Application of Clustering Machine Learning Algorithms for Geometallurgical Modeling

Abstract

Geometallurgy modeling has experienced a significant evolution in recent years. Various approaches have been explored to enhance the modeling process and gain deeper insights into the characteristics of a mineral deposit. Machine learning (ML) has emerged as one of the approaches that gained significant traction within this subject area. The objective of this thesis is to develop a data-driven clustering methodology for the generation, analysis, and validation of geometallurgical domains within the Escondida mineral deposit. A systematic review of the literature shows that clustering models are the most suitable ML methods for domain estimation within a mineral deposit. The thesis applies and evaluates three clustering machine learning models (K-means, Hierarchical Clustering, and Gaussian Mixture Models (GMM)) for domain estimation of the Escondida deposit. The proposed methodology includes Principal Component Analysis (PCA) and evaluates clustering methods using three performance metrics: Silhouette Score, Calinski-Harabasz Index, and Davies-Bouldin Index. Based on these metrics, K-means is identified as the best-performing method for this case study, with the optimal number of clusters being k = 3. K-means is then subjected to further validation processes, demonstrating consistent clustering behavior across different dataset conditions. The obtained clusters are compared with the current geometallurgical characteristics of the deposit, identifying similar spatial behavior in terms of hardness, mineralogy, and recovery-dominant zones. This thesis demonstrates that a data-driven approach can effectively capture and represent geometallurgical characteristics of a mineral deposit. It offers new insights into geometallurgical estimations and a complementary data-driven tool for geometallurgical domain generation.