A data-driven prediction of residual carbon dioxide under different porous media structures

Abstract

The capture and storage of carbon dioxide has been the most prevalent method on mitigating CO₂ emissions in the atmosphere caused by global energy demands so as to combat climate change and global warning. Deep saline formations provide a vast area for underground storage of carbon dioxide by which immobile gas can be stored permanently through residual trapping. Residual trapping renders CO₂ immobile in form of clusters under equilibrium conditions within the pore spaces. Pore network modeling and machine learning techniques have been applied to further investigate the influence of porous media structural properties on residual gas trapping. Multiphase flow behavior in CO₂ – brine systems was investigated on 28 porous media samples representing sandstone, carbonates and sandpacks. The porous media information was extracted from Micro CT images of these samples to create pore networks and obtain structural properties. The porous media structural properties and simulation results were used as input data for machine learning models. SVM and XGBoost were deployed to predict residual saturation and classify the rock structures on their trapping potential. Pore size and throat size distribution, pore – throat ratio and connectivity impacted the fluid transport behavior within the porous media. High capillary pressure was observed in samples with small pores and throats distributed non-uniformly in the media such as sandstone and carbonates. Higher residual trapping was observed in larger pore-throat ratios which is conducive for snapoff events. The more the connectivity in the porous media, less trapping was observed as connectivity affects the ability of fluid to escape the media from one end to the other. The prediction of residual carbon dioxide was successful for both machine learning models with XGBoost model performing better than SVR in regression with a correlation coefficient of 0.97. Residual CO₂ saturation was classified into high (>0.5) and low (<0.5) to create a binary classification on which XGBoost performed better than SVC with a 100 % accuracy. Saturation data showed to have less impact in both regression and classification in machine learning whereas the structural properties showed more contribution to successful predictions with connectivity significantly affecting the machine learning performance.