Introduction ` a la m´ ethode de sensibilit´ e topologique pour l’identification de d´ efauts.
Aspects th´ eoriques et num´ eriques
Marc Bonnet1
1UMA (Dept. of Appl. Math.), POems, UMR 7231 CNRS-INRIA-ENSTA ENSTA, PARIS, France
mbonnet@ensta.fr
Formation Probl`emes inverses - Contrˆole des EDP, Amiens, 5-6 septembre 2011
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 1 / 148
Plan
1. Introduction
2. Time-independent (Laplace, elasticity)
3. Waves, frequency domain (Helmholtz, elastodynamics) 4. Waves, time domain
5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
To-do list
◮ champ total dans lippmann-schwinger
◮ estimations au moyen deδa
◮ separer inner and outer expansions
◮ definir¯v mais pas¯u
◮ define polarization tensor from far-field representation
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 3 / 148
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) Helmholtz
Elastodynamics
Heuristics and partial justification for limiting situation Numerical examples
4. Waves, time domain Numerical examples
Experimental studies (Dominguez and Gibiat, Tixier and Guzina) 5. Crack identification
Numerical examples
6. Higher-order topological sensitivity Numerical examples
7. Further reading
Forward and inverse problems
Ωtrue= Ω\Btrue
S
q Sobs
Sobs
Γtrue Btrue
JLS(ΩΓ) = 1 2
Z
Sobs
|uΓ−uobs(ξ)|2 dΩξ
J(ΩΓ) = Z
Sobs
ϕ(uΓ(ξ),ξ) dΩξ
where uobs(ξ) =utrue(ξ)|Sobs+ (noise)
◮ Insulated or conductive inhomogeneities (heat transfer, electrostatics)
◮ Sound-hard or penetrable objects (linear acoustics)
◮ Cavities, inclusions or cracks (elasticity)
◮ Electromagnetic inverse scattering ...
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 5 / 148
Background
◮ Extensive literature on (acoustic, electromagnetic, elastic) inverse scattering e.g. books by Colton & Kress (93), Ramm (92), Cakoni & Colton (06), Kirsch (96), Kirsch and Grinberg (08)
◮ Iterative solution techniques (optimization-based, Newton): facilitated by sensitivity formulae, e.g.:
δJ = Z
Γ
b(u,ˆu;C, ρ)vndS (shape sensitivity) withˆu: adjoint solutionsuch thatσ[ˆu].n=∂uϕ
vn: normal perturbation of unknown surface Γ
Extension to combined shape-material sensitivity: b(u,u;ˆ C, ρ,∆C,∆ρ)
◮ Combined optimization and level-set approaches (Dorn and Lesselier 06, topical review)
MBEng. Anal. with Bound. Elem.(1995)
S. Nintcheu Fata, B. Guzina, MB,Comput. Mech.(2003) MB and B. Guzina,J. Comput. Phys. (2008)
Optimization-based identification: example (inclusion, elastodynamics)
0 10 20 30 40 50
Iteration k 0
0.2 0.4 0.6 0.8 1
Parameter
c1/c1true c2/c2true c3/c
3 true J/J0 Centroidal coordinates
0 10 20 30 40 50
Iteration k 0
0.2 0.4 0.6 0.8 1 1.2
Parameter
α1/α 1 true α2/α2true α3/α3true J/J0 Semi-axes
0 10 20 30 40 50
Iteration k -2
-1 0 1 2 3 4
Parameter
µ/µtrue ν/νtrue ρ/ρtrue J/J0
Material properties
^
−3
−1 −2
0 ξ1/d
−2
−1 ξ3/d 1.5
2 2.5
3 3.5
0 1 ξ2/d
Initial
True 0 True Initial
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
µ/µ
M. Bonnet, B. B. Guzina,J. Comput. Phys (2009)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 7 / 148
Background
◮ Iterative solution techniques:
Computationally expensive: ≥1forward solution(s) per iteration;
Success strongly depends on choice of initial guess.
◮ Non-iterative, fast preliminary ‘probing’ (or ‘sampling’) methods (Potthast 06, IP topical review):
(a) Linear sampling (Colton, Kirsch 96)
Extension to near-field elastodynamic data (Guzina, Nintcheu Fata 04);
(b) Topological sensitivity
Mathematical studies (Sokolowski-Zochowski 97, Guillaume-Sid Idris 04) Asymptotic methods (Maz’ya-Nazarov-Plamenevskij 91, Il’in 92,
Vogelius-Volkov 00, Ammari-Kang 04...)
Initially proposed for topological optimization (Eschenauer et al. 1994);
Inverse problems (Guzina, MB 04, Feijoo 04, Masmoudi 05, MB 06...)
Linear sampling
Sobs
S
B f
◮ Scattering operator:
uobs(x) = [R(B)f](x) (x∈Sobs)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 9 / 148
Linear sampling
z Sobs
S
B g
◮ Linear sampling equation (withG: Green’s tensor):
∃?g(ξ), kG(z;x).e−[R(B)gεz](x)k ≤ε (x∈Sobs) z∈B =⇒ lim
z→∂BkgεzkL2(S)=∞ z6∈B =⇒ lim
ε→0kgεzkL2(S)=∞
◮ Defect indicator function: z7→1/kgεzkL2(S)
Topological sensitivity
Sobs
S
B f
Ba(z) z z
◮ Small-obstacle expansion of cost function:
J(Ba(z)) =J(∅) +δ(a)T(z) +o(δ(a)) whereδ(a)quantifies the perturbation behavior abouta= 0;
T(z)is thetopological derivativeofJ; zacts as a sampling point.
