Machine Learning-Assisted Discovery of Chiral Azobenzene Dopants for Liquid Crystals

Cholesteric liquid crystals possess unique properties, such as selective reflection and circular polarization, due to their helical structures. The helical structures are determined by the helical twisting power (HTP) of chiral dopants. Chiral azobenzene compounds can change their HTP through photoisomerization, enabling photoresponsive tuning of the helical structure. For device applications, a large change in HTP value (ΔHTP) is desirable. In this study, we report a screening workflow to generate and identify chiral azobenzene compounds with high ΔHTP using machine learning (ML) and molecular fingerprint feature importance analysis. We trained ML classification models using ΔHTP values of 35 azobenzene compounds. Furthermore, virtual screening was performed on 11,184 azobenzene derivatives. The screened azobenzene was synthesized and was found to exhibit the highest ΔHTP of 43 [1/μm] among those in the training dataset. These results demonstrate that combining machine learning and feature analysis is a powerful approach for designing chiral compounds.