas is taken as the reference state. As to the reference length, to calculate the
ZND detonations, we let x
0
= L
1/2
, the half- reaction length, defined as the distance from
the shock front to the point where a half of the reactant is consumed by the combustion.
For the PDE plume calculations, we let x
0
= R, where R is the radius of the thruster tube.
The heat release q
0
are made dimensionless by RT
0
. The activation energy E
+
is made
dimensionless by R
u
T
0
.
Thus, in addition to Eqs. (2.4), we have the following definition of the
dimensionless variables (denoted by a bar):
0 0
/ x RT
K
K =
,
0
T R
E
E
u
+
+
=
,
0
0
0
RT
q
q =
. (2.11)

14
By substituting these dimensionless variables into the governing equations, a set of
dimensionless Euler equations is obtained, which remain unchanged in the form as
compared with Eqs. (2.5)-(2.6), or (2.9)-(2.10) for the PDE simulations, except the source
term :
) / exp( T E Z K
+
=
. (2.12)
Note that the universal gas constant R
u
in the exponential function has been absorbed into
the dimensionless activation energy
+
E . Since the derived dimensionless equations are
identical to their dimensional counterparts except Eq. (2.12), the bar for dimensionless
variables will be dropped in the following discussions.
2.3 The ZND Solution
Although the classical Zeldovich, von Neumann, and Doering (ZND) solution has
been widely cited, I have not been able to find the complete derivation in the literature.
As a part of this dissertation work, the complete ZND solution has been carefully derived
based on incomplete information and hints provided in the open literature. The successful
use of the classical ZND solution for code validation and as the initial condition is critical
to the present Ph.D. research.
In the ZND model for detonation, we assume that (i) the flow is one dimensional;
(ii) transport effects, i.e., heat conduction, radiation, diffusion, and viscosity, are

15
neglected; (iii) the reaction rate is null ahead of the shocks and at a finite rate
immediately behind the shock; (iv) chemical reactions are modeled by a one-step,
irreversible reaction step at a finite rate; and (v) all thermodynamic variables other than
the chemical composition are in local thermodynamic equilibrium. Refer to Fig. 2.1 for a
schematic of the ZND detonation wave.

Figure 2.1: A schematic of the ZND detonation wave.

16
Consider the one-dimensional Euler equations for chemically reacting flow in the
dimensional form,
m
m m
x
f
t
u
=

. (2.13)
In a vector form, the flow variables and fluxes are

=
Z
E
u
u
m

] [ ,

+
+
=
uZ
u p E
p u
u
f
m

) (
] [
2
,

0
0
0
] [
m
, (2.14)
where the total energy 2 /
2
u e E + = and internal energy
c i
e e e + = .
To simplify the analysis, the polytropic assumptions are invoked (refer to the
discussion right below Eq. (2.7)). Because the gas mixture is ideal, we have C
p
= R/(1-)
and C
v
= R/(1-) with = C
p
/C
v
and R = C
p
- C
v
. Furthermore, ) 1 /( = = pv T C e
v i

where v = 1/ is the specific volume.
We also assume that the reacting gas mixture is composed of only two species: A
as the reactant and B as the product. Let
fB fA
h h q =
0
where h
fA
and h
fB
are the heat of
formation of species A and B. The specific enthalpy of the reacting gas mixture is
0 0
) 1 ( ) ( Zq RT T C Zq T C h Z Zh T T C h
v p fB fA ref p
+ + = + = + + = , (2.15)
where the heat of formation of species B is defined as
ref p fB
T C h = . Therefore, we have
0
Zq e
c
= and
0
) 1 /( Zq pv e + = . We consider a simple one-step irreversible
reaction between the reactant (A) and the product (B): B A
f
K
. According to the law
of mass action,

17
[ ] [ ] A K dt A d
f
= / , and [ ] [ ] A K dt B d
f
= / , (2.16)
where [A], [B] are the mole concentration of species A and B, respectively. According to
the Arrhenius kinetics, ) / exp( T R E K K
u f
+
= . The source term in the species equation
due to chemical reaction can be written as
Z T R E K
u
) / exp(
+
= . (2.17)
To be consistent with the classical ZND model, we reformulate the species equation
based on the mass fraction of the product instead of the reactant. Let = 1-Z be the
mass fraction of the product, which is frequently referred to as the progress variable since
= 1 indicates completion of the chemical reaction. The species equation becomes

+
T R
E
K
x
u
t
u
exp ) 1 (
) (


. (2.18)
In this case, the definition of internal energy needs further clarification. Let the
specific enthalpy of the gas mixture be
0
( ) (1 )
p ref fA fB p
h C T T h h C T q = + + = . (2.19)
Here the heat of formation of species A is defined as
ref p fA
T C h = . Note that when
formulating in terms of Z, Eq. (2.15) implies h
fB
= C
p
T
ref
. As a result, we have
0 0 0
1 1
q
pv
q RT
RT
pv q T C e
p

= = . (2.20)
And the total energy E becomes
2 1
2
0
u
q
pv
E +

. (2.21)

18
We assume that the detonation wave propagates at a constant velocity D, and
u =u-D is the fluid velocity in the shock frame. We then transform the coordinates such
that the spatial origin is on the shock front. The equations of this steady problem in the
new coordinate system become
0
) (
=
dx
u d
, (2.22a)
0
) (
2
=
+
dx
p u d
, (2.22b)
[ ]
0
) (
=
+
dx
u p E d
, (2.22c)
) 1 ( exp
) (

=
+
T R
E
K
dx
u d
u
. (2.22d)
For convenience, the carat on the top of velocity u is dropped in the following
discussions.
To proceed, we substitute the continuity equation (2.22a) into the energy equation
(2.22c), and we have
0
2
2
=

+ +
u
pv e
dx
d
, (2.23)
where v = 1/ is the specific volume. Integrating the continuity (2.22a), momentum
(2.22b), and energy (2.23) equations gives the classical Rayleigh line relation:
v v
p p
u

=
0
0 2
) ( , or
v v
p p
v u

=
0
0 2 2
. (2.24)

19
This relation states that the mass flow rate u, which is a constant due to mass
conservation, is equal to the square root of the negative slope of a line in the (p, v) plane
connecting the initial state and an evolving state. Because mass and momentum must
conserve, a process evolving on a (p, v) plane is restricted to occur along a straight line,
i.e., the Rayleigh line. Moreover, because (u)
2
is always positive, a steady combustion
process described on a (p, v) plane cannot be simultaneously lower or higher in both
pressure and specific volume.
We then substitute the Rayleigh line relation into the integral form of the energy
equation (2.22c), and we have
)
1 1
)( (
2
1
0
0 0

+ = p p e e , or ) )( (
2
1
0 0 0
v v p p e e + = . (2.25)
Consider Eq. (2.20) and let = 0 for the unburned state. We then obtain the Hugoniot
curve:
0 ) )( (
2
1
1
0 0 0
0 0
= + +

q v v p p
v p pv

. (2.26)
Based on a given starting state, the above equation provides a set of hyperbolic curves,
which restricts the evolution of the thermodynamic properties on a (p, v) plane.
Depends on the evolution of the progressive variable , a specific Hugoniot curve could
be identified, and viable solutions must locate at the intersection points of the Rayleigh
line and Hugoniot curves.

20
Next, we nondimensionalize the above formulation for further analyses, refer to
Eqs. (2.4) and (2.11). The dimensionless variables, denoted by the superscript *, for the
unburned state, denoted by 0, can be derived as
1
*
0
*
0
*
0
= = = T p ,
0
*
0
M u = , and ) 1 /( 1
*
0
= e . (2.27)
Thus, the equation of state changes to
* * *
T v p = . (2.28)
The continuity equation becomes
* *
0
* *
D u u = = . (2.29)
The dimensionless Rayleigh line relation, i.e., Eq. (2.24), becomes
) 1 /( ) 1 (
* * 2
0
v p M = , (2.30a)
and the dimensionless Hugoniot curve relation is
0
2
) 1 )( 1 (
1
) 1 (
*
0
* * * *
=
+
+

q
v p v p

. (2.30b)
For convenience, the superscript * and the subscript 0 in M
0
denoting the Mach
number of the unburned gas will be dropped in the following discussions. We remark that
in commonplace textbooks for combustion, Eqs. (2.30) would complete the discussions
of the classical Rankine-Hugoniot (R-H) relations. Usually, a graphic illustration of
viable solutions on the (p, v) plane is provided. Refer to Fig. 2.2. However, we would
continue to derive the analytical solutions for p and v based on the use of Eqs. (2.30a-b).

21

Figure 2.2: Diagram of the Rayleigh lines and Hugoniot curves with different .
By the given values for the , M, and , two equations of Eqs. (2.30) determine
two unknowns, i.e., p and v. To proceed, we have the following equation for p from the
Rayleigh line equation, Eq. (2.30a),
2
) 1 ( 1 M v p + = . (2.31)
Substitute Eq. (2.31) into the Hugoniot curve relations, Eq. (2.30b), and we have
0 ) 1 )( 2 (
2
1
1
1 ) 1 (
0
2 2
2 2 2
= + +

+
q v v M M
v M v M

. (2.32)
By solving the above second-order polynomial equation for v and simplifying the result,
we have
[ ] ) ( 1
) 1 (
1
2
2

w
M
M
v
+
+
= , (2.33)

22
where

=
0
1 ) (
q
, (2.34a)
2 2
2 2
) 1 ( 2
) 1 (
M
M

, (2.34b)
2
2
1
1
M
w
M

=
+
. (2.35)
Note that both and w are constants, and () is a function of only. To proceed, we
substitute Eq. (2.33) into the Eq. (2.31), and we have the solution for p
[ ] ) ( 1
1
1
2

w
M
p
+
+
= . (2.36)
Based on the dimensionless equation of state, Eq. (2.28), we multiply Eq. (2.33) to Eq.
(2.36), to obtain the solution of T as
[ ]
2 2
2 2
2 2
) ( ) ( ) 1 ( 1
) 1 (
) 1 (

w w
M
M
pv T
+
+
= = . (2.37)
Clearly, solution of p, v, and T are functions of the progressive variable only. Moreover,
Eqs. (2.33)-(2.37) have two roots denoting the strong and weak detonation solutions
(refer to the two intersections denoted by S and W in Fig. 2.2). For the Chapman-Jouquet
(CJ) detonation, only one solution exists, and we have w() = 0 for = 1. According to
Eq. (2.35), a sonic unburned flow would render w = 0, which would also render
undetermined solution for (). Refer to Eqs. (2.34a-b). Therefore, the reasonable
solution for the CJ detonation is to let (1) = 0, or q
0
.

23
The above discussions show that all thermodynamic variables of the detonating
gas could be calculated as function of the combustion progressive variable . To proceed,
we illustrate the calculation of flow variable inside the reaction zone. We consider the
dimensionless species equation at a steady state:

=
+
*
*
*
*
*
exp
) 1 (

T
E
u
K
dx
d
. (2.38)
Refer to Eq. (2.22d) for the dimensional species equation. For convenience, the
superscript * denoting the dimensionless variables will be dropped in the following
discussions. According to Eq. (2.29), u = -Dv in Eq. (2.38), where v can be expressed by
Eq. (2.33).
To proceed, we must first determine the pre-exponential factor K and detonation
speed D before we integrate Eq. (2.38). To calculate K, we use the rate equation, and by
rearranging the Eq. (2.38),

d T E
u
dx K K ) / exp(
1
2 / 1
0
1
0
+


= = . (2.39)
Note that the integration of the progressive variable is performed from 0 to 1/2.
Therefore, the length scale x
0
is set to be equal to the half reaction zone length. Equation
(2.39) can be solved by numerical integration. We use Simpsons method. To calculate
the detonation velocity D, we use the definition of the overdriven factor f:

24
2
) / (
CJ
D D f = . (2.40)
If the CJ shock speed D
CJ
and the over-driven factor are given, the shock speed D is
determined. To calculate D
CJ
, we let (1) = 0, and substitute = q
0
into Eq. (2.34b) to
obtain
0 1 ) 2 / 1 ( 2
2 2 4
= + +
CJ CJ
M M , (2.41)
where
/ ) 1 ( 2
0
2 2
q = . (2.42)
By solving Eq. (2.41), we have
4
1
2
1
2 2
2

+ + =
CJ
M . (2.43)
Here, we dropped the root with the negative sign, which is the solution at the deflagration
branch. By using Eqs. (2.27) and (2.29), we have
2 2
CJ CJ
M D = . (2.44)
For a given overdriven factor f, D can be calculated by Eq. (2.40). With K and D
determined, the right hand side of the species equation, Eq. (2.38), is a function of only.
Integration of the ODE, i.e., Eq. (2.38) over a spatial domain gives the spatial distribution
of the mass fraction of the combustion product, leading to straightforward calculations of
spatial distribution of all thermodynamic variables in the reaction zone, including p, v,
and T.

25
To recap, the input parameters to the above analytical model are , q
0
, E
+
, and f,
and the calculation procedure for the reaction zone structure is summarized in the
following.
(i) The detonation velocity D is calculated by Eqs. (2.40), (2.43), and (2.44).
(ii) With the known value of D, Eq. (2.39) is integrated to obtain K.
(iii) By integrating Eq. (2.38), the distribution of inside the reaction zone
could be calculated.
(iv) With the given profile, the distribution of p, v, and T inside the reaction
zone can be calculated by Eqs. (2.33), (2.36), and (2.37), respectively.
A typical ZND solution is shown in Fig. 2.3, including species distributions,
pressure, density, and velocity. The input parameters are = 1.2, q
0
= 50, E
+
= 50, and f =
1.6.

26


(a) (b)

(c) (d)
Figure 2.3: Flow variable profiles of a one-dimensional ZND detonation wave from the
analytical solution: (a) mass fraction of reactant; (b) pressure; (c) density, and (d) velocity.
Parameters: = 1.2, q
0
= 50, E
+
= 50, and f = 1.6.

27
CHAPTER 3
THE SPACE-TIME CESE METHOD
3.1 Conventional Finite Volume Method
Finite volume methods are formulated according to flux balance over a fixed
spatial domain. The conservation laws state that the rate of change of the total amount of
a substance contained in a fixed spatial domain, i.e., the control volume V, is equal to the
flux of that substance across the boundary of V, denoted as S(V). Consider the differential
form of a conservation law as follows:
0
u
t

+ =

f , (3.1)
where u is density of the conserved flow variable, f is the spatial flux vector. By applying
Reynoldss transport theorem to the above equation, one can obtain the integral form as:
( )
0
V S V
udV d
t

+ =

f s
v
, (3.2)

28
where dV is a spatial volume element in V, ds = d n with d and n being the area and the
unit outward normal vector of a surface element on S(V) respectively. By integrating Eq.
(3.2) over the time interval (t
s
, t
f
), we have
( )
( )
0
f
s
f s
t
V V t S V
t t t t
udV udV d dt
= =
( (
+ =


f s
v
. (3.3)
The discretization of Eq. (3.3) is the focus of the finite-volume methods.
Lets consider the one dimensional case first. Let time and space (length) be the
two orthogonal coordinates of a space-time system, i.e., x
1
= x and x
2
= t. They constitute
a two-dimensional Euclidean space E
2
. Define ( , ) f u h , then by using the Gauss
divergence theorem, Eq. (3.1) becomes
( )
0
S R
d =

h s
v
, (3.4)
where R is a space-time region in E
2
and S(R) is the boundary of R, ds = d n with d and
n being the area and the unit outward normal vector of a surface element on S(R)
respectively. Equation (3.4) states that the total space-time flux h leaving the space-time
volume R through S(R) vanishes. Refer to Fig. 3.1 for a schematic of Eq. (3.4). Note that
all mathematical operations can be carried out as though E
N
were an ordinary Euclidean
space. For the current case N = 2.

29
r+dr
r
dr
ds
V
S(V)
t
x

Figure 3.1: A schematic of the space-time integral.
Figure 3.2(a) shows the arrangement of conservation cells for the one spatial
dimensional problem in the finite volume methods. Note that conventional finite volume
methods discretize and perform integration on the spatial dimensions only, as shown in
Eq. (3.3). Thus fixed control volume in space is required. Due to this constraint, the shape
of the conservation cells must be rectangular on the x-t plane. Refer to Fig. 3.2(a). The
mesh points are usually placed at the center of the spatial mesh, i.e., on the boundary of
the space-time conservation cells (marked by dots in Fig. 3.2(a)). The conservation cells
must stack up exactly on top of each other; no staggering in time is allowed. This
arrangement results in vertical interfaces extended in time between the neighboring
conservation cells. Across the interface, flow information travels in both directions. As
such, interfacial fluxes generally must be evaluated by interpolating the data from the

30
mesh points embedded in these two cells. Determining how this interpolation should be
carried out properly under temporally evolving solution is a difficult problem. Usually an
upwind bias method or a Riemann solver must be employed to calculate the nonlinear
fluxes.

(a)

(b)
Figure 3.2: Space-time meshes by the conventional finite volume methods in (a) one
and (b) two spatial dimensions.

31
To proceed, consider a rectangular space-time region R with vertices ABCD on
the x-t plane, as depicted in Fig. 3.3. Then the surface of R, i.e., S(R) is formed by the line
segments AB, BC, CD, and DA. Let t = t
s
at CD, t = t
f
at AB, x = x
s
at BC, and x = x
f
at DA.
Then because ( , ) f u = h , Eq. (3.4) implies
0
f f f f
s s s s
f s f f
x x t t
x x t t
t t t t x x x x
udx udx fdt fdt
= = = =
( ( ( (
+ =
( ( ( (


. (3.5)

Figure 3.3: Space-time geometry of the conventional finite volume methods in E
2
.
The discretization in muilti-spatial dimensions is performed similarly. Refer to
Fig. 3.2(b) for the two-dimensional case. A conservation cell is a uniform-cross-section
cylinder in the space-time domain. Note that though a circular cross section is shown, the
cross section could be of any shape. Usually its rectangular or triangular for structured

32
and unstructured meshes, respectively. Each conservation cell is a cylinder in space-time
with its spatial projection being the control volume in space and its top and bottom faces
representing two constant time levels. Again, because the control volume is a fixed
spatial domain, these conservation cells generally are stacked up exactly on the top of
each other. With this arrangement, the vertical interface that separates any two
neighboring cells will always be sandwiched between two neighboring columns of mesh
points (as shown in Fig. 3.2(a) for one-dimensional case). As such, the difficulty in
evaluating the flux at the vertical interface of two neighboring conservation cells always
exists. As will be shown, with a new space-time discretization and arrangement of
conservation cells adopted in the space-time CESE method, the above difficulty can be
bypassed completely.
3.2 The CESE Method
In contrast to the conventional finite volume methods, the space-time CESE
method treats space and time equally. The integration of Eq. (3.4) is performed on both
space and time dimensions to solve marching variables

in the new time levels. Moreover,
the CESE method has separate definitions of Conservation Element (CE) and Solution
Element (SE). SEs are used to discretize the space-time domain, and the conservation
laws are performed over CEs to solve the unknowns. CEs are non-overlapping space-time
domains such that (i) the whole computational domain can be filled by the union of all

33
CEs; (ii) flux conservation is enforced over each CE or over a union of several
neighboring CEs; and (iii) inside a CE, flow discontinuity is allowed. SEs are
non-overlapping space-time domains such that (i) an SE does not necessarily coincide
with a CE; (ii) the union of all SEs does not have to fill the whole computational domain;
(iii) flow variables and fluxes could be discontinuous across interfaces of neighboring
SEs; and (iv) within an SE, flow variable and fluxes are assumed continuous, and they
are approximated by the first-order Taylor series expansion in both space and time.
Consider the one dimensional case:
0
u f
t x

+ =

. (3.6)
Refer to Fig. 3.4(a), the space-time domain is divided into many non-overlapping
rhombus SEs. Each SE is associated with a grid point (marked by a dot). In the SE, flow
variables are assumed continuous and a Taylor series is used to discretize the equation.
For any (x, t) SE (j, n), u(x, t), f(x, t) and h(x, t), are approximated by
*
( , ; , ) u x t j n ,
*
( , ; , ) f x t j n , and
*
( , ; , ) x t j n h , respectively. Let
*
( , ; , ) ( ) ( ) ( ) ( )
n n n n
j x j j t j
u x t j n u u x x u t t + + , (3.7)
where (x
j
, t
n
) are the coordinates of the mesh point (j, n) at the center of SE(j, n). Note
that
n
j
u , ( )
n
x j
u , and ( )
n
t j
u , as the numerical analogues of the values of u, u
x
, u
t

respectively, are constant in SE(j, n).

34

(a)

(b)
Figure 3.4: Schematics of the CESE method in one spatial dimension. (a) Zigzagging SEs.
(b) Integration over CE to solve u and u
x
at the new time level.

35
Let
n
j
f and ( ')
n
j
f denote the value of f and df/du when u assumes the value of
n
j
u .
Let
( ) ( ') ( )
n n n
x j j x j
f f u , and ( ) ( ') ( )
n n n
t j j t j
f f u . (3.8)
Because
f f u
x u x

=

, and
f f u
t u t

=

, (3.9)
( )
n
x j
f and ( )
n
t j
f can be considered as the numerical analogues of the value of f/x and
f/t at (x
j
, t
n
), respectively. As a result, we assume that
*
( , ; , ) ( ) ( ) ( ) ( )
n n n n
j x j j t j
f x t j n f f x x f t t + + . (3.10)
Because h = (f, u), we also assume that
* * *
( , ; , ) ( ( , ; , ), ( , ; , )) x t j n f x t j n u x t j n h . (3.11)
Note that, by their definitions,
n
j
f and ( ')
n
j
f are functions of
n
j
u ; ( )
n
x j
f are functions
of
n
j
u and ( )
n
x j
u ; and ( )
n
t j
f are functions of
n
j
u and ( )
n
t j
u .
Moreover, we assume that, for any (x, t) SE (j, n), u =
*
( , ; , ) u x t j n and f =
*
( , ; , ) f x t j n satisfy Eq. (3.6), i.e.,
* *
( , ; , ) ( , ; , )
0
u x t j n f x t j n
t x

+ =

. (3.12)
According to Eqs. (3.7) and (3.10), the above equation is equivalent to
( ) ( )
n n
t j x j
u f = . (3.13)

36
Since ( )
n
x j
f are functions of
n
j
u and ( )
n
x j
u , Eq. (3.13) implies that ( )
n
t j
u are also
functions of
n
j
u and ( )
n
x j
u . Thus, we conclude that the only independent discrete
variables needed to solve are
n
j
u and ( )
n
x j
u .
To proceed, let the space-time domain be divided into non-overlapping rectangular
CEs. Refer to Fig. 3.4. Each CE is associated with a mesh point. We then employ local
space-time flux conservation over CEs to solve the unknowns. Figure 3.4(b) shows the CE
and SE associated with mesh points (j, n), and two neighborinh SE associated with mesh
points (j-1, n-1) and (j+1, n-1)on the previous time level. We assume that u
*
and u
x
*
at mesh
points (j-1, n-1) and (j+1, n-1) are known and used to calculate
n
j
u and ( )
n
x j
u at the new
time level n. Note that the line segments forming the boundary of CE(j, n) belong to the
three SEs respectively. As a result, by imposing flux conservation over the CE, i.e.,
*
( ( , ))
0
S CE j n
d =

h s
v
, (3.14)
one obtains
1/ 2 1/ 2 1/ 2 1/ 2
1/ 2 1/ 2 1/ 2 1/ 2
( ) / 2
n n n n n
j j j j j
u u u s s

+ +
= + + , (3.15)
where
2
( / 4)( ) ( / ) ( / 4 )( )
n n n n
j x j j t j
s x u t x f t x f = + + , (3.16)
Given the values of the marching variables at t = t
n-1/2
, the value of u
j
n
can be calculated
explicitly.

37
To solve (u
mx
)
j
n
at point (j, n), central differencing is performed:
( ) [( ) ( ) ] / 2
n n n
x j x j x j
u u u
+
= + , (3.17)
where
1/ 2
( ) ( ) /( / 2)
n n n
x j j j
u u u x

= , (3.18)
1/ 2 1/ 2
1/ 2 1/ 2 1/ 2
( / 2)( )
n n n
j j t j
u u t u


= + . (3.19)
For flows with discontinuities, Eq. (3.17) is replaced by a re-weighting procedure to add
artificial damping:
( ) (( ) , ( ) , )
n n n
x j x j x j
u W u u
+
= , (3.20)
where the function W is defined as
( , , )
x x x x
W x x
x x

+ +
+
+
+
=
+
. (3.21)
is an adjustable constant, and usually = 1 or 2.
Compared with the conventional finite volume method, there is no fixed spatial
domain constraint in the CESE method. Thus, the SEs could be arranged in a stagging
way. In time marching calculations, the flow variables at neighboring cells leapfrog each
other. Across oblique interfaces of neighboring SEs, discontinuities of u
*
(and h
*
) are
allowed. Flow information propagates only in one direction, i.e., toward the future, through

38
each oblique interface. While within the SEs, the unknows (including the flow cariables
and their derivatives) are distributed linearly. Thus, the calculation of the space-time flux
passing these interfaces of CEs is simple and straightforward. No Riemann problem is
encountered for the nonlinear equations.
Note that there is a family of schemes constructed with the CESE method. The
above introduced method is celled a- scheme because of the introduced adjustable
constant . In the a- scheme, u
x
is not solved by the above mentioned differencing method.
Instead, the two unknowns u and u
x
are treated equivalent and solved simultaneously by
constructing two CEs within the same three SEs. Its straightforward to construct two
equations to solve two unknowns.

Figure 3.5: CE
+
and CE
-
for solving u and u
x
at the new time level in the a- scheme.

39
To proceed we briefly discuss the extension to two spatial dimensions. For
equations in two spatial dimensions, the SEs and CEs are in a three-dimensional Euclidian
space E
3.
With three unknowns u, u
x
, and u
y
now, according to the idea of the a- scheme,
three equations must be provided to solve the new flow variables at the new time step. Thus
triangle meshes is used. Figure 3.6 shows the mesh arrangement for the two spatial
dimensions. Note that the unknowns are located at the centroid of each triangle.
Similarly, tetrahedron meshes are used in three spatial dimensional cases and four
unknowns (u, u
x
, u
y
, and u
z
) at each mesh node are calculated for three spatial dimensional
equations. Details of the original CESE schemes in multiple spatial dimensions please refer
to Wang and Chang (1999).

Figure 3.6: Grid points in the x-y plane for the original CESE method.

