Convolutional neural network for coded metasurface inverse design

Abstract

A coded metasurface design framework based on convolutional neural networks and fully connected networks is constructed to achieve an efficient reverse design process. By feeding the structure coding matrix into the forward prediction neural network, the network can quickly infer the transmission spectrum of the metasurface corresponding to the structure coding matrix in milliseconds. On the other hand, the reverse design network can effectively learn and grasp the deep relationship between transmission spectrum and metasurface. When the desired target transmission spectrum is input into the reverse design network, it can efficiently generate the metasurface structure matrix that meets the specific requirements. Compared with the traditional simulation design method, the proposed scheme greatly reduces the design time and improves the work efficiency.