Learning-Based BD-RIS Architecture Design for WET Systems

This paper investigates a wireless energy transfer (WET) system assisted by a beyond-diagonal reconfigurable intelligent surface (BD-RIS), in which the BD-RIS topology, its scattering matrix, and the transmit beamforming are jointly optimized. To address the resulting non-convex problem, an unsupervised learning approach based on two neural networks trained in two stages, followed by local refinement, is proposed. The proposed solution aims to maximize the harvested energy while reducing the architectural complexity of the BD-RIS. Simulation results show energy performance close to that of the \textit{Fully Connected} architecture, but with a significantly smaller number of active connections.