Emergent networks: Predicting strain localization and fracture network development

How can we estimate the timing of the next large earthquake? The ability to estimate when the next large earthquake will occur at a particular location (i.e., Los Angeles) would provide immediate societal and economic benefits. Observations of natural, crustal earthquakes, and laboratory earthquakes indicate that the precursory processes tend to accelerate in activity leading up to the dynamic, macroscopic, system-scale failure of a system. In the previous year, we have used laboratory experiments to quantitatively describe and characterize these precursory processes that signal the onset of earthquake preparation. We then used machine learning to predict the timing of laboratory and crustal earthquakes. These analyses indicate which characteristics of fracture networks and strain fields provide the greatest predictive power of the timing of earthquakes. In particular, the dilative strain, and opening of the fractures, provide the best predictive power of the timing of failure in the laboratory. This result suggests that we should focus crustal monitoring efforts on the dilative strain of the crust, and the related changes in geophysical properties, rather than the more commonly monitored shear strain.

Our experimental time series of fracture networks and strain fields provide unique access to coalescing fractures and localizing strain at spatiotemporal resolutions previously unavailable. This access enables quantifying the dynamics of the transition from distributed to localized deformation. The geophysical community lacks a quantitative understanding of the criteria that govern the transition from distributed to localized deformation. To address this gap, we will apply spatial clustering statistics, machine learning, and numerical modeling. The spatial clustering results will quantify the localization process with clear, concise, and quantifiable metrics, and thereby provide a unifying framework to describe the localization process that leads to macroscopic failure. This analysis aims to determine if the clustering statistics of fracture and/or strain networks predicts the time to macroscopic failure. The machine learning analyses will predict the volume by which a fracture grows, magnitude of strain localization, and time to macroscopic failure. These analyses aim to isolate the criteria that govern fracture propagation and coalescence, and strain localization. Determining the factors that exert the greatest impact on fault network evolution may help focus seismic hazard assessments of natural fault systems. The numerical modelling analyses will help determine if the conclusions gleaned from the cm-scale experimental clustering and machine learning applications apply to km-scales and seismogenic depths, and will enable visualizing the evolving stress field. Comparison of the accuracy of the machine learning predictions that do and do not use information about the stress field will quantify the importance of characterizing this parameter. This quantification could help justify the cost of field measurements of the stress field in seismically active areas, or indicate that the stress field is not critical to characterize when predicting fault interaction.