Google’s new AI weather model, GenCast, offers state-of-the-art forecasting accuracy using a diffusion model adapted to Earth’s spherical geometry. Unlike its predecessor, GraphCast, which provided a single prediction, GenCast generates an ensemble of 50+ forecasts to capture uncertainty and predict multiple possible weather scenarios.
GenCast outperforms traditional systems like ECMWF’s ENS, providing more accurate forecasts for day-to-day and extreme weather up to 15 days in advance. It excels at predicting tropical cyclones, extreme temperatures and high winds, creating forecasts in just 8 minutes on a Google Cloud TPU v5 compared to hours on supercomputers.
Now open-source, GenCast’s code and weights are available for integration into research and modeling workflows, with real-time and historical forecasts coming soon.
Source: 9to5Google
