Seismic Analysis Informs Geothermal and Carbon Storage Projects
Two subsets of the renewable energy field—carbon storage and geothermal projects—can benefit from new machine learning-based 3D seismo-acoustic mapping techniques developed by Oak Ridge National Laboratory researchers.
Researchers out of Tennessee-based Oak Ridge National Laboratory (ORNL) have developed advanced data processing algorithms to generate high-resolution 3D images of underground subsurfaces via seismic and acoustic sensing.
Oak Ridge National Laboratory’s seismic data analysis techniques carry utility beyond geothermal and carbon stage projects, assisting with nuclear nonproliferation efforts. Researchers set up seismic sensors around ORNL’s research reactor to determine its operating state. Image used courtesy of ORNL by Carlos Jones
Earthquake data, machine learning techniques, and other methods can reveal insight into the causes of seismic activity to understand how operations like fracking change the state of stress when energy production involves extracting or injecting fluid/material underground.
ORNL’s advanced algorithms are practical for understanding subsurface conditions before choosing sites and drilling boreholes for renewable energy projects in carbon storage and geothermal energy. In the former case, when carbon dioxide is pumped for storage in underground caves or reservoirs, it’s necessary to monitor fluid movement and the fracture’s progression. Similarly, in geothermal plants, which involve creating a network of stable fractures to draw hot water from the Earth’s depths, accurate mapping could help boost the site’s capacity.
How the Technique Works
Chengping Chai, an ORNL geophysicist, employed advanced seismic data algorithms and machine learning methods to develop high-quality image processing tools. Combined with earthquake records, this offers insight into seismic activity caused by human-led activities like fracking, which can serve as analytical tools for planning and managing clean energy projects.
Using a machine learning-enhanced seismic monitoring workflow, Chai and his team demonstrated the ability to monitor both natural and induced seismicity at different spatial scales. It was applied to the Oklahoma region as an example for 62-mile scale sites (or 100 kilometers) and to another site with a length scale of 32.8 feet (10 meters). The research resulted in high-precision seismicity catalogs and high-resolution 3D imaging of the subsurface structure.
ORNL’s research also demonstrates time-saving benefits. Human-operated high-precision seismic analysis typically takes months to complete. But for a single step of the analysis, Chai’s deep-learning approach found more seismic events and reduced the turnaround time by 99.9%. The model took 38 minutes on a laptop to process some 235,000 body-wave seismograms in the Oklahoma region. It would take a human analyst 118 days to perform the same phase-picking task.
3D seismic event data and locations: The top row shows sequences from Oklahoma’s Waynoka earthquake in 2016, while the bottom row covers the 2016 Pawnee earthquake. The left column (“a” and “d”) displays locations with deep learning picks, and the middle column (“b” and “e”) shows manual picks. The right column features the original geological locations. Image used courtesy of ORNL (page 9)
Seismic and acoustic sensing offer utility in helping energy projects understand what’s happening under the surface as fissures progress and the earth moves. ORNL researcher Monica Maceira said seismic data is critical for alerting operators to safety issues.
Maceira and Chai aim to produce higher-resolution data at a reduced timescale, ultimately shooting for real-time updates and allowing faster decision-making around adjusting operations to prevent earthquakes.
Seismic and Acoustic Sensing in Other Applications
Beyond geothermal and carbon storage projects, ORNL’s technology can monitor nuclear reactors—an advantage in the nuclear nonproliferation field, as parties must verify whether a reactor operates according to the relevant declaration. ORNL’s machine learning models can gauge the operating condition of nuclear reactors from a distance. For example, the researchers could remotely determine whether ORNL’s High Flux Isotope Reactor was running. They refined the technology to perform the same functions for Idaho National Laboratory’s Advanced Test Reactor.
A 2022 paper in Seismological Research Letters explains how the researchers visualized and analyzed continuous data at a single seismo-acoustic station about 50 meters (164 feet) away from a research reactor. They developed a workflow with two machine learning models to measure operational states (on, off, and transition) and power levels (several steps between 10% and 90%). The study achieved an accuracy level of 0.98 to determine the on/off status. The transition state and power levels proved more challenging, with a 0.66 minimum accuracy.
One hurdle the researchers faced in their work involved devising ways to get around large equipment emitting vibration sources that would interfere with the data. The researchers set up seismic stations around the reactor’s cooling tower and determined the exact location of vibration signals via amplitude and polarization analysis. The resulting research paper was published in the Bulletin of the Seismological Society of America earlier this year.
