Structural mechanism-based intelligent capacity prediction methods for concrete-encased CFST columns

Abstract

Intelligent capacity prediction using machine learning (ML) approach has been gradually employed in structural design. Due to the presence of complicated interactions and multiple factors that strongly influence the capacity of concrete-encased concrete-filled steel tube (CFST), implementing ML methods for such members is expected to be more challenging and requires great care. However, most current ML approaches rely heavily on data laws and ignore the underlying physical mechanisms. The physical mechanisms of proposed models may thus fail due to over-fitting, especially when the training data is not broad enough to reflect the actual laws. In this study, an innovative framework is proposed for combining the data-driven models with physical mechanisms to estimate the axial compression capacity of concrete-encased CFST, in which the mechanism verification is integrated into the optimization of model hyperparameters. A refined finite element analysis (FEA) model is established and validated, based on which a comprehensive database consisting of 143 experimental and 1560 simulated samples is constructed. Six single ML algorithms and two ensemble methods are adopted and compared, with their underlying mechanisms further revealed by the shapley additive explanation (SHAP) approach. The results reveal that all models developed through proposed modeling method demonstrate overall good performance and reflect structural mechanisms to a fair extent, where the XGBoost model outperforms others in terms of accuracy and dispersion. Besides, inputting combined parameters instead of basic ones is found to effectively solve the multi-factor problems using ML approaches.