DeepMind's WeatherNext buys cyclone forecasters an extra day

Google DeepMind has open-sourced WeatherNext, a family of AI models designed to predict tropical cyclone track, intensity, and wind structure. By tackling both global atmospheric movement and local storm physics in a single model, it delivers three-day forecasts that match the accuracy of traditional two-day forecasts—effectively buying forecasters an extra 24 hours of lead time, an advance equivalent to roughly a decade of meteorological progress.

The release includes the code and model weights on GitHub for both WeatherNext Cyclones and WeatherNext 2, along with a lightweight WeatherNext 2-mini model.

What it does

Cyclone forecasting has always been stuck in an awkward compromise. Tracking where a storm goes requires global models that map massive atmospheric currents. But gauging how intense it will get demands high-resolution local models focusing on the micro-thermodynamics inside the storm core.

WeatherNext bridges that divide. It forecasts the entire lifecycle—path, intensity, and wind structure—up to 15 days in advance. To capture rare and volatile rapid-intensification events, the model uses Functional Generative Networks to simulate an ensemble of 1,000 distinct probability scenarios at once.

This is not purely theoretical. During the 2025 hurricane season, an earlier version helped the US National Hurricane Center (NHC) predict the rapid intensification and Jamaica landfall of Hurricane Melissa.

How it works and hardware requirements

The system was trained end-to-end on nearly 20 terabytes of global atmospheric data alongside the IBTrACS database covering nearly 5,000 historical storms.

The real surprise is how coarse its inputs can be. Traditional intensity models demand ultra-fine spatial data. WeatherNext Cyclones achieves state-of-the-art results with 28x28 km inputs—about 100 times coarser than physics-based setups. The smaller 2-mini version operates at an even coarser 111x111 km resolution.

Because of that lower data overhead, generating a full 15-day forecast takes less than a minute on a single Google TPU. Researchers can pull down the weights from GitHub, while developers can test WeatherNext 2-mini directly inside a free Google Colab notebook. Non-technical users can also inspect the outputs interactively on the updated Weather Lab portal.

Why it matters

If you work in disaster response, atmospheric research, or climate risk modeling, this is worth testing immediately. Traditionally, running a 1,000-member ensemble at this accuracy would devour massive supercomputing clusters for hours; DeepMind just shrunk that footprint down to a single TPU and a minute of compute.

That said, exactly how the model pulls off such fine-grained intensity forecasts from 28 km data remains an open scientific puzzle that even DeepMind says needs further study. It is a massive weapon for operational desks, but human forecasters will still need to keep their hands on the wheel.