Solving Time Domain Electromagnetic Forward and Inverse Problems using a Differentiable Programming Platform

Deep-learning techniques have been widely applied in scientific forward and inverse modeling. Recent advances in high-performance tensor processing hardware and software also provide new opportunities for accelerated linear algebra calculations. In this paper, a trainable recurrent neural network (RNN) is deployed to formulate the electromagnetic propagation, solve the Maxwell’s equations and inverse problems on Pytorch—one of the most state-of-the-art differentiable programming platforms. Due to the specific performance-focused design of PyTorch, the computation efficiency is substantially improved compared to Matlab. Moreover, by setting the trainable weights of RNN as the material-related parameters, an inverse problem can be solved through training the network. Numerical simulation demonstrates the efficiency and effectiveness of our method for forward and inverse modeling.