Professional Documents
Culture Documents
This work describes a new method for building a statistical model for the structural
uncertainties in the k − ω turbulence model. An inverse RANS problem is solved using the
adjoint method to determine the turbulent viscosity that produces the flow field closest to
that predicted by direct numerical simulation. We describe the difference in the turbulent
viscosity field inferred by the inverse RANS problem and turbulent viscosity field predicted
by RANS as a Gaussian random field, and develop a statistical model of this random field
using maximum likelihood estimation. The resulting statistical model is used to propagate
uncertainty to engineering quantities of interest using non-intrusive techniques. Results
for turbulent flow in a periodic straight walled channel are presented and analyzed.
Nomenclature
J Objective function
u RANS velocity field
UDN S DNS mean velocity field
νT Turbulent viscosity
ν Laminar viscosity
p RANS pressure field
δu Velocity perturbation field
δp Pressure perturbation field
δνT Turbulent viscosity perturbation field
δν Laminar viscosity perturbation field
δJ Objective function perturbation
û Adjoint velocity field
p̂ Adjoint pressure field
L Likelihood function
σ Process variance
λ Correlation length
uτ Friction velocity
δ Channel half-width
νT,eff DNS effective turbulent viscosity
I. Introduction
When the effects of turbulence on an engineering system’s performance must be determined, the most
popular approach is to solve the Reynolds averaged Navier-Stokes (RANS) equations, which employ RANS
models to estimate the effects of turbulence. Since RANS models seek to compute only the statistically
averaged flow field, relatively coarse meshes and large timesteps can be used as compared to higher fidelity
methods such as large eddy simulation or direct numerical simulation (DNS). The reduced computational
∗ PhD student, Department of Aeronautics and Astronautics, Room 37-442, MIT, Cambridge, MA 02139, AIAA Student
Member
† Professor, Department of Aeronautics and Astronautics, Room 37-408, MIT, Cambridge, MA 02139, AIAA Member
1 of 12
American
Copyright © 2011 by the American Institute of Aeronautics and Institute
Astronautics, Inc. All of Aeronautics
rights reserved. and Astronautics
effort required to solve the RANS equations has made them a popular choice for computing flows where the
effects of turbulence must be included, but a limited amount of computational power is available. Despite
this popularity, the flow field computed using RANS models can exhibit significant uncertainty, which in turn
leads to uncertainty in the quantities of interest.1 These uncertainties are referred to as model or structural
uncertainties. In engineering applications, this uncertainty can greatly complicate the design process, since
it may be difficult to determine whether a perceived increase in performance is due to uncertainty in the
computed quantities of interest. This motivates the need for a priori estimates of the structural uncertainties
in RANS models.
Due to the importance of RANS models in industry, many attempts have been made to quantify the
structural uncertainties in RANS simulations. The work of Platteeuw et. al. uses a collection of experimental
results and direct numerical simulations to determine the distributions of the closure coefficients of the k −
model. These uncertainties are then propagated using the Probabilistic Collocation Method.2 The emphasis
of this work is on the efficient propagation of uncertainty rather than the characterization of the sources of
uncertainty. For example, the assumed distributions of some parameters must be guessed, due to a lack of
available experimental data. The work of Oliver and Moser focuses more on the characterization of these
sources of uncertainty.3 In this work, the closure coefficients of the the Spalart-Allmaras RANS model are
treated as random variables. The probability distributions of these tuning parameters are calculated by
solving a Bayesian inverse problem based on experimental calibration data. Gaussian noise is added to
the random model parameters on the state level, and the magnitude and correlation length of this noise is
estimated by solving a separate Bayesian inverse problem. Their results agree well with experimental data,
but their method is restricted to simple flows since only a small set of parameters is included in the inverse
model.
