Prediction of nitrogen solubility in ionic liquids by machine learning methods based on COSMO-derived descriptors

The inertness of nitrogen (N2) results in extremely low N2 solubility in most solvents. However, in some important reactions, such as the nitrogen reduction reaction (NRR) driven by light or electrical energy, sufficient N2 feedstock in electrolyte allows the reaction to proceed more smoothly. In this context, the study concerning N2 solubility is of great significance. Previous research has demonstrated that ionic liquids (ILs) have many special characteristics and can be taken as green solvents to enhance N2 solubility. Herein, two quantitative structur-e-property relationship (QSPR) models are established to predict the N2 solubility in ILs by combining machine learning methods of random forest (RF) and gradient boosting regressor (GBR) with ionic fragments contribution (IFC) method. Firstly, a database with 385 N2 solubilities in 38 ILs at various temperatures and pressures is collected, in which the ILs are separated into 27 ionic fragments. Then COSMO-derived descriptors for the cations and anions of ILs are calculated and used as input variables. In specific, the determination of coefficient (R2) for the training sets by RF-IFC and GBR-IFC is 0.9983 and 0.9999, respectively. The application examples of N2 solubility in 1-hexyl-3-methyl-imidazolium tris(pentafluoroethyl)trifluorophosphate ([HMIM][eFAP]) confirm the availability and usefulness of RF-IFC and GBR-IFC models, indicating that both two models are feasible and reliable for predicting N2 solubility in ILs, and hopefully can be used to design and screen ILs in the future.

Automated summary

This paper predicts the solubility of nitrogen gas in ionic liquids (ILs) using two quantitative structure-property relationship (QSPR) models. By combining machine learning methods and ionic fragments contribution method, the accuracy and reliability of the prediction models are improved.