Today, liquid crystals (LCs) are common materials used in the flat screen displays that we see all around us. Yet, these displays only use one type of liquid crystal, the nematic phase, a phase which is characterised solely by orientational order of shape anisotropic molecules. In fact, there is a whole zoo of about 25 different liquid crystal phases that have been discovered over the last 130 years, which all vary in their ordering phenomena, displaying short-range one- and two-dimensional positional order, orthogonal vs tilted phases, different polar subphases, which show paraelectricity, ferro-, ferri-, and antiferroelectricity, all the way to the many different soft crystal materials. The plethora of liquid crystal phases, all separated thermodynamically by 1st or 2nd order phase transitions, allows for a wealth of applications far beyond displays.
These liquid crystal phases of novel compounds are generally characterised by texture observation in polarised optical microscopy (POM), which takes much practice and a good amount of expertise. And even then, additional methods are often needed, such as differential scanning calorimetry (DSC), x-ray diffraction and other more sophisticated methodologies. We show that machine learning can be employed to help researchers in characterising liquid crystals by automating texture characterisation with convolutional neural networks (CNNs) and inception algorithms. We discuss the machine learning performance with respect to number of layers, inception blocks, image augmentation etc. for different types of liquid crystal series, (i) chiral nematic-fluid smectic -hexatic smectic, (ii) orthogonal smectic and soft crystal phases, (iii) ferroelectric subphases, and (iv) novel nematic structures, including the much anticipated ferroelectric nematic phase. Accuracies in phase identification lie between 95-100%, depending on the amount of data used, the capacity of the algorithm and the phases involved.
We further demonstrate the detection of discontinuous and continuous phase transitions, as well as the localisation of topological defects. Finally, we discuss the determination of liquid crystal physical parameters from electro-optic curves.