Graph Neural Network Frameworks for Pore-scale Modelling and Characterization of Subsurface Porous Media

This thesis examines Graph Neural Network-based methodologies for pore-scale modelling and characterization of subsurface porous media, which is a vital aspect of petroleum and energy resources engineering. It presents innovative frameworks for predicting properties that influence microscopic transport behavior in porous media. It addresses the gaps imposed by traditional numerical methods, such as user bias and computational requirements, and by other deep learning techniques, such as Convolutional Neural Networks, including their inflexibility towards inputs of variable sizes and sparse data representations. First, an overview of the workflow of microscopic porous media modelling and characterization is presented, including micro-CT image acquisition, image pre-processing, image segmentation, and pore scale numerical and deep learning methods. Second, Pore-GNN, a GNN-based framework that predicts flow properties in porous media using micro-CT images, is presented. It encompasses the utilization of traditional Convolutional Neural Network-based feature extractors to construct graphs, and training various graph convolutional layers to predict the flow-based properties. Third, MicroGraphNets are presented as an approach to automate microscopic wettability characterization by employing message-passing Graph Neural Networks to predict the surface and interfacial properties, namely, surface roughness, in-situ interfacial curvatures, and contact angles. Finally, the applications of Graph Neural Networks were extended to the simulation of single- and multiphase fluid flow in porous media. In this approach, message-passing Graph Neural Networks were trained to predict the flow dynamics of the next time step from the current time-step dynamics. The results of these studies demonstrate the notable capabilities of Graph Neural Networks to replace traditional numerical and deep learning approaches and provide a versatile solution to mitigate their shortcomings. The main findings are summarized as follows: (1) Pore-GNN ...