What question did this study set out to answer?

This study aims to create an innovative framework for designing concrete mixes using manufactured sand and data-driven methods.

May 9, 2026Open Access

Interpretable XGBoost framework for multi-objective manufactured sand concrete mix design

LZLihao ZhangNorth China University of Technology TLTie LiHunan International Economics University

Puntos clave

This study aims to create an innovative framework for designing concrete mixes using manufactured sand and data-driven methods.
Integrated XGBoost performance prediction with SHAP interpretability and NSGA-II optimization.
Utilized 1200 laboratory datasets and validated with field data of 6350 m3.
Applied quantile gradient boosting for uncertainty quantification.
XGBoost model achieved R2=0.989 and RMSE=0.789 MPa for 28-day compressive strength.
Identified water-binder ratio as the primary predictor contributing 80.7% to model performance.
NSGA-II optimization reduced carbon emissions by 11.6% and material costs by 3.2% while exceeding C40 strength requirements.

Resumen

The global shortage of natural aggregates and the rise of manufactured sand (M-sand) necessitate innovative approaches for concrete mix proportioning. This study proposes an integrated, data-driven decision-support framework combining XGBoost-based performance prediction, SHAP-based interpretability, quantile gradient boosting for uncertainty quantification, and NSGA-II multi-objective optimization. Based on 1200 original laboratory datasets (C30-C45 grade, northern China, limestone stone powder) validated against 6350 m3 of field placement, the XGBoost model achieves test-set R2=0.989 and RMSE = 0.789 MPa for 28-day compressive strength—a 39.2% improvement over a W/B linear regression baseline and 90.1% over the JGJ 55-2011 Abrams formula. SHAP analysis identifies water-binder ratio as the dominant predictor (80.7% mean absolute contribution), with stone powder content exhibiting a mechanistically explained three-stage response. Quantile gradient boosting provides 90% prediction intervals; empirical coverage falls below the nominal target (PICP = 68.8% for 28-day strength), reported honestly as a current limitation. NSGA-II Pareto optimization identifies solutions reducing carbon emissions by 11.6% and material cost by 3.2% while exceeding C40 requirements by ≥ 17%. Field validation confirmed mean prediction error below 5%, with a lab-to-field transfer function (R2=0.94). This framework provides a replicable methodological template; direct model transfer requires domain-specific retraining.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo