Scholarlink

<p>Digital rock physics (DRP) is often constrained by the high cost and limited accessibility of pore-scale X-ray computed tomography (CT) imaging. This study proposes a conditional diffusion model (CDM) to generate pore-scale rock CT images under multi-level conditions using classifier-free guidance and a residual U-Net noise predictor. Two condition settings are investigated: Under weakly conditioned generation, porosity is prescribed as the condition, and 100 CT images are generated for each target porosity under different classifier-free guidance strengths γ. The generated results are quantified using porosity statistics and distributional distance metrics computed from the normalized gray-level frequency histograms. An optimal balance is achieved at guidance strength <i>γ</i> = 3, where the Wasserstein distance between normalized gray-level histograms is reduced to 3.973 ± 0.0718. The extrapolation behavior is further examined by sampling at out-of-distribution porosity targets, revealing increased sensitivity to guidance strength beyond the training distribution. Under strongly conditioned generation, phase composition maps of pore, clay, and quartz are incorporated as condition to guide the reverse diffusion process. Evaluation over 1000 generated images demonstrates high structural similarity, with a mean SSIM of 0.9393 ± 0.0030. Finally, 3D reconstructions using Marching Cubes demonstrate voxel-level and mesh-level consistency between generated and original samples. These results show the proposed CDM provides an effective promote for digital rock physics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Generative Modeling of Digital Rock CT Images Within a Conditional Diffusion Framework

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Zhe-Yu Yang,
  • Zhi Zhao,
  • Xiao-Ping Zhou

摘要

Digital rock physics (DRP) is often constrained by the high cost and limited accessibility of pore-scale X-ray computed tomography (CT) imaging. This study proposes a conditional diffusion model (CDM) to generate pore-scale rock CT images under multi-level conditions using classifier-free guidance and a residual U-Net noise predictor. Two condition settings are investigated: Under weakly conditioned generation, porosity is prescribed as the condition, and 100 CT images are generated for each target porosity under different classifier-free guidance strengths γ. The generated results are quantified using porosity statistics and distributional distance metrics computed from the normalized gray-level frequency histograms. An optimal balance is achieved at guidance strength γ = 3, where the Wasserstein distance between normalized gray-level histograms is reduced to 3.973 ± 0.0718. The extrapolation behavior is further examined by sampling at out-of-distribution porosity targets, revealing increased sensitivity to guidance strength beyond the training distribution. Under strongly conditioned generation, phase composition maps of pore, clay, and quartz are incorporated as condition to guide the reverse diffusion process. Evaluation over 1000 generated images demonstrates high structural similarity, with a mean SSIM of 0.9393 ± 0.0030. Finally, 3D reconstructions using Marching Cubes demonstrate voxel-level and mesh-level consistency between generated and original samples. These results show the proposed CDM provides an effective promote for digital rock physics.