Google has introduced an experimental AI model for predicting tropical cyclones, developed in collaboration with DeepMind and Google Research. The company launched an interactive site, Weather Lab, where users can explore how the latest AI models compare to traditional physical approaches to weather forecasting. The model’s feature is its ability to create fifty different cyclone development scenarios, considering its trajectory, size, and intensity up to fifteen days in advance.
Google is collaborating with the National Hurricane Center (NHC) in the USA, which evaluates the model’s effectiveness based on real storm observations in 2023 and 2024. In these tests, the five-day cyclone trajectory forecast was on average one hundred forty kilometers closer to the actual hurricane path than predictions from the well-known ENS physical model by ECMWF. This is a significant improvement, usually achieved only after decades of gradual changes in classical methods.
The Google model is trained on massive ERA5 archives, containing hundreds of millions of observations from around the world, combined with forecasts from traditional models. It can not only predict the cyclone’s appearance but also assess its intensity and the likelihood of landfall, as was the case during the analysis of Cyclone Alfred, for which the model predicted weakening and the approach to the Australian coast a week in advance.
Google is engaging experts from the University of Colorado, the UK Met Office, the University of Tokyo, and the Japanese company Weathernews Inc. to improve its models. At the same time, the company emphasizes that Weather Lab is currently only a research tool and is not intended for use as an official source of forecasts for the general public.
