Study on a Recurrent Convolutional Neural Network Based FDTD Method

Abstract

A finite difference time domain method (FDTD) based on recurrent convolutional neural network (RCNN) is studied in this paper. The finite difference operator in FDTD can be described as a convolution operator in CNN, and the time marching scheme can be described using the framework of recurrent neural network (RNN). Based on this analysis, a recurrent convolutional neural network can be directly set up that can rigorously solve a given FDTD problem. The network coefficients are derived from the FDTD formulation and the training process is not needed. This study links the FDTD method with the artificial neural network such that some electromagnetic problems can be solved on the deep learning framework. Benefiting from the recent development of software and hardware in artificial neural networks, this RCNN-FDTD method can be easily implemented in parallel across multiple computing platforms. Numerical results demonstrate that the RCNN-FDTD can solve scattering problems with a good parallel efficiency on massively parallel architectures and the same accuracy as the traditional FDTD.