Vous êtes sur la page 1sur 9

Partially-Averaged Navier-Stokes

Model for Turbulence: A


Reynolds-Averaged
Navier-Stokes to Direct Numerical
Sharath S. Girimaji
Department of Aerospace Engineering,
Simulation Bridging Method
Texas A&M University,
College Station, TX 77843 A turbulence bridging method purported for any filter-width or scale resolutionfully
averaged to completely resolvedis developed. The method is given the name partially
averaged Navier-Stokes (PANS) method. In PANS, the model filter width (extent of partial
averaging) is controlled through two parameters: the unresolved-to-total ratios of kinetic
energy f k and dissipation f . The PANS closure model is derived formally from the
Reynolds-averaged Navier-Stokes (RANS) model equations by addressing the following
question: if RANS represents the closure for fully averaged statistics, what is the corre-
sponding closure for partially averaged statistics? The PANS equations vary smoothly
from RANS equations to Navier-Stokes (direct numerical simulation) equations, depend-
ing on the values of the filter-width control parameters. Preliminary results are very
encouraging. DOI: 10.1115/1.2151207

1 Introduction vestigation of fundamental premise or model validity. It is there-


fore important, and timely, to develop bridging models from first
Speziale 1 was among the first to propose a new turbulence
principles.
model paradigm that combines the advantages of Reynolds-
averaged Navier-Stokes RANS method with those of large-eddy We develop a bridging method inspired by the modeling para-
simulations LES. Speziale suggested that such a variable-filter digm of Speziale 1. The method is given the name partially
model must have the following qualities: it must function as i a averaged Navier-Stokes PANS model and is purported for any
reasonably sophisticated RANS model when the intended cutoff filter widthfrom RANS to direct numerical simulations DNS.
wave number is in the largest scales of motion, ii a LES subgrid We start with the premise that RANS closure provides a model for
closure if the desired cutoff is in the inertial scales, and iii DNS the physical effect of averaging over all scales of fluctuations.
when all scales of motion are resolved. The development of such Therefore, we propose the model for the physical effect of partial
a model was at first a mere academic exercise, but soon its im- averaging must be incumbent in the RANS closure. The objective
mense practical utility became apparent. In a variety of engineer- of our work is to extract, from a given RANS, the latent PANS
ing flows involving large-scale unsteadiness, it became clear that model for partial averaging. The PANS development involves two
RANS approach was physically inadequate and LES was compu- important steps. The first step is to identify the physically appro-
tationally too expensive. The need emerged for a turbulence priate filter-width control parameters. We reason that the degree of
model that resolved only the largest dynamically important fluc- filtering can be best quantified by ratios of unresolved-to-total
tuating scales without the burden of having to resolve inertial kinetic energy and dissipation. For a PANS simulation, these
scales. Speziales pursuit of a turbulence closure method that filter-control parameters must be specified. Given the control pa-
could function effectively at all levels of scale resolution or filter rameters, the second step is to derive the PANS equations from
width quickly attained engineering relevance. Presently, this ap- the parent RANS closure by asking the question: if the RANS
proach is called a bridging or hybrid method. Currently, several equations represent the closure of the fully averaged turbulence
bridging strategies are under consideration: e.g., detached-eddy statistics, what is the implied closure for partially averaged statis-
simulations DES 2; hybrid RANS/LES; limited numerical tics? In PANS more scales of motion will be resolved than in the
scales method LNS 3; unsteady Reynolds-averaged Navier parent RANS, depending on the specified values of the filter-
Stokes method URANS 4; and very large-eddy simulations control parameters. If these ratios are set to unity, RANS equa-
VLES 1. The objective of these bridging models is to extract tions are recovered. If, on the other hand, the ratios are set to zero,
important large-scale unsteady features of a turbulent flow at the PANS equations reduce to Navier-Stokes as required of a
minimal computational expense. The mixed methods resolve more bridging model. Clearly, the PANS model will inherit much of its
scales of motion than RANS and significantly lesser scales than a physics from the parent RANS closure. The better the parent
typical LES. Compelled by immediate engineering needs, the RANS, the superior will be the corresponding PANS for partially
mixed methods are being increasingly used without adequate in- averaged fields. However, for any given RANS model, the PANS
computation can be expected to yield improved performance at
higher computational expense as more scales of motion more
Contributed by the Applied Mechanics Division of ASME for publication in the
JOURNAL OF APPLIED MECHANICS. Manuscript received February 5, 2004; final manu- scales are resolved exactly. Apart from the model development,
script received November 8, 2005. Review conducted by D. Siginer. Discussion on we will also present some preliminary results for validation.
the paper should be addressed to the Editor, Prof. Robert M. McMeeking, Journal of The PANS method is distinct from standard LES in three im-
Applied Mechanics, Department of Mechanical and Environmental Engineering,
University of California-Santa Barbara, Santa Barbara, CA 93106-5070, and will be
portant aspects: i the decomposition of the velocity field is based
accepted until four months after final publication of the paper itself in the ASME on kinetic energy content rather than on cutoff wave number; ii
JOURNAL OF APPLIED MECHANICS. the filter demarcating calculated resolved and modeled unre-

