Predicting the compressive strength of high-performance concrete using an interpretable machine learning model.

Area of Science:

• Civil Engineering
• Materials Science
• Data Science

Background:

• Accurate concrete strength prediction is crucial for construction efficiency and quality assurance.
• Current methods may lack interpretability, hindering understanding of influencing factors.
• High-performance concrete (HPC) requires precise strength estimations for optimal application.

Purpose of the Study:

• To introduce an interpretable machine learning (ML) framework for predicting the compressive strength of HPC.
• To enhance the accuracy of concrete strength predictions by incorporating derived features.
• To identify the most influential features affecting HPC compressive strength.

Main Methods:

• Developed and compared four interpretable ML models: Random Forest (RF), AdaBoost, XGBoost, and LightGBM.
• Integrated feature derivation and random search with 5-fold cross-validation for hyperparameter optimization.
• Applied SHapley Additive exPlanations (SHAP) to analyze feature importance in the LightGBM model.

Main Results:

• The study successfully constructed interpretable ML models for HPC compressive strength prediction.
• Age, water/cement ratio, slag, and water content were identified as key predictors.
• Superplastic/cement ratio, slag/cement ratio, and ash/cement ratio showed nonsignificant impacts on predicted strength.

Conclusions:

• Interpretable ML models, particularly LightGBM, offer reliable predictions for HPC compressive strength.
• Feature engineering and advanced ML techniques improve prediction accuracy and provide insights.
• Understanding key influencing factors enables better material selection and construction practices.