Neural general circulation models for weather and climate

General circulation models (GCMs) are crucial for weather and climate prediction. This paper introduces NeuralGCM, a hybrid GCM combining a differentiable atmospheric dynamics solver with machine-learning components. NeuralGCM achieves accuracy comparable to leading machine-learning and physics-based methods for deterministic and ensemble weather forecasting (1-15 days) and climate simulation (multiple decades). It offers significant computational advantages over conventional GCMs, though extrapolation to significantly different future climates remains a limitation. The results demonstrate the potential of end-to-end deep learning for enhancing large-scale physical simulations of the Earth system.

Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Klöwer, James Lottes, Stephan Rasp, Peter Düben, Sam Hatfield, Peter Battaglia, Álvaro Sanchez-González, Matthew Willson, Michael P. Brenner, Stephan Hoyer