40
Note that in the present work, the a- scheme is used and extended to solve Euler
equations in two and three spatial dimensions. Hence only one CE is constructed to solve
u and the spatial derivatives are obtained by the differencing method. Therefore, different
with the original mesh type settings, quadrilateral and tetrahedron cells are used in two-
and three- dimensional cases, respectively. See the following sections for details.
3.3 The Modified CESE Method
Consider the one-dimensional Euler equations for reacting flows with source term
for chemical reactions, refer to Eqs. (2.13)-(2.14):
m m
m
u f
t x


+ =

, m = 1, 2, 3, 4. (2.13)
Note
m
is the function of u
m
, m = 1, 2, 3, 4. Let x
1
= x, x
2
= t and ( , )
m m m
f u h , then by
using the Gauss divergence theorem, we have the similar equation as Eq. (3.4) except the
source term on the right hand side (RHS):
( )
m m
S R R
d dR =

h s
v
, m = 1, 2, 3, 4. (3.22)
By the same definition of SE and CE as shown in Fig. 3.4, for any (x, t) SE (j, n), u
m
(x,
t), f
m
(x, t) and h
m
(x, t), are approximated similarly by
*
( , ; , ) u x t j n ,
*
( , ; , ) f x t j n , and
*
( , ; , ) x t j n h , respectively. By the linear distribution assumption within SEs, we have
*
( , ; , ) ( ) ( ) ( ) ( ) ( )
n n n n
m m j mx j j mt j
u x t j n u u x x u t t + + . (3.23)

41
Let ( )
n
m j
f and
,
( )
n
m l j
f denote the value of f
m
and f
m
/u
l
, m, l = 1, 2, 3, 4,
respectively, when u
m
assumes the value of ( )
n
m j
u . Let
4
,
1
( ) ( ) ( )
n n n
mx j m l j lx j
l
f f u
=

, m = 1, 2, 3, 4, (3.24a)
and
4
,
1
( ) ( ) ( )
n n n
mt j m l j lt j
l
f f u
=

, m = 1, 2, 3, 4. (3.24b)
Because
4
1
m m l
l
l
f f u
x u x
=

=

, and
4
1
m m l
l
l
f f u
t u t
=

=

, (3.25)
( )
n
mx j
f and ( )
n
mt j
f can be considered as the numerical analogues of the value of f
m
/x
and f
m
/t at (x
j
, t
n
), respectively. As a result, we assume
*
( , ; , ) ( ) ( ) ( ) ( ) ( )
n n n n
m m j mx j j mt j
f x t j n f f x x f t t = + + , (3.26)
and
* * *
( , ; , ) ( ( , ; , ), ( , ; , ))
m m m
x t j n f x t j n u x t j n = h . (3.27)
Note that, by their definitions, for any m = 1, 2, 3, 4, ( )
n
m j
f and
,
( )
n
m l j
f are functions of
( )
n
m j
u ; ( )
n
mx j
f are functions of ( )
n
m j
u and ( )
n
mx j
u ; and ( )
n
mt j
f are functions of ( )
n
m j
u
and ( )
n
mt j
u .
Assume that, for any (x, t) SE (j, n), u
m
=
*
( , ; , )
m
u x t j n and f
m
=
*
( , ; , )
m
f x t j n
satisfy Eq. (2. ), i.e.,

42
* *
*
( , ; , ) ( , ; , )
( , ; , )
m m
m
u x t j n f x t j n
x t j n
t x


+ =

, m = 1,2, 3, 4. (3.28)
According to Eqs. (3.23) and (3.26), and further assume that
*
m
is constant within SE(j,
n), i.e.,
*
( , ; , )
m
x t j n =( )
n
m j
, the above equation is equivalent to
( ) ( ) ( )
n n n
mt j mx j m j
u f = + . (3.29)
Since ( )
n
mx j
f are functions of ( )
n
m j
u and ( )
n
mx j
u ; and ( )
n
m j
is also a function of
( )
n
m j
u , Eq. (3.29) implies that ( )
n
mt j
u are also functions of ( )
n
m j
u and ( )
n
mx j
u . Thus,
the only independent discrete variables needed to solve are ( )
n
m j
u and ( )
n
mx j
u .
To proceed, we employ local space-time flux balance over CE(j, n) to solve the
unknowns. Refer to Fig. 3.4(b). Assume that
*
m
u and
*
mx
u at mesh points (j-1, n-1) and
(j+1, n-1) are known and used to calculate ( )
n
m j
u and ( )
n
mx j
u at the new time level n. By
enforcing flux balance over CE(j, n), i.e.,
* *
( ( , )) ( , )
m m
S CE j n CE j n
d dV =

h s
v
, (3.30)
one obtains
( ) ( )
4
n n
m j m j
t
u

=
1/ 2 1/ 2 1/ 2 1/ 2 1/ 2 1/ 2
1/ 2 1/ 2 1/ 2 1/ 2 1/ 2 1/ 2
1
( ) ( ) ( ) ( ) ( ) ( )
2 4 4
n n n n n n
m j m j m j m j m j m j
t t
u u s s

+ + +
(
+ + + +
(

,
(3.31)
where
2
( ) ( / 4)( ) ( / )( ) ( / 4 )( )
n n n n
m j mx j m j mt j
s x u t x f t x f = + + . (3.32)

43
Given the values of the marching variables at mesh points (j-1/2, n-1/2) and (j+1/2, n-1/2),
the RHS of Eq. (3.31) is known. Since ( )
n
m j
is a function of ( )
n
m j
u , ( )
n
m j
u can then be
solved with Newtons iteration method. Note that only the values of
1/ 2
1/ 2
( )
n
mt j
f

are
required, which needs the values of
1/ 2
1/ 2
( )
n
mt j
u

as input. Thus, Eqs. (3.29) only need to be

carried out on SE(j-1/2, n-1/2) and SE(j+1/2, n-1/2).
To solve ( )
n
mx j
u at point (j, n), a similar central differencing is used:
( ) [( ) ( ) ] / 2
n n n
mx j mx j mx j
u u u
+
= + , (3.33)
where

1/ 2
( ) [( ) ( ) ] /( / 2)
n n n
mx j m j m j
u u u x

= , (3.34)
1/ 2 1/ 2
1/ 2 1/ 2 1/ 2
( ) ( ) ( / 2)( )
n n n
m j m j mt j
u u t u


= + . (3.35)
For flows with discontinuities, Eq. (3.33) is replaced by a re-weighting procedure to add
artificial damping:
( ) (( ) , ( ) , )
n n n
mx j mx j mx j
u W u u
+
= , (3.36)
where the function W is defined as in Eq. (3.21). is an adjustable constant, and
usually = 1 or 2.
In simulating combustion dynamics, the time scale of the source term is much
smaller than the characteristic time scale of the convection term, and the equation set is
stiff. The above method with CE and SE defined as in Fig. 3.4 is useful for treating the

44
conservation laws with non-stiff source terms. However, it could fail in treating stiff
source terms. With stiff source terms, the dominant effect in the overall flux balance, i.e.,
Eq. (3.31) is caused by the source terms. In particular, the stiff source terms would
directly influence the calculation of ( )
n
mt j
u (Eq. (3.29)), and hence the calculation of
( )
n
mt j
f (Eq. (3.24b)). As such, any small difference between the value of u
m
(and/or u
mx
)
at mesh points (j-1/2, n-1/2) and (j+1/2, n-1/2) will be amplified by the stiffness factor.
As a result, huge differences occur between
1/ 2
1/ 2
( )
n
mt j
f

and
1/ 2
1/ 2
( )
n
mt j
f

+
, and between
1/ 2
1/ 2
( )
n
m j

and
1/ 2
1/ 2
( )
n
m j

+
. Thus the Newtons method could fail to solve this stiff
relaxation system.
Due to the above difficulty, a modification to the original method has been
developed to avoid the amplification effects. The new treatment was based on
redistributing the space-time regions such that all source term effects are hinged on the
mesh point at the new time level.
Figure 3.7(b) shows the new layout of the CE and the SE associated with grid
point (j, n). The SE is constituted by the rectangle ABB'A' and the line segments QQ on
the top of ABB'A' as well as its immediate neighborhoods. The CE is the rectangle ABB'A'.
Note that in Fig. 3.7(b), Q, Q' and Q share the same spatial projection, so do A and A',
and B and B'. The superscript prime and quotation mark denote the time level n-1/2 and
n+1/2, respectively. Besides the SE(j, n), two more neighboring SEs are also shown to
illustrate the belonging (to SEs) of the three parts of CE(j, n).

45

(a)

(b)
Figure 3.7: Schematics of the modified CESE method in one spatial dimension: (a) the
staggered space-time mesh, (b) SE (j, n), shown as the yellow part, and CE (j, n).

46
With this new construction of SEs, we proceed to perform integration as that in
the original CESE method, and have Eqs. (3.23), (3.26), and (3.27). However, the
evaluation of
1/ 2
1/ 2
( )
n
mt j
u

and
1/ 2
1/ 2
( )
n
mt j
u

+
differs from the original method by simply using
1/ 2 1/ 2
1/ 2 1/ 2
( ) ( )
n n
mt j mx j
u f


= , (3.37)
and no source term effect is included. Note that these temporal derivatives are used only
along the line segments sticking out on the top of the rectangular region. The area of the
line segment is zero due to the re-distribution of the SEs. This layout excludes the
influence of the stiff source term on time level n-1/2.
Therefore, in the calculation of the overall flux balance over CE(j, n), i.e., Eq.
(3.30), the integration of the source term is totally based on the flow properties at the
mesh point (j, n), and one obtains
1/ 2 1/ 2 1/ 2 1/ 2
1/ 2 1/ 2 1/ 2 1/ 2
1
( ) ( ) ( ) ( ) ( ) ( )
2 2
n n n n n n
m j m j m j m j m j m j
t
u u u s s

+ +

( = + +

, (3.38)
where
2
( ) ( / 4)( ) ( / )( ) ( / 4 )( )
n n n n
m j mx j m j mt j
s x u t x f t x f = + + . (3.39)
Similar to that in the original method, for any m = 1, 2, 3, 4, ( )
n
m j
f and ( )
n
m j
are
functions of ( )
n
m j
u ; ( )
n
mt j
f is a function of ( )
n
m j
u and ( )
n
mx j
u . Thus, given the values
of the marching variables at t = t
n-1/2
, Eq. (3.38) is a nonlinear equation of ( )
n
m j
u . Again,
the Newtons iteration method is used to solve it.

47
Based the above treatment, we avoid the amplification effect of the stiff source
term. The effect of the source term calculation on the flow properties from the time level
n-1/2 was not involved. The layout of the modified SEs helps to eliminate the stiffness
problem. The union of all the CEs still covers the whole space-time domain without
overlapping, thus the source term effect satisfies the local and global flux balance in an
integral sense.
We remark that for faithful simulation of the physical phenomena, one should
resolve the time and space scales associated with the source terms. However, this may
require huge computer resources for multi-dimensional simulations and probably exceed
what the present facilities can provide.
The algorithm for solving ( )
n
mx j
u in the modified method remains unchanged,
i.e., by Eqs. (3.33)-(3.36).
3.4 Two-Dimensional Euler Solver
Consider the standard conservation form of the two-dimensional unsteady Euler
equations for reacting flows with source term for chemical reactions:
m m m
m
u f g
t x y

+ + =

, m = 1, 2, 3, 4, 5, (2.9)

48
where f
m
and g
m
, are explicit functions of the independent flow variables u
m
. Let x
1
= x, x
2

= y, and x
3
= t be the coordinates of a three-dimensional Euclidean space E
3
. We have
( )
m m
S R V
d dR =

h s
v
, m = 1, 2, 3, 4, 5, (3.40)
where S(R) and ds were defined following Eq. (3.4) and ( , , )
m m m m
f g u h .
3.4.1 Conservation Elements and Solution Elements
In two spatial dimensions, the computational domain on the xy plane is divided
into non-overlapping convex quadrilaterals and any two neighboring quadrilaterals share
a common side. Refer to Fig. 3.8(a). Vertices and centroids of quadrilaterals are marked
by dots and circles, respectively. Q is the centroid of a typical quadrilateral B
1
B
2
B
3
B
4
.
Note the underline differentiates the points in the computational domain on the xy plane
with those in the space-time domain as to be introduced below. Points A
1
, A
2
, A
3
, and A
4
,
respectively, are the centroids of the four neighboring quadrilaterals of the quadrilateral
B
1
B
2
B
3
B
4
. Point Q
*
(marked by a cross in Fig. 3.8(a)), is the centroid of the polygon
A
1
B
1
A
2
B
2
A
3
B
3
A
4
B
4
. Hereafter, point Q
*
, which generally does not coincide with point Q,
is referred to as the solution point associated with the centroid Q. Similarly, points A
1
*
,
A
2
*
, A
3
*
, and A
4
*
, which are also marked by crosses, are the solution points associated
with the centroids A
1
, A
2
, A
3
, and A
4
, respectively.


49

(a)

(b)
Figure 3.8: The space-time mesh in two spatial dimensions: (a) grid points in the x-y
plane, (b) SE and CE for the two-dimensional scheme.

50
To proceed, consider the space-time mesh shown in Fig. 3.8(b). Here t = nt at
the nth time level, where n = 0, 1/2, 1, 3/2, . For a given n, Q, Q, and Q, respectively,
denote the points on the time levels n, n-1/2, and n+1/2 with point Q being their common
spatial projection. Other space-time mesh points in Fig. 3.8(b) are defined similarly. In
particular, Q
*
, A
1
*
, A
2
*
, A
3
*
, and A
4
*
, by definition, lie on the nth time level and are the
space-time solution mesh points associated with points Q, A
1
, A
2
, A
3
, and A
4
, respectively.
Q
*
,
'*
1
A ,
'*
2
A ,
'*
3
A , and
'*
4
A , lie on the time level n-1/2 and are the space-time solution
mesh points associated with points Q, A
1
, A
2
, A
3
, and A
4
, respectively.
With the above preliminaries, we are ready to discuss the geometry of the CE and
the SE associated with point Q
*
. The numerical solution of the flow variables u
m
at the
nth time level are calculated based on the known flow solution at points in the time level
n-1/2, denoted by superscript prime. To integrate Eq. (3.40), four Basic Conservation
Elements (BCEs) of point Q
*
are constructed and denoted by BCE
l
(Q), with l = 1, 2, 3,
and 4. These four BCEs are defined to be the space-time cylinders A
1
B
1
QB
4
A
1
B
1
QB
4
,
A
2
B
2
QB
1
A
2
B
2
QB
1
, A
3
B
3
QB
2
A
3
B
3
QB
2
, and A
4
B
4
QB
3
A
4
B
4
QB
3
, respectively. The
compounded conservation element of point Q, denoted by CE(Q), is defined to be the
space-time cylinder A
1
B
1
A
2
B
2
A
3
B
3
A
4
B
4
-A
1
B
1
A
2
B
2
A
3
B
3
A
4
B
4

, i.e., the union of the
above four BCEs. Moreover, the SE of point Q
*
, denoted by SE(Q
*
), is defined as the

51
union of CE(Q) and four plane segments QQB
1
B
1
, QQB
2
B
2
, QQB
3
B
3
, and
QQB
4
B
4
as well as their immediate neighborhoods.
3.4.2 Approximations with a Solution Element
To proceed, denote the set of the space-time mesh points whose spatial
projections are the centroids of quadrilaterals and the set of the space-time mesh points
whose spatial projections are the solution points, depicted in Fig. 3.3(a), by and
*
,
respectively. For any Q
*

*
and any (x, y, t)SE(Q
*
), The flow variables and flux
vectors, i.e., u
m
(x, y, t), f
m
(x, y, t), g
m
(x, y, t), and h
m
(x, y, t), are approximated to their
numerical counterparts, i.e., u
m
*
(x, y, t; Q
*
), f
m
*
(x, y, t; Q
*
), g
m
*
(x, y, t; Q
*
), and h
m
*
(x, y, t;
Q
*
), respectively, based on the first-order Taylor series expansion with respect to Q
*
(x
Q*
,
y
Q*
, t
n
). Specifically, for any m = 1, 2, 3, 4, 5, let
* *
* * * * * *
( , , ; ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
n
m m Q mx Q Q my Q Q mt Q
u x y t Q u u x x u y y u t t + + + ,
(3.41)
where (x
Q*
, y
Q*
, t
n
) are the coordinates of the space-time solution mesh point Q
*
and
*
( )
m Q
u ,
*
( )
mx Q
u ,
*
( )
my Q
u , and
*
( )
mz Q
u , which are constant in SE(Q
*
), are the numerical
analogues of the values of u
m
, /
m
u x , /
m
u y , and /
m
u t at point Q
*
,
respectively.

52
Based on the chain rule,
*
( )
mx Q
f ,
*
( )
mx Q
g ,
*
( )
my Q
f ,
*
( )
my Q
g ,
*
( )
mt Q
f , and
*
( )
mt Q
g are defined in a similar way as ( )
n
mx j
f and ( )
n
mt j
f are defined in the 1D case.
Refer to Eq. (3.24). We then define
*
* * * * * *
( , , ; *) ( ) ( ) ( ) ( ) ( ) ( ) ( )
n
m m Q mx Q Q my Q Q mt Q
f x y t Q f f x x f y y f t t + + + ,
(3.42)
*
* * * * * *
( , , ; *) ( ) ( ) ( ) ( ) ( ) ( ) ( )
n
m m Q mx Q Q my Q Q mt Q
g x y t Q g g x x g y y g t t + + + ,
(3.43)
and
* * * *
( , , ; *) ( ( , , ; *), ( , , ; *), ( , , ; *))
m m m m
x y t Q f x y t Q g x y t Q u x y t Q h . (3.44)
similarly for any m = 1, 2, 3, 4, 5.
Note that by definitions
*
( )
m Q
f ,
*
( )
m Q
g ,
*
( )
mx Q
f ,
*
( )
mx Q
g ,
*
( )
my Q
f ;
*
( )
mt Q
f
and
*
( )
mt Q
g are functions of
*
( )
m Q
u ,
*
( )
mx Q
u ,
*
( )
my Q
u and
*
( )
mt Q
u only, for any m = 1,
2, 3, 4, 5.
To proceed, we assume that for any (x, y, t)SE(Q
*
), and any m = 1, 2, 3, 4, 5,
* * * * * *
*
( , , ; ) ( , , ; ) ( , , ; )
( )
m m m
m
u x y t Q f x y t Q g x y t Q
Q
t x y


+ + =

, (3.45)

53
which is the numerical analogue of Eq. (2. ). Note that the source term
*
m
is assumed
constant within SE(Q
*
) and the value of
*
m
is determined by
*
( )
m Q
u only. With the aid
of Eqs. (3.41)-(3.43), Eq. (3.45) implies that for any m = 1, 2, 3, 4, 5,
* * * *
( ) ( ) ( ) ( )
mt Q mx Q my Q m Q
u f g = + . (3.46)
However, like in the 1D case, to solve Euler equations with stiff source term, for any l =
1, 2, 3, 4,
'*
( )
l
mt
A
u are evaluated excluding the source term effect, i.e.,
'* '* '*
( ) ( ) ( )
l l l
mt mx mx
A A A
u f g = . (3.47)
Using the above equations, it can be shown that the only independent discrete solution
variables associated with the space-time solution point Q
*
are
*
( )
m Q
u ,
*
( )
mx Q
u and
*
( )
my Q
u , m = 1, 2, 3, 4, 5, as similar to the 1D case.
3.4.3 Evaluation of u
m

To evaluate space-time flux in E
3
, consider the following preliminaries first. Let
be a space-time plane segment lying within SE(Q
*
). Let A be the area of , (x
c
,

y
c
, t
c
) be
the coordinates of the centroid of , and n be a unit vector normal to . Then, because
* *
( , , ; )
m
u x y t Q ,
* *
( , , ; )
m
f x y t Q , and
* *
( , , ; )
m
g x y t Q are linear in x, y, and t, Eq. (2.9)
implies that
* * *
( , , ; )
m m c c c
d x y t Q A

h s h n , (3.48)
where ds = d n with d being the area of a surface element on .

54
The boundary of CE(Q) belongs to the union of SE(Q
*
) and SE(
'*
l
A ), l = 1, 2, 3, 4.
Specifically, (i) the octagon A
1
B
1
A
2
B
2
A
3
B
3
A
4
B
4
belongs to SE(Q
*
); (ii) the quadrilaterals
A
1
B
1
QB
4
, A
1
B
4
B
4
A
1
, and A
1
B
1
B
1
A
1
belong to SE(
'*
1
A ); (iii) the quadrilaterals
A
2
B
2
QB
1
, A
2
B
1
B
1
A
2
, and A
2
B
2
B
2
A
2
belong to SE(
'*
2
A ); (iv) the quadrilaterals
A
3
B
3
QB
2
, A
3
B
2
B
2
A
3
, and A
3
B
3
B
3
A
3
belong to SE(
'*
3
A ); and (v) the quadrilaterals
A
4
B
4
QB
3
, A
4
B
3
B
3
A
4
, and A
4
B
4
B
4
A
4
belong to SE(
'*
4
A ).
To proceed, we evaluate the surface vector (i.e., the unit outward normal vector
multiplied by the area) of every boundary face of CE(Q). Let S denote the area of the
octagon A
1
B
1
A
2
B
2
A
3
B
3
A
4
B
4
, then the surface vector of the top face of CE(Q) is (0, 0, S)
because the unit outward normal vector of this face is (0, 0, 1).
Then we consider the bottom face of CE(Q), which is constituted by four
quadrilaterals, namely, A
1
B
1
QB
4
, A
2
B
2
QB
1
, A
3
B
3
QB
2
, and A
4
B
4
QB
3
. Let (x
l
, y
l
)
and S
l
, l = 1; 2; 3; 4, denote the spatial coordinates of the centroids and the areas of the
above four quadrilaterals, respectively. Then for any l = 1, 2, 3, 4, (x
l
, y
l
, t
n-1/2
) are the
coordinates of the above four centroids, and (0, 0, -S
l
) are the surface vectors of the above
four quadrilaterals, respectively. Furthermore, because the area of the bottom face of
CE(Q) is identical to that of the top face, one concludes that
4
1
l
l
S S
=
=

.
Finally consider the side faces of CE(Q), i.e., A
1
B
4
B
4
A
1
, A
1
B
1
B
1
A
1
, A
2
B
1
B
1
A
2
,
A
2
B
2
B
2
A
2
, A
3
B
2
B
2
A
3
, A
3
B
3
B
3
A
3
, A
4
B
3
B
3
A
4
, and A
4
B
4
B
4
A
4
, which belong to SE(
'*
l
A ), l
= 1, 2, 3, 4, respectively. Let the eight side faces be assigned the indices (k, l),

55
respectively. Then the (1, l) and (2, l) side faces belong to SE(
'*
l
A ). Note the spatial
projection of each side face is a line segment on the x-y plane. Let
l
k
, ( , )
l l
kx ky
n n , and
( , )
l l
k k
x y , respectively, denote the length, the unit outward normal, and the coordinates of
the midpoint of the spatial projection (on the x-y plane) of the (k, l) side face. Then,
because each side face is sandwiched between the (n-1/2)th and the nth time levels, one
concludes that the surface vector and the coordinates of the centroid of the (k, l) side face,
are given by ( / 2) ( , , 0)
l l l
k kx ky
t n n and ( , , / 4)
l l n
k k
x y t t , respectively.
By using Eq. (3.48), the flux of
*
m
h leaving each face of CE(Q) can then be
evaluated in terms of the independent marching variables at points Q
*
and
'*
l
A , l = 1, 2, 3,
4. For example, because (x
Q*
, y
Q*
, t
n
) are the coordinates of the centroid Q
*
of the top face
of CE(Q);
* *
* *
( , , ; )
n
m Q Q
u x y t Q =
*
( )
m Q
u (see Eq. (3.41)); and the surface vector of the
top face is (0, 0, S), the flux of
*
m
h leaving CE(Q) through its top face is
*
( )
m Q
u S.
To proceed, we employ local space-time flux balance over CE(Q):
*
( ( )) ( )
m m
S CE Q CE Q
d dV =

h s
v
, m = 1, 2, 3, 4, 5. (3.49)
With the aid of the above discussion, it can be shown that
4
* *
1
1
( ) ( )
2
l
m Q m Q m
l
t
u R
S
=
| |
|
\ .

, m = 1, 2, 3, 4, 5, (3.50)

56
where, for any l = 1, 2, 3, 4,
l
m
R =
* 1/ 2 '*
( , , ; )
l l l n
m l
S u x y t A

2
* '* * '*
1
, , ; , , ;
2 4 4
l l l l n l l l n
k kx m k k l ky m k k l
k
t t t
n f x y t A n g x y t A
=
( | | | |

| | (
\ . \ .

. (3.51)
Note that the functions
* '*
( , , ; )
m l
u x y t A ,
* '*
( , , ; )
m l
f x y t A , and
* '*
( , , ; )
m l
g x y t A are defined
using Eqs. (3.41), (3.42), and (3.43), respectively, with the symbols Q
*
and t
n
in these
equations being replaced by
'*
l
A and t
n-1/2
, respectively. Given the values of the
marching variables at t = t
n-1/2
, the RHS of Eq. (3.50) can be readily calculated. Note
*
( )
m Q
is a function of
*
( )
m Q
u . Therefore
*
( )
m Q
u can then be solved with Newtons
iteration method.
3.4.4 Evaluation of u
mx
and u
my

Like in the 1D case,
*
( )
mx Q
u and
*
( )
my Q
u are evaluated with a similar
finite-difference approach. First, we perform a spatial translation of the quadrilateral
* * * *
1 2 3 4
A A A A so that the centroid of the resulting new quadrilateral
1 2 3 4
o o o o
A A A A coincides
with Q
*
. Refer to Fig. 3.9. Let the centroid of the quadrilateral
* * * *
1 2 3 4
A A A A and its spatial
coordinates be denoted by A
*
and (
*
A
x ,
*
A
y ), respectively. Then for any l = 1, 2, 3, 4,
(
o
l
A
x ,
o
l
A
y ), the spatial coordinates of
o
l
A are given by
* * * o
l l
A A Q A
x x x x = + , and
* * * o
l l
A A Q A
y y y y = + . (3.52)

57

Figure 3.9: Spatial translation of the quadrilateral
* * * *
1 2 3 4
A A A A .
To proceed, let
* '*
( ) ( , , ; )
o o o
l l l
n
m m l
A A A
u u x y t A , m = 1, 2, 3, 4, 5, l = 1, 2, 3, 4. (3.53)
Next, for any m = 1, 2, 3, 4, 5, consider the three points in the xyu space with the
coordinates (
*
Q
x ,
*
Q
y ,
*
( )
m
Q
u ), (
1
o
A
x ,
1
o
A
y ,
1
( )
o
m
A
u ), and (
2
o
A
x ,
2
o
A
y ,
2
( )
o
m
A
u ), respectively. The
values of / u x and / u y on the plane that intercepts the three points are given by
( )
*
( )
/
l
mx x
Q
u and
( )
*
( )
/
l
my y
Q
u ( 0 ), (3.54)
where
* *
1 1
* *
2 2
o o
o o
A Q A Q
A Q A Q
x x y y
x x y y

, (3.55a)
* *
1 1
* *
2 2
( ) ( )
( ) ( )
o o
o o
m m
A Q A Q
x
m m
A Q A Q
u u y y
u u y y

, (3.55b)

58
and
* *
1 1
* *
2 2
( ) ( )
( ) ( )
o o
o o
m m
A Q A Q
x
m m
A Q A Q
u u x x
u u x x

. (3.55c)
Note that = 0 if and only if the spatial projections of
1
o
A ,
2
o
A and Q
*
are collinear.
Similarly,
( )
*
( ) k
mx
Q
u and
( )
*
( ) k
my
Q
u , k = 2, 3, 4, are defined, respectively, by replacing the
points
1
o
A and
2
o
A in the above operations with
2
o
A and
3
o
A ,
3
o
A and
4
o
A , and
4
o
A
and
1
o
A , respectively.
With the above preliminaries, for each m = 1, 2, 3, 4, 5,
*
( )
mx Q
u and
*
( )
my Q
u can
then be evaluated by
( )
*
4
( )
*
1
1
( )
4
k
mx Q mx
Q
k
u u
=

, and
( )
*
4
( )
*
1
1
( )
4
k
my Q my
Q
k
u u
=

. (3.56)
For a flow with discontinuities, the above equation may be replaced by a re-weighting
procedure, i.e.,
( ) ( )
( )
*
4
( ) ( )
1
* 4
( )
1
( )
k k
m mx
Q
k
mx Q
k
m
k
W u
u
W
=
=
(
(

and
( ) ( )
( )
*
4
( ) ( )
1
* 4
( )
1
( )
k k
m my
Q
k
my Q
k
m
k
W u
u
W
=
=
(
(

, (3.57)
where 0 is an adjustable constant,
( ) ( )
* *
2 2
( ) ( ) k k
mk mx my
Q Q
u u
( (
+
( (

, m = 1, 2, 3, 4, 5, k = 1, 2, 3, 4, (3.58)

59
and
(1)
2 3 4 m m m m
W ,
(2)
3 4 1 m m m m
W ,
(3)
4 1 2 m m m m
W ,
(4)
1 2 3 m m m m
W .
(3.59)
In particular, if for any k = 1, 2, 3, 4,
mk
= 0, then
*
( )
mx Q
u and
*
( )
my Q
u are set to be 0.
Usually = 1 or = 2, and Eqs. (3.57) reduce to Eqs. (3.56) if = 0. Note that in practice
a small positive number like 10
-60
is added to the denominators that appear in Eqs. (3.57)
to avoid dividing by zero.
3.5 Three-Dimensional Euler Solver
For 3D unsteady Euler problems, Eqs. (2. ) and (3.40) are replaced by
m m m m
m
u f g q
t x y z

+ + + =

, m = 1, 2, 3, 4, 5, 6, (2. )
and
( )
m m
S R V
d dR =

h s
v
, m = 1, 2, 3, 4, 5, 6 (3.60)
respectively, where ( , , , )
m m m m m
f g q u h , with x
1
= x, x
2
= y, x
3
= z, and x
4
= t being the
coordinates of a four-dimensional Euclidean space E
4
.

60
3.5.1 Conservation Elements and Solution Elements
The spatial computational domain is divided into non-overlapping convex
hexahedrons and any two neighboring hexahedrons share a common face. In Fig. 3.10, Q
is the centroid of a typical hexahedron B
1
B
2
B
3
B
4
B
5
B
6
B
7
B
8
(hereafter referred to as
hexahedron Q). Each of the central hexahedrons six neighboring hexahedrons is
arbitrarily assigned an identification index l = 1, 2, , 6. The centroid of the lth neighbor
is denoted by A
l
(hereafter the lth neighbor is referred to as hexahedron A
l
). As an
example, Fig. 3.10 shows the hexahedrons Q and A
1
with common face B
1
B
4
B
8
B
5
.

Figure 3.10: A schematic of the 3D spatial mesh in the x-y-z space.

61
The idea of definitions of CE and SE, and the evaluation of the marching
variables are similar to that in 2D cases. However, due to the complexity in illustrating
objects in 3D space, we have the following geometric definitions first.
(i) The twelve triangles formed by point Q and the two end points of any of the
twelve edges of hexahedron Q is arbitrarily assigned an index j = 1, 2, 3, , 12 and
denoted by (j). For example, in Fig. 3.10, QB
1
B
2
is denoted by (1).
(ii) For any l = 1, 2, , 6, the four triangles formed by point A
l
and the two end
points of any of the four edges of the quadrilateral interface that separates hexahedrons Q
and A
l
is arbitrarily assigned an index k = 1, 2, 3, 4, and denoted by (k, l). For example,
in Fig. 3.10, A
1
B
1
B
4
, A
1
B
4
B
8
, A
1
B
8
B
5
, and A
1
B
5
B
1
are denoted by (1, 1), (2, 1),
(3, 1), and (4, 1), respectively.
(iii) The 24-faced polyhedron B
1
B
2
B
3
B
4
B
5
B
6
B
7
B
8
A
1
A
2
A
3
A
4
A
5
A
6
and its centroid
are denoted by V(24) and Q
*
, respectively. Hereafter Q
*
(marked by a cross in Fig. 3.10)
is referred to as the solution point associated with point Q.
(iv) For any l = 1, 2, , 6, the octahedron formed by points Q and A
l
and the four
vertices of the quadrilateral interface that separates that separates hexahedrons Q and A
l
is
denoted by V(8; l).

