Answers are grounded in this paper and cite their sources.
Significant wave height is a key statistic for marine operations. Physical constraints must be considered for data-driven machine learning models to ensure consistency with the underlying wave dynamics. We develop a residual neural network to predict significant wave height in the waters surrounding Japan. The predictors are significant wave height, mean wave period, mean wave direction, model bathymetry, 10 metre wind speed, and 10 metre wind direction. We enhance the model with two physics-informed considerations. Firstly, swells propagating into the region of interest result in unobserved external wave information which can degrade prediction accuracy. To combat this, we use the mean group speed to construct a border that insulates against boundary-condition uncertainty from incoming swells, reducing the mean-squared testing error near domain boundaries by up to 9.5%. Secondly, we adapt the size of the spatial convolutional kernels to account for wave speed. We estimate mean group speeds from the data to define two physics-based spatial kernel sizes. The adaptation of the convolutional kernels further reduces the mean-squared testing error by 6.5%. Finally, a Shapley value analysis indicates that model directional data are unnecessary as wave and wind directionality are implicitly learned by the model via spatiotemporal trends in the data.
No takes yet. Share an insight, caveat, or question.
Surtel et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: