Near-Field Beamforming for Terahertz Communications With NLG and Massive MIMO

This paper investigates deep learning-based near-field beamforming for Terahertz (THz) wideband massive MIMO systems, addressing beam-splitting effects, severe path loss, and hardware constraints inherent to THz frequencies. The proposed framework integrates quantized phase shifters (PS) and time-delay (TD) units within a partially connected hybrid beamforming architecture, enabling more efficient and frequency-adaptive beamforming across multiple subcarriers. Then, a reinforcement learning-based optimization is used to jointly configure phase shifts and time delays, significantly enhancing beamforming efficiency while reducing computational complexity. After that, an optimization problem is formulated aimed at maximizing the average signal-to-noise ratio (SNR) across subcarriers and develops a novel decomposition scheme to separately optimize phase shifters and TD units, allowing for more practical hardware implementation. A reinforcement learning framework inspired by actor-critic network is further employed to efficiently search for optimal phase configurations, leveraging a signal model-based critic network that reduces computational overhead of natural language generation (NLG) based networks. Meanwhile, a low-complexity, geometry-assisted algorithm is introduced to determine TD unit configurations, mitigating beam-splitting effects and ensuring consistent phase alignment across subcarriers. Finally, simulation results are provided to demonstrate that the proposed TD-PS hybrid architecture achieves a 5 dB improvement in beamforming gain over the phase-shifter-only scheme while maintaining robust performance across a 95 GHz to 105 GHz frequency range. Additionally, compared to traditional exhaustive search-based beamforming optimization, the reinforcement learning-based phase shifter design reduces training iterations by 80%, making the proposed scheme computationally feasible for large-scale antenna arrays.