Google DeepMind and Google Research have introduced an artificial intelligence model capable of extending tropical cyclone forecasts by a full day, giving authorities extra time to prepare for severe storms.
The new technology, named WeatherNext, can produce three-day predictions with accuracy comparable to what previous weather models achieved over a two-day timeframe. The study detailing the technology was published in the scientific journal Nature.
WeatherNext forecasts not only the trajectory of a storm, indicating where it will travel, but also predicts storm intensity and surrounding wind structure. Tropical cyclones, which include hurricanes and typhoons, rank among the most destructive weather phenomena on Earth.
According to data presented by Google, tropical cyclones have caused more than 700,000 deaths and $1.4 trillion in economic damage worldwide over the past 50 years. Meteorologists emphasized that each additional hour of warning expands the crucial window needed to evacuate populations, deploy emergency responders and protect vulnerable infrastructure.
Combining global models and localized storm data
The WeatherNext system was trained on nearly 20 terabytes of global atmospheric data alongside the International Best Track Archive for Climate Stewardship database, known as IBTrACS, which contains historical records for almost 5,000 storms. By analyzing this vast repository, the model learns atmospheric patterns and key indicators associated with extreme weather events.
The technology addresses a long-standing challenge in meteorological forecasting. A storm's overall trajectory is governed by large atmospheric currents best tracked by global weather models, while its intensity is driven by smaller, localized phenomena near the storm core that typically require high-resolution modeling.
WeatherNext unifies both global and localized approaches into a single artificial intelligence architecture. Researchers highlighted that the system calculates intensity predictions using atmospheric data mapped at a resolution of 28 by 28 kilometres. Google noted that this grid is approximately 100 times coarser than the resolution required by traditional forecasting models.
A smaller model variant, designated WeatherNext 2-mini, operates on an even lower resolution of 111 by 111 kilometres. Researchers reported that this compact version also demonstrated high performance during testing.
Simulating atmospheric scenarios and hurricane tracking
Rather than generating a single deterministic path, the WeatherNext system produces multiple alternative scenarios for a single weather event. This ensemble approach accounts for the natural uncertainty inherent in atmospheric predictions.
Google stated that a complete 15-day forecast can be calculated in under one minute using a specialized hardware processor. During the previous hurricane season, the system evaluated 50 scenarios simultaneously. Google reported that it has now scaled the model to run 1,000 simultaneous scenarios, enabling it to detect rare but dangerous events such as rapid storm intensification.
The system was deployed during the 2025 hurricane season. According to Google, WeatherNext assisted the National Hurricane Center in the United States by forecasting the rapid intensification of Hurricane Melissa before it struck Jamaica, allowing forecasters to issue public warnings earlier.
Open source release and research tools
Following the publication of the research in Nature, Google released the source code and model weights for WeatherNext Cyclones and WeatherNext 2 free of charge to researchers, meteorological agencies and international organizations.
Google also made WeatherNext 2-mini available for public evaluation through a sample notebook on Google Colab, allowing it to run on a single Tensor Processing Unit. The initiative is intended to help external researchers build specialized regional forecasting tools. In addition, Google maintains Weather Lab, a visualization platform displaying model predictions for cyclone tracks, temperature, rainfall and wind speed.
Despite these results, Google noted that researchers do not yet fully understand how the artificial intelligence model achieves such high predictive accuracy while utilizing relatively coarse data resolution. Google stated that identifying the underlying mechanism remains an open scientific question.
