Weather forecasts help the public plan for daily life and prepare for safety in case of storms and other atmospheric conditions. A reliable two-week prediction that could save government resources is a holy grail that has eluded weather scientists since the 1960s—until now. Google says its DeepMind division’s new GenCast AI weather forecast agent uses machine learning to outshine the world’s top operational ensemble systems with 15-day global weather forecasts.
GenCast uses a “probabilistic” weather model to process information faster and more accurately in predicting weather patterns, generating two-week forecasts in eight minutes. Until now, the longest reliable forecast was one week. Google’s last attempt at developing a weather AI model led to reliable estimates of 10 days. Extending the forecasts to 15 days would give the public enough room for preparation to minimize the impact of storms and harsh weather conditions on communities.
Testing GenCast Against the ECMWF EPS
The European Center for Medium-Range Weather Forecasts (ECMWF) tested GenCast against its Ensemble Prediction System (EPS). Twenty-three ECMWF member states and 12 cooperating states rely on ECMWF EPS for their weather forecasts.
In the test, both ECMWF EPS and GenCast performed 15-day forecasts on a designated set of 1,320 global wind speeds, temperatures, and other atmospheric features. Google’s AI model proved quick, consistent, and more accurate, outperforming ECMWF EPS weather predictions on 97.2 percent of all reported variable, lead time and level combinations.
“I’m a little bit reluctant to say it, but it’s like we’ve made decades worth of improvements in one year,” DeepMind’s Rémi Lam told the New York Times. “We’re seeing really, really rapid progress.” Unlike conventional weather forecasting that uses supercomputers, GenCast runs on smaller machines to process data from patterns in the atmosphere to create predictions with astonishing results.
AI has helped weather scientists accelerate the processes of analyzing and predicting patterns in the natural chaos of the atmosphere. GenCast’s forecasts are not deterministic, but probabilistic, offering only ranges of the likelihood that rain is coming to a certain area at a given time. The European Center for Medium-Range Weather Forecast has already adopted some of GenCast’s features.
Learn more about other industry’s applications for artificial intelligence technology.