Learning Subseasonal-to-Seasonal Global Ocean Forecasting on a Hierarchical Triangle Mesh

The accuracy of medium-range weather predictions has steadily improved over the years. However, risk mitigation strategies for weather extremes require longer-term forecasts. Subseasonal-to-seasonal forecasts address this need, traditionally relying on ensembles of physics-based atmosphere-ocean coupled simulations. More recently, data-driven models emerged as a competitive alternative. Still, these models lack a description of the ocean, whose coupling becomes critical at these timescales. To address this issue, we propose a data-driven model for subseasonal-to-seasonal global ocean forecasting based on graph neural networks. We describe and motivate architectural choices, and present improvements based on analogies with finite element methods required to deal with irregular domains.