Professional Documents
Culture Documents
Turb PDF
Turb PDF
net/publication/318562275
CITATIONS READS
0 4,056
1 author:
Ideen Sadrehaghighi
CFD Open Series
23 PUBLICATIONS 24 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Ideen Sadrehaghighi on 05 July 2018.
Turbulence Modeling
A Review
Turbulence
contemporary abstaract
By Nancy Eckels
ANNAPOLIS, MD
2
Contents
1 Introduction .................................................................................................................................. 6
List of Tables:
Table 3.1 Constant values for κ-ε turbulence model equations ............................................................... 23
Table 3.2 Constant values for SST κ-ω turbulence model equations..................................................... 27
Table 4.1 Summery of Turbulence Strategies ................................................................................................. 58
Table 5.1 Advantages and Disadvantages of different Turbulence Models ........................................ 62
4
List of Figures:
Figure 2.1 Flow Pattern Pass a Circle..................................................................................................................... 8
Figure 2.2 Large Eddies in a Turbulent Boundary Layer ............................................................................ 10
Figure 2.3 Velocity Profile for Laminar vs Turbulent Flows ..................................................................... 11
Figure 2.4 Turbulence Transition Region ......................................................................................................... 11
Figure 2.5 Snapshots of the Stream-Wise Velocity Component for (A) Adiabatic Wall , (B)
Heated Wall, and (C) Cooled Wall ............................................................................................................................. 12
Figure 3.1 Relationship between averaged variables; a) Steady flow b) Un-Steady flow ...... 15
Figure 3.2 Hierarchy of Turbulence Models..................................................................................................... 18
Figure 3.3 Lift Coefficient at Stall (AoA) against Number of Cells........................................................... 28
Figure 3.4 Comparison between experimental data [Abbott et al.]and three different turbulent
models simulation results of the Lift Coefficient for NACA 0012 Airfoil .................................................. 29
Figure 3.5 Comparison Between Experimental Data for Transitional Boundary Layer and
Different Turbulent Models on the Drag Coefficient of NACA-0012 Airfoil ........................................... 30
Figure 3.6 Contours of velocity magnitude at 9° (Top) and 16° (Bottom) AoA with the Spalart-
Allmaras turbulence model ......................................................................................................................................... 31
Figure 3.7 Zones in Turbulent B. L. for a typical Incompressible flow over a smooth flat plate 33
Figure 3.8 Turbulence and near Wall Function .............................................................................................. 34
Figure 3.9 Schematics of hybrid LES-RANS scheme (upper) and two-layer zonal scheme
(lower) ................................................................................................................................................................................. 35
Figure 4.1 Schematic Representation of the Computational Domain ................................................... 40
Figure 4.3 Comparison of Wall Pressure Coefficient (Cp) .......................................................................... 41
Figure 4.2 Streamline Patterns at Four Different Time Intervals for one Vortex Shedding
Cycle ...................................................................................................................................................................................... 41
Figure 4.4 Variation of lift Coefficient with Time ........................................................................................... 41
Figure 4.5 Difference Between the Filtered Velocity and the Instantaneous Velocity .................. 42
Figure 4.6 A velocity field produced by a (LES) of homogeneous decaying turbulence................ 43
Figure 4.7 Vorticity Prompted by the Wake Passing Cycle ........................................................................ 46
Figure 4.8 Time Averaged (RANS) vs Instantaneous (DES) Simulation over a backup step ....... 47
Figure 4.9 Integrated RANS-LES Computations in Gas Turbines: Compressor-Diffuser, ............. 49
Figure 4.11 Compressor and combustor: RANS and LES axial velocity, mid-passage (Courtesy
of Medic et al.) ................................................................................................................................................................... 50
Figure 4.10 DES and effect of Grid Density on the Wake flow ................................................................. 51
Figure 4.12 Comparison of Pressure contours on planner cuts for RANS and DES models ........ 52
Figure 4.13 Vorticity contours from spectral DNS at two Reynolds numbers (Spalart et al.
2008). ................................................................................................................................................................................... 53
Figure 4.14 The sound radiated by a Mach 1.9 circular jet ........................................................................ 54
Figure 4.15 Instantaneous Contours of Stream-Wise Density Gradients from DNS ....................... 55
Figure 4.16 Stream Lines in Square Channel with Nonlinear Constitutive –( Courtesy of
[speziale, 1987]) .............................................................................................................................................................. 56
Figure 4.17 Simulation of Flow Past Circular Cylinder by Various Approaches (Shur et al.,
1996; Travin et al., 2000) ............................................................................................................................................. 57
5
6
1 Introduction
Turbulence has been the victim of many colorful descriptions over the years, from Lamb’s (1916)
scholarly “chief outstanding difficulty of our subject” to Bradshaw’s (1994) inspired “invention of the
Devil on the seventh day of creation.” This apparent frustration results largely from the mixture of
chaos and order and the wide range of length and time scales that turbulent flows possess1. The three
key elements of CFD are algorithm development, grid generation and turbulence modelling.
Turbulence is inherently three-dimensional and time dependent, and an enormous amount of
information is thus required to completely describe a Introduction Turbulent Flow. This is beyond the
capability of the existing computers for virtually all practical flows. Thus, some kind of approximate
and statistical method, called a turbulence model, is needed.
Complexity of different turbulence models may vary strongly depends on the details one wants to
observe and investigate by carrying out such numerical simulations. N-S equation is inherently
nonlinear, time-dependent, three-dimensional PDE. Turbulence could be thought of as instability
of laminar flow that occurs at high Reynolds numbers. Such instabilities origin form interactions
between non-linear inertial terms and viscous terms in N-S equation. These interactions are
rotational, fully time-dependent and fully three-dimensional. Rotational and three-dimensional
interactions are mutually connected via vortex stretching. Vortex stretching is not possible in two
dimensional space. That is also why no satisfactory two-dimensional approximations for turbulent
phenomena are available. Furthermore turbulence is thought of as random process in time.
Therefore no deterministic approach is possible. Certain properties could be learned about
turbulence using statistical methods. These introduce certain correlation functions among flow
variables. However it is impossible to determine these correlations in advance2.
Another important feature of a turbulent flow is that vortex structures move along the flow. Their
lifetime is usually very long. Hence certain turbulent quantities cannot be specified as local. This
simply means that upstream history of the flow is also important of great importance. In short, the
turbulence is severely restriction the calculation of CFD in aerospace design process where the
inability to reliably predict turbulent flows with significant regions of separation3.
Presently, turbulence modelling based on Reynolds-Averaged Navier Stokes (RANS) equations is the
most common and practical approach for turbulence simulation. RANS are time-averaged
modification of Navier-Stokes equations and turbulence models are semi-empirical mathematical
relations that are used to predict the general effect of turbulence. The objective of turbulence
modelling is to develop equations that will predict the time-averaged velocity, pressure, and
temperature fields without calculating the complete turbulent flow pattern as a function of time.
Unfortunately, there is no single universally accepted turbulence model that works for all flows and
all regimes. Therefore, users have to use engineering judgement to choose from a number of different
alternatives sine the accuracy and effectiveness of each model varies depending on the application.
1 P., Moin and K., Mahesh, “Direct Numerical Simulation: A Tool in Turbulence Research”, Annual Rev. Fluid Mech.
1998. 30:539–78.
2 J., SODJA, “Turbulence models in CFD”, University of Ljubljana, Faculty for mathematics and physics,
2 Turbulence Essentials
2.1 Physical Perspectives
While the fluid elements are smooth and regular for Laminar flow pass a circle as shown in Figure
2.1 (a), the Turbulent elements are irregular in which various quantities show a random variation
with time and space as in Figure 2.1 (b). Therefore, a statically distinct averaging of values can be
distinguished. Also, because of this agitated motion in turbulent flow, the higher energy fluid
elements from the outer region of the flow are pumped close to surface. The diffusion rate of a scalar
quantity is usually greater in a turbulent flow than in a laminar. As the result, the frictional effects
are more severe for a turbulent flow. An outstanding feature of turbulent flow, as opposite to laminar
flow, is that molecules move in chaotic fashion along complex irregular path. The strong chaotic motion
causes the various layers of fluid to mix together intensively. Because of increase momentum and
energy exchange between the molecules and solid walls, turbulent flow leads at some conditions to
higher skin friction and heat transfer as compared to laminar case. It could be argued as that because
of the agitated motion in turbulent flow, the higher energy fluid elements from the outer regions of
flow are pumped close to surface, and hence, average flow velocity near solid surface is larger for
turbulent flow in comparison to laminar as depicted in Figure 2.3.
4 David C. Wilcox, ”Turbulence Modeling for CFD”, 1993, 1994 by DCW Industries, Inc.
9
excited length and time scales. The irregular nature of turbulence stands in contrast to laminar
motion, because the fluid was imagined to flow in smooth laminae, or layers. Virtually all flows of
practical engineering interest are turbulent. Turbulent flows always occur when the Reynolds
number is large. For slightly viscous fluids such as water and air, large Reynolds number corresponds
to anything stronger than a small swirl or a puff of wind. Careful analysis of solutions to the Navier-
Stokes equation, or more typically to its boundary-layer form, show that turbulence develops as an
instability of laminar flow. The features contributing to complexity of turbulence flow are;
Enhance Diffusion
Stability Criteria
Time Dependency
Eddies and Spectral Length
Non-Linearity effects
Separation
Enhanced Diffusion
Perhaps the most important feature of turbulence from an engineering point of view is its enhanced
diffusivity. Turbulent diffusion greatly enhances the transfer of mass, momentum and energy.
