Concrete strength prediction using domain-driven feature engineering and explainable LightGBM

Abstract

Accurately predicting the compressive strength of concrete is essential for optimizing mix designs, ensuring structural integrity, and reducing construction costs. This study integrates domain-informed feature engineering with optimized tree-based ensemble models to benchmark concrete strength prediction on a widely recognized Yeh dataset. We expanded the baseline feature set with 11 engineered variables, evaluating five architectures: Random Forest, XGBoost, LightGBM, Extra Trees, and a Voting Regressor. Among these, LightGBM emerged as the top-performing architecture with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} of 0.959 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RMSE$$\end{document} of 3.47 MPa. SHAP analysis revealed that the newly engineered features were the most influential predictors, with Age-Cement Interaction and Cementitious-to-Water Ratio identified as the primary drivers of model accuracy. To facilitate practical adoption, the optimized framework was deployed via a PyQt5-based desktop interface and a FastAPI web service, enabling mix design optimization across different user environments.