A data-driven framework for mix design optimization of limestone calcined clay cement concrete using machine learning and experimental validation

Limestone calcined clay cement (LC³) offers a promising binder for reducing clinker in concrete; however, its compressive strength is influenced by both binder chemistry and mix proportions. Although many studies have investigated LC³, an integrated approach that combines machine learning prediction, optimization, and experimental validation for mix design is still limited. This study proposes a metaheuristic-assisted machine learning based framework to predict the compressive strength of LC³ concrete, and to identify an optimal mix design. The optimized mix was then subjected to preliminary experimental validation. A literature database of 172 LC³ concrete mixtures compiled from 42 published studies was used, with 19 input variables describing oxide compositions of Ordinary Portland cement (OPC), calcined clay (CC), and limestone powder (LSP), together with key mix parameters. Four ensemble learners (CatBoost, XGBoost, Random Forest, and LightGBM) were trained using an 80/20 train–test split. Five-fold cross-validation was applied to ensure model reliability. CatBoost provided the best generalization on the test set compared to other models studied. SHAP results revealed that the water-to-binder ratio (W/B) was the most important factor in determining strength, followed by the content of coarse aggregate (CA) content, sand-to-binder ratio (S/B), and LSP content. The trained CatBoost model was coupled with a Gray Wolf Optimizer (GWO) to obtain a feasible optimal combination. The strength was predicted to be 57.21 MPa. The average strength of 100 mm cubes was 51.54 MPa with standard deviation of 0.21 MPa. The proposed framework has the potential to be used for data-driven design of LC³ concretes.