Excavation-Induced Fault Instability: A Machine Learning Perspective

Excavation-induced fault instability has been known as a major barrier for underground engineering in deep rocks. A comprehensive understanding of unloading-induced stress changes on a pre-existing fault is a critical clue to reveal the mechanism of excavation-induced fault instability. Here we established a machine learning model based on eXtreme Gradient Boosting (XGBoost) to predict changes in normal stress, shear stress, and Coulomb failure stress along the fault due to tunnel excavation. We first created the training datasets based on discrete-element modeling and tested machine learning models to select a better performing model. We then conducted a relative importance analysis and showed that the horizontal stress on the model and the coordinates along the fault are two critical factors to predict the stress changes. We used the XGBoost model to further investigate the fault-slip rockburst during the construction of Jinping II Hydropower station and demonstrated the relationships between the stress changes and the failure locations. Finally, we discussed an interesting correlation between the stress changes (reduction-dominated and rotation-dominated) and the failure locations (initiation and termination) along the fault, which is crucial to understand the mechanism of excavation-induced fault instability and to forecast the fault failure during tunnel excavation.