Exploring a general convolutional neural network-based prediction model for critical casting diameter of metallic glasses

Abstract

Metallic glasses (MGs) as emerging amorphous materials have attracted considerable interest due to their excellent mechanical, physical, and chemical properties. However, the poor glass-forming ability (GFA) of MGs makes them difficult to produce large specimen sizes required by engineering applications. Traditional design on MGs mainly relies on empirical rules, thus being limited in the accuracy and generality, in turn leading to slow development of MGs. Motivated by the issue, we hope to utilize the powerful learning capability of deep learning to mine the relationship between the metallic glass structure and the critical casting diameter (Dmax) such that develop a universal and accurate tool to aid the experimental investigation. Based on a reliable metallic glass dataset of 1121 unique alloys collected from literature, we introduce a periodic table representation (PTR) strategy to characterize the MG structure, which only needs the information of alloy composition such that can avoid feature engineering involving domain knowledge. Based on the image representation, we accordingly construct a convolutional neural network to effectively extract the structure feature from PTR, and then follow a fully connected neural network to realize the Dmax prediction. In addition, two data augmentation strategies are introduced to address the dependence of deep learning on big data, through which the model performance is indeed improved. In particular, the pairwise difference regression (PADRE) strategy exhibits better performance than the Mixup augmentation way, as PADRE targets the difference between sample pairs such that can to some extent drop the systematic error of experimental determination. The principal component analysis (PCA) analysis further confirms the effectiveness of the PTR representation and the powerful capacity of CNN to extract structural information from PTR. Benefited from the technical advantages, our deep learning model achieves satisfactory performance with R2 of 0.822 for unseen samples in the independent test set, which outperforms two competitive models using traditional machine learning algorithms coupled with hand-selection features, further demonstrating the advantage of our model in the universality and accuracy.