properties of thin dielectric objects from far field data
Noam Zeev
Colaboration: Dr. Fioralba Cakoni
Old Dominion University Department of Mathematics and Statistics
AIP 2009
July 23, 2009
Outline
1 Direct problem for thin dielectric objects
2 Inverse Problem - far-field
3 Numerical results from far field data
Outline
1 Direct problem for thin dielectric objects
2 Inverse Problem - far-field
3 Numerical results from far field data
Applications
Important problems in nondestructive testing
detection of flaws in materials in specific (typically thin) areas, as well as the determination of the integrity of thin coatings
In the study of optical devices in communication networks In the detection of thin air pockets inside structures.
we investigate the inverse problem of using far field time harmonic electromagnetic measurements to determine the shape and information about the thickness and physical properties of a thin dielectric film embedded in an inhomogeneous background.
Direct problem
Assume
1 The obstacle is an infinitely long thin cylinder with an arc as cross section,
2 The electric field is polarized parallel to the cylinder axis.
After factoring out the terme−iωt, whereωis the frequency, the only component of the total electric fieldusatisfies the
Helmholtz equation
∆u+k2u=0
in the exterior of the cylinder, wherek is the wave number.
Formulation of the problem
1. LetΓ⊂R2be an oriented piecewise smooth nonintersecting arc .
2. The total fieldu=us+ui satisfies,
∆u+k2u =0 in R2\Γ ∂u
∂ν
=0 on Γ
[u]−iλ∂u+
∂ν =0 on Γ
r→∞lim
√r ∂us
∂r −ikus
=0 whereλinvolves some physical properties ofΓ.
Direct problem
In general, givenf ∈H−1/2(Γ)andh∈H−1/2(Γ)find v ∈Hloc1 (R2\Γ)satisfying
∆v +k2n(x)v =0 in R2\Γ ∂v
∂ν
=f on Γ
[v]−iλ∂v+
∂ν =h on Γ
r→∞lim
√r ∂v
∂r −ikv
=0
wheren(x)represent the inhomogeneity of the medium.
Direct problem
Well posedness of the direct problem,i.e.,
1 Existence
2 Uniqueness
3 The solution is continuously dependent of the data.
Approaches:
1 Integral equations - BEM
2 Variational formulation - FEM
Direct problem - Via integral equations
Theorem The direct scattering problem has a unique solution v which satisfies
kvkH1(BR\Γ)≤C
kfkH−1/2(Γ)+khkH−1/2(Γ)
x ∈R2\Γ where the positive constantC depends onRbut not onf andh.
Proof of existence:
LetDbe the region bounded by the extension∂DofΓ. Note that[v]∈H˜1/2(Γ)and∂v
∂ν
∈H˜−1/2(Γ).
LetG(x,y)be the radiating Green’s function of the background medium satisfying
∆G(x,y) +k2n(x)G(x,y) =δ(x−y).
From the Greens representation formula we have
v(x) =
Z
∂D
∂v(y)
∂νy G(x,y)dsy − Z
∂D
v(y) ∂
∂νyG(x,y)dsy, x ∈D
− Z
∂D
∂v(y)
∂νy
G(x,y)dsy + Z
∂D
v(y) ∂
∂νy
G(x,y)dsy, x ∈R2\D where
H˜1/2(Γ) :={v ∈H1/2(Γ) :suppv ⊆Γ}
Using the jumps relations of the single- and double-layer potentials across the boundary∂D, we obtain that the jump[v] satisfies
(i
λI+TΓ) [v] = i λh+
KΓ0 − I
2
f (∗) where the operatorsKΓ0 : ˜H−1/2(Γ)→H−1/2(Γ)and TΓ: ˜H1/2(Γ)→H−1/2(Γ)are defined by
KΓ0ψ (x) :=
Z
Γ
ψ(y) ∂
∂νxG(x,y)dsy forx ∈Γ, (TΓψ) (x) := ∂
∂νx Z
Γ
ψ(y) ∂
∂νyG(x,y)dsy forx ∈Γ respectively.
To solve(∗)we observe that(λiI+TΓ) : ˜H1/2(Γ)→H−1/2(Γ)is a Fredholm operator of index zero. Therefore, it suffices to prove only the injectivity of λiI+TΓ.
To this end letξ∈H˜1/2(Γ)satisfies i
λI+TΓ
ξ =0 and define the following potential
w(x) = Z
Γ
ξ(y) ∂
∂νyG(x,y)dsy forx ∈R2\Γ.
