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

Abstract

<p>The ability to detect precursors of dynamic failure in brittle rocks has key implications for hazard forecasting at the field scale. In recent years, laboratory scale rock deformation experiments are providing a wealth of information on the physics of the fracture process ranging from fracture nucleation, crack growth and damage accumulation, to crack coalescence and strain localization. Parametric analysis of laboratory Acoustic Emission (AE) data has revealed periodic trends and precursory behaviour of the rupture source mechanisms as a fault zone enucleates and develops, suggesting these processes are somehow repeatable and forecastable. However, due to the inherent anisotropy of rock media and the range of environmental conditions in which deformation occurs, finding full consistency between AE datasets and a prediction of rupture mechanisms from AE analysis is still an open goal. Here we apply a Time Delay Neural Network (TDNN) to Acoustic Emission (AE) sets recorded during conventional triaxial rock deformation tests. We forecast the Time-to-Failure using the discrete, non-continuous timeseries of AE rate, amplitude, focal mechanism and forward scattering properties. 4x10 cm samples of Alzo granite, a homogeneous medium-grained plutonic rock from NW Italy with an initial porosity as low as 0.72%, were triaxially deformed at strain rates of 3.6mm/hr under dry conditions until dynamic failure at confining pressures of 5, 10, 20 and 40 MPa respectively. Each sample was positioned inside an engineered rubber jacket fitted with ports where an array of twelve 1 MHz single-component Piezo-Electric Transducers were embedded, allowing to record AE during the experimentation. Several parameters were considered for the TDNN training: AE rate, deformation stages prior failure (elasticity, inelasticity and coalescence), AE amplitude, source mechanisms and scattering. All these parameters are key indicators of the evolving damage in the medium. Our training input consists of simplified timeseries of the previously discussed AE parameters from the experiments carried out at the lowest confining pressure (5 MPa). The inputs are classified as the stress-until failure and strain-until-failure for each AE. Once trained we then simulate the model on the untrained datasets to test it as a forecasting tool at higher confinements. At each step the model is simulated on AE data from the previous 0.2% of strain. At 10 MPa we observe a reliable forecast of failure that starts with the anelastic phase and becomes more accurate during strain-softening. At higher confining pressure, an increased limit of forecasting the solution is observed and interpreted with more complexity in the coalescence process. Despite these limitations, the model shows that when trained even on a limited input it is able to forecast dynamic failure in unseen data with surprising accuracy. Future studies should investigate AE spatial distribution for the TDNN training.</p>