an optimization-based framework for controlling...
TRANSCRIPT
An Optimization-based Frameworkfor Controlling Discretization Errorthrough Anisotropic h-Adaptation
Masayuki Yano and David Darmofal
Aerospace Computational Design LaboratoryMassachusetts Institute of Technology
Acknowledgments: Dr. Steven Allmaras, ProjectX Team, The Boeing Company
March 24, 2011
M. Yano (MIT) FEF 2011 March 24, 2011 1 / 22
Outline
1 Introduction
2 Optimization Framework : Anisotropic h-Adaptation
3 Results : L2 Error Control
4 Results : Functional Output Error Control
5 Conclusion
M. Yano (MIT) FEF 2011 March 24, 2011 1 / 22
Objective and Approach
Objective: Realize the full potential of high order methods throughautomated mesh adaptation and error control
Our choice of technologies:Discontinuous Galerkin (DG) method
hp-flexible on unstructured meshesConvergence limited by regularity
- shocks, geometric singularities,turbulence models
Simplex mesh adaptationComplex geometriesArbitrarily oriented anisotropy
Adjoint-based error estimatorError estimate / localization foroutput of interest (e.g. lift, drag)
M. Yano (MIT) FEF 2011 March 24, 2011 2 / 22
Functional Error Estimation
Problem: Find J(u) with u ∈ V s.t.
R(u, v) = 0 ∀v ∈ V
Dual-weighted residual (DWR) error representation [Becker,2001]
E ≡ Jh,p(uh,p)− J(u) = Rh,p(uh,p, ψ), ψ = adjoint
Local error indicator
ηK ≡ Rh,p(uh,p, ψh,p+1|K )
Error indicator assigns 1 value to each elementOnly sufficient for 1 parameter adaptation
type nparam parameters ηK sufficientisotropic h 1 size 4
isotropic hp 2 size, poly. order 8
anisotropic h 3 (6) size, shape 8
anisotropic hp 4 (7) size, shape, poly. order 8
M. Yano (MIT) FEF 2011 March 24, 2011 3 / 22
Review of Anisotropic h-Adaptation
Need additional info to make anisotropy decisionPrevious work on output-based anisotropic h-adaptation
Method p > 1 ψ ani. simplexMach Hessian [Venditti, 2003] 8 8 4Mach “Hessian” [Fidkowski, 2007] 4 8 4Mach jump [Hartmann, 2008] 4 8 8Edge comp. ref [Park, 2008][Sun, 2009] 4∗ 4∗ 4∗
Comp. ref. [Georgoulis, 2009][Ceze, 2010] 4 4 8Flux reconstruction + C.V. [Loseille, 2010] 8 4 4Anisotropic error estimate [Richter, 2010] 8 4 8
Contribution: Anisotropic h-adaptation framework for simplexmeshes that implicitly synthesizes information about
primal and adjoint solution behaviorsinterpolation ordersolution regularity
M. Yano (MIT) FEF 2011 March 24, 2011 4 / 22
Outline
1 Introduction
2 Optimization Framework : Anisotropic h-Adaptation
3 Results : L2 Error Control
4 Results : Functional Output Error Control
5 Conclusion
M. Yano (MIT) FEF 2011 March 24, 2011 4 / 22
Problem Definition / Example
Objective: Develop “optimal” anisotropic meshExample problem: L2-projection error control of boundary layer
ED(Th) = ‖u − uh,p‖2L2(Ω) where u = exp (−x1/δ)
CD(Th) = DOF(Th)
u = exp(−x1/δ) optimal isotropic mesh optimal anisotropic mesh(δ = 0.01) E
12 = 1.6× 10−4 E
12 = 5.1× 10−8
*each mesh has ≈ 200 elements (p = 3)
M. Yano (MIT) FEF 2011 March 24, 2011 5 / 22
Continuous Relaxation
(Intractable) discrete optimization problem
Th,opt = arg minThED(Th) s.t. CD(Th) = DOFtarget
Approximability of Th encoded in metric tensor field
M(x) ∈ Sym+d (d × d SPD matrix)
Continuous relaxation [cf. Loseille, 2009]
Mopt = arg minME(M) s.t. C(M) = DOFtarget
Metric Field M Mesh Th
Mesh Gen.