Journal of Applied Mechanics Copyright 2006 by ASME MAY 2006, Vol. 73 / 413

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


solved motion is implied rather than explicit, requiring no filter-
ing operation during computation; and iii the SFS subfilter
scale constitutive relationship is independent of grid spacing, i.e.,
Tij =

xk
Vi,V j,Vk + p,Vi jk + p,V jik
Vi,V j
xk

the physical resolution filter width is decoupled from numerical 5
resolution. Clearly, the numerical resolution must be in commen- The subfilter kinetic energy and dissipation associated with any
surate with the desired filter width. Although distinctly different operator angular brackets can be defined as
from LES, the PANS method shares some similarities with the
currently popular URANS unsteady RANS method 5,6. The
similarities and differences between URANS and PANS is dis-
cussed in detail in 7.
1
Ku = Vi,Vi;
2
u = Vi Vi
,
x j x j
6

The outline of the paper is as follows. After the introduction Throughout, subscript u indicate PANS unresolved statistics.
Sec. 1, we present the general hybrid modeling issues in Sec. 2. When the averaging is over all scales of motion denoted by
In Sec. 3, the premise of the PANS model is first formulated. This over bar, the filtered velocity becomes the mean velocity and the
is followed by the development of the PANS closure equations. SFS stress reduces to Reynolds stress-RVi , V j: where RA , B
Preliminary results are presented in Sec. 4 to demonstrate the
= AB AB. The RANS statistics are related to their PANS counter-
potential capability of PANS. We close with a brief discussion in
part according to 8
Sec. 5.
V j = V j = U j ; RVi,V j ViV j ViV j = Vi,V j + RUi,U j
7
2 Modeling Fundamentals
The RANS kinetic energy and dissipation are denoted by K and .
We start with the instantaneous incompressible flow equations:
Averaging over all scales leads Eqs. 2 and 5 to collapse to the
RANS equations. Thus, the form of the statistical hierarchy of
Vi Vi p 2V i 2 p Vi V j
+ Vj = + ; = 1 Eqs. (2)(5) is invariant of the filter, when the filtered equations
t x j xi x jx j x i x i x j xi are expressed in terms of the generalized central moments 8.
Consider partially averaged flow variables-Ui = Vi; This feature of the Navier-Stokes equation is given the name
averaging-invariance property. Much like tensor invariance and
pU = p-where the angular brackets denote an arbitrary implicit Galilean invariance requirements on RANS models, the averaging
or explicit filter that is constant preserving and commutes with invariance imposes a crucial constraint on bridging models.
spatial and temporal differentiation. Partial averaging corresponds Based on the objective of bridging models, we propose the
to filtering a portion of the fluctuating scales. Throughout the pa- following constraints on the PANS model: i The bridging model
per the words filtering and averaging will be used synonymously. must vary smoothly and seamlessly from RANS to DNS as the
The partially averaged Navier-Stokes equation can be written as filter-control parameters are varied the Speziale 1 guideline.
8 The URANS approach has not been able to demonstrate a smooth
variation 6. ii The filter-width control parameters that quantify
Ui Ui Vi,V j pU 2U i the extent of averaging in a bridging model must be identified. For
+ Uj + = + ;
t x j x j xi x jx j example, in URANS, the extent of implied filtering is not clear.
iii The form of the bridging model closures must be invariant of
2 pU Ui U j 2Vi,V j the filter when expressed in terms of the appropriate generalized
= + 2 central moments. This constraint rules out many of the hybrid
x i x i x j x i x i x j methods that switch from RANS to LES, depending on some local
In Eq. 2, Vi , V j is the generalized central second moment. length scale.
Generalized central second and third moments of random vari-
ables A, B, and C are defined as 8
3 PANS Model Development
A,B = AB AB
The goal now is to develop a SFS stress Vi , V j closure
model that accurately reflects the extent of partial averaging and is
A,B,C = ABC AB,C BC,A consistent with the guidelines formulated in Sec. 2. The
CA,B ABC 3 averaging-invariance principle permits the use of LES- or RANS-
type closure modeling. Most of the popular LES closures are al-
The subfilter scale SFS stress Vi , V j is a positive semi-definite gebraic in nature zero-equation models and are best suited for
tensor exhibiting much of the same properties as Reynolds stress. cutoff in the small inertial scales of motion. On the other hand,
Its evolution equation can be written without any further assump- RANS models are purported for averaging over all scales of mo-
tions as 8 tion and, hence, carry sufficient physics to accurately represent
partially averaged fields. Furthermore, years of modeling experi-
Vi,V j Vi,V j ence is incumbent in the RANS models and they have been tested
+ Uk = Pij + ij Dij + Tij 4
t xk widely in important engineering flows. Important issues, such as
tensor invariance, realizability, and effects of extra strain rate due
In Eq. 4, the various terms on the right-hand side rhs can be to rotation, buoyancy, etc., are best addressed at the level of
classified as production Pij, pressure-correlation ij, dissipa- RANS closure using well-tested techniques 9. As the PANS
tion Dij, and transport Tij of SFS stress: bridging model is purported for use at all degrees of physical
resolution including RANS, we suggest that there be at least as
U j Ui much physics in the PANS models as in some of the more ad-
Pij = Vi,Vk V j,Vk ; ij = 2p,Sij; vanced two-equation RANS models. Therefore, we base the
xk xk
PANS models on the RANS paradigm. In this paper, we will


develop the theoretical framework for adapting RANS models to
Ui U j the PANS method. Although the development here is at the level
Dij = 2 ,
xk xk of the two-equation closure, the methodology is equally valid for

