Prediction of Nernst coefficient of S-containing compounds between fuel and ionic liquid phases in the extractive desulfurization using linear and supported vector machine (SVM) methods: QSPR-based machine learning

Background

The presence of sulfur-containing compounds (SCCs) in the refinery streams has significant economic and environmental implications. Great attention has been focused on finding the proper green solvents like ionic liquids (ILs¹) with efficient extraction performance as a matter of concern of environmental issues.

Methods

The research aimed to develop a predictive model using QSPR to forecast the partition coefficient of dibenzothiophene by ILs from n-dodecane. Utilizing a dataset of 54 ILs and their partition coefficients for DBT, the study employed two methods to optimize ILs structures and compared linear (GA-MLR) and non-linear (LS-SVM) models, with non-linear models showing higher accuracy. After initial modeling and assessing the primary dataset of 54 ILs, yielding an R2 parameter of 0.39 for the test set, the dataset was divided into smaller clusters for further analysis. Three additional clusters were investigated. The second cluster comprised 14 ILs with identical cations and varying anions, modeled with two descriptors. The third cluster, consisting of 21 ILs with imidazolium cations and diverse anions, was modeled with three descriptors. Lastly, the fourth cluster, comprising 26 ILs with different cations but the same anion, was also modeled with three descriptors.

Significant findings

The MLR model yielded R2 values of 0.98, 0.85, and 0.93 for the test sets of the second, third, and fourth clusters respectively. Effective descriptors, including cation polarizability and alkyl branch length, were examined for their impact on partition coefficient and desulfurization efficiency. This research aids in enhancing EDS processes with ILs, advancing more efficient desulfurization technologies.