Professional Documents
Culture Documents
Simulations
Pierre Sagaut
Institut Jean Le Rond dAlembert, Universite Pierre et Marie Curie - Paris 6,
4 place Jussieu -case 162, F-75252 Paris cedex 5, France
pierre.sagaut@upmc.fr
1 Introduction
Real-life CFD applications involve a large number of choices that are to be
made in order to setup the computational congurations of the problem.
These choices include the physical models and related arbitrary constants,
boundary conditions, initial conditions as well as tuning parameters in the
numerical methods. In many cases, all physical and geometrical parameters
are not exactly known and in some instances are not known at all. This lack of
information raises the issue of taking into account uncertainties in the CFD
process, modeling it and measuring the dependency/sensitivity of the results
with respect to these uncertainties. Consequently, this approach implies that
instead of seeking a single deterministic solution, we are now interested in
recovering a continuous description of the space of possible solutions spanned
by uncertain parameters.
This article aims at providing a survey of recent progress made in the
eld of uncertainty and error quantication and propagation in CFD, the
emphasis being put on Large-Eddy Simulation (LES) of turbulent ows. All
issues addressed below are also relevant to Reynolds-Averaged Numerical
Simulations (RANS) and hybrid RANS/LES methods. Section 2 illustrates
why, even in a simple academic turbulent ow, arbitrary parameters that appear in turbulence models must be considered as uncertain parameters, since
their values are ow-dependent. The concept of robust model, i.e. whose
sensitivity to the tuning of arbitrary parameters is minimal, is then presented in Section 3. The issue of representing the space of solutions spanned
by possible variations of the computational setup parameters is then addressed. Section 4 rst illustrates the use of the generalized Polynomial Chaos
(gPC) [Xiu and Karniadakis (2002), Xiu and Karniadakis (2003)] method on
an academic case and then exemplies the use of the Kriging method
[Krige (1951)] on an engineering problem.
20
P. Sagaut
4/3
C 3/4
C(L/, ReL ) =
1
,
(1)
C
where the spectrum shape is dened as
E(k) = K0 2/3 k 5/3 fL (kL)f (k),
(2)
+
3/4
4
x1/3 G2 (x)dx
.
3 0
(4)
21
0.2
Cs
0.15
0.1
0.05
0
0
10
10
/
10
Fig. 1. Smagorinsky constant as a function of the ratio between cuto length and
Kolmogorov scale. Solid line: exact value; dotted line: asymptotic value; dashed
line: linear and quadratic remapping.
22
P. Sagaut
G
(k,
t))dk
dt
0
LES
DN S
0
p (N, C) =
.
(6)
2
T kc
p G2 (k)E
k
(k,
t))dk
dt
DN
S
0
0
Fig. 2. Error for the Smagorinsky model as a function of the number of grid points
N and the value of the Smagorinsky constant. Shadded regions denote optimality
regions for the mode constant, in which the committed error is within 20 percents
of the minimum possible error on the same grid. Symbols: square: p= -1; circle:
p=0; right-pointing triangle: p=1; left-pointing triangle: p=2.
23
Choosing p=-1,0 or 2, one can put the emphasis on the error committed
on the integral lengthscale, the resolved kinetic energy and the resolved enstrophy, respectively. Since in practical applications one can be interested in
predicting several parameters at the same time with satisfactory accuracy, it
is convenient to introduce the following multiobjective error measure
(N,
C)/
(N,
C(p,
N
)
p
p
p=1,0,1,2
,
(7)
(N, C) =
24
P. Sagaut
Fig. 4. Uncertainty error bars on the 643 grid resolved turbulent kinetic energy
spectrum computed using a Smagorinsky model with uncertain constant: envelope
of possible solutions
25
3000
2500
24
22
20
18
16
14
12
10
8
6
4
2
0
1500
2000
1000
500
0
0
0.005
0.01
0.015
0.02
0.025
E(k)
0.03
0.035
0.04
0.045
Fig. 5. pdf of the turbulent kinetic energy E(k) in DHIT with uncertain Smagorinsky constant following a beta distribution
Fig. 6. Flow conguration for the hot jet exhaust of an aeronautical engine cooling
system
that are optimal for the representation of the random inputs to the problem.
