Climate & weather · 22 July 2024
A hybrid neural general circulation model for weather and climate
NeuralGCM combined a differentiable dynamical core with learned physics and produced weather forecasts competitive with leading models, as well as climate simulations that tracked observational metrics over multiple decades.
- Field
- Climate & weather
- Date
- 22 July 2024
- AI system
- NeuralGCM
- Institution
- Google Research, with ECMWF, MIT, Harvard and Google DeepMind
- Published in
- Nature
- Identifier
- 10.1038/s41586-024-07744-y
General circulation models are the foundation of weather and climate prediction: physics-based simulators that solve large-scale atmospheric dynamics and use tuned schemes for unresolved processes such as clouds. Pure machine-learning weather models had already reached high skill on short forecasts, but had not, the authors argued, demonstrated improved ensemble forecasts or enough stability for long climate simulations.
NeuralGCM is a hybrid. A differentiable dynamical core solves the governing equations for large-scale motion, while a neural network parameterizes unresolved physics. Because the whole model is differentiable, the learned components can be trained end-to-end on forecast trajectories, in contact with the dynamics, rather than being trained separately and inserted later.
The paper reports that NeuralGCM is competitive with leading machine-learning models for one- to ten-day forecasts, and with the ECMWF ensemble system for one- to fifteen-day forecasts. With prescribed sea-surface temperature, it can track climate metrics for multiple decades. At 140-kilometre resolution, climate forecasts show emergent phenomena such as realistic frequency and trajectories of tropical cyclones. The authors also state a clear limit: the model does not extrapolate to substantially different future climates.