Joint Beamforming for RIS-Assisted Hybrid MIMO Systems Based on Model-Driven Deep Learning

Hybrid beamforming technology is pivotal for achieving high-rate communications in multiple-input multiple-output (MIMO) systems. However, its performance struggles, particularly when blockages interfere with the direct link between transceivers. Reconfigurable intelligent surfaces (RISs) that facilitate transmission across blockages can mitigate this problem. Nevertheless, practical implementation of existing algorithms for RIS-assisted hybrid beamforming is hindered by numerous iterations to converge. This paper integrates a RIS to assist hybrid beamformers at the transceivers and tackles this challenge by adopting a model-driven deep learning network, which speeds up convergence and reduces computational overhead of iterative algorithms. Specifically, the RIS and hybrid beamforming are jointly designed to maximize spectral efficiency. The joint optimization problem is decoupled into three sub-problems for simplicity, which are separately solved by Riemannian manifold optimization (RMO). The iterative RMO algorithm is unfolded into a model-driven network and learns the key algorithm parameters with trainable variables. Simulation results demonstrate the performance advantages of our proposed network compared with other existing schemes, while the complexity of our proposed network is lower than that of the competitors as well.