full
search.png

Search

close.png

Cancel

arrow
Volume 36, Issue 3
Physically Guided Neural Network Based on Transfer Learning (TL-PGNN) for Hypersonic Heat Flux Prediction

Biao Chen, Jifa Zhang, Shuai Zhang, Xiaoxiao Song & Yao Zheng

DOI: 10.4208/cicp.OA-2023-0308

Commun. Comput. Phys., 36 (2024), pp. 651-672.

Published online: 2024-10

Preview Purchase PDF 114 2044

Cited by

google scholar semantic scholar
Export citation
  • Abstract

Hypersonic vehicles create high-temperature environments, thus causing air molecules to undergo chemical reactions. As a result, the nonlinearity of the mapping relationship between heat flux and inflow condition is strengthened. Accurately predicting the heat flux of real gases using machine learning methods relies on an amount of training data of real gas, but the computational cost is often unacceptable. To solve the issue stated, we propose a novel neural network named Physically Guided Neural Network based on Transfer Learning (TL-PGNN). We design a new network architecture to depict physical laws and use heat flow data of ideal gases to enhance the network’s capability to depict heat flow from only a few real gas samples. Experiments of sphere demonstrate that, compared to ordinary DNN and PGNN, applying TL-PGNN decreases the mean L1 error by 72.66% and 24.53% on the test set, respectively.

  • Keywords

TL-PGNN, transfer learning, heat flux prediction, small datasets.

  • AMS Subject Headings

76K05, 68T05

  • Copyright

COPYRIGHT: © Global Science Press

  • Email address
  • BibTex
  • RIS
  • TXT
@Article{CiCP-36-651, author = {Chen , BiaoZhang , JifaZhang , ShuaiSong , Xiaoxiao and Zheng , Yao}, title = {Physically Guided Neural Network Based on Transfer Learning (TL-PGNN) for Hypersonic Heat Flux Prediction}, journal = {Communications in Computational Physics}, year = {2024}, volume = {36}, number = {3}, pages = {651--672}, abstract = {

Hypersonic vehicles create high-temperature environments, thus causing air molecules to undergo chemical reactions. As a result, the nonlinearity of the mapping relationship between heat flux and inflow condition is strengthened. Accurately predicting the heat flux of real gases using machine learning methods relies on an amount of training data of real gas, but the computational cost is often unacceptable. To solve the issue stated, we propose a novel neural network named Physically Guided Neural Network based on Transfer Learning (TL-PGNN). We design a new network architecture to depict physical laws and use heat flow data of ideal gases to enhance the network’s capability to depict heat flow from only a few real gas samples. Experiments of sphere demonstrate that, compared to ordinary DNN and PGNN, applying TL-PGNN decreases the mean L1 error by 72.66% and 24.53% on the test set, respectively.

}, issn = {1991-7120}, doi = {https://doi.org/10.4208/cicp.OA-2023-0308}, url = {http://global-sci.org/intro/article_detail/cicp/23455.html} }
TY - JOUR T1 - Physically Guided Neural Network Based on Transfer Learning (TL-PGNN) for Hypersonic Heat Flux Prediction AU - Chen , Biao AU - Zhang , Jifa AU - Zhang , Shuai AU - Song , Xiaoxiao AU - Zheng , Yao JO - Communications in Computational Physics VL - 3 SP - 651 EP - 672 PY - 2024 DA - 2024/10 SN - 36 DO - http://doi.org/10.4208/cicp.OA-2023-0308 UR - https://global-sci.org/intro/article_detail/cicp/23455.html KW - TL-PGNN, transfer learning, heat flux prediction, small datasets. AB -

Hypersonic vehicles create high-temperature environments, thus causing air molecules to undergo chemical reactions. As a result, the nonlinearity of the mapping relationship between heat flux and inflow condition is strengthened. Accurately predicting the heat flux of real gases using machine learning methods relies on an amount of training data of real gas, but the computational cost is often unacceptable. To solve the issue stated, we propose a novel neural network named Physically Guided Neural Network based on Transfer Learning (TL-PGNN). We design a new network architecture to depict physical laws and use heat flow data of ideal gases to enhance the network’s capability to depict heat flow from only a few real gas samples. Experiments of sphere demonstrate that, compared to ordinary DNN and PGNN, applying TL-PGNN decreases the mean L1 error by 72.66% and 24.53% on the test set, respectively.

Biao Chen, Jifa Zhang, Shuai Zhang, Xiaoxiao Song & Yao Zheng. (2024). Physically Guided Neural Network Based on Transfer Learning (TL-PGNN) for Hypersonic Heat Flux Prediction. Communications in Computational Physics. 36 (3). 651-672. doi:10.4208/cicp.OA-2023-0308
Copy to clipboard
BibteX RIS TXT
The citation has been copied to your clipboard