Chemical structure and thermodynamic properties based models for estimating nitrous oxide solubility in ionic Liquids: Equations of state and Machine learning approaches

Abstract

In past decades, nitrous oxide (N2O), a strong greenhouse gas, has become a serious transdisciplinary issue. As a result, removing N2O utilizing strong green solvents like ionic liquids (ILs) has emerged as a popular method of lowering N2O levels in the environment. The ability to accurately estimate N2O solubility in ILs provides a deeper understanding of the ILs'performance as a solvent for the elimination and management of this hazardous gaseous contaminant in the atmosphere. For the implementation of future IL-based separation procedures at massive scales, reliable calculation of this critical factoris required. The purpose of this research was to develop reliable intelligent networks that can estimate N2O solubility in diverse ILs. To this end, four powerful intelligent models including deep belief network (DBN), categorical boosting algorithm (Cat-Boost), extreme learning machine (ELM), and extreme gradient boosting (XGB) were developed based on two distinct methods, (I): Chemical structure-based and (II): Thermodynamic properties-based methods. Also, different equations of state (EOSs) were employed to compare their performance with smart models. The acquired findings indicate that the novel approaches appropriatelyestimate the solubility of N2O in ILs. Furthermore, the XGB approach was discovered to be the superior forecasting technique in both methods ((I): R2 = 0.9999 and RMSE = 0.0016, (II): R2 = 0.9998 and RMSE = 0.0025). The sensitivity analysis of the XGB models revealed that pressure has the greatest effect on the solubility values in both strategies with an absolute relevancy factor value of 0.74, and concerning the chemical structures of ionic liquids, the –SO2 substructure has the greatest effect on N2O solubility with an absolute relevancy factor value of 0.4. In addition, significant superiority of machine learning models over EOSs was observed. Finally, the Leverage approach was used to demonstrate the reliability of the novel paradigm, showing over 96 % of data are into the paradigm's applicability range.