Learning-Based Beamforming for Joint Rate and Energy Harvesting in RIS-Aided Systems

In this work, we propose a deep learning-based beamforming framework that simultaneously maximizes data rates and energy harvesting (EH) in Reconfigurable Intelligent Surface (RIS)-assisted Terahertz (THz) Massive Multiple-Input Multiple-Output (MIMO) systems. Operating within the Rate-Splitting Multiple Access (RSMA) paradigm and a Simultaneous Wireless Information and Power Transfer (SWIPT) environment, the design integrates a dual-layer power splitting strategy with a Transformer-based deep unfolding network for generating digital beamformers. This architecture enables multi-objective optimization, effectively balancing spectrum efficiency (SE) and energy harvesting (EH) performance. The Transformer model dynamically predicts digital beamformers while jointly incorporating EH-related parameters into its learning process. Simulation results reveal significant improvements in SE and EH compared to conventional RSMA and Space Division Multiple Access (SDMA) baselines.