AI-driven quality classification of concrete after high-temperature exposure: a hybrid framework for structural integrity assessment

Abstract Accurately assessing the residual performance of fire-damaged concrete is crucial for structural safety but challenging due to non-linear degradation. Traditional methods often suffer from data scarcity and lack objective quality criteria, creating a prediction-decision gap. This study introduces a novel AI-driven quality classification framework, shifting from numerical prediction to actionable decision-making. We established a comprehensive database (334 data points) integrating compressive strength and ultrasonic pulse velocity (UPV). To address data scarcity, the SMOTE algorithm expanded the dataset to 3,006 points. Our key innovation is a hybrid AI approach: K-means clustering identified inherent data patterns, followed by a Decision Tree (DT) to establish objective, rule-based criteria for four quality grades (Safety, Caution, Warning, Danger). Six machine learning models were rigorously evaluated using an independent test set (80/20 split). While Gradient Boosting achieved the highest accuracy (97.6%), the DT model was optimal, balancing high accuracy (96.1%) with superior interpretability (white-box model) and computational efficiency. UPV and temperature were confirmed as the dominant factors. This framework provides a reliable and practical tool for the immediate assessment of fire-damaged concrete structures.