Deep Learning Applications in Reconstructing and Upscaling Porous Medium and Flow Functions
Open Access
- Author:
- Ren, Zihan
- Graduate Program:
- Energy and Mineral Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- December 02, 2024
- Committee Members:
- Shimin Liu, Program Head/Chair
Sharon Huang, Outside Unit & Field Member
Sanjay Srinivasan, Chair & Dissertation Advisor
Russell Johns, Major Field Member
Dustin Crandall, Special Member
Luis Ayala H, Major & Minor Field Member - Keywords:
- Geology CO2 storage
Deep Learning
3D Structure Generation
Two Phase Flow
Porous Medium
Upscaling - Abstract:
- Accurate representation of flow functions, particularly absolute permeability and relative permeability is crucial for understanding subsurface fluid behavior in Geologic Carbon Storage (GCS) applications. While micro-CT scanning and digital rock physics have improved our understanding of multiphase flow processes through image simulation techniques such as pore network modeling, lattice Boltzmann, or direct image simulation, their limited scale and high acquisition costs pose significant challenges in bridging the gap between detailed pore-scale physics and field-scale reservoir characteristics. Such disconnection has long prevented the effective upscaling of flow functions for practical reservoir simulations. This thesis presents a systematic investigation into methodologies for upscaling both porous medium and flow functions through the reconstruction of porous media conditioned to larger-scale rock properties. We present methodologies evolving from simple statistical correlations between flow functions and field-scale characteristics to increasingly sophisticated deep-learning approaches such as GANs and transformers. We first explore GAN-based reconstruction methods with property conditioning, which successfully generate synthetic structures but face limitations in capturing long-range spatial dependencies. This exploration led to our key innovation: a transformer-based architecture for multi-dimensional porous media reconstruction through an autoregressive approach. From a pure image reconstruction perspective, the transformer-based approach adopts a robust multi-token generation strategy for high-dimensional spatial object reconstruction involving multiple image patches, which could inspire research related to video generation, 3D object reconstruction, and similar applications. In the context of porous media, it provides a viable approach for spatial upscaling and enables arbitrary-size generation. Although the current prototype is only a baby transformer model trained on a total of 6 micro-CT scans (each with a voxel resolution of $636^3$ voxels), this approach demonstrates its capability in generating spatially coherent porous media that honor field-scale constraints while accurately predicting both absolute and relative permeability. This success validates our hypothesis that autoregressive modeling can effectively capture the complex spatial relationships crucial for accurate flow function prediction. Our work serves as a fundamental stepping stone toward a more holistic approach to solving upscaling challenges in subsurface characterization. The methodologies developed here extend beyond pore-scale reconstruction, offering potential applications in digital material reconstruction, inverse problems, and reservoir model generation.