Explainable Hybrid Ensemble Learning for Sustainable Concrete Compressive Strength Prediction and Mix Design Optimization

Abstract

Accurate prediction of concrete compressive strength is essential for ensuring structural reliability while promoting sustainable construction practices through the efficient utilization of supplementary cementitious materials. This study proposes an explainable machine learning framework for predicting concrete compressive strength and optimizing sustainable concrete mix designs using a dataset comprising 1,030 concrete mixtures. Following data preprocessing and feature engineering, several machine learning algorithms, including Linear Regression, Support Vector Regression, Random Forest, Gradient Boosting, Extra Trees, LightGBM, and XGBoost, were evaluated. The results demonstrated that XGBoost achieved the highest predictive performance, attaining an R² value of 0.9205, MAE of 3.0829 MPa, and RMSE of 4.4687 MPa. To enhance model interpretability, SHapley Additive exPlanations (SHAP) were employed to quantify the influence of input variables on compressive strength predictions. The explainability analysis revealed that curing age, water-binder ratio, and binder content were the most influential factors governing strength development. Furthermore, a sustainability-oriented optimization framework was implemented to identify environmentally favorable concrete mixtures with reduced cement consumption and increased utilization of fly ash and blast furnace slag. The optimal sustainable mixture achieved a sustainability score of 0.736 while maintaining a predicted compressive strength of 33.22 MPa. The findings demonstrate that explainable machine learning provides a reliable and interpretable decision-support tool for sustainable concrete design, contributing to low-carbon construction and resource-efficient infrastructure development.