parametric manifold models for validation; verification for...
TRANSCRIPT
Parametric Manifold Modelsfor Validation;
Verification forParametric Manifold Models
Anthony T PateraMassachusetts Institute of Technology
NRC Symposium on VVUQWashington, DC28 March 2012
Acknowledgements Collaborators
P Huynh* D Knezevic* M Dihlmann
J Eftang E Rønquist*
S Vallaghe K Urban
T Tonn
Also S Boyaval, A-L Gerner, M Grepl, B Haasdonk, J Hesthaven, KC Hoang,C Le Bris, T Leurent, GR Liu, Y Maday, C Nguyen, J Penn, A Quarteroni,G Rozza, K Steih, K Veroy, K Willcox, and M Yano.
NRC Symposium on VVUQ AT Patera 2
Acknowledgements Sponsors
AFOSR/Office of Secretary of Defense
Office of Naval Research
MIT-Singapore International Design Center
Deshpande Center for Technological Innovation
NRC Symposium on VVUQ AT Patera 3
Acknowledgements AFOSR/OSD MURI(F Fahroo)
Multi-Scale Fusion of Information UncertaintyQuantification and Management
in Large-Scale Simulations
Brown: K Karniadakis, J Hesthaven, B RozovskyCaltech: T HouCornell: N Zabaras
MIT: A Patera, K Willcox
NRC Symposium on VVUQ AT Patera 4
FrequentistValidation Framework
with“Exact”† Model
†Exact in the sense of verification, not validation.
NRC Symposium on VVUQ AT Patera 5
“Exact” Model
Given µ ∈ D ⊂ RP , µ: parameter P -vector
umodel( · ;µ) ∈ X(Ω) ,
for Ω a domain in Rd.
Note the field umodel(x;µ) may derive from aparametrized partial differential equation.†
†In an example we also consider time-dependent fields.
NRC Symposium on VVUQ AT Patera 6
Measurements of Physical System
Given measurement functionals,
`measi : X → R, 1 ≤ i ≤ m ,
assume experiments yield outputs
Y expi = `meas
i (uphy)︸ ︷︷ ︸Y
phyi
+ ε(ωi), 1 ≤ i ≤ m ,
for ε(ω) a zero-mean Gaussian process.
Note uphy(x) serves to “interpret” Y phyi .
NRC Symposium on VVUQ AT Patera 7
Validation Hypothesis
For given µ ∈ D, δ ≥ 0
H(µ):
For all µ′ ∈ Nδ(µ) ,
‖Y phy − Y model(µ′)‖ ≤ δ .
Here Y modeli (µ′) = `meas
i (umodel( · ;µ′)), 1 ≤ i ≤ m;Nδ(µ) a neighborhood in D about µ.
Note: H(µ) FALSE, ∀µ ∈ D ⇒ “unmodeled physics.”
NRC Symposium on VVUQ AT Patera 8
Statistical Test
For any µ ∈ D,
REJECT H(µ) vs. ACCEPT H(µ)
if and only if
Fε(Y exp, Y model(µ)) > Bδ(µ) ;
require PrErrorType I ≤ 1− γ.†
† H(µ) TRUE: PrREJECT H(µ) ≤ 1− γ.
NRC Symposium on VVUQ AT Patera 9
FrequentistValidation Framework
withApproximate Model
NRC Symposium on VVUQ AT Patera 10
Approximate Model
Given µ ∈ D ⊂ RP , µ: parameter P -vector
umodel( · ;µ) ∈ X(Ω) ⊂ X(Ω) ,
where u(µ) ≈ umodel(µ).
Introduce Y modeli (µ) = `meas
i (u( · ;µ)), 1 ≤ i ≤ m.
NRC Symposium on VVUQ AT Patera 11
Measurements of Physical System
Given measurement functionals,
`measi : X → R, 1 ≤ i ≤ m ,
assume experiments yield outputs
Y expi = `meas
i (uphy)︸ ︷︷ ︸Y
phyi
+ ε(ωi), 1 ≤ i ≤ m ,
for ε(ω) a zero-mean Gaussian process.
Note uphy(x) is an “interpretation” of Y phyi .
NRC Symposium on VVUQ AT Patera 12
Validation Hypothesis
For given µ ∈ D, δ ≥ 0
H(µ):
For all µ′ ∈ Nδ(µ) ,
‖Y phy − Y model(µ′)‖ ≤ δ .
