Work

A deep learning framework for microstructural analysis of nano-modified cementitious composites using metal intrusion and BSE imaging

Year: 2025

Type: article

Abstract: Nanomodification is a promising technology for advancing construction materials toward multifunctionality, intelligence, and sustainability. However, current assessments of its effects on microstructure modification require improvements in objectivity, quantification, and pattern generalization. This study proposes a novel characterization-analysis framework, integrating metal intrusion technology with deep learning to enable expertise-independent extraction and evaluation of microstructure characteristics. Using image-based algorithms, the optimization of nanopore structure in cementitious waste rockfill material is first described. Based on a dataset comprising over 4000 microscopic images, the proposed deep learning model achieved a maximum 82.0 % accuracy at a 39.7 μm × 39.7 μm observation scale in distinguishing microstructure images of samples with various water-cement ratios and graphene oxide (GO) reinforcement. Porosity and fractal dimension show weak correlation with classification accuracy, suggesting insufficient description of these parameters on microstructure characteristics. The class activation mapping algorithm further revealed that the model consistently prioritized pore structure identification. The deep Taylor decomposition (DTD) algorithm extracted microstructure features that concentrated on the pore distribution near hydration products, where the GO groups exhibited denser and less continuous pore structure. Finally, a coefficient of variation matrix was employed to fuse micropores image data with DTD features data to generate typical pore probability distribution maps. Nanomodified pore structures exhibit discretized spatial distributions and lower overall pore probabilities, especially at low water-cement ratios. The established framework paves the way for intelligent, automated analysis of nanomodified microstructure, offering significant potential for future construction engineering applications of nanomaterials and deep learning technologies.

Source: Case Studies in Construction Materials

DOI: 10.1016/j.cscm.2025.e05694