In recent tests, Google's GraphCast outperformed the European Center for Medium-Range Weather Forecasts (ECMWF) system in forecast accuracy.
Google's AI-powered GraphCast system is expected to revolutionize the meteorology industry. |
Specifically, in a study published in the journal Science, GraphCast was able to make more accurate predictions for 90% of the 1,380 parameters tested, including temperature, pressure, wind speed and direction, and humidity.
Previously, in September 2023, GraphCast predicted that Hurricane Lee would make landfall on the coast of Nova Scotia, Canada, nine days before the event occurred, while traditional meteorological forecasting tools only predicted six days in advance. In addition, they proved to be less accurate in terms of the time and location of landfall.
Research shows: “Google's GraphCast can predict hundreds of weather variables for 10 days around the world in less than a minute.”
The GraphCast model combines machine learning algorithms and “graph neural networks” (GNNs) - an architecture for processing spatially structured data.
The system is trained using over 40 years of ECMWF-archived meteorological data. GNN enables rapid forecast generation using minimal computing resources.
GraphCast’s primary mission is to predict the interactions between atmospheric conditions at different locations around the globe. But the GraphCast system is not yet capable of providing the complex information that is crucial for forecasting weather events like hurricanes.
DeepMind researchers also expressed confidence in the model’s ability to scale to different types of weather systems. A test version of GraphCast is now available on the ECMWF website.
Source
Comment (0)