Teaching neural networks to predict ship movements with built-in uncertainty

Predicting ship trajectories from sparse, irregular tracking data is crucial for maritime safety, but standard neural approaches don't quantify uncertainty reliably. This work adds structure to Bayesian Neural ODEs by imposing Gaussian process priors on the vector field rather than just weights, encoding smoothness and locality. Combined with probabilistic multiple shooting for long trajectories, the method delivers better-calibrated uncertainty estimates on real AIS datasets.