Loading [MathJax]/jax/output/SVG/jax.js
Research article

Application of neural networks to inverse elastic scattering problems with near-field measurements

  • Received: 26 August 2023 Revised: 12 October 2023 Accepted: 24 October 2023 Published: 01 November 2023
  • This paper is concerned with the application of a machine learning approach to inverse elastic scattering problems via neural networks. In the forward problem, the displacements are approximated by linear combinations of the fundamental tensors of the Cauchy-Navier equations of elasticity, which are expressed in terms of sources placed inside the elastic solid. From the near-field measurement data, a two-layer neural network method consisting of a gated recurrent unit to gate recurrent unit has been used to reconstruct the shape of an unknown elastic body. Moreover, the convergence of the method is proved. Finally, the feasibility and effectiveness of the presented method are examined through numerical examples.

    Citation: Yao Sun, Lijuan He, Bo Chen. Application of neural networks to inverse elastic scattering problems with near-field measurements[J]. Electronic Research Archive, 2023, 31(11): 7000-7020. doi: 10.3934/era.2023355

    Related Papers:

    [1] Xinlin Cao, Huaian Diao, Jinhong Li . Some recent progress on inverse scattering problems within general polyhedral geometry. Electronic Research Archive, 2021, 29(1): 1753-1782. doi: 10.3934/era.2020090
    [2] Yan Chang, Yukun Guo . Simultaneous recovery of an obstacle and its excitation sources from near-field scattering data. Electronic Research Archive, 2022, 30(4): 1296-1321. doi: 10.3934/era.2022068
    [3] Deyue Zhang, Yukun Guo . Some recent developments in the unique determinations in phaseless inverse acoustic scattering theory. Electronic Research Archive, 2021, 29(2): 2149-2165. doi: 10.3934/era.2020110
    [4] Yujie Wang, Enxi Zheng, Wenyan Wang . A hybrid method for the interior inverse scattering problem. Electronic Research Archive, 2023, 31(6): 3322-3342. doi: 10.3934/era.2023168
    [5] Weishi Yin, Jiawei Ge, Pinchao Meng, Fuheng Qu . A neural network method for the inverse scattering problem of impenetrable cavities. Electronic Research Archive, 2020, 28(2): 1123-1142. doi: 10.3934/era.2020062
    [6] John Daugherty, Nate Kaduk, Elena Morgan, Dinh-Liem Nguyen, Peyton Snidanko, Trung Truong . On fast reconstruction of periodic structures with partial scattering data. Electronic Research Archive, 2024, 32(11): 6481-6502. doi: 10.3934/era.2024303
    [7] Talaat Abdelhamid, F. Khayat, H. Zayeni, Rongliang Chen . Levenberg-Marquardt method for identifying Young's modulus of the elasticity imaging inverse problem. Electronic Research Archive, 2022, 30(4): 1532-1557. doi: 10.3934/era.2022079
    [8] Xiaoping Fang, Youjun Deng, Wing-Yan Tsui, Zaiyun Zhang . On simultaneous recovery of sources/obstacles and surrounding mediums by boundary measurements. Electronic Research Archive, 2020, 28(3): 1239-1255. doi: 10.3934/era.2020068
    [9] Yiyuan Qian, Kai Zhang, Jingzhi Li, Xiaoshen Wang . Adaptive neural network surrogate model for solving the implied volatility of time-dependent American option via Bayesian inference. Electronic Research Archive, 2022, 30(6): 2335-2355. doi: 10.3934/era.2022119
    [10] Shiqi Ma . On recent progress of single-realization recoveries of random Schrödinger systems. Electronic Research Archive, 2021, 29(3): 2391-2415. doi: 10.3934/era.2020121
  • This paper is concerned with the application of a machine learning approach to inverse elastic scattering problems via neural networks. In the forward problem, the displacements are approximated by linear combinations of the fundamental tensors of the Cauchy-Navier equations of elasticity, which are expressed in terms of sources placed inside the elastic solid. From the near-field measurement data, a two-layer neural network method consisting of a gated recurrent unit to gate recurrent unit has been used to reconstruct the shape of an unknown elastic body. Moreover, the convergence of the method is proved. Finally, the feasibility and effectiveness of the presented method are examined through numerical examples.



    Direct and inverse time-harmonic elastic wave scattering by a bounded body [1] has a wide range of applications, such as radar/sonar, geophysical exploration, medical imaging, seismology and materials sciences [2,3,4,5]. The forward problem is to determine and approximate the diffracted field given an incident wave and the elastic body. We refer the reader to, e.g., [6,7] for the existence and uniqueness of forward solutions via the variational approach. The inverse scattering problem (see e.g., [8,9,10]) is to determine the physical or geometric information of the scatterer based on knowledge of the scattered field. This paper is concerned with both direct and inverse problems.

    In recent years, numerical approaches for position and shape reconstruction in many elastic scattering problems have drawn the attention of a large number of scholars. In [11], Arens proved that in two dimensions the shape of an elastically rigid body can be determined from the information of compressional and shear far-field patterns for all incident elastic plane waves with distinct directions at a fixed frequency. In [7], Bao et al. investigated the direct and inverse problems for a penetrable anisotropic elastic body embedded in a homogeneous and isotropic background medium. They show the uniqueness and existence of weak solutions by using the Fredholm alternative together with properties of the Dirichlet-to-Neumann map in both two and three dimensions. Moreover, an iterative method was designed to recover the interface from the multi-frequency near-field data. Charalambopoulos et al. [12] evaluated the inverse elastic scattering problem by using penetrable isotropic bodies embedded in an isotropic host environment via the factorization method. Hu et al. [13] investigated some uniqueness results for a ball or a convex Lipschitz polyhedron and adapted the factorization method to recover the shape of a rigid body from the scattered S-waves or P-waves corresponding to all incident plane shear or pressure waves. Recently, a reverse time migration method was proposed by Chen and Huang [14]. In [15], Li et al. presented an iterative method to solve the inverse elastic scattering problem. The domain derivative expresses the rate of change with respect to the surface displacement. A continuation method using multiple frequency data was established for the inverse problem.

