An interactive ensemble learning approach for predicting ultimate axial load capacity in concrete-filled steel tube column

Determining the ultimate axial compressive load-carrying capacity (UACLC) of square/rectangular concrete-filled steel tube (S/RCFST) short columns is crucial for maintaining structural integrity in civil engineering projects. This investigation presents an interactive ensemble learning approach utilizing five machine learning algorithms: Decision Trees (DT), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Boosted Regression Trees (BRT), and Categorical Gradient Boosting (CatBoost). The study employed a comprehensive dataset comprising 932 experimental samples to train and validate 491 models, with each model undergoing hyperparameter optimization. Among these, the CatBoost model exhibited superior performance, achieving a training R2 of 0.999 and a testing R2 of 0.984 when configured with optimal hyperparameters (learning rate = 0.1, maximum depth = 5, and 2000 estimators). The model's accuracy was further demonstrated by a testing weighted mean absolute percentage error (WMAPE) of 0.071 and a root mean square error (RMSE) of 0.316 MPa. Monotonicity analysis confirmed the consistency of the model's predictions, revealing that material properties such as concrete compressive strength and steel yield stress had the most significant impact on UACLC predictions. To facilitate practical application, the researchers developed a graphical user interface (GUI) enabling real-time predictions, which allows engineers to integrate the model into their structural design processes. This innovative framework provides a robust and highly precise tool for predicting UACLC, addressing significant limitations in conventional methods and enhancing the practical design of S/RCFST columns.