CO2/brine relative permeability estimation using effective rock/fluid properties: A machine learning-based approach

Relative permeability is a crucial parameter in CO2 storage as it describes the fluids' transport in porous media and residual saturations. CO2 injection into saline aquifers leads to CO2 dissolution in brine and subsequent formation of carbonic acid, resulting in complicated reactions with the rock minerals. These reactions include mineral dissolution, mineral precipitation or silicate weathering, depending on the composition and texture of the reservoir rock. As a consequence, these reactions bring about alterations in key rock characteristics including mineralogy, porosity, absolute permeability, and wettability. The main objective of this study is to develop a CO2/brine relative permeability model using machine learning (ML) algorithms based on a more comprehensive, yet representative, set of input parameters including rock/fluid properties and their compositions, instead of time-consuming experiments. In this study, experimental relative permeability data were data mined from the literature, along with some additional fluid and rock properties including absolute permeability, porosity, fluid density, fluid viscosity, CO2/brine interfacial tension, CO2 solubility, rock minerals’ concentration, brine salinity and composition, and fluid saturations. The dataset was pre-processed, and missing data were identified and populated using applicable correlations. Various supervised ML approaches were tried to predict the relative permeability values, including Decision Tree Regression (DTR), Gaussian Process Regression (GPR), Linear Regression (LR), and Bayesian Ridge Regression (BRR). In a comparative analysis, the same ML algorithms were trained and tested, incorporating only three typical input parameters that have been repeatedly used recently in the literature for relative permeability modelling (i.e., porosity, absolute permeability, and fluids saturation). Results indicate that there was a good agreement between the predicted and experimental relative permeability data for models developed using DTR and GPR algorithms when all input parameters were incorporated. However, the models were not nearly as accurate when only three conventionally agreed input parameters were used. A sensitivity analysis was also performed to determine the most accurate input parameters and their impact on relative permeability predictions. According to the sensitivity analysis results, the most important input parameters are brine salinity, CO2 and water viscosity, porosity, absolute permeability, and iron and halite content of the rock samples. Overall, the results indicate the critical role of rock-fluid interactions and composition, as well as fluid properties, in predicting relative permeability using ML-based approaches, underscoring their significance for future studies.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.