Quantifying the precursors to brittle failure in rocks using synchrotron imaging and machine learning

Predicting the onset of system-size failure in rocks represents a fundamental goal in assessing earthquake hazard. On the field, seismological, geodetic and other monitoring data may record precursors to earthquakes. In laboratory experiments, such precursors often rely on monitoring acoustic emissions and this technique has some limitations in terms of spatial resolution and the lack of detection of aseismic strain. To overcome these challenges, we have performed a series of forty rock deformation experiments where we imaged, using synchrotron X-ray microtomography, rock samples as they deformed until brittle failure, at in situ conditions of pressure, high spatial micrometer spatial resolution, and through time. On the one hand, direct processing of the X-ray tomograms allow visualizing how precursory microfractures nucleate, grow, and coalesce until failure. From these data, we propose to characterize brittle failure as a critical phase transition, with evidence of several power-laws that characterize fracture growth. On the other hand, digital volume correlation techniques quantify the evolution of the local strain field inside each sample. We analysed the statistical properties of these strain fields using several machine learning techniques to predict the main parameters that control fracture growth (length, volume, shape, distance to the nearest fracture), and the features of the strain field that best predict the distance to failure. Our rock deformation experimental results show that, under laboratory conditions, precursors to brittle deformation exist. These precursors show predictable evolution when approaching system-size brittle deformation and we demonstrate that specific components of the strain field characterize this evolution to failure.