Channel Estimation for Indoor Terahertz UM-MIMO: A Deep Learning Perspective for 6G Applications

Abstract

The emergence of terahertz (THz) communication in ultra-massive multiple-input multiple-output (UM-MIMO) systems presents new challenges for accurate and efficient channel estimation, particularly under hybrid-field propagation conditions. Conventional estimation techniques struggle to meet the demands of such high-dimensional systems, especially in the presence of limited radio frequency (RF) chains and mixed near- and far-field effects. To address these limitations, this paper proposes a deep learning-based framework that combines a fully connected neural network (FCNN) for linear channel estimation with a convolutional neural network (CNN) for non-linear refinement. The architecture is designed to adapt to diverse propagation environments while maintaining computational efficiency. Simulation studies based on realistic THz scenarios demonstrate that the proposed approach significantly improves estimation accuracy, achieving up to 90% reduction in normalized mean squared error (NMSE) compared to traditional and advanced estimation techniques. The robustness of the model under varying signal-to-noise ratios and noise power levels underscores its potential for deployment in future 6G THz communication networks.