“Neural General Circulation Models for Weather and Climate,” by Dmitrii Kochkov, Janni Yuval, Stephan Hoyer and colleagues at Google with the European Centre for Medium-Range Weather Forecasts, was posted to arXiv in November 2023 and published in Nature in 2024. It introduced NeuralGCM, a hybrid approach that differs from purely data-driven models like GraphCast.
NeuralGCM keeps a traditional differentiable solver for the large-scale fluid dynamics of the atmosphere and replaces only the hard-to-model small-scale physics (such as clouds and precipitation) with learned neural components, training the whole system end to end. The paper reports that this hybrid is competitive with leading machine-learning models for short-range forecasts, matches physics-based ensemble methods for one-to-fifteen-day predictions, and, unlike many pure ML models, remains stable enough to run multi-decade climate simulations.
The significance is that weather forecasting and climate projection have different demands: forecasts need accuracy over days, while climate runs must stay physically consistent over decades. By blending physics with learning, NeuralGCM aimed to serve both, suggesting that the future of Earth-system modeling may be hybrid rather than a wholesale replacement of physics with neural networks.