414 / Vol. 73, MAY 2006 Transactions of the ASME

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


the adaptation of more sophisticated RANS models, including eling of the unresolved scales.
second-moment closures SMC and algebraic Reynolds stress The ratio of the PANS to RANS subscript t characteristic
models ARSM. scales are
The RANS two-equation model used in this development is
Uu Lu f 1.5
k Tu f k
2 = fk; = ; = ;
K Ut Lt f Tt f
t = C Sij

uu lu tu


= f 0.25
; = f 0.25
; = f 0.50
11
K K t K ut lt tt
+ U j =P+
t x j x j k x j In the above,


t
+ U j

x j
= Ce1
P
K
2
Ce2 +
t

K x j x j
8
fk =
Ku
K
, f =
u

12

are the ratios of the unresolved-to-total kinetic energy and


where P =RVi , V jVi / x j and are the RANS production and unresolved-to-total dissipation. It is, therefore, natural to quantify
dissipation. Recall that Vi Ui. The values for the various coeffi- the extent of PANS filtering using f k and f . These two parameters
cients used in our works are C = 0.9; Ce1 = 1.44; Ce2 = 1.92; k are the partial average-quantification and, hence, filter-control pa-
= 1.0; and = 1.3. rameters. Note that turbulence physicslarge scales contain most
At the two-equation closure level, the PANS development in- of the kinetic energy and much of the dissipation occurs in the
volves three important steps: i identification of the PANS filter- smallest scalesdictates 0 f k f 1. The smaller the f k, the
control parameters; ii postulation/development of the SFS stress finer is the filter: f k = 1 represents RANS and f k = 0 indicates DNS.
constitutive relationship; and iii derivation of the unresolved ki- Unity value of f implies that the RANS and PANS unresolved
netic energy and dissipation evolution equations. The constitutive small scales are identical. Smaller values of f would require the
relation determines the anisotropy of the SFS stress, and the evo- resolution of dissipative scales of motion.
lution equations set the energy level of the unresolved field. Our It is generally desirable to have the filter width vary as a func-
consideration will be restricted to the Boussinessq constitutive tion of space as in DES. Rapid spatial variation of a filter in
relationship. Nonlinear constitutive relations and algebraic Rey- physical space will be considered in future works as it introduces
nolds stress closures can be used instead of the Boussinesq ap- an additional complication of a filtering operation that does not
proximation without affecting the remainder of the development. commute with spatial differentiation. In this paper, we will as-
sume that the spatio-temporal variation of the implied PANS filter
3.1 Identification of Filter Control Parameter. In LES, the is slow compared to that of flow variables. Averaging invariance
filter width is closely linked to the numerical grid size. The larger can be a useful constraint even when the filters are slowly varying
the cutoff wave number, the more accurate the calculation as more functions of space.
scales are resolved exactly. The cutoff wave number enters the
calculations via explicit filtering and subgrid-stress constitutive 3.2 Constitutive Relationship. When the Boussinesq consti-
relationship and thus controls the physical resolution. For PANS it tutive approximation is invoked in conjunction with averaging-
is important to identify the corresponding physical resolution pa- invariance requirement for arbitrary filters, we get
rameter that i quantifies the filtering and ii appears in the clo-
K2u
sure model to ensure that the required filtering is achieved. Vi,V j = uSij ; where u = Cu 13
In many of the current LES closures, such as Smagorisnsky u
model, it is tacitly assumed that the unresolved scales are in equi- The objective of PANS dynamics is to reduce the effective vis-
librium with the resolved scales. Furthermore, it is assumed that cosity from its RANS value according to the desired filter width,
the production of the unresolved energy is equal to dissipation. leading to the liberation of more scales of motion. It would appear
These assumptions permit the estimation of the subgrid stress in from Eq. 13 that such a reduction can be accomplished by i
terms of the cutoff wave number and resolved strain. Thus, the simply assigning Cu a smaller value than RANS-C and/or ii
cutoff wave number can serve to characterize the unresolved manipulating the levels of Ku and u in such a manner that the
scales and quantify the resolution. However, these LES closure
K2u / u is smaller than its RANS counterpart. Fixed-point analysis
assumptions are completely invalid if the cutoff is in energy-
of the PANS/RANS equations can provide a cause-effect relation-
containing production scales or large inertial scales. For such a
ship between these parameters and the resulting stresses. Such an
cutoff, we need a more detailed characterization of the unresolved
analysis of the RANS equations has been performed in 9. Simi-
scales. Based on our experience in RANS, we deem that adequate
lar analysis of the PANS equations are performed in 7, which
characterization of the PANS scale requires, at a minimum, the
will be discussed further in the Results section. The conclusions
knowledge of the kinetic energy and dissipation of the unresolved
from the two studies are:
scales. Knowing these quantities, the characteristic length, veloc-
ity, and time scales Lu , Uu , Tu of the largest unresolved-scales 1. The asymptotic unresolved kinetic energy growth rate is
can be estimated governed only by the values of C1 and C2. The value of C


does not affect the normalized kinetic energy growth rate.
3 K1.5
u Ku Furthermore, C1 and C2 also affect the anisotropy of the
Uu = Ku ; Lu = ; Tu = 9 stresses.
2 u u
2. The value of Cu affect only the anisotropy, but not the
The unresolved dissipative motion can be characterized using un- energetics. This may appear counterintuitive, but the follow-
resolved dissipation and viscosity ing sequence of events leads to that behavior. A reduced Cu


value not only reduces production of turbulent kinetic en-
0.75 ergy but also leads to reduced production of dissipation. Re-
uu = u0.25 ; lu = ; tu = 10 duced production of dissipation results in lower values of
0.25
u
dissipation. Lower dissipation, in turn, leads to higher-than-
We suggest that the extra knowledge of these scales not known in desired levels of unresolved kinetic energy. Thus, the desired
LES is crucial for accurate characterization and subsequent mod- effect of partial averaging, which is to lower the levels

