DeepMind's cyclone AI sees a day further ahead using data 100x coarser

Google DeepMind has dropped a new weather model that breaks a decades-old tradeoff in cyclone forecasting, and the researchers admit they aren’t entirely sure how it works. WeatherNext Cyclones, or WN-C, predicts both the track and intensity of tropical storms in a single system, something that’s long forced forecasters to choose between two specialized tools.

The model can see about one day further into the future than leading operational systems. For a five-day forecast, WN-C’s estimated storm center is off by an average of 230 kilometers. That stacks up against 370 kilometers for the European Centre for Medium-Range Weather Forecasts’ ensemble and 335 kilometers for DeepMind’s own GenCast. On three-day intensity forecasts, it’s 3.75 knots more accurate than NOAA’s specialized hurricane model. Since June 2025, the forecasts have been running live on Google’s Weather Lab, and during Hurricane Melissa the system helped the National Hurricane Center spot rapid intensification in time.

What’s genuinely weird is the resolution. WN-C works with a data grid where each point covers about 28 kilometers — roughly a hundred times coarser than the regional models that specialize in intensity. Even a stripped-down variant at 111 kilometers per grid point holds its own. “High resolution is not a strict prerequisite for state-of-the-art intensity forecasting,” the authors write in a paper published in Nature. The coarse data apparently encodes more storm-strength information than anyone assumed, and how the model extracts it remains an open research question.

The speed comes from swapping GenCast’s diffusion method for Functional Generative Networks. Diffusion needs dozens of passes through the neural network per forecast step; FGN gets by with one, making it eight times faster. A 15-day forecast runs in under a minute on a single Google AI chip, so the team scaled ensemble runs from 50 to 1,000 parallel forecasts. That many scenarios could catch rare extreme events that smaller ensembles miss. But traditional physics-based models aren’t going anywhere — when WN-C was added to the NHC’s consensus mix, intensity forecasts only improved by about 6 percent, suggesting the old guard still has a strong hand in that fight.

The code and model weights are public on GitHub, with a mini variant running on a free Colab notebook. It’s a rare move of genuine openness from a major lab, and it puts a powerful cyclone forecasting tool in the hands of anyone with a browser.