For the first time, artificial intelligence (AI) was able to predict the weather much more accurately than meteorologists using a traditional system of collecting and analyzing information.
According to a study published in the journal Science, Google DeepMind’s GraphCast neural network outperformed the European Center for Medium-Range Weather Forecasts in forecast accuracy.
GraphCast can identify hazardous weather events without even being trained to find them. With the integration of a simple cyclone tracker, the model predicts cyclone movement more accurately than the HRES method. With the climate becoming increasingly unpredictable, timely and accurate forecasts will be critical when planning to deal with the threat of natural disasters.
GraphCast produces forecasts at a resolution of 0.25° latitude and longitude. In other words, the Earth is divided into a million areas, each of which produces a forecast with five variables on the Earth’s surface and six atmospheric indicators that cover the planet’s atmosphere in three dimensions at 37 levels. Variables include temperature, wind, humidity, precipitation and sea level pressure. Geopotential is also taken into account – gravitational potential energy per unit mass at a specified point relative to sea level. In testing, the GraphCast model outperformed the most accurate deterministic systems by 90% on 1,380 test objects. In the troposphere, the lower layer of the atmosphere, GraphCast forecasts were more accurate than HRES for 99.7% of test variables. At the same time, the model demonstrates high efficiency: a ten-day forecast is completed in less than a minute on one Google TPU v4 machine, while the traditional approach requires several hours of supercomputer work with hundreds of machines.
In the study, the DeepMind team said its model should not be seen as a replacement for traditional forecasting models, but rather as proof that AI-based weather forecasting systems have the “potential to complement and improve upon currently existing best practices.”