关键词: 大数据挖掘, MOFid, 高通量筛选, 化学战剂, 机器学习, 分子模拟, 金属有机框架

To efficiently capture low-concentration chemical warfare agents (CWAs), which pose severe threats to human health and the environment, this study employed an AI big-data-driven high-throughput computational screening strategy to systematically analyze and evaluate the capture and separation performance of tens of thousands of metal-organic framework (MOF) adsorbents for trace CWAs in air. Grand Canonical Monte Carlo (GCMC) simulations were employed to evaluate the CWAs uptake of 15333 Computation-ready experimental MOFs (CoRE-MOFs) under a gas mixture containing N₂, O₂, and toxic gas at 298 K and 101.325 kPa. A trade-off score (TSN) combining selectivity and adsorption capacity was introduced as a comprehensive performance metric. Predictive models were constructed using six algorithms (Decision Tree, Random Forest, Gradient Boosting Regression Tree, Extreme Gradient Boosting (XGB), Backpropagation Neural Network, and Light Gradient Boosting Machine) based on seven key descriptors of MOFs— namely, Largest Cavity Diameter (LCD), density of MOF (ρ_MOF), Pore-Limiting Diameter (PLD), porosity (φ), Volumetric Surface Area (VSA), Henry Coefficient (K), and heat of adsorption (Q⁰_st). And the XGB algorithm yielded the best predictive performance, with the R² up to 0.923. ​​SHAP analysis revealed that K was the most critical descriptor for CWAs adsoprtion, with optimal φ and VSA ranges for maximal sarin and soman adsorption​​ being 0.7-0.8 and 2000-2500 m²/cm³, respectively​​. Then, the structural commonalities of the top 1% high-performance MOFs were analyzed by integrating the XGB algorithm with the MOFid standardized identifier. The results reveal that the synergistic effect between open transition metal sites (​​particularly Cr, Nb, Al, In, Li​​) and rigid organic linkers of MOFs enhances the adsorption affinity towards CWAs. Meanwhile, the high occurrence probability of specific topological structures (​​such as sql and kgm​​) indicates that they are conducive to forming favorable pore structures, thus enhancing the adsorption effect. ​​Guided by these insights, 26 novel MOFs were rational designed for enhanced CWAs capture. This study, employing AI big-data mining driven by MOFid method, provides critical guidance for optimizing MOF adsorption performance and screening highly efficient materials, thereby facilitating the efficient capture of trace chemical warfare agents in air ​​and advancing the development of protective equipment​​.

Key words: AI big-data mining, MOFid, High-throughput screening, Chemical warfare agents, Machine learning, Molecular simulation, Metal-organic frameworks