AI weather modeling being developed by U Chicago could help create more accurate forecasts

Scientists at the University of Chicago are working on AI trained to model weather which could help create better, more accurate forecasts.

We're all familiar with the unpredictability of daily weather; a forecast may say one thing, but what actually happens may be something else.

Gail Williams, a Chicago Public Schools crossing guard whose post is at 66th and California Avenue, knows perhaps better than most that a forecast that isn't perfect isn't fun.

"Chicago weather is very unique, you have to prepare. You have to prepare for Chicago weather," she said. "It's terrible, especially when you get rained on and you get soaking wet."

Researchers at the University of Chicago are working to change that.

"Until a few years ago, whenever you saw a weather forecast on TV or on your phone, it came from a very complicated model of thousands of equations solved on super computers," explained Dr. Pedram Hassanzadeh.

Hassanzadeh is on the team working to develop AI weather models.

"They're energy-efficient, they're cost-efficient, they are very fast," he said.

Artificial intelligence is already providing to be a powerful tool in the medical field. A study of nearly a half million mammogram patients in Germany found that radiologists using AI support during breast cancer screening had a 17% higher cancer detection rate than doctors screening just with their own eyes, and have fewer false alarms.

AI accomplishes that using pattern recognition, scanning millions of pixels and examining subtleties that humans may miss.

So, Dr. Hassanzadeh had a thought.

"Maybe we can go back to that idea, and do weather forecasting as a pattern recognition problem," he said.

Researchers trained AI weather models by digesting 40 years of high-quality weather observations. The model scans what the weather patterns look like now, finds a similar picture it's seen before, and makes its prediction based on what happened last time.

"You can do weather forecasts for days and weeks," Hassanzadeh said.

CBS News Chicago's First Alert weather team already uses AI models to predict the weather every day, alongside the traditional supercomputer forecasts. The maps they produce look the same, and they've actually proven quite accurate.

But AI has limits. Since it's trained on weather in the past, can it forecast more extreme storms in a changing climate? That's a real concern for meteorologists working every day, and Hassanzadeh and his colleagues at the University of Chicago.

"That's a real concern," he said. "Can AI models actually get these kinds of events, that they haven't seen anything like them in the past?"

So he and his team developed a new model that marries the two: the speed and efficiency of AI with the old school ability to forecast record-setting storms.

"I think it's a really exciting time for meteorology, for climate science, to have this powerful set of tools in our arsenal to do better prediction," Hassanzadeh said.

And in turn, these more accurate forecasts will help people like Gail as she tries to keep her students safe.

"We can call the neighbors and tell them, put an extra sweater on this child, or that. That would help out a lot," she said.

The new University of Chicago weather prediction model has been tested, but is a long way from being used daily by weather forecasters.

The University of Chicago team that developed the forecast model blending AI weather prediction with physics-based methods was featured in a recent article in Society for Industrial and Applied Mathematics (SIAM) News.