How DeepMind's WeatherNext Is Changing Cyclone Forecasting

Tropical cyclones claim thousands of lives and cause billions in damage each year. Every precious hour of accurate warning can mean the difference between orderly evacuation and disaster. On August 6, 2026, Google's DeepMind introduced WeatherNext Cyclones, an AI system that dramatically advances the ability to forecast these deadly storms.

WeatherNext takes a big step forward by predicting a cyclone's track, intensity, and wind structure all at once in a single model rather than piecing together separate models. This is a major improvement over the traditional approach of using coarse global models to predict a storm's path and specialized local models to predict its strength.

WeatherNext was trained on nearly 20 terabytes of atmospheric data and the historical database of tropical cyclones, which contains records of more than 5,000 past storms. The model's core component, Functional Generative Networks (FGNs), allows it to quickly generate a large number of possible storm outcomes, which helps forecasters understand the likelihood of dangerous, rare events.

Despite operating at a coarser resolution than traditional physics-based systems, WeatherNext matches or exceeds their accuracy. In tests using historical cyclones from 2023 to 2025, the model provided an extra day of reliable warning compared to leading operational systems. A three-day WeatherNext forecast was as accurate as previous systems' two-day forecasts.

At the five-day mark, the model's average track error was about 230 kilometers, while other systems were around 370 and 335 kilometers respectively. When blended with official forecasts, WeatherNext improved track accuracy by 28% and intensity accuracy by 6%. During the 2025 Atlantic hurricane season, the U.S. National Hurricane Center used WeatherNext to help forecast Hurricane Melissa's rapid intensification and its landfall in Jamaica.

DeepMind has released the model's code and pretrained weights under an Apache 2.0 license, enabling researchers worldwide to build upon the work. The model is intended to be a tool to support forecasters, not replace them. WeatherNext represents a significant leap forward in cyclone forecasting, providing more time for preparation, evacuation, and response.