August 6, 2026 Science
The WeatherNext team
WeatherNext enables highly accurate cyclone forecasting and can provide an extra day of lead time for warnings. Today, we are open-sourcing the model.
Forecasting how dangerous cyclones will develop has long been a challenge, because every hour matters. Tropical cyclones—also known as hurricanes or typhoons—are among the most destructive weather phenomena on Earth. Over the past 50 years, tropical cyclones have killed more than 700,000 people worldwide and caused $1.4 trillion in economic damage. For forecasters, issuing timely and accurate warnings has always been a race against the clock.
Today, in a paper published in Nature, we demonstrate that the WeatherNext AI model achieves industry-leading accuracy in forecasting cyclone tracks, intensity, and wind-field structure. On average, our model gives forecasters an additional day of useful lead time: its three-day forecasts are as effective as the two-day forecasts produced by previous models. This level of improvement is roughly equivalent to a decade of progress in meteorology.
This collaborative research brought together AI researchers and engineers from Google DeepMind and Google Research, as well as expert forecasters from the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and meteorological agencies around the world.
Our research has already had real-world impact. During the 2025 hurricane season, our model helped the NHC make a historic forecast for Hurricane Melissa, accurately predicting the storm’s rapid intensification and landfall in Jamaica. This enabled the NHC to issue timely warnings and gave local teams critical time to prepare. This year, we are continuing our collaboration and can now generate 1,000 possible scenarios for each cyclone to support forecasters in their decision-making.
Weather affects everyone. Given its far-reaching impact, we are now open-sourcing the WeatherNext 2 and WeatherNext Cyclones models used during hurricane season. By making this technology publicly available, we hope to empower the research community and expand AI’s impact in building more resilient communities—whether by giving local forecasters the tools they need to prepare for natural disasters, supporting the growth of renewable energy, or helping anticipate extreme weather.
How WeatherNext forecasts weather and cyclones
Starting from the global atmospheric conditions during Hurricane Milton in October 2024, WeatherNext Cyclones iteratively forecasts global weather patterns and fine-scale cyclone tracks, with a lead time of up to 15 days. Running an ensemble forecast with 1,000 members generates localized probability maps for wind speeds ranging from tropical-storm strength to hurricane strength.
Cyclone forecasting typically involves a trade-off between two fundamentally different modeling techniques. A cyclone’s track—where it will go—is guided by atmospheric flows on a massive global scale, which have traditionally been best modeled using lower-resolution global models. Its intensity—how strong it will become, however—is driven by highly localized, fine-scale thermodynamic processes around its core, which are best modeled using specialized high-resolution regional models.
Our WeatherNext models bridge this gap, improving both overall global weather forecasting and cyclone forecasting. It is a single AI model capable of forecasting tropical cyclone tracks, intensity, and wind-field structure with industry-leading accuracy. This breakthrough is enabled by its unique training method, model architecture, and approach to processing low-resolution inputs.
We evaluated WeatherNext Cyclones using historical cyclones from 2023 to 2024 and benchmarked its deterministic and probabilistic forecasting performance against other leading weather models. On average, WeatherNext Cyclones provides more than a full additional day (24 hours) of lead time when forecasting cyclone tracks, intensity, and wind-field structure.
The model is jointly trained on two different data modalities: global weather dynamics data and expert-curated historical cyclone observations. Through end-to-end training on nearly 20 TB of global atmospheric data and the historical IBTrACS database covering nearly 5,000 historical storms, the model learned to represent complex atmospheric patterns and extreme weather.
Cyclone forecast accuracy has steadily improved over the past several decades. The figure shows the three-day forecast accuracy of ECMWF-ENS track forecasts (a) and HWRF intensity forecasts (b) in past years, as well as how WeatherNext Cyclones achieves a step change in track and intensity forecast accuracy. Based on the trend over the past 20 years, this improvement is equivalent to a decade of progress.
Our model uses Functional Generative Networks (FGN) to efficiently generate multiple distinct forecasts, capturing the inherent uncertainty of weather. We can now generate a single 15-day forecast on a TPU in under a minute, enabling forecasters to quickly assess the probability distribution of potentially catastrophic tail risks. Last year, our system generated 50 forecasts at a time, comparable in scale to global physics-based models. This year, we expanded the ensemble to 1,000 members, enabling it to capture rare but high-impact scenarios such as rapid intensification events—one of which occurred during Hurricane Melissa in 2025.
Until now, extremely high spatial resolution has generally been considered the primary factor in improving intensity forecast accuracy. However, WeatherNext Cyclones requires data at a resolution of just 28 × 28 kilometers—100 times lower than that of traditional models. A smaller version of the model, WeatherNext 2-mini, operates at an even lower resolution of 111 × 111 kilometers and still demonstrates excellent performance. This surprised scientists; fully understanding why our model can generate such accurate forecasts at this resolution remains an open research question. We hope to work with the research community to find the answer.
Making WeatherNext available to the research community
Alongside the publication of the Nature paper, we are open-sourcing the code and model weights, which anyone can use and build upon for free. This includes academic research, operational forecasting, and the development of more specialized and localized models. We hope to accelerate progress across the global weather research community, helping meteorological agencies, researchers, and nonprofit organizations better forecast a wide range of weather events and make critical decisions that protect lives and infrastructure.
We will also release two related sets of models: WeatherNext Cyclones, which was run during hurricane season and whose results are presented in the paper; and the subsequent updated version, WeatherNext 2, which we put into operational use in October. We will also release WeatherNext 2-mini, a compact version of the model that can run on a single TPU in a free public Colab notebook.
You can explore