◮ Heuristic idea: find regions inΩwhere field T ismost negative.
◮ Proposed defect indicator function:
z7→ T(z)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 11 / 148
Concept of topological sensitivity
◮ Sensitivity analysis toolinitially proposed for topological optimization (Eschenaueret al94; Sokolowski, Zochowski 99; Garreauet al01; Allaireet al05. . . )
Allaire, de Gournay, Jouve, Toader (2005)
Example (Topological optimisation)
0 20 40 60 80 100
Iterations
4 4.5 5 5.5 6
Compliance
Th`ese X-EADS “Optimisation g´eometrique et topologique du drapage des composites” (2010–
2013, G. Delgado, dir. G. Allaire) — Algorithm combining topological and shape derivatives
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 13 / 148
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) Helmholtz
Elastodynamics
Heuristics and partial justification for limiting situation Numerical examples
4. Waves, time domain Numerical examples
Experimental studies (Dominguez and Gibiat, Tixier and Guzina) 5. Crack identification
Numerical examples
6. Higher-order topological sensitivity Numerical examples
7. Further reading
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) 4. Waves, time domain
5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 15 / 148
Background problem (BP)
◮ Media endowed with scalar (thermal, electrostatic...) conductivityκ Generalization to anisotropic (tensorial) conductivity not addressed, but relatively straightforward
◮ Primary field variableuis scalar (temperature, electrostatic potential...) Background problem (κ: background conductivity): defines the physical state in the absence of inclusion
div (κ∇u) +f = 0 inΩ, κ∂nu=g onSN, u=uDon SD
i.e., in weak form:
Findu∈W(uD), hu,wiκΩ=F(w), ∀w∈W0
withW(uD) =
w∈H1(Ω), w=uDonSD . W0=W(0)and hu,wiκΩ:=
Z
Ω
κ∇u·∇w dV, F(w) :=
Z
Ω
fw dV+ Z
SN
gwdS
Generic transmission problem (TP)
◮ Generic (possibly multiple) inclusionB⊂Ωand conductivityκB: κB = [1−χ(B)]κ+χ(B)κ⋆=κ+χ(B)∆κ
◮ Transmission problem:
div (κB∇uB) +f = 0inΩ, κ∂nu=g on SN, u=uDonSD
i.e., in weak form:
FinduB∈W(uD), huB,wiκΩ+huB,wi∆κB =F(w), ∀w∈W0
◮ Both formulations implicitly enforce the transmission conditions uB |+=uB |− and κ∂nuB |+=κ⋆∂nuB |− on∂B
◮ TP in terms of solution perturbationvB :=uB−u:
FindvB∈W0, hvB,wiκΩ+hvB,wi∆κB =−hu,wi∆κB , ∀w∈W0
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 17 / 148
Free-space transmission problem (FSTP)
◮ Auxiliary TP for an inclusionBembedded in an infinite mediumΩ =R3 subjected to a given remote uniform gradientE∈R3 through the background solutionu(ξ) =E·ξ=:ψ[E](ξ), i.e.:
( div (κB∇uB[E]) = 0inR3,
uB[E](ξ)−ψ[E](ξ) =O(kξk−2)askξk → ∞ withκB:=κ+χ(B)∆κ.
◮ Weak formulation for the solution perturbationvB:=uB−ψ[E]:
FindvB∈H1(R3), hvB,wiκR3 =−hψ[E],wi∆κB , ∀w∈H1(R3)
Cost functional
Consider generic cost functionals of the form J(κB) =J(uB) with J(w) :=
Z
Ω
ϕV(·,w) dV + Z
∂Ω
ϕS(·,w) dS with twice-differentiable densitiesw 7→ϕV(·,w)andw 7→ϕS(·,w)
◮ Example 1: weighted least-squares output residuals (identification based on overdetermined datag,uobs onSN)
J(w) = 1 2
Z
SN
Q|w−uobs|2dS
◮ Example 2: potential energy (expressed as boundary integrals) J(w) = 1
2 Z
SD
(κ∂nw)uDdS−1 2
Z
SN
gwdS
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 19 / 148
Concept of topological sensitivity
SN
SD
Ba
Ω⊂RD
(κ) Ba z
(κ⋆)
Ba(z) =z+aB
Definition (Topological derivative)
Assume thatJ(ua)can be expanded in the form
J(κa) =J(κ) +δ(a)T(z) +o(δ(a))
whereδ(a)is such thatδ(a) =o(1)asa→0 and characterizes the small-inclusion behavior ofJ(κB).
Then, the coefficientT(z)is called the topological derivative ofJ atz∈Ω.
◮ T(z)also depends on shapeB and conductivity contrastβ:=κ⋆/κ;
◮ Terminology forT varies, with “gradient” or “sensitivity” sometimes used instead of “derivative”.
Small-inclusion asymptotic behavior of
uaEstablished using a domain integral equation (DIE) formulation of supportBa
◮ The Green’s functionG is defined by div κ∇G(·,x)
+δ(· −x) = 0 in Ω G(·,x) = 0 on SD κ∂nG(·,x) = 0 on SN
(x∈Ω)
and verifies
hG(·,x),wiκΩ=w(x) x∈Ω, ∀w∈W0∩C0(Ω)
◮ SplitG into singular and nonsingular parts:
G(ξ,x) =G(ξ−x) +GC(ξ,x)
whereG(r) = 1/(4πkrk)is the (singular) full-space fundamental solution and the complementary Green’s functionGC is bounded atξ=x (and in factC∞ forξ,x∈Ωby virtue of solving a BVP with regular boundary data).
