Graph neural network-based structural classification of glass-forming liquids and its interpretation via self-attention mechanism

TL;DR

This paper uses graph neural networks with self-attention to analyze and interpret the structural changes in glass-forming liquids across different temperatures, revealing key indicators of disorder.

Contribution

It introduces a GNN-based approach with self-attention for automatic feature extraction and interpretability in studying glass-forming liquids.

Findings

High attention coefficients correlate with increased structural disorder.

GNN effectively captures temperature-dependent structural changes.

Insights into microscopic mechanisms of glass formation.

Abstract

Glass-forming liquids exhibit slow dynamics below their melting temperatures, maintaining an amorphous structure reminiscent of normal liquids. Distinguishing microscopic structures in the supercooled and high-temperature regimes remains a debated topic. Building on recent advances in machine learning, particularly Graph Neural Networks (GNNs), our study automatically extracts features, unveiling fundamental mechanisms driving structural changes at varying temperatures. We employ the self-attention mechanism to generate attention coefficients that quantify the importance of connections between graph nodes, providing insights into the rationale behind GNN predictions. Exploring structural changes with decreasing temperature through the GNN+self-attention using physically-defined structural descriptors, including the bond-orientational order parameter, Voronoi cell volume, and coordination number, we identify strong correlations between high attention coefficients and more disordered structures as a key indicator of variations in glass-forming liquids.