Machine learning-based classification of laboratory fault peak slip-rate trends using acoustic emission catalogs

Fluid injection into subsurface reservoirs may lead to reactivation of natural faults, inducing slip at varying rates depending on the in-situ stress state, fluid pressure, and hydraulic properties. As a laboratory analogue of injection-induced earthquakes, we conducted four sets of fluid-driven fault slip experiments on cylindrical Bentheim sandstone samples containing pre-cut faults, subjected to different fluid injection rates and initial confining pressures. Across these experiments, we observe two distinct behaviors in the evolution of peak slip rate, which either increases or remains approximately constant during subsequent injection cycles. We attribute these contrasting mechanical responses to the combined effects of elevated pre-stress, pore pressure evolution, and the development of a localized gouge layer. Since progressive failure during stimulation campaigns may eventually result in larger events, recognizing a progressive evolution of slip rates may assist in adapting fluid injection strategies. To this end, we investigate induced seismicity patterns to study whether the subsequent peak slip rate will progressively increase or remain constant. For this purpose, we train a Convolutional Neural Network (CNN) classifier with acoustic emission (AE) catalogs as the input data. The classifier utilizes features derived from sequences of 100 consecutive AE events and their inter-event times, three-dimensional locations, and magnitudes within an arbitrary time window of a given injection cycle to classify the subsequent peak slip-rate trend. Results reveal that the model achieves a balanced classification accuracy of > 99%, indicating that the peak slip-rate trend can be classified with a high reliability. Feature importance analysis further reveals that the normal distance to the fault plane of AE hypocenter locations is the dominant factor controlling classification performance, indicating that deformation localization governs the slip response. Our results suggest that the evolution of fault slip rates may be preceded by detectable and potentially forecastable seismicity patterns, at least for the trends observed in our experiments. This framework may potentially be extended and contribute to the improvement of the warning protocols to mitigate induced seismicity in enhanced geothermal systems.