Apparent stresses often develop in turbulent flows that are several orders of magnitude larger than
in corresponding laminar flows.
Stability Criteria
To analyze the stability of laminar flows, virtually all methods begin by linearizing the equations of
motion. Although some degree of success can be achieved in predicting the onset of instabilities that
ultimately lead to turbulence with linear theories, the inherent nonlinearity of the Navier-Stokes
equation precludes a complete analytical description of the actual transition process, let alone the
fully-turbulent state. For a real (i.e., viscous) fluid, the instabilities result from interaction between
the Navier-Stokes equation's nonlinear inertial terms and viscous terms. The interaction is very
complex because it is rotational, fully 3-D and time dependent. The strongly rotational nature of
turbulence goes hand-in-hand with its three dimensionality. Vigorous stretching of vortex lines is
required to maintain the ever-present fluctuating vorticity in a turbulent flow. Vortex stretching is
absent in two-dimensional flows so that turbulence must be three dimensional. This inherent 3-D
means there are no satisfactory 2-D approximations and this is one of the reasons turbulence remains
the most noteworthy unsolved scientific problem of the twentieth century.
Time Dependency
The time-dependent nature of turbulence also contributes to its intractability. The additional
complexity goes beyond the introduction of an additional dimension. Turbulence is characterized by
random fluctuations thus obviating a deterministic approach to the problem. Rather, we must use
statistical methods. On the one hand, this aspect is not really a problem from the engineer's view.
Even if we had a complete time history of a turbulent flow, we would usually integrate the flow
properties of interest over time to extract time-averages. On the other hand, time averaging
operations lead to statistical correlations in the equations of motion that cannot be determined a
priori. This is the classical closure problem, which is the primary focus of this text. In principle, the
time-dependent, three-dimensional Navier-Stokes equation contains all of the physics of a given
turbulent flow. That this is true follows from the fact that turbulence is a continuum phenomenon. As
noted, "Even the smallest scales occurring in a turbulent flow are ordinarily far larger than any
molecular length scale”. Nevertheless, the smallest scales of turbulence are still extremely small.
10
They are generally many orders of magnitude smaller than the largest scales of turbulence, the latter
being of the same order of magnitude as the dimension of the object about which the fluid is flowing.
Furthermore, the ratio of smallest to largest scales decreases rapidly as the Reynolds number
increases. To make an accurate numerical simulation (i.e., a full time dependent 3-D solution) of a
turbulent flow, all physically relevant scales must be resolved. While more and more progress is
being made with such simulations, computers of the early 1990's have insufficient memory and speed
to solve any turbulent flow problem of practical interest.
Non-Linearity Effects
The nonlinearity of the Navier-Stokes equation leads to interactions between fluctuations of differing
wavelengths and directions. As discussed above, the wavelengths of the motion usually extend all the
way from a maximum comparable to the width of the flow to a minimum fixed by viscous dissipation
of energy. The main physical process that spreads the motion over a wide range of wavelengths is
vortex stretching. The turbulence gains energy if the vortex elements are primarily oriented in a
direction in which the mean velocity gradients can stretch them. Most importantly, wavelengths that
are not too small compared to the mean-flow width interact most strongly with the mean flow.
Consequently, the larger-scale turbulent motion carries most of the energy and is mainly responsible
for the enhanced diffusivity and attending stresses. In turn, the larger eddies randomly stretch the
vortex elements that comprise the smaller eddies, cascading energy to them5.
5 David C. Wilcox, “Turbulence Modeling for CFD”, Copyright © 1993, 1994 by DCW Industries.
11
2.3 Transition
The transition from laminar to turbulent is not subtle and consists of several processes, each subject
to intense research interests on their own right. The transition and stability of the laminar flow seem
to be dependent on a critical value of Reynolds number (Figure 2.4) as
ρ V x cr
Re cr 2100 Eq. 2.1
μ
The equations governing a turbulent flow are precisely the same as for laminar flow; however,
solutions is clearly more complicated in this region. That is due to the introduction of new terms and
issues with closure. Two general approach could be envisioned. First a more direct approach with
extreme spatial discretization to capture all the flow eddies near the wall. This is prohibitly CPU
intensive, even with current computing powers. To counter that, the use of some modeling (empirical
or otherwise) in the vicinity of wall region is advocated. Rest of this report organized as subsequent.
Since subject of Turbulence is very involved, therefore, the physical aspects of formulation, namely
Reynolds Stress formulation, and how to derive them represented first. Followed by different aspects
of Turbulence modeling.
The numerous factors contributing to transition from laminar to turbulent flow in a fluid. The intent
has been to provide general background information on the various transition phenomena rather
than to make a study of the problem in depth. Included are the effects on transition of such factors as
pressure gradient, surface temperature, Mach number, and two- and three-dimensional types of
surface roughness [A. L, Braslow]7. The effect of wall heat transfer was investigated by [Shadloo and
Hadjadj]8 through high-resolution DNS , among other factors. As in the case of adiabatic wall
(Figure 2.5-a), stream-wise-elongated streaks are visible in the laminar and transitional regions for
both heated (Figure 2.5-b) and cooled cases (Figure 2.5-c). However, they locally break down and
create turbulent spots further downstream for the heated case when compared with the adiabatic
wall. Therefore, the wall heating stabilizes the flow and postpones the transition.
Figure 2.5 Snapshots of the Stream-Wise Velocity Component for (A) Adiabatic Wall , (B) Heated
Wall, and (C) Cooled Wall
t 0 Δt
1
f
Δt f dt
t0
and f 0 , u i u i ui , ρ ρ ρ , p p p , T T T Eq. 3.1
It is required that Δt be large compared to the period of random fluctuations associated with
turbulence, but small with respect to the time constant for any slow variations in flow field associated
with ordinary unsteady flows. In conventional Reynolds decomposition, the randomly changing flow
variations are replaced by the time average plus fluctuations (see Figure 3.1) and would be discuss
later for Unsteady RANS (URANS) formulations. The resulting of decomposition of variables and time
averaging the entire equation, the continuity, momentum, and energy yield to
ρ
M ass : (ρ u j ρu j )
t x j
M omentum : (ρ u i ρu i ) (ρ u i u j u i ρu j )
t x j
p
( τ u ρu ρ u u ρu u )
x x ij j i i j i j
i j
(c ρT c ρ T)
p p
Energy : ρc p u j T
t x
j
p p p T
uj u j k ρc p Tu j c p ρTu j
t x j x j x j x j
ui u
where τij τij i
x j x j
u u j 2 u k 1 i j
τij μ i δ ij δ ij
3 x
and
jx x 0 i j
i k
Eq. 3.2
15
Figure 3.1 Relationship between averaged variables; a) Steady flow b) Un-Steady flow
1
U
U( ) d
0
, U U u Eq. 3.3
Note that the dependent variables are now not only a function of space but also a function of time as
well. Be advised that this concept of time, is different from the time step (t) of mean value. In essence
if you solve your equations with one global time step, which is used in every cell, and if value of the
time step is small enough then you will be able to capture fluctuations, or unsteady behavior in the
MEAN quantities. In other words your solution is time accurate. Steady RANS, or RANS, marches the
solution with a local optimized time step for each cell, and hence is not time accurate, you will get a
faster solution, and it will be steady state. For URANS we have:
Even if the results from URANS are unsteady, one is often interested only in time averaged flow as
denoted as <Ū>, which means that we can decompose the results from URANS as a time averaged
part <Ū>, a resolved fluctuation uˈ, and the modeled turbulent fluctuation, u̎ , i.e.
overbar has been dropped from the mean values. This convention will be used throughout here for
obvious reasons.
Closure
Problem
Eq. 3.6
Eq. 3.6 presents the fundamental problem of turbulence. In order to compute all the mean-flow
properties of the turbulent flow we need a reasonably accurate way to compute the Reynolds stress
uʹiuʹj. This is the fundamental reason for the need of the turbulence models. The complete set of the
Reynolds Averaged Navier-Stokes (RANS) equations are not presented here because they can be
found from many references9. Scalar transport equations are also needed, for example to describe
the transport of the concentration of species or the mass fraction of species. Their exact formulation
can be found in10. For incompressible flow with constant properties and no body force the
momentum and energy
9 P. Kaurinkoski and A. Hellsten, FINFLO: The Parallel Multi-Block Flow Solver, Report A-17, Laboratory of
Aerodynamics, Helsinki University of Technology, Espoo, Finland, 1998.
10 P. Kaurinkoski, “Development of an Equation of State for an Arbitrary Mixture of Thermally Perfect Gases to the
FINFLO flow solver”, Report No B-48, Series B, Laboratory of Aerodynamics, Helsinki University of Technology,
Finland, 1995.
17
∂ ̅
∂T ∂
−(∇. q)Lam = (k ) and − (∇. q)Turb = (ρC𝑝 ̅̅̅̅̅̅
T ′ 𝑢𝑗′ )
∂xj ∂xj ∂xj
Eq. 3.7
The Reynolds equation cannot be solved in the form given because the new apparent turbulent
stresses and heat-flux quantities must be viewed as new unknowns. To proceed further, we need to
find additional equations involving the new unknowns to make assumption regarding the
relationship between the new apparent turbulence quantities and time-mean flow variables. This is
known as closure problem which is most commonly handled through turbulence modeling to be
discussed next.
u i u j 2 u k
ρui uj μ T δij μ T
ρ κ κ ui ui /2 Eq. 3.8
x x 3
j i x k
Where μT is the turbulent viscosity, κ is the kinematic energy of turbulence are key modeling aspects
associated with so called turbulence closure problem. Reynolds stresses with deformation rate, and
are related to viscosity, mean rate of deformation, and turbulent kinetic energy with Boussinesq’s
proposal expressed can be used to calculate Reynold’s stresses in the final step of turbulence
modelling. It is seen from this equation that the Reynold’s stresses are considered proportional to
the dissipation rate reduced by the eddy turbulent kinetic energy. (The Kronecker delta ensures that
the normal Reynolds stresses are each appropriately accounted for). It can also be seen that the
kinetic energy allocates an equal third for each normal stress component (isotropic assumption).
