Artificial Neural Network Models for Determining the Load-Bearing Capacity of Eccentrically Compressed Short Concrete-Filled Steel Tubular Columns

Abstract

Artificial neural networks (ANN) have a great promise in predicting the load-bearing capacity of building structures. The purpose of this work was to develop ANN models to determine the ultimate load of eccentrically compressed concrete-filled steel tubular (CFST) columns of circular cross-sections, which operated on the widest possible range of input parameters. Short columns were considered for which the amount of deflection does not affect the bending moment. A feedforward network was selected as the neural network type. The input parameters of the neural networks were the outer diameter of the columns, the thickness of the pipe wall, the yield strength of steel, the compressive strength of concrete and the relative eccentricity. Artificial neural networks were trained on synthetic data generated based on a theoretical model of the limit equilibrium of CFST columns. Two ANN models were created. When training the first model, the ultimate loads were determined at a given eccentricity of the axial force without taking into account additional random eccentricity. When training the second model, additional random eccentricity was taken into account. The total volume of the training dataset was 179,025 samples. Such a large training dataset size has never been used before. The training dataset covers a wide range of changes in the characteristics of the pipe metal and concrete of the core, pipe diameters and wall thicknesses, as well as eccentricities of the axial force. The trained models are characterized by high mean square error (MSE) scores. The correlation coefficients between the predicted and target values are very close to 1. The ANN models were tested on experimental data for 81 eccentrically compressed samples presented in five different works and 265 centrally compressed samples presented in twenty-six papers.