Interpretable machine learning of metal‐ion adsorption by <scp>MOFs</scp> in aqueous systems: Cross‐study predictions and design rules

Abstract Metal–organic frameworks (MOFs) with tunable structural and physicochemical properties are promising adsorbents for removing toxic metal ions from aqueous systems. However, the interplay among MOF characteristics, ion properties, and operating conditions complicates rational design and prediction. Here, we assemble a cross‐study dataset of 209 adsorption‐capacity measurements from ten peer‐reviewed studies, comprising experimental adsorption conditions (solution pH, initial concentration C 0 , adsorbent dose), MOF‐specific properties (Brunauer‐Emmett‐Teller (BET) surface area, pore size, pore volume, pH pzc ), and pollutant‐specific properties (oxidation state, atomic mass, electronegativity, ionization energy, atomic radius). Four machine‐learning regressors were trained and evaluated; XGBoost achieved the best performance ( R 2 train = 0.9999, R 2 test = 0.9513). Model interpretability using Shapley additive explanations (SHAP) identified C 0 as the dominant factor, followed by solution pH in relation to pH pzc ; BET surface area and pore size also contributed substantially, while oxidation state and atomic mass had moderate effects. Collectively, these results demonstrate that interpretable machine learning can uncover governing factors in MOF–ion adsorption and provide practical design rules for tuning operating conditions and selecting MOFs for metal‐ion removal from water.