This is the reason for the inherent inaccuracy of the κ-ε model, making it incapable of describing
anisotropic flow. Other scalar flow properties such as mass and heat can also be modelled using time-
averaged values.
To obtain mathematical closure, the Reynolds stress terms must be related to mean flow properties
either empirically or through flow model which allows calculation of this relationship. References
are also sometimes made to the order of the closure. According to this terminology, a 1st order
closure evaluates the Reynolds stresses through functions of the mean velocity and geometry alone,
as the case above (linear). A 2nd order closure employs a solution to a modeled form of transport
partial differential equations for one or more of the characteristics of turbulence (non-linear). More
complicated quadratic or cubic closures are also available through literature.
Mixing Length
0-Eqaution
Baldwin-Lomax
Spalart-
1-Equation
1st - order Almars Standard
(linear)
Reynolds Realizable
Average 2-Equation
(RANS) SST
Reynolds
Turbulence Stress κ-ω models
DES 2nd - order Standard
Models
(non-linear) (Wicox)
Algebric Stress
DNS
μ T ρu T l where uT
2
3
1
3
ui ui Eq. 3.9
It can be seen that the turbulent viscosity is a product of density and two new variables representing
turbulent velocity and turbulent length scale. Turbulent velocity uT can be described as the typical
velocity occurring in the largest eddies and can also be related to the same eddies. Turbulent kinetic
energy according to turbulent length scale is the average length. The new variables, uT and l form the
11Anderson, Dale A; Tannehill, John C; Plecher Richard H; 1984,”Computational Fluid Mechanics and Heat
Transfer”, Hemisphere Publishing Corporation.
19
basis for the “two-equation‟ k-ε turbulence model, meaning that in addition to the RANS equations,
two more equations are required to solve for turbulent velocity and turbulent length using the model.
The length scale l (for large eddies) is used in the k-ε model to define the length scale ε (for small
eddies), for which a transport equation is used in the model, and represents the dissipation of the
turbulent kinetic energy. The dissipation is expressed as:
κ 2/3 μ T ui u j
ε as related to viscosity ε Eq. 3.10
l ρ x κ x κ
As seen in the second expression for ε, dissipation of the turbulent kinetic energy κ is proportional to
the rate of deformation of eddies. Other scalar flow properties such as mass and heat can also be
modelled using time-averaged values. Similar to the turbulent momentum transport’s
proportionality to average velocity gradients, turbulent scalar transport is proportional to mean
scalar value gradients and can be expressed as
ρu i Γ T (9.16) Eq. 3.11
x i
Where ГT refers to turbulent (eddy) diffusivity. As can be seen from above expression, the turbulent
scalar property transport occurs with the same mechanism as in transport of momentum (mixing of
eddies, represented by ГT). For this reason, it can be assumed by Reynold’s analogy that the value of
ГT is similar to μT, the turbulent viscosity. The ratio of μT to ГT is defined as the Prandtl/Schmidt
number σT, and has a value which is normally constant with a value around unity. The next section
presents the κ-ε model and the two extra k and ε transport equations (PDEs) for closing the system
of time-averaged RANS equations. The model is based on the mechanisms causing changes to
turbulent kinetic energy (i.e. turbulent viscosity and velocity fluctuations).
u
μ T ρu T l or μ T ρl 2 Eq. 3.12
y
Where l a mixing length can be thought of as a transverse distance over which the particles maintain
their original momentum. For 3D thin shear layers, Prandtl’s equation is usually interpreted as:
12 David C. Wilcox, “Turbulence Modeling for CFD”, Copyright © 1993, 1994 by DCW Industries, Inc.
20
1/2
2 u
2
w
2
This formula treats the turbulent viscosity as a scalar and gives qualitively correct results, specially
near the wall. The evaluation of l in the mixing length model varies with the type of flow being
considered, wall boundary layer, jet, wakes, etc. For flow along a solid surface (internal or external),
good results are observed by evaluating l according to
/A y( τ w /ρ w )1/2
linner κy(1 e y ) , louter C1δ and y Eq. 3.14
νw
Where linner predicts the inner region close to wall and Iouter exceeds linner. The κ is the von Karman
constant as 0.41 and A+ is the damping constant usually set to 26. The expression for linner is
reasonable for producing the inner law-of-the-wall region of turbulent flow and louter produces the
outer “wake-like” region. These two zones are depicted in Figure 3.7 which shows a typical velocity
distribution for an incompressible turbulent boundary layer on a smooth impermeable plate using
“law-of-the-wall” coordinates.
3.4.1.3 Baldwin-Lomax
[Baldwin & Lomax -1978] proposed an alternative algebraic model to eliminate some of the difficulty
in defining a turbulence length scale from the shear-layer thickness. It is a two-layer algebraic 0-
equation model which gives the eddy viscosity μt as a function of the local boundary layer velocity
profile. The model is suitable for high-speed flows with thin attached boundary-layers, typically
present in aerospace and turbo machinery applications. It is also commonly used in quick design
iterations where robustness is more important than capturing all details of the flow physics. The
Baldwin-Lomax model is not suitable for cases with large separated regions and significant
curvature/rotation effects.
One-Equation Model
While employing a much simpler approach than two-equation or second-order closure models, one-
equation models have been somewhat unpopular and have not showed a great deal of success. One
notable exception was the model formulated by [Bradshaw, Ferris, and Atwell -1967], whose model
was tested against the best experimental data of the day at the 1968 Stanford Conference on
Computation and Turbulent Boundary Layers. There has been some renewed interest in the last
several years due to the ease with which one-equation models can be solved numerically, relative to
more complex two-equation or second-order closure models. An obvious shortcoming of algebraic
methods is that μT and uT is zero at the center of for example pipe line cases. The mixing-length model
can be fixed up to overcome this using
μ T C k ρl (κ)1/2
For 2D incompressible thin - shear layer Eq. 3.15
Dκ μ T κ
2
u C ρ(κ) 3/2
μ μ T D
Dt y Prκ y y l
21
Where Prκ is defined as Prandtl number for turbulence kinetic energy (≃1.0) and CD ≃ 0.164.
3.4.2.1 K Model
The complete derivation is obtained in two equation model.
D ν
P ε ν T
x Eq. 3.16
Dt x j σ j
On the left-hand side is the quantity’s material derivative. On the right-hand side are one or more
production terms, a dissipation term, a diffusion term dependent on molecular viscosity, and another
13Alessandro Corsini, Giovanni Delibra, and Anthony G. Sheard, “A Critical Review of Computational Methods
and Their Application in Industrial Fan Design”, Hindawi Publishing Corporation, ISRN Mechanical Engineering,
Volume 2013, Article ID 625175.
22
given as the turbulent viscosity’s function, corrected using the Prandtl number 𝜎𝜙. The primary
difference between the 𝑘-𝜀 and 𝑘-𝜔 models is the different trend of 𝜀 and 𝜔 at the wall and the
definition of the wall boundary conditions for the same variables. When one studies normalized
values of 𝜀 and 𝜔 for an attached flow, it is evident that 𝜔 is less dependent on the Reynolds number
than 𝜀 in the wall’s near vicinity.
There is a general consensus within the computational fluid dynamics community that the 𝑘-𝜀 model
better reproduces the energy cascade of large-scale structures in the main flow core, whilst the 𝑘-𝜔
model performs better near the wall14. A realization that 𝑘-𝜀 models perform better in the
main flow whilst the 𝑘-𝜔 models perform better near the wall leads to the natural conclusion that,
ideally, one would use the two models in combination. It was observed that it is possible to combine
𝑘-𝜀 and 𝑘-𝜔 models as one can reformulate every two-equation model into every other by changing
model coefficients. This realization has enabled engineers to formulate the 𝑘-𝜔 shear stress transport
(𝑘-𝜔 SST) model that solves the equation for 𝜔 near the wall and 𝜀 elsewhere. The use of two-
equation models has become established within the industrial community.
𝛛(𝛒𝛋) 𝛛(𝛒𝐮̅ 𝐢 𝛋) 𝛛 𝝁𝑻 𝝏𝜿 ̅𝐢
𝛛𝐮
+ = [ ̅̅̅̅̅̅
] − 𝛒𝐮 ′ 𝐮′ .
𝐢 𝐣 − 𝛒.
⏟ 𝛆
⏟𝛛𝐭 ⏟ 𝛛𝐱 𝐢 𝛛𝐱 𝐢 𝝈𝜿 𝝏𝒙𝒊
⏟ ⏟ 𝛛𝐱 𝐣
𝟓
𝟏 𝟐 𝟑 𝟒
Eq. 3.18
Where is a constant turbulent Schmidt number for κ. The terms (1)-(5) of the turbulent kinetic energy
k transport equation, can be interpreted as the following:
14 Alessandro Corsini, Giovanni Delibra, and Anthony G. Sheard, “A Critical Review of Computational Methods
and Their Application in Industrial Fan Design”, Hindawi Publishing Corporation, ISRN Mechanical Engineering,
Volume 2013, Article ID 625175.
