arXiv:2505.04148v2 Announce Type: replace-cross Abstract: This paper proposes a novel non-terrestrial communication architecture that integrates Rate-Splitting Multiple Access (RSMA) with a UAV-mounted Beyond-Diagonal Active Reconfigurable Intelligent Surface (BD-ARIS) under the coverage of a Low Earth Orbit (LEO) satellite. The BD-ARIS employs a group-connected structure to provide signal amplification and enhanced beamforming flexibility, while RSMA splits user messages into common and private parts to improve SE and interference management. A joint optimization of satellite beamforming, UAV placement, power allocation, and RIS configuration is formulated to maximize the system overall energy efficiency. To solve the resulting non-convex, high-dimensional problem, we first develop a model-based iterative algorithm using Dinkelbach transform, weighted minimum mean-square error reformulation, successive convex approximation, and sequential quadratic programming. We then propose two deep reinforcement learning (DRL) approaches: a standard trust-region policy optimization (TRPO) baseline and a novel Dinkelbach-Guided TRPO (DG-TRPO) scheme. DG-TRPO is initialized via supervised pretraining on Dinkelbach solutions and subsequently fine-tuned with TRPO in noisy, time-varying environments. Simulation results show that DG-TRPO closely matches the model-based benchmark while providing sub-millisecond inference, making it suitable for real-time LEO operation. It also exhibits superior robustness to CSI uncertainty and satellite mobility, and the RSMA BD-ARIS framework significantly outperforms benchmarks employing NOMA or conventional RIS designs. These results establish the proposed architecture and DG-TRPO algorithm as a scalable, energy-efficient solution for future 6G massive IoT non-terrestrial networks.