Witze (2025) This AI model ‘studied’ physics — and learnt to forecast extreme weather

Identification

• Journal: Nature
• Year: 2025
• Date: 2025-12-11
• Authors: Alexandra Witze
• DOI: 10.1038/d41586-025-04055-8

Research Groups

• Boston University, Massachusetts (Jacob Landsberg)
• University of Chicago, Illinois (Pedram Hassanzadeh, Qiang Sun, Alexander Wikner)
• Google DeepMind, London (GraphCast model development)
• NVIDIA, Santa Clara, California (FourCastNet model development)
• Huawei Cloud, Shenzhen (Pangu-Weather model development)
• Laboratory of Dynamic Meteorology, Paris (Amaury Lancelin)

Short Summary

This paper explores a novel hybrid approach combining artificial intelligence (AI) models with conventional physics-based climate models and mathematical tools to forecast extreme weather events more effectively. This method demonstrates the ability to predict extreme heatwaves as accurately as traditional models but significantly faster, addressing AI's limitations with unprecedented events.

Objective

• To develop and test a hybrid AI-physics modeling approach that can accurately and rapidly forecast extreme weather events, particularly those not encountered in historical training data, by integrating conventional climate models and mathematical tools for rare event statistics.

Study Configuration

• Spatial Scale: Global (e.g., tropical cyclones across ocean basins, mid-latitude heatwaves in Chicago and France, extreme rainfall in Dubai, deep freeze in Texas, heatwave in Moscow).
• Temporal Scale: Short-term forecasting (e.g., 8 days in advance for extreme rainfall), and potential for long-term climate change scenario analysis for events occurring once every 1,000 years.

Methodology and Data

• Models used:
  • GraphCast (Google DeepMind)
  • FourCastNet (NVIDIA)
  • Pangu-Weather (Huawei Cloud)
  • Conventional physics-based global climate models
  • Mathematical methods for analyzing the statistics of rare events
  • Hybrid approach: AI models combined with physics-based climate models and rare event statistics.
• Data sources:
  • Past observations (used for AI model training data, typically 40 years)
  • Climate model simulations

Main Results

• The hybrid AI-physics approach successfully simulated the probabilities of extreme heatwaves as accurately as older, non-AI methods, but with significantly faster computational speed.
• The GraphCast model accurately forecast an extreme rainfall event in Dubai eight days before it occurred.
• While FourCastNet struggled to forecast the strongest tropical cyclones without extreme storms in its training set, it demonstrated the ability to learn from storms in one ocean basin and apply that knowledge to others.
• The Pangu-Weather model, when combined with a physics-based global climate model and rare event statistics, effectively predicted mid-latitude heatwave probabilities faster than the climate model alone.
• The AI component of the hybrid model helps identify scenarios most likely to lead to extreme weather, guiding the conventional climate model to focus its simulations, thereby accelerating the forecasting process.

Contributions

• Introduces a pioneering hybrid modeling framework that merges AI capabilities with established physics-based climate models and rare event statistics, overcoming a key limitation of AI in forecasting unprecedented extreme weather.
• Demonstrates a significant improvement in the speed of extreme weather forecasting (specifically for heatwaves) while maintaining accuracy comparable to conventional, slower methods.
• Provides a proof-of-concept for a method that can potentially improve the reliability of long-term climate change scenario predictions for extreme events, which have historically been highly uncertain.
• Highlights the potential for AI models to learn transferable patterns across different geographical regions (e.g., ocean basins) to enhance regional forecasts.

Funding

Not explicitly mentioned in the provided news article.

Citation

@article{Witze2025This,
  author = {Witze, Alexandra},
  title = {This AI model ‘studied’ physics — and learnt to forecast extreme weather},
  journal = {Nature},
  year = {2025},
  doi = {10.1038/d41586-025-04055-8},
  url = {https://doi.org/10.1038/d41586-025-04055-8}
}