Neural General Circulation Models for Weather and Climate
Abstract
Atmospheric modeling is crucial for predicting weather and understanding the climate system. Recent deep learning models such as Pangu and GraphCast have shown that pure machine learning (ML) can be competitive with standard numerical weather prediction models for global weather forecasting. However, pure ML models are hard to interpret, may produce non-physical and blurry forecasts, and cannot be straightforwardly combined with existing models for other parts of the Earth system.
An alternative hybrid approach uses ML components to replace uncertain parameterizations in combination with governing equations for resolved scales. Although hybrid atmospheric models have shown some success, existing attempts encounter challenges such as instabilities, climate drift, applicability limited to idealized scenarios, and modest improvements on longer time scales in realistic scenarios. Some limitations of hybrid models stem from their current training methodology, an "offline" approach, where the ML component remains unaware of dynamics during training. We introduce an innovative modeling paradigm applicable to both weather forecasting and climate modeling. We built a differentiable atmospheric model that unifies ML techniques with numerical solvers for the governing equations. This approach enables "online training," allowing the optimization of ML tunable parameters while considering the interaction between the dynamical core and the ML component. Our approach, the Neural General Circulation Model (Neural GCM), achieves state-of-the-art accuracy for 10-day weather forecasts while providing a realistic spectrum (sharp prediction) for any lead time. Moreover, Neural GCM models are capable of running for extended timescales, while simulating emergent phenomena like seasonal cycles, monsoon circulation, and tropical cyclone formation. We showcase the potential of hybrid models in decadal simulations and use Neural GCMs in an AMIP-like configuration to simulate realistic historical warming trends.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2023
- Bibcode:
- 2023AGUFMNG31A..06H