Toward predicting CO2 loading capacity in monoethanolamine (MEA) aqueous solutions using deep belief network

Abstract

The viability of CO₂ capture projects, particularly through absorption with monoethanolamine (MEA) and other commercial absorbents, strongly depends on the CO₂ loading capacity. Therefore, comprehending the impact of variables on the CO₂ loading capacity of MEA is crucial in designing CO₂ capture units, which can be further optimized through multi-objective optimization. To this end, four machine learning models—Bagging Regression (BR), Categorical Boosting (CatBoost), Deep Belief Network (DBN), and Gaussian Process Regression with Rational Quadratic kernel function (GPR-RQ)—were utilized to predict the CO₂ loading capacity of MEA aqueous solutions. Temperature, partial pressure of CO₂, and MEA concentration were inputted into the intelligent network to calculate the CO₂ loading capacity. The binary values of R² and standard deviation (SD), which were 0.9889 and 0.0628 for Bagging Regression, 0.9932 and 0.06586 for CatBoost, 0.9957 and 0.0588 for GPR-RQ, and 0.9971 and 0.0329 for DBN, confirm that DBN has the highest accuracy in statistical analysis, followed by GPR-RQ, CatBoost, and Bagging Regression. Additionally, graphical methods like scattered plots and relative deviation plots corroborate the superior performance of the DBN model over all other intelligent techniques. By conducting a relevancy factor analysis on DBN outcomes, sensitivity analysis demonstrates that pressure has the most significant influence among the inputs. Furthermore, the Leverage technique affirms that the DBN model has a substantial degree of validity in forecasting the CO₂ loading capacity of MEA. Finally, 3-D image plots were systematically examined to analyze the binary interactive effect of (temperature, CO₂ partial pressure), (temperature, MEA concentration), and (CO₂ partial pressure, MEA concentration) on the carbon absorption efficiency, which is essential to reach the net-zero emission purpose.