Overview
- The Science Advances study published Thursday found leading AI systems — GraphCast, Pangu-Weather and Fuxi — underpredicted the intensity and number of extreme events compared with Europe’s physics-based HRES model from ECMWF.
- Researchers trained the AI on weather from 1979 to 2017 and checked forecasts against real extremes in 2018 to 2020, where errors grew as events became rarer and stronger.
- The shortfall stems from training data with few record heat waves, cold snaps or severe storms, while numerical models encode physical laws that better handle rare conditions.
- Germany’s weather service DWD began running the AI model AICON for three-hour updates in early March, yet experts caution current AI is not ready to anchor severe-weather alerts.
- Scientists point to probabilistic ensembles that produce many plausible forecasts, hybrid designs that build in physics, and synthetic extreme-event data as paths to improve performance as climate change drives harsher weather.