Deep learning directed synthesis of fluid ferroelectric materials

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.

💡 Research Summary

The authors present a closed‑loop, data‑driven workflow for the discovery of organic fluid ferroelectric (ferroelectric nematic, NF) materials, integrating deep‑learning models with synthetic planning and experimental validation. They first assembled a curated dataset of more than 700 small‑molecule liquid‑crystal compounds, comprising every reported NF material up to October 2025 and an additional 200 unpublished entries, each annotated with SMILES strings and measured phase‑transition temperatures. Molecular graphs were generated from SMILES using RDKit, and three types of graph neural networks (GNNs) were trained: (i) binary classifiers for NF presence, (ii) regression models for transition temperature, and (iii) a graph variational autoencoder (VAE) for de‑novo molecule generation. The classifiers achieved up to 95 % accuracy, while the best regression models yielded root‑mean‑square errors (RMSE) as low as 11 K.

The VAE, a modified MGVAE architecture, learned a probabilistic latent space from the experimental liquid‑crystal molecules. Sampling this space and decoding the vectors produced 15 413 novel molecular graphs, which were canonicalized to SMILES and deduplicated. Structural analysis showed that the generated set retained the overall carbon‑count distribution of the training data but explored broader ranges of ring count, fluorination, and introduced new functional groups such as ketones, indicating genuine exploration beyond memorization.

For screening, the three highest‑performing classifiers and three lowest‑RMSE regressors were combined into ensembles. A molecule received a positive NF label only if at least nine of twelve classifiers voted “yes”; the mean predicted transition temperature from the regression ensemble was used for ranking. This dual‑filtering dramatically reduced the candidate pool while preserving high‑confidence NF prospects.

Synthetic feasibility was assessed by linking the filtered candidates to a retrosynthesis engine (based on publicly available tools) and to a digitized inventory of reagents stocked in the authors’ laboratory. The engine proposed synthetic routes using only commercially available precursors, and human chemists reviewed the suggestions to produce a synthesis‑ready longlist. Eleven compounds were selected for experimental work, balancing predicted NF likelihood and structural diversity.

All eleven compounds were synthesized using standard organic reactions, purified by flash chromatography, and characterized by ¹H/¹³C/¹⁹F NMR, HRMS, and HPLC (purity > 99 %). Their liquid‑crystal behavior was probed by polarized optical microscopy (POM) and differential scanning calorimetry (DSC) in mixtures containing 90 wt % of the known NF host DIO and 10 wt % of the new guest. Transition temperatures were extrapolated from these mixtures. Every synthesized molecule displayed an NF phase, confirming the model predictions. The experimental transition temperatures correlated with the predicted values with an RMSE of ≈ 25 K; predictions were most accurate for molecules structurally similar to the training set, while larger deviations occurred for out‑of‑distribution structures.

The study demonstrates that (1) graph‑based deep learning can reliably predict both the presence of a complex mesophase and its transition temperature from molecular structure; (2) a variational autoencoder can generate chemically realistic yet novel candidates that occupy adjacent regions of chemical space; (3) ensemble consensus provides a practical confidence metric for prioritizing synthesis; and (4) integration with automated retrosynthesis and inventory data can rapidly translate virtual designs into laboratory‑ready targets. Limitations include the modest size of the training dataset, which restricts extrapolation to highly dissimilar chemistries, and the residual temperature prediction error for such outliers. Future work will expand the dataset, incorporate physics‑based constraints (e.g., quantum‑chemical dipole calculations), and explore multi‑task learning to further improve accuracy and discovery speed.

Overall, this work establishes the first end‑to‑end, AI‑driven pipeline that moves from raw data to experimentally verified fluid ferroelectric materials, offering a blueprint for autonomous design of functional soft matter and accelerating the development of next‑generation electro‑optic and energy‑conversion technologies.