    The method of fundamental solutions (MFS) is an effective method for solving both direct and inverse problems governed by partial differential equations (PDEs). Since it is very simple and easy to program, it has been attracting an increasing amount of attention in recent years. Various engineering problems, including the acoustics scattering problems [16,17] and the acoustics-elastic scattering problem [18], can be solved by the MFS. The MFS can also be used to solve the inverse acoustic scattering problems (see e.g., [19,20,21]). Alves et al. [22] used the MFS to solve non-homogeneous elastic wave equations.

    In recent years, neural networks, one of the most popular schemes in machine learning, have been applied in many research studies, including inverse scattering problems [23,24]. The seq2seq neural network has been applied for natural language processing and machine translation in the literature [25,26]. In [27], a new SwitchNet with a sparse neural network structure was proposed to solve the wave equation-based backscattering problems. Neural networks are especially suitable for dealing with inaccurate and vague information processing problems involving many parameters and constraints, and they can fully approximate complex nonlinear maps. Yin et al. [23] proposed a two-layer neural network method to recover the location and shape of an obstacle by using phaseless far-field data.

    In this paper, we consider inverse elastic scattering problems related to recovering the shape of a rigid body and a cavity by applying a two-layer gated controlled recurrent unit (GRU) neural network. The convergence of the presented method will be verified. The contributions of this paper are summarized as follows.

    (1) We propose a fast numerical scheme for solving the direct problem;

    (2) We prove the convergence of a two-layer GRU neural network;

    (3) We test the accuracy of the reconstruction method by numerical examples.

    The rest of the paper is arranged as follows. In Section 2, we formulate the elastic scattering problem and describe the numerical scheme for forward solutions. In Section 3 we introduce the architecture of the neural network. In Section 4, we perform theoretical analysis for the convergence of the network. In Section 5, numerical experiments are presented to verify the efficiency and effectiveness of the neural network.

    Let DΩ be a two-dimensional elastic body with the smooth boundary D, where ΩR2 is a subdomain. The exterior domain Dc:=R2¯D is supposed to be connected and filled with a homogeneous and isotropic elastic medium. The displacement u(x,t) satisfies

    μu(x,t)+(λ+μ)u(x,t)ρ2u(x,t)t2=0.

    In general, the displacement u(x,t) is time harmonic and given by u(x,t)=eiωtv(x) (where i is the imaginary unit); thus, the displacement v(x) satisfies the Navier equation

    μΔv+(λ+μ)v+ρω2v=0    in R2¯D, (2.1)

    where ω>0 is the angular frequency, λ,μ are the Lamé constants satisfying that μ>0 and λ+μ>0, and ρ is the density function, which is assumed to be constant.

    The forward problem considered in this paper is to give an incident elastic plane wave uincident=deikpdx+deiκsdx with d=(cosϑ,sinϑ) and d=(sinϑ,cosϑ),ϑ[0,2π), as well as to find the scattered field v=(v1,v2) that satisfies the Navier equation.

    By the well-known Hodge decomposition, any solution v to (2.1) can be decomposed into

    v=vp+vs,

    where vp:=(1/k2p)graddivv is known as the compressional part of v and vs:=(1/k2s)graddivv, with grad=(2,1) and divv=2v1+1v2 as the shear part. Here, ks=ωρμ is the shear wave number and kp=ωρλ+2μ is the compressional wave number.

    For a rigid body D, the scattered field v satisfies the first (Dirichlet) kind of boundary condition, i.e., it holds that

    v=uincident   on  D (2.2)

    so that the total displacement is zero.

    If D is a cavity, the following second (Neumann) kind of boundary condition holds:

    tnv=t   on  D, (2.3)

    with t:=tnuincident. Here, the traction operator tn is defined by

    (tnv)(x)={[(λ+2μ)v1x1+λv2x2]n1+μ[v1x2+v2x1]n2}ˆe1+{μ[v2x1+v1x2]n1+[λv1x1+(λ+2μ)v2x2]n2}ˆe2, (2.4)

    where ˆei, i=1,2, designates the unit vector along the xi axis and n=(n1,n2) denotes the unit normal direction point into Dc.

    To ensure uniqueness, we require the scattered field v to satisfy the two-dimensional Kupradze-Sommerfeld radiation condition at infinity, that is,

    limrr1/2(rvpikpvp)=0,  limrr1/2(rvsiksvs)=0,r:=|x|.

    In this paper, we will employ the MFS (see [28] for the acoustic scattering problem) to calculate the forward solution v. The main idea of the MFS is to approximate the scattered field in the solution domain Dc by using a linear combination of fundamental solutions [18,29] to the Navier equation (2.1) with N source points yjD, i.e.,

    v(x)vN(x)=Nj=1U(x,yj)[ajbj],xDc. (2.5)

    Here, aj,bj are the coefficients to be determined by the boundary condition on D and U(x,y) is Green's tensor of the two-dimensional Lamˊe system of the form

    U(x,y)=Φ(ks|xy|)μI+1ρω2(2x212x1x22x2x12x22)(Φ(ks|xy|)Φ(kp|xy|)),

    where I is the 2-by-2 identity matrix and

    Φ(k|xy|)=i4H(1)0(k|xy|),k=kp,ks

    is the fundamental solution of the Helmholtz equation associated with the wavenumber k. Here, H(1)0 is the Hankel function of the first kind of order zero.

    Remark 2.1. Following the idea in [29], the solution v(x) of this problem can be approximated by the single-layer potential function, with ψ(L2(B))2, as follows:

    Sψ(x)=BU(x,y)ψ(y)dsy,xDc, (2.6)

    where ¯BD. If we discretize the integral operator S, we can get a linear combination of the fundamental solutions with respect to some source points located on B with the unknown density function ψ(yj). It can be seen that (2.5) is a certain discrete form of (2.6) with some weight coefficients.

    We respectively define the compressional and shear parts of U(x,y) as follows:

    Up(x,y)=1ρω2(2x212x1x22x2x12x22)Φ(kp|xy|)

    and

    Us(x,y)=Φ(ks|xy|)μI+1ρω2(2x212x1x22x2x12x22)Φ(ks|xy|).

