Multi-fidelity neural optimization machine for Digital Twins

Jie Chen, Changyu Meng, Yi Gao, Yongming Liu

Research output: Contribution to journalArticlepeer-review

6 Scopus citations

Abstract

Digital Twins (DTs) are widely used for design, manufacturing, prognostics, and decision support for operations. One critical challenge in optimizing DTs usually involves multi-fidelity (MF) models and data, such as multi-resolution computational simulation and experimental testing. The MF strategies provide advantages of high accuracy with low computational and experimental cost in DTs. A novel MF optimization framework is proposed in this paper to demonstrate and validate its potential application in DTs. First, MF data aggregation using convolutional neural networks (MDACNN) is introduced to integrate low-fidelity (LF) and high-fidelity (HF) models and data. It can fully utilize the LF data to learn the relationship across multiple fidelities. With MDACNN, predictions can be made with high accuracy compared to HF models. Next, MDACNN is integrated into the neural optimization machine (NOM), an optimization framework based on NNs. NOM is explicitly designed for optimizing NN objective functions based on the stochastic gradient descent method. The integrated model is named MF neural optimization machine (MFNOM). A numerical example is presented to illustrate the procedure of implementing MFNOM. Two engineering applications are presented for both MF simulation models and experimental data. The first problem focuses on the multi-resolution finite element simulation for structures and materials. Coarse and fine meshes are applied for simulation. The properties of multi-phase heterogeneous materials are optimized to minimize the stress in the simulation domain. The second problem investigates the internal defects in additively manufacturing. Low/high resolution and full/partial field scanning data are utilized to build DTs. MFNOM is used to design pore size and orientation to reduce the risk caused by irregular pores.

Original languageEnglish (US)
Article number340
JournalStructural and Multidisciplinary Optimization
Volume65
Issue number12
DOIs
StatePublished - Dec 2022

Keywords

  • Additive manufacturing
  • Digital Twins
  • Multi-fidelity
  • Neural networks
  • Optimization

ASJC Scopus subject areas

  • Software
  • Control and Optimization
  • Control and Systems Engineering
  • Computer Science Applications
  • Computer Graphics and Computer-Aided Design

Fingerprint

Dive into the research topics of 'Multi-fidelity neural optimization machine for Digital Twins'. Together they form a unique fingerprint.

Cite this