Crystallization prediction and reverse engineering framework construction for mold flux based on machine learning methods

To rapidly quantify the crystallization characteristics of different mold flux compositions, five ensemble tree models, trained on a dataset from public literature and experiments conducted in the laboratory, were used to predict the initial crystallization temperature. The results show that the R-2 values for LightGBM, XGBoost, CatBoost, RandomForest and Extra-Trees in the test set are 0.904, 0.919, 0.889, 0.881 and 0.899, respectively, indicating that the models fit the data well. SHAP analysis shows that the cooling rate, CaO, B2O3, CaO/SiO2, SiO2 and Na2O are key features influencing the label. The trained model captures the variation trend of cooling rate well and exhibits a highly consistent increasing trend of SHAP values when the basicity is < 1.6. furthermore, a reverse engineering framework was proposed for the automatic design of mold flux. In the case of reasonable performance requirements of mold flux and just 30 iterations, the framework can design an appropriate mold flux composition based on its desired properties. The relative errors between the designed and desired properties of mold flux are less than 3 %, and the time required is under 1 minute.