Other efforts have sought to improve the accuracy of RANS models by comparing to direct numerical
simulation. Since DNS resolves all of the relevant scales of turbulent motion, the results are extremely high
fidelity, and have thus been used to determine the accuracy of turbulence models. Some recent examples
include the work of Venayagamoorthy et. al., where the results of direct numerical simulation are used to
develop trends for the various tuning parameters of the k − model for stratified flows.4 They note that the
DNS results do not always present clear trends, and that it may be up to the modeler to choose the trend
they feel most appropriate. Kim et al. provide a detailed comparison between the results of DNS with a
variety of RANS models for turbulent mixed convection. They conclude that some models are superior in
capturing the effects of buoyancy, and that the performance of these models is highly sensitive to the choice
of tuning parameters. Comparisons like these shed significant light on the uncertainties in RANS models,
but typically must be performed on a case by case basis.
We propose a predictive method for quantifying the structural uncertainties in RANS simulations which
does not rely upon calibration data from experimental measurements. Instead, the structural uncertainties
are estimated by comparing the results of RANS simulations to the results of direct numerical simulations.
We use adjoint based inverse modeling to compute the “true” turbulent viscosity field from ensemble averaged
DNS results. Compared to Bayesian inversion, adjoint based inversion is significantly more computationally
efficient for such large scale problems. A statistical model for the discrepancy between the true turbulent
viscosity and that computed using RANS is constructed using maximum likelihood estimation. Rather than
treating the model input parameters as random variables, we model the discrepancy as a Gaussian random
field, and use maximum likelihood estimation to determine the covariance function that is most likely to have
generated the observed results. This statistical model for the structural uncertainties is used to determine
the uncertainty in engineering quantities of interest.
It is important to note that, while this work focuses on quantifying uncertainties in the k − ω turbu-
lence model, our approach is entirely generalizable to any eddy viscosity model, both linear and nonlinear.
Throughout the process of constructing the statistical model, we only consider the output of the turbulence
model, namely the turbulent viscosity field. Since all eddy viscosity models use the turbulent viscosity to
relate the Reynolds stresses to the mean rate of strain, our approach applies independent of the method by
which the turbulent viscosity field is computed. We have chosen to focus on the Wilcox k − ω turbulence
model due to its popularity in industry and its ease of use.6
Section II describes our approach for modeling the structural uncertainties in RANS turbulence models.
Section III presents the results of applying our methodology to model the structural uncertainties in turbulent
flow through a straight walled channel. Section IV describes future extensions of this work.
2 of 12
In the optimization iterations, the turbulent viscosity is decoupled from the transport scalars and is treated
as a parameter to be optimized. In each iteration, we solve the mean flow equations with an additional
prescribed turbulent viscosity. This optimization problem is ill-posed due to the existence of regions where
the mean velocity gradient is zero, e.g. at the centerline of an axisymmetric flow field. In these regions, the
solution is insensitive to variations in the turbulent viscosity, and the problem becomes ill-posed. We adopt
a regularization procedure whereby an additional regularization term is added to the objective function
to ensure that the problem remains well-posed. The regularization term used in this work measures the
smoothness of the turbulent viscosity field. With the addition of the regularization term, the optimization
problem then takes the form:
min ||u(νT ) − UDN S ||2L2 + ε||∇νT ||2L2 s.t. νT ≥ 0. (3)
νT
The regularization scaling parameter ε is chosen to be small relative to the geometric length scales of the
problem to ensure that the optimized turbulent viscosity field is not overly smoothed.
3 of 12
∇ · δu = 0. (6)
Note that the density is taken to be unity everywhere. We introduce adjoint variables û and p̂, corresponding
to the adjoint velocity and adjoint pressure, respectively. The first variation of the objective function can
be written as
Z
δJ = 2(u − UDN S ) · δu. (7)
Ω
The regularization term is not incorporated directly into the adjoint equations, since the sensitivity of the
regularization term with respect to the turbulent viscosity is computed directly. Dotting the adjoint velocity
û into equation (5), and adding the product of the adjoint pressure p̂ to equation (6) gives
Integrating by parts over the domain Ω and adding the result to equation (7) gives
Z Z
δJ = 2(u − UDN S ) · δu + (δν + δνT )∇u : ∇û
Ω Ω
Z
+ δu · (−u · ∇û + ∇u · û − ν∇2 û + ∇p̂) − δp∇ · û
Ω
Z
+ δp(û · ~n − ν((∇δu) · û) · ~n − δν((∇u) · û) · ~n. (9)
∂Ω
Finally, we set all terms involving δu or δp to zero to obtain equations (10) and (11).
