Prediction of Compressive Strength of Concrete Using Explainable Machine Learning Models

Area of Science:

• Materials Science
• Civil Engineering
• Data Science

Background:

• Accurate prediction of concrete compressive strength is crucial for structural integrity and quality control.
• Existing empirical formulas struggle with the complex, non-linear relationships inherent in concrete mix design.
• Machine learning offers a powerful alternative for modeling these intricate interactions.

Purpose of the Study:

• To develop and compare interpretable machine learning models for predicting concrete compressive strength.
• To identify key factors influencing concrete strength and their interactions.
• To provide a data-driven tool for optimizing concrete mix proportions.

Main Methods:

• Utilized Gradient Boosting Trees, Random Forest, and Backpropagation Neural Networks.
• Employed Bayesian optimization for hyperparameter tuning.
• Applied SHAP (SHapley Additive exPlanations) analysis for feature importance and interpretability.
• Trained and validated models on 223 sets of concrete compression test data.

Main Results:

• The CatBoost model demonstrated superior predictive performance with R² of 0.9388, RMSE of 2.7131 MPa, and MAPE of 5.45%.
• SHAP analysis revealed cement content as the most influential factor, followed by water content, water reducer, fly ash, and aggregates.
• The CatBoost model significantly outperformed the industry standard empirical formula in specific scenarios.

Conclusions:

• Interpretable machine learning, particularly the CatBoost model, provides a highly accurate and reliable method for predicting concrete compressive strength.
• Understanding feature contributions through SHAP analysis enables data-driven optimization of concrete mix designs.
• This approach offers a valuable tool for enhancing engineering design and quality assurance in concrete construction.