Diffusion Model‐Based Controlled Reconstruction Method for 3D Fractured Rock Masses

Abstract

Accurate, cost‐effective, and controllable reconstruction of three‐dimensional (3D) digital samples of fractured rock masses is crucial for understanding the macro‐mechanical effects of fractures. Traditional reconstruction methods are not only expensive but also limited by sample representativeness, often failing to capture the true complexity of fractures. In light of these limitations, this paper introduces an innovative method based on diffusion models, termed Text2Rock, for the reconstruction of 3D fractured rock mass digital samples with enhanced controllability. By incorporating a CLIP model and an improved U‐Net3D network architecture, our approach enables control over key fracture characterization parameters such as fractal dimension, fracture intensity, and Feret's maximum diameter, thereby enhancing both the accuracy and controllability of the reconstructed samples. Reconstructed samples demonstrate that the proposed method can generate digital rock samples that closely adhere to the specified parameters, with reliability validated through two‐point correlation coefficients and statistical analyses. Furthermore, the paper illustrates the application of these reconstructed samples in uniaxial compression numerical simulations, providing a new framework for investigating the macro‐mechanical properties of fractured rock masses.