A Convolutional Neural Network-Based Method for Accurate Computation of Scattered Fields From Reconfigurable Intelligent Surfaces

Abstract

The cell-to-cell coupling in a reconfigurable intelligent surface (RIS) is very different from a periodic structure, where coupling effects can be precisely evaluated via full-wave analysis with periodic boundary conditions. We propose a novel method based on convolutional neural networks (CNNs), to predict the contribution of mutual coupling on the near-zone tangential electric field of every RIS unit cell that characterizes its scattering. Our CNN model incorporates an attention mechanism based on the squeeze-and-excitation block module, enhancing its capability to discern and quantify coupling effects, especially from neighboring cells surrounding the unit cell of interest. The predictions of the model enable the computation of RIS scattered fields, fully accounting for the aperiodic nature of an RIS. Comparisons to finite-element analysis confirm that our computed fields are accurate at any point and for any RIS configuration. Furthermore, our trained model can be re-trained through transfer learning, to accurately and efficiently predict cell-to-cell coupling under different incident wave conditions, utilizing only a reduced number of training data. Therefore, the proposed method is a valuable tool for various practical applications, such as synthesizing RIS scattered field patterns and evaluating the performance of RIS-enabled channels.