(169bl) Maxwell-Stefan Diffusivities of Oil-CO2 Mixtures in Nanopores: Physics and Machine Learning Models

Transport properties of oil and CO₂ confined in nanoscale pores are key inputs for physics-based modeling of CO₂-enhanced oil recovery (EOR) in unconventional reservoirs. Due to confinement and fluid-wall interactions, oil and CO₂ molecules exist as heterogeneous mixtures across nanopores. Their transport properties can deviate from those of bulk mixtures, but the corresponding models for nanoconfined liquid mixtures are lacking. This study investigates the Maxwell-Stefan (MS) diffusivities of CO₂-C10 (1: CO₂; 2: C10) mixtures in nanopore under the compositions relevant to CO₂ Huff-n-Puff by molecular dynamics (MD) simulations.

In the composition space explored here, D₁₂ (characterizing the CO₂-C10 interaction) is insensitive to the mixture composition in contrast to that without nanoconfinement. D_1,s (CO₂-wall interaction) increases sharply with CO₂ loading, while a nonmonotonic dependence on C10 loading is observed on the D_2,s (C10-wall interaction). Also, surprisingly, opposite to the expectation for dense fluid mixtures confined in nanopores, D_2,s is negative. These observations can ultimately be traced to the fact that CO₂ molecules have far stronger wall affinity than C10 molecules to the calcite wall, resulting in significantly heterogeneous density distribution and ultra-low mobility of the adsorbed CO₂ molecules.

As MD simulations are computationally expensive, a Multi-Output Gaussian Process machine learning model is developed as a surrogate model to predict the MS diffusivities in the vast composition space efficiently. Trained from a limited MD dataset, the model achieves a less than 10% relative root mean square error in the specified composition space.