Unraveling the Physics of Earthquake Precursors
Using Ultrasonic Imaging and Physics-Informed Machine Learning

Dr. Jacques Rivière, Assistant Professor

Department of Engineering Science and Mechanics

Pennsylvania State University

University Park, PA 16801

Improving our understanding of earthquakes is critical to help assess seismic hazard and has broad implications for numerous subsurface activities related to energy extraction (oil/gas, geothermal), CO₂ sequestration and waste repositories. The heterogeneous nature of Earth’s crust, which leads to non-uniform stress states and complex fault systems, is often cited as a major impediment in interpreting field results, assessing seismic risk, or even one day, predicting the occurrence of earthquakes. Despite these challenges, recent work at the laboratory scale shows that machine learning (ML) algorithms can successfully predict the timing and magnitude of laboratory quakes using radiated ultrasonic waves that emanate from the faults (equivalent to seismic waves in nature). Existing studies show that laboratory quakes are preceded by systematic changes in wave properties. Inspired by these new developments, the first objective of the proposed work is to reveal the underlying physics that allows ML-based prediction of laboratory quakes. In the long term, advanced ML models – if generalizable – could provide a pathway to improving seismic risk assessment (or even predicting earthquakes) in nature. In the short- to medium-term, the use of advanced ML models in laboratory settings – where large datasets under well-constrained conditions are readily available – represents a unique opportunity to improve our understanding of earthquake physics and in turn, better understand why/how the reported predictions work. Here, I propose to build a ML framework informed by current knowledge on constitutive laws governing frictional instability. If successful, the proposed physics-informed ML framework would require less training data for equally good predictions, and make predictions more generalizable, for instance when tested outside the bounds of the training set (e.g., for real earthquakes). The second objective of the proposed work is to illuminate the role of fault heterogeneity and off-fault rock damage. High-resolution imaging of laboratory-scale faults, through the use of dense arrays of ultrasonic sensors and in combination with the physics-informed ML framework, will allow us to visualize the mechanics of earthquake nucleation and to build process-based models that connect pre-seismic activity (or lack of) to the precursory changes in wave velocity/attenuation occurring on the fault plane and surrounding host rock, in relation with heterogeneity on the fault plane. In sum, the combination of physics-informed ML models with high resolution imaging of laboratory faults will provide a physical basis that can in turn improve our interpretation of field observations as well as inform future use of ML approaches in the field.

This research was selected for funding by the Office of Basic Energy Sciences.