Join on-line via Zoom: https://wse.zoom.us/j/99813484575
“Impact on the Structure of Turbulence by Different Axial Casing Grooves in the Tip Region of an Axial Compressor Rotor Passage”
Presented by SUBHRA SHANKHA KOLEY
(Adviser: Prof. Joseph Katz)
Stereo PIV measurements performed in a refractive index matched facility examine the effect of axial casing grooves (ACGs) on the structure of turbulence in the tip region of an axial compressor rotor. The ACGs delay the onset of stall at low flowrates by entraining the Tip Leakage Vortex (TLV), and by causing periodic changes to incident angle as their outflow impinges on the rotor blade. However, ACGs typically cause undesirable loss of efficiency at design flowrates. Interactions of the tip flow with ACGs modifies the magnitude and spatial distribution of the highly anisotropic and inhomogeneous components of the turbulent kinetic energy (TKE). Owing to TLV entrainment, at low flowrate the ACGs reduce the turbulence in the passage compared to that of the smooth endwall, but the anisotropy varies with the groove geometry. Still, the TKE is high in the TLV, the shear layer separating the backward leakage flow from the main passage flow, and near the corner of the grooves. At high flowrates, interactions of the TLV with secondary flows generated by typical grooves increase the tip region turbulence. This adverse effect and associated efficiency loss can be mitigated using grooves that minimize the injection of secondary flows into the passage at high flowrates. Most of the observed trends can be explained by examining the spatial distribution of the turbulence production rates. Such understanding elucidates the different mechanisms involved and provides a unique database for modelling turbulence in the passage.
“Autoencoder and Neural Networks to Learn the Reduced Order Dynamics in Multiphase Flows”
Presented by CRISTINA MARTIN LINARES
(Adviser: Profs. Gretar Tryggvason & Yannis Kevrekidis)
We explore the use of neural networks (NN) to learn a grey box model of the averaged 1-dimensional PDE that governs multiphase flows in a 2-dimensional vertical channel. The data is generated using direct numerical simulations . The covariance method is used to perform proper orthogonal decomposition (POD) on the velocity and void fraction. The selected POD modes are further reduced through an autoencoder. The selected minimum number of embeddings have a one-to-one correspondence with the first few POD modes. The reconstruction of the variables filters the data so we can learn an effective ODE. In addition to learning an ODE for the nonlinearly reduced representation of the system, we employ a grey-box PDE modelling approach. Here, we use a NN to approximate unknown closure terms in an otherwise theoretically-grounded PDE.