DeepMind’s GraphCast AI model outperforms traditional weather forecasting methods
Artificial intelligence has surpassed traditional weather forecasting methods for the first time, writes Financial Times.
It is a meteorological AI model created by Google DeepMind developers. It is called GraphCast and is capable of preparing a 10-day weather forecast in 1 minute.
Evaluation of the model showed that it is more accurate than the leading traditional forecasting system used by the European Center for Medium-Range Weather Forecasts (ECMWF).
GraphCast outperformed this system with a 3-10 day forecast in 90% of cases. For this purpose, 1380 indicators were analyzed, including temperature, pressure, wind speed and direction, and humidity.
“We find GraphCast to be consistently better than other machine learning models, Huawei’s Pangu-Weather and Nvidia’s FourCastNet, and more accurate than our own forecasting system in many respects,” the ECMWF noted.
In its work, GraphCast uses a machine learning architecture based on ECMWF data on global weather conditions over the past 40 years. As an example of a successful forecast, DeepMind mentioned Hurricane Lee in the North Atlantic in September.
The AI was able to correctly predict that Lee would reach the shores of Nova Scotia in Canada 9 days before it happened. Using traditional approaches, the calculation was made 6 days before the hurricane reached the Canadian province.