Metallurgical Mechanism Guided Machine Learning to Predict Slag Entrapment Behavior during Ladle Refining with Bottom Blowing

This study analyzed the impact of gas flow rate, oil layer thickness, and purging plug position on oil eye area and oil entrapment depth during ladle refining. To this end, a single-plug stirred water model system was used to experimentally investigate the dynamics of slag entrapment, which plays a vital role in purifying molten steel. High-definition cameras captured oil eye images and processed them through binarization. Oil eye area data were obtained using pixel point counting, establishing a database correlating slag-eye area with oil entrapment depth. The dataset was divided into 90 pct training and 10 pct validation data. The best hyperparameter combinations and established decision tree (DT) model were determined using grid search and cross-validation methods. The DT model successfully realized the prediction function for both parameters and achieved a high accuracy of 97.318 pct for the oil eye area and 90.624 pct for oil entrapment depth based on experimental data. A thermodynamic diagram and feature importance were employed to analyze these factors' effects on the experimental results. Furthermore, the decision-making process of the model was explored through Individual Conditional Expectation (ICE), Partial Dependence Plot (PDP), and Shapley Additive Explanations (SHAP) diagrams. The proposed DT model's feasibility was validated, exhibiting its strong adaptability in predicting oil entrapment and proving its effectiveness for practical applications within a plant setting.