Predicting the glass formation of metallic glasses using machine learning approaches

Abstract

Glass-forming ability (GFA) has always been the key point for the researchers of metallic glasses (MGs) during past decades. Compared to traditional trial-and-error approaches and empirical criteria, machine learning (ML) approaches exhibit superior advantages in fast prediction speed and universal characteristic. To predict the glass formation of metallic glasses whether a composition alloy could form into MGs or no-MGs, we established five different feature subsets through the feature theoretical calculation and feature selection, and then developed machine learning models using different classifier algorithms, hyperparameters and feature subsets, and investigated their predictive performance via accuracy, F1 score, and area under curve of receiver operating characteristic (AUC). The results show that random forest (RF) classifier algorithm with tuned hyperparameters and 5G-12 feature subset presents the best overall performance. We also validated the classification ability of our optimal model on composition subspaces and the extrapolation performance in multicomponent alloy systems. The results imlpy that our optimal model exhibits relatively good overall performance in the classification of MGs and no-MGs and provides a meaningful reference for generalizing to multicomponent alloys. This work indicates great potential in searching for new candidate MGs in a vast composition space via ML approaches.