Multi-objective optimization of the amine-based natural gas-fired post-combustion carbon capture process using machine learning surrogate modeling

The rapid increase in atmospheric CO2 concentrations has increased the need for efficient mitigation strategies, with post-combustion carbon capture (PCC) remaining the most practical option for retrofitting existing power plants. Amine-based absorption is the most mature PCC technology, yet its performance is strongly affected by operating conditions, solvent properties, and complex interactions, which limit the applicability of traditional modeling approaches. This study develops an integrated framework combining rigorous Aspen Plus simulation, machine learning (ML), and multi-objective optimization (NSGA-II) to evaluate and enhance PCC performance in NG-fired systems. A dataset of 4426 simulation points was generated across five amines using a Python–Aspen interface. We applied six ML algorithms (ANN, TabPFN, PSO-XGBoost, PSO-GBR, PSO-CatBoost, PSO-AdaBoost), among which the ANN delivered the most balanced accuracy for predicting CO2 recovery, CO2 purity, and specific energy consumption (SEC) (R2 > 0.993), and TabPFN followed ANN as second place. The ANN surrogate, combined with a surrogate XGBoost model for CO2–amine loading prediction, was coupled with NSGA-II to explore trade-offs between capture efficiency, purity, and energy demand. Aspen-validated Pareto solutions (average error of less than 10% for all targets in the feasible solutions) revealed an optimal operating point with 97.6% recovery, 99.1% purity, and a reduced SEC of 10,914 kJ/kg from its initial value (43% energy reduction), with diglycolamine (35.9 wt%) emerging as the most promising solvent among the five examined solvents for the studied process from a technical perspective. The workflow demonstrates an ML-assisted pathway for accelerating PCC process optimization for the NG-fired power plants.