NeuralGCM

Neural General Circulation Models for Weather and Climate

NeuralGCM is a Python library for building hybrid ML/physics atmospheric models for weather and climate simulation.

What is NeuralGCM

NeuralGCM (Neural General Circulation Models) is an innovative Python library developed to merge the strengths of machine learning (ML) with the robust framework of traditional physics-based models for weather and climate prediction. General circulation models (GCMs) have long been the cornerstone of atmospheric simulation, relying on numerical solvers to predict large-scale dynamics and finely tuned models to represent small-scale processes like cloud formation. However, while GCMs excel in long-term forecasting, recent advancements in ML have shown that data-driven approaches can achieve comparable or superior accuracy in short-term deterministic weather forecasting.

How to Install NeuralGCM

pip install neuralgcm

It is recommended to run NeuralGCM models on a computer with a GPU or a TPU for the best performance.

Why is NeuralGCM Important?

High level structure of the the NeuralGCM models. Figure 1 from the NeuralGCM paper. — Figure 1. High level structure of NeuralGCM models. (Dmitrii Kochkov et al., 2024)

The Hybrid Approach

NeuralGCM sets itself apart by combining a differentiable solver for atmospheric dynamics with advanced ML components, creating a hybrid model that excels in both weather and climate forecasting. This approach leverages the strengths of both methodologies—using ML for rapid, data-driven predictions, while relying on traditional physics to ensure stability and accuracy over longer periods.

Performance and Capabilities

NeuralGCM has demonstrated remarkable capabilities in various forecasting scenarios. For short-term weather forecasts (1-10 days), it competes with leading ML models. In ensemble forecasting, a critical component for uncertainty estimation, it performs on par with the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble prediction for up to 15 days. In climate simulations, NeuralGCM accurately tracks global mean temperature trends over multiple decades, even capturing complex phenomena like the frequency and paths of tropical cyclones with high-resolution (140 km) modeling.

Efficiency and Impact

One of the most significant advantages of NeuralGCM is its computational efficiency. By integrating ML, it achieves the same tasks as traditional GCMs with orders of magnitude less computational resources. This efficiency not only reduces the cost and time of running simulations but also opens the door to more frequent and detailed forecasts.

Looking Ahead

NeuralGCM represents a significant leap forward in the field of geospatial artificial intelligence. By harmonizing the power of ML with the foundational principles of atmospheric physics, it provides a new tool for researchers and forecasters. Whether for daily weather predictions or long-term climate analysis, NeuralGCM is set to enhance our understanding and predictive capabilities of the Earth system.

Citation

Dmitrii Kochkov, Janni Yuval, 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. https://doi.org/10.1038/s41586-024-07744-y

‌