Explainable Machine Learning for Multicomponent Concrete: Predictive Modeling and Feature Interaction Insights.

Area of Science:

• Materials Science
• Civil Engineering
• Computer Science

Background:

• Multicomponent concrete performance evaluation relies on subjective expert judgment and lengthy monitoring.
• Artificial intelligence (AI) and machine learning (ML) offer advanced data analysis capabilities for building science.
• Addressing uncertainties from human factors is crucial for reliable concrete performance assessment.

Purpose of the Study:

• To investigate key factors influencing concrete compressive strength using diverse machine learning techniques.
• To enhance the interpretability of ML models in concrete science through SHAP analysis.
• To provide a data-driven foundation for optimizing concrete performance and sustainability.

Main Methods:

• Application of various machine learning algorithms: linear regression, polynomial regression, Decision Tree, Random Forest, ExtraTrees, AdaBoost, CatBoost, XGBoost, and TabPFN.
• Utilizing SHAP (SHapley Additive exPlanations) analysis for feature importance and interaction uncovering.
• Data-driven investigation into the influencing factors of multicomponent concrete compressive strength.

Main Results:

• Identification of critical factors affecting concrete compressive strength through ML model analysis.
• Enhanced understanding of feature importance and interactions within multicomponent concrete systems via SHAP.
• Demonstration of ML's capability to provide predictive insights into concrete performance.

Conclusions:

• Machine learning provides a robust, data-driven methodology for evaluating concrete compressive strength.
• SHAP analysis offers valuable interpretability, revealing underlying mechanisms in concrete material science.
• The findings support improved engineering design, construction decision-making, and the development of sustainable concrete materials.