Estimation of CO2 solubility in ionic liquids using connectionist tools based on thermodynamic and structural characteristics

Ionic Liquids (ILs) are a promising alternative to conventional amine-based solvents in CO2 absorption processes. Relatively high cost and viscous nature of ILs tempt researchers to find appropriate combination of cations and anions for targeting a competitive absorbent close to amines in terms of CO2 absorption capacity. Conducting extensive experimental studies appears to be time-consuming and expensive for a large number of common CO2/IL systems. One of the fast and reliable approaches to predict the solubility of CO2 in ILs is Machine Learning (ML)-based models or smart tools where the thermodynamic-based and structure-based property relationships can be explored. Four ML methods including Least Squares Support Vector Machine (LSSVM), Decision Tree (DT), Random Forest (RF), and Multilinear Regression (MLR) are employed in this study to obtain CO2 solubility in a structurally diverse set of ILs based on thermodynamic properties and Quantitative Structure-Activity Relationship (QSPR) model. In this paper, two datasets of CO2 solubility (taken from the literature) at various operating conditions are used; one model (or dataset) considers critical properties, molecular weight, and acentric factor of pure ILs as the input information, and the second one includes structural descriptors of cations and anions as the input parameters. Among different types of descriptors, the most important input variables (e.g., Homo-Lumo fraction and Disps (cation)) are selected using Genetic Algorithm (GA)-MLR method. The predictive ability of the introduced models is also analyzed using a four-fold external cross-validation procedure. A great match between the predicted values and experimental measurements is attained while using RF and DT techniques developed based on descriptors and thermodynamics properties. The structural descriptors-based models are more accurate and robust than those built on the critical properties. The feature selection approach is also applied to identify the most effective cation and anion descriptors. For both thermodynamic and QSPR modeling approaches, pressure has the maximum relative importance in the estimation of CO2 solubility in ILs. In the absence of temperature and pressure impacts, the critical pressure and Homo-Lumo fraction have the highest contribution to the thermodynamic - based and structural-based tools, respectively.