Machine learning-driven intelligent screening system for synergistic compound reagent in coal slime flotation

To address the technical bottlenecks of low screening efficiency and complex action mechanisms in the design of compound collectors for coal slime flotation, this study proposes a machine learning-based intelligent screening method for compound reagents. By constructing a dataset comprising 950 experimental groups and integrating the XGBoost ensemble algorithm with interpretability analysis tools, the method achieves high-accuracy prediction of flotation performance (training set MAE = 0.4592, validation set MAE = 0.4923, testing set MAE = 0.5268). SHAP feature importance analysis indicates that reagent type has the most significant impact on combustible recovery, while reagent mixing ratios also modulate model predictions to a certain extent. Further analysis based on SHAP interaction values reveals that certain collector combinations exhibit synergistic enhancement effects at specific mixing ratios, indicating that the model can effectively identify and decouple nonlinear coupling behaviors within mixed reagent systems. At the mechanistic level, the “adsorption-complementarity model” and “molecular assembly effect model” are proposed to elucidate the cooperative mechanisms: polar components selectively anchor hydrophilic sites on coal surfaces via hydrogen bonding, while non-polar components self-assemble along molecular backbones to form interfacial hydrophobic structures, thereby synergistically enhancing overall hydrophobicity. The bidirectional prediction framework integrates forward performance prediction with reverse optimization of reagent formulations based on target functions, enabling controllable design of compound schemes through interpretability analysis. Compared with traditional trial-and-error methods, this strategy significantly reduces experimental workload and greatly improves the screening efficiency of compound collectors. This approach overcomes the limitations of empirical and costly traditional design methods, providing a data-driven paradigm for intelligent collector optimization in complex coal slime systems. The results indicate that the model exhibits good generalization ability under single-coal conditions; however, the microscopic interaction mechanisms among reagents require further validation, and its applicability across different coal types remains limited. Future work may involve the integration of multi-source flotation data to enhance reagent system modeling and promote the development of intelligent mineral processing technologies.