Predicting heavy metal impacts on soil ammonia oxidation during sludge land application: An interpretable framework integrating mechanistic models and machine learning

Heavy metal (HM)-induced ecological risks hinder the sustainable land application of treated municipal sludge, particularly impacting sensitive biogeochemical processes such as ammonia oxidation. Reliably predicting the long-term effects of HMs on ammonia oxidation remains challenging due to the complex interplay among metal bioavailability, microbial adaptation, and nitrogen transformations. In this study, a hybrid mechanistic-machine learning framework was developed to predict ammonia oxidation dynamics in sludge-amended soils under stress from Cd, Cr, Cu, Ni, Pb, and Zn. The mechanistic model, which incorporated time-delayed response of ammonia-oxidizing bacteria (AOB) to NH₄⁺ availability and bioavailable metal-specific inhibition, accurately simulated the dynamics of NH₄⁺ and ammonia oxidation rates (V_AOB^NH₄+) over two rounds of sludge application (R² = 0.9288-0.9506). Genetic algorithm optimization quantified the half-inhibitory content of bioavailable HMs for AOB (K_metal, mg·kg⁻¹), which followed the order: Cd (3.61) < Ni (19.54) < Cu (49.66) < Cr (58.43) < Pb (78.03) < Zn (134.10). By leveraging mechanistic data augmentation, the extreme gradient boosting and random forest models outperformed feedforward neural networks and support vector regression, achieving test R² > 0.96 for NH₄⁺ and test R² > 0.81 for V_AOB^NH₄+. Shapley additive explanations analysis suggested hormetic-like patterns of Cu and Pb for V_AOB^NH₄+ prediction, with transition points at bioavailable content of 27.08 and 30.18 mg kg⁻¹ in sludge-amended soils, respectively, whereas Zn above 73.04 mg kg⁻¹ contributed negatively to V_AOB^NH₄+. This study provides a novel and interpretable modeling framework for assessing HM-induced ecological risks in sludge-amended soils under AOB-dominated conditions.