WeatherNext: AI model achieves breakthrough in forecasting cyclones

Every hour matters when a cyclone is bearing down on a coastline, and for decades forecasters have had to choose between two flawed approaches: coarse global models good at tracking where a storm is headed, or specialized local models better at gauging how strong it will get. Google DeepMind says it has closed that gap with WeatherNext, a single AI model detailed in a new Nature paper that predicts a cyclone's track, intensity, and wind structure all at once, and does it with enough lead time to matter.

The headline number is stark: three-day forecasts from WeatherNext now match the accuracy that older models could only manage for two days out. DeepMind frames that jump as roughly a decade's worth of meteorological progress compressed into one model. It's not just a lab result either — during the 2025 hurricane season, the system helped the National Hurricane Center anticipate Hurricane Melissa's rapid intensification and its landfall in Jamaica, giving ground teams more time to prepare before the storm hit.

What makes this work technically odd is the resolution question. Conventional wisdom held that sharper intensity forecasts required very fine-grained spatial data. WeatherNext Cyclones instead runs on data roughly 100 times coarser than traditional models, at 28x28km, and still produces state-of-the-art results. A smaller variant, WeatherNext 2-mini, runs at an even coarser 111x111km and still performs well enough to surprise the researchers themselves — they admit they don't fully understand why it works this well at such low resolution, and they're treating that as an open question for the wider research community to help answer.

Under the hood, the model was trained end-to-end on close to 20 terabytes of global atmospheric data alongside the IBTrACS historical storm database, which covers nearly 5,000 past cyclones. It uses something DeepMind calls Functional Generative Networks to generate ensembles of possible outcomes cheaply — a full 15-day forecast in under a minute on a TPU. Last year the system produced 50 scenarios at a time; this hurricane season it scaled that to 1,000, specifically to catch rare but dangerous events like the kind of rapid intensification Melissa showed.

DeepMind is now releasing the code and weights for WeatherNext 2, WeatherNext Cyclones, and the lightweight WeatherNext 2-mini, the last of which can run on a single TPU through a free public Colab notebook. The stated goal is to hand these tools to meteorological agencies, academic researchers, and nonprofits so they can build their own localized or specialized versions, and to let anyone explore current forecasts through the newly redesigned Weather Lab interface, which now shows global weather variables alongside cyclone tracks.