Web-post buckling resistance prediction models of stainless-steel cellular beams using machine learning algorithms

Abstract

Cellular steel beams are increasingly used in construction projects due to their decorative and economical characteristics, enabling long spans that reduce the number of columns and footings in the structure, thus shortening construction times and reducing infrastructure costs. The objective of this paper is to predict the ultimate strength, and to realize an accurate design method for determining the shear buckling of the web of stainless-steel cellular beams using machine learning. Several machine learning algorithms are trained using dataset generated from validated finite element models (FEM), including artificial neural networks, decision trees and random forests. All these models performed remarkably well, with coefficients of determination R2 greater than 0.9. The artificial neural network stood out for its superior predictive capacity, offering the best results. In particular, the ANN model with 8 neurons produced very accurate predictions for estimating the web-post buckling strength. In conclusion, a formula based on artificial neural networks (ANN) was presented and proved to be highly accurate, with a regression value (R2) equal to 0.99823, and mean absolute error (MAE), root mean square error (RMSE) values equal to 11.19 and 17.29 respectively. The formula based on artificial neural networks can therefore be used as a design tool.