Prediction of load-bearing capacity of sigma section CFS beam-column members using ensemble and deep learning algorithms

Abstract

In this study, the bearing capacity of CFS sigma-section (S) beam-column members are predicted using three different ensemble learning (EL) algorithms and three deep learning (DL) algorithms. A validated detailed finite element model was created, and a comprehensive dataset was generated by performing 2552 finite element analyzes (FEAs) with different cross-sectional properties and loading conditions. The dataset was divided into train (80 %) and test (20 %) sets and hyperparameter optimization was performed on the train set using the GridSearchCV method. Then, analyses were performed on train and test sets. In addition, SHAP, ICE and PDP analyses were performed to determine the effect of input characteristics. The most accurate model for predicting the load-bearing capacity of the members is Extreme Gradient Boosting Regressor (XGR) with an R² value of 0.995. According to SHAP analysis identified major eccentricity and local buckling critical load factor as key influencing parameters. As a result of the study, the load-bearing capacity of S-section CFS members under axial force and uniaxial\biaxial bending were predicted with high accuracy. Finally, using the computational speed of the XGR model, the load bearing capacity of S-section CFS members under axial force and uniaxial\biaxial bending was estimated in less than one second with the help of a designed graphical user interface.