# Otto-von-Guericke-Univ.

Magdeburg

Turbulence

According to an apocryphal story Werner Heisenberg was asked what he would ask God, given the opportunity. His reply was: "When I meet God, I am going to ask him two questions: Why relativity? And why turbulence? I really believe he will have an answer for the first."

**Numerische Strömungsmechanik / Computational Fluid Dynamics (CFD): Turbulence
**

Dominique Thévenin/Gábor Janiga Lehrstuhl für Strömungsmechanik und Strömungstechnik (LSS)

Dominique.Thevenin@ovgu.de Gabor.Janiga@ovgu.de

Organization

!Introduction

!Basics of turbulence !Direct Numerical Simulation !Reynolds-Averaged Navier-Stokes models !The « middle way »? Large-Eddy Simulation and Detached-Eddy Simulation

General introduction

!More than 90% of all industrial applications involving ﬂuids take place in a turbulent regime !Whenever turbulence is present, it plays a very important role and controls key properties: " mixing of different components " heat transfer at the wall " drag " ... ! It is therefore essential to understand turbulence and to be able to model it correctly

**Starting system of equations
**

!Working in the ﬁeld of Computational Fluid Dynamics, a good starting point: the incompressible Navier-Stokes equations !These equations take basically 4 variables into account !They are fully coupled and extremely stiff (very broad range of spatial and temporal scales) !They represent already a (quite accurate) modeling of the physical reality !They contain real turbulence (see later DNS)

Governing equations

Conservation of mass:

∂ ∂ρ + (ρui ) = 0 ∂t ∂ xi

Momentum equations (3 equations):

∂ ∂ ∂p (ρui ) + (ρui uj ) = − + ∂t ∂ xj ∂ xi

+

∂ µ ∂ xj

∂ ui ∂ uj 2 ∂ ul + − δij ∂ xj ∂ xi 3 ∂ xl

Complexity scale

!These equations represent quite well complex turbulent ﬂows !Many possibilities exist to write down such equations, depending on the acceptable cost/complexity ratio: three-dimensional compressible with chemical reactions multiphase ﬂows including radiation model cost & complexity scale non-equilibrium thermodynamics ...

Organization

!Introduction

!Basics of turbulence

!Direct Numerical Simulation !Reynolds-Averaged Navier-Stokes models !The « middle way »? Large-Eddy Simulation and Detached-Eddy Simulation

Back to the roots…

Experimental results

Turbulent boundary layer

**Energy transfer for turbulence
**

energy of turbulence E(k) size of turbulent structure large vortices small vortices

**Energy transfer for turbulence
**

L E(k) ld energy transfer between scales, no creation, no destruction

cascade of turbulent energy (power law -5/3)

log

log

wave number k of turbulent structure

mean ﬂow feeds ke turbulence, characteristic length = size of domain

kd

k

dissipation of turbulence by viscosity, at Kolmogorov length

**The life of turbulence: example
**

dissipation through energy transfer between scales, viscosity at Kolmogorov no creation, no destruction scale methane

**What does it imply?
**

!There is a minimum and a maximum spatial scale for any given turbulent computation: " minimum: Kolmogorov length " maximum: size of the device !Similarly, there is a minimum and a maximum temporal scale for any given turbulent computation: " minimum (fastest): characteristic time of sound propagation " maximum (slowest): total duration of the process, or time needed to get « established » ﬂow (at least in a statistical sense)

air

mean ﬂow feeds turbulence through instabilities, characteristic size = size of the domain

Organization

!Introduction !Basics of turbulence

**Getting the idea of DNS...
**

!Experiments and theory are limited. We need numerical simulations as a complementary source of information, but without introducing doubtful hypotheses This is « possible »: all spatial and temporal scales are bounded! We « just » need to solve everything exactly: Direct Simulation!

