Machine learning applications for studying the structural behaviour of cold-formed steel columns with web openings

Abstract

This study presents a data-driven machine learning (ML) approach to study the structural behaviour of cold-formed steel (CFS) columns with web openings. A total of 23 experimental results from literature were employed for validation purposes. It was shown that the validation ratio is reasonable as the average ratio of experimental to FEA strengths (FEXP/FFEA) is 1.07. Afterwards, a total of 15,000 data points for training selected ML algorithms are generated from elasto plastic finite element analysis (FEA), which incorporates both initial geometric imperfections and residual stresses. The input features of this study are the overall lip width of the section, web height, section thickness, length of channel section, hole and edge configuration (number, spacing, radius). A total of six machine learning (ML) algorithms, namely, XGBoost, Decision tree, Artificial Neural Network, Random Forest, Lasso Regression, and Linear Regression, are evaluated to examine the structural behaviour of CFS columns with web openings. 10-fold cross-validations are performed on selected ML algorithms. It was found that the proposed XGBoost model outperformed other previously described machine learning (ML) algorithms in comparison. The XGBoost algorithm produced the best accurate predictions (99%) with the shortest training time. In addition, the XGBoost model has the lowest mean root squared error and mean absolute error. An investigation of the importance of input factors found that lip width of the section and length of channel section were the most relevant features.