Insights into metal glass forming ability based on data-driven analysis

Abstract

Scientists have extensively studied metallic glasses (MGs) for their excellent properties and potential applications. However, the limited glass forming ability (GFA) of MGs poses a significant challenge in engineering fabrication, severely limiting their scope of application. Herein, we collected data from 167 peer-reviewed research articles and 2 commonly used datasets and constructed a comprehensive dataset containing 2085 data points. We employed gradient boosting algorithmic models, such as XGBoost, LGBoost, and AdaBoost, for selecting the best feature set and constructing the most effective model from among 118 feature parameters through techniques such as three-step feature selection method, ten-fold cross-validation, and grid search. As the uneven distribution of the original dataset significantly impacts the model’s performance, we implemented five nonlinear transformation techniques and three oversampling techniques to improve the model. The results indicated that the nonlinear and oversampling techniques improved the model’s performance by improving the distribution balance. Finally, we interpreted the constructed models through Shapley additive explanation, providing relevant rules concerning GFA. The techniques employed and the rules presented herein contribute to the statistical understanding of GFA and may facilitate the discovery of novel MGs.

Keywords

• Metallic glass
• Glass forming ability
• Machine learning
• Data distribution imbalance
• Model interpretation