Quantized Deep Learning Channel Model and Estimation for RIS-gMIMO Communication

Reconfigurable intelligent surfaces (RISs) and multiuser gigantic multiple-input multiple-output (MU-gMIMO) systems are key technologies for enabling sixth-generation (6G) networks. Their numerous advantages include minimal path losses, high energy efficiency (EE), high spectrum efficiency (SE), high data rates, and compatibility with line-of-sight (LoS) and non-LoS (NLoS) paths. However, RIS-gMIMO faces numerous challenges, including pilot overhead during beam training due to a combined radiation field, high training overhead due to the cascaded channels between transceivers, inaccurate channel state information (CSI) due to the rapidly changing RIS-user equipment (UE) channel, and low-accuracy channel estimation caused by semipassive RISs. With semipassive RIS-gMIMO communications, we present a novel quantized deep learning (qDL) channel model. This proposed channel model is constructed via a radio frequency (RF) chain matrix, a combined radiation field, and a truncated activation output. To enhance the feature extraction performance and reduce the loss of the model, a novel qDL-based channel estimation scheme is also proposed to concurrently utilize denoising multilayer perceptron (DnMLP) units to satisfy the imposed sparsity constraint. The qDL scheme outperforms the previously developed benchmark schemes in terms of accuracy and performance according to the normalized mean squared error (NMSE) of the simulation results.