Deep-learning analysis of microstructural deterioration in rocks exposed to high temperatures

Determining the high-temperature history of rocks and evaluating their associated deterioration levels are essential for the stable and efficient functioning of geological engineering projects. This research introduces a precise, time-efficient, and cost-effective approach that integrates metal intrusion technology, backscattered electron (BSE) imaging, and the ResNet50 deep-learning algorithm to differentiate high-temperature histories. The damage characteristics in the microstructure of rocks subjected to different temperature treatments are successfully extracted. The results show that, compared to previously reported convolutional neural networks (CNN) training and classification methods, the proposed ResNet50 algorithm improves identification accuracy by over 10%, achieving up to 98% accuracy in classifying degraded sandstone treated at temperatures ranging from 25 °C to 1000 °C. More importantly, through feature extraction of sandstone specimens after high-temperature deterioration, the ResNet50 algorithm demonstrates a superior ability to locate microscopic damage characteristics associated with different temperatures - an achievement rarely reported in previous research. For sandstone specimens exposed to 200 °C–600 °C, the extracted features primarily highlight the opening of primary pores and changes in rock particle morphology. In contrast, as the treated temperature exceeds 600 °C, the extracted features predominantly reflect thermal damage fracture, whose area first diffuses and then concentrates, aligning closely with the thermal damage theory of rock. The findings of this study not only advance a deep learning-based approach for identifying rock deterioration after high-temperature exposure but also deepen the understanding of the relation between rock microstructural characteristics and high-temperature deterioration.