Evaluation of machine learning models for load-carrying capacity assessment of semi-rigid steel structures

Abstract

The paper investigates the potential application of machine learning methods to estimate the load-carrying capacity of semi-rigid connected steel structures. The database is developed using the advanced analysis based on beam-column and zero-length elements. The input variables are member cross-sections and parameters of the high-order nonlinear functions describing semi-rigid connection behaviors. Twelve machine learning algorithms, including three linear regression models, support vector machines, deep learning and seven tree-based ensemble algorithms, were evaluated. Three practical semi-rigid connected steel (including two-dimensional planar and three-dimensional space) structures were studied. The numerical results revealed that the distribution of the load-carrying capacity responses of the structures was non-normal and heavy-tailed. The class of linear regression methods was less efficient than the tree-based ensemble algorithms. Among the seven tree-based ensemble approaches, the extreme gradient boosting method presented the best performance in determining not only the lowest (mean squared, mean absolute and mean absolute percentage) errors but also the highest coefficient of determination. It was followed by the light gradient boosting machines and categorical gradient boosting algorithms. Three linear regression methods (linear regression, Lasso and Ridge), light gradient boosting machines, categorical gradient boosting and extreme gradient boosting provided the competitive computing cost. The extreme gradient boosting algorithm was therefore recommended for the accurate prediction of load-carrying capacities of semi-rigid connected steel frames.