Classification of liquid crystal textures using convolutional neural networks

ABSTRACT

We investigate the application of convolutional neural networks (CNNs) to the classification of liquid crystal phases from images of their experimental textures. Three CNN classifier model types (Sequential, Inception and ResNet50) are tuned and trained on five individual phase group datasets. The complete dataset includes images of the cholesteric phase, chiral fluid smectic A and C phases and hexatic smectic I and F phases, all extracted from polarised microscopy videos of various liquid crystalline compounds. Three binary classification tasks, each including two liquid crystal phases, provide the foundational demonstration of CNN model viability. The average test set accuracies obtained are approximately (95 ± 2)%. More complex multi-phase datasets are also created and investigated, with a three-phase cholesteric, fluid smectic, and hexatic smectic set, in addition to a set containing all five individual phases. The average test set accuracies for these classification tasks are (85 ± 2)%.

GRAPHICAL ABSTRACT

KEYWORDS:

• Machine learning
• convolutional neural networks
• liquid crystals
• texture characterisation
• phase transitions

Disclosure statement

No potential conflict of interest was reported by the author(s).