    Then U(x,y)=Up(x,y)+Us(x,y). Thus we have

    vp(x)=Nj=1Up(x,yj)[ajbj],xDc, (2.7)

    and

    vs(x)=Nj=1Us(x,yj)[ajbj],xDc. (2.8)

    Recalling the definitions of the traction vector (2.4) and the fundamental solution to the Navier equation, one can approximate the traction vector on the boundary D as follows:

    tnv(x)tnvN(x)=Nj=1T(x,yj)[ajbj]. (2.9)

    Here, the matrix T(x,yj) is defined by

    T(x,yj)=(T11(x,yj)T12(x,yj)T21(x,yj)T22(x,yj))

    with the entries

    T11(x,yj)=[(λ+2μ)U11(x,yj)x1+λU21(x,yj)x2]n1+μ(U11(x,yj)x2+U21(x,yj)x1)n2,
    T12(x,yj)=[(λ+2μ)U12(x,yj)x1+λU22(x,yj)x2]n1+μ(U12(x,yj)x2+U22(x,yj)x1)n2,
    T21(x,yj)=μ(U11(x,yj)x2+U21(x,yj)x1)n1+[λU11(x,yj)x1+(λ+2μ)U21(x,yj)x2]n2,
    T22(x,yj)=μ(U12(x,yj)x2+U22(x,yj)x1)n1+[λU12(x,yj)x1+(λ+2μ)U22(x,yj)x2]n2.

    To determine the coefficients aj,bj,j=1,2,...N, we employ the collocation method by choosing M collocation points {xj}Mj=1D and impose the boundary condition on D. Set d:=(a1,,aN,b1,,bN) to be the vector of 2N unknown parameters. It can be calculated by solving a linear system with 2M equations of the form

    Ad=h, (2.10)

    where the matrix AC2M×2N on the left-hand side and the vector hC2M×1 on the right-hand side are defined as follows depending on the boundary condition on D.

    ● Dirichlet boundary condition (rigid body)

    Ai,j=U11(xi,yj),Ai,j+N=U12(xi,yj),Ai+M,j=U21(xi,yj),Ai+M,j+N=U22(xi,yj),
    hi=u1incident(xi),hi+M=u2incident(xi),i=1,,M,j=1,,N.

    ● Neumann boundary condition (cavity)

    Ai,j=T11(xi,yj),Ai,j+N=T12(xi,yj),Ai+M,j=T21(xi,yj),Ai+M,j+N=T22(xi,yj),
    hi=t1(xi),hi+M=t2(xi),i=1,,M,j=1,,N.

    In our numerics, we choose the number M=N in all examples. The algebraic linear system (2.10) is solved directly by using Matlab with the command d=Ah. Then we can get the scattered near-field data v on Ω from (2.5). One can also get the shear part vs|Ω and compressional part vp|Ω of v by replacing Green's tensor U by its shear and compressional parts in (2.7) and (2.8), respectively. These synthetic data will be used as the measurement data in our inverse problem. In our examples below, we suppose, for simplicity that the boundary Ω can be parameterized as follows:

    Ω={c(θ)=R(cosθ,sinθ)+(a,b):θ(0,2π]},R,a,bR.

    This means that the measurement data are taken on the circle centered at (a,b)R2 with the radius R>0.

    Definition 2.1. Let ˆdm:=R(cosθm,sinθm)+(a,b),m=1,2,,N1; indicate the measurement positions where N1N and θm(0,2π]. Given the obstacle D and the compressional and shear wave-numbers kp,ks, respectively, we define the measurement set I=(I(1),I(2),,I(T)) with T=2N1 as follows:

    I(n)=vp(ˆdn,D),I(n+N1)=vs(ˆdn,D),n=1,2,,N1.

    The inverse problem in this paper is to recover the shape of the obstacle or the cavity, namely D, by applying knowledge of the near-field data on Ω associated with a fixed incident field:

    (vp(ˆdn,Ω),vs(ˆdn,Ω))D.

    Assumption 2.1. The boundary of the obstacle D is assumed to take the form

    D:x(θ)=p(θ)(cosθ,sinθ),0θ2π,

    where p(θ) permits the following (truncated) Fourier representation:

    p(θ)=a02+Qm=1amcos(mθ)+Qm=1bmsin(mθ), (2.11)

    where QN. Below we define the set of the Fourier coefficients a0,am and bmm=1,2,,Q, by y={y(1),y(2),,y(N)} with N=2Q+1. In the remaining part of this paper, we shall consider the inverse problem of recovering y from our measurement data set.

    In this section, we introduce the main components of the neural network first. Then how to construct the neural network to deal with the inverse elastic scattering problem is given.

    Figure 1 gives a simple map of the exhibition of its working status. We introduce the GRU cells (see e.g., [23,30]), i.e., the principal element of the update network. This network consists of many GRU neurons.

    Figure 1.  Internal structure of GRU neuron.

    (1) The cell receives a stimulus, i.e., an input I(t)(t=1,2,,T). Then the cell will respond to the stimulus. We describe this response as an activation function acting on the previous output h(t1) and the present stimulus I(t) to produce a reset gate r(t), which is to control candidate status ˜h(t) based on whether it depends on the prior status of learning.

    r(t)=σ(wr[I(t),h(t1)]) (3.1)
    ˜h(t1)=h(t1)r(t) (3.2)
    ˜h(t)=tanh(w˜h[I(t),˜h(t1)]) (3.3)

    (2) It preserves historical output h(t1) and presents input I(t) to obtain update gate z(t), which can be used to choose and forget memory. In other words, it can forget some information passed down in h(t1) and select and memorize the information in ˜h(t).

    z(t)=σ(wz[I(t),h(t1)]) (3.4)
    h(t)=(1z(t))h(t1)z(t)˜h(t). (3.5)

    In order to see the notations clearly, we list the notations used in the neural network in Table 1.

