Predicting eVTOL flight paths with physics-informed AI

What is it about?

Urban Air Mobility (UAM) aims to move people and goods across cities using electric vertical take-off and landing (eVTOL) aircraft. Flying in cities is tricky because winds and turbulence change quickly between buildings, making flight paths hard to predict. Our study uses a physics-informed neural network (PINN), a type of AI that learns from data and obeys the laws of motion. We train the model on simulated eVTOL flights from NASA's Generic UAM model dataset. During training, the AI is penalized whenever its predictions break known physics, which keeps it 'honest.' To show how trustworthy each prediction is, we also quantify uncertainty in two ways: (1) epistemic uncertainty (what the model doesn’t know due to limited data) using an ensemble of models, and (2) aleatoric uncertainty (random noise in the environment) using a probabilistic output layer. Together, these produce confidence bands around the predicted trajectory. We compare our approach to standard neural networks and Gaussian Mixture Regression and find that PINNs are more accurate and more reliable, especially during challenging phase transitions. The result is a practical method that can make eVTOL operations safer by providing both an accurate trajectory predictions and a measure of confidence.

Featured Image

Photo by Hyundai Motor Group on Unsplash

Photo by Hyundai Motor Group on Unsplash

Why is it important?

Safe UAM operations depends on accurate, real-time flight predictions and clear knowledge of when predictions may be wrong. Purely data-driven models can drift in fast changing urban winds, while traditional physics models often ignore uncertainty. Our PINN bridges this gap: it couples flight dynamics equations with machine learning and outputs confidence bands that expand during aggressive maneuvers and tighten in steady flight. That combination of physics consistency plus quantified uncertainty, is what regulators, air traffic managers, and autonomy stacks need to plan safe separations, avoid conflicts, and certify systems. Beyond eVTOLs, the same framework can accelerate trustworthy AI for drones, advanced air mobility corridors, and other safety critical vehicles.

Perspectives