Conditional generative AI for high-fidelity synthesis of hydrating cementitious microstructures
Minfei Liang (TU Delft - Materials and Environment, University of Oxford)
Kun Feng (Southwest Jiaotong University)
Jinbao Xie (TU Delft - Materials and Environment)
Yuyang Wei (University of Oxford)
Sonia Contera (University of Oxford)
Erik Schlangen (TU Delft - Materials and Environment)
Branko Šavija (TU Delft - Materials and Environment)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Portland cement paste has a highly heterogenous evolving microstructure that complicates the development of stronger and greener cementitious materials. Microstructure is the fundamental input of multiscale studies on material behaviors. Herein, we propose a conditional generative AI framework for synthesizing high-fidelity 3D microstructures of hydrating cement paste (1–28 days) with varying water-to-cement ratios and Blaine fineness values. A latent diffusion transformer, operating within a compact two-stage latent space derived via a vector quantized variational autoencoder, efficiently captures and reproduces experimentally measured microstructural patterns. Statistical analyses confirm strong consistency in grey value distributions, micromechanical properties, hydration phase evolution, and particle size distributions, with only minor boundary-related discrepancies. Validation using a pretrained classifier further corroborates the fidelity of generated microstructures. This approach provides a robust tool for realistic cement paste microstructure generation, supporting multiscale modeling and advancing the design of sustainable cementitious materials.
help_outline