Deep learning techniques for the localization and classification of liquid crystal phase transitions

Abstract

Deep Learning techniques such as supervised learning with convolutional neural networks and inception models were applied to phase transitions of liquid crystals to identify transition temperatures and the respective phases involved. In this context achiral as well as chiral systems were studied involving the isotropic liquid, the nematic phase of solely orientational order, fluid smectic phases with one-dimensional positional order and hexatic phases with local two-dimensional positional, so-called bond-orientational order. Discontinuous phase transition of 1st order as well as continuous 2nd order transitions were investigated. It is demonstrated that simpler transitions, namely Iso-N, Iso-N*, and N-SmA can accurately be identified for all unseen test movies studied. For more subtle transitions, such as SmA*-SmC*, SmC*-SmI*, and SmI*-SmF*, proof-of-principle evidence is provided, demonstrating the capability of deep learning techniques to identify even those transitions, despite some incorrectly characterized test movies. Overall, we demonstrate that with the provision of a substantial and varied dataset of textures there is no principal reason why one could not develop generalizable deep learning techniques to automate the identification of liquid crystal phase sequences of novel compounds.