(398b) From Components to Mixtures: Equivariant Machine Learning Models for Predicting VLE, Vapor Pressure, and Azeotropes

Industry relies on accurate models to predict the behavior of gas and liquid mixtures in separation processes. Traditional solution activity models like NRTL require regression on specific data, limiting their applicability to novel systems. Existing alternatives like UNIFAC use group additivity to predict behavior in novel systems, but they struggle with uncommon mixtures and fail to capture interactions between nearby functional groups.

In this study, we introduce a novel equivariant machine learning architecture, used to train generalizable VLE models. These models are trained on diverse VLE datasets, including binary VLE compositions, pure component vapor pressures, and infinite dilution activity coefficient data using a dataset of nearly 500,000 datapoints. Our model enables internal prediction of pure-component vapor pressures, extending applicability to include components where pure component vapor pressure is unavailable. Equivariance ensures predictions remain consistent regardless of component order and improves data efficiency. By using thermodynamically consistent architectures, we output model forms aligned with traditional equations, ensuring realistic model predictions.

We benchmark our model performance against UNIFAC. We also demonstrate predictions for extended datasets where VLE compositions are known but tabulated pure component vapor pressures are unavailable. The model shows improved generalization to novel and extended mixture conditions and accurately predicts azeotrope locations.