Electromagnetic Scattering Solver for Metal Nanostructures via Deep Learning

Accurate predictions of near-field scattering of metal nanoparticles is an important mission in modern computational optics. The traditional difference algorithm usually requires a lot of computing resources and cumbersome iterative operation, which is not suitable for real-time scenarios. In recent years, deep learning as a new technology is widely used in various computational physics problems. In this paper, a neural network based on U-net is proposed to predict the near-field scattering of metal nanoparticles. The network architecture is applicable both to the two-dimensional and three-dimensional scattering problems in the ultraviolet band. As long as sufficient training samples are provided, the framework can predict the scattering field with high accuracy. Compared with the traditional method (frequency domain finite difference), its calculation acceleration is about three orders of magnitude, so we are confident that it can provide the possibility for real-time prediction of the field quantity. In addition, the network also has a strong generalization ability, which can accurately deal with the scattering problem of the real-site samples which are different from the training set samples, so it is possible to study the interaction between light and complex scatterers.