Is AI better at predicting the weather than a supercomputer?

Google has put together a system based on AI that was consistently more accurate than the current gold-standard supercomputer-based system used to forecast the weather. GraphCast was trained on 40 years of weather data from satellite images, weather stations and radar. It uses that data plus the weather six hours ago plus the current weather to predict what the forecast will be in six hours. It can then build off of that in six-hour increments until it gets a forecast up to 10 days out.

GraphCast does these predictions for more than a million grid points across the surface of the Earth and at 37 different altitudes in the atmosphere. And it can run on a single Google TPU v4 machine as opposed to other current weather models that use multiple supercomputers.

When Google tested GraphCast against one of those models, the gold-standard simulation system High Resolution Forecast (HRES), it blew the competition out of the water. In under a minute GraphCast made 10-day weather predictions that were more accurate than what HRES put out on 90 percent of variables and lead times. When working specifically on forecasts for the troposphere, the lowest layer of the atmosphere and the one that affects surface weather the most, GraphCast was better than HRES 99.7 percent of the time.