Artificial intelligence (AI) weather forecasts remain experimental, but Google DeepMind’s machine-learning model, called GraphCast, has recently produced forecasts that have been better than numerical forecast models and other AI forecasts. The model is so simple that it can be run on a typical desktop computer in minutes. While still experimental, this is potentially promising to the future of AI weather forecasting, which runs thousands of times faster than the current standard numerical models, which run on supercomputers and involve complex physics.
GraphCast forecasts 11 weather variables – five close to Earth’s surface and six far above the surface, and has proven it could be useful for forecasting extreme temperatures, severe weather, tropical cyclones, and short-range rainfall. However, AI models still need numerical forecast models for training purposes—GraphCast used nearly 40 years of data for training—and it is not completely understood how these AI models actually work, so reliability is questionable. The training itself takes a lot of time and can also amplify bias.
ForecastWatch is the standard for measuring forecast accuracy and we hope Google will consider adding its GraphCast forecasts into ForecastWatch for robust accuracy comparison.
For more, see this article at Nature.com based on a November 14 article published at Science that introduces and describes GraphCast.
