Developing novel glasses with new, improved properties and functionalities is crucial to address major challenges in energy, communications and infrastructure. A data-driven machine learning (ML) technique is a potential method to accelerate the discovery in novel glass; however, there are some intrinsic limitations embedded in such ML models. For example, the use of traditional ML requires the data to be (i) available, (ii) complete, (iii) consistent, (iv) accurate and (v) representative. Moreover, these “blind” ML models usually generate good predictions when interpolate existing dataset, but yield poor performances when extrapolating, which prevents the efficient exploration of new unknown compositional domains. Here, I will present a new “topology-informed machine learning” framework to overcome these limitations. We show that the topological description of the connectivity of the atomic network greatly improves the accuracy of predictive model for extrapolation, compared with the traditional “blind” machine learning models. We show that the combination of atomic topology knowledge and machine learning models can be a promising route to predict the stiffness of glasses.

Author: Kai Yang, University of California, Los Angeles