Over the past 50 years, tropical cyclones have killed more than 700,000 people and caused $1.4 trillion in economic losses. On August 6, Google DeepMind released the code and model weights of an AI system that forecasts these storms better than anything else in operation, and made it free for anyone to use.
The system is called WeatherNext, and the release came alongside a paper in Nature showing it achieved state-of-the-art accuracy in predicting a cyclone’s track, intensity, and wind structure. The headline result: WeatherNext gives forecasters roughly an extra day of warning. Its three-day forecasts are as accurate as what previous models managed for only two days out, an improvement DeepMind says is equivalent to about a decade of meteorological progress arriving all at once.
The forecast that helped Jamaica prepare
This is not a lab result waiting for a real-world test. During the 2025 hurricane season, WeatherNext helped the US National Hurricane Center predict Hurricane Melissa’s rapid intensification and its landfall in Jamaica, enabling the centre to issue an advance warning that gave teams on the ground critical time to prepare.
The model was built by AI researchers at Google DeepMind and Google Research working with expert forecasters at the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, the UK Met Office, and weather agencies around the world. That collaboration shows in the design. The system was trained end-to-end on nearly 20 terabytes of global atmospheric data alongside IBTrACS, an expert-curated database spanning close to 5,000 historical storms.
Solving forecasting’s oldest trade-off
Cyclone prediction has always forced meteorologists to choose between two kinds of models. A storm’s track is steered by massive global atmospheric currents, best captured by coarse worldwide models. Its intensity is driven by fine-scale thermodynamic processes around the storm’s core, best captured by specialised high-resolution local models. No single system did both well.
WeatherNext does both in one model, and it does so at a resolution that has surprised the scientists involved. The system works with input data at 28 by 28 kilometres, roughly 100 times coarser than traditional intensity models require. A smaller version, WeatherNext 2-mini, performs well at an even coarser 111 kilometres. Quite how the models achieve accurate intensity forecasts from such coarse data remains an open research question, which DeepMind says it hopes the wider community will now help answer.
Speed matters too. A single 15-day forecast takes less than a minute on one of Google’s TPU chips. Last year the system produced 50 simultaneous predictions per storm, matching conventional physics ensembles. This year it generates 1,000 possible scenarios for each cyclone, wide enough to capture rare but devastating outcomes like rapid intensification, the phenomenon that made Melissa so dangerous.
What open sourcing actually unlocks
The release includes WeatherNext Cyclones, the version that ran during the hurricane season, and WeatherNext 2, a broader update operationalised in October. Both the code and trained weights are on GitHub under an open licence. The mini version runs on a single TPU in a free public Colab notebook, which puts a competitive cyclone model within reach of a university lab, a national weather service in a developing country, or a curious graduate student.
That distribution question is arguably the real story. The countries most exposed to tropical cyclones are frequently the ones least able to afford supercomputer-scale forecasting infrastructure. An accurate model that runs cheaply, and that local agencies can adapt to their own coastlines, changes who gets access to life-saving lead time. DeepMind is pitching the release at exactly this audience, alongside uses in renewable energy planning and extreme weather anticipation.
Anyone can also explore the forecasts directly through Weather Lab, DeepMind’s visualisation platform, which was refreshed with the release to show temperature, precipitation, and wind predictions alongside cyclone tracks.
The season ahead
The Atlantic hurricane season peaks between August and October, which makes the timing of this release more than symbolic. Forecasters will have the strongest cyclone model yet published available as open infrastructure while the most dangerous months of the year unfold.
The larger test is what the research community builds on top. DeepMind has invited meteorological agencies and researchers to partner with it and extend the models, and the unexplained accuracy at coarse resolution gives atmospheric scientists a genuine puzzle to work on. Better answers there could compound into better forecasts for every kind of extreme weather, not just the storms with names.
For more coverage of AI and climate technology, visit Mylistingo.
