Ore/waste identification in underground mining through geochemical calibration of drilling data using machine learning techniques

Abstract

Chemical X-ray fluorescence (XRF) analyses of drill cuttings and measurement-while-drilling (MWD) records were jointly collected in two production levels with different mineralogical characteristics of a narrow vein underground mine. More than 840 chemical samples and 23,000 MWD records with geological and geotechnical accompanying data constitute the database used in this study. A classification methodology and a site-specific model were developed following two approaches: a k-means clustering heuristic algorithm to define rock classes based on their physicochemical features, and a machine learning (ML) ensemble model trained with those classes to distinguish ore and waste as function of drilling parameters. Oversampling and under sampling techniques are required due to the unbalanced size of the rock classes. The prediction success rate (F1 score) is about 70% for the ore when associated with a high content of silica, and about 86% for the waste. When continuous drilling chips sampling is carried out, the operator bias is reduced and the F1 scores exceed 80% for all classes using different ML techniques. The study presents a complete analysis of the correlations between rock mineralogical characteristics and their mechanical response during drilling, highlighting the benefits that MWD-based information can bring to the analysis-while-drilling (AWD) for ore grading.