GraphCast: Is AI weather model the forecasting future?

AUSTIN (KXAN) — Google entered the field of weather forecasting with the debut of their GraphCast AI weather computer model.

KXAN Meteorologist Nick Bannin spoke with UT Austin Professor Liang Yang about the model and its benefits and potential flaws.

KXAN Meteorologist Nick Bannin: Liang Yang, UT professor at the Department of Earth and Planetary Sciences joins us. So Google recently announced their AI weather computer model known as GraphCast. What do you make of it so far?

Liang Yang, UT Austin Professor: Yeah, this is revolutionary news, and, indeed, Google's GraphCast is a very intriguing development in AI weather for weather modeling. And this actually one of the recent developments, following other similar developments elsewhere, like in China, and in Europe. So this leverages advanced technologies to predict weather patterns. My initial impression suggests this holds promise, especially considering Google's track record in AI applications.

Bannin: Now, how did this new GraphCast model do against some of the best weather models out there?

Yang: So you know, typically, in weather forecasting, we look at the next few hours and the next few days, and up to one week and this is typical weather forecasting. And for this critical test, published by Google's GraphCast, they look at this 'ten-day' forecast, which is, in the weather forecasting business, we call Medium Range weather forecast...and for this 10-day forecast, GraphCast is doing a very good job. Overall it is more accurate, over the scores or metrics they have selected...and this is very encouraging.

Bannin: What do you think the drawbacks or negatives of AI are for forecasting? And what areas would we lean more on human meteorologists than we would on AI?

Yang: In some events in the future, which is unprecedented...which have never happened before, and AI does not have that kind of training, so therefore AI would have no way to predict that part. [It] would have to struggle with that unprecedented event or repeated changes that deviate from historical patterns. In this case, meteorologists will continue to provide context, interpretation and adaptability in such cases.