Prediction of the load-displacement and local buckling behavior of hollow structural sections using Deep Neural Networks (DNN)

The increased use of high-strength steel (HSS) in structures has recently led to the need for a closer look at shortcomings in current code provisions and to the development of new design methodologies. This is mainly due to the traditional separation of analysis and verification, whereby both are dependent on the cross-sectional slenderness and corresponding classification into categories, affecting HSS hollow sec-tions in particular. Therefore, advanced finite element GMNIA (geometrically material nonlinear analysis with imperfections) simulations provide often an alternative. Nevertheless, these methods are computa-tionally time intensive and not generally suitable for use in a design. For this reason, various efforts are being undertaken to combine the computational efficiency of beam-element models with the ability to account for slenderness-dependent deformation capacities and nonlinear redistribution of internal forces in a structural truss or frame, leading to the development of advanced inelastic methods e.g. CSM and GBT.

The proposed paper, presents a novel approach to carry out a beam-element analysis that accounts for the nonlinear load-displacement behavior of sections of various local slenderness: the DNN-DSM, which makes use of machine learning techniques (deep-neural-networks - DNN) to predict the nonlinear stiff-ness matrix terms in a beam-element formulation for implementation in the Direct-Stiffness-Method (DSM). Based on DNN-models trained on an extensive pool of nonlinear (GMNIA) shell element results, the first implementations of this method are able to accurately predict the nonlinear load-displacement behavior of various sections with great accuracy. The paper will explain these first findings and give an outlook on the planned future work.

Learning Objectives:
Understanding of Deep-Neural-Networks.

View content