WeatherNext: AI Model Achieves Breakthrough in Cyclone Forecasting

Cyclone forecasting is a long-standing challenge where every hour counts. Tropical cyclones have caused over 700,000 deaths and $1.4 trillion in economic losses globally in the past 50 years. In a paper published in Nature, Google DeepMind demonstrates that WeatherNext achieves state-of-the-art accuracy in predicting cyclone track, intensity, and wind structure. On average, the model provides forecasters with an extra day of predictive accuracy: its 3-day forecasts are comparable to the 2-day forecasts of previous models, an improvement roughly equivalent to a decade of meteorological progress.

This breakthrough was achieved through unique training, architecture, and handling of low-resolution inputs. The model was trained end-to-end on nearly 20 TB of global atmospheric data and the IBTrACS database containing nearly 5,000 historical storms, learning complex atmospheric patterns and extreme weather modeling.

The research has already had real-world impact. During the 2025 hurricane season, WeatherNext helped the U.S. National Hurricane Center (NHC) make a historic forecast for Hurricane Melissa, predicting its rapid intensification and landfall in Jamaica, enabling early warnings that bought critical preparation time for ground teams. This year, the team continues to collaborate, now predicting 1,000 possible scenarios per cyclone to support forecasters' decisions.

This collaboration brings together AI researchers and engineers from Google DeepMind and Google Research, along with expert forecasters from the U.S. National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and meteorological agencies worldwide.

Traditionally, cyclone forecasting required a trade-off between two different modeling techniques: track is driven by large-scale global atmospheric circulation, suited to coarser global models, while intensity is driven by local fine-scale thermodynamic processes near the core, suited to high-resolution local models. WeatherNext bridges this gap with a single AI model, improving forecast accuracy for both global weather and cyclones.

WeatherNext Cyclones requires only 28x28 km resolution, 100 times coarser than traditional models, yet performs exceptionally well. The smaller version, WeatherNext 2-mini, also performs well at an even coarser resolution of 111x111 km. This phenomenon has surprised scientists and remains an open research question.

The model uses Functional Generative Networks (FGNs) to efficiently generate ensembles of different forecasts, capturing the inherent uncertainty of weather. Now, generating a single 15-day forecast on a TPU takes less than a minute, enabling forecasters to quickly assess the probability distribution of potentially catastrophic tail risks. Last year, the system generated 50 forecasts at a time, comparable to global physical models; this year, the ensemble size has expanded to 1,000 members to capture rare but important scenarios, such as the rapid intensification event during Hurricane Melissa in 2025.

Google DeepMind is open-sourcing WeatherNext 2 and WeatherNext Cyclones models, along with code and weights, for anyone to use freely, including for academic research, business forecasting, or developing more specialized local models. Additionally, WeatherNext 2-mini, a compact version that runs on a single TPU, is released with a free public Colab notebook.

The Weather Lab interface has been refreshed, expanding global weather forecasts and cyclone track visualizations, allowing users to view predictions for temperature, precipitation, wind speed, and more. Weather Lab and WeatherNext models are part of Google Earth AI.