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

Jiang Xiao, Zean Tian, Kenli Li

TL;DR

Researchers have developed Geo-GNN, a geometry-enhanced graph neural network that predicts how particles move in glassy systems using only their static starting positions. By introducing the concept of "dynamic smoothness"—the tendency of neighboring particles to move similarly or differently over time—the model outperforms state-of-the-art baselines, particularly in the challenging short-term "caging" regime.

What Problem Does This Paper Address?

Predicting the dynamics of glass is one of the "holy grails" of condensed matter physics. Unlike crystals, glasses have no obvious structural order, yet their viscosity can skyrocket by orders of magnitude with tiny temperature changes.

A key feature of this transition is dynamic heterogeneity: some particles move fast while others move slow, often clustering in "cooperatively rearranging regions." While Graph Neural Networks (GNNs) have recently shown promise in predicting this motion from static snapshots, they face a fundamental hurdle: the low-pass filter problem. Standard GNNs tend to average the features of neighboring nodes. This works for long-term dynamics where motion is "smooth" (neighbors move together), but it fails miserably for short-term dynamics where particles are trapped in "cages" and may have very different propensities for motion compared to their neighbors.

Methods

The authors propose Geo-GNN, which tackles the smoothness problem through two main innovations:

  1. Geometry Feature Encoder: Standard GNNs are often limited by the "1-WL test," meaning they can't distinguish between certain different local structures if they only look at distances. Geo-GNN uses triplet-to-edge layers that incorporate both inter-particle distances and angles. This high-dimensional geometric information allows the model to capture "high-frequency" structural patterns that simpler models miss.
  2. Adaptive Aggregation: Instead of simple averaging (which acts as a low-pass filter), Geo-GNN uses a self-attention mechanism with a tanh activation function. This allows for negative attention weights, meaning the model can explicitly learn that a particle is different from its neighbors (non-smoothness), rather than just a blurred version of them.
  3. Node-to-Node Aggregation: The final layer explores the dissimilarity between a node’s own features and its local environment, effectively acting as a high-pass filter to highlight dynamic outliers.

Datasets

The model was evaluated on the Kob-Andersen (KA) system, a standard benchmark in glass physics consisting of a 3D mixture of 4,096 Lennard-Jones particles.

  • Training/Testing: 800 independent configurations (400 for training, 400 for testing).
  • Target: "Dynamic propensity"—the average displacement of a particle over a specific timescale t, calculated from multiple simulations starting from the same configuration but with different initial velocities.

Key Insights and Ideas

  • Smoothness is Time-Dependent: Using the Rayleigh quotient (a tool from graph signal processing), the authors proved that glassy dynamics start "non-smooth" at short timescales (particles vibrating in cages) and gradually become "smooth" at long timescales (particles moving cooperatively).
  • The Angle Advantage: Distances alone are insufficient to describe the local environment of a glass. Adding angular information (three-body correlations) is critical for the model to "see" the structural differences that lead to non-smooth motion.
  • Implicit vs. Explicit Constraints: While other recent models (like BOTAN) try to force the model to learn relative motion explicitly, Geo-GNN learns the underlying smoothness patterns implicitly through its architecture, making it more flexible across different timescales.

Key Results

  • Superior Accuracy: Geo-GNN achieved higher Pearson correlation coefficients than standard GNNs, SE(3)-GNNs, and Support Vector Machines across all examined timescales.
  • Caging Regime Breakthrough: In the short-time "caging" regime ($S(q, t) = 0.8$), Geo-GNN reached a correlation of ~0.798, significantly outperforming the standard GNN (~0.515).
  • Ablation Success: Removing the geometric encoder or the adaptive aggregator led to significant performance drops, confirming that both high-dimensional geometry and the ability to handle non-smoothness are essential.

Limitations and Future Work

While Geo-GNN is a major step forward, the authors note a few areas for improvement:

  • The Deepest Cage: The model is still slightly less effective at the very earliest stages of the "cage" regime, where motion is most chaotic.
  • Inherent Structures: The authors suggest that incorporating "inherent structures" (the local energy minima of a configuration) or explicit "cage structures" could further refine predictions.
  • Causality: The ultimate goal remains unresolved—exactly which structural features cause a particle to be fast or slow? Geo-GNN provides the lens to investigate this, but the "smoking gun" structural descriptor is still being sought.

Paper Citation: Jiang, X., Tian, Z., & Li, K. (2023). Geometry-enhanced graph neural network for learning the smoothness of glassy dynamics from static structure. arXiv:2211.12832v2 [cond-mat.dis-nn].

Summary updated

Abstract

Understanding the dynamic processes of the glassy system continues to be challenging. Recent advances have shown the power of graph neural networks (GNNs) for determining the correlation between structure and dynamics in the glassy system. These methods treat the glassy system as a topological graph. However, the inherent "smoothness" property of the dynamics on the graph (variation of dynamics over the graph) is ignored, resulting in a deteriorated performance of GNN for dynamic predictions over various time scales. In this paper, we first present an experimental investigation assessing the smoothness patterns of particle dynamics from the graph perspective. The results indicate that the long-time dynamics exhibit a smoother property on the graph, while the short-time dynamics reveal a distinctly non-smooth pattern. To establish the relationship between the static structure and dynamics with different smoothness patterns, we propose a novel geometry-enhanced graph neural network (Geo-GNN) model. The Geo-GNN architecture consists of a geometry feature encoder that incorporates rotation-invariant distances and angles to enhance geometric feature learning and a geometry-enhanced aggregation block that can adaptively learn the underlying smoothness patterns during message passing. Experimental results demonstrate that our method is able to capture the inherent smoothness pattern of particles and outperforms state-of-the-art baselines in predicting the dynamics across all time scales. A subsequent study has revealed that geometric features are critical for the model to accurately capture smoothness patterns in dynamics. Our research not only refines the method for predicting glassy dynamics but also provides a new lens through which to investigate the issue of causality regarding dynamical heterogeneity in glasses.
BibTeX
@article{xiao2023geometry,
  author = {Jiang Xiao and Zean Tian and Kenli Li},
  title = {{Geometry-enhanced graph neural network for learning the smoothness of glassy dynamics from static structure}},
  journal = {The Journal of Chemical Physics},
  year = {2023},
  doi = {10.1063/5.0162463},
}

Loading related content...

Figure