Google's GenCast AI can make fast and accurate weather predictions

Google has debuted a weather prediction AI model that can predict weather faster and more accurately than the top operational system, the European Centre for Medium-Range Weather Forecasts (ECMWF), up to 15 days in advance. 

In a paper published in the scientific journal Nature, Google argues that current prediction models rely on numerical weather prediction systems (NWP) that simulate and "approximately solve the equations that model atmospheric dynamics." 

While these provide a decent estimate on developing scenarios, each simulation must be run individually, forcing weather agencies to run multiple simulations for each possible scenario. The system is also "prone to errors" and "slow to run," making it time-consuming for engineers. 

GenCast can generate "global 15-day ensemble forecasts" that are more accurate. It also takes 8 minutes to create a single 15-day forecast, and several "forecasts can be generated in parallel." 

In contrast to Google's last forecast model, which could only do a "single, best estimate of future weather," GenCast can process 50 or more predictions at once. Google trained GenCast on four decades of weather data from ECMWF's ERA5 archive. The company says the model was fed weather patterns at 0.25° resolution, a measure of the spatial resolution in a variety of data sets. The model was trained on historical weather data up to 2018. 

GenCast proved to be 97.2% more accurate than ECMWF's systems and "on 99.8% at lead times greater than 36 hours," when tested on data from 2019.