A Machine Learning Approach for Predicting Defluorination of Per- and Polyfluoroalkyl Substances (PFAS) for Their Efficient Treatment and Removal

Abstract

We present the first application of machine learning on per- and polyfluoroalkyl substances (PFAS) for predicting and rationalizing carbon-fluorine (C–F) bond dissociation energies to aid in their efficient treatment and removal. Using a variety of machine learning algorithms (including Random Forest, Least Absolute Shrinkage and Selection Operator Regression, and Feed-forward Neural Networks), we were able to obtain extremely accurate predictions for C–F bond dissociation energies (with deviations less than 0.70 kcal/mol) that are <i>within chemical accuracy</i> of the PFAS reference data. In addition, we show that our machine learning approach is extremely efficient (requiring less than 10 minutes to train the data and less than a second to predict the C–F bond dissociation energy of a new compound) and only needs knowledge of the simple chemical connectivity in a PFAS structure to yield reliable results – without recourse to a computationally expensive quantum mechanical calculation or a three-dimensional structure. Finally, we present an unsupervised machine learning algorithm that can automatically classify and rationalize chemical trends in PFAS structures that would otherwise have been difficult to humanly visualize/process manually. Collectively, these studies (1) comprise the first application of machine learning techniques for PFAS structures to predict/rationalize C–F bond dissociation energies and (2) show immense promise for assisting experimentalists in the <i>targeted</i> defluorination of specific bonds in PFAS structures (or other unknown environmental contaminants) of increasing complexity.