Abstract
Ultra-high-performance fiber-reinforced concrete (UHPFRC) exhibits exceptional tensile properties, but its tensile strength is highly dependent on fiber distribution, orientation, and count, making accurate strength estimation challenging. This study introduces a novel approach in which tensile strength estimation is achieved by analyzing fiber characteristics at predicted cracking locations using deep learning. Using X-ray computed tomography (CT) and image analysis techniques, the fiber orientation factor (μ(0)) and average efficiency factor ((μ(1))(-)) were determined at predicted cracking locations. A deep learning model (YOLOv11) was trained to identify regions with a defective distribution, achieving a mean Average Precision (mAP@0.5) of 0.87, demonstrating its high reliability in predicting cracking locations. The overall cracking location prediction success rate was 73% for strain-hardening specimens. The estimated tensile strength was then compared with uniaxial tensile test (UTT) results, revealing an average experiment-estimation error of 5.72% and an average theory-estimation error of 3.34% for strain-hardening specimens, whereas strain-softening specimens exhibited significantly higher errors, with an average experiment-estimation error of 43.09% and an average theory-estimation error of 15.73%. These findings highlight the strong correlation between fiber count, cracking behavior, and tensile strength in UHPFRC, offering a trustworthy, non-destructive framework for estimating tensile performance in UHPFRC elements.