Tech Explained: How UC scientists are putting AI to the test in Simple Terms

Nervous about artificial intelligence? You’re not alone: half of Americans now say they’re more concerned than excited about the ways AI is changing daily life. But most people surveyed also recognize that AI is changing some things for the better. Take science: with its ability to detect never-before-seen patterns in complex systems, AI is speeding the search for everything from new medicines to new galaxies. 

We talked to researchers across the University of California who are using AI to advance health, energy, agriculture and meteorology. It turns out they’re as skeptical of AI as anyone — and that’s actually a good thing. Even as they’re pushing the boundaries of research and discovery, UC scientists are asking thoughtful questions about the transformation this technology brings: Can we trust AI to make high-stakes decisions? How can we ensure its benefits outweigh its risks? And how can we make sure human wisdom and well-being stays at the center of scientific progress?

Why extreme weather is tricky for AI to predict

Each time you glance at a weather app, you’re tapping into one of humanity’s most impressive technological feats: a global network of instruments collecting hundreds of billions of observations each day and beaming the data to supercomputers, which take hours to solve trillions of equations approximating the physics that govern all atmospheric and oceanic dynamics on Earth. 

This all makes weather forecasting one of the most computationally demanding tasks humanity routinely undertakes, says Ashesh Chattopadhyay, an applied mathematician at UC Santa Cruz. Chattopadhyay’s research aims to use AI to see further into the future more accurately, using a fraction of the time, energy and computing power of today’s methods.

Using supercomputers at the UC-managed Lawrence Berkeley National Laboratory, Chattopadhyay worked with NVIDIA, CalTech and Rice University to develop FourCastNet, the first AI that can go toe-to-toe with traditional forecasting methods for accuracy and range. Like ChatGPT taking in text and images from the internet to suggest a response to a user’s prompt, FourCastNet looks at the past 40 years of weather and predicts what will happen next.

Because it’s not solving trillions of equations from scratch each time, AI can generate a forecast in seconds instead of hours, using thousands to hundreds of thousands of times less computing power. The venerable European Centre for Medium-Range Weather Forecasting is now using FourCastNet and similar tools from Google and Huawei in its daily operations.

But recent research from Chattopadhyay’s group, including collaborators at the University of Chicago and NYU, suggests it’s probably too soon to turn forecasting over to AI entirely.

“AI works great for day-to-day weather over, say, Houston,” Chattopadhyay says. “But what about when Houston is facing something that’s never been seen in recorded history, like Hurricane Harvey?” The 2017 storm dumped over five feet of rain on parts of south Texas, a once-every-two-millennia event. 

“The fact that a storm like that can happen is embedded in the physics of the system, so the traditional physics-based models predicted it, which is the great thing about them,” Chattopadhyay says. If an AI model is only trained on data going back 40 years, would it have been able to predict Harvey?

To find out, Chattopadhyay’s group trained a version of FourCastNet using decades of weather observations, but filtered out all the hurricanes stronger than a Category 2. Then they fed it an atmospheric condition that they knew would generate a Category 5 hurricane in a few days. The AI model clocked the storm, but seriously underestimated its intensity, predicting it would top out at a Category 2.

“We found that it couldn’t really extrapolate beyond what it had seen in its training data,” Chattopadhyay says. “Despite how good these models are with routine weather, getting the extremes right is still a problem. And those extremes are actually the thing scientists and forecasters care most about.”

The study flagged an important consideration for meteorologists integrating AI into their predictions, and pointed Chattopadhyay and his colleagues towards their next task for improving AI forecasts. Now they’re experimenting with integrating algorithms designed to model longer-term climate trends into the shorter-term pipeline for forecasting weather, an approach that seems to boost the capacity of AI to foresee dangerous, unprecedented storms. 

More from UC Santa Cruz: AI is good at weather forecasting. Can it predict freak weather events?