The Google DeepMind team presented GenCast, an advanced artificial intelligence model capable of forecasting weather and extreme weather events with high accuracy. According to the company, this technology provides the world's best forecasts for up to 15 days.
GenCast is a diffusion model, a type of generative artificial intelligence used to create images, videos, and sounds. It is adapted to the spherical geometry of the Earth and learns to generate probable future weather scenarios, given actual weather conditions as input.
To train the model, Google used four decades of meteorological data from the ERA5 archive created by the European Center for Medium-Range Weather Forecasts (ECMWF), up to and including 2018. The archive contains information on variables such as temperature, wind speed, and atmospheric pressure at different altitudes.
Google conducted a test comparing GenCast with ESN, the best integrated forecasting system from ECMWF, which is used daily by many national and local institutions. The results of a comprehensive test of forecasts of various variables for different periods, 1320 combinations in total, showed that GenCast was 97.2% more accurate than ESN, or 99.8% for forecasts over 36 hours.
Another advantage of GenCast compared to other weather forecasting systems is that it can produce a 15-day forecast in just 8 minutes. According to the company, this requires only one Tensor Processing Unit (TPU) v5. A TPU is an integrated circuit specially designed by Google for neural network machine learning. While traditional systems would take hours even on a supercomputer with thousands of processors to make the same prediction.
Google also published a scientific paper in the journal Nature. In addition, GenCast is an open-source model, and the company has shared it on GitHub. This model will also be used to inform users in Search and Maps.