Exploring Metal-Organic Framework Design Strategies for CO₂ Capture Using Explainable Artificial Intelligence

Metal-organic frameworks (MOFs) have emerged as promising adsorbents for CO₂ capture due to their high surface area, tunable pore size, and exceptional chemical functionality. However, the identification of optimal MOFs is hindered by the absence of efficient and interpretable high-throughput screening methods, which are capable of addressing the complexity and diversity of MOF structures. In this study, we developed predictive models for the CO₂ adsorption capacity and CO₂/N₂ selectivity of MOFs, incorporating a wide range of pore structures, topological features, and organic linkers. These models are based on a gradient-enhanced regression framework. Through hyperparameter optimization, the best-performing model achieved a mean absolute error of 0.0792 mmol/g for the CO₂ adsorption capacity and 1.7464 mmol/g for the CO₂/N₂ selectivity. Shapley additive explanation analysis identified void fraction as the most influential factor governing both adsorption capacity and selectivity. Specifically, a void fraction in the range of 0.10-0.30 provides the greatest positive impact on the CO₂ adsorption capacity, whereas a void fraction between 0.04 and 0.24 has the most beneficial effect on selectivity. Furthermore, specific functional groups, particularly aromatic rings, enhance adsorption efficiency, while the presence of halogen and metal atoms exerts a negative impact on performance. Other pore structural characteristics, along with topological and organic linker properties, were also found to impact the adsorption performance. The significant role of topological design in enhancing adsorption behavior has also been highlighted, indicating its critical influence on adsorption efficiency. These findings provide valuable insights for the rational design of MOF materials for CO₂ capture.