Google DeepMind: WeatherNext Redefines Cyclone Forecasting

Google Deepmind: WeatherNext Redefines Cyclone Forecasting

WeatherNext from Google Deepmind predicts cyclone trajectories and intensity simultaneously. Deepmind's AI forecasts tropical cyclones with greater accuracy than specialized models, while using data that is approximately one hundred times less detailed. The way this is achieved remains unclear even to the developers.

Google Deepmind introduces WeatherNext Cyclones, or WN-C, an AI system for forecasting tropical cyclones that can anticipate about one day further into the future than leading operational models. This improvement corresponds to the advancements made by traditional weather models over the past decade.

The model was developed in collaboration with the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere, and the UK Met Office. Since June 2025, forecasts have been broadcast live on Google’s Weather Lab. During Hurricane Melissa, which made landfall in Jamaica in 2025, the model helped the NHC predict the rapid intensification of the storm in time, according to Deepmind. This occurs when a storm gains at least 30 knots (approximately 34 mph) in wind speed within 24 hours.

WN-C Solves a Decades-Old Trade-Off

Cyclone forecasting has long suffered from a trade-off. Global models like the ECMWF ensemble system are effective at predicting trajectory but too coarse for intensity. Specialized regional models like the Hurricane Analysis and Forecast System (HAFS) provide more accurate intensity readings but lose precision on trajectory. WN-C manages both in a single system, according to a paper published in Nature.

For a five-day forecast, the estimated position of the storm's center is off by an average of 230 kilometers, compared to 370 kilometers for the ENS and 335 kilometers for Deepmind's previous model, GenCast. For three-day intensity forecasts, WN-C is 3.75 knots (approximately 4.3 mph) more accurate than HAFS.

WN-C also achieves more than double the results of the ENS and GenCast on probabilistic storm intensity forecasts over many time frames. For the probability of winds at 64 knots, the threshold at which a storm reaches hurricane strength, the model provides a more actionable value for decision-making than the ENS, according to Deepmind.

Coarse Data, More Accurate Forecasts

WN-C operates with a data grid where each point covers approximately 28 kilometers, which is about one hundred times less detailed than specialized regional models. Even a compact variant with 111 kilometers per grid point produces competitive results, according to Deepmind.

The authors write that "high resolution is not a strict requirement for state-of-the-art intensity forecasting." Coarse weather data seems to contain more information about storm strength than anyone had thought. How the models generate such accurate forecasts at this resolution remains an open research question.

Functional Generative Networks Replace Diffusion

WN-C uses Functional Generative Networks (FGN). GenCast relies on diffusion, a method better known from AI image generators, and requires dozens of passes through the neural network per forecasting step. FGN only requires a single pass, making it eight times faster.

To capture uncertainty, the system does not add noise to individual pixels as most approaches do. Instead, it injects noise into the control layers of the network. This produces different but physically coherent scenarios rather than just noisy variations of the same output.

Training on two very different data sources has been crucial. The team used nearly 20 terabytes of global atmospheric data from ECMWF analyses, as well as a carefully curated database of about 5,000 historical cyclones. To allow the model to learn from both sources together, the tabular data of storms, including position, intensity, and extent, are projected onto the same geographic grid as the weather data.

This enables the network to directly predict trajectory, intensity, and wind radii. Ablation studies in the paper show that this joint training is what drives the improvements.

1,000 Scenarios per Storm

A 15-day forecast runs in under a minute on one of Google’s AI chips, allowing Deepmind to increase the number of parallel forecasts from 50 to 1,000. Large ensembles like this could help capture rare extreme events more reliably. The Critical Success Index, which weighs correct warnings against false alarms and missed events, improves from below 0.3 to 0.5, according to the paper.

The blue line of WN-C reaches the error levels of the ENS, GenCast, and HAFS about one day later in each case, representing an additional day of warning time in practice. At the five-day horizon, the position error is 230 kilometers instead of 370.

A Complement, Not a Replacement

Traditional numerical models remain important, according to Deepmind. The NHC typically combines the results of several physics-based models into a weighted average for its official forecasts, divided between trajectory (TVCN) and intensity (IVCN). In a simulated weighted addition to these consensus models, WN-C improves trajectory forecasts by an average of 28%. For intensity, the gain is about 6%. The smaller intensity improvement shows that traditional models still contribute significantly in this area.

Position errors from the ENS and intensity errors from the HWRF have been slowly decreasing since 2005. The blue data points from 2023 show that WN-C has gained as much in a single step as traditional models have managed in about a decade.

A Model Years in the Making

Deepmind has made all of this publicly available, publishing the code and weights for WeatherNext 2 and WeatherNext Cyclones on GitHub. The mini variant runs on a single TPU in a free Colab notebook. For official warnings, the authors still direct users to national weather services. WN-C is intended to support forecasters, not replace them.

Last year, Deepmind introduced WeatherNext 2, the underlying weather AI that is now open-source along with the cyclone variant. In June 2025, Deepmind and Google Research launched Weather Lab, a platform designed for cyclone forecasting whose experimental model became WeatherNext Cyclones. The collaboration with the NHC and CIRA was already in place at that time.

Before that, Deepmind released GenCast in late 2024, the first probabilistic weather model to outperform the entire ECMWF, also open-source. WN-C replaces GenCast's diffusion approach with the faster FGN method and expands the scope to tropical cyclones.

However, Google’s track record in forecasting has not always been as solid. Its Android Earthquake Alerts system severely underestimated earthquakes in Turkey in 2023, rating a 7.8 earthquake at only 4.5 to 4.9 and sending out only 469 critical alerts when millions should have been warned.

This scientific achievement comes during a turbulent time for Deepmind. CEO Demis Hassabis has recently stepped back from daily operations to become the Chief Scientist of Alphabet, while Chief Researcher Jeff Dean has left after 27 years to co-found the AI startup Discovery Loop with several other senior researchers.