Entropy Production in Non-Gaussian Active Matter: A Unified Fluctuation Theorem and Deep Learning Framework

Abstract: We present a general framework for deriving entropy production rates in active matter systems driven by non-Gaussian active fluctuations. Employing the probability flow equivalence technique, we rigorously derive an entropy production decomposition formula and demonstrate that the entropy production, $\Delta s_\mathrm{tot}$, satisfies the integral fluctuation theorem $\langle \exp[ -\Delta s_\mathrm{tot} + B_\mathrm{act}] \rangle = 1$ and the generalized second law of thermodynamics $\langle \Delta s_\mathrm{tot}\rangle \geq\langle B_\mathrm{act}\rangle$, where $B_\mathrm{act}$ is a path-dependent random variable associated with the active fluctuations. Our result holds generally for arbitrary initial conditions, encompassing both steady-state and transient finite-time regimes. In the limiting case where active fluctuations are absent (i.e., $B_\mathrm{act} \equiv 0$), the theorem reduces to the well-established results in stochastic thermodynamics. Building on the theoretical foundation, we propose a deep learning-based methodology to efficiently compute the entropy production, utilizing the L\'{e}vy score function we proposed. To illustrate the validity of this approach, we apply it to two representative systems: a Brownian particle in a periodic active bath and an active polymer system consisting of an active Brownian particle cross-linker interacting with passive Brownian beads. Our results provide a unified framework for analyzing entropy production in active matter systems while offering practical computational tools for investigating complex nonequilibrium behaviors.