Here Y modeli (µ′) = `meas
i (umodel( · ;µ′)), 1 ≤ i ≤ m;Nδ(µ) a neighborhood in D about µ.
Note: H(µ) FALSE, ∀µ ∈ D ⇒ “unmodeled physics.”
NRC Symposium on VVUQ AT Patera 13
Statistical Test
For any µ ∈ D,
REJECT H(µ) vs. ACCEPT H(µ)
if and only if
Fε(Y exp, Y model(µ)) > Bδ(µ)Verification
; (≥ Bδ(µ))
require PrErrorType I ≤ 1− γ.†
† H(µ) TRUE: PrREJECT H(µ) ≤ 1− γ.
NRC Symposium on VVUQ AT Patera 14
Statistical Test
For many µ ∈ D,
REJECT H(µ) vs. ACCEPT H(µ)
if and only if
Fε(Y exp, Y model(µ)) > Bδ(µ)Verification
; (≥ Bδ(µ))
require PrErrorType I ≤ 1− γ.†
† H(µ) TRUE: PrREJECT H(µ) ≤ 1− γ.
NRC Symposium on VVUQ AT Patera 15
Parametric Manifold Models
Approximation Space
NRC Symposium on VVUQ AT Patera 16
Parametric ManifoldMh
Introduce uh(µ) ∈ Xh ⊂ X: a highly accuratefinite element (FE) surrogate for umodel(µ).
. ..Greedy Samplingsnapshots
uuhh((˜µµ11))uuhh((˜µµ22))
uuhh((˜µµNN ))
XXhh
MMhh ≡≡ uuhh((µµ)) µµ ∈∈ DD
Approximation Space: XD = spanuh(µn), 1 ≤ n ≤ N.NRC Symposium on VVUQ AT Patera 17
Parametric Manifold Models
Approximation Space
Reduced Basis (RB)
NRC Symposium on VVUQ AT Patera 18
Galerkin Projection
Given weak form (parametrized PDE),
µ ∈ D → A(w, v;µ), ∀ w, v in X(Ω) ,†
uRB ∈ XD satisfies
A(uRB, v;µ) = 0, ∀ v ∈ XD .
Advantage: N = dim(XD) dim(Xh) .
†Note umodel(µ) ∈ X: A(umodel(µ), v;µ) = 0, ∀ v ∈ X;uh(µ) ∈ Xh: A(uh(µ), v;µ) = 0, ∀ v ∈ Xh.
NRC Symposium on VVUQ AT Patera 19
A Posteriori Error Estimation
For any µ ∈ D,
‖uRB(µ)− umodel(µ)‖X≤ ‖uRB(µ)− uh(µ)‖X + ‖uh(µ)− umodel(µ)‖X≈ ‖uRB(µ)− uh(µ)‖X ; OFFLINE-ONLINE
provide rigorous error bound
‖uRB(µ)− uh(µ)‖ ≤ ∆hRB(µ) ,
where ∆hRB(µ) ≡ Ch
stability(µ) ‖A(uRB, · ;µ)‖(Xh)′ .
NRC Symposium on VVUQ AT Patera 20
OFFLINE-ONLINE Computational Strategy
OFFLINE Construct XD (once): expensive.
ONLINE Evaluate µ→ uRB, ∆hRB: inexpensive.
In the many-query (or real-time) context, amortizeOFFLINE effort over many ONLINE evaluations.
Note: ONLINE cost independent of dim(Xh),hence choose Xh conservatively rich.
NRC Symposium on VVUQ AT Patera 21
Parametric Manifold Models
Approximation Space
Reduced Basis (RB)
Empirical Interpolation (EI)
NRC Symposium on VVUQ AT Patera 22
Collocation
OFFLINE: Given XD, identifyquasi optimal interpolation points xk;associated characteristic functions Vk ∈ XD;
for 1 ≤ k ≤ N .
ONLINE: For µ ∈ D, form uEI(µ) ∈ XD as
uEI(x;µ) =N∑k=1
uh(xk;µ)Vk(x) .†
Advantage: Ω ⊂ Rd replaced by xkk=1,...,N .
†In practice, EI applied to functions of approximations of uh(µ).
