Intelligent Inference of 3D Seepage Channels and Prediction of Seepage Behavior for Engineered Cementitious Composites Based on Diffusion-Based Generative Modeling

The post-cracking seepage behavior of Engineered Cementitious Composites (ECC) is critical to their service security and durability. However, conventional methods are not capable of in-situ seepage assessment due to the difficulty in characterizing internal fissures from observable surface cracks. This study was to develop an intelligent approach for inferring 3D seepage channels from surface cracks and accurately predicting the seepage performance, thereby providing a novel pathway for in-situ durability evaluation of cracked ECC structures.

A computer vision-based approach was employed to achieve a high-precision characterization of surface microcracks and internal 3D fissures. A dual pre-modification deep learning strategy was proposed for semantic segmentation of surface cracks, significantly improving the accuracy to 99.87%. For internal fissure characterization, a transformer-based super-resolution model coupled with a fine segmentation network was developed to enhance computed tomography (CT) voxel resolution by 4×4×4 times, enabling extraction of fine fissures less than 50 μm in width. Also, a novel diffusion-based generative model was designed to intelligently infer 3D internal fissures from 2D surface cracks. The model could integrate optical flow mechanisms to ensure structural coherence and physical plausibility. Furthermore, the Lattice Boltzmann Method (LBM) was utilized to simulate seepage flow within both real and generated fissures, considering complex geometric features and fiber blocking effects, to validate the permeability performance of the inferred fissures.

The proposed diffusion model effectively generates 3D fissures that closely resemble real ones obtained from CT. The key geometric parameters (i.e., tortuosity, roughness, and average width) are differed by less than 5% between generated and real fissures. The results of the LBM simulations further demonstrate that a ratio of hydraulic width to geometric width for generated fissures has an error of less than 9%, compared to real ones. These results indicate that the intelligently inferred fissures can reliably replicate the seepage behavior of actual ECC cracks, enabling to accurate in-situ permeability assessment based solely on surface crack information.

This study presented an integrated framework combining computer vision, diffusion generative modeling, and LBM for intelligent inference and seepage prediction of 3D seepage channels in post-cracked ECC. The key achievements could include, i.e., 1) highly accurate surface crack segmentation; 2) enhanced internal fissure characterization overcoming CT resolution limitations; 3) effective 3D fissure inference from 2D surface cracks with high geometric fidelity; and 4) validated seepage performance of generated fissures via the LBM simulation. The proposed method could offer a promising tool for in-situ durability assessment of ECC structures. A future work could be needed to extend an approach to multiple cracks and validate it under broader experimental conditions.