8 November 2025 in Aerospace Science and Technology
18 August 2026 in SSRN Electronic Journal
Accurate multi-aircraft trajectory prediction helps air traffic managers anticipate congestion, coordinate traffic flows, and detect emerging separation risks in dense terminal airspace. Thetask is difficult because aircraft motions combine individual intent with responses to surrounding traffic, effects that existing predictors usually encode implicitly through discrete prediction steps. This study develops a physics-informed neural ordinary differential equation framework that combines explicit interaction modeling with continuous-time kinematics. A spatiotemporal graph encoder summarizes the observed traffic scene, and a dual-branch latent evolver separates the primary aircraft’s motion intent from its response to neighboring aircraft. The social branch applies distance-aware attention to relative geometry and closing speed. A second-order decoder transforms the two latent representations into acceleration fields and integrates velocity and position through a structured state equation. The framework is trained on heterogeneous traffic before the encoder and latent evolver are frozen and the differential equation module is adapted to four terminal classes. One shared parameter set serves all classes at inference. Experiments use five days of broadcast surveillance data from the Guangdong–Hong Kong–Macao Greater Bay Area. Under a common specialist comparison protocol, runs with three random seeds give an average displacement error of 3.36±0.20 km and final displacement error of 7.63±0.57 km. These values are 11.9% and 8.7% lower, respectively, than those of the strongest evaluated learned baseline. The framework also provides the best balance between alert precision and recall in phase-aware separation warning. Ablation experiments examine the principal design choices,while attention analysis, acceleration-field decomposition, and controlled neighbor removal characterize responses across classes. The findings indicate that combining explicit interaction structure with continuous-time kinematics can improve forecast reliability and support earlier, more transparent tactical and traffic management decisions in complex terminal airspace.