Neural network modelling of mechanical joints for the application in large-scale crash analyses

Abstract

This paper presents an artificial neural network (NN) modelling approach to represent a connector model in large-scale finite element explicit crash simulations. The NN model was established to describe the local force–deformation response of point connectors in automotive applications, namely self-piercing rivets and flow-drill-screws. The study is limited to two-sheet connections and to the use of feedforward NNs. Successive loading and unloading of the joints is not studied. Various architectures and complexities of the feedforward NNs were evaluated and trained based on data generated from a constraint model found in the literature. This forms a proof of concept for implementing a modelling technique not based on physics-motivated constitutive equations. The impact of the network complexity and training data diversity was investigated. The NN model was implemented as a cohesive zone model for incremental force prediction in an explicit finite element code. In order to have a wide selection of joint types, five different joint configurations including self-piercing rivets (SPR) and flow-drill screws (FDS) were investigated. Numerical results from the NN model were compared to physical tests from all joint configurations. It was shown that a rather basic machine learning technique like a feedforward NN was able to reproduce path-dependent force–deformation behaviour for the application in explicit FE solvers.