    Table 1.  The notations used in GRU neurons.
    Notations
    r(t) The reset gate
    z(t) The update gate
    wr The reset gate weight
    w˜h The candidate status weight
    wz The update gate weight
    σ(x)=11+ex. A sigmoid activation function
    tanh(x)=exexex+ex The hyperbolic tangent activation function
    h(t) The GRU output
    The tensor product of matrices
    Plus

     | Show Table
    DownLoad: CSV

    The whole model consists of three parts, namely, the encoding end, the middle status and the decoding end. The role of the encoding end is to extract data features, and the intermediate state serves to connect the encoding and decoding ends. The function of the decoder is to decode the information obtained by the encoder to obtain the output sequence. Next, the relevant details of the network structure are illustrated as below.

    Step 1: The scattered data features are extracted by using GRU neurons on the encoding end, i.e.,

    h1(t):=ζt(I(t)  h1(t1)),  t=1,2,...,T,

    where ζt signifies the tth GRU cells in the course of the work. That is, it is equivalent to extracting the input data features by using (3.1)–(3.5). Then the first layer extracts the scattered data as h1=(h1(1),h1(2),,h1(T)).

    Step 2: Extracting features again.

    Take the features extracted for the first time, i.e., h1=(h1(1),h1(2),,h1(T)) as input to extract the features again by means of (3.1)–(3.5):

    h2(t):=ζt(h1(t)  h2(t1)),  t=1,2,...,T.

    Step 3: First decoding process on the decoding terminal.

    The decoder side will use N GRU neurons to decode the second layer output information h2(T) and the obstacle boundary Fourier expansion coefficients y=(y(1),y(2),,y(N)):

    h3(1)=ζ1(y(1)  h2(T)),
    h3(n)=ζn(y(n)  h2(n1)),  n=2,...,N,

    where ζn signifies operation of the nth GRU. It plays the same role as ζt in Step 1.

    Step 4: Second decoding process at the decoder end.

    Get the output h4 for the fourth layer by decoding the output h3=(h3(1),h3(2),,h3(N)) of the third layer. That is, this layer will preserve profitable information and discard unprofitable information

    h4(1)=ζ1(h3(1)  h2(0)),
    h4(n):=ζn(h3(n)  h4(n1)),  n=2,...,N.

    Finally, the latest output from the decoding side will be converted into a predictive output by the fully connected layer. That is to say, we will calculate ˆy=(ˆy(1),ˆy(2),,ˆy(N)), i.e., the predicted output of the network by using the following formula

    ˆy(n)=woh4(n).

    This section is devoted to the convergence analysis of the neural network.

    Definition 4.1. Let {y(n)}Nn=1 be the initial values of the boundary curve and {ˆy(n)}Nn=1 be the network prediction output. Define the common mean absolute value loss (error) function as

    L(w)=1NNn=1|y(n)ˆy(n)|. (4.1)

    When y(n)ˆy(n)0, the derivative of the loss function (4.1) with respect to the weight w in (3.1)–(3.5) is given by

    L(w)w=1NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)w). (4.2)

    For the sake of convenience, we introduce some symbols. The input weights and the internal weights of every layer are gathered in the weight matrices WI and WL. We denote W=(WI,WL). The specific forms are as follows

    WI=(W1I,W2I,W3I,W4I),
    WL=(W1L,W2L,W3L,W4L)=(wr1      w˜h1      wz1wr2      w˜h2      wz2wr3      w˜h3      wz3wr4      w˜h4      wz4).

    Here, WI has the same form as WL.

    Next, we write the input of every layer as

    u1(t)=(I(t)h1(t1)),    u2(t)=(h1(t)h2(t1)),
    u3(n)=(y(n)h3(n1)),    u4(n)=(h3(n)h4(n1)).

    The input vectors for each layer of the encoding and decoding ends are given as follows:

    (v1(t)v2(t)v3(n)v4(n))=(W1u1(t)W2u2(t)W3u3(n)W4u4(n))=(W1II(t)+W1Lh1(t1)W2Ih1(t)+W2Lh2(t1)W3Iy(n)+W3Lh3(t1)W4Ih3(n)+W4Lh4(n1)).

    Definition 4.2. Given a vector α=(α1,α2,,αn), the activation function acts on a vector as follows:

    σ(α)=(σ(α1),σ(α2),,σ(αn)).

    The derivatives of the activation function acts similarly

    σ(α)=(σ(α1),σ(α2),,σ(αn)),
    σ(α)=(σ(α1),σ(α2),,σ(αn)).

    In other words, the activation function acts on a vector componentwise.

    The weight gradient of the fourth-layer neural network update gate can be calculated directly:

    ˆy(n)wz4=ˆy(n)h4(n)  ((1z4(n))  h4(n1)  z4(n)  ˜h(n))σ(v4(n))× (u4(n)+W4L h4(n1)wz4). (4.3)

    From the initial condition constant ˆy(0)0, we have

    ˆy(0)w=0. (4.4)

    For any given initial weight w0, the weight sequence {wk} can be generated from an iterative formula:

    wk+1=wk+Δwk, (4.5)

    where

    Δwk=γL(wk)w (4.6)

    and γ is the learning rate.

    Assumption 4.1. ([31,32]) |σ(a)|,|σ(a)|,|σ(a)| and |tanh(a)|,|tanh(a)|,|tanh(a)| are uniformly bounded for aR.

    Assumption 4.2. ([31,32]) WkL(k=1,2,) is bounded in the learning course described by (4.5).

    In order to check the convergence of the neural network, we claim the following theorem.

    Theorem 4.1. The loss function is defined as the formula (4.1), and the weight sequence {wk} can be generated from the iterative formula (4.5). If Assumptions 4.1 and 4.2 are valid, then we have the following:

    (1) L(wk+1)L(wk);

    (2)  L0 s.t. limkL(wk)=L.

