An Accurate and Efficient FBPINN Framework for Time-Domain Electromagnetic Simulation in Heterogeneous Media

This work proposes an accurate and efficient Finite Basis Physics-Informed Neural Networks (FBPINN) framework for predicting electromagnetic wave propagation in two-dimensional heterogeneous media. Unlike traditional Physics-Informed Neural Networks (PINN), FBPINN decomposes the computational domain into multiple overlapping subdomains, each governed by an independent neural network trained on time-domain Maxwell’s equations. This approach significantly improves computational efficiency and scalability. To ensure accuracy, we introduce a novel method to enforce field continuity at material interfaces, addressing a key challenge in heterogeneous media modeling. Additionally, we propose a progressive activation scheduler that dynamically adjusts training dynamics, substantially reducing computational cost while accelerating convergence. Extensive numerical experiments validate the superior performance of FBPINN, demonstrating its effectiveness in both accuracy and efficiency.