Modeling solubility of oxygen in ionic liquids: Chemical structure-based Machine Learning Systems Compared to Equations of State

The solubility of different gases such as O2 in different liquids and finding the proper solvent for the gas sepa-ration process is one of the important concerns that has received much attention in recent decades. Ionic liquids (ILs) have received much interest in recent years as a prospective category of suitable solvents for gas separation operations. In this study, the solubility of oxygen in ILs has been estimated using powerful machine learning approaches including Deep belief network (DBN), Categorical boosting (Cat-Boost), Multivariate adaptive regression splines (MARS), and Extreme gradient boosting (XGB). Although temperature, pressure, and critical properties are among the properties of ILs that affect the solubility of gases, especially oxygen, in ILs, the chemical substructure of anions and cations also have a significant effect on the properties of ILs because the substructures can change the properties of ILs and create new solubility properties. In this regard, two different strategies including (I): Chemical structure-based and (II): Thermodynamic properties-based models have been used in model development. The results obtained from two separate methods show that in the first strategy, the DBN model has the most accurate predictions with the coefficient of determination (R2) and the root mean square error (RMSE) values of 0.9976 and 0.00341, respectively, while in the second strategy, the XGB model has the best performance with the R2 and RMSE values of 0.9998 and 0.00095, respectively. Also, the comparison of smart models and equations of state (EOSs) shows that the accuracy of smart models compared to EOSs is remarkable. Sensitivity analysis of the DBN and XGB models in the first and second strategies, respectively, shows that pressure has the greatest effect on the O2 solubility among operational parameters in both strategies, and among ionic liquid type and thermodynamic properties, the -Cl substructure and critical pressure have the greatest effect on the solubility of oxygen in ILs, respectively. Trend model analysis also shows that increasing temperature decreases the O2 solubility while increasing pressure increases the O2 solubility. Additionally, replacing the alkyl group with the ether group in the cation's chain reduces the oxygen solubility. The group error analysis also represents that the proposed models have less accuracy at higher values of the input pa-rameters. Finally, the results of the leverage technique show that more than 98% of the data are in the valid region. The findings of this study can provide a better view to control some chemical reactions such as oxidation and provide suitable solutions.

Automated summary

This study used machine learning methods such as Deep belief network (DBN), Categorical boosting (Cat-Boost), Multivariate adaptive regression splines (MARS), and Extreme gradient boosting (XGB) to estimate the solubility of oxygen in ionic liquids (ILs). The results showed that the DBN model performed the best in the first strategy, while the XGB model performed the best in the second strategy. It was also found that pressure had the greatest effect on the solubility of oxygen in ILs.