Novel Machine Learning Model Correlating CO2 Equilibrium Solubility in Three Tertiary Amines

In this work, new artificial neural network (ANN) models were developed and correlated with the CO2 equilibrium solubility in three new tertiary amines of 1-dimethyl-amino-2-propanol (1DEA2P), 1-dimethylamino-2-propanol (1DMA2P), and 1-(2-hydroxyethyl)-piperidine (1-(2-HE)PP). The predicted data of CO2 solubility extracted from the newly developed ANN model were consistent with the experimentally observed results. It was shown that only some newly developed ANN models could predict CO2 solubility with acceptable accuracy. To better estimate the observed CO2 equilibrium solubility, a novel machine learning model of XGBoost was proposed and established to correlate the experimental data. XGBoost could satisfactorily represent the CO2 solubility in all three tertiary amines, with a mean absolute percentage error (MAPE) of 3.77% for 1DMA2P, 0.29% for 1DEA2P, and 0.70% for 1-(2-HE)PP. The performance of both the newly developed ANN models and XGBoost in terms of MAPE was discussed as well in this work. By comparing the ANN models and XGBoost model in terms of MAPE, it could be concluded that XGBoost exhibited a much better prediction of CO2 solubility for all three amines, which could be considered a potential model for CO2 solubility estimation in amine solutions.