Soft-computing models for predicting plastic viscosity and interface yield stress of fresh concrete

Area of Science:

• Civil Engineering
• Materials Science
• Computational Science

Background:

• Fresh concrete properties, specifically interface yield stress and plastic viscosity, are critical for its pumpability.
• On-site testing for these properties is time-consuming and resource-intensive, necessitating alternative methods.

Purpose of the Study:

• To develop accurate predictive models for concrete's plastic viscosity and interface yield stress.
• To compare the performance of various machine learning algorithms and a statistical technique for this prediction task.

Main Methods:

• Utilized Random Forest Regression (RFR), Gene Expression Programming (GEP), K-nearest Neighbor (KNN), Extreme Gradient Boosting (XGB), and Multi Linear Regression (MLR).
• Models were trained using published literature data with six input parameters (e.g., cement, water, time after mixing) and two output parameters.
• Performance was evaluated using error metrics, k-fold validation, and residual analysis. SHAP and ICE analyses were performed on the best model.

Main Results:

• Extreme Gradient Boosting (XGB) demonstrated the highest accuracy in predicting both plastic viscosity and interface yield stress.
• Gene Expression Programming (GEP) was the only algorithm that provided an empirical equation.
• Water, cement content, and time after mixing were identified as the most influential parameters for predicting fresh concrete properties using XGB.

Conclusions:

• Machine learning, particularly XGBoost, offers a highly accurate and efficient method for predicting fresh concrete properties, reducing the need for extensive on-site testing.
• A graphical user interface was developed to facilitate the practical application of these predictive models in the civil engineering industry.