Evaluation of machine learning techniques for capacity prediction of cold-formed steel beams subjected to bending

Abstract

Stiffened Cold-Formed Steel (CFS) sections often exhibit intricate nonlinear behaviors attributable to factors such as flexure effects and excessive slenderness. Traditional design methodologies, including the direct stiffness method, may inadequately capture these subtleties, potentially resulting in conservative or suboptimal designs. This study aimed to evaluate the performance of various machine learning algorithms, including simple and ensemble models, to predict the bending capacity of stiffened and unstiffened cold-formed beams in pure bending. A parametric study was conducted based on verified finite element analysis, and the machine learning algorithms were utilized to develop a unified capacity prediction method. The performance of six classical machine learning algorithms and four ensemble models were compared. The findings demonstrate that ensemble models, including AdaBoost, Gradient Boosting, Random Forest, and Extra Trees, outperform simple machine learning models in predicting the bending capacity of CFS beams. Moreover, introducing the stacking ensemble technique, using six different base models selectively, resulted in better performance than the individual baseline models. The approach addressed the nonlinearity pattern in the dataset caused by the flexure effect and excessive slenderness. The study suggests that adopting the proposed numerical and machine learning techniques could be a reliable method for predicting the structural behaviour and conducting cost-effective design of CFS beams, compared to the traditional analytical methods.