Development of a Coupled Ocean-Atmosphere Emulator

In recent years, Earth system models that use machine learning (ML) techniques have shown promise for forecasting Earth’s weather when trained on high-quality atmospheric reanalysis datasets such as ERA5. Examples of such models include PanguWeather, GraphCast, FourCastNet, NeuralGCM, and AIFS, which can produce forecast skill similar to state-of-the-art physical models. However, very few of the existing models consider the problem of coupled prediction involving the interior ocean, ice, land, and atmosphere. Here, we present progress towards building a coupled ocean-ice-land-atmosphere emulator by leveraging advances made in building atmosphere-only emulators. In particular, we build on the capabilities of the Graph Neural Network (GNN) architecture employed in GraphCast and additionally include ocean variables by training the network on the coupled ocean-atmosphere “UFS-Replay” dataset. This dataset uses the NOAA coupled UFS forecast model “nudged” towards ERA5 and ORAS5 to produce a consistent ocean-atmosphere dataset similar to a true reanalysis. Our approach treats the Earth system as a whole by using a single multivariate state vector that includes variables from multiple Earth system components, which should explicitly capture cross-component interactions, instead of using a coupler between separate ocean and atmosphere models. Results from initial tests and future work will be presented.