Hierarchical Descriptor-Driven Machine Learning Reveals the Structure–Performance Landscape of Hydrogen-Bonded Organic Frameworks for CO 2 /N 2 Separation

Abstract Hydrogen-bonded organic frameworks (HOFs) are emerging as promising adsorbents for postcombustion CO2 capture owing to their mild self-assembly, solution processability, and facile regeneration, yet the structural and chemical factors governing their separation performance remain poorly resolved. Here, we report an integrated computational workflow combining high-throughput GCMC simulation, interpretable machine learning, and atomistic mechanism analysis to evaluate 84,261 HOFs for CO2/N2 separation under simulated flue-gas conditions. After benchmarking against experimental isotherms, a hierarchical four-level descriptor set spanning geometric, thermodynamic, elemental, and chemical/electronic features was constructed. Among eight algorithms, XGBoost delivered the highest accuracy (R2 = 0.93), and both feature importance and SHAP analyses identified geometric and thermodynamic descriptors as the dominant determinants of selectivity. High-performance HOFs converge on a common signature of smaller pores, denser frameworks, and enriched polar hydrogen-bond-donating groups (amino and carboxyl moieties), which cooperatively strengthen framework-CO2 interactions and widen the CO2/N2 heat-of-adsorption gap. Three-dimensional density mapping and geometry-based hydrogen-bond statistics confirm that preferential framework-CO2 hydrogen bonding is the key microscopic driving force, whereas N2 remains uniformly dispersed and forms essentially no effective hydrogen bonds. This work provides a high-fidelity predictive model and transferable design rules that advance HOF-based CO2/N2 separation from isolated breakthroughs toward mechanism-guided systematic design.