◮ Note thatG and∇G are homogeneous of degree -1 and -2, respectively:
G(λr) =|λ|−1G(r), ∇G(λr) =|λ|−2sgn(λ)∇G(r)
(r∈R3\ {0}, λ∈R\ {0})
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 21 / 148
Small-inclusion asymptotic behavior of
ua◮ Domain integral equation (DIE) for the TP:
combine weak formulations for the TP (withw =G(·,x)) and the Green’s function (withw =va):
(hva,G(·,x)iκΩ+hva,G(·,x)i∆κBa =−hu,wi∆κBa
hG(·,x),vaiκΩ=va(x) (x∈Ω) to obtain
La[va](x) =−hu,G(·,x)i∆κBa
x∈Ba (integral equation forva) x∈Ω\B¯a (integral representation forva) with the linear operatorLa defined by La[w](x) :=w(x) +hw,G(·,x)i∆κBa i.e. (in expanded form):
La[w](x) =w(x) + Z
Ba
∆κ∇w·∇G(·,x) dV
◮ DIE for the FSTP:(similarly)
LB[vB](x) =−hψ[E],G(·,x)i∆κB with LB[w](x) :=w(x) +hw,G(·,x)i∆κB
Small-inclusion asymptotic behavior of
ua◮ Perturbation due to small inclusion governed by domain integralof supportBa
La[va](x) =−hu,G(·,x)i∆κBa Facilitates investigation of limiting behavior asa→0
◮ Introduce normalized coordinates
(a) (ξ,x) =z+a(¯ξ,¯x), (b) dVξ =a3d ¯Vξ¯ (ξ∈Ba,¯ξ∈ B)
Asymptotic behavior of va asa→0: InsideBa, one has (inner expansion) va(x) =avB[∇u(z)](¯x) +o(a) ξ∈Ba,ξ¯∈ B
Moreover, at any fixed locationx6=z (outer expansion), one has va(x) =−a3∇1G(z,x)·A·∇u(z) +o(a3) x6=z
whereA=A(κ, κ⋆,B)is the (second-order) polarization tensor, defined by A·E=
Z
B
∆κ∇uB[E] dV = Z
B
∆κ(E+∇vB[E]) dV ∀E∈R3 (uB[E] =ψ[E] +vB[E]: solution to the FSTP)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 23 / 148
Small-inclusion asymptotic behavior of
ua(proof sketch 1/2)
Proof based on seeking and exploiting the limiting form of the DIE, and using scaled coordinates.(i) Asymptotic behavior of the Green’s function:
∇1G(ξ,x) =a−2∇G(¯ξ−¯x) +∇1GC(z,z) +o(1) (ii) Defining¯va(¯ξ) :=va(z+a¯ξ), ¯u(¯ξ) :=u(z+a¯ξ), one finds
∇va(ξ) =a−1∇v¯a(¯ξ), ∇¯u(¯ξ) =∇u(z) +o(1) (iii) Substitute above expansions and rescale coordinates in DIE:
La[va](x) =LB[¯va](¯x) +aO(k∇v¯ak2,B) hu,G(·,x)i∆κBa =ahψ[∇u(z)],G(·,¯x)i∆κB +o(a) (iv) Make ansatz ¯va(¯ξ) =aV(¯ξ) +o(a)and isolate lowest-order (ina)
contributions in DIE, which correspond to the FSTP:
LB[V](¯ξ) =−hψ[∇u(z)],G(·,¯x)i∆κB Hence: V(¯ξ) =vB[∇u(z)](¯ξ) ¯ξ∈ B
Small-inclusion asymptotic behavior of
ua(proof sketch 2/2)
(v) Now, considerx6=zandasmall enough. In that case, one has∇1G(ξ,x) =∇1G(z,x) +o(1), ξ∈Ba,x6∈Ba. (vi) Substitute¯va(¯ξ) =aV(¯ξ) +o(a)and the above into the integral
representation outsideBa, to obtain va(x) =−a3∇1G(z,x)·
Z
B
∆κ∇vB[∇u(z)](¯ξ) d ¯Vξ¯+o(a3)
=−a3∇1G(z,x)·A·∇u(z) +o(a3)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 25 / 148
Topological derivative of cost functionals
(i) Expand cost functional w.r.t. va:J(κa) =J(κ) +J′(u;va) +o(kvak2,Ω+kvak2,∂Ω) with J′(u;w) :=
Z
Ω
∇2ϕV(·,u)w dV+ Z
SN
∇2ϕS(·,u)wdS (ii) Interpret J′(u;va)as component of a weak formulation with test function va
and define theadjoint solutionˆuby
Find ˆu∈W0, hˆu,wiκΩ=J′(u;w), ∀w∈W0
(iii) Reciprocity between transmission and adjoint problems:
hva,uiˆ κΩ+hva,ˆui∆κBa =−hu,uiˆ ∆κBa , hˆu,vaiκΩ=J′(u;va) to obtain J′(u;va) =−hˆu,ui∆κBa − hˆu,vai∆κBa =−hˆu,uai∆κBa (iv) Exploit known limiting behavior ofva in Ba (inner expansion), yielding
J′(u;va) =−hˆu,uai∆κBa =−a3∇u(z)·A·ˆ ∇u(z) +o(a3) Topological expansion ofJ: The topological expansion ofJ is given by
J(κa) =J(κ) +a3T(z) +o(a3), T(z) =−∇ˆu(z)·A·∇u(z)
Remarks on topological derivative and polarization tensor
(i) Ais symmetric and has the same sign as∆κ(Cedyo-Fengya, Volkov,Vogelius, 1998);
(ii) By superposition:
V(¯x) =vB[∇u(z)](¯x) = X3
i=1
ei·∇u(z)
vB[ei](¯x) One thus needs only compute thevB[ei], which are independent onz.
