End-to-End Deep Learning for CSI Estimation and Feedback in Large-Scale RIS-Assisted Wireless Systems

Abstract

To fully harness the potential of reconfigurable intelligent surfaces (RIS)-assisted communication systems, accurate channel estimation and reliable acquisition of full channel state information (CSI) are essential. However, this task becomes increasingly challenging due to the high dimensionality of the channel, particularly in large-scale RIS deployments with hundreds or thousands of reflecting elements. In this paper, we propose a deep learning (DL)-based end-to-end framework that jointly performs CSI estimation and feedback for RIS-assisted wireless systems. The proposed framework first employs a denoising super-resolution network (DSRNet) to reconstruct full CSI from available partial CSI. This is followed by a novel encoder-decoder-based CSI feedback network, termed DCFNet, where the encoder compresses the estimated CSI into a compact codeword, and the decoder reconstructs the CSI using residual denoising and channel refinement sub-blocks. By jointly optimizing CSI estimation, compression, and reconstruction, the proposed framework reduces the mismatch between separately trained estimation and feedback modules. The simulation results indicate that the proposed E2E-CsiNet outperforms representative CSI feedback methods and deep learning–based end-to-end estimation and feedback baselines. It achieves superior accuracy in CSI reconstruction. In addition, DCFNet achieves a favorable accuracy–complexity tradeoff. It reduces trainable parameters and memory usage compared with several feedback baselines while maintaining a practical inference time. These results highlight the effectiveness of the proposed framework for scalable, large-scale RIS CSI acquisition, even under conditions of partial CSI availability and noisy feedback transmission.