Deep learning directed synthesis of fluid ferroelectric materials.

We present a data-to-molecule deep-learning workflow for the targeted design of organic fluid ferroelectrics and validate it through the synthesis and experimental characterisation of machine-learning generated candidate materials. Fluid ferroelectrics, a 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. Without well-validated design rules, development of new fluid ferroelectric materials remains challenging. We curate a comprehensive dataset of all known longitudinally polar liquid-crystal materials and train graph neural networks to predict ferroelectric nematic transition temperatures. A graph variational autoencoder generates de novo structures which are filtered by an ensemble of 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 narrows the design space to a synthesis-ready longlist. Eleven machine-learning-selected candidates were synthesised and characterised, with one molecule exhibiting a modulated antiferroelectric phase. Extrapolated ferroelectric nematic transitions were obtained, compared against model predictions, and used to augment the original dataset. Further iteration using graph neural networks and XGBoost models produced a second ranked set of candidates; chemical synthesis yields fluid ferroelectric materials through a practical, closed loop-workflow.