Stable and Generalizable Neural Network for Spatiotemporal Electromagnetic Modeling

Although rigorous numerical methods, such as the finite-difference time-domain (FDTD) method, can solve time-domain Maxwell's equations with high accuracy, they still face significant challenges regarding stability and computational efficiency. In this paper, a novel stable and generalizable neural network for spatiotemporal electromagnetic modeling (SGNN-STEM) is proposed to enable inference stability with larger stepping time interval than that of the conventional FDTD, and accurate prediction of spatiotemporal electromagnetic fields comparable to FDTD. The results show that the proposed SGNN-STEM exhibits strong generalization capabilities for both varying initial conditions and geometric configurations. Furthermore, its effectiveness across different boundary conditions is validated through various numerical results. Benefiting from the inference stability property, the SGNN-STEM approach showcases superiority in computational efficiency when simulation times reach 400 or larger. Regarding accuracy, the normalized mean squared error (NMSE) of the proposed method is below 0.1 percent in scenarios involving a polygonal scatterer. In the modeling of a scattering array, the NMSEs are 0.5 percent and 9.9 percent under absorbing boundary conditions and perfect electric conductor boundary conditions, respectively. Finally, two practical cases are provided to exhibit the possible utility of the proposed method.