Viscosity and melting temperature prediction of mold fluxes based on explainable machine learning and SHapley additive exPlanations

Viscosity and melting temperature are indispensable properties for mold flux design and evaluation, significant investment is required for measurement. In this paper, viscosity and melting temperature predictive model for mold fluxes were established and trained based on 3300 groups of data and four representative machine learning algorithms. The gradient boosting regression tree (GBRT) algorithm-based model performed best prediction with the determination coefficient R2 of 0.969 and 0.900 for viscosity and melting temperature. It also far outperforms the widely used viscosity and melting temperature prediction models, with a least mean deviation of 7.06% and 0.8%. SHapley Additive exPlanations (SHAP) analysis revealed that Al2O3, followed by SiO2 showed the strongest positive correlation with viscosity, Na2O has the greatest negative contribution to melting temperature. Ternary viscosity and melting temperature distribution diagrams were constructed to visualize the prediction results. Insights from this study will highly benefit computer-aided design of mold fluxes with desired properties.