How do you bet millions on heat you cannot measure

A 2024 circulation test underway at the DOE-funded Utah FORGE research site. Data from this location will help Portland State University researchers test new AI technology aimed at cutting the high costs of geothermal exploration. (PC: Eric Larson, Flash Point SLC, courtesy Utah FORGE)

Imagine being asked to spend millions of dollars drilling a well, based on a prediction of how hot the rock is two miles down. Predict too low, and you walk away from a site that would have worked. Predict too high, and you drill a well that never pays for itself. And the deeper you go, which is where the heat is, the less certain the prediction gets.

That is the bet the geothermal industry makes every time it picks a site. Heat underground can be turned into electricity that runs day and night, in any weather, which is exactly what the grid needs as demand from data centers climbs. But the only way to know the temperature down there for certain is to drill, and drilling is the expense you are trying to justify in the first place.

John Lipor thinks AI can make that bet less of a gamble, and the U.S. Department of Energy has decided the idea is worth testing. PSU has been selected to lead ARISE, one of the first projects chosen under DOE's Genesis Mission, a national initiative to put AI to work on energy and scientific discovery. Lipor, Wedge Vision Associate Professor of electrical and computer engineering, leads a team that includes Stanford University, the U.S. Geological Survey and the company 400C Energy.

Lipor is not a geologist. He is a machine learning researcher who studies how to decide which measurement is worth taking when every measurement costs money, work supported by an NSF CAREER Award and a DARPA Young Faculty Award. The geothermal turn came from proximity. The USGS has a presence on PSU's campus, and that connection led to Lipor co-leading the USGS geothermal machine learning team with research hydrologist Erick Burns since 2021. Burns is now a collaborator on ARISE.

The system the team is building does three things. It predicts underground temperature. It converts the uncertainty in that prediction into dollars rather than degrees, so a developer sees the range of prices a site might deliver. And it recommends where to measure next.

"We use AI to make that bet less of a gamble," Lipor said.

Example sorting of the Great Basin into regions that are geologically similar using the ARID algorithm. AI models trained within regions may have lower uncertainty in temperature predictions, leading to a clearer understanding of the cost to develop geothermal energy. (Courtesy of John Lipor)

The middle piece is PSU's, and it started as a master's thesis. Graduate student Joshua Sills developed the algorithm, called ARID, which sorts the country into zones that are geologically alike. One model trained on the whole country has to describe the Nevada desert and the Appalachian foothills at once, and ends up imprecise about both. Give each zone its own model and each one only has to be right about one kind of place. Sills continues on the Genesis project.

"When I was writing my thesis, ARID was an idea I was trying to prove could work," Sills said. "Seeing it become part of a national project, something people might actually use to make geothermal cheaper, is not what I pictured when I started. I'm excited to find out how far we can take it."

The nine-month first phase is a test with a number attached. The team is aiming to narrow the range on those cost estimates by at least 10 percent compared with the approach used now. To check it, they will replay the real history of measurements at the DOE-funded Utah FORGE research site and compare what their system would have recommended against what the engineers there actually did.

Oregon is in the picture, though not first. The project's deliverables include a new underground temperature map for the state. Utah remains the priority site, and the Newberry volcanic area in central Oregon is a candidate for the longer term rather than a near-term target.

If the first phase holds up, the team wants to build the work into something a developer could actually use to plan an exploration campaign. The models and code will be released publicly either way.