Deep learning-based 3D reconstruction of cement hydration: Topological densification of C-S-H and morphological growth of CH

Quantifying the 3D spatial evolution of hydration products is challenging due to stereological limitations inherent to 2D imaging and insufficient spatial resolution of macroscopic averaging methods. This study characterizes the spatiotemporal evolution of C-S-H and CH using SliceGAN deep learning reconstruction. The volumetric accuracy was rigorously validated against independent measurements from thermogravimetric analysis (TGA) and quantitative X-ray diffraction (QXRD). The 3D topological features and spatial distribution of the hydrated paste were analyzed at ages of 1, 7, 28, and 196 days. Results indicate that the C-S-H gel evolves from a loose, disconnected network into a dense, continuous skeleton as hydration proceeds. The local thickness of the C-S-H walls increases consistently, reflecting the progressive densification and space-filling nature of the gel growth. Meanwhile, the CH phase exhibits continuous coarsening, with characteristic lengths increasing from ~0.8 μm to ~4.5 μm. CH maintains a plate-like geometry (aspect ratio > 2.0) with random spatial orientation. Spatial analysis of the generated models suggests that the volumetric growth of CH drives a progressive shift in their geometric centroids relative to the C-S-H interface, implying that the modeled phase growth effectively occupies capillary voids. The statistically reconstructed CH volumes align with the converted TGA volumetric results with an absolute deviation of approximately 1 vol.% at maturity (D28, D196), and the volumetric growth of C-S-H exhibits a trend consistent with experimental metrics. These findings confirm the framework's reliability for quantitative 3D spatial characterization of cementitious materials.