Electric Flux Density Learning Method for Solving 3-D Electromagnetic Scattering Problems

• Published in: IEEE transactions on antennas and propagation 2022-07, Vol.70 (7), p.5144-5155
• Main Authors: Yin, Tiantian, Wang, Chao-Fu, Xu, Kuiwen, Zhou, Yulong, Zhong, Yu, Chen, Xudong
• Format: Article
• Language: English
• Subjects: Deep learning (DL); Dielectrics; Electric fields; Electric flux; electromagnetic (EM) field; Electromagnetic scattering; Flux density; Integral equations; Learning; Mathematical models; Method of moments; Neural networks; Teaching methods; Testing; Three-dimensional displays; volume integral equation (VIE); Volume integral equations
• ISSN: 0018-926X; 1558-2221
• DOI: 10.1109/TAP.2022.3145486

Inspired by a discretized formulation resulting from volume integral equation and method of moments, we propose an electric flux density learning method (EFDLM) using cascaded neural networks to solve 3-D electromagnetic (EM) scattering problems that involve lossless dielectric objects. The inputs of the proposed EFDLM consist of the contrast of the objects, the projections of incident field, and the first-order scattered field onto the testing functions, and the output is chosen as the normalized electric flux density. Analyses on the computational complexity, computation time, and memory usage of the EFDLM are conducted to fully understand its fundamental features. Numerical simulations clearly show that the proposed method outperforms black-box learning method, which chooses the contrast and incident field as its inputs and the total electric field as its output. It is also demonstrated that the EFDLM is able to solve the scattering problems of dielectric objects with higher contrasts by increasing the number of subnetworks. Further, the pros and cons of the proposed learning approach for solving EM scattering problems are discussed, where some caveats are provided to avoid using learning approaches in a black-box way.