An AI Cyclone Model Delivers a Day More Forecasting Skill, on Average

A new artificial-intelligence weather model has matched the accuracy of leading cyclone forecasts roughly one day earlier, while predicting a storm's path, strength and wind footprint in one system. The result could give professional forecasters more useful time to judge dangerous possibilities, but it does not promise every community an extra 24 hours of warning.

KEY NUMBER
Across track, intensity and wind-radius forecasts, the model gained an average of at least 24 hours of useful lead time over leading operational baselines.

Why one additional day can matter

Tropical-cyclone decisions are not made from a single track line. Emergency teams must weigh where destructive winds could reach, how quickly a storm could intensify and how uncertain the forecast remains. An earlier dependable signal can help agencies stage crews, refine evacuation plans and communicate risk before roads and shelters become crowded.

The important advance is therefore not simply that an AI drew a better path. In the peer-reviewed Nature study, WeatherNext Cyclones predicted track, intensity and wind radii together, then generated many plausible futures instead of presenting one outcome as certain.

What the researchers tested

The model was evaluated on tropical cyclones from 2023 through 2025 and compared with leading operational systems. The researchers report an average lead-time advantage of a day or more across the three central tasks. Google DeepMind's technical announcement says the system's three-day forecasts matched the accuracy that previous models reached at two days.

Training combined nearly 20 terabytes of global atmospheric data with IBTrACS, an expert-curated record spanning almost 5,000 historical storms. The model can project global weather and cyclone behavior as far as 15 days ahead. It was also run live during the 2025 Atlantic hurricane season, so the paper includes more than a purely retrospective laboratory exercise.

How coarse data produced detailed storm guidance

Conventional forecasting often involves a trade-off. Global models see the broad atmospheric currents that steer a cyclone, while high-resolution regional models resolve more of the compact processes that influence strength. WeatherNext Cyclones was co-trained to learn both the planet-scale weather pattern and the storm-specific features from inputs spaced about 28 kilometres apart.

The system uses a generative approach to create an ensemble, a collection of possible forecasts that represents uncertainty. It can generate as many as 1,000 scenarios, compared with about 50 in many conventional ensembles, and produce a 15-day forecast in under a minute on a tensor processing unit. Google has released the WeatherNext code and model documentation, including weights for operational and compact versions.

What happens next

Forecasters now need to learn when the model is most trustworthy and where it fails. The most useful operational product may be a consensus that combines WeatherNext Cyclones with physics-based systems, rather than a contest that chooses one model. The paper reports that adding its forecasts to a weighted consensus improved overall skill.

Open access to the code should make wider testing easier, although running an operational forecasting service still requires data pipelines, computing resources and local expertise. As independent coverage of the release notes, the public model forecasts out to 15 days, but a long horizon should be read as a widening range of possibilities, not a precise two-week warning.

The takeaway

WeatherNext Cyclones is a meaningful forecasting result because it tackles where a storm goes, how strong it becomes and how wide its winds spread in one operational system. Its value will be measured not by replacing meteorologists, but by whether it gives them earlier, better-calibrated evidence for decisions when uncertainty is most dangerous.