Unveiling the Phase Diagram and Reaction Paths of the Active Model B with the Deep Minimum Action Method

TL;DR

This paper uses a deep neural network implementation of the geometric minimum action method to analyze the phase diagram and reaction paths of the active Model B, revealing how active matter systems transition between phases.

Contribution

It introduces a novel deep learning approach to compute the phase diagram and reaction paths of active matter, uncovering unconventional nucleation mechanisms and stability properties.

Findings

Mean escape time from phase-separated state is exponentially extensive in system size.

Escape time from homogeneous state remains finite regardless of size.

Active terms enhance homogeneous phase stability, reducing phase separation in large systems.

Abstract

Nonequilibrium phase transitions are notably difficult to analyze because their mechanisms depend on the system's dynamics in a complex way due to the lack of time-reversal symmetry. To complicate matters, the system's steady-state distribution is unknown in general. Here, the phase diagram of the active Model B is computed with a deep neural network implementation of the geometric minimum action method (gMAM). This approach unveils the unconventional reaction paths and nucleation mechanism in dimensions 1, 2 and 3, by which the system switches between the homogeneous and inhomogeneous phases in the binodal region. Our main findings are: (i) the mean time to escape the phase-separated state is (exponentially) extensive in the system size $L$, but it increases non-monotonically with $L$ in dimension 1; (ii) the mean time to escape the homogeneous state is always finite, in line with the recent work of Cates and Nardini~[Phys. Rev. Lett. 130, 098203]; (iii) at fixed $L$, the active term increases the stability of the homogeneous phase, eventually destroying the phase separation in the binodal for large but finite systems. Our results are particularly relevant for active matter systems in which the number of constituents hardly goes beyond $10^7$ and where finite-size effects matter.