Journal of Applied Mechanics MAY 2006, Vol. 73 / 415

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


PANS kinetic energy and effective viscosity cannot be ac-
complished by merely lowering Cu value. Tku =
u Ku
x j ku x j
21

Based on the findings of the fixed-point analysis of the model where ku is the corresponding Prandtl number, which needs to be
equations, we will not modify the value of Cu determined.
C u = C 14 To complete the modeling of ku, the transport of Ku due to the
resolved velocity fluctuations U j U j must be known. For this
This decision is further justified in our related asymptotic analysis transport term, two of the simplest closures that are consistent
work 7, which is discussed later in the paper. In the future, we with the averaging-invariance requirement are suggested. The first
will consider the modification of Cu as a function of partial av- is based on the assumption that the resolved fluctuations do not
eraging. This can introduce other filter control parameters. contribute to SFS energy transport ZT, zero-transport model
3.3 PANS Closure Equations. We now develop the evolution
equations for Ku and u as functions of f k and f . The logic fol- Ku
U j U j =0 22
lows from the simple question: if Eq. 8 represents the kinetic x j
energy and dissipation of the fully averaged fluctuating field, then
what are the equations for specified fractions of the energy and This simplification leads to from 20 and 21


dissipation.
Taking the trace of 5 leads to the following averaging- u f Ku u Ku
Tku = = 23
invariant form of the unresolved kinetic-energy evolution equation x j k f 2k x j x j ku x j
8
Thus, the zero-transport models lead to
Ku Ku Ui
+ Uj = Pu u + Tku where Pu = Vi,V j 15
t x j x j f 2k
ku k ZT assumption 24
The transport term Tku needs closure modeling. Although the f
form of the unresolved kinetic energy equation is dictated by the The second proposal is, in some sense, the maximum transport
averaging-invariance constraint, the content of this equation MT model in which the transport is proportional to the eddy
comes from the requirement that the ratio of unresolved-to-total viscosity of the resolved fluctuations, r = t u
kinetic energies must be f k. Since f k is a constant, the evolution of
RANS and PANS kinetic energies are related according to
Ku
+ Uj
Ku
= fk
K
+ UjK
16
U j U j
Ku
=
r Ku

x j x j k x j
25

t x j t x j
From Eqs. 20 and 21, this leads to the following closure for the
This Ku evolution is following the mean RANS velocity field. unresolved-kinetic energy Prandtl number
The evolution of Ku following PANS filtered velocity field can be
written as ku k MT assumption 26
Ku
t
+ Uj
Ku
x j
= fk
Ku
t
+ Uj
Ku
x j

+ U j U j
Ku
x j
The two Prandtl-number assumptions are assessed in ongoing
work and will be discussed further in the Results section.


We now develop PANS-unresolved dissipation u equation
t K Ku
= fk P + + U j U j from the requirement that PANS-RANS dissipation ratio is f .
x j k x j x j This requirement yields


17
u u
where the substitution for the term in the square brackets has been + U j =0 27
made from Eq. 8. To achieve the required f k, Eq. (17) must be t x j
consistent with (15). From these equations, we can write leading to

Pu u + Tku = f k P+
t K

x j k x j
+ U j U j
Ku
x j
u
+ U j
u
x j
= f

+ U j

x j

t t


18
P 2 t
The relationship between various RANS terms and the corre- = f Ce1 Ce2 + 28
sponding PANS terms are now evident. Equating the source/sink K K x j x j
terms local processes and noting = u / f , we get
As RANS variables are not known, to close the u equation all
1 u terms must be expressed in PANS variables U j , Ku , u , Pu and
Pu u = f kP , implying P = Pu u + 19 PANS parameters f k , f . We then have
fk f
This relationship will be used later to close the PANS dissipation
equation. Comparison of the advection/transport terms nonlocal
processes in Eq. 18 leads to noting Ku = f kK
u
t
+ Uj
u
x j
= Ce1 f k
Pu
Ku
Ce2
f k 2u
+
t u
f Ku x j x j

+ U j U j
u
x j

Tku
t Ku

x j k x j

U j U j
Ku
x j The RANS production P is obtained from Eq. 19. Again, the u
29

=
u f Ku
x j k f k x j
2
U j U j
Ku
x j
20
transport due to the resolved scalesthe U j U j termis the
only term that requires modeling. For the resolved-velocity trans-
port term, we again propose two closure options: ZT model as in
The averaging-invariance principle permits a transport closure Eq. 22 and MT model as in Eq. 25. Finally, the fully closed
model of the following form: equation can be written as

416 / Vol. 73, MAY 2006 Transactions of the ASME

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


u
t
+ Uj
u
x j
= Ce1 f kPu u

fk ffk
f f k
u
Ku
Ce2
f k 2u
f Ku Settings
Table 1 PANS simulation setting in

Choice
FLUENT

+
t u
x j x j
+ U j U j
u
x j
Simulation type
Solver
3D, Unsteady
Segregated, implicit


Temporal discretization Second order
P u u 2u u u Turbulence model k- 2eqn. with modified PANS parameters
= Ce1 C*e2 + 30 Pressure PRESTO
Ku K u x j u x j Pressure-velocity coupling SIMPLE
Momentum Second-order upwind
where Turbulent kinetic energy Second-order upwind
Turbulent dissipation rate Second-order upwind
Boundary Conditions
fk f 2k Inlet Velocity
C*e2 Ce1 + Ce2 Ce1; u for ZT; Outlet Outflow
f f Top wall No-slip wall
Bottom wall No-slip wall
Lateral Periodicity
u = for MT 31

The two-equation PANS model is given by Eqs. 15 and 30.


