Professional Documents
Culture Documents
Hirsch and Tartinville - 2009 - Reynolds-Averaged Navier-Stokes Modelling For Industrial Applications and Some Challenging Issues
Hirsch and Tartinville - 2009 - Reynolds-Averaged Navier-Stokes Modelling For Industrial Applications and Some Challenging Issues
Hirsch and Tartinville - 2009 - Reynolds-Averaged Navier-Stokes Modelling For Industrial Applications and Some Challenging Issues
Reynolds-Averaged Navier-Stokes modelling for industrial applications and some challenging issues
Charles Hirscha,b* and Benoit Tartinvilleb
a
Vrije Uuniversiteit Brussel, Department of Fluid Mechanics, Brussels, Belgium; bNumeca Int., Av. Franklin Roosevelt,
5, Brussels 1050, Belgium
(Received 28 October 2008; final version received 22 January 2009)
CFD simulations with the Reynolds-Averaged Navier-Stokes (RANS) model have become the widespread standard
in industry, forming the underlying base for numerous design procedures. This raises the issue of the reliability of the
associated turbulence models and to a lesser extent of the numerical accuracy and associated errors. The article
provides an overview of representative industrial applications and questions raised concerning identified weaknesses
of current turbulence models.
Keywords: turbulence modelling; discretisation errors; artificial dissipation; turbulent mixing
. Refine the grids in regions with high gradients, also to the issue of keeping the load balance of the
such as boundary layers, leading edges of airfoils partitions with massive parallel simulations.
and any region where large changes in flow On the other hand, with implicit methods, fully
properties might occur. coupled Jacobians for the momentum and turbulence
. Make sure that the number of points in the equations together should be considered, for theore-
boundary layers is sufficient for the expected tical reasons. However, taking into account the full
accuracy; more than 10 points over the inner part Jacobians raises serious problems of robustness,
of the boundary layer thickness should be because of the nonlinear properties of the source
considered. terms. Hence, we need simplifications, such as linear-
. Monitor the grid quality by adequate mesh isations, or reduction to point implicit approach, or to
parameters, available in most of the grid a full decoupling of the turbulence model equations,
generators, such as aspect ratio, internal angle, which is widely applied.
concavity, skewness, negative volume.
Figure 2. Comparison of RANS simulations with different turbulence models for the OBI axisymmetric diffuser. The top figure
(including streamlines in red and the velocity magnitude in colour) shows the position and extends of the separation region,
whereas the bottom figure compares calculated and measured pressure distribution, wall shear stress at the bottom wall, and
velocity profiles at the four positions indicated in the lower insert. (Courtesy Haase et al. 2006.)
measured pressure distribution and wall shear stress at the tested turbulence models, in that the velocity
the bottom wall, velocity profiles at the positions profile in the downstream duct of the diffuser is
indicated in the lower insert. From these figures it experimentally fully recovered, whereas the calculated
appears that the k-e and k-o models are unable to profiles still show remaining effects of their earlier
properly capture the recirculation region and that separation. Furthermore, the pressure recovery along
more recent turbulence models such as the SST and v2- the inclined wall is over-estimated by all the models.
f models produce a maximum backward velocity that is This is in accordance with the observed underestima-
closer to experimental data. Such a finding has also tion of the velocity above this wall, resulting from an
been observed by Iaccarino (2000) when using k-e and underestimated extends of the separation region.
v2-f models. It can be observed that the Spalart– Furthermore, LES of this test case can also be
Allmaras turbulence model results do match pretty found in the literature (Kaltenbach et al. 1999). They
well on the measured velocity profile at the down- show a very good agreement between experimental
stream corner of the inclined wall. The last figure is data and computed results in the recirculation region.
noteworthy, as it demonstrates a clear weakness of all For instance, the location of the separation and
International Journal of Computational Fluid Dynamics 299
Figure 3. Typical mesh for the UCAM wind tunnel, showing here half of the tunnel section, with the upper wall removed for
display.
reattachment points are accurately predicted when 3D geometry of the wind tunnel has been meshed. A
compared with data from Buice and Eaton. (1997). block-structured mesh with about 3.3 6 106 cells has
The agreement between results from LES without been generated. The clustering at the solid walls has
sidewalls and the experimental data also confirms the been defined to have yþ close to unity at the first inner
nearly two-dimensional nature of the flow, although cell and the maximum expansion ratio in the boundary
separation is often tri-dimensional. Because the velo- layer is about 1.2 (Figure 1).
city profiles far downstream of the diffuser are not The Euranus flow solver from Numeca Interna-
shown in this article, we cannot state if the velocity tional has been used to perform these simulations. It is
recovery is well captured by LES. This test case clearly a structured multi-block Navier-Stokes code using
shows that current turbulence models need serious finite volume approach. Central-space discretisation is
improvements for prediction of separated flows, as well employed together with Jameson type artificial dis-
as for the prediction of the mixing-recovery effects. We sipation. A four-stage Runge-Kutta scheme is selected
can hope that the gained knowledge on turbulence for the temporal discretisation. Multi-grid, local time
from advanced DNS and LES will contribute to this stepping and implicit residual smoothing are also used
highly needed improvement of current turbulence to speed-up the convergence.
