Predicting long-term carbonation in cement-based materials is crucial for assessing the durability, carbon capture potential, and service life of concrete structures. To enable machine learning-based predictions, two databases have been compiled, each containing data on carbonation depth in cement-based systems under natural and accelerated conditions. Each database includes measurements from over 1000 paste, mortar, and concrete mixes collected over the past 65 years using the phenolphthalein spray method. The datasets cover a wide range of variables, including the chemical composition and physical properties of binders, the type and replacement level of supplementary cementitious materials, mix compositions, fresh and hardened properties, and curing and carbonation conditions. The relationships between these variables and carbonation coefficients were analysed using Spearman's rank correlation, analysis of variance (ANOVA) on rank integrated with Post Hoc test and a CatBoost machine learning model. The relative contributions of different variables to carbonation were further analysed using SHapley Additive exPlanations (SHAP) values from the CatBoost model. The findings indicate that the water-to-CaO ratio (w/CaO ratio) is the most significant factor influencing carbonation under both natural and accelerated conditions. This is followed by the carbonation environment, curing condition, and the aggregate-to-paste ratio (Agg./paste ratio), all showing strong correlations with the carbonation coefficient (k). The insights gained will guide the selection and handling of variables in future machine learning models for predicting carbonation in cement-based materials.