The form of the equations are clearly invariant to averaging. The 3.3.3 Specification of Filter-Control Parameter. In an LES
filter-control parameter manifests only via the modified model co- computation, the cutoff wave number must be specified in order to
efficients. In the PANS equations derivation, the form of the con- perform a simulation. Generally, the cutoff wave number is taken
stitutive relationship is never invoked. This derivation is com- to be inversely proportional to the grid size. Hence, in LES the
pletely valid for any linear or nonlinear including explicit physical resolution and the numerical resolution are closely
algebraic Reynolds stress model constitutive relationship. coupled. Such a coupling is possible only when the LES closure
In the absence of transport and in the limit of f approaching assumptions discussed earlier are valid. In PANS, as in RANS, the
unity, this result is similar to the LES dissipation transport model turbulence closure model is decoupled from the numerical resolu-
of 10. That model 10 was derived for homogeneous turbulence tion. In PANS the filter control parameters must be specified for a
in spectral space using splitting technique. In that method 10, given simulation. The prescription of f k and f would depend on
there is a distinct demarcation between the resolved and unre- the desired physical resolution and affordability of numerical
solved field and the split must be in the inertial range. The current resolution. Unless the dissipation scales are to be resolved, f can
derivation does not invoke such spectral splitting and by introduc- be set to unity. For high Reynolds number simulations, when the
ing a second control parameter f we extend the applicability into intended cutoff is in the inertial range, f k is the only resolution
dissipation range, resulting in a complete bridging model. Also, parameter of relevance. Clearly, the smaller the f k, the larger is the
we derive the model in more realistic inhomogeneous turbulent grid resolution required. The computational grid must be incom-
flows, leading to modifications in turbulent Prandtl numbers. mensurate with the chosen filter width. To some extent, the length
Thus, the present derivation can be considered a significant gen- scale in Eq. 9 can provide guidance for grid spacing. In general,
eralization of the important result of 10. for given f k and f , the grid should be chosen from a grid
3.3.1 Reynolds Number Sensitivity. A unique feature of PANS resolution/independence study as is routinely done in RANS
subgrid closure is its sensitivity to Reynolds number. Most sub- calculations.
grid closure models e.g., all versions of the Smagorinsky model
are not directly sensitive to the Reynolds number of the unre- 4 Results and Discussion
solved scales. For example, the closures cannot distinguish be- Any new turbulence model, such as the PANS, must be subject
tween two flows with similar resolved fields but vastly different to a wide range of tests. Our objective is to compare the parent
unresolved fields. Narrow low Reynolds number or wide high RANS results to PANS results of various filter widths. The goal is
Reynolds number unresolved energy spectra elicit the same clo- to establish, with decreasing f k, that PANS i leads to liberation
sure from these models as long as the resolved strain rates are of more and more scales of motion and ii yields mean flow and
identical. On the contrary, in PANS closure, the extent of the turbulence correlation statistics that are more accurate than the
unresolved spectrum or, equivalently, the separation between the parent RANS. Clearly, the degree of accuracy will depend on the
energy containing and dissipative scales can be imposed on the quality of the parent RANS model. Here, we will employ the
model by suitably specifying f . For example, in high Reynolds standard two-equation K- RANS model, which is known to be
number flows in which the dissipative scales are not resolved f is inadequate in complex flows. We will then demonstrate that the
specified as unity. At the other extreme, the physics of low Rey- predictive capability of the corresponding PANS improves
nolds number flows, in which there is no discernible separation of substantially with decreasing f k, thereby validating the bridging
scales, can be accurately captured by f k f . The physics of mod- methodology.
erate Reynolds numbers can be capture by suitably setting f in In general, the validity of a model must be judged based on its
the range 1 , f k. collective performance over a series of benchmark problems.
Here, we present both a priori asymptotic analysis and prelimi-
3.3.2 Near-Wall Treatment. The near-wall modeling is an im- nary a posteriori flow-computation examinations of the PANS
portant issue in LES and any non-Reynolds averaging approach. A method. While a flow computation can be used to evaluate a
satisfactory and efficient affordable near-wall resolution wall model in that single flow, fixed-point/asymptotic analysis provide
treatment has proved elusive even in such a mature methodology valuable insight into the inherent dynamic character of a model
as LES. A detailed investigation of the PANS unsteady near-wall equation. Thus, asymptotic analysis is equally important for as-
treatment will be undertaken in the future. Here, we propose a sessing the suitability of a model for the purported applications.
simple approximation. In the near-wall region, there is no separa-
tion of scales, as the Reynolds number is low, leading to f k f 4.1 Asymptotic Analysis. In the RANS method, analytical
and MT model for the Prandtl numbers. This collapses the PANS fixed-point analysis has been very useful for calibrating and as-
equations almost back to RANS equations. Therefore, as a first sessing closure coefficients and model behavior. Since, PANS is
step, the wall-boundary treatment for PANS can be similar to that inherently unsteady, establishing the nature of its asymptotic time-
of RANS. dependent behavior when subject to time-dependent forcing is im-

Journal of Applied Mechanics MAY 2006, Vol. 73 / 417

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Fig. 2 Time-step sensitivity study. In all the computations fk
= 0.6 and grid size is 95 106

