Self-Supervised Deep Learning Design for MU-SIMO Beyond-Diagonal RIS

Abstract

[Abstract]: Beyond-diagonal reconfigurable intelligent surfaces (BD-RISs) have emerged as a promising technology for wireless communications. BD-RIS enhances the capabilities of conventional RIS architectures by enabling full coupling among surface elements connected by adjustable reactances. However, leveraging such extended flexibility requires effectively addressing the added complexity in optimization tasks. In this paper, we propose a self-supervised (SS) deep learning (DL) approach for BD-RIS optimization in multi-user (MU) single-input multiple-output (SIMO) uplink systems. Unlike supervised methods that require expensive optimization solvers for training labels, the learning process of our approach is guided by a loss function derived from the optimization objective function. The proposed neural network (NN) model learns to map the channel state information (CSI) to the BD-RIS scattering matrices that maximize the system sum-rate, while guaranteeing the matrices’ structural constraints. Simulation results show that our approach outperforms traditional optimization-based methods in realistic 3GPP channels, while exhibiting lower computational complexity and maintaining comparable user fairness to state-of-the-art baselines.