Machine learning-assisted exploration of covalent organic frameworks for short-chain per- and polyfluoroalkyl substances (PFAS) removal from water

Short-chain per- and polyfluoroalkyl substances (PFAS) are increasingly being adopted as alternatives to long-chain PFAS, which have been regulated or banned in several countries due to their persistence in water and soil and associated human health risks. However, short-chain PFAS are also environmentally persistent and highly mobile, and research on their adsorptive removal from water remains scarce. This challenge is compounded by the poor performance of conventional adsorbents for short-chain PFAS. In this study, we employed machine learning-assisted computational molecular modeling to investigate the removal of perfluorobutanoic acid (PFBA), a representative short-chain PFAS, using covalent organic frameworks (COFs). We examined the structure-performance relationship governing PFBA selectivity in the presence of water. Our results show that the type of linkage plays a critical role in PFBA selectivity. Specifically, triazine-, borazine-, boronate ester-, and borosilicate-based COFs exhibit the highest PFBA selectivity. However, because boron-based COFs are known to have poor water stability, triazine-based COFs emerge as the most promising candidates for PFBA removal from water. The impact of fluorine functionalization depends strongly on the linkage type: it enhances selectivity in azine- and imine-based COFs but reduces it in symmetric hexagonal linkages such as borazine, triazine, and boroxine. This reduction is attributed to fluorine's high electronegativity disrupting the electron density distribution in these symmetric structures.