Machine learning-based performance prediction and intelligent screening of compound collector for low-rank coal with varying slime contents

The traditional trial-and-error development of compound collectors for low-rank coal (LRC) flotation is inefficient, and the process is susceptible to slime interference. To address this, a data-driven machine learning (ML) strategy was proposed to predict and screen optimal collectors for LRC with varying slime contents. Based on 504 flotation experiments, a dataset was constructed using combustible recovery as the evaluation metric. Six ML models, including XGBoost, CatBoost, LightGBM, GBDT, RF, and AdaBoost, were evaluated. XGBoost achieved the best performance (test R2 = 0.935, RMSE = 2.649). SHAP analysis identified collector type and slime content as the dominant factors and revealed nonlinear slime effects, component-specific contributions, and potential synergistic interactions among collector components. The XGBoost model accurately predicted combustible recovery under new conditions, with a maximum absolute error of 2.72%. Consequently, optimal compound collectors for LRC were screened and designed. C3+C8, C3+C6, and C4+C6 were recommended for low, medium, and high slime contents. Zeta potential and FTIR analyses indicated changes in surface charge characteristics and surface functional groups after collector treatment, providing supporting evidence for the effectiveness of the model-selected formulations. This work establishes a robust methodology for the data-driven design and efficient screening of flotation reagents, providing a practical route for optimizing LRC flotation.