Title: PhD%20Preliminary%20Oral%20Exam%20CHARACTERIZATION%20AND%20PREDICTION%20OF%20CFD%20SIMULATION%20UNCERTAINITIES
1PhD Preliminary Oral ExamCHARACTERIZATION AND
PREDICTION OF CFD SIMULATION UNCERTAINITIES
- by
- Serhat Hosder
- Chair Dr. Bernard Grossman
- Committee Members
- Dr. Raphael T. Haftka Dr. William H. Mason
- Dr. Reece Neel Dr. Rimon Arieli
- Department of Aerospace and Ocean Engineering
- Virginia Tech.
- Blacksburg, VA
-
-
2Outline of the Presentation
- Introduction
- Classification of CFD Simulation Uncertainties
- Objective of the Present Work
- Previous Studies
- Transonic Diffuser Case
- Results, findings and discussion about different
sources of uncertainty - Conclusions
3Introduction (1)
- The Computational Fluid Dynamics (CFD) as an
aero/hydrodynamic analysis and design tool - Increasingly being used in multidisciplinary
design and optimization (MDO) problems - Different levels of fidelity (from linear
potential solvers to RANS codes) - CFD results have a certain level of uncertainty
originating from different sources - Sources and magnitudes of the uncertainty
important to assess the accuracy of the results
4Introduction (2)
Drag Polar Results for DLR F-4 Wing at M0.75,
Rec3x106 (taken from 1st AIAA Drag Prediction
Workshop (DPW), Ref. 1)
5Classification of CFD Simulation Uncertainties
- Physical Modeling Uncertainty
- PDEs describing the flow (Euler, Thin-Layer N-S,
Full N-S, etc.) - Boundary and initial conditions (B.C and I.C)
- Auxiliary physical models (turbulence models,
thermodynamic models, etc.) - Uncertainty due to Discretization Error
- Numerical replacement of PDEs and continuum B.C
with algebraic equations - Consistency and Stability of PDEs
- Spatial (grid) and temporal resolution
- Uncertainty due to Iterative Convergence Error
6Definition of Uncertainty and Error
- Oberkampf and Trucano (Ref. 2) defined
-
- Uncertainty as a potential deficiency in any
phase or activity of modeling process that is due
to the lack of knowledge (uncertainty of
turbulence models, geometric dimensions,
thermo-physical parameters, etc.) -
- Error as a recognizable deficiency in any phase
or activity of modeling and simulation - Discretization errors can be estimated with
certain methods by providing certain conditions - In this work, well refer the inaccuracy in the
CFD simulations due different sources as
uncertainty -
-
7Objective of the Present Work
- Characterize different sources of CFD simulation
uncertainties - Consider different test cases
- Apply different grids, solution
schemes/parameters, and physical models - Try to quantify/predict the magnitude and the
relative importance of each uncertainty - Compare the magnitudes of CFD simulation
uncertainties with other sources of uncertainty
(geometric uncertainty, uncertainty in flow
parameters, etc.)
8Previous Studies
- Previous CFD related studies mainly focused on
discretization and iterative convergence error
estimations - Grid Convergence Index (GCI) by Roache (Ref. 3)
- Discretization Error of Mixed-Order Schemes by C.
D. Roy (Ref. 4) - Trucano and Hill (Ref. 5) proposed statistical
based validation metrics for Engineering and
Scientific Models - Hemsch (Ref. 6) performed statistical analysis of
CFD solutions from 1st AIAA DPW. - Kim (Ref. 7) made statistical modeling of
simulation errors (from poorly converged
optimization runs) and their reduction via
response surface techniques
9Description of Transonic Diffuser Test Case (1)
- Known as Sajben Transonic Diffuser case in CFD
Validation studies - Top wall described by an analytical equation
- Although geometry is simple, the flow-field is
complex. - The Shock strength and the location determined
by exit-pressure-to-inlet-total pressure ratio
Pe/P0i - Pe/P0i0.72 (Strong shock case), Pe/P0i0.82
(Weak shock case),
10Description of Transonic Diffuser Test Case (2)
Mach contours for the weak shock case
Mach contours for the strong shock case
11Simulation Code, Solution Parameters, and Grids
(1)
- General Aerodynamic Simulation Program (GASP)
- 3-D, structured, multi-block, finite-volume, RANS
code - Inviscid fluxes calculated by upwind-biased 3rd