The neglect of transport reduces the model equations to coupled


ordinary differential equations that can be easily computed. In 7,
the homogeneous equations are forced with a periodic resolved
velocity gradient and the unresolved statistics are calculated by
numerical integration. The time-dependent shear forcing used is of
the form
St = Sm + S f sin2t 32
In Eq. 32, Sm corresponds to the mean shear; S f represents the
approximate amplitude of resolved fluctuating shear; and is the
frequency of fluctuation. Because of the periodic forcing, PANS
production, dissipation, and Reynolds stress anisotropies exhibit
time dependence. The time-averaged behavior as a function f k
for various amplitudes and frequencies of forcing are examined.
The time-averaged values of the various turbulence quantities are
defined as


t0+T
1
Q = Qtdt 33
T t0

where T is of the order of five eddy turnover times. The main


results from 7 are
Fig. 1 Grid-size sensitivity study in a 2D flow past a cylinder in 1. The temporal behavior of PANS kinetic energy, dissipation,
2D domain. a Centerline velocity profile from various PANS
fk = 0.6 computations on different grids. The time step in all and Reynolds stress depends strongly on the amplitude S f
computations is 0.008 s. b Similar plot for fk = 0.8 computa- and frequency of forcing.
tions. The time step in all these computations is 0.025 s. 2. The time-averaged values of the unresolved quantities are,
however, nearly constant after a short initial transition pe-
riod. The time-averaged value of production-to-dissipation
portant. Much like fixed-point analysis of RANS equations, this ratio which determines the energetics and shear anisotropy
analytical examination can reveal important intrinsic features of b12 progressively decrease in magnitude with smaller f ks,
the PANS equations. A detailed asymptotic investigation of PANS indicating that the desired filtering effect has been achieved.
is performed in 7, and we report the important results here. 3. The time-averaged values are nearly independent of S f and
The homogeneous PANS equations are Eqs. 15 and 30 with- .
out the transport terms. Transport only transfers energy from place 4. The time-averaged values are, however, strong functions of
to place in physical space and does not contribute to overall pro- f k. These values are nearly identical to the constant forcing
duction or dissipation of energy. Hence, the neglect of transport case when S f = 0.
terms does not alter the overall energetics of the turbulent field. 5. Simple fixed-point analysis of the constant-forcing PANS

Table 2 Flow past a square cylinder

Case Grid Re/ 103 St CDmean CDrms CLrms

URANS 95 107 21 22 0.132 1.44 0.0022 0.27


PANS f k = 0.7 95 107 21 22 0.133 1.77 0.129 0.994
PANS f k = 0.4 95 107 21 22 0.130 1.97 0.216 1.19
LES 185 105 25 22 0.1260.132 2.032.32 0.160.20 1.231.54
Exp 21.4 0.132 2.1

418 / Vol. 73, MAY 2006 Transactions of the ASME

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


shows that, asymptotically, Pu / u varies linearly with
f kgoing from the RANS value to DNS value of unity:


Pu
u lim t
=
C*e2 1
Ce1 1
=1+
f k Ce2 Ce1
f Ce1 1
34

Thus, the PANS SFS velocity field progressively gets weaker and
isotropic with smaller f k as required. On the contrary in URANS
wherein the model coefficients are constants at RANS values,
the unresolved scales will be as energetic as the RANS scales
irrespective of implied resolution as was found by 7. This ex-
plains the high energy levels of the unresolved field found in the
computation of 5.
4.2 Flow Simulations. Currently, several large-scale PANS
computations of flow past square cylinder, high and low-Reynolds
number flow past circular cylinder, flow in a square cavity, and
flow over a surface-mounted cube are underway. In all these stud-
ies, the commercial CFD code FLUENT is used with the model
coefficients appropriately modified, depending on the specified
filter-control parameter. The choice of FLUENT as the computa-
tional platform is motivated by the reason that it is the most
widely used commercial code for engineering applications. It is
important that the capabilities of a new engineering model, such
as the PANS, be demonstrated on FLUENT rather than on a special-
purpose code. The various solvers used in the FLUENT simulations
are detailed in Table 1. The standard wall-function approach is
used for the first near-wall grid point.
In these studies, detailed PANS results mean flow and Rey-
nolds stress profiles are being compared to those of other meth-
ods, such as URANS, DES, and LES. The main conclusions are:
i The ZT model is most appropriate for the turbulent transport in
high Reynolds number flows. The MT model is better for low
Reynolds number flows. ii When the Reynolds number is
smallno scale separationf k and f must be approximately
equal for best agreement with data. iii PANS is an effective
bridging model: the computed results go from RANS to LES/
DNS/experimental data as f k is reduced from unity to zero. All
these studies will be published in detail elsewhere. In support of
the PANS theory developed in this paper, we present selected
preliminary results from simulations of i flow past a circular
cylinder, ii flow past a square cylinder, and iii flow in a three-
dimensional 3D cavity. The results clearly demonstrate PANS
ability to function as a bridging model. Before the results are
presented, we first perform a grid-dependence study.
4.2.1 Grid and Time-Step Dependence Study. The PANS
model closure for various filter widths is developed without any
reference to numerical issues. However, in any computation, the
numerical grid size must be in accordance with the model filter
width. Much like the Navier-Stokes and RANS equations, the
PANS closure model of a given filter width is amenable to a
formal grid convergence study. In our implementation, we use the
wall-function boundary condition at solid boundaries. As the wall
function is dependent on the distance from the wall, our PANS
simulations cannot be totally grid independent. However, we will
demonstrate that the calculations are only very weakly dependent
on grid size at reasonably fine resolution. For this demonstration
we employ two-dimensional 2D flow past a square cylinder.
Although two-dimensional simulations cannot capture turbulence
physics, they can serve to establish the grid convergence proper-
ties of the model equations at reasonable cost.
Two sets of PANS simulations of f k values of 0.8 and 0.6 are
performed. The f value for both sets of simulations is unity. For
each f k, computations are performed on several grids. The coars-
est grid used is 85 85, and the finest is 400 400. Various time-
Fig. 3 Profiles of normal Reynolds stresses along the center- steps ranging from 0.025 to 0.00625 are used. The flow Reynolds
line from various calculations: fk = 1.0, ---; fk = 0.7 ; fk = 0.4, number in these computations is 140,000. Periodic boundary con-
; LES, ; Durao expt., ; Lyn expt., ditions are imposed in the spanwise direction in three-

