Coupling bayesian optimization and multi-paradigm ensemble learning to predict lead remediation in microbially...

Microbially Induced Carbonate Precipitation (MICP), recognized as a highly promising green remediation technology, exhibits lead (Pb) immobilization efficiency governed by the complex and nonlinear coupling of multiple biogeochemical factors. Consequently, achieving precise optimization through traditional experimental approaches remains a formidable challenge. To address this challenge, this study established a data driven framework coupling Bayesian Optimization (BO) with SHapley Additive exPlanations (SHAP). Based on a meticulously curated dataset of 168 high quality experimental records, we systematically evaluated six base models and four Stacking ensemble strategies, while comparing three hyperparameter optimization algorithms. The results indicate that, in small sample scenarios, a refined parameter optimization strategy outperforms complex model stacking. The Bayesian optimized Random Forest (BO-RF) model exhibited the superior generalization capability (R²=0.9035, RMSE=7.988). Furthermore, SHAP analysis successfully decoded the black box of the model, identifying pH, urea dosage, and initial lead concentration as the three dominant factors. The model, for the first time, quantitatively revealed the bell-shaped biphasic response of urea concentration, successfully reconstructing the underlying physicochemical logic of classical enzymatic kinetics (Michaelis Menten and Haldane mechanisms) and microenvironmental Pb speciation evolution (Biotic Ligand Model, BLM) within the algorithmic framework. More importantly, the identified global optimal thresholds (pH > 8.0, urea ≈ 20 g) precisely point to the formation boundary of thermodynamically stable calcite from statistical and thermodynamic perspectives, effectively averting the risks of explosive precipitation and metastable vaterite formation associated with excessive substrate concentrations. This study not only provides a high fidelity predictive tool (the MICP-Pb AI platform) but also elucidates the critical process windows for ensuring crystalline stability and acid resistance, thereby achieving the synergistic optimization of rapid Pb removal and long term secure sequestration.