Metal intrusion enhanced deep learning-based high temperature deterioration analysis of rock materials

Abstract

The deep learning-based method demonstrates superior capability in high-temperature deterioration analysis of rock materials through scanning electron microscopy (SEM) characterization. However, the blurred boundaries between rock particles and pore structure presented in SEM images always affect the training efficiency. Hence, in this study, the metal intrusion technology and backscattered electron (BSE) observations are applied to assist a deep learning-based model to analyze the deterioration of the rock materials exposed to different levels of high temperatures. >2000 micro images for each level were inputted and trained to distinguish the features of rock after 5 temperature levels deterioration. The results reveal that the classification accuracy on the heat deterioration of the rock materials achieved 94.3%. The classification accuracy under the optimized observation area (from 240 pixels × 240 pixels to 280 pixels × 280 pixels) is stable over 90%, balancing the high precision and training efficiency. Moreover, the interpretable high-temperature degradation characteristics of the micro-damage on rock particles and pore regions can be extracted, enabling further analysis of the degradation process of high-temperature treated rocks. We expect this work will inspire a high-precision, wide-application and cost-efficient method for the deterioration analysis of rock materials in underground projects.