Prediction of fracture initiation in cohesive soils based on data mining modelling and large-scale laboratory verification

Many empirical and analytical methods have been proposed to predict fracturing pressure in cohesive soils. Most of them take into account three to four specific influencing factors and rely on the assumption of a failure mode. In this study, a novel data-mining approach based on the XGBoost algorithm is investigated for predicting fracture initiation in cohesive soils. This has the advantage of handling multiple influencing factors simultaneously, without pre-determining a failure mode. A dataset of 416 samples consisting of 14 distinct features was herein collected from past studies, and used for developing a regressor and a classifier model for fracturing pressure prediction and failure mode classification respectively. The results show that the intrinsic characteristics of the soil govern the failure mode while the fracturing pressure is more sensitive to the stress state. The XGBoost-based model was also tested against conventional approaches, as well as a similar machine learning algorithm namely random forest model. Additionally, several large-scale triaxial fracturing tests and an in-situ case study were carried out to further verify the generalization ability and applicability of the proposed data mining approach, and the results indicate a superior performance of the XGBoost model.