The Phase Behavior of CO₂ and H₂O mixtures using Ab Initio-based Machine Learning Models

CO₂ and H₂O mixtures are essential to a wide range of natural and engineered systems, from planetary atmospheres to carbon capture technologies. One promising carbon storage method involves sequestration in saline aquifers, where extreme temperature and pressure conditions challenge existing modeling techniques. Traditional ab initio methods, though highly accurate, are computationally prohibitive for the large size and time scales required in direct coexistence simulations. Meanwhile, classical molecular dynamics (MD) simulations often lack the necessary transferability, and experimental data at relevant conditions are sometimes scarce and, in the case of dissolution rate, highly inconsistent. In this work, we address these limitations by developing machine learning models that replicate the accuracy of ab initio methods, specifically density functional theory (DFT), while significantly reducing computational costs. However, many of these functionals can only accurately model one of the molecules in the mixture. Thus, after evaluating various DFT functionals for their ability to reproduce higher-accuracy coupled cluster results, we train machine learning models on the selected functionals. These machine learning models enable us to simulate large-scale systems and perform direct coexistence simulations under these conditions. Our approach allows us to predict key properties such as VLE, diffusion, viscosity, interfacial tension, and solubility, providing crucial insights.