AI weather models struggle to predict extreme events: Swiss study

Artificial intelligence-based weather forecasting models struggle to accurately predict extreme events, often underestimating their intensity and frequency, according to a report Thursday by Swissinfo.

The study led by the University of Geneva compared leading AI systems, GraphCast, Pangu-Weather and Fuxi, with the physical HRES model of the European Centre for Medium-Range Weather Forecasts.

Researchers found that AI models were “systematically wrong” when it came to record-breaking events.

Extreme cold spells were predicted as less intense than they actually were, while heat waves and strong winds were also underestimated. In addition, the report said, such events were forecast less frequently than they occurred.

“These results highlight a central challenge for the use of AI in the prediction of weather events with a large societal impact,” said the authors.

The researchers said the limitations stem from how AI systems are trained, relying on historical data and patterns. Rare or unprecedented events, which are increasingly common due to climate change, fall outside the range.

By contrast, traditional physical models simulate atmospheric behavior based on natural laws, allowing them to better capture extreme scenarios, even though not previously observed, according to the research.

The findings come as extreme weather events intensify globally, increasing the need for reliable forecasts to protect lives and infrastructure.

Despite the shortcomings, researchers said AI still offers “vast new possibilities” and called for hybrid systems combining machine learning with physical models to improve forecasting accuracy.