Convolutional neural network-based prediction of hardness in bulk metallic glasses with small data

Abstract

This study applies deep learning to predict Vickers hardness in bulk metallic glasses (BMGs) using limited datasets, addressing key challenges in materials informatics. Leveraging a convolutional neural network (CNN) model based solely on compositional features, we bypass traditional feature selection. Trained on 418 BMG samples across 10 cross-validation subsets, the model achieved strong predictive performance, with a peak R² score of 0.983 and RMSE of 55.814 in the CV₃ subset, highlighting the CNN's ability to capture composition-property relationships. Validation on unseen compositions confirmed the model's robustness, closely matching experimental values. Additionally, a pseudo-ternary diagram for Zr-Al-Co alloys was constructed, visually mapping composition to hardness. This work underscores the viability of CNNs for small datasets, advancing data-driven methods for BMG hardness prediction and materials design.