In a year of extreme weather—record heat, sudden hurricanes, and surprising intensity—accurate, timely forecasts matter. Traditional computer models can take hours or days to crunch data and produce a single projected outcome.
Google DeepMind’s new AI, GenCast, instead generates multiple forecasts within minutes to predict weather up to 15 days in advance.
GenCast is a “generative” AI model, resembling systems like ChatGPT. Rather than offering just one forecast, it imagines a range of future weather scenarios—an ensemble—and estimates their likelihood.
This approach outperformed the gold-standard European Center for Medium-Range Weather Forecasts model (ENS) 90+ percent of the time, including during Typhoon Hagibis in Japan, showing a clear path several days before landfall.
Unlike physics-based methods that simulate atmospheric processes in detail (requiring enormous computing power), GenCast detects patterns from decades of global weather data. By rapidly producing numerous possible outcomes, the AI embraces the weather’s inherent uncertainty (“the butterfly effect”) rather than ignoring it. The result is faster forecasts—minutes instead of hours—and, in many cases, more accurate predictions, including for extreme weather.
Such advances could help communities better prepare for dangerous storms, facilitate planning for renewable energy like wind power, and ultimately give meteorologists and emergency agencies valuable lead time for life-saving alerts. While the AI depends on foundational climate science, DeepMind hopes further community input will improve GenCast and, in turn, benefit global weather forecasting.