| dc.description.abstract | Earthquake prediction is a highly important goal in geoscience. In this study we present usage of machine learning to predict distance to failure in rocks, a problem adjacent to earthquake prediction. We use two machine learning techniques, XGBoost and Neural Networks, to predict the strain distance to failure in 15 rock deformation experiments. In these experiments on six different rock types, we use the local strain components calculated with digital volume correlation (DVC) to predict the normalized macroscopic axial strain, i.e., the distance to failure. We use Shapley Additive Explanation (SHAP) to quantify the impact of each feature on our models, and transfer learning between rock types to constrain the generalizability of each model. We combine data from multiple experiments to generate models with increased generalizability. In this study, the importance of dilation in predicting macroscopic failure is about double the importance of the shear strain or contraction. We found that the differences in failure mechanisms between rock types produces lower transfer scores, and that brittle failure in rocks carry differences even for rock types expected to deform with similar mechanisms. The evolution of the strain components is critical to the model performance: all models with systematic evolution towards failure performed with strong or moderately strong correlation between the predicted and observed values. Lastly, we highlight the increase in model performance when the models use data from multiple experiments, rather than individual experiments. | eng |