Earth and Mineral Sciences

Machine learning to monitor stored CO2 saves cost and time, researchers report

First study using data from the field demonstrates that machine learning can drastically reduce the seismic sources needed to monitor sequestered carbon dioxide

Carbon sequestration could potentially bridge some areas of transportation and manufacturing that are more difficult to shift to carbon neutral sources because the greenhouse gases can be stored deep in the earth. Yet the technology is costly. New work from researchers at Penn State demonstrates that machine learning could greatly reduce the long-term costs of monitoring carbon sequestration sites. Credit: Pixabay. All Rights Reserved.

UNIVERSITY PARK, Pa. — Incorporating field data for the first time, researchers at Penn State demonstrated machine learning can be a powerful and cost-effective tool for monitoring sequestered carbon dioxide (CO2), overcoming a hurdle for the burgeoning technology aimed at combating climate change.

Carbon sequestration, already in use in at more than a dozen sites in the U.S., according to the Congressional Budget Office, could be a solution for more difficult areas of decarbonization such as transportation, because it stores greenhouse gases in a super compressed, liquid form below ground for hundreds of years. It’s critical, though, that the CO2 doesn’t disturb the aquifers and other areas above — or worse, reach the surface — as it would negate the carbon sink. These sites require millions of dollars in seismic monitoring throughout their lifespan and existing methods for tracking stored CO2 — and detecting possible leaks — take weeks or longer, the researchers said.

This new approach, detailed in the International Journal of Greenhouse Gas Control, shows machine learning can create efficient, cost-effective solutions to better monitor stored CO2 with a small amount of seismic data. These solutions can be processed within seconds, at a fraction of the cost of traditional seismic monitoring methods that need hundreds of seismic devices implanted thousands of feet below ground.

Zi Xian Leong, who earned his doctorate in geosciences from Penn State in 2023, developed the method with his then adviser, Tieyuan Zhu, associate professor of geosciences.

The U.S. Department of Energy (DOE) asked the researchers to investigate more cost-effective approaches to monitoring CO2 stored at the Frio-II site in Texas and provided the team with field data from the site. Yet, there was just one problem. Instead of utilizing many seismic sources — devices that emit seismic waves similar to sonar transducers to characterize inaccessible areas — as is typical in seismic surveys to assess the CO2 plume and surrounding rock, the researchers had access to only one source.

That prompted Leong, now employed at Chevron in Houston, Texas, to explore machine learning as a means to overcome the limitation posed by having only a single seismic source.

Leong said he knew that machine learning was useful for solving complex problems — including those related to geosciences. Researchers developed the approach by first training the machine learning algorithm by feeding in training dataset which encompasses the geology of its surrounding area, rock physics and fluid dynamics for liquefied CO2. From there, they computationally generated thousands of possible scenarios of how the stored gas could change and how those changes would trigger on a seismic record. This process took months to develop, but the effort resulted in the ability to monitor the stored CO2 in just seconds.

“When you compare our results with those from traditional methods, you get a nearly perfect match,” Leong said. “That this approach was able to work with a single seismic source shows that machine learning could be the future of monitoring stored CO2.”

The lone seismic source could track the movement and volume of the stored CO2 in nearly real time, meaning that a single seismic source could infer what was happening with the entire plume of stored greenhouse gas.

For example, a CO2 storage site would typically have dozens of seismic sources and receivers spanning the approximate area and depth that it comprises. It takes significant time and computing cost for researchers to pull all of the data together to have a clear picture of the entire site and infer any changes in movement of the CO2. By training the algorithm on simulated data about the surrounding site, it could reliably predict how the CO2 behaved across the site with only the data from the single source.

“Traditional methods for monitoring sequestered CO2 are slow and costly. If you have a leak, you need to know quickly and you need to know how much,” Zhu said. “This machine learning approach speeds up the process for detection while dramatically lowering the costs. We’ve really improved the efficiency of monitoring sequestered carbon and opened the doors for using this advancing technology while also showing that this technology doesn’t just work in the lab. It can also be applied in the field.”

Zhu said although the Texas site was small compared to the size of some sites eyed for sequestration, it is easily scalable.

“While these estimations are more accurate when they are informed by more sources, this method shows that it’s possible to monitor CO2 using far fewer sources,” Zhu said. “And this method can incorporate more sources to improve reliability, still at the fraction of the cost and time of existing methods.”

Zhu said the method could be used for storing other gases, such as hydrogen, or also spotting inefficiencies in geothermal power generation.

Alexander Sun, of the University of Texas at Austin, contributed to this research. It was funded by DOE.

Last Updated February 26, 2024

Contact