The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2604
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T08:01:30.397084Z
status: published
lens: trace
sector: climate
headline: Evaluating the Predictability of Selected Weather Extremes with Aurora, an AI Weather Forecast Model
dek: 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 lea…
gain_title: In a study of 16 extreme-weather cases, Microsoft's Aurora model matched the Texas 2021 freeze footprint at 1 day (overlap score 0.952) and tracked Hurricane Ian with a 13.5 km mean error.
problem_title: In the same study, Aurora's Texas 2021 freeze overlap fell to 0.002 at 14 days, it badly missed Typhoon Hinnamnor's recurvature, and its storm intensity ran off by up to ~10 hPa.
trace_subject: Aurora AI model forecasts of high-impact weather extremes by lead time
gain_reading: In a study of 16 extreme-weather cases, Microsoft's Aurora model matched the Texas 2021 freeze footprint at 1 day (overlap score 0.952) and tracked Hurricane Ian with a 13.5 km mean error.
gain_evidence: Aurora showed strong short-range (1–7 day) skill
problem_reading: In the same study, Aurora's Texas 2021 freeze overlap fell to 0.002 at 14 days, it badly missed Typhoon Hinnamnor's recurvature, and its storm intensity ran off by up to ~10 hPa.
problem_evidence: Event-specific failures include a severe recurvature forecast failure for Hinnamnor, TC intensity biases
quick_read: Researchers at Arizona State and Seoul National universities tested Microsoft's Aurora forecasting model on 16 hand-picked extreme events: tropical cyclones, freezes, heatwaves, atmospheric rivers and floods. Up to about 7 days ahead it performed well. For the Texas 2021 freeze, the overlap between forecast and observed sub-zero area was 0.952 at 1 day and 0.831 at 7 days, and Hurricane Ian's track was off by 13.5 km at 1 day.

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.
limitation: The conclusions rest on a small, hand-picked set of events (four cyclone cases), verified against reanalysis and best-track data, and Aurora was not compared directly with operational forecast models such as IFS or GFS.
tag: Dual reading
key_points: Tropical cyclone track errors for Hurricane Ian were 13.5 km at 1 day and 35.3 km at 3 days; Cyclone Amphan's were 22.6 km and 30.6 km. | Typhoon Hinnamnor, which recurved into the midlatitude westerlies, was a clear failure: Aurora's tracker failed at 1 day and 7-day mean track error was 170.7 km. | Storm intensity errors reached up to ~10 hPa or several m/s in some runs, which the authors call large enough to matter operationally. | For the 2018 Beast from the East freeze, 14-day overlap fell to 0.157 and the forecast ran 6.88 °C too warm, while pattern correlation stayed at 0.875. | Heatwaves showed a cold bias instead: the British Columbia 2021 heat dome was 10.63 °C too cool at 14 days, with overlap at 0.024. | For the California 2022 atmospheric river, local pattern correlation was 0.95 at 1 day but 0.50 at 7 days, with the regional moisture-transport bias at −314.94 kg/m/s. | Precipitation skill was weakest for post-2020 events: 1-day pattern correlation was 0.39 for Western Europe 2021 and 0.03 for Appalachia 2022, versus above 0.53 for Pakistan 2010 and Sudan 2020. | Forecasts for the four cyclone cases took about 6 minutes of GPU time on a single Nvidia A100.
rundown: 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.
sources:
- peer_reviewed | Atmosphere | https://doi.org/10.3390/atmos17080716 | 2026-07-23
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

a7dab4d5a461312c7b9e694a24b7c5e49c798cacf29883b7c5411ea3ec626b27

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000