An interpretable XGBoost-SHAP machine learning model for reliable prediction of mechanical properties in waste foundry sand-based eco-friendly concrete

Construction and development projects worldwide heavily rely on concrete as their primary building material, making it an essential component of global infrastructure and growth. The various key ingredients that compose concrete contribute differently to its overall environmental footprint. Moreover, rapid urban and industrial growth has strained ecological systems and depleted resources, necessitating environmentally friendly substitutes for traditional concrete ingredients. Waste foundry sand has emerged as a potential replacement for natural sand in concrete mixtures, offering a sustainable option. Since evaluating waste foundry sand effects on concrete through laboratory methods is resource-intensive, therefore, this research employs six distinct ensemble boosting algorithms including CatBoost, XGBoost, HistGBRT, NGBoost, LightGBM, Adaboost and MLR to predict key properties of concrete made with waste foundry sand: split tensile strength, compressive strength, and elastic modulus. This investigation utilized SHapley additive exPlanations (SHAP) method to precisely illustrate feature interdependencies, quantify their complex relationships, and establish a hierarchy of importance. The models' performance was then assessed using multiple robust metrics, including mean absolute error (MAE), root mean squared error (RMSE), correlation coefficient (R), Nash-Sutcliffe efficiency coefficient (NSE), Kling-Gupta efficiency (KGE), and Willmott's Index (WI). The findings reveal that the XGBoost model excels in estimating the compressive strength, achieving an RMSE of 2.845 MPa and an R-value of 0.958. CatBoost follows as the second-best performer, with an RMSE of 3.083 MPa and an R-value of 0.952. For the elastic modulus estimation, XGBoost again outperforms other models, yielding an RMSE of 0.992 GPa and an R-value of 0.990, while NGBoost secures the second position with an RMSE of 1.256 GPa and an R-value of 0.985. In predicting the split tensile strength, XGBoost once more demonstrates superior accuracy, recording an RMSE of 0.296 MPa and an R-value of 0.925, closely follows by LightGBM with an RMSE of 0.302 MPa and an R-value of 0.926. This research demonstrates the effectiveness of XGBoost algorithms in predicting key properties of concrete with waste foundry sand when relevant factors are considered.