Machine learning-based glass formation prediction in multicomponent alloys

Abstract Metallic glasses (MGs) have attracted considerable academic attention owing to their unique properties and great application prospects. Unlike other glassy materials, such as oxide glasses, MGs have limited glass forming-ability (GFA) that often leads to failure during new MG development. Although intensive studies have proposed various parameters and criteria enabling the evaluation of the GFA of MG samples, achieving accurate predictions of glass formation before the actual MG sample synthesis remains a great challenge and an open topic. In this study, we investigated the glass formation through the data-driven machine learning technique and trained a backpropagation neural network model based on a dataset assembled from thousands of ternary alloys. Applying the well-trained model, we accurately identified the MG and non-MG classes. More importantly, our model can effectively predict glass-formation likelihood of multicomponent alloys and locate the probable MG compositions without any prior experiment, thereby directing the MG design. From the model's predictions, we discovered several new MGs in the ribbon form. Glass-formation likelihood reveals the correlation with the thermodynamic and topological parameters, which provides insights into the GFA of MGs.