Nonequilibrium entropy from density estimation

Entropy is a central concept in physics, but can be challenging to calculate even for systems that are easily simulated. This is exacerbated out of equilibrium, where generally little is known about the distribution characterizing simulated configurations. However, modern machine learning algorithms can estimate the probability density characterizing an ensemble of images, given nothing more than sample images assumed to be drawn from this distribution. We show that by mapping system configurations to images, such approaches can be adapted to the efficient estimation of the density, and therefore the entropy, from simulated or experimental data. We then use this idea to obtain entropic limit cycles in a kinetic Ising model driven by an oscillating magnetic field. Despite being a global probe, we demonstrate that this allows us to identify and characterize stochastic dynamics at parameters near the dynamical phase transition.