AI models outperform traditional weather forecasting, but won't replace them yet
Artificial intelligence models are proving to be faster, cheaper, and more accurate than traditional weather forecasting models, but experts caution against complete reliance on them.
According to Christopher Dickson, a technical meteorologist at WeatherBELL Analytics, AI has evolved to mimic human-like intelligence. The new iterations of AI use deep learning trained on LLMs or large language models.
In the most general sense, AI is human-like intelligence by machines. But recently, AI almost entirely refers to something as Language Models.
Deep learning trains AI to mimic the human brain, teaching machines to learn and think like humans.
"In the context of weather, that training looks like historical weather data," Dickson said.
He added that typically, around 40 years of weather data is used in this process.
Meteorologists traditionally use forecast models like the GFS or the European to create accurate forecasts. These models use a different approach than AI.
"In the traditional approaches, we do not learn from the data but prescribe a physical set of physical rules that explains how the atmosphere behaves," Maier-Gerber said.
The question arises — which is better? Physics-based models or AI?
"There's a famous song from the Swedish pop music group ABBA. The winner takes it all. and I think at the moment we see that AI models outperform the traditional physics-based models, based on many standard scores," Maier-Gerber said.
AI models are not only more accurate but also cheaper and faster.
"A traditional model takes several hours to run, "Dixon said. "An AI model with the right hardware can run in less than a minute. So that's a crazy, crazy difference. I mean, from my point of view, like my jaw dropped when I saw that, like, I didn't believe it."
AI can also improve hurricane season predictions.
Maier-Gerber said, "I've checked the representation of tropical cyclones in these models and we see that we get significant improvements in the prediction, especially of the track."
He cited the example of Hurricane Lee last year, where Google's AI forecast model Graphcast accurately predicted Lee would hit Canada three days ahead of traditional models. However, AI models are not without their challenges, such as biases in historical data.
Dickson said he is encouraging skepticism at this point and suggested approaching this hurricane season as a bit of a scientific experiment.
Despite the advantages of AI, it doesn't seem likely that AI models will replace existing models anytime soon.
"From an operational perspective," Maier-Gerber said. "I would never drop the physics-based models."
Dickson agreed, adding, "In terms of traditional human forecasting, I think it's just going to be another tool that we use to forecast."
He emphasized the importance of the human element in high-stakes situations like hurricanes.
"The human element, in terms of communicating and interpreting the weather models, is everything," he said.
"I would rather — and I think a lot of people would rather — tune into their local station to get the forecast when it comes to something that's dangerous to them and their community, than just purely trusting an AI forecast," Dixon said.
Count on NBC 2 all hurricane season to keep you informed of what's happening and what to expect so you can be prepared and stay safe.
