A goal-oriented reduced basis method for wave equation in inverse analysis
K. C. Hoang*, P. Kerfriden, SPA. Bordas
Institute of Mechanics and Advanced Materials (IMAM) Cardiff School of Engineering
Cardiff University
25 – 27 March 2013
Acknowledgments
Collaborators
P Kerfriden SPA Bordas
Sponsors:
European Research Council Starting Independent Research Grant (ERC Stg grant agreement No. 279578) entitled“Towards real time multiscale simulation of cutting in non-linear materials with
applications to surgical simulation and computer guided surgery”.
Outline
Motivation Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results Inverse Analysis
Numerical Results Conclusion
Contents
Motivation
Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results Inverse Analysis
Numerical Results Conclusion
Motivation Problem Description
The in-vitro dental implant model
Real-time evaluation?
Purpose: rapid identification of the material properties of the interfacial tissue in dental implant systems by NDE (nondestructive evaluation) in the time domain.
Motivation Problem Description
FEM model
[Hoang, K. C., et al. “Rapid identification of material properties of the interface tissue in dental implant systems using reduced basis method.” (2013).]
Motivation Problem Description
Applied force & output
0 0.2 0.4 0.6 0.8 1
x 10−3 0
0.2 0.4 0.6 0.8 1 1.2
Time (s)
Applied force (N)
0 0.2 0.4 0.6 0.8 1
x 10−3
−10
−5 0 5x 10−3
Time (s) Displacement sx (mm)
Our code ABAQUS
Motivation Problem Description
Measurement protocol & parametrized model
Given displacement measurement at output area, determine Young’s modulus E
Rayleigh damping coefficient β
of the interfacial tissue (which caused such displacement)?
===============
Input parameters: µ≡(µ1, µ2) = (E, β) µ∈ D ≡[1,25]MPa×[5×10−6,5×10−5] Field variable: displacement field u(µ, t)
Output: displacement responses at output area s(µ, t)
Motivation Problem Description
Governing equation
Dynamic linear elasticity equilibrium
∂σij
∂xj
+bi =ρ∂2ui
∂t2
Constitutive relation (isotropic material)
σij =Cijkl
∂uk
∂xl +β ∂
∂t
∂uk
∂xl
Initial conditions: ui(x,0) = 0 and ∂ui
∂t (x,0) = 0 Dirichlet boundary condition: ui = 0 on∂ΩD
Motivation Problem Description
Problem definition
Forward Problem
GivenE and β (and applied forceg(t)) calculate output s(µ, tk), 1 ≤k ≤K.
Inverse Problem
Given (noisy) experimental measurements at output point
“determine” the unknown parametersµ∗ = (E∗, β∗).
Contents
Motivation
Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results Inverse Analysis
Numerical Results Conclusion
Contents
Motivation Methodology
Finite Element Approximation
Reduced Basis Approximation Numerical Results
Inverse Analysis Numerical Results Conclusion
Contents
Motivation Methodology
Finite Element Approximation
Reduced Basis Approximation
Numerical Results Inverse Analysis
Numerical Results Conclusion
Methodology Reduced Basis method
Critical observation
To approximateu(µ, tk) and hence s(µ, tk), we need not represent every possible functionin Y.
Methodology Reduced Basis method
Critical observation...
Rather we pay our attention to the solution manifold Mu=
u(µ, tk),1 ≤k≤K|µ∈ D .
YN =spanζn =u(µn, tkn),1 ≤n≤N
Contents
Motivation Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results
Inverse Analysis Numerical Results Conclusion
RB Approximation Numerical Results
Parameters
Parameters:
µ≡(E, β) ∈ D
D ≡[1×106,25×106]Pa×[5×10−6,5×10−5]⊂RP=2.
Configuration:
9479 nodes with 50388linear tetrahedral elements.
N = 26343, ∆t= 2×10−6s, T = 1×10−3s, K = 500.
kwk2Y =a(w, w; ¯µ) +m(w, w; ¯µ)
¯
µ= (13×106Pa,2.75×10−5).
Qm = 1, Qa = 2, Qc= 2.
RB Approximation Numerical Results
RB outputs
RB outputs with N = 2,10; µtest = (10×106Pa,1×10−5).
