Accurate VLE Predictions via COSMO-RS-Guided Deep Learning Models: Solubility and Selectivity in Physical Solvent...

Carbon capture through physical solvents reduces energy consumption and lowers environmental impact compared with conventional chemical absorption methods. Typical properties for solvent screening are solubility and selectivity. However, they require accurate prediction of vapor-liquid equilibrium (VLE), which remains a critical challenge due to the lack of enough available experimental data. This could be supplemented by in silico data prediction, provided that current prediction models are improved as this paper intends. When modeling physical solvents, a challenge arises due to the dominant role of nonbonding interactions and molecular geometry. For this purpose, a machine learning pipeline is developed using VLE results obtained from the quantum chemical-based thermodynamic model COnductor-like Screening MOdel for Real Solvents (COSMO-RS) and experimental data. A directed message passing neural network (D-MPNN) architecture is employed, leveraging molecular representations, additional features, and transfer learning to refine predictions. Two models, solubility and selectivity, are pretrained over 30,000 COSMO-RS simulated data points and fine-tuned with experimental VLE data sets for CO₂ and common gas impurities (H₂S, CH₄, N₂, and H₂), respectively. The models' accuracy is significantly improved over that of COSMO alone by correcting bias in total pressure predictions. Experimental trends are successfully reproduced in the test data, confirming the physical consistency of the models. Sensitivity analysis confirms that molecular features have the highest impact on estimations, while the scaling effect of additional features is essential for accuracy. These results demonstrate the potential of the proposed methodology to systematically screen and optimize an extensive range of physical solvents on the basis of their chemical structure for carbon capture applications, reducing the reliance on costly and time-consuming experimental measurements.