Compressive strength prediction of recycled aggregate concrete based on DE-XGBoost algorithm model and explainable machine learning

Abstract In the current concrete performance prediction research, common machine learning algorithms include random forest (RF), gradient boosting decision tree (GBDT), LightGBM (Light Gradient Boosting Machine), extreme gradient boosting tree (XGBoost), etc. Based on the compressive strength data of recycled aggregate concrete at different ages under different mixing ratios, this paper establishes a combined model (DE-XGBoost) of differential evolution algorithm (DE) and extreme gradient boosting tree algorithm (XGBoost), using XGBoost, Light GBM, GBDT, RF models as comparisons, and R², RMSE, MSE and MAE as model evaluation indexes. Use interpretability machine learning algorithms (SHAPs) to explore the feature importance of model input features to output results. The results show that: (1) In a single model, the prediction accuracy of XGBoost model (R²=0.9599) is better than that of Light GBM (R²=0.9493), GBDT (R²=0.9459), RF (R²=0.9321) model; (2) The prediction accuracy of the DE-XGBoost combination model is improved by 1.12% compared with the XGBoost model, and the RMSE, MSE and MAE values are reduced by 16.09%, 29.60% and 27.77%, respectively.