Intelligent design of limit states for recycled aggregate concrete filled steel tubular columns

Abstract

The defects of RAC and confinement effect have a significant influence on the performance of RACFST. But traditional model based on linear regression are insufficient to evaluate this multi-factor coupling mechanism. This study proposed an intelligent prediction procedure for the limit state of RACFST (ultimate bearing capacity, peak strain and stress–strain model) based on machine learning. A database containing 176 circular and 126 rectangular RACFST samples was employed to assess the applicability of existing specification, and was then used to train 8 ML models. The ML predictive models showed higher accuracy than 6 existing equations in the specifications. The XGBoost model had the highest prediction accuracy (predicted/test values of ultimate bearing capacity was 1.0062/0.9981 for rectangular/circular, respectively). The underlying mechanisms of the design parameters on the ultimate bearing capacity of RACFST were revealed employing the SHapley Additive exPlanations. The design method of RACFST ultimate bearing capacity applicable to engineering design was derived by analyzing the key parameters, and the prediction results obtained from XGB model were used to improve the stress–strain model. The results show that the design procedure proposed in this study combines the ML models with the traditional physical model can effectively guide the intelligent design of RACFST.