Abstract

Soil contamination by heavy metals endangers the environment, food security, and public health alike. Existing analytical approaches to estimate how metals bind to soil particles typically employ Langmuir and Freundlich isotherms; yet these models cannot adequately represent the intricate, nonlinear interactions among soil chemistries and the dissolved metals. Instead, we applied Artificial Neural Networks (ANNs) to cope with the inherent complexity. A series of feed-forward networks received, as input layers, key physicochemical variables soil pH, cation exchange capacity (CEC), organic matter percentage, clay fraction supplemented with initial concentrations of lead (Pb), cadmium (Cd), copper (Cu), and zinc (Zn). Hidden layers captured the multivariate nonlinear spectra, and the output layer quantified adsorbed mass for each species. The network training employed diverse datasets curated from laboratory batch experiments and reached R² coefficients surpassing 0.95 in nearly all validation folds. Imbalanced or distorted data cases, often compromising conventional models, did not derail the ANNs. The speed with which realistic field samples can be processed renders the approach practical for frequent, nationwide screening. By incorporating ANNs into geochemical diagnostics, the research provides authorities and industries with a low-cost, yet robust, forecasting protocol for assessing the binding potential of legacy or emerging contaminants, thus guiding targeted analytical follow-ups and optimizing decontamination design schemes.