# Glasses and machine learning

In some ways the field of glass physics suffers from an embarrassment of riches in the sense that a wide variety of seemingly very different ideas and theoretical methods are able to capture some features of glassy dynamics with varying levels of predictive or descriptive power. An incomplete list includes: the random first order transition entropy crisis approach (RFOT) built on spin-glass-like ideas; microscopic force-level theories based on collective density fluctuations (mode coupling theory) and activated particle hopping (nonlinear Langevin equation theory); coarse-grained dynamic facilitation models based on diffusing mobility fields and phenomenological kinetic constraints; potential-energy-landscape approaches; and the concept of frustration-limited domains. At the same time, issues as fundamental as whether the controlling mechanisms of the transition are dynamic, structural, and/or thermodynamic, and whether dynamic heterogeneity is important to first order for the massive slowing down of structural relaxation, remain highly contested.

We approached this problem by using techniques from the machine learning community to precisely quantify the relationship between local structure and propensity to rearrange in simulations of several model glassy systems. Using support vector machines and taking the unusual step of interpreting the distance to the hyperplane as a physically relevant quantity, we found a scalar order parameter, “softness”, which proved to be much more predictive of local rearrangement dynamics than existing measure. Even more surprising, we showed that in Kob-Andersen mixtures this scalar order parameter decomposed the relaxation dynamics into a set of individually Arrhenius processes — that is, a given value of softness had a one-to-one correspondence with a local energy barrier to rearrangements.

We then used this (and other!) machine learning approach to try to understand glassy thin films. The challenge there is the existence of often large length scales associated with dynamic mobility gradients (and changes in the glass transition temperature); efforts to find accompanying structural length scales near interfaces have typically come up short. Despite the successes of softness for bulk glasses, we found no correlation structure and changing dynamics near glassy interfaces. Rather than this being a single data point favoring kinetic models, by considering a broad spectrum of very generic descriptors for a particles local structural environment, we were able to conclusively rule out extremely broad classes of theoretical models connecting structure and dynamics at interfaces. This pair of studies nicely shows how machine learning techniques can be used to test assumptions about the relative importance of structural and kinetic explanations of glassy behavior in different contexts.