One-shot robust optimisation with grid adaptation using adjoint … · 2012. 4. 18. · Adjoint...
33
Munich 2012 A. Jaworski, L.Laniewski and J. Rokicki One-shot robust optimisation with grid adaptation using adjoint sensitivities A. Jaworski, L.Laniewski and J. Rokicki Institute of Aeronautics and Applied Mechanics Warsaw University of Technology FLOWHEAD Research Project 7 Framework Programme Munich 28.03.2012
Transcript of One-shot robust optimisation with grid adaptation using adjoint … · 2012. 4. 18. · Adjoint...
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
One-shot robust optimisation with gridadaptation using adjoint sensitivities
A. Jaworski, Ł.Łaniewski and J. Rokicki
Institute of Aeronautics and Applied MechanicsWarsaw University of Technology
FLOWHEAD Research Project7 Framework Programme
Munich 28.03.2012
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Motivation
I Aerodynamic optimisation is limited bycomputational cost
I Grid adaptation can reduce computational costI Robust optimal design is needed
Research tasks:
I Develop one-shot optimisation coupled with adjointbased grid adaptation
I Develop robust optimisation coupled with gridadaptation
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Motivation
I Aerodynamic optimisation is limited bycomputational cost
I Grid adaptation can reduce computational costI Robust optimal design is needed
Research tasks:
I Develop one-shot optimisation coupled with adjointbased grid adaptation
I Develop robust optimisation coupled with gridadaptation
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Research tasks
I One-shot optimisation coupled with adaptation
I Reduced cost of robust optimisation
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Adjoint solver
(∂R∂Q
)Tv =
(∂L∂Q
)T,
ATv = g.
dLdα
=∂L∂α
+ gTu =∂L∂α
+ vTf
1. Implicit adjoint solver developed by WUTI solving ATv = g using sparse JacobianI 0.1 factor of CPU consumption compared to primal
calculation
2. ANSYS Fluent v14 adjoint solver
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Adaptation indicator
I Adaptation indicator: adjoint · Hesjan
ek =n∑i
(|vi| · |hTHih|) ·Vi (1)
where: i - flow variable, k - node
I Adaptation indicator - scaling
ak =
(ek ·Nδlim
)ω(2)
I Separate ω for coarsening and refinement gives betterconvergence
I New edge length hk:
hnew = hold ·1ak
hmin < hnew < hmax (3)
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Adaptation - comparison with uniformlyrefined grids, M = 2.0
142000 nodes, 19000 nodes.
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Adaptation - comparison with uniformlyrefined grids, M = 2.0
142000 nodes, 19000 nodes.
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Adaptation - comparison with uniformlyrefined grids, M = 2.0
142000 nodes, 19000 nodes.
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Adaptation efficiency
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Optimisation + adaptationWave Rider testcase
I Mach = 2.0, inviscid flowI Optimisation task - minimize drag with constant liftI Gradient based L-BFGS-B optimiserI 4 design variables
source: boeing.com
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Optimisation + adaptationWave Rider testcase
I Mach = 2.0, inviscid flowI Optimisation task - minimize drag with constant liftI Gradient based L-BFGS-B optimiserI 4 design variables
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Optimisation + adaptation, error = const
iteration: 1 iteration: 2 iteration: 3
iteration: 4 iteration: 5 iteration: 34
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Optimisation + adaptation, error = constConvergence history:
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
One-shot method
Cost of single design iteration depends on:I mesh size, number of nodesI solution accuracy, residual stop criteria
Total cost of optimisation can decreased by:I solving CFD with minimal acceptable accuracyI using mesh with lowest acceptable number of nodes
One-shot Wolfe condition:I estimation of accuracy acceptable by optimiser:
acc ∼ min(
log( |g|
gmin
), log
( |∆f|∆fmin
))
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
One-shot method
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Wave-rider: one-shot + adaptation
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
One-shot + adaptation, comparison
Difference in final design between one-shot and constanterror.
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
One-shot + adaptation2D sband testcase
I Optimisation task - minimize pressure dropI Gradient based L-BFGS-B optimiser, 14 des. var.I ANSYS Fluent v14 adjoint solverI Laminar flow, Re = 300I Simplified adaptation - adjoint solver convergence
problems
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Sband: one-shot + adaptation, performance
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Sband: one-shot + adaptation, performance
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Summary one-shot+adaptation coupling
I Faster optimisation from 10 to 100 times is obtainedI Adaptation is keeping accuracy at desired levelI Coupling one-shot method with adaptation can
significantly reduce overall optimisation cost
ProblemsI One-shot performance is case-sensitive and depends
on user input parametersI Optimisation algorithm performane is affected,
Hessian approximation is affected by changingaccuracy of functional and gradient
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Robust Optimisation Framework — Data-flow
ConfigurationDesignDesignDesignCFD
Scheduling
and calculationO
ptimization
Ne
w d
esi
gn
Ne
w d
esi
gn
Old
de
sig
ns
Dataset
Sampling strategy DoE
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Robust Optimisation Framework —Generating new designs
InnerLoop
Models Sampling criterionEI, REI, EHVI
New design OptimizerNSGA-II / L-BFGS-B
N/A
de
sig
ns
Ca
lcu
late
d d
esi
gn
s
Mo
de
ls
Model Constructionfor objectives ans constraints
Dataset
Model extension
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Robust optimisation concept
I Kriging approximation is smoothing the functionalI Uncertainty level provided by a user
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Wave-rider — Optimization loop
Mesh gen.Green
Mesh morphing
OptimizationFramework CFD+Adjoint
optFrame
Metricgeneration
Error control
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Wave-rider Testcase
I M = 2.0, inviscid flowI Kriging based optimiserI 7 design variables
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Wave-rider — Optimization convergence
I 7 desing var. 200 runs. 94 crashed.
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Renault side-mirror testcase
I Minimisation of turbulence in the wakeI Kriging based optimiserI CAD based parametrisation, 8 desing var.
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Side-mirror — Optimization Loop
Mesh gen.Star-CCM+
Geometrygeneration
CAD ModelCatia V5
OptimizationFramework
CFDStar-CCM+
WUT
PACAGrid Cluster (INRIA)
SIREHNA
Interface provided by Renault
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Side-mirror — Optimization run
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Side-mirror — Geometry comparison
Munich 2012
A. Jaworski,Ł.Łaniewskiand J. Rokicki
Adaptation
Opt+adapt
One-shot +adapt
Summary
Robust
Summary
Summary - Kriging based robust optimisation
I Robust design with respect to random parametervariations
I Easy integration with any toolchain (e.g. CADparametrisation)
I Capable of using derivative informationI Faster global optimisation if compared to genetic
algorithms
Limitations
I Moderate number of design var. (up to 30)