**!Direct Numerical Simulation
**

!Reynolds-Averaged Navier-Stokes models !The « middle way »? Large-Eddy Simulation and Detached-Eddy Simulation

**Estimation for DNS
**

ReL N (# of nodes in each direction) 104 214 498 1.260 3.360 9.218 N3(total nodes) M (time steps) CPU Time

**Status and use of DNS
**

!Why use Direct Numerical Simulations ? " Because there is no other tool allowing to investigate turbulence in such detail. DNS is (should be?) a “numerical experiment”. But the computing cost is very high... !Is it possible in practice to obtain useful results? " Yes, certainly! Of course, a lot of work is necessary to obtain a sufﬁcient quality. And many limitations remain! Not an industrial tool yet for full-scale installations…

94 375 1.500 6.000 24.000 96.000

1.1×106 1.0×107 1.2×108 2.0×109

1.2×103 3.3×103 9.2×103 2.6×104

20 9 13 20 90 5.000

min h days months years years

3.8×1010 7.4×104 7.8×1011 2.1×105

Pope: Turbulent Flows (2000) - assuming 1 gigaﬂop

Organization

!Introduction !Basics of turbulence !Direct Numerical Simulation

**Time averaging for nonstationary turbulence
**

u ¯i (x, t) = 1 T

t+T

ui (x, t) dt,

t

T1

T

T2

**!Reynolds-Averaged Navier-Stokes models
**

!The « middle way »? Large-Eddy Simulation and Detached-Eddy Simulation

Kolmogorov hypothesis: separate each variable in a mean part and a ﬂuctuating part:

ui (x, t) = u ¯i (x, t) + ui (x, t)

mean ﬂuctuating

Mean equations

!Use this decomposition for each variable

Mean equations: mass

¯+φ φ=φ

Incompressible ﬂow

!Introduce it in the original system, and develop !You obtain products (mean x mean), (mean x ﬂuctuation), (ﬂuctuation x ﬂuctuation), (triple products) !Take the average of the resulting full equations !Simplify by taking into account that

∂ (ux ) ∂ (uy ) ∂ (uz ) + + =0 ∂x ∂y ∂z

¯=φ ¯ φ ¯ =φ−φ ¯=φ ¯−φ ¯=0 φ

ux = u ¯ x + ux uy = u ¯ y + uy uz = u ¯ z + uz

Mean equations: mass

**Mean equations: momentum
**

¯+φ φ=φ

ui = u ¯ i + ui

uy ) ∂ (¯ uz ) ∂ (ux ) ∂ uy ∂ (uz ) ∂ (¯ ux ) ∂ (¯ + + + + + =0 ∂x ∂y ∂z ∂x ∂y ∂z

Momentum equations:

∂ (¯ ux ) ∂ (¯ uy ) ∂ (¯ uz ) + + =0 ∂x ∂y ∂z

Same as starting equation, but with mean variables only!

∂ ∂ ∂p (ρu ¯i ) + (ρu ¯i u ¯j ) = − + ∂t ∂ xj ∂ xi ∂u ¯i ∂u ¯j 2 ∂u ¯l ∂ µ + − δij + ∂ xj ∂ xj ∂ xi 3 ∂ xl

+

∂ (−ρui uj ) ∂ xj

Contain (unknown !) correlations #closure problem #a (very difficult) modeling is required !

**Turbulence models for RANS
**

!The oldest models, but which are still mostly in use today, are based on the idea of a turbulent viscosity (Boussinesq) !This is a purely hypothetical analogy with the classical stresstensor!

**Reynolds-averaged Navier-Stokes
**

Conservation of mass:

∂ ∂ρ + (ρu ¯i ) = 0 ∂t ∂ xi

−ρui uj = µt

∂u ¯j ∂u ¯i + ∂ xj ∂ xi

−

2 3

ρk + µt

∂u ¯i ∂ xi

Momentum equations:

δij

!But the question remains: how to compute this turbulent viscosity, which is problem-dependent and NOT a ﬂuid parameter????

