Enabling extrapolation of Young’s modulus of CaO-Al2O3-SiO2 ternary glasses by topology-informed machine learning

Abstract

The application of machine learning (ML) in material discovery, particularly in the design of novel materials like glasses, has shown considerable promise. However, the efficacy of data-driven ML approaches is often hindered by the limited volume and representativeness of material datasets. While these approaches demonstrate notable success in interpolating data, they tend to perform inadequately in extrapolation tasks, which are crucial in the context of material discovery. In this study, we address this challenge by incorporating topological knowledge, derived from the atomic structures of glasses, to inform ML models with physics-based insights. To showcase this approach, we focus on predicting Young's modulus of CaO-Al₂O₃-SiO₂ glasses. By leveraging the topological information, i.e., the fractions of bond-stretching and bond-bending constraints, we transform a non-linear composition-property mapping to a higher-linearity topology-property mapping to improve the extrapolation abilities of ML models. Our results demonstrate that the topology-informed ML approach maintains comparable prediction accuracy within the training domain while significantly improving performance in extrapolating the Young’s modulus of glasses beyond the training domain. Therefore, our topology-informed approach can offer a more efficient and expedited pathway towards the discovery of new glass materials in unexplored domains.