Deep Learning Assisted Hybrid Beamforming for THz Hybrid Near and Far Field Communications

For reconfigurable intelligent surface (RIS)-assisted terahertz systems, a hybrid beamforming scheme under a mixedfield environment is investigated. A deep learning-based framework that integrates channel estimation and beamforming is proposed to maximize the weighted sum-rate. To obtain independent channel state information (CSI), a positioning-based approach is employed, and a mixed-field channel estimation model is trained using deep learning techniques. Inspired by the gradient descent (GD) method, a system model capable of iteratively updating the target matrix is designed, which optimizes the beamforming based on the estimated CSI to maximize the weighted sum-rate. Simulation results demonstrate that the proposed method significantly outperforms traditional schemes in weighted sum-rate and approaches the performance under perfect CSI conditions.