15 Tennekes, H.; Lumley, J. L. “A First Course in Turbulence”, MIT Press, Cambridge, MA (1972).
23
The above equation for transport of κ along with the equation for transport of ε, constitute the two
additional transport equations to be solved in addition to the RANS equations in the κ-ε turbulence
model. The next section presents the equation for transport of ε.
𝛛 𝛛 𝛛 𝛍𝐓 𝛛𝛆 𝛆 ̅𝐢
𝛛𝐮 𝛆𝟐
+ (𝛒𝐮
̅ 𝐢 𝛆) = [ ] + 𝐂𝟏 (−𝛒𝐮𝐮. ) − 𝐂𝟐 𝛒
⏟ ⏟
𝛛𝐭 𝛛𝐱 𝐢 𝛛𝐱 𝐢 𝛔𝛆 𝛛𝐱 𝐢
⏟ 𝛋
⏟ 𝛛𝐱 𝐢 ⏟ 𝛋
𝟏 𝟐 𝟑 𝟒 𝟓
Eq. 3.19
Similar to the transport equation for κ, the transport equation for ε includes the terms 1 - 5:
And C1, C2, σε are constants of the ε equation (see Table 3.1). Alternatively, we inscribe both
equations in a more compact form, using Lagrangian Derivatives terms D/Dt :
Dκ 1 μT κ μ T 2
μ Sij ε
Dt x j ρ σk x j ρ
Dε ε μ 1 μT ε
(C1 T Sij2 C 2 ε) μ Eq. 3.20
Dt κ ρ x j ρ σε x j
κ2 1 u u j
where μ T ρCμ and Sij i
ε 2 x j x i
Cμ ρ(κ) 2
μT Eq. 3.21
ε
Despite the enthusiasm which is noted from time to time over two equation model, it is perhaps
appropriate to point out two major restriction on this type of models. Since the two equation model
basically turbulent viscosity models which assumes that the Boussinesq approximation holds.
Therefore, its validates depends to Boussinesq approximation. In algebraic methods, μT is a local
function whereas in two equation model is a more general and complex functioned governing by two
additional PDEs. The second shortcoming is the need to make assumptions in evaluating the various
24
terms in model transport equation especially third order turbulent correlations16. The same short
coming plagues all other higher order closure attempts, so there is no magic bullet.
16Anderson, Dale A; Tannehill, John C; Plecher Richard H; 1984,”Computational Fluid Mechanics and Heat
Transfer”, Hemisphere Publishing Corporation.
Florian R. Menter, “Improved Two-Equation k-ωTurbulence Models for Aerodynamic Flows”, NASA Technical
17
properties of the scheme. A second problem is that ε does not go to zero at a nonslip surface. There
is a significant number of alternative models that have been developed to overcome the shortcomings
of the κ-ε model. One of the most successful, with respect to both, accuracy and robustness, is the κ-
ω model of Wilcox18. It solves one equation for the turbulent kinetic energy k and a second equation
for the specific turbulent dissipation rate (or turbulence frequency) ω. The model performs
significantly better under adverse pressure-gradient conditions than the κ-ε model although it is the
authors experience that an even higher sensitivity to strong adverse pressure-gradients would be
desirable. Another strong-point of the model is the simplicity of its formulation in the viscous
sublayer. The model does not employ damping functions and has straightforward Dirichlet boundary
conditions. This leads to significant advantages in numerical stability.
18 W'llcox, D. C., "Reassessment of the Scale-Determining Equation for Advanced Turbulence Models," AIAA
Journal, Vol.26, Nov. 1988, pp.1299-1310.
19 Menter, F.R., “Two-equation eddy-viscosity turbulence models for engineering applications”, AIAA-Journal.,
Dκ ρκ κ ui
ρP β*ρωκ μ σ κ with P τ ij
Dt x j ω x j xj
Eq. 3.22
Dω γω ρκ ω ρσ d κ ω
P - βρω 2 μ σ ω
Dt κ x j ω x j ω x j x j
For recommendations for the values of the different parameters, see Wilcox (2008)21.
21 Wikipedia.
22 Menter, F.R., “Multiscale model for turbulent flows”, In 24th Fluid Dynamics Conference. American Institute of
Aeronautics and Astronautics, 1993.
23 Menter, F.R., “Two-equation eddy-viscosity turbulence models for engineering applications”, AIAA-Journal.,
𝛛(𝛒𝛋) 𝛛 𝛛 𝛍𝐓
+ (𝛒𝐮
̅ 𝐢 𝛋) = [(𝛍 + ) 𝛁𝛋 ] + 𝐏
⏟ 𝛃∗ 𝛒𝛋𝛚
𝛋−⏟
⏟𝛛𝐭 𝛛𝐱 𝐢
⏟ 𝛛𝐱 𝐢
⏟ 𝛔𝛋
𝟒 𝟓
𝟏 𝟐 𝟑
̅ 𝐢 𝛛𝐮
𝛛𝐮 ̅𝐢 𝟐 ̅𝐢
𝛛𝐮
𝐰𝐡𝐞𝐫𝐞 𝐏𝛋 = (𝟐𝛍𝐓 . − 𝛒𝛋 𝛅 )
𝛛𝐱 𝐣 𝛛𝐱 𝐣 𝟑 𝛛𝐱 𝐣 𝐢𝐣
Eq. 3.23
The terms (1) - (5) in above expression, the turbulent kinetic energy k transport equation for the SST
Omega turbulence model, can be interpreted as the following:
𝛛(𝛒𝛚) 𝛛(𝛒𝐮̅ 𝐢 𝛚) 𝛛 𝛍𝐓
+ = [(𝛍 + ) 𝛁𝛚] +
⏟𝛛𝐭 ⏟ 𝛛𝐱 𝐢 𝛛𝐱
⏟𝐢 𝛔𝛚,𝟏
𝟏 𝟐 𝟑
̅ 𝐢 𝛛𝐮
𝛛𝐮 ̅𝐢 𝟐 ̅𝐢
𝛛𝐮 𝛒 𝛛𝛋 𝛛𝛚
𝛄𝟐 (𝟐𝛒 . − 𝛒𝛚 𝛃𝟐 𝛒𝛚𝟐 + 𝟐
𝛅𝐢𝐣 ) − ⏟
⏟ 𝛛𝐱 𝐣 𝛛𝐱 𝐣 𝟑 𝛛𝐱 𝐣 𝛔
⏟𝛚,𝟐 𝛛𝐱 𝛋 𝛛𝐱 𝜿
𝟓
𝟒 𝟔
Eq. 3.24
The general description for each of the terms in (1) to (6) are the usual terms for accumulation,
convection, diffusion, production, and dissipation of ω. The last term (6) is called a “cross-diffusion‟
term, an additional source term, and has a role in the transition of the modelling from ε to ω. The
constants for the Mentor SST κ-ω turbulence model are
listed in β* β2 σκ σω,1 σω,2 Υ2
Table 3.2. Additional modifications have been made to 0.09 0.083 1.0 2.0 1.17 0.44
the model for performance optimization. There are
blending functions added to improve the numerical Table 3.2 Constant values for SST κ-ω
stability and make a smoother transition between the turbulence model equations
two models. There have also been limiting functions
made to control the eddy viscosity in wake region and adverse pressure flows25. The κ-ω model is
similar to κ-ε, but it solves for ω, the specific rate of dissipation of kinetic energy. It also uses wall
functions and therefore has comparable memory requirements. Additionally, it has more difficulty
converging and is quite sensitive to the initial guess at the solution. Hence, the κ-ε model is often used
first to find an initial condition for solving the κ-ω model. The κ-ω model is useful in many cases
25Versteeg, H K; Malalasekera, W.”An Introduction to Computational Fluid Dynamics, The Finite Volume
Method”, Second edition, Pearson Education Limited, Essex, England (2007).
28
where the κ-ε model is not accurate, such as internal flows, flows that exhibit strong curvature,
separated flows, and jets26.
26 Walter Frei, “Which Turbulence Model Should I Choose for My CFD Application? “, COMSOL Blog, September
16, 2013.
27 D. C. Eleni, Tsavalos I. Athanasios and Margaris P. Dionissios, “Evaluation of the turbulence models for the
simulation of the flow over a National Advisory Committee for Aeronautics (NACA) 0012 airfoil”, Journal of
Mechanical Engineering Research, March 2012.
28 Abbott IH, Von Doenhoff AE. “Theory of Wing Sections”. Dover Publishing, New York.
29
sufficient to establish a grid independent solution. The domain height was set to approximately 20
chord lengths, and the height of the first cell adjacent to the surface was set to 10-5, corresponding to
a maximum y+ of approximately 0.2. A y+ of this size should be sufficient to properly resolve the inner
parts of the boundary layer. In order to include the transition effects in the aerodynamic coefficients
calculation and get accurate results for the drag coefficient, a new method was used. The transition
point from laminar to turbulent flow on the airfoil was determined and the computational mesh was
split in two regions, a laminar and a turbulent region. To calculate the transition point the following
procedure was used. A random value for the transition point (xtr) was chosen and the computational
domain was split at that point with a perpendicular line. The problem was simulated by defining the
left region as laminar and the right as turbulent zone.
Figure 3.4 Comparison between experimental data [Abbott et al.]and three different turbulent
models simulation results of the Lift Coefficient for NACA 0012 Airfoil
component of the net force acting normal to the incoming flow stream is known as the lift force and
the component of the net force acting parallel to the incoming flow stream is known as the drag force.
