Neural General Circulation Models

• DOI: 10.48550/arxiv.2311.07222
• Publication Date: 2023-01-01
• Authors: Dmitrii Kochkov; Janni Yuval; Ian Langmore; Peter Nørgaard; Jamie Smith; Griffin Mooers; Milan Klöwer; James Lottes; Stephan Rasp; Peter Düben; Sam Hatfield; Peter Battaglia; Álvaro Sánchez‐Go...; Matthew Willson; Michael P. Brenner; Stephan Hoyer
• Keywords: Model output statistics

General circulation models (GCMs) are the foundation of weather and climate prediction. GCMs physics-based simulators which combine a numerical solver for large-scale dynamics with tuned representations small-scale processes such as cloud formation. Recently, machine learning (ML) trained on reanalysis data achieved comparable or better skill than deterministic forecasting. However, these have not demonstrated improved ensemble forecasts, shown sufficient stability long-term simulations. Here we present first GCM that combines differentiable atmospheric ML components, show it can generate forecasts weather, par best methods. NeuralGCM is competitive 1-10 day European Centre Medium-Range Weather Forecasts prediction 1-15 forecasts. With prescribed sea surface temperature, accurately track metrics global mean temperature multiple decades, 140 km resolution exhibit emergent phenomena realistic frequency trajectories tropical cyclones. For both climate, our approach offers orders magnitude computational savings over conventional GCMs. Our results end-to-end deep compatible tasks performed by GCMs, enhance physical simulations essential understanding predicting Earth system.