Variational optimization approach for reconstruction of dielectric permittivity and conductivity functions using partial boundary measurements

We present a variational optimization approach for the solution of a coefficient inverse problem of simultaneous reconstruction of the dielectric permittivity and conductivity functions in time-dependent Maxwell’s system using limited boundary observations of the electric field. The variational optimization approach is based on constructing a weak form of a Lagrangian which allows to use finite element based reconstruction algorithms. The optimality conditions for the Lagrangian and stability estimate for the adjoint problem are derived, as well as Frechét differentiability of it and of the regularized Tikhonov functional are also presented. Two- and three-dimensional numerical studies confirm our theoretical investigations.

💡 Research Summary

The paper introduces a variational optimization framework for the simultaneous reconstruction of spatially varying dielectric permittivity ε(x) and electrical conductivity σ(x) in a time‑dependent Maxwell system, using only limited boundary measurements of the electric field. Starting from the full Maxwell equations, the authors specialize to non‑magnetic, isotropic media (μ_r ≡ 1) and derive a stabilized second‑order wave‑type equation for the electric field E (equations 2.8a‑c). The inverse problem is to determine ε and σ inside a bounded domain Ω from observations ˜E_obs on a portion of the space‑time boundary Γ_T, possibly contaminated with noise of level δ.

To tackle the ill‑posedness, a Tikhonov functional F(ε,σ) is defined (3.1) comprising a data‑misfit term and regularization terms weighted by parameters γ_ε and γ_σ. Because E depends implicitly on ε and σ, direct gradient descent on F is impractical. The authors therefore construct a Lagrangian L(E,λ,ε,σ) that incorporates the forward Maxwell equation as a constraint via an adjoint (Lagrange multiplier) field λ. Lemma 3.1 presents the weak form of L, obtained by integrating by parts and allowing lower regularity for the state and adjoint variables.

The core analytical contributions are:

1. Derivation of the first‑order optimality conditions (3.8‑3.11) by computing the Fréchet derivative of L with respect to each variable. These conditions couple the forward problem, the adjoint problem, and the gradient of the regularization terms.
2. Proof of stability estimates for the adjoint problem (Section 4), establishing boundedness of λ in terms of the data noise and the forward solution, which is essential for the well‑posedness of the gradient computation.
3. Demonstration of Fréchet differentiability of both the adjoint operator and the regularized Tikhonov functional in the product space U = H¹_E × H¹_λ × C_ε × C_σ. This guarantees that the gradient of F is Lipschitz continuous, a prerequisite for convergence analysis of gradient‑based algorithms.
4. Existence of minimizers for the non‑regularized and regularized functionals (Section 6), ensuring that the optimization problem admits a solution.

On the computational side, the authors propose two gradient‑based algorithms: a standard Conjugate Gradient Algorithm (CGA) and an Adaptive Conjugate Gradient Algorithm (ACGA) that updates the regularization parameters during iteration. The domain is split into a finite‑difference region (Ω_FDM) and a finite‑element region (Ω_FEM) to exploit the efficiency of FDM for wave propagation and the flexibility of FEM for handling complex geometries and material heterogeneities. Time discretization follows a staggered scheme compatible with the second‑order wave equation, while spatial discretization uses linear Lagrange elements for both state and adjoint fields.

Numerical experiments are conducted in both two and three dimensions. In 2‑D, synthetic objects with known ε and σ are reconstructed from data collected on only 30 % of the boundary, with 5 % Gaussian noise added. Both CGA and ACGA recover the material parameters with errors below 3 %. In 3‑D, realistic models of malignant melanoma (MM) at 6 GHz, taken from recent biomedical literature, are used. The dielectric contrast between tumor and healthy tissue is significant, making MM an ideal test case for microwave imaging. Using back‑scattered electric field data from roughly a quarter of the boundary, the algorithms accurately reconstruct both the shape and the complex permittivity distribution; ACGA converges roughly 1.8 times faster than CGA.

The results demonstrate that the variational approach can handle partial, noisy boundary data and still produce stable, high‑resolution reconstructions of both ε and σ. The paper emphasizes potential applications in non‑invasive microwave medical imaging (e.g., breast cancer, melanoma detection) and in the detection of improvised explosive devices, where only limited access to the target’s surface is possible.

In conclusion, the study provides a rigorous mathematical foundation (optimality conditions, stability, differentiability, existence) together with practical algorithms that are validated numerically. Future work suggested includes extending the framework to incorporate unknown source terms, multi‑frequency data, and experimental validation with real measurement systems.