AI is Google’s new favorite hammer and the next nail in its path is weather forecasting. The company is introducing GenCast, a “high-resolution AI ensemble model,” which was detailed in a article published in Nature.
Accurate weather forecasts are important for everything from your daily life to disaster preparedness and even renewable energy. And GenCast beats the best current system, ECMWF‘s ENS, forecast up to 25 days in advance.
GenCast is a diffusion model, similar to those you may have seen in AI image generators. However, this is specifically adapted to the geometry of the Earth. It drew on four decades of historical data from the ECMWF archives.
To test it, Google trained GenCast on historical weather data through 2018, ran 1,320 different forecasts for 2019, and compared its results to those from ENS and real weather. GenCast was more accurate than ENS 97.2% of the time, up to 99.8% more accurate for forecasts 36 hours or more ahead.
Here is a demo. Google commissioned GenCast to predict the path of Typhoon Hagibiswhich hit Japan in 2019. You can see in red the path taken by the typhoon, in blue the possible paths predicted by Google’s AI model. At 7 days apart, they are fairly spread out, but they narrow in the actual path as the typhoon gets closer to landfall.
GenCast predicts the path of Typhoon Hagibis
Giving local authorities more time to prepare for severe weather is one use case. GenCast can also predict wind speed near wind farms, weather at solar farms, and more.
GenCast is an “ensemble model,” meaning it produces more than 50 predictions with different probabilities. Such a prediction covering a 15-day period can be generated in 8 minutes on a Google Cloud TPU v5, Google says. Multiple predictions can be made in parallel. Meanwhile, a traditional weather forecasting model takes hours on a supercomputer.
Google releases GenCast as an open model and shares its code and weights. The company plans to continue to cooperate with weather forecasting agencies and scientists to further improve future forecasts.