Efficient post-combustion carbon capture remains a significant challenge in the global transition toward low-carbon energy systems. Among the available separation technologies, amine-based absorption is widely adopted in industrial CO 2 scrubbing due to its high selectivity and technological maturity, with methyldiethanolamine (MDEA) serving as a benchmark solvent for large-scale applications. In this study, a comparative machine learning framework is developed to predict CO 2 absorption capacity in MDEA-based nanofluid systems, integrating data-driven intelligence with environmental process modeling. A comprehensive database of experimental data was curated, covering graphene oxide (GO), Fe 3 O 4 , and carbon nanotube (CNT) nanofluids over a wide range of operating conditions. Six supervised algorithms—K-Nearest Neighbors (KNN), Random Forest (RF), Gradient Boosting (GB), XGBoost, LightGBM, and a stacking ensemble with LightGBM as the meta-learner—were trained and optimized using tenfold cross-validation. All models exhibited strong predictive performance (R 2 > 0.96), while the stacking ensemble achieved the highest accuracy with MAE = 0.021 mol kg −1 , RMSE = 0.036 mol kg −1 , AARD = 2.21%, and R 2 = 0.992. Feature importance and SHAP analyses identified temperature and CO 2 pressure as the dominant variables governing absorption behavior, followed by MDEA and nanoparticle concentrations. The proposed framework enables rapid solvent screening, optimization of operating windows, and digital twin integration for industrial CO 2 capture systems, providing a scalable pathway toward energy-efficient and cleaner production-oriented carbon capture technologies.
Gao et al. (2026) studied this question.