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. We trained a Time Delay Neural Networks (TDNN) on key seismic attributes derived from AE, such: Event rate, i.e. the negative log time difference between successive events; Amplitude, i.e. the average max amplitude of all waveforms for each single event AE; Source mechanism estimated from first-motion polarity spheres (King et al., JGR, 2021); Seismic scattering, i.e the ratio between high and low frequency peak delays (King et al., GJI, 2022); Vp/Vs ratios from vertical P-wave velocities and horizontal S-wave velocities for individual AE (King et al., GJI, 2023).

These timeseries are then classified by the TDNN as variations in stress and strain (target parameters). TDNN require continuous, regularly sampled data but AE are discrete and irregular. To transform the training data for the TDNN, parameters are smoothed in a weighted moving window of 100 AE events, where weighting is given towards high amplitude events that occur close in space together. Data processing is applied to waveform data from all experimental condition. Despite the inherent complexity in the raw data, clear increasing or decreasing trends are repeated at different experimental conditions.

Hyperparameters for the neural network are optimised using a Genetic Algorithm (GA) by evaluating the misfit between training target (mechanical data) and model output. Each model is trained on the 10 MPa dataset and validated on the 40 MPa dataset. Roles are reversed and the results summed. This approach ensures consistent trends in the training data (waveform parameters) whilst reducing bias towards a particular dataset. We then investigate 120 configurations for the training data following a ‘leave-one-out’ strategy. E.g., a model is trained on 5, 10 and 20 MPa datasets whilst omitting the Event rate parameter. The model is then validated on the 40 MPa dataset.

Model output on validation datasets demonstrate that the TDNN can classify AE-derived parameters as increasing variations in stress and strain. 10 and 40 MPa demonstrate the best fit and are likely linked to the GA optimisation, highlighting biases driven by the training data. Forecasting results for strain and stress reveal notable over- and under-estimations of values. However, both 10 and 40 MPa are generally accurate to within 20% further highlighting the feasibility of using a TDNN for forecasting the development of new fracture under conventional triaxial conditions.