Google Launches AI-Powered Weather Lab, Releases Experimental AI Cyclone Model

Google DeepMind and Google Research launched a public preview of Weather Lab on Thursday. It is an interactive website where the company will share its artificial intelligence (AI) weather models and share weather predictions based on their output. The Mountain View-based tech giant has also released its latest experimental AI-based tropical cyclone model. This model is said to be able to predict a cyclone's formation, track, intensity, size, and shape up to 15 days in advance. Notably, the company says a scientific validation of the AI model is currently pending.

Google Releases New AI Model to Predict Cyclones

In a blog post, DeepMind announced the launch of the new Weather Lab website and detailed its new cyclone-focused AI model. The website shows live and historical cyclone predictions using both AI weather models and physics-based models from the European Centre for Medium-Range Weather Forecasts (ECMWF).

Google DeepMind highlighted that on the website, several AI models, such as the WeatherNext Graph, WeatherNext Gen, and the new cyclone model, run in real-time to analyse weather data and make predictions. Additionally, Weather Lab also contains more than two years of historical AI-generated predictions that researchers can download to evaluate the efficiency of the models.

Weather Lab also allows users to compare predictions from different AI and physics-based models. Notably, the company emphasises that the website is a research tool and is not meant to provide official warnings.

Coming to the new AI-based cyclone model, Google has published a pre-print version of its paper. However, it is yet to be peer reviewed. For scientific validation from the research community, Google has partnered with the US National Hurricane Center (NHC).

DeepMind says that in traditional cyclone prediction, two different physics-based models are used. A global low-resolution model predicts cyclone tracks, which requires analysing the atmospheric steering currents, whereas a regional high-resolution model is used to track a cyclone's intensity, which requires observing the complex turbulent processes within and around its compact core.

The new AI model is said to solve this dual-approach problem by unifying both cyclone track and intensity prediction. As per the post, the model is trained on both the “reanalysis dataset that reconstructs past weather over the entire Earth from millions of observations, and a specialised database containing key information about the track, intensity, size and wind radii of nearly 5,000 observed cyclones from the past 45 years.”

Highlighting an example, DeepMind said that the model was deployed in the North Atlantic and East Pacific basins between 2023-24 for testing, and during the time, its five-day cyclone track prediction was, on average, 140km closer to the true location compared to the prediction of ECMWF's ENS model. Additionally, the company claimed that the cyclone model's results, based on internal testing, are at least on par with physics-based models.