A geometry-enhanced graph neural network for learning the smoothness of glassy dynamics from static structure

Area of Science:

• Physics
• Materials Science
• Computer Science

Background:

• Modeling glassy system dynamics is a long-standing challenge in physics.
• Graph Neural Networks (GNNs) show promise in capturing particle dynamics from structural data.
• Existing GNNs neglect explicit consideration of dynamic patterns among neighboring particles.

Purpose of the Study:

• To introduce a novel dynamical parameter,
• smoothness
• based on graph signal processing.
• To propose a Geometry-enhanced Graph Neural Network (Geo-GNN) to learn dynamic smoothness.
• To improve the correlation between structure and dynamics in glassy systems, especially on short timescales.

Main Methods:

• Developed a novel dynamical parameter "smoothness" using graph signal processing theory.
• Proposed a Geometry-enhanced Graph Neural Network (Geo-GNN) architecture.
• Employed ablation studies to validate individual components of the Geo-GNN.

Main Results:

• The proposed Geo-GNN method outperforms state-of-the-art baselines in predicting glassy dynamics.
• The "smoothness" parameter effectively captures dynamic patterns from a graph perspective.
• Ablation studies confirmed the contribution of each component to learning dynamic smoothness.

Conclusions:

• The Geo-GNN effectively models glassy dynamics by incorporating dynamic smoothness.
• This work enhances the understanding of the relationship between static structure and dynamics in glassy materials.
• The proposed approach offers a more accurate method for predicting particle dynamics in complex systems.