Image fragmented learning for data-driven topology design

Yang, Yusibo; Yaji, Kentaro; Yamasaki, Shintaro et al.

Structural and Multidisciplinary Optimization, 2025, 68(8), 148

Number of Access:22（2025-10-05 09:44 Counts）

Identifier to cite or link to this item: https://hdl.handle.net/11094/102593

Link to primary information 

File Format Terms of use Size Views Date.Available Description information
StructMutltidiscipOpt_68_8 pdf None 5.38 MB 14 2025.08.29  
Creative Commons : Attribution

Item Information

Output File Export EndNote Basic Export Mendeley

Title
Image fragmented learning for data-driven topology design
Creator
CreatorName
Yang, Yusibo
NRID
1000090779373 1000090779373
ORCID
0000-0002-4309-2043 0000-0002-4309-2043
researchmap permalink
yajiken0609 yajiken0609
Scopus Author ID
55787947500 55787947500
CreatorName
Yaji, Kentaro
NRID
1000070581601 1000070581601
CreatorName
Yamasaki, Shintaro
NRID
1000010228992 1000010228992
ORCID
0000-0002-2197-1012 0000-0002-2197-1012
researchmap permalink
read0014111 read0014111
Scopus Author ID
7404057902 7404057902
CreatorName
Fujita, Kikuo
Subject
Data-driven topology design
Image fragmented learning
Multi-objective optimization
Topology optimization
Description
This version of the article has been accepted for publication, after peer review (when applicable) and is subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: https://doi.org/10.1007/s00158-025-04054-3.
Abstract
This paper proposes a data-driven topology design (DDTD) framework, incorporating imagefragmentedlearning that leverages the technique of dividing an image into smaller segments for learning each fragment. This framework is designed to tackle the challenges of high-dimensional, multi-objective optimization problems. Original DDTD methods leverage the sensitivity-free nature and high capacity of deep generative models to effectively address strongly nonlinear problems. However, their training effectiveness significantly diminishes as input size exceeds a certain threshold, which poses challenges in maintaining the high degrees of freedom crucial for accurately representing complex structures. To address this limitation, we split a trained conditional generative adversarial network into two interconnected modules: the first performs dimensionality reduction, compressing high-dimensional data into a lower-dimensional representation, which is then fed into a variational autoencoder (VAE) to generate new low-dimensional data. The second module reconstructs the generated low-dimensional data back into the high-dimensional design space. The effectiveness of the proposed approach is demonstrated through two case studies: the optimization of an L-bracket design problem and a turbulent heat transfer design problem, both involving design variables at a scale unattainable by the conventional VAE-based method.
Publisher
Springer Science and Business Media Deutschland GmbH
Source Title
Structural and Multidisciplinary Optimization
Volume (Issue)
68(8)
Page
148
Date of Issued
2025-08-12
Language
English
Handle URL
https://hdl.handle.net/11094/102593
PISSN
1615147X
eISSN
16151488
Relation.isIdenticalTo
DOI (Publisher Version)
https://doi.org/10.1007/s00158-025-04054-3
Access Rights
open access
CopyRight
This article is licensed under a Creative Commons Attribution 4.0 International License.
oaire:version
Version of Record
Category
学術雑誌論文 Journal Article