Performance evaluation of the machine learning approaches in modeling of CO2 equilibrium absorption in Piperazine aqueous solution

Abstract

Accurate knowledge of equilibrium solubility of CO2 in solvents is crucial for optimal design and operation of the absorption-based CO2 capture systems. This study aims at evaluation of the performance of four machine learning approaches including artificial neural network (ANN), adaptive neuro-fuzzy inference system (ANFIS), least squares support vector machine (LSSVM), and adaptive boosting in conjunction with the classification and regression tree (AdaBoost-CART) in calculating the CO2 equilibrium absorption in Piperazine (PZ) aqueous solution. To this end, an extensive databank covering wide ranges of temperature, pressure, and solvent concentration was collected from literature. With an overall absolute relative deviation in percent (AARD%) equal to 18.69, 15.99, and 16.23, the results from the error analysis reveal that none of the ANN, ANFIS, and LSSVM methods are capable to reliably model the equilibrium system of CO2+water+PZ. On the other hand, the AdaBoost-CART model shows much higher accuracy and robustness in representing the equilibrium solubility of CO2 in PZ solution. The AdaBoost-CART model reproduces the experimental targets with an AARD%=0.93 and a R2=0.9934 which demonstrates excellent agreement between the outputs and corresponding experimental values.