Recognition of rock materials after high-temperature deterioration based on SEM images via deep learning

Abstract

Accurately identifying the high-temperature history experienced by rocks is essential for understanding their behaviour and predicting properties. However, current approaches are limited by the heterogeneity of rocks, test scale and costs. Here, we proposed an economical, efficient and accurate approach to identifying the rocks after high-temperature deterioration via deep learning. This deep learning-based method exhibited superior abilities in distinguishing the heat-treated rock. Using a scanning electron microscopy (SEM) image covering a size of 14.6 μm × 14.6 μm, the high-temperature deterioration history of rocks can be recognized with an accuracy of 80.2%. Features such as cracks, rock patterns, and cleavage steps in SEM images would further improve the recognition accuracy. For example, SEM images with higher fractal box dimensions show a higher recognization accuracy, especially for temperatures under 600 °C. Besides, using the deep Taylor decomposition algorithm, the high-temperature deterioration regions of the rocks in the microscale were successfully located, extracted, and characterized for the first time. This study highlights the vast potential of the deep learning-based approach in damage deterioration identification of rock after high temperature, which significantly extends the application of deep learning in underground projects.