The curves of the lift and the drag coefficient are shown for various angles of attack, computed with
three turbulence models and compared with experimental data. Figure 3.4 shows that at low angles
of attack, the dimensionless lift coefficient increased linearly with angle of attack. Flow was attached
to the airfoil throughout this regime. At an angle of attack of roughly 15 to 16°, the flow on the upper
30
surface of the airfoil began to separate and a condition known as stall began to develop. All three
models had a good agreement with the experimental data at angles of attack from -10 to 10° and the
same behavior at all angles of attack until stall. It was obvious that the Spalart-Allmaras turbulence
model had the same behavior with the experimental data as well as after stall angle.
Near stall, disagreement between the data was shown. The lift coefficient peaked and the drag
coefficient increased as stall increased. The predicted drag coefficients were higher than the
experimental data (Figure 3.5). This over prediction of drag was expected since the actual airfoil has
laminar flow over the forward half. The turbulence models cannot calculate the transition point from
laminar to turbulent and consider that the boundary layer is turbulent throughout its length. From
theory, the turbulent boundary layer carries more energy and is much greater than at the viscous
boundary layer, which carries less energy. The computational results must be compared with
experimental data of a fully turbulent boundary layer. This was done only for CD as CL is less sensitive
to the transition point.
[ Johansen]29 contained experimental data of CD for the NACA 0012 airfoil and Re = 3×106, where the
boundary layer formed around the airfoil is fully turbulent. Figure 3.5 shows the curves of CD for
various angles of attack, compared with experimental data for fully turbulent boundary layer30. The
values of from the three turbulence
models were very close to
experimental data for the fully
turbulent boundary layer. The most
accurate model was the κ-ω SST
model, next came the Spalart-
Allmaras, and latest in precision
was the Realizable κ-ε .
In order to get more accurate results,
the computational domain could be
split into two different domains to run
mixed laminar and turbulent flow. The
disadvantages of this approach were
that the accuracy of simulations
depends on the ability to accurately
guess the transition location, and a
new grid must be generated if the
transition point had to change
[Silisteanu-Botez]31. If the transition Figure 3.5 Comparison Between Experimental Data for
point is known, the grid can easily be Transitional Boundary Layer and Different Turbulent Models
split in two with a vertical line that on the Drag Coefficient of NACA-0012 Airfoil
passes through this point and then
laminar and turbulent zones are defined. The results of this method at angle of attack a=0 and
operating at Re = 1×106, 2×106, 3×106, 4×106 and 5×106. Initially, was calculated for a fully turbulent
boundary layer and compared with CD experimental data from NASA [McCroskey]32. Then,
29 Johansen J. “Prediction of Laminar/Turbulent Transition in Airfoil Flows”. RISE National Laboratory, Roskilde,
Denmark, 1997.
30 Comparison Between Different Turbulent Models and Experimental Data obtained by [Abbott & Von
Doenhoff] and [Johansen] for Transitional Boundary Layer on the Drag Coefficient of NACA-0012 Airfoil.
31 Silisteanu PD, Botez RM. “Transition flow occurrence estimation new method”. 48th AIAA Aerospace Science
Research and Technology Activity, Nasa Technical Memorandum, 42: 285-330, 1987.
31
simulations were made with the split grid for the five Reynolds numbers. The computational results
for the fully turbulent boundary layer agreed very well with the corresponding experimental data.
The discrepancy between the Drag Coefficient and experimental data from [McCroskey] for fully
turbulent boundary layer was up to 5.6%. On the other hand, the comparison between the simulation
results with the split grid and the experimental data from [McCroskey] for transitional boundary
layer showed an excellent agreement, with maximum error of about 3.6%. It was also observed that
as the Reynolds number increased, the Drag Coefficient decreased. When the boundary layer was
fully turbulent the reduction of was more intense and when there was a transition from laminar to
turbulent, was reduced to a much
lower rate. It is worth noting that
this process of calculating the
transition point is quite simple
when the angle of attack is zero
because the flow is symmetric and
the transition point is the same
above and below the airfoil. At
nonzero angles of attack the process
is more complicated because
transition points are different for
the upper and lower surface of the
airfoil. With different AoA, the
pressure on the lower surface of the
airfoil was greater than that of the
incoming flow stream and as a
result it effectively “pushed” the
airfoil upward, normal to the
incoming flow stream. On the other
hand, the components of the
pressure distribution parallel to the
incoming flow stream tended to
slow the velocity of the incoming
flow relative to the airfoil, as do the
viscous stresses.
Contours of velocity components at
angles of attack 9 and 16° are also
shown in Figure 3.6. The trailing
edge stagnation point moved
Figure 3.6 Contours of velocity magnitude at 9° (Top) and 16°
slightly forward on the airfoil at low
(Bottom) AoA with the Spalart-Allmaras turbulence model
angles of attack and it jumped
rapidly to leading edge at stall angle.
A stagnation point is a point in a flow field where the local velocity of the fluid is zero. The upper
surface of the airfoil experienced a higher velocity compared to the lower surface. That was expected
from the pressure distribution. As the angle of attack increased the upper surface velocity was much
higher than the velocity of the lower surface. For further information, please consult the [Eleni et
al.]33.
33D. C. Eleni, Tsavalos I. Athanasios and Margaris P. Dionissios, “Evaluation of the turbulence models for the
simulation of the flow over a National Advisory Committee for Aeronautics (NACA) 0012 airfoil”, Journal of
Mechanical Engineering Research, March 2012.
32
3.4.4.3 Conclusions
This paper showed the behavior of the 4-digit symmetric airfoil NACA 0012 at various angles of
attack. The most appropriate turbulence model for these simulations was the κ-ω SST two-equation
model, which had a good agreement with the published experimental data of other investigators for
a wider range of angles of attack. The predicted drag coefficients were higher than the existing
experimental data from reliable sources. This over prediction of drag was expected since the actual
airfoil has laminar flow over the forward half. The computational results from the three turbulence
models were compared with experimental data where the boundary layer formed around the airfoil
is fully turbulent and they agreed well. Afterwards, the transition point from laminar to turbulent
regime was predicted, the computational grid split in two regions, a laminar and a turbulent region,
and then new simulations were realized. By this method, the computational results agreed very well
with corresponding experimental data34.
Variation in velocity etc. are predominantly normal to the wall, leading to one dimensional
behavior.
Effects of pressure gradients and body forces are negligible, leading to uniform shear stress
in the layer.
Shear stress and velocity vectors are aligned and unidirectional through the layer.
A balance exist between turbulence energy production and dissipation.
There is a linear variation of turbulence length scales.
For flow along solid surfaces (internal or external), a good approximation is observed by evaluating
the expressions for linner is responsible for producing the inner law of the wall region of turbulent flow
and louter produces the outer wake-like region. These two zones are illustrated in Figure 3.7 which
depicts a typical velocity distribution for an incompressible turbulent boundary layer on a smooth
impermeable plate.
34 See Previous.
33
Figure 3.7 Zones in Turbulent B. L. for a typical Incompressible flow over a smooth flat plate
2 - Low-Reynolds-Number Method
Resolves the details of the boundary layer profile by using very small mesh length scales in the
direction normal to the wall (very thin inflation layers). Note that the low-Re method does not refer
to the device Reynolds number, but to the turbulent Reynolds number, which is low in the viscous
sublayer. This method can therefore be used even in simulations with very high device Reynolds
numbers, as long as the viscous sublayer has been resolved. The computations are extended through
the viscosity-affected sublayer close to the wall. The low-Re approach requires a very fine mesh in
the near-wall zone and correspondingly large number of nodes. Computer-storage and run-time
requirements are higher than those of the wall-function approach and care must be taken to ensure
good numerical resolution in the near-wall region to capture the rapid variation in variables. To
reduce the resolution requirements, an automatic wall treatment was developed which allows a
gradual switch between wall functions and low-Reynolds number grids, without a loss in accuracy.
34
3.4.5.2 Case Study - Wall-modelling strategies in Large Eddy Simulation of separated high-
Reynolds-number flows
If ReL (say, that based on the boundary layer thickness) exceeds roughly 105, the resource
requirements are entirely dominated by the need to resolve the near-wall region (the inner layer),
with the number of nodes rising roughly as N ~ Re2.5L. Simulations for practical configurations would
not, in most circumstances, be undertaken today with meshes exceeding 10-50M nodes,
corresponding to ReL ~ 5*105, yet this is still a very modest Reynolds number in practice [Tessicini,
and Leschziner]35. The general approach taken in recent years towards addressing the problem of
near-wall resolution has been to combine RANS modelling near the wall with LES in the outer flow.
The key premise underpinning this strategy is that it should allow near-wall numerical cells to be
used that have far higher aspect ratios than those required by LES, typically 500-1000, relative to 50
in wall-resolving LES. Substantial savings could thus be made by using much coarser stream wise and
span wise meshes than are dictated by the LES constraints. Efforts currently have focused on two
particular methods, one being a hybrid LES-RANS scheme and the other being a two-layer zonal
schemes. The difference between them is explained by reference to Figure 3.9, which conveys the
manner by which the LES and RANS regions communicate numerically.
