Understanding and predicting concrete carbonation are significant in designing durability and maintaining the service life of reinforced concrete structures. However, this purpose can hardly be reached because of complex carbonation mechanisms depending on various variables such as cement content (C), fly ash content (FA), water content (W), concentration of CO 2 , relative humidity (RH), temperature (T C), and exposition time (Time). This investigation proposes a machine learning (ML) approach including eight ML algorithms such as four single ML models: XGB, GB, RF, and SVM, and four hybrid ML models: XGB_RRHC, GB_RRHC, RF_RRHC, and SVM_RRHC for investigating and predicting concrete carbonation depth containing fly ash. To achieve this purpose, a dedicated database consisting of 688 samples and seven input variables is built, and the performance of eight machine learning models is compared. Single ML model Extreme Gradient Boosting (XGB) using the default hyperparameters exhibited the highest performance with R 2 = 0.9770, RMSE = 2.2725 and MAE = 1.5218. Shapley Additive exPlanations (SHAP) identifies the most influential feature and order of feature effect on concrete carbonation depth. The first four important features can be sorted in order: time of exposition > cement content > water content > CO 2 concentration. Moreover, at a higher value of exposition time, water content, CO 2 concentration, fly ash content, temperature and relative humidity, the carbonation depth of concrete increases. Using high content of cement can reduce the carbonation depth of concrete.