Google's DeepMind Unveils WeatherNext 2, an AI Weather-Forecasting System Delivering Faster and More Accurate Global Forecasts
Google DeepMind has introduced WeatherNext 2, an advanced AI weather-forecasting system that generates global forecasts eight times faster than traditional tools. It produces hundreds of scenario options from a single starting point, with each computed in under one minute using a TPU.
WeatherNext 2 employs Functional Generative Networks (FGN), a new modeling approach detailed in a June paper. This method is trained on single-variable forecasts while learning inter-variable interactions to predict regional weather patterns and cyclone behavior. Compared to its predecessor, WeatherNext Gen, WeatherNext 2 delivers eightfold speed increases, sixfold higher resolution, and improved time-step granularity moving from six to one hour.
The system outperforms GenCast forecasts on 99.9% of tested variables, with FGN-based forecasts demonstrating average improvements of 8.7% (average-pooled CRPS) and 7.5% (max-pooled CRPS). Cyclone forecasting has also improved significantly, with ensemble-mean predictions reducing position errors by about 24 hours of lead time for three- to five-day forecasts. A 12-hour timestep variant, while less effective than the six-hour version, still outperforms GenCast beyond two days.
WeatherNext 2 is currently deployed in Google Search, Gemini, Pixel Weather, and the Maps Weather API, with plans for broader rollout. The aim is to provide faster, more accurate weather forecasts to assist agencies in managing severe weather and to support decision-making across supply chains, energy grids, and agriculture, particularly through better calibration when predictions cover larger regions.
