Deep Learning-Based Hybrid Precoding for RIS-Aided Broadband Terahertz Communication Systems in the Face of Beam Squint

In this letter, we design a wideband hybrid reconfigurable intelligent surface network (WHRIS-Net) based on deep learning for reconfigurable intelligent surface (RIS)-assisted terahertz massive multiple-input multiple-output systems with beam squint. Firstly, the mean channel covariance matrices (MCCMs) from the base station to the RIS and from the RIS to the user are used as the inputs of WHRIS-Net. Then, a Phase-Net is applied to calculate the frequency-independent analog precoder and the phases of RIS elements. Finally, a Digital-Net designs the frequency-dependent digital precoder utilizing the MCCMs and the outputs of Phase-Net. Numerical simulations validate the effectiveness of the WHRIS-Net in terms of the sum rate and demonstrate its promising performance gain despite reduced average running time and feedback overhead.