Can AI weather models predict out-of-distribution gray swan tropical cyclones?

Area of Science:

• Artificial Intelligence
• Meteorology
• Climate Science

Background:

• Predicting rare, high-impact weather extremes, termed gray swans, is a critical challenge for AI weather models.
• A key question is whether AI can extrapolate from observed weather events to predict unseen extreme events.

Purpose of the Study:

• To investigate the ability of the AI weather model FourCastNet to predict extreme weather events (Category 5 tropical cyclones).
• To determine if FourCastNet can extrapolate from weaker events to forecast stronger, unobserved extreme events.

Main Methods:

• Trained FourCastNet on ERA5 data (1979-2015), with and without Category 3-5 tropical cyclones (TCs), globally and by basin.
• Tested model performance on Category 5 TCs from 2018-2023, assessing forecasting accuracy.

Main Results:

• Models trained without extreme TCs could not accurately forecast Category 5 TCs, demonstrating a lack of extrapolation ability.
• Models trained without TCs in a specific basin showed some skill in forecasting Category 5 TCs within that basin, indicating cross-basin generalization.

Conclusions:

• Current AI weather models, like FourCastNet, cannot reliably predict gray swan events without specific training data.
• Novel AI learning strategies are required for accurate prediction of the rarest and most impactful weather extremes.