End-to-End DNN-based Joint Combining Matrix Optimisation and Channel Estimation in Near-Field Large-Scale THz Communications

Terahertz (THz) communication has emerged as a critical technology for next-generation wireless communication systems due to its exceptionally wide bandwidths and ultrahigh data rates. However, channel estimation in THz bands faces severe challenges, especially in near-field regions where planarwave assumptions break down. Conventional compressive sensing (CS) and machine learning (ML) based channel estimation methods often struggle with near-field intricacies and large-scale antenna arrays, necessitating novel estimation techniques. Motivated by the success of autoencoder-based channel estimation approaches in end-to-end system design, this work proposes a novel deep neural network (DNN)-based THz channel estimation technique tailored for near-field THz communication systems with a large-scale Widely-Spaced Multi-Subarray (WSMS) architecture. By exploiting near-field properties and leveraging the spatial diversity provided by WSMS and large-scale arrays, the proposed framework achieved a $\mathbf{6 ~ d B}$ improvement in normalised mean square error (NMSE). Simulation results confirm that this technique outperforms state-of-the-art channel estimation schemes, thus paving the way for more efficient and accurate near-field THz channel estimation in future wireless communication systems.