Inverse Elastic Scattering

• Inverse elastic scattering is the process of retrieving unknown geometry and material properties from measured elastic waves using the Navier equation framework.
• Direct sampling, factorization, and monotonicity methods enable unique and non-iterative domain reconstruction even under limited data and noise.
• Iterative and statistical algorithms refine boundary recovery and address challenges such as phaseless data, limited aperture, and random media inhomogeneities.

Inverse elastic scattering concerns the determination of unknown properties—geometry, boundary conditions, or medium inhomogeneities—of a target region or object, from measured wavefields generated by the scattering of elastic waves. These problems are central to geophysics, non-destructive evaluation, medical elastography, and many other fields. The challenges are driven by the complexity of the Navier system, the coupling of compressional (P-) and shear (S-) waves, the strong ill-posedness of the inverse map, and the multi-scale, multi-type nature of scatterers or inhomogeneities.

1. Mathematical Foundations of Inverse Elastic Scattering

Inverse elastic scattering is formulated via the time-harmonic Navier equation for a displacement field $u(x)$ in a homogeneous or piecewise-homogeneous isotropic elastic background, characterized by Lamé parameters $\lambda,\mu$ and possibly variable density. A prototypical forward problem reads

$$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$

with appropriate boundary conditions on the unknown boundary $\partial D$ (Dirichlet, Neumann, or Robin). The central quantities for inversion are far-field patterns of scattered P- and S-waves, or near-field data measured on a surface.

The inverse problem is to recover unknown domains $D$, internal parameters (potentials, inclusions), or boundary conditions from partial or complete knowledge of wavefield data. The nonlinearity arises from the domain dependence in parameter-to-data maps, and ill-posedness manifests in non-uniqueness, instability, and high sensitivity to data errors.

2. Uniqueness and Direct Sampling Methods

Advances in uniqueness theory for inverse elastic scattering, particularly under minimal data regimes, provide rigorous foundations for practical imaging schemes. For a single incident plane wave (fixed frequency, direction, polarization), the boundary $\partial D$ and its boundary condition (Dirichlet, Neumann, or Robin) can be uniquely identified from the far-field pattern, as established for regular as well as polygonal obstacles (Elschner et al., 2019, Liu, 2016). This result is built on a novel reflection principle for the Navier equation:

• If $u=0$ on a line segment, $u$ can be analytically continued by a nonlocal reflection.
• Leveraging Rellich-type uniqueness and propagation of zero nodal sets, the inverse map from the far-field pattern for a single wave is injective for connected polygonal obstacles.

For practical inversion, direct "sampling" methods construct indicator functions that test whether a given point (or domain) is inside the true scatterer, based only on a single far-field pattern. These indicators are typically quadratic forms involving inner products of the data with test functions parametrized by the sampling points. The peaks of these functions correspond to the location(s) and, in multi-scale settings, possibly the shapes of scatterers (Hu et al., 2013). For example, for small, extended, or multi-scale rigid bodies,

$$I^1(z) = \frac{|\int_{S^2} u_p^\infty(\hat x)\cdot[\hat x\otimes\hat x\,e^{-i k_p \hat x \cdot z}]\,d\hat x|^2}{\|u_p^\infty\|^2_{L^2(S^2)}}$$

achieves its maximum when $z$ coincides with the scatterer’s center.

3. Factorization and Monotonicity-Based Inversion

The factorization method has become a leading theoretical and algorithmic framework for elastic inverse problems. The far-field operator $\lambda,\mu$0, mapping incident Herglotz densities to far-field patterns, admits explicit operator factorizations. For the obstacle (Dirichlet) case,

$\lambda,\mu$1

where $\lambda,\mu$2 encapsulates the map from boundary data to far-field patterns, and $\lambda,\mu$3 is the adjoint of the elastic single-layer operator (Bai et al., 5 Jun 2025, Li et al., 2024, Elschner et al., 2019). For cavities or Neumann cases, the factorization involves the hypersingular conormal double-layer operator.

This structure enables the construction of efficient range tests (via Picard's criterion) for determining domain inclusion. For $\lambda,\mu$4 in the search region, the indicator

$\lambda,\mu$5

(where $\lambda,\mu$6 are eigenpairs of $\lambda,\mu$7) is large for $\lambda,\mu$8 inside the obstacle and small outside. These ideas are extended to monotonicity-based tests where the sign (definiteness) and multiplicity of positive eigenvalues of certain operator combinations provide a geometric characterization of inclusions (Bai et al., 5 Jun 2025). Such tests are robust: the number of positive eigenvalues is stable under small operator perturbations from noise.

Monotonicity and factorization methods do not require initial guesses or iterations, providing non-iterative frameworks with full theoretical guarantees under broad geometric and regularity assumptions.

