Predicting dynamics from structure in a sodium silicate glass

Abstract Understanding the dynamics of atoms in glasses is crucial for unraveling the origin of relaxation and the glass transition as well as predicting transport properties. However, identifying the structural features controlling atom dynamics in glasses remains challenging. Recently, machine learning models based on graph neural networks (GNNs) have successfully been used to predict future dynamics, but these prior studies focused primarily on model systems such as Kob–Andersen-type Lennard–Jones mixtures. This study investigates the use of local descriptors, GNN models, and molecular dynamics simulations to clarify the atomics dynamics in a realistic glass system (sodium silicate) across varying time scales. By harnessing the capabilities of different structural representations, we develop effective models for predicting the dynamics of sodium ions within the glassy silicate network, based solely on the initial atom positions. We further demonstrate the viability of our approach through comparison to previously proposed methods. Our findings pave the way for designing new glass formulations with tailored dynamical properties (e.g., as glassy electrolytes for batteries). Impact statement Glass science has long grappled with understanding the fundamental nature and origin of glassy dynamics. The governing principles of atomic dynamics in glasses remain elusive as it is not obvious what to look for in the glass structure. While previous studies have focused on simplified model systems, we demonstrate for the first time that machine learning models can be used to accurately predict multi-time scale atomic dynamics in a complex oxide glass (sodium silicate) from the static atomic structure. By comparing different machine learning architectures, we establish that graph neural networks outperform conventional structural descriptors for dynamics prediction, with graph representations being able to effectively capture the complex multibody correlations that govern dynamics. Our findings show that the future dynamics in oxide glasses on time scales up to nanoseconds are at least partially encoded in the initial glassy configuration itself, showing that glassy dynamics is not a completely stochastic process. The capability to predict dynamics from structure has major implications as it could provide new tools for rational design of glassy materials with tailored dynamical properties and functionalities, possibly accelerating development of advanced glasses for applications in areas such as solid-state batteries and nuclear waste immobilization. Graphical abstract