Process-oriented evaluation of machine learning and physics-based models for wave parameter prediction under extreme conditions in a fetch-limited sea

Machine learning approaches are increasingly used for wave prediction, yet their ability to represent the physical processes governing extreme waves in fetch-limited seas remains insufficiently understood. This study evaluates four machine learning models for reconstructing significant wave height (Hs), peak wave period (Tp), and mean wave direction (θ) at the Östergarnsholm site in the Baltic Sea, using ERA5 atmospheric forcing and benchmarking against a physics-based wave model (WAM). Results are interpreted through a physically informed framework linking model skill to temporal memory, spatial integration, and physical constraints. Recurrent networks reproduce Hs with skill comparable to WAM (r = 0.95, RMSE = 0.20 m), highlighting the importance of temporal memory. In contrast, Tp shows strong dependence on spatial forcing, with multi-point recurrent models outperforming WAM (RMSE = 0.81 s vs 0.95 s), reflecting improved representation of basin-scale storm evolution. Spatial predictors have a dual effect: they improve performance in tree-based models by providing missing large-scale context, but can reduce robustness in high-dimensional recurrent networks under limited training data. θ remains challenging, although recurrent models achieve comparable skill to the physical model. These findings show that model performance depends on the alignment between model structure and underlying physical processes.