4. Iterative and Newton-Type Reconstruction Algorithms

In scenarios requiring shape optimization or fine geometry recovery, iterative algorithms based on Newton or gradient descent methods are prominent. These methods iteratively update a parametric representation of the obstacle boundary based on minimizing the misfit between model-predicted and measured data (Chang et al., 2023, Li et al., 2017, Li et al., 2018). Critical steps involve:

• Parametrizing the boundary (e.g., via star-shaped Fourier expansions or spherical harmonics).
• Solving the forward scattering problem, often via boundary integral equations, at each iterate.
• Computing Frechét derivatives of the forward map analytically (domain derivatives) or via adjoints, often invoking the analyticity or regularity of boundary integrals.
• Employing regularization, e.g., Tikhonov penalty in the update step, to control ill-posedness of the inversion step.

For example, algorithms using analytic Fourier-Bessel expansions provide explicit formulas for the field and its derivatives on any candidate curve, enabling a "one-shot" Newton update for the boundary curve (Chang et al., 2023). These approaches avoid repeated full PDE solves and feature rapid convergence under data with moderate noise.

Frequency-continuation strategies, where lower-frequency data are used to recover coarse shapes followed by high-frequency data for fine features, improve stability and accuracy (Li et al., 2017).

5. Statistical and Random Media Inversion

Inverse elastic scattering with random inhomogeneities in the medium is formulated in terms of random fields, with the principal question being the recovery of statistical descriptors (covariance, principal symbol) of the unknown. For microlocally isotropic Gaussian random potentials $\lambda,\mu$9 with covariance operator $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$0 of order $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$1, the key result is that the principal symbol $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$2 of $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$3 is uniquely recovered from the frequency-averaged high-frequency amplitude data measured at a distance from the potential. The main analytical tools are:

• Well-posedness of the Lippmann-Schwinger equation with distributional $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$4 (Li et al., 2020, Li et al., 2021, Li et al., 2018).
• Use of the leading Born approximation term and its frequency-averaged energy.
• Stationary phase and microlocal analysis to connect measured intensity to $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$5 through explicit Abel-type (in 2D) or Newtonian potential (in 3D) integral equations.

For example, for 2D random potentials,

$$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$6

almost surely, allowing unique recovery of $$\mu\,\Delta u + (\lambda+\mu)\nabla(\nabla\cdot u) + \omega^2 \rho\,u = \text{source}, \quad \text{in }\mathbb{R}^d\backslash\overline{D},$$7 by inverting the explicit kernel.

These results demonstrate that single realization, high-frequency "self-averaging" enables unique statistical inversion, underlining the power of microlocal techniques for random elastic inverse problems.

6. Surface and Rough Object Inverse Elastic Scattering

Inverse elastic scattering for unbounded or periodic rough surfaces is a structurally distinct problem, often with measured data available only on a line segment in the near- or far-field (Hu et al., 2018, Zhu et al., 2020, Diao et al., 2017). These problems are cast as boundary-value or variational problems with noncompact geometry. Imaging methods in this context include:

• Direct inner-product-based indicator functions constructed from measured near-field data and superpositions of Green's function or incident fields (Hu et al., 2018).
• Non-iterative sampling methods solving modified near-field equations based on a reference surface, exploiting blow-up behavior of solution norms at the true surface (Zhu et al., 2020).
• For small-amplitude surface perturbations, explicit inversion formulae via field-transformation expansions and least-squares recovery of the surface modulation (Diao et al., 2017).

These methods are computationally efficient and robust with respect to noise, enabling stable reconstruction even with significant data uncertainty.

7. Challenges: Phaseless Data, Limited Aperture, and Non-standard Settings

Phaseless inverse elastic scattering, where only the amplitude of the far-field data is known, presents intrinsic non-uniqueness due to translation invariance. Techniques to overcome this include:

• Reference object methods, in which a known body is introduced to break invariance, allowing classical reconstruction of both location and shape (Dong et al., 2018, Ji et al., 2018).
• Phase retrieval from multi-source or multi-strength data, using geometric constraints and algebraic inversion to reconstruct complex field information from magnitudes only (Ji et al., 2018).
• Factorization and monotonicity methods adapted to limited aperture or partial data (Li et al., 2024), where theoretical justification and numerical stability are maintained provided the data acquisition geometry covers critical directions.

These sections synthesize and reference key published work:

• Uniqueness with minimal data and direct sampling: (Elschner et al., 2019, Liu, 2016, Hu et al., 2013).
• Factorization, monotonicity, and indicator-based approaches: (Bai et al., 5 Jun 2025, Li et al., 2024, Elschner et al., 2019).
• Newton-type iterative methods, variational inversion, and frequency-continuation: (Chang et al., 2023, Li et al., 2017, Li et al., 2018).
• Random media inversion and microlocal analysis: (Li et al., 2020, Li et al., 2021, Li et al., 2018).
• Rough surface and periodic structure imaging: (Zhu et al., 2020, Hu et al., 2018, Diao et al., 2017).
• Phaseless data and phase retrieval: (Ji et al., 2018, Dong et al., 2018).
• Multi-particle and three-dimensional extensions: (Lai et al., 2022, Li et al., 2017).

This body of work establishes a comprehensive, mathematically rigorous, and computationally diverse foundation for the field of inverse elastic scattering, with methodologies tailored to the breadth of geometric, physical, and data-collection scenarios encountered in science and engineering.