Neural General Circulation Models for Weather and Climate
1. Observation and Context
Traditional physics-based computer simulations—called General Circulation Models (GCMs)—are the main tool used to predict global weather and long-term climate. GCMs use math equations to model large-scale atmospheric movements, but they rely on hand-tuned formulas for small-scale events like clouds and rain. This makes them computationally slow and prone to bias. Recently, pure machine learning (AI) models trained on historical weather data have outperformed GCMs at short-term, single-scenario weather forecasts at a fraction of the computational cost. However, pure AI models suffer from major flaws: their predictions become blurry and physically unrealizable over longer periods, they cannot accurately predict weather uncertainties using ensemble forecasts, and they struggle to simulate multi-decade climate trends without breaking down or drifting off track.
2. Hypothesis
Combining a physics-based mathematical solver for large-scale air movement with AI neural networks to handle complex, small-scale physics into a single, fully differentiable hybrid model will allow for “online” end-to-end training. This hybrid “NeuralGCM” model will generate precise short-term weather forecasts, realistic multi-scenario ensemble predictions, and stable, decades-long climate simulations while running thousands of times faster than traditional physics models.
3. Experiment and Methodology
The researchers created NeuralGCM, which merges two main parts:
- Differentiable Dynamical Core: A physics engine built using Python (JAX) that directly solves the mathematical equations governing large-scale fluid motion, thermodynamics, gravity, and Earth’s rotation.
- Learned Physics Module: A neural network that calculates how unresolved local processes (such as cloud formation, radiation, and precipitation) alter single vertical columns of the atmosphere.
Because the physics core is differentiable, the AI module was trained “online”—learning directly while interacting with the physics equations over simulated forecast rollouts gradually expanding from 6 hours up to 5 days. The team built both deterministic (single forecast) and stochastic (randomness-injected for ensemble forecasting) versions of NeuralGCM at varying grid resolutions (2.8∘, 1.4∘, and 0.7∘). They evaluated NeuralGCM using historical ERA5 weather reanalysis data against top pure-AI models (GraphCast and Pangu-Weather) and operational physics models (ECMWF).
4. Results and Data
- Deterministic Weather Forecasting: NeuralGCM-0.7° matched or exceeded the accuracy of top pure-AI models (like GraphCast) and traditional physical models for 1-to-10-day global forecasts. Crucially, unlike pure-AI models, NeuralGCM kept forecasts sharp without losing physical balance (such as geostrophic wind balances).
- Ensemble Forecasts: NeuralGCM-ENS (the multi-scenario stochastic version) achieved higher accuracy scores (better CRPS) than ECMWF’s gold-standard physics-based ensemble system across 1-to-15-day lead times.
- Climate Simulations: In multi-decade runs, NeuralGCM remained stable without crashing or drifting. It successfully tracked global warming temperature trends over 40-year tests, captured seasonal cycles like monsoons, and naturally generated realistic tropical cyclone (hurricane) counts and paths.
- Speed and Efficiency: NeuralGCM ran 3 to 5 orders of magnitude faster than conventional high-resolution GCMs. A 1.4∘ NeuralGCM model simulated 70,000 days of weather in just 24 hours on a single TPU processor.
5. Conclusion and Climate Impact
NeuralGCM demonstrates that combining physical law solvers with machine learning creates a flexible, highly efficient tool for atmospheric science. By drastically lowering computation costs while maintaining high physical accuracy and long-term stability, NeuralGCM makes massive multi-scenario weather forecasting and rapid climate assessment accessible, improving our ability to predict severe weather disasters and understand global climate change.
6. Full Citation
Kochkov, D., Yuval, J., Langmore, I., Norgaard, P., Smith, J., Mooers, G., Klöwer, M., Lottes, J., Rasp, S., Düben, P., Hatfield, S., Battaglia, P., Sanchez-Gonzalez, A., Willson, M., Brenner, M. P., & Hoyer, S. (2024). Neural general circulation models for weather and climate. Nature, 632(8027), 1060–1066. https://doi.org/10.1038/s41586-024-07744-y