Вы находитесь на странице: 1из 19

Large Eddy Simulations

Goal of LES: provide flow quantities of interest with


comparable accuracy as DNS but at significantly
reduced computational cost (if RANS does not work)

Separation

Heat transfer
Free turbulence

Wall bounded
turbulence

Why LES and not RANS?


LES predicts transition behavior, wake mixing (Medic, Joo, Kalitzin, Sharma UTRC)
Mid-span section of Purdue transonic compressor rotor, Re = 740K, 135M cells, Tu=0.8%
Traditionally, fully turbulent RANS is
used in compressor simulations

Dt+ < 0.3


Dy+ < 0.1
Ds+ < 80
Dr+ < 10

transition

(x-xTE)/Bx=-0.041

RANS with transition model


improves predictions
LES allows transition prediction
and provides better wake mixing

(x-xTE)/Bx=0.0034

(x-xTE)/Bx=0.085

(x-xTE)/Bx=0.41

Research projects during 2012 CTR Summer Program


have contributed to the stated goal of LES by:

Introducing and investigating new SGS models for


momentum, mixing, and heat transfer
Developing wall models for LES for boundary layer
flows
Assessing and using LES capabilities for specific flows
where RANS techniques are inadequate, e.g., flows
and heat transfer for turbine blades, MAV, etc.
Exploiting particle methods as a complementary
approach to LES for increasing computational
efficiency for turbulence simulations

LES fundamentals
Goals:
Assess LES capabilities for simulating flows with
separation, laminar, transitional, and turbulent
regions (Francois Cadieux, Taraneh Sayadi, J. Andrzej
Domaradzki, Sanjeeb Bose, Curtis Hamman)
Develop physics based off-wall boundary condition
for LES of turbulent boundary layers (Ricardo GarciaMayoral, Brian Pierce, James Wallace)
Develop and evaluate new SGS models (Bing-Chen
Wang, Guillaume Balarac, Mohammad Saeedi, Sanjeeb
Bose, Curtis Hamman)

Goal: assess LES capabilities to reduce computational requirements for simulating


separated flows significantly, to O(1%) of DNS resolution
Benchmark problem: laminar separation bubble, Re=5x104

DNS databases:
Sayadi (Nx x Ny x Nz=1536 x 300 x 128 = 60 million points; x+ =10, y+ =0.5, z + =8)
Spalart and Strelets, JFM 403, 2000; mesh size about 15 million points

4.5

Why LES? Because RANS for this flow are not reliable.

1.
2.
3.
4.
5.

DNS: 60 million points


LES with dynamic model: 4% of DNS
No model DNS: 4% of DNS
No model DNS: 2% of DNS
No model DNS: 1% of DNS

Cp

Cf

Different investigators (Florent Duchaine and Franck Nicoud)


using a different code arrived at similar conclusions

Off-wall boundary condition at the top of the buffer layer at


y+ = 100
Statistical properties of turbulent spots in boundary layers with bypass transition
are remarkably like those in developed turbulent boundary layers.
Park et al., Phys Fl. 24, 2012 (from 2010 CTR Summer Program).
Motivating basis for creating an off-wall boundary condition (BC) at the top of the
buffer layer for a turbulent boundary layer LES.
BC to be created as a lower dimensional space-time representation of the flow at
the top of a minimal flow unit taken from the transitional flow.

K-type transition
Sayadi et al., 2012

PROJECT STEPS & ACHIEVEMENTS


Determine minimum flow unit size to give good statistics at y+ = 100 in transitional

and developed turbulent flow regions.


Obtain Fourier space-time representation of these x-z planes and explore required
modes to adequately characterize the flow.
Obtain the temporal modes using Dynamic Mode Decomposition (DMD) for these
planes.

u fluctuations

y+ = 100
Repeating minimal unit B. C.
Turbulent channel flow

u fluctuations

Yy+ = 250
DNS using this off-wall B.C.

