Prediction and screening of glass properties based on high-throughput molecular dynamics simulations and machine learning

Abstract

The design of novel glass materials has been plagued by inefficient Edisonian "trial-and-error" discovery approaches. The machine learning (ML) algorithm was combined with molecular dynamics method to predict the properties of glass overcome the defect of single method. In this paper, six ML algorithms were systematically studied using the data set generated by molecular dynamics. It is demonstrated that this method has excellent and reliable prediction performance in the whole synthetic space. The ML model and the hyperparameter suitable for glass property prediction were selected by comparing the performance of the model after tuning the hyperparameter. The ML model was utilized for rapid and low-cost screening of components (105) of multicomponent glass to develop a database of component attributes as an example of its possible application. Five samples were prepared with random components for experimental validation, and the relative errors of densities and elastic moduli were less than 10%. This study can readily be extended to anticipate various compositional-property combinations, thus replacing empirical approaches of glass property prediction with related properties and applications.