Implied Metric
M. Yano (MIT) FEF 2011 March 24, 2011 6 / 22
Continuous Optimization
Problem Statement: Seek optimal metric field
Mopt = arg minME(M) s.t. C(M) = DOFtarget
Assume locality of error and cost functionals
E(M) =
∫Ω
e(x ,M(x))dx and C(M) =
∫Ω
c(M(x))dx
local error function : e(·, ·) : Rd × Sym+d → R+
local cost function : c(M) = Cp√
det(M)
First order optimality
∂e∂M
(x ,M(x))− λ ∂c∂M
(x ,M(x)) = 0 ∀x ∈ Ω
Need to estimate behavior of e(M)
M. Yano (MIT) FEF 2011 March 24, 2011 7 / 22
Local Sampling
Split edges of K0 to form Ki ’s, i = 0, . . . ,mObtain ηKi on split config.
Freeze neighbor statesSolve local problem
Obtain metricMKi implied by split config.Associate ηKi ≈ e(MKi )|Ki |
original edge 1 split edge 2 split edge 3 split uniform splitMK0 , ηK0 MK1 , ηK1 MK2 , ηK2 MK3 , ηK3 MKu , ηKu
M. Yano (MIT) FEF 2011 March 24, 2011 8 / 22
Manipulation of SPD Matrix Space
Original problem on Sym+d
∂e∂M
(x ,M)− λ ∂c∂M
(x ,M) = 0 M∈ Sym+d
SPD matrices define cone in Euclidean spaceHard to manipulate SPD matrices directly
Consider substitutionsM(Σ) =M1/2
0 exp(Σ)M1/20 ⇒ M(Σ) ∈ Sym+
d ∀Σ ∈ Symde(f ) = e0 exp(f ) ⇒ e(f ) ∈ R+ ∀f ∈ Rc(g) = c0 exp(g) ⇒ c(g) ∈ R+ ∀g ∈ R
New problem on Symd
e0∂f∂Σ
(x ,Σ)− λc0∂g∂Σ
(x ,Σ) = 0 Σ ∈ Symd
Manipulation of symmetric matricesApproximate f (Σ) and move toward optimality
M. Yano (MIT) FEF 2011 March 24, 2011 9 / 22
Local Error Model / Parameter Identification
Construct linear approximation of f (Σ)
f (Σ) = tr(RΣ), R ∈ Symd = rate tensor
R is generalization of convergence rate to tensor spacetr(R) : sensitivity to uniform scalingR − tr(R)
d I : sensitivity to shape change
Identify R from local samples MKi , ηKimi=1 − ΣK1 −...
− ΣKm −
R11
2R12
R22
=
f (ΣK1 )...
f (ΣKm )
Σ(Ki ) = log(M−1/2
0 MKiM−1/20 ) and f (ΣKi ) = log(ηKi/ηK0 )
M. Yano (MIT) FEF 2011 March 24, 2011 10 / 22
Discrete Optimality Condition
Discrete Optimality Conditions (Anisotropic)Global isotropic error balance : ∃λ ∈ R s.t.
ηK tr(RK )− λd2 dof(K ) = 0 ∀K ∈ Th
Local anisotropic optimality :RK ∝ I ∀K ∈ Th
e.g. R 6= I : more reduction by increasingM11 thanM22⇒ decrease h1 and increase h2
RK ,11 RK ,22
M. Yano (MIT) FEF 2011 March 24, 2011 11 / 22
Algorithm
1 Obtain primal solution uh,p on T (i)h,p
2 Obtain adjoint solution ψh,p+1 on T (i)h,p+1 (for DWR)
3 Solve local problems and obtain MKi , ηKimi=0, ∀K ∈ T (i)h
4 Construct local error function e(M) ≈ e(M; MKi , ηKimi=0)
5 Update metric fieldM(i+1)
6 Generate new mesh T (i+1) fromM(i+1) and go to 1
Adapt 0 : E12 = 9.4× 10−3 Adapt 2 : E
12 = 1.8× 10−5 Adapt 7 : E
12 = 7.9× 10−8
*each mesh has ≈ 200 elements (p = 3)
M. Yano (MIT) FEF 2011 March 24, 2011 12 / 22
Assessment of Boundary Layer L2 Error Control
Function: u = exp(−x1/δ) (δ = 0.01)Analytical optimal h⊥ vs. x1
h⊥,opt = C exp(k⊥,optx
), k⊥,opt =
1δ(p + 3/2)
Weaker grading for higher p
0 0.02 0.04 0.06 0.08 0.110
−4
10−3
10−2
10−1
kperp
=43.2
kperp
=44.1
kperp
=43.5
kperp
=42.8
x
h⊥
dof=500dof=1000
dof=2000dof=4000
0 0.02 0.04 0.06 0.08 0.110
−4
10−3
10−2
10−1
kperp
=24.9
kperp
=24.9
kperp
=22.9
kperp
=22.7
x
h⊥
dof=1000dof=2000
dof=4000dof=8000
