Deep Learning for the Design of Random Coding Metasurfaces

In this paper, a deep learning-based method for random coding metasurface design is proposed. This method involves constructing a residual convolutional neural network to achieve forward spectrum prediction and inverse structure design of random coding metasurfaces. The proposed forward network model can quickly predict the absorption spectrum of any given coding structure, with a computational speed approximately 100 times faster than traditional full-wave simulators. In addition, by combining the direct binary search algorithm (DBS), the inverse network model is able to accurately design random coding structures with arbitrary absorption spectrum in the near-infrared wavelength range. This approach provides a new perspective for the fast design of random coding metasurfaces.