(iii) Spherical inclusion(i.e. Bis unit sphere): FSTP solution is given by vB[E](¯x) = 3
2 +βE·¯x withβ:=κ⋆/κ (¯x∈ B).
Hence:
A(κ, κ⋆,B) = 4πκ(β−1) β+ 2 I ;
(iv) Analytical FSTP solutions available for other shapes (e.g. ellipsoids);
otherwise, solve numerically (once and for all) the normalized DIE with E=ei (i= 1,2,3);
(v) Special caseβ= 0: insulated(i.e. impenetrable) inclusion.
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 27 / 148
Remarks on topological derivative and polarization tensor
(vi) Similar analysis available for (penetrable or insulated) inclusions in 2Dmedia, yielding cost function asymptotics of the form J(κa) =J(κ) +a2T(z) +o(a2)
(vii) If homogeneous Dirichlet conditions are instead considered on∂Ba, one finds cost function asymptotics of the form (e.g. Guillaume, Sid Idris, 2002)
J(κa) =J(κ) +aT(z) +o(a) (three-dimensional case)
=J(κ)− 1
lnaT(z) +o−1 lna
(two-dimensional case) (viii) Cost function example 1 (output least squares): adjoint loads are (as usual)
measurement residuals, since J(uB) =1
2 Z
SN
Q|uB−uobs|2dS =⇒ J′(u;w) = Z
SN
Q(w−uobs)wdS (ix) Cost function example 2 (potential energy, withuD= 0):
J(uB) =−1 2
Z
SN
guBdS =⇒ J′(u;w) =−1 2
Z
SN
gwdS
=⇒ uˆ=−u/2 =⇒ T(z) = (1/2)∇u·A·∇u Use sign properties ofAfor interpretation
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) 4. Waves, time domain
5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 29 / 148
Background solution
Background solution (C: background elasticity tensor):
div (C:ε[u]) +f=0in Ω, t[u] =gonSN, u=uDonSD
where
ε[u] = 1
2(∇u+∇uT) (linearized strain tensor), t[u] = (C:ε[u])·n (traction vector)
In weak form:
Findu∈W(uD), hu,wiCΩ=F(w), ∀w∈W0 withW(uD) :=
v∈H1(Ω;R3),v=uDonSD . W0=W(0)and hu,wiCD :=
Z
D
ε[u] :C:ε[w] dV = Z
D
∇u:C:∇wdV, F(w) =
Z
Ω
f·wdV+ Z
SN
g·wdS
Generic transmission problem (TP)
◮ Generic TP:(possibly multiple) inclusionB⊂Ωand elasticity tensor CB = [1−χ(B)]C+χ(B)C⋆=C+χ(B)∆C
div (CB:ε[uB]) +f=0in Ω, t[uB] =gonSN, uB =uDonSD
i.e., in weak form:
Findua∈W(uD), hua,wiCΩ+hua,wi∆CBa =F(w), ∀w∈W0
Both formulations implicitly enforce the perfect-bonding conditions ua|+=ua|− and t
ua|+
=t⋆ ua|−
on∂B TP in terms of solution perturbationvB :=uB−u:
FindvB ∈W0, hvB,wiCΩ+hvB,wi∆CB =−hu,wi∆CB , ∀w∈W0
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 31 / 148
Free-space transmission problem (FSTP)
◮ Auxiliary TP for an inclusionBembedded in an infinite mediumΩ =R3 subjected to a given remote uniform strainE∈R3,3 through the background solutionu(ξ) =E·ξ=:ψ[E](ξ), i.e.:
( div (CB:∇uB[E]) =0 inR3, uB[E](ξ)−ψ[E](ξ) =O(kξk−2) askξk → ∞ withCB:=C+χ(B)∆C.
◮ Weak formulation for the solution perturbationvB:=uB−ψ[E]:
FindvB∈H1(R3;R3), hvB[E],wiCR3 =−hψ[E],wi∆CB , ∀w∈H1(R3;R3)
Cost functional
Cost functional:J(CB) =J(uB) with J(w) :=
Z
Ω
ϕV(·,w) dV + Z
∂Ω
ϕS(·,w) dS with twice-differentiable densitiesw7→ϕV(·,w)andw7→ϕS(·,w)
Topological derivative ofJ: sought by considering the limiting behavior of j(a) :=J(Ca) =J(ua).
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 33 / 148
Small-inclusion asymptotic behavior of
ua◮ The elastostatic Green’s tensor is defined by div C:ε[G(·,x)]
+δ(· −x)I=0 in Ω G(·,x) =0 onSD
t[G(·,x)] =0 onSN
(x∈Ω)
and verifies
hG(·,x),wiCΩ=w(x) x∈Ω, ∀w∈W0∩C0(Ω)
◮ SplitG into singular and nonsingular parts:
G(ξ,x) =G(ξ−x) +GC(ξ,x)
whereG(r)is the (singular) full-space (Kelvin) fundamental solution and the complementary Green’s tensorGC is bounded atξ=x (and in factC∞for ξ,x∈Ωby virtue of solving a BVP with regular boundary data).
