Hypersonic trajectory prediction based on partially observable information

The precise prediction of reentry trajectories under sparse telemetry conditions presents a formidable challenge in aerospace engineering, critically impacting mission safety and recovery reliability. While contemporary approaches leveraging numerical simulations and machine learning demonstrate complementary advantages, they fundamentally struggle to cope with sparse observation data and strong nonlinear dynamics arising from active energy management strategies. To address these limitations, we propose Deformable Cross-Attention Residual LSTM (DCAR), a physics-informed deep learning architecture that synergistically integrates energy management pattern recognition with adaptive spatiotemporal feature extraction. The extensive evaluation of the trajectory of Common Aero Vehicles (CAVs) has demonstrated the outstanding performance of DCAR, and these advances have established a new paradigm for robust reentry prediction systems.