Abstract
Mustard gas (HD) is a highly toxic blistering chemical warfare agent that poses risks to health and safety. To enable safer investigation, experimental studies commonly use simulants such as diethyl sulfide (DES) as substitutes. However, molecular simulation studies on HD simulants remain limited. This work integrates molecular simulations, machine learning (ML), and metal-organic framework identifier (MOFid) analysis to identify hydrophobic metal-organic frameworks (MOFs) with DES adsorption capacity and explore structure-performance relationships. Univariate analysis evaluates the effects of key structural descriptors on adsorption behavior. Four ML models, coupled with three interpretability techniques-permutation feature importance (PFI), shapley additive explanations (SHAP), and partial dependence plots (PDP)-are employed to assess the influence of structural features on performance. Notably, extending PDP analysis to two- and three-dimensional spaces allows synergistic effects between structural descriptors to be explicitly visualized, which cannot be resolved using traditional one-dimensional PDPs. MOFid analysis over 30,000 hydrophobic MOFs further reveals performance-dependent distributions of metal nodes, linkers, and topologies. High-performing MOFs are enriched in vanadium nodes and rna topologies, suggesting synergistic contributions, while ten rare organic linkers appear exclusively in top candidates, indicating performance-determining roles. Together, these results define a data-driven framework integrating simulation, ML, and structural informatics for efficient screening, guiding hazardous-gas adsorption MOF design and improving interpretability.