Machine learning-derived PFAS classifications for streamlining property calculations and experiments to support development of degradation approaches

Several thousand poly- and perfluoroalkyl substances (PFASs) have been developed and implemented for a wide variety of applications. Furthermore, the Organisation for Economic Co-operation and Development (OECD) recently put forth an expansive definition of PFAS. These substances exhibit highly desirable properties; however, there unique structural feature – a high number of carbon-fluorine bonds – leads to persistence of these compounds in the environment. Due to the large number of different PFAS and use cases, experimental methods are unable to exhaustively characterize these PFAS and fully understand their environmental interactions and fate. Computational chemistry methods can help close this data gap, but experimental data is still vital for validate computational predictions and calibrate models. We have developed a novel classification scheme in order to create subsets of PFAS to streamline calculations and experiment such that data representative of large number of PFASs can be generated from a smaller set. Using predictions and data collected by us, we explore different subsets generated under different ML models. Ultimately, the work of this effort is to support PFAS degradation approaches by providing necessary chemical properties.