Metasurface design with a complex residual neural network.

Abstract

In recent years, researchers have made great progress in solving complex electromagnetic field computing problems by using deep learning methods. However, the approaches found in literature were devoted to solving the real-number problem of electromagnetic field calculations. For the complex number problem, there was no good solution. Here, we proposed an advanced computation method for metasurfaces based on a complex residual neural network (CRNN). We predicted the scattering (S)21 parameters of a cylindrical structure in the range of 1.2 to 1.7µm wavelengths. By providing a set of cylindrical structure parameters, we could quickly predict the S 21 parameters with CRNN and design a metalens, which proved the ability of the proposed method. In addition, our method can also be extended to the calculation of electromagnetic fields where the speed of the calculation of the complex number of metasurfaces should be accelerated.