Dynamics of machine-learned softness in supercooled liquids

Previous work has shown that machine learning can identify a local structural variable, called softness, which is predictive of particle-scale dynamics in many disordered systems. In simulations of supercooled liquids, this quantity has been associated with a local energy barrier to rearrangement, and has been found to be strongly descriptive of structural aging out of equilibrium, remaining predictive of particle rearrangements and the structural relaxation time throughout aging. Thus, a theory of how softness evolves in time makes predictions about the aging of a glass out of equilibrium. Here we develop a phenomenological model for how the softness of particles evolves in time in and out of equilibrium. We test the predictions of this model against the aging behaviour and temperature dependence of observables in our MD simulations of a Kob-Andersen Lennard-Jones glass.