Artificial intelligence-based investigation of fault slip induced by stress unloading during geo-energy extraction

Abstract

Seismic events triggered by stress unloading during geo-energy extraction activities have become a key focus in both seismological research and engineering safety. This study presents a novel application of waveform neural networks, combining unsupervised and supervised learning techniques to classify and characterize fractures in laboratory-induced seismic events. Initially, A neural network model was initially developed that is capable of extracting time-frequency features from waveforms through unsupervised training on 1.2 million Acoustic Emission waveforms. Subsequently, this model was fine-tuned using a labeled dataset obtained from Brazilian split and uniaxial compression tests. The final result was a highly accurate model, achieving an accuracy rate of 97.6%. By applying this refined model, insights have been gained into the complex fault slip behaviors induced by geo-energy extraction activities. Our findings reveal that fluid infiltration at the onset triggers low-energy, shear-induced fractures in low-stress fault regions, which then escalate into tensile fractures during critical sliding in high-stress areas. Key precursors to fluidinduced seismicity have been identified, providing a major advance in early seismic hazard detection. These insights are essential for monitoring and early warning of induced seismicity during geo-energy extraction activities. Our work contributes significantly to improving the safety and efficiency of geo-energy extraction, including geothermal, shale gas, and conventional hydrocarbon production.