Physics-Informed Neural Models for Uncertain Trajectory Prediction in Urban Air Mobility

This study proposes a physics-informed neural network (PINN) framework for trajectory prediction and uncertainty quantification in urban air mobility (UAM), with a particular focus on electric vertical takeoff and landing (eVTOL) vehicles. UAM operations are subject to complex and variable urban environments, requiring accurate and reliable real-time trajectory forecasting. By embedding governing physical laws into the learning architecture, the proposed PINN model adheres to eVTOL flight dynamics, thereby improving prediction fidelity and offering more consistent uncertainty estimates. Comparative evaluations using NASA’s UAM simulation dataset demonstrate that the PINN approach outperforms conventional neural networks and Gaussian mixture regression in both accuracy and robustness. Additionally, the model enhances interpretability, making it particularly suitable for safety-critical applications. These results underscore the potential of physics-informed learning to support the development of more dependable and efficient trajectory planning tools for emerging UAM systems.