From CT scans to surface metrics: A novel workflow for rock fracture roughness evaluation at sample scale

This study presents a robust deep learning-based workflow for fracture segmentation and subsequent Joint Roughness Coefficient computation from X-ray computed tomography data. The core of the approach is a convolutional neural network trained on a heterogeneous multi-material fracture dataset, capable of accurately segmenting even sub-millimetric and low-contrast features without domain-specific retraining. This network demonstrated superior continuity, robustness, and precision compared with classical image-processing approaches, identifying a higher number of fractures and providing a more complete representation of the fracture network. In the proposed workflow, the convolutional neural network output is converted into a high-resolution three-dimensional point cloud, from which surface roughness is computed directly using the widely adopted Joint Roughness Coefficient metric. The method is validated against: (i) point clouds derived from high-resolution digital microscope images, and (ii) digitized profiles acquired with the Barton comb profilometer. This multi-method validation captures microscale roughness variability and ensures robustness across different acquisition modalities. The results confirm that the CNN-based CT workflow provides a non-destructive, scalable, and objective solution for fracture roughness characterization, particularly suitable for intact core samples and inaccessible discontinuities.