Explainable machine learning for predicting drying shrinkage in mortars with ground recycled concrete cement

Drying shrinkage (Ԑ_sh) is a critical time-dependent deformation that governs crack susceptibility, durability, and long-term performance of cementitious materials. Ground recycled concrete cement (GRC) has emerged as a sustainable supplementary cementitious material (SCM), yet its influence on Ԑ_sh remains insufficiently understood. This study develops an explainable machine learning framework to predict and interpret the Ԑsh of mortar incorporating GRC, explicitly considering oxide-level chemical composition, including calcium oxide (CaO), silicon dioxide (SiO₂), aluminium oxide (Al₂O₃), iron oxide (Fe₂O₃), and particle size rather than treating GRC as a bulk dosage variable. A comprehensive dataset compiled from published experimental studies, including mix design, curing age, aggregate content, and GRC chemical and physical properties, was used to train and benchmark three ensemble learning algorithms: Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost). CatBoost achieved superior generalization performance with a coefficient of determination of 0.992 and the lowest prediction errors. Mixture-grouped cross-validation further confirmed its robustness against potential data leakage among related mixtures. Post-hoc interpretation using SHapley Additive exPlanations (SHAP), Individual Conditional Expectation (ICE), and Local Interpretable Model-agnostic Explanations (LIME) identified curing age as the dominant Ԑ_sh driver, followed by GRC chemical composition and SCMs such as fly ash. The proposed explainable machine learning framework integrates accurate prediction with global- and local-level explanations, enabling shrinkage-aware design of sustainable cementitious materials incorporating recycled constituents and supporting data-informed materials engineering.