Google is releasing its cyclone forecast AI system to the public. The model, called WeatherNext, is now open source. Researchers say it gives an extra day of warning before major storms strike.

A Race Against Time

Tropical cyclones remain among the most destructive forces on the planet. Over the past 50 years, they have caused more than 700,000 deaths and $1.4 trillion in economic damage worldwide. For that reason, every hour of lead time matters for forecasters trying to issue timely warnings.

According to a new paper published in Nature, the WeatherNext model has achieved state-of-the-art results in predicting a cyclone’s track, intensity, and wind structure. On average, the system provides forecasters with an extra day of accuracy. As a result, three-day forecasts from WeatherNext now match the reliability that older models could only achieve two days out. Researchers describe this leap as equivalent to roughly a decade of typical progress in meteorology.

A Global Collaboration

The project brought together AI researchers and engineers from Google DeepMind and Google Research. Meanwhile, expert forecasters from the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office also contributed, alongside weather agencies from around the world.

Notably, the cyclone forecast model has already proven itself in the field. During the 2025 hurricane season, it helped the NHC make a historic call ahead of Hurricane Melissa. By predicting the storm’s rapid intensification and eventual landfall in Jamaica, the model gave teams on the ground extra time to prepare. This year, the researchers have pushed further, now generating 1,000 possible scenarios for each storm to support forecaster decision-making.

Why Open Source It Now

Because weather affects everyone, the team decided to release WeatherNext 2 and WeatherNext Cyclones, the versions used throughout the hurricane season, to the public. In doing so, they hope to strengthen the research community and expand AI’s role in building more resilient communities. That includes equipping local forecasters, supporting renewable energy planning, and helping regions anticipate extreme weather events.

Solving a Long-Standing Trade-Off

Traditionally, predicting cyclones required two separate approaches. A storm’s track, driven by large-scale atmospheric currents, has typically been modeled using coarse global systems. On the other hand, a storm’s intensity depends on fine-scale processes near its core, which usually call for specialized, high-resolution local models.

WeatherNext bridges that divide. It is a single AI model capable of forecasting a cyclone’s track, intensity, and wind structure simultaneously, with state-of-the-art accuracy. It achieves this through its training approach, architecture, and its handling of lower-resolution input data.

When benchmarked against historical cyclones from 2023 to 2024, WeatherNext Cyclones outperformed other leading weather models. On average, it delivered more than a full 24-hour lead-time advantage across track, intensity, and wind structure predictions.

How the Model Was Trained

The system was co-trained on two very different types of data: global weather dynamics and expert-curated historical cyclone records. Specifically, it learned from nearly 20 terabytes of global atmospheric data, combined with the IBTrACS database, which spans almost 5,000 historical storms. Through this process, the model absorbed complex atmospheric patterns and learned how extreme weather behaves.

Furthermore, cyclone forecasting accuracy has been improving steadily for decades. Comparisons show that WeatherNext Cyclones represents a genuine step change, in both track and intensity accuracy, equal to about a decade of normal progress based on 20-year trends.

Faster, Bigger Ensembles

To capture uncertainty, the model relies on Functional Generative Networks (FGNs), which efficiently generate ensembles of possible outcomes. As a result, a full 15-day forecast can now be produced in under a minute on a single TPU. This speed allows forecasters to quickly assess the odds of potentially devastating tail-risk scenarios.

Last year, the system produced 50 predictions at once, matching the output of global physics-based models. This year, however, the team scaled the ensemble up to 1,000 members. That expansion allows the model to capture rare but consequential events, such as the rapid intensification seen during Hurricane Melissa.

A Surprising Resolution Finding

Until now, scientists generally assumed that high spatial resolution was essential for accurate intensity forecasts. Surprisingly, though, WeatherNext Cyclones only requires data at a 28x28km resolution, roughly 100 times coarser than traditional models. Even more strikingly, a smaller version called WeatherNext 2-mini performs well at just 111x111km resolution. Researchers admit this result has surprised the scientific community, and exactly why the models perform so well at coarse resolution remains an open question.

Opening the Doors to Researchers

Alongside the Nature publication, the team is releasing the code and model weights for free public use. This applies to academic research, operational forecasting, and the development of more specialized, localized models. The goal, researchers say, is to accelerate progress across the global weather community.

Two related model sets are being released together: WeatherNext Cyclones, which operated throughout the hurricane season, and WeatherNext 2, a later update introduced in October. In addition, WeatherNext 2-mini, a lightweight version, can run on a single TPU through a free public Colab notebook.

Interested users can also explore live forecasts on Weather Lab, which was recently redesigned. The updated platform now includes global weather forecasts alongside cyclone tracks, letting users view temperature, precipitation, and wind speed predictions in a single interface. Both Weather Lab and WeatherNext are part of the broader Google Earth AI initiative.

Looking Ahead

Overall, the team describes this as a historic breakthrough, gaining more than a full day of lead time for cyclone prediction. As future storm seasons approach, researchers are inviting other scientists, meteorological agencies, and experts to build on the open-source models and explore the tools on Weather Lab. By pairing advanced machine learning with human forecasting expertise, the team hopes to build a collaborative ecosystem capable of saving lives as the climate continues to change.

For official storm warnings, readers are advised to consult their local meteorological agency or national weather service. Nonetheless, the release of this cyclone forecast system marks a significant step in making advanced storm prediction tools available to forecasters and researchers worldwide.

Related post

View all