Machine learning prediction and experimental exploration of liquid-state density and viscosity for rare earth alloys

Abstract

The liquid-state density and viscosity of multicomponent alloys are essential thermophysical properties for computational materials science. Nevertheless, such physicochemical parameters in the high-temperature liquid state are difficult to measure due to their strong chemical activity. In this work, based on our measured thermophysical properties datasets, including Fe–Nd–B, Fe–Dy–B, and Fe–Tb–B based rare-earth alloys, the random forest, support vector machine, and deep neural network models were established. It was found that support vector machine models displayed the highest prediction accuracies of 0.973 and 0.986 in the density and viscosity test sets. The temperature dependence of density and viscosity for liquid Fe76Nd5Tb3B16 and Fe78Nd10Dy3Tb3B6 alloys, with maximum undercoolings of 198 and 218 K (0.14 TL), was measured through electrostatic and electromagnetic levitation techniques, respectively. The experimental data showed satisfactory determination coefficients of 0.817 and 0.921 with the calculated values by support vector machine models, indicating the high accuracy of machine learning models in predicting liquid-state properties for rare-earth alloys.