(148f) Generalized Activity Coefficient Models for VLE Mixture Predictions
Accurate prediction of the phase behavior properties of chemical species and their mixtures is essential for the design of chemical processes involving separation operations. Predicting phase equilibrium properties using reliable models offers an attractive alternative to costly and time consuming experimental measurements. In this work, we focus on liquid activity coefficient modeling.
Activity coefficient is a basic phase equilibria property that accounts for liquid mixture deviations from ideal behavior. Although a number of activity coefficient models exist in the literature, their use is limited by the availability of experimental data. Among these models, non-random two, liquid (NRTL), universal quasi-chemical (UNIQAUC) and WILSON are widely used to correlate fluid phase equilibrium data. These models require two or three adjustable interaction parameters that are determined through regression of experimental data for a specific system. Typically, to facilitate a priori predictions, group-contribution methods have been employed to generalize the UNIQAUC and Wilson models.
Our recent research has focused on developing an alternative method for generalizing the interaction parameters of the NRTL, UNIQUAC and WILSON models. Specifically, we applied theory-framed quantitative structure-property relationship (QSPR) modeling approach. In this modeling approach, theoretical frameworks, such as the NRTL, are used to develop the behavior models, and QSPR to generalize the substance-specific parameters of the models.
A database of 960 binary VLE data was employed to develop the QSPR models. Data regression analyses were performed to determine the interaction parameters of the NRTL, UNIQUAC and WILSON models. The structural descriptors of the molecules were generated and used in developing the QSPR models to predict the regressed interaction parameters. The predictive capabilities of the generalized models were assessed for phase equilibria properties including pressure, temperature, vapor mole fractions and equilibrium constants. The developed QSPR models provided phase equilibria property predictions within two times the errors obtained through the data regression analyses. The overall property predictions from the QSPR models are comparable to that of the results from the group-contribution method (UNIFAC-2006), which has over 4,000 model parameters. In contrast, the QSPR model has about 300 model parameters (neural network weights and biases). Thus, our methodology provides a priori and easy implementable models with comparable property predictions to those the UNIFAC model.