Deep learning directed synthesis of fluid ferroelectric materials

Abstract

Fluid ferroelectrics, a recently discovered class of liquid crystals that exhibit switchable, long-range polar order, offer opportunities in ultrafast electro-optic technologies, responsive soft matter, and next-generation energy materials. Yet their discovery has relied almost entirely on intuition and chance, limiting progress in the field. Here we develop and experimentally validate a deep-learning data-to-molecule pipeline that enables the targeted design and synthesis of new organic fluid ferroelectrics. We curate a comprehensive dataset of all known longitudinally polar liquid-crystal materials and train graph neural networks that predict ferroelectric behaviour with up to 95% accuracy and achieve root mean square errors as low as 11 K for transition temperatures. A graph variational autoencoder generates de novo molecular structures which are filtered using an ensemble of high-performing classifiers and regressors to identify candidates with predicted ferroelectric nematic behaviour and accessible transition temperatures. Integration with a computational retrosynthesis engine and a digitised chemical inventory further narrows the design space to a synthesis-ready longlist. 11 candidates were synthesised and characterized through established mixture-based extrapolation methods. From which extrapolated ferroelectric nematic transitions were compared against neural network predictions. The experimental verification of novel materials augments the original dataset with quality feedback data thus aiding future research. These results demonstrate a practical, closed-loop approach to discovering synthesizable fluid ferroelectrics, marking a step toward autonomous design of functional soft materials.