An explainable AI-driven framework for predicting carbonation depth in recycled aggregate concrete under environmental exposure

Abstract

Accurate prediction of carbonation depth (CD) is crucial for assessing the long-term durability of recycled aggregate concrete (RAC); however, it remains challenging due to the significant nonlinearity between material composition and environmental exposure conditions. This study proposes a robust and explainable AI framework to predict CD in RAC, utilizing advanced hybrid and ensemble learning methods. A database comprising experimental records from previous studies was developed, including 12 key input variables related to mixture design and environmental exposure. Six machine learning models were implemented and evaluated: XGB-GWO, SVR-PSO, RF-BO, ANN-GA, CNN-LSTM, and ENSEMBLE-AVG. These models combine advanced learning algorithms with optimization strategies to improve predictive accuracy and generalization. Statistical evaluation indicates that XGB-GWO offers the best individual model performance, while the ENSEMBLE-AVG strategy enhances robustness and stability by lowering overall prediction uncertainty. Sensitivity analysis shows that exposure time (ET) and water-to-binder ratio (WB) are the key factors influencing carbonation depth, with aggregate density (DA) and superplasticizer content (SP) following, in the order of ET > WB > DA > SP. An interactive GUI that incorporates PDP-ICE analysis effectively visualizes nonlinear relationships between input variables and carbonation progression, improving model interpretability and practical use. The proposed framework provides a dependable, clear, and decision-focused tool for assessing durability and designing sustainable RAC structures.