Thermal expansion prediction in oxide glasses via graph neural networks with temperature-encoded virtual nodes

M Huang and JX Xiong and YQ Peng and HJ Mao and ZF Liu and W Li and FL Wang and WJ Zhang and XY Chen, MATERIALS TODAY PHYSICS, 59, 101943 (2025).

DOI: 10.1016/j.mtphys.2025.101943

Predicting thermal expansion coefficient (CTE) of oxide glasses is challenging due to over-squashing and over-smoothing in traditional graph neural network when capturing long-range dependencies. Here, we develop PPPM-GNN, a physics-informed framework that introduces a temperature-encoded virtual global node to directly integrate global structural information and broadcast thermal information to all atomic nodes. Our method provides accurate prediction of classical molecular simulation calculated properties for quaternary oxide glass systems after being trained. Further, the framework is interpretable because attention visualization reveals physically meaningful learning patterns evolving from localized interaction to task-specific pathway across layers. Meanwhile, we illustrate how this formation facilitates strong generalization capabilities in previously unexplored compositional domains. This framework established a foundation for tackling challenging materials science problems that require modeling of thermal effect and distant atomic interaction.

Return to Publications page