The lack of reliable code-based predictive models for the residual properties of recycled aggregate concrete (RAC) at elevated temperatures hinders its use in fire-prone applications. Therefore, this study developed and compared 8 machine learning (ML) models using comprehensive parameters to estimate the residual compressive and tensile strength of RAC under elevated temperatures. The developed ML models included linear regression, K-nearest neighbour, support vector regression, classification and regression tree, random forest, adaptive boosting, gradient boosting, and extreme gradient boosting (XGB). A total of 156 experimental datasets for residual compressive strength and 132 for residual tensile strength were collected from the literature. The compiled datasets included 15 input features related to concrete ingredients, aggregate properties, heating mechanisms, and the age at which concrete was heated. The results demonstrated that the XGB model provided the most accurate predictions with the least statistical errors. It achieved a coefficient of determination of 98.34% for residual compressive strength and 97.35% for residual tensile strength. The root mean square errors were 0.032 and 0.043 on the test datasets for residual compressive and tensile strength, respectively. Additionally, the XGB models recorded the lowest residual errors for both responses compared to all ML models. In addition, the SHapley Additive exPlanation (SHAP) framework demonstrated that the temperature had the highest impact on the residual compressive and tensile strength of RAC compared to the concrete ingredients. The XGB models also outperformed the existing analytical models in predicting the residual compressive and tensile strength. Furthermore, the established XGB models for the residual compressive and tensile strength of RAC were deployed as a single graphical user interface (GUI) tool, accessible to users at the same web address. The developed GUI showed accurate predictions, with actual-to-predicted ratios between 1.03 and 1.05 for both responses.
Abushanab et al. (Sun,) studied this question.