Abstract
Viscosity is a critical thermophysical property that influences the castability of alloys, but it is hard to be experimentally determined due to the high temperature and easy oxidation. Employing theoretical models to predict the thermodynamic properties of alloys has always been the pursuit goal for alloy design, but there are challenges in predicting multicomponent alloys with traditional theoretical models. In this study, five different machine learning algorithms were used to construct a composition-temperature-viscosity prediction model for multicomponent alloys using 867 sets of viscosity experimental data collected in the literature. The melting temperatures (T) and solute contents of Mg, Al, Cu, Si, and Fe were utilized as model inputs, while the viscosity values were taken as model outputs. The outcomes suggest that the random forest regression (RFR) algorithm delivers excellent predictive performance, with root mean square error (RMSE) on the test set being 0.168 and the coefficient of determination (R2) being 0.984. The Pearson correlation analysis reveals a significant positive correlation between the viscosity and the content of Fe and Cu. On the contrary, Si and Mg exhibit a negative correlation with viscosity. SHapley Additive exPlanations (SHAP) analysis uncovers the critical ranges for input features (T > 1500 K, xCu < 21at.%, Fe-free, or xSi > 3.8at.%) that are significant for the design of low-viscosity alloys. Furthermore, the relation between fluidity and viscosity is investigated and optimized by regulating silicon content and solidification processes, while the established viscosity-composition-temperature mathematical model provides a theoretical basis for predicting and controlling fluidity.