An element-wise machine learning strategy to predict glass-forming range of ternary alloys enabled by comprehensive data

Abstract

Many works have reported machine learning (ML) to predict the glass-forming range (GFR) of metallic glasses. However, the datasets used to train the model were mostly imbalanced and clustered around only a small portion of the composition space that made the model hard to extrapolate. In this work, the generalization deficiency of the ML model was addressed by combining combinatorial materials chip (CMC) high-throughput (HiTp) experimentation and ML modelling, and the importance of comprehensive data was highlighted. By training the models with the HiTp dataset and the published stacked dataset, it is found that 1) the data imbalance can easily lead to a skewed model, and 2) the comprehensive and balanced data distribution over the composition space dominates the model's performance. An element-wise ML strategy that the number of predictable ternary systems could be the combination of the dominant elements was demonstrated, which could dramatically accelerate the identification of GFR.