the study
Author published on June 12, 2025
Weather Lab Team
We have launched Weather Lab, featuring experimental cyclone predictions, and partnered with the National Hurricane Center to support forecasting and warnings for this cyclone season.
Tropical cyclones are extremely dangerous, risking lives and making catastrophic communities a koichi. And over the past 50 years, they have caused economic losses of $1.4 trillion.
Also known as hurricanes and typhoons, these vast, spinning storms form over warm seawater. They are notoriously very sensitive to even small differences in atmospheric conditions and are difficult to predict accurately. However, improving the accuracy of cyclone predictions can help protect communities through more effective disaster preparation and previous evacuation.
Today, Google Deepmind and Google Research are launching Weather Lab, an interactive website for sharing artificial intelligence (AI) weather models. Weather Lab features modern AI-based tropical cyclone models based on stochastic neural networks. This model can predict cyclone formation, track, strength, size and shape. This generates 50 possible scenarios up to 15 days in advance.
Animation showing predictions from experimental cyclone models. Our model (in blue) accurately predicted the pathways of Hondo and Garance, a cyclone south of Madagascar. Our model also photographed the paths of Jude and Ibonne in the Indian Ocean cyclone. This will robustly predict storm weather areas for about seven days in the future, eventually intensifying into a tropical cyclone.
We have released a new paper describing our core weather models and provide archives at the Historic Cyclone Track Data Weather Lab for evaluation and backtesting.
Internal tests show that model predictions for cyclone tracks and strength are as accurate and often more accurate as current physics-based methods. We are partnering with the National Hurricane Center (NHC), which assesses cyclone risk in the Atlantic and Eastern Pacific Basin, to scientifically examine our approaches and outputs.
NHC expert predictors are looking at live predictions from experimental AI models along with other physics-based models and observations. We hope that this data will help improve NHC forecasts and provide early and more accurate warnings of the risks associated with tropical cyclones.
Weather Lab Live and Historical Cyclone Prediction
Weather Lab shows live and historical cyclone predictions for a variety of AI weather models, along with physics-based models for European Medium-Range Weather Forecast (ECMWF). Some of the AI weather models are run in real-time on Weathernext graphs, Weathernext Gen, and the latest experimental cyclone models. It also launches Weather Lab with over two years of historical predictions for download and analysis by experts and researchers, allowing external evaluation of models on all seafloors.
An animation showing the model’s predictions for Cyclone Alfred when it was a Category 3 Cyclone in the Coral Sea. The model’s ensemble average forecast (Bold Blue Line) correctly predicted a rapid weakening of Cyclone Alfred, which was rapidly weakened by tropical storms near Brisbane, Australia, seven days later and a final landfall.
Weather Lab users can explore and compare predictions from a variety of AI and physics-based models. Read together, these predictions will help weather agencies and emergency services experts better predict the path and strength of the cyclone. This could help experts and decision makers prepare for a variety of scenarios, share news of the risks involved, and support decisions to manage the impact of the cyclone.
It is important to emphasize that Weather Lab is a research tool. The live predictions shown are generated by the model still being developed and are not official warnings. Keep this in mind when using tools, such as supporting decisions based on predictions generated by Weather Lab. For official weather forecasts and warnings, please refer to your local weather agency or the National Weather Service.
AI-driven cyclone prediction
In physical-based cyclone predictions, the approximations required to meet operational demands mean that it is difficult for a single model to excel at predicting both the tracks of a cyclone and its strength. This is because cyclone trucks are dominated by vast atmospheric steering currents, while cyclone strength depends on complex turbulent processes within and around compact cores. Global low-resolution models are best for predicting cyclone tracks, but do not capture the fine scale processes that determine cyclone intensity. Therefore, a high-resolution model for the region is required.
The experimental cyclone model is a single system that overcomes this trade-off, with internal evaluations showing cutting-edge accuracy of both cyclone tracks and strength. They are trained to model two different types of data. It is a specialized database that contains a vast reanalytical dataset that reconstructs the global past weather from millions of observations, as well as important information on tracks, strength, size and wind radii for around 5,000 cyclones over the past 45 years.
Modeling analytical and cyclone data together greatly improves cyclone prediction capabilities. For example, the initial assessment of observed hurricane data for the NHC showed that in 2023 and 2024 tests in the North Atlantic and Eastern Pacific Basin, 5-day cyclone track predictions averaged closer to true cyclone positions of 140 km than the ENS, the leading global physics-based ensemble model from the ECMWF. This is comparable to the accuracy of ENS’s 3.5-day predictions. This is a 1.5 day improvement that usually took over 10 years.
Previous AI weather models struggled with calculating cyclone strength, but the experimental cyclone models exceeded the average intensity error of the National Oceanic and Atmospheric Administration (NOAA) Hurricane Analysis and Prediction System (HAFS). Preliminary tests also show that model predictions for size and wind radius are comparable to physics-based baselines.
Here we visualize track and intensity prediction errors compared to ENS and HAF, and present the results of the average performance assessment of the experimental cyclone model up to 5 days ago.
Evaluation of track and intensity predictions of experimental cyclone models compared to the main physics-based models ENS and HAFS-A. In our assessment, we use the NHC best track as ground truth and follow a homogeneous verification protocol.
More useful data for decision makers
In addition to the NHC, we have worked closely with the Atmospheric Cooperative Institute (CIRA) at Colorado State University. Dr. Kate Musgrave, a research scientist at CIRA, and her team evaluated our model and found that they had the skills to “compared or better than the operating model that is best suited to track and strength. Musgrave said, “We look forward to seeing results from real-time forecasts during the 2025 hurricane season.” We are also working with experts from the UK’s Met Office, the University of Tokyo, Weathernews Inc. in Japan, and other experts to improve our models.
Our new experimental tropical cyclone model is the latest milestone in a series of pioneering Weathernext research. By responsibly sharing AI weather models through Weather Lab, we continue to collect important feedback from weather agents and emergency services experts on how our technology can improve official forecasts and inform life-saving decisions.
Acknowledgments
This research was jointly developed by Google Deepmind and Google Research.
We would like to thank collaborators NOAA’s NHC, CIRA, the Met Office in the UK, the University of Tokyo, Weathernews Inc. in Japan, Fox Weather Bryan Norcross, and other trusted tester partners who have shared invaluable feedback throughout the development of Weather Lab.
