Predicting compressive strength of CCFST columns with core concrete defects: An interpretable and unified machine learning approach

Kaizhong Xie, Dong Liang, Quanguo Wang, Yi Zhou, Xianyan Luo,
Predicting compressive strength of CCFST columns with core concrete defects: An interpretable and unified machine learning approach,
Structures,
Volume 81,
2025,
110375,
ISSN 2352-0124,
https://doi.org/10.1016/j.istruc.2025.110375.
(https://www.sciencedirect.com/science/article/pii/S2352012425021903)
Abstract: This study developed an interpretable, unified machine learning model to predict the compressive strength of circular concrete-filled steel tube (CCFST) columns, incorporating experimental uncertainties and defect types. Firstly, a database of 309 experimental tests on CCFST columns with differing defect types was established using the developed database screening and quantification criteria. The statistical properties and variability of features within this database was subsequently analyzed. Secondly, a physical gating mechanism for the unified model was established based on defect parameters. Monte Carlo techniques were employed to split the data into training and testing sets. The prediction performance and robustness of the unified models based on five machine learning models (SVR, RF, ANN, XGBoost, and GPR) were then evaluated. The prediction results were compared with empirical formulas from the literature. Finally, SHAP was utilized to reveal the independent contributions and coupling effects of each feature on the CCFST column bearing capacity. The results demonstrate that the proposed models achieve high predictive accuracy, with RMSE values below 0.4 and R2 ranging from 0.952 to 0.997, significantly surpassing those obtained from formulas. Notably, the ANN and GPR models demonstrated exceptional performance and robustness. Diameter and thickness significantly influenced compressive strength, with influence scores of over 0.4 and 0.25, respectively. Interaction effects between large defects and the arc length substantially reduced the compressive strength of the structure by roughly 35 % and 18 %, thus highlighting the necessity of limiting defect distribution to enhance and refine traditional design practices.
Keywords: Circular concrete-filled steel tube (CCFST); Core concrete defects; Machine learning; Explainable artificial intelligence; Compressive strength