DeepMind develops breakthrough AI-driven weather software

[Courtesy of Science/Remi Lam et al.]

DeepMind Technologies Ltd., a subsidiary of Alphabet Inc., has developed a new artificial intelligence-driven weather software that outperforms existing weather forecast models. The new AI software reportedly predicts weather more accurately and quickly than the current weather prediction models used by meteorological agencies worldwide, forecasting the weather across the globe for the next 10 days in about a minute.

This latest development was announced by Dr. Remy Lam and the research team at DeepMind in the international academic journal “Science” on Tuesday.

The current weather prediction models generally used by meteorological agencies worldwide are based on numerical forecast models. These models calculate the upcoming weather by computing the state and movement of the atmosphere using various pieces of data that is inputted into supercomputers. But existing models fall short of being 100 percent accurate as they do not fully understand atmospheric characteristics.

In contrast to traditional numerical forecast models, DeepMind’s weather AI was trained on atmospheric variability. According to the research team, the system analyzed 40 years’ worth of weather data from 1979 to 2017 to understand weather variability.

GraphCast can predict “hundreds of weather variables over 10 days at a 0.25-degree resolution globally in under one minute,” making it approximately 1,000 to 10,000 times faster than existing models in terms of processing speeds.

The accuracy of forecasts provided by GraphCast significantly surpasses that of meteorological prediction models in various countries. For instance, in a comparison with the European Centre for Medium-Range Weather Forecasts (ECMWF) prediction model, GraphCast showed a winning rate of 99 percent after 12,000 rounds of competition.