An End-to-End Neural Network for Complex Electromagnetic Simulations

Although many numerical methods can accurately solve time-domain electromagnetic (EM) simulation problems, such as finite-difference time-domain (FDTD), the computational demands are usually significant for complex scenarios. In this letter, we investigate the feasibility of applying deep learning technology to accurately solving EM forward simulations. Based on an end-to-end neural network framework, a convolutional neural network is used to extract features of scatters, and a long short-term memory neural network is used to predict EM distributions. To ensure the accuracy of the framework, especially when dealing with complicated phenomena, principal component analysis is employed to compress data sets before training. Numerical experiments show that the proposed scheme can predict EM field distributions efficiently and accurately for complex scenarios containing scatters of different materials, locations, geometrical shapes, and random numbers. The average relative mean square error is around 8. 05e−5 for scenarios with certain number of scatters and 2.25e−4 for random number of scatters respectively, which outperforms other neural network frameworks. Meanwhile, compared with FDTD, the time speedup is around 1528 times.