The Weather and Climate of a Deep Learning Earth System Model

We explore a Deep Learning Earth System Model (DLESyM) which couples Deep Learning Weather Prediction and Ocean models. DLESyM predicts atmospheric fields every 6 hours at roughly 110 km horizontal resolution. The atmospheric component simulates the evolution of 9 fields while the ocean component predicts sea surface temperature. The unique coupling mechanism facilitates the exchange of relevant information between the two components throughout roll-outs. The coupler’s design supports component models with different temporal resolution, architectures, and predicted fields.

DLESyM provides an incredibly robust representation of our Earth System. While many other models which are designed to simulate weather and climate can manifest aphysical behavior during long autoregressive rollouts [1][2], DLESyM can be run for over 1,000 years while maintaining high fidelity to our current Earth System. Numerical climate models require extensive tuning and spin up to arrive at an equilibrium atmospheric state from which other experiments can be run [3]. Even after these costly efforts, state of the art models still suffer from model drift. DLESyM exhibits no drift or “coupling shock” even over 1,000 year roll-outs. DLESyM also conserves important quantities such as atmospheric mass without any explicit constraints imposed during training.

We find that the climate internal to long DLESyM roll-outs maintains high fidelity to observed climate. A particularly compelling aspect of this fidelity is presence across timescales: from weather to climate. To demonstrate this we examine the behavior of DLESyM in the extratropical (ET) Northern Hemisphere. On weather timescales DLESyM simulates sharp ET cyclones. DLESyM roll-outs feature blocking events which govern weekly-monthly atmospheric anomalies. Finally compounding these phenomena over many years expresses a climate, internal to DLESyM, with high fidelity to our observed climate. We find this fidelity stronger in DLESyM than historical simulations from leading CMIP6 participants. Perhaps most intriguing is that DLESyM is only trained to optimize atmospheric prediction over 24 hours, an impressive example of emergent behavior.