DeepMind’s WeatherNext Gives Hurricane Forecasters Extra Day
A Nature paper published Thursday shows Google DeepMind's WeatherNext AI model predicts tropical cyclone tracks and intensity a full day earlier than existing models, a gain that historically took a decade of conventional research to achieve.
👁 116 reads
Google DeepMind’s WeatherNext model can predict tropical cyclone intensity and track a full day earlier than existing forecasting tools, according to a paper published Thursday in Nature - a lead-time gain the researchers say would historically have required a decade of incremental work.
- WeatherNext predicted Hurricane Melissa would hit Jamaica as a Category 5 storm five days before landfall, with 80 percent confidence - the first time the National Hurricane Center forecast a Cat 5 from a Cat 1 stage
- On average the model gives forecasters one extra day of lead time, meaning its three-day forecasts match the accuracy of previous models' two-day forecasts
- The model was trained on broad weather data alongside scarce cyclone data, letting it handle both large-scale track prediction and small-scale intensity prediction that earlier AI models could not manage
The practical stakes became clear in October 2025, when a storm system formed over the Caribbean and weather models disagreed on its path. WeatherNext predicted with 80 percent confidence, five days before landfall, that the system would strike Jamaica as a Category 5 hurricane. It did. Hurricane Melissa caused widespread flooding and landslides, but the early warning gave communities extra time to prepare. It was also the first occasion on which the US National Hurricane Center successfully forecast a Category 5 storm while it was still at Category 1 intensity.
Mike Brennan, director of the National Hurricane Center, put the operational value plainly:
“Time is really golden when it comes to those types of decisions, so the ability to push forecast accuracy out as much as a day beyond what we’ve previously been able to do is really valuable.”
The technical challenge WeatherNext addresses is a long-standing split in hurricane modeling. Predicting a storm’s track requires global-scale atmospheric data - cold fronts, prevailing winds. Predicting its intensity demands fine-grained local data on ocean and atmospheric conditions at the storm itself. Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and a paper co-author, noted that earlier AI models handled track reasonably well but that “intensity they could not do well at all.” WeatherNext was trained simultaneously on broad weather data and the comparatively sparse cyclone record, giving it leverage on both problems.
Ferran Alet, a research scientist at Google DeepMind and a lead author, described the approach: “We don’t have that much cyclone data, but we have a lot of weather data. So what we did was train a model to be both good at weather as well as cyclones.” The results were strong enough that the team was initially skeptical they would hold in live forecasts - and then they did.
Practitioner note: WeatherNext is a reminder that the highest-value AI gains often come from data strategy rather than raw model size. Training on abundant adjacent data - general weather - to compensate for a scarce target domain - cyclones - is a transferable design choice. Developers building specialized prediction tools in any data-sparse domain can apply the same logic: identify a richer related dataset, train jointly, and test on retrospective cases before live deployment.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →
More on this from QUANTSLANT
news·OpenAI Reveals Timeline of AI Agents’ Accidental Hugging Face Attack