In recent years, the continuous development of 3D reconstruction methods for porous media has promoted the integrated evolution of image modeling, performance prediction, and multiscale microstructure analysis. Although traditional reconstruction approaches achieve satisfactory matches for low-order statistical descriptors, they often struggle to accurately reproduce realistic geometric morphology and topological characteristics across multiple scales. Meanwhile, traditional generative adversarial network (GAN) methods exhibit remarkable advantages in handling complex porous structures (e.g., cement-based materials), yet are prone to issues such as mode collapse and gradient saturation. To address the aforementioned challenges, a high-fidelity 3D reconstruction method based on deep convolutional generative adversarial networks (DCGANs) is proposed in this paper, incorporating an autoencoder module within the generator to enhance multiscale feature representation. The model introduces a porosity constraint term in the adversarial learning framework and integrates Wasserstein generative adversarial networks with gradient penalty (WGAN-GP), significantly improving training stability and convergence performance. Trained on CT images of new silicoaluminate cement paste at different curing ages, the model demonstrates good performance in both geometric morphology and statistical feature consistency of the reconstructed results.

The samples were prepared using the new silicoaluminate (limestone and calcined clay cement) system with 55% (mass fraction) Portland cement, 30% calcined clay, 15% limestone powder, and a water-to-binder ratio of 0.42. High-resolution micro-CT images were obtained at curing ages of 3, 28, 90 d, and 180 d. The images were preprocessed using Gaussian filtering and threshold segmentation to obtain binarized pore structures. A DCGAN architecture was developed by integrating 3D generative adversarial networks (GANs) with an autoencoder module. The framework incorporated gradient penalty to establish a WGAN-GP based on the Wasserstein distance, thereby constraining the Lipschitz continuity of the discriminator. A porosity constraint term was introduced into the adversarial learning framework to ensure global statistical consistency of the reconstructed microstructures. The reconstructed 3D volumes were validated against experimental microstructures using Minkowski functionals (porosity, surface area, Euler characteristic) and spatial correlation descriptors including the two-point correlation function (S2), lineal-path function, two-point cluster function, and surface correlation function.

The results demonstrate that the proposed model effectively reconstructs pore morphology and distribution characteristics across different curing ages, with the porosity of reconstructed volumes showing good consistency with measured samples. The introduction of a gradient penalty term in the loss function enhanced the training stability and convergence efficiency of the model. By incorporating porosity as a regularization term in the adversarial learning framework, the model was effectively guided to learn the porosity evolution trends across different curing ages, thereby improving the consistency between generated structures and experimental data. The generator employed a combined ReLU and GELU activation mechanism, which contributed to stable generator training. The integration of Instance Normalization in the discriminator further improved the quality of generated results and the generalization capability of the model. Comparative analyses based on Minkowski functionals and statistical descriptors revealed that the generated structures exhibit high agreement with the original CT images in terms of geometric morphology, connectivity, and spatial correlation. Compared to conventional reconstruction methods, the proposed approach improves computational efficiency by approximately 150% while maintaining structural fidelity.

This study constructed an efficient 3D generative adversarial network (GAN) model based on CT images of cement paste, achieving high-precision 3D reconstruction of pore structures at different curing ages. The proposed model demonstrates good stability and robustness during training and achieves high-quality reconstruction of pore structures in cement-based materials while effectively preserving structural diversity and morphological realism.