Inferring Internal Structures of Porous Media From Orthogonal External Surfaces: A Novel 2D‐to‐3D Reconstruction Method Based on a Deep Conditional Diffusion Model

Abstract Diffusion models, a class of generative models renowned for producing realistic images, hold significant promise for reconstructing complex three‐dimensional (3D) porous media. Nevertheless, existing approaches predominantly generate stochastic microstructures visually resembling the training data but often struggle to accurately recover opaque internal architectures from sparse external observations. Recognizing that external surfaces of porous media may encode critical insights about their internal microstructures, we introduce an Attenuation‐based Conditional Diffusion Model (AB‐CDM) to establish this relationship. AB‐CDM translates external surface image conditions into detailed volumetric reconstructions, achieving unprecedented fidelity in both voxel‐level statistics and spatial correlation functions across diverse rock materials, including the widely used DeePore data set and two real‐core volumes. We found that AB‐CDM can effectively reconstruct complex internal microstructures from any three mutually orthogonal external surfaces. This capability arises from a set of learned axis‐dependent attenuation parameters that automatically regulate how surface information decays along each spatial direction and directly influence the feature maps within the 3D U‐Net. To the best of our knowledge, this study is the first to elucidate the intrinsic link between opaque internal microstructures and multiple external surface images through a 3D conditional generative model, thereby enabling data‐driven 3D reconstruction of porous media directly from 2D observations, with broad implications for digital rock physics and geoscientific modeling.