Machine learning-driven source apportionment and health risk assessment of multi-source groundwater pollution in...

Industrial expansion releases multiple pollutants into groundwater, whose migration in highly permeable aquifers is governed by rapid flow through coarse sediments or fractured bedrock. This dynamic flow disperses and blends contaminant plumes from different sources, creating overlapping chemical signatures that mask their origins. Consequently, precise source identification is substantially hindered, posing significant challenges for groundwater management. This study integrates hydrogeochemical analysis with self-organizing maps to resolve multi-source groundwater pollution and assess health and environmental risks. The SOM approach effectively characterizes complex multi-source contamination and enables accurate source apportionment. Complementarily, hydrogeochemical analysis enhances the interpretability of the machine learning model and provides constraints that further refine pollution identification. Results indicate dominant anthropogenic impacts including industrial, domestic and agricultural activities, with five clusters were defined. Cluster 1 and Cluster 3 represent severely and moderately polluted zones, respectively, primarily attributed to vanadium-titanium magnetite mining and processing. Cluster 2 exhibits signatures of industrial activities on ammonium production and manure leachate, while Cluster 4 reflects domestic sewage impacts. Cluster 5 demonstrates reductive dissolution of Fe/Mn minerals under reducing conditions. Notably, industrial-sourced Mn poses elevated non-carcinogenic risks, exceeding safety thresholds in 1.57% of exposed adults and 9.74% of exposed children. Given the proximity of pollution hotspots to residential areas, targeted source controls are imperative. Our framework establishes a robust methodology for source apportionment in complex groundwater systems, enabling prioritized risk management through pollution assessment and tailored mitigation strategies.