(nominal) order spatially accurate Roe-flux
scheme - All viscous terms were modeled (full N-S)
- Implicit time integration to reach steady-state
solution with Gauss-Seidel algorithm
12Simulation Code, Solution Parameters, and Grids
(2)
- Flux-Limiters
- Van Albadas limiter
- Min-Mod limiter
- Turbulence Models
- Spalart-Allmaras (Sp-Al)
- k-w (Wilcox, 1998 version)
- Grids Generated by an algebraic mesh generator
- Grid 1 (g1) 41x26x2
- Grid 2 (g2) 81x51x2
- Grid 3 (g3) 161x101x2
- Grid 4 (g4) 321x201x2
- Grid 5 (g5) 641x401x2 (Used only for Sp-Al,
Min-Mod, strong shock case) - y 0.53 (for g2) and y 0.26 (for g3) at the
bottom wall
13Output Variables (1)
Nozzle efficiency, neff H0i Total enthalpy
at the inlet He Enthalpy at the exit Hes
Exit enthalpy at the state that would be reached
by isentropic expansion to the actual pressure
at the exit
Throat height
14Output Variables (2)
- Orthogonal Distance Error, En
- A measure of error in wall pressures between the
experiment and the curve representing the CFD
results
Pc Wall pressure obtained from CFD
calculations Pexp Experimental Wall Pressure
Value Nexp Total number of experimental points
(Nexp36) di Orthogonal distance from the ith
experimental data point to Pc(x) curve
15Uncertainty due to iterative convergence error (1)
- Normalized L2 Norm Residual of the energy
equation for the case with Sp-Al turbulence
model, Van-Albada and Min-Mod limiters at the
strong shock case.
- Same convergence behavior with respect to the
limiters observed for the k-w case.
16Uncertainty due to iterative convergence error (2)
Poor L2 norm convergence does not seem to effect
the convergence of the neff results at different
grid levels
17Uncertainty due to iterative convergence error (3)
Roy and Blottner (Ref. 8) proposed a method to
estimate, the iterative convergence error at
time level (cycle) n
Assuming exponential decrease for
Need three time levels in the exponential region
where
18Uncertainty due to discretization error (1)
For each case with a different turbulence model,
grid level (resolution) and the flux-limiter
affect the magnitude of the discretization error
- The effect of the limiter observed at grid
levels g1 and g2 - At grid levels g3 and g4, the effect is much
smaller
19Uncertainty due to discretization error (2)
- Richardsons extrapolation method
-
h a measure of grid spacing p The order of the
method.
- Assumptions needed to use Richardsons method
- Grid resolution is in the asymptotic region
- The order of the spatial accuracy, p should be
known. Usually observed order of spatial accuracy
is different than the nominal value. The observed
order should be determined. - Monotonic grid convergence. Mixed-order schemes
can cause non-monotonic convergence. Roy (Ref. 4)
proposed a method for for the discretization
error estimate of mixed-order schemes. -
20Uncertainty due to discretization error (3)
21Uncertainty due to discretization error (4)
- p values are dependent on the grid levels used
- However the difference between the (neff)exact
values are small compared to overall uncertainty
22Uncertainty due to discretization error (5)
- The uncertainty due to discretization error is
bigger for the cases with strong shock compared
to the weak shock results at each grid level. The
flow structure has significant effect on the
discretization error. - For the monotonic cases, largest errors occur for
the Sp-Al, Min-Mod, strong shock case and the
smallest errors are obtained for the k-w ,
Van-Albada, weak shock case - Non-monotonic convergence behavior for the cases
with k-w and the strong shock as the mesh is
refined
23Uncertainty due to discretization error (6)
24Uncertainty due to discretization error (7)
- Noise due to discretization error observed at
grid levels 1 and 2. - Noise error small compared to the systematic
discretization error between each grid level.
However, this can be important for gradient-based
optimization. - Kim (Ref. 7) successfully modeled the the noise
error due to poor convergence of the optimization
runs by fitting a probability distribution
(Weibull) to the error. - The noise error can be reduced via response
surface modeling.
