Evaluating the Predictability of Selected Weather Extremes with Aurora, an AI Weather Forecast Model

Artificial intelligence weather models achieve forecast skill comparable to numerical weather prediction at far lower computational cost, yet their reliability for high-impact extremes remains largely uncharacterized. We evaluate Aurora, a state-of-the-art deterministic AI model, using an event-based framework spanning tropical cyclones, freezes, heatwaves, atmospheric rivers, and extreme precipitation at lead times from 1 to 21 days. Aurora demonstrates strong short-range (1–7 day) skill: mean tropical cyclone track errors of 20–60 km at 1–3 day leads, high spatial agreement for temperature extremes (IoU ≥ 0.78), and accurate atmospheric river structure reproduction. Beyond 7–10 days, amplitude collapses as surface fields regress toward climatology, consistent with theoretical Lorenz predictability limits, while large-scale circulation patterns remain moderately skillful (pattern correlations 0.57–0.85 at 14–21 days for temperature extremes). This pattern–amplitude divergence, where synoptic-scale structure persists but threshold-based extremes collapse, is the central finding; event-specific failures include catastrophic TC recurvature errors, systematic intensity underestimation, and pronounced in-sample versus out-of-sample precipitation skill degradation. Aurora provides reliable deterministic guidance within 7–10 days, positioning it as a computational anchor for hybrid probabilistic forecasting systems rather than a standalone operational replacement.