Google launches Weather Lab with an experimental AI cyclone model

Google DeepMind and Google Research launched Weather Lab, a public website for exploring the company’s experimental AI weather models, built around a new cyclone-forecasting system. Rather than simulating atmospheric physics on a supercomputer, as conventional forecast models do, the system uses a probabilistic approach that introduces random perturbations to produce a set of 50 possible trajectories for a given storm in a single step, letting forecasters see a spread of likely outcomes rather than one deterministic track.

Google said the model showed state-of-the-art accuracy on preliminary internal evaluation for both cyclone track and intensity, and cited examples of it identifying storm paths — including cyclones Jude and Ivone in the Indian Ocean — nearly a week before the storms had fully formed. The model was developed with roughly two months of feedback from trusted testers, including forecasters at the US National Hurricane Center, who worked with the DeepMind team to shape features such as an “expert mode” for exploring alternative cyclogenesis scenarios; one forecaster was quoted describing the model’s performance as potentially significant for operational practice.

Google was explicit that the tool remained experimental and was not intended to replace physics-based forecasting used operationally by national weather agencies, positioning Weather Lab instead as a public testbed feeding into that established process. The launch extended DeepMind’s earlier weather work — including its GraphCast and, later, WeatherNext models — into a specific, high-stakes application area, and continued a broader pattern across 2025 of frontier labs pairing general-purpose model releases with narrower science and public-safety applications.