A quantum chemistry-driven machine learning model for predicting solubility of carbon dioxide in ionic liquids

Ionic liquids (ILs) are promising eco-friendly solvents for carbon dioxide (CO2) dissolution and capture. Utilizing deep neural network modeling to accelerate the design and screening of ILs can contribute to promoting green and sustainable development. In this study, a quantitative structure-property relationship (QSPR) model was constructed to link the structure of ionic liquids with their CO2 solvation ability. The deep neural network model was driven using two environmental descriptors, temperature and pressure, as well as 16 quantum chemical descriptors calculated from the Conductor-like Screening Model for Real Solvents (COSMO-RS). This model uniquely utilizes the sparsity of IL σ-profile curves for descriptor classification. The study explored the impact on machine learning modeling using two data splitting methods: “point-based” and “component-based”. The former randomly divides the entire dataset into a training set and a test set, yielding Coefficient of Determination(R2), Root Mean Square Error(RMSE), and Mean Absolute Error(MAE) values of 0.9904, 0.0216, and 0.0133, respectively, on the test set. The latter splits the dataset based on the type of ILs into set1 and set2, yielding R2, RMSE, and MAE values of 0.9297, 0.0631, and 0.0450, respectively, on the test set. The model was further validated and explained using Applicability Domain (AD) and SHapley Additive exPlanations (SHAP) methods. This model provides accurate predictions of CO2 solubility in ILs and offers guidance for designing ILs for CO2 capture.

Engineering Applications of Artificial Intelligence