Predicting fault reactivation and macroscopic failure in discrete element method simulations of restraining and releasing step overs

Abstract

To assist the identification of precursors to ruptures in segmented fault networks, we build discrete element method simulations of two parallel faults that are underlapping or overlapping, in releasing or restraining steps. We use machine learning models to predict the timing of the reactivation of the faults, and the timing of macroscopic failure identified from the peak shear stress acting on the boundaries. The machine learning models use the evolving three-dimensional components of the strain and velocity fields to make these predictions. The models depend on the same characteristics, regardless of the preexisting geometry: the component of the velocity vector parallel to the applied loading direction, vx, the shear strain component εxy, and the second invariant of the strain deviator tensor, J2. The results suggest that crustal monitoring of strike-slip systems in which the loading is approximately parallel to the fault strike, with releasing and restraining steps and overlapping and underlapping faults, may focus on the same set of strain and velocity components. However, the results indicate that the key characteristics that control the timing of fault reactivation depend on the distance from the preexisting faults. When the models use data within one fault half-length of the preexisting fault, the predictions primarily depend on εxy and vx, whereas when they use data outside this region, they primarily depend on vx. Monitoring efforts that focus on the near-fault deformation field may benefit from tracking information that helps estimate εxy and vx, in contrast to monitoring efforts further from the main faults. Models developed with data further from the faults perform worse than models developed with all of the data, but better than models developed with near fault data, consistent with observations of distributed and subsequently localizing low-magnitude seismicity.