Approach and application to transfer heterogeneous simulation data from finite element analysis to neural networks. Issue 1 (14th December 2020)

• Record Type: Journal Article
• Authors: Spruegel, Tobias C; Bickel, Sebastian; Schleich, Benjamin; Wartzack, Sandro
• Is Part Of: Journal of computational design and engineering. Volume 8:Issue 1(2021)
• Page Start: 298
• Page End: 315
• Publication Date: 2020-12-14
• Subjects: plausibility check -- finite element simulation -- spherical detector surface -- deep neural network -- convolutional neural network -- simulation DNA; Engineering -- Data processing -- Periodicals; Computer-aided design -- Periodicals; Computer-aided design; Engineering -- Data processing; Electronic journals; Electronic journals; Periodicals; 620.0042
• DOI: 10.1093/jcde/qwaa079
• Languages: English
• ISSNs: 2288-4300

Abstract

Abstract: The simulation of product behavior is a vital part of current virtual product development. It can be expected that soon there will be more product simulations due to the availability of easy-to-use finite element analysis software and computational power. Consequently, the amount of accessible new simulation data adds up to the already existing amount. However, even when using easy-to-use finite element software tools, errors can occur during the setup of finite element simulations, and users should be warned about certain mistakes by automatic algorithms. To use the vast amount of available finite element simulations for a data-driven finite element support tool, in this paper, a methodology will be presented to transform different finite element simulations to unified matrices. The procedure is based on the projection of nodes onto a detector sphere, which is converted into a matrix in the next step. The generated matrices represent the simulation and can be described as the DNA of a finite element simulation. They can be used as an input for any machine learning model, such as convolutional neural networks. The essential steps of preprocessing the data and an application with a large dataset are part of this contribution. The trained network can then be used for an automatic plausibility check for new simulations, based on the previous simulation data from the past. This can result in a tool for automatic plausibility checks and can be the backbone for a feedback system for less experienced users. Graphical Abstract: