Predicting multi-heavy-metal adsorption on bimetallic zeolitic imidazolate frameworks via interpretable machine learning: From latin hypercube design to web implementation.

Efficient remediation of heavy metal-contaminated water requires systematic evaluation of adsorbent performance under complex environmental conditions. In this study, an interpretable machine learning approach was applied to predict multi-metal adsorption on bimetallic zeolitic imidazolate frameworks (ZIFs) and to analyze the contributions of material properties, aqueous chemistry, and metal ion characteristics. Three structurally distinct bimetallic ZIF-derived materials were synthesized and evaluated via adsorption experiments designed using Latin Hypercube Sampling (LHS) over a wide range of environmentally relevant conditions. Among five machine learning models, a neural network optimized using grid search achieved the best predictive performance (R2 = 0.98) under strict data leakage control. Shapley Additive Explanations (SHAP) analysis quantified the relative importance of system parameters, revealing that the ionic properties of heavy metals (48.1 %) showed a stronger influence on adsorption capacity than experimental conditions (39.9 %) or adsorbent characteristics (12.0 %). Pore-related parameters were found to influence adsorption mainly through diffusionaccessibility and pore density rather than pore size alone. A web-based prediction platform was further developed, yielding a mean absolute percentage error (MAPE) of 18 % on independent validation data. This work offers both fundamental insights into heavy metal-ZIF interactions and a practical framework for developing targeted water remediation technologies.