Machine learning with new functional structure descriptors for design and screening of ionic liquids in CO2 efficient capture

Abstract

Carbon dioxide emission reduction, conversion and utilization are the hot and difficult issues in the world. As a new kind of green solvents, ionic liquids (ILs) are widely used in CO2 capture and conversion, but there are various kinds of ILs (more than 1018). How to select and screen the appropriate ILs for CO2 capture is an urgent problem to be solved. Therefore, it is of great significance to establish the Quantitative Structure-Property Relationships (QSPR) of ILs for CO2 capture. From the practical point of view of ILs design and synthesis, a new functional structure descriptor (FSD) based on group contribution method (GC) was constructed. At the same time, the idea of increasing dimension to increase accuracy in traditional machine learning is changed, and the feasibility of reducing the dimension under the condition of ensuring accuracy is examined. A dimensionless molecular descriptor CORE is constructed. Based on these two new molecular descriptors, we discussed the performance of six common ensemble learning models (CatBoost, LightGBM, XGBoost, GBDT, RF and AdaBoost) for CO2 solubility in ILs. It is shown that all ensemble learning models can achieve good performance, but CatBoost model is the most outstanding. The R² of 0.9945 and MAE of 0.0108 for CatBoost-FSD model is achieved, while R² and MAE is 0.9925 and 0.0120 for CatBoost-CORE model, respectively. The interpretability of CatBoost-CORE model is analyzed, and the key features are determined. Based on the CORE descriptor, the best experimental conditions are obtained, and nine kinds of ILs with superior performance are recommended.