Benchmarking Physics-Informed Neural Networks for Time-Domain Electromagnetic Simulations

Abstract

We benchmark physics-informed neural networks (PINNs) for time-domain electromagnetic simulations, systematically addressing fundamental questions on their accuracy, stability and computational efficiency. A U-Net is trained to solve the wave equation for the electric field in the time domain, in an unsupervised manner. The performance of the physics-informed U-Net is evaluated by varying its architecture, while monitoring its accuracy and stability. Thus, we propose a counterpart of the familiar and insightful dispersion and stability analysis of conventional numerical techniques (such as FDTD) for PINNs. We demonstrate the significance of this approach for building effective and accurate PINNs for time-domain electromagnetic simulations.