Predicting rock mass strength from drilling data using synergistic unsupervised and supervised machine learning approaches

Abstract Extracting rock mass strength properties from existing data like Measurement While Drilling (MWD) is important to reduce the cost of additional geological and geotechnical surveys. This study presents an approach that combines clustering (unsupervised learning) and classification algorithms to identify similar rock groups for their prediction. The dataset comprises 272,272 MWD from 2,790 drill holes, split into 215,401 data points (2,332 drill holes) for cross-validation, and another 215,401 data points, from 558 previously unseen drill holes for testing. Principal component analysis (PCA) and clustering algorithms such as K-means, Gaussian mixture, C Fuzy, and hierarchical clustering were employed to group rocks with similar MWD parameters. The combination of PCA and k-means clustering provides good cluster quality which best describes the different rock strength characteristics (clusters), as revealed by geological investigation and coring data. After identifying the rock categories, Extra Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost) approaches were used to develop classification models for rock strength prediction. The XGBoost model achieved the best and most reliable performance with accuracy, precision, recall, and F1 score exceeding 98% on the test set. This study highlights the synergetic benefits of combining unsupervised and supervised machine learning techniques to predict rock mass conditions, especially in scenarios with limited geological information or unavailable labeled data.