Deep reinforcement learning for beam pointing control in ground-to-LEO laser communications under time-varying turbulence

Atmospheric turbulence significantly limits the performance of ground-to-satellite optical uplinks by causing beam wandering and scintillation. This study presents a reinforcement learning (RL)-based framework for adaptive beam pointing to mitigate these effects. A high-fidelity wave optics simulation with angular spectrum propagation and autoregressive phase screens was used to model uplink turbulence. A recurrent soft actor–critic (RSAC) algorithm with short-wave infrared (SWIR) sensor inputs was trained under baseline turbulence condition (C²_n = 10⁻¹⁴m^−2/3) to dynamically adjust the transmitted beam pointing at the optical ground station (OGS) and maximize received optical power. Compared to static pointing, the RL-based approach achieved 4–8 % higher mean received power and reduced power variance by 17–54 %. These results demonstrate that RL can effectively exploit temporal turbulence correlations, offering a robust solution for real-time adaptive beam control in free-space optical communication uplinks.