Glassy Dynamics from First-Principles Simulations

TL;DR

This paper employs machine learning-accelerated first-principles molecular dynamics to explore the microscopic origins of viscosity increase in glass-forming liquids, revealing the role of dynamically correlated molecules and differentiating features of glassy dynamics.

Contribution

It introduces a machine learning approach to simulate glass-formers with first-principles accuracy, providing new insights into the microscopic mechanisms of glassy dynamics.

Findings

Viscosity increase correlates with the number of dynamically correlated molecules $N^*$.

Physical aging is linked to $N^*$, but relaxation stretching is not.

Machine learning accelerates accurate simulations of complex glass-forming liquids.

Abstract

The microscopic understanding of the dramatic increase in viscosity of liquids when cooled towards the glass transition is a major unresolved issue in condensed matter physics. Here, we use machine learning methods to accelerate molecular dynamics simulations with first-principles accuracy for the glass-former toluene. We show that the increase in viscosity is intimately linked to the increasing number of dynamically correlated molecules $N^*$. While certain hallmark features of glassy dynamics, like physical aging, are linked to $N^*$ as well, others, like relaxation stretching, are not.