Machine learning for the yield prediction of CO2 cyclization reaction catalyzed by the ionic liquids

Ionic liquids are one of the excellent catalysts for the CO2 cycloaddition reaction, which is an effective means to realize CO2 utilization and alleviate environmental problems. However, the design of ionic liquids catalysts has great blindness due to the diversity of their anion and cation structures. Herein, we collected a database of 866 samples for CO2 cyclization with ionic liquids as catalysts and established the yield prediction model with various machine learning regression algorithms. Together with density function theory (DFT) calculated mo-lecular electronic properties and the experimental conditions as the input descriptors, random forest (RF) model has better prediction accuracy than support vector regression (SVR) and multilayer perceptron (MLP) for the whole data. When the dataset was subdivided into different subsets based on substrates and ionic liquids, the prediction accuracy of the model was also improved. For the imidazole subset with no additional solvent or additive added, the RF model achieves good accuracy with the R2 of 0.80 for the test data. Moreover, the shapely additive explanation (SHAP) method was used to interpret the ML models. The strategy of refining the de-scriptors and dataset used in our work provides guidance for establishing highly reliable machine learning models in the chemistry reactions.

Automated summary

Ionic liquids are effective catalysts for CO2 cycloaddition reactions, but their design lacks guidance due to the diversity of their structures. In this study, we collected a database of 866 samples and established a yield prediction model using machine learning regression algorithms. The random forest model showed better prediction accuracy compared to support vector regression and multilayer perceptron models, especially when the dataset was subdivided into subsets based on substrates and ionic liquids. Our work also provided guidance for establishing reliable machine learning models in chemistry reactions.