Machine Learning-Driven Prediction of Glass-Forming Ability in Fe-Based Bulk Metallic Glasses Using Thermophysical Features and Data Augmentation

The identification of suitable alloy compositions for the formation of bulk metallic glasses (BMGs) is a key challenge in materials science. In this study, we developed machine learning (ML) models to predict the critical casting diameter (Dmax) of Fe-based BMGs, enabling rapid assessment of glass-forming ability (GFA) using composition-based and calculated thermophysical features. Three datasets were constructed: one based on alloy molar fractions, one using thermophysical quantities calculated via the CALPHAD method, and another utilizing Magpie-derived features. The performance of various ML models was evaluated, including support vector machines (SVM), XGBoost, and ensemble methods. Models trained on thermophysical features outperformed those using only molar fractions, with XGBoost and SVM models achieving test R2 scores of up to 0.63 and 0.60, respectively. Magpie features yielded similar results but required a larger feature set. To enhance predictive accuracy, we explored data augmentation using the PADRE method and a modified version (PADRE-2). While PADRE-2 demonstrated slight improvements and reduced data redundancy, the overall performance gains were limited. The best-performing model was an ensemble combining SVM and XGBoost models trained on thermophysical and Magpie features, achieving an R2 score of 0.69 and MAE of 0.69, comparable to published results obtained from larger datasets. However, predictions for high Dmax values remain challenging, highlighting the need for further refinement. This study underscores the potential of leveraging thermophysical features and advanced ML techniques for GFA prediction and the design of new Fe-based BMGs.