Improved weather forecasts
At the moment we are deep in the clutches of Storm Darragh as it batters the UK. We have currently lost one fence panel and there may be more casualties. I looks like I will get to experience the “joy” soon of watching my son play football in high winds and torrential rain.
The British are well known for talking about the weather. It is a common way to start a conversation, especially with people you don’t know well. This may be down to how different the weather can be across such a relatively small country.
In a recent article, the researchers at Google DeepMind have documented the results they have had making weather forecasts compared to the current approach of using supercomputers.
They trained their model on weather data up to 2018 and then compared its forecast for 2019 with the actual weather data. They found it was more accurate than the supercomputer predictions 97.2 percent of the time and it produced the results much quicker.
It takes a single Google Cloud TPU v5 just 8 minutes to produce one 15-day forecast in GenC ast’s ensemble, and every forecast in the ensemble can be generated simultaneously, in parallel. Traditional physics-based ensemble forecasts such as those produced by ENS, at 0.2° or 0.1° resolution, take hours on a supercomputer with tens of thousands of processors.
It has also had good results predicting extreme events such as cyclones.
Is AI going to lead to the end of the supercomputer for weather forecasting?
Links
GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy