Artificial intelligence-based weather forecasting models remain less reliable than traditional numerical systems when predicting extreme weather events, according to new research from the University of Geneva and the Karlsruhe Institute of Technology.
The study, published in Science Advances, found that while AI models can outperform conventional approaches under typical conditions, they struggle to accurately predict the intensity and frequency of extreme weather events such as heatwaves, heavy rainfall and strong winds.
Meteorological forecasting has long relied on numerical models that simulate atmospheric conditions using physical laws and large datasets collected from satellites, weather stations and aircraft. Systems such as the High Resolution Forecast (HRES), used by the European Centre for Medium-Range Weather Forecasts (ECMWF), generate forecasts for multiple countries but require significant computing power.
Recent advances in AI have introduced alternative approaches that reduce computational costs and processing time. However, the research highlights key limitations when these models are applied to rare or unprecedented events.
“The introduction, three years ago, of the first models based on artificial intelligence has opened the way to simplifying processes and reducing their costs,” said Sebastian Engelke, full professor at the University of Geneva. “But the main problem with AI models is their difficulty in generalizing beyond the data on which they were trained.”
The study found that AI models consistently produced larger errors than traditional systems when forecasting extreme temperatures and wind conditions up to 10 days in advance.
Zhongwei Zhang, postdoctoral researcher at the Karlsruhe Institute of Technology and first author of the study, said this limitation stems from the historical datasets used to train AI systems. “They tend to be limited to extreme values already observed in the past, as if they had an implicit ceiling,” he explained. “By contrast, conventional models, based on the laws of atmospheric physics, are not constrained by this limitation and can theoretically represent unprecedented situations.”
The findings come as extreme weather events become more frequent and severe due to climate change, increasing the importance of accurate forecasting for early warning systems and disaster preparedness.
Researchers said the results underline the need for further development and evaluation of AI-based forecasting tools before they can be relied on independently in operational environments.
Related news, AI model improves real-time prediction of wildfire spread