Small-inclusion asymptotic behavior of
ua◮ Applying 3-D Fourier transform, one has in fact G(r) = 1
(2π)3 Z
R3
exp(iη·r)N(η) dV(η) (r∈R3\ {0}) withN(η) =K−1(η)in terms of theacoustic tensorK(η), defined by Kik(η) =Cijkℓηjηℓ
◮ This implies that∇G is homogeneous of degree -2:
∇G(λr) =|λ|−3λ∇G(r) =|λ|−2sgn(λ)∇G(r)
(r∈R3\ {0}, λ∈R\ {0})
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 35 / 148
Small-inclusion asymptotic behavior of
ua◮ Domain integral equation (DIE) for the TP:
combine weak formulations for the TP (withw =G(·,x)) and the Green’s function (withw =va):
(hvB,G(·,x)iCΩ+hvB,G(·,x)i∆CBa =−hu,G(·,x)i∆CBa
hG(·,x),vaiCΩ=va(x) (x∈Ω) to obtain
La[va](x) =−hu,G(·,x)i∆CBa
x∈Ba (integral equation forva) x∈Ω\B¯a (integral representation forva) with the linear operatorLa defined by La[w](x) =w(x) +hw,G(·,x)i∆CBa i.e. (in expanded form):
La[w](x) =w(x) + Z
Ba
∇w: ∆C:∇G(·,x) dV
◮ DIE for the FSTP:(similarly)
LB[vB](x) =−hψ[E],G(·,x)i∆CB with LB[w](x) :=w(x) +hw,G(·,x)i∆CB
Small-inclusion asymptotic behavior of
uaAsymptotic behavior of va asa→0: InsideBa, one has (inner expansion) va(x) =avB[∇u(z)](¯x) +o(a) x∈Ba,¯x∈ B
Moreover, at any fixed locationx6=z (outer expansion), one has va(x) =−a3∇1G(z,x) :A:∇u(z) +o(a3) (x6=z)
whereA=A(C,C⋆,B)is the (fourth-order) elastic moment tensor, defined by A:E=
Z
B
∆C:∇uB[E] dV = Z
B
∆C: (E+∇vB[E]) dV ∀E∈R3,3 (uB[E] =ψ[E] +vB[E]: solution to the FSTP)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 37 / 148
Small-inclusion asymptotic behavior of
ua(proof sketch 1/2)
Proof based on seeking and exploiting the limiting form of the DIE, and using scaled coordinates.(i) Asymptotic behavior of the Green’s tensor:
∇1G(ξ,x) =a−2∇G(¯ξ−¯x) +∇1GC(z,z) +o(1) (ii) Defining¯va(¯ξ) :=va(z+a¯ξ),¯u(¯ξ) :=u(z+a¯ξ), one finds
∇¯va(¯ξ) =a−1∇va(ξ), ∇¯u(¯ξ) =∇u(z) +o(1) (iii) Substitute above expansions and rescale coordinates in DIE:
La[va](x) =LB[¯va](¯x) +aO(k∇¯vak2,B) hu,G(·,x)i∆CBa =ahψ[∇u(z)],G(·,¯x)i∆CB +o(a) (iv) Make ansatz ¯va(¯ξ) =aV(¯ξ) +o(a)and isolate lowest-order (ina)
contributions in DIE, which correspond to the FSTP:
LB[V](¯ξ) =−hψ[∇u(z)],G(·,¯x)i∆CB Hence: V(¯ξ) =vB[∇u(z)](¯ξ) ξ¯∈ B
Small-inclusion asymptotic behavior of
ua(proof sketch 2/2)
(v) Now, considerx6=zandasmall enough. In that case, one has∇1G(ξ,x) =∇1G(z,x) +o(1), ξ∈Ba,x6∈Ba. (vi) Substitute¯va(¯ξ) =aV(¯ξ) +o(a)and the above into the integral
representation outsideBa, to obtain va(x) =−a3∇1G(z,x) :
Z
B
∆C:∇uB[∇u(z)](¯ξ) d ¯Vξ¯+o(a3)
=−a3∇1G(z,x) :A:∇u(z) +o(a3)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 39 / 148
Topological derivative of cost functionals
(i) Expand cost functional w.r.t. va:J(Ca) =J(C) +J′(u;va) +o(kvak2,Ω+kvak2,∂Ω) with J′(u;w) :=
Z
Ω
∇2ϕV(·,u)·wdV + Z
SN
∇2ϕS(·,u)·wdS (ii) Interpret J′(u;va)as component of a weak formulation with test functionva
and define theadjoint solutionˆuby
Find ˆu∈W0, hˆu,wiCΩ=J′(u;w), ∀w∈W0
(iii) Reciprocity between transmission and adjoint problems:
hva,uiˆ CΩ+hva,ˆui∆CBa =−hu,uiˆ ∆CBa , hˆu,vaiCΩ=J′(u;va) to obtain J′(u;va) =−hˆu,ui∆CBa − hˆu,vai∆CBa =−hˆu,uai∆CBa (iv) Exploit known limiting behavior ofva in Ba (inner expansion), yielding
J′(u;va) =−hˆu,uai∆CBa =−a3∇ˆu(z) :A:∇u(z) +o(a3) Topological expansion ofJ: The topological expansion ofJ is given by
J(Ca) =J(C) +a3T(z) +o(a3), T(z) =−∇u(z) :A:∇u(z)ˆ
Elastic moment tensor (anisotropic elastic inclusion)
1. The EMTAhas major and minor symmetries: for any (E,E′)∈R3,3×R3,3, one has
E′:A:E=E:A:E′ (major symmetry) E′:A:E=E′:A:ET =E′T:A:E (minor symmetry) 2. By superposition, one has
VB[∇u(z)] = X
1≤i≤j≤3
Eij:∇u(z) VB[Eij]
with Eij=ei⊗ei (i=j), Eij= 2−1/2(ei⊗ej+ej⊗ei) (i6=j) One thus needs only compute thevB[Eij](6 auxiliary solutions independent onz) once and for all.
