An efficient procedure for prediction of the load-displacement curve of CFDST columns

Abstract

This paper proposes a novel procedure for prediction of both load-displacement curve and load-carrying capacity of concrete-filled double-skin steel tube (CFDST) columns under uniaxial compression by using convolutional neural network (CNN)-based regression and Nelder-Mead methods. Firstly, hybrid databases collected from experiments in literature and generated from finite element analyses are employed to build the proposed CNN-based model. The accuracy of the proposed model is described through a comparison between predictive results of the proposed model and unseen data. Two machine learning models, including eXtreme Gradient Boosting and Multilayer Perceptron, are adopted for comparison. It can be observed that the CNN-based model provides the most accurate predictions for both the load-displacement curve and axial compression capacity of CFDST columns in both experimental and numerical databases. An efficient procedure is developed to calibrate the preliminary load-displacement curve estimated by the CNN-based model, and to notably enhance its smoothness and performance. Adjusted formulae (based on well-known equations) are obtained for predicting the load-displacement curve of CFDST columns. The hyperparameters of these formulae are optimized using the Nelder-Mead method. It is indicated that the adjusted load-displacement curves obtained from the proposed procedure outperform the preliminary curves estimated by the CNN-based model. A sensitivity analysis was conducted to investigate the model's performance in predicting the load-displacement curves of CFDST columns with variations of input variables within stochastic environments. Finally, a cloud-based graphical user interface is developed to provide a convenient tool for users to predict axial load-displacement responses of CFDST columns without prior programming knowledge.