Documente Academic
Documente Profesional
Documente Cultură
1 Introduction case all the solutions reported were unsuccessful in predicting the
measured data in the bend region and the resolved structure of the
Computational Fluid Dynamics tools are becoming standard in
flow field was significantly affected by the choice of the turbu-
many fields of engineering involving flow of gases and liquids;
numerical simulations are used both in the design phase to select lence model.
between different concepts and in the production phase to analyze The uncertainties associated with 共i兲 different computational
performance. Industrial CFD applications require high flexibility grids, 共ii兲 boundary conditions definition, 共iii兲 convergence, and
in the grid–generation procedure for complex configurations, 共i v 兲 numerical schemes do not allow drawing specific conclusions
short turn around time, and easy–to–use environments. At about the codes used, other than the usual conclusion that further
present, several commercial packages are available for the CFD research into more advanced turbulence models for use in com-
industrial community; these packages are usually integrated sys- mercial CFD codes is required 关1兴.
tems which include a mesh generator, a flow solver, and a visual- In order to carry out a fair comparison between different CFD
ization tool. Often the numerical techniques adopted in these CFD codes and to establish definitive conclusions on the state–of–the–
codes are well accepted algorithms published in the open litera- art of commercial CFD codes, all the differences 共i-i v 兲 must be
ture; the selection of one technique with respect to others is usu- fully addressed and, if possible, eliminated. In the present work,
ally based on robustness and reliability. an effort has been made to control all these parameters. The codes
There have been few attempts in the literature to compare the available for comparison are CFX, Fluent, and Star-CD. The ob-
performance of these codes; laminar and turbulent test cases have jective is to compare their predictive capabilities for the simula-
been proposed to several CFD code vendors by the Coordinating tion of a turbulent separated flow. Several turbulence closures
Group for Computational Fluid Dynamics, of the Fluids Engineer- 共and near-wall treatments兲 are available in these codes ranging
ing Division of ASME关1兴. A series of five benchmark problems from k-⑀-type models to full Reynolds stress models. The main
were calculated, with all the mesh generation and simulations per- focus of the work is on two models: the k-⑀ low-Reynolds model
formed by the vendors themselves; only two of the problems re- by Launder and Sharma 关3兴 and the v ⬘ 2 ⫺ f by Durbin 关4兴. In
quired turbulent simulations. The first problem is the flow around addition, results obtained using different closures are reported.
a square cylinder; the flow is unsteady and all the codes predicted The k-⑀ model is well described in the literature and has been
reasonably well the measured Strouhal number. However, poor widely used. Its implementation poses some challenges and it re-
accuracy resulted in the prediction of the details of the wake flow quires the solution of two transport equations with numerically
field. It was also noted that, depending on the code used 共and stiff source terms. This model is available in all the codes consid-
assuming grid-converged results兲 the same k-⑀ model predicted ered and, although it is not expected to be extremely accurate 关5兴,
very different results. The reasons for this difference can be dif- it provides common ground for comparisons between different
ferent grids, no demonstration of grid convergence, different codes.
implementations of the models, and different boundary conditions. The v ⬘ 2 ⫺ f model 共implemented in a NASA research code兲 has
It must also be pointed out that the prediction for this problem is been already successfully used for simulating separated flows 关4兴,
strongly affected by the treatment of the stagnation point region. three dimensional configurations 关6兴 and flows with heat transfer
As shown by Durbin 关2兴, the k-⑀ models predict a spurious high 关7兴. It is rather complex involving the solution of four differential
level of turbulent kinetic energy in this region. equations 共three transport equations plus an Helmotz-type equa-
The other turbulent problem reported by Freitas 关1兴 was the tions兲.
three-dimensional developing flow in a 180 degrees bend. In this The test case analyzed in this study is a two-dimensional tur-
bulent flow in a diffuser. Due to the adverse pressure gradient the
Contributed by the Fluids Engineering Division for publication in the JOURNAL flow is separated and a large recirculation bubble is generated.
