Automated and Explainable Machine Learning Framework for Predicting the Fracture Toughness of Fiber-Reinforced Ultra-High-Performance Concrete Subjected to Freeze–Thaw Damage

The extended longevity of fiber-reinforced ultra-high-performance concrete (FR-UHPC) in frigid regions remains incompletely comprehended, especially concerning the deterioration of mode-I fracture toughness (KIC) due to repeated freeze–thaw cycles (FTCs). While prior investigations have predominantly concentrated on traditional mechanical attributes such as compressive, tensile, and flexural strengths, there is a scarcity of predictive frameworks that can precisely simulate the intricate nonlinear interactions between mixture composition, fiber features, porosity, and freeze–thaw damage impacting fracture toughness. To bridge this deficiency, this research introduces an integrated and interpretable automated machine learning framework designed to forecast the fracture toughness of FR-UHPC exposed to as many as 300 FTCs. A detailed experimental database consisting of 756 laboratory-evaluated specimens with varying fiber types, fiber geometries, mixture proportions, curing durations, and freeze–thaw exposure levels was compiled and utilized to methodically assess twelve state-of-the-art machine learning algorithms under a standardized optimization and validation protocol. Unlike traditional methods relying on manual model selection, the proposed framework leverages the Tree-based Pipeline Optimization Tool (TPOT) to automatically determine the optimal predictive pipeline and hyperparameters while removing user bias. Model resilience was tested using Grid Search and five-fold cross-validation, whereas SHAP analysis was employed to provide physically interpretable feature attribution aligned with fracture mechanics principles. Among all evaluated models, TPOT delivered the highest predictive performance, achieving an R2 of 0.95 and an RMSE of 0.10 MPa·m1ᐟ2 on the hold-out dataset, while cross-validation yielded R2 values between 0.92 and 0.96 with RMSE values ranging from 0.08 to 0.12 MPa·m1ᐟ2, showcasing excellent robustness and generalization capability. SHAP analysis additionally identified freeze–thaw cycles and fiber volume content as the key factors influencing fracture toughness degradation. The proposed framework pushes forward the state of the art by integrating automated machine learning, comprehensive model benchmarking, and explainable artificial intelligence within a unified predictive framework, thus providing an accurate, interpretable, and practically applicable tool for durability-oriented design and long-term performance assessment of FR-UHPC infrastructure exposed to severe freeze–thaw environments.