The following is an excerpt.
PhD candidate Victor André presented a novel approach to mechanical joint modelling at this year’s LS-DYNA conference.
Victor André´s PhD topic is modelling techniques for multilayered joints, and at LS-DYNA, he presented an artificial neural network (NN) modelling approach. Here, the NNs represent mechanical fasteners in large-scale finite element (FE) crash simulations. Victor André’s approach is that machine learning (ML) could increase the accuracy of joint models in FE analyses.
NEURAL NETWORKS FORMECHANICAL JOINTS
The German researcher earned his master’s degree at Technical University Dresden in 2018, in a joint thesis project including 6 months internship at Audi AG. The motivation behind the PhD project at CASA is the multitude of different materials, parts, and pieces found in modern car bodies or other complex structures. Various joining techniques for such mixed parts are a crucial design challenge. In some cases, welding is not applicable because the materials are too dissimilar. Then, other techniques like adhesive bonding or mechanical fastening can be considered. However, modelling the joining process, especially the characterization of mechanical behaviour like deformation, damage, and fracture, is a complex task.
READ ALSO: Simulations Getting Closer To an Actual Vehicle Crash
THE TRADE-OFF BETWEEN ACCURACY AND COMPUTATIONAL TIME
The model Victor André presented at LS-DYNA describes the local force-deformation response of self-piercing rivets and flow-drill-screws in automotive applications. Typically, the basis for large-scale predictions of a final product is a combination of physical testing and numerical methods. In the design process of structures, finite element analysis with large structural elements is used to increase computational efficiency. However, Victor André states that this always means a trade-off between accuracy and computational time.
INCREASED ACCURACY OF JOINT MODELS
According to the PhD candidate, NN-modeling could increase the accuracy of joint models in FE analyses.
Large amounts of data could be generated by detailed mesoscopic joint characterization and used to train a neural network. The NN then predicts the mechanical joint response throughout the large-scale analysis.
Rather than having specific mathematical equations for only one type of joint, this approach gives more flexibility for fitting different joints. Victor André states that this novel work forms a proof of concept for implementing a NN modelling technique not based on physics-motivated constitutive equations.
SIMULATING DEFORMATION BEHAVIOUR
More than 830 participants joined the LS-DYNA 2021, the main event relating to LS-DYNA in Europe. Also, this was the first-ever hybrid conference. In total, 600 of the participants joined online.
LS-DYNA is one of the world´s leading finite element software systems, capable of simulating the complex deformation behaviour of structures. It is a widely used tool in the automotive and aerospace industry and construction, civil engineering, military, manufacturing, and bioengineering. Nine of the ten largest car companies and 7 out of the worldwide largest automotive suppliers are on the company’s customer list.
HEAD IMPACT AND SPACE DEBRIS
In all, 4 PhD candidates and postdocs from CASA gave presentations in Ulm:
Postdoc Andria Antoniou: Survey of four material models for ballistic simulations of high-strength concrete.
Postdoc Karoline Osnes: Modelling of Fracture Initiation and Post-fracture behaviour of head impact on car windshields.
Rannveig Færgestad: Modelling and simulation of hypervelocity impact against debris shields for spacecraft protection.
Also, three former PhD students at SFI CASA, Henrik Granum. Jens Kristian Holmen and Joakim Johnsen, presented papers at LS-DYNA. They are all employed by the spin off-company Enodo AS. Jens Kristian Holmen is also a researcher at NTNU.
