Data-driven rational design of functionalized COFs for PFAS remediation via DFT and machine learning integration

Per- and polyfluoroalkyl substances (PFAS) are persistent pollutants that resist conventional water treatment methods, raising serious environmental and health concerns. Covalent organic frameworks (COFs), due to their high porosity and tunable chemistry, have emerged as promising materials for PFAS remediation. AB-COF in particular present a robust, synthetically accessible, and highly stable scaffold, for designing ideal COFs for targeted functionalization and PFAS remediation. However, studies on the rational design of effective COFs with optimal binding properties for efficient and selective PFAS removal are still lacking. Herein, we present an integrated framework that combines density functional theory (DFT) with machine learning (ML) to accelerate the discovery of high-performance COFs for PFAS removal. Functionalized AB-COFs were systematically modeled and their interactions with representative PFAS molecules were evaluated using DFT-derived structural, electronic, and thermodynamic descriptors including pore topology, ionization energy, energy gap, and adsorption enthalpy. These features were used to train fourteen machine learning algorithms for both regression and classification. To address data limitations, synthetic augmentation was employed, enabling a marked improvement in predictive accuracy and generalizations. Consequently, Ensemble and boosting-based ML models, especially Extra Trees, CatBoost, and XGBoost, achieved remarkable predictive accuracy after synthetic data augmentation, with R² values exceeding 0.99 and RMSE as low as 1.0 kJ mol⁻¹. SHAP analysis further revealed that adsorption enthalpy and Gibbs free energy are the dominant predictors of PFAS uptake, while pore size and surface area exerted secondary effects. Notably, COFs functionalized with hydrophilic and anionic groups (–SO₃H, –PO₃H, –SiOH) achieved the highest PFAS adsorption energies (up to ∼710 kJ mol⁻¹). These results establish a data-driven and physically interpretable approach for designing and optimizing high-performance COF adsorbents for PFAS remediation and could be further explored.