A Machine Learning-Assisted Liquid Crystal Droplet Array Platform for the Sensitive and Selective Detection of Per- and Polyfluoroalkyl Substances (PFAS) in Water

We report a machine learning (ML)-assisted liquid crystal (LC) droplet array platform for the detection of per- and polyfluoroalkyl substances (PFAS) in water. Our approach uses an autoencoder network to process thousands of images obtained from arrays of microscale droplets of thermotropic LCs. The latent space obtained using the autoencoder contains significant information that enables sensitive and selective detection of two amphiphilic PFAS [perfluorooctanoic acid (PFOA) and perfluorooctanesulfonic acid (PFOS)] at concentrations as low as parts-per-trillion (ppt) in ultrapure water, municipal tap water, and simulated river water containing dissolved organic matter. Despite the absence of visually discernible changes in the optical outputs of LC arrays at low PFAS concentrations, this approach accurately predicts their presence, even in water containing interfering molecules. We also demonstrate the use of transfer learning to differentiate between PFOA, PFOS, and PFOA/PFOS mixtures, showcasing the potential for practical environmental monitoring. This platform permits identification of PFOA and PFOS below the maximum contaminant levels (4 ppt) established by the U.S. Environmental Protection Agency. Our approach is compatible with automated printing, treatment, and high-throughput optical and ML analysis and could provide a basis for the development of low-cost sensors to monitor PFAS and other amphiphilic contaminants in real-world water samples.