NRC Symposium on VVUQ AT Patera 23
Parameter Manifold Models
Approximation Space
Reduced Basis (RB)
Empirical Interpolation (EI)
Dirty Laundry
NRC Symposium on VVUQ AT Patera 24
Issues
Difficulty Palliative
Stability Constant, Chstability Successive Constraint
P 1: µ ∈ D ⊂ RP Domain Decomposition in Ω
D Large: µ ∈ D ⊂ RP Domain Decomposition in D
Nonpolynomial Nonlinearity Empirical Interpolation
Long-Time Integration Space-Time Inf-Sup
NRC Symposium on VVUQ AT Patera 25
Numerical Example
NRC Symposium on VVUQ AT Patera 26
Cast of Characters
“Exact” Model:
(µ1, µ2) ∈ D → umodel(t, x;µ) ∈ X
FE Surrogate: dim(Xh) ≈ 3,000
(µ1, µ2) ∈ D → uh(t, x;µ) ∈ Xh
RB Approximate Model: dim(XD) ≤ 40 †
(µ1, µ2) ∈ D → uRB(t, x;µ) ∈ XD, ∆hRB(t, µ)
Physical System: uphy(t, x)
†RB 100× faster (ONLINE) than FE.
NRC Symposium on VVUQ AT Patera 27
Case I: Model
1100
1100
ΩΩ
ΓΓinin
zero
flux
zero
flux
ΓΓ00
ΓΓ∞∞
ΩΩmeme
00.5.500.5.5
ΩΩ11µµ11 ΩΩ00
µµ00 == 11ΩΩ22µµ22
∂∂uummoodeldel
∂∂nn== 11
YY momodedellii ((µµ)) ==
11
——ΩΩmeme —— ΩΩmeme
uumomodedell((ttii,, xx;;µµ)) ddxx
∂∂uummoodeldel
∂∂tt−−∇∇ ·· ((µµkk∇∇uumomodeldel)) == 00 in ΩΩkk
Parametrization: µ ≡ (µ1, µ2) ∈ [0.2, 5]2 ≡ D.
NRC Symposium on VVUQ AT Patera 28
Case I: Physical System(fully modeled physics)
1100
1100
ΩΩ
ΓΓininze
ro fl
uxze
ro fl
ux
ΓΓ00
ΓΓ∞∞ΩΩmeme
00.5.500.5.5
ΩΩ11ΩΩ00ΩΩ22
∂∂uuphphyy
∂∂nn== 11
αα11 == 44αα22 == 11 αα00 == 11
∂∂uuphphyy
∂∂tt−−∇∇ ·· ((ααkk∇∇uuphphyy)) == 00 inin ΩΩkk
YY expexpii ==
11
——ΩΩmmee —— ΩΩmeme
uuphphyy((ttii,, xx)) ddxx
++ ((ttii;;ωωii))unknown
Gaussian process
uphy(t, x) = umodel(t, x;µ∗ ≡ (4, 1)(α1,α2)
) †.†In practice, umodel(t, x;µ∗) replaced by uh(t, x;µ∗).
NRC Symposium on VVUQ AT Patera 29
Case I: Validation(noise: ≈ 1%, uncorrelated)
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50.51
1.52
2.53
3.54
4.55
µµ∗∗
µµ11
µµ22
PrPr REJECT HH ((µµ∗∗ ≤≤ 00..0505))reject HH when HH is true
REJECT
ACCEPT
NRC Symposium on VVUQ AT Patera 30
Case I: Effort
Experiment (synthetic):
m = 10 repeats× 100 measurements in time
= 1,000 data in total.
Computation: 10,000 ONLINE evaluations
µ→ Y model(µ), ∆hRB(µ), Bδ(µ)
⇒ ACCEPT or REJECT H(µ).
NRC Symposium on VVUQ AT Patera 31
Case II: Model
1100
1100
ΩΩ
ΓΓinin
zero
flux
zero
flux
ΓΓ00
ΓΓ∞∞
ΩΩmeme
00.5.500.5.5
ΩΩ11µµ11 ΩΩ00
µµ00 == 11ΩΩ22µµ22
∂∂uummoodeldel
∂∂nn== 11
YY momodedellii ((µµ)) ==
11
——ΩΩmeme —— ΩΩmeme
uumomodedell((ttii,, xx;;µµ)) ddxx
∂∂uummoodeldel
∂∂tt−−∇∇ ·· ((µµkk∇∇uumomodeldel)) == 00 in ΩΩkk
Parametrization: µ ≡ (µ1, µ2) ∈ [0.2, 5]2 ≡ D.
