Assessment of groundwater vulnerability to heavy metals in four aquifers using machine learning algorithms

Using a machine learning framework, this study investigates the spatial distribution and key environmental factors of heavy metal contamination (iron, nickel, lead, copper) in four groundwater aquifers of Isfahan province during the water year 2023-2024. A total of 150 wells were sampled and metal concentrations were determined using ICP-MS, AAS, VGA, and Mercury Analyzer methods in accordance with WHO and Iranian standards. The maximum observed concentrations of iron, nickel, lead and copper were approximately 48, 44.1, 2.9 and 11.2 mg/L respectively, with the peak concentrations of iron and copper in the Damaneh - Daran aquifers, nickel in Bouin and lead in Chadegan. Random Forest (RF) and Support Vector Machine (SVM) models were used, and in RF, 100 trees were used for accurate predictions. Multiple collinearity between environmental predictors, including soil properties, unsaturated and saturated zones, hydraulic parameters, slope, groundwater level, and aquifer depth was assessed through variance inflation factor (VIF), all of which were below 10. Model interpretation showed that soil properties and groundwater level had the greatest influence in RF, while the unsaturated layer was dominant in SVM. Iron decreased with increasing aquifer depth, pore thickness, and water table, while soil permeability and slope increased iron accumulation. Nickel was higher in shallow, shallow, and low-conductivity areas, while lead increased with depth and slope, indicating a nonlinear dependence on hydraulic and soil properties. Copper was positively correlated with soil permeability and negatively correlated with water table. Spatial predictions showed that the Bouin aquifer showed the highest iron and nickel (more than 40 and more than 30 mg/L), lead reached about 44 mg/L in Chadegan, and copper peaked in Bouin from southeast to northwest. RF outperformed SVM by achieving an accuracy of 0.7874, sensitivity of 0.7448, and specificity of 0.8243, while SVM performed poorly. This study innovatively combines machine learning models with the parameters of the DRASTIC analytical model to assess and predict heavy metal contamination in the aquifers of Isfahan province. Overall, the results confirm the nonlinear hydrogeological controls on heavy metal distribution and demonstrate the high capability of RF for reliable prediction of groundwater contamination. This approach provides a transferable method for groundwater quality assessment and supports sustainable aquifer management in arid and semi-arid regions.