Activities

Home > Activity > Newsletter > 2023.10

Newsletter  2023.10  Index

Theme : "AJK FED 2023"

  1. Preface
    Hyun Jin PARK, Shoichi MATSUDA, Chungpyo HONG
  2. Evaluation of railway vehicles' resistance against strong crosswinds and its application for safe railway operation
    Yayoi MISU (East Japan Railway Company)
  3. Order within Turbulence
    Susumu GOTO (Osaka University), Yutaro MOTOORI (Osaka University)
  4. LES/Lagrangian-particle-simulation of a Reactive Turbulent Planar Jet
    Jiabao Xing (Nagoya University),Tomoaki WATANABE (Nagoya University),and Koji NAGATA (Kyoto University)
  5. Unsteady Characteristics of Tip Leakage Vortex Cavitation in the Occurrence of Cavitation Instability in Liquid Rocket Inducer
    Koki TAMURA (Tohoku University),Yuto NAKURA (Tohoku University), Satoshi KAWASAKI (Japan Aerospace Exploration Agency), Yuka IGA (Tohoku University)
  6. Water Condensation in PEMFCs at Nano-scale: Insights through Lattice DFT simulations
    Clint John Cortes OTIC (The University of Tokyo), Masazumi ARAO (FC-Cubic), Masashi MATSUMOTO (FC-Cubic), Hideto IMAI (FC-Cubic), Ikuya KINEFUCHI (The University of Tokyo)
  7. Reconstruction of Fluid Stress Field from Flow Birefringence using Physics-Informed Convolutional Encoder-Decoder (PICED)
    Daichi IGARASHI (Tokyo University of Agriculture and Technology), Shun MIYATAKE (Tokyo University of Agriculture and Technology), Jingzu YEE (Tokyo University of Agriculture and Technology), Yoshiyuki TAGAWA (Tokyo University of Agriculture and Technology)
  8. Determination of Permeability in the Volume Penalisation Method with a Smooth Mask Function
    Taichi TSUJIMOTO (Osaka University), Yuta NAKAO (Osaka University), Takuya TSUJI (Osaka University), Toshitsugu TANAKA (Osaka University), Kimiaki WASHINO (Osaka University)

 

Reconstruction of Fluid Stress Field from Flow Birefringence using Physics-Informed Convolutional Encoder-Decoder (PICED)


Daichi IGARASHI
Shun MIYATAKE
Jingzu YEE
Yoshiyuki TAGAWA
Tokyo University of
Agriculture and Technology

Abstract

A photoelastic method is expected to be a new fluid measurement technique that enables non-contact and direct measurement of the stress field. However, the measured birefringence values are integrated values along the optical axis direction, which can only be reconstructed using a complex matrix calculation. Therefore, we propose a machine learning-based photoelasticity approach as an alternative to reconstructing the fluid stress field. For this purpose, we construct and train a Physics-Informed Convolutional Encoder-Decoder (PICED) to predict the stress field from the flow birefringence field. For the loss function, “data loss” is calculated by the mean squared error (MSE) from the correct image while “residual loss” is calculated from the continuity equation and viscosity law. Such loss function enables a physics-informed prediction. The results show that the model successfully predicted the trends and values of the stress field with a mean absolute percentage error relative to answer data as low as 1.74 %. These results show that the model has correctly predicted with high accuracy even for untrained interpolated data. Moreover, the results of PICED are more accurate than the results trained with only “data loss”. This is because PICED considers the physical equation and predicts the values smoothly by considering the continuity equation. Such findings indicate the successful construction of a machine learning model that considers the physical phenomena and makes more accurate predictions. Furthermore, the results obtained in this study mark a major advance toward the non-contact and direct stress measurement of a fluid flow.

Key words

Fluid stress field, Photoelasticity, Convolutional Encoder-Decoder, Physics-informed neural network.

Figures


Fig. 1
(a) Schematic of the automated experimental setup (b) Examples of the images of (i) retardation and (ii) orientation captured by the polarization camera.


Fig. 2
Architecture of the Physics-Informed Convolutional Encoder-Decoder (PICED).

 
Fig. 3 Velocity u, v and stress intensity σxy: calculation results of CFD, prediction results of CNN and PICED.

Last Update:10.13.2023