Predicting ultimate compressive strength of CFST columns under axial compression using an interpretable optimized RF-Transformer model

Concrete-filled steel tube (CFST) columns exhibit complex interactions between the steel tube and concrete during the loading process, making the accurate prediction of their axial compressive capacity particularly challenging. To improve the accuracy and applicability of predicting the axial compressive strength for circular CFST columns, this study proposes a novel interpretable optimized ensemble learning model. Initially, a comprehensive database is constructed using experimental test data of CFST columns, including feature parameters such as geometric parameters and material properties, followed by Pearson correlation analysis to quantify the relationships among the feature variables. Subsequently, an optimized ensemble learning model integrating random forest (RF) and Transformer architectures is developed for predicting the axial compressive strength of circular CFST columns, in which the Tianji’s horse racing optimization algorithm is employed to automatically tune the model hyperparameters. Finally, the SHAP method is utilized to offer interpretations for both global and local perspectives of the proposed model. Various evaluation indicators are employed to compare the predictive performance of the proposed ensemble learning model with that of empirical methods and other machine learning models. The results demonstrate that the proposed ensemble learning model outperforms both empirical formulas and competing machine learning models. In addition, the findings reveal that ensemble strategies do not necessarily guarantee better performance than individual models, and model performance is highly sensitive to hyperparameter configurations. Notably, the column diameter, steel tube thickness, and concrete compressive strength exert significant impacts on the axial compressive strength of circular CFST columns.