Google DeepMind, the leading artificial intelligence company, has made a groundbreaking development in weather forecasting. Their newly developed machine learning algorithm, GraphCast, claims to accurately predict weather conditions faster and more precisely than traditional methods, including those used by supercomputers. […]
Google DeepMind, the leading artificial intelligence company, has made a groundbreaking development in weather forecasting. Their newly developed machine learning algorithm, GraphCast, claims to accurately predict weather conditions faster and more precisely than traditional methods, including those used by supercomputers.
GraphCast has surpassed the High Resolution Forecast (HRES) system, managed by the European Center for Medium-Range Weather Forecasts (ECMWF), by generating more accurate ten-day weather forecasts within minutes instead of hours. In fact, GraphCast outperformed the ECMWF in over 99% of the weather variables across 90% of the 1,300 test regions, according to a study published in the prestigious journal Science on November 14th.
However, researchers point out that GraphCast is not perfect. The algorithm operates as a black box, meaning that the artificial intelligence cannot explain its decision-making process or demonstrate its workings. It should be regarded as a complementary tool rather than a replacement for existing forecasting methods.
Traditionally, weather predictions relied on inputting data into complex physical models and utilizing supercomputers for simulations. These predictions heavily depend on intricate details within the models, making them computationally intensive and costly.
However, machine learning-based weather models have the potential to provide more cost-effective solutions as they require less computing power and deliver faster results. For the development of GraphCast, researchers trained the algorithm using 38 years of global meteorological data up until 2017. The algorithm identified patterns between variables such as air pressure, temperature, wind, and humidity, even some that were not fully understood by the researchers themselves. After training, the model generated accurate ten-day forecasts in less than a minute using global estimates from meteorological data in 2018. When tested alongside the ECMWF’s high-resolution forecast, which employs conventional physical models, GraphCast consistently provided more accurate predictions in over 90% of the 12,000 data points used.
Another impressive feature of GraphCast is its ability to forecast extreme weather events, including heatwaves, cold spells, and tropical storms. When the highest layers of Earth’s atmosphere were removed from the meteorological data, the accuracy of GraphCast in predicting these extreme events surged to over 99%.