3. C⋆=0, i.e. ∆C=−C, corresponds to atraction-free cavity.
Thus, small-cavity asymptotics is special case of small-inclusion asymptotics.
Beretta, Bonnetier, Francini, Mazzucato (2011); B, Delgado (in preparation, 2011) Isotropic case: Ammari, Kang, Nakamura, Tanuma (2002); Guzina, Chikichev (2006)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 41 / 148
Elastic moment tensor (anisotropic elastic inclusion)
4. LetBbe an ellipsoid (of principal semiaxesa1,a2,a3). The Eshelby tensor S(a1,a2,a3)ofBis given by (Mura, 1987):
Sijmn(a1,a2,a3) = 1 8πCpqmn
Z 1
−1
du Z 2π
0
Gipjq(η(u, θ)) +Gjpiq(η(u, θ)) dθ
Gijkℓ(η) :=ηkηℓNij(η), η(u, θ) :=p
1−u2[a−11 cosθe1+a−12 sinθe2] +a−13 ue3
Then,Ahas the explicit expression (valid for arbitrary anisotropicC andC⋆) A(C,C⋆,a1,a2,a3) =C:
C+ ∆C:S(a1,a2,a3)−1
: ∆C
S(a1,a2,a3)known in closed form forisotropicC (elliptic integrals involved in the case of ellipsoids).
5. For ellipsoidal cavities, one has A(C,a1,a2,a3) =
S(a1,a2,a3)−I−1
:C
6. Closed-form expression for spherical cavity (i.e. whenBis unit sphere):
A=−2µh 3(1 +ν)
2(1−2ν)2J +15(1−ν) 7−5ν Ki
(3J =I⊗I, K=I−J)
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) 4. Waves, time domain
5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 43 / 148
Numerical example
Ba Ω
ΓD g
ΓN
L= 2 H= 1
α= arctan(0.5) O
z= (1/3,2/3)
J(Ba) =−1 2
Z
SN
g·uadS (compliance)
Isotropic tests: C=
1.34 0.57 0.
0.57 1.34 0.
0. 0. 0.38
, C⋆= 10−9C
Anisotropic tests: C=
1. 0.5 0.
0.5 2. 0.
0. 0. 0.04
, C⋆=
3. 0.4 0.
0.4 1.5 0.
0. 0. 0.03
.
B, Delgado (in preparation, 2011)
Numerical example
0 0.02 0.04 0.06 0.08 0.1 0.12
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Characteristic size a
Relative error e(a)
Anisotropic, a1/a2=2, angle = pi/4
Isotropic, a1/a2=1
Anisotropic, a1/a2=2, angle = 0 Isotropic, a1/a2=2, angle = 0 Isotropic, a1/a2=2, angle = pi/4
e(a) = |J(Ca)−J(C)−a2T(z)|
|a2T(z)|
B, Delgado (in preparation, 2011)
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 45 / 148
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) 4. Waves, time domain
5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
Energy-based cost functional
◮ ’Neumann’ and ’Dirichlet’ background fields:
FinduN∈WN(uD), huN,wiCΩ=Fno(w) +Fo(w), ∀w∈W0N. FinduD∈WD(uD), huD,wiCΩ=Fno(w), ∀w∈W0D. whereWN(uD) =W(uD),W0N=W0and
WD(uD) :=
v∈H1(Ω;R3),v=uDonSD, v=uobson So , W0D:=WD(0)
◮ ’Neumann’ and ’Dirichlet’ fields for small trial inclusion:
FindvNa ∈W0N, hvNa,wiCΩ+hvNa,wi∆CBa =−huN,wi∆CBa, ∀w∈W0N. FindvDa ∈W0D, hvDa,wiCΩ+hvDa,wi∆CBa =−huD,wi∆CBa , ∀w∈W0D.
◮ Energy discrepancy functional:
E(Ca) =E(uNa,uDa,Ca) = 1
2huNa−uDa,uNa−uDaiCΩa
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 47 / 148
Energy-based cost functional
Topological derivative ofE: it is given by TE(z) = 1
2ε[uN](z) :A:ε[uN](z)−1
2ε[uD](z) :A:ε[uD](z)
◮ Energy-based cost functionals are very useful for identification of (possibly heterogeneous) material parameters, see e.g. Kohn, Vogelius (1987), Kohn, McKenney (1989), Constantinescu (1995), Chavent, Kunisch, Roberts (1996).
◮ Specific treatment needed to findTE, because (i)ϕV(·,∇w,CB)rather than ϕV(·,w), and (ii) va=O(a)butva =O(1)insideBa (so care needed in doing expansions abouta= 0).
