Accurate prediction of temperature-induced degradation in concrete is essential for improving structural fire safety and supporting reliable post-fire engineering decisions. However, previous studies have generally focused on conventional machine learning applications or limited optimization strategies, while integrated frameworks combining systematic input screening, robust validation, large-scale metaheuristic optimization, and interpretable analysis remain limited. This study aims to develop a comprehensive predictive framework for estimating the temperature-induced weight loss and compressive strength of concrete using advanced machine learning techniques. First, a detailed collinearity analysis was performed to filter the input dataset, eliminate redundant correlations, and improve statistical reliability. For modeling consistency, all fiber-containing mixtures were treated as polymer-fiber systems, and fiber-related variables were interpreted as polymer-fiber descriptors. To reduce overfitting and ensure robust validation, 5-fold cross-validation was applied during training, while 23% of the dataset was reserved as a strictly independent test set. In addition, 25 metaheuristic algorithms were evaluated under a standardized computational budget of 5000 function evaluations to perform neural architecture search. The results showed that the Marine Predators Algorithm (MPA), Symbiotic Organisms Search (SOS), and Kepler Optimization Algorithm (KOA) achieved superior convergence behavior in optimizing hybrid Levenberg–Marquardt-trained networks. SHapley Additive exPlanations (SHAP)-based sensitivity analysis further revealed that matrix-related properties, particularly unit weight and water absorption capacity, were the dominant drivers of thermal degradation. Overall, the proposed framework provides not only a robust benchmarking platform for predictive modeling but also a practically relevant and interpretable tool for post-fire structural assessment and thermally resilient concrete design.
Beytekin et al. (Thu,) studied this question.