Overview
- WeatherNext 2 generates hundreds of forecast scenarios from a single starting point in under a minute on one TPU, an approach that replaces hours of supercomputer runs.
- Google reports eightfold speed gains over its prior model with hourly resolution and useful skill out to roughly 15 days, including improved tropical‑storm track predictions reported to about three days.
- Forecasts now power Search, Pixel Weather, Gemini and the Google Maps Weather API, with data available in Earth Engine and BigQuery plus an early‑access program on Vertex AI for custom modeling.
- The model uses a Functional Generative Network that injects targeted randomness and trains on single‑variable “marginals” to infer coherent, physically realistic multi‑variable “joints.”
- DeepMind acknowledges challenges with rare precipitation extremes due to training data gaps, and researchers call for independent, multi‑season benchmarking as rivals like ECMWF, Nvidia, Huawei and Microsoft advance their own AI systems.