Using this approach, a continuous reconstruction of the space of possible solutions is obtained using a restricted set of usual LES realizations. The analysis
of decaying turbulence with an uncertain Smagorinsky model has been performed. We thus treat the Smagorinsky constant as an uncertain input to the
stochastic problem, and its probability distribution is assumed. In order to
26
P. Sagaut
Maximum and minimum values
Exp 50037
Standard LES
3
X/D=1
2.5
2
Z/D
Z/D
1.5
1.5
0.5
0.5
-0.5
0.5
U/U0
0
0.6
1.5
0.7
0.8
0.9
U/U0
1.1
2.5
2.5
Z/D
Z/D
1.5
1.5
0.5
0.5
-0.1
0.1
0.2
V/U0
-0.05
V/U0
0.05
0.1
Z/D
3
2.5
Z/D
3
2.5
1.5
1.5
0.5
0.5
X/D=8
2.5
0.1
W/U0
0.2
0.3
0
-0.2
-0.1
W/U0
0.1
27
X/D=8
0.6
0.6
0.5
0.5
0.4
0.4
0.7
0.7
0.3
0.3
0.2
0.2
0.1
0.1
0.02
0.04
0.06
0.02
Y=0
1
0.9
Exp 50037
Maximum and m
Standard LES
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
functions, showing the great complexity of the LES solution response to random variability in the Smagorinsky constant.
We now illustrate the use of the response surface approach on a practical engineering problem, namely the separated ow at the exhaust of
the cooling system in a aeronautical engine (see gure 6). Most numerical
methods used for practical engineering purposes involve articial dissipation. However, it is known that numerical dissipation and subgrid/turbulence
model induced dissipation are in competition. They must be tuned in an
ad hoc manner to adapt to the case in order to recover the best possible results (e.g. [Garnier et al.(2001), Ciardi et al. (2005)]). In the following
example [Jouhaud et al. (2008)], it is chosen to retain both the Smagorinsky constant CS and the articial fourth-order dissipation parameter smu4
which appears in Jamesons scheme as uncertain optimization parameters. A
variability range of 30% around the standard values Cs = 0.18 and
smu4 = 0.01 is considered. The sensitivity of the mean ow with respect
to these two parameters is illustrated in Figs. 7 and 8, which compare the
standard LES solution (i.e. the LES solution computed using the standard
parameters), experimental data, and the extreme values retrieved from the
response surface built for the mean ow solution at every grid point. The sensitivity of the solution is directly related to the dierences between the two
28
P. Sagaut
5 Concluding Remarks
The present paper aimed at presenting recent results dealing with unresolved
scale modelling for unsteady simulation of turbulent ows. The emphasis was
put on LES, but proposed analysis tools can be extended to hybrid RANSLES method in a straightforward manner. In a rst step, the concept of
inertial range consistency was introduced. Let us notice that this non-trivial
issue is of direct interest for practical application, since it was shown that the
usual value of the Smagorinsky constant can be used if and only if L/ >
20 30 and that / > 100, i.e. for coarse grids in ows such that L/ >
2000 3000. Since similar criteria may be dened for hybrid approaches,
one may wonder if model tuning on test cases with nearly-innite Reynolds
number is of practical interest for VLES models which involve several tuning
parameters.
A second step is to design robust models, i.e. models which will lead
to very good results even if some parameters of the simulation (grid resolution, ) are changed. Using the error map approach and the concept
of inertial-range consistency, it will be shown that robust, nearly-optimal
subgrid models can be designed, which satisfy the three basic modelling
constraints: 1/ physical constistency 2/ robustness 3/ numerical viability
[Meyers and Sagaut (2006), Meyers et al. (2006)]. In a third step, a deeper
insight in the solution sensitivity was gained using the gPC approach. One
may wonder if DHIT is a relevant test case for model validation, since it is
a very simple turbulent ow. Let us rst remark that theoretical analysis
needs relevant, unambiguous test cases. A last point is therefore the reliability of some famous test cases for LES-like model validation, such as the plane
channel ow. It was recently shown that the error associated with coarse-grid
plane channel DNS exhibits a complex non-linear behavior, which can lead
to misleading interpretations of the results in [Meyers and Sagaut (2007)].
