Transformer-Driven Beam Control via Reconfigurable Antenna Arrays for Terahertz UAV Communications

Terahertz (THz) unmanned aerial vehicle (UAV)-to-UAV communication offers multi-gigabit-per-second throughput through large available bandwidths, but requires highly directional beams to overcome severe path loss and atmospheric attenuation at these frequencies. On the other hand, the rapid and irregular mobility of UAVs results in frequent beam misalignment, which reduces link capacity. To address these challenges, this paper introduces COntext-Reasoned Transformer for Efficient Beam IndeXing (CORTEX) as a two-stage antenna-aware deep learning framework, which enables beam control for THz UAV communications via reconfigurable antenna arrays. CORTEX learns beam selection directly from 255 full-wave electromagnetic radiation patterns, generated by an 8×8 row-selective microstrip patch antenna array, operating at 272.5 GHz, thereby explicitly incorporating antenna radiation characteristics into the learning process. In Stage I, future angular coordinates are predicted from UAV mobility sequences using a spherical geometry-aware Wrapped Cauchy loss. In Stage II, these predictions are mapped onto discrete antenna configurations through codebook cross-attention, preserving realistic sidelobe structures that enhance robustness under transient misalignment. Comprehensive evaluations considering five representative UAV mobility profiles, including real flight traces and other realistic trajectories, demonstrate that the proposed method achieves near optimal performance with a mean angular prediction error of 0.87° and average capacity performance of 93.78% of the exhaustive search capacity. Moreover, CORTEX is more than three times faster than the exhaustive search in inference time, incurring only 0.36 ms latency, promising highly effective beam control for sustaining reliable THz UAV links under challenging aerial mobility conditions.