    Proof. First, we can also use the method employed in [32], which adopts the induction argument to calculate Δv4(wk,n) as follows:

    Δv4(wk,n)=(u4(n)+W4L h4(n1)wz4)Δwk+δ(wk,n),Δv(wk,n)CΔwk, (4.7)

    where δ(wk,n) will be given in the following remark. Then, apply the Taylor expansion to v4(wk,n); we have

    L(wk+1)L(wk)=1NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1)  z4(n)  ˜h(n))σ(v4(n))Δv4(wk,n)12NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1) z4(n)  ˜h(n))(σ(τ(wk,n)))2(Δv4(wk,n))212NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1) z4(n)  ˜h(n))σ(τ(wk,n))(Δv4(wk,n))2.

    Next, substituting (4.7) into the first order term of Δv4(wk,n) of the above equation, we have

    L(wk+1)L(wk)=1NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1)z4(n)  ˜h(n))σ(v4(n))(u4(n)+W4L h4(n1)wz4)Δwk+δ(wk,n))12NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1) z4(n)  ˜h(n))(σ(τ(wk,n)))2(Δv4(wk,n))212NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1) z4(n)  ˜h(n))σ(τ(wk,n))(Δv4(wk,n))2.

    Then, combining (4.2), (4.3) and (4.6), we will get

    L(wk+1)L(wk)=1γΔwk21NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1)z4(n)  ˜h(n))σ(v4(n))δ(wk,n)12NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1) z4(n)  ˜h(n))(σ(τ(wk,n)))2(Δv4(wk,n))212NNn=1y(n)ˆy(n)|y(n)ˆy(n)|(ˆy(n)h4(n))((1z4(n)) h4(n1) z4(n)  ˜h(n))σ(τ(wk,n))(Δv4(wk,n))2=1γΔwk2+φ(wk).

    By Assumptions 4.1 and 4.2, together with the Cauchy-Schwarz inequality, we obtain that |φ(wk)|CΔwk2. Hence L(wk+1)L(wk)(1γ+C)Δwk2. If the learning rate is sufficiently small and satisfies that 0<γ<1/C (C is a constant), then the first conclusion of Theorem 4.1 is established.

    Since L(wk)0 is monotonically bounded, there exists an L such that limkL(wk)=L. The proof is completed.

    Remark 4.1. In (4.7), we have

    δ(wk,n)=n1d=1nl=nd+1WkL((1z4(l)) h4(l1) z4(l)˜h(l))σ(v4(l))ΔWkL×((1z4(nd)) h4(nd1) z4(nd)˜h(nd))σ(v4(nd))Δv(wk,nd)+12n1d=0nl=nd+1WkL((1z4(l)) h4(l1) z4(l)˜h(l))σ(v4(l))×ΔWkL×(((1z4(nd)) h4(nd1) z4(nd)˜h(nd))(σ(τ(nd)))2+((1z4(nd)) h4(nd1) z4(nd)˜h(nd)σ(τ(nd))))(Δv(wk,nd))2,

    with

    nl=nd+1WkL((1z4(l)) h4(l1) z4(l)˜h(l))σ(v4(l))=1.

    when d=1. Here, τ(nd) is the intermediate quantity for v(wk+1,n) and v(wk,n). From (4.3) and (4.4), we have

    h4(n)wz4=n1d=0nl=nd+1((1z4(l)) h4(l1) z4(l)˜h(l))σ(v4(l))×
    W4L((1z4(nd)) h4(nd1) z4(nd)
    ˜h(nd))×σ(v4(nd))u(nd).

    When d=0, nl=nd+1W4L((1z4(l)) h4(l1) z4(l)˜h(l))σ(v4(l))=1.

    In this section, we will give some numerical examples to show the effectiveness of the presented method.

    The scattered data described by ID=(I(1),I(2),,I(T)) corresponding to the obstacle D are given by Definition 2.1. The shape parameters yD=(y(1),y(2),,y(N)) are defined as in Assumption 2.1. We denote the training data set by

    T={(ID,yD),DD},

    where D denotes the set of shape parameters. For the elements in T, ID is also updated as the shape D changes, and we randomly selected 2% of the scattered data as the testing set to get the final prediction results.

    The inverse problem can be solved in two steps. First, for a given shape D, the elastic wave equation (2.1) is solved by using the MFS described in Section 2. Second, the network was trained to identify the obstacle from the scattered data, i.e., the network was trained to approximate the Fourier coefficients of the obstacle.

    To test the sensitivity to the noise, we suppose that noisy data is generated by adding Gaussian random noise pointwise, i.e.,

    Iδ(i)=I(i)+δI(i)ξ(i),

    where δ0 denotes the noise level and the random variable ξ(i) is the standard Gaussian distribution.

    In all numerical examples, the parameters are given as follows.

    (1) We shall use the Lamé constants λ=2 and μ=1 and the frequency ω=3, ρ=1.

    (2) The symbol "-" signifies the exact boundary of an obstacle, and "--" signifies the forecast boundary of an obstacle.

    (3) The number of neurons of the GRU neural network is generally set as 2p(pR+), and the network parameters were determined through a large number of numerical experiments. The number of GRU neurons was set as 128. Dropout was set as 0.5. Batch processing was set as 500. And the number of iterations is 40.

    Remark 5.1. The scattering data are complex numbers and the current neural network method can only deal with real numbers; thus, we took out the real and imaginary parts of the scattering data as the scattering data information in the training data set. In addition, we need to use a random function to generate random numbers as the shape parameters in the training data set. This is the preparation of the training set. After the training set is prepared, the network model is trained by using the data in the training set through the two-layer network which has been built. Finally the output of the network, i.e., the predicted values of the shape parameters, is obtained by using the trained model to use the test data set.

    Example 5.1 (Dirichlet boundary condition). First, the scattered data used in the training network are noise free, and the scattered data are collected on a circle with the radius R=3 centered at the origin. We chose to investigate the numerical recovery results by changing the number of measurements T. The d=(1,0) is the incident direction. The boundaries of the scatterers used in this example are parameterized as follows.

    The first case is a heart-shaped domain parameterized by

    x(θ)=(1.1+1.6cos2(θ2)sin(θ2))(cosθ,sinθ).

    The second case is a round-square domain parameterized by

    x(θ)=(cos10θ+sin10θ)110(cosθ,sinθ).

