Enhanced Full-Wave Inverse Scattering Solver Using FDTD-Equivalent CNNs

A full-wave inverse scattering solver based on finite-difference time-domain (FDTD)-embedded neural networks is proposed in this paper. By introducing a convolutional neural network (CNN)-accelerated FDTD forward solver that reformulates FDTD operators as GPU-implemented convolutional kernels, we significantly enhance the speed of the forward process in the inverse scattering problem. The compatibility between CNN and automatic differentiation (AD) makes gradient computation during the optimization process fast and straightforward. Our framework is training-free, leveraging the computational efficiency of machine learning (ML) platforms while maintaining the interpretability and generalizability of the physical solver.