Accurate VLE Predictions via COSMO-RS-Guided Deep Learning Models: Solubility and Selectivity in Physical Solvent Systems for Carbon Capture.
Accurate VLE predictions for carbon capture solvents using COSMO-RS-guided deep learning models. Solubility and selectivity models outperform COSMO alone, unlocking systematic screening of physical solvents to reduce energy and environmental impact.
This study developed a machine learning pipeline to improve the prediction of vapor-liquid equilibrium (VLE) properties for physical solvents used in carbon capture. The approach combines the quantum chemistry-based COSMO-RS model with a directed message passing neural network (D-MPNN) architecture, leveraging molecular representations and additional features. Two models were trained on over 30,000 COSMO-RS simulated data points and fine-tuned with experimental VLE data for CO2 and common gas impurities. The models significantly outperformed COSMO alone, accurately reproducing experimental trends. Sensitivity analysis confirmed the importance of molecular features and the scaling effect of additional features for model accuracy. This methodology enables systematic screening and optimization of physical solvents for carbon capture, reducing reliance on costly experiments.