Dynamic Risk Anticipation for Airborne Satellite Resilience through Physics-Based Safe Reinforcement Learning

Airborne satellite communication is increasingly used in safety-critical aviation operations. Existing switching schemes often rely on fixed thresholds, particularly vertical-rate indicators, to separate stable cruise from dynamic flight conditions. This design can miss risk when the vertical rate stays moderate but anisotropic turbulence alters aircraft attitude dynamics and antenna pointing error. We address this problem by treating airborne satellite switching as a resilience engineering problem. We develop a physics-based risk metric from a noise-driven escape model that describes nonmonotonic risk under anisotropic turbulence and use it to guide an adaptive switching controller. A control barrier function safety layer constrains actions that move the system toward elevated-risk states. The framework is evaluated using operational data from ten commercial flights across five aircraft types; final policy metrics are computed on two held-out synchronized flights, with additional cross-aircraft consistency checks reported in the appendices. The proposed policy reduces risk exposure and recovery time while keeping communication availability close to the fixed-threshold baseline, and it adjusts switching thresholds before channel degradation occurs, indicating anticipatory rather than purely reactive behavior. An ablation analysis shows that both the risk metric and the safety filter contribute to the observed resilience gains. These findings suggest that physically interpretable risk sensing can help airborne satellite links respond earlier to elevated-risk conditions and recover more effectively after disturbances.