Tao Jia

Comprehensive monitoring of per- and polyfluoroalkyl substances (PFASs) in drinking water sources is challenging because sampling intensity, analytical coverage, reporting limits, and well construction vary among monitoring programs. We developed a machine learning pipeline to characterize environmental and pollution source susceptibility to PFAS occurrence in municipal and public supply groundwater wells across California. The assembled covariate inventory contained 7951 public supply wells; 7938 wells contributed 60,744 first-classifiable well–compound records for the eight-compound modeling analysis. Occurrence was defined as a concentration greater than 2 ng/L at the first-classifiable test for each well–compound pair. Eight compounds met prespecified support criteria: PFOS, PFHxS, PFOA, PFBS, PFHxA, PFPeA, PFBA, and PFHpA. Separate binary classifiers were evaluated using five spatially blocked folds and a forward temporal test for first-tested wells from 2024 onward. Across the eight compounds, calibrated Extra Trees achieved a mean spatial ROC-AUC of 0.722 and a forward temporal ROC-AUC of 0.669. Mean PR-AUC was 0.350 for spatial validation (mean positive rate, 16.2%) and 0.223 for forward temporal validation (mean positive rate, 9.6%), which indicates moderate discrimination that supports ranking wells for confirmatory sampling but not classifying individual wells. The spatial ensemble generated 1 km screening surfaces for PFOS, PFOA, and PFHxS with explicit model support and extrapolation diagnostics. These surfaces represent susceptibility under a typical public supply well scenario and are intended to prioritize confirmatory monitoring, not establish contamination at unsampled wells, attribute occurrence to specific facilities, or validate 1 km positional accuracy.