Convolutional Neural Network analysis of optical texture patterns in liquid-crystal skyrmions

TL;DR

This paper demonstrates that convolutional neural networks can analyze optical images of liquid crystal skyrmions to accurately predict system parameters, offering a new approach for material characterization and potential applications.

Contribution

It introduces a machine learning method focusing on skyrmion regions in optical images to efficiently predict key liquid crystal parameters, reducing computational costs.

Findings

CNNs accurately predict free energy, cholesteric pitch, and electric field strength from skyrmion images.

Focusing on skyrmion regions enhances computational efficiency and prediction accuracy.

The approach enables advanced material characterization using optical microscopy and machine learning.

Abstract

Liquid crystals are known for their optical birefringence, a property that gives rise to intricate patterns and colors when viewed in a microscope between crossed polarisers. Resulting images are rich in geometric patterns and serve as valuable fingerprints of the liquid crystal's intrinsic properties. By using machine learning techniques, it is possible to extract from the images information about, e.g., liquid crystal elastic constants, the scalar order parameter, local orientation of the director, etc. Machine learning can also be employed to identify phase transitions and classify different liquid crystalline phases and topological defects. In addition to well studied singular defects such as point or line disclinations, liquid crystals can also host non-singular solitonic defects such as skyrmions, hopfions, and torons. The solitons, with their localised and stable configurations, offer an alternative view into material properties and behaviour of liquid crystals. In this study, we demonstrate that the optical signatures of skyrmions can be utilised effectively in machine learning to predict important system parameters. Our method focuses specifically on the skyrmion-localised regions, reducing significantly the computational cost. By training convolutional neural networks on simulated polarised optical microscopy images of liquid crystal skyrmions, we showcase the ability of trained networks to accurately predict several selected parameters such as the free energy, cholesteric pitch, and strength of applied electric fields. This study highlights the importance of localized topologically arrested order parameter configurations for materials characterisation research empowered by state-of-the-art data science methods, and may pave the way for the development of advanced skyrmion-based applications.