    The polar diameter has the following (truncated) Fourier representation: a02+Qm=1amcos(mθ). We take Q=6, and the idea is to reconstruct the Fourier coefficients am. The coefficients used in the training process, i.e., a0[0.5,4], a1,a2[0.5,0.5] and am[0.05,0.05],m=3,4,5,6, are uniformly distributed random numbers.

    Figure 2 presents the recovery results for a fixed number T{4000, 20,000, 40,000}. From the numerical recovery results in Figure 2, we can see that the accuracy in the numerical results will be improved by increasing the number T.

    Figure 2.  Numerical results for recovering the heart-shaped and round-square domain, obtained with noise free data and different numbers T {4000,20,000,40,000} for Example 5.1.

    Second, in addition to the usual rounding errors, we also contaminated the forward scattered data by adding random noise. We chose to investigate the stability of the inversion schemes by changing the noise level δ. Figure 3 presents the recovery results for the noise level δ{5%,10%,15%}. From the fourth column in Figure 3, we can see that the errors are becoming smaller as the noise level decreases. We can also see that the numerical reconstruction results are poorest where the body has high curvature. This meets our expectations, since more expansion coefficients are required at the corner.

    Figure 3.  Numerical results for recovering the heart-shaped and round-square domain, obtained with different noise levels δ{5%,10%,15%} and T = 40,000 for Example 5.1.

    Example 5.2 (Neumann boundary condition). For comparison, in this example we consider the scattering problem by cavities, i.e., the Neumann boundary condition is imposed. In this example, we will consider a heart-shaped domain parameterized by

    x(θ)=(1.1+1.6cos2(θ2)sin(θ2))(cosθ,sinθ).

    The training sets are the same as in Example 5.1.

    We chose the direction d=(1,0). The scattered data were collected on a circle with the radius R=3 centered at the origin. The numerical recovery results for the noise level δ{0%,5%,10%,15%} with a fixed number T=40,000 are shown in Figure 4. From this, we observe that the locations and shapes are still well captured. From the fourth row in Figure 4, we can see that the errors are becoming smaller as the noise level decreases. We can also see that the numerical reconstruction results are poorest where the body has high curvature. This further shows that our methods are independent of the physical properties of the underlying scatterers.

    Figure 4.  Numerical results for recovering the heart-shaped domain, obtained with different noise levels δ{0%,5%,10%,15%} and T=40,000 for Example 5.2.

    Example 5.3 (Neumann boundary condition). In this example, we will consider the case of an apple-shaped domain parameterized by

    x(θ)=(1+sin2θ5+4cosθ)(cosθ,sinθ).

    The polar diameter has the following (truncated) Fourier representation a02+Qm=1bmsin(mθ). We take Q=6, and the idea is to reconstruct the Fourier coefficients a0 and bm. The coefficients used in the training process, i.e., a0[0.5,3], b1,b2,b3,[0.5,0.5] and b4,b5,b6,[0.1,0.1], are uniformly distributed random numbers.

    We chose the direction d=(1,0). The scattered data were collected on a circle with the radius R=3 centered at the origin. The numerical recovery results for the noise level δ{5%,10%,15%} with a fixed number T=40,000 are shown in Figure 5. The error for all testing sets obtained with different noise levels δ{5%,10%,15%} and T=40,000 is demonstrated by the fourth column in Figure 5. From this, we observe that the locations and shapes are still well captured.

    Figure 5.  Numerical results for recovering the apple-shaped domain obtained with different noise levels δ{5%,10%,15%} and T=40,000 for Example 5.3.

    The above examples give the effectiveness of the presented method for the case of a single obstacle or cavity. The following example will give the case of two disjoint components. Meanwhile, we should know that D is the case of two disjoint components as the a priori information. The Fourier coefficients were tracked as two independent groups.

    Example 5.4 (Mixed boundary conditions). As shown in Figure 6, we consider a scatterer with two disjoint components. We set the scatterer to be a combination of a triangle-shaped domain and a peanut-shaped domain. Suppose that D1 is a big triangle-shaped domain of the Dirichlet kind, parameterized by

    x(θ)=0.8(2+0.3cosθ)(cosθ,sinθ).

    D2 is a mini peanut-shaped cavity. The boundary D2 is parameterized by

    x(θ)=0.053cos2θ+1(cosθ,sinθ)+(2,1).

    We chose the direction d=(32,12) in this example. The scattered data were collected on a circle with the radius R=3 centered at the origin. Figure 6 shows the numerical results obtained from the knowledge of the compressional near-field measurements, the shear near-field measurements and the full near-field measurements with δ=10%. From these subfigures, we observe that the locations and shapes can be captured. This further shows that the presented method is independent of the physical properties of the underlying scatterers, even for two disconnected obstacles with different boundary conditions. We also note that the numerical error is worse than that in the above examples.

    Figure 6.  Numerical results for recovering a Dirichlet-type boundary D1 and a Neumann-type boundary D2 shown in Example 5.4 with δ=10%,T=40,000.

    Example 5.5 (Two obstacles). As shown in Figures 7 and 8, the underling scatterer D is given as the union of two disjoint obstacles D=D1D2. We consider two cases in this example.

    Figure 7.  Numerical recovery results for Dirichlet boundary condition on D for Example 5.5 with T=40,000.
    Figure 8.  Numerical recovery results for Neumann boundary condition on D for Example 5.5 with T=40,000.

    Dirichlet boundary condition: We chose the direction d=(1,0). D1 is a peanut-shaped domain, and the boundary D1 is parameterized by

    x(t)=0.53cos2t+1(cost,sint)+(0,3).

    D2 is a heart-shaped domain, and the boundary D2 is parameterized by

    x(θ)=0.5(1.1+1.6cos2(θ2)sin(θ2))(cosθ,sinθ)+(3,0).

    Neumann boundary condition: We chose the direction d=(22,22). D1 is a round-square domain, and the boundary D1 is parameterized by

    D1=0.5(cos10θ+sin10θ)110(cosθ,sinθ).

