Exploring the link between glassy dynamics and structure using machine learning

Abstract

One of the most intriguing features of glassy fluids is their heterogeneous dynamics: the system spontaneously forms areas of fast and slow moving particles. Recent research has shown that local structure can be used as a powerful predictor for future dynamics when combined with machine learning techniques. In this thesis we examine the relation between local structure and the dynamical heterogeneity in a glassy system consisting of spheres of two different sizes. We measure the dynamics of this system using event-driven molecular dynamics simulations, and train machine learning techniques to make predictions about these dynamics. We begin the thesis by comparing three machine learning techniques, namely linear regression, neural networks and graph neural networks, that are trained to predict the dynamic propensity based on the same set of parameters that capture the local structure of a particle’s environment. We conclude that linear regression is the preferred method, since it is fast, robust and not sensitive to overfitting. Thereafter, we examine several methods to improve the propensity prediction made by linear regression. The most significant improvement is obtained by providing the algorithm with information about the center of the cage that is formed by neighbouring particles and that acts as a temporary trap for each particle. Providing the linear regression algorithm with this cage center information, leads to impressive improvements in the accuracy of the propensity prediction at times associated with the caging regime. Additionally, we examine the anisotropy in the movements of particles. We find that both collective local drift and preferential directions in cage escape are sources for directionality in the dynamic propensity. Finally, we examine the dynamics of particles in the context of the slow and fast moving regions. We observe that at time scales close to the relaxation time, fast particles preferentially move parallel to the boundary between fast and slow regions. Intriguingly, at short times we find that the earliest particles to escape their cages, are located at the boundary and not, as expected, within the fast regions.