An Extra Day of Warning: Google DeepMind Open-Sources WeatherNext, the AI That Jumped Hurricane Forecasting a Decade Ahead

Every hour counts when a hurricane is bearing down on a coastline. Tropical cyclones — hurricanes and typhoons — have killed more than 700,000 people and caused $1.4 trillion in economic losses worldwide over the past 50 years, and the difference between a two-day evacuation window and a three-day one is measured in lives. That is why the numbers in Google DeepMind’s latest results are turning heads across meteorology: the company’s WeatherNext AI model doesn’t just edge out conventional forecasts. On average, it gives forecasters an extra full day of predictive accuracy — three-day forecasts as good as what prior models could only deliver two days out.

As the New York Times reported on August 28, analysis by the company’s DeepMind unit shows the AI-enabled model delivers accurate forecasts “a day or more before conventional models.” Google’s own researchers frame the gain more bluntly: measured against the improvement trends of the last 20 years, this single model represents roughly a decade of meteorological progress.

A single model for a two-model problem

The result, published in Nature on August 6, 2026, solves a trade-off that has structured cyclone forecasting for decades. A storm’s track — where it goes — is steered by planetary-scale atmospheric currents, which coarse global models handle well. Its intensity — how strong it gets — is driven by fine-scale thermodynamic processes in the storm’s core, which historically demanded specialized high-resolution local models. Forecasters had to stitch the two together.

WeatherNext bridges the gap. It is one neural network that predicts track, intensity, and wind structure with state-of-the-art accuracy, evaluated on historical cyclones from 2023–2024 against the top operational weather models in the world. The work was co-developed by Google DeepMind and Google Research, with expert forecasters at the US National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office as partners and co-authors.

The training recipe combined two data modalities end-to-end: roughly 20 terabytes of global atmospheric data and the expert-curated IBTrACS database covering nearly 5,000 historical storms. The architecture uses Functional Generative Networks (FGNs) to generate ensembles — swarms of alternative forecasts that capture the inherent uncertainty of the atmosphere rather than pretending it doesn’t exist.

1,000 scenarios per storm, in under a minute

The operational numbers are as striking as the accuracy ones. A single 15-day forecast now runs in less than one minute on a single TPU chip. Last year the system produced 50 ensemble members at a time, roughly matching global physics models. This hurricane season it scaled to 1,000 possible scenarios per cyclone, which is what makes rare-but-catastrophic tail risks — like rapid intensification — visible to a forecaster instead of buried in an average.

That capability has already produced a historic save. During the 2025 season, the model helped the NHC forecast Hurricane Melissa’s rapid intensification and landfall in Jamaica, giving the agency confidence to issue warnings early and teams on the ground critical time to prepare. The NHC’s own 2025 verification report documents the episode.

Perhaps the most scientifically provocative detail: WeatherNext Cyclones achieves its intensity accuracy from input data at just 28×28 km resolution — about 100× coarser than traditional models. The assumption that ever-higher spatial resolution drives intensity forecasts is baked into decades of modeling strategy, and Google’s researchers admit it is now “an open research question” how the model does so much with so little. A compact sibling, WeatherNext 2-mini, runs at an even coarser 111×111 km — and fits in a free public Colab notebook on a single TPU.

Now open source — code and weights

Alongside the Nature paper, Google is open-sourcing the code and model weights for both WeatherNext Cyclones (the model that ran during the hurricane season) and WeatherNext 2 (a later update operationalized in October). The license terms invite academic research, operational forecasting by national agencies, and specialized localized derivatives. For meteorological services in storm-prone countries that could never fund an ECMWF-class modeling pipeline, that is the real story: frontier forecasting capability, free to download.

The public-facing Weather Lab was also refreshed with a new interface, expanding beyond cyclone tracks to global forecasts of temperature, precipitation, and wind speed in a single view. Both Weather Lab and the WeatherNext models sit under the Google Earth AI umbrella.

Why it matters

Two larger threads run through this release. The first is the ongoing collapse of the boundary between AI research and operational science. WeatherNext follows GraphCast (2023) and GenCast (2024) out of DeepMind, each one eating a chunk of a domain — medium-range forecasting, ensemble prediction, now cyclone prediction — that physics-based supercomputer models had owned since the 1980s. When a Nature paper plus open weights lets any agency reproduce a decade of forecast-skill gains, the economics of national weather services change.

The second is what “an extra day” actually buys. Evacuation decisions are made 48–72 hours out; rapid intensification near landfall is what kills, because it outruns the warning cycle. A model whose three-day forecast carries yesterday’s two-day error bars, running 1,000 scenarios in seconds, turns the hardest forecasting problem in operational meteorology into a probability map a human forecaster can actually reason over. Google is careful to note the human half of that loop — the model doesn’t replace forecasters, it arms them.

As the 2026 Atlantic season approaches its statistical peak, that combination — open weights, a minute-per-forecast, and a verified record from Melissa — makes WeatherNext the most consequential AI story in weather since AI forecasting itself arrived. The next decade of hurricane prediction may already be sitting in a Colab notebook.

Sources for this article are listed in the frontmatter and rendered below.