OF FLUIDS ENGINEERING. Manuscript received by the Fluids Engineering Division
October 16, 2000; revised manuscript received May 21, 2001. Associate Editor: This problem has been selected because a very reliable experi-
I. Celik. mental database is available. Moreover, a detailed Large Eddy
Journal of Fluids Engineering Copyright © 2001 by ASME DECEMBER 2001, Vol. 123 Õ 819
Simulation study was carried out at the Center for Turbulence 2.3 Star-CD v3.1. Star-CD v3.1 is a CFD computer code
Research and the resulting numerical database is also available for developed and marketed by Computational Dynamics Ltd. The
comparison 关8兴. code solves the three-dimensional NS equations on unstructured
The objective of the paper is to present a detailed and careful meshes; various linear and non-linear two-equation turbulence
comparison of the simulations performed using three commercial models are available 关13兴.
CFD codes. Although the flow under investigation is geometri- Star-CD uses the SIMPLE technique for velocity-pressure cor-
cally simple, it is challenging for turbulence modeling and must rection and a PCG method to solve the implicit system of equa-
be considered as a necessary step to evaluate the merits of differ- tions; several first and high order spatial discretization schemes
ent approaches. In addition, its simplicity allows to control all the can be used including QUICK.
numerical parameters involved in the simulations and to under-
stand the causes of discrepancy between the codes.
The three CFD codes used are briefly introduced in the next 3 Turbulence Modeling
section; the turbulence models with the governing equations are
presented in Sec. 3, while results and comparison are in Sec. 4. Several turbulence models are available in the codes presented
in the previous section. Most of them are derived from the stan-
2 Numerical Method dard k-⑀ model 关14兴 with different treatments of the wall region.
The low-Reynolds model of Launder and Sharma 关3兴 and the
The steady Navier-Stokes 共NS兲 equations for an incompressible
fluid are considered: v ⬘ 2 ⫺ f model 关4兴 are the focus of this work, and are described in
detail. Additional simulations are performed with the k-⑀ Two-
ui Layer Model 关15兴, the cubic Non-Linear Eddy-Viscosity 共NLEV兲
⫽0 (1) k-⑀ Model 关16兴, and the Differential Reynolds Stress Model 关13兴.
xi
The Launder and Sharma k-⑀ model is available as a standard
ui
u j
⫽
xi x j 冋
共⫹t兲
u j
x j
⫺册p
x j
(2)
option in all the codes 共a slightly different damping function is
employed in Star-CD兲. The v ⬘ 2 ⫺ f model has been implemented
using the User Defined Subroutines in each of the codes.
where u i are the mean velocity components, p is the pressure, and
and t the laminar and turbulent viscosity, respectively. Addi- 3.1 Low-Reynolds k-⑀ Model. The k-⑀ model was intro-
tional equations for turbulent quantities are considered to compute duced by Launder and Spalding 关14兴. The high Reynolds number
the eddy viscosity, and are explained in the following section. version is obtained by neglecting all the terms containing the ki-
All the codes solve the discretized equations in a segregated nematic viscosity. In the proximity of solid walls, viscous effects
manner, with the SIMPLE 共Semi-Implicit Method for Pressure- become important and this assumption no longer holds. Several
Linked Equations兲 algorithm, or its ‘‘consistent’’ variant, SIM- modifications have been proposed: in the two-layer formulation
PLEC 关9兴, used to achieve the pressure-velocity coupling for sta- 关15兴, a simpler model is used close to the wall 共usually a one-
bility. In the SIMPLE algorithm, the continuity equation 共1兲 is equation model兲 and then the eddy viscosity is patched at a certain
converted into a discrete Poisson equation for pressure. The dif- distance from the wall; both Fluent and Star-CD offer this option.