B, Delgado (in preparation, 2011)
1. Introduction
2. Time-independent (Laplace, elasticity) Scalar (conductivity) problems
Elasticity
Numerical example
Energy-based cost functional
3. Waves, frequency domain (Helmholtz, elastodynamics) Helmholtz
Elastodynamics
Heuristics and partial justification for limiting situation Numerical examples
4. Waves, time domain Numerical examples
Experimental studies (Dominguez and Gibiat, Tixier and Guzina) 5. Crack identification
Numerical examples
6. Higher-order topological sensitivity Numerical examples
7. Further reading
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 49 / 148
1. Introduction
2. Time-independent (Laplace, elasticity)
3. Waves, frequency domain (Helmholtz, elastodynamics) Helmholtz
Elastodynamics
Heuristics and partial justification for limiting situation Numerical examples
4. Waves, time domain 5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
Acoustic transmission problem
Scattering of an acoustic wave by a penetrable inclusionB (ρ⋆,c⋆) embedded in an acoustic mediumΩ(ρ,c).
β:=ρ/ρ⋆ mass density ratio γ:=c/c⋆ velocity ratio η:= (ρc2)/(ρc2)⋆ bulk modulus ratio Letm:= 1 + (β−1)χ(B)andn:= 1 + (η−1)χ(B).
◮ Field equation (assuming homogeneous inclusion and background) on acoustic pressurep:
div (m∇uB) +nk2uB+f = 0 (in Ω)
◮ Transmission conditions implied (continuity of pressure and normal velocity across∂B)
uB |+=uB |− and ρ−1∂nuB |+= (ρ⋆)−1∂nuB |− on∂B
Martin, P.A. (2003), Acoustic scattering by inhomogeneous obstacles. SIAM J. Appl. Math.
64:297-308
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 51 / 148
Weak formulation of TP
Notations: LetW(uD) :=w∈H1(Ω), w=uDonSD . W0=W(0)and hu,wimD :=
Z
D
m∇u·∇w dV (u,w)nD:=
Z
D
nuw dV hu,wiD :=
Z
D
∇u·∇w dV (u,w)D:=
Z
D
uwdV F(w) =
Z
Ω
fw dV+ Z
SN
gwdS
◮ Background solution(weak formulation):
Findu∈W(uD) hu,wiΩ−k2(u,w)Ω=F(w), ∀w∈W0
For unbounded-media idealizations (Ω =R3),uis an arbitrary solution to (∆ +k2)u= 0(e.g. a plane wave), in particular not constrained by radiation conditions.
◮ TP(weak formulation), in terms of the solution perturbation (scattered field) vB :=uB−u:
FindvB∈W0, hvB,wimΩB−k2(vB,w)nΩB =−hu,wi∆mB +k2(u,w)∆nB , ∀w∈W0
Cost functional
J(mB,nB) =J(uB) with J(w) :=
Z
Ω
ϕV(·,w) dV+ Z
∂Ω
ϕS(·,w) dS with twice-differentiable real-valued densitiesw 7→ϕV(·,w)andw 7→ϕS(·,w).
Example (output least-squares for defect identification):
ϕV(·,w) = 0, ϕS(·,w) =1
2(w−uobs)Q(·)(w−uobs)
Topological derivative ofJ: sought by considering the limiting behavior of j(a) :=J(ma,na) =J(ua)
withma:= 1 + (β−1)χ(Ba)andna:= 1 + (η−1)χ(Ba)
Ba(z)
z p
SN SD
Sobs
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 53 / 148
Small-inclusion asymptotic behavior of
ua◮ Green’s function: defined, forx∈Ω, by (∆ +k2)G(·,x;k) +δ(· −x) = 0 (in Ω),
G(·,x;k) = 0 (onSD), ∂nG(·,x;k) = 0 (onSN)
◮ SplitG into singular and nonsingular parts:
G(ξ,x;k) =G(ξ−x;k) +GC(ξ,x;k)
whereG(r;k)is the (singular) Helmholtz full-space fundamental solution:
G(r;k) = eikr
4πr (r =krk)
◮ Moreover, one has
G(r;k) =G(r) +GC(r;k)
whereG(r) = 1/(4πr)is the Laplace fundamental solution andGC(r;k)is nonsingular forr=0
Small-inclusion asymptotic behavior of
ua Domain integral equation (DIE) for the TP:Combining weak formulations for the TP (withw =G(·,x;k)) and the Green’s function (withw =va) yields a DIE of Lippmann-Schwinger type:
Lka[va](x) =−hu,G(·,x;k)i∆mBa +k2(u,G(·,x;k))∆nBa
x∈Ba (integral equation forva) x∈Ω\B¯a (integral representation forva) with the linear operatorLka defined by
Lka[w](x) =w(x) +hw,G(·,x;k)i∆mBa −k2(w,G(·,x;k))∆nBa i.e. (in expanded form):
La[w](x) =w(x) + Z
Ba
(β−1)∇w·∇G(·,x;k)−(η−1)k2wG(·,x;k) dV
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 55 / 148
Small-inclusion asymptotic behavior of
uaLka[va](x) =−hu,G(·,x)i∆mBa +k2(u,G(·,x))∆nBa (LS)
◮ Coordinate scaling:
ξ=z+a¯ξ, x=z+a¯x, dVξ =a3d ¯Vξ¯, ∇ξ=a−1∇ξ¯
◮ Scaling properties ofG(ξ,x;k)(by homogeneity properties ofG(ξ−x)):
G(ξ,x;k) =a−1G(¯ξ−¯x) +O(1)
∇1G(ξ,x;k) =a−2∇G(¯ξ−¯x) +O(1)
◮ Leading contribution in (LS) using ansatz va(ξ) =aV(¯ξ) +o(a) inBa : Lka[va](x) =aLB[V](¯x) +o(a1)
hu,G(·,x)i∆mBa −k2(u,G(·,x))∆nBa =ahψ[∇u(z)],G(·,¯x)i∆mB +o(a1) Hence, retaining only theO(a)terms,V solves again thestatic FSTP for the normalized inclusion:
LB[V](¯x) =−hψ[∇u(z)],G(·,¯x)i∆mB (¯x∈ B)
Small-inclusion asymptotic behavior of
ua InsideBa, one has (inner expansion)va(x) =avB[∇u(z)](¯x) +o(a) x∈Ba,¯x∈ B wherevB[E]is again the solution of thestaticFSTP.