ApproachingΓand using the jump relations for double layer potential we obtain
∂w+
∂ν = ∂
∂νx
Z
Γ
ξ(y) ∂
∂νy
G(x,y)dSy =TΓ(ξ),
[w] =ξ and
∂w
∂ν
=0 Hence we have
[w]−iλ∂w+
∂ν =ξ−iλTΓξ =−iλ i
λI+TΓ
ξ =0.
Thereforewsatisfies the direct problem withf =h=0, and from uniqueness resultsw =0 inR2\Γwhich finally implies [w] =ξ =0.
Note: knowing[v], we are able to computev+,v−, ∂v∂ν+, ∂v∂ν−.
Direct Problem- Numerical results
For numerical purposes we need to knowG, we first assume n(x) =1.
Operator
i λI−TΓ
[u] = i λh Techniques used:
1 Cosine substitution introduced by Yan and Sloan (1988) for TΓ.
2 We used a Quadrature method developed by Chapko and Kress (for Single layer potential). Adapted by Lars Monch toTΓ.
n Reu∞((1,0),(1,0)) Imu∞((1,0),(1,0)) 8 0.071205428169135 0.081916845574100 16 0.242512388888656 0.360640284054226 k=3 32 0.253705871997084 0.346918547805065 64 0.253705867911025 0.346918558878099 16 0.049017043906720 0.282768227362883 k=5 32 0.338614266128542 0.400542203216391 64 0.339344419678081 0.400032127366243 128 0.339344419678121 0.400032127366210 16 0.168887350799880 0.224115244323497 32 0.110221269971698 0.382784494259534 k=10 64 0.482325291178517 0.519975395707543 128 0.482676124016400 0.519608279401000 256 0.482676124016448 0.519608279400957
Outline
1 Direct problem for thin dielectric objects
2 Inverse Problem - far-field
3 Numerical results from far field data
Inverse Problem - far-field
Inverse problem
We want to detectΓandλfrom the knowledge of the far-field datau∞(ˆx,d), wherexˆandd are on the unit circle. Here we assumen(x) =1 inR2.
The scattered field has the asymptotic behavior us(x) = eikr
√r u∞(ˆx,d) +O(r−3/2) asr → ∞, whereˆx = |x|x .
Inverse Problem - Uniqueness results
Theorem.AssumeΓ1andΓ2are two scattering obstacles with corresponding surface impedanceλ1andλ2such that for a fixed wave number the far field patterns coincide for all incident directions d, thenΓ1= Γ2andλ1=λ2.
Inverse Problem - far-field
Thefar field operatorF :L2(S)→L2(S)is defined by (Fg)(ˆx) :=
Z
S
u∞(ˆx,d)g(d)ds(d) TheHerglotz wave functionis given by
vg(x) :=
Z
S
g(d)eikx·dds(d)
whereg∈L2(S)is called the kernel ofvg.
Inverse Problem - far-field
By superposition we have the following relation (Fg) =B(Hg)
whereH:L2(S)→H−1/2(Γ)is defined by Hg:=iλ∂vg
∂ν
andB:H−1/2(Γ)→L2(S)takesh∈H−1/2(Γ)to the far field patternu∞of the solution to the direct problem withf :=0 and h
Inverse Problem - far-field
Forβ ∈H˜1/2(Γ)we construct the double layer potential D(β)(x) :=
Z
Γ
β(y) ∂
∂νy
Φ(x,y)ds(y)
whereΦ(x,y)is the fundamental solution of∆v+k2v =0. D has far field patternγFβwhereγ = √eiπ/4
8πk and Fβ :=
Z
Γ
β(y)∂e−ikˆx·y
∂νy ds(y)
Inverse Problem - far-field
Lemma The compact operatorsF: ˜H1/2(Γ)→L2(S)and H:L2(S)→H−1/2(Γ)are injective and have dense range, provided that there does not exist a Herglotz wave function such that its normal derivative vanishes onΓ.
Inverse Problem - far-field
Fβ =γ−1B(I−iλTΓ)β
which implies the following factorization of the far field operator Fg=γF(I−iλTΓ)−1Hg, g∈L2(S)
LemmaFor any piecewise smooth non intersecting arcLand a functionβL∈H˜1/2(L)we defineΦL∞∈L2(S)by
ΦL∞(ˆx) :=
Z
L
βL(y) ∂
∂νye−ikˆx·ydsy.
Then,ΦL∞(ˆx)∈Range(F)if and only ifL⊂Γ.
Proof
IfL⊂Γ, thenΦL∞(ˆx)∈Range(F).
IfL6⊂Γ, assume on the contrary, thatΦL∞(ˆx)∈Range(F), i.e.
there existsβ ∈H˜1/2(Γ)such that ΦL∞(ˆx) =
Z
Γ
β(y) ∂
∂νy
e−ikˆx·ydsy.
