Interpretable Artificial Intelligence Framework for Predicting Concrete Creep Compliance Based on CatBoost and SHAP

ABSTRACT Creep compliance of concrete is an important parameter to evaluate the long‐term deformation of concrete structures under sustained load. Based on the Northwestern University concrete creep and shrinkage database, this paper establishes a prediction and interpretation framework based on CatBoost. After data filtering, feature engineering and logarithmic transformation of target variables, four regression models of CatBoost, XGBoost, LightGBM and CNN‐LSTM were trained by using the same input features and the same training set and test set. The performance of the model was evaluated by R 2 , RMSE and MAPE, and SHapley Additive exPlanations (SHAP) analysis was used to explain the predictive behavior of CatBoost. The test results show that CatBoost has the best overall prediction performance, with R 2 of 0.912, RMSE of 7.335, and MAPE of 17.6%. The measured and predicted values of CatBoost are mainly distributed near the ideal fitting line, indicating that it has more stable prediction ability than XGBoost, LightGBM and CNN‐LSTM. The prediction results of representative specimens further show that CatBoost can reproduce the development process of creep compliance with loading duration. SHAP analysis shows that 28‐day compressive strength, water‐binder ratio and loading duration are the three most important variables affecting the prediction results of the model. Higher compressive strength generally decreases the predicted creep compliance, while higher water‐binder ratio and longer loading duration increase the predicted value. The results demonstrate that the proposed framework can provide both accurate prediction and interpretable analysis for concrete creep compliance.