Comparative Study of Ship Motion Prediction Models: Data-Driven Physics-Based vs Pure Machine Learning
Two data-driven routes to vessel trajectory prediction go head-to-head: optimising the parameters of a physics-based hydrodynamic model versus a pure feed-forward neural network. Trained and 5-fold cross-validated on 90 simulated trajectories, the physics-based-with-optimisation approach wins — roughly 40% lower summarised error and 35% tighter consistency than the black-box model.
Abstract
Accurate ship motion prediction is essential for safety and efficiency in maritime navigation. Data-driven techniques offer potential accuracy improvements to ship model prediction, yet their utility and reliability are being explored. This study compares two data-driven approaches for vessel trajectory prediction: a data-driven optimization of physics-based model parameters and a purely machine learning approach. Both models are trained and evaluated based on a dataset comprising of 90 trajectories generated via a vessel simulator. The physics-based model is based on hydrodynamic principles while key parameters are fitted, using constrained nonlinear least squares, to tailor its prediction accuracy to the dataset. For the machine learning approach, a feed-forward neural network learns motion patterns directly from data without prior domain knowledge. A 5-fold cross-validation is employed for both models ensuring robust evaluation. The models' prediction accuracy is quantified using: (i) Euclidean Distance, (ii) Euclidean Distance with Heading Penalization, and (iii) Custom Vessel Distance Measure. The results across the dataset show that the data-driven physics-based model achieves higher prediction accuracy (40 per cent lower summarized absolute error) and consistency (35 per cent lower summarized error spread). Individual prediction scenarios are examined to highlight the trade-offs between accuracy, reliability and adaptability of the two approaches.
Index terms: autonomous shipping, machine learning, parameter optimization, data-driven models.
Why it matters
For autonomous and assisted navigation, a motion model has to be both accurate and trustworthy in conditions it has never seen. This study shows that grounding a model in hydrodynamics and then fitting its open parameters to data beats a pure black-box network — not only on average error, but on consistency. That consistency is what matters when the prediction feeds a collision-avoidance or control decision, where an occasional large miss is far more costly than a slightly higher average error.
Cite this paper
Michail Mathioudakis, Petros Iatropoulos, Theodoros Stouraitis, Christos Papandreou, Antonis Nikitakis, Stavros Paschalakis and Konstantinos Kyriakopoulos (2025). Comparative Study of Ship Motion Prediction Models: Data-Driven Physics-Based vs Pure Machine Learning. IEEE Symposium on Maritime Informatics & Robotics (IEEE Maritime 2025), Ermoupoli, Syros, Greece. https://blueautonomy.gr/insights/papers/comparative-ship-motion-prediction/