Applicability of 2D algorithms for 3D characterization in digital rocks physics: an example of a machine learning-based super resolution image generation

Abstract Digital rock physics is based on imaging, segmentation and numerical computations of samples. Due to challenges regarding the handling a large 3-dimensional (3D) sample, 2D algorithms have always been attractive. However, in algorithms, efficiency pore structures third direction generated 3D sample questionable. We used four individually captured µCT-images given Berea sandstone with different resolutions (12.922, 9.499, 5.775, 3.436 µm) evaluate super-resolution images by multistep Super Resolution Double-U-Net (SRDUN), algorithm. Results show that unrealistic features form due section-wise reconstruction images. To overcome this issue, we suggest generate three samples using SRDUN directions then use one two strategies: compute average (reconstruction averaging) or segment one-directional combine them together (binary combination). numerically physical properties (porosity, connected porosity, P- S-wave velocity, permeability formation factor) these models. reveal compared samples, harmonic averaging leads more similar original sample. On other hand, trends can be calculated binary combination strategy generating low, medium high porosity These are compatible obtained from averaged as long scale difference between input output small enough (less than about 3 our case). By increasing difference, dispersed results obtained.