0 0.2 0.4 0.6 0.8 1
x 10−3
−10
−5 0 5x 10−3
Time (s)
Displacement (mm)
sRB sFEM
0 0.2 0.4 0.6 0.8 1
x 10−3
−10
−5 0 5x 10−3
Time (s)
Displacement (mm)
sRB sFEM
RB Approximation Numerical Results
Convergent rate (GO POD-Greedy)
We consider:
Ξtrain ∈ D, ntrain = 1225.
D ≡[1×106,25×106]Pa×[5×10−6,5×10−5].
M = 5, Nmax= 100.
Max relative RB errors:
max,relu = max
µ∈Ξtrainu(µ); max,rels = max
µ∈Ξtrains(µ).
N max,relu max,rels 10 2.6273e-01 6.5563e-02 20 2.2871e-01 3.3993e-03 40 6.2021e-02 4.2840e-04 60 4.1866e-02 1.6040e-04 80 2.0117e-02 6.7330e-05
RB Approximation Numerical Results
Computational time
N tRB(online) tF EM α =tF EM/tRB(online) (sec) (sec)
10 0.0172 29 1686
30 0.0202 29 1435
40 0.0222 29 1306
60 0.0295 29 983
(Offline computational time to run the POD–Greedy procedure ≈24 hrs.)
0CPU: processor Intel(R) Core(TM) i7-3930K CPU @3.20GHz 3.20GHz, RAM 32GB, 64-bit Operating System.
Contents
Motivation Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results
Inverse Analysis
Numerical Results Conclusion
Methodology Inverse analysis
Optimization problem
Find the “unknown” parameter µ∗ that minimizes the objective function:
S(µ) =
K
X
i=1
[sN,i(µ)−smeasurei ]2 =rTr,
where the residual
ri =sN,i−smeasurei . Parametersµ∗ that minimizes S(µ) should satisfy
K
X
i=1
2∂sN,i
∂µj (sN,i−smeasurei ) = 2∂rT
∂µjr = 2JTr = 2v = 0, j = 1,2.
Jacobian matrix Jij = ∂ri
∂µj, i= 1, . . . , K; j = 1,2, . . .
Contents
Motivation Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results Inverse Analysis
Numerical Results
Conclusion
Methodology GA + Levenberg-Marquardt
Confidence regions
95% confidence ellipses:
µmeasure≡(8×106Pa,8×10−6).
7.985 7.99 7.995 8 8.005 8.01 8.015
x 106 7.9
7.92 7.94 7.96 7.98 8 8.02 8.04 8.06 8.08x 10−6
E (Pa)
β
7.99 7.995 8 8.005 8.01
x 106 7.94
7.96 7.98 8 8.02 8.04 8.06 8.08
8.1x 10−6
E (Pa)
β
100 points 500 points 1000 points
(a) 500 random tests (b) Cases
0Randomly created {smeasure}−−→ {µLMF compute}. {µcompute}is then covered by a 95% confidence ellipse using the Principal Component Analysis
Methodology GA + Levenberg-Marquardt
More results...
Strue of regular(4×4) grid pattern over [1×106,25×106Pa]×[5×10−6,5×10−5].
0 0.5 1 1.5 2 2.5
x 107 1
2 3 4 5 6x 10−5
E (Pa)
β
0 0.5 1 1.5 2 2.5
x 107 1
2 3 4 5 6x 10−5
E (Pa)
β
(a) pe = 7%noise added (b) pe = 20% noise added.
Methodology GA + Levenberg-Marquardt
Computational time
Number of Average number Number of RB calls Total RB calls by GA of LM iterations in each LM iteration RB calls
200 40 3 m= 320
Total RB calls CPU time for each solver Total computation time1 m= 320 tF EM 29 (sec) m×tF EM 155 (mins)
tRB(online) 0.3405 (sec) m×tRB(online) 7.104 (sec)
0Computational time for a (GA+LMF) model using FEM and RB as forward solvers (for one particularµmeasure).
Contents
Motivation Methodology
Finite Element Approximation Reduced Basis Approximation
Numerical Results Inverse Analysis
Numerical Results