EMWP-RNN: A Physics-Encoded Recurrent Neural Network for Wave Propagation in Plasmas
Yilan Qin, Haiyang Fu, Feng Xu, Ya‐Qiu Jin
Abstract
Electromagnetic (EM) wave propagation and inversion in complex time-varying medium is a challenging problem, particularly for plasma applications. We extend the EM wave–plasma coupling physics computation mapping to the recurrent neural network (RNN). The system can be trained to learn inhomogeneous time-varying magnetized plasma parameters from temporal scattered field. As a proof-of-concept demonstration, a physics-encoded RNN has been verified, which encodes Maxwell's vector wave equation describing the multiphysics coupling system into the standard RNN architecture. The results demonstrate that time-varying plasma parameter inversion can be accomplished using only a few sets of transmitted electric fields. This model is interpretable and computationally efficient, benefiting from optimization strategies provided by deep learning, which may be extended for various EM–plasma interaction applications and beyond.