Machine Learning-Based Classification of Copper Recovery by Flotation: A Comparative Evaluation of Logistic Regression and Support Vector Machine Models

Machine learning (ML) provides data-driven approaches for characterizing complex hydrodynamic conditions in mineral flotation. This study proposes a reproducible ML workflow for multiclass classification and knowledge discovery of flotation bubble-size regimes using the Sauter mean bubble diameter (D32) as a hydrodynamic indicator. Experimental data from a laboratory-scale flotation column were curated to obtain 353 observations comprising nine hydrodynamic, physicochemical, and frother-related predictors. The experimental platform was configured to reproduce flotation characteristics and operating conditions representative of Chilean copper-processing applications, strengthening the industrial relevance of the generated dataset. Continuous D32 measurements were formalized into three domain-informed regimes: Low (D32 < 0.9 mm), Normal (0.9 ≤ D32 ≤ 1.4 mm), and High (D32 > 1.4 mm). Logistic Regression (LR) and Support Vector Machine (SVM) classifiers were evaluated using a stratified 70/30 training–test partition, with model optimization restricted to the training subset. The radial basis function SVM achieved the best performance on the independent test subset, with an accuracy of 0.9528, precision of 0.9621, recall of 0.9621, and an F1-score of 0.9510, compared with an accuracy of 0.9434 and an F1-score of 0.9384 for LR. Correlation analysis identified gas holdup and frother concentration as the variables with the strongest linear associations with D32. Both models effectively discriminated bubble-size regimes, with RBF-SVM providing a modest but consistent improvement. The proposed workflow integrates domain-informed class formalization, supervised classification, and process-oriented interpretation, providing a reproducible basis for knowledge discovery, data-driven characterization, and future monitoring of flotation hydrodynamics.