A New Digital Reconstruction Method of Porous Materials Based on Diffusion Model

Porous materials are ubiquitously present in natural environments and engineering applications, where their pore structural characteristics serve as critical factors influencing functional performance. Consequently, quantitative characterization of porous structures have attracted multidisciplinary attention, providing theoretical support for aerospace engineering, electrochemical engineering, functional materials, and biochips. The rapid advancement of computational technologies has enabled digital reconstruction as effective tool for characterizing the microstructures of porous materials. Machine learning-based methodologies have established novel pathways for digital reconstruction of porous materials. Thus, a new structural generation method for porous materials is proposed based on diffusion model. By using the Bentheimer sandstone 2D slices from the digital rocks super-resolution dataset, the image generation is implemented through denoising diffusion models following data preprocessing. Both Fréchet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS) metrics are used to evaluate the generated porous images. The results demonstrate the superior performance of diffusion models in generating high-fidelity images, achieving FID and LPIPS scores of 284.6933 and 0.165 respectively. The generated porous images exhibit enhanced structural authenticity compared to conventional methods. The present model provide a comprehensive framework for structural reconstruction of porous materials.