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

Artificial intelligence (AI) 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 present an event-based diagnostic evaluation of Aurora, a deterministic AI model, across 16 case studies chosen for physical diversity rather than statistical representativeness, spanning tropical cyclones (TCs), freezes, heatwaves, atmospheric rivers (ARs), and extreme precipitation at lead times from 1 to 21 days. Aurora showed strong short-range (1–7 day) skill: TC track and landfall positions were accurate for well-behaved systems, temperature extremes achieved high spatial agreement, and the atmospheric river structure was reproduced faithfully. This study’s central finding is a pattern–amplitude divergence: beyond 7 days, large-scale circulation patterns remained moderately skillful even as surface amplitudes weakened toward climatological values. Event-specific failures include a severe recurvature forecast failure for Hinnamnor, TC intensity biases, and a pronounced in-sample versus out-of-sample precipitation skill gap that is substantially confounded by event-type differences (large-scale monsoon vs. mesoscale-convective/cutoff-low regimes), so the gap cannot be attributed to training-period recency alone. Across the events examined here, Aurora provides reliable deterministic guidance within 7 days. We recommend deploying Aurora as a rapid ensemble generation and regime-identification tool alongside physics-based numerical weather prediction at short-to-medium range, with its directional intensity and amplitude biases corrected through post-processing before standalone use in operational warnings.