ResNet networks for plausibility detection in Finite Element simulations

Bickel S, Schleich B, Wartzack S (2022)

Publication Type: Conference contribution

Publication year: 2022

Publisher: The Design Society

Conference Proceedings Title: Proceedings of NordDesign 2022: How Product and Manufacturing Design Enable Sustainable Companies and Societies

Event location: Copenhagen DK

ISBN: 9781912254170

Abstract

Nowadays, the efficient execution of product design and process planning activities without the extensive use of Finite Element (FE) simulation is hardly conceivable. The ability to virtually test components at an early stage can reduce costs while significantly increas-ing product quality. In addition, the current business environment promotes ever shorter development times coupled with a greater variety of products. This combination poses the risk that more and more inexperienced users from the field of product design have to perform simulation and validation tasks. An idea to support the less experienced users is to automatically check their FE models and simulations for plausibility. This helps to con-duct simulations independently by alerting the designers when major errors occur. One approach to the automatic plausibility check of FE simulations utilizes simulations that have already been calculated as a database and applies them to train a Deep Learning model that classifies the new simulations into plausible and non-plausible. This requires converting the existing data into a uniform format by the projection method, so it is read-able for a Deep Learning network. However, for this method to be applied in an industrial environment, a high recognition accuracy is required for unknown simulations. Therefore, the goal of this paper is to investigate the ability of a new Deep Learning architecture to check the plausibility of FE simulations. The objective is to achieve the highest possible recognition accuracy. So far, mainly serial network types have been used for this procedure, which will now be extended by the application of ResNet networks. These have different paths in their structure through the addition of skip-connections, allowing for a theoretically better-trained model. This will be analysed together with a new dataset of simulations. These components for the structural mechanical simulation are geometrically varied and served to evaluate the new network types and their achieved accuracy.

Authors with CRIS profile

Sebastian Bickel Chair of Engineering Design Benjamin Schleich Chair of Engineering Design Sandro Wartzack Chair of Engineering Design

How to cite

APA:

Bickel, S., Schleich, B., & Wartzack, S. (2022). ResNet networks for plausibility detection in Finite Element simulations. In N.H. Mortensen, C.T. Hansen, M. Deininger (Eds.), Proceedings of NordDesign 2022: How Product and Manufacturing Design Enable Sustainable Companies and Societies. Copenhagen, DK: The Design Society.

MLA:

Bickel, Sebastian, Benjamin Schleich, and Sandro Wartzack. "ResNet networks for plausibility detection in Finite Element simulations." Proceedings of the NordDesign 2022: How Product and Manufacturing Design Enable Sustainable Companies and Societies, Copenhagen Ed. N.H. Mortensen, C.T. Hansen, M. Deininger, The Design Society, 2022.

BibTeX: Download