62
In the space-time computational domain, again we assume that at the nth time
level t = nt (n = 0, 1/2, 1, 3/2, ) and let Q, Q', and Q, respectively, be the points on
the nth, (n1/2)th, and (n+1/2)th time levels with point Q being their common spatial
projection. Other space-time mesh points such as Q
*
and
'*
Q ; B
k
,
'
k
B , and
"
k
B , k = 1,
2, , 8; and A
l
,
*
l
A ,
'
l
A , and
'*
l
A , l = 1, 2, , 6, are defined similarly. Due to due to the
difficulty in visualizing objects in E
4
, the above defined points are not shown. Refer to
Fig. 3. for their counterparts in E
3
. Also we have the following definitions and
discussions on objects in E
4
.
A hyperplane in E
4
is defined by a linear equation
1 2 3 4 0
a x a y a z a t a + + + = ,
2 2 2 2
1 2 3 4
(( ) ( ) ( ) ( ) 0) a a a a + + + , (3.61)
where a
k
, k = 0, 1, 2, 3, 4, are constants. As a result, a hyperplane in E
4
is a
three-dimensional subspace with unit normal vector
1 2 3 4
2 2 2 2
1 2 3 4
( , , , )
( ) ( ) ( ) ( )
a a a a
a a a a
=
+ + +
n . (3.62)
Note that a hyperplane segment, by definition, is a bounded region of a hyperplane. There
are two types of hyperplane segments involved in the definition of SEs (see below).
(i) A hyperplane segment of type I, denoted by (V; t
c
), is formed by all the points
(x, y, z, t) that satisfy the conditions t = t
c
and (x, y, z) V, where t
c
is a constant and V

63
is a 3D spatial region. It can be shown that: the area of (V; t
c
) is the volume of V; the
unit normal vector to (V; t
c
) is (0, 0, 0, 1); and the coordinates of the centroid of (V;
t
c
) are (x
c
, y
c
, z
c
, t
c
), where (x
c
, y
c
, z
c
) are the coordinates of the centroid of V.
(ii) In contrast, a hyperplane segment of type II, denoted by (S; t
-
, t
+
), is formed
by all the points (x, y, z, t) that satisfy the conditions (x, y, z) S and t t t
+
, where
S denotes a spatial plane segment and t
-
and t
+
(t
-
< t
+
) are constants. Note that every
point (x, y, z) on the spatial plane segment S satisfies a linear equation of the form
1 2 3 0
c x c y c z c + + = ,
2 2 2
1 2 3
(( ) ( ) ( ) 0) c c c + + , (3.63)
where c
k
, k = 0, 1, 2, 3, are constants. It can be shown that: the area of (S; t
-
, t
+
) is the
area of S multiplied by (t
+
- t
-
); the unit normal vector to (S; t
-
, t
+
) is (n, 0), where n is
the unit normal to the spatial plane segment S, i.e.,
1 2 3
2 2 2
1 2 3
( , , )
( ) ( ) ( )
c c c
c c c
=
+ +
n ; (3.64)
the coordinates of the centroid of (S; t
-
, t
+
) are (x
c
, y
c
, z
c
, (t
-
+t
+
)/2), where (x
c
, y
c
, z
c
) are
the coordinates of the centroid of S.
(ii) A hypercylinder in E
4
, denoted by (V; t
-
, t
+
), is formed by all the points (x, y,
z, t) that satisfy the conditions (x, y, z) V and t t t
+
, where V is a
three-dimensional spatial region and t
-
and t
+
(t
-
< t
+
) are constants.

64
With the above preliminaries, the six BCEs of point Q, denoted by CE
l
(Q), l = 1,
2, , 6, are defined to be the hypercylinders (V(8; l); t
n-1/2
, t
n
), l = 1, 2, , 6,
respectively. The CE of point Q, denoted by CE(Q), is defined to be (V(24); t
n-1/2
, t
n
),
i.e., the union of the above six BCEs. Moreover, the solution element of point Q
*
,
denoted by SE(Q
*
), is defined to be the union of CE(Q), and ((j); t
n
, t
n+1/2
), j = 1, 2,
3, , 12, and their immediate neighborhoods.
3.5.2 Approximations within a Solution Element
To proceed, denote the set of the space-time mesh points whose spatial
projections are the centroids of the hexahedrons that fill the 3D spatial computational
domain by , and denote the set of their corresponding solution points by
*
. For any
Q
*

*
and (x, y, z, t) SE(Q
*
), u
m
(x, y, z, t), f
m
(x, y, z, t), g
m
(x, y, z, t), q
m
(x, y, z, t),
and h
m
(x, y, z, t) are approximated by
*
m
u (x, y, z, t; Q
*
),
*
m
f (x, y, z, t; Q
*
),
*
m
g (x, y, z, t;
Q
*
),
*
m
q (x, y, z, t; Q
*
), and
*
m
h (x, y, z, t; Q
*
), respectively (see below). For any m = 1, 2,
3, 4, 5, 6, let
* *
* * * * *
( , , , ; ) ( ) ( ) ( ) ( ) ( )
m m Q mx Q Q my Q Q
u x y z t Q u u x x u y y + +
* * *
( ) ( ) ( ) ( )
n
mz Q Q mt Q
u z z u t t + + , (3.65)
* *
* * * * *
( , , , ; ) ( ) ( ) ( ) ( ) ( )
m m Q mx Q Q my Q Q
f x y z t Q f f x x f y y + +
* * *
( ) ( ) ( ) ( )
n
mz Q Q mt Q
f z z f t t + + , (3.66)

65
* *
* * * * *
( , , , ; ) ( ) ( ) ( ) ( ) ( )
m m Q mx Q Q my Q Q
g x y z t Q g g x x g y y + +
* * *
( ) ( ) ( ) ( )
n
mz Q Q mt Q
g z z g t t + + , (3.67)
* *
* * * * *
( , , , ; ) ( ) ( ) ( ) ( ) ( )
m m Q mx Q Q my Q Q
q x y z t Q q q x x q y y + +
* * *
( ) ( ) ( ) ( )
n
mz Q Q mt Q
q z z q t t + + , (3.68)
and
* * * * * *
( , , , ; ) ( ( , , , ; ), ( , , , ; ),
m m m
x y z t Q f x y z t Q g x y z t Q h
* * * *
( , , , ; ), ( , , , ; ))
m m
q x y z t Q u x y z t Q , (3.69)
be the 3D extension of Eqs. (3.41)-(3.44). Note that, in this section any notation that has a
2D counterpart and is not explicitly defined is assumed to be defined similarly.
In addition, we assume that, for any (x, y, z, t) SE(Q
*
), and any m = 1, 2, , 6,
* * * *
( , , , ; *) ( , , , ; *) ( , , , ; *) ( , , , ; *)
m m m m
u x y z t Q f x y z t Q g x y z t Q q x y z t Q
t x y z

+ + +


( *)
m
Q = . (3.70)
Note that the source term
*
m
is assumed constant within SE(Q
*
) and the value of
*
m

is determined by
*
( )
m Q
u only. Thus, for any m = 1, 2, , 6,
* * * * *
( ) ( ) ( ) ( ) ( )
mt Q mx Q my Q mz Q m Q
u f g q = + . (3.71)

66
However, like in the 1D case, to solve Euler equations with stiff source term, for any l =
1, 2, 3, 4,
'*
( )
l
mt
A
u are evaluated excluding the source term effect, i.e.,
'* '* '* '*
( ) ( ) ( ) ( )
l l l l
mt mx my mz
A A A A
u f g q = . (3.72)
Using the equations given above, it can be shown that, for the current 3D case, the only
independent marching variables associated with the space-time solution point Q
*
, are
*
( )
m Q
u ,
*
( )
mx Q
u ,
*
( )
my Q
u and
*
( )
mz Q
u , m = 1, 2, , 6.
3.5.3 Evaluation of u
m

To evaluate space-time flux in E
4
, consider the following preliminaries first. Let
be a hyperplane segment lying within SE(Q
*
). Let A be the area of , (x
c
,

y
c
, z
c
, t
c
) be the
coordinates of the centroid of , and n be a unit vector normal to . Then it can be shown
that
* * *
( , , , ; )
m m c c c c
d x y z t Q A

h s h n, (3.73)
where ds = d n with d being the area of a surface element on .
The boundary of CE(Q) is formed by the top face (V(24); t
n
), the bottom
face (V(24); t
n-1/2
), and the 24 side faces ((k, l); t
n-1/2
, t
n
), k = 1, 2, 3, 4 and l = 1,
2, , 6. Since V(24) is the union of V(8; l), l = 1, 2, , 6, the bottom face is the union of
((V(8; l); t
n-1/2
)), l = 1, 2, , 6. From the above observations, one concludes that the
boundary of CE(Q) belongs to the union of SE(Q
*
) and SE(
'*
l
A ), l = 1, 2, , 6.

67
Specifically, (V(24); t
n
) belongs to SE(Q
*
), and for each l = 1, 2, , 6, (V(8; l); t
n-1/2
)
and ((k, l); t
n-1/2
, t
n
), k = 1, 2, 3, 4, belong to SE(
'*
l
A ).
To proceed, we evaluate the surface vector (i.e., the unit outward normal vector
multiplied by the area) of every boundary face of CE(Q) first. Let V denote the volume
of V(24), i.e., the area of the top face (V(24); t
n
) of CE(Q), then because the unit
outward normal vector of this face is (0, 0, 0, 1), its surface vector is (0, 0, 0, V).
Then consider the bottom face of CE(Q). For any l = 1, 2, , 6, let V
l
and (x
l
, y
l
,
z
l
) denote the volume and the spatial coordinates of the centroid of any V(8; l),
respectively. Then the surface vector and the coordinates of the centroid of (V(8; l); t

n-1/2
) are (0, 0, 0, - V
l
) and (x
l
, y
l
, z
l
, t
n-1/2
), respectively. Moreover, since the bottom face
is the union of ((V(8; l); t
n-1/2
)), l = 1, 2, , 6, we have
6
1
l
l
V V
=
=

.
Finally consider the side faces. Let
l
k
S , (
l
kx
n ,
l
ky
n ,
l
kz
n ), and (
l
k
x ,
l
k
y ,
l
k
z ),
respectively, denote the area, the spatial unit outward normal vector, and the coordinates
of the centroid of any (k, l). Then, based on the discussion on the hyperplane segment of
type II, one obtains that the surface vector, and the coordinates of the centroid of the
side face ((k, l); t
n-1/2
, t
n
), respectively, are (t/2)
l
k
S (
l
kx
n ,
l
ky
n ,
l
kz
n , 0) and (
l
k
x ,
l
k
y ,
l
k
z , / 4
n
t t ).
By using Eq. (3.73), the flux of
*
m
h leaving each boundary face of CE(Q) can
then be evaluated in terms of the independent marching variables at points Q
*
and
'*
l
A , l
= 1, 2, , 6. For example, because (x
Q*
, y
Q*
, z
Q*
, t
n
) are the coordinates of the centroid Q
*


68
of the top face of CE(Q);
* *
* * *
( , , , ; )
n
m Q Q Q
u x y z t Q =
*
( )
m Q
u (see Eq. (3.65)); and the
surface vector of the top face is (0, 0, 0, V), the flux of
*
m
h leaving CE(Q) through its
top face is
*
( )
m Q
u V.
To proceed, we employ local space-time flux balance over CE(Q):
*
( ( )) ( )
m m
S CE Q CE Q
d dV =

h s
v
, m = 1, 2, , 6. (3.74)
With the aid of the above preliminaries, it can be shown that
6
* *
1
1
( ) ( )
2
l
m Q m Q m
l
t
u R
V
=
| |
|
\ .

, m = 1, 2, , 6, (3.75)
where, for any m, l = 1, 2, , 6,
l
m
R =
* 1/ 2 '*
( , , , ; )
l l l l n
m l
V u x y z t A

4
* '*
1
[ ( , , , / 4; )
2
l l l l l n
k kx m k k k l
k
t
S n f x y z t t A
=

* '* * '*
( , , , / 4; ) ( , , , / 4; )]
l l l l n l l l l n
ky m k k k l kz m k k k l
n g x y z t t A n q x y z t t A . (3.76)
Note that the functions
* '*
( , , , ; )
m l
u x y z t A ,
* '*
( , , , ; )
m l
f x y z t A ,
* '*
( , , , ; )
m l
g x y z t A , and
* '*
( , , , ; )
m l
q x y z t A are defined using Eqs. (3.65)-(3.68), respectively, with the symbols Q
*

and t
n
in these equations being replaced by
'*
l
A and t
n-1/2
, respectively. Given the values
of the marching variables at t = t
n-1/2
, the RHS of Eq. (3.75) can be readily calculated.
Note
*
( )
m Q
is a function of
*
( )
m Q
u . Therefore
*
( )
m Q
u can then be solved with
Newtons iteration method.

69
3.5.4 Evaluation of u
mx
, u
my
and u
mz

Like in the 2D case,
*
( )
mx Q
u and
*
( )
my Q
u are evaluated with a similar
finite-difference approach. First, we perform a spatial translation of the polyhedron
* * * * * *
1 2 3 4 5 6
A A A A A A so that the centroid of the resulting new polyhedron
1 2 3 4 5 6
o o o o o o
A A A A A A
coincides with Q
*
. Let the centroid of the polyhedron
* * * * * *
1 2 3 4 5 6
A A A A A A and its spatial
coordinates be denoted by A
*
and (
*
A
x ,
*
A
y ,
*
A
z ), respectively, and let x = x
Q*
- x
A*
, y
= y
Q*
- y
A*
, and z = z
Q*
- z
A*
. Then (
o
l
A
x ,
o
l
A
y ,
o
l
A
z ), the spatial coordinates of
o
l
A , l = 1,
2, , 6, are given by
* o
l l
A A
x x x = + ,
* o
l l
A A
y y y = + , and
* o
l l
A A
z z z = + . (3.77)
To proceed, let
* '*
( ) ( , , , ; )
o o o o
l l l l
n
m m l
A A A A
u u x y z t A , (3.78)
*
( ) ( )
o
l
l
m m m
A Q
u u u , (3.79)
and
* o
l
l
A Q
x x x ,
* o
l
l
A Q
y y y ,
* o
l
l
A Q
z z z , (3.80)
for any m, l = 1, 2, , 6.
Next, consider the vertex B
1
depicted in Fig. 3.10, which is the common vertex of
hexahedron Q and three of its neighbors. Let the identification indices l of these three

70
neighbors be 1, 2, and 3. Then, for any m = 1, 2, , 6, consider the four points in the
xyzu space with the coordinates (x
Q*
, y
Q*
, z
Q*
,
*
( )
m Q
u ) and (
o
l
A
x ,
o
l
A
y ,
o
l
A
z , ( )
o
l
m
A
u ), l
= 1, 2, 3. It can be shown that the values of u / x, u / y, and u / z on the hyperplane
that intercepts the above four points are given by
( )
*
( )
/
l
mx x
Q
u ,
( )
*
( )
/
l
my y
Q
u ,
( )
*
( )
/
l
mz z
Q
u ( 0 ), (3.81)
where
1 1 1
2 2 2
3 3 3
x y z
x y z
x y z




, (3.82)
and
1
1 1
2
2 2
3
3 3
m
x m
m
u y z
u y z
u y z




,
1
1 1
2
2 2
3
3 3
m
y m
m
x u z
x u z
x u z




,
1
1 1
2
2 2
3
3 3
m
z m
m
x y u
x y u
x y u




.
(3.83)
Note that = 0 if and only if the spatial projections of
1
o
A ,
2
o
A ,
3
o
A , and Q
*
are
coplanar. Similarly, for each k = 2, 3, , 8,
( )
*
( )
k
mx Q
u ,
( )
*
( )
k
my Q
u , and
( )
*
( )
k
mz Q
u , are defined
by replacing B
1
with B
k
.
With the above preliminaries, for each m = 1, 2, , 6,
*
( )
mx Q
u ,
*
( )
my Q
u , and
*
( )
mz Q
u can then be evaluated by

71
( )
*
8
( )
*
1
1
( )
8
k
mx Q mx
Q
k
u u
=

,
( )
*
8
( )
*
1
1
( )
8
k
my Q my
Q
k
u u
=

,
( )
*
8
( )
*
1
1
( )
8
k
mz Q mz
Q
k
u u
=

.
(3.84)
For a flow with discontinuities, the above equation may be replaced by a re-weighting
procedure, i.e.,
( ) ( )
( )
*
8
( ) ( )
1
* 8
( )
1
( )
k k
m mx
Q
k
mx Q
k
m
k
W u
u
W
=
=
(
(

,
( ) ( )
( )
*
8
( ) ( )
1
* 8
( )
1
( )
k k
m my
Q
k
my Q
k
m
k
W u
u
W
=
=
(
(

, and
( ) ( )
( )
*
8
( ) ( )
1
* 8
( )
1
( )
k k
m mz
Q
k
mz Q
k
m
k
W u
u
W
=
=
(
(

, (3.85)
where 0 is an adjustable constant,
( ) ( ) ( )
* * *
2 2 2
( ) ( ) ( ) k k k
mk mx my mz
Q Q Q
u u u
( ( (
+ +
( ( (

, m = 1, 2,, 6, k = 1, 2,, 8,
(3.86)
and for each k,
( ) k
m
W is the production of
1 m
,
2 m
, ,
8 m
excluding
mk
.
In particular, if for any k = 1, 2, , 8,
mk
= 0, then
*
( )
mx Q
u and
*
( )
my Q
u are set
to be 0. Usually = 1 or = 2, and Eqs. (3.85) reduce to Eqs. (3.84) if = 0.


72
CHAPTER 4
PARALLEL COMPUTATION
4.1 Introduction
For large size two- and three-dimensional flow simulations, the demand for the
computer resource of large memory and reasonable turn-around time is much more than a
single personal computer (PC) can provide. To solve this problem, one can employ
powerful super computers, which have been employed in scientific computing for tens of
years. However, they may be too expensive to small research groups. Fortunately,
benefiting from the dramatically fast improvement of the performance of PCs and
network equipments, having a homemade and affordable super computer exclusively for
our own group is not a dream anymore. A Beowulf system, which is constructed with the
available PC components, is such an implementation. This type of computer systems has
become more and more popular.
There isn't a specific computer or a software package for the so-called Beowulf
system. A Beowulf system is a cluster of PCs interconnected by a Local Area Network
(LAN). These PCs are running an open-sourced UNIX-like operating system, such as a
Linux distribution or the Free BSD, and executing parallel applications, powered by an


73
industry standard parallel programming models, such as Message Passing Interface (MPI).
In this setting, these PCs could work together, share their memory, provide united
computing power, and solve the same problem concurrently and efficiently. Please refer
to Fig. 4.1 for a schematic of a Beowulf system.

Figure 4.1: A schematic of a Beowulf cluster.
Nowadays, a high-performance PC costs less than two thousand US dollars.
Compared with the commercial super computers, this price is about one percent or less.
With more than twenty years of continuously improvement, a typical PC nowadays has
very respectable computing power. Intel and AMD manufacture processors which could
run at a frequency nearly 4 GHz with many power enhancement features, e.g.,


74
hyper-threading technology, multiple-branch prediction, dataflow analysis, speculative
execution, multiple independent bus, etc. With the help of a fast system bus and a fast
memory sub system, which provides a bandwidth of up to more than 10 GB/s to match
the speed of modern CPUs, a PC could outperform many much more expensive UNIX
based workstations. Because of the flexibility of network connection, a PC cluster is
highly scalable. PC nodes could be easily added to the cluster for more computing power,
or simply replace the malfunctioning ones. A large cluster could be divided into smaller
ones. Or, on the contrary, small clusters could be combined to form a large cluater.
Moreover, due to the inherent compatibility of commodity PC components, parts for a
Beowulf cluster constituted by PCs are readily available. Thus, such a system is easy to
be maintained or upgraded.
Because of its availability, scalability and cost-effectiveness, a Beowulf cluster
could be a viable platform for CFD applications. For example, a cluster with 24 PC nodes
and 1 GB memory on each node has total 24 GB memory. The hardware of the cluster
costs about 30K dollars. This modest system is capable of simulating fairly large CFD
problems, with more than 30 millions mesh nodes. Depending on the nature of the
problems, the turn around time of such a typical calculation ranges from one week to less
than a month.
Effective use of a Beowulf system requires a proper distribution of simulation
tasks among the available processing nodes. A common approach is to decompose the
computational domain into a number of sub-domains, followed by assigning the
computational tasks of each sub-domain to a computer node. All computer nodes execute


75
the same CFD solver simultaneously but on different sub-domains. These computer
nodes communicate with each other to exchange intermediate numerical results on the
sub-domain boundaries at the end of each iteration step. Finally, the solution on all the
sub-domains is consolidated to produce the complete solution.
Optimized domain decomposition would balance the computational workload and
memory occupancy among processing nodes, hence benefit the system performance. It
should be done carefully such that no node is overloaded while other nodes are starving
for work. In addition to evenly distributed task assignments, the minimization of the
communication among nodes, which is usually the source of computation overhead, is
also critical to the overall performance of the parallel computing. For CFD applications,
the above two requirements are the main concern of domain decomposition.
It is intuitively and straightforward to split a regular domains, usually just a
simple geometric partition. For irregular computational domains, we use an open-sourced
software METIS (http://www-users.cs.umn.edu/~karypis/metis/) for domain
decomposition. Refer to Fig. 4.2 for examples of domain decomposition.


76

(a)


(b)
Continued
Figure 4.2: Domain decomposition examples: (a) mesh for a simple three-dimensional
domain of a box, (b) tow-dimensional mesh for a flow-over-cylinder problem, (c) a
three-dimensional mesh of an annular tube; different colors represent different
sub-domains.


77
Figure 4.2 continued

(c)


78
4.2 Hardware System
We have built several Beowulf systems since 1999. The one we are currently
using was built in summer 2003, which consists of thirty-two computer nodes, each of
them has an Intel Pentium IV 2.8 GHz CPU, 1 GB dual channel DDR 400 memory and an
onboard gigabit network card. The computational tasks are assigned by the head node,
which is in charge of the running of the whole parallel program. All computer nodes are
connected with other nodes with a Gigabit Ethernet LAN. The Keyboard/Video/Mouse
(KVM) switches provide the access to a specific node of the cluster with the consoles.
Outside access to the cluster is via the switch/router or the additional network interface on
the head node. The system is connected to the OSU campus network. With the network
utilities, one can access and control every node remotely via the head node. However, for
the security reason, only the secured shell (SSH) protocol is supported at the headnode
and the visit from outside campus may be limited.
There are plenty of hardware components readily available in the market for
building the computer cluster. We must choose suitable parts for our CFD applications.
The most important factors to be considered for building the cluster are (i) reliability, (ii)
performance and (iii) cost-effectiveness. To build a reliable system we usually choose
well-known and credible brands for the core parts and adopt only mature techniques or
those being tested by ourselves. To save money, we do not buy the fastest processors
available on the market, which are usually overpriced due to low availability and
manufacturer marketing strategy. For example, in 2003, the newly developed 3.2 GHz


79
Pentium CPU may be a perfect choice. We decided to use the 2.8 GHz Pentium IV CPU
and saved more than 50% for the CPU, motherboard, and the chipset.
We spend necessary money on the core components such as high quality memory
and motherboards. To achieve decent performance, the most important thing is matching
the speed of the components, which is also a way to get saving. The speed of the system
is not only determined by the processors, it also relies on the supporting sub-systems like
memory, system bus, hard drive, network, etc. All above parts cooperate to execute the
assigned task. Weak link at any point could compromise the overall performance of the
whole system. If a processor is starving for instructions and data, and the memory and
system bus cannot feed the processor fast enough, the investment to the processor is a
waste. Therefore, we must look into the system to find the bottlenecks. However, it
depends on the work performed on the system. What kind of task is it? Is it a
CPU-intensive application or a memory-intensive application? In other words, we must
analyze our applications and optimize the construction of the system to achieve high
performance. In purchasing less important parts and/or peripheral devices, we employed
cheaper parts to save money.
According to our analysis and experiences, our applications in two- and
three-dimensional CFD simulations are high memory consuming and high network
bandwidth consuming. This is the reason why for Intel based system (like our current
system) we prefer single CPU nodes, instead of dual CPU nodes. That is, although
current memory technique can offer a bandwidth of up to more than 10 GB/s, it may not
be enough to feed two processors, due to the limitation of the traditional structure of the


80
Intels x86 systems. However, with the application of new system structure, such as the
hyper transport technology developed by AMD, which enable currently fastest
microprocessors to use system memory more efficiently, it has become realistic for us to
put two or even more CPUs into one box. New generation of memory like DDR2, which
can provide a higher bandwidth than the predecessors also increases the possibility of
multi-processors motherboard. To match the speed of the fast processor and the fast
memory sub-system, and to accelerate the access to the network file system (NSF) from
all the nodes, fast hard drives (usually with SCSI interface though relatively expensive),
must be used in the head nodes.
Another concern is the network performance. Parallel computational in CFD
simulations require frequent exchange of huge amount of data among computer nodes.
High-speed communication is imperative. Fortunately, gigabit networking is quite
affordable since two to three years ago. However, it is not just a no-brainer selection. The
truth is, that unless your system provides a wide enough channel from the network
interface to the processor and memory sub-system, high performance network cannot
work with its full power and therefore the network connection is inefficient. For example,
in a traditional Intels x86 system, due to the limited bandwidth of the PCI bus and the
narrow link between the north bridge and the south bridge, a gigabit network card cannot
exert its strength, though it may be more than enough to most office use. Refer to the
cyan line in Fig. 4.3. The gigabit network card can only offer a throughput of about 600
Mbps due to the limit of the system structure. One may consider using more advanced


81
and complex motherboards to support server-level processors, which usually offer high
bandwidth and PCI-X bus, but this solution is fairly expensive.

Figure 4.3: Network performance of two gigabit Ethernet systems with Intels traditional
x86 architecture and CSA, respectively, and a fast Ethernet system with single, double
and triple fast Ethernet network interface cards (NICs) installed on each computer node,
where the channel bonding technology is used for the multiple card cases. Benchmarked
by netperf 2.1.


82
In summer 2003 when we built the current system, the new announced Intel 875P
chipset and the Communication Streaming Architecture (CSA) which attaches the
network interface directly to the north bridge, provided us an alternate solution which has
been proved successful on our system. Refer to the black line in Fig. 4.3 for the
performance. New developed chipsets and system structure nowadays provides more
options. The key point is that the system bandwidth must meet the requirement of the fast
processors, network and graphics sub-systems to enable their full power.

Figure 4.4: A schematic of channel bonding technology applied on a Fast Ethernet system
with dual network cards used on each node.


83
There is another way to increase network bandwidth. With the support of
channel-bonding technology in recent Linux kernel, one can bond multiple Ethernet
cards into a faster virtual network interface to get higher data transmission rate. Using
channel bonding, the data is divided into halves (if two cards are used in each end) before
sending. After transferred via two separate networks, the two parts are integrated again at
the destination node to complete the delivery. Refer to Fig. 4.4 for a schematic of channel
bonding. We had successful employed this technology on one of our previous systems in
2001. Refer to Fig. 4.3 for the benchmark of the network performance. The precondition
is that the system bus must be wide enough to handle these multiple network channels.
Recent development of AMD 64-bit processors and supporting chipsets provides
us even more choices in building a Beowulf system for super-fast float point calculation
capability and the almost unlimited memory capacity support. Usually a 32-bit system
can only support up to 4 GB memory on a single system. On the other hand, theoretically,
a 64-bit system could support 16 million Terabytes, i.e., almost unlimited memory. As the
information technology is dashing ahead continuously, the computing power of a
Beowulf cluster will continuously progress by riding on the technical revolution and
development.
The author is not going to review all the decision we made and all the techniques
we employed to build the current system. The above discussion is just to present some
general guidelines in building a fast, cost-effective system for our specific CFD work.
Because the information technology develops so fast, a current new technology may be
obsolete and quickly replaced by a futuristic technology in a very short period of time.


