Predicting water solubility in ionic liquids using machine learning towards design of hydro-philic/phobic ionic liquids

In this work, a database containing the solubility data for water in 16,137 ionic liquids (ILs), which were formed by a combination of the most commonly used 163 cations (in nine groups) and 99 anions, were analyzed using machine learning. The water solubility in ILs was computed by COnductor-like Screening MOdel for Realistic Solvents (COSMO-RS) while the molecular descriptors for the individual cations and anions were determined by semi-empirical PM3 method. Association rule mining, decision tree and multilayer fully connected neural network (a deep learning model) were employed as machine learning techniques. The association rule mining analysis clearly identified the descriptors leading to low water solubility in ILs, while the decision tree analysis provided heuristic rules for the selection of cations and anions to form ILs with low water capacity. The prediction accuracy of fully connected neural network model was also high; even the model constructed from a small fraction of data was successful to predict the water solubility in other ILs in the dataset indicating that the anionic and cationic descriptors used were sufficient to represent the performance of ILs. The classification ability of decision trees was verified by the experimental water solubility data for 49 ILs extracted from 13 published papers in literature; the decision tree model correctly classified the experimental solubility of 46 of these ILs. The deep learning predictions of solubility were also in agreement with the experimental data within the accuracy level of COSMO-RS calculations. It was generally found that the anionic descriptors were more influential to predict the water capacity of ILs, while the cationic descriptors made limited contribution. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

Machine learning techniques were employed to accurately predict water solubility in ionic liquids, identifying key descriptors for low solubility. The deep learning model demonstrated high accuracy in predicting solubility and performed well in experimental data validation. When predicting water capacity of ILs, anionic descriptors were found to be more influential than cationic descriptors.