TopoRSNet: A Unique Approach to Maintaining Topological Features for Digital Rock Image Rescaling With Minimal Quality Degradation
Abstract The composition, microstructure, and physical behavior of rock at the core‐to‐pore scale are commonly visualized and characterized using 3D X‐ray microcomputed tomography (micro‐CT). However, a key geophysical problem in micro‐CT rock characterization is how to efficiently handle and proces...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2025-06-01
|
| Series: | Journal of Geophysical Research: Machine Learning and Computation |
| Subjects: | |
| Online Access: | https://doi.org/10.1029/2024JH000557 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract The composition, microstructure, and physical behavior of rock at the core‐to‐pore scale are commonly visualized and characterized using 3D X‐ray microcomputed tomography (micro‐CT). However, a key geophysical problem in micro‐CT rock characterization is how to efficiently handle and process large numbers of high‐resolution 3D rock images, and how to preserve multiscale features for characterization at coarse resolution. Typically, a single high‐resolution 3D micro‐CT image can be extremely large with more than 1010 voxels, and in the case of dynamic scans, hundreds of such large images will be generated, posing challenges in storage, downstream analysis, modeling, and their application to deep learning. Existing image rescaling methods face challenges such as information loss during down‐sampling and limited feature recovery during up‐sampling. Herein, we present TopoRSNet, an image rescaling network designed to adjust micro‐CT images to an optimal size and resolution while preserving global and local representative features. Feature preservation is ensured by introducing three feature‐based loss functions: adversarial loss, a new feature consistency loss, and a novel persistent homology loss (PH loss). Combined with a pixel consistency loss, we assure the preservation of both pixels and features during rescaling. The efficacy of TopoRSNet is validated on two common geological rocks, and the results are compared to other rescaling methods. This method enables scalable analysis of multiscale rock features, allowing for broader integration of 3D imaging in geoscientific modeling, simulation, and machine learning workflows. |
|---|---|
| ISSN: | 2993-5210 |