Hence by Rellich’s lemma and the unique continuation principle we have that the potentials
ΦL(x) = Z
L
βL(y) ∂
∂νyΦ(x,y)dsy and D(x) = Z
Γ
β(y) ∂
∂νyΦ(x,y)dsy coincide inR2\(Γ∪L).
Proof
Now letx0∈L,x0∈/ Γ, and letB(x0)be a small ball with center atx0such thatB(x0)∩Γ =∅. HenceDis analytic inB(x0) whileΦLhas a singularity atx0which is a contradiction. This
proves the lemma.
Inverse Problem - far-field
Theorem.IfF is the far field operator corresponding to the direct scattering problem withui(x) =eikx·d, the following is true:
1 IfL⊂Γthen for every >0 there exists a solution gL∈L2(S)of the inequality
kFgL−ΦL∞kL2(S)≤ such thatHgL
converges to a well defined function in H−1/2(Γ).
2 IfL6⊂Γthen for every >0 allgL∈L2(S)satisfying kFgL−ΦL∞kL2(S)≤
are such that
→0limkgLkL2(S) =∞ and lim
→0kHgL
kH−1/2(Γ)=∞
Outline
1 Direct problem for thin dielectric objects
2 Inverse Problem - far-field
3 Numerical results from far field data
Numerical results from far field data
Having
u∞=
N
X
n=−N
u∞,nexp(inθ)
we then add random noise to the Fourier coefficients ofu∞to obtain the approximate far-field pattern
u∞=
N
X
n=−N
u∞,a,nexp(inθ)
where
u∞,a,n=u∞,n(1+χn) withχn a random variable in[−1,1]and=.05.
Numerical results from far field data
The inversion scheme is based in solving the following ill-posed first kind equation
Z
S
u∞(ˆx,d)g(d)ds(d) = Z
L
βL(y) ∂
∂νye−ikˆx·ydsy
for an arbitrary open arc L andβL∈H˜1/2(L).
ifL⊂Γwe can find a boundedg ∈L2(S)
ifL6⊂Γall approximates solutions are unbounded.
Numerical results from far field data
In this study we search for the crack by takingLto be a small segment centered at a sampling pointz with unit normalνz and βLa sequence that converges toδ(x −z).
Hence we want to solve Z
S
u∞(ˆx,d)gz,νz(d)ds(d) =−ikνz·x eˆ −ikˆx·z. (∗) Ifz ∈Γandνz coincides with the normal vector toΓatz, then we can find a boundedgz,νz ∈L2(S)that solves(∗). Otherwise all suchgz,νz are unbounded.
Numerical results from far field data
In order to solve(∗)we use Tikhonov regularization and the Morozov discrepancy principle to deal with the severe ill-posedness of this equation.
Rewriting the above equation as an ill-condition matrix equation Agz,νz =fx,νz, where νz ∈[0, π]
We use the singular value decomposition ofA, i.e.A=UΛV∗, is then equivalent to solvingΛV∗gz,νz =U∗fz,νz.
Numerical results from far field data
The Tikhonov regularization leads to the problem of solving mingz,νz∈RnkΛV∗gz,νz −ρz,νzk2l2+αkgz,νzk2l2
whereρz,νz = (ρ1, ρ2, ..., ρn)T =U∗fz,νz whereαis chosen using the Morozov descrepancy principle.
Numerical results from far field data
The numerical procedure to locate the crack is the following:
we consider a uniform grid of sampling points{zi}i=1..N, and for each sampling point we choose a finite number of anglesνz, and compute for eachzi, andνi
G(zi, νi) =kfz,νzkl2/kgz,νzkl2
Numerical results from far field data
formula f(t)=[!2+t,2t]
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1
0 1
2
3 4
5
Figure:Original Crack.
Numerical results from far field data
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
formula f(t)=[t, .5*cos(t*pi/2)+.2*sin(t*pi/2)!0.1cos(3*t*pi/2)]
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1
0 1
2
3
4
5
Figure:Original Crack.
Numerical results from far field data
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
formula f(t)=[2sin((t+4/3)*6pi/4)+2, sin((t+4/3)*3pi/4)]
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1 0
1
2
3 4
5
Figure:Original Crack.
Numerical results from far field data
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
Having foundgandΓby solving the far field equation, it is now possible to use the boundary conditions
[u] +iλ∂u+
∂ν to determineλ. (Zeev’s PhD thesis)
Numerical results from far field data
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
2.2/(dvg/dnu) for non linear crack lab=2
λ
Figure:λreconstructed from far field data. The solid line represent the exact solution.
