Data-driven machine learning prediction of glass transition temperature and the glass-forming ability of metallic glasses

Abstract

The limited glass-forming ability (GFA) poses a significant challenge for the practical applications of metallic glasses (MGs). The development of high-GFA MGs typically involves trial-and-error processes to screen materials with a large critical diameter (D_max), which serves as a criterion for determining the GFA. The formation and stability of MGs are influenced by the glass transition temperature (T_g). Over the past decade, the emergence of machine learning (ML) has shown great promise in the exploration of high-GFA materials. However, the contribution of material features to T_g and D_max predictions, as well as their correlations, remains ambiguous, posing a challenge to achieving high prediction accuracy. Herein, we present a comprehensive dataset consisting of 1764 datapoints for T_g and 1296 datapoints for D_max. The governing rules for GFA have been established through feature significance analysis. The light gradient boosting (LGB) model exhibits remarkable accuracy in predicting T_g, utilizing sixteen features, achieving a coefficient of determination (R²) score of 0.984 and a root mean square error (RMSE) of 20.196 K. An integrated ML model, based on the weighted voting of three basic models, is developed to enhance the accuracy of D_max prediction, achieving an R² score of 0.767 and an RMSE of 2.331 mm. Additionally, a GFA rule is proposed to explore materials with large D_max values, defined by satisfying the criteria of a thermal conductivity difference ranging from 0.60 to 1.32 and an entropy density exceeding 1.05. Our work provides valuable insights into T_g and D_max predictions and facilitates the exploration of potential high-GFA MGs through the implementation of a well-established ML model and GFA rules.