Machine learning-based strength prediction for circular concrete-filled double-skin steel tubular columns under axial compression

Circular concrete-filled double skin steel tubular (CFDST) columns show great potential in various infrastructures due to their excellent structural performance and easy construction process. As the rapid development of infrastructure towards large scales and manifold functions, CFDST columns within an extended scope of dimensions and sectional configurations will be applied in engineering practice. This trend poses challenges to both the accuracy and efficiency of traditional strength prediction methods including design equations, experimental fitting and numerical simulations. Based on experimental datasets obtained from existing research, machine learning (ML) methods pave a path for understanding the complex relations between key parameters and the axial compressive ultimate strength of CFDST columns. This paper attempts to develop reliable ML-based models for predicting the axial compressive ultimate strength of CFDST columns. Three advanced ML algorithms, namely Back Propagation Neural Network (BPNN), Support Vector Regression (SVR) and Gaussian Process Regression (GPR), were adopted to develop prediction models by training an experimental database of 162 CFDST specimens from existing literature which covers a wide scope of parameters including high diameter-to-thickness ratios and large hollow ratios. Particularly, Particle Swarm Optimization (PSO) method was used to optimize the hyperparameters of ML algorithms for enhancing the prediction accuracy. The performances of the developed prediction models were evaluated through an in-depth comparison analysis. It is found that the ML-based models could predict the strengths of CFDST columns with higher accuracies and wider applicable ranges than existing design methods; PSO-enhanced GPR (PSO-GPR) model achieved the most improved performance for strength prediction. Further validation on PSO-GPR model was conducted based on experimental data from recently tested large-dimension CFDST specimens and the results demonstrate the wide applicability of the developed model.