WeatherNext: AI model achieves breakthrough in forecasting cyclones
2026-08-20 · Google DeepMind
WeatherNext: AI Model Achieves Breakthrough in Forecasting Cyclones
Major Achievement
The WeatherNext AI model developed by Google DeepMind and Google Research has achieved state-of-the-art performance in tropical cyclone forecasting. According to a paper published in *Nature*, the model significantly outperforms previous systems in predicting a cyclone’s track, intensity, and wind structure.
On average, WeatherNext provides forecasters with an extra day of predictive accuracy. Its three-day forecasts are as accurate as what prior models could only achieve at the two-day range. This improvement corresponds to roughly a decade of conventional meteorological progress.
Real-World Impact
The model has already demonstrated practical value. During the 2025 hurricane season, WeatherNext enabled the National Hurricane Center (NHC) to issue a historic forecast for Hurricane Melissa, accurately predicting its rapid intensification and landfall in Jamaica. This advance warning gave emergency teams on the ground critical preparation time.
This year, the system now generates 1,000 possible scenarios for each cyclone to better support forecasters in assessing risks and decision-making.
How WeatherNext Works
Predicting cyclones has traditionally required a trade-off between two modeling approaches:
- Track forecasting: Best performed by coarse global models that capture large-scale atmospheric steering currents.
- Intensity and structure forecasting: Best performed by high-resolution local models focused on fine-scale thermodynamic processes near the storm’s core.
WeatherNext bridges this gap with a single AI model. Starting from global atmospheric conditions (for example, during Hurricane Milton in October 2024), it iteratively predicts both large-scale weather patterns and fine-scale cyclone evolution up to 15 days ahead. A 1,000-member ensemble produces localized probability maps of tropical storm to hurricane-force winds.
Training and Technical Innovation
The model was co-trained on two distinct data modalities:
- Nearly 20 terabytes of global atmospheric data.
- The IBTrACS database of nearly 5,000 historical tropical cyclones.
It employs Functional Generative Networks (FGNs) to efficiently generate diverse ensemble forecasts that capture inherent weather uncertainty. A complete 15-day forecast can be produced in less than a minute on a TPU.
The ensemble size has been scaled from 50 members last year to 1,000 members this year, significantly improving the capture of rare but high-impact events such as rapid intensification.
Surprising Performance at Low Resolution
Evaluations on historical cyclones from 2023–2024 show WeatherNext delivers more than 24 hours of additional lead time for track, intensity, and wind structure forecasts compared to leading models.
Notably, the model achieves these results using only 28×28 km resolution input data — approximately 100 times coarser than traditional intensity models. Even the smaller WeatherNext 2-mini variant performs well at 111×111 km resolution. Understanding exactly how the model extracts such accurate predictions at coarse scales remains an open research question.
Open-Sourcing the Models
Alongside the *Nature* paper, the team is releasing both WeatherNext 2 and WeatherNext Cyclones models as open source. By making this technology freely available, the collaborators aim to empower the global research community, support local forecasters in disaster preparedness, accelerate renewable energy applications, and build greater societal resilience to extreme weather.