A Machine Learning Approach to Lithology Classification in Mining Using Measurement While Drilling and Exploration Data

Abstract Accurate lithology classification is essential as variations in rock formations can significantly impact the cost and efficiency of mineral exploration and mining. Initial exploration maps provide insights into subsurface formations, though typically collected at widely spaced intervals. This study examines the use of early exploration data and Measurement While Drilling (MWD) data for lithology prediction through machine learning. The research specifically evaluates the benefit of incorporating spatial coordinates with MWD parameters to enhance classification accuracy, using support vector machine (SVM), random forest (RF), and extra gradient boosting (XGBoost) classifiers with tenfold cross-validation. The models were trained on 235,501 data points of six MWD parameters from 308 drill holes. The effects of raw (imbalanced) versus Synthetic Minority Oversampling Technique combined with Edited Nearest Neighbor (SMOTEENN) (balanced) data were analyzed, along with a comparison between random and spatial data splits. Results indicate that SMOTEENN-balanced data paired with a spatial split strategy consistently improved model stability, with the XGBoost model achieving the highest performance, with a precision of 95.60% and an F1 score of 94.41% on unseen data. Additionally, the study revealed that integrating spatial coordinates of drilling locations consistently enhanced lithology classification, with a notable F1 score improvement of 27.97% using XGBoost. The findings highlight the value of combining spatial coordinates and MWD data for improved lithology classification and offer potential support for geological modeling and sustainable mining practices.