Predictive Models for Physical Properties of Bi-phasic Reaction Compounds

Knowledge of the thermophysical properties of compounds is important in several chemical processes such as bio-fuel synthesis. Although experimental determination of these properties is the most reliable technique, experiments are time consuming and costly. Hence, a need exists for reliable models capable of providing a priori property predictions of diverse compounds in the absence of experimental data. Quantitative structure–property relationship (QSPR) modeling has the potential to provide reliable property estimates based on detailed chemical structure information. Many of the existing QSPR models in the literature are linear in nature and therefore do not lead to accurate predictions. Further, the existing QSPR models are limited to estimating properties at a single temperature.

This work is focused on developing improved generalized QSPR models for temperature-independent and temperature-dependent properties. To illustrate the general methodology employed, normal boiling point (NBP) and surface tension (ST) are selected as the temperature-independent and temperature-dependent properties, respectively. For ST modeling, the scaled-variable-reduced-coordinate (SVRC) model, based on corresponding state theories (CST) and scaling laws was used as the theoretical backbone. QSPR techniques were then used to predict for the compound-dependent model parameter of the SVRC model. The percentage absolute deviation (%AAD) values for the training and external test sets for NBP modeling are 2.1% and 4.5%, respectively. For the ST models, the %AAD values are 1.9% and 2.6% for the training and external test sets, respectively. Both these models perform better than existing models in the literature.


Non-linear QSPR models perform significantly better than linear models.

Theory-framed QSPR modeling is effective in generalizing surface tension models for different temperatures.

The models from the current work are able to predict the NBP temperatures with a %AAD of 2.5%, and ST values with a %AAD of 2.6%, on external test sets that have not been used for training.