Metaheuristic-optimized Machine Learning for Mechanical Property Prediction in Eco-friendly Rubberized Concrete

Abstract

The disposal of discarded tires presents a global environmental challenge, but recycling them into rubber particles for concrete offers a sustainable solution, reducing waste while creating eco-friendly construction materials. However, experimental studies in this field are resource-intensive, requiring significant time and financial investment. This research develops 99 machine learning (ML) models, optimized using three improved metaheuristic algorithms, to predict the mechanical properties of rubberized concrete. The eight ML models include Adaptive Boosting, Artificial Neural Networks, Decision Tree, Gradient Boosting, K-Nearest Neighbors, Random Forest, Support Vector Machines, and Extreme Gradient Boosting, refined through the Improved Hybrid Growth Optimizer, Improved Ray Optimization, and Improved Sand Cat Swarm Optimization algorithms. A dataset of 315 experimental cases, incorporating six input variables, rubber size, rubber weight, cement content, water content, coarse aggregate, and fine aggregate, was analyzed. In addressing missing data for certain mechanical properties, a two-level sequential artificial neural network was employed. The study revealed that the XGBoost-IRO hybrid model excelled in predicting compressive and tensile strength, while the AdaBoost-ISCSO ensemble was best for flexural strength, and the XGBoost-IHGO model performed optimally for modulus of elasticity. Partial Dependence Plots and SHAP analysis highlighted the complex relationships between input variables and mechanical properties, confirming the significance of all input features. Validation through Taylor diagrams and error distribution further confirmed the reliability of the models in predicting all mechanical properties.