Predicting the compressive strength of high-performance concrete via the DR-CatBoost model

To predict the compressive strength of high-performance concrete (HPC-CS), machine learning (ML) technology has been widely applied. The traditional categorical boosting (CatBoost) model is robust against overfitting and predicting HPC-CS. However, the model hyperparameter values need to be optimized and verified, and the prediction accuracy needs to be improved. Thus, on the basis of the CatBoost model with derived features and a random search (DR) method, this study introduces a DR-CatBoost model for accurate HPC-CS prediction and employs the SHapley Additive exPlanations (SHAP) model to perform interpretability analysis on the DR-CatBoost model. These findings indicate that, compared with those of the CatBoost model, the DR-CatBoost model achieves superior accuracy and more stable performance in predicting HPC-CS. Compared with those of the traditional CatBoost model, the RMSE, MAPE, and MAE of the DR-CatBoost model decrease by 9.0 %, 1 %, and 9.3 %, respectively, and R 2 of the DR-CatBoost model increases by 0.01. In the DR-CatBoost prediction model, age, the water/cement ratio, and water are features that have considerable effects on HPC-CS prediction. The model effectively reveals the nonlinear relationship of the characteristic parameters, which can provide a basis for the engineering design of concrete.