Separating Injection‐Driven and Earthquake‐Driven Induced Seismicity by Combining a Fully Coupled Poroelastic Model With Interpretable Machine Learning

• Published in: Geophysical research letters 2024-09, Vol.51 (18), p.n/a
• Main Authors: Hill, R. G., Trugman, D. T., Weingarten, M.
• Format: Article
• Language: English
• Subjects: Algorithms; Aquifers; Brines; Catalogues; Clustering; Deformation; Earthquakes; Fluid injection; induced seismicity; Injection; Learning algorithms; Machine learning; nearest neighbor distance; paradox valley unit; Paradoxes; Physics; poroelastic; Poroelasticity; Pressure changes; Seismic activity; Seismicity; Static deformation; Statistical analysis; Statistical models; Valleys; Wastewater
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2024GL109802

In areas of induced seismicity, earthquakes can be triggered by stress changes due to fluid injection and static deformation from fault slip. Here we present a method to distinguish between injection‐driven and earthquake‐driven triggering of induced seismicity by combining a calibrated, fully coupled, poroelastic stress model of wastewater injection with interpretation of a machine learning algorithm trained on both earthquake catalog and modeled stress features. We investigate seismicity from Paradox Valley, Colorado as an ideal test case: a single, high‐pressure injector that has induced thousands of earthquakes since 1991. Using feature importance analysis, we find that injection‐driven earthquakes are approximately 22± $\pm $5% of the total catalog but act as background events that can trigger subsequent aftershocks. Injection‐driven events also have distinct spatiotemporal clustering properties with a larger b‐value, closer proximity to the well, and earlier occurrence in the injection history. Generalization of our technique can help characterize triggering processes in other regions where induced seismicity occurs. Plain Language Summary The Paradox Valley Unit, Colorado in the central United States has had a remarkable increase in seismicity coincident with over 8 million cubic meters of brine fluid injection since 1991, inducing thousands of earthquakes within an aquifer 4.5 km below the surface. We use a physics‐based model of the Earth combined with statistical and machine learning techniques to help discern which earthquakes are triggered by other earthquakes and which earthquakes are directly triggered by the stress changes from the well as well as their comparative characteristics. Discerning which earthquakes are directly caused from pressure changes due to the fluid injected by the well can inform our understanding of earthquake physics and provide useful information to operators of energy production sites. Key Points Physics‐based models with interpretable machine learning can identify likely triggering mechanism of earthquakes within an induced sequence With this technique, we show that only 22土5% of earthquakes in Paradox Valley are primarily injection‐driven Injection‐driven events have a larger b‐value, are closer to the well, and occur earlier in the injection history