In less than a month of training, Google DeepMind's new AI model, GraphCast, has already shown it can predict atmospheric conditions up to ten days in advance with precision surpassing ECMWF, the European numerical weather prediction agency.
This performance boost, achieved through a graph-based neural network that exploits Copernicus satellite data and ground observations, could transform how weather services plan their alerts, bulletins, and backup models.
What the study reveals: GraphCast's performance and reach
The study published by DeepMind shows that GraphCast maintains a statistical superiority over ECMWF for 10-day forecasts of key variables – 2m temperature, precipitation, surface wind, and atmospheric pressure. On validation sets, GraphCast's mean squared error is reduced by several points compared to weather models, resulting in more reliable forecasts for medium-amplitude phenomena.
The model covers the entire globe with a spatial resolution of 0.25° and a time step of three hours, allowing observation of synoptic developments with a level of detail usually reserved for regional models. This granularity, combined with the network's ability to learn non-linear interactions between distant points, improves representation of cold fronts, tropical cyclones, and low-pressure systems.
Moreover, GraphCast demonstrates increased resilience to rare events – such as heatwaves or heavy rainfall – where classical models tend to underestimate intensity. According to weather models, the ability to anticipate these extremes could reduce forecast uncertainties by several days, providing additional lead time for decision-makers.