84
However, the key idea here in making an efficient and balanced system is that no obvious
bottleneck subsystem should exist. To date, the bottleneck of the system for the overall
performance of CFD applications have related to the networking speed. Identify the
performance bottleneck will always critical in building a high-performance and
cost-effective computer cluster.
4.3 Software and Programming
The software environment of our calculations usually includes three parts: (i) the
operating system, (ii) the compiler, and (iii) the MPI library. Commercial cluster software
packages, e.g., Portland Group, may contain additional components such as debugging
tools and workload and throughput manager. However, these peripheral utilities are
beyond the present discussion. Most clusters use a certain version of distributed Linux as
the operating system, e.g., RedHat and SuSE. Linux is a free as open-sourced code. In
general, it is reliable, sturdy, powerful, and very efficient in executing binary code.
Our CFD codes were written in high-level languages. To date, all our production
codes were written by Fortran 95. In our group, we are in the process of rewriting the
codes by using C++. The use of high-level language requires good compilers to exert
all the strength of the hardware platform. There are many free compilers from GNU,
which is the project of Free Software Foundation. One could also try commercial
compilers for special features provided by vendors. On top of the compiler, a parallel
programming model is required. For cluster computing, we use MPI. Two popular
implementations of MPI are MPICH and Lam/MPI. Both of them offer a group of


85
subroutines in the setting of Fortran and/or C++ to support data exchange among
processors.
In parallel computing, the value of the speedup ratio is used to assess the system
performance in solving specific problem. The speedup ratio is the reciprocal of the ratio
of the time used in solving an identical problem by single process and multiple processes.
For example, a half time saving parallel performance designates a speedup of 2. Table 4.1
shows the wall clock time used for 80 iterations by our Euler solver for simulating the
two-dimensional PDE problem on our cluster, and the corresponding speedup, when 1, 4,
8, and 16 processor nodes are employed, respectively. The speedup and
number-of-processes relation usually presents a rising curve beginning from the (1, 1)
point on the x-y coordinate system. Refer to Fig. 4.5. The ideal linear speed-up (along the
line y = x or y/x = 1) is usually impossible due to the communication overhead.
Sometimes the curve even turns downwards because of poor memory or network
bandwidth. We have seen this kind of poor performance by using certain dual-CPU
system, in which the memory bandwidth could not feed two CPUs on board and the
system performance suffered greatly when a large amount of data are being read and
write. Thus, building a high-performance and balanced system and tuning it up are
critical to successful parallel computing.


86

Processes Time for 80 iterations (s) Speedup
1 501 1
4 131 3.82
8 67 7.48
16 38 13.18
Table 4.1: Parallel computing performance.

Figure 4.5: A schematic of the speedup.


87
Compared with a regular sequential program, writing a parallel program needs
more skills and efforts. The program must be well organized in task load and data
exchange to maximize the performance. If not done carefully, the program may suffer
waiting for the data to be received in execution, and the overall performance of the
program could be compromised. Evenly-distributed task load could be achieved by
careful and accurate domain decomposition. Having shortest shared boundaries between
neighboring sub-domains will minimize the amount of data to be exchanged. The
developed program must be organized and written accordingly.
To improve the performance in communication, first we increase the network
bandwidth by using appropriate hardware configuration in conjunction with some tuning
work. In the CFD code, we have to use the so-called non-blocking communication
technology, which allows data massage passing being carried out simultaneously with
respect to number crunching. In using non-blocking feature in MPI, a proper message
size is critical to the communication performance as well as the accuracy in execution.
Refer to the black line in Fig. 4.3 for a typical performance curve, illustrating the network
throughput as a function of the message size. As can be seen, the network performance
obviously depends on the message size. To get high performance, larger size messages
are favored. This could be due to the network latency, and it can hardly be improved with
a faster transmitting rate. Since memory access is usually much faster than data
transmission via network, if necessary, data should be packed before sending and
unpacked after receiving the data to achieve large message size.


88
As a final note, CFD schemes based on explicit time marching method may
benefit more from the above discussed domain decomposition and parallel computing
than the implicit methods. The Space-Time CESE method, as an explicit time marching
method, is very suitable for parallel computing. For details of the CESE scheme, please
refer to Chapter 3.
4.4 System Specifications
As an integral part of my Ph.D. training, I have built four generations of clusters
in the past six years; three systems in Wayne State and one at OSU. These parallel
systems not only saved considerable time for calculation, bus also made it possible for
me to perform large-size simulations by using millions of mesh cells. This experience
also convinced me that the use of parallel computing technology is indispensable for
future CFD simulations. The experience I obtained from building these systems will
benefit the further development in building and maintaining such home-made super
computers our CFD group in the future.
Since there is no standard cluster system, and the performance difference between
one cluster to that of another cluster could be quite dramatic, my experience in building
these clusters could be summarized in the following statement. As a CFD expert, one has
to consider the option of building a small system exclusively for his or her use. To say the
least, since none of the commercial cluster is tailored for CFD use, one has to know how
to customize or upgrade a commercial system available on the market. Applications in
different fields or for different purposes may require very different system configurations.


89
Even for the same purpose, e.g., CFD, the optimized configuration of the system will
depend on the CFD numerical scheme or method employed.
In the following, I provided the details of all four systems that I built. The
information here could be used as reference for people who want to conduct large-size
CFD calculations. The change from the earlier generation to the latest one also shows the
development of the Beowulf computer cluster in the past several years.
The first cluster was built in the year 1999, consisting of two nodes of Dell
workstation with dual Intel Xeon 450 MHz CPU. The cost was 24K US dollars. The
purpose was mainly to investigate the feasibility and the performance. Some simple
parallel programs were tested on this platform. Refer to Fig. 4.6 for the picture and Table
4.2 for the specifications. We note all dollar amounts refer to the costs of hardware only.
We always bought components and assembled the systems by ourselves.

Figure 4.6: The 1
st
generation cluster in 1999.


90

CPU: Dual Xeon 450 MHz CPU
Memory: 768 MB SDRAM ECC
Network card: 3Com 905 Fast Ethernet card
Operating
System:
SuSE Linux 6.1
Compiler: GNU Fortran compiler (included in all Linux distribution)
MPI: MPICH 1.2
Table 4.2: Specifications of the1
st
generation cluster in 1999.
The second-generation cluster is a homemade fully functional system, including
eight dual CPU (Intel Pentium III 450 MHz) computer nodes and a Fast Ethernet LAN
at 100 Mbps. It was built in late 1999 with 25K US dollars. Refer to Fig. 4.7 for the
pictures and Table 4.3 for the specifications. This cluster, as the second-generation
cluster in our group, was used for code development and real calculations, including
simulations of the PDE plume dynamics and three-dimensional ZND detonation
problems. With the aid of this powerful machine, we were able to conduct very large size
calculations within a relatively short time. For example, in May 2000, we ran a
three-dimensional ZND detonation calculation by using a mesh of 1.8 million mesh cells.
The typical running time for one case was only about twelve hours.
Generally speaking the performance of the cluster was satisfactory. We then
found out that the network throughput is the bottleneck of the system performance. For
the above mentioned mesh, usually message packages of several megabytes need to be


91
constantly exchanged between every sender/receiver pair (computer nodes which handle
neighboring sub-domains) after every time iteration step. This bottleneck problem could
be solved partially with improved mesh decomposition, but essentially a faster network is
required. Using a gigabit network was not practical at that time due to the cost. Therefore,
we turned to the novel and fantastic channel bonding technology, which consolidates
multiple network interfaces to produce double or triple throughput. The second Fast
Ethernet card was then plugged into every computer node and a second LAN was formed
separately besides the existing one. After some software installation and tuning (channel
bonding was not supported by the Linux kernel in 2000), the new system performs well
and we resolved our bandwidth problem completely. Refer to Fig. 4.3 for the network
performance comparison.

Fig. 4.7: The 2
nd
generation cluster in 2000.


92
We also tested different communication mode with this cluster (before the
network upgrade). For example, with 4, 8, 16 processes and blocking mode used for 20
iterations in the three-dimensional ZND calculation, the wall clock time is 56, 55, and 90
seconds, respectively. While with the non-blocking mode used, the time became 50, 28,
and 23 seconds, respectively.
CPU: Dual Intel Pentium III 450MHz
Memory: 512 MB SDRAM ECC
Hard drive: 13 GB IDE for master node, 6 GB IDE for slave nodes
Motherboard: Gigabyte GA-6BXD with Intel 440BX chipset
Network card: Dual 3Com 905C Fast Ethernet (100 Mbps) with channel bonding
technology
Network
Switch:
Dual SMC-EZ1024FDT 16-port 10/100 Mbps
KVM Switch: Iogear GCS138
Operating
System:
RedHat Linux 7.1
Compiler: PGI Cluster Development Kit (CDK) 3.0
MPI: MPICH 1.2
Table 4.3: Specifications of the 2
nd
generation cluster in 2000.


93
The third cluster, built in 2002 with 22K US dollars, consisted of eight computer
nodes (equipped with Xeon 2.0 GHz CPU) and a gigabit network. Refer to Fig. 4.8 for
the pictures and Table 4.4 for the specifications. The SCSI hard drive on the head node
provides reliable and fast network file system (NFS) service for other nodes to share the
program and data. The integrated 8 GB memory made it suitable for calculations on
meshes with multi-million mesh cells. For example, in December 2002, we calculated a
transverse injection in a supersonic flow with 5.5 million mesh cells. The cluster system
finished the job within one day. In January 2003, a PDE Plume dynamics calculation on a
fine mesh with 6.4 million mesh cells lasted 88 hours for one cycle of firing (roughly
83K time steps) and we ran about 10 cycles of PDE firing. The system ran continuously
for several weeks and the performance was quite stable.



Figure 4.8: The 3
rd
generation cluster in 2002.


94
This cluster improved a lot as compared with the previous two clusters by using
the workstation level CPU and the gigabit Ethernet. The application of the first
generation of dual channel DDR memory was a plus. However, the bandwidth for
memory access was not enough to support the dual CPU. Though 266 MHz memory
chips were used, the chipset (Intel 7500) only supported the chipset in running on a
frequency of 200 MHz, which provides a memory bandwidth of 3.2 GB/s on the dual
channel. This could be good for a single CPU system, but it was not enough to
accommodate the second CPU.
When the computer nodes were equipped with the second CPU (feature supported
by the motherboard and the chipset), the performance was disappointing. For example,
for the three-dimensional calculation on a mesh with 1.8 million cells, an eight node
single CPU system spent 58.8 seconds for 20 time iterations, while the same work load
cost a four node dual CPU system (with the same CPU) 70.3 seconds. Note for regular
decomposition on a structured mesh, the two processes running in the same dual CPU
node are usually handling the neighboring sub-domains and dont need to exchange data
over the network. Thus the network load was roughly the same compared with that in the
single CPU case. Therefore, we concluded that the performance drop was due to the
limited memory bandwidth. This result also made us consider more on the speed match
of the subsystems, or the balance configuration, of the whole system.


95

CPU: Intel Xeon 2.0 GHz
Memory: 1 GB PC2100 DDR registered ECC, dual channel
Hard drive: 37 GB SCSI for master node, 40 GB IDE for slave nodes
Motherboard: Tyan S2720 UGN/GN with Intel E7500 chipset
Network card: onboard Intel Gigabit Ethernet controller 82554GC
Network
Switch:
Netgear GS516T 16-port Gigabit switch
KVM Switch: Aten CS1016 16-port
Operating
System:
RedHat Linux 7.3
Compiler: Intel Fortran compiler 6.0 for Linux
MPI: LAM/MPI 6.5
Table 4.4: Specifications of the 3
rd
generation cluster in 2002.
The fourth cluster, the one we are currently using, is built in summer 2003 with a
cost of 60K US dollars. Its equipped with 32 computer nodes with an Intel Pentium 2.8
GHz CPU and 1 GB PC3200 DDR memory on each node. Refer to Fig. 4.9 for the
pictures and Table 4.4 for the specifications. The cluster was designed to support two-
and three-dimensional simulations by using very fine meshes, e.g., up to 50 million mesh
cells.
Aided by this powerful system, we have done many three-dimensional
calculations, including detonation wave, implosion/explosion processes for detonation
initiation, supersonic cavity flows, and jet noise. The largest mesh we have tested is a
structured 40 million mesh-cell mesh for the three-dimensional ZND detonation


96
calculation. For a small mesh with couple of million cells, most work could be done
within one day by using this system. For example, using a mesh with 2.88 million cells, a
3D detonation simulation in an annulus tube needs about twenty hours. A 3D jet noise
cancellation calculation with 8 million grids can complete a firing cycle (0.11 millisecond,
or 2560 time iterations) within five hours, after the stable baseline state has been reached.


Figure 4.9: The 4
th
generation cluster in 2003 in OSU.
Based on the disappointing test of the performance of dual CPU systems used in
the third-generation cluster, computer nodes of the present cluster have only one CPU. To
obtain satisfying network performance, we chose the motherboard with the new Intel
875P chipset and real CSA network controller. There were not too many options in the
market in summer 2003 since this chipset was very new at that time. Many motherboard
manufacturers simply omitted the integration of network controller or attached the


97
controller to the PCI bus to save cost. After the hardware integration, we also have to
wait for the development of qualified Linux driver of the network controller because it
was new too. Anyway, the satisfying benchmark of the new system was the best reward
for our effort in designing the system.
CPU: Intel Pentium IV 2.8GHz
Memory: 1 GB PC3200 DDR, dual channel
Hard drive: 73 GB SCSI for master node, 80 GB IDE for slave nodes
Motherboard: Gigabyte GA-8KNXP, with Intel 875P chipset
Network card: onboard Intel 82547EI Gigabit Ethernet controller
Network
Switch:
Two SMC 8624T 24-port Gigabit switch*
KVM Switch: Aten ACS-1216AL (x2)
Operating
System:
RedHat Linux 9
Compiler: Intel Fortran Compiler 7.1 for Linux
MPI: LAM/MPI 7.0
Note *: upgraded to a Force 10 E300 network switch in Feb. 2004.
Table 4.5: Specifications of the 4
th
generation cluster in 2003.
At the time we built the cluster we had to use two 24 port SMC network switches
to interconnect the 32 computer nodes because there were no affordable network switch
with 32 gigabit ports or more in the market. Thus the 32 nodes are divided into two


98
groups with separate network switches. The bandwidth of the two switches was 4 GB/s,
which was not very good but enough for calculations on medium size meshes or
structured meshes if the domain decomposition was done carefully. This situation,
however, was improved in February 2004 when we replaced the old switches with a
Force 10 switch, which has 36 gigabit ports and may be extended to 72 ports. This
upgrade costs us another 52K US dollars. Though its expensive and made the cluster, it
also made the transition to large size unstructured mesh smooth.
Our next goal is to routinely perform numerical simulations by using meshes up
to 100 million cells. In the future, we plan to employ the AMD Opteron 64 bit CPU to
build the next cluster. AMDs Optreron and Athlon CPUs are proved to be very good for
the float point operation. With the HyperTransport technology, each CPU can have its
own memory and the bandwidth among CPUs, and the memory sub-system rate could be
up to 12.8 GB/s, which makes it possible for us to build dual CPU systems for CFD
calculations.
As to the networking, modern workstation-level motherboards usually include
two gigabit network controllers. Though we can improve the throughput by channel
bonding, it may be unnecessary because with the mesh size doubled, the size of the
boundary between neighboring sub-domains increases only about 58%. As a result, as the
number of the mesh nodes increases, the demand of high-speed network becomes less
important. Moreover, very affordable entry-level network switch nowadays could offer
32 to 48 gigabit ports. While for a cluster with a size more than 60, one may still have to
turn to high-level network switches such as Force 10. Anyway the information


99
technology grows so fast that it will certainly benefit the development of the CFD
applications continuously in the future.
4.5 Useful Links
Some links are listed below, which may provide useful information in building a
Beowulf cluster:
Beowulf: http://www.beowulf.org/
Linux: http://www.linux.org/
Red Hat Linux: http://www.redhat.com/
SuSE Linux: http://www.novell.com/linux/suse/
LAM/MPI: http://www.lam-mpi.org/
MPICH: http://www-unix.mcs.anl.gov/mpi/mpich/
METIS: http://www-users.cs.umn.edu/~karypis/metis/
Intel Compiler: http://www.intel.com/software/products/compilers/flin/
GNU Compiler: http://www.g95.org/
PGI CDK: http://www.pgroup.com/
Netperf: http://www.netperf.org/
Hardware Review: http://www.tomshardware.com/

100
CHAPTER 5
DETONATIONS
5.1 Introduction
Detonations are efficient processes of converting stored chemical energy in fuels
to useful thermal and mechanical energy for various applications. In scramjet engines,
standing detonation waves are the primary combustion mode and the associated blast
waves also enhance fuel/air mixing. In Pulsed Detonation Engines (PDE), the thrust
power is directly from the high-pressure region behind the traveling detonation wave.
The existence of the detonation waves was first recognized by Mallard and Le
Chatelier (1881) during their studies of flame propagation. Shortly afterwards, by
focusing on the thermodynamic relations between the unburned and fully burned states,
Chapman (1899) and Jouguet (1905) proposed the first theoretical model to explain the
underlying physics of detonation waves. The structure of the detonation wave was
independently proposed by Zeldovich (1940), von Neumann (1942), and Doering (1943),
i.e., the ZND model, in which they postulated a steady detonation wave consisting of a
non-reacting shock front followed by a finite-rate reaction zone. Refer to Fig. 2.1 for the
pressure profile of a ZND detonation wave. The spike at the shock front is the von
101
Neumann spike (von Neumann, 1942) and D denotes the velocity of the whole wave
structure. This important insight provided a steppingstone for further studies of
detonations in the past decades. Later, experimental evidence, however, showed that
detonation waves were often unstable with transverse wave structure, and the pressure
spike is unsteady with averaged values much higher than that predicted by the ZND
model.
Direct calculations of traveling detonation waves have always been challenging.
Fickett and Wood (1966) pioneered the numerical calculations of detonations by solving
the one-dimensional Euler equations for chemically reacting flows by using the method
of characteristics in conjunction with a shock-fitting method. Longitudinal instability
waves were accurately simulated. Taki and Fujiwara (1978) reported the first
two-dimensional simulation of detonations. They applied van Leers upwind method
(1969) for shock capturing and a two-step finite-rate model to model the chemical
reactions and the associated ignition delay. Oran, Kailasanath, and coworkers (Oran et al.,
1978, 1981; Kailasanath et al., 1985, 1986; Gamezo, 1999) applied the Flux-Corrected
Transport (FCT) method (Boris and Book, 1973) to calculate detonations and reported
detailed structures of two-dimensional detonations, including transverse waves, Mach
stems, and fish-scale soot trace. Bourlioux and Majda (1995) developed an accurate
method, composed of a high-order upwind scheme, a front tracking method, and an
adaptive refinement algorithm to calculate detonation waves. Quirk (1994) addressed a
particular deficiency of the upwind schemes when solving detonations. He also proposed
a remedy to overcome the so-called carbuncle effects. Papalexandris (1997) developed an
102
unsplit upwind method based on integrating the flow equations along the characteristic
manifolds in the space-time domain. He applied this specialized method to model oblique
detonation waves.
The first three-dimensional simulation of detonations was performed by Fujiwara
and Reddy (1989). They studied detonation propagation in co-axial annular tubes. By
using the FCT method, Williams et al. (1996) simulated three-dimensional detonations in
a square channel. They reported the evolving transverse waves and Mach stems in a
rectangular pattern. By using the CESE method, Zhang et al. (2001) and He et al. (2001)
reported the first simulated diagonal instability mode of three-dimensional detonations in
a square duct. By using a Total Variation Diminishing (TVD) scheme (Yee, 1987), Tsuboi
et al. (2002) presented similar results in their three-dimensional calculations. In all above
three-dimensional calculations, however, the computational domains were very small,
and the discussions of the numerical results were focused on the flow structure of one or
a handful of detonation cells.
In this chapter, we report direct calculations of two- and three-dimensional
detonations by using the space-time CESE. The objectives of the present work are
twofold. First, we wish to demonstrate the capability of the CESE method for calculating
detonation waves. With a moderate mesh resolution, we will show that the CESE method
is capable of capturing salient features of a propagating detonation wave in two- and
three-dimensional calculations. Second, we will show that the structure of detonation
waves is essentially three-dimensional and one has to conduct three-dimensional
simulation to catch basic features of detonation waves, e.g., the cell size of the soot traces.
103
We will clearly present the differences between the two-, and three-dimensional results.
The detonation examples and the CESE results here could be employed as benchmark
cases to test the performance of other CFD algorithms.
5.2. One-Dimensional Detonations
As a steppingstone for calculating detonation waves in multiple spatial
dimensions, we first conduct calculations of two one-dimensional detonation problems: (i)
a piston problem, which is initialized by pushing a piston into the reactant gas, and (ii)
the instability problem, in which the analytical ZND solution is used as the initial
condition with certain flow parameters chosen to simulate the known flow instability
phenomena.
Consider a long tube filled with stagnant reactant gas. A piston on one end of the
tube moves into the gas at a constant speed. Choose the piston face as the origin of the
spatial coordinate. In this coordinate system, the reactant gas charges into a closed-end
tube at a constant speed. The reflected shock wave from the piston face raises the gas
temperature and ignites the reactant. Refer to Fig. 5.1. In the present study, the specific
heat ratio 2 . 1 = , the heat of formation 50
0
= q , and the activation energy . 50 =
+
E
Results of the linear-stability analysis (Erpenbeck, 1964; Fickett and Wood, 1966;
Bourlioux and Majda, 1995) show that there is a critical overdriven factor f, below which
the detonation is unstable. For the above parameters, the critical overdriven factor is f
*
=
1.73. Two cases are considered: (i) a stable detonation with f = 2, and (ii) an unstable
detonation at f = 1.6. For the stable case, the pressure history of shock front is shown in
104
Fig. 5.2(a), in which the peak pressure of the established detonation wave remains
constant after the initiation process. For the unstable case, Fig. 5.2(b) shows galloping
pressures at the shock front when t > 50. Essentially, a longitudinal wave bounces back
and forth, leading to pulsating peak pressures at the shock front. These results, including
the values of the pulsating peak pressures and frequency of the instability wave, agree
well with Fickett and Woods results (1966).

Figure 5.1: A schematic of the piston problem.
105

(a)

(b)
Figure 5.2: Dimensionless shock front pressure history in the piston problem for 2 . 1 = ,
50
0
= q , 50
+
= E , and (a) f = 2.0, (b) f = 1.6. The horizontal line is the ZND solution.
106
We then consider the instability problem. The analytical solution of an overdriven
ZND wave is used as the initial condition. Refer to Section 2.3 for details of the ZND
analytical solution. For parameters 2 . 1 = , 50
0
= q , and 50 =
+
E , two cases are
calculated: (i) stable detonation with f = 1.8 and (ii) unstable detonation with f = 1.6. For
the stable detonation f = 1.8, Fig. 5.3(a) shows the history of shock front pressure. As can
be seen, it converges to a constant value, i.e., the shock front pressure of the ZND
analytical solution, after several cycles in the beginning.
Moreover, Fig. 5.4 shows the calculated p-v diagram as compared with the
theoretical solution. The initial condition of p = v =1 is located at the lower right corner
of the figure. The solid lines are the Hugoniot curves at various mass fraction of the
product , and the dotted line is the Rayleigh line for the specific over driven factor f. The
intersection of a Hugoniot curve and the Rayleigh line is the analytical solution of the
state corresponding to the specific across the detonation wave. The numerical result is
denoted by circles. The analytical shock path, denoted by the dashed line, is calculated by
the analytical solution of the one-dimensional Navier-Stokes equations (White, 1974).
Chemical reactions and heat release occur along the Rayleigh line between the two
Hugoniot curves of = 0 (the solution of the von Neumann spike) and = 1 (the final
product). Across the jump condition of the shock front, there are two CFD solution points.
The numerical solution shows that the artificial damping by the central differencing and
the reweighing procedure in the CESE method produces flow solution very close to the
real-gas effects inside the shock wave predicted by the one-dimensional Navier Stokes
equations.
107

(a)

(b)
Figure 5.3: Dimensionless time history of the shock front pressure in the instability
problems for 2 . 1 = , 50
0
= q , 50 =
+
E , and (a) f = 1.8, (b) f = 1.6, with the mesh
resolution 50 pts/L
1/2
. The horizontal line is the ZND solution.
108
Figure 5.3(b) shows the history of shock front pressures for an unstable
detonation wave with f = 1.6, 2 . 1 = , 50
0
= q , and 50 =
+
E . The wave length of the
oscillating pressures agrees well with that reported by Bourlioux and Majda (1995).

Figure 5.4: The dimensionless p-v diagram of the detonation for
parameters 2 . 1 = , 50
0
= q , 50 =
+
E and f = 1.8.
109
To assess the numerical accuracy of the CESE method, the same calculation is
repeated by using 5, 10, 20 and 50 grid nodes to resolve the length of the half reaction zone.
Figure 5.5 shows the peak pressures obtained by various upwind schemes as compared
with the CESE method. The x axis in Fig. 5.5 is the relative mesh spacing, defined as 5/n
with n as the number of mesh nodes for the half-reaction zone. According to Fickett and
Wood (1966), the maximum peak pressure with the above flow parameters is 98.6. As
shown in Fig. 5.5, when 5 mesh nodes per half reaction zone is used, the CESE method
produces a peak pressure about 98.2 (except the first several cycles). When fine meshes
are used, i.e., data points closer to the y axis, nuermcial results converge to 98.6.

Figure 5.5: Variation of the dimensionless peak pressure with mesh resolution for various
schemes. The relative mesh spacing is defined as 5/n with n as the number of mesh nodes
for the half-reaction zone. Flow parameters: 2 . 1 = , 50
0
= q , 50 =
+
E , and f = 1.6.
110
5.3 Two-Dimensional Detonation Waves
For the planar detonation wave, we consider a two-dimensional detonation wave
propagating into a quiescent medium. The reaction zone is initialized by the analytical
solution of the ZND wave with sinusoidal perturbations on the shock front. The flow
parameters are q
0
= 50, E
+
= 50, = 1.2, and f = 1.6. The duct is vertical and the
detonation wave is propagating upwards to consume the reactant. Let the coordinate
system fixed on the shack front. Thus, the reactant charges into the detonation front
downwards at a constant speed in this coordinate system. The width of the computational
domain is 7.5, and the height is 9.0 with the unit length as the length of the half reaction
zone. The flow conditions on the upper boundary surface are fixed according to the
condition of the unburned gas. Periodic boundary condition is imposed at the two lateral
boundaries. 54,000 quadrilateral cells are used for the computational domain.
Figure 5.6 shows the snapshots of the contour of mass fraction of the reactant,
pressure, temperature, and vorticity. The flow field is much more complex than that of
the one-dimensional detonations. The shock front is characterized by mushroom-shaped
incident shocks interacting with Mach stems. The width of the Mach stem changes
periodically due to the moving triple points at the shock front. At each collision of two
triple points, a pair of vortices, with opposite rotational directions, are created and
propagate downstream. Due to these vortices, unburned reactant is engulfed into the
reaction zone and forms unburned pockets behind the reaction zone. The continuous
burning of the unburned pockets shows the phenomena of explosions inside explosion
and greatly extends the effective reaction zone.
111

(a) (b)


(c) (d)
Figure 5.6: Two-dimensional detonation waves: (a) mass fraction of reactant, (b) pressure,
(c) vorticity and (d) temperature. The detonation wave travels upwards. The unburned gas
is on the top. The burned gas is on the bottom. Flow parameters: f = 1.6, = 1.2, q
0
= 50,
E
+
= 50.
5.4 Three-Dimensional Detonation Waves in a Duct
In this section, we report three-dimensional simulations of planar detonations in
square tubes. We will focus on the calculated cell size of the detonation soot trace, which
is the most distinct flow feature of detonation waves. First, we consider a detonation
wave in a square duct with cross section area comparable with the detonation cell size.
Second, planar detonations in a duct with much larger cross section area are considered.
112
For various flow conditions, including activation energies and heat release amount, we
examine the corresponding changes in the calculated cell sizes. Finally, we will calculate
the averaged cell size of the soot trace on the side walls. This result will be compared
with the measured data. We will also compare the three-dimensional results with the
results of the corresponding two-dimensional calculations.
In all three-dimensional calculations, the analytical solution with a small spatial
perturbation on the shock front is employed as the initial condition. The detonation waves
travel upwards to consume the fresh fuel/oxidant mixture. The upper boundary conditions
are set to be the flow conditions of the unburned gas. The reflective wall boundary
condition is used on the four lateral side walls. On the bottom surface, we impose the
non-reflective boundary condition.
In a small duct with cross section of 8 8, we conduct three-dimensional
simulation of a planar detonation wave. The unit length is the length of the half reaction
zone. A structured mesh of 6.4 million hexahedral cells was employed. The detonation
parameters are f = 1.6, = 1.2, E
+
= 50, q
0
= 50. According to the classical theories, these
parameters would trigger flow instability. Figure 5.7 shows snapshots of the pressure
contour at two different time: t
1
= 12.0 and t
2
= 13.5..
Figure 5.8 shows the temperature contour and the contour of the mass fraction of
the reactant at time t
2
. The side view of the three-dimensional contours resembles that of
a two-dimensional detonation.
113

(a)

(b)
Figure 5.7: Snapshots of pressure contour at different time for a detonation wave
propagating in a square duct: (a) t = 12.0, (b) t = 13.5. The detonation wave travels
upwards. The unburned gas is on the top. The burned gas is on the bottom. The cross
section size is 88. The unit length is the half reaction length. Flow parameters: f = 1.6,
= 1.2, q
0
= 50, E
+
= 50.
114

(a)

