University of California, Los Angeles,
Keywords: glass, machine learning, gaussian process regression
Summary:Data-driven modeling based on machine learning (ML) offers a promising route to develop robust composition-property models in glasses. However, traditional ML shows several limitations: (i) it requires a large amount of consistent data, (ii) it has a poor ability for extrapolation far from the training set, and (iii) it can potentially violate physics laws. To address these limitations, we present a new topology-informed ML framework and apply it toward predicting the stiffness of silicate glasses. We show that the incorporation of a topological description of the glass network greatly enhances the accuracy of the developed models with respect to traditional “blind” ML. Importantly, the network topology acts as a reduced-dimensionality parameter, which partially overcomes the “curse of dimensionality” that typically affects data-driven modeling. In turn, the decrease in the dimensionality of the model greatly increases its ability to extrapolate predictions far from the training set. Overall, this method offers a promising route toward the development of robust models enabling the discovery of new glasses with improved properties.