Two-Dimensional Electromagnetic Solver Based on Deep Learning Technique

• Published in: IEEE journal on multiscale and multiphysics computational techniques Vol. 5; pp. 83 - 88
• Main Authors: Qi, Shutong, Wang, Yinpeng, Li, Yongzhong, Wu, Xuan, Ren, Qiang, Ren, Yi
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 2-D Maxwell's equation; Algorithms; Artificial neural networks; Computer architecture; Convolutional neural network (CNN); Decoding; Deep learning; Finite difference method; Finite difference methods; finite-difference frequency-domain (FDFD) method; Frequency-domain analysis; Geometry; Incident waves; Machine learning; Two dimensional displays; Wave propagation
• ISSN: 2379-8815; 2379-8815
• DOI: 10.1109/JMMCT.2020.2995811

Although the deep learning technique has been introduced into computational physics in recent years, the feasibility of applying it to solve electromagnetic (EM) scattering field from arbitrary scatters remains open. In this article, the convolutional neural network (CNN) has been employed to predict the EM field scattered by complex geometries under plane-wave illumination. The 2-D finite-difference frequency-domain (FDFD) algorithm, wrapped by a module to randomly generate complex scatters from basic geometries, is employed to produce training data for the network. The multichannel end-to-end CNN is modified and combined with residual architecture and skip connection, which can speed up convergence and optimize network performance, to form the EM-net. The well-trained EM-net has good performance in this problem since it is compatible with different shapes, multiple kinds of materials, and different propagation directions of the incident waves. The effectiveness of the proposed EM-net has been validated by numerical experiments, and the average numerical error can be as small as 1.23%. Meanwhile, its speedup ratio over the FDFD method is as large as 2000.