Generation of cement paste microstructure using machine learning models

The microstructure of cement paste determines the overall performance of concrete and therefore obtaining the microstructure is an essential step in concrete studies. Traditional methods to obtain the microstructure, such as scanning electron microscopy (SEM) and X-ray computed tomography (XCT), are time-consuming and expensive. Herein we propose using Denoising Diffusion Probabilistic Models (DDPM) to synthesize realistic microstructures of cement paste. A DDPM with a U-Net architecture is employed to generate high-fidelity microstructure images that closely resemble those derived from SEM. The synthesized images are subjected to comprehensive image analysis, phase segmentation, and micromechanical analysis to validate their accuracy. Findings demonstrate that DDPM-generated microstructures not only visually match the original microstructures but also exhibit similar greyscale statistics, phase assemblage, phase connectivity, and micromechanical properties. This approach offers a cost-effective and efficient alternative for generating microstructure data, facilitating advanced multiscale computational studies of cement paste properties. • This paper de monstrates the use of Denoising Diffusion Probabilistic Models (DDPM) to synthesize high-fidelity microstructures of cement paste, providing a cost-effective alternative to SEM and XCT. • Synthesized microstructures were validated through greyscale statistical analysis, phase segmentation, and micromechanical property analysis, showing close resemblance to original SEM-derived microstructures. • Generated microstructures exhibit logical distributions of LD-CSH, HD-CSH, and UHC phases, aligning with classical Portland cement hydration theories and accurately reflecting micromechanical properties. • The DDPM approach improves efficiency in microstructure generation without the need for phase assumptions, offering advantages over traditional methods.