A Neural Network based Electromagnetic Simulator

Abstract

Simulating electromagnetic problems using the finite difference method or the finite element method can lead to large systems of linear equations which need to be solved. Often in the design process, while fine tuning, few system parameters are changed, while the overall system remains largely the same. The system is simulated repeatedly to find the optimal parameters, which can be a time-consuming process. In this paper we propose a new method that uses a neural network trained on a lot of variations of similar problems that can be used to get a quick estimation of the system’s response to small changes in the parameters. Rather than attempting to solve the electromagnetic problem with a neural network, which has been done before, we focus on getting an extremely fast, but also accurate estimation. A concrete example problem is demonstrated through the simulations of a coaxial a cable with varying inner conductor shapes. Details about the design of the neural network regarding the choice of hyperparameters and the network’s architecture are given. Additionally, an evaluation shows the performance of different proposed neural network architectures.