models. At the nozzle entry, the total pressure and
temperature are imposed in accordance with figures
provided by UCAM. The walls are assumed adiabatic
Shock-boundary layer interactions in internal flows and smooth. The outlet, where the static pressure is
This case is part of the current UFAST EU project imposed, is located 330 mm downstream of the
(http://www.ufast.gda.pl/) aimed at the investigation measured shock position.
of unsteady effects of shock wave induced separation. According to information provided by UCAM the
Experimental data are obtained at the University of exact location of the laminar to turbulent transition is
Cambridge (UCAM) by the team of Dr. Olger unknown. Because the location of the shock wave
Babinsky on the interaction of a planar normal shock should strongly depend on the thickness of the
with a turbulent boundary layer (Bruce and Babinsky turbulent boundary layer, it has been decided to adapt
2008). Three cases are considered: Mach number 1.3, the outlet pressure in order to match the location of the
1.4 and 1.5. At a Mach number of about 1.3, the experimental shock wave.
boundary layer remains attached; on the other hand at
the higher Mach numbers of 1.4 and 1.5, a separation
appears downstream of the shock. The Reynolds Application uncertainties
number based on the boundary layer thickness Because of the lack of precise information or to
30 mm upstream of the shock is of the order of simplifications (for instance in the geometry) several
200,000. uncertainties can potentially affect the results of the
CFD calculation and its comparison to experimental
data. Two major uncertainties can be listed:
Grid and computation settings
To have a proper development of the boundary layer . According to report from UCAM, the location
downstream of the sonic throat, the nozzle was of the laminar to turbulent transition is
included in the computational domain and the full unknown.
300 C. Hirsch and B. Tartinville
. The inlet boundary conditions for the turbulent appears also that the solution is symmetrical on
quantities are unknown. coarser grid levels, and becomes non-symmetrical
when using a refined mesh. Furthermore, a series of
numerical tests have been performed to verify the
The surprising appearance of non-symmetrical solutions sensitivity of the non-symmetrical solutions to numer-
At the higher Mach numbers 1.4 and 1.5 involving ical schemes and parameters. None of the tests
separations, CFD results obtained with different suppressed the non-symmetrical behaviour.
codes exhibit nonsymmetrical solutions, depending To investigate if the non-symmetrical pattern is
on the turbulence model used, whereas the geo- only linked to the Spalart–Allmaras turbulence model
metry is symmetrical while the experiments seem to a series of simulations have been conducted with six
indicate symmetrical flow patterns, although slight different turbulence models: Baldwin–Lomax, Spalart–
asymmetries cannot totally be excluded, as seen from Allmaras, k-e, k-o, SST and v2-f. All the other
Figure 4. parameters of these simulations are identical. Because
CFD results on the UCAM configuration have even the back-pressure is maintained fixed for all these
been computed with a variety of turbulence models. At simulations, the shock location depends on the
the lowest Mach, whatever turbulence model is used all predicted location of the laminar to turbulent transi-
the results display a symmetrical behaviour. At Mach tion, and could be different from the experimental
1.4, the Spalart–Allmaras and SST turbulence model shock location at x ¼ 0.6 m.
produce a non-symmetrical solution, whereas the k-e The convergence history of all these simulations is
model leads to a symmetrical solution. At the highest displayed in Figure 5. The first finding is that the
Mach of 1.5, no symmetrical solution was obtained. It Baldwin–Lomax model cannot converge at the same
Figure 4. Oil flow visualisation from UCAM (M ¼ 1.5) (left) and IMP (M ¼ 1.45) (right).
level as the other models. It appears that two models corner separation. Finally, and in order to further check
produce a convergence in two steps: the Spalart– the possible causes of the non-symmetrical patterns, a
Allmaras and the SST, appearing to converge first to simulation has been performed on quarter of the
a symmetrical solution, followed by an increase in computational domain (assuming two planes of sym-
the residuals which further converge to the non- metry) with Spalart–Allmaras and k-e turbulence
symmetrical solution. The other models converge models. Though the latter model produces a lambda
continuously towards a residual of 1076. The v2-f shock that is close to the observed one, the lambda shock
model has a more erratic behaviour with a residual produced by the former is too thick and reaches the
drop between 1075 and 1076. All calculations are symmetry plane. Furthermore, the corner flow separa-
performed in single precision. tion is much more intense when using the Spalart–
The solutions are displayed on Figures 6 and 7 and Allmaras model compared to the k-e model. This means
two important recirculation zones can also be observed that the non-symmetrical behaviour can be because
in the corners, showing that the flow is highly tri- of a complex interplay between the turbulence model,
dimensional. The asymmetry seems to be driven by the the corner flow separation, and the lambda shock.
excessive corner separation and several questions still A similar behaviour was observed on another test
remain unanswered. Even if the experiments would case with a curved nozzle run at the IMP in Gdansk,
indicate some level of asymmetry, it remains far less Poland; see Doerffer (2007) for some details. Experi-
compared to the CFD results, which reveal some major mental data at Mach 1.35 and 1.45 for a flat wall
shortcomings in the prediction of separation, as already nozzle (UFAST case), an accelerating nozzle and
mentioned with the OBI diffuser case. In particular, the curved nozzle have been collected over the last years.
interaction length in CFD seems much greater than in Calculations from IMP also showed CFD results on
the experiments; the separation length in CFD seems the flat wall nozzle at Mach 1.45 with a nonsymmetrical
too large and too much effective blockage tends to be pattern when using the Spalart–Allmaras turbulence
predicted by the turbulence models. model, while results at Mach 1.35 are symmetrical.