NRC Symposium on VVUQ AT Patera 32
Case II: Physical System(unmodeled physics: crack)
1100
1100
ΩΩ
ΓΓinin
zero
flux
zero
flux
ΓΓ00
ΓΓ∞∞
ΩΩmeme
00.5.500.5.5
ΩΩ11ΩΩ00ΩΩ22
∂∂uuphphyy
∂∂nn== 11
αα11 == 44αα22 == 11 αα00 == 11
∂∂uuphphyy
∂∂tt−−∇∇ ·· ((ααkk∇∇uuphphyy)) == 00 inin ΩΩkk
YY expexpii ==
11
——ΩΩmmee —— ΩΩmeme
uuphphyy((ttii,, xx)) ddxx
++ ((ttii;;ωωii))unknown
Gaussian process
ΓΓcracracckk
uphy(t, x) 6= umodel(t, x;µ) for any µ ∈ D.NRC Symposium on VVUQ AT Patera 33
Case II: Validation(noise: ≈ 1%, uncorrelated)
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.51
1.52
2.53
3.54
4.55
µµ11
µµ22
REJECT
ACCEPT
µµ∗∗“ ”
Crack Length = 1: crack conflated withlower conductivity medium.
NRC Symposium on VVUQ AT Patera 34
Case II: Validation(noise: ≈ 1%, uncorrelated)
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.51
1.52
2.53
3.54
4.55
REJECT
µµ∗∗“ ”
µµ11
µµ22
Crack Length = 1.5: unmodeled physics detected.
NRC Symposium on VVUQ AT Patera 35
AFOSR/OSD MURI
Multi-Scale Fusion of Information UncertaintyQuantification and Management
in Large-Scale Simulations
Brown: K Karniadakis, J Hesthaven, B RozovskyCaltech: T HouCornell: N Zabaras
MIT: A Patera, K Willcox
NRC Symposium on VVUQ AT Patera 36
MURI Architecture
while (model insufficient)
propose/update model: µPDE, SPDE;
develop Reduced Order Model (ROM);ISOMAP, ANOVA, RB/EI, MEPC, . . .
verify ROM;
validate model⇐ Fε (data, ROM);
end
design under uncertainty⇐ ACCEPTED H(µ);
NRC Symposium on VVUQ AT Patera 37
MURI Domains & Applications
Thermofluids
Acoustics (for Noise Mitigation . . . )
Electromagnetics (for Stealth . . . )
(Multiscale) Materials
NRC Symposium on VVUQ AT Patera 38
References
NRC Symposium on VVUQ AT Patera 39
Frequentist Validation
V Chew, Confidence, prediction, and tolerance regions for the multivariatenormal distribution. J Am Stat Ass 61(315): 605–617, 1966.
O Balci, R Sargent, Validation of simulation models via simultaneousconfidence intervals. Am J Math Manage Sci 4: 375–406, 1984.
J McFarland, S Mahadevan, Multivariate significance testing and modelcalibration under uncertainty. Comput Methods Appl Mech Engrg 197(29– 32):2467–2479, 2008.
DBP Huynh, DJ Knezevic, AT Patera, Certified reduced basis model validation:A frequentistic uncertainty framework. Comput Methods Appl Mech Engrg201–204:13–24, 2012.
NRC Symposium on VVUQ AT Patera 40
Parametric Manifold Models
AK Noor, Recent advances in reduction methods for nonlinear problems.Comput Struct 13:31–44, 1981.
BO Almroth, P Stern, FA Brogan, Automatic choice of global shape functions instructural analysis. AIAA J 16:525–528, 1978.
TA Porsching, Estimation of the error in the reduced basis method solution ofnonlinear equations. Math Comput 45(172):487–496, 1985.
L Machiels, Y Maday, IB Oliveira, AT Patera, DV Rovas, Output Bounds forReduced-Basis Approximations of Symmetric Positive Definite EigenvalueProblems. CR Acad Sci Paris Series I 331:153–158, 2000.
G Rozza, DBP Huynh, AT Patera, Reduced Basis Approximation and APosteriori Error Estimation for Affinely Parametrized Elliptic Coercive PartialDifferential Equations — Application to Transport and Continuum Mechanics.Arch Comput Methods Eng 15(3):229–275, 2008.
M Barrault, Y Maday, NC Nguyen, AT Patera, An ‘Empirical Interpolation’Method: Application to Efficient Reduced-Basis Discretization of PartialDifferential Equations. CR Acad Sci Paris Series I 339:667–672, 2004.
NRC Symposium on VVUQ AT Patera 41
ATP Research Group:
augustine.mit.edu
AFOSR/OSD/MURI:
www.dam.brown.edu/AFOSR MURI 2009.htm
NRC Symposium on VVUQ AT Patera 42