Averaging Local Structure to Predict the Dynamic Propensity in Supercooled Liquids.

Area of Science:

• Condensed matter physics
• Materials science
• Computational chemistry

Background:

• Predicting the local dynamics of supercooled liquids from local structure is a significant challenge in glassy materials science.
• Machine learning, particularly Graph Neural Networks (GNNs), has shown promise but often requires complex models with numerous parameters (e.g., 10^5).

Purpose of the Study:

• To develop a more computationally efficient model for predicting supercooled liquid dynamics.
• To identify the essential structural features driving predictive accuracy in GNNs.

Main Methods:

• Proposed a simplified model focusing on averaged structural features around neighboring particles, inspired by GNNs.
• Applied the model to fit the dynamic propensity of Kob-Andersen and binary hard-sphere mixtures.
• Investigated the importance of radial and angular descriptors in predicting dynamics.

Main Results:

• Developed a significantly more efficient model with approximately 1000 fit parameters.
• Achieved predictive power comparable to more complex GNN models.
• Successfully fitted the dynamics of benchmark liquid models.

Conclusions:

• A simplified approach considering averaged local and neighbor-averaged structural features is sufficient for accurate dynamic propensity prediction.
• This efficient model offers a computationally tractable alternative for understanding glassy material dynamics.
• Radial and angular descriptors play crucial roles in the dynamics of these liquid systems.