Building a hurricane AI in Jamaica, on one GPU.

Jamaica sits in the Atlantic hurricane belt, and hurricane forecasting is one of the places AI has already earned a role. When Hurricane Melissa made landfall in Jamaica in October 2025, the National Hurricane Center drew on Google DeepMind’s WeatherNext for its forecast of the storm’s rapid intensification.

Systems like that are built by large teams, at 0.25° resolution, on decades of data and enormous compute. NatureV1 asks a narrower question: what can one developer, one GPU and public data produce — and can the result be honest about exactly how good it is? Every part of it is open: the engine, the model, the training pipeline and the scoring.

Don’t make the model learn geometry. Most weather transformers flatten the globe into tokens and let the model work out geography from data. NatureV1 runs on iHELIX, an engine where every attention head has a physical radius — from 120 km for fronts and convection to 1,600 km for whole weather systems — measured along the curved Earth, and where the vertical axis is log-pressure. Distances, areas and the sphere are given, so the parameters go on the atmosphere.

Spend scarce labels on few parameters. The whole Atlantic best-track record is on the order of 55,000 points, and some fields are far thinner — radius of maximum wind only entered the record in 2021. So the backbone learns from about 2×10¹¹ values of free reanalysis, then freezes, and only about 1.2 million parameters ever see a hurricane label: 16,219 samples from 584 storms. Even the track is not learned from scratch. It starts from a physics-only steering track — the path the surrounding winds would carry the storm along — and learns only the correction.

A forecast must be able to say it doesn’t know. Every output carries an uncertainty. An output that was never trained is never shown, and a forecast run from an input the model was never trained on is flagged as untrustworthy.

The whole run took about one week of wall-clock time on a single RTX PRO 6000 (96 GB) in a hosted notebook. Only three to four days of that was GPU time. The rest was the cost of the platform: sessions end after 12 hours or 90 idle minutes and the disk is wiped, so every restart re-pulled checkpoints and staged data from Hugging Face before training could resume. Every stage resumes from a one-minute checkpoint, so no restart lost more than a minute of training.

NatureV1 never gives one track for a hurricane. It proposes six scenarios and puts a probability on each. With six scenarios, an even split is 16.7%.

Midway through training, a hindcast of Hurricane Dorian (2019) gave its top scenario 17% — an even split — and its track sat within 10 km of the steering track at every lead. It had not yet learned to tell one path from another. The released model gives its top scenario 26%, and its learned track was 99 km off at 24 hours, against 177 for steering.

One storm proves nothing, so the held-out seasons are what count: across 59 storms the final forecast beat steering at every lead, and the full set of scenarios turned out to be honest — the region it called 90% likely held the real storm 87–92% of the time.

• Hurricane track beats a physics-only steering track and straight-line extrapolation at every lead on the held-out seasons: 163 km at 24 hours, 421 km at 48, 739 km at 72.
• On an unseen El Niño year, 2023, it beat steering for all seven hurricanes at 24, 48 and 72 hours, and its margin over steering grew — 29% better at 72 hours, against 25% on the held-out seasons.
• Landfall probability is skilful, with an AUC of 0.81–0.86.
• Hurricane wind beats “it stays as strong as it is now” from 24 hours on.
• The full-atmosphere rollout beats climatology on all 84 field–lead pairs scored, and has skill against persistence and climatology out to 72 hours.

Before release the model went through a red team — a deliberate attempt to break it. Most of what it found is the kind of thing a model page usually leaves out:

• It is 2.5–4.5× less accurate than the NHC, which averaged about 66 km at 24 hours and 168 km at 72 over 2018–2022.
• It underestimates strong storms — by 10–18 kt for major hurricanes. It forecast about 70 kt for storms that reached 145–160 kt.
• It cannot warn of explosive intensification. Hurricane Otis (2023) went from 50 kt to a 130 kt Category 5 in 24 hours; NatureV1 forecast 45 kt.
• Its rapid-intensification probabilities are far too high. When it said 55%, it happened 38% of the time.
• Garbage in, plausible out. Random noise in place of its inputs produced a normal-looking forecast that passed every check.
• A single scenario’s cone is over-confident. Its “95%” cone held the storm 60–87% of the time; the cone has to be drawn from all six scenarios.

Every one of those is listed, with the numbers, on the NatureV1 page.

Two tests were still running at release: six storms outside the Atlantic — in the West Pacific, the Southern Hemisphere and the East Pacific — and a 10-day rollout to see whether the atmosphere stays stable past the 72 hours it was trained on. Their results will go on the model page, whatever they show.