Enhancing glass property predictions through ab initio‐derived descriptors

Abstract

The performance of ab initio descriptors derived from density functional theory simulations is systematically investigated in comparison to traditional compositional descriptors for the ability to predict glass properties utilizing machine learning algorithms. Two datasets are used for this purpose: an extensive, publicly available database involving a wide range of oxide glasses, and a small in‐house dataset covering a broader collection of inorganic glasses from metallic to non‐metallic materials. For the larger dataset, it was demonstrated that ab initio descriptors offer a substantial reduction in input dimensionality while retaining nearly equivalent predictive performance when compared to the compositional descriptors. The combination of ab initio and compositional descriptors showed an improvement in prediction accuracy. For the smaller dataset, the ab initio‐derived descriptors performed significantly better than the compositional descriptors, providing a valuable tool to improve glass property prediction in settings where the availability of data is limited. Furthermore, ab initio‐derived descriptors are not only computationally inexpensive and allow extrapolation beyond the training composition space but also facilitate model interpretation.