Prediction of the local buckling strength and load‐displacement behaviour of SHS and RHS members using Deep Neural Networks (DNN) – Introduction to the Deep Neural Network Direct Stiffness Method (DNN‐DSM)

Abstract

The traditional separation of analysis and verification during the structural design of steel structures is a known source of conservatism and inaccuracy, as the true deformation/rotation capacity of sections and the redistribution of internal forces in systems remains only vaguely known in many cases. This particularly affects structures made of high‐strength steel, since often sections would need to be classified as slender, thus disallowing the possibility to account for plasticity and stress redistribution. Shell‐element FEM‐models with material nonlinearities and imperfections would be suitable to overcome this separation and increase the accuracy and economy of designs, yet are computationally intensive and impractical for design of whole structures. In this paper, a novel approach for carrying out a computationally economical beam‐element analysis that accounts for the nonlinear load‐displacement behaviour of sections of various local slenderness is presented: the ”DNN‐DSM“, which makes use of machine learning techniques (deep neural networks – DNN) to predict the nonlinear stiffness matrix terms in a beam‐element formulation for implementation in the Direct Stiffness Method. Based on trained DNN models from an extensive pool of nonlinear (GMNIA) shell element results. The motivation, general features, and first implementations of this method in the sense of a ”proof‐of‐concept“, for the case of hollow‐section truss members, are presented in the paper, as well as an outlook on the method's on‐going, full implementation.