TY - CONF
T1 - Super-resolution analysis with machine learning for low-resolution flow data
AU - Fukami, Kai
AU - Fukagata, Koji
AU - Taira, Kunihiko
N1 - Funding Information:
Kai Fukami and Koji Fukagata are supported by the Japan Society for the Promotion of Science (KAKENHI grant number: 18H03758). Kunihiko Taira acknowledges support from ARO (grant number: W911NF-17-1-0118) and AFOSR (grant number: FA9550-16-1-0650).
Publisher Copyright:
© 2019 International Symposium on Turbulence and Shear Flow Phenomena, TSFP. All rights reserved.
PY - 2019
Y1 - 2019
N2 - Machine-learned super-resolution is performed to reconstruct the high-resolution flow (HR) field from low-resolution (LR) fluid flow data. As preliminary tests, we use two-dimensional cylinder and NACA0012 airfoil wake flow fields and observe good agreement with reference HR data. Next, we apply two machine-learned architectures based on the convolutional neural network (CNN) for two-dimensional decaying isotropic turbulence. The HR data sets are obtained from direct numerical simulation (DNS) and LR data sets are generated by max and average pooling operations. In this work, we present the hybrid Down-sampled Skip-Connection Multi-Scale (DSC/MS) model, which can reconstruct the flow field accurately from coarse input flow field data. Towards the end of the paper, we discuss the possibility of a machine-learned model for super-resolution in experimental and computational fluid dynamics.
AB - Machine-learned super-resolution is performed to reconstruct the high-resolution flow (HR) field from low-resolution (LR) fluid flow data. As preliminary tests, we use two-dimensional cylinder and NACA0012 airfoil wake flow fields and observe good agreement with reference HR data. Next, we apply two machine-learned architectures based on the convolutional neural network (CNN) for two-dimensional decaying isotropic turbulence. The HR data sets are obtained from direct numerical simulation (DNS) and LR data sets are generated by max and average pooling operations. In this work, we present the hybrid Down-sampled Skip-Connection Multi-Scale (DSC/MS) model, which can reconstruct the flow field accurately from coarse input flow field data. Towards the end of the paper, we discuss the possibility of a machine-learned model for super-resolution in experimental and computational fluid dynamics.
UR - http://www.scopus.com/inward/record.url?scp=85084024207&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85084024207&partnerID=8YFLogxK
M3 - Paper
AN - SCOPUS:85084024207
T2 - 11th International Symposium on Turbulence and Shear Flow Phenomena, TSFP 2019
Y2 - 30 July 2019 through 2 August 2019
ER -