ABSTRACT:

Gray swan weather extremes, which are physically possible but so rare that they are absent from the training dataset, are a major concern for AI weather models, especially as climate change introduces unprecedented conditions. An important open question is whether and how these models can forecast such out-of-distribution events. Here, we combine controlled training experiments using FourCastNet with forecast diagnostics of state-of-the-art AI models (GraphCast, AIFS) to identify the learning mechanism behind successful gray swan forecasts. We train independent versions of FourCastNet on the ERA5 dataset with Category 3-5 tropical cyclones removed, either globally or only over one ocean basin. The version trained without any strong cyclones cannot accurately forecast Category 5 storms, indicating that the model cannot extrapolate from weaker events. However, the version trained with strong cyclones removed from only one basin retains skill in that basin, suggesting that the model can learn from similar events in other regions, a mechanism we term "translocation" as opposed to "extrapolation". This mechanism explains the surprising skill of GraphCast and AIFS in forecasting Dubai's unprecedented April 2024 rainfall, which had nearly twice the training set's highest rainfall in that region, up to 8 days ahead. Receptive field analyses show that these models use highly non-local information, and fine-tuning GraphCast with precipitation of dynamically similar events capped outside Dubai substantially degrades its forecast for this event, providing causal evidence for translocation. Evidence of extrapolation is not found: even in-distribution tail events are underestimated, due to both data imbalance and spectral bias. These findings demonstrate the potential of AI models to forecast regional gray swans and highlight the innovations needed for globally unprecedented extremes.

Reference:

1. Sun, Y.Q., Hassanzadeh, P., Zand, M., Chattopadhyay, A., Weare, J. and Abbot, D.S., 2025. Can AI weather models predict out-of-distribution gray swan tropical cyclones?. Proceedings of the National Academy of Sciences, 122(21), p.e2420914122.

2. Sun, Y.Q., Hassanzadeh, P., Shaw, T. and Pahlavan, H.A., 2026. Predicting beyond training data via extrapolation versus translocation: AI weather models and Dubai's unprecedented 2024 rainfall. arXiv preprint arXiv:2505.10241.