The hybrid method uses a single computational domain. Within a predefined layer near the wall, that
can be prescribed in terms of y+, RANS equations are solved using one-equation or two-equation
eddy-viscosity models that are dynamically adjusted so to comply with continuity of eddy viscosity
across the interface, νRANS= νLES, beyond which a LES sub-grid-scale model is used. To achieve this
compatibility, the RANS model coefficients at the interface is determined by comparison of the RANS
viscosity, containing the RANS-determined turbulence energy and dissipation rate at the interface,
to the LES viscosity at the interface. The variation of the coefficients from the interface to the wall is
then prescribed analytically, based on observations derived from a-priori wall-resolved LES
performed in channel flows. The zonal method uses two overlapping grids across the near-wall layer.
The LES grid extends to the wall, but is relatively coarse, maintaining cell-aspect-ratio constraints
35F. Tessicini, M.A. Leschziner, “Wall-modelling strategies in large eddy simulation of separated high- Reynolds-
number flows”.
35
appropriate to LES. Within the near-wall layer, a separate grid is inserted, which is refined towards
the wall, typically to a wall-nearest node located at y+=O(1). Within that layer, parabolized RANS
equations are solved for the wall-parallel-velocity components, using a simple algebraic turbulence
model - for example, a mixing-length model.
Figure 3.9 Schematics of hybrid LES-RANS scheme (upper) and two-layer zonal scheme (lower)
36
The Boussinesq eddy-viscosity approximation assumes the principal axes of the Reynolds-stress
tensor, τij, are coincident with those of the mean strain-rate tensor, Sij, at all points in a turbulent flow.
This is the analog of Stokes approximation for laminar flows. The coefficient of proportionality
between τij and Sij is the eddy viscosity, μT, which is linear. Unlike the molecular viscosity which is a
property of the fluid, the eddy viscosity depends upon many details of the flow under consideration.
It is affected by the shape and nature (e.g., roughness height) of any solid boundaries, freestream
turbulence intensity, and, perhaps most significantly, flow history effects. Experimental evidence
indicates that flow history effects on τij often persist for long distances in a turbulent flow, thus
casting doubt on the validity of a simple linear relationship between τij and Sij. Next, we outline
several flows for which the Boussinesq approximation yields a completely unsatisfactory description.
ui u j 2
ρu i uj μ T δ ij ρ κ or
x x 3
j i
Eq. 4.1
2 1 u u j
ρu i uj TSij δ ij κ where Sij i
3 2 x j x i
The above isotropic relation assumes that the principal axis of the Reynolds stress tensor S͞ ij coincides
37
with that of the mean strain rate. The standard κ-ε model does not take into account the anisotropic
effects and fails to represent the complex interaction mechanisms between Reynolds stresses and
the mean velocity field. For example, the linear model fails to mimic the effects related to streamline
curvature, secondary motion, or flow with extra strain rates. These anisotropic effects can be
predicted by introducing a nonlinear expression for the Reynolds stresses as given in the following
expression36:
𝟐 ̅𝛋
𝛒 ̅𝛋
𝛒
̅̅̅̅̅̅̅
−𝛒𝐮 ′ 𝐮′ = − 𝛒
𝐢 𝐣 ̅ 𝛋𝛅 𝐢𝐣 + 𝟐𝛍 ̅
𝐒
𝐓 𝐢𝐣 − 𝐁 ̅
𝐒 ̅
𝐒 𝛅 − 𝐂 𝐒 𝐒 −
⏟𝟑 ⏟𝛚𝟐 𝐦𝐧 𝐧𝐦 𝐢𝐣 𝛚𝟐 𝐢𝐤 𝐤𝐣
𝐁𝐨𝐮𝐬𝐬𝐢𝐧𝐞𝐬𝐪 𝑵𝒐𝒏−𝑳𝒊𝒏𝒆𝒂𝒓 𝑻𝒓𝒆𝒎 𝟏
̅𝛋
𝛒 ̅𝛋
𝛒 ̅𝛋
𝛒
𝐃 𝟐 (𝐒̅𝐢𝐤 𝛀
̅ 𝐤𝐣 + 𝐒̅𝐣𝐤 𝛀
̅ 𝐤𝐢 ) − 𝐅 ̅ 𝐦𝐧 𝛀
𝛀 ̅ 𝐧𝐦 𝛅𝐢𝐣 − 𝐆 ̅ 𝛀
𝛀 ̅
⏟𝛚 𝛚𝟐 𝛚𝟐 𝐢𝐤 𝐤𝐣
𝑵𝒐𝒏−𝑳𝒊𝒏𝒆𝒂𝒓 𝑻𝒆𝒓𝒎 𝟐
𝛋 𝛋𝟑 ̅ 𝛛𝐮
𝟏 𝛛𝐮 ̅ ̅𝐣
̅ 𝐢 𝛛𝐮
𝟏 𝛛𝐮
𝐰𝐡𝐞𝐫𝐞 ≈ , ̅𝐢𝐣 = ( 𝐢 + 𝐣 )
𝐒 ̅ 𝐢𝐣 =
𝐚𝐧𝐝 𝛀 ( − )
𝛚𝟐 𝛆𝟐 𝟐 𝛛𝐱 𝐣 𝛛𝐱 𝐢 𝟐 𝛛𝐱 𝐣 𝛛𝐱 𝐢
Eq. 4.2
Where B, C, D, F, and G are closure coefficients. These coefficients of the non-linear terms should be
carefully determined because they are expected to influence the physical accuracy and numerical
performance of the model. Here, the coefficients are adjusted through the consideration of the
anisotropy in simple shear flows detailed by [Champagne et al.]37 and [Harris et al.]38 Solving this
equation is a daunting task, but there are assumptions (as the case with all turbulence models) which
can be made, depending to the case. We don’t get into different modeling but for an excellent
discussion, readers should refer to39. (See Eq. 4.2).
Case Study – 3D Simulation of Flow Past a Cylinder using Nonlinear Turbulence Model
The flow past a cylinder of circular cross section has been the subject of interest for industrial
researchers as well as scientists, because of its wide range of applications. To cite a few examples,
flow in bridge piers, chimney stacks, and tower structures in civil engineering; electrodes in chemical
engineering; nuclear fuel rods in the atomic field and heat exchanger tubes in thermal engineering,
etc., fall under this subject of study. Although the geometry is simple, the flow has complicated
features such as stagnation points, laminar boundary-layer separation, turbulent shear layers,
periodic vortex shedding, and wakes. Even though there is much literature available on numerical
simulation of laminar flow past a two-dimensional circular cylinder at low Reynolds number, a focus
on practical high Reynolds numbers is less. This could be due to the complexity of formulating
Reynolds stresses in turbulent flows. The majority of turbulent flow calculations carried out in earlier
days used two equation models such as the standard κ-ε model (hereafter referred as SKE) and the
κ-ω model, because of their robustness, computational efficiency, and completeness. In the classical
SKE model, the turbulent kinetic energy (κ) and the turbulent kinetic energy dissipation rate (ε) were
36 Ichiro Kimura, and Takashi Hosoda, ”A non-linear κ-ε model with realizability for prediction of flows around
bluff bodies”, Int. J. Numerical Meth. Fluids 2003; 42:813–837 (DOI: 10.1002/_d.540).
37 Champagne FH, Harris VG, Corrsin S. ,”Experiments on nearly homogeneous turbulent shear flow”. Journal of
calculated using modeled transport equations separately for k and e along with the Boussinesq eddy
viscosity approximation.
40 T. B. Gatski and C. G. Speziale,“On Explicit Algebraic Stress Models for Complex Turbulent Flows”, J. Fluid Mech.,
vol. 254, pp. 59–78, 1993.
41 T. J. Craft, B. E. Launder, and K. Suga, “Development and Applications of a Cubic Eddy-Viscosity Model of
Turbulence”, Int. J. Heat Fluid flow, vol. 17, pp. 108–115, 1996.
42 T. H. Shih, J. Zhu, and J. L. Lumley, “A New Reynolds Stress Algebraic Equation Model”, Computer. Meth. Appl.
Regime with Various Turbulence Models”, Int. J. Numerical Methods Fluids, vol. 35, pp. 763–784, 2000.
44 M. Saghafian, P. K. Stansby, M. S. Saidi, and D. D. Asplay, “Simulation of Turbulent Flows around a Circular
Cylinder Using Non-Linear Eddy-Viscosity Modeling: Steady and Oscillatory Ambient Flows”, J. Fluids Structure,
vol. 17, pp. 1213–1236, 2003.
45 R. B. Jennifer, “Verification Testing in Computational Fluid Dynamics: An Example Using Reynolds-Averaged
Navier-Stokes Methods for Two Dimensional Flow in the Near Wake of a Circular Cylinder”, Int. J. Numerical Meth.
Fluids, vol. 43, pp. 1371–1389, 2003.
46 L. Kimura and T. Hosoda, “A Non-linear κ-ε Model with Reliability for Prediction of Flows around Bluff Bodies”,
Body by Non-linear Model”, Int. J. Numerical Meth. Heat Fluid Flow, vol. 16, pp. 660–673, 2006.
