New structure-based models for the prediction of normal boiling point temperature of ternary azeotropes

Abstract

Recently, development of the QSPR models for mixtures has received much attention. The QSPR modeling of mixtures requires the use of appropriate mixture descriptors. In this study, 12 mathematical equations were considered to compute mixture descriptors from the individual components for the prediction of normal boiling points of 78 ternary azeotropic mixtures. Multiple linear regression (MLR) was employed to build all QSPR models. Memorized_ACO algorithm was employed for subset variable selection. An ensemble model was also constructed using averaging strategy to improve the predictability of the final QSAR model. The models have been validated by a test set comprised of 24 ternary azeotropes and by different statistical tests. The resulted ensemble QSPR model had R2training, R2test, and q2 of 0.97, 0.95, and 0.96, respectively. Mean absolute error (MAE) as a good indicator of model performance were found to be 3.06 and 3.52 for training and testing sets, respectively.