Google DeepMind and Google Research built an AI weather model that called Hurricane Melissa's Category 5 landfall in Jamaica with 80% confidence five days before it hit. That prediction, made in October 2025 while the storm was still forming over the Caribbean Sea, gave forecasters and emergency planners a full extra day of lead time compared with the numerical models the field has relied on for decades. A peer-reviewed paper describing the system, called WeatherNext, was published Thursday in Nature.
The headline result is a one-day gain in forecast accuracy. WeatherNext's three-days-out predictions are as accurate as the two-days-out predictions of the previous generation of models. Historically, moving forecast skill forward by a day has taken roughly a decade of research. Mike Brennan, director of the US National Hurricane Center, said the shift matters at the operational level, where evacuations, supply staging, and resource pre-positioning are all decisions bound by the clock.
The technical problem WeatherNext had to solve is that cyclones are rare. Machine learning models need volume, and there simply aren't enough hurricanes in the historical record to train a system from scratch on cyclone data alone. Google DeepMind's team got around this by training the model on general weather data first, then tuning it to perform on cyclones — treating hurricane prediction as a specialization of a broader weather task.
Key facts
- 01WeatherNext predicted Hurricane Melissa's Category 5 landfall in Jamaica with 80% confidence five days before impact in October 2025.
- 02The model's three-day forecasts match the accuracy of prior models at two days out — one extra day of lead time on average.
- 03WeatherNext now generates 1,000 storm scenarios per cyclone, up from 50 last year, capturing a wider range of trajectories.
- 04Google DeepMind is open-sourcing the WeatherNext models used during hurricane season for outside researchers to build on.
- 05Melissa marked the first time the National Hurricane Center called a Category 5 hurricane while the storm was still at Category 1.
Hurricanes are also hard because they operate at two spatial scales at once. Predicting a storm's track requires global-scale inputs like cold fronts and prevailing winds. Predicting intensity requires fine-grained local data on ocean temperature and atmospheric conditions. Kate Musgrave of the Cooperative Institute for Research in the Atmosphere, an author on the paper, said earlier AI models handled track reasonably well but consistently failed on intensity — a gap that matters because a Category 1 and a Category 5 are different events entirely.
Melissa was the field test. Before WeatherNext was used in live forecasts, researchers ran it on retrospective data and found it outperformed the incumbent models by wide margins. When forecasters plugged it into real-time operations during hurricane season, the performance held. Melissa marked the first time the National Hurricane Center was able to identify a storm as a Category 5 threat while it was still classified as a Category 1.
“The results were so good that we were skeptical that we would actually see that in the real-time demonstration.”— Kate Musgrave, Tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere
Even the researchers who built WeatherNext can't fully explain why it works as well as it does. The model uses lower-resolution atmospheric data than traditional numerical simulations require, yet still produces sharper intensity forecasts. Ferran Alet, a research scientist at Google DeepMind and a lead author on the paper, said the community was surprised that coarse-resolution inputs contain more signal than physicists had assumed. He framed the finding as a hint that something in the atmospheric dynamics is not yet understood, not as a black box to be dismissed.
WeatherNext also doesn't produce a single forecast. It generates a distribution of possible storm evolutions, capturing the sensitivity of cyclone paths to small perturbations. Last year the model produced 50 scenarios per storm; the current version produces 1,000. Musgrave said numerical models cannot run that many scenarios with existing compute budgets, so the ensemble approach is a genuine capability difference rather than a marginal improvement.
Brennan was careful to frame WeatherNext as one tool among many. No single model wins every season or every storm, and forecasters still cross-reference outputs from multiple systems. He added that the ultimate value of any forecast is in impact translation — turning a track and an intensity into evacuation orders, shelter openings, and infrastructure prep — and that remains a human judgment call.
Google DeepMind is open-sourcing the WeatherNext models used during hurricane season, releasing the weights for outside researchers to test, extend, and probe. Alet said he expects the research community to uncover additional insights into cyclone dynamics by dissecting the model, particularly the question of what signal the low-resolution inputs actually carry.
The broader lesson from WeatherNext is that AI models are starting to outperform physics-based simulations on problems the physics community treated as compute-bound. Weather forecasting has been one of the flagship applications of supercomputing since the 1950s, and the numerical models running at national weather centers cost tens of millions per year to operate. If a learned model trained on the same historical archives can match or beat those systems while running on far less hardware — and generate 20x more ensemble scenarios in the process — the economics of operational forecasting shift. That reframes AI's role in the physical sciences from a novelty to a serious contender for the core simulation stack, with implications well beyond hurricanes.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
