| About this Abstract |
| Meeting |
NUMISHEET 2022: The 12th International Conference on Numerical Simulation of 3D Sheet Metal Forming Processes
|
| Symposium
|
NUMISHEET 2022
|
| Presentation Title |
On the Potential of Convolutional Neural Networks for Estimating Structure-property Relationships |
| Author(s) |
Julian Heidenreich, Maysam Gorji, Dirk Mohr |
| On-Site Speaker (Planned) |
Maysam Gorji |
| Abstract Scope |
During the past years, neural networks drew more and more interest from the solid mechanics community. They have proven to provide a powerful framework, in particular for constitutive modeling. This work takes first tentative steps towards the use of convolutional neural network based modelling techniques to estimate structure-property relationships. The present work investigates the initial yield of two-dimensional architected materials. In future applications, the neural network will be trained based on physical experiments. In the course of this work, they have been replaced by virtual experiments relying on numerical simulations. The computational framework is structured as an encoder- decoder network and leverages the effectiveness of convolutional neural networks to directly translate geometrical data into mechanically significant quantities. More precisely, the constructed network manages to replicate shape distortions of the yield surface for numerous hole configurations as well as various types of perforation. Furthermore, it accurately predicts the orientation dependent material response for varying degrees of anisotropy. Additionally, the capabilities of the computational framework can be extended by adding another internal variable accounting for time-dependent material behavior. This extension allows the network to accurately mimic the evolution of constant energy surfaces over the course of the progressing deformation applied to the architected materials. The fact that the network solely relies on geometrical information leads to the strong conjecture that the neural network is capable to extract and encapsulate all geometrical and mechanically significant information into a small number of scalar variables. |
| Proceedings Inclusion? |
Definite: At-meeting proceedings |