438 INTRINSIC CHARACTERIZATION OF SAG MILL FEED ORE THROUGH DEEP CLUSTERING BASED ON CONVOLUTIONAL VARIATIONAL AUTOENCODER (CVAE) AND GAUSSIAN MIXTURE MODEL (GMM) *D. Medina1, R. Martinez2, J. Campos3, A. Muñoz4 1Department of Advanced Analytics, Minera Chinalco Perú S.A., Peru, (*Presenting author: dmedina@chinalco.com.pe) 2Department of Advanced Analytics, Minera Chinalco Perú S.A., Peru, 3Department of Advanced Analytics, Minera Chinalco Perú S.A., Peru, 3Production Management, Minera Chinalco Perú S.A., Peru ABSTRACT This work presents a clustering approach that integrates Deep Clustering techniques and probabilistic models for the characterization of ore fed into the SAG mill. The methodology combines a Convolutional Variational Autoencoder (CVAE), used to learn nonlinear latent representations and temporal dependencies within 1.5-hour windows, with a Gaussian Mixture Model (GMM) for the generation of probabilistic clusters that reflect distinct ore conditions and their impact on plant operation. The study is based on 15-minute resolution data collected between January 2023 and June 2025 from the Operational Decisions Just in Time (ODJIT) system, the dataset includes geological variables (alteration: Intrusive A, Intrusive B, Hornfels, Serpentine Skarn, Actinolite Skarn; mineral: Magnetite), metallurgical variables (Bond, DWI), and geomechanical variables (Hardness). Experiments with a 4-cluster configuration show that the Deep Clustering approach effectively captures the mineralogical, metallurgical, and geomechanical heterogeneity of the SAG mill feed. It identifies patterns associated with lithologies and minerals without relying on traditional linear assumptions. The combination of CVAE for latent representation and GMM for probabilistic clustering constitutes the main contribution of this work, enabling the identification of robust and statistically interpretable clusters, and proving to be an effective strategy to represent the complex geological variability of ore fed into the SAG mill. The 4-cluster model provides an adequate balance between interpretability and innovation, establishing itself as a tool for geometallurgical characterization and decision support in milling processes. KEYWORDS Deep clustering, Convolutional Variational Autoencoder (CVAE), Gaussian Mixture Model (GMM), SAG Mill Feed Characterization. 1. INTRODUCTION Semi-autogenous grinding (SAG) milling is one of the most energy-intensive and operationally variable stages in copper processing. Mill performance is strongly influenced by nonlinear interactions among ore characteristics and operating conditions, including mineralogy, hardness, and solids content. Traditional linear monitoring and clustering approaches, such as Principal Component Analysis (PCA), often fail to capture the nonlinear and non-Gaussian behavior inherent to complex geological systems (Saldana, et al., 2026). The Toromocho deposit presents high geological and mineralogical variability, directly impacting SAG mill performance. The availability of integrated mine–plant data through the Operational Decisions
RkJQdWJsaXNoZXIy MTM0Mzk2