A computer vision-based model for predicting the glass-forming ability of alloys without feature engineering

Abstract

Traditional empirical rule-based methods for metallic glass (MG) design exhibit limitations. Machine learning (ML) prediction of glass-forming ability (GFA) offers an effective solution. Most existing ML models require feature engineering, which introduces additional workloads. This study exclusively uses elemental composition as input and constructs images via outer product-based transformations to map elemental interactions, surpassing prior transformation methods. We enhanced the ResNet18 model with parallel multi-scale convolution, improving prediction performance without feature engineering. Results demonstrate the model's superiority over alternatives. To address data scarcity in deep learning (DL), Mixup outperformed Gaussian noise (GN) in data augmentation, achieving a coefficient of determination ($R^2$) of 0.86. Finally, the model's generalization capability was validated through comparisons between predicted and experimental values in Mg-Cu-Gd and Zr-Al-Cu alloy systems.