CSI-Free Deep Learning-Based Hardware-Aware Precoding for Wideband Near-Field THz Systems

Abstract

This letter investigates hybrid analog precoding for wideband near-field terahertz (THz) ultra-massive multiple-input multiple-output (UM-MIMO) systems under beam squint and practical true time delay (TTD) and phase shifter (PS) hardware limitations. We formulate the joint optimization of TTD and PS coefficients as a non-convex mixed-integer problem subject to TTD range/resolution and PS quantization constraints. To address this problem, we propose a hardware-aware unsupervised deep learning (DL) framework that operates directly on measurements, requiring channel state information (CSI) only during offline training while enabling CSI-free inference. The framework employs a hardware-aware loss function that maximizes near-field array gain while enforcing TTD and PS constraints. Simulation results demonstrate superior performance over recent schemes under practical TTD and PS limitations.