Detecting Liquid–Liquid Phase Separations Using Molecular Dynamics Simulations and Spectral Clustering

A stringent test of the accuracy of empirical force fields is reproducing the phase diagram of bulk phases and mixtures. Exploring the phase diagram of mixtures requires the detection of phase boundaries and critical points. In contrast to most solid-liquid transitions, in which a global order parameter (average density) can be used to discriminate between two phases, some demixing transitions entail relatively subtle changes in the local environment of each molecule. In such cases, finite sampling errors and finite-size effects make the identification of trends in local order parameters extremely challenging. Here we analyze one such example, namely a methanol/hexane mixture, and compute several local and global structural properties. We simulate the system at various temperatures and study the structural changes associated with demixing. We show that despite a seemingly continuous transformation between mixed and demixed states, the topological properties of the H-bond network change abruptly as the system crosses the demixing line. In particular, by using spectral clustering, we show that the distribution of cluster sizes develops a fat tail (as expected from percolation theory) in the vicinity of the critical point. We illustrate a simple criterion to identify this behavior, which results from the emergence of large system-spanning clusters from a collection of aggregates. We further tested the spectral clustering analysis on a Lennard-Jones system as a standard example of a system with no H-bonds, and also, in this case, we were able to detect the demixing transition.