Machine learning-aided channel estimation and hybrid combining for ultra-massive MIMO THz UAV communications

Official URL: https://dx.doi.org/10.1109/BalkanCom71095.2026.11605049

Terahertz (THz) communication is considered a key enabling technology for future sixth-generation (6G) wireless networks due to its potential to support extremely high data rates. In this paper, we investigate an uplink THz communication scenario with unmanned aerial vehicles (UAVs) as user equipment and a master UAV equipped with an ultra-massive multiple-input multiple-output (MIMO) array employing an antenna-ofsubarrays (AoSA) architecture. While this architecture reduces hardware complexity, it introduces challenges related to compressed channel observations and high spatial correlation, which complicate channel estimation and signal detection. To address these challenges, machine learning-based channel estimation using the Fixed-Point Network Orthogonal Approximate Message Passing (FPN-OAMP) algorithm is applied, and hybrid analog-digital combining schemes are investigated. In addition, a combining method based on Simultaneous Perturbation Stochastic Approximation (SPSA) is proposed to optimize the analog combiner with respect to the detector performance. Simulation results show that the proposed approach improves channel estimation accuracy and achieves significant BER improvements under realistic conditions. The proposed framework is particularly relevant for emerging 6G non-terrestrial networks, where UAV-assisted THz communications are expected to play a key role.