Comparative Evaluation of Machine Learning Models for Predicting Grinding Throughput in Mining Operations

Huamani J.A.R.; Fernando Lozano Fernandez J.; Milla E.C.; Ventocilla C.L.R. · IEEE Access (2026) · DOI: 10.1109/access.2026.3719319

Resumen

Accurate grinding throughput prediction, expressed in tons per hour (tph), plays an important role in maximizing the economic value of comminution circuits. This study integrates geometallurgical data, drill-and-blast parameters, and plant operating variables to estimate tph in near-real time using machine learning algorithms. After data preparation, including outlier analysis, logarithmic transformations, and feature selection, several supervised learning models were evaluated. The Extra Trees Regressor achieved the best predictive performance, with an R2 of 0.956 on the test set and an average fold-wise R2 of 0.9553±0.0015. Additional robustness analyses, including sequential validation and outlier sensitivity evaluation, confirmed the stability of the proposed model under different data partitions and operating conditions. SHapley Additive exPlanations (SHAP) and partial dependence analyses showed that fines fraction and ore hardness indicators, including SAG Power Index (SPI), Bond Work Index (BWI), and uniaxial compressive strength (UCS), were the most influential variables affecting throughput prediction, supporting the physical interpretability of the Mine-to-Mill framework. Although the study was developed using data from a single mining operation, the results demonstrate the potential of machine learning techniques to support industrial decision-making and future integration into industrial advisory monitoring systems for mining process optimization.

Instituciones: Universidad Nacional de Ingeniería