280897 A Theory-Framed Quantitative Structure-Property Relationship (QSPR) Model for Liquid-Liquid Equilibrium
A thorough understanding of phase behavior properties of chemicals is essential for designing and optimizing processes that involve separation of components from a mixture. The non-random two, liquid (NRTL) model is an activity coefficient model used widely in phase equilibria calculations, which has three adjustable parameters that are determined through regression of experimental data for a specific system. Generalizing a model can reduce time, money and effort required to carry out experiments. This work focuses on application of a theory-framed quantitative structure-property relationship (QSPR) modeling approach. A theoretical framework is employed to develop the behavior models, and the QSPR model is used to generalize the substance-specific parameters.
A database of 342 binary low temperature (10 – 40°C) LLE data was employed in this work. Data regression analysis was performed to determine the parameters of the NRTL model. The structural descriptors of the molecules were generated and used in developing a QSPR model to estimate the regressed NRTL parameters. The regression analysis yielded %AADs of 2, 13, 11, 17 for mole fractions of component 1 in phase 2 (component 1 rich) and phase 1 (component 2 rich) and partition coefficient of components 2 and 1, respectively.
The newly developed QSPR model yielded predictions with 8, 37, 40 and 44 %AAD for the mole fractions and partition coefficients, respectively. These errors are approximately 2 to 3 times the errors found from the regression analysis. There is also a failure rate associated with the property predictions due to convergence issues, which in the case of our model was 8%. The application of the popular and often employed UNIFAC model resulted in 3 to 7 times the errors obtained through regression analysis and a failure rate of 35%. The smaller prediction errors as well as the lower failure rate of the QSPR model demonstrates the superiority of the model in providing improved and reliable predictions, as well as an enhanced range of applicability when compared to the UNIFAC model for LLE systems.