39
u i u i u i u j P (ui uj ) 1 u i
2
0 , Re Eq. 4.3
x i t i x j x i x j x i x j
where xi is the spatial coordinate, t is the time, u͞i is the ensemble-averaged velocity, ͞u´i is the
fluctuating velocity, and P is the averaged pressure divided by the density. As a result of ensemble-
averaging process, further unknowns are introduced into the momentum equations by means of
Reynolds stresses ͞u´iu´j . In engineering flows, closure approximation using two-equation models
for u
͞ ´ju´j have gained popularity because of their simplicity. In this article the study is confined to the
κ-ε model, which employs additional transport equations for turbulent kinetic energy κ and its
dissipation rate ε, and they are given as
κ κu j u ν t κ
ui uj i ε ν
t x j x j x j σ
κ x j
Eq. 4.4
ε εu j ε ε2 ν t ε
C1 u i u j C 2 ν
t x j κ κ x j σ
ε x j
Where κ is the turbulent kinetic energy, is the turbulent kinetic energy dissipation rate, ν is the fluid
kinematic viscosity, and νt is the eddy viscosity. C1, C2, σκ, and σε are the model constants given by
Table 3.1.
on a subcritical Reynolds number showed the dominant span wise scales having wavelengths of
approximately three to four cylinder diameters in the Reynolds number range 180 < Re < 240. After
this Reynolds number, the wavelength shortens to nearly one diameter. In our simulation, we have
taken a span wise length of 4D along the z direction, and periodic boundary condition is enforced on
the boundary. This extent is slightly larger than that used in LES and DNS calculations. A non-
equilibrium wall function approach is used to capture the adverse pressure gradient effect.
48M. Breuer, “Large Eddy Simulation of the Subcritical Flow past a Circular Cylinder: Numerical and Modeling
Aspects”, Int. J. Numerical Meth. Fluids, vol. 28, pp. 1281–1302, 1998.
41
4.2.1.6 Conclusions
In the present work, 3-D unsteady computation of flow past a circular cylinder at subcritical Reynolds
number has been performed using a nonlinear κ-ε model to evaluate its applicability. The same test
case was simulated with 2DNLKE and its 3D counterpart 3DSKE to understand the effectiveness of
the present model. The bulk parameters and the wake velocity recovery match well with
experimental data and LES results. Since the grid requirement is not as severe as in LES and the
number of cycles required to do averaging is also less, computational cost associated with the present
model is very much less. For high-Re flows and flows encountered in practical engineering
applications, there is a restriction on the mesh size and the LES technique is prohibitively expensive.
Encouraging performance of the present NLKE model suggests that it could be used as an alternative
tool in these situations. Further improvement in the prediction may be possible by making the model
fully cubic form.
best suited for transient simulations. The main difference between conventional turbulence
modeling and LES is the averaging procedure. The LES technique does not involve the use of
ensemble average; rather it consists in applying a spatial filter to N-S equations.
Filter Definition
Mathematically, one may think of separating the velocity field into a resolved and sub-grid part. The
resolved part of the field represent the "large" eddies, while the sub-grid part of the velocity
represent the "small scales" whose effect on the resolved field is included through the sub-grid-scale
model49. This is called explicit filtering and Figure 4.5 illustrates the difference between the filtered
velocity ūi and the instantaneous velocity ux. formally, one may think of filtering as the convolution
of a function with a filtering kernel G:
Δ = L3 Δ = L /32 Δ = L /16
or its generalization, the cubic root of the cell volume. In case of anisotropic grids, definition tends to
provide a fairly low value50. For this reason, the quadratic mean is used as following
49From Wikipedia.
50Jochen Frӧhlich , Dominic von Terzi,” Hybrid LES/RANS methods for the simulation of turbulent flows”,
Progress in Aerospace Sciences 44 (2008) 349– 377.
44
1/2
2x 2y 2z
g Eq. 4.7
3
which is advocated in some publications. Other authors favor the maximum
u i u i u i u j 1 p 2ui
0 , ν Eq. 4.9
x i x j x j ρ x i x jx j
u i u i u j 1 p 2u i
ν or
t x j ρ x i x jx j
Eq. 4.10
u i u i u j 1 p 2 ui
ν
t x j ρ x i x jx j
This equation models the change of time of the filtered variable u͞i . Since the unfiltered variable ui
are not known, it is impossible to directly calculate
𝛛𝐮𝐢 𝐮𝐣 ̅𝐢𝐮
𝛛𝐮 ̅𝐣
≠
⏟𝛛𝐱 𝐣 ⏟𝛛𝐱 𝐣
𝐍𝐨𝐭 𝐊𝐧𝐨𝐰𝐧 𝐊𝐧𝐨𝐰𝐧
Eq. 4.11
However the quantity on the right is known. Substituting:
45
̅𝐢𝐮
̅ 𝐢 𝛛𝐮
𝛛𝐮 ̅𝐣 𝟏 𝛛𝐩 𝛛𝟐 𝐮̅𝐢 ̅̅̅̅̅
𝛛𝐮 𝐢 𝐮𝐣 ̅𝐢𝐮
𝛛𝐮 ̅𝐣
+ =− +𝛎 − −
𝛛𝐭 𝛛𝐱 𝐣 𝛒 𝛛𝐱 𝐢 𝛛𝐱 𝐣 𝛛𝐱 𝐣 ⏟𝛛𝐱 𝐣 𝛛𝐱 𝐣
𝛛𝛕𝐢𝐣
( 𝛛𝐱 𝐣 )
𝐰𝐡𝐞𝐫𝐞 𝛕𝐢𝐣 = ̅̅̅̅̅
𝐮 ̅𝐢𝐮
⏟𝐢 𝐮𝐣 − 𝐮 ̅𝐣
𝐒𝐆𝐒 𝐌𝐨𝐝𝐞𝐥𝐢𝐧𝐠
Eq. 4.12
The result is a set of LES equations. In addition, the filtered governing equation for a passive scalar φ
such as mixture fraction or temperature, can be written as
q ij
t
x j
u j
J
x j
x j
Eq. 4.13
Where the diffusive flux of and is the sub-filter stress tensor for the scalar. The filtered diffusive flux
is unclosed, unless a particular form is assumed for it (e.g. a gradient diffusion model ∂
q ij u j u j Eq. 4.14
And can similarly be split up into contributions from interactions between various scales. This sub-
filter tensor also requires a sub-filter model.
3-7643-2723-1.
46
RANS
u
t
( u u ) ν ν T u u
T
p
Eq. 4.15
LES
u
t
(u u ) ν ν SGS u u
T
p
It is obvious, the only change is in the dynamic viscosity determination, νT and νSGS. Or, the main
difference being that in RANS the unclosed term is a function of the turbulent kinetic energy and the
turbulent dissipation rate whereas in LES the closure term is dependent on the length scale of the
numerical grid. So in RANS the results are independent of the grid resolution!
and usually the DES needs more refine mesh that RANS. Another point of view is that RANS can only
give a time averaged mean value for velocity field since it is based on time averaging. In fact velocity
54 Grinstein, Fernando, Margolin, Len, Rider, William, “Implicit large eddy simulation”, Cambridge University
Press. ISBN 978-0-521-86982-9, 2007.
55 Sagaut, Pierre, “ Large Eddy Simulation for Incompressible Flows”, (3rd Ed.), Springer, 2006.
56 “Large-eddy simulation of unsteady surface pressure over a LP turbine due to interactions of passing wakes and
57Jochen Frӧhlich , Dominic von Terzi,” Hybrid LES/RANS methods for the simulation of turbulent flows”,
Progress in Aerospace Sciences 44 (2008) 349– 377.
48
ui ui u j p ν u i τ ijRANS
t x j x j x j x j
x j
Eq. 4.16
u i u i u j p νu i τ ij
DES
t x j x j x j x j x j
ui u
τ ijRANS f , κ, ε, C , τ ijDES f i , Δ g , C Eq. 4.17
x i x i
where C is a model constant, κ the turbulent kinetic energy, and ε the turbulent dissipation rate. The
latter two are determined from other relations. For LES based on the Smagorinsky model uses a
relation like where Δg is a length scale related to the numerical grid, Since there exist many variants
of LES and RANS models we define the following: a model qualifies as an LES model if it explicitly
involves in one or the other way the step size of the computational grid. RANS models, in contrast,
only depend on physical quantities, including geometric features like the wall distance.
u i u i u j p νu i τ ijModel
t x j x j x j x j
x j Eq. 4.18
A transition from LES to RANS can be achieved in several ways. One possibility is blending, i.e. by a
weighted sum of a RANS model and an LES the models according to
In this fashion, f RANS and f LES are local blending coefficients determined by the local value of a given
criterion. Another strategy is to use a pure LES model in one part of the domain and a pure RANS
model in the remainder, so that a boundary between a RANS zone and an LES zone can be specified
at each instant in time. The transport equation for the velocity, however, is the same in both zones
with no particular adjustment other than switching the model term at the interface. This way the
computed resolved velocity is continuous. We term this strategy Interfacing LES and RANS.
Furthermore, if the interface is constant in time, it is called a hard interface. If it changes in time
depending on the computed solution, it is termed a soft interface.
Segregated modeling is the counterpart to unified modeling as LES is employed in one part of the
computational domain, while RANS is used in the remainder. With segregated modeling, however,
the resolved quantities are no more continuous at the interfaces. Instead, almost stand-alone LES and
RANS computations are performed in their respective subdomains which are then coupled via
49
appropriate boundary conditions. Except for laminar flows, the solution is discontinuous at these
interfaces. This avoids any gradual transition in some gray area characteristic of unified turbulence
models. Segregated modeling allows for embedded LES by designing a configuration where in an
otherwise RANS simulation a specific region is selected to be treated with LES with full two-way
coupling between the zones58.
58 Jochen Frӧhlich , Dominic von Terzi,” Hybrid LES/RANS methods for the simulation of turbulent flows”,
Progress in Aerospace Sciences 44 (2008) 349– 377.
59 J. U. Schlüter, X. Wu, S. Kim, J. J. Alonso, and H. Pitsch, AIAA-2004-369, 42nd Aerospace Sciences Meeting and
61 G. Medic, D. You AND G. Kalitzin, “An approach for coupling RANS and LES in integrated computations of jet
engines”, Center for Turbulence Research, Annual Research Briefs, 2006.