4 of 12
where yi and yj are spatial coordinates. The parameters σ and λ are not known a priori, but must be
determined using statistical analysis. The squared exponential covariance function represents the belief that
the log-discrepancy varies smoothly in space.
To estimate the parameters of the covariance function, we employ maximum likelihood estimation (MLE).
This approach seeks to determine the set of parameters that is most likely to have generated the observed
turbulent viscosity discrepancy. Since we model the discrepancy as a Gaussian random field, the probability
density function of the discrepancy is described by a zero mean multivariate Gaussian, that is:
1 1 T −1
fX (x|σ, λ) = exp − x Σ(σ, λ) x , (15)
(2π)k/2 |Σ(σ, λ)|1/2 2
where Σ(σ, λ) is the covariance matrix, and k is the dimension of the random vector of discrepancies X,
i.e. the number of nodes in the mesh. The likelihood function L can be thought of as the unnormalized
probability distribution of the parameter set taking particular values, conditioned on the observed data x,
and is computed directly from the conditional probability fX (x|σ, λ):10
Here, x is the observed turbulent viscosity log-discrepancy field. To determine the parameter set (σ, λ)
that is most likely to have generated the realized discrepancy field, we determine the parameter set that
maximizes the likelihood function. For computational convenience, we maximize the log-likelihood function
log(L), which is obviously monotonically related to the likelihood function.
5 of 12
where the (λi , xi (y)) are eigenvalue/eigenvector pairs of the covariance matrix, and φi (θ) ∼ N (0, 1) are
independent normally distributed random variables.11,12 The number of K-L modes NK-L used to construct
the K-L expansion depends on the decay rate of the λi , and usually is much smaller than dim (y). Since the
log-discrepancy typically varies smoothly in space, we can approximate the full K-L expansion quite well
with very small NK-L . To generate a sample, we sample the φi (θ) from the standard normal, and reconstruct
X(y, θ) from the K-L expansion.
III. Results
In this section, we present results for turbulent flow through a straight walled channel. For this test
case, we perform the inverse modeling and statistical inference steps for flow at a specified friction Reynolds
number, and propagate these results to determine the uncertainty in the flow solution at higher friction
Reynolds numbers.
6 of 12
0
10
-1
10
-2
10
J
-3
10
-4
10
-5
10
-6
10
0 20 40 60 80 100
Iteration
Figure 2 shows the results of the optimization procedure for the straight walled channel. The objective
function value decreases from an initial value J = 6.3127 × 10−1 to J = 4.6796 × 10−6 after 100 optimization
iterations. The path taken by the L-BFGS algorithm is shown in figure 1. The initial velocity profile predicted
by the Wilcox k − ω model is lower everywhere except very close to the wall in the log law region, with
a maximum relative error of approximately 10%. The optimized velocity profile matches the DNS velocity
profile very well, with a maximum relative error of approximately 1%. The figure on the right depicts the
initial and optimized turbulent viscosity profile. The DNS viscosity profile represents the effective turbulent
viscosity computed using a simple force balance relation:
−1
1 ∂UDN S
νT,eff = , (21)
(1 − y/δ) ∂y
where the velocity gradient values have been provided in the DNS database. We observe that the optimized
turbulent viscosity profile is nearly identical to the DNS effective turbulent viscosity, even near the channel
centerline where the solution is relatively insensitive to changes in the turbulent viscosity.
