Prediction of biphasic separation in CO2 absorption using a molecular surface information-based machine learning model

Abstract

Carbon dioxide capture technologies have become a focus to overcome global warming. Biphasic absorbents are one of the promising approaches for energy-saving CO₂ capture processes. These biphasic absorbents are mainly composed of a mixed solvent composed of alkanolamine and organic solvents like glycol ether or alcohol. However, screening experiments of the mixed-solvent absorbents are required to search for biphasic absorbents due to their complicated phase behavior. In this work, we developed a prediction method for the phase states of the mixed-solvent absorbents using a quantum calculation and machine learning models, including random forest, logistic regression, and support vector machine models. There are 61 mixed-solvent absorbents containing alkanolamine/glycol ether or alcohol in the dataset. The machine learning models successfully predicted the phase states of the mixed-solvent absorbents before and after CO₂ absorption with accuracies of more than 90%. Furthermore, we analyzed the contributions of explanatory variables for prediction using the learned model. As a result, we found that molecular surface charge of the amine species is more important than those of the other organic solvents to determine the phase behaviors during CO₂ absorption.