Quantitative Identification of Mesoscopic Failure Mechanism in Granite by Deep Learning Method Based on SEM Images

Tensile and shear fractures are fundamental brittle fractures that are usually observed in rock failure processes. Investigating the mesoscopic morphology of shear and tensile fractures is useful for revealing the macroscopic failure mechanism of rock. This paper presented a quantitative method based on scanning electron microscopy (SEM) images and deep learning to identify the mesoscopic failure mechanism of granite, which can obtain the distribution of tensile and shear fractures on failure surfaces. For this purpose, preset angle shear and direct tensile tests were conducted to obtain the shear and tensile fracture surfaces, which were observed by SEM. These SEM images were cropped to form three image databases with different fields of view. Deep learning models (AlexNet) were developed based on training and validation images. Testing performances suggested that the developed AlexNet models had superior capabilities to identify tensile and shear fracture surfaces (accuracy 96–98%). The characteristics of tensile and shear fractures learned by AlexNet models were extracted by the integrated gradients algorithm. Additionally, AlexNet models were implemented to quantitatively evaluate the distribution of tensile and shear fractures in rock fragmentations caused by uniaxial compression load. The evaluation results showed that the proportion of shear fracture in the shear cone was 66.1–95.8%, and the proportion of tensile fracture in the spalling piece was 75–94.9%. The results verified the application of the proposed method, which was beneficial to prove the hypothesis of the failure mechanism of rock under uniaxial compression.