Research on seismic performance prediction of CFST latticed column-composite box girder joint based on machine learning

This paper presents an experimental investigation on the joints between four-limb concrete-filled steel tubular (CFST) latticed column and composite box girder under low-cycle reciprocating loading. The load-displacement hysteresis curve and skeleton curve of the joint were obtained, as well as the failure mode of the joint. Based on the joint test, a finite element (FE) model was established and validated with the experimental results. The validated FE model was used to obtain the load-displacement skeleton curves for the 140 joint specimens under different parameter combinations, which facilitated the development of a machine learning based predictive model to evaluate the effects of various parameters on the seismic performance of the joints. Six parameters on the skeleton curve of the joint was studied by mRMR, which are the concrete strength, yield strength of transverse diaphragm plate, steel bar diameter, axial compression ratio, concrete slab thickness, and yield strength of steel box girder. Five machine learning (ML) algorithms were used to predict the joint's skeleton curve considering the six parameters. The skeleton curve, positive and reverse stiffness and ultimate load obtained from the predictive models and test results were compared. The results show that the numerical results match well with the experimental results, and thus the FE model can be used to develop a database. Among the six parameters, the axial compression ratio has the greatest influence on the skeleton curve, while the steel bar diameter has the least impact. Among the five machine learning (ML) algorithms, the XGBoost algorithm consistently achieved the lowest errors across three metrics, demonstrating a better performance in prediction. With a prediction accuracy approaching 98 %, it is shown to be suitable for predicting the skeleton curves.