It is important to note the importance of the regularization term for this problem. Equation (20) and
the homogeneous Neumann boundary condition enforced at y/δ = 1, which arises due to the symmetry of
the problem, imply that the sensitivity gradient of J at the channel centerline is identically zero. Physically,
this agrees with the intuition that changing the viscosity in regions where the velocity gradient is zero does
not affect the resulting flow field. This means that the optimization routine will never change the value of
the turbulent viscosity at y/δ = 1, and the resulting optimization problem is ill-posed. This ill-posedness
manifests itself in the form of oscillations in the optimized turbulent viscosity profile near the channel
centerline. The plot shown at the bottom of figure 2 demonstrates this issue. The optimized turbulent
viscosity profile shows good agreement until y/δ = 0.4, where oscillations appear and grow up to y/δ = 1.0.
Since the velocity gradient is small in the region 0.4 < y/δ < 1.0, the oscillations in the viscosity field do not
significantly affect the computed velocity profile. However, since we ultimately model the discrepancy in the
turbulent viscosity field, these oscillations will impact our statistical model. The regularization term remedies
this issue by introducing a nonzero gradient at y/δ = 1. To determine the proper value of the regularization
parameter, the value of ε was increased until significant improvement was made in the agreement between
the DNS effective and RANS optimized viscosity fields after 100 optimization steps. Ultimately, a value of
ε = 1.0 × 10−4 was selected. As seen in figure 1, most of the change in the objective function J is made
during the first fifty optimization iterations, where the magnitude of J is much larger than ε. Once the DNS
and RANS velocity profiles match and J is small compared to ε, the regularization term becomes dominant,
and further iterations damp the oscillations in the viscosity field.
7 of 12
0.06
10
T
u
0.04
5
0.02
00.0 0.2 0.4 0.6 0.8 1.0 0.000.0 0.2 0.4 0.6 0.8 1.0
y/ y/
(a) Velocity profile (b) Viscosity profile with regularization
0.10
Initial
DNS
0.08
Optimized
0.06
T
0.04
0.02
Figure 2. Initial and optimized velocity and viscosity profiles compared to DNS results.
8 of 12
0.6
0.4
)
0.2
k
T
log( 0.0
T )
log(
0.2
0.4
0.6
8 7 6 5 4 3 2 1 0
log(y/)
in figure 3, we determine the set of parameters that maximizes the log-likelihood function. The log-likelihood
function is computed as
N
(X T vΣi 2
1X i )
log(L) = − log(σΣ )− , (22)
2 i=1 λi
i i
where σΣ and vΣ are the singular values and singular vectors of the covariance matrix, respectively. Clearly,
if any of the singular values of Σ are zero, the value of log(L) is not well-defined. To address this issue, we
assume that a small error e has been made in the estimation of the true turbulent viscosity field, so that the
log-discrepancy is actually given by
! !
νT? + e νT? e
X = log k−ω
≈ log k−ω
+ ?. (23)
νT νT νT
In computing the log-likelihood function, we add (e/νT? )2 to the diagonal of the covariance matrix Σ, since
the error term relates to the variance of the Gaussian field. The value of e is chosen to be small relative to
the largest singular value of Σ. In this work, we have chosen e = 10−6 . Decreasing e below this value does
not change the estimated parameter set.
In general, the log-likelihood function is nonlinear in the parameter set. In that case, determining the
parameter set that maximizes the log-likelihood requires some sort of gradient-free optimization method.
For this work, since the dimension of the parameter set is small, we simply plot the log-likelihood function
for a large number of parameter sets and observe where the maximum value occurs. Figure 4 shows a plot
of the log-likelihood function as a function of the parameter set (σ, λ). We note that the parameter set
(σ, λ) = (0.1898, 0.1532) maximizes the log-likelihood function, and this set is used in the statistical model.
