Study of 16 extreme-weather events finds AI model Aurora reliable for 7 days, then weak

Beyond a week the model kept the large-scale circulation pattern but lost the strength of the extreme. In the Texas freeze the overlap dropped to 0.002 at 14 days even though pattern correlation was still 0.728. The authors say the practical horizon for useful extreme-event guidance is about 7–10 days.

The authors recommend using Aurora as a fast ensemble-generation and regime-identification tool next to physics-based forecasts. They say its intensity and amplitude biases should be corrected through post-processing before it is used alone in operational warnings. Events were chosen for physical variety, not statistical representativeness, so the results are diagnostic rather than a ranking against operational systems.

The team ran the pretrained 0.25° Aurora model with no fine-tuning from ERA5 states, across 16 events from 1 to 21 days ahead. Four were cyclones (Sandy, Amphan, Ian, Hinnamnor), two freezes, two heatwaves, two atmospheric rivers and four floods. Forecasts were checked against ERA5 reanalysis and best-track records, and precipitation came from an auxiliary decoder verified against the same dataset it was trained on, so that check is not fully independent.

In the 1–7 day range the results were solid. Freeze overlap scores ran from 0.83 to 0.98 with pattern correlations above 0.90, Hurricane Sandy's 3-day mean track error was 77.3 km, and Aurora reproduced blocking ridges and jet displacement for the freeze events. Short-lead biases were small enough that the authors think post-processing could correct them.

At 14–21 days the pattern emerged that the authors call a pattern–amplitude divergence. Forecasts drifted toward climatology, which made RMSE look deceptively low (2.82 °C for Beast from the East at 14 days) while threshold overlap collapsed. For the Texas freeze, spatial extent at 14 days was 0.2%. The authors say this matches intrinsic predictability limits of midlatitude weather, not only a model flaw.

Training period mattered less clearly than hoped. Precipitation skill was higher for events inside Aurora's 1979–2020 training period, but the authors say the gap is substantially confounded by event type (large-scale monsoon versus mesoscale convective and cutoff-low regimes), so it cannot be blamed on training recency alone.