Abstract
Accurately estimating reservoir rock properties and understanding capillary trapping mechanisms are crucial for fluid storage and flow modeling in porous media, particularly for applications such as carbon dioxide sequestration (CCS) and underground hydrogen storage (UHS). This thesis presents a comprehensive workflow that combines machine learning techniques and pore-network modeling approach to predict petrophysical properties and assess the impact of microscopic pore structures on capillary trapping. In the first part, a convolutional neural network (CNN) framework is used to predict rock properties—such as porosity, throat area, and pore surface area—from micro-computed tomography (micro-CT) X-ray images. It has been observed that incorporating a greater variety of rock sample types during model training enhances the accuracy of predicting the properties of unknown rock samples. The model, trained on Bentheimer and Castlegate sandstone samples, achieved highest accuracy in predicting the properties of Leopard sandstone, with mean absolute percentage errors of 2.19%, 3.04%, and 6.08% for porosity, pore surface area, and throat area, respectively. Additionally, a regression model using extreme gradient boosting (XGBoost) predicted absolute permeability with an R² of 0.813 for Leopard sandstone, demonstrating the importance of pore network parameters like tortuosity in determining permeability. In the second part, this study evaluates the role of pore structures and networks in UHS by analyzing capillary trapping through a pore-network model and Land's trapping coefficient. The Land constant showed a wide variation in the model constant, ranging from 0.2 to 2.7, depending on the pore properties of the rock sample. The analysis revealed that carbonate rocks exhibit higher Land’s constants compared to sandstone, implying lower hydrogen trapping efficiency in carbonate reservoirs. Field-scale reservoir simulations further quantified the fraction of injected hydrogen gas trapped due to capillary forces, linking pore-scale properties to reservoir-scale behavior. Together, these methods offer an efficient and scalable approach for predicting the morphological and flow properties of porous media, contributing valuable insights for selecting optimal sites for CCS and UHS.
Date of publication
Fall 12-14-2024
Document Type
Thesis
Language
english
Persistent identifier
http://hdl.handle.net/10950/4794
Committee members
Dr. Fernando Resende, Dr. Aaditya Khanal, Dr. Prabha Sundaravadivel
Degree
Master's in Mechanical Engineering
Recommended Citation
Khan, Md Irfan, "Application of digital rock physics and machine learning to improve the understanding of fluid flow behavior through porous media" (2024). Mechanical Engineering Theses. Paper 39.
http://hdl.handle.net/10950/4794