Hybrid intelligence approach for performance estimation of rectangular CFST columns under different loading conditions

Abstract

The ultimate bearing capacity (UBC) is a crucial design parameter of compressed rectangular concrete-filled steel tube (RCFST) members. Past extensive experimental studies have demonstrated that the UBC of RCFST has a highly nonlinear relationship with its constitutes due to the sophisticated interaction mechanism between steel tube and concrete core. 1) Although massive experimental data are available in literatures, conventional methods still struggle in accurately modeling and predicting the UBC of RCFST based on existing dataset. 2) RCFST columns with various loading conditions are treated differently in an analytical method, which does not satisfy the consistency of modern design. In this study, a hybrid data-driven model is developed to predict the UBC of both axially and eccentrically loaded RCFST using multilayer perceptron (MLP) neural network integrated with beetle antennae search (BAS) algorithm, abbreviated as BAS-MLP. MLP is developed using existing experimental data, while BAS is introduced to optimize the initial connection weights and biases of MLP to further enhance its nonlinear mapping capacity. In the proposed BAS-MLP model, input variables include geometrical and material properties of RCFST columns with different loads, and the model output is UBC. The proposed model is evaluated using additional testing datasets, and the prediction accuracy of BAS-MLP is verified by comparing obtained results with those from baseline models using conventional modeling methods. A global sensitivity analysis is further conducted to measure the contribution of each input variable affecting the UBC of RCFST. The results show that the adaptive BAS-MLP model performs better than several baseline models and conventional methods (analytical methods and numerical simulation), and geometrical properties of RCFST columns are highly correlated with UBC. As a result, it can be regarded as an effective tool to predict the mechanical behavior of RCFST columns and other composite members, thereby reducing the cost and time in material planning and design.