Prediction of CO2 solubility in pyridinium-based ionic liquids implementing new descriptor-based chemoinformatics models

Abstract

Ionic liquids (ILs) are regarded as unique, attractive kinds of solvents, which can be utilized in carbon dioxide (CO2) capture processes. The preparation and design of such processes need simple and accurate models to predict solubility properties. This study incorporates the potential of the molecular descriptors, structural information, and its direct effect on the estimation of CO2 solubility in pyridinium-based IL mixtures. Using a collection of 430 experimental measurements, two different feed-forward back-propagation neural network models with three and five input variables were developed. In the first scenario, molecular weight, absorption temperature, and equilibrium pressure were considered as the model input. In contrast, the second scenario presents a new descriptor-based chemoinformatics model with the input data of the structural information for the estimation of the CO2 solubility in pyridinium-based ILs. To depict the structural clue of the various solvents, the presence of ether groups as a categoric factor, and the number of carbon in hydrocarbon chain and, or ether groups in functionalized pyridinium-based ILs were identified as the two input parameters in our novel descriptive model. The network architecture, including the neurons’ number, training, and transfer functions, is optimized. The statistical analysis of the obtained results illustrated that the developed molecular descriptor-based model, with the estimated Root Mean Square Error (RMSE) and R2 of 9.06-E03 and 0.995 for the test data, can be employed for the detailed assessment of the CO2 loading capacities of pyridinium-based IL solutions. The authors believe that the out-perform of the ready-made descriptor-based model can be helpful as a guided example for all kinds of descriptor-based modeling campaigns.