EquiNet: Modeling Vapor-Liquid Equilibrium Using Simultaneous Neural-Network Predictions for Activity Coefficient and Vapor Pressure

In this work, we present EquiNet: a robust neural network model for predicting vapor-liquid phase equilibrium in novel binary systems through the estimation of activity coefficients, which are central to separation processes. By integrating classic functional forms for the excess Gibbs free energy into our model structure, the trained model enforces the Gibbs-Duhem relation on all model outputs, ensuring physically realistic predictions. The model simultaneously predicts component vapor pressures, enabling direct training on experimental composition data rather than inferred activity coefficients. Given molecule SMILES representations, liquid-phase composition, and temperature data as inputs, the model can predict vapor phase compositions and pressure. Comparisons to alternative activity coefficient predictors, the UNIFAC group additivity method and the HANNA machine-learning model, show superior performance by EquiNet across a broad spectrum of binary mixtures. Our results further highlight the value of incorporating the extended dataset made accessible through the model’s internal vapor pressure prediction module, enabling broader applicability and higher accuracy. Finally, we share predictions on a public benchmark dataset for vapor-liquid equilibrium compositions, providing a foundation for further development and transparent evaluation of machine-learning approaches in thermodynamic modeling.