Digital-Twin-Enabled Satellite–UAV–RIS Cooperative Networks: A Reinforcement-Learning Framework for Joint Trajectory and Passive Beamforming Optimization

Abstract

Abstract—Sixth-generation (6G) networks must deliver ubiquitous, high-rate, and fair connectivity in regions where terrestrial infrastructure is sparse, damaged, or economically unviable. This paper proposes and validates a cooperative Satellite–UAV–RIS–User architecture in which a digital twin (DT) continuously mirrors the physical radio environment to drive two coupled decisions: (i) real-time user localization from received-signal-strength (RSSI) fingerprints and (ii) joint optimization of unmanned aerial vehicle (UAV) hover position and reconfigurable intelligent surface (RIS) phase/beamforming configuration. A Random Forest digital-twin classifier localizes users into service zones from seven-dimensional Wi-Fi RSSI fingerprints with 98.0% accuracy and a weighted F1-score of 0.980, evaluated on the UCI Wireless Indoor Localization benchmark (2,000 samples, four rooms). A second Random Forest surrogate learns the mapping between UAV three-dimensional position and achievable sum rate, which is embedded as the reward signal of a tabular Q-learning agent searching a 9×9 spatial grid for the throughput- and fairness-maximizing UAV waypoint. The RIS layer applies phase-optimized passive beamforming with adaptive power allocation across a 64-element panel. Monte Carlo simulation over 120 digital-twin-localized users shows that the proposed architecture achieves an average per-user rate of 417.9 Mbps and an aggregate sum rate of 50.1 Gbps, a 162.4% gain over a satellite-only baseline and a 19.1% gain over a fixed-position UAV relay without RIS assistance while sustaining a Jain fairness index above 0.999 and eliminating outage at practical throughput thresholds where the satellite-only baseline is in complete outage. A six-way ablation and a qualitative comparison against six representative 2022–2024 RIS/UAV/NTN architectures confirm that the distinguishing advantage is the closed-loop DT-driven co-design of mobility and reflection, rather than either optimization performed independently.