Explainable machine learning and ensemble models for predicting fresh properties of self consolidating concrete

Area of Science:

• Civil Engineering
• Materials Science
• Computational Science

Background:

• Accurate prediction of fresh self-consolidating concrete (SCC) properties is vital for construction efficiency and performance.
• Traditional experimental methods for testing SCC are time-consuming, expensive, and prone to errors.
• Developing reliable predictive models can streamline concrete mix design and application.

Purpose of the Study:

• To develop and compare machine learning (ML) models for predicting SCC fresh properties (slump flow and V-funnel time).
• To assess the predictive accuracy of Gene Expression Programming (GEP), Deep Neural Networks (DNN), Decision Trees (DT), Support Vector Machines (SVM), and Random Forests (RF).
• To enhance model interpretability using SHAP and PDP to understand mix design variable influences.

Main Methods:

• A comprehensive dataset of 348 SCC mix designs was curated from 176 studies.
• Five ML models (GEP, DNN, DT, SVM, RF) were trained and tested (85% training, 15% testing).
• Shapley Additive explanations (SHAP) and Partial Dependence Plots (PDP) were used for model interpretability.

Main Results:

• GEP and DNN models demonstrated the highest predictive accuracy.
• GEP achieved R² values up to 0.957 for V-funnel time and 0.915 for slump flow.
• DNN achieved R² values up to 0.950 for V-funnel time and 0.911 for slump flow.

Conclusions:

• Advanced ML models, particularly GEP and DNN, can reliably predict SCC fresh properties.
• These ML approaches offer a viable alternative to extensive laboratory testing, reducing time and cost.
• The findings support data-driven optimization of SCC mix designs in contemporary construction.