Neural General Circulation Models for Weather and Climate

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

2024-03-08

Abstract:General circulation models (GCMs) are the foundation of weather and climate prediction. GCMs are physics-based simulators which combine a numerical solver for large-scale dynamics with tuned representations for small-scale processes such as cloud formation. Recently, machine learning (ML) models trained on reanalysis data achieved comparable or better skill than GCMs for deterministic weather forecasting. However, these models have not demonstrated improved ensemble forecasts, or shown sufficient stability for long-term weather and climate simulations. Here we present the first GCM that combines a differentiable solver for atmospheric dynamics with ML components, and show that it can generate forecasts of deterministic weather, ensemble weather and climate on par with the best ML and physics-based methods. NeuralGCM is competitive with ML models for 1-10 day forecasts, and with the European Centre for Medium-Range Weather Forecasts ensemble prediction for 1-15 day forecasts. With prescribed sea surface temperature, NeuralGCM can accurately track climate metrics such as global mean temperature for multiple decades, and climate forecasts with 140 km resolution exhibit emergent phenomena such as realistic frequency and trajectories of tropical cyclones. For both weather and climate, our approach offers orders of magnitude computational savings over conventional GCMs. Our results show that end-to-end deep learning is compatible with tasks performed by conventional GCMs, and can enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system.

Atmospheric and Oceanic Physics,Machine Learning,Computational Physics

What problem does this paper attempt to address?

The paper aims to address several key issues in weather and climate forecasting:

1. **Improving Forecast Accuracy**: Traditional General Circulation Models (GCMs) perform well in weather and climate forecasting but have inherent errors and biases. Machine Learning (ML) models excel in short-term deterministic weather forecasts but have not yet shown significant advantages in ensemble forecasting and long-term climate simulations.

2. **Integrating Machine Learning with Physical Models**: The paper proposes a new hybrid model—NeuralGCM, which combines a differentiable dynamics core with machine learning components. This model can generate forecast results comparable to the best existing physical models and machine learning methods across different time scales.

3. **Enhancing Computational Efficiency**: NeuralGCM has significant advantages in computational cost compared to traditional GCMs, especially in short-term weather forecasts (1-10 days) and medium-term weather forecasts (1-15 days).

4. **Stability in Long-term Climate Simulations**: NeuralGCM can perform climate simulations spanning years to decades and accurately track climate indicators such as global average temperature. It also demonstrates phenomena like the realistic frequency and trajectories of tropical cyclones.

Through these improvements, the paper demonstrates that end-to-end deep learning can be compatible with the tasks of traditional General Circulation Models and can enhance the effectiveness of large-scale physical simulations, which is crucial for understanding and predicting the Earth system.