Numerical results - Impedance reconstruction
Goal:Solve the ill posed, nonlinear equation,
F(λ) =u∞ (∗)
WhereF mapsλ(x)onto the far field patternu∞for a fixed incident fieldui.
Linearizing(∗) Solve
Fλ0h+Fλ =u∞
forh, using Newton type methods.
Improve an approximate of the surface impedanceλr into λ˜r =λr +h.
The scattered field,
us =K(I−iλTΓ)−1Rui whereRui =iλ∂u∂νi. the corresponding far field,
u∞=K∞(I−iλTΓ)−1Rui We computeFλ0h,
Fλ0h=K∞ψh where
(I−iλTΓ)ψh=−ihTΓ(I−iλTΓ)−1Rui+ih∂ui
∂ν.
Numerical results - Impedance reconstruction
Leth(x) =PN
k=0akbk(x). we want to solve
N
X
k=0
akFλ0bk +Fλ =u∞
for realak. We considered Chebyshev polynomials bk(x) =cos(kcos−1(x)), and radial basis functions
bk(x) =e−c|x−ζk|2, withx ∈[−1,1],c >0 andζk ∈[−1,1].
We use a collocation method with respect tomequidistant pointsˆx1...xˆm in the unit circle, i.e. we solve
N
X
k=0
ak(Fλ0bk)(ˆxj) =u∞(ˆxj)−Fλ(ˆxj) j =1, ...,m for realak.
We apply Tikhonov regularization, then we minimize the following equivalent problem,
C(p,a0, ...,aN) :=
p
X
k=0
m
X
i=1
bijkaj−cik
2+α
N
X
j=0
θjka2j
wherebijk = (Fλ0bj)(ˆxi;dk),cik =u∞(ˆxi;dk)−Fλ(ˆxi;dk),αis the regularization parameter andθjk are positive weights.
Numerical results - Impedance reconstruction
We solve the linear system
p
X
k=1 N
X
l=0 m
X
i=1
b¯ijkbilkal+α
p
X
k=1
θjkaj =
p
X
k=1 m
X
i=1
b¯ijkcik j=0, ...,N.
Choseα=2−k, wherek is the number of iteration.
Usedn=64 quadrature points in the forward solver and n=32 in the backward solver.
Used one incident direction.
Numerical results from far field data
(a) (b)
Figure:Panel (a) shows a plot ofλ(x)(in red) and the initial guess λ0(x)(in blue) chosen along the crack. Panel (b) shows a plot ofλ(x) (in red) and lambda reconstructed (in blue) after 10 iterations with Newton type methods. One incident field was used, 5%noise,N=3.
Numerical results from far field data
(a) (b)
Figure:Panel (a) shows a plot ofλ(x)and the initial guessλ0(x) chosen along the crack. Panel (b) shows a plot ofλ(x)and lambda reconstructed after 9 iterations with Newton type methods. One incident field was used, 5%noise,N=3.
Numerical results from far field data
(a) (b)
Figure:Panel (a) shows a plot ofλ(x)and the initial guessλ0(x) chosen along the crack. Panel (b) shows a plot ofλ(x)and lambda reconstructed after 12 iterations with Newton type methods. One incident field was used, 5%noise,N=3.
Conclusions:
Have succesfully reconstructed complex thin objects (with no a priori information) using LSM from the knowledge of noisy far-field data.
Reconstructed their physical properties using a Newton type method from the knowledge of noisy far-field data.
This work have been done for Near field data.
N. Zeev and Fioralba Cakoni, The identification of thin objects from far field or near field scattering data, SIAM J.
APPL. MATH., Vol. 69, No. 4, pp. 1024−1042.
THANK YOU
Numerical results from far field data
formula f(t)=[!2+t,2t]
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1
0 1
2
3 4
5
Figure:Original Crack.
Numerical results from far field data
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1 0
1
2 3
4
5
0.2 0.25 0.3 0.35 0.4 0.45 0.5
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
formula f(t)=[t, .5*cos(t*pi/2)+.2*sin(t*pi/2)!0.1cos(3*t*pi/2)]
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1
0 1
2
3
4
5
Figure:Original Crack.
Numerical results from far field data
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1
0
1
2
3
4
5 0.15
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
formula f(t)=[2sin((t+4/3)*6pi/4)+2, sin((t+4/3)*3pi/4)]
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1 0
1
2
3 4
5
Figure:Original Crack.
Numerical results from far field data
!
!
!5 !4 !3 !2 !1 0 1 2 3 4 5
!5
!4
!3
!2
!1 0 1 2 3 4 5
0(1 0(15 0(2 0(25
Figure:Crack reconstructed from far field data, wavenumberk =5.
Numerical results from far field data
Figure:Crack reconstructed from far field data, wavenumberk =5.