Abstract

The construction industry’s shift towards sustainable practices has spurred interest in innovative materials, with Recycled Aggregate Concrete (RAC) standing out as a notable candidate. This material leverages recycled aggregates to mitigate waste, conserve resources, and reduce environmental impact. However, the accurate prediction of RAC’s compressive strength (CS) is challenging due to its intricate composition and variable material properties. To address this, artificial intelligence (AI) models are increasingly being used for their ability to uncover complex data patterns. This study offers a detailed comparison of ten advanced AI models for predicting RAC CS, including Artificial Neural Networks, Support Vector Regression, Decision Tree Regression, Random Forest Regression, k-Nearest Neighbors, Lasso, AdaBoost, Bagging, XGBoost, and CatBoost models. Each model is fine-tuned through hyperparameter optimization to enhance predictive accuracy. Additionally, SHAP (SHapley Additive exPlanations) algorithms are employed to interpret the models, providing insights into feature importance. The results demonstrate that all models achieved R² values exceeding 75%, with the CatBoost model attaining the highest R² value of 91% on the testing set. The CatBoost model also recorded the lowest error indices, with an MAE of 2.79 and an RMSE of 4.045, making it the most effective model for predicting RAC strength. SHAP analysis identified cement, water, sand, and RA water absorption as key features influencing RAC strength. This study underscores the potential of AI models in advancing the predictability and performance of sustainable construction materials.