    D2 is a kite-shaped domain, and the boundary D2 is parameterized by

    D2=0.4(cosθ+0.65cos2θ0.65,1.5sinθ)+(1.3,0).

    The numerical recovery results for the noise level δ{0%,10%} with a fixed number T=40,000 are shown in Figures 7 and 8. From this, we observe that the locations and shapes are still well captured.

    Example 5.6 (Dirichlet boundary condition). The last example considered here is a different option of the physical parameters. The Lamé constants are given as λ=3.88,μ=2.56 and the frequency ω=π. The scattered data were collected on a circle with the radius R=3 centered at the origin. The incident direction is d=(12,32). The boundary of the scatterer is a heart-shaped domain parameterized by

    x(θ)=0.5(1.1+1.6cos2(θ2)sin(θ2))(cosθ,sinθ).

    Figure 9 presents the recovery results for a fixed number T=40,000 with the noise level δ{5%,10%,15%}. It can be seen that the recovery results are independent of the physical parameters.

    Figure 9.  Numerical results for recovering the heart-shaped domain obtained with different noise levels δ{5%,10%,15%} and T=40,000 for Example 5.6.

    In this paper, we applied a neural network trained on solutions from the MFS to solve inverse elastic wave scattering problems. Both convex and non-convex geometries with smooth boundaries have been investigated. The numerical results confirm the effectiveness of the presented method. From the presented examples, we can see that the presented method is independent of the physical properties of the underlying scatterers. When the boundary consists of two parts, this method also gives a good approximation of the exact boundary curves.

    The authors declare that they have not used Artificial Intelligence (AI) tools in the creation of this article.

    We would like to thank the reviewers for their valuable comments that helped to improve the quality of the submitted manuscript. We also thank professor G. Hu for the valuable comments on this paper. The research was supported by the Natural Science Foundation of China (Nos. 12101603, 11501566), the Fundamental Research Funds for the Central Universities (No. 3122019159) and Tianjin Education Commission Research Project (No. 2022KJ072).

    The authors declare that there are no conflicts of interest.



