2D-to-3D Digital Rock Reconstruction Using a Diffusion Model Constrained by Multimodal Information

The application of generative AI in the field of digital rock analysis has significantly advanced the reconstruction of digital twins, especially in the intelligent generation of three-dimensional core samples. However, research on the reconstruction of reliable rocks in highly heterogeneous and severely missing strata remains relatively scarce. This paper proposes a digital core reconstruction method based on a latent diffusion model, aiming to utilize limited two-dimensional slices to achieve three-dimensional core reconstruction. This method captures highly heterogeneous information of the micro-pores in the generative direction through channel embedding while achieving multi-output. Simultaneously, multimodal information that controls pore morphology is embedded into the model’s network layers at multiple scales, enhancing the model’s controllability in recovering digital cores. Test results on cores of different intervals and types indicate that the introduction of multimodal information enhances the stability and reliability of core reconstruction. Our model demonstrates generalization ability, effectively reconstructing highly heterogeneous core samples, providing a foundation for expanding digital core libraries and achieving controlled generation of rock transition zones.