Furthermore, to check the flow pattern without any Calculations with three flow solvers Fine/Turbo,
shock wave, a simulation has been conducted on the SPARC and Fluent all produce a non-symmetrical
fine mesh with the Spalart–Allmaras model with an solution. When using either a coarser mesh, or the
increased back pressure. The flow does not display any Speziale et al. turbulence model, the solution remains
non-symmetrical pattern indicating that the non- symmetrical. Computations performed by adding
symmetrical pattern is due to the shock-boundary chamfers at the flat walls with different thickness of
layer interaction and the induced three-dimensional 0.01, 1 and 5 mm – to be compared to the size of the
tunnel 100 mm – show that the nonsymmetrical Doerffer, P., 2007. European research on unsteady effects of
solution is suppressed with a thick enough chamfer. shock wave induced separation UFAST-project. In:
Proceedings of the 8th international symposium on
This confirms that the asymmetrical solution is because experimental and computational aerothermodynamics of
of the corner flow separation at the shock location. internal flows, ISAIF8-0051, Lyon, July 2007.
Haase, W., et al., eds., 2006. FLOMANIA - a European
initiative on flow physics modelling. Notes on numerical
Conclusions fluid mechanics and multidisciplinary design (NNFM),
Vol. 94 XI. Berlin: Springer Verlag.
RANS simulations are the current reference for Hirsch, C., 2007. Numerical computation of internal and
industrial CFD simulations in complex configurations, external flows: the fundamentals of computational fluid
and although many valuable results are obtained, dynamics. Elsevier Butterworth Heinemann, Hardbound,
including applications towards CFD-based shape 680 pages.
optimisation, severe deficiencies have been identified Iaccarino, G., 2000. Prediction of the turbulent flow in a
diffuser with commercial CFD codes. Center for turbu-
with all of the current turbulence models. lence research. Annual research briefs, 271–278.
The most significant ones are related to the Kaltenbach, H.J., et al., 1999. Study of the flow in a planar
inaccurate prediction of 2D and 3D separation and asymmetric diffuser using large eddy simulations. Journal
to an insufficient incorporation of mixing effects, as of Fluid Mechanics, 390, 150–185.
shown on the presented geometrical simple test cases. Ladeinde, F., et al., 2006. The first high-order CFD
simulation of aircraft: challenges and opportunities. In:
It is expected that the accumulation of DNS and LES 44th AIAA Aerospace Sciences Meeting and Exhibition.
databases will allow significant improvements in the Reno, NV, USA. AIAA-Paper-2006–1526.
near future. Lien, F.S. and Kalitzin, G., 2001. Computations of transonic
flow with the v2-f turbulence model. International Journal
of Heat and Fluid Flow, 22, 53–61.
References Menter, F.R., 1994. Two-equations eddy-viscosity turbu-
Bruce, P.J.K. and Babinsky, H., 2008. Unsteady shock wave lence models for engineering applications. AIAA Journal,
dynamics. Journal of Fluid Mechanics, 603, 463–473. 32, 1598–1605.
Buice, C.U. and Eaton, J.L., 1997. Experimental investiga- Obi, S., Aoki, K., and Masuda, S., 1993. Experimental
tion of flow through an asymmetric plane diffuser. and computational study of turbulent separating
Report No.TSD-107, Thermosciences Division, Depart- flow in an asymmetric plane diffuser. In: Ninth Sympo-
ment of Mechanical Engineering, Stanford University, sium on turbulent shear flows, 16–19 August, Kyoto,
Stanford, CA, USA. Japan, 305.
International Journal of Computational Fluid Dynamics 303
Spalart, P.R. and Allmaras, S.R., 1992. A one-equation Wolkov, A., Hirsch, C., and Leonard, B., 2007. Discontin-
turbulence model for aerodynamic flows. AIAA Paper, uous Galerkin method on unstructured hexahedral grids
92–0439. for 3D Euler and Navier-Stokes equations. In: 18th
Venkatakrishnan, V., et al., 2003. Higher order schemes for AIAA CFD conference, AIAA paper 2007–4078.
the compressible Navier-Stokes equations. In: 16th AIAA Yang, Z. and Shih, T.H., 1993. A k-e model for turbulence
computational fluid dynamics conference, 23–26 June, and transitional boundary layer. In: R.M.C. So, ed. Near
Orlando, FL. AIAA Paper 2003–3987. wall turbulent flows, 165–175.
Wilcox, D.C., 1988. Reassessment of the scale-determining
equation for advected turbulence model. AIAA Journal,
26, 1299–1310.