Modelling of carbon dioxide solubility in ionic liquids at sub and supercritical conditions by neural networks and mathematical regressions

Multiple Linear Regression (MLR), Multiple Quadratic Regression (MQR), Radial Basis Network (RB), and MultiLayer Perceptron (MLP) Neural Network (NN) models were explored for the prediction of the CO2 solubility in 1-n-ethyl-3-methylimidazolium hexafluorophosphate, 1-n-hexyl-3-methylimidazolium hexafluorophosphate, 1-n-butyl-3-methylimidazolium tetrafluoroborate, 1-n-hexyl-3-methylimidazolium tetrafluoroborate and 1-n-octyl-3-methylimidazolium tetrafluoroborate ionic liquids (ILs) at sub and supercritical conditions. The models fitting performance was analyzed calculating statistical parameters (adjusted correlation coefficient, mean prediction error (MPE) and estimated standard deviation). To verify how well the models fit the data, analysis of residuals from the interpolative models by graphical and numerical methods and central tendency and statistical dispersion test were applied. For every model tested, the MPE values in all five studied systems is less than 12, 8, 2.2, 2.1% for MLR, MQR, MLP and RB models, respectively. Taking the numerical analysis of residuals from non linear models into account, there are no correlation between residuals and CO2 solubility. Using new databases reported in literature, the optimized interpolated models were tested. The MPE values calculated by the non linear models were less than 3.3%.