Artificial Intelligence falls short in predicting weather extremes

As climate change intensifies heatwaves, extreme rainfall and violent storms, the ability to forecast exceptional weather events is becoming increasingly critical. A study conducted by Sebastian Engelke, Professor at Research Institute for Statistics and Information Science and Zhongwei Zhang, former postdoctoral researcher, in collaboration with the Karlsruhe Institute of Technology, shows that traditional weather forecasting models currently remain more reliable than artificial intelligence–based models for anticipating extreme phenomena.

While AI models perform very well under typical weather conditions, they struggle to predict the intensity and frequency of exceptional events. This limitation stems from their difficulty in generalizing beyond the data on which they were trained. “AI models tend to be confined to extreme values already observed in the past, as if they were subject to an implicit ceiling,” explains Zhongwei Zhang. By contrast, traditional numerical models, based on the laws of atmospheric physics, can theoretically represent unprecedented situations.

Sebastian Engelke notes that “the recent introduction of artificial intelligence offers significant opportunities to reduce the costs and energy consumption of weather forecasting, but it does not yet allow us to dispense with physics-based models when it comes to extreme events.”

Published in Science Advances, these findings highlight the need for continued evaluation and improvement of AI-based models before they can be used autonomously in early warning systems and disaster management.

> Read the article published in Science Advances: Physics-based models outperform AI weather forecasts of record-breaking extremes

> Read the University of Geneva press release

2026