Exploring glassy dynamics with Markov state models from graph dynamical neural networks

TL;DR

This paper uses machine learning to develop a Markov state model from graph neural networks, revealing long-timescale dynamics and local packing fluctuations in glassy materials, consistent with free volume theories.

Contribution

It introduces a novel machine learning approach to construct Markov state models that capture slow dynamics and local heterogeneities in glass formers.

Findings

Transition timescales exceed conventional alpha-relaxation times.

The MSM reveals a fragile to strong crossover at the glass transition.

The learned state map correlates with local packing fluctuations and Voronoi volumes.

Abstract

Amorphous materials exhibit structural heterogeneities that relax only on long timescales. Using machine learning techniques, we construct a Markov state model (MSM) for model glass formers that coarse-grains the dynamics into a low-dimensional space, in which transitions occur with rates corresponding to the slowest modes of the system. The transition timescale between states is more than an order of magnitude larger than the conventional alpha-relaxation time, and reveals a fragile to strong crossover at the glass transition. The learned map of states assigned to the particles exhibits correlations of a few molecular diameters both at liquid and glassy temperatures. We show that the MSM effectively constructs a map of scaled excess Voronoi volume, and the free energy difference between the two states is given exactly by the entropy of the these distributions. These results resonate with classic free volume theories of the glass transition, singling out local packing fluctuations as the slowest relaxing features.