Interpretable machine learning reveals geometry-dominated design rules for atmospheric CO2 capture in metal-organic frameworks.

Direct air capture (DAC) requires sorbents capable of binding CO2 at a partial pressure of 0.0004 bar, where conventional flue-gas screening criteria may not define absorption performance. Machine-learning (ML) surrogates for Grand Canonical Monte Carlo (GCMC) simulation can accelerate screening of metal-organic frameworks (MOFs) for DAC applications. However, the relative performance of compositional featurization approaches under DAC conditions remains insufficiently explored. In this study, CO2 uptake at 0.0004 bar and 298 K was derived for 2523 experimentally realized CoRE-MOF-2019 frameworks by Langmuir fitting of NIST GCMC isotherms yielding a mean fit R2 of 0.987. Two ML pipelines were developed using the same data and splits, and combined CoRE-MOF geometric descriptors with either 93 Matminer compositional features or 53 JARVIS-CFID chemical-interaction features. Optuna-optimized gradient-boosting regressor and LightGBM models achieved test-set R2 of 0.82 and 0.83 for Matminer and JARVIS-CFID feature spaces, respectively. External validation on hold-out sets comprising 839 frameworks yielded R2 values of 0.76-0.81 with Pearson correlation coefficients above 0.92. Feature-importance, SHAP, and LIME analyses consistently showed that volumetric surface area, Langmuir saturation capacity, and pore dimensions carry more than 90% of model attribution, while compositional descriptors contribute an additional contribution of approximately 7%. Partial-dependence analysis revealed a largest cavity diameter of 7-8 Å and pore-limiting diameter of 5 Å, combined with volumetric surface area above 1500 m2cm-3, as key thresholds associated with higher CO2 uptake. These results indicate that pore geometry is dominant in determinant of CO2 uptake under DAC conditions, while chemical composition provides secondary effect.