ferential equations are linearized and solved implicitly in se- In the damping functions approach 关17兴 algebraic functions are
quence: starting with the pressure equation 共predictor stage兲, fol- introduced to correct the behavior of turbulent quantities close to
lowed by the momentum equations and the pressure correction the wall. Several different choices are available in the open litera-
equation 共corrector stage兲. The equations for the scalars 共turbulent ture. All the codes have built-in damping function models; in par-
quantities兲 are solved after the updating of both pressure and ve- ticular, Fluent has six different versions available. In this work,
locity components. Within this loop, the linearized equations for the model introduced by Launder and Sharma 关3兴, which is avail-
each variable, as they arise, are treated using a linear system able in all the codes, was used.
solver 共i.e., multigrid, Preconditioned Conjugate Gradient, PCG, The k-⑀ equations are:
冋冉 冊 册
etc.兲.
A brief description of the codes is given in the next subsections k t k
ui ⫽ P⫺ ⑀ ⫹ ⫹ ⫺D (3)
with emphasis given only to the features required for this study. xi x j k x j
冋冉 冊 册
All the codes allow the implementation of customized models
through User Defined Subroutines. ⑀ f 1 C ⑀ 1 P⫺ f 2 C ⑀ 2 ⑀ t ⑀
ui ⫽ ⫹ ⫹ ⫹E (4)
xi T x j ⑀ x j
2.1 CFX v4.3. CFX v4.3 is a CFD computer code devel-
oped and marketed by AEA Technologies. The code solves the The eddy viscosity is obtained from
three-dimensional NS equations on structured multiblock grids for
both compressible and incompressible flows. Various turbulence t ⫽C f kT (5)
models are available ranging from two-equation to complete Dif- The damping functions f 1 , f 2 , and f , the timescale T and the
ferential Reynolds Stress Models 共DRSM兲. CFX uses a SIMPLEC extra source terms D and E are:
pressure-correction scheme 共SIMPLE is also available兲, and sev-
eral spatial discretizations which include first-order Upwind Dif- f 1 ⫽1 (6)
ferencing 共UD兲 and QUICK 关10兴; central differencing is used for 2
⫺Re T
the pressure. The linear system arising at each iteration is then f 2 ⫽1⫺0.3e (7)
solved using a Preconditioned Conjugate Gradient technique. ⫺3.4/ 共 1⫹0.02ReT 兲 2
f ⫽e (8)
2.2 Fluent v5.3. Fluent v5.3 is a CFD computer code devel- T⫽k/ ⑀ (9)
冉冑冊
oped and marketed by Fluent Inc. The code provides mesh flex-
2
ibility by unstructured meshes. Turbulence closures range from k
one-equation turbulence model up to DRSM 关11兴. D⫽2 (10)
xn
Fluent employs the SIMPLEC technique and an algebraic mul-
tigrid linear system solver to update the solution at each iteration.
The QUICK spatial discretization technique is available among
others. In particular, a second-order Total Variation Diminishing
E⫽2 t 冉 冊 2u i
x jxk
2
(11)
共TVD兲 limited discretization for the pressure in the Poisson equa- where ReT⫽k2/⑀ is the turbulent Reynolds number and x n is the
tion is used 关12兴. direction normal to walls.
