Natural carbonation of concrete: a data-driven analysis

This study presents a data-driven analysis of long-term natural carbonation in concrete, using a newly compiled database comprising 1079 mixes and 8194 carbonation depth measurements over 65 years, representing one of the largest natural carbonation datasets assembled to date. Four different tree-based machine learning models (CatBoost, XGBoost, Random Forest and Decision Tree) were evaluated for predicting carbonation rate (k), with CatBoost emerging as the most effective. Partial dependence and SHAP analyses were used to quantify the relative importance of 13 features influencing k, including binder composition, mix proportion, curing and exposure conditions. The water-to-calcium oxide (w/CaO) ratio emerges as the most important feature, encapsulating the effects of porosity, carbonatable content, and supplementary cementitious material (SCM) type and content. Higher SCM replacement levels increase carbonation, while aggregate-related features show comparatively less pronounced effects. Carbonation environment is a more important feature than curing condition in predicting k. By combining machine learning with interpretable SHAP analysis, the model independently recognised trends consistent with literature findings. This study offers a comprehensive dataset and robustly selected features to facilitate future machine learning-based predictions of concrete carbonation.