We run Weather Lab with our experimental cyclone forecasts, and cooperate with the National Hurricane Center to support their forecasts and warnings this cyclone season.
Tropical cyclones are extremely dangerous, life -threatening and destructive communities. And in the last 50 years they have caused USD 1.4 trillion of economic losses.
These extensive, rotating storms, also known as hurricanes or typhuna, form over warm ocean waters – powered by heat, moisture and convection. They are very sensitive to even small differences in weather conditions, which makes them extremely difficult to forecast accurately. However, improving the accuracy of cyclone forecasts can help protect community through more effective readiness for a disaster and earlier evacuation.
Today, Google Deepmind and Google Research run Weather Lab, an interactive website to share our weather models of artificial intelligence (AI). Weather Lab contains our latest experimental tropical cyclone based on artificial intelligence, based on stochastic neural networks. This model can predict the formation, path, intensity, size and shape of the cyclone – generating 50 possible scenarios, up to 15 days.
Animation showing the forecast from our experimental cyclone model. Our model (in blue) thoroughly predicted the paths of Honde and Garance cyclones, south of Madagascar, during their activity. Our model also captured the paths of Jude and Ivone cyclones in the Indian Ocean, almost seven days in the future, solidly anticipating the areas of storm weather that would eventually intensify in the tropical cyclones.
We published new paper Describing our basic weather model and provides the archive in the Weather Historical Cyclone Track laboratory to assess and reverse test.
Internal tests show that the forecasts of our model regarding the path and intensity of the cyclone are so accurate and often more accurate than more accurate thanIN Current methods based on physics. We work with the National Hurricane Center (NHC) who assess the risk of cyclone in the Atlantic Pool and the Eastern Pacific to scientifically confirm our approach and results.
NHC expert forecasts now see live forecasts from our experimental AI models, as well as other models and observations based on physics. We hope that these data will help improve NHC forecasts and provide earlier and more accurate warnings regarding threats related to tropical cyclones.
Vivid and historical forecasts of the Weather Lab cyclone
Weather Lab shows live and historical cyclone forecasts for various AI weather models, as well as models based on physics from the European Medium range forecasts (ECMWF). Several of our AI weather models operate in real time: Weathernext chart, Weathernext Gen and our latest experimental cyclone model. We also launch the Vetota laborator with over two years historical forecasts for experts and researchers for download and analysis, enabling external assessments of our models in all ocean pools.
Animation showing the anticipation of our model for Alfred's cyclone when it was a cyclone of category 3 in the coral sea. Model Model Model set (bold blue line) correctly anticipated the rapid weakening of Alfred cyclone to the status of a tropical storm and final landing near Brisbane in Australia, seven days later, with high probability of landing somewhere along the coast of Queensland.
Weather Lab users can explore and compare forecasts from various AI and physics models. After reading these forecasts, we can help weather agencies and emergency experts better to predict the path and intensity of the cyclone. This can help experts and decision -makers better prepare for various scenarios, share the risk of risk and support decisions regarding the management of cyclone influence.
It is important to emphasize that Weather Lab is a research tool. The expected forecasts are generated by models are still being developed and are not official warnings. Remember this when using the tool, including supporting decisions based on forecasts generated by Weather Lab. Official weather forecasts and warnings can be found at the Local Meteorological Agency or the National Weather Service.
Cyclone forecasted AI
In predicting a cyclone based on physics, the approximations required to meet the operational requirements mean that a single model is difficult to predict both the cyclone track and its intensity. This is due to the fact that the cyclone path is regulated by extensive atmospheric control currents, while the intensity of the cyclone depends on the complex turbulent processes within its compact core and around it. Global low resolution models work best in predicting cyclone tracks, but do not capture the dictation processes of cyclone intensity, which is why regional high -resolution models are needed.
Our experimental cyclone model is a single system that overcomes this compromise, and our internal assessments show the latest accuracy for both the cyclone path and intensity. It is trained for modeling two separate types of data: a huge set of re -analysis data, which reconstructs the weather in the past all over the Earth from millions of observations, and a specialized database containing key information about the track, intensity, size and rays of the wind almost 5,000 observed cyclones from the last 45 years.
Modeling of data analysis and data data significantly improves the possibilities of cyclone forecasting. For example, our preliminary assessments of the NHC Hurricane Data, in the 2023 and 2024 test years, in the North Atlantic and Eastern Pacific, showed that the 5-day forecast of the cyclone tracks in ECMWF is on average 140 km closer to the real location of the cyclone than Ens-Global Model of Physics. This is comparable to the accuracy of the 3.5-day ENE-1.5-day improvement, which it usually took For over a decade to achieve.
While the previous AI weather models fought for the calculation of the intensity of the cyclone, our experimental cyclone model exceeded the average intensity error of the national ocean and atmospheric administration (Noaa) A hurricane analysis system and forecasts (Sea), leading regional model of high resolution physics. Preliminary tests also show that forecasts of the size and rays of the wind of our model are comparable to physics based on physics.
Here, we visualize the errors of forecasting path and intensity and show the results of the average performance assessment of our experimental cyclone model up to five days earlier, compared to ENS and HAF.
Assessment of the forecasts of our and intensity of our experimental cyclone model compared to leading models based on ENS and HAFS physics. Our assessments use the best NHC bags as a ground truth and are in line with their homogeneous verification protocol.
More useful data for decision -makers
In addition to NHC, we worked closely with the Cooperative Institute for Research in the Atmosphere (DEFER) at Colorado State University. Dr. Kate Musgrave, a scientist from CIRA and her team, assessed our model and said that he has “comparable or greater skills than the best operational models for tracking and intensity.” Musgrave said: “We are looking forward to confirmation of these results from real -time forecasts during the 2025 Hurricane season.” We also worked with UK Met OfficeIN University of TokyoJapan Weathernews Inc. And other experts to improve our models.
Our new experimental model of tropical cyclone is the latest milestone in our series of pioneering Weathernext tests. By sharing our AI weather models via Weather Lab, we will continue to gather important feedback from weather agency and emergency experts on how our technology can improve official forecasts and inform about decisions regarding life saving.
Thanks
These studies were developed by Google Deepmind and Google Research.
We would like to thank our colleagues of NHC NHC Noaa, Cira, British Met office, University of Tokyo, Japanese Weathernews Inc., Bryan Norcross from Fox Weather and other trusted partners of testers who shared invaluable feedback during the development of Weather Lab.
