🌎 AI fails in the face of extreme weather phenomena

Record heat waves, torrential rains, supercell thunderstorms: extreme events are intensifying under the effect of climate change, with heavy human and economic consequences. Artificial intelligence models are revolutionizing weather forecasts. But can they anticipate these exceptional episodes?

A team from the University of Geneva (UNIGE) and the Karlsruhe Institute of Technology shows that traditional numerical models remain more reliable for predicting extreme phenomena, even though AI models outperform them in common situations. These results are published in Science Advances.


To predict the weather for days or weeks ahead, meteorologists rely on simulations generated by complex mathematical models. Fueled by large amounts of data - collected by weather stations, satellites or aircraft - they apply the laws of physics to this information to simulate the future state of the atmosphere.

The European Centre for Medium-Range Weather Forecasts, for example, uses a model called High RESolution forecast or "HRES". It provides simulations to 35 European countries based on this model.

While this method is reliable and robust, it is also expensive and energy-intensive: it requires a large fleet of supercomputers capable of solving millions of equations several times a day. "The introduction, about three years ago, of the first models based on artificial intelligence, alongside the traditional numerical approach, paved the way for simplifying processes and reducing their costs," explains Sebastian Engelke, professor at the Research Institute for Statistics and Information Science of the Geneva School of Economics (GSEM) of UNIGE.

But is this AI-based approach capable of predicting the occurrence, up to ten days in advance, of extreme events that are often unprecedented? In a recent study, Sebastian Engelke's team shows that AI outperforms traditional models - in this case HRES - in predicting common situations, but that it makes larger errors than HRES in 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 spans from 1979 to 2017. They thus tend to limit themselves to extreme values already observed in the past, as if they were hitting an implicit ceiling. Conversely, conventional models, based on atmospheric physics, are not constrained by this limit and can theoretically represent unprecedented situations," explains Zhongwei Zhang, a 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 weather models based on artificial intelligence when it comes to extrapolating beyond their training domain and forecasting record-breaking weather events. They underline the need to continue their evaluation and improvement before they can be used autonomously in early warning and disaster management systems.