Our WeatherNext 2 AI model demonstrated a massive leap forward in predicting cyclones.

WeatherNext, an AI model developed by Google DeepMind and Google Research, has made a significant breakthrough in forecasting cyclones. This achievement enables accurate cyclone forecasts that can provide an extra day of warning. The model is now open sourced, aiming to empower the research community and enhance AI's role in building resilient communities.

Cyclones, such as hurricanes or typhoons, pose a considerable threat, causing over 700,000 deaths and $1.4 trillion in economic losses globally in the past 50 years. Accurate forecasting is crucial, as timely and precise warnings are essential to mitigate the impact of these destructive weather phenomena.

The WeatherNext model achieves state-of-the-art accuracy in predicting a cyclone's track, intensity, and wind structure. It offers an average of an additional day of predictive accuracy, meaning that a three-day forecast is as reliable as previously possible for just two days. This improvement equates to a decade's worth of meteorological progress.

To create this groundbreaking model, researchers from Google DeepMind and Google Research collaborated with expert forecasters from various institutions, including the National Hurricane Center, the UK Met Office, and weather agencies worldwide.

The model's success is attributed to its innovative training, architecture, and approach to handling low-resolution inputs. It was trained end-to-end on nearly 20 terabytes of global atmospheric data and historical cyclone observations, spanning over 5,000 storms. By utilizing Functional Generative Networks (FGNs), the model produces ensembles of different predictions, capturing the inherent uncertainty of weather events.

This allows the generation of a single 15-day forecast in less than a minute on a TPU, enabling forecasters to swiftly evaluate potential tail risks.

In the 2025 hurricane season, WeatherNext significantly impacted the National Hurricane Center's forecasting efforts. It aided in predicting Hurricane Melissa's rapid intensification and landfall in Jamaica, providing the center with critical time to prepare and issue an advance warning. Building on this success, the team now predicts 1,000 possible scenarios for each cyclone, assisting forecasters in making informed decisions.

WeatherNext models come in different sizes, with WeatherNext 2-mini operating at a coarser resolution (111x111km) while maintaining excellent performance. This surprising discovery has prompted further research to understand the factors contributing to the model's accuracy at this resolution.

By opening the WeatherNext model to the public, Google aims to spur progress across the global weather community, benefiting meteorological agencies, researchers, and nonprofits. The organization is releasing various versions of the model, including WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini. WeatherNext 2-mini, in particular, can run on a single TPU within a free public Colab notebook.

The model's predictions for temperature, precipitation, wind speed, and more are accessible through Weather Lab, a refreshed platform with an expanded interface and global weather forecasts. This collaboration between tech giants and the research community marks a historic breakthrough in cyclone forecasting, providing more time to prepare for and mitigate the devastating effects of these natural disasters.