Ensemble Learning for Predicting Concrete Compressive Strength

Machine learning techniques, particularly ensemble learning, are transforming the construction industry by significantly enhancing the accuracy of concrete compressive strength predictions. This study explores various ensemble learning methods to predict compressive strength using two distinct datasets: a real-world dataset from Advanced Construction Technology Services (ACTS) and a laboratory dataset from the UCI Machine Learning repository. We evaluated multiple ensemble algorithms, including Random Forest, Histogram-based Gradient Boosting, LightGBM, XGBoost, CatBoost, and Stacking, whose performance was assessed using mean squared error (MSE), root mean square error (RMSE), and the coefficient of determination ($R^2$). The results revealed that for the field dataset, Random Forest, CatBoost, and Stacking achieved the best predictive performance, while for the laboratory dataset, Stacking and CatBoost were the top performing models. To further enhance predictive accuracy, one of the best performing algorithms was selected for each dataset and optimized using random search and nested cross-validation to ensure robust model validation. Additionally, SHapley Additive Explanations (SHAP) were used to interpret the importance of features and analyze the influence of different input variables on the predictions of compressive strength. Our findings indicate that models trained on laboratory data generally outperformed those trained on field data. The optimized models achieved the best overall performance, reinforcing the value of model optimization in ensemble learning and offering practical insights to improve the reliability of concrete compressive strength predictions in the construction industry.