(378i) Modeling Vapor Liquid Equilibrium Using Machine Learning

Industries rely heavily on accurate equations to predict the behavior of gas and liquid mixtures for separation processes. However, due to reliance on specific data for regression, existing equations such as NRTL cannot produce models on novel mixtures. This limitation has prompted the development of alternative approaches such as UNIFAC, a group additivity model which uses functional group compositions to predict properties without requiring previous data of a specific mixture. UNIFAC does require pairwise interaction parameters to produce good models, but these parameters may not exist for uncommon pairs. UNIFAC is hence constrained by its reliance on group additivity and the rigid functional group categories, which don't allow nearby functional groups to influence each other.

In this study, we provide an alternative to UNIFAC by leveraging machine learning models on known VLE databases. Our approach aims to offer a more flexible representation compared to traditional and group additivity models, allowing projection to novel mixture pairs and considering neighboring functional groups in the calculation. We apply different model structures, utilizing the functional forms of traditional models as the model predictive output, thereby maintaining thermodynamic consistency of predictions and ensuring a smooth prediction output.

Furthermore, we incorporate simultaneous prediction of vapor pressure in the model, enabling predictions for species outside the range of available Antoine’s coefficients. We benchmark our trained model against UNIFAC and other traditional models trained on the same data. The model is then evaluated by testing the predicted location of azeotropes against known locations. Our results demonstrate a greater ability to generalize novel mixtures and condition extensions of existing mixtures. We also include uncertainty quantification into the model to estimate the bounds of applicability for model prediction.