WeatherNext 2: Our most advanced weather forecasting model (blog.google)

Google DeepMind and Google Research have released WeatherNext 2, a major upgrade to their AI weather-forecasting suite that generates hundreds of coherent scenarios up to 8× faster and with temporal resolution down to one hour. The model is already powering Google products (Search, Gemini, Pixel Weather) and the Maps Platform Weather API, and forecast data is available in Earth Engine and BigQuery; an early-access inference option is on Vertex AI. Experimental cyclone predictions and multi-scenario outputs aim to give agencies and businesses actionable worst-case planning rather than a single deterministic forecast. Technically, WeatherNext 2 introduces a Functional Generative Network (FGN) approach that injects noise in function space and trains independent neural nets to produce physically realistic, interdependent forecast ensembles. Crucially, the model is trained only on marginals (individual variables like temperature, wind, humidity) yet learns to produce skillful joint predictions (spatially coherent systems such as heat waves or wind-farm outputs). Each ensemble member can be produced in under a minute on a single TPU—where physics-based models would take hours—and the model outperforms the prior WeatherNext on 99.9% of variables and lead times (0–15 days), improving probabilistic accuracy and practical decision-making for weather-sensitive operations.