Overview
- The model generates hundreds of scenarios from a single starting point in under a minute on one TPU, delivering eight-times faster outputs at one-hour resolution.
- A new Functional Generative Network injects targeted noise to create physically coherent ensembles, improving joint system predictions and probabilistic accuracy.
- WeatherNext 2 outperforms the prior model on 99.9% of tested variables and lead times up to 15 days, including temperature, wind, humidity and pressure.
- Forecasts now power Search, Gemini, Pixel Weather and the Google Maps Weather API, with full Maps app integration set to arrive in the coming weeks.
- Data is available in Earth Engine and BigQuery with Vertex AI early access for custom use, and research shows sharper tropical cyclone tracks that cut ensemble-mean position errors by roughly a day at three- to five-day leads, with known challenges on rare extreme rain and snow still under study.