Moreover, at any fixed locationx6=z (outer expansion), one has va(x) =−a3
∇1G(·,x;k)·A·∇u−k2(η−1)|B|G(·,x;k)u
(z) +o(a3) (x6=z) whereA=A(β,B)is again the polarization tensor associated with the normalized staticFSTP.
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 57 / 148
Topological derivative of cost functionals
(i) Expand cost functional w.r.t. va:J(ma,na) =J(1,1) +J′(u;va) +o(kvak2,Ω+kvak2,∂Ω)
(ii) Interpret J′(u;va)as component of a weak formulation with test function va and define theadjoint solutionˆuby
Find ˆu∈W0, hˆu,wiΩ−k2(ˆu,w)Ω=J′(u;w), ∀w∈W0
(iii) Reciprocity between transmission and adjoint problems:
hva,ˆuimΩa−k2(va,u)ˆ nΩa =−hu,ˆui∆mBa +k2(u,u)ˆ ∆nBa hˆu,vaiΩ−k2(ˆu,va)Ω=J′(u;va)
to obtain J′(u;va) =−hˆu,uai∆mBa +k2(ˆu,ua)∆nBa
(iv) Exploit known limiting behavior ofva in Ba (inner expansion), yielding J′(u;va) =−a3
∇u(z)·A·ˆ ∇u(z)−k2(η−1)|B|ˆu(z)u(z)
+o(a3) Topological derivative of J:
T(z) =−∇ˆu(z)·A·∇u(z) +k2(η−1)|B|ˆu(z)u(z)
Comments
SinceA(β,B)is associated with the normalizedstaticFSTP, the following hold:
◮ Ais symmetric;
◮ By superposition:
V(¯x) =vB[∇u(z)](¯x) = X3
i=1
ei·∇u(z)
vB[ei](¯x) One thus needs only compute thevB[ei], which are independent onz.
◮ Spherical vanishing inclusion(i.e. Bis unit sphere): The FSTP solution is given by
vB[E](¯x) = 3
2 +βE·¯x withβ :=κ⋆/κ (¯x∈ B).
Hence:
A(β,B) = 4π(1−β) 2 +β I;
◮ FSTP solutions analytically available for other shapes (e.g. ellipsoids);
otherwise, solve numerically (and once and for all) the normalized DIE with E=ei (i= 1,2,3);
◮ The special caseβ= 0, i.e. ρ⋆=∞, corresponds to theimpenetrable inclusion.
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 59 / 148
1. Introduction
2. Time-independent (Laplace, elasticity)
3. Waves, frequency domain (Helmholtz, elastodynamics) Helmholtz
Elastodynamics
Heuristics and partial justification for limiting situation Numerical examples
4. Waves, time domain 5. Crack identification
6. Higher-order topological sensitivity 7. Further reading
Background solution
◮ Background solution (C, ρ: background elasticity tensor and mass density):
div (C:ε[u]) +ρω2u+f=0in Ω, t[u] =gon SN, u=uDonSD
In weak form:
Findu∈W(uD), hu,wiCΩ−ω2(u,w)ρΩ=F(w), ∀w∈W0
withW(uD) :=
v∈H1(R3;C3),v=uDonSD . W0=W(0)and hu,wiCD :=
Z
D
ε[u] :C:ε[w] dV = Z
D
∇u:C:∇wdV, (u,w)ρΩ:=
Z
D
ρu·wdV F(w) =
Z
Ω
f·wdV + Z
SN
g·wdS
◮ For unbounded-media idealizations,uis an arbitrary solution to the Navier PDE (e.g. a plane wave), which in particular is not constrained by radiation conditions.
Marc Bonnet (POems, ENSTA) Sensibilit´e topologique et identification 61 / 148
Transmission problem (TP)
(possibly multiple) inclusionB⊂ΩwithCB = [1−χ(B)]C+χ(B)C⋆ =C+χ(B)∆C (elasticity tensor) ρB = [1−χ(B)]ρ+χ(B)ρ⋆ =ρ+χ(B)∆ρ (mass density) TP:
div (CB:ε[uB]) +ρBω2u+f=0in Ω, t[uB] =g onSN, uB =uDonSD i.e., in weak form and in terms of the solution perturbationvB :=uB−u:
FindvB∈W0, hvB,wiCΩB−ω2(vB,w)ρΩB =−hu,wi∆CB +ω2(u,w)∆ρB , ∀w∈W0
This weak form is also applicable to scattering in unbounded media.