Machine-Learning-Enabled Interpretation of PFAS Adsorption Mechanisms Using the Freundlich Coefficient ( <i>K</i> _F ) as the Target Response

Per- and polyfluoroalkyl substances (PFAS) show strongly compound- and condition-dependent adsorption, creating uncertainty in activated-carbon treatment design, particularly for short-chain PFAS with weak and variable affinity. Here, an interpretable machine learning framework is developed to predict the Freundlich adsorption coefficient ( K F ) as a screening-level indicator of adsorption affinity within a harmonized PFAS–adsorbent–condition data set. A literature database of 256 extracted records was curated into 238 harmonized PFAS–adsorbent–condition entries (I F = 0–85.60). Data were split into 70% training, 15% validation, and 15% testing, with 3-fold cross-validation ( k = 3) used for hyperparameter optimization. Seven models (SVR, RF, GB, XGBoost, ANN, and KNN) were evaluated using R 2, RMSE, and MAE. Performance increased from conventional to advanced models: XGBoost was the best traditional baseline ( R 2 test = 0.77). Among the evaluated models, KNN showed the highest predictive performance (KNN: R 2 test = 0.89). SHAP interpretation identified influential descriptor domains consistent with known adsorption behavior, with pH, adsorbent identity, and temperature as key operational controls and molecular descriptors governing size/dispersion effects. Chain-length heatmaps of log 10 ( K F ) revealed distinct adsorption regimes and data set gaps, supporting predictive infilling for missing PFAS–adsorbent combinations. These combined approaches help investigate and control PFAS in the environment by supporting source tracking while accurately predicting K F and key drivers for faster screening and improved treatment design.