Machine Learning Methods and Computationally Efficient Techniques in Digital Rock Analysis

Digital Rock Analysis involves (1) 3D X-ray 𝜇CT imaging and processing, (2) identifying and segmenting the minerals, and (3) performing flow simulation to obtain upscalable petrophysical parameters. Limitations exist at each step, primarily: (1) the resolution and Field of View (FOV), (2) bias and accuracy of identification and segmentation, and (3) the accuracy and computational intensity of direct simulation. These limitations are surpassed with machine learning and efficient simulation techniques. Super Resolution Convolutional Neural Networks (SRCNNs) and Enhanced Deep Generative Adversarial Networks (EDSRGANs) are shown in 2D and 3D to compensate for resolution-FOV limitations. SRCNNs boost resolution and recover edge sharpness, while EDSRGANs also recover texture. The noise reduction of SRCNNs precondition for image segmentation. Physical accuracy measured by phase topology and permeability achieves the closest match with EDSRGAN. Generalisation with augmentation shows high adaptability to noise and blur. Regenerated under-resolution features and comparison with SEM images shows consistency with underlying geometry. A custom formulated Deep CNN, U-ResNet and other networks are trained to perform 3D multi-mineral segmentation to eliminate user-bias, manual tuning, and algorithmic limitations inherent in traditional methods. U-ResNet performs most accurately and reliably, achieving the highest voxelwise accuracy and most consistent physical accuracy measured by calculating the topology of segmented mineral phases and comparing single and multi-phase direct flow simulations. Several techniques are proposed for efficient single and multi-phase flow at steady-state conditions. Single-phase flow in large images can be estimated using a Dual Grid Domain Decomposition (DGDD) that significantly reduces memory computational requirements, allowing workstations to solve supercomputer size problems. Multi-phase flow can be accelerated with a Morphologically Coupled Multi-phase Lattice Boltzmann Method (MorphLBM), rapidly computing capillary dominated flows, typically 5x faster using a Shell Aggregation morphing method. A U-net CNN can also rapidly estimate steady-state velocity fields, used as-is or as preconditioner in direct LBM simulation (ML-LBM). Similarly, the same acceleration procedure can also be coupled to Pore Network Models and Semi-Analytical Solvers to form accelerated direct simulation techniques. At each step of the Digital Rock workflow, machine learning methods and efficient techniques enhance results past physical limits and/or boost performance of traditional techniques.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2020

Author dc:creator

• Wang, Ying Da

• Deep Learning
• Porous Media
• Digital Rock
• Machine Learning
• Fluid Mechanics
• Image Processing

Statement dc:rights

• open access
• CC BY-NC-ND 3.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by-nc-nd/3.0/au/

Language dc:language

EN

Identifier

https://doi.org/10.26190/unsworks/22141

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/70165