A Deep Learning-Based Hybrid Precoding With True-Time-Delay for THz Massive MIMO Systems

Terahertz (THz) Massive multiple-input multiple-output (MIMO) has been considered to be a crucial technology for improving the spectrum efficiency of 6G communications. However, in THz massive MIMO systems, it is difficult for classical hybrid precoding to handle the beam split effect and the non-convex optimization problem under imperfect channel state information (CSI). Therefore, to address these issues, this paper proposes a novel hybrid precoding scheme based on a true-time-delay (TTD) enabled deep neural network (DNN) that learns in an unsupervised manner for THz massive MIMO systems. Specifically, a TTD-based hybrid precoding structure is first used to prevent serious loss of performance due to the beam division effect. Then, to solve the non-convex optimization problem, an unsupervised learning DNN mode hybrid precoding scheme under imperfect CSI is designed. The proposed scheme exhibits good robustness under imperfect CSI conditions and excellent performance, which is verified by simulation results.