Inverse Screening of Metal-Organic Frameworks for Low-to-Mid-Pressure Carbon Dioxide Adsorption using Geometric, AP-RDF, and RAC Descriptors

Abstract

Metal-Organic Frameworks (MOFs) are promising materials for carbon dioxide (CO₂) capture because of their high porosity, tunable chemistry, and structural diversity. However, the enormous design space of MOFs makes exhaustive experimental and computational screening prohibitively expensive. Consequently, there is a growing need for computational approaches that can rapidly identify promising materials while reducing the cost of materials discovery. This thesis investigates the use of geometric, revised autocorrelation (RAC), and atomic property-weighted radial distribution function (AP-RDF) descriptors for predicting low-to-mid-pressure CO₂ adsorption using machine learning (ML) models. The predictive capabilities of different descriptor combinations are systematically evaluated to establish an accurate surrogate modeling framework. The best-performing model, selected based on predictive accuracy, is subsequently used for interpretability analyses to quantify the relative contribution of different descriptor groups to model predictions and to examine how descriptor importance varies across pressure conditions. These findings facilitate a deeper understanding of the factors influencing CO₂ adsorption predictions. The best-performing XGBoost model using the combined geometric, RAC, and AP-RDF descriptor set achieved R² values of 0.739 and 0.904 at 0.015 and 0.15 bar, respectively. Descriptor ablation and SHAP analyses showed that geometric descriptors provided the strongest overall predictive contribution, while AP-RDF and RAC descriptors supplied complementary information; surrogate-guided screening further demonstrated that high-performing MOFs could be efficiently recovered from a large candidate pool using limited training data. Building upon these insights, surrogate-guided inverse screening is employed to efficiently recover top-performing MOFs from a large candidate pool of 208,849 MOFs using limited training data. The descriptor-level insights obtained from the interpretability analyses are further evaluated through chemistry-aware matched-analog screening, which tests whether AP-RDF and RAC information can help identify higher working capacity candidates among MOFs with similar pore geometry compared with geometry-only selection. Across 234 matched neighborhoods, chemistry-aware guided selection achieved a median working capacity gain of 5.85 wt%, compared with 1.20 wt% for geometry-only selection and 0.15 wt% for the repeated-random benchmark. Predicted and actual guided gains showed strong agreement (Spearman correlation = 0.90), while the grouped chemistry-aware contribution was also strongly associated with actual guided gain (Spearman correlation = 0.89). These results demonstrate that the developed surrogate framework can support both large-scale inverse screening of MOF databases and the selection of higher working capacity candidates among geometrically similar MOFs, while providing descriptor-level information that helps interpret the model predictions and prioritize promising candidates for subsequent simulation or experimental evaluation.