Rapid Prediction of the Viscosity of Mold Flux by Ensemble Tree Models

Viscosity is a crucial property for mold flux. To realize the rapid prediction of viscosity, different ensemble tree models were employed to learn and predict the viscosity of mold flux, based on an established viscosity database. Results show that LightGBM, XGBoost, and CatBoost models achieve the coefficient of determination (R²) values of 0.916, 0.919 and 0.908 on the test set after bayesian optimization, which are much better than RandomForest (0.859) and Extra-Trees (0.889). SHAP feature ranking reveals that F⁻, temperature, SiO₂ and Al₂O₃ are the primary factors influencing the viscosity of mold flux. The predictive ability of the three trained boosting models, Iida empirical formula and the commercial software Factsage were evaluated by both data from test set and viscometer measurement. The results suggest that the proposed boosting models demonstrate a promising predictive performance in prediction of mold flux viscosity.