Prediction of concrete compressive performance based on the ensemble learning method

A total of 1,030 experimental datasets on concrete compressive strength were collected. Using concrete mixture components as input features and compressive strength as the output, four ensemble learning models (RF, AdaBoost, GBRT, and XGBoost) were established. Grid search and k-fold cross-validation (CV) were employed to identify optimal hyperparameters for each model. Data augmentation was further applied to enhance model accuracy. The rationality and precision of the models were evaluated by comparing their coefficients of determination (R 2 ). Comparative analysis with traditional machine learning models revealed that the XGBoost ensemble model exhibited the best predictive performance. Building on these results, the SHAP (SHapley Additive exPlanations) method was applied to interpret the models, quantifying the influence of individual factors on concrete compressive strength. Furthermore, the interaction effects of multiple variables on concrete strength were systematically characterized.