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

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Materials Science

Background:

• Physical solvents offer energy-efficient carbon capture but require accurate vapor-liquid equilibrium (VLE) predictions.
• Limited experimental data hinders the development of reliable predictive models for solvent screening.
• Understanding nonbonding interactions and molecular geometry is crucial for modeling physical solvents.

Purpose of the Study:

• To develop an improved in silico method for predicting VLE data of physical solvents for carbon capture.
• To enhance the accuracy of solubility and selectivity predictions for CO2 and common gas impurities.
• To reduce the reliance on extensive experimental measurements in solvent discovery.

Main Methods:

• A machine learning pipeline combining quantum chemical (COSMO-RS) and experimental VLE data.
• Utilized a directed message passing neural network (D-MPNN) architecture with molecular representations and transfer learning.
• Pretrained models on 30,000 COSMO-RS data points and fine-tuned with experimental data for CO2 and impurities (H2S, CH4, N2, H2).

Main Results:

• Significantly improved prediction accuracy by correcting bias in total pressure predictions compared to COSMO-RS alone.
• Successfully reproduced experimental trends in test data, confirming the physical consistency of the developed models.
• Sensitivity analysis indicated molecular features and additional feature scaling are critical for accurate estimations.

Conclusions:

• The proposed machine learning methodology can systematically screen and optimize physical solvents for carbon capture based on chemical structure.
• This approach reduces the need for costly and time-consuming experimental measurements in solvent development.
• The findings pave the way for accelerated discovery of efficient physical solvents for carbon capture applications.