62 Jochen Frӧhlich , Dominic von Terzi,” Hybrid LES/RANS methods for the simulation of turbulent flows”,
63Kyle D. Squires, Les Application in Aerodynamics”, School of Mechanical, Aerospace, Chemical and Materials
Engineering Arizona State University Tempe, Arizona, USA.
52
Figure 4.12 Comparison of Pressure contours on planner cuts for RANS and DES models
u L
N 3DNS Re9/4 Re Eq. 4.21
Where N is the number of points along a given mesh direction with increments h and ú is the root
53
mean square (RMS) of velocity. Hence, the memory storage requirement in a DNS grows very fast
with the Reynolds number64. In addition, given the very large memory necessary, the integration of
the solution in time must be done by an explicit method. This means that in order to be accurate, the
integration, for most discretization methods, must be done with a time step, Δt, small enough such
that a fluid particle moves only a fraction of the mesh spacing h in each step. Therefore, it remains
limited to very simple cases. Filtering requires a finer grid than implicit filtering, and the
computational cost increases with [Sagaut (2006)] covers LES numeric in greater detail65. [Sarkar
and Voke]66 carried out an LES study of interactions of passing wakes and in flexional boundary layer
over a low-pressure turbine blade and Figure 4.7 shows flow structures due to the complex
interactions of passing wakes and the separated shear layer. Evidently, as it appears in Eq. (9.2) and
above discussion, the DNS method is highly dependent to Reynolds number which will be debated in
the next section.
Figure 4.13 Vorticity contours from spectral DNS at two Reynolds numbers (Spalart et al. 2008).
64 From Wikipedia.
65 Sagaut, Pierre (2006). “Large Eddy Simulation for Incompressible Flows”, (3rd ed.), Springer.
66 “Large-eddy simulation of unsteady surface pressure over a LP turbine due to interactions of passing wakes and
in flexional boundary layer”, J Turbo machine, 128 (2) (2006), pp. 221–231
67 Gary N. Coleman and Richard D. Sandberg, “A Primer on Direct Numerical Simulation of Turbulence – Methods,
Numerical Considerations
The obligation of having to resolve all spatial and temporal scales of the turbulence requires that
numerical errors be monitored and controlled. As a result, DNS has historically not used commercial
CFD packages, but specially-written codes, optimized for the flow-types of interest68. The need for
DNS algorithms to be efficient; that is, to have a high ratio of accuracy to computational cost which is
particularly important. There are a number of strategies that DNS codes have employed to do this,
including finite-volume, finite-element, and discrete-vortex as well as B-spline methods. But central
to development of an efficient algorithm for DNS are two methods; Spectral and Finite Difference.
68 Gary N. Coleman and Richard D. Sandberg, “A Primer on Direct Numerical Simulation of Turbulence –
Methods, Procedures and Guidelines”, Technical Report AFM-09/01a (March 2010).
69 TamCKW. “Computational aeroacoustics: issues and methods”, AIAA pap. 95–0677.
70 Lele SK. 1992, “Compact finite difference schemes with spectral-like resolution”, J. Comp. Phys. 103:16–42.
71 Parviz Moin and Krishnan Mahesh, “ Direct Numerical Simulation: A Tool in Turbulence Research”, Annu. Rev.
72P.R. Spalart, “Strategies for turbulence modelling and simulations”, International Journal of Heat and Fluid
Flow 21 (2000) 252±263.
56
73 Speziale, C.G.,” On nonlinear K- l and K - ԑ models of turbulence”. J. Fluid Mech. 178, 459, 1987.
74 Travin, A., Shur, M., Strelets, M., Spalart, P.R., 2000. “Detached-eddy simulations past a circular cylinder”.
Flow, Turb. Combust, to appear.
75 P.R. Spalart, “Strategies for turbulence modelling and simulations”, International Journal of Heat and Fluid
Figure 4.17 Simulation of Flow Past Circular Cylinder by Various Approaches (Shur et al., 1996;
Travin et al., 2000)
per step would be the same, and the time step could only rise by a modest amount. Conclusive tests
of the filter-grid relationship would crucially depend on the definition of a figure of merit.
Re-
Name Aim Unsteady Empiricism Grid Size
dependence
2DURANS Numerical Yes Weak Strong 105
3DRANS Numerical No Weak Strong 107
3DURANS Numerical Yes Weak Strong 107
DES Hybrid Yes Weak Strong 108
LES Hybrid Yes Weak Weak 1011.5
QDNS Physical Yes Strong Weak 1015
DNS Numerical Yes Strong None 1016
Outlook
Progress in numerical methods and computers is intensifying the challenge for turbulence
treatments, to provide a useful level of accuracy in slightly or massively separated flows over fairly
complex geometries at very high Reynolds numbers. This is desirable in the near future, especially
the jet-engine industry. In addition, the needs of non-specialist users and automatic optimizers
dictate a very high robustness. Flows with shallow or no separation appear to be within the reach of
the current steady RANS methods or their finely calibrated derivatives, incorporating modest
improvements such as nonlinear constitutive relations. For such flows, transition prediction with
generality, accuracy, and robustness may prove more challenging than turbulence prediction. With
massive separation, it appears possible we will give up RANS, steady or unsteady. This will probably
be the major debate of the next few years. The alternative is a derivative of LES, in which the largest,
unsteady, geometry-dependent eddies are simulated and (for most purposes) “discarded” by an
averaging process. We have to balance our ambitions with cost considerations, and a table brief the
issue was tentatively provided.
A major consideration is whether LES is practical for the entire boundary layer, and it was strongly
argued that this will not be the case, in the foreseeable future. This forces hybrid methods, with quasi-
steady RANS in the boundary layer. In this paper, LES was effectively defined as a simulation in which
the turbulence model is tuned to the grid spacing, and RANS as the opposite. Other more subtle
definitions probably exist, but this one seems to classify almost all the studies to date. Speziale's
hybrid proposal involves the grid spacing and the Kolmogorov length scale but, surprisingly, not the
internal length scale of the RANS turbulence model; thus, it is difficult to classify [Speziale]76. The
proposal of [Aubrun et al.]77. is also hybrid, as it leads to combining “modelled” and “resolved”
Reynolds stresses, but the modelled stresses do not scale (and vanish) with the grid spacing as they
do in LES 78. Variations on the now-running DES proposal clearly have a wide window of opportunity.
76 Speziale, C.G., “Turbulence modeling for time-dependent RANS and VLES”, a review.AIAA J. 36 (2), 173, 1998.
77 Aubrun, S., Kao, P.L., HaMinh, H., Boisson, H, “The semi deterministic approach as way to study coherent
structures. Case of a turbulent flow behind a backward-facing step”, Proceedings of the Fourth International
Symposium on Engineering Turbulence Modelling and Measurements, Corsica, Elsevier, Amsterdam, 1999.
78 See Previous.
59
The plausible spread of hybrid methods highlights the permanence of a partnership between
empiricism and numerical power in turbulence prediction at full-size Reynolds numbers. This
demands a balance in funding and in publication space. Since hybrid methods offer flexibility when
setting the boundary between “RANS region” and “LES regions”, the more capable the RANS
component is, the lower the cost of the hybrid calculation will be. Therefore, the switch to LES in
some regions does not remove the incentive to further the RANS technology. This scene also raises
the issue of which core of experiments and DNS will be the foundation of the empirical component of
the system. As ever, we will need simple flows for calibration of the RANS sub-system, and more
complex flows for validation of the full CFD system.
In general, the approach for solving turbulent flow equations can roughly be divided into four classes
as Direct Numerical Simulation (DNS), Reynolds Average N-S equations (RANS), Large Eddy
Simulation (LES) and Detached Eddy Simulation (DES), as 79:
79 Wikipedia.
61
1. Decide whether to solve all equations simultaneously or to solve the model equations and
mean-flow equations sequentially.
2. If the preferred option is to solve all equations simultaneously, determine the changes to the
flux-Jacobian matrices.
3. Make provision for handling source terms.
In principle, solving all equations simultaneously will yield the most rapidly convergent scheme in
the number of iterations, but not necessarily in CPU time. However, the coupling between the
turbulence-model equations and the mean-flow equations appears to be relatively weak. The
primary coupling is through the diffusion terms, and the eddy viscosity is usually treated as a
constant in forming the viscous-flux Jacobian matrix. Limited experience to date seems to indicate
64
there is little advantage to solving all equations simultaneously as opposed to solving the model
equations and mean-flow equations sequentially. If all equations are solved simultaneously, the basic
system of equations for 1-D κ-ω model would be as follows:
Q
(F Fv ) S
t x
0
4 u 0
ρ ρu μ τ
3 x
xx
ρu ρu 2 p 0
4 u
u μ τ xx q̂ x
0
Q ρ E , F ( ρ E p)u , Fv 3 x , S u
ρκ ρ uκ κ τ xx x β ρ ωκ
(μ σ μ T ) ω u
ρ ω ρ uω x
α xx
τ β ρ ω 2
(μ σμ T ) ω κ x
x
1 4 u 2
E e u 2 κ and τ xx μ T ρκ
2 3 x 3
I F Fv S F Fv
δx ΔQ δ x (F Fv ) S , where
n n n
and are the Jacobian
Δt Q Q Q Q Q
Eq. 5.1
Since dS/dQ is a diagonal matrix and its diagonal elements are always negative, its contribution is
guaranteed to enhance diagonal dominance of the matrix multiplying ΔQ.