Image segmentation-based quantification of phase distribution in coarse aggregate ultra-high performance concrete and its impact on mechanical properties

Abstract

This study investigates the mechanical performance and phase distribution characteristics of coarse aggregate ultra-high performance concrete (CA-UHPC) using vertically cast specimens (100 mm × 400 mm × 1000 mm). A series of mixes with varying coarse aggregate (CA) and steel fiber contents were designed to analyze their effects on vertical distribution, compressive and flexural properties. Advanced image recognition techniques—including a hybrid ResNet50+U-Net deep learning model with cross-attention mechanisms—were developed to quantify CA distribution, while morphology-based image processing was used to evaluate steel fiber dispersion. Results showed significant downward migration of both CA and steel fiber during casting, leading to mechanical property gradients along specimen height. The maximum differences in compressive and flexural strength reached 52.53 MPa and 16.77 MPa, respectively. Specimens with 17 % CA and 2.5 % steel fiber exhibited relatively uniform internal distribution and favorable mechanical performance, suggesting a synergistic interaction between the CA-induced skeletal framework and the fiber bridging mechanism. In contrast, excessive CA content (30 %) led to fiber clustering and reduced ductility, while insufficient fiber dosage resulted in brittle failure. These findings offer a novel perspective for evaluating material homogeneity and may inform future mix design strategies for CA-UHPC in large-scale structural applications.