Regularized SGS Gradient Model for Scalar Transport


The gradient model comes
from a mathematical
approximation of the filter.
Give a physical interpretation of the gradient
model to improve the predicted scales transfer.

stretching

The GSSGS transfer of the gradient model is (analytical result)

Our proposal is to keep only the forward scatter


term as SGS model resulting in a regularized gradient
model.

rotation
compression

Minimal Error Tool for SGS Models


Model performance is first evaluated through a priori tests
(based on filtered DNS of HIT)
Quadratic error of models are computed from filtered DNS of isotropic turbulence.
From optimal estimation theory,
the minimal error which is
expected for a given
model can be evaluated.

Quadratic errors as a function of the filter size

No correction of a model
could lead to smaller error
than its irreducible error.

This is an efficient tool


to evaluate the ability of
improvement of a model

[Dotted lines: optimal estimators]

LES of passive scalar in forced isotropic turbulence


Model behavior in a posteriori tests:

(5123-DNS of velocity and 323-LES of scalar)


Statistical scalar variance decay with time

Regularization procedure allows to


correct the unphysical behavior of the
classic gradient model
Dynamic gradient model is less
dissipative than the dynamic
Smagorinsky model
Scalar variance spectrum (t=2)

Good spectrum prediction even at


smallest resolved scales
Model validated in turbulent jet
flow and for various grids and
Schmidt numbers

Dynamic Full Linear Tensor-Diffusivity SGS Heat Flux Models


(DLT) Wang et al.

2
h j CE D | S |
CS D Sjk
x j
x k
2

Features of this scalar-flux model:


Tensor diffusivity;
Allows for backscatter;
SGS scalar flux is not aligned with the scalar gradient;
Typical to a dynamic modelling approach, it may still need clipping
in order to achieve numerical stability.
Dynamic Regularized Gradient Model Balarac et al.

Heated Cylinder Test Case

Objective: to predict mean and fluctuating Nusselt numbers at


Re= 3000 and 8900, and compare with the experimental data of
Nakamura & Igarashi (IJHFF, 2004).

Preliminary LES results by Sanjeeb Bose at CTR: scalar isosurface


(Re=8900, 10 million CVs). SGS stress model: dynamic Vreman model
(DVM), SGS scalar-flux model: DLT).

Mesh Adaptation (Adapt Tool) and Numerical Results

Mesh adaptation at the high Reynolds number (Re=8900).

Comparison of the LES model predictions with


the experimental data (data averaged over
approximately 10 cylinder sheddings).
Necessity for mesh adaptation at high Reynolds number
(demonstrated using the predicted mean Nusselt number).

Simulation performed by Sanjeeb Bose using


CharLES

A priori analysis of Dynamic Nonlinear Stress Model: Euler Angle Analysis


e e S

Perfect alignment
No rotation
0
q : Singular

Four SGS Stress Models Tested


Smagorinsky type models
Similarity model
Dynamic two-parameter mixed model (DTPMM)
Dynamic nonlinear SGS stress model (DNM)
Wang et al.

*ij CSij CW ij CNij

e S
Example: Smagorinsky type models

A priori test in turbulent boundary layer

e S
Euler axis & angle characteristic of a linear
constitutive relationship (Boussinesq).

75

(deg)

Turbbulent Region, Re = 1750


Line and time averaged (100 snapshots)

70

ij f (sij )

Rotation matrix:

e S

R ij e'i e j cos(e'i , e j )
Euler angle :

65

Euler angle

e e S

60

55

tr(R ij ) 1 cos

50

Euler axis q:

45

DNS (Wu & Moin, 2012)


DNM (Wang et al.)
DTPMM (Morinishi & Vasilyev)
Similarity (Liu & Meneveau)

q (R T R ) /( 2 sin )

e S

e
Euler axis & angle describing the attitude of eigen-frame
of ij with respect to eigen-frame of sij.

40

10

10

y+

10

Time- and spanwise-averaged wall-normal profiles of the


Euler angle at Re =1750. a priori LES based on channel flow
DNS data of Wu & Moin (JFM, 2012)

Accomplishments for LES Fundamentals


A benchmark case of a laminar separation bubble
followed by a nonequilibrium turbulent boundary
layer was proposed and adopted by several
groups for testing LES models
Established that LES for the benchmark case with
O(1%) of DNS resolution are feasible
Reduced-order off-wall boundary conditions for
LES were explored by employing simulation
databases for transitional flows
Improved SGS models were developed and tested
in isotropic turbulence and heated cylinder

Вам также может понравиться