Phase diagram construction of Poly (ethylene oxide) (PEO) and 1-butyl-3-methylimidazolium tetrafluoroborate ([BMIM][BF₄]) mixture with local affinity vector and unsupervised machine learning

We use machine learning methods to study the phase behavior of polymers in ionic liquids. We study a mixture of Poly (ethylene oxide) (PEO) and 1-butyl-3-methylimidazolium tetrafluoroborate ([BMIM][BF₄]), which exhibits lower critical solution temperature (LCST) behavior. The complexity of the system makes traditional Monte Carlo methods for the phase behavior computationally challenging. We develop a feature vector called ‘local affinity’ for use in unsupervised machine learning methods and use autoencoders to determine the phase behavior. The method is quantitatively accurate for model systems, such as Ising models and polymer blends, for which the phase diagram is knowns. For the PEO/[BMIM][BF₄] system the method makes predictions which can be tested via experiment. We expect that this method will suggest new way to study complex molecular systems with unusual phase behavior.