Unified Approach for Estimating Axial-Load Capacity of Concrete-Filled Double-Skin Steel Tubular Columns of Multiple Shapes Using Nonlinear FE Models and Artificial Neural Networks

Concrete-filled double-skin steel tubular (CFDST) columns are a modern generation of composite columns optimized for a high strength-to-weight ratio with concrete sandwiched in the annulus between the inner and outer steel-tube skins. The cross-sectional shape of steel tubes can influence the confined concrete behavior and thus the axial load-bearing capacity of CFDST columns. This study proposed a unified approach using artificial neural networks (ANNs) to predict the refined axial load-bearing capacity of CFDST columns of multiple shapes based on combinations of circular and square steel tubes, i.e., circle-circle, circle-square, square-square, and square-circle forms. A total of 233 CFDST columns (82 tested specimens and 151 hypothetical columns) were used to formulate a detailed case study. The hypothetical columns were generated using calibrated nonlinear finite element modeling. Then, ANNs were trained using a subset of these hypothetical columns to map their geometric and material properties with the ultimate axial-load capacity. The 133 CFDST columns (i.e., 82 documented tests and the remaining 51 hypothetical columns) were finally used to check the trained ANN's prediction accuracy and generalization ability. The analytical formulation based on the optimized ANN architecture showed similar axial-load capacity prediction accuracy (about 8% mean absolute error) across all cross-sectional shapes of the CFDST columns in the testing set. Lastly, for the practical utility of the ANN-based model, prediction adjustment factors were proposed in this study for making conservative axial load-bearing capacity estimations within a targeted error margin.