Physics-informed neural networks for learning fluid flows with symmetry

We suggest symmetric variational physics-informed neural networks (symmetric VPINN) to learn the symmetric fluid flow and physical properties of fluids from a limited set of data. Symmetric VPINN is based on the VPINN framework and guarantees the symmetry of the solutions by modifying the network ar...

Full description

Saved in:
Bibliographic Details
Published in:The Korean journal of chemical engineering 2023, 40(9), 282, pp.2119-2127
Main Authors: Kim, Younghyeon, Kwak, Hyungyeol, Nam, Jaewook
Format: Article
Language:English
Subjects:
Biotechnology
Catalysis
Chemistry
Chemistry and Materials Science
Computer architecture
Exact solutions
Fluid flow
Industrial Chemistry/Chemical Engineering
Inverse problems
Laminar flow
Materials Science
Neural networks
Physical properties
Power law
Process Safety
Process Systems Engineering
Symmetry
Velocity distribution
화학공학
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:We suggest symmetric variational physics-informed neural networks (symmetric VPINN) to learn the symmetric fluid flow and physical properties of fluids from a limited set of data. Symmetric VPINN is based on the VPINN framework and guarantees the symmetry of the solutions by modifying the network architecture. The effectiveness of the symmetric VPINN is demonstrated by predicting the velocity profiles and power-law fluid properties in the Poiseuille flow of a parallel channel. Symmetric VPINN models robustly and accurately learn power-law fluid flow in both forward and inverse problems. We demonstrate that the symmetric VPINN can be particularly useful when the power-law index is small and the data are extremely limited. The modified network architecture in the symmetric VPINN guides the neural network towards an exact solution by reinforcing symmetry. We show that symmetric VPINN is effective in obtaining unknown physical properties in practical experiments where data are scarce, suggesting the possibility of introducing known conditions of the system directly into the network structure to improve the accuracy of the network.
ISSN:0256-1115
1975-7220
DOI:10.1007/s11814-023-1420-4