WeatherNext: AI model achieves breakthrough in forecasting cyclones

Google DeepMind has announced WeatherNext, an AI model that achieves state-of-the-art accuracy in predicting the track, intensity, and wind structure of tropical cyclones. The model provides forecasters with an average of one additional day of predictive accuracy, meaning three-day forecasts are now as accurate as previous models' two-day forecasts.

State-of-the-Art Cyclone Forecasting Accuracy

WeatherNext Cyclones delivers a lead time advantage of more than 24 hours for predicting cyclone tracks, intensity, and wind structure. According to research published in Nature, this improvement is equivalent to approximately ten years of meteorological progress based on trends from the last 20 years.

Key performance benchmarks include:

Position Error: WeatherNext Cyclones shows significantly lower position error (approximately 100 km) compared to ENS models for 3-day forecasts between 2023 and 2025.
Intensity Error: The model achieves lower intensity error (approximately 11 kt) than HWRF models for the same period.

Technical Architecture and Training

WeatherNext bridges the gap between global atmospheric currents (which steer a cyclone's track) and localized thermodynamic processes (which drive intensity). It achieves this as a single AI model rather than requiring two distinct modeling techniques.

Training Data and Modalities

The model was co-trained on two distinct data modalities:

Global Weather Dynamics: Nearly 20 terabytes of global atmospheric data.
Historical Observations: The IBTrACS database, spanning nearly 5,000 historical storms.

Ensemble Forecasting and Uncertainty

WeatherNext utilizes Functional Generative Networks (FGNs) to produce ensembles of predictions to capture weather uncertainty. The system has scaled from 50 predictions per run to 1,000-member ensembles. This allows the model to capture rare but consequential "tail-risks," such as the rapid intensification events seen during Hurricane Melissa in 2025.

Computational Efficiency and Resolution

WeatherNext can generate a single 15-day forecast in less than a minute on a TPU. Notably, the model achieves high accuracy using data with a resolution of 28x28km—100 times coarser than traditional models. A compact version, WeatherNext 2-mini, operates at an even coarser 111x111km resolution while maintaining strong performance.

Real-World Impact and Open Source Release

WeatherNext has already demonstrated operational utility. During the 2025 hurricane season, it helped the National Hurricane Center (NHC) predict the rapid intensification and landfall of Hurricane Melissa in Jamaica, enabling critical advance warnings.

To accelerate global research and resilience, Google DeepMind is open-sourcing the following:

WeatherNext Cyclones: The model used during the hurricane season.
WeatherNext 2: An updated version operationalized in October.
WeatherNext 2-mini: A compact version capable of running on a single TPU via a public Colab notebook.

Integration with Google Earth AI

WeatherNext models are part of Google Earth AI and are integrated into Weather Lab. The platform now allows users to visualize predictions for wind speed, temperature, and precipitation alongside cyclone tracks in a single view.