Improved Physics-Driven Neural Network to Solve Inverse Scattering Problems

TL;DR

This work tackles electromagnetic inverse scattering problems, which are notoriously ill-posed and data-limited, by introducing IPDNN, a physics-driven neural framework that embeds physical constraints into a lightweight network.The key innovations are a trainable Gaussian-localized activation (GLOW) to stabilize training, a dynamic subregion mechanism to focus computation on relevant regions, and transfer learning to adapt to practical task variations with reduced data.Across simulations and experiments, IPDNN achieves superior reconstruction accuracy, robustness to noise, and computational efficiency compared with state-of-the-art methods, while preserving physical interpretability.The approach blends the interpretability of iterative physics-based methods with fast neural inference, offering practical impact for nondestructive testing and related electromagnetic imaging tasks.

Abstract

This paper presents an improved physics-driven neural network (IPDNN) framework for solving electromagnetic inverse scattering problems (ISPs). A new Gaussian-localized oscillation-suppressing window (GLOW) activation function is introduced to stabilize convergence and enable a lightweight yet accurate network architecture. A dynamic scatter subregion identification strategy is further developed to adaptively refine the computational domain, preventing missed detections and reducing computational cost. Moreover, transfer learning is incorporated to extend the solver's applicability to practical scenarios, integrating the physical interpretability of iterative algorithms with the real-time inference capability of neural networks. Numerical simulations and experimental results demonstrate that the proposed solver achieves superior reconstruction accuracy, robustness, and efficiency compared with existing state-of-the-art methods.

Figures (12)

• Figure 1: Sketch of the concerned two-dimensional inverse scattering problems.
• Figure 2: Variation of GLOW function value when (a) fixing $\sigma=1$ and changing $c$ and (b) fixing $c=1$ and changing $\sigma$.
• Figure 3: The schematic diagram of the applied neural network architecture.
• Figure 4: The flow chart for the dynamic scatter subregion identification method.
• Figure 5: Based on the experimental dataset (a) “dielTMdec8f”, (b) “FoamTwinDiel”, (c) “FoamDielExt” and (d) “FoamDielInt”, the imaging results from the solvers with PDNN Du2025PDNN and fully connected layer (denoted by "FCN") neural network architecture equipped with ReLU, LeakyReLU, Tanh, Softsign and GLOW activation function, respectively.