In the past half-century, tropical cyclone disasters have resulted in an estimated 779,000 fatalities worldwide and economic damages totaling USD 1.4 trillion. Typhoons and similar disasters have consistently posed a threat to coastal regions. Thus, the ability to predict their trajectories, intensities, and wind radii earlier and with greater precision is of paramount importance. Google DeepMind, along with other collaborating institutions, has published a paper in Nature, unveiling the AI model WeatherNext Cyclones. This innovative model is capable of generating global weather and cyclone scenario forecasts extending up to 15 days, encompassing predictions for path, intensity, and wind radius. Remarkably, it achieves accuracy levels comparable to existing models an average of one day or more ahead of time. Furthermore, it can simultaneously forecast typhoon paths, intensities, and wind structures within a single framework, as well as generate multiple future scenarios to account for forecast uncertainties.

The model was trained using two comprehensive datasets: ERA5 reanalysis data from 1979 to 2018 and historical tropical cyclone data sourced from the IBTrACS database. Through end-to-end training on nearly 20TB of global atmospheric data and a repository of nearly 5,000 historical storms, the model has learned to recognize patterns in atmospheric evolution. When benchmarked against leading operational weather models, WeatherNext Cyclones exhibited consistent advantages in predicting path, intensity, and wind radius. Its intensity prediction capabilities even outperformed those of the high-resolution regional model HAFS. Moreover, integrating WeatherNext Cyclones with existing models can further enhance overall forecast performance.

The research team has made the relevant code, pre-trained weights, and a lightweight version of the model publicly accessible. However, they stress that the model is currently an experimental tool and should not be used as a substitute for official warnings.