Decoding microstructures of fault carbonate rocks with X-ray microtomography: A deep learning approach to fabric classification and analysis

Fault zones in carbonate rocks exhibit distinct microstructural fabrics that develop different microstructures with increasing deformation, going from the outer zone towards the fault core. These fabrics can be effectively characterized using X-ray micro-computed tomography (XRμCT), a powerful imaging technique that supports a wide range of analyses, from morphometric measurements (e.g., pore size distribution, fractures orientation) to digital rock physics (i.e., virtual experiments on 3D volumes). However, the need for an automated, user-independent tool to classify these microstructures is crucial for large-scale studies. Furthermore, a fully quantitative classification of fault rock fabrics provides valuable insights into the extent and nature of deformation within these rocks. In this study, we present a deep learning-based supervised neural network designed to automate the classification of fault rock microstructures. This system offers rapid, quantitative, and scalable analysis of XRμCT data, facilitating the identification and classification of fabrics of brittle fault limestone rocks with high precision. The network was trained and validated on purpose collected datasets representing specific fabrics, then it was successfully used on different limestone fault rocks collected from the same area or obtained from the literature. The results show that the software can reliably classify fault rock fabrics affected by brittle deformation into three primary categories, each representing a distinct stage of deformation: fractured limestone, breccia, and cataclasite. The network assigns identification probabilities to each image, which can then be visualized in a ternary diagram for intuitive comparison and interpretation. This classification system streamlines fabric analysis and provides a quantitative measure of the degree of deformation within the rock. This automated classification tool paves the way for advanced studies on the anisotropic properties of fault rocks, enabling high-throughput analysis and enhancing our understanding of fault zone mechanics. • Machine learning was used to identify fault rock microstructures • Deep neural networks can quantitatively characterize fault rock textures • The presented microstructure classification approach proved fast and versatile.