Characterizing seasonal changes in perfluoroalkyl substances in lake sediment by environmental magnetism from a...

Seasonal mechanisms controlling perfluorinated substances (PFASs) in lake sediments remain poorly understood. This study employed interpretable machine learning (ML) to analyze seasonal variations of PFASs in a large shallow lake, and interpretable methods to quantify the influence of environmental and magnetic parameters. PFAS concentrations were higher in spring than in the other three seasons (p < 0.01); however, there was no significant spatial distribution (p > 0.05). Redundancy analyses showed that environmental factors explained the distribution of PFASs (R² = 0.47), with PFOA, PFOS, PFBA, and ΣPFASs mainly associated with particulate matter entering the lake, while PFBS and PFHxS were linked to industrial sources. Multiple linear regression revealed a strong correlation between PFASs and environmental factors (R² = 0.68), with TC, χ_ARM, and χ_ARM/χ showing negative correlations, and χ and SIRM showing positive correlations. Compared with XGBoost, CatBoost, SVM, and Bayesian neural networks (BNN), the random forest demonstrated the best performance for different PFASs simulation (R² = 0.85 on average). Incorporating spatial information into ML enhanced the R² by >20 %. Lake environmental chemical factors predominated (58.87 %), followed by magnetic factors (32.57 %) on PFASs simulation, based on feature importance. Partial dependence plots showed that when total carbon <6 g/kg or χ_ARM/χ < 10 × 10³A/m, they were negatively correlated with PFAS concentration. Force plots indicated PFAS increasing attributable to exogenous particle inputs, and the decreasing in autumn and winter may reflect internal transformations and particle movements. This study deciphered complex seasonal PFAS dynamics in sediments, and highlights the lifecycle mechanism requiring consideration in PFAS seasonal distribution and influence factors.