Advancing PFAS Detection through Machine Learning Prediction of 19F NMR Spectra

Per- and polyfluoroalkyl substances (PFAS) are persistent environmental pollutants with diverse structures. To further advance the impact assessment and remediation technology for PFAS pollution, new approaches for identifying emerging PFAS are necessary. While 19F nuclear magnetic resonance (NMR) spectroscopy has unique advantages in analyzing PFAS in complex sample matrices, the spectra interpretation remains challenging due to the limited reference data. Herein, we address this significant knowledge gap using machine learning (ML) to predict 19F NMR chemical shifts of PFAS. We curated a data set comprising 3616 chemical shifts from 647 fluorinated compounds, explored various atomic feature descriptors for modeling, and evaluated multiple ML algorithms. The feed-forward neural network (FFNN) model performed the best, achieving a mean absolute error of 2.40 ppm on the test data set. Notably, 49% of predictions had errors <1.0 ppm and 86% < 5.0 ppm. The model predicted the chemical shifts of novel F atom configurations with up to 90% lower average errors than a database-driven method. We also developed a confidence level system (1–6) to provide error estimation for each prediction and guide future data set expansion toward low-confidence structures. The utility of the model was further validated through (i) prediction of NMR spectra of novel PFAS compounds, (ii) assistance in peak assignment, and (iii) structural clarification of an unknown PFAS in a real wastewater sample. Overall, this study demonstrates the advantages of ML and offers a practical predictive tool to support PFAS analysis.