Developing machine learning models for ionic conductivity of imidazolium-based ionic liquids

Abstract

In this study, we developed two machine learning models: - support vector machine (SVM) and artificial neural network (ANN) - to correlate ionic conductivity of pure ionic liquids based on the imidazolium cations using the data acquired from the NIST ILThermo database. Both models were shown to successfully capture the entire range of ionic conductivity spanning six orders of magnitude over a temperature range of 275–475 K with relatively low statistical uncertainty. Due to slightly better performance, ANN was used to predict the ionic conductivity for 1102 ionic liquids formed from every possible combination of 29 cations and 38 anions contained in the database. The procedure led to the generation of many ionic liquids for which the ionic conductivity was estimated to be greater than 1 S/m. The ionic liquid dimethylimidazolium dicyanamide, not present in the original dataset, was identified to exhibit the ionic conductivity of 3.70 S/m, roughly 30% higher than the highest conductivity reported for any ionic liquid at 298 K in the database. The ANN model was also found to accurately predict the ionic conductivity for several ionic liquid-ionic liquid mixtures, for which experimental data are available. Encouraged by this result, we calculated ionic conductivity for all the possible binary ionic liquid-ionic liquid mixtures based on the cations and anions in the dataset. The model predictions revealed a large number of ionic liquid mixtures systems exhibiting nonideal behavior where a maximum or minimum in the ionic conductivity was observed as a function of composition, similar to trends seen in binary ionic liquid mixture of water or conventional solvents with ionic liquids.