Machine learning for data-driven insights into CO₂ adsorption on amorphous porous organic polymers

Increasing CO₂ emissions demand advanced carbon capture and storage technologies. Among various approaches, the adsorption of CO₂ on porous organic polymers (POPs) is particularly promising due to their low density, high surface area, tunable pore structure, and excellent thermal and chemical stability. However, optimizing CO₂ uptake is challenging because the quantitative relationships between material properties and adsorption capacity remain unclear. Although machine learning (ML) algorithms have improved predictive performance, many models offer limited actionable insights for material design due to poor interpretability. In this study, we apply four supervised ML models random forest, light gradient boosting, extreme gradient boosting, and support vector machines to predict the CO₂ adsorption capacity of amorphous POPs using a comprehensive dataset (8 inputs and 737 data points) that integrates textural properties, elemental composition, and operating conditions. The extreme gradient boosting model achieved the best performance (R² = 0.995; RMSE = 0.056; MAE = 0.0321). Beyond prediction, we employ SHapley Additive exPlanations, permutation importance, and uni‑factorial partial dependence analysis to quantitatively elucidate the role of individual descriptors. Our results reveal that operating conditions and textural features (e.g., BET surface area and micropore volume) exert a greater influence on CO₂ uptake than elemental composition. These data-driven insights provide a roadmap for the rational design of next-generation POP adsorbents for efficient carbon capture.