T⫽max 冋 冑册
k
⑀
,6
⑀
(15)
D⫽0 (16)
E⫽0 (17)
The additional equations model the turbulence velocity scale v ⬘ 2 ,
and its production, k f :
ui
v ⬘2
xi
⑀
⫽k f ⫺6 v ⬘ 2 ⫹
k x j
冋冉 冊 册
⫹
t v ⬘2
k x j
(18)
2 f 共 2/3⫺ v ⬘ 2 /k 兲 P 5 v ⬘ 2 /k
f ⫺L 2 ⫽C 1 ⫹C 2 ⫹ (19)
x jx j T k T
where L is the length scale, defined as
L 2 ⫽C L2 max 冋 冑册
k3 2
,C
⑀2
3
⑀
(20)
Fig. 1 Asymmetric diffuser geometry
The eddy-viscosity damping is provided in this case by the
presence of v ⬘ 2 共Eq. 共14兲兲 instead of k in Eq. 共5兲. In other words,
the amount of damping is controlled by the ratio between v ⬘ 2 and
k 共instead of the turbulent Reynolds number, Ret , like in the k-⑀
model兲 which is a measure of the turbulence anisotropy 关4兴. The
other important feature of the v ⬘ 2 ⫺ f model is the nonlocality
arising from the solution of an elliptic equation for f . Fig. 2 Computational grid—detail of the channel-diffuser
The v ⬘ 2 ⫺ f model has been implemented by the author in the connection
three CFD codes described above. Four additional scalars are
solved and the diffusion, source, and convective terms are speci- Table 1 Numerical parameters used for the simulations. Leg-
end: U,V: Mean Velocity Components; P: Pressure; TQ: Turbu-
fied according to Eqs. 共3兲-共4兲 and 共18兲-共19兲. The eddy viscosity is
lent Quantities; CD: Central Differencing; UD: Upwind Differ-
then computed according to Eq. 共5兲 and used in the mean flow encing; TVD: Total Variation Diminishing.
Eqs. 共1兲-共2兲.
Spatial Discretization Pressure Correction Under-Relaxation
4 Results U,V P TQ U,V P TQ
Steady flow in asymmetric, two-dimensional diffuser is consid-
CFX QUICK CD UD SIMPLEC 0.65 1 0.6
ered. This problem was a test-case for the 8th ERCOFTAC/IAHR/ Fluent QUICK TVD UD SIMPLEC 0.65 1 0.6
COST Workshop on Refined Turbulence modeling in Espoo, Fin- Star-CD QUICK CD UD SIMPLE 0.70 0.2 0.6
land, 17-18 June 1999.
Fig. 3 Convergence history „L ⴥ norm…. Left column: v ⬘ 2 À f model; right column: low-Reynolds k -⑀ model.
„a… CFX v4.3; „b… fluent v5.3; „c… star-CD v3.1.
Fig. 4 Mean streamwise velocity—CFX. Contour levels MinÄÀ0.05; maxÄ1.0, ⌬Ä0.05 „dashed lines negative
values….
ticular, the separation zone is captured 共even if the maximum especially in turbulent kinetic energy. The very good agreement
intensity of the recirculating velocity is underestimated兲. The pre- obtained by using the v ⬘ 2 ⫺ f suggests that the differences are not
dictions using the k-⑀ model are in poor agreement with the data related to the numerical techniques used to discretize the equa-
because the model fails to respond correctly to the adverse pres- tions but to the implementation of the models. For example, dif-
sure gradient and misses the separation completely. The compari- ferent approximations of the terms in 共10兲-共11兲 could lead to the
sons reported in Fig. 6 for the turbulent kinetic energy confirm the mentioned differences.
quality of the v ⬘ 2 ⫺ f predictions as compared to the k-⑀. The peak In particular, it is worth noting that StarCD and Fluent results
of the turbulent intensity is very well predicted by the v ⬘ 2 ⫺ f are closer to each other 共especially for the turbulent kinetic en-
model in the diffuser. However, in the recovery region 共after the ergy兲 than they are with CFX. This may be related to the fact that
reattachement兲 the model underestimates the level of kinetic en- both are unstructured mesh codes 共whereas CFD is a structured
ergy. This is consistent with the v ⬘ 2 ⫺ f calculations shown by grid solver兲 and they deal similarly with the issues 共mentioned at
Durbin 关4兴, the LES results reported by Kaltenbach et al. 关8兴 and the end of Sec. 3.1兲 related to the computation of the cross deriva-
with the recent computations presented by Apsley and Leschziner tives in the term D 共Eq. 共10兲兲 and the evaluation of the wall
关5兴 using quadratic and cubic nonlinear k-⑀ models. Possible rea- normal direction. It is also useful to add that the use of the stan-
sons for this disagreement are the presence of strong three- dard f damping function available in StarCD 共instead of the one
dimensional effects after the flow reattachement and strong non- reported in 共8兲兲 leads to somewhat different results which no
equilibrium effects which cannot be correctly accounted for in longer agree with the Fluent results.
