WeatherNext Cyclones Gives Forecasters an Extra Day on Tropical Storms
Google DeepMind's WeatherNext Cyclones gives forecasters roughly an extra day of warning on tropical cyclones, tested operationally during the 2025 Atlantic hurricane season with Hurricane Melissa. The code is open-sourced under Apache 2.0.
On August 6, Google DeepMind published WeatherNext Cyclones in Nature — an AI model that produces state-of-the-art ensemble forecasts for track, intensity, and wind radii of tropical cyclones, offering what the paper describes as an average lead time advantage of a day or more over leading operational systems. That gap, the authors note, is comparable to roughly a decade of conventional progress in tropical cyclone forecasting — a domain where even marginal improvements translate directly into additional hours of evacuation and preparation time.
The model works differently from the physics-based numerical systems it’s measured against. Rather than simulating fluid dynamics on a fine-resolution grid, WeatherNext Cyclones uses functional generative networks to produce 1,000 possible weather predictions, processing a 15-day forecast in under a minute on a TPU despite operating on a coarser 28×28 km input resolution. The ensemble approach means forecasters don’t get a single deterministic track — they get a probabilistic spread that better reflects the genuine uncertainty inherent in predicting where a storm will go and how strong it will be when it gets there.
Three-day forecasts produced by the model are as accurate as what prior systems achieved at two days, and that extra day matters most when a storm is intensifying rapidly and time to act is shortest. During the 2025 Atlantic hurricane season, the National Hurricane Center used WeatherNext Cyclones operationally — and the model helped the NHC make a historic forecast for Hurricane Melissa by predicting rapid intensification and landfall in Jamaica. A forecast like that, issued in time for meaningful response, is the kind of event that stress-tests whether a model actually works when pressure — literal and figurative — is highest.
Google DeepMind has open-sourced the code and model weights under Apache 2.0 across three variants released on GitHub, including a mini model that runs on a free Google Colab notebook. The mini model’s accessibility is worth noting: it means the base technology isn’t locked behind institutional compute budgets or proprietary licensing, and researchers at meteorological services in countries that can’t afford enterprise hardware can still run experiments with the core architecture. Apache 2.0 is one of the most permissive open-source licenses available, so downstream use — including commercial applications — doesn’t require permission from DeepMind.
The Nature paper positions the improvement as comparable to a decade of operational development progress, which is the kind of claim that invites careful scrutiny. Traditional numerical weather prediction models have been refined over half a century of sustained investment by agencies like NOAA, the ECMWF, and national met offices; the idea that a generative model can match or exceed that trajectory in a single step is a strong assertion. The open-sourcing helps: whether the improvement holds up will depend on independent teams reproducing the results in operational settings, not on any single benchmark.
What’s less ambiguous is the operational evidence. The NHC’s use of the model during an actual hurricane season — and the specific track record on Melissa’s rapid intensification and Jamaica landfall — gives WeatherNext Cyclones something most AI weather research lacks: a real-world test case where getting the forecast wrong had immediate, tangible consequences. The gap between promising results in a Nature paper and a tool that actually changes how a forecast center makes decisions on a Tuesday night is large, and this model appears to have closed it.
Whether the technology reshapes cyclone forecasting broadly depends on adoption beyond the handful of agencies that have already piloted it, and on whether the probabilistic ensemble output integrates cleanly into the decision workflows forecasters already use. The open-source release lowers one barrier to that adoption. Whether the rest follows will be a story told over the next several hurricane seasons — starting, in practice, with the one already underway.