AI-driven prediction of carbonation depth in recycled aggregate concrete: Influential parameters and nonlinear interactions from machine learning perspectives

To advance predictive modeling in concrete durability, this study presents a transparent, AI-based framework specifically designed to infer carbonation depth (CD) in recycled aggregate concrete (RAC). By employing a rigorously curated dataset comprising 579 test results across 16 independent sources, the model integrates 12 influential variables, including the material properties and environmental factors. A decision forest regression algorithm, fine-tuned via exhaustive grid search, exhibited robust performance with coefficients of determination of 0.924, 0.884, and 0.839 for the training, validation, and testing sets, respectively. Root mean square error values of 1.93 mm, 2.38 mm, and 2.61 mm further confirm its accuracy. SHAP-based sensitivity analysis revealed that carbonation behavior is predominantly driven by exposure time, water-to-binder ratio, aggregate density, CO 2 concentration, and superplasticizer content, together accounting for over 83 % of the model’s explanatory power. Marked increases in CD were observed when WB exceeded 0.55 and CO 2 levels surpassed 15 %, while greater cement content (>450 kg/m³) and coarse aggregates of high density (>2600 kg/m³) reduced carbonation risks. With an inference speed ranging from 82.1 to 248.1 ms, the model is well-suited for deployment in real-time monitoring systems, bridging the gap between predictive accuracy and practical engineering utility in durability management.