Predicting Density and Viscosity for Liquid Metals and Alloys Using Machine Learning

Abstract

Density and viscosity are fundamental properties for liquid metals and alloys, however, accurate prediction of their values for multicomponent systems is still very challenging. In an effort to construct accurate density and viscosity models for multicomponent liquid alloys, machine learning methods (random forest and neural network) and conventional regression methods (linear, Arrhenius and MYEGA) were investigated in this work. The models were optimized based on a very large and comprehensive experimental database, which consisted of 34,232 density data and 19,211 viscosity data in 50 elements. The unary, binary, and ternary data was used as training dataset, the multicomponent data as testing dataset. For density modeling, the regression method results in very low testing error, whereas both random forest and neural network methods suffer from overfitting. The inclusion of an artificial 10-order equiatomic density dataset calculated by linear regression model significantly reduces the overfitting in machine learning density models, with testing error reduced from 0.231 g⋅cm⁻³ to 0.156 g⋅cm⁻³ for neural network model. For viscosity modeling, the random forest model suffers significantly from overfitting; the neural network model has relatively low overall training and testing errors than regression method, but the high viscosity region is underfitted. Augmentation of high viscosity data significantly improves the underfitting for the neural network viscosity model.