single–point closures. The results using the k-⑀ model completely Finally in Fig. 7 the skin friction coefficients on the lower and
fail to capture the asymmetric development of the turbulent ki- upper walls are reported. The separation bubble on the curved
netic energy and underestimate its magnitude in the diffuser. wall is indicated by a negative skin friction from x/H⬇7 to
The three codes show some differences when the same k-⑀ x/H⬇30; the v ⬘ 2 ⫺ f model predicts the bubble in very close
model is invoked. The disparities are in the mean velocity and agreement with the experiments. The k-⑀ model fails to predict
any separation 共as already noted兲. In addition, the three codes employed to enforce the positivity of selected quantities 共turbulent
predicts quite different friction levels when the k-⑀ closure is em- variables, for example兲 and to improve convergence quality.
ployed. Additional simulations are performed using different turbulence
In order to assess the grid sensitivity of the results presented, models to explore capabilities of the CFD codes tested and the
additional calculations on a refined grid were carried out. The grid results are presented in Fig. 9. The standard k-⑀ model with the
was obtained by doubling the number of points in the two direc- two-layer near-wall treatment gives results which are closely com-
tions. The comparison between coarse and fine grid is reported in parable to the predictions presented in Figs. 5 and 6 共using the
Fig. 8, in terms of mean flow velocity and turbulent kinetic energy damping functions兲. The separation on the bottom wall is not cap-
profiles using Fluent and the low-Reynolds k-⑀ model. The results tured and the asymmetry in the turbulent kinetic energy profiles is
show that a grid converged solution has been reached. Similar very small. The two-layer treatment of the near-wall regions is
comparisons are obtained for the other two codes. In addition, the available in both Fluent and StarCD and the results are compa-
results obtained using a high-order upwind discretization for the rable. On the other hand, the Nonlinear version of the Launder
turbulent quantities are also reported on the same plots. The dif- and Sharma k-⑀ model 共available only in StarCD兲 captures the
ference, in this case, is very small being the flow dominated by separation and gives a reasonably good representation of the tur-
turbulence generation. This conclusion does not apply to more bulent kinetic energy. The results are in agreement with the ex-
complex situations where the use of high-order differencing for periments and close to the predictions of the v ⬘ 2 ⫺ f model. These
the turbulent equations is mandatory. results are also in agreement with the data reported in the work by
The grid convergence study shows that the results obtained are Apsley and Leschziner 关5兴.
not dependent on the grid and therefore, the differences in the Finally, results are also presented for calculations with the
streamwise velocity profiles in Fig. 5 and in the turbulent kinetic DRSM model in low-Reynolds number form. This model is avail-
energy in Fig. 6, are not directly related to discretization accuracy able in Fluent 共in CFX only a High-Reynolds number version is
or to the presence of artificial dissipation. One possible cause of available兲. The DRSM predictions do not show the expected im-
the discrepancy between the codes is the presence of limiters/ provement with respect to the standard k-⑀ model. This could be
smoothers in the solution procedure. These operators are usually partly related to the near-wall treatment based on the two-layer
approach. Calculations performed in a similar configuration with dure based on the SIMPLE technique has been used. In terms of
the high-Reynolds version of the DRSM in Fluent 关11兴 demon- convergence behavior, all the codes reach the steady-state ap-
strated good agreement with the experimental measurements. proximately in the same number of iterations, regardless of the
It is worth noting that the two-layer k-⑀ required about the same turbulence model used. The accuracy of the calculations as com-
amount of CPU as the Launder and Sharma model presented be- pared to the experimental and LES data is very good using the
fore and the convergence behavior was very similar. On the other v ⬘ 2 ⫺ f model. The length of the recirculation region is captured to
hand, a slight increase in computational time is associated with within 6 percent and the skin friction on both walls agree reason-
the NLEV model and 25 percent more iterations were required to ably well with the data. The negative velocity in the separation
achieve the same drop in the residuals. The DRSM simulation bubble is slightly underestimated. The results using the k-⑀ model
required a CPU comparable with the v ⬘ 2 ⫺ f one 共the number of do not show any recirculation. The flow is fully attached and this
differential equations to be solved is the same for two-dimensional leads to a severe underprediction of the maximum velocity in the
problems兲 but almost twice as many iterations were required to diffuser.
