Record-breaking heatwaves, torrential rainfall and supercell thunderstorms: extreme events are intensifying under the influence of climate change, with major human and economic consequences. Artificial intelligence models are revolutionizing weather forecasting. But can they anticipate such exceptional events? A team from the University of Geneva (UNIGE) and the Karlsruhe Institute of Technology (KIT) shows that, to date, traditional numerical models remain more reliable for predicting extreme phenomena, even though AI models outperform them under typical conditions. These findings are published in Science Advances.

To forecast the weather in the coming days or weeks, meteorologists rely on simulations generated by complex mathematical models. Powered by vast amounts of data—collected from weather stations, satellites, and aircraft—these models apply the laws of physics to simulate the future state of the atmosphere. The European Centre for Medium-Range Weather Forecasts, for instance, uses a model known as the High Resolution Forecast, or ‘‘HRES’’, to provide simulations to 35 countries across the continent.

While this method is reliable and robust, it is also costly and energy-intensive, as it requires extensive supercomputing infrastructure capable of solving millions of equations several times a day. ‘‘The introduction, three years ago, of the first models based on artificial intelligence, alongside the traditional numerical approach, has opened the way to simplifying processes and reducing their costs,’’ explains Sebastian Engelke, full professor at the Research Institute for Statistics and Information Science at the UNIGE Geneva School of Economics and Management (GSEM).

But is this AI-based approach capable of predicting the occurrence of often unprecedented extreme events up to ten days in advance? In a recent study, Sebastian Engelke'steam shows that AI outperforms traditional models—specifically HRES—when forecasting typical conditions, but consistently makes larger errors than HRES when predicting the intensity and frequency of extreme temperatures and winds.

‘‘The main problem with AI models is their difficulty in generalizing beyond the data on which they were trained, which span the period from 1979 to 2017. As a result, they tend to be limited to extreme values already observed in the past, as if they had an implicit ceiling. By contrast, convential models, based on the laws of atmospheric physics, are not constrained by this limitation and can theoretically represent unprecedented situations,” explains Zhongwei Zhang, former postdoctoral researcher in Sebastian Engelke’s team, now affiliated with the Institute of Statistics at the Karlsruhe Institute of Technology, and first author of the study.

These results highlight the current limitations of AI-based weather models when it comes to extrapolating beyond their training domain and predicting record-breaking weather events. They underscore the need for continued evaluation and improvement before such models can be used autonomously in early warning systems and disaster management.