Extrapolation of AI weather models to stronger out-of-distribution extremes

Determine whether artificial intelligence weather forecasting models trained on reanalysis data (for example, FourCastNet trained on ERA5, 1979–2015) can extrapolate from weaker in-distribution weather events to forecast stronger out-of-distribution extreme events, specifically Category 5 tropical cyclones characterized by mean sea-level pressure below 970 hPa.

Background

The paper investigates whether state-of-the-art AI weather models can forecast gray swan events—physically possible but historically unobserved extremes—by conducting controlled experiments with FourCastNet. Training datasets are constructed with stronger tropical cyclones removed to test out-of-distribution generalization. The authors’ overarching motivation is to assess if models trained on weaker events can learn physical relationships that extend to stronger, unseen extremes.

This question is central to the utility of AI weather/climate emulators for early warning and statistical estimation of rare, high-impact events. While the paper presents evidence that FourCastNet cannot extrapolate to Category 5 tropical cyclones when such events are absent from training, the broader question stated in the abstract remains general to AI models and motivates further research across architectures and event types.