Improved Physics-Driven Neural Network to Solve Inverse Scattering Problems

This article presents an improved physics-driven neural network (IPDNN) framework for solving electromagnetic inverse scattering problems (ISPs). Unlike data-driven solvers that require large labeled datasets and often suffer from limited generalization, the proposed method performs instance-wise optimization under electromagnetic governing equations and measured scattered-field constraints. Three improvements over the original PDNN framework are introduced. First, a Gaussian-localized oscillation-suppressing window (GLOW) activation function is designed to stabilize convergence and enable a lightweight yet accurate network architecture. Second, a dynamic scatter subregion identification strategy is further developed to adaptively refine the computational domain, preventing missed detections and reducing computational cost. Third, transfer learning is incorporated to extend the solver’s applicability to practical scenarios, integrating the physical interpretability of iterative algorithms with the approximation capability of neural networks. Numerical and experimental results on representative Fresnel datasets show that the proposed IPDNN achieves up to a 57% reduction in reconstruction error while maintaining robust performance under noise levels up to 50%. Under the current implementation, the proposed FCN-based realization also requires substantially less training memory than the original PDNN. These results demonstrate that IPDNN provides a more accurate and practically extensible physics-driven framework for electromagnetic inverse scattering.