(b)
Figure 5.8: Snapshots of different contour for a detonation wave propagating in a square
duct: (a) temperature, (b) concentration of the reactant. t = 13.5. Cross section size: 88.
Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
= 50.
115
The triple points in the two-dimensional detonations become triple lines and the
interactions between the triple lines are much more complex than that in the
two-dimensional modeling of the same detonation wave. Moreover, we found that the
peak pressures in three-dimensional detonations are much higher than that in
two-dimensional calculations for the same flow conditions. The calculated
three-dimensional detonation structure is similar to that reported by Williams et al.
(1996).
Depending on the imposed perturbations to the ZND solution at the initial stage of
the computation, three-dimensional structure of unstable detonations in the square duct
with a cross section about the size of the detonation cell size can be classified into two
fundamental types: a rectangular structure and a diagonal structure.
The rectangular structure consists of two two-dimensional waves. The
two-dimensional waves are orthogonal to each other and they travel independently.
Therefore, the rectangular detonation structure may be described as two superimposed
two-dimensional detonations. Figures 5.9(a, b, c) are the results of the rectangular
instability structure. Snapshots of two-dimensional pressure contour on a cross section
near the shock front are plotted.
In contrast to the rectangular mode, the diagonal structure in three-dimensional
detonations is the essential mechanism in actual three-dimensional detonations with triple
points moving parallel to the diagonal lines of the cross section of the square tube. The
axes of the transverse waves are canted at 45
o
with respect to the wall. Figure 5.9(d, e, f)
116
shows snapshots of two-dimensional pressure contour on a cross section near the shock
front. Alternating diagonal waves can be clearly seen. According to our numerical results,
the instability in the diagonal mode always shows up after the flow imposed by the initial
conditions fades away. For certain initial condition employed to set up the rectangular
mode, the flow instability tends to transform to the diagonal mode after the initial stage of
the evolution. This is consistent with finding reported by Hanana et al. (2001). We remark
that if one sets up the calculation by taking advantage of symmetry with respect to the
central x-z and y-z planes, e.g., Williams et al. (1996), one will only obtain the instability
wave in the rectangular mode.
117



(a) (d)


(b) (e)


(c) (f)
Figure 5.9: Snapshots of pressure contour on a cross section near the shock front of
detonation waves propagating in a square duct: (a), (b), (c) rectangular mode; (d), (e), (f)
diagonal mode. Cross section size: 88. Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
=
50.
118
Figure 5.10(a) shows calculated soot trace of the detonation of rectangular mode
on one side wall. The soot trace is made visible by a continuous record of the shape of the
shock front. The soot trace is the path of the triple points etched on the wall. In our
numerical result, the classical cellular structure can be clearly seen. Figure 5.10(b) shows
the result of a diagonal wave. Note that the cellular structure can not be fully spread out
due to the constraint of the duct size. The spinning detonation heads could be observed.
Refer to Fig. 5.9(d), (e), (f). With an increase of the duct size, a fully developed and
stable diagonal wave could be obtained. Refer to Fig. 5.10(c). Compared with the
dual-headed detonation shown in Fig. 5.9(d), (e), (f) and 5.10(b), this wave in Fig. 5.10(c)
is of four heads.
In the duct with a larger cross section area, multiple-headed detonation waves will
be sustained. The computational domain size is 40 40 40, and 9.6 millions of
hexahedral cells are used. Figure 5.11 shows the pressure contours and iso-surfaces. As
can be seen, multiple transverse waves collide with each other and interact with the wall.
As a result, a multi-cell structure is shown in the soot trace recorded on the side wall.
Qualitatively, the structure of the calculated soot trace agrees well with the smoked foil
record obtained in experiments (Strehlow, 1970). Note in all previous three-dimensional
calculations in this section, the detonation parameters are f = 1.6, = 1.2, E
+
= 50, q
0
=
50.
119
(a) (b) (c)
Figure 5.10: Numerical soot trace on the side wall: (a) rectangular mode, (b)
under-expanded diagonal mode, (c) fully developed diagonal mode with four heads. The
detonation waves travel upwards. Cross section size: 88 for cases (a) and (b), 2020 for
case (c). Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
= 50.
120


(a) (d)


(b) (e)


(c) (f)
Figure 5.11: Snapshots at different time: (a), (b), (c) pressure iso-surfaces; (d), (e), (f)
pressure contour on a cross section near the wave front. The detonation wave travels
upwards. Cross section size: 4040. Flow parameters: f = 1.6, = 1.2, q
0
= 50, E
+
= 50.
121
To proceed, we set the detonation parameters as = 1.2, q
0
= 50, E
+
= 50, and f =
1. With the overdriven factor as unity, we model a CJ detonation. Figure 5.12(a) shows
the calculated soot trace on one side wall. To investigate the influence of detonation
parameters on the cell size of the soot trace, we recalculate the same case with different
activation energies and heats of formation. Figures 5.12 and 5.13 show the results of this
investigation. In general, larger q
0
results in larger cell size, and larger E
+
leads to smaller
cell size. With E
+
decreased from 50 to 35, the cells become about three times larger.
To validate our three-dimensional calculation, we compare the calculated cell
sized with the experimental data reported by Guirao et al. (1982) and Kaneshige et al.
(1997). Figure 5.14 shows the calculated soot traces of the H
2
/air detonation waves with
four different equivalence ratios, i.e., = 0.7, 0.8, 1.0, and 1.4.

122


(a)

(b)

(c)
123
Figure 5.12: Numerical soot trace on one side wall. The detonation waves travel upwards.
Cross section size: 4040. Flow parameters: f = 1, = 1.2, E
+
= 35, (a) q
0
= 30; (b) q
0
=
40; (c) q
0
= 60.

(a) (b)

(c) (d)
Figure 5.13: Numerical soot trace on one side wall. The detonation waves travel upwards.
Cross section size: 4040. Flow parameters: f = 1, = 1.2, q
0
= 50, (a) E
+
= 20; (b) E
+
=
40; (c) E
+
= 45; (d) E
+
= 50.
124

(a) (b)
(c) (d)
Figure 5.14: Numerical soot trace on one side wall. The detonation waves travel upwards.
Cross section size: 4040. The reactants are H
2
-Air mixtures at: (a) = 0.7; (b) = 0.8;
(c) = 1.0; (d) = 1.4.
125
The parameters used to construct the one-dimensional ZND solution, which is
used as part of initial conditions, are determined by running the CEA program (Gordon
and McBride, 1996; McBride and Gordon, 1996) with the experimental condition as
input data, including the value of and the state conditions, e.g., pressure and
temperature of the unburned gases. One can also specify the mole fraction of H
2
, O
2
and
N
2
in the unburned gases as input. In the settings of the model equation employed, the
detonation wave can be specified by four parameters, i.e., , q
0,
f and E
+
, where f is
overdriven factor. Refer to Section 2.3. Since the detonation wave is unconfined, the CJ
detonation is anticipated. Therefore, f = 1. The values of specific heat C
p
and specific
heat ratio are obtained from the output of the CEA program directly. For the H
2
/air
reactions with the condition of p = 1 atm, T = 293 K, and = 1.0, the CEA output shows
= 1.1641 and C
p
= 3.3336 kJ/kgK for the burned gases. Because the reacting gases are
assumed polytropic, the above values for C
p
and are used for the calculation. Thus the
gas constant R of the reacting gases can be readily calculated. With the known wave
velocity from the output of the CEA program, one can obtain D
CJ
= 5.30. Using Eq. (2.44)
leads to M
CJ
= 4.91. Having and M
CJ
known, q
0
= 36.35 can be solved from the classical
CJ velocity equation. Refer to Eqs. (2.41) and (2.42). Parameters E
+
and K depend on the
reaction rate, and determine the width of reaction zone. In the present work, the value of
K is obtained from numerical integration. Refer to Eq. (2.39). The value of E
+
can only
be determined by experiments. Consider the hydrogen-oxygen system. Though the whole
reaction system is very complicated, it could be considered taking place in two stages: the
induction stage and the recombination stage. According to Fickett and Davis (1979), the
126
effective activation energy for the induction stage is that of the reaction
2
H O OH O + + , because it is the slowest and thus the rate-determining step. The
activation energy of this reaction is 69.1 kJ/gmol, which translates to a dimensionless
value E
+
= 28.4. Note for a given heat release q
0
and overdriven factor f, the activation
energy E
+
must be larger than some transition value to trigger the instability. The ZND
waves could be unstable in multi-dimensions at transverse direction without necessarily
unstable in the normal direction corresponding to a one-dimensional stability analysis
(Bourlioux and Majda, 1995).
To compare the dimensionless numerical results with the experimental data, both
the numerical and experimental data are normalized based on the minimal
numerical/experimental cell width obtained at the stoichiometric condition = 1.0.
Figure 5.15 shows the cell width equivalence ratio correlations, including the case of
= 2. As can be seen, the numerical result compared favorably with the experimental data
(Guirao et al., 1982; Kaneshige et al., 1997). When the stochiometric fuel/air ratio is used,
i.e., = 1.0, the smallest cell size is obtained.
To further investigate the results, we conduct a two-dimensional calculation for
the case of = 1.4 with identical flow conditions as that in the three-dimensional
calculation. The numerical soot trace is shown in Fig. 5.16. Compared with its
three-dimensional counterpart, i.e., Fig. 5.14(d), the averaged cell width in
two-dimensional case is about 65% of that in the three-dimensional results. Similar
calculations of two-dimensional detonations consistently show significantly smaller cell
widths than that of the corresponding three-dimensional simulations.
127

Figure 5.15: Cell width versus equivalence ratio for detonations in the H
2
-Air mixtures.
Cell width is normalized based on the minimal numerical/experimental cell width
obtained at the stoichiometric condition ( = 1.0).

Figure 5.16: Numerical soot trace of the two-dimensional calculation of a detonation
propagating in a H
2
-Air mixture at = 1.4. The detonation wave travels upwards.
128
Figure 5.17 shows the calculated time histories of the peak pressures at the shock
front of a CJ detonation in a H
2
/air mixture at = 1.4. For the same flow condition, three
sets of the results are plotted: (i) the ZND analytical solution, (ii) the two-dimensional
CESE solution, and (iii) the three-dimensional CESE solution. The peak pressures of the
two-dimensional detonation are about 2.5 times of that of the analytical solution because
of continuously moving and colliding triple points at the shock front. For the same
detonation wave, the maximum peak pressures of the three-dimensional simulation are
markedly higher than that in the two-dimensional calculation. Moreover, the patterns and
frequencies of the pressure fluctuations are distinctly different between the two- and the
three-dimensional calculations. Two-dimensional results show a regular cyclical pattern
with lower frequency, while the three-dimensional results are random with higher
frequencies, because of much more complex three-dimensional structures.
In general, we found that the flow structure of the three-dimensional detonation is
very similar to that of the two-dimensional ones. In particular, the rectangular mode in the
three-dimensional detonations resembles the transverse waves in the two-dimensional
detonations. Nevertheless, significance differences in key flow parameters were found
between the two and three-dimensional simulations, including the values of the maximum
pressure peaks and the frequencies of the pressure fluctuations. We conclude that one has
to resort to three-dimensional calculations to catch the quantitative features of
propagating detonation waves.
129

Figure 5.17: Time history of the pressure peaks at the shock front for two- and
three-dimensional calculations of a CJ detonation wave in a H
2
-Air mixture at = 1.4.
The horizontal dashed line represents the von Neumann pressure peak calculated by using
the ZND analytical solution. The peak pressure and time are both dimensionless. The
reference pressure is the pressure of the unburned gas. Refer to Eq. (2.4) for the
nondimensionalization of p and t.
130
5.5 Detonations in Circular and Annular Tubes
All currently used internal combustors are based on the steady-state deflagration
and the reacting flows are subsonic. Under certain conditions, e.g., the lean premixed
combustion for pollutant reduction in turbojet engine, unsteady deflagration has been
employed. However, as an equally viable combustion mode, detonation has not been
employed as the primary combustion mode for power-generation in propulsion devices
except for the recent development of Pulsed Detonation Engine (PDE). While numerous
studies on the PDE concept have demonstrated its promising applications to propulsion
devices, a typical PDE prototype relies on repetitive detonation initiation processes to
generate detonation waves at 30 to 100 Hz to sustain the propulsion power. A robust
detonation initiation mechanism by using a minimum amount of the sensitized gas, e.g.,
oxygen, and ignition energy has been the key technical issue under intense studies in all
PDE programs.
Researchers are now investigating the possibility of developing a continuous
detonation wave engine by using a spinning detonation wave around an annular chamber.
Since it is a continuous detonation wave, no repetitive detonation initiation processes, as
that in a standard PDE, is needed.
Since late 50s, a group of Russian scientists at the Lavrentev Institute of
Hydrodynamics, Siberian Division, Russian Academy of Sciences, Novosibirsk, have
been developing a novel detonation-based annular combustor (Mikhailov and Topchiyan,
1965; Bykovskii and Mitrofanov, 1980, 2000; Bykovskii and Vedernikov, 1996;
Nikolaev and Topchiyan, 1977). They considered combustion of acetylene-oxygen
131
mixtures in annular chambers with constriction of the exit cross section. Acetylene and
oxygen were injected into the combustor through a narrow radial slot and an upstream
axial slot, respectively. The reacting flows inside the combustor were subsonic with
respect to the axis of the annulus. Experimental observation indicated that conventional
turbulent subsonic combustion of the mixture prevailed in a large part of the annular
combustor. However, further investigation showed distinct wave structure occurred at the
leading edge of the reaction zone and the combustion mode was actually related to a
spinning detonation wave along the circumferences of the annulus. The reaction zone was
apparently continuously initiated by the rotating detonation wave with velocities close to
the speed of the sound of the combustion products, i.e., the CJ velocity of the detonation
for the given fuel/air mixture. In the past decades, they reported various flow
geometries and operational conditions of the abovementioned transverse detonation
waves (TDW) in annular cylindrical chambers. Combustion of both gaseous and
two-phase propellants was considered in their works.
A typical Russian TDW annular detonation combustor, tested by Voitsekhovskii
and coworkers, is shown in Fig. 5.18. The outer diameter of the annulus ranges from
40 to 100 mm. Oxygen was used as the oxidizer, which was injected into the
combustor from an annular slot of 0.2 mm wide at the top of the combustor. In the
downstream, a continuous subsonic O
2
flows at an averaged velocity of about 23 to
26 m/sec, which occupies the entire cross section of the combustor. Acetylene was used
as the fuel, which was injected into the combustor from uniformly distributed
orifices over the circumference of the inner wall of the annular channel at a
downstream location as shown in Fig. 5.18. The width of the annular channel ahead
132
of the fuel injection was about 5 mm. The TDW in the annulus chamber was
initiated by burning a wire inside the chamber by running an electric current
through it.

Figure 5.18: The TDW combustor geometry and the set-up.
The most remarkable feature of these combustors is that the axial flow was
subsonic while the detonation waves traveled tangentially in the azimuthal
direction in the annular chamber. It is imperative to understand why the detonation
wave would be confined to spin at a near CJ velocity only in the azimuthal
direction without progressing upstream in the axial direction.
Moreover, theoretical and numerical analysis on such spinning detonation
combustor developed by the Russian group does not exist. To date, the coexistence
of spinning detonation waves and relatively low-speed axial flows is an assertion
133
based on a set of experimental conditions, unverified by any other research groups.
Therefore, high-fidelity CFD simulation of the reacting flow in the annular
chamber is desired to understand the operation for such remarkable detonating
flows and to help possible future applications of such a detonation-based
combustor as a compact and powerful propulsion source.
We have simulated several spinning detonating flows as shown below. The
detonation parameters are set as = 1.2, q
0
= 50, E
+
= 50, and f = 1.2. In all calculations,
the analytical solution with a small spatial perturbation on the shock front is employed as
the initial condition. The detonation waves travel horizontally from left (along the x axis)
to right to consume the fresh fuel/oxidant mixture. The right boundary conditions are set
to be the flow conditions of the unburned gas. The non-reflective boundary condition is
imposed on the left boundary. The reflective wall boundary condition is used on the side
walls. For up to 5 million mesh cells the turn-around time is less than 20 wall hours on
our current 32-node computer cluster.
The first simulation is a single-head spinning detonation wave in a circular tube. The
radius of the computational domain is 2. The length is 10. The unit length is the half
reaction length. An unstructured mesh with 1.9 million hexahedron cells is used. Refer to Fig.
5.19 for a schematic of the mesh. In the tube, the confluence line of the transverse wave lies
along a tube radius, and rotates about the tube axis. It intersects with the tube wall and
traces out a helical track as the front progresses, giving rise to the name spinning
detonation. Figure 5.20 shows the calculated soot trace on the tube wall. This result is
consistent with the experimental pictures. It also reminds us the similar spinning pattern in
134
square ducts with small cross-section areas. The comparison between the experimental
picture and the numerical result is shown in Fig. 5.21 where the numerical result is the
iso-surfaces of the calculated species concentration of combustion reactant. Single spin, with
just one transverse wave, occurs near the detonation limit for the tube employed. The
single strong, nearly steady transverse wave without triple-point collisions makes
this perhaps the cleanest three-dimensional detonation system.

Figure 5.19: A schematic of the mesh used for the three-dimensional calculations on
spinning detonation in a circular tube. The tube radius is 2. The unit length is the half
reaction length.
135

Figure 5.20: Calculated soot trace on the tube wall of a single head spinning detonation wave.
The tube radius is 2. The unit length is the half reaction length. Flow parameters: f = 1.2,
= 1.2, q
0
= 50, E
+
= 50.


Figure 5.21: A schematic of a single-head detonation wave and comparison between the
experimental picture and the numerical result. Tube radius: 2. Flow parameters: f = 1.2, =
1.2, q
0
= 50, E
+
= 50.
136
Figure 5.22 is a schematic for the structure of a single-head spinning detonation.
The plot is in the rotating frame attached to the triple point of the detonation wave. Thus the
flow is steady. The segment BC of the transverse wave is an overdriven detonation
propagating into the unburned gas mixture, which has passed through the incident shock
denoted by I. The Mach stem at A is tilted so that the normal velocity of the material into it
corresponds to an overdriven detonation. Consequently its reaction zone is very short.
Proceeding around the tube (upwards in Fig. 5.22) on the leading shock, the reaction zone
lags farther and farther behind until it joins the transverse wave at C. Behind the front, we
observed two interleaved helical strips of reactants and combustion products. Strip 2 passes
through the reaction zone and strip 5 passes through the transverse wave. There is a
considerable gradient of all flow variables across Strip 2. The hottest material of all is that
on the left edge of Strip 2, which has passed through the mach stem just above point A. The
rearward extension of BC is a shock wave, which rapidly decays to an acoustic wave.
137

Figure 5.22: A schematic of the single-head spinning detonation in a rotating frame.
Figures 5.23 show a series of snapshots of pressure iso-surfaces of the
single-headed spinning detonation wave calculated by the three-dimensional CESE code.
The detonation front rotates in the counterclockwise direction. Complex flow pattern
behind the reaction zone can be clearly seen. The most interesting feature is the spiral
pattern of the transverse pressure wave lagging behind the shock front.
Figures 5.24 show the pressure contours on the sliced x-z and x-y planes at
different time, respectively. On the x-z plots, the incident shock, Mach stem, and
transverse wave are captured. Figure 5.25 shows the snapshots of pressure contour on the
wall with y-axis being the length on the circumferential direction. With the advance of
time, the track of the triple point can be clearly seen.
138


t = 34.74 t = 35.13

t = 35.52 t = 35.91

t = 36.30 t = 36.69


t = 37.08 t = 37.50
Figure 5.23: Pressure iso-surfaces of a spinning detonation at different time. Tube radius:
2. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.
139
t = 32.4
t = 36.30
Figure 5.24: Snapshots of pressure contour on the sliced x-z and x-y planes at different
time. Tube radius: 2. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.

140


t = 34.74 t = 35.13 t = 35.52

t = 35.91 t = 36.30 t = 36.69


t = 37.08 t = 37.50
Figure 5.25: Snapshots of pressure contour on the tube wall of a spinning detonation at
different time. Tube radius: 2. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.
141
Figures 5.26 show snapshots of the detonations with more than one transverse
waves. The radius and length of the computational domain are 6 and 10, respectively. The
mesh size is 1.9 million. Several traveling modes of the unstable detonations exist in this
case. Some of the triple points rotate around the circumference of the round tube. Other
triple points travel diagonally across the tube. The transverse waves, incident wave, and
Mach stem can be seen in Fig. 5.26(c). The interactions between these triple points give
rise to the soot trace on the wall with the diamond-shaped pattern of moderate regularity
more like those in rectangular tubes. Refer to Fig. 5.27.
Finally, the detonation in an annular chamber is investigated. The external and
internal radius is set as 6 and 3, respectively. The length is 10. A schematic of the domain
and the mesh is shown in Fig. 5.28. The mesh consists of 2.88 million hexahedron cells. The
numerical results are shown in Figs. 5.29 -31. As can be seen in the pressure plots, there are
many transverse waves traveling simultaneously in the annular space and the flow is very
complex. These waves could be divided into two orthogonal systems as an analogy to those
detonation waves in rectangular ducts. One family of waves moves along the circumference
and the other along a radius. The transverse waves collide with each other and with the wall,
leaving the similar cellular structure as that shown in previous cases. Refer to Fig. 5.31. There
is not available experimental result to compare. However, the numerical results are consistent
with the analysis by Ficket and Wood (1979).

142

(a)

(b) (c)
Figure 5.26: Calculated multi-headed detonation wave in a round tube: (a) pressure
iso-surfaces; (b) pressure contour on the sliced y-z plane; (c) pressure contour on the sliced
x-y plane. Tube radius: 6. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.
143

Figure 5.27: Calculated soot trace on the tube wall of a multi-headed detonation wave.
Tube radius: 6. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.

Figure 5.28: A schematic of the mesh for the 3D calculations on detonations in an annular
tube. The external and internal radius is 6 and 3, respectively. The unit length is the half
reaction length.
144


t = 34.7 t = 35.1


t = 35.5 t = 35.9


t = 36.3 t = 36.7


t = 37.1 t = 37.5
Figure 5.29: Pressure iso-surfaces of an annular detonation at different time. External and
internal radius: 6 and 3. Flow parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.
145


Figure 5.30: Calculated pressure contour on the y-z and x-z planes of the multi-headed
detonation wave in an annular chamber. External and internal radius: 6 and 3. Flow
parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.


Figure 5.31: Numerical soot trace on the outer and inner tube walls of the multi-headed
detonation wave in an annular chamber. External and internal radius: 6 and 3. Flow
parameters: f = 1.2, = 1.2, q
0
= 50, E
+
= 50.

146
CHAPTER 6
IMPLOSION AND EXPLOSION
6.1 Introduction
It is highly desirable that the envisioned PDE could use a regular aviation fuel to
support detonation. However, it is difficult to detonate the fuel-air mixtures of regular
aviation fuels, e.g., the JA fuel. One method of initiating these mixtures is via the use of a
pre-detonator or initiator tube, in which one uses a small amount of highly detonable
sensitized gases. Spark ignition of this mixture leads to a detonation wave, which in turn
could propagate into the main detonation chamber and initiate detonation waves in the
fuel-air mixture in the main combustion chamber.
Although the detonation could be easily initiated in the initiator, the subsequent
transmission of the detonation from the initiator to the larger main chamber requires
suitable flow arrangement to ensure robust detonation initiation at the main chamber. In
this setting, implosion could be useful to facilitate the transmission process. The idea is to

147
set up an implosion in the main detonation chamber by using the blast waves provided by
the initiator.
Implosion is a very powerful mechanism which has been used to initiation violent
explosion. In an implosion process, the collapsing shock front compresses the gas
mixture adiabatically as it flows into a decreasing area, leading to extremely high energy
density at the focal region. In a successful arrangement, the compression would raise the
post-detonation pressure to be higher than the Chapman-Jouguet (CJ) pressure, giving
rise to an overdriven detonation wave. Such a shock focusing mechanism could reduce
the required sensitized initiator gas and the additional energy deposition for successful
detonation initiation and transmission processes.
Murray et al. (2001) proposed the transmitting detonations from an initiator into a
larger tube through various arrangements of annular orifices. Detonation waves, passing
the annular orifice, generated imploding toroidal waves in the back of the orifice. Aided
by high pressures and temperatures of imploding shocks, robust detonations could be
re-initiated in the main chamber. Jackson and Shepherd (2002) and Jackson et al. (2003)
have successfully developed an efficient imploding detonation initiator based on the use
of an array of small channels which were distributed evenly on the surface of a circular
tube. By merging multiple jets and wave fronts transmitted through the channels
simultaneously, they generated an imploding shock front in a circular chamber and
detonated the C
3
H
8
/air mixtures at room temperature (298 K).

148
Analytical solutions of imploding and exploding blast waves have been available
in the self-similar regimes, approached by the evolving flows after the initial flow
structure fades away. Based on dimensional analyses, the classical solution of implosion
was provided by Guderley (1942). Self-similar solutions of explosion were developed by
Taylor (1950) and Sedov (1959). Later, these classical solution have been refined by
Sharma and Radha (1995), Toque (2001), and Van Dyke and Guttmann (1982). Note that
in these solutions, the central region, from where the blast wave emanates, is without any
flow structure. This is true when the blast wave has propagated far away from the center
and all wave structures in the initiation process have been dissipated. Other theoretical
and experimental studies of imploding/exploding blast waves include the work by Boyer
(1960), Friedman (1961), Baker (1973), Schwendeman and Whitham (1987), and Glass
and Sislian (1994). There have been few satisfactory implosion experiments with useful
data due to the extreme difficulty in breaking a cylindrical diaphragm instantaneously and
simultaneously in different directions.
Many numerical simulations of implosions and explosions have been conducted.
Brode (1955, 1959) reported simulations of spherical blast waves. His numerical results
were compared with an analytical solution based on an isothermal profile. In his
numerical studies of detonation with a spherical charge of TNT,

he showed a second
shock in the focal region. Flores and Holt

(1981) employed Glimms method (1965) to
simulate explosions that were generated by a pressurized sphere in water. Charrier and

149
Tessieras (1986) studied a cylindrical explosion in air by using a front-tracking technique.
Oran and DeVore (1994) used the Flux Correction Transport (FCT) method (Boris and
Book, 1973; Book et al., 1975) and a connection machine to solve imploding detonation
waves. Payne (1956) used Laxs method (1954) to simulate converging cylindrical
shocks. Falcovitz and Birman (1994) used a Godunov-type scheme (Ben-Artzi and
Falvovitz, 1984) and studied various discontinuities in an explosion and implosion flow.
Sod (1977) proposed an operator splitting method to overcome the singularity problem at
the focal center. He (1996) used a second-order upwind method to simulate direct
initiation of detonations. Menikoff et al. (1996) showed that it was important to resolve
the curved detonation front and that a cell size on the order of 0.1 mm or less was
required to resolve the reaction zone of common detonation waves.
The objective of the present work in this chapter is to analyze the complex
structures of converging shocks and the subsequent detonation initiation. The space-time
CESE method was used to solve the two- and three-dimensional Euler equations for
chemically reactive flows in cylindrical and Cartesian coordinates, respectively. In
two-dimensional calculations, both reactive flow and non-reactive flows are investigated.
First, we consider circular imploding flows with small perturbations on the shock front.
The perturbations are employed to mimic the effect of merging supersonic jets, similar to
that by Jackson and Shepherd (2002) and Jackson et al. (2003). We then consider
implosion of several polygonal shock fronts. In three-dimensional calculations, the

150
non-reactive implosion/explosion process generated from a similar polygonal shock front
is studied. We also compared the results of different settings of initial condition.
6.2 Two-Dimensional Implosion and Explosion
For two-dimensional calculations of implosions and explosions, consider a
circular computational domain. Figure 6.1 is a schematic of the computational domain
and the associated mesh. The dimensionless radius of the computational domain is 1.0.
Because of flow symmetry, the actual computational domain is a quarter of the shown
circular domain. All numerical results to be presented were calculated by using the same
mesh, which is composed of about 4.3 million quadrilaterals. In all calculations, the
initial condition includes two separate regions with P
out
/P
in
= 20. Temperature is uniform
for the whole domain and
out
/
in
= 20.
Four implosion cases are presented: (i) Non-reacting flow of a square front. The
length of each edge is 1.0. (ii) Non-reacting flows of an octagonal front. The circum-
radius of the octagon is 0.7. (iii) Non-reacting flows of a circular front with sinusoidal
perturbations on the shape of the front. The averaged radius of the front is 2/3. The
amplitude of the perturbation is 2% of the radius of the front. Thirty-two cycles of
sinusoidal perturbations are imposed around the circular front. (iv) Reacting flow of a
circular front with the same geometry and similar perturbations as that in case (iii).


151

Figure 6.1: A schematic of the mesh for 2D implosion/explosion calculations.
The calculations were performed by using a parallelized CESE Euler solver,
running on our 32-node PC cluster. The simulation ends when the main shock reflects
from the center and exits the outer boundary of the computational domain. Typical run
time of one calculation is about 4 hours on a wall clock.
Figures 6.2-6.5 show snapshots of pressure contour at various times for the
abovementioned four cases. Note at different time, the range of pressure value could be
of large difference. To spread the contour colors on the full color map for better visual
effect, different pressure levels are chosen for snapshots at different time. Thus, there is
not a global legend. In each case, when the diaphragm separating the driving and driven
sections is lift, the primary shock wave, followed by a contact discontinuity, implodes to

152
the center. In the opposite direction, a rarefaction wave moves away from the center. In
the imploding phase, pressure and velocity behind the imploding shock steadily increase
due to the decreasing area. Moreover, flow evolution is symmetric with respect to the
bisector of each vertex of the polygon. In other words, the bisectors behave like inviscid
walls. Evolution of each side of a polygonal shock front may be interpreted as a planar
shock entering into a channel of convergent walls, or a wedge-shaped cavity. Imploding
shocks interact with the bisector and form Mach reflection. Refer to Figs. 6.7. In general,
the evolution of each planar section of the polygonal shock front resembles the scenario
of a shock-on-wedge problem, in which an incident shock interacts with a wedge with the
bisector corresponding to a wedge surface. In this process, the Mach stem of the Mach
reflection propagates with a wave speed faster than that of the initial shock. As a result,
the Mach stem increases in size, while the length of the shock front decreases to zero.
During this evolution, the number of the vertices of the imploding polygon doubles that
of the initial polygon. Finally, a new polygon is formed with a similar shape but the
orientation is rotated. Seemingly, the original polygon, with a shrunk area, was recovered
with a rotated orientation. The sides of this smaller polygon are composed of Mach stems.
However, it is known that Mach stems are curved. Thus, the regenerated polygonal
shocks have curved sides. One may have multiple regenerations of the polygonal front
during the imploding process. Refer to Figs. 6.6 for the evolution of the polygonal shock.

