New artificial intelligence model improves prediction of weather uncertainties and threats, providing faster and more accurate forecasts up to 15 days in advance
The weather affects us all – shaping our decisions, our safety and our way of life. As climate change causes more and more extreme weather events, accurate and reliable forecasts are more important than ever. However, the weather cannot be predicted perfectly, and forecasts are particularly uncertain over periods longer than a few days.
Because a perfect weather forecast is impossible, scientists and weather agencies use probabilistic ensemble forecasts, in which a model predicts a range of likely weather scenarios. Such aggregate forecasts are more useful than relying on a single forecast because they provide decision-makers with a more complete picture of possible weather conditions in the coming days and weeks and the likelihood of each scenario.
Today in the newspaper published in NatureIntroducing GenCast, our new high-resolution (0.25°) AI ensemble model. GenCast provides better forecasts of both daily weather and extreme events than the leading operating system, the European Center for Medium-Range Weather Forecasts” (ECMWF) ENS, with a maximum of 15 days' notice. We will be publishing our model code, weights and predictions to support the broader weather forecasting community.
The evolution of AI weather models
GenCast is a critical advancement in AI-based weather forecasting that builds on our previous weather model, which was deterministic and provided a single best estimate of future weather. In contrast, a GenCast forecast consists of a collection of 50 or more forecasts, each representing a possible weather trajectory.
GenCast is a diffusion model, a type of generative artificial intelligence model, that underlies recent rapid advances in image, video and music generation. However, GenCast differs from them in that it is adapted to the spherical geometry of the Earth and learns to accurately generate a complex probability distribution of future weather scenarios when the latest weather condition is given as input.
To train GenCast, we fed it four decades of historical weather data from ECMWF ERA5 archive. This data includes variables such as temperature, wind speed and pressure at various altitudes. The model learned global weather patterns with a resolution of 0.25° directly from processed weather data.
Setting a new standard for weather forecasting
To rigorously evaluate GenCast's performance, we trained it on historical weather data through 2018 and tested it on data from 2019. GenCast demonstrated better forecasting skills than ECMWF's ENS, the best operational ensemble forecast system on which many national and local decisions depend every day.
We tested both systems comprehensively, analyzing forecasts of various variables at different lead times – a total of 1,320 combinations. GenCast was more accurate than ENS for 97.2% of these targets and 99.8% for turnaround times longer than 36 hours.