achieve convergence. An effort was made to control all aspects of the simulations so
that the same results were expected using different codes. In par-
5 Conclusions ticular, the implementation of the v ⬘ 2 ⫺ f turbulence model was
A comparison between three CFD commercial codes, namely carried out the same way in the three codes; indeed v ⬘ 2 ⫺ f results
CFX, Fluent, and Star-CD, is reported for turbulent flow in a do show an almost perfect agreement between CFX and Star-CD
planar asymmetric diffuser. Two turbulence models have been with Fluent being slightly more dissipative. The results using the
used. The first is the low-Reynolds number k-⑀ model 共with Laun- k-⑀ model, on the other hand, show strong sensitivity to the code
der and Sharma damping functions兲 which is available as a stan- used. The model formulation is exactly the one proposed by Laun-
dard feature in the codes. The second model is the v ⬘ 2 ⫺ f model der and Sharma, but the results are different 共especially in terms
that has been implemented through the User Defined Routines in of turbulent quantities and friction coefficients兲. This may be due
the three codes. to differences in implementation details which are not specified in
The same grid and the same spatial discretization have been the user manuals. In general, the differences between the k-⑀ re-
used for all the simulations. In addition, a similar iterative proce- sults are much larger than those obtained using v ⬘ 2 ⫺ f , suggest-
ing that the differences are less due to details of the numerical physical model in an industrial tool is an available option for CFD
procedures in the codes than to the implementation of the turbu- practitioners thus shortening the distance between published re-
lence models. search work and real–world applications.
Today, one of the challenges in using commercial CFD codes is
to choose between several physical/numerical models available.
The cross comparison presented in this work proved that the basic Acknowledgments
numerical techniques 共default options兲 are reliable and deliver the The author wishes to thank A. Ooi for providing the initial
expected performance in terms of accuracy and convergence at implementation of the v ⬘ 2 ⫺ f model in Fluent, and M. Fatica for
least when the computational grid and the boundary conditions are providing the LES data for the diffuser; in addition, support for
defined carefully. On the other hand, the selection of the correct
the implementation of the v ⬘ 2 ⫺ f model was provided by P. Ma-
physical model 共in this case the turbulence model兲 is crucial for
lan 共Fluent Inc.兲, S. Jonnavithula 共Adapco Ltd.兲, and H. Pordal
the success of the simulations. Using one of the available turbu-
共AEA Technologies Inc.兲. Discussion with P. Moin, P. Durbin, and
lence models the results were not accurate and, in addition, not
G. Medic are appreciated.
reproducible using different codes.
The v ⬘ 2 ⫺ f model was implemented in CFX, Fluent, and
StarCD only using the User Defined Routine feature. Even if the References
model is rather complex 共involving three differential transport 关1兴 Freitas, C. J., 1995, ‘‘Perspective: Selected Benchmarks From Commercial
equations and a Helmotz-like equation兲 no particular difficulty CFD Codes,’’ ASME J. Fluids Eng., 117, p. 210–218.
was faced by the author. The performance of the codes was not 关2兴 Durbin, P. A., 1996, ‘‘On the k-⑀ Stagnation Point Anomaly,’’ Int. J. Heat Fluid
Flow, 17, pp. 89–91.
compromised when compared with built-in models, and the ex- 关3兴 Launder, B. E., and Sharma, A., 1974, ‘‘Application of the Energy-Dissipation
pected accuracy level was reached with all the commercial codes Model of Turbulence to the Calculation of Flow Near a Spinning Disk,’’ Lett.
tested. This demonstrates that the implementation of a customized Heat Mass Transfer 1, pp. 131–138.