∂ ∂ ∂p (ρu ¯i ) + (ρu ¯i u ¯j ) = − + ∂t ∂ xj ∂ xi ∂u ¯i ∂u ¯j 2 ∂u ¯l ∂ µ + − δij + ∂ xj ∂ xj ∂ xi 3 ∂ xl

−ρui uj = µt ∂u ¯j ∂u ¯i + ∂ xj ∂ xi 2 3 ρk + µt

+

∂ (−ρui uj ) ∂ xj

−

∂u ¯i ∂ xi

δij

**Turbulence models for RANS
**

!Zero-equation model: mixing length of Prandtl

**Turbulence models for RANS
**

!One-equation model: not much used any more, but useful to develop two-equation models. !Introduce the kinetic energy of turbulence k:

with

Standard k-ε-Modell

Standard k-ω-Modell

Model constants: Model constants:

Alternatives to RANS?

!Up to now, nobody has been able to propose a really satisfactory closure. Many possibilities have been tested, but the results lack generality, need many « constants »,... !The resulting models are not really « predictive » !But this is still the best compromise for industrial purposes (k-!), more than 90% of all computations! Hands-On…

Reynolds-Stress-Modell (RSM)

Other possibilities ?

!Avoid the closure at this level. Write transport equations for the ﬂuctuations, introduce higher-moment ﬂuctuations, close at that level... and hope it will sometime converge! RSM-models. !Avoid completely the closure problem, which means you do not want to average the equations... Leads to DNS.

Organization

!Introduction !Basics of turbulence !Direct Numerical Simulation !Reynolds-Averaged Navier-Stokes models

**Key idea of LES
**

size of turbulent structure large vortices small vortices energy of turbulence E(k) cascade of turbulent energy wave number k of turbulent structure

!The « middle way »? Large-Eddy Simulation and Detached-Eddy Simulation

large scales are computed exactly (DNS)

small scales (dissipative) are modelled (back to RANS)

**Key idea of LES (2)
**

small scales (dissipative) are modelled

Advantages of LES

L E(k) ld

cut-off scale " model (=RANS)

methane

air

exact

(« = DNS ») kd ke k

large scales are computed exactly

Comments on LES

!First parameter of the LES computation: cut-off scale " !Limiting cases: " if " = Kolmogorov length: this is a true DNS " if " = size of the device: this is a « RANS-type » computation !Advantage for modeling: " the modelled scales are the smallest ones. In that case, the hypothesis of homogeneous isotropic turbulence is OK! " they contain relatively little energy. Small inaccuracies should not change much the larger (interesting) scales !Advantage for computation: when computer power increases, simply reduces ", increasing the quality of the computation

Filtering of equations

!The best way to take into account the fact that only large-scale quantities are computed is by introducing ﬁltering !For example, for the velocity variable, the full (exact) value is represented by u, while the ﬁltered value, corresponding to the resolved large scales, is written u !The ﬁltering procedure is completely deﬁned by giving a ﬁlter function H, which is used to compute the ﬁltered values knowing the exact ones !It is then possible to ﬁlter the full Navier-Stokes equations in order to obtain the resulting system of equations for the ﬁltered variables, solved by the LES computations !This procedure is very similar to the Komogorov hypothesis (separation: mean/ﬂuctuating parts). But in space, not in time!

Filtered equations

∂u ¯j =0 ∂ xj

∂ (¯ ui u ¯j ) 1 ∂p ¯ ∂ ∂u ¯i + =− + ν ∂t ∂ xj ρ0 ∂ xi ∂ xj

continuity momentum (3 eqs.)

Comparison

∂u ¯i ∂u ¯j + ∂ xj ∂ xi

+ Tij

The subgrid-scale tensor T is similar to the Reynolds-stress tensor introduced in the RANS turbulent decomposition

Tij = u ¯i u ¯ j − ui uj

This tensor must be modeled, since we do not know the correlation term...

Comparison

RANS: statistical approach Averaged values Increasing need for models

To summarize

(Real) Navier-Stokes equations

LES: filtered approach Filtered values Increasing CPU cost DNS: exact approach Local & instantaneous values

direct, but with introduce Kolmogorov bad resolution hypothesis, separate (large time-step, mean and ﬂuctuation few grid points) (closure problem) k – !,… OK if laminar industry

LES help

direct, AND accurate resolution DNS

Reynolds stress model help something really « cheap and good » ??