Better Great Lakes forecasts: U-M and NOAA pair AI with human judgment

By Eric Shaw

In a team-based-learning classroom on the University of Michigan’s central campus, the work looked like any coding workshop: small clusters of three and four bent over laptops, water bottles and coffee cups crowding the tables, a workshop leader leaning in over someone’s shoulder to point at a line on the screen.

But the screens told a more specific story.

One showed a Google Colab notebook open to a step labeled “Understand the FVCOM unstructured mesh,” part of a hydrodynamic model used to represent how water moves through the lakes. Another showed a programming environment mid-setup. During the hands-on exercises, scientists in NOAA polos worked through the same problems as the new postdoctoral researchers at the next table, beside another group from the U.S. Army Corps of Engineers.

That mix is the point. The researchers from the University of Michigan and the federal scientists from NOAA’s Great Lakes Environmental Research Laboratory who filled the room earlier this month do not usually collaborate by conference call. Many of them share a building in Ann Arbor, where staff from the U-M Cooperative Institute for Great Lakes Research, or CIGLR, work alongside NOAA colleagues, pulling from the same datasets and asking the same questions about five lakes the region cannot afford to misunderstand.

The occasion was a workshop on how to clean sensor data, test forecasting models and check machine-learning outputs against what scientists already know about how the lakes behave.

AI tools are entering environmental forecasting, and Great Lakes researchers, from veteran federal scientists to new postdocs, need to know how to use them, and how to question them, before they become routine. Over the course of the day, experts from CIGLR, NOAA and the Michigan Institute for Data and AI in Society (MIDAS) worked through the building blocks: where Great Lakes data comes from — buoys, satellites, atmospheric records — and its limits and uncertainties; how the region’s operational forecast models work; and how to set up a computing environment, then visualize and analyze the data hands-on.

None of it was about any one model. It was about a field deciding, together, when AI belongs in Great Lakes science and when it does not, and about U-M helping lead that conversation.

Why the lakes, why now

“Even those of us who have lived here our whole lives underestimate the magnitude of the lakes,” said Mary Ogdahl, CIGLR’s managing director.

That magnitude is hard to overstate. The Great Lakes hold about 21% of the world’s surface fresh water and 84% of North America’s. Over 30 million people live in the basin, and the lakes supply drinking water to more than 40 million people across the United States and Canada. Michigan does not study this resource from a distance: the state touches four of the five lakes and contains nearly two-thirds of the basin’s shoreline, some 3,288 miles of it.

And the lakes are changing in ways that make better tools urgent. Since 1951, the region’s annual average air temperature has risen 2.9 degrees Fahrenheit; separately, Lake Superior’s summer surface water temperature has climbed 4.8 degrees since 1979, the largest such increase of any of the lakes. Ice cover has generally declined since the 1990s, even as high-ice years still occur. As warming and drought strain other parts of the country, the region’s freshwater abundance becomes both more important and more complicated to manage.

Understanding those dynamics is what better data tools are for. The Great Lakes are watched by an array of instruments: buoys, gliders, autonomous underwater vehicles and satellites that feed forecasts for navigation, water-quality alerts and the harmful algal blooms that recur in places like Lake Erie and Saginaw Bay.

Those forecasts are not academic exercises: water-treatment managers, beach-health officials, shippers and the communities watching a bloom spread all act on them. AI is being explored as part of how that flood of readings could be turned into a more reliable forecast, meaning earlier warning, fewer false alarms and predictions a scientist can still check against the physics of the lake.

Teaching researchers to handle it well is not a side project; it is the work.

Why the partnership matters

The reason U-M can bring this mix of people together — veteran NOAA scientists, university faculty and new postdocs, learning side by side — is because of how CIGLR is built. U-M administers the institute, NOAA scientists share its space and its priorities, and a wider set of partner agencies put the resulting science to use.

CIGLR’s role is anchored by a five-year, $53 million NOAA cooperative agreement with the university, the latest in a series of multiyear cooperative agreements NOAA has awarded since 1989.

Built into that agreement is a mandate to train. In the previous five-year period alone, CIGLR helped provide career training to more than 250 students and postdoctoral researchers at 33 institutions.

That training mission took shape through a National Science Foundation-funded grant to CIGLR, said Ogdahl. The collaboration with U-M’s Michigan Institute for Data and AI in Society, or MIDAS, is what made it work.

“CIGLR brings the day-to-day expertise in the data, modeling and processing,” Ogdahl said. “MIDAS brings the latest technologies and the responsible use of AI in science.”

CIGLR knows what a buoy reading means, how a lake stratifies and mixes, where a model is likely to break; MIDAS brings the methods and the discipline to document uncertainty and avoid black-box forecasts, predictions whose reasoning can’t be inspected or checked against the physics of the lake.

“That’s one of the really powerful things about CIGLR, it’s an academic and federal-government collaboration,” Ogdahl said. “We work hand in hand, side by side, every day with our NOAA colleagues.”

That structure is designed to move an idea from a university lab into federal hands and, as Ogdahl describes it, to open doors a single university could not.

Through the NOAA relationship, CIGLR also works with the U.S. Environmental Protection Agency, U.S. Army Corps of Engineers, U.S. Coast Guard and others, agencies with roles in navigation, water quality and safety across the basin, she said.

The academic side reaches just as wide: a consortium that now includes 10 university partners and five private-sector organizations, with U-M as administrative lead.

The result is a network linking universities, businesses, NGOs and the federal agencies responsible for forecasting and managing the lakes around a single shared concern: the health of the lakes and the people who depend on them.

For Jing Liu, MIDAS’s executive director, that convening power is what the moment calls for. AI is moving quickly through science, and the harder question is no longer what the tools can do but whether researchers can tell when an output is trustworthy, she said.

“The University of Michigan is in a position to shape science so that it serves society, because we can bring together expertise from many disciplines,” Liu said. “The real question isn’t just whether we adopt these tools. It’s whether we use them responsibly, in ways that have lasting value.”

What they’re describing is a deliberately unglamorous vision of AI. A model trained mainly on past data, without enough grounding in lake physics, can miss exactly the conditions forecasters most need to catch: a record-warm winter, an unfamiliar bloom or an extreme outside the record it learned from. If researchers can’t see how such a model reached its answer, they can’t tell when to trust it.

The emerging Great Lakes AI effort rests on a simple insistence: a careful researcher should be able to check a model’s output against the physics of the lake. The goal is not faster models for their own sake. It is research that holds up, with expert judgment kept at the center.

What the partnership produces

Federal Great Lakes research money pays for people who can turn messy lake data into forecasts that agencies can use — and who know how to question those forecasts before anyone relies on them.

CIGLR researchers and NOAA colleagues developed a 3D harmful-algal-bloom tracker whose output was later transitioned into NOAA’s operational Lake Erie HAB Forecast, a tool managers depend on each summer. The workshop sits upstream of work like that, sharpening the judgment of the people, across the university and the agencies, who will build and vet the next tools.

For now, the clearest evidence of that return is the people. The researchers leaning over those laptops — NOAA scientists, university faculty, new postdocs and data specialists working alongside them — are the ones who will carry Great Lakes science through warmer water, shifting ice and heavier flows of data.

“It’s not just about publishing,” Liu said. “It’s about doing something that creates real knowledge and real returns for society.”