Neural general circulation models for weather and climate.

Area of Science:

• Earth System Science
• Climate Modeling
• Machine Learning Applications

Background:

• General Circulation Models (GCMs) are crucial for weather and climate prediction, relying on physics-based simulations.
• Recent advancements show machine-learning models matching or exceeding GCMs in deterministic weather forecasting.
• Existing machine-learning models lack stability for long-term climate simulations and haven't improved ensemble forecasts.

Purpose of the Study:

• To develop a novel General Circulation Model (GCM) that merges differentiable physics-based dynamics with machine learning.
• To evaluate the GCM's performance in deterministic weather forecasting, ensemble forecasting, and long-term climate simulations.
• To assess the computational efficiency and stability of the new hybrid GCM.

Main Methods:

• Developed a hybrid GCM, termed NeuralGCM, combining a differentiable atmospheric dynamics solver with machine-learning components.
• Trained and evaluated NeuralGCM on reanalysis data for weather forecasting (1-15 days).
• Assessed NeuralGCM's climate simulation capabilities over multiple decades with prescribed sea surface temperatures at 140-km resolution.

Main Results:

• NeuralGCM demonstrates competitive skill with state-of-the-art machine-learning models for short-term weather forecasts (1-10 days).
• It matches the European Centre for Medium-Range Weather Forecasts ensemble prediction for medium-term forecasts (1-15 days).
• NeuralGCM accurately tracks decadal climate metrics and simulates emergent phenomena like tropical cyclones, offering significant computational savings.

Conclusions:

• End-to-end deep learning is compatible with and can enhance traditional GCM tasks for Earth system science.
• The hybrid approach provides a computationally efficient alternative for weather and climate prediction.
• NeuralGCM shows promise for large-scale physical simulations essential for understanding Earth's climate system.