Weather forecasting, a practice that dates back to ancient times, is being given a high-tech twist. Artificial intelligence is now being applied to predict the weather — and it could make forecasts more accurate.
Weather forecasting is already a complex business. Giant supercomputers, some the size of a school bus, typically do the heavy lifting to predict the weather using physics simulations that take hours to compute, and are mostly in the possession of government organizations, such as the European Centre for Medium-Range Weather Forecasts (ECMWF).
But that could soon change. Artificial intelligence, a force disrupting many industries, could soon do the same for weather forecasting. Google announced its GraphCast program on Tuesday, which uses machine learning to predict the weather and is a part of the company’s DeepMind artificial intelligence research lab. In a media release, Google said the program offers 10-day weather predictions in under one minute and with “unprecedented accuracy.”
“There’s been tremendous progress the last year,” Steve Easterbrook, a computer science professor and director of the University of Toronto’s School of the Environment, told Global News.
AI is used to predict weather in much the same way it operates in other applications, such as the personal assistant ChatGPT. Google’s GraphCast takes terabytes worth of weather data from the last four decades and uses the past trends to make a prediction of how the current conditions will play out, Easterbrook said.
Much like how ChatGPT doesn’t actually know what it is saying but is predicting which word would likely follow the next based on literature it has consumed, GraphCast does the same and is actually a “mimic” of what the supercomputers take hours to do, according to Easterbrook. And it is able to come up with results — fast.
“It’s memorizing all of the patterns that you might see in a traditional weather forecasting model,” Easterbrook said. “It’s able to spit out results almost instantly.”
That speed is what has almost guaranteed programs like GraphCast a role in weather forecasting, Easterbrook said. He predicts traditional forecasters will likely use both AI and supercomputers together to improve their forecasting.
Global News’ head meteorologist, Anthony Farnell, said that he’s excited about the potential of AI in weather forecasting and has even heard that it is beating the vast majority of regular numeric models. He said its speed could help meteorologists adjust on the fly, and it could help with evacuations in the case of weather emergencies and an assessment of who could be at risk.
Easterbrook noted that AI has the potential to detect major weather systems earlier than supercomputers, and Google says that GraphCast accurately predicted nine days in advance that Hurricane Lee would make landfall in Nova Scotia, compared to six days in advance from traditional weather forecasts.
Farnell, though, isn’t worried about being replaced by a robot.
“It doesn’t tell you the full story,” he said. “You still need experts to basically say, ‘this is what is going to happen.'”
Easterbrook warned that there are still limitations to the use of AI in weather forecasting and also some risks. For one, it is better at predicting weather in macro scenarios, not on a local level. And since machine learning weather models are mostly untested in operational settings, such as when there’s an extreme weather event, that could leave questions about how accurate it could be.
“It’s harder to use data from the past to predict what these extreme weather events are going to be like in the future,” he said. With weather records broken more and more frequently, such as heat records this past summer, there is little comparison to what was experienced in the past.
Easterbrook also said that AI weather models can sometimes violate the laws of physics in their predictions since the technology doesn’t actually know anything about the physics of weather events, similar to how ChatGPT isn’t actually talking but predicting words. (Google employees said that the company’s Bard personal assistant could sometimes be a pathological liar.)
“You have to watch out for things like that,” Easterbrook said.
The typical hyperbole behind such Google products could also put some in danger, he noted, especially if an unwarranted degree of confidence is given to predicting extreme weather events.
The cheaper cost of AI weather models — GraphCast is open-sourced, meaning anyone can apply it — also opens the door to more privatization in weather forecasting, which could prove to have its own risks. In the tech world of move-fast-and-break-things, putting out an inaccurate weather forecast that looks legitimate is a “huge risk,” Easterbrook said.
“We’re going to see a privatization of weather forecasting services and it would be like a closed box,” he said. “It’s a massive danger in this industry.”