Explainable Machine Learning for Structural Materials: SHAP-Enhanced Ensemble Prediction of Concrete Compressive Strength

Accurate prediction of concrete compressive strength is critical for structural design and quality assurance, yet traditional methods struggle with complex non-linear relationships between mixture proportions and mechanical properties. This study proposes a two-layer ensemble learning framework integrated with SHAP-based interpretability for transparent strength forecasting. The architecture employs seven diverse base learners (Random Forest, K-Nearest Neighbors, Deep Neural Network, Stacking, CatBoost, Support Vector Machine, Gaussian Process Regression) in the first layer, followed by a LightGBM meta-learner that adaptively synthesizes predictions to achieve superior accuracy (R2=0.948, RMSE=4.161 MPa). SHAP analysis quantifies global feature importance, identifying age (mean |SHAP|=1.5) and water content (mean |SHAP|=1.2) as dominant factors, while waterfall plots provide instance-level explanations for mixture optimization. Experimental validation on 1,030 concrete specimens demonstrates the framework's dual capability of delivering accurate predictions and actionable engineering insights, advancing interpretable AI applications in structural materials design.