C. Uncertainty propagation
For each friction Reynolds number considered, 500 Monte Carlo simulations were performed to propagate
the uncertainty. Sample turbulent viscosity profiles are generated by sampling from the Gaussian random
field with the parameter set determined using MLE. Figure 6 shows five sample turbulent viscosity profiles
and the corresponding sample velocity profiles for flow at Reτ = 180. We observe that the turbulent sample
viscosity fields vary smoothly in space. Figure 5 shows the mean and variance of the computed samples. The
solid blue line represents the mean velocity profile computed from the Monte Carlo samples. We note that
9 of 12
100
0.14
0.12 50
0.1 0
log(L)
λ
0.08 −50
0.06 −100
0.04 −150
0.02 −200
Figure 4. Contours of log-likelihood function, showing maximum value at (σ, λ) = (0.1898, 0.1532).
the DNS velocity profile mostly falls within the 2σ error bars (the shaded pink regions). The error bars grow
larger towards the channel centerline, reflecting the fact that the level of uncertainty in the velocity profile
near the wall is small relative to the uncertainty near the centerline. This agrees with the results presented
in figure 2, which show that the velocity discrepancy between the RANS and DNS solution is small very
near the wall, and remains nearly constant outside of this region.
Acknowledgments
The authors would like to thank Professor Steven Johnson at MIT for his helpful suggestions regarding
the NLopt code. This work was funded by Pratt and Whitney and a subcontract of the DOE Predictive
Science Academic Alliance Program (PSAAP) from Stanford to MIT.
References
1 Revell, A., Iaccarino, G., and Wu, X., “Advanced RANS Modeling of Wingtip Vortex Flows,” Annual Research Briefs,
Center for Turbulence Research, NASA-AMES, 2006, pp. 73-85.
2 Platteeuw, P.D.A., Loeven G.J.A., and Bijl H., “Uncertainty Quantification Applied to the k − Model of Turbulence
Using the Probabilistic Collocation Method,” 49th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and
Materials Conference, 2008, AIAA Paper 2008-2015.
3 Oliver, T., and Moser., R., “Uncertainty Quantification for RANS Turbulence Model Predictions,” American Physical
10 of 12
20 20
15 15
u
u
10 10
5 +2σ 5 +2σ
−2σ −2σ
Mean Mean
DNS DNS
00.0 0.2 0.4 0.6 0.8 1.0 00.0 0.2 0.4 0.6 0.8 1.0
y/δ y/δ
(a) Reτ = 180 (b) Reτ = 395
25
20
15
u
10
5 +2σ
−2σ
Mean
DNS
00.0 0.2 0.4 0.6 0.8 1.0
y/δ
(c) Reτ = 590
Figure 5. Monte Carlo simulation results for three friction Reynolds numbers.
0.12 20
0.10
15
0.08
T
u
0.06 10
0.04
0.02
0.00 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
y/ y/
(a) Turbulent viscosity (b) Velocity
Figure 6. Realizations of turbulent viscosity and velocity from Monte Carlo simulation at Reτ = 180.
11 of 12
Flows Based on DNS Data,” Annual Research Briefs, Center for Turbulence Research, NASA-AMES, 2003, pp. 127-138.
5 Kim, W.S., He, S., and Jackson, J.D., “Assessment by Comparison with DNS DATA of Turbulence Models Used in
Simulations of Mixed Convection,” International Journal of Heat and Mass Transfer, Vol. 51, 2008, pp. 1293-1312.
6 Wilcox, D. C., Turbulence Modeling for CFD, Griffin Printing, Glendale, CA, 1993.
7 Nocedal, J., “Updating Quasi-Newton Matrices with Limited Storage,” Mathematics of Computation, Vol. 35, 1980, pp.
773-782.
8 Johnson, S. G., “The NLopt nonlinear-optimization package,” http://ab-initio.mit.edu/nlopt.
9 Jameson, A., “Aerodynamic Shape Optimization Using the Adjoint Method,” Von Karman Institute Lecture Series
90-100.
11 Le Maı̂tre, O.P., and Knio O.M., Spectral Methods for Uncertainty Quantification: With Applications to Computational
Manufacturing Error of Transonic Airfoil,” 49th AIAA Aerospace Sciences Meeting, 2011, AIAA Paper 2011-658.
13 Moser, R. D., Kim, J., and Mansour, N. N., “Direct Numerical Simulation of Turbulent Channel Flow up to Re = 590,”
τ
Physics of Fluids, Vol. 11, No. 4, 1999, pp. 943-945.
12 of 12