Deep Learning Earth System Models for Sub-seasonal to Seasonal (S2S) Forecasting

Recent advances have produced deep learning weather prediction (DLWP) models with performance comparable state-of-the-art numerical weather prediction models at spatial resolutions between ¼° and 1° latitude-longitude and forecast lead times up to 10 days. These models are very computationally efficient, suggesting the possibility of using them in huge ensembles that might improve S2S forecasting – particularly if the ensembles are well calibrated and properly capture extreme events in the tails of the forecast probability distributions.

Nevertheless, many of these DLWP models produce increasingly smooth forecasts at longer lead times and their suitability for the direct Monte Carlo generation of useful large ensembles on S2S time scales is unclear. Here we present first results from a parsimonious deep learning earth-system model with coupled atmospheric and ocean modules. As with more traditional numerical models, free- running forecasts continue to simulate realistic atmospheric states at multi-year lead times, and interestingly, they are stabilized by the ocean-atmosphere coupling. We examine the model’s ensemble performance on El Niño forecasts and at shorter lead times. We conclude by exploring the potential to expand this approach to a full earth-system model.