Prediction of the solubility of fluorinated gases in ionic liquids by machine learning with COSMO-RS-based descriptors

The substantial emission of fluorinated gases (F-gases) has exacerbated global climate change. Ionic liquids (ILs) are a type of promising green solvents for capturing F-gases. For efficient screening of IL candidates, this study proposed two machine learning (ML) models multilayer perceptron (MLP) and support vector regression (SVR) based on the critical descriptors from the thermodynamic model COSMO-RS to predict the solubility of F-gases in ILs for the first time. Over 4000 experimental solubility data (25F-gases and 52 ILs) collected were utilized to build the dataset. The used COSMO-RS-based descriptors consist of distribution of surface shielding charge density (sigma-profiles) of ILs and F-gases, molecular surface area and molecular volume of anions, cations and Fgases. In addition, the input descriptors include temperature (T), pressure (P) and molecular weight of anions, cations and F-gases. The results indicate that MLP exhibits the better prediction capability compared to SVR, with the average absolute relative deviation (AARD) of 10.16% and coefficient of determination (R2) of 0.9956, respectively. The generalization performance of the MLP model was successfully evaluated by predicting the solubility data of F-gas trans-1,3,3,3-tetrafluoropropene (R1234ze(E)), which are not been learned, with an average AARD of 14.28%. This demonstrates that the developed MLP possesses strong generalization ability, and provide a reliable reference for the screening task-specific ILs for high-efficiency capture of F-gases.