Data-driven prediction of dielectric anisotropy in nematic liquid crystals

We curate a large-scale dataset of low frequency dielectric anisotropy values for low molecular weight liquid crystals. Using this dataset, we demonstrate that supervised machine-learning models can predict dielectric anisotropy with substantially improved accuracy (RMSE 2.6) compared to estimates obtained from the Maier-Meier relations using molecular properties from both the widely used semiempirical AM1 method (RMSE 9.7) and the modern r2scan-3c composite method (RMSE 11.2). Realising the potential of machine learning techniques for liquid crystalline materials requires carefully curated data to be accessible, and on this basis, we propose a simple and standard template for reporting data.