Machine learning-fugacity framework reveals the fate and environmental drivers of per- and polyfluoroalkyl substances...

Mountainous plateaus may act as condensers for persistent organic pollutants; however, the multimedia fate of per- and polyfluoroalkyl substances (PFAS) and the influence of environmental factors across the Third Pole-Tibetan Plateau remain largely unexplored. In this study, the occurrence and distribution of PFAS across multiple environmental media of the Yarlung Tsangpo River were comprehensively investigated, and the applicability of machine learning (ML) models for predicting contamination levels and identifying key drivers was evaluated using meteorological, water quality and anthropogenic factors as inputs. A multimedia fugacity model was established to simulate PFAS migration pathways and fate. Short-chain PFBA and PFBS dominated in water (ΣPFAS concentration: 1.05-20.4 ng/L), whereas 6:2 fluorotelomer sulfonate and C7HFPO-TrA prevailed in sediments and soils (ΣPFAS concentration: n. d.-17.3 and 0.03-5.67 ng/g dw, respectively). Increasing altitude was associated with elevated short-chain PFAS levels in both water and soils, suggesting a potential altitude-related compositional pattern. ML models demonstrated that meteorological and spatial geographic variables successfully predicted PFAS occurrence in water, with precipitation, mean annual temperature and solar radiation identified as the key factors. The fugacity model indicated that transport flux between water and sediment constituted the preferential migration pathway. Riverine sediment served as the dominant sink, intercepting 26.6% of the downstream PFAS flux. Based on previous reports that glacier runoff can release PFAS into high-altitude rivers, increasing glacier meltwater inputs under climate change may enhance PFAS accumulation in mountainous river sediments. This study presents a hybrid framework integrating field observations, ML and fugacity modelling to elucidate the environmental fate of PFAS under extreme alpine conditions, offering a framework that may be transferable to tracing and predicting trace organic contaminants in fragile, data-scarce ecosystems.