Abstract

The construction industry is shifting towards sustainability, emphasizing the need for innovative materials. Recycled Aggregate Concrete (RAC), utilizing recycled aggregates, emerges as a promising eco-friendly solution to minimize waste and resource utilization. However, accurately predicting its compressive strength (CS) is challenging due to varying composition and properties. This study addresses this issue by employing machine learning models, specifically five gradient boosting algorithms: Gradient Boosting Machine (GBM), LightGBM, XGBoost, Categorical Gradient Boost (CGB), and HistGradientBoosting (HGB). A total of 314 mixes from relevant published literature were aggregated to train the models. These models are meticulously fine-tuned through hyperparameter optimization for optimal predictive performance. The study also introduces SHAP (SHapley Additive exPlanations) algorithms for model interpretability, elucidating feature contributions to predictions. The results revealed that among the five gradient boosting models, CGB demonstrated the highest R2 value of 92% on the testing set, while LightGBM exhibited the lowest Coefficient of Determination (R2) value of 88%. Additionally, CGB achieved the lowest Root Mean Square Error (RMSE) of approximately 4.05, whereas XGBoost showed the highest RMSE of around 4.8. Furthermore, for Mean Absolute Error (MAE), LightGBM recorded the lowest value of approximately 3.16, while HGB yielded the highest MAE of about 3.8. The SHAP analyses reveal influential features impacting RAC strength, highlighting the significance of cement, water, sand, and recycled aggregate water absorption in predicting RAC compressive strength.