Decadal wave reconstruction in the Mediterranean Sea with graph neural networks

Abstract

Accurate simulation and prediction of ocean waves are essential for coastal risk management and climate studies. Deep learning has shown promising results for wave modeling, but most approaches still operate on regular grids and on forecasting time scales, and do not generalize to unstructured discretization or to long time horizons. Here we present WaveGraph, a model based on Graph Neural Networks (GNNs) that emulates basin-scale wave dynamics directly on unstructured meshes with high resolution along the coasts (up to 2-3 km). Trained on bias-corrected simulation data over the Mediterranean Sea, WaveGraph uses a multiscale architecture combining the unstructured model mesh with a uniform graph, allowing simultaneous representation of local coastal interactions and large-scale wave dynamics. The model reconstructs the evolution of significant wave height, mean period, and mean direction, and is applied autoregressively for a continuous 17-year period without reinitialization or drift. Validation against buoy and satellite observations shows skill comparable to the input data set, and ablation experiments indicate that wind forcing drives most of the long-term stability while wave history improves swell-driven and basin-scale dynamics. These results show that GNNs can provide stable and efficient emulators of spectral wave models on unstructured domains, enabling decadal wave reconstructions.