    [1] G. Bao, T. Yin, Recent progress on the study of direct and inverse elastic scattering problems (in Chinese), Sci. Sin. Math., 47 (2017), 1103–1118. https://doi.org/10.1360/N012016-00198 doi: 10.1360/N012016-00198
    [2] D. Colton, R. Kress, Inverse Acoustic and Electromagnetic Scattering Theory, Springer Nature, Berlin, 2019. https://doi.org/10.1007/978-3-030-30351-8
    [3] F. Cakoni, D. Colton, A Qualitative Approach to Inverse Scattering Theory, Springer, US, 2014. https://doi.org/10.1007/978-1-4614-8827-9
    [4] F. Cakoni, D. Colton, Qualitative Methods in Inverse Scattering Theory, Springer, Vienna, 2006. https://doi.org/10.1007/3-540-31230-7
    [5] H. Ammari, E. Bretin, J. Garnier, H. Kang, H. Lee, A. Wahab, Mathematical Methods in Elasticity Imaging, Princeton University Press, New Jersey, 2015.
    [6] J. H. Bramble, J. E. Pasciak, A note on the existence and uniqueness of solutions of frequency domain elastic wave problems: a priori estimates in H1, J. Math. Anal. Appl., 345 (2008), 396–404. https://doi.org/10.1016/j.jmaa.2008.04.028 doi: 10.1016/j.jmaa.2008.04.028
    [7] G. Bao, G. Hu, J. Sun, T. Yin, Direct and inverse elastic scattering from anisotropic media, J. Math. Pures Appl., 117 (2018), 263–301. https://doi.org/10.1016/j.matpur.2018.01.007 doi: 10.1016/j.matpur.2018.01.007
    [8] H. Liu, M. Petrini, L. Rondi, J. Xiao, Stable determination of sound-hard polyhedral scatterers by a minimal number of scattering measurements, J. Differ. Equations, 262 (2017), 1631–1670. https://doi.org/10.1016/j.jde.2016.10.021 doi: 10.1016/j.jde.2016.10.021
    [9] J. Li, H. Liu, Y. Wang, Recovering an electromagnetic obstacle by a few phaseless backscattering measurements, Inverse Probl., 333 (2017), 035011. https://doi.org/10.1088/1361-6420/aa5bf3 doi: 10.1088/1361-6420/aa5bf3
    [10] J. Li, H. Liu, Q. Wang, Enhanced multilevel linear sampling methods for inverse scattering problems, J. Comput. Phys., 257 (2014), 554–571. https://doi.org/10.1016/j.jcp.2013.09.048 doi: 10.1016/j.jcp.2013.09.048
    [11] T. Arens, Linear sampling methods for 2D inverse elastic wave scattering, Inverse Probl., 17 (2001), 1445–1464. https://doi.org/10.1088/0266-5611/17/5/314 doi: 10.1088/0266-5611/17/5/314
    [12] A. Charalambopoulos, A. Kirsch, K. Anagnostopoulos, D. Gintides, K. Kiriaki, The factorization method in inverse elastic scattering from penetrable bodies, Inverse Probl., 23 (2007), 27–51. https://doi.org/10.1088/0266-5611/23/1/002 doi: 10.1088/0266-5611/23/1/002
    [13] G. Hu, A. Kirsch, M. Sini, Some inverse problems arising from elastic scattering by rigid obstacles, Inverse Probl., 29 (2013), 015009. https://doi.org/10.1088/0266-5611/29/1/015009 doi: 10.1088/0266-5611/29/1/015009
    [14] Z. Chen, G. Huang, Reverse time migration for extended obstacles: elastic waves (in Chinese), Sci. Sin. Math., 45 (2015), 1103–1114. https://doi.org/10.1360/N012014-00097 doi: 10.1360/N012014-00097
    [15] P. Li, Y. Wang, Z. Wang, Y. Zhao, Inverse obstacle scattering for elastic waves, Inverse Probl., 32 (2016), 115018. https://doi.org/10.1088/0266-5611/32/11/115018 doi: 10.1088/0266-5611/32/11/115018
    [16] J. Antˊonio, A. Tadeu, L. Godinho, A three-dimensional acoustics model using the method of fundamental solutions, Eng. Anal. Boundary Elem., 32 (2008), 525–531. https://doi.org/10.1016/j.enganabound.2007.10.008 doi: 10.1016/j.enganabound.2007.10.008
    [17] A. Karageorghis, B. T. Johansson, D. Lesnic, The method of fundamental solutions for the identification of a sound-soft obstacle in inverse acoustic scattering, Appl. Numer. Math., 62 (2012), 1767–1780. https://doi.org/10.1016/j.apnum.2012.05.011 doi: 10.1016/j.apnum.2012.05.011
    [18] Y. Sun, X. Lu, B. Chen, The method of fundamental solutions for the high frequency acoustic-elastic problem and its relationship to a pure acoustic problem, Eng. Anal. Boundary Elem., 156 (2023), 299–310. https://doi.org/10.1016/j.enganabound.2023.08.010 doi: 10.1016/j.enganabound.2023.08.010
    [19] A. Karageorghis, D. Lesnic, L. Marin, The MFS for the identification of a sound-soft interior acoustic scatterer, Eng. Anal. Boundary Elem., 83 (2017), 107–112. https://doi.org/10.1016/j.enganabound.2017.07.021 doi: 10.1016/j.enganabound.2017.07.021
    [20] A. Karageorghis, D. Lesnic, L. Marin, The method of fundamental solutions for the identification of a scatterer with impedance boundary condition in interior inverse acoustic scattering, Eng. Anal. Boundary Elem., 92 (2018), 218–224. https://doi.org/10.1016/j.enganabound.2017.07.005 doi: 10.1016/j.enganabound.2017.07.005
    [21] D. Colton, R. Kress, Using fundamental solutions in inverse scattering, Inverse Probl., 22 (2006), R49. https://doi.org/10.1088/0266-5611/22/3/R01 doi: 10.1088/0266-5611/22/3/R01
    [22] C. J. S. Alves, N. F. M. Martins, S. S. Valtchev, Extending the method of fundamental solutions to non-homogeneous elastic wave problems, Appl. Numer. Math., 115 (2017), 299–313. https://doi.org/10.1016/j.apnum.2016.06.002 doi: 10.1016/j.apnum.2016.06.002
    [23] W. Yin, W. Yang, H. Liu, A neural network scheme for recovering scattering obstacles with limited phaseless far-field data, J. Comput. Phys., 417 (2020), 109594. https://doi.org/10.1016/j.jcp.2020.109594 doi: 10.1016/j.jcp.2020.109594
    [24] Z. Wei, X. Chen, Deep-learning schemes for full-wave nonliner inverse scattering problems, IEEE Trans. Geosci. Remote Sens., 57 (2019), 1849–1860. https://doi.org/10.1109/TGRS.2018.2869221 doi: 10.1109/TGRS.2018.2869221
    [25] E. Chung, J. G. Park, Sentence-chain based Seq2seq model for corpus expansion, ETRI J., 39 (2017), 455–466. https://doi.org/10.4218/etrij.17.0116.0074 doi: 10.4218/etrij.17.0116.0074
    [26] J. Torres, C. Vaca, L. Terán, C. L. Abad, Seq2Seq models for recommending short text conversations, Expert Syst. Appl., 150 (2020), 113270. https://doi.org/10.1016/j.eswa.2020.113270 doi: 10.1016/j.eswa.2020.113270
    [27] Y. Khoo, L. Ying, SwitchNet: a neural network model for forward and inverse scattering problems, SIAM J. Sci. Comput., 41 (2019), A3182–A3201. https://doi.org/10.1137/18M1222399 doi: 10.1137/18M1222399
    [28] G. Fairweather, A. Karageorghis, P. A. Martin, The method of fundamental solutions for scattering and radiation problems, Eng. Anal. Boundary Elem., 27 (2003), 759–769. https://doi.org/10.1016/S0955-7997(03)00017-1 doi: 10.1016/S0955-7997(03)00017-1
    [29] Y. Sun, F. Ma, A meshless method for the Cauchy problem in linear elastodynamics, Appl. Anal., 93 (2014), 2647–2667. https://doi.org/10.1080/00036811.2014.882913 doi: 10.1080/00036811.2014.882913
    [30] P. Zhang, P. Meng, W. Yin, H. Liu, A neural network method for time-dependent inverse source problem with limited-aperture data, J. Comput. Appl. Math., 421 (2023), 114842. https://doi.org/10.1016/j.cam.2022.114842 doi: 10.1016/j.cam.2022.114842
    [31] D. Xu, Z. Li, W. Wu, X. Ding, D. Qu, Convergence of gradient method descent algorithm for recurrent neuron, Int. Symp. Neural Network, 4493 (2007), 117–122. https://doi.org/10.1007/978-3-540-72395-0_16 doi: 10.1007/978-3-540-72395-0_16
    [32] D. Xu, Z. Li, W. Wu, Convergence of gradient method for fully recurrent neural network, Soft Comput., 14 (2010), 245–250. https://doi.org/10.1007/s00500-009-0398-0 doi: 10.1007/s00500-009-0398-0
  • This article has been cited by:

    1. Yao Sun, Yating Wang, A highly accurate indirect boundary integral equation solution for three dimensional elastic scattering problem, 2024, 159, 09557997, 402, 10.1016/j.enganabound.2023.12.015
    2. Yao Sun, Pan Wang, Bo Chen, A decomposition method for two and three dimensional fluid-solid interaction scattering problem, 2024, 165, 08981221, 106, 10.1016/j.camwa.2024.04.012
    3. Yule Lin, Xiaoyi Yan, Jiguang Sun, Juan Liu, Deep Neural Network-Oriented Indicator Method for Inverse Scattering Problems Using Partial Data, 2024, 12, 2227-7390, 522, 10.3390/math12040522
  • Reader Comments
  • © 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(1200) PDF downloads(54) Cited by(3)

Figures and Tables

Figures(9)  /  Tables(1)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog