AI Model Predicts Hidden Damage in Recycled Concrete Used in Cold Climates
Researchers have developed a machine learning system that can predict structural weakening in recycled aggregate concrete exposed to freeze-thaw cycles—a critical safety issue for infrastructure in northern regions. The model identifies hidden damage patterns that traditional inspection methods miss, enabling builders and engineers to prevent costly failures before they occur.
Originaltitel: Mechanism-driven modeling of bond softening in recycled aggregate concrete after extreme freeze-thaw cycles: An explainable Boosting approach
Recycled aggregate concrete (RAC) offers significant environmental and economic benefits by mitigating construction waste and pollution. Reinforced concrete structures in extremely cold regions are susceptible to complex damage mechanisms driven by the interaction of freeze-thaw cycles (FTC) and mechanical loading. Consequently, prioritizing the bonding performance between RAC and rebar is critical for maintaining structural integrity. The pullout load will form a “softening zone” around the rebar, which produces radial cracks spreading outward in a complex stress state. Determining the extent of the softening zone is essential to improving bonding performance. This study investigates the formation mechanisms and influencing factors of the softening zone after freeze-thaw exposure, employing an advanced computational framework that integrates a modified softening sleeve ring theory with five Boosting machine learning algorithms. SHAP and PDP-2D analyses are employed to interpret the effects of multiple features on softening radius ( r ), bond strength ( τ u ), peak slip ( s u ), and failure patterns ( M ). XGBoost demonstrated optimal regression performance ( R 2 values: r = 0.950 , τ u = 0.933 , s u = 0.932 ). SHAP results indicate that r is mainly affected by water-cement ratio, rebar diameter, and minimum freeze-thaw temperature; τ u by tensile strength, maximum size of the recycled aggregate, and modulus of elasticity; and s u by compressive strength, sand ratio, and maximum size of the recycled aggregate. Classification performance was enhanced with SMOTE, with CatBoost achieving the highest classification test accuracy (96.2 %), and compressive strength was the most influential factor across failure patterns. PDP-2D results indicated that r decreases and stabilizes with increasing FTC and is significantly affected by the coupling of other variables. The developed Boosting model offers high accuracy and interpretability and can guide the engineering design of RAC structures in extremely cold environments.