This study investigates the differences in flexural behavior of ultra-high-performance concrete (UHPC) arising from variations in test methods and key experimental parameters. Flexural tensile tests were conducted on 51 specimens representing 17 combinations of test variables, including steel fiber length (13 mm and 19.5 mm), specimen cross-sectional dimensions (75×75 mm, 100×100 mm, and 150×150 mm), presence or absence of a notch, and loading configuration (three-point and four-point loading). The tests were performed in accordance with ASTM C1609 and EN 14651, and both deflection and crack mouth opening displacement (CMOD) were normalized by the span length to compare the influence of each parameter. The notched specimens demonstrated significantly improved reliability, exhibiting up to an 8.4-fold reduction in standard deviation due to the consistent initiation of cracking. Regarding size effects, the 75×75 mm specimens showed an overestimation of flexural performance due to the wall effect of fiber distribution, whereas the 100×100 mm and 150×150 mm specimens exhibited similar flexural responses. The comparison of loading configurations revealed that three-point loading produced up to 11.7% higher flexural tensile strength than four-point loading, attributable to concentrated moment–shear interaction and the combined effects of fiber bridging and shear resistance mechanisms. In addition, specimens with longer steel fibers (19.5 mm) exhibited 5.2–9.7% higher flexural performance than those with shorter fibers (13 mm), which is attributed to enhanced interfacial bonding and improved crack dispersion capacity.

This study investigates the differences in flexural behavior of ultra-high-performance concrete (UHPC) arising from variations in test methods and key experimental parameters. Flexural tensile tests were conducted on 51 specimens representing 17 combinations of test variables, including steel fiber length (13 mm and 19.5 mm), specimen cross-sectional dimensions (75 × 75 mm, 100 × 100 mm, and 150 × 150 mm), presence or absence of a notch, and loading configuration (three-point and four-point loading). The tests were performed in accordance with ASTM C1609 and EN 14651, and both deflection and crack mouth opening displacement (CMOD) were normalized by the span length to compare the influence of each parameter. The notched specimens demonstrated significantly improved reliability, exhibiting up to an 8.4-fold reduction in standard deviation due to the consistent initiation of cracking. Regarding size effects, the 75 × 75 mm specimens showed an overestimation of flexural performance due to the wall effect of fiber distribution, whereas the 100 × 100 mm and 150 × 150 mm specimens exhibited similar flexural responses. The comparison of loading configurations revealed that three-point loading produced up to 11.7% higher flexural tensile strength than four-point loading, attributable to concentrated moment–shear interaction and the combined effects of fiber bridging and shear resistance mechanisms. In addition, specimens with longer steel fibers (19.5 mm) exhibited 5.2–9.7% higher flexural performance than those with shorter fibers (13 mm), which is attributed to enhanced interfacial bonding and improved crack dispersion capacity.

Reliable estimation of the peak shear strength of squat reinforced concrete (RC) walls remains challenging because their resistance is governed by interacting concrete and reinforcement mechanisms. Conventional strut-and-tie models (STMs) employ fixed coefficients, whereas purely data-driven machine learning (ML) models do not explicitly represent the underlying force-transfer mechanisms. To address these limitations, this study develops a mechanics-informed graph neural network-guided strut-and-tie model (STM-GNN). The framework embeds the STM force-transfer topology within a five-node graph representing the wall panel, diagonal concrete strut, boundary element, horizontal truss, and vertical truss. The GNN learns specimen-specific correction factors for the concrete strut-tie, horizontal-truss, and vertical-truss contributions rather than applying a single correction to the total shear strength. Physics-based constraints enforce non-negative resistance components, bounded correction factors, and compliance with the theoretical concrete-strut capacity. An ensemble of STM-GNN models estimates specimen-level uncertainty, while SHAP analysis provides feature-level interpretation. The framework was developed and evaluated using 614 RC squat-wall specimens compiled from experimental studies. On the held-out test set, the ensemble STM-GNN achieved an (R2) of 0.978, an MAE of 58.74 kN, and an RMSE of 88.80 kN. The mean ratio was 0.974, with a coefficient of variation of 11.57%; 79.12% and 95.60% of the predictions were within ±15% and ±30% of the experimental strengths, respectively. Compared with the calibrated classical STM, the proposed model reduced the MAE and RMSE by 35.5% and 39.1%, respectively, and outperformed eight benchmark analytical models. SHAP analysis identified wall length, wall height, wall thickness, and boundary-element width as the dominant variables, with effects consistent with established shear-transfer behavior. The proposed STM-GNN therefore improves predictive accuracy while preserving mechanism-level interpretability and providing uncertainty-informed shear-strength estimates.