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 is detailed in a paper published in Nature.
Accurate weather forecasting is important for anything from your daily life to disaster preparedness and even renewable energy. And Gencast outperforms the current top system, ECMWF’s ENS, in forecasts up to 25 days in advance.
Gencast is a diffusion model, similar to what you may have seen in AI image generators. However, it is specifically tuned to Earth’s geometry. It was trained on four decades of historical data from ECMWF’s archives.
To test this, Google trained Gencast on historical weather data dating back to 2018 and ran 1,320 different forecasts for 2019 and compared its output to ENS and actual weather. Gencast was more accurate than ENS in 97.2% of the cases, up to 99.8% for forecasts 36 hours ahead or longer.
Here is a demo. Google tasked Gencast with predicting the path of Typhoon Hagibis that hit Japan in 2019. You can see the path taken by the storm in red, with the possible path predicted by Google’s AI model in blue. Over a 7-day period, they have spread out significantly, but as the storm approaches landfall, they narrow in on the actual path.
Gencast is predicting the path of Typhoon Hagibis
Giving local officials more time to prepare for severe weather is one use case. Gencast can also predict wind speed near wind farms, weather at solar farms, etc.
Gencast is an “ensemble model”, meaning it makes 50+ predictions with different probabilities. Google says that one such forecast with a 15-day forecast can be generated in 8 minutes on Google Cloud TPU v5. Multiple predictions can be made in parallel. Meanwhile, a traditional weather forecast model takes hours on a supercomputer.
Google is releasing Zencast as an open model and sharing its code and weights. The company plans to continue collaborating with weather forecasting agencies and scientists to further improve future forecasts.
