Insights on earthquake nucleation revealed by numerical simulation and unsupervised machine learning of laboratory-scale earthquake

Abstract

Abstract Understanding earthquake nucleation is vital for predicting and mitigating seismic events, saving lives, and enhancing construction practices in earthquake-prone areas. Cascade triggering and preslip triggering are prevalent theories, posing challenges in differentiation based on field observations. Our study employs a novel unsupervised machine learning pipeline, integrating macroscopic- and grain-scale data from stick-slip experiments in a discrete element method (DEM) framework. Running 27 simulations, we cluster foreshocks and mainshocks separately and assess their correlation. The study supports the cascade triggering model on the macro-scale, as we did not observe any scaling between nucleation parameters and the mainshock size. On the other hand, further grain-scale analysis identifies that, separate from Coulomb stress transfer, there is an additional mechanism related to shear stress accumulation on the fault, which is likely the preslip triggering. Overall, while foreshocks may not directly influence the trend at which contact force evolves, they could prime the fault for dynamic rupture by increasing the proportion of contacts accumulating shear stresses. Our findings infer the possible coexistence of the two theorized mechanisms.

Keywords

• Machine learning
• Clustering analysis
• DEM
• Grain-scale
• Laboratory earthquake
• Earthquake nucleation