Machine Learning Unveils 86,000 Hidden Earthquakes Beneath Yellowstone

Researchers have uncovered approximately 86,000 previously undetected earthquakes beneath Yellowstone National Park, a tenfold increase over prior records, by applying machine learning techniques to seismic data spanning 2008 to 2022. This significant discovery, detailed in the July 18, 2025, issue of Science Advances, offers new insights into the park's subterranean activity and underscores the potential of artificial intelligence in geoscience.

The Yellowstone Caldera, a vast volcanic depression encompassing parts of Wyoming, Idaho, and Montana, is renowned for its geothermal features and seismic activity. Traditionally, the region records between 1,500 and 2,500 measurable earthquakes annually, most of which are minor, measuring magnitude 3 or weaker.

In this study, a collaborative team from Western University in Canada, Universidad Industrial de Santander in Colombia, and the U.S. Geological Survey reanalyzed 15 years of seismic data using advanced machine learning algorithms. Their efforts expanded the historical earthquake catalog to 86,276 events, significantly enhancing the understanding of Yellowstone's seismic behavior.

A key finding is that more than half of these earthquakes occurred in swarms—clusters of small, interconnected quakes within confined areas over short periods. These swarms often migrated upward over time, likely driven by hydrothermal fluids rising through the Earth's crust. The study also revealed that these swarms occurred along relatively immature, rough fault structures, contrasting with more mature fault zones observed in regions like Southern California. This roughness was quantified using fractal geometry, measuring the self-similarity and spatial irregularity of seismic activity.

"By understanding patterns of seismicity, like earthquake swarms, we can improve safety measures, better inform the public about potential risks, and even guide geothermal energy development away from danger in areas with promising heat flow," said Bing Li, an engineering professor at Western University and co-author of the study.

The application of machine learning allowed for the detection and magnitude assignment of approximately ten times more seismic events than previously recorded, demonstrating the potential of AI in improving seismic monitoring. Prior to the application of machine learning, earthquakes were generally detected through manual inspection by trained experts. This process takes time, is cost-intensive, and often detects fewer events than possible now with machine learning.

Yellowstone has a history of significant earthquake swarms. For instance, in 1985, more than 3,000 earthquakes were recorded over several months. Between 1983 and 2008, over 70 smaller swarms were detected. The 2008–2009 Yellowstone Lake earthquake swarm involved over 800 earthquakes in a ten-day period.

This enhanced understanding of Yellowstone's seismic activity can improve volcanic hazard assessments, leading to better public safety measures and early warning systems. Insights into the patterns of seismicity can guide geothermal energy development, helping to avoid areas with unstable fault zones while utilizing regions with high underground heat flow. The successful application of machine learning in this study sets a precedent for its use in other volcanic regions, potentially revolutionizing seismic monitoring and hazard assessment worldwide.

This study marks a significant advancement in our understanding of Yellowstone's seismic activity. By leveraging machine learning, researchers have unveiled a more complex and active seismic landscape than previously recognized. These findings have profound implications for volcanic hazard assessment, public safety, and geothermal energy development.