Interpretable constitutive compressive stress-strain model for rubberized aggregate concrete – Integrating comprehensive empirical database and efficient XGBoost ensemble learning

Abstract

This study addresses the critical challenge of estimating compressive stress in rubberized recycled aggregate concrete, a key concern for sustainable construction. By integrating empirical data and machine learning techniques, particularly Extra Gradient Boosting (XGBoost), this research aims to enhance predictive models for structural performance assurance. Utilizing a dataset comprising 3840 stress-strain responses collected from 60 rubberized concrete mixtures, with 12 input variables, the study employs statistical analyses and regression to validate the model's performance. The XGBoost model demonstrates robust predictive capability, with low mean absolute error values (1.551 for testing and 0.967 for training) and high a20−index (0.917 for testing and 0.933 for training), effectively predicting compressive stress. Key factors influencing compressive stress were identified through Shapley's additive explanations. The developed model offers a practical tool for predicting stress-strain behavior in rubberized concrete formulations, with implications for mix design optimization and testing cost reduction. Additionally, the introduction of a user-friendly GUI facilitates interaction with the developed XGBoost model, further enhancing its practical application in sustainable construction practices. This study also includes the comprehensive database utilized in modeling, along with the developed Python code, provided as supplementary files. Future directions include exploring the influence of different types and sizes of rubber aggregates on concrete, particularly in high-strength and ultra-high-strength mixtures, and investigating the impact of outliers on model prediction performance.

Keywords

• Constitutive modeling
• Machine learning
• Recycled aggregate
• Rubberized concrete
• Sustainability