Surrogate Model for Structural Analysis Simulation using Graph Convolutional Network (Unrefereed Workshop Manuscript)

• Item type: SIG Technical Reports(1)
• 公開日: 2021-07-13
• タイトル: Surrogate Model for Structural Analysis Simulation using Graph Convolutional Network (Unrefereed Workshop Manuscript)
• タイトル / 言語: en
• タイトル: Surrogate Model for Structural Analysis Simulation using Graph Convolutional Network (Unrefereed Workshop Manuscript)
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 深層学習
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Fujitsu Ltd.
• 著者所属: Fujitsu Ltd.
• 著者所属: Fujitsu Ltd.
• 著者所属: The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Fujitsu Ltd.
• 著者所属(英): en
• 著者所属(英): Fujitsu Ltd.
• 著者所属(英): en
• 著者所属(英): Fujitsu Ltd.
• 著者所属(英): en
• 著者所属(英): The University of Tokyo
• 著者名: Amir, Haderbache; Koichi, Shirahata; Tsuguchika, Tabaru; Hiroshi, Okuda
• 著者名(英): Amir, Haderbache; Koichi, Shirahata; Tsuguchika, Tabaru; Hiroshi, Okuda
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Deep learning surrogate models can replace simulation solver computations by a low-latency inference. However, current surrogate models have important drawbacks such as inefficient memory usage when mapping mesh into regular-grids or limited predictive capabilities to specific input conditions. We propose an efficient surrogate model based on graph convolutional neural network designed for any finite-element mesh and boundary conditions. The accuracy is guaranteed by a mapping of the FEA data into the GCN graph while scalability is achieved by using the principal component analysis on the stiffness matrix. Our technique achieves significant speed-up and maintain accuracy with 1e-03 precision compared with HPC simulations.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Deep learning surrogate models can replace simulation solver computations by a low-latency inference. However, current surrogate models have important drawbacks such as inefficient memory usage when mapping mesh into regular-grids or limited predictive capabilities to specific input conditions. We propose an efficient surrogate model based on graph convolutional neural network designed for any finite-element mesh and boundary conditions. The accuracy is guaranteed by a mapping of the FEA data into the GCN graph while scalability is achieved by using the principal component analysis on the stiffness matrix. Our technique achieves significant speed-up and maintain accuracy with 1e-03 precision compared with HPC simulations.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10463942
• 書誌情報: 研究報告ハイパフォーマンスコンピューティング（HPC）; 巻 2021-HPC-180, 号 10, p. 1-9, 発行日 2021-07-13
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8841
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会