(58f) The Development of Machine Learning, Group Contribution and Molecular Modeling Approach to Screen Physical Solvents for Gas Separation

In this work, we focus on screening and designing solvents to physically and selectively absorb one gas from the other in a gas mixture. Specifically, a hybrid computational approach was developed to systematically screen solvents to separate CO₂ from a fuel gas mixture which contains CO₂, H₂ and H₂O. We have identified a solvent which is hydrophobic, non-foaming, less volatile and less viscous than the commercial Selexol solvent. Additionally, this solvent exhibits a large CO₂ loading and large CO₂/H₂ loading selectivity compared with Selexol. Due to its very high hydrophobicity, water presence in the gas stream will not affect CO₂ loading in this solvent. In contrast, water presence in the gas stream will adversely decrease CO₂ loading and CO₂/H₂ loading selectivity by 2-3 times in the hydrophilic Selexol. Along with other favorable properties for this solvent, such as the large overall CO₂ mass transfer coefficient obtained from simulations and physical model estimation and the reasonable price, this solvent is promising to be used in CO₂ pre-combustion capture applications and is being experimentally tested at NETL. Additionally, some solvents with much higher CO₂ loadings than the commercial methanol solvent used at low temperatures have also been identified. From our in-house database and molecular simulations, the largest CO₂ loading in any organic compound at 298 K was estimated to be 11 mol/MPa.L and the theoretical minimal CO₂ loading was approximately calculated to be 0.4 mol/MPa.L. A literature survey of experimental CO₂ loadings in ~100 compounds shows that most of those CO₂ loadings are between 0.4 - 11 mol/MPa.L; the largest experimental CO₂ loading obtained from the literature survey is about 3.3 mol/MPa.L, which is about 3.3 times smaller than the theoretical largest value of 11 mol/MPa.L. This result indicates that there is still much room for improvement to find better solvents with higher CO₂ loadings. Finally, some findings obtained from machine learning algorithms including artificial neural network will also be presented. It was found that the H₂ loading in a solvent correlates almost linearly with the solvent free volume fraction. For CO₂, a relationship between CO₂ loading and three other factors, that is, the solvent free volume fraction, effective CO₂-solvent interaction by including solvent molar volume, and functional availability was built. The best machine learning model gives a CO₂ solubility of about 0.4 mol/MPa.L different than the experimental data. Although more accurate models are being pursued, this preliminary model can be used for screening purposes to select solvents with large CO₂ loadings.