153



Figure 6.2: Snapshots of pressure contour of a square front implosion at time = 0, 0.1, 0.2,
0.25, 0.3, 0.4, 0.5, 0.65, and 0.8. Domain radius: 1. Circumradius of the square front:
0.707, i.e., edge length 1.0. P
out
/P
in
= 20.
out
/
in
= 20. Contour levels are different for
snapshots at different time. The color map without levels is shown on the bottom.

154




Figure 6.3: Snapshots of pressure contour of an octagonal front implosion at time = 0,
0.09, 0.18, 0.27, 0.315, 0.36, 0.495, 0.63, and 0.72. Domain radius: 1. Circumradius of
the octagonal front: 0.7. P
out
/P
in
= 20.
out
/
in
= 20. Contour levels are different for
snapshots at different time. The color map without levels is shown on the bottom.

155






Figure 6.4: Snapshots of pressure contour of a circular front implosion at time = 0, 0.09,
0.18, 0.27, 0.36, 0.45, 0.585, 0.72, 0.90, respectively. Domain radius: 1. Radius of the
circular front: 0.667. P
out
/P
in
= 20.
out
/
in
= 20. Contour levels are different for snapshots
at different time. The color map without levels is shown on the bottom.

156






Figure 6.5: Snapshots of pressure contour of a circular front implosion with chemical
reaction at time = 0, 0.09, 0.18, 0.27, 0.315, 0.36, 0.405, 0.54, 0.90, respectively. Domain
radius: 1. Radius of the circular front: 0.667. P
out
/P
in
= 20.
out
/
in
= 20. Detonation
parameters: f = 1, = 1.4, q
0
= 50, E
+
= 20. Contour levels are different for snapshots at
different time. The color map without levels is shown on the bottom.

157


Figure 6.6: The evolution of the polygonal shock. In each polygonal section, the flow
field is a planar shock wave entering a converging channel. As time evolves, Mach stem
will take over the initial planar shock. Thus, the polygon is regenerated and rotated

158
Figure 6.7 shows the time histories of temperature and pressure at the focal region
for three non-reacting cases. The dimensionless temperature in Fig. 6.7 is defined as T/T
in
,
and the dimensionless pressure is defined as P/P
in
. The subscript in denotes the driven
section in the initiation of the imploding flows. The x axis is the dimensionless time.
Refer to Eq. (2.4). For a realistic condition with T
in
= 300 K, P
in
= 0.2 atm, and domain
radius R = 0.38 m, the unit of time is about 1.3 ms.
The case of square front has lower maximum pressure and temperature because of
stronger Mach reflections and large incident angles as compared with the other cases. The
maximum pressures and temperatures of the two cases of the octagonal front and the
circular front are comparable. However, the circular front generates a longer
high-pressure plateau.
Moreover, the results of both circular front and the octagonal front show a
secondary implosion. This happens because of the interaction between the reflected and
exploding main shock and the imploding contact discontinuity. After reflection from the
center, the main shock wave moves outwards and clashes with the imploding contact
discontinuity. This collision creates a new shock wave, which implodes to the focal
center. Meanwhile, this interaction would slow down the discontinuity wave. The main
shock wave is less affected and it continues the exploding motion after the interaction.
The implosion of the second shock causes the second peak in both the pressure and the
temperature histories at the focal point.

159
Similar to the reflected main shock, the second shock reflects from the focal
center and collides with the contact discontinuity wave again and generates the third
imploding shock wave. Although it is much weaker than the first one, the second peak of
in the pressure and temperature histories at the focal point could be observed in Fig. 6.7
for the cases with circular front and octagonal front. The secondary implosion in the
octagonal front case occurs later than that in the circular front case. The circular front
case also shows an even much weaker third implosion. On the other hand, the case of the
square front does not show a clear third implosion. After the third shock implodes and
reflects from the focal point, it could collide with the contact discontinuity wave again.
At this point, the contact discontinuity wave is weakened and its motion slowed down.
Details of this interaction and the follow-up are too weak to be observed in Fig. 6.7.
Figure 6.8 shows the time histories of pressure and temperature at the focal point
of a successful detonation initiation processes by imploding circular shock front with
perturbations on the initial front surface. The chemical reaction is modeled by a one-step
irreversible reaction, i.e., the two-dimensional version of Eqs. (2.6)-(2.8). In the settings
of this simplified model equation, the resulting detonation wave can be specified by the
following four parameters: the specific heat ratio = 1.4, the heat release q
0
= 50, the
activation energy E
+
= 20, and the overdriven factor f = 1 (CJ detonation anticipated for
unconfined waves). Note these parameters could be related with real chemical reactions
with the aids of the NASA CEA program (Gordon and McBride, 1996; McBride and

160
Gordon, 1996) and some empirical global models (Westbrook and Dryer, 1981). As can
be seen, similar to the previous results without chemical reaction, consecutive shock
implosions with multiple peaks are evident. However, the secondary and third shocks are
stronger than the first one. Moreover, the resulting central pressure and temperature
plateau is much higher for a much longer time period. This is due to the initiated
detonation in the focal region. The overall pattern is much more complex as compared to
the one-dimensional cases (Wang et al., 2005).

161

(a)

(b)
Figure 6.7: Dimensionless time histories of (a) temperature, and (b) pressure, at the focal
point for the three non-reacting cases. P
out
/P
in
= 20.
out
/
in
= 20. Reference temperature
and pressure are the initial values in the driven section. Refer to Eq. (2.4) for the
definition of dimensionless variables.

162

(a)

(b)
Figure 6.8: Dimensional time histories of (a) temperature, and (b) pressure, at the focal
point for the imploding reacting flow. P
out
/P
in
= 20.
out
/
in
= 20. Detonation parameters: f
= 1, = 1.4, q
0
= 50, E
+
= 20. Reference temperature and pressure are the values in the
driven section in the initiation stage. Refer to Eq. (2.4) for the definition of dimensionless
variables.

163
6.3 Three-Dimensional Implosion and Explosion
In this section, we report three-dimensional calculation of the implosion/explosion
processes in the main detonation chamber. Refer to the experiments by Jackson et al.
(2003). This is an extension to the above two-dimensional calculations.
Compared to the two-dimensional calculations, the non-reflective boundary
condition on the outer circular boundary is replaced by a wall boundary condition as the
end wall of the chamber. In the planar view, the wall boundary condition is imposed at
the left end of the chamber. The right end of the chamber is open, where the
non-reflective boundary condition is imposed. Similar to the two-dimensional simulations,
the initial condition is composed of two distinct regions, i.e., the driving part and the
driven part with different pressures and densities. In all calculations following, P
out
/P
in
=
20 and temperature is uniform at 300 K. Thus, in all cases,
out
/
in
= 20.
In this section, results of non-reacting imploding flows with different initial
conditions will be presented. In the first cases, all lengths are made dimensionless by
setting the radius of the chamber as unity. The length of the circular detonation chamber
is 8. The x axis is the central line of the detonation tube. The initiating section is located
between x = 1 and x = 3. The diaphragms separating the driving gases and the driven
gases consist of three parts: an axial diaphragm located in the initiating section with the
projection on the y-z plane being an octagon with its circumradius being equal to 0.75,
and two lateral diaphragms at x = 1 and x = 3 covering the region between the octagon

164
and the tube wall. An unstructured mesh with 4.1 million cells is used. Refer to Fig. 6.9
for the schematic of the mesh and the geometry. Refer to the two snapshots on the top of
Figs. 6.10 for the settings of the initial conditions.

Figure 6.9: A schematic of the mesh for 3D implosion/explosion problems.

165
Figures 6.10 show a sequence of snapshots of pressure contour after breaking the
diaphragm. Two-dimensional plots on the y-z plane at x = 2 and the x-z plane at y = 0 are
presented. The initial pressure front has an octagonal front. The pressure contours on the
plane x = 2 resemble the corresponding two-dimensional results. However, being not
confined on the axial direction, the imploding shock not only converges to the center, but
also propagates into the fore and aft regions. The shock also reflects from the dead end of
the chamber. In this case, there is no chemical reaction.
In the other cases, the size of the computational domain is double of that of the
first case. The radius of the chamber is 2 and the length is 16. The pressurized section at
the initial condition is now located between x = 2 and x = 4. Other settings of the initial
condition are similar to that of the first case. However, different size of the driving part is
studied with the circum-radius of the octagon being 30% and 70% of the chamber radius.
These two flows (referred to as Case 2 and Case 3, respectively) are carried out on the
same mesh of 4.1 million cells. Moreover, to investigate the effect of mesh density, we
also performed a calculation of case 2 on a larger mesh with 10.0 million cells, denoted
by Case 4.
Figures 6.11 show the result of Case 2. The flow pattern is quite similar to that of
Case 1 as shown in Fig. 6.10. The quantitative differences between the latter three cases
are shown in Figs. 6.12. The temperature and pressure are dimensionless values based on

166
the initial values in the driven section. As can be seen, compared with Case 2, the longer
converging distance of the shock in case 3 leads to a longer time to reach the focus.
However, higher central pressure is obtained in Case 3. The higher maximum central
pressure in Case 4 reflects the effect of the refined mesh. Compared with Case 2, the
second peak is also successfully caught in the history of the central pressure of Case 4. In
addition, the second peak can be seen in the central pressure and temperature histories of
Case 3. However, the strength of the second peak is weaker than the first one, and the
maximum central pressure and temperature obtained in these three cases are far below
that in the two-dimensional case. This could be due to (i) relatively low mesh density on
the y-z plane and (ii) lack of flow confinement on both ends of the detonation chamber in
the three-dimensional flows.


167










Continued
Figure 6.10: Snapshots of pressure contour on the sliced planes x = 2 and y = 0 of an
octagonal front implosion at time = 0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.5, 0.75, 1.25.
Tube radius: 1. Circumradius of the octagonal front: 0.75. P
out
/P
in
= 20.
out
/
in
= 20.

168
Figure 6.10 continued














169










Continued
Figure 6.11: Snapshots of pressure contour on the sliced planes x = 4 and y = 0 of an
octagonal front implosion at time = 0, 0.24, 0.4, 0.56, 0.8, 1.8, 2.8, 4.0. Tube radius: 2.
Circumradius of the octagonal front: 0.3. P
out
/P
in
= 20.
out
/
in
= 20.

170
Figure 6.11 continued




171

(a)

(b)
Figure 6.12: Dimensionless time histories of (a) temperature, and (b) pressure, at the
focal point for the three cases with different initial condition and mesh size. Tube radius:
2. P
out
/P
in
= 20.
out
/
in
= 20. Reference temperature and pressure are the initial values in
the driven section. Refer to Eq. (2.4) for the definition of dimensionless variables.

172
CHAPTER 7
PULSE DETONATION ENGINE AEROACOUSTICS
7.1 Introduction
Detonation could be a very efficient combustion mode to convert stored chemical
energy in fuels to useful thermal and mechanical energy for practical applications. A
pulsed detonation engine (PDE) is a novel propulsion concept based on the use of the high
pressures generated by repetitive detonation waves in a detonation chamber (Kailasanath,
2001; Kailasanath et al., 1999; Li et al., 2000; Ebrahimi et al., 2000; Schauer et al., 2001).
The calculation in this chapter aims to support ground tests of a PDE prototype at NASA
Glenn.
In this particular PDE concept, the following process repeats itself at a frequency
about 60 Hz. First, the PDE tube is filled with a propane/air mixture. After the detonation is
initiated at the closed end of the thrust tube, the detonation wave propagates toward the
tube exit. The resulting high-pressure region behind the detonation wave provides the
thrust force for the aircraft. The PDE tube exit is kept open throughout the fueling and
firing process. After the filling process is completed, the pressure of the unburned gases is

173
generally in equilibrium with the surrounding atmosphere. Because the flow is unconfined,
the burned gases and the shock wave travel at the sonic Chapman-Jouguet (CJ) velocity.
After the detonation blow-down process, an inert gas is used to purge the detonation
chamber before refilling the tube with the fuel/air mixture. A graphical depiction of the
PDE cycle is shown in Fig. 7.1.

Figure 7.1: A typical PDE cycle.

174
In the fully burned region following the detonation front, the pressure rise is about
15 to 20 times of that in the unburned gases. The pressure peak of the von Neumann spike
at the shock front could be 50 to 100 times of that of the unburned gases. However, the high
pressure of the von Neumann spike does not produce a useful thrust force. Because the
traveling detonation wave and burned gases behind the shock move away from the closed
end of the thrust tube, the high pressure reached by the equilibrium CJ condition cannot be
fully utilized for propulsion. Following the detonation wave, the burned gases in the
closed-end region expand and pressure reduces. Typically, pressure in the closed-end
region during the firing process is only about 5 to 8 times the pressure level observed in the
unburned gases. This is the source of the propulsion force for the PDE.
Following the detonation wave, the burned gases leave the PDE tube and expand
into the surrounding atmosphere. Due to the momentum of accelerating gases,
over-expansion occurs at the tube exit. As a result, expansion waves, originated from the
PDE tube exit, would move to the forward and aft quadrants of the PDE tube, as well as
upstream into the PDE tube. Thus the pressure at the closed-end region would temporarily
reduce to a value lower than the surrounding atmospheric pressure. This low-pressure
condition inside the PDE tube is detrimental to propulsion power, but it could be useful in
the refueling process for the next firing. To ensure successful detonation initiation,
experimentalists generally recommend purging the tube with an inert gas before each
refueling. By removing undesired combustion products as well as unsteady

175
expansion/compression waves, the purging procedure refreshes the flow condition inside
the chamber, and thus avoids fuel leakage to the surrounding atmosphere during the
refueling process.
During the detonation process, the shock wave travels at a supersonic speed into the
unburned gases, and the flow conditions outside the tube exit would have no impact to the
traveling detonation wave. However, the blow-down and refueling processes strongly
depend on the boundary conditions outside the PDE exit, which in turn cannot be fully
described without knowing the details of the PDE plume dynamics. Moreover, gas
blow-down and refueling processes occupy more than 90% of the time in each firing cycle.
Well-managed blow-down and refueling processes are the keys to providing continuous
strong detonations. Consequently, it is imperative to simulate the plume dynamics as an
integral part of the overall calculation in order to assess the performance of a PDE. Because
the existing jet is pulsating, significant pressure fluctuations in both forward and aft
quadrants of the PDE are anticipated. The main concern would be vibrations of the
fuselage and wings of the aircraft due to periodic pressure loading. Moreover, propagation
of the acoustic waves to the far field is also of concern.
A typical approach to Computational Aero Acoustics (CAA) for propulsion
systems is a two-step procedure: (i) CFD solution of the Reynolds Averaged Navier-
Stokes (RANS) equations for the background mean flow at a steady state, and (ii) the

176
solution of a linear wave equation for acoustic propagation based on the known mean flows.
However, for the PDE plume dynamics, the background flow is highly unsteady with large
nonlinear flow structures. The above approach cannot be applied and simulation of the
PDE plume dynamics is indeed a challenging CAA problem.
By using an upwind method, Cambier and Adelman (1988) and Eidelman et al.
(1990) conducted numerical studies on the propulsion performance of various PDE
concepts. Kailasanath et al. (2002) employed the Flux Correction Transport (FCT) method
in conjunction with a simplified two-step chemistry model and reported a series of
performance simulation. Povinelli and Yungster (2002) carried out cycle analyses of a
PDE by using the Total Variation Diminishing (TVD) method.
In this chapter, the CESE code is used to perform high-fidelity simulations of
pulsating PDE plumes. The CESE method is capable of capturing shocks and acoustic
waves simultaneously, while the magnitude of the pressure jump across the shock wave
could be several orders in magnitude greater than the pressure fluctuation in the acoustic
waves. In dealing with the CAA calculations, we took the advantage of the unique
advantage of the CESE method in treating non-reflective boundary condition. Since all
flow information must propagate into the future, the non-reflective boundary-condition
treatment is very simple in the sense that one simply imposes space-time flux conservation
by integration over conservation elements near the boundary of the computational domain.

177
The high-resolution capabilities and the simple treatment for non-reflective boundary
conditions of the CESE method are critical to the success of the present calculations. The
rest of the chapter is organized as follows. Section 7.2 discusses the initiation condition
setup for detonation wave. Obviously, we cannot model the detonation initiation process in
the present calculations. Instead, a one-dimensional detonation flow profile calculated
based on the ZND model is employed. Section 7.3 shows the numerical solutions of the
near-field plume dynamics. Section 7.4 shows the far-field solution. Section 7.5 provides
concluding remarks for the chapter.

178
7.2 Initial Condition of Detonation Wave
Premixed gas mixture of fuel and air burns inside the PDE chamber, and
detonation waves travel through the thrust tube and exit to the ambient atmosphere. The
flow field outside of the PDE tube is to be investigated. The numerical results at specific
locations are recorded and compared with experimental data.
At the beginning of each firing, 85% of the tube is filled with a propane/air
mixture at an equivalence ratio = 1.3. The initial chamber pressure is 15.0 psia (103.4
kPa), and temperature is 510 R (283.3 K). Since the detonation wave is unconfined, we
first run the NASA CEA program (Gordon and McBride, 1996; McBride and Gordon,
1996) to obtain the basic flow parameters of the detonation wave. The CJ detonation
velocity V
CJ
= 6200 ft/sec (1890 m/s) and shock pressure jump P
2
/P
1
= 16.667, where the
subscript 1 refers to the unburned condition and the subscript 2 refers to the fully burned
condition. The PDE tube operates at 60 Hz, i.e., 16.67 ms for one cycle. When t = 0, the
initial detonation wave is located at 1.04 cm from the closed end. Refer to Fig. 7.2. The
one-dimensional ZND analytical solution is mapped to the axisymmetric mesh of the
present calculations. Eight mesh nodes along central line are used to describe the ZND
detonation profile. Refer to the details of the ZND solution in Chapter 2.

179

Figure 7.2: A schematic of setting up the initial condition.
The parameters used to construct the one-dimensional ZND solution are
determined by running the CEA program (Gordon and McBride, 1996; McBride and
Gordon, 1996) with the experimental condition as input data. The output data from CEA
program are also compared with experimental data for checking consistence.
In the settings of the simplified model equation employed, the detonation wave
can be specified by four parameters, i.e., the specific heat ratio, q
0
the heat release, f the
overdriven factor, and E
+
the activation energy. Since the detonation wave is unconfined,
the CJ detonation is anticipated and f = 1. For the specific heat ratio, the CEA program
shows = 1.2015 and C
p
= 2.1064 kJ/kgK for the burned gases. Because the reacting
gases are assumed polytropic, the above values of C
p
and are used in the calculations.
Thus the gas constant R of the reacting gases can be readily calculated. With the known

180
wave velocity from the output of the CEA program, one can obtain M
CJ
= 5.450. Having
and M
CJ
known, q
0
= 37.563 can be solved from the classical CJ velocity equation.
Refer to Eqs. (2.41) and (2.42). Parameters E
+
and K depend on the reaction rate, and
determine the width of reaction zone. In the present work, the value of K is obtained from
numerical integration. Refer to Eq. (2.39). The value of E
+
can only be determined by
experiments. Though the whole reaction system of C-H-O is very complicated, empirical
global models have been developed to capture the overall behavior in a sequence of
global or quasi-global steps. Westbrook and Dryer (1981) presented and evaluated
one-step, two-step, and multiple-step global kinetics for a wide variety of hydrocarbons.
According to Westbrook and Dryer (1981), the activation energy of
3 8 2 2 2
C H 5O 3CO 4H O + + , a single-step global reaction for oxidation of C
3
H
8
, is 30
kcal/gmole (125.5 kJ/gmole). Thus, one can calculate the dimensionless activation energy
E
+
and obtain 53.3, provided the temperature of the unburned gases is 283.3 K. Refer to
Eq. (2.11).
To support the prototype design and experiments of the PDE, our modeling works
were divided into two phases. The first phase is to investigate the near-field dynamics
behind the PDE tube exit, which is featured by shock waves interacting with complex
vortex rings. The second phase is to model the far field, in which the strength of the
acoustic waves is of great concern.


181
7.3 Near-Field Calculations
Figure 7.3 shows the computational domain for near field calculations. Since the
flow field is symmetric with respect to the thrust vector, the region above the symmetric
line is used as the computational domain. The length and radius of the computational
domain are 12.5 ft (3.81 m) and 4.5 ft (1.37 m), respectively. The spatial domain is divided
into 300K non-uniform quadrilateral cells, with clustering near the tube exit. Inside the
PDE tube, the mesh size is about 23,000.

Figure 7.3: A schematic of computational domain for the near-field calculations.

182
To examine the details of the flow features, we recorded time histories of pressure
and density at thirty-six sampling points. Experimentalists determined the locations of the
sampling points. The collected data are compared with NASAs experimental data. As
shown in Fig. 7.4, sampling points are distributed on a plane parallel to the thrust vector.
The position of the plane is offset by a distance of 5 and 7/8 inches (14.92 cm) from the
thrust vector to avoid direct impact to the pressure transducers by the detonation waves.
The thirty-six sampling points are divided into four groups, located along four lines on
the plane, namely, L-1, L-2, L-3, and L-4. All lines lie on the checking-point plane. The
L-1 line is parallel with the thrust vector, with its origin corresponding to the PDE tube
exit on the sampling point plane. Ten sampling points are distributed along L-1 with
distances from origin of 4, 5.66, 8, 11.32, 16, 22.64, 32, 45.28, 64, and 90.56 inches (or
10.16, 14.38, 20.32, 28.75, 40.64, 57.51 81.28, 115.01, 162.56, 230.02 cm). Sampling
points on L-2, L-3 and L-4 have the same corresponding distances from their origins. The
L-2, L-3, and L-4 lines are rotated counterclockwise from L-1 with the angles of 30, 60,
and 90, respectively. Note that there are only 8 sampling points on L-3 and L-4.
We applied the CESE axisymmetric Euler solver to perform the calculations. We
then map all the checking-point locations to the computational domain by using the
Pythagoras theorem. Refer to the triangle in Fig. 7.4. Once the sampling point locations are
identified on the numerical domain, we simply take the numerical solutions at the grid
points closest to the sampling points.

183

Figure 7.4: A schematic of the sampling point distribution in the computational domain
for near field calculations. Domain size: 12.5 ft in length, 4.5 ft in radius.
The detonation wave is ignited at the closed end of the thrust tube, and travels
from left to right. Figure 7.5 shows the time history of pressure at the closed end of the
thrust tube under a single firing condition. When the detonation is initiated at the closed
end, there is a big pressure jump up to about 17.235 bars (1723.5 kPa), which is the
equilibrium CJ pressure of the propane/air detonation under the prescribed condition.

184

Figure 7.5: Time history of pressure at the closed end of the PDE tube.
The high pressure, however, is sustained for only a very short period of time
due to the induced flow expansion behind the traveling detonation wave. Pressure at the
closed end quickly reduces to only about 6.5 bars (650 kPa), and maintain a constant
level until the reflected expansion wave reaches the closed end of the PDE tube at about t
= 0.9 ms. Although 0.9 ms is less than 6% of one cycle (16.67 ms), it is the time period
when the PDE tube is producing thrust forces. Further examination shows that it takes

185
only about 0.361 ms for the detonation wave to travel through the tube. Thus, it takes
about 0.54 ms for the expansion wave to reach the closed end of the tube after the
detonation wave has left the PDE tube. Thus, the time period when the detonation wave
is traveling inside the PDE tube, is only about 2 percent of the PDE cycle (16.67 ms).
After the expansion wave has reached the closed end of the PDE tube, pressure
declines steeply. When t = 3.4 ms, the pressure at the closed-end of the tube reaches its
lowest value, which is less than half of the atmospheric pressure. This is the ideal time to
begin the refueling/purging process for the next firing cycle.
Figure 7.6 shows three snapshots of the pressure contour after the detonation
wave has left the thrust tube. The detonation wave quickly quenches and becomes a
non-reactive shock wave. The pressurized gases quickly expand in a spherical manner
forward and aft of the PDE tube. The expansion wave also travels upstream into the PDE
tube. Wilson and Paxson (2002) proposed a heuristic and insightful theoretical model for
the PDE plume dynamics. In the model, they pointed out the spherical expansion wave as
the final stage of the PDE plume expansion. The present CFD results strongly support
their proposed model.

186



Figure 7.6: Three snapshots of pressure contour of a PDE plume at different time of the
initial stage of expansion. Domain size: 12.5 ft in length, 4.5 ft in radius. t = 0.642, 1.124,
and 1.927 ms, respectively.

187
Figure 7.7 illustrates the vortex/shock interactions in the near field. Due to a
strong jet, a vortex ring is developed, which functions as a convergent/divergent nozzle.
The jet accelerates through this fluidic nozzle. Shown in Fig. 7.7(a), the supersonic jet is
adjusted to the surrounding air through a bow shock, which stands on the divergent
section of the fluid nozzle. At a later stage of the wave expansion, due to decreasing mass
flow rate, the vortex ring shrinks in radius. Refer to Fig. 7.7(b). In this transient process,
the shock wave would occasionally stand on the converging section of the fluid nozzle.
Under this condition, the jet behaves just like an un-started inlet and the shock wave
travels quickly upstream into the PDE tube. The above discussions illustrated the
mechanism responsible for the oscillating pressure waves at the closed end of the PDE
tube from 3.4 to 10 ms as shown in Fig. 7.5.
Figure 7.8 shows pressure histories at the sampling points 2 and 6 in L-1. At Point
2, which is closer to the tube exit, one dominant pressure peak is followed by a pressure
dip. At Point 6, the pressure peak and dip dissipate. Time histories of pressures at all
sampling points show similar pattern, i.e., one pressure peak followed by a trough and
subsequent minor oscillations. The farther away from the PDE tube exit, the pressure
peaks and troughs are smoother due to the decay of the wave. Figure 7.9 shows the time
histories of pressure and density at Point 2 on L-1 for three cycles of firing. Here pressure
trace shows little variations from one cycle to the other. However, time history of density
shows significant cycle-to-cycle variations.

188

(a)

(b)

(c)
Figure 7.7: Vortex-shock interaction in the near field of a PDE plume.

189

(a)

(b)
Figure 7.8: Time histories of pressure at two sampling points located at (a) (5.66, 5.875)
and (b) (22.64, 5.875).

190

(a)

(b)
Figure 7.9: Time histories of pressure and density at the location (5.66 in, 0.49 ft) for a
three-cycle calculation.

191
Peak pressures and their corresponding time for the PDE wave passing all
sampling points are shown in Fig. 7.10(a). The experimental data on L-1, denoted by
square symbols, are also included. Numerical results compare favorably with the
experimental data. The wave speeds of the pressure peaks and troughs are also checked
and compared with the experimental data. Refer to Fig. 7.10(b). To calculate the wave
velocity, we perform numerical sampling to identify the time when the pressure peaks
reach the certain mesh nodes. By checking a group of mesh nodes in the vicinity of the
sampling points, we then divided the displacements between the mesh nodes by time lag
to calculate the wave speed at the sampling point.
Once out of the PDE tube, the hot burned gases quench and the associated wave
speed decrease accordingly. Note that the wave speed at the first sampling point, which is
only 7.11 inches away from the PDE exit, is about 2,200 ft/sec (671 m/s). However the
CJ detonation speed is over 6,000 ft/sec (1830 m/s) inside the PDE chamber. The shock
wave experiences tremendous quenching and thus slow-down in a very short distance.

192

(a)

(b)
Figure 7.10: Calculated wave characteristics at sampling points along L-1, L-2, L-3 and
L-4: (a) peak pressure, and (b) the wave speed. The x axis is time in unit of s
(microsecond). The experimental data is measured on L-1 only.

193
Unlike the experimental results, the decay of the wave speeds predicted by the CFD
calculation is not monotonic. Instead, significant fluctuations of wave speeds can be
discerned. This may be caused by complex vortex structure in the PDE plume. Moreover,
numerical calculations were performed based on the assumption that the ambient
atmosphere is clean and at the specified initial condition, which may not be true for the
testing environment after several testing firings.
7.4 Far-Field Calculations
In this phase of modeling work, the computational domain is extended to 57 ft
(17.37 m) in length (including the tube) and 30 ft (9.14 m) in the radial direction. The
whole computational domain is divided into 6.4 million non-uniform quadrilateral cells,
clustered near the tube exit. Inside the PDE tube, the mesh keeps unchanged as compared
with that used in the near-field simulations.
It took us about 25 days of continuous running to simulate 6 cycles of PDE firing
on an 8-node Xeon cluster. The long running time was mainly due to (i) the large size of
the computational domain, (ii) a long simulated time period, and (iii) the necessary small
time steps. To examine the details of the flow features, we recorded the temporal histories
of pressure at a score of sampling points, located on concentric circles at 5, 10, 15, 20, ,
50 ft (or 1.52 m, 3.05 m, 5.22 m, 60.96 m, , 15.24 m) away from the detonation tube
exit. The sampling points on each circle are distributed evenly with off-axis angles at 0,

194
3, 6, 9, , 45. The detonation tube exit is taken as the origin of the coordinates. Refer
to Fig. 7.11 for the distribution of the sampling points.

