Group Contribution Based Graph Convolution Network: Predicting Vapor–Liquid Equilibrium with COSMO-SAC-ML

The conductor-like screening model with the segment activity coefficient (COSMO-SAC) model enables us to predict the liquid phase behavior based on quantum chemical calculations when experimental information is unavailable. However, the quantum chemical COSMO calculations require significant computational resources (software, hardware, experience, and time) for successful results. In this work, we suggest machine learning models for replacing COSMO calculations to alleviate the computational burden. The machine learning (ML) models include graph convolutional neural networks. The ML models were constructed based on the molecular level structural information and previous COSMO calculation databases. Additional COSMO calculations were generated to train several missing functional groups, which are not included in the existing databases. The ML prediction abilities were shown by comparison of generated cavity volume and sigma profile with the original COSMO-SAC. The vapor–liquid equilibrium (VLE) prediction results were also compared with the original COSMO-SAC and the UNIQUAC functional-group activity coefficients (UNIFAC) model. It is shown that the suggested ML models can predict VLE with reasonable accuracy comparable with the original COSMO-SAC and UNIFAC. The new ML models can effectively replace the time-consuming COSMO-SAC calculations and experimental data-dependent UNIFAC models. Graphical Abstract