Google DeepMind Researchers Develop AI to Better Predict Hurricane Paths and Intensity

Tropical cyclone forecasting has long faced a trade-off between scale and detail: global models excel at capturing large-scale circulation but struggle to resolve small-scale processes in the storm core; regional high-resolution models (such as HAFS) can better represent these processes but require substantial computational resources and have limited coverage. Previous global AI models also had similar shortcomings—they could accurately predict tracks but often underestimated peak wind speeds.

WN-C employs a joint training strategy: it simultaneously uses global gridded atmospheric data and a specialized database of nearly 5,000 tropical cyclone observations (IBTrACS), enabling the model to learn both large-scale atmospheric behavior and the characteristics of real cyclones. Notably, WN-C's atmospheric input resolution is about 0.25 degrees (approximately 28 km near the equator), yet it still produces competitive intensity forecasts, suggesting that extremely high spatial resolution may not be necessary for accurate intensity prediction.

In tests on tropical cyclones from 2023 to 2025, WN-C achieved an average track error of about 230 km over the five-day forecast period, compared to 370 km for ECMWF's ENS system and 335 km for GenCast. To reach the same 230 km accuracy, ENS would need a lead time of about 3.75 days, meaning WN-C provides approximately 30 hours of additional warning time.

For intensity, WN-C's average intensity forecast at three days is 3.75 knots more accurate than NOAA's HAFS model. The study states that these improvements are equivalent to about a decade of progress in traditional numerical weather prediction.

WN-C is an ensemble model, with a standard forecast containing 50 members, but due to its computational speed, it can generate ensembles of up to 1000 members. Large ensembles are particularly important for estimating rare but costly events, such as extreme wind conditions days ahead. The study found that increasing the ensemble from 50 to 1000 members is especially useful for estimating such events.

The model can also estimate the probability of a location experiencing winds of 34, 50, or 64 knots or more. The study shows that WN-C's predictive uncertainty is generally well calibrated, meaning the ensemble spread closely matches the actual errors.

Rapid intensification (an increase in maximum sustained winds of at least 30 knots within 24 hours) is one of the most challenging problems in hurricane forecasting, as storms can quickly become dangerous before landfall. WN-C achieves a better balance between detecting rapid intensification events and avoiding false alarms, with its critical success index improving from below 0.3 for comparison models to about 0.5.

This improvement has significant implications for emergency management, as hurricane forecasts need to predict not only the center track but also the storm's intensity and the extent of dangerous wind fields.

The research is led by Google DeepMind, with collaborators including Google Research, the U.S. National Hurricane Center, Colorado State University, and the UK Met Office. The paper has been peer-reviewed and accepted by Nature, but it is not yet formally published.

The team states that the model has begun transitioning from retrospective testing to operational forecasting, but the specific deployment scope and practical application are not yet clear. Computational speed is one advantage: WN-C generates forecasts about 10 times faster than GenCast and several orders of magnitude faster than traditional numerical weather prediction systems.