Predicting PFAS exposure risks from rural private wells using integrated mechanistic and machine-learned Bayesian network models

Abstract

Per- and polyfluoroalkyl substances (PFAS) pose complex challenges for drinking water management, particularly in rural communities reliant on private wells that fall outside federal regulatory oversight. As regulators and communities grapple with evolving standards and limited resources, there is a critical need for scalable, accurate, and resource-efficient models to assess PFAS exposure risks. This study compares traditional mechanistic groundwater models to machine-learned Bayesian networks and hybrid approaches that integrate both model types. Using hexafluoropropylene oxide dimer acid, commonly known as GenX, contamination near North Carolina's Fayetteville Works Plant as a case study, we evaluate each model's (1) ability to identify wells exceeding health-based GenX thresholds, (2) robustness to incomplete data, and (3) development and calibration time. Results demonstrate that Bayesian networks match or exceed mechanistic models in classification performance, while offering superior robustness to missing data and varying decision thresholds, along with decreased modeling resource demands. Hybrid models incorporating mechanistic outputs into Bayesian networks achieve the highest performance, showing that mechanistic insights can improve Bayesian network predictions. These findings support a flexible modeling framework that can adapt to changing regulations and emerging contaminants without recalibrating full process-based models, offering practical tools for regulators and communities to advance PFAS risk management.