Matrix-specific PFAS source allocation machine learning-based models: Identifying differential indicators in soil and...

PER: and polyfluoroalkyl substances (PFAS) pose a significant environmental and human health risk due to their persistence, bioaccumulation, mobility, and toxicity. Current PFAS source-allocation methods often demand data on many compounds while overlooking differences in matrix type nature, leading to high costs and potential errors. In this study, matrix-specific machine-learning classifiers were developed to discriminate PFAS originating from aqueous film-forming foam (AFFF) versus non-AFFF sources in soil and water. A comprehensive dataset was compiled from peer-reviewed literature (2012-2024), comprising 673 soil samples (437 AFFF-impacted, 263 non-AFFF), 520 water samples (379 AFFF-impacted, 141 non-AFFF), and 111 original AFFF formulations. Log-transformed concentrations of 12 legacy PFAS compounds were analyzed using fifteen classification algorithms through H2O.AutoML platform. The optimal water model (Gradient Boosting Machine) achieved an area under the curve (AUC) of 0.9864, accuracy of 0.8929, sensitivity of 0.9286, and specificity of 0.8571. The optimal soil model (Distributed Random Forest) achieved an AUC of 0.9936, accuracy of 0.9787, sensitivity of 0.9787, and specificity of 0.9787. Feature-importance analysis revealed PFOS, PFHxS, and PFPeS as the strongest water indicators, and PFHxS, PFPeA, and PFOS in soil. Stepwise variable reduction showed that source allocation accuracy above 0.92 can be maintained using only nine PFAS indicators in water and six in soil, significantly reducing the data requirements. This matrix-specific approach recognizes how PFAS compounds behave differently across environmental media, while identifying key "sentinel" indicators reduces analytical burden without sacrificing classification performance. This targeted methodology enhances forensic source tracking capabilities while providing more efficient guidance for remediation efforts.