Google has launched GenCast, a new AI-based weather forecasting tool that predicts conditions up to 25 days ahead with impressive accuracy.
Built on a "high-resolution AI ensemble model," it outperforms traditional methods like ECMWF’s ENS and could transform disaster planning and energy management.
GenCast uses a diffusion model, similar to those in AI image generators, but adapted for Earth's unique geometry.
Trained on 40 years of historical weather data, it can produce over 50 predictions for any forecast.
Each prediction is made in just 8 minutes using Google Cloud TPU v5, compared to the hours needed by older systems.
Here’s what you should know:
GenCast is 97.2% more accurate than ENS and reaches 99.8% accuracy for forecasts 36 hours or longer.
A 15-day forecast takes 8 minutes, compared to hours on supercomputers.
Google is sharing GenCast’s code and weights to work with researchers and agencies.
The overachiever of forecasts
Google tested GenCast by predicting the path of Typhoon Hagibis, which hit Japan in 2019.
The model’s accuracy improved as the storm got closer to land, giving authorities more time to prepare.
GenCast can also predict conditions near wind and solar farms, supporting better energy management.
Google is making GenCast open source, encouraging collaboration with scientists and agencies to refine and improve forecasts.
Finally, no more jokes about weather forecasts being wrong?