An end-to-end explainable graph neural networks-based composition to mechanical properties prediction framework for bulk metallic glasses

Abstract

Accurate prediction of the properties of bulk metallic glasses (BMGs) can provide an important guideline for the design of novel BMGs. While various machine learning (ML) models have been employed to predict the properties of BMGs, feature engineering is typically necessary to choose suitable descriptors based on domain knowledge or experience. In this work, an end-to-end generic framework has been proposed based on graph neural networks (GNNs) for composition-to-property prediction of BMGs. Firstly, an innovative graph representation of alloy compositions is designed. Then, two classes of GNNs have been developed to predict the fracture strength and plastic strain of BMGs. The R2 values for the optimal model on the test set were 0.963 and 0.801, respectively. Additionally, the optimal model has been fine-tuned using transfer learning for the problem of skewed distributions in the plastic strain dataset. As a result, the R2 scores on the test set improved significantly by 23.8% and 6.24%, respectively. Finally, a Hybrid Explainer has been developed to explain the entire prediction process of the model. The results of this study indicate that the proposed GNNs models may be informative for the rational design of BMGs.