Using AE based Machine Learning Approaches to Forecast Rupture during Rock Deformation Laboratory Experiments

Parametric analysis of laboratory Acoustic Emission (AE) during rock deformation laboratory experiments has revealed periodic trends and precursory behaviour of the rupture source, as crack damage nucleates, it grows and coalesces into a fault zone. Due to the heterogeneity of rocks and the different effective pressures, finding a full prediction of rupture mechanisms is still an open goal.

4x10cm cylindrical samples of Alzo granite were triaxially deformed at confining pressures of 5-40 MPa, while AE are recorded by an array of twelve 1MHz Piezo-Electric Transducers. AE are then post-processed to derive attributes and parameters. We aim to identify what are our most important parameters, and more interestingly, when they are most relevant for predicting when the rock will fail.

Time Delay Neural Networks (TDNN) have shown promise in forecasting failure when using AE-derived parameters. We trained a TDNN with 5 key parameters: 1) AE event rate, i.e. the number of events obtained during the incremental deformation (strain); 2) AE amplitude, i.e. maximum amplitude of S-waves, 3) AE source mechanisms inferred by the source radiation patterns to categorize events and obtain source orientations of mixed-mode type mechanisms; 4) Seismic scattering, i.e. the ratio between the low frequency (LF, 50-500 kHz) and high frequency (HF, 500-1000 kHz) peak delay (PD) values for individual AE and 5) Bulk elastic S-wave velocity measured at intervals throughout the experiment along the ray-paths created by transmitters and receivers. As each parameter investigates a specific mechanical aspect, taken together they provide information on deformation, fracturing and the evolving state of the background medium as failure is approached. These timeseries are then classified by the TDNN as variations in stress and strain (target parameters).

We are currently assessing the importance of individual parameters by omitting one at a time from the training routine. The more important the omitted parameter, the larger the misfit will be when comparing the network output and the target timeseries. The omission analysis determines what are the most important parameters to use when training a neural network to predict dynamic failure. Results are strongly dependent on the methods used to define the training parameters, but several trends are emerging. Event rate and amplitude differently influence predictions of stress and strain. Event rate appears relevant only in the early deformation phases, while amplitude seems much more significant during the coalescence/propagation phase. Seismic scattering and source mechanisms also show an early relevance, interpreted as due 1) to the breakup of low frequency surface waves as microcracks begin to coalesce and 2) bursts of tensile events in the enucleation phase and an increase at ~80% UCS, likely related to the crack propagation. Similarly, there is a clear pivot in the importance of seismic velocity during the early stage, but it emerges a progressive increase ~40% UCS whose origin is unclear. We are currently determining if these variations are directly related to the mechanics of the fault zone or are simply an artifact of the processing.