Figure 7.11: A schematic of the computational domain and the distribution of sampling
points for far field calculations. Domain size: 57 ft in length, 30 ft in radius.

195
Figure 7.12 shows three snapshots of the pressure contour of the PDE plume at
the far field. In general, the near-field plume associated with each firing includes a
spherical shock fronts, a transitory jet, a vortex ring, and complex interactions between
the slow-moving vortex ring and the approaching shock front of the following firings. As
can be seen, following each spherical shock front, the gas jet and its peripheral vortex
ring are soon approached by the shock front generated by the next firing. Further
downstream, both the leading shock front and the vortex ring decay to weak pressure
waves.
Figure 7.13(a) shows pressure history at 3 sampling points located 10 ft away
from the tube exit, with off-axis angles 0, 5, and 30, respectively. As can be seen, at
each sampling point, there is a pressure peak followed by a pressure dip and subsequent
minor oscillations. The same pattern of the pressure history occurs at almost all of the
sampling points. Refer to Fig. 7.13(b) for the pressure histories at two sampling points at
a distance of 50 ft away from the detonation tube exit. Note that the wave amplitude is
much smaller than that at the distance of 10 ft. Moreover, at the locations very close to
the central line, a higher peak and a lower trough due to the interaction between the shock
and the vortex ring can be observed.

196

(a) t = 30 ms

(b) t = 50 ms
Continued
Figure 7.12: Three snapshots of pressure contour of a PDE plume at different time in the
far field. Domain size: 57 ft in length, 30 ft in radius.

197
Figure 7.12 continued

(c ) t = 80 ms

198

(a)

(b)
Figure 7.13: Time histories of pressure at various locations on the circles at (a) 10 ft and
(b) 50 ft away from the PDE tube exit.

199
In general, at some distance, the values of the peak pressure at the point located
on the centerline is larger than that at all other sampling points on the same circle.
However, the lowest value of the pressure trough may not occur at the sampling point on
the centerline as compared to other sampling points with the same distance from the
detonation tube exit. Instead, in the near-field area, it usually occurs at a location a little
off from the central line due to the existence of the vortex rings.
Away from the PDE tube, the shock fronts and the vortex rings decay to pressure
waves. At 10 feet away from the PDE tube exit, the amplitude of wave becomes about
44% of the ambient pressure. At about 30 ft away from the PDE tube exit, the wave
amplitude drops to about 5% of the ambient pressure. The change of wave amplitude
along the central line is plotted in Fig. 7.14. The y axis is the logarithm of the pressure
amplitude and the x axis is the distance from the tube exit. The sharp drop at the distance
of about 35 ft is attributed to the lack of the influence of vortex rings, which move very
slowly. The first vortex ring reaches about 35 ft away from the tube exit after 6 firings.
Near the end of the present calculation, the first vortex ring nearly stops propagating in
the axial direction. Since a pressure wave with amplitude of less than 3% of the
background pressure could be considered linear, Fig. 7.14 shows that 33 ft away from the
PDE tube exit, the flow becomes linear waves.

200

Figure 7.14: Amplitude of pressure fluctuations along the central line at different distance
from the PDE tube exit. The horizontal line denotes linear wave.

201
Figure 7.15 presents the profile of the sound strength in dB at the sampling points.
The sound strength is defined as P
dB
= 20 log [(P
Peak
- P

)/P
dB_Ref
], where P
Peak
is the peak
pressure recorded, P

is the background pressure of the flow field, the reference level

P
dB_Ref
is usually taken as the threshold of hearing, i.e., 0.02 mPa. As can be seen, the
detonation wave is about 144 dB at after 50 ft away from the PDE tube exit. The
decaying trend with respect to the increasing distance from the PDE is apparent. We were
informed that the sound strength data agrees well with the experimental result.

Figure 7.15: Sound strength (in dB) at different distance from the PDE tube exit.

202
7.5 Concluding Remarks
In this chapter, the CESE method has been employed to study the complex plume
dynamics of a PDE prototype fueled by propane/air mixtures. The flow condition was
chosen to mimic realistic PDE operating conditions. Numerical results show that in about
6% of a PDE cycle, the engine is actively producing thrust forces. The time for traveling
detonation wave through the PDE tube is less than 3% of one cycle. The rest of the time is
occupied by the gas blow-down, purging, and refueling processes, which heavily depend
on the flow conditions outside the PDE tube.
Outside the PDE tube, flow expands and accelerates through a vortex ring, which
behaves like a fluidic nozzle. Due to unsteadiness of the shock/vortex interactions,
alternating expansion and compression occur in the near field of the PDE. Away from the
PDE tube, a spherical expansion is the dominant flow feature. The wave strength reduces
quickly when the distance to the tube exit increases. After about 33 ft of propagation, the
waves become linear. However, the acoustic waves could be still as loud as nearly 150
dB at a distance of 50 ft away.


203
CHAPTER 8
CONCLUSIONS
This chapter summarizes the findings and the achievements of the present work.
Suggestions to the further development of the CESE code for modeling flows through
advanced propulsion systems are also tabulated. The chapter is divided to two sections: (i)
the achievements and (ii) suggestions.
8.1 Achievements and Findings
The most notable achievement of the present dissertation work is the development
of a high-quality CFD computer code based on the CESE method. The results show the
newly developed code can be routinely applied to simulate two- and three-dimensional
reacting and non-reacting flows using meshes composed of millions of cells. The
successful code development and applications also showed that the CESE method is
indeed a highly accurate and efficient CFD method. The synergy of this novel numerical
method and the cost-effective parallel computation could point to a new direction for
high performance computing for three-dimensional unsteady flows.

204
In this dissertation, although the applications of the code has been confined to
model detonation related flow fields, this state-of-the-art code can be readily used to
provide high-fidelity solutions of unsteady flows in various applications. In what follows,
we list the specific contributions by the present dissertation work.
1. The original space-time CESE method has been extended to use quadrilateral and
hexahedral cells to construct unstructured meshes for simulations of two- and
three-dimensional flows, respectively. The code can take an arbitrary unstructured
mesh generated by using a commercial code, e.g., Gambit in Fluent, and partition the
mesh for parallel computation. To support the distribution and collection of data for
parallel computing, extensive codes have been developed to support domain
decomposition, solution animation, and post processing for data analyses. The
structure of the software was designed based on the same principle of a commercial
CFD code. Extensive input and output facilities have been developed to connect the
code with commercial pre and post processors.
2. A 32-node cluster parallel computer was built as the hardware platform of the
parallel computing. Instead of purchasing a commercial cluster computer, we have
been building and maintaining cluster computers since 1999. The current 32-node
cluster computer used for the dissertation work is the 4
th
generation of the parallel

205
computers that we have built. Direct access to the hardware allows me to fine tune
my code for high performance calculations.
3. We have reported the one-, two- and three-dimensional simulations of detonation
waves by using the CESE method. In one-dimensional calculations, known features
of steady detonations and unsteady detonations with a galloping pattern were
successfully simulated. The results compared well with the results of classical
stability analyses. The mesh refinement study showed that based on the use of the
CESE method only 5 mesh nodes were needed to resolve the half reaction zone for
accurate solutions. For planar two-dimensional detonations, salient features have
been crisply resolved, including transverse waves, triple points, Mach stems, counter
rotating vortices, and unburned pockets inside the reaction zone. For
three-dimensional calculations, the flow structure is very similar to that of the
two-dimensional ones. However, further studies showed significance differences in
key flow parameters between the two and three-dimensional simulations. For the
same detonation wave, the averaged cell width of the two-dimensional result is about
30 to 40% smaller than that of the three-dimensional results. Moreover, both
amplitudes and the frequencies of the pressure fluctuations at the shock fronts of the
three-dimensional simulations are much higher than that of the two-dimensional
simulations, which in turn, as repeatedly pointed out in the literature, are also
significantly higher than that predicted by the classical one-dimensional ZND

206
solution. Therefore, we conclude that one has to resort to three-dimensional
calculations to catch the quantitative features of propagating detonation waves.
4. Direct numerical calculations of two- and three-dimensional implosions and
explosions have been conducted. In two-dimensional calculations, we focused on the
calculations of polygonal converging shock fronts. In the implosion phase, numerical
images showed the regenerations of rotated and smaller polygonal forms. Time
histories of flow variables at the focal point show the double implosion mechanism.
A successful detonation initiation was observed in one case with chemical reaction
involved. In three-dimensional calculations similar double implosion mechanism can
be seen. These results clearly demonstrate that the space-time CESE method is
capable to catch salient features of complex implosion/explosion flows. The results
also show that the mesh density is important to catch detailed flow features.
5. The complex plume dynamics of a propane/air PDE has been studied in the present
research. The flow condition was chosen to mimic realistic PDE operating conditions.
Numerical results show that in about 6% of a PDE cycle the engine is actively
producing thrust forces. The time for traveling detonation wave through the PDE tube
is less than 3% of one cycle. The rest of the time is occupied by the blow-down and
refueling processes, which heavily depend on the flow conditions outside the PDE
tube. Outside the PDE tube, flow expands and accelerates through a vortex ring, which
behaves like a fluid funnel. Due to unsteadiness of the shock/vortex interactions,

207
alternating expansion and compression occur in the near field of the PDE plume.
Away from the PDE tube, a spherical expansion is the dominant flow feature. The
wave strength reduces quickly when the distance to the tube exit increases. After
about 33 ft of propagation, the waves become linear. However, the acoustic waves
could be still as loud as nearly 150 dB even at the distance of 50 ft. Moreover, all
salient features of complex shock/shock and shock/vortex interactions are shown.
Many fine flow features are crisply resolved.
8.2 Recommendations for Future Work
In this section, specific improvements which could be implemented in the further
development of the CESE codes are tabulated. Based on our current results and
experience, we have the following suggestions for the future works:
1. In the present work, chemical reactions in detonation waves were modeled by using a
single-step irreversible finite-rate equation, because our main concern is the overall
code structure and high resolution of aerodynamics. To obtain more accurate
quantitative results in simulating chemically reacting flows, a realistic chemistry
model will be needed. A finite-rate chemistry model composed of multiple species
and multiple reactions steps could be used. However, inclusion of multiple species
will increase the number of equations to be solved. In addition, the vastly different
time scales among reaction steps also introduce the stiffness problem in coupling the

208
reactions with the flow equations. Compared with the simple model used in the
present work, the cost of including detailed chemistry model could be overwhelming.
It might be impractical for three-dimensional calculations of unsteady flows. An
effective way of incorporating realistic chemistry model is an important issue for the
further development of the CESE code.
2. Since the viscosity effect is unimportant in detonation, only Euler solver is used in
the present research. Our numerical results have shown the accuracy and efficiency
of the Euler solver. However, in other applications, the viscosity effect may be
important. In the calculation of imploding flows, it is shown that the shock front
converges to a singular point. The value of flow properties could be extremely high if
a fine mesh is used. The finer the mesh is, the higher central temperature and
pressure we get. There is no upper limit. Thus, the remedy for accurate simulation is
to use the Navier-Stokes solver in the central area, where the viscosity effect could
be involved to damp the singularity. Note the Navier-Stokes solver does not have to
be used on the whole domain. It is only a remedy to solve the singular point problem
encountered in the central point. In the future, we suggest that the NS solver need to
be thoroughly checked and validated. There are several ways of implementing the
viscous effects and the pros and cons of each approach need to be evaluated.
3. As shown in the calculations of three-dimensional implosion/explosion processes,
the mesh density may be critical to get accurate numerical results, e.g., the strength

209
of the second and the third imploding waves. Although the CESE method can
produce very accurate numerical results, it is still necessary to employ high
resolution meshes in local region where the flow has high gradients. Therefore, we
suggest that one should seriously consider implementing local-mesh-refinement
method to the present CESE code. On top of a relative coarse mesh, one could
impose patches of fine meshes to provide local resolution. In this case, a new data
structure will need to be developed for parallel computation.
4. To solve unsteady hyperbolic systems, the Courant-Friedrichs-Lewy (CFL) condition
is an important stability criterion. This is also true for the CESE method. Usually the
iteration time step must be small enough to keep the scheme stable. However, if the
time iteration step employed is too small, i.e., a very small CFL number, a much
longer running time will be needed and inaccurate numerical results will occur. In
other words, the solution could be smeared in the area where the flow properties are
of large gradient. Recently, Chang (2002) proposed to solve this problem by a new
numerical scheme, i.e., the Courant number insensitive CESE scheme, which has
been successfully used to solve several one- and two-dimensional problems. This
improvement should be included in the future development of the current
three-dimensional CESE solver.
5. The pre- and post-processing may be further developed for the three-dimensional
CESE code. The issue is that although we could perform calculations by using a very

210
large mesh of more than 10 million mesh nodes, we always have to feed the entire
data set back to one single workstation in order to do post processing for solution
animation or data analyses. Due to the huge amount of data produced by the 3D
solver and pre-processor, one cannot use all the data for post processing. There are
two solutions for the problem: (i) one could develop a parallel post processing
software, and (ii) one could use a smaller set of data for post processing. While there
are several research and commercial groups who are in the process of developing the
parallel post processor, we have no interest in writing our own code for post
processing for plotting the results. In the current work, we took the second approach
and employed data interpolation and extrapolation to decrease the size of the output
data. However, the software of this process is still far from perfect. An accurate and
efficient method to produce high-quality and compact output data is desired.
Moreover, unstructured mesh decomposition mainly relies on METIS in the present
work. Though it could be done manually on domains of simple geometry, in most of
cases, METIS is needed for high-quality domain decomposition for computational
domains of complex shapes. However, METIS is a sequential program and it is
incapable to perform domain decomposition for really large meshes of more than 100
millions of cells, because of the limit of the memory on a single computer and the
impractically long running time. Therefore, a parallelized, high-quality mesh
decomposer is desired for the future simulations using very large meshes.

211
BIBLIOGRAPHY
Aki, T. and Higashino, F., 1990, A Numerical Study on Implosion of Polygonally
Interacting Shocks and Consecutive Explosion in a Box, Current topics in shock waves;
17th Proceedings of the International Symposium on Shock Waves and Shock Tubes, pp.
167-172.
Austin, J. M. and Shepherd, J. E., 2000, Detonations in Hydrocarbon Fuel Blends,
Explosion Dynamics Laboratory Report FM99-6.
Baker, W. E., 1973, Explosions in Air, Wilfred Baker Engineering, San Antonio, TX,
Chap. 7.
Bauwens, L. and Williams, D. N., 1996, Computation and Visualization of Detailed
Detonation Structures, 34th Aerospace Sciences Meeting and Exhibit, AIAA Paper
96-0343.
Bauwens, L., 2000, Ignition Between a Shock and a Contact Surface: Influence of the
Downstream Temperature, Proceedings of the Combustion Institute, Vol. 28, pp.
653-661.
Bauwens, L. and Liang, Z., 2002, Shock Formation Ahead of Hot Spots, Proceedings
of the Combustion Institute, Vol. 29, pp. 2795-2802.
Bauwens, L. and Williams, D. N., 2002, Three-Dimensional Detonation Cells, 40th
AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper 2002-0774.
Belokon, V. A., Petrukhin, A. I. and Proskuryakov, V. A., 1965, Entrance of a Strong
Shock Wave into a Wedge-Like Cavity, Soviet Physics Jetp, Vol. 21, No. 1, pp. 33-40.


212
Ben-Artzi, M. and Falvovitz, J., 1984, A Second-Order Godunov-Type Scheme for
Compressible Fluid Dynamics, Journal of Computational Physics, Vol. 55, No. 1, pp.
1-32.
Book, D. L., Boris, J. P. and Hain, K., 1975, Flux-Corrected Transport II.
Generalizations of the Method, Journal of Computational Physics, Vol. 18, No. 3, pp.
248-283.
Boris, J. P. and Book, D. L., 1973, Flux-Corrected Transport. I. SHASTA, A Fluid
Transport Algorithm That Works, Journal of Computational Physics, Vol. 11, No. 1, pp.
38-69.
Boris, J. P. and Oran, E. S., 1987, The Numerical Simulation of Compressible Reactive
Flows, 19th Fluid Dynamics, Plasma Dynamics and Lasers Conference, AIAA Paper
87-1323.
Bourlioux, A., Majda, A. J. and Roytburd, V., 1991, Theoretical and Numerical
Structure for Unstable One-Dimensional Detonations, SIAM Journal of Applied
Mathematics, Vol. 51, pp. 303-343.
Bourlioux, A. and Majda, A. J., 1992, Theoretical and Numerical Structure for Unstable
Two-Dimensional Detonations, Combustion and Flame, Vol. 90, pp. 211-229.
Bourlioux, A. and Majda, A. J., 1995, Theoretical and Numerical Structure of Unstable
Detonations, Philosophical Transactions of the Royal Society of London, Series A, Vol.
350, pp. 29-68.
Boyer, D. W., 1960, An Experimental Study of the Explosion Generated by a
Pressurized Sphere, Journal of Fluid Mechanics, Vol. 9, pp. 401-429.
Bratkovich, T. E., Aarnio, M. J., Williams, J. T. and Bussing, T. R. A., 1997, An
Introduction to Pulse Detonation Rocket Engines (PDREs), 33rd
AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA Paper 97-2742.
Brode, H. L., 1955, Numerical Solution of Spherical Blast Waves, Journal of Applied
Physics, Vol. 26, pp. 766-775.
Brode, H. L., 1959, BlastWave from a Spherical Charge, Physics of Fluids, Vol. 2, No.
2, pp. 217-229.

213
Bussing, T. R. A. and Pappas, G., 1993, A Novel Hydrogen Fueled Propulsion System,
NASA Langley Research Center Project Summary, SBIR Phase I Final Report.
Bussing, T. and Pappas, G., 1994, An Introduction to Pulse Detonation Engines, 32nd
Aerospace Sciences Meeting and Exhibit, AIAA Paper 94-0263.
Bussing, T. R. A., Hinkey, J. B. and Kaye, L., 1994, Pulse Detonation Engine
Preliminary Design Considerations, 30th AIAA/ASME/SAE/ASEE Joint Propulsion
Conference and Exhibit, AIAA Paper 94-3220.
Butler, D. S., 1954, Converging Spherical and Cylindrical Shocks, Armament Research
and Development Establishment, Ministry of Supply No. 54/54.
Bykovskii, F. A. and Mitrofanov, V. V., 1980, Detonation Combustion of a Gas Mixture
in a Cylindrical Chamber, Combustion, Explosion, and Shock Waves, Vol. 16, No. 5, pp.
570-578.
Bykovskii, F. A. and Vedernikov, E. F., 1996, Discharge Coefficients of Nozzles and of
Their Combinations in Forward and Reverse Flow, Journal of Applied Mechanics and
Technical Physics, Vol. 37, No. 4, pp. 541-546.
Bykovskii, F. A. and Mitrofanov, V. V., 2000, A Continuous Spin Detonation in Liquid
Fuel Sprays, Control of Detonation Processes, edited by S. F. G. Roy, D. Netzer, and A.
Borisov, Elex-KM Publishers, Moscow, pp. 209-211.
Bykovskii, F. A. and Vedernikov, E. F., 2003, Continuous Detonation of a Subsonic
Flow of a Propellant, Combustion, Explosion, and Shock Waves, Vol. 39, No. 3, pp.
323-334.
Cambier, J. L. and Adelman, H. G., 1988, Preliminary Numerical Simulations of a Pulse
Detonation Wave Engine, AIAA Paper 88-2960.
Chang, E. and Kailasanath, K., 2000, Shock Wave Interactions with Particles and Fuel
Droplets, 38th AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper 2000-0319.
Chang, S. C., 1995, The Method of Space-Time Conservation Element and Solution
Element - A New Approach for Solving the Navier-Stokes and Euler Equations, Journal
of Computational Physics, Vol. 119, pp. 295-324.


214
Chang, S. C., Wang, X. Y. and Chow, C. Y., 1999, The Space-Time Conservation
Element and Solution Element Method: A New High-Resolution and Genuinely
Multidimensional Paradigm for Solving Conservation Laws, Journal of Computational
Physics, Vol. 156, pp. 89-136.
Chang, S. C., Yu, S. T., Himansu, A., Wang, X. Y., Chow, C. Y. and Loh, C. Y., 1999,
The Method of Space-Time Conservation Element and Solution Element - A New
Paradigm for Numerical Solution of Conservation Laws, Computational Fluid
Dynamics Review 1998, edited by M. Hafez and K. Oshima, World Scientific Publisher,
pp. 206-240.
Chang, S.-C., Wang, X. Y. and To, W. M., 2000, Application of the Space-Time
Conservation Element and Solution Element Method to One-Dimensional
Convection-Diffusion Problems, Journal of Computational Physics, Vol. 165, No. 1, pp.
189-215.
Chang, S. C., 2002, Courant Number Insensitive CE/SE Schemes, 38th
AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA Paper
2002-3890.
Chang, S. C., Himansu, A., Loh, C. Y., Wang, X. Y. and Yu, S. T., 2003, Robust and
Simple Non-Reflecting Boundary Conditions for the Euler Equations - A New Approach
Based on the Space-Time CE/SE Method, NASA Technical Report, TM-2003-212495.
Chang, S. C. and Wang, X. Y., 2003, Multi-Dimensional Courant Number Insensitive
CE/SE Euler Solvers for Applications Involving Highly Nonuniform Meshes, 39th
AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA Paper
2003-5285.
Chao, T. W., Wintenberger, E. and Shepherd, J. E., 2001, On the Design of Pulse
Detonation Engines, GALCIT Report FM00-7.
Chapman, D. L., 1899, On the Rate of Explosion in Gases, Philosophical Magazine,
Vol. 47, pp. 90-104.
Charrier, P. and Tessieras, B., 1986, On Front-Tracking Methods Applied to Hyperbolic
Systems of Nonlinear Conservation Laws, SIAM Journal on Numerical Analysis, Vol.
23, No. 3, pp. 461-472.


215
Chitsomboon, T., Kumar, A. and Tiwari, S. N., 1987, Numerical Study of Finite-Rate
Supersonic Combustion Using Parabolized Equations, AIAA Paper 87-0088.
Cooper, M., Jackson, S. and Shepherd, J. E., 2000, Effect of Deflagration-to-Detonation
Transition on Pulse Detonation Engine Impulse, GALCIT Report FM00-3.
Cullen, R. E., Nicholls, J. A. and Ragland, K. W., 1966, Feasibility Studies of a Rotating
Detonation Wave Rocket Motor, Journal of Spacecraft and Rockets, Vol. 3, No. 6, pp.
893-898.
De Neef, T. and Nechtman, C., 1978, Numerical Study of the Flow due to a Cylindrical
Implosion, Journal of Computational Physics, Vol. 6, pp. 185-202.
DeBonis, J. R., Trefny, C. J. and C. J. Steffen, J., 1999, Inlet Development for a Rocket
Based Combined Cycle, Single Stage to Orbit Vehicle Using Computational Fluid
Dynamics, 35th AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit,
AIAA Paper 99-2239.
Devore, C. R. and Oran, E. S., 1992, The Stability of Imploding Detonations in the
Geometrical Shock Dynamics (CCW) Model, Physics of Fluids A: Fluid Dynamics, Vol.
4, No. 4, pp. 835-844.
Dionne, J. P., Ng, H. D. and Lee, J. H. S., 2000, Transient Development of
Friction-Induced Low-Velocity Detonations, Proceedings of the Combustion Institute,
Vol. 28, pp. 645-651.
Doering, W., 1943, On Detonation Processes in Gases, Annalen der Physik, Vol. 43, pp.
421-436.
Dryer, F. L., 1991, The Phenomenology of Modeling Combustion Chemistry, Section
5.1: One- and Two-Step Global Mechanisms, Fossil Fuel Combustion A Source Book,
edited by Bartok, W. and Sarofim, A. F., Wiley InterScience, New York, Chap. 5.
Ebrahimi, H. B., Mohanraj, R. and Merkle, C. L., 2000, Multi-Level Analysis of Pulsed
Detonation Engines, 36th AIAA/ASME/SAE/ASEE Joint Propulsion Conference and
Exhibit, AIAA Paper 2000-3589.
Edwards, D. H., Parry, D. J. and Jones, A. T., 1966, On the coupling between spinning
detonation and oscillations behind the wave, British Journal of Applied Physics, Vol. 17,
pp. 1507-1510.

216
Eidelman, S., Grossmann, W. and Lottati, I., 1990, Computational Analysis of Pulsed
Detonation Engines and Applications, AIAA Paper 90-0460.
Eidelman, S., Grossmann, W. and Lottati, I., 1990, Air-Breathing Pulsed Detonation
Concept: A Numerical Study, AIAA Paper 90-2420.
Erpenbeck, J. J., 1962, Stability of Steady-State Equilibrium Detonations, The Physics
of Fluids, Vol. 5, pp. 604-614.
Erpenbeck, J. J., 1964, Stability of idealized One-Reaction Detonations, The Physics of
Fluids, Vol. 7, pp. 684-696.
Erpenbeck, J. J., 1966, Detonation Stability for Disturbances of Small Transverse
Wavelength, The Physics of Fluids, Vol. 9, No. 7, pp. 1293-1306.
Eto, K., Tsuboi, N. and Hayashi, A. K., 2004, Three-Dimensional CJ Detonation in
Rectangular Tube, 42nd AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper
2004-311.
Falcovitz, J. and Birman, A., 1994, A Singularities Tracking Conservation Laws
Scheme for Compressible Duct Flows, Journal of Computational Physics, Vol. 115, No.
2, pp. 431-439.
Fickett, W. and Wood, W. W., 1966, Flow Calculations for Pulsating One-Dimensional
Detonations, Physics of Fluids, Vol. 9, No. 5, pp. 903-916.
Fickett, W. and Davis, W. C., 1979, Detonation, University of California Press, Berkeley,
California, pp. 13-15, 42-51, 151-153, Chapter 7.
Flores, J. and Holt, M., 1981, Glimm's Method Applied to Underwater Explosions,
Journal of Computational Physics, Vol. 44, No. 2, pp. 377-387.
Friedman, M. P., 1961, A Simplified Analysis of Spherical and Cylindrical Blast
Waves, Journal of Fluid Mechanics, Vol. 11, pp. 1-15.
Fujiwara, T. and Reddy, K. V., 1989, Propagation Mechanism of Detonation -
Three-Dimensional Phenomen, Memoirs of the Faculty of Engineering, Nagoya
University, Vol. 41, No. 1, pp. 93-111.


217
Gamezo, V. N., Desbordes, D. and Oran, E. S., 1999, Formation and Evolution of
Two-dimensional Cellular Detonations, Combustion and Flame, Vol. 116, pp. 154-165.
Gamezo, V. N., Vasilev, A. A., Khokhlov, A. M. and Oran, E. S., 2000, Fine Cellular
Structures Produced by Marginal Detonations, Proceedings of the Combustion Institute,
Vol. 28, pp. 611-617.
Gamezo, V. N., Khokhlov, A. M. and Oran, E. S., 2002, Effects of Wakes on
Shock-Flame Interactions and Deflagration-to-Detonation Transition, Proceedings of the
Combustion Institute, Vol. 29, pp. 2803-2808.
Glass, I. I. and Sislian, J. P., 1994, Non-Stationary Flows and Shock Waves, Oxford
Science Publications, Chap. 12.
Glimm, J., 1965, Solutions in the Large For Nonlinear Hyperbolic Systems of
Equations, Communications on Pure and Applied Mathematics, Vol. 18, No. 4, pp.
697-715.
Gordon, S. and McBride, B. J., 1994, Computer Program for Calculation of Complex
Chemical Equilibrium Compositions and Applications. I. Analysis, NASA Reference
Publication, NASA RP-1311.
Gruber, M. R., Nejad, A. S., Chen, T. H. and Dutton, J. C., 1997, Compressibility
effects in supersonic transverse injection flowfields, Physics of Fluids, Vol. 9, No. 5, pp.
1448-1461.
Guderley, G., 1942, Kugelige and Zylindrische Verdichtungsstosse in der Nake des
Kugelmittelpunktes Lzw der Zylinderachse, Luftfahrtforschung, Vol. 19, pp. 302-312.
Guirao, C. M., Knystautas, R., Lee, J. H., Benedick, W. and Berman, M., 1982,
Hydrogen-Air Detonations, Proceedings of the Combustion Institute, Vol. 19, pp.
583-590.
Guo, C., Thomas, G., Li, J. and Zhang, D., 2002, Experimental study of gaseous
detonation propagation over acoustically absorbing walls, Shock Waves, Vol. 11, pp.
353-359.
Hanana, M., Lefebvre, M. H. and Van Tiggelen, P. J., 2000, Preliminary Experimental
Investigation of the Pressure Evolution in Detonation Cells, Experimental Thermal and
Fluid Science, Vol. 21, pp. 64-70.

