3D solid digital and numerical modeling of multimineral heterogeneous rocks based on deep learning

Mesostructure has a significant effect on the macromechanical response of rock. However, it is challenging to precisely describe the 3D mesostructure of multimineral rock. Physical equipment such as Computed Tomography cannot distinguish different mineral grains with similar densities, and the 3D heterogeneous rock model reconstructed by the parametric method is severely distorted. In this study, to reconstruct the complex structure of multimineral rock and verify its characterization ability, a deep learning method is adopted to generate 3D digital models for six types of granites, namely, maple red, Sanbao red, Wulian red, yellow rust stone, sesame gray and tan brown, based on the randomness and self-similarity of texture structures. A comparison of the slices of natural granite, the first-order digital model and the second-order digital model shows that the digital models have high texture similarities with natural granite in terms of histogram matching, cosine distance and fingerprint characteristics. High geometric similarities are also observed in grain morphology and proportion. In addition, it is proven that the 3D homogeneous or 2D heterogeneous model cannot represent the mechanical properties of the complex 3D heterogeneous model under uniaxial compression. Finally, the second-order digital model can characterize the axial deformation and compressive strength of the first-order digital model under uniaxial compression. Therefore, the digital model can well characterize the texture and geometric characteristics and has good accuracy and stability in simulating the mechanical properties of natural multimineral granite.

Automated summary

This study adopts a deep learning method to generate 3D digital models for six types of granites, and compares them with natural granites to evaluate their characterization ability. The results show high texture and geometric similarities between the digital models and natural granites. It is also demonstrated that 3D homogeneous or 2D heterogeneous models cannot accurately represent the mechanical properties of complex 3D heterogeneous models. The second-order digital model is found to effectively characterize the mechanical properties of the first-order digital model.