A Review of Recent Machine Learning Approaches for Inverse Scattering Problems

Advances in computational methods, particularly (ML) and (DL), have opened new avenues for tackling (ISP) due to their flexibility, adaptability, and robustness to noise. Beyond improving accuracy, these methods also enable real-time applications, making ML an increasingly valuable tool for ISP. In this work, we present a comprehensive review of ML-based approaches to ISP published from 2017 to the present. The literature is classified into two main categories: agnostic methods, which do not incorporate physics-based information, and physics-informed approaches, which embed physical constraints either through the loss function or in the network architecture. We compare existing contributions across several dimensions, including physical model, scatterer type, ML model, loss function design, incidence data, homogeneity, (BC), dataset characteristics, prior information, and claims. Special attention is given to innovative methodologies and works with significant scientific impact. The reviewed studies reveal a progressive convergence between data-driven and physics-based modeling, with both agnostic and physics-informed strategies continuing to coexist, each addressing distinct requirements in ISP. High-impact publications and recent advances indicate that future research will refine these complementary approaches and explore innovative integrations, combining physical constraints, attention mechanisms, and hybrid architectures to enhance interpretability, robustness, and scalability.