Generation of Vertical Profiles with Neural Ordinary Differential Equations Trained on Open Trajectory Data

Abstract

Recent advances in Neural Ordinary Differential Equations (Neural ODE) have shown that high-fidelity aircraft dynamics can be learned from flight recorder data, but such proprietary datasets remain largely inaccessible. In this study, we extend these principles to open trajectory data by training a Neural ODE model on Automatic Dependent Surveillance–Broadcast (ADS-B) and Mode~S Enhanced Surveillance (EHS) information retrieved from the OpenSky Network. The model focuses on reconstructing the vertical dynamics of transport aircraft using only openly available surveillance variables. The resulting framework learns continuous-time dynamics that remain physically consistent through embedded kinematic relations, demonstrating that realistic vertical profiles can be generated solely from open surveillance data. This work contributes to reproducible, data-driven performance modelling and supports the broader adoption of open, physics-guided learning methods in aviation research.