Packing Efficiency Governs CO₂/H₂ Selectivity in Metal–Organic Frameworks: Uncertainty-Guided Machine Learning Screening of 278,778 Structures Reveals Topology-Level Reticular Design Rules

This Zenodo archive contains the data, trained models, candidate rankings, and analysis outputs associated with the manuscript “Packing Efficiency Governs CO₂/H₂ Selectivity in Metal–Organic Frameworks: Uncertainty-Guided Machine Learning Screening of 278,778 Structures Reveals Topology-Level Reticular Design Rules.” The study develops a reproducible machine-learning screening workflow for identifying metal–organic frameworks suitable for pre-combustion CO₂/H₂ separation. After data cleaning, 278,778 ARC-MOF structures were screened at 298 K and 40 bar using XGBoost-based surrogate models trained on a 77-dimensional descriptor set combining geometric descriptors, RAC principal components, RDF principal components, and REPEAT charge statistics. Four adsorption targets are predicted: CO₂ uptake, working capacity, CO₂/H₂ selectivity, and heat of adsorption. The archive supports full reproducibility of the manuscript workflow, including feature engineering, fixed train/test splits, model training, baseline comparison, cross-validation, split conformal uncertainty quantification, SHAP-based interpretation, Pareto-front analysis, topology–selectivity mapping, synthesizability assessment, weight-sensitivity testing, and final candidate prioritisation. The trained models achieve strong held-out test performance, with R² = 0.981 for CO₂ uptake, R² = 0.985 for working capacity, R² = 0.975 for CO₂/H₂ selectivity in log-space, and R² = 0.817 for heat of adsorption using a stacking ensemble. Split conformal prediction provides calibrated uncertainty intervals, allowing screened MOFs to be ranked by both predicted performance and prediction confidence. The central finding is that CO₂/H₂ selectivity is governed more strongly by packing efficiency than by pore diameter, while working capacity peaks near POAVAg ≈ 1.5 cm³ g⁻¹ and weakens beyond approximately 2 cm³ g⁻¹. Database-scale topology analysis identifies fof and fsc nets as high-selectivity structural families, with median CO₂/H₂ selectivities of 172.0 and 145.5, respectively, compared with a database-wide median of 89. These results provide topology-level reticular design rules for future MOF synthesis and validation. The screening workflow identifies an 803-structure viable candidate pool, a 4-member working-capacity/selectivity Pareto front, and 50 priority MOF candidates selected by four-target scalarisation. Among these candidates, 72% are rated High synthesizability based on metal-node tier and framework-family precedent. Candidate-level comparison against ARC-MOF GCMC values shows no systematic overprediction in the high-performance tail. This archive is intended to serve as an open, reusable database for extending uncertainty-guided machine-learning screening of MOFs toward carbon capture, hydrogen purification, adsorption-based separation, reticular materials design, and explainable materials informatics. All screening values are benchmarked against ARC-MOF GCMC data at 298 K; therefore, absolute capacities should be temperature-corrected before industrial process-level interpretation.