Journal of Applied Mechanics MAY 2006, Vol. 73 / 419

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Table 3 Flow past a circular cylinder

Case ReD Grid t* Cd Cpb St sep

PANS f k = 1.0 1.4 10 5


85 85 0.21 1.02 0.511 0.238 105
PANS f k = 0.7 1.4 105 150 135 42 0.0525 1.35 0.769 0.26 99.77
PANS f k = 0.5 1.4 105 150 135 42 0.0525 1.37 0.8 0.26 95.86
LES 13, Case 1.4 105 325 325 64 2.0 104 1.454 1.764 0.204 95
B1, Dynamic
DES 14, 1.4 105 150 109 42 0.05 1.08 1.04 0.21 77
Case LS8
Experimental 15 1.4 105 1.237 1.21 0.179 77

dimensional simulations and wall condition on the cylinder sur- lower resolution than the one used here. However, here our com-
face and transverse boundaries. The inflow is a uniform flow with putations are performed on the same grid to isolate the effect of f k
2% turbulence intensity. The fully developed condition is used at in the model.
the outflow. The computed flow is inherently unsteady, and the
flow velocities are averaged after statistical stationarity is 4.2.3 Flow Past Circular Cylinder. The PANS results are
established. again compared to LES 13, DES 14, and experiment 15. The
The mean streamwise velocity profiles along the centerline prediction of coefficient of drag Cd, the coefficient of back pres-
from various f k = 0.6 simulations are shown in Fig. 1a. The time sure C pb and the angle of separation sep again improve sig-
step in all these simulations of various grid sizes is 0.008 s. Simi- nificantly with smaller f k values Table 3. In Fig. 4, the surface
lar profiles from the various f k = 0.8 computations time-step C p distribution is shown for the various PANS calculations and
0.025 are given in Fig. 1b. It is found that for resolution better other competing methods. It is seen that URANS performs very
than 95 107, the mean flows solutions on various grids are con- poorly. The PANS results approach the experimental results with
fined to a narrow range of values. The extremes of the observed decreasing f k. The f k = 0.5 PANS case is nearly as accurate as an
range of behavior are shown in the figures. It is seen that the LES computation with ten times as many grid nodes and more
bands of f k = 0.8 and f k = 0.6 cases are distinctly different. The accurate than the DES calculation Fig. 4. The conclusion is
difference between the results on various grids for a given f k is clearly that the PANS can go from RANS to experiments
very small compared to the difference between simulations of smoothly as f k value is decreasedthe hallmark of a bridging
different f k values. For, example, all the f k = 0.6 simulations pre- model.
dict a much smaller separation bubble than the f k = 0.8 cases. 4.2.4 Three-Dimensional Driven Cavity Flow. Computations
Clearly, the PANS results are reasonably grid insensitive in this of the 3D driven cavity show that the RANS model captures the
range. mean flow profiles adequately well. Thus, PANS simulations with
Next, in Fig. 2, we present the results from f k = 0.6 simulations decreasing f k do not lead to significant improvements in predict-
of different time steps but identical grid size 95 107. The be- ing the mean flow variables. However, as seen in Fig. 5, with
havior is again constrained to a narrow band of values. This again
demonstrates reasonable insensitivity to time step, provided the
time increment is 0.025 s.
4.2.2 Three-Dimensional Flow Past Square Cylinder. In the
three-dimensional flow past a square cylinder study, we perform
PANS simulations with f k values of 1, 0.7 and 0.4. Guided by the
grid and time-step sensitivity study, we employ a grid of size 95
107 21 and time step 0.025 s for adequate accuracy at reason-
able cost. It is found that this grid is just adequate for the f k
= 0.4 case and finer than necessary for the other cases. Therefore,
any difference between the various f k computations are purely a
consequence of changing filter width and not numerical in origin.
The PANS results of various filter widths are compared to
URANS, LES 11, and experimental 12 data. The f k = 1 com-
putation can be viewed as the standard two-equation URANS
method. We find that with decreasing f k decreasing filter width,
the PANS results approach the experimental values for all
quantitiesStrouhal number, coefficient of lift, coefficient of
dragconsidered. These results are tabulated in Table 2. We fur-
ther compare the PANS centerline mean-streamwise velocity pro-
file obtained with various f k and with experimental results and
LES. It is found figure not shown that reducing f k leads to im- Fig. 4 Profiles of Coefficient of pressure distribution along the
proved agreement with experimental data. In Fig. 3, we compare cylinder surface for various fk values: , Experimental data
the Reynolds stress component uu of various PANS simula- from Achenbach for flow Reynolds number of ReD = 1 105; ,
LES data from Wang et al. at ReD 106; , DES data from Travin
tions to experimental and LES data. The results go from unaccept-
et al. 14 at ReD 1.4 105 data from run LS 8 corresponding to
able in the f k = 1 case to better than LES agreement in the f k laminar separation case. For laminar separation to occur the
= 0.4 case. It should be cautioned that the accuracy of low f k turbulent viscosity is set to zero at the inflow boundary; ,
comes, in general, at the expense of increased computational cost. data from PANS with fk = 0.5; , data from PANS with fk = 0.7;
For example, the f k = 0.7 and 1.0 computations require markedly ---, data from PANS with fk = 1.0.

