| dc.description.abstract | There are many complex factors that govern the development of fracture networks and the timing of macroscopic failure in rock. So far, we lack a unifying theory to predict fracture development in heterogeneous rock, and the corresponding timing of macroscopic failure. To better understand the factors that most strongly influence fracture development and impending macroscopic failure, we analyze the characteristics of fracture networks in rock under increasing differential stress. These characteristics describe the fracture volume, fracture orientation, fracture length, fracture aperture, and spacing between fractures in a network. We train extreme gradient boosting (XGBoost) machine-learning models with these features to predict the change in fracture volume in local subvolumes throughout the rock (i.e., local failure) and the stress distance (as a proxy for time) to macroscopic failure (i.e., global failure). We train models on data from eight individual experiments on several rock types: Carrara marble, Westerly granite, and monzonite. The resulting models exhibit a wide range of R²-values, with scores up to 0.99 for some experiments. We examine the Shapley Additive Explanation, SHAP, values to determine which fracture network characteristics exert the strongest impact on local and global failure. When the models predict the change in total fracture volume in sub-volume, the volume of individual fractures has the highest feature importance, followed by the fracture orientation, aperture, and fracture. We observe that subvolumes that decrease in fracture volume are correlated to larger fracture lengths and apertures. In most cases, we find that a high volume of individual fractures is also associated with subvolumes that decrease in fracture volume. When the models predict the stress distance to failure, the fracture orientation is the most important feature, followed by the minimum distance between fractures, the fracture length, the fracture aperture, and the volume of individual fractures. In models where the fracture aperture has a high importance, there is a negative trend in the fracture aperture as failure approaches. Changes in fracture length with approaching failure varied by rock type. The minimum distance between fractures decreases as we approach failure, implying increased localization. We also observe an increase in the volume of individual fractures. As failure approaches, the fractures orient their shortest axis (the eigenvector of their smallest eigenvalue) to a mean angle of 64° from the maximum compression direction, consistent with Mohr-Coulomb theory. | eng |