Deep learning approach for topology optimization
- Authors
- Kim, Kyeong-Hwan; Gu, Hye-Ji; Han, Seog-Young
- Issue Date
- Oct-2018
- Publisher
- 한국생산제조학회
- Keywords
- Deep learning; Machine learning; Topology optimization; Generative Adversarial Networks (GAN)
- Citation
- Proceedings of the International Conference of Manufacturing Technology Engineers (ICMTE) 2018, pp.17 - 17
- Indexed
- OTHER
- Journal Title
- Proceedings of the International Conference of Manufacturing Technology Engineers (ICMTE) 2018
- Start Page
- 17
- End Page
- 17
- URI
- https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/15986
- Abstract
- The purpose of this paper is to predict the optimal topology of a structure using deep learning. Deep learning is a field of machine learning that uses Deep Neural Networks (DNN) to train feature data. Generative Adversarial Networks (GAN) were used for automatic topology optimization. GAN is a generative modeling technique that generates a new image with generator and discriminator networks opposing each other. Previously used algorithms are computationally complex and time-consuming because they use element sensitivity for topology optimization. Topology optimization using finite element method and element sensitivity may lead to checkerboard pattern, mesh dependency and local convergence. In this study, since the learning is performed from the optimal topology image dataset with a grid of 28×28 size, the sensitivity except for creating the learning dataset is not used. As the result, the optimal topology can be independently predicted without the above problems. It is verified through numerical experiments that the proposed approach can successfully predict optimal topologies for the forces applied to the untrained positions.
- Files in This Item
-
Go to Link
- Appears in
Collections - 서울 공과대학 > 서울 기계공학부 > 1. Journal Articles
![qrcode](https://api.qrserver.com/v1/create-qr-code/?size=55x55&data=https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/15986)
Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.