Using Deep Learning for Coupled Atmosphere-Ocean Modelling

We present a Deep Learning Earth System Model for S2S forecasting. Building on successful atmospheric models [1], we create a Deep Learning Ocean Model (DLOM) designed to couple with our Deep Learning Weather Prediction (DLWP) models. The DLOM forecasts sea surface temperature (SST) and height (SSH). DLOMs use deep learning techniques as in our DLWP models but are configured with different architectures and slower time stepping (2-day resolution). These differences reflect distinct dynamical properties of the ocean and atmosphere. As with the atmospheric model, the SST module can be recursively stepped forward to give realistic results over a full annual cycle and shows rapid performance compared to traditional numerical circulation models. Our DLOMs and DLWP models are trained to learn atmosphere-ocean coupling. We forecast both as one continuously evolving Earth System.

The performance of both modules is diagnosed in their coupled and uncoupled modes. We assess the improvements gained by our coupling strategy by examining individual synoptic events, skill scores, and long-range prediction of subseasonal to seasonal (S2S) and seasonal modes of earth system variability. A particular focus is on El Nino Southern Oscillation (ENSO) forecasts within the SST module.

[1] Weyn, Jonathan A., Dale R. Durran, Rich Caruana, and Nathaniel Cresswell-Clay. “Sub-Seasonal Forecasting With a Large Ensemble of Deep-Learning Weather Prediction Models.” Journal of Advances in Modeling Earth Systems 13, no. 7 (2021): e2021MS002502. https://doi.org/10.1029/2021MS002502.

Figure Caption: Top panel shows monthly averaged SST anomalies in the Nino3 region forecasted by the uncoupled DLOM ensemble initialized July 1, 2015 (dash-dot yellow) and monthly averaged ERA5 SST anomalies in the Nino3 region (black solid). DLOM models are forecasted at 2-day resolution and trained to optimize forecasts over the first 9 days. Bottom panel shows the same analysis using the UK MetOffice ensemble forecasting system (dash-dot orange) and observational analysis.