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

Area of Science:

• Environmental Chemistry
• Analytical Chemistry
• Computational Chemistry

Background:

• Per- and polyfluoroalkyl substances (PFAS) are persistent environmental pollutants requiring advanced identification methods.
• Interpreting ¹⁹F NMR spectra for PFAS analysis is challenging due to limited reference data.
• New approaches are needed to advance PFAS impact assessment and remediation technologies.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting ¹⁹F NMR chemical shifts of PFAS.
• To address the knowledge gap in PFAS spectral data interpretation.
• To provide a practical tool for identifying emerging PFAS and supporting remediation.

Main Methods:

• Curated a dataset of 3616 chemical shifts from 647 fluorinated compounds.
• Explored various atomic feature descriptors and evaluated multiple ML algorithms.
• Developed a feed-forward neural network (FFNN) model and a confidence level system.

Main Results:

• The FFNN model achieved a mean absolute error of 2.40 ppm on test data, with 49% of predictions having errors <1.0 ppm.
• The model predicted chemical shifts for novel PFAS structures with up to 90% lower average errors than database methods.
• Validated utility through prediction of novel PFAS spectra, peak assignment assistance, and structural clarification in wastewater.

Conclusions:

• Machine learning offers a powerful approach to overcome challenges in PFAS ¹⁹F NMR spectral interpretation.
• The developed predictive tool can significantly support the identification and analysis of PFAS in environmental samples.
• This study advances PFAS remediation and impact assessment by providing a practical computational solution.