Google DeepMind unveils AI Weather Lab and new cyclone forecasting model
text_fields
Google DeepMind, in collaboration with Google Research, has launched a public preview of Weather Lab, an online platform designed to showcase the company’s artificial intelligence (AI)-driven weather forecasting models.
This includes a newly developed experimental model aimed specifically at predicting tropical cyclones.
While promising in its capabilities, the company clarified that the Weather Lab is a research initiative and not intended for issuing official weather alerts.
The AI-powered cyclone model, which remains under scientific review, is designed to forecast the formation, path, strength, size, and shape of tropical cyclones as far as 15 days in advance. DeepMind noted that a preprint of its research paper has been made available and that validation is underway in partnership with the U.S. National Hurricane Center (NHC).
According to a blog post from DeepMind, Weather Lab integrates various models—including WeatherNext Graph, WeatherNext Gen, and the new cyclone model—to provide real-time analysis and predictions. The platform also includes more than two years of archived AI forecasts, allowing researchers to study and evaluate the performance of these systems. Users can compare AI-generated outputs with traditional physics-based forecasts, such as those from the European Centre for Medium-Range Weather Forecasts (ECMWF).
In traditional forecasting, cyclone tracking and intensity assessments are split between two distinct physics-based models—one for low-resolution global tracking and another for high-resolution regional intensity analysis. The AI model, however, addresses both aspects within a single system.
DeepMind said the model was trained using a combination of reanalysis datasets and a specialised archive of nearly 5,000 cyclone events over the past 45 years. During recent testing across the North Atlantic and East Pacific regions (2023–2024), the AI model’s five-day forecast placed cyclone locations about 140 km closer to their actual paths than predictions by ECMWF’s ENS model.
The company claims that, based on internal evaluations, its AI model performs at least on par with leading physics-based systems.