420 / Vol. 73, MAY 2006 Transactions of the ASME

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Fig. 5 Vorticity structure in 3D driven cavity flow computations with various fk values

decreasing f k more fluctuating scales are liberated and the vortic- 3 Batten, P., Goldberg, U., and Chakravarthy, S., 2002, LNS-An Approach
Towards Embedded LES, AIAA Paper No. 2002-0427.
ity and fluctuating flow fields are significantly different. Although 4 Khorrami, M. R., Singer, B., and Berkman, M. E., 2002, Time-Accurate
no quantitative measure of the flow topography is available for Simulations and Acoustic Analysis of Slat Free Shear Layer, AIAA J., 407,
comparison, the qualitative features of f k = 0.2 case agrees best pp. 12841291.
with experimental data. For example, the number of vortex pairs 5 Khorrami, M. R., Singer, B., and Lockard, D., 2002, Time-Accurate Simula-
tions and Acoustic Analysis of Slat Free Shear Layer: Part 2, AIAA/CEAS
observed in the f k = 0.2 case matches the number observed in ex-
Aeroacoustics Conference and Exhibit, Breckenridge, CO, AIAA Paper No.
periments 16,17. 2002-2579.
Based on these preliminary results, it can be reasonably con- 6 Hedges, L. S., Travin, A. K., and Spalart, P. S., 2002, Detached-Eddy Simu-
cluded that the PANS methodology proposed in this paper satisfies lation Over a Simplified Landing Gear, ASME J. Fluids Eng., 124, pp. 413
the requirements of a bridging model. 423.
7 Girimaji, S. S., Jeong, E., and Srinivasan, R., 2006, PANS Method for Tur-
bulence: Fixed Point Analysis and Comparison With URANS, ASME J. Appl.
5 Conclusion Mech., 73, pp. 422429.
Following the turbulence modeling paradigm proposed by Spe- 8 Germano, M., 1992, Turbulence: The Filtering Approach, J. Fluid Mech.,
238, pp. 325336.
ziale 1, we formulate and develop the partially averaged Navier- 9 Girimaji, S. S., 2000, Pressure-Strain Correlation Modeling of Complex Tur-
Stokes PANS bridging technique for modeling turbulent flows. bulent Flows, J. Fluid Mech., 422, pp. 91123.
The PANS closure model is formally extracted from the RANS 10 Schiestel, R., 1987, Multiple Time-Scale Modeling of Turbulent Flows in
closure, for given filter-control parameters. Many important prac- One-Point Closures, Phys. Fluids, 30, pp. 722731.
11 Sohankar, A., and Davidson, L., 2000, Large Eddy Simulations of Flow Past
tical issues, such as near-wall PANS modeling, a priori strategy
a Square Cylinder: Comparison of Different Sub-Grid Scale Models, ASME
for specifying f k and f , and precise model for the turbulent trans- J. Fluids Eng., 122, pp. 3947.
port term, need further investigation. Yet, there is reasonable the- 12 Lyn, D. A., Einav, S., Rodi, W., and Park, J. H., 1992, A Laser-Doppler
oretical justification and adequate preliminary computational evi- Velocimetry Study of the Ensemble-Averaged Characteristics of the Turbulent
dence to be cautiously optimistic about this new methodology. Wake of a Square Cylinder, J. Fluid Mech., 304, pp. 285319.
13 Breuer, M. A., 1995, A Challenging Test Case for Large Eddy Simulation:
High-Reynolds Number Circular Cylinder Flow, Int. J. Heat Fluid Flow, 21,
Acknowledgment pp. 648654.
I would like to dedicate this paper to Professor Charles Gregory 14 Travin, A., Shur, M., Strelets, M., and Spalart, P., 1999, Detached Eddy
Simulations Past a Circular Cylinder, Flow, Turbul. Combust., 63, pp. 293
Speziale, a good friend and mentor. This material is based on 313.
work supported by NASA under Grant No. NAG-01-005. 15 Cantwell, B., and Coles, D., 1983, An Experimental Study of Entrainment
and Transport in the Turbulent Near Wake of a Circular Cylinder, J. Fluid
Mech., 136, pp. 321374.
References 16 Jordan, S. A., and Ragab, S. A., 1994, On the Unsteady and Turbulent Char-
1 Speziale, C. G., 1996, Computing Non-Equilibrium Flows With Time- acteristics of Three-Dimensional Shear-Driven Cavity Flow, ASME J. Fluids
Dependent RANS and VLES, 15th ICNMFD. Eng., 116, pp. 439449.
2 Spalart, P. R., 2000, Trends in Turbulence Treatments, Fluids 2000, Denver, 17 Prasad, A. K., and Kosef, J. R., 1989, Reynolds Number and End-Wall Ef-
AIAA Paper No. 2000-2306. fects on a Lid-Driven Cavity Flow, Phys. Fluids, 1, pp. 208218.

Journal of Applied Mechanics MAY 2006, Vol. 73 / 421

Downloaded From: http://appliedmechanics.asmedigitalcollection.asme.org/ on 09/12/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Vous aimerez peut-être aussi