A Fast and Generalizable ML-Assisted Framework for Full-Wave Inverse Scattering

Abstract

This article proposes a novel machine learning (ML)-assisted framework for solving full-wave inverse scattering problems (ISPs) in inhomogeneous, high-contrast media. Traditional deterministic algorithms used to solve such ISPs face significant challenges due to their high computational cost, inherent nonlinearity, and strong ill-posedness. Recently, the introduction of ML methods has enabled the development of rapid solutions to this problem. However, these solutions’ limited out-of-distribution (OOD) generalization capabilities pose significant challenges for practical applications. To address these challenges, we propose a novel pathway to combine ML models with full-wave inversion (FWI). In this framework, ML models serve as auxiliary tools, supplying prior knowledge for use in FWI. A mathematically guaranteed bounds-generation algorithm is proposed to bridge ML models with FWI, and a limited-memory Broyden-Fletcher–Goldfarb-Shanno algorithm with bound constraints (L-BFGS-B) is introduced in FWI to incorporate these bounds. In contrast to the existing research, our framework leverages ML to enhance computational speed while preserving the interpretability and broad applicability of physical models, making it outstanding for OOD samples. We validate the framework across three numerical datasets and conducted rigorous ablation studies on each component to confirm its contributions. To further assess the robustness of the framework, we perform a noise stability study under perturbed conditions. In addition, we extend the framework to multifrequency and time-domain inversion schemes, thereby demonstrating its broad applicability across diverse FWI tasks. We also integrate transfer learning techniques to highlight the framework’s strong compatibility with emerging ML techniques.