Machine learning-driven high-throughput prediction of Hg0 capture in MOFs and decipherment of structure-property relationships.

Abstract

Metal-organic frameworks (MOFs) show potential for Hg0 adsorption via tunable pore structures and chemical functionalities. However, efficient screening of MOF adsorbents remains challenging due to vast material space and unclear structure-property relationships. Herein, a Hg0 adsorption database of 6528 MOF samples was established, and machine learning models were developed for predicting adsorption performance. Among tested models, extreme gradient boosting (XGBoost) exhibited the best prediction on the independent test set, achieving a coefficient of determination (R2) of 0.893. SHAP analysis revealed that energy descriptors, including logarithmic Henry coefficient (log(KH)) and adsorption enthalpy (ΔHads), dominated Hg0 adsorption, while structural descriptors further regulated adsorption behavior. Combined model interpretation revealed a condition-dependent interaction between adsorption affinity and pore characteristics. High-throughput screening revealed characteristic descriptor intervals for high-performing MOFs, including log(KH)=-3.62 to -2.61, ΔHads = -34.60 to -25.21 kJ·mol-1, and VF = 0.67-0.78. MOFid-based structural analysis identified significant enrichment of Cr, Fe, Co nodes and nbo topology, alongside preferred acs and tbo topologies, as well as nitrogen-containing aromatic conjugated heterocyclic ligands in high-performing MOFs. These findings clarify the roles of metal nodes, topology, and ligand chemistry in Hg0 adsorption and provide a machine learning-driven strategy for rational design of MOF-based mercury removal adsorbents.