218
Hanana, M., Lefebvre, M. H. and Van Tiggelen, P. J., 2001, Pressure profile in
detonation cells with rectangular and diagonal structures, Shock Waves, Vol. 11, pp.
77-88.
Harten, A., Osher, S. J., Engquist, B. E. and Chakravarthy, S. R., 1986, Some Results on
Uniformly High-order Accurate Essentially Non-oscillatory Schemes, Journal of
Applied Numerical Mathematics, Vol. 2, pp. 347-377.
He, H., Zhang, Z. C., Yu, S. T. J. and Chang, S. C., 2001, Three-Dimensional
Simulations of Detonations by the CE/SE Method Using a Low-Cost Beowulf Cluster,
39th AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper 2001-0140.
He, H., Zhang, Z.-C., Yu, S.-T. J. and Jorgenson, P. C. E., 2002, Application of the
CESE Method to PDE Plume Dynamics Using a Beowulf Cluster, 40th AIAA Aerospace
Sciences Meeting and Exhibit, AIAA Paper 2002-0614.
He, H. and Yu, S.-T. J., 2003, Three-Dimensional Simulation of Transverse Injection in
a Supersonic Flow by the CESE Method, 41st AIAA Aerospace Sciences Meeting and
Exhibit, AIAA Paper 2003-0375.
He, H., Yu, S.-T. J. and Zhang, Z.-C., 2005, Direct Calculations of One-, Two-, and
Three-Dimensional Detonations by the CESE Method, 43rd AIAA Aerospace Sciences
Meeting and Exhibit, AIAA Paper 2005-0229.
He, L., 1996, Theoretical Determination of the Critical Conditions for the Direct
Initiation of Detonation in Hydrogen-Oxygen Mixtures, Combustion and Flame, Vol.
104, pp. 401-418.
Hedges, K. R. and Hill, P. G., 1974, Compressible Flow Ejectors. Part I. Development
of a Finite Difference Flow Model, Journal of Fluids Engineering (Transactions of the
ASME), Vol. 96, No. 3, pp. 272-281.
Hedges, K. R. and Hill, P. G., 1974, Compressible Flow Ejectors. Part II: Flow Field
Measurements and Analysis, Journal of Fluids Engineering (Transactions of the ASME),
Vol. 96, No. 3, pp. 282-288.
Heiser, W. H. and Pratt, D. T., 1994, Hypersonic Airbreathing Propulsion, AIAA Inc.,
Washington DC.


219
Heiser, W. H. and Pratt, D. T., 2002, Thermodynamic Cycle Analysis of Pulse
Detonation Engines, Journal of Propulsion and Power, Vol. 18, No. 1, pp. 68-76.
Helman, D., Shreeve, R. P. and Eidelman, S., 1986, Detonation Pulse Engine, AIAA
Paper 86-1683.
Higgins, A. J., Radulescu, M. I. and Lee, J. H. S., 1998, Initiation of Cylindrical
Detonation by Rapid Energy Deposition along a Line, Proceedings of the Combustion
Institute, Vol. 27, pp. 2215-2223.
Hill, P. G. and Peterson, C. R., 1965, Mechanics and Thermodynamics of Propulsion,
Addison-Wesley Publishing Company, Reading, Massachusetts.
Hinkey, J. B., Bussing, T. R. A. and Kaye, L., 1995, Shock Tube Experiments for the
Development of a Hydrogen-Fueled Pulse Detonation Engine, 31st
AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA Paper 95-2578.
Hu, X. Y., Zhang, D. L. and Jiang, Z. L., 2002, Analytical study of idealized
two-dimensional cellular detonations, Shock Waves, Vol. 11, pp. 475-480.
Hwang, P., Fedkiw, R. P., Merriman, B., Aslam, T. D., Karagozian, A. R. and Osher, S.
J., 2000, Numerical Resolution of Pulsating Detonation Waves, Combustion Theory
Modeling, Vol. 4, No. 3, pp. 217-240.
Im, K.-S., Yu, S.-T. J., Kim, C.-K., Chang, S.-C. and Jorgenson, P. C. E., 2002,
Application of the CESE Method to Detonation with Realistic Finite-Rate Chemistry,
41st Aerospace Sciences Meeting & Exhibit, AIAA Paper 2002-1020.
Im, K.-S., Lai, M.-C., Yu, S.-T. J. and Mathe son, R. R., 2004, Simulation of Spray
Transfer Processes in Electrostatic Rotary Bell Sprayer, Journal of Fluids Engineering
(Transactions of the ASME), Vol. 126, No. 3, pp. 449-456.
Jackson, S. and Shepherd, J., 2002, Initiation Systems for Pulse Detonation Engines,
38th AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA Paper
2002-3627.
Jackson, S., Grunthaner, M. and Shepherd, J., 2003, Wave Implosion as an Initiation
Mechanism for Pulse Detonation Engines, 39th AIAA/ASME/SAE/ASEE Joint
Propulsion Conference and Exhibit, AIAA Paper 2003-4820.

220
Jones, S. M. and Welch, G. E., 1996, Performance Benefits for Wave Rotor-Topped Gas
Turbine Engines, NASA Technical Memorandum, NASA TM-107193.
Jouguet, E., 1906, On the Propagation of Chemical Reactions in Gases, Journal de
Mathmatiques Pures et Appliques, Vol. 1, pp. 347-425, continued in Vol. 2, pp. 5-85.
Kailasanath, K. and Oran, E. S., 1983, Ignition of Flamelets Behind Incident Shock
Waves and Transition to Detonation, Combustion Science and Technology, Vol. 34, pp.
345-362.
Kailasanath, K., Oran, E. S., Boris, J. P. and Young, T. R., 1985, Determination of
Detonation Cell Size and the Role of Transverse Waves in Two-Dimensional
Detonations, Combustion and Flame, Vol. 61, pp. 199-209.
Kailasanath, K. and Oran, E. S., 1985, Power-Energy Relations for the Direct Initiation
of Gaseous Detonations, Dynamics of Shock Waves, Explosions, and Detonations, edited
by J. R. Bowen, N. Manson, A. K. Oppenheim and R. I. Soloukhin, Vol. 94 of Progress
in Astronautics and Aeronautics, pp. 38-54.
Kailasanath, K., Oran, E. S. and Boris, J. P., 1986, Numerical Simulations of the
Initiation of Gaseous Detonations in Shock Tubes, Shock Waves and Shock Tubes,
edited by Bershader and Hanson, Stanford University Press, Stanford, California, pp.
367-373.
Kailasanath, K., 1999, Applications of Detonations to Propulsion: A Review, 37th
AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper 99-1067.
Kailasanath, K., Patnaik, G. and Li, C., 1999, Computational Studies of Pulse
Detonation Engines: A Status Report, 35th AIAA/ASME/SAE/ASEE Joint Propulsion
Conference and Exhibit, AIAA Paper 99-2634.
Kailasanath, K. and Patnaik, G., 2000, Performance Estimates of Pulse Detonation
Engines, Proceedings of the Combustion Institute, Vol. 28, pp. 595-601.
Kailasanath, K., 2001, A Review of PDE Research: Performance Estimates, 39th
Aerospace Sciences Meeting and Exhibit, AIAA Paper 2001-0474.
Kailasanath, K., Patnaik, G. and Li, C., 2002, The Flowfield and Performance of Pulse
Detonation Engines, Proceedings of the Combustion Institute, Vol. 29, pp. 2855-2862.

221
Kamiyama, Y. and Matsuo, A., 2000, Flow Features of Shock-Induced Combustion
Around Cylindrical Projectiles, Proceedings of the Combustion Institute, Vol. 28, pp.
671-677.
Kaneshige, M. and Shepherd, J. E., Detonation Database, Technical Report FM97-8,
GALCIT, July 1997. See also the electronic hypertext version at
http://www.galcit.caltech.edu/detn_db/html/.
Kasahara, J., Takazawa, K., Arai, T., Tanahashi, Y., Chiba, S. and Matsuo, A., 2002,
Experimental Investigations of Momentum and Heat Transfer in Pulse Detonation
Engines, Proceedings of the Combustion Institute, Vol. 29, pp. 2847-2854.
Kawai, S. and Fujiwara, T., 2002, Numerical Analysis of 1st and 2nd Cycles of
Oxyhygrogen PDE, 40th Aerospace Sciences Meeting & Exhibit, AIAA Paper
2002-0929.
Korobeinikov, V. P., Levin, V. A., Markov, V. V. and Chernyi, G. G., 1972,
Propagation of Blast Wave in a Combustible Gas, Astronautica Acta, Vol. 17, pp.
529-537.
Krzycki, L. J., 1962, Performance Characteristics on an Intermittent-Detonation
Device, U.S. Naval Ordnance Test Station NAVWEPS Report 7655, AD-284312.
Kuznetsov, M. S., Alekseev, V. I., Dorofeev, S. B., Matsukov, I. D. and Boccio, J. L.,
1998, Detonation Propagation, Decay, and Reinitiation in Nonuniform Gaseous
Mixtures, Proceedings of the Combustion Institute, Vol. 27, pp. 2241-2247.
Lax, P. D., 1954, Weak solutions of Nonlinear Hyperbolic Equations and their
Numerical Computation, Communications on Pure and Applied Mathematics, Vol. 7,
No. 1, pp. 159-193.
Lee, H. I. and Stewart, D. S., 1990, Calculation of linear detonation instability:
one-dimensional instability of plane detonation, Journal of Fluid Mechanics, Vol. 216,
pp. 103-132.
Lee, J. H. and Lee, B. H. K., 1965, Cylindrical Imploding Shock Waves, The 5th Shock
Tube Symposium, U.S. Naval Ordnance Laboratory, White Oak, Silver Spring, Maryland,
pp. 2148-2152.


222
Li, C. and Kailasanath, K., 2000, Detonation Transmission and Transition in Channels
of Different Sizes, Proceedings of the Combustion Institute, Vol. 28, pp. 603-609.
Li, C., Kailasanath, K. and Patnaik, G., 2000, A Numerical Study of Flow Field
Evolution in a Pulse Detonation Engine, 38th AIAA Aerospace Sciences Meeting and
Exhibit, AIAA Paper 2000-0314.
Loh, C. Y., Chang, S. C., Scott, J. R., and Yu, S. T., 1995, Application of the Method of
Space-Time Conservation Element and Solution Element to Aeroacoustics Problems, A
Collection of Technical Papers, Volume II, 6th International Symposium on
Computational Fluid Dynamics, pp. 713-718,.
Lu, T., Law, C. K. and Ju, Y., 2003, Some Aspects of Chemical Kinetics in
Chapman-Jouguet Detonation: Induction Length Analysis, Journal of Propulsion and
Power, Vol. 19, pp. 901-907.
Mallard, E. and Le Chatelier, H., 1881, On the Propagation Velocity of Burning in
Gaseous Explosive Mixtures, Comptes Rendus Hebdomadaire des Scances de
l'Acadmie des Sciences, Vol. 93, pp. 145-148.
McBride, B. J. and Gordon, S., 1996, Computer Program for Calculation of Complex
Chemical Equilibrium Compositions and Applications. II. User's Manual and Program
Description, NASA Reference Publication, NASA RP-1311-P2.
Menikoff, R., Lackner, K. S. and Bukiet, B. G., 1996, Modeling Flows with Curved
Detonation Waves, Combustion and Flame, Vol. 104, No. 3, pp. 219-240.
Mikhailov, V. V. and Topchiyan, M. E., 1965, Study of Continuous Detonation in an
Annular Channel, Combustion, Explosion, and Shock Waves, Vol. 1, No. 4, pp. 12-14.
Murray, S. B., Thibault, P. A., Zhang, F., Bjerketvedt, D., Sulmistras, A., Thomas, G. O.,
Jenssen, A. and Moen, I. O., 2001, The Role of Energy Distribution on the Transmission
of Detonation, High-Speed Deflagration and Detonation: Fundamentals and Control,
edited by G. D. Roy, Frolov, S. M., Netzer, D. W., and Borisov, A. A., Elex-KM
Publishers, Moscow, Russia, pp. 139-162.
Nalim, M. R., 1995, Preliminary Assessment of Combustion Modes for Internal
Combustion Wave Rotors, 31st AIAA/ASME/SAE/ASEE Joint Propulsion Conference
and Exhibit, AIAA Paper 95-2801.

223
Nicholls, J. A., Wilkinson, H. R. and Morrison, R. B., 1957, Intermittent Detonation as a
Thrust-Producing Mechanism, Jet Propulsion, Vol. pp. 534-541.
Nikolaev, Y. A. and Topchiyan, M. E., 1977, Analysis of Equilibrium Flows in
Detonation Waves in Gases, Combustion, Explosion, and Shock Waves, Vol. 13, No. 3,
pp. 327-337.
Oran, E., Young, T. and Boris, J., 1978, Application of Time-Dependent Numerical
Methods to the Description of Reactive Shocks, Proceedings of the Combustion Institute,
Vol. 17, pp. 43-54.
Oran, E. S., Boris, J. P. and Young, T., 1981, Numerical Simulations of Detonations in
Hydrogen-Air and Methane-Air Mixtures, Proceedings of the Combustion Institute, Vol.
18, pp. 1641-1649.
Oran, E. S. and Boris, J. P., 1982, Weak and Strong Ignition. II. Sensitivity of the
Hydrogen-Oxygen System, Combustion and Flame, Vol. 48, No. 2, pp. 149-161.
Oran, E. S., Young, T. R., Boris, J. P. and Cohen, A., 1982, Weak and Strong Ignition. I.
Numerical Simulations of Shock Tube Experiments, Combustion and Flame, Vol. 48,
No. 2, pp. 135-148.
Oran, E. S. and DeVore, C. R., 1994, The Stability of Imploding Detonations: Results of
Numerical Simulations, Physics of Fluids, Vol. 6, No. 1, pp. 369-380.
Papalexandris, M. V., 1997, Unsplit Numerical Schemes for Hyperbolic Systems of
Conservation Laws with Source Terms, Ph.D. Dissertation, California Institute of
Technology.
Park, S.-J., Yu, S.-T., Lai, M.-C., Chang, S.-C. and Jorgenson, P. C. E., 1998, Direct
Calculations of Stable and Unstable ZND Detonations by the Space-Time Conservation
Element and Solution Element Method, 34th AIAA/ASME/SAE/ASEE Joint Propulsion
Conference and Exhibit, AIAA Paper 98-3212.
Park, S. J., Yu, S. T., Lai, M. C., Chang, S. C. and Jorgenson, P. C. E., 1999, Numerical
Calculation of Unstable Detonations by the Space-Time Conservation Element and
Solution Element Method, 37th AIAA Aerospace Sciences Meeting and Exhibit, AIAA
Paper 99-0491.


224
Paxson, D. E., 1993, A Comparison Between Numerically Modelled and Experimentally
Measured Loss Mechanisms in Wave Rotors, 29th AIAA/ASME/SAE/ASEE Joint
Propulsion Conference and Exhibit, AIAA Paper 93-2522.
Paxson, D. E., 1995, A Numerical Model for Dynamic Wave Rotor Analysis, 31st
Joint Propulsion Conference and Exhibit cosponsored by AIAA, ASME, SAE and ASEE,
AIAA Paper 95-2800.
Paxson, D. E., 1996, A Numerical Investigation of the Startup Transient in a Wave
Rotor, NASA Technical Memorandum, NASA TM-107196.
Payne, R. B., 1956, A Numerical Method for a Converging Cylindrical Shock, Journal
of Fluid Mechanics, Vol. 2, pp. 185-200.
Pegg, R. J., Couch, B. D. and Hunter, L. G., 1996, Pulse Detonation Engine Air
Induction System Analysis, 32nd AIAA/ASME/SAE/ASEE Joint Propulsion Conference
and Exhibit, AIAA Paper 96-2918.
Povinelli, L. A. and Yungster, S., 2002, Airbreathing Pulse Detonation Engine
Performance, NASA Technical Memorandum, NASA TM-2002-211575.
Povinelli, L. A. and Yungster, S., 2003, Real Gas Effects on the Performance of
Hydrocarbon-Fueled Pulse Detonation Engines, 41th AIAA Aerospace Sciences Meeting
and Exhibit, AIAA Paper 2003-0712.
Qin, J. R., Yu, S.-T. J., Zhang, Z.-C. and Lai, M.-C., 2001, Simulation of Cavitating
Flows by the Space-Time CESE Method, 39th Aerospace Sciences Meeting and Exhibit,
AIAA Paper 2001-1043.
Quirk, J. J., 1994, Godunov-Type Schemes Applied to Detonation Flows, Combustion
in High-Speed Flows, edited by T. L. J. a. A. K. J. Buckmaster, Kluwer Academic
Publishers, Dordrecht, pp. 575-596,
Radulescu, M. I., Higgins, A. J., Lee, J. H. S. and Murray, S. B., 2000, On the Explosion
Length Invariance in Direct Initiation of Detonation, Proceedings of the Combustion
Institute, Vol. 28, pp. 637-644.
Radulescu, M. I., Ng, H. D., Lee, J. H. S. and Varatharajan, B., 2002, The Effect of
Argon Dilution on the Stability of Acetylene/Oxygen Detonations, Proceedings of the
Combustion Institute, Vol. 29, pp. 2825-2831.

225
Rockwell, D. and Naudascher, E., 1978, Review - Self-Sustaining Oscillations of Flow
Past Cavities, Transaction of ASME, Vol. 100, pp. 152-165.
Romano, M. P., Radulescu, M. I., Higgins, A. J., Lee, J. H. S., Pitz, W. J. and Westbrook,
C. K., 2002, Sensitization of Hydrocarbon-Oxygen Mixtures to Detonation via
Cool-Flame Oxidation, Proceedings of the Combustion Institute, Vol. 29, pp.
2833-2838.
Sarno, R. L. and Franke, M. E., 1994, Suppression of Flow-Induced Pressure
Oscillations in Cavities, Journal of Aircraft, Vol. 31, No. 1, pp. 90-96.
Schauer, F., Stutrud, J. and Bradley, R., 2001, Detonation Initiation Studies and
Performance Results for Pulse Detonation Engine Applications, 39th AIAA Aerospace
Sciences Meeting and Exhibi, AIAA Paper 2001-1129.
Schwendeman, D. W. and Whitham, G. B., 1987, On Converging Shock Waves,
Proceedings of the Royal Society of London. Series A, Mathematical and Physical
Sciences, Vol. 413, No. 1845, pp. 297-311.
Sharma, V. D. and Radha, C. H., 1995, Similarity Solutions for Converging Shocks in a
Relaxing Gas, International Journal of Engineering Science, Vol. 33, No. 4, pp.
535-553.
Shen, P. I.-w. and Adamson, T. C., Jr., 1972, Theoretical Analysis of a Rotating
Two-Phase Detonation in Liquid Rocket Motors, Astronautica Acta, Vol. 17, pp.
715-728.
Shinn, J. L., Yee, H. C. and Uenishi, K., 1987, Extension of a Semi-Implicit Shock
Capturing Scheme for Three-Dimensional Fully Coupled Chemically Reactive Flows in
Generalized Coordinate Systems, AIAA Paper 87-1577.
Shuen, S. and Yoon, S., 1989, Numerical Study of Chemically Reacting Flows Using a
Lower-Upper Symmetric Successive Overrelaxation Scheme, AIAA Journal, Vol. 27,
No. 12, pp. 1752-1760.
Sod, G. A., 1977, A Numerical Study of a Converging Cylindrical Shock, Journal of
Fluid Mechanics, Vol. 83, No. 4, pp. 785-794.
Stanyukovich, K. P., 1960, Unsteady Motion of Continuous Media, Pergamon Press,
Oxford.

226
Sterling, J., Ghorbanian, K. and Sobota, T., 1996, Enhanced Combustion Pulsejet
Engines for Mach 0 to 3 Applications, AIAA Paper 96-2687.
Strehlow, R. A., 1970, Multi-Dimensional Detonation Wave Structure, Astronautica
Acta, Vol. 15, pp. 345-357.
Takayama, K., Onodera, O. and Hoshizawa, Y., 1984, Experiments on the Stability of
Converging Cylindrical Shock Waves, Theoretical and Applied Mechanics, Vol. 32, pp.
117-127.
Takayama, K., Kleine, H. and Gronig, H., 1987, An Experimental Investigation of the
Stability of Converging Cylindrical Shock Waves in Air, Experiments in Fluids, Vol. 5,
pp. 315-322.
Taki, S. and Fujiwara, T., 1978, Numerical Analysis of Two-Dimensional Nonsteady
Detonations, AIAA Journal, Vol. 16, No. 1, pp. 73-77.
Taki, S. and Fujiwara, T., 1981, Numerical Simulation of Triple Shock Behavior of
Gaseous Detonation, Proceedings of the Combustion Institute, Vol. 18, pp. 1671-1681.
Tam, C. K. W. and Block, P. J. W., 1978, On the Tones and Pressure Oscillations
Induced by Flow over Rectangular Cavities, Journal of Fluid Mechanics, Vol. 89, No. 2,
pp. 373-399.
Taylor, G., 1950, The Dynamics of the Combustion Products Behind Plane and
Spherical Detonation Fronts in Explosives, Proceedings of the Royal Society of London.
Series A, Mathematical and Physical Sciences, Vol. 200, No. 1061, pp. 235-247.
Taylor, G., 1950, The Formation of a Blast Wave by a Very Intense Explosion. I.
Theoretical Discussion, Proceedings of the Royal Society of London. Series A,
Mathematical and Physical Sciences, Vol. 201, No. 1065, pp. 159-174.
Taylor, G., 1950, The Formation of a Blast Wave by a Very Intense Explosion. II. The
Atomic Explosion of 1945, Proceedings of the Royal Society of London. Series A,
Mathematical and Physical Sciences, Vol. 201, No. 1065, pp. 175-186.
Thomas, G. O. and Williams, R. L., 2002, Detonation interaction with wedges and
bends, Shock Waves, Vol. 11, pp. 481-492.


227
Tishkoff, J. M., Drummond, J. P., Edwards, T. and Nejad, A. S., 1997, Future Direction
of Supersonic Combustion Research: Air Force/NASA Workshop on Supersonic
Combustion, 35th Aerospace Sciences Meeting and Exhibit, AIAA Paper 97-1017.
Toque, N., 2001, Self-Similar Implosion of a Continuous Stratified Medium, Shock
Waves, Vol. 11, No. 3, pp. 157-165.
Toro, E. F., 1999, Riemann Solvers and Numerical Methods for Fluid Dynamics, 2nd ed.,
Springer, London.
Tsuboi, N., Hayashi, A. K. and Matsumoto, Y., 2000, Three-Dimensional Parallel
Simulation of Cornstarch-Oxygen Two-Phase Detonation, Shock Waves, Vol. 10, pp.
277-285.
Tsuboi, N., Katoh, S. and Hayashi, A. K., 2002, Three-Dimensional Numerical
Simulation for Hydrogen/Air Detonation: Rectangular and Diagonal Structures,
Proceedings of the Combustion Institute, Vol. 29, pp. 2783-2788.
Van Dyke, M. and Guttmann, A. J., 1982, The Converging Shock Wave from a
Spherical or Cylindrical Piston, Journal of Fluid Mechanics, Vol. 120, pp. 451-462.
Van Leer, B., 1969, Stabilization of Difference Schemes for the Equations of Inviscid
Incompressible Flow by Artificial Diffusion, Journal of Computational Physics, Vol. 3,
No. pp. 473-485.
Varsi, G., Back, L. H. and Kim, K., 1976, Blast Wave in a Nozzle for Propulsive
Application, Acta Astronautica, Vol. 3, No. Pergamon Press, pp. 141-156.
von Neumann, J., 1942, Theory of Detonation Waves, John von Neumann: Collected
Works, edited by Taub, A. H., Macmillan, New York.
Wang, B. and Yu, S.-T. J., 2004, Direct Calculation of Spherical Detonation Initiation of
H2/O2/Ar Mixtures by the CESE Method, 42nd AIAA Aerospace Sciences Meeting and
Exhibit, AIAA Paper 2004-0793.
Wang, B., He, H. and Yu, S.-T. J., 2005, Direct Calculation of Wave Implosion for
Detonation Initiation, AIAA Journal, Vol. 43, No. 10, pp. 2157-2169.


228
Wang, X. Y. and Chang, S. C., 1999, A 2D Non-Splitting Unstructured Triangular Mesh
Euler Solver Based on the Space-Time Conservation Element and Solution Element
Method, Computational Fluid Dynamics Journal, Vol. 8, pp. 309-325.
Welch, G. E., Jones, S. M. and Paxson, D. E., 1995, Wave Rotor-Enhanced Gas Turbine
Engines, 31st AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA
Paper 95-2799.
Welch, G. E., 1996, Two-Dimensional Computational Model for Wave Rotor Flow
Dynamics, NASA Technical Memorandum, NASA TM-107192.
Westbrook, C. K. and Dryer, F. L., 1981, Simplified Reaction Mechanisms for the
Oxidation of Hydrocarbon Fuels in Flames, Combustion Science and Technology, Vol.
27, pp. 31-43.
White, F. M., 1974, Viscous Fluid Flow, 1st ed., McGraw-Hill Book Co., New York,
pp. 195-202.
Williams, D. N., Bauwens, L. and Oran, E. S., 1996, Detailed Structure and Propagation
of Three-Dimensional Detonations, Proceedings of the Combustion Institute, Vol. 26, pp.
2991-2998.
Wilson, J., 1996, An Experimental Determination of Losses in a 3-Port Wave Rotor,
NASA Contractor Report, NASA CR-198456.
Wilson, J. and Paxson, D. E., 2002, On the Exit Boundary Condition for
One-dimensional Calculations of Pulsed Detonation Engine Performance, NASA
Technical Memorandum, NASA TM-2002-211299.
Wortman, A., Othmer, P. and Rostafinski, W., 1992, Detonation Duct Gas Generator
Demonstration Program, AIAA Paper 92-3147.
Yee, H. C., 1987, Upwind and Symmetric Shock-Capturing Schemes, NASA Technical
Memorandum, NASA TM-89464.
Yu, K. H., Wilson, K. J. and Schadow, K. C., 2001, Effect of Flame-Holding Cavities
on Supersonic-Combustion Performance, Journal of Propulsion and Power, Vol. 17, No.
6, pp. 1287-1295.


229
Yu, S.-T. and Chang, S.-C., 1997, Treatments of Stiff Source Terms in Conservation
Laws by the Method of Space-Time Conservation Element and Solution Element, 35th
Aerospace Sciences Meeting and Exhibit, AIAA Paper 97-0435.
Yu, S.-T. J. and Chang, S.-C., 1997, Applications of the Space-Time Conservation
Element and Solution Element Method to Unsteady Chemically Reacting Flows, AIAA
Paper 97-2099.
Zhang, M., Yu, S.-T. J., Lin, S.-C. H., Chang, S.-C. and Blankson, I., 2004, Solving
Magnetohydrodynamic Equations Without Special Treatment for Divergence-Free
Magnetic Field, AIAA Journal, Vol. 42, No. 12, pp. 2605-2607.
Zhang, X. and Edwards, J. A., 1992, Experimental Investigation of Supersonic Flow
over Two Cavities in Tandem, AIAA Journal, Vol. 30, No. 5, pp. 1182-1190.
Zhang, Z. C. and Yu, S. T., 1999, Shock Capturing without Riemann Solver---A
Modified Space-Time CESE Method for Conservation Laws, AIAA Paper 99-0904.
Zhang, Z. C., Yu, S. T., Chang, S. C., Himansu, A. and Jorgenson, P., 1999, A Modified
Space-Time CESE Method for Euler and Navier-Stokes Equations, AIAA Paper
99-3277.
Zhang, Z. C., Yu, S. T. J., Wang, X. Y., Chang, S. C., Himansu, A. and Jorgenson, P. C.
E., 2000, The CESE Method for Navier-Stokes Equations Using Unstructured Meshes
for Flows at All Speeds, AIAA Paper 2000-0393.
Zhang, Z. C., Yu, S. T. J., He, H. and Chang, S. C., 2001, Direct Calculations of Two-
and Three-Dimensional Detonations by an Extended CE/SE Method, 39th AIAA
Aerospace Sciences Meeting and Exhibit, AIAA Paper 2001-0476.
Zhang, Z.-C., Yu, S. T. J., He, H. and Jorgenson, P. C. E., 2001, Direct Calculations of
Plume Dynamics of a Pulse Detonation Engine by the Space-Time CE/SE Method, 37th
AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, AIAA Paper
2001-3614.
Zhang, Z. C., Yu, S. T. J. and Chang, S. C., 2002, A Space-Time Conservation Element
and Solution Element Method for Solving the Two- and Three-Dimensional Unsteady
Euler Equations Using Quadrilateral and Hexahedral Meshes, Journal of Computational
Physics, Vol. 175, pp. 168-199.

230
Zeldovich, Ya. B., 1940, On the Theory of the Propagation of Detonation in Gaseous
Systems, Zhurnal Eksperimentalnoi i Teoreticheskoi Fiziki, Vol. 10, pp. 542-568.
Zukoski, E. E. and Spaid, F. W., 1964, Secondary Injection of Gases into a Supersonic
Flow, AIAA Journal, Vol. 2, No. 10, pp. 1689-1696.