Machine learning modeling of the CO2 solubility in ionic liquids by using σ-profile descriptors
J. Laakso, Ali Ebrahimpoor Gorji, Petri Uusi–Kyyny, Ville Alopaeus
Abstract
• Machine learning models were developed by using σ profile descriptors and dataset containing 9864 covering 124 ionic liquids. • CO 2 solubilities were predicted for 1444 unstudied ionic liquids and structural trends ofions where observed. • Descriptors importance highlights the non-polar section of the σ profile as important factor. • The σ profile-based descriptors offer a better generalization than descriptors from group contribution. The solubility of carbon dioxide (CO 2 ) in solvents is important for carbon capture and utilization technologies, with ionic liquids (ILs) being promising due to their ability to capture CO 2 . Since the number of possible ILs is huge, predicting CO 2 solubility during solvent screening is essential. In this work, various machine learning (ML) models including multiple linear regression, artificial neural network, and random forest, were developed by using 9864 data points covering 124 ILs and descriptors from the σ-profile for predicting CO 2 solubility in ILs. The random forest model produced the best performance (R 2 = 0.9754 and MAE = 0.0257). We estimated the importance of the descriptors, highlighting that those with non-polar characteristics of the σ-profile are important. Lastly, we predicted CO 2 solubilities for 1444 unstudied ILs. The combination of ML with the σ-profile descriptors offers great generalizability for predicting CO 2 solubility in ILs. This enables IL screening for CO 2 -related applications.