Numerical comparison among POD-ROM (3D)
We investigated the POD closure models for the moderate turbulent flow: 3D flow past a cylinder at Re = 1,000. Collecting 1,000 snapshots on the velocity filed over the time interval [0, 75], we employed the POD method and achieved the leading 6 POD modes that capture 84% of the system's kinetic energy. All POD-ROMs tested here were generated based on this set of basis and integrated over the time interval [0, 150].
Simulations are shown below. For clarity, only five isosurfaces are drawn.
Direct Numerical Simulation - obtained from CFD and considered as banchmark solutions. |
|
|
POD-Galerkin ROM (no turbulence modeling) - it seems to add unphysical structures and, clearly, the numerical result is inappropriated. |
|
|
Mixing-Length ROM (linear turbulence modeling) - dominated structures are missed after a while. It dues to the constant artificial viscosity (both spatially and temporally independent) that is added to the ROM eliminates some of the DNS structures. |
|
|
VMS POD-ROM (nonlinear trubulence modeling) - Variational Multiscales mthod is extened to the POD setting. Artifical viscosity (spatially dependent) has been introduced to the small scales retained in the ROM. Dominated sturctures are clearly perserved. |
|
|
DS POD-ROM (nonlinear trubulence modeling) - Dynamic subgrid-scale method is extended to the POD setting. Artifical viscosity (both spatially and temporally dependent) has been introduced to the ROM. Dominated sturctures are clearly perserved. |
|