Characterizing Ore Particle Size for Sag Mill Feed Control

Abstract

This thesis presents a comprehensive, data-driven study for predicting the particle size distribution (PSD) at the feed of a semi-autogenous grinding (SAG) mill. A consistent PSD is essential for efficient energy use, stable throughput, and effective downstream processing. However, feed variability, caused by ore heterogeneity, stockpile segregation, and reactive blending, makes prediction challenging. The first part of this work reviews the current state of research and industrial practice, synthesizing insights from over 45 peer-reviewed studies on ore blending, stockpile management, and the application of machine learning in mineral processing. The review identifies a gap between the growing availability of high-frequency sensor data and its limited use in real-time predictive tools, highlighting the need for systems that forecast PSD and support data-informed operational understanding. Building on these insights, the second part develops and validates a machine learning framework using two years of high-resolution operational data from a copper mining operation. The methodology integrates unsupervised clustering with Random Forest regression to forecast key PSD metrics (F10–F90 and TOPSIZE) based on variables such as feeder speeds, stockpile levels, and ore throughput. Cluster-specific models capture nonlinear, regime-dependent behaviors with high accuracy ($R^2 > 0.90$). In addition to prediction, the thesis presents a data-driven sensitivity analysis to evaluate how changes in input variables, such as individual feeder rates, influence PSD outcomes. This analysis shows that coarser PSD metrics, such as F70 and TOPSIZE, are more responsive to input changes than finer ones. A variability analysis further demonstrates the model’s ability to quantify and explain fluctuations in PSD, with predicted distributions exhibiting up to a 15% reduction in standard deviation compared to historical data. All components are integrated into a web-based application that allows users to explore model behavior, visualize PSD forecasts, and assess the impact of operating scenarios in real time. This work advances the field by demonstrating how machine learning and clustering can be effectively applied to model and interpret SAG mill feed variability. The resulting framework offers a practical approach for enhancing predictive insight in mineral processing operations.