p = 1 (k⊥,opt = 40.0) p = 3 (k⊥,opt = 22.2)M. Yano (MIT) FEF 2011 March 24, 2011 13 / 22
Assessment of Singular Corner L2 Error Control
Function: u = rα sin[α(θ + π
2
)](α = 2/3)
Analytical optimal h vs. r
hopt(r) = Cr kopt , kopt = 1− α + 1p + 2
Stronger grading for higher p
10−3
10−2
10−1
100
10−3
10−2
10−1
k=0.43
k=0.44
k=0.44
k=0.44
r
h
dof=500
dof=1000
dof=2000
dof=4000
10−3
10−2
10−1
100
10−3
10−2
10−1
k=0.68
k=0.67
k=0.67
k=0.67
r
h
dof=500
dof=1000
dof=2000
dof=4000
p = 1 (kopt = 0.44) p = 3 (kopt = 0.67)M. Yano (MIT) FEF 2011 March 24, 2011 14 / 22
Outline
1 Introduction
2 Optimization Framework : Anisotropic h-Adaptation
3 Results : L2 Error Control
4 Results : Functional Output Error Control
5 Conclusion
M. Yano (MIT) FEF 2011 March 24, 2011 14 / 22
RAE2822 Transonic AirfoilRANS-SA : M∞ = 0.729, Rec = 6.5× 106, α = 2.31; p = 2
Objective: Application to transsonic turbulent flowFunctional : DragComparison : Mach-based anisotropy (∇p+1M)
Mach
Mass adjoint
Drag error
20000 30000 40000 60000
10−7
10−6
10−5
10−4
10−3
dof
|J −
Jre
f|
mach ani.
unified
M. Yano (MIT) FEF 2011 March 24, 2011 15 / 22
RAE2822 Transonic AirfoilRANS-SA : M∞ = 0.729, Rec = 6.5× 106, α = 2.31; p = 2
Anisotropic resolution of featuresPrimal: BLs, wake, shockAdjoint: BLs, stagnation streamline, shock-induced perturbations
Unified (p = 2, dof = 60k) Mach anisotropy (p = 2, dof = 60k)
M. Yano (MIT) FEF 2011 March 24, 2011 16 / 22
NACA0006 Shock PropagationEuler : M∞ = 2.0; p = 2
Objective: Application to sonic-boom predictionFunctional : Pressure 50c below airfoil, J =
∫l(p − p∞)2ds
Pressure
Mass adjoint
Pressure line output error
10000 20000 30000 40000
10−4
10−3
10−2
10−1
100
dof
|J −
Jre
f|/|J
ref|
mach ani.
unified
M. Yano (MIT) FEF 2011 March 24, 2011 17 / 22
NACA0006 Shock PropagationEuler : M∞ = 2.0; p = 2
Unified (p = 2, dof = 40k) Mach anisotropy (p = 2, dof = 40k)
M. Yano (MIT) FEF 2011 March 24, 2011 18 / 22
Three-Element MDA AirfoilRANS-SA : M∞ = 0.2, Rec = 9× 106, α = 8.10 ; p = 2
Objective: Application to complex turbulent flowFunctional : Drag
Mach
Mass adjoint
Drag error
30000 45000 60000 90000
10−5
10−4
10−3
dof
|J −
Jre
f|
mach ani.
unified
M. Yano (MIT) FEF 2011 March 24, 2011 19 / 22
Three-Element MDA AirfoilM∞ = 0.2, Rec = 9× 106, α = 8.10; p = 2
Unified (p = 2, dof = 90k)
Mach anisotropy (p = 2, dof = 90k)
M. Yano (MIT) FEF 2011 March 24, 2011 20 / 22
Outline
1 Introduction
2 Optimization Framework : Anisotropic h-Adaptation
3 Results : L2 Error Control
4 Results : Functional Output Error Control
5 Conclusion
M. Yano (MIT) FEF 2011 March 24, 2011 20 / 22
Conclusion and Ongoing Work
Conclusion:Presented optimization framework for anisotropic h-adaptationthat implicitly synthesizes information about
primal and adjoint solution behaviorsinterpolation ordersolution regularity
Verified optimality of algorithm for L2 error controlCompared performance of algorithm against primal derivativebased anisotropy detection
Ongoing Work:Extension to hp-adaptationImproving robustness of error estimator for highly nonlinearproblems
M. Yano (MIT) FEF 2011 March 24, 2011 21 / 22