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

Area of Science:

• Civil Engineering
• Materials Science
• Artificial Intelligence in Engineering

Background:

• Global natural aggregate scarcity drives demand for manufactured sand (M-sand) in concrete.
• Efficient concrete mix proportioning is crucial for performance and sustainability.
• Existing methods may not fully leverage data-driven insights for M-sand concrete.

Purpose of the Study:

• To develop an integrated, data-driven decision-support framework for concrete mix proportioning with M-sand.
• To enhance prediction accuracy and interpretability of concrete performance.
• To optimize mixes for reduced environmental impact and cost while meeting strength requirements.

Main Methods:

• Utilized XGBoost for performance prediction and SHAP for interpretability.
• Employed quantile gradient boosting for uncertainty quantification.
• Applied NSGA-II multi-objective optimization for mix design.
• Validated models with extensive laboratory data and field placement.

Main Results:

• XGBoost model achieved R²=0.989 for 28-day compressive strength, significantly outperforming baseline models.
• SHAP analysis identified water-binder ratio as the primary predictor.
• Optimized mixes reduced carbon emissions by 11.6% and cost by 3.2%, exceeding C40 strength.
• Field validation showed prediction errors below 5% with a lab-to-field transfer function (R²=0.94).

Conclusions:

• The proposed data-driven framework offers a replicable template for M-sand concrete mix design.
• The integrated approach successfully balances performance, cost, and environmental considerations.
• Domain-specific retraining is necessary for direct model transfer to different contexts.