Recurrent Transformer Neural Network-Based Hybrid Far- and Near-Field THz Channel Estimation

The use of terahertz, extremely large-scale multiple-input multiple-output (XL-MIMO) systems in 6G is driven by its potential to deliver unprecedented data rates, alleviate spectrum congestion, and improve network efficiency and performance. Due to the increase in antenna aperture and carrier frequency of such systems, the Rayleigh distance increases dramatically, what leads to users being located in both the near-field and conventional far-field regions. To construct the spatial precoders with high accuracy, the base station needs to estimate the channel with high precision. This task is challenging due to the combination of near- and far-field responses and the limited number of digital chains in hybrid beamforming systems. This paper proposes utilizing the recurrent block transformer model to solve this problem. We demonstrate that a single transformer block with state memory can be trained and recurrently applied for hybrid-field channel estimation for the required number of iterations. Simulation results show significant gains over state-of-the-art solutions present in the literature.