Modeling of CO2 solubility in piperazine (PZ) and diethanolamine (DEA) solution via machine learning approach and response surface methodology

Abstract

With the help of machine-learning algorithms, data-driven models have become increasingly capable of predicting CO2 solubility. As part of this study, two machine learning approaches are evaluated: artificial neural networks (ANNs) and support vector machines (SVM), as well as response surface methodology (RSM) to calculate CO2 equilibrium in aqueous solutions containing piperazine (PZ) and diethanolamine (DEA). Correlations are useful for predicting the solubility of CO2 in the liquid phase (PZ + DEA) in the temperature (303, 323, 343.2 K) and various CO2 partial pressures (100–1000 KPa). The optimization SVM tested multiple kernel functions, such as linear, quadratic, cubic, and gaussian, alongside different optimizers. The cubic kernel function was found proper for training SVM. The optimum multilayer perceptron (MLP) structure in Levenberg-Marquardt algorithm for CO2 solubility is created with ten neurons in one hidden layer. It was found that the MLP network had the greatest mean square error (MSE) afterward 7 epochs, equivalent to 0.000128, and the coefficient of determination (R2) was 0.99947. There was a coefficient of determination of over 0.99 for all three models, indicating that all three models have excellent prediction capabilities.

Keywords

• Carbon dioxide
• Solubility
• ANN modeling
• RSM
• SVM