Machine learning-based prediction of crack mouth opening displacement in ultra-high-performance concrete

Ultra-high-performance concrete (UHPC) has become an essential construction material due to its exceptional strength, durability, and crack-resistance properties, making it well-suited for long-span bridges, protective structures, and demanding infrastructure applications. However, accurately predicting crack mouth opening displacement (CMOD) in fiber-reinforced UHPC (FR-UHPC) presents significant challenges, as traditional empirical and physics-based models struggle to capture the complex nonlinear relationships between mix design, fiber geometry, and structural parameters. This study assembled a comprehensive experimental database containing eleven mix and material variables that control post-cracking behavior. Nine advanced machine learning algorithms were trained and evaluated using fivefold cross-validation, including kernel-based regressors, ensemble methods, deep neural networks, and the innovative tabular prior-data fitted networks (TabPFN). The modela were ensured interpretability through SHapley Additive exPlanations (SHAP), sensitivity analysis, contour mapping, and interaction diagrams. These analyses consistently showed that fiber volume (FV) and fiber length (FL) were the primary factors controlling CMOD, while silica fume content (SF), fly ash content (FA), initial notch depth (a₀), and water-to-binder ratio (w/b) had secondary effects. Feature selection improved predictive performance by narrowing the input space to the six most influential variables. With this optimized configuration, TabPFN delivered exceptional accuracy (R2 = 0.942, RMSE < 0.072 mm), surpassing ensemble methods (XGBoost, RFR, GBR) and kernel-based approaches (SVR, NuSVR, GPR). Model predictions were validated against experimental CMOD curves, and bootstrap resampling generated 95% confidence intervals, demonstrating that TabPFN provided both high precision and reliable uncertainty estimates. This research contributes three main innovations: (i) creation of a detailed experimental database for FR-UHPC fracture behavior, (ii) first application of TabPFN for CMOD prediction, and (iii) combined focus on interpretability and uncertainty quantification. The transparent, uncertainty-aware predictive framework developed here connects data-driven modeling with fracture mechanics, providing practical tools for designing and optimizing resilient UHPC structures.