Simulation, prediction and optimization of heavy metal adsorption by metal-organic frameworks with machine learning

The unique structures and complex characteristics of Metal-organic frame (MOFs) obscure understanding the processes and mechanisms of heavy metal (HM) removal. This study established an interpretable machine learning (ML) framework predicting adsorption capacities for six HMs across 89 MOF composites using a dataset of 1225 points with 44 features. The optimized combined gradient boosting decision tree model achieved exceptional accuracy (test R² = 0.921–0.962; Zn(II) external validation R² = 0.914). The integration of Shapley additive explanations with partial dependence plots was employed to interpret model results. And the importance of adsorption conditions, synthesis parameters, adsorbent properties, HM characteristics, and functional groups were 38.99 %, 20.39 %, 19.69 %, 12.19 %, and 8.74 %, respectively. Adsorption exhibited a triphasic response to pH: enhanced within pH 2–6, moderately reduced at pH 6–8, and inhibited above pH 8–10, driven by multi-parameter interactions. The synthesis conditions of MOF were optimized, including optimal pore size matching (0.5–3 nm), temperature-controlled crystallinity (100–200 °C), and drying time limitation (<10 h). The framework bridges ML predictions with adsorption chemistry, enabling data-driven material optimization for environmental remediation.