Using machine learning to estimate the riverine fluxes of per- and polyfluoroalkyl substances to estuaries in 2000-2022, China

Per- and polyfluoroalkyl substances are persistent contaminants of global concern. China released an estimated 250 t of perfluorooctanoic acid (PFOA) and its salts during 2004-2012, but riverine transport to coastal waters remains poorly quantified. We compiled 4978 concentration records from 136 publications covering more than 100 rivers nationwide. The analysis based on the database revealed that PFOA dominated the PFAS profiles (median: 12.80 ng/L), followed by perfluorobutanoic acid (PFBA, 4.00 ng/L), while the remaining compounds were substantially lower (0.34-2.27 ng/L). Moreover, substantial spatial heterogeneity in PFAS concentrations was observed across Chinese rivers. We further developed machine learning models to predict nationwide riverine PFAS concentrations at six time points (2000-2022 at 5-year intervals). The random forest models achieved test-set adjusted R² values ranging from 0.44 to 0.75. By combining these predictions with river discharge data, the average annual seaward fluxes of five major PFAS-namely PFOA, perfluorooctanesulfonic acid (PFOS), perfluorohexanesulfonic acid (PFHxS), PFBA, and perfluorobutanesulfonic acid (PFBS)-were estimated to be 45.59, 17.45, 4.78, 9.27, and 8.24 t/yr, respectively. Fluxes exhibited pronounced spatial heterogeneity, with eastern and southeastern coastal rivers contributing the majority of inputs. Temporal variations were compound-specific and likely related to changes in production, usage patterns, and regulatory controls. Riverine PFOS fluxes represented approximately 18%-27% of reported PFOS emissions, a lower proportion than that for PFOA. Our results provide useful information for the regulation of both legacy and short-chain PFAS from inland rivers to coastal waters.