25Uncertainty due to turbulence models (1)
- Uncertainty due to turbulence modeling (in
general physical modeling) should be investigated
after estimation of the discretization and
iterative convergence error. - Difficult to totally separate physical modeling
errors from discretization errors - Validation of the Engineering and Scientific
Models deals with accuracy of the physical model - Need high-quality experimental data
26Uncertainty due to turbulence models (2)
- Orthogonal distance error, En is used for
comparison of CFD results with the experiment
En for each case is scaled with the maximum
value obtained for k-w , Min-Mod, strong shock
case
27Uncertainty due to turbulence models (3)
For each case (strong shock or weak shock), best
match with the experiment is obtained with
different turbulence models at different grid
levels
28Uncertainty due to turbulence models (4)
- Experimental uncertainty should be considered
- With the experimental geometry, a perfect match
with CFD and experiment can be observed upstream
of the shock - Upstream of the shock, discrepancy between CFD
simulations and experiment is most likely due to
the experimental uncertainty
29Uncertainty due to turbulence models (5)
- A better way of using En for this example would
be to evaluate it only downstream of the shock - The discretization and iterative convergence
error should be estimated for En in a similar
way used for the nozzle efficiency - An estimate of exact value of (En ) can be used
for approximating the uncertainty due to
turbulence models - The relative uncertainty due to the selection of
turbulence models can also be investigated by
using (neff)exact values obtained by Richardsons
extrapolation
30Uncertainty due to turbulence models (6)
- Hills and Trucano (Ref. 5) proposed a Maximum
Likelihood based model validation metric to test
the accuracy of the model predictions - Uncertainty in the experimental measurements and
the model parameters are considered - Model parameters
- Material properties
- Geometry
- Boundary or Initial Conditions
- This method requires prior knowledge about the
measurement and the model parameter uncertainty
(modeling with probabilistic distributions) - Looks for statistically significant evidence that
the model validations are not consistent with the
experimental measurements
31Uncertainty due to turbulence models (7)
- PDF(d) PDF of measurement vector occurrence
- PDF(p) PDF of model parameter vector
occurrence - PDF(d, p) PDF(d) x PDF(p)
- Find the maximum likely values for the mode of
the measurements d and the model parameters p - Find the maximum value of Joint PDF via
optimization - Evaluate the probability of obtaining a smaller
PDF assuming that the model is correct - If this value is bigger than the level of
significance that we assumed for rejecting a good
model, than the model predictions are consistent
with the experiment
32Uncertainty due to turbulence models (8)
- Possible application to test the accuracy of the
turbulence models - Takes into account the experimental uncertainty
- Requires prior knowledge of uncertainty in the
measurements and the model parameters - Selection of model parameters
- No simple relationship with the model parameters
and the output quantities. Using response surface
techniques may be needed to find a functional
form.
33Additional Test Cases
- Need more cases to generalize the results
obtained in Transonic Diffuser Case - Next possible case Steady, turbulent, flow
around an airfoil (RAE2822 or NACA0012) - Consider transonic and subsonic cases
- Consider a range of AOA
- Output quantities to monitor Cl, Cd, Cp
distributions - Orthogonal distance error may be used for
characterizing Cp distributions - Consider a case with a more complex geometry
34Conclusions (1)
- Different sources of uncertainty in CFD
simulations should be investigated separately. - Discretization and iterative convergence errors
can be estimated by certain methods in certain
conditions - Limiters affect the iterative convergence and the
discretization error. - L2 norm convergence affected by the use of
different limiters - Poor L2 norm convergence do not seem to affect
the neff results - Asymptotic Grid convergence hard to obtain
- Flow structure has a strong effect on the
magnitude of the discretization error. - Iterative convergence error small compared to the
discretization error - Uncertainty due to turbulence model should be
investigated after the estimation of
discretization and iterative convergence error.
35Conclusions (2)
- Comparison with the experiment is needed to
determine the accuracy of the turbulence models - Experimental uncertainty should be considered
possibly by using a statistical method - More cases need to be analyzed to generalize the
results
36References
- Levy, D. W., Zickuhr, T., Vassberg, J., Agrawal
S., Wahls. R. A., Pirzadeh, S., Hemsch, M. J.,
Summary of Data from the First AIAA CFD Drag
Prediction Workshop, AIAA Paper 2002-0841,
January 2002 - Oberkampf, W. L. and Trucano, T. G., Validation
Methodology in Computational Fluid Dynamics.
AIAA Paper 2000-2549, June 2000 - Roache, P. J. Verication and Validation in
Computational Science and Engineering.Hermosa
Publishers, Albuquerque, New Mexico, 1998. - Roy, C. J., Grid Convergence Error Analysis for
Mixed-Order Numerical Schemes, AIAA Paper
2001-2606, June 2001 - Hills, R. G. and Trucano, T. G., Statistical
Validation of Engineering and Scientific Models
A Maximum Likelihood Based Metric, Sandia
National Loboratories, SAND2001-1783 - Hemsch, M. J., Statistical Analysis of CFD
Solutions from the Drag Prediction Workshop, AIAA
Paper 2002-0842, January 2002 - Kim, H., Statistical Modeling of Simulation
Errors and Their Reduction Via Response Surface
Techniques, PhD dissertation, VPISU, June 2001 - Roy, C. J. and Blottner F. G., Assesment of
One-and Two-Equation Turbulence Models for
Hypersonic Transitional Flows, Journal of
Spacecraft and Rockets, Vol.38, No. 5,
September-October 2001