A Sea Surface Model for Coupled Data-Driven S2S Forecasting

Abstract

Data-driven modelling of the atmosphere has rapidly become a vibrant area of research. Recent studies have shown these models have the ability to outperform existing state-of-the-artnumerical weather prediction models. Many of these efforts, however, remain targeted at short range forecasts (within 2 weeks). We propose using recent advancements in machine learning to extend the window of predictive skill to the seasonal to subseasonal timescales (2-10 weeks). To do this we believe capturing couplings between Earth system components is necessary. Toward this end we have developed an entirely data-driven sea surface model. Our model predicts global sea surface temperature at daily resolution and can be run iterative like traditional circulation models. We find that even without atmospheric influence, our ocean model can produce skillful forecasts, consistently beating persistence and outperforming a climatological forecast out to 60 days. We also succeed in predicting the extreme El Nino event of 2015 at extended leadtimes. Interestingly, our models can run freely for over a year without producing unstable behavior even though they have no prescribed physical constraints such as conservation of energy. Furthermore we show that adding information about the atmosphere can significantly improve upon model performance suggesting that a these architectures are capable of learning coupled atmosphere-ocean interactions. Our study is an important step toward developing a fully coupled Deep Learning Earth System Model.