- DeepMind’s machine learning model takes less than a minute to predict future climate around the world more accurately than other approaches
Google’s artificial intelligence division, DeepMind, has managed to create a machine learning model that allows weather forecasting in a matter of minutes, instead of hours like conventional models.
The best thing about GraphCast is that it can be run simply from a desktop PC, making more accurate predictions from home.
GraphCast outperforms conventional and AI-based approaches in most global weather forecasting tasks.
The researchers first trained the model using estimates of past global climate made between 1979 and 2017 using physical models. This allowed GraphCast to learn links between climate variables such as air pressure, wind, temperature and humidity.
Making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine. In comparison, a 10-day forecast using a conventional approach, such as HRES, can require hours of calculation on a supercomputer with hundreds of machines.
In a comprehensive performance evaluation against the standard deterministic system, GraphCast provided more accurate predictions on more than 90% of 1,380 test variables and predicted delivery times
The European Center for Medium-Range Weather Forecasts (ECMWF) is already using GraphCast, and other meteorological agencies are also developing their models based on the graph neural network architecture proposed by Google.
In the study published in the journal Science, GraphCast predicted the state of 5 meteorological variables close to the Earth’s surface, such as air temperature 2 meters above the ground, and 6 atmospheric variables, such as wind speed, further away from the surface of the Earth.
It was also useful in predicting severe weather events, such as the path of tropical cyclones and extreme heat and cold events.