Explainable stacking-based hybrid machine learning for predicting uni-axial creep deformation in concrete

Abstract

To address the complexity of modeling concrete creep behavior and the limitations of traditional models, this study proposes a data-driven hybrid machine learning model for accurate prediction of creep deformation. The Northwestern University creep database is preprocessed to identify the most influential factors, and a stacking-based hybrid model is developed by combining five ensemble tree-based algorithms with an artificial neural network. Bayesian optimization, implemented via the Hyperopt library, is employed for hyperparameter tuning, ensuring optimal model performance. A 10-fold cross-validation is conducted to demonstrate the model's strong generalization capability. The hybrid model outperforms standalone base estimators, achieving a coefficient of determination (R²) of 0.960 on the testing set. SHapley Additive exPlanations are used to interpret the model's predictions globally and locally, revealing factor importance consistent with experimental findings. A comparison with three widely used traditional models, the Comité Européen du Béton (CEB) Model Code 90–99, Fédération Internationale du Béton (fib) Model Code 2010, and the B4 model on selected testing subsets demonstrates the superiority of the proposed model across six evaluation metrics. The prediction of various creep strains closely aligns with experimentally measured values, further validating the model's accuracy and effectiveness in predicting different types of creep deformations.