Abstract
We present a memory-efficient algorithm for significantly enhancing the quality of segmented 3D micro-Computed Tomography (micro-CT) images of rocks using a Machine Learning (ML) Generative Model. The proposed model achieves a 16 × increase in resolution and corrects inaccuracies in segmentation caused by the overlapping X-ray attenuation in micro-CT measurements across different minerals. The generative model employed is a 3D Octree-Based Progressive Growing Deep Convolutional Wasserstein Generative Adversarial Network with Gradient Penalty (3D OB PG DC WGAN-GP). To address the challenge of extremely high memory consumption inherent in standard PyTorch 3D Convolutional (Conv3D) layers, which is a significant constraint in 3D Super-Resolution (SR) applications, we implemented an Octree structure within the 3D Progressive Growing Generator (3D PG G) model. This enabled the use of memory-efficient 3D Octree-Based Convolutional layers provided by the open-source Minkowski Engine library. The adoption of the octree structure was pivotal in overcoming the long-standing memory bottleneck in volumetric deep learning, making it possible to reach 16 × Super-Resolution in 3D, a scale that is challenging to attain due to cubic memory scaling. For training, we utilized segmented 3D Low-Resolution (LR) micro-CT images along with unpaired segmented complementary 2D High-Resolution (HR) Laser Scanning Microscope (LSM) images. Post-training, we achieved high-quality, segmented 3D SR images with resolutions improved from 7 to 0.44 µm/voxel and accurate segmentation of constituent minerals. Validated on Berea sandstone, this framework demonstrates substantial improvements in pore characterization and mineral differentiation, which are key factors for accurate Digital Rock Physics (DRP) simulations. The proposed algorithm advances the feasibility of large-scale, high-resolution 3D reconstructions and offers a robust solution to one of the primary computational limitations in modern geoscientific imaging.