Machine learning a time-local fluctuation theorem for nonequilibrium steady states

Abstract: Fluctuation theorems (FTs) quantify the thermodynamic reversibility of a system, and for deterministic systems they are defined in terms of the dissipation function. However, in a nonequilibrium steady state of deterministic dynamics, the phase space distribution is unknown, making the dissipation function difficult to evaluate without extra information. As such, steady state FTs for deterministic systems to date have required either that the trajectory segment of interest is relatively long, or that information is available about the entire trajectory surrounding that segment. In this work, it is shown that a simple machine learning model trained to predict whether a given steady state trajectory segment is being played forward or backward in time calculates a function that satisfies an FT and relies solely on information within the segment of interest. The FT is satisfied even for very short trajectory segments where the approximate relation derived from theory breaks down, for systems far from equilibrium, and for various nonequilibrium dynamics. It is further demonstrated that any function that is a well-calibrated predictor of time's arrow must satisfy a fluctuation theorem, and that a local FT can be derived which depends only on local dissipation and its correlations with the surrounding non-local dissipation.