These authors show that coarse-grid DNS may, in some cases, lead to exact prediction of usual test parameters such as skin friction, mean centreline
velocity and peak of streamwise turbulence intensity.
29
The last point dealt with the representation of the uncertainty in the
solution which originates in the lack of knowledge of complex ows. The
response surface approach is observed to be an ecient tool for that purpose.
Response surfaces can be used toinvestigate the sensitivity of the solution,
but also to draw error maps at lowcost is some reference data are available.
Acknowledgements. The author would like to warmly acknowledge the
fruitful and enlightning collaborations which lead to the results presented
here. The error map approach and the robust modelling theory were developped in close collaboration with Dr. Johann Meyers. The Kriging-based
results were obatined with Dr. Jean-Christophe Jouhaud. Dr. Didier Lucor
was the main contributor to the development of the gPC-based researches
presented here.
References
[Ciardi et al. (2005)] Ciardi, M., Sagaut, P., Klein, M., Dawes, W.N.: A dynamic
nite-volume scheme for large-eddy simulation on unstructured grids. Journal
of Computational Physics 210, 632655 (2005)
[Garnier et al.(2001)] Garnier, E., Sagaut, P., Deville, M.: A class of explicit
ENO lters with application to unsteady ows. Journal of Computational
Physics 170, 184204 (2001)
[Jouhaud et al. (2006)] Jouhaud, J.C., Sagaut, P., Labeyrie, B.: A Kriging approach for CFD/wind tunnel data comparison. Journal of Fluids Engineering 128(4), 847855 (2006)
[Jouhaud et al. (2008)] Jouhaud, J.C., Sagaut, P., Eneaux, C., Laurenceau, J.: Sensitivity analysis and multi-objective optimization for LES numerical parameters. Journal of Fluids Engineering 130, 021401 (2008)
[Krige (1951)] Krige, D.G.: A statistical approach to some basic mine valuations
problems on the Witwatersrand. Journal of Chemical, Metallurgy and Mining
Society of South Africa 52, 119139 (1951)
[Lilly (1967)] Lilly, D.K.: The representation of small-scale turbulence in numerical simulation experiments. In: Proceedings of the IBM Scientic Computing
Symposium on Environmental Sciences, Yorktown Heights, USA (1967)
[Lu and Sagaut (2007)] Lu, S.Y., Sagaut, P.: Direct sensitivity analysis for smooth
unsteady compressible ows using complex dierentiation. International Journal for Numerical Methods in Fluids 53(12), 18631886 (2007)
[Lucor et al. (2007)] Lucor, D., Meyers, J., Sagaut, P.: Sensitivity analysis of LES
to subgrid-scale-model parametric uncertainty using polynomial chaos. Journal
of Fluid Mechanics 585, 255279 (2007)
[Meyers and Sagaut (2006)] Meyers, J., Sagaut, P.: On the model coecient for the
standard and the variational multiscale Smagorinsky model. Journal of Fluid
Mechanics 569, 287319 (2006)
[Meyers et al. (2006)] Meyers, J., Sagaut, P., Geurts, B.J.: Optimal model parameters for multi-objective large-eddy simulations. Physics of Fluids 18, 095103
(2006)
30
P. Sagaut
[Meyers and Sagaut (2007)] Meyers, J., Sagaut, P.: Is plane channel ow a friendly
test-case for the testing of LES subgrid scale models? Physics of Fluids 19,
048105 (2007)
[Porte-Agel et al. (2000)] Porte-Agel, F., Meneveau, C., Parlange, M.: A scaledependent dynamic model for large-eddy simulation: application to a neutral
atmospheric boundary layer. Journal of Fluid Mechanics 415, 261284 (2000)
[Xiu and Karniadakis (2002)] Xiu, D., Karniadakis, G.E.: The Wiener-Askey Polynomial Chaos for stochastic dierential equations. SIAM Journal of Scienitifc
Computing 24, 137167 (2002)
[Xiu and Karniadakis (2003)] Xiu, D., Karniadakis, G.E.: Modeling uncertainty in
ow simulations vi generalized Polynomial Chaos. Journal of Computational
Physics 187, 137167 (2003)