Experimental study of tight reservoir rock failure process based on unsupervised machine learning

Understanding the failure process of tight reservoir rocks is essential for reservoir stimulation with hydraulic fracturing. The acoustic emission (AE) technique has proven to be an effective tool to monitor fracture propagation and thus characterize the rock failure process. In the present study, AE is employed to investigate different rock failure processes under uniaxial compression. We use four typical tight reservoir rocks from oil and gas production fields in China. We employed an unsupervised machine learning method to classify the recorded AE waveforms and evaluated the results via two score systems: elbow and silhouette scores. The cluster number is approximately constant for all samples, indicating that this method provides a more precise and reliable interpretation of the rock failure process. Machine learning results demonstrate that AE events could be distinguished into three clusters, which could relate to the mechanisms of the microscopic ruptures, i.e., tensile, shear, and mixed cracking types. Our results reveal that cracks formed under low-stress conditions are predominantly in tensile failure mode, and the failure mode transit into shear ruptures before the compression peak strength. In addition, different strength of weak planes could diversify the process of tensile-to-shear rupture transition by affecting the local stress concentration. This research may help understand the failure mechanism in tight reservoir rocks and shed light on further hydraulic fracturing technology in reservoir development.