Optimized data-driven machine learning models for axial strength prediction of rectangular CFST columns

Abstract

The nonlinear material interaction in concrete-filled steel tube (CFST) significantly contributes to its excellent axial compression behaviour, which in the meantime results in highly complex relationships between various specimen parameters and the column strength. Machine learning (ML) technique, being excellent in capturing complicated data mapping, is therefore applied in this study to predict the axial compression strength of rectangular CFST columns. A comprehensive test database containing 1,641 rectangular CFST samples is established. The key input parameters for ML models are identified through both correlation analysis and mechanical principles, highlighting the effects of sectional configuration and column slenderness. Strength prediction models are established applying five mainstream ML methods including the back-propagation neural network, the radial basis function neural network, the adaptive neuro-fuzzy inference system, the Gaussian process regression and the M5′ model tree. The prediction reasonableness and stability of established models are comprehensively evaluated and compared. Outcomes reveal that the established ML models exhibit higher prediction accuracies and wider applicable ranges than current design standards, whilst the prediction stability of ML models is highly affected by the quantity of the available data set. Finally, ML models are optimized by dividing the training database as per mechanical principles, and a prediction-error-ratio amending method based on ML algorithm is also proposed to further improve the model accuracy.