An AI developed by Google Deepmind manages to outperform current medium-term weather forecast models. It’s called GraphCast and it requires just a fraction of the computing power currently required to deliver precise results in record time.
This will also interest you
[EN VIDÉO] Have the weather forecasts for 2050 already come true? What could the weather forecast look like in 2050? In collaboration with the WMO, the organization…
If you find that weather forecasts are not always reliable, especially in the medium term, artificial intelligence (AI) should help you change your mind. According to a recent paper published in the journal Science, a team at GoogleGoogle Deepmind has developed a new 10-day climate prediction program called GraphCast.
GraphCast is powered by AI and almost always outperforms existing forecasting tools. Better yet, this AI only uses a fraction of the computing power of the current system (HRES). The difference is that GraphCast’s AI has absorbed decades of weather information, as well as around 40 years of data from satellites, weather stations and radars. Current processes work differently. Since enormous computing power requires hundreds of machines on a supercomputer, 10-day forecasts are made using huge databases. They contain data on thermodynamics, fluid dynamics and other criteria related to atmospheric conditions.
Example of a 10-day GraphCast forecast showing humidity, surface temperature and wind speed. © GraphCast, Deepmind
More precise and less energy consuming than the current model
Only one machine of the same type of supercomputer is required for GraphCast to provide a very precise medium-term forecast in less than a minute. Unlike the current model, GraphCast was even more accurate in over 99.7% of tests by limiting the scope of the analysis to the lowest part of the atmosphere, where the most visible weather events occur. Since its debut in September at the European Center for Medium-Range Weather Forecasts (ECMWF), GraphCast has been able to predict the exact trajectory of Hurricane Lee, which hit Nova Scotia, nine days in advance, three days less than the HRES. GraphCast is open source and its experiments continue with the ECMWF alongside the current HRES.