Deep Learning-Based Composite Channel Estimation for Reconfigurable Intelligent Surfaces-Aided Systems

The performance gains offered by reconfigurable intelligent surfaces (RIS) in wireless networks critically depend on the availability and accuracy of channel information. Channel estimation becomes particularly challenging when the RIS comprises passive elements without sensing or processing capabilities, as the channels cannot be measured directly and must instead be inferred indirectly by an external, non-colocated controller. In this work, we propose a deep learning-based approach for composite channel estimation in RIS-assisted communication systems. Unlike traditional methods that estimate the direct and reflected channels separately, our approach integrates the effects of both paths into a unified model and employs deep learning to capture the complex nonlinear interactions introduced by RIS phase shifts. This method reduces pilot overhead and enhances the efficiency of channel estimation. Simulation results reveal that the proposed method significantly outperforms other existing methods in RIS-assisted communication systems.