Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract: A review of using large-eddy simulation (LES) in computational fluid dynamic studies of internal combustion engines is presented. Background material on turbulence modelling, LES approaches, specifically for engines, and the expectations of LES results are
discussed. The major modelling approaches for turbulence, combustion, scalars, and liquid
sprays are discussed. In each of these areas, a taxonomy is presented for the various types of
models appropriate for engines. Advantages, disadvantages, and examples of use in the literature are described for the various types of models. Several recent examples of engine studies
using LES are discussed. Recommendations and future prospects are included.
Keywords: LES, engines, CFD, turbulence, combustion, sprays
INTRODUCTION
It is generally agreed that the next generation of turbulence modelling in computational fluid dynamics
(CFD) for many applications will be some form of
large-eddy simulation (LES). For the appropriate
applications, LES can offer significant advantages
over traditional Reynolds Averaged Navier Stokes
(RANS) modelling approaches. For example, in
internal combustion (IC) reciprocating engines, LES
can be used to study cycle-to-cycle variability, provide more design sensitivity for investigating both
geometrical and operational changes, and produce
more detailed and accurate results. There are also
characteristics of IC engines, such as inherent
unsteadiness and a moderately sized domain, that
are well suited to LES. This is not to say that LES will
replace RANS. There are pluses and minuses for
both methods and users should pick the appropriate
tool for the topics being studied. However, as inexpensive computing power increases, the ability to
use LES in IC engine simulations is increasing.
As LES gains in capability, there is the potential for
a larger set of people using the models and a broader
application of LES to engines. In addition, LES in IC
engines is new, and there are potential uncertainties
C J Rutland
(1)
Here, the overbar symbol represents the spatial filtering in LES or the ensemble averaging in RANS.
For engines, density varies significantly and the
overbar represents a mass weighted (or Favre) filtering or averaging [8]. Then, ui is usually called the
mean velocity, although more formally it is the filtered velocity in LES. In both LES and RANS, the
overbar represents an averaging process designed to
reduce the range of eddy sizes or length scales
in the flow so that ui can be represented on a computational grid appropriate for engines. An
(2)
(3)
(4)
C J Rutland
C J Rutland
Emphasis
Number of grid cells
Numerical methods
Fraction of kinetic energy
resolved on grid
Scientific LES
Engineering LES
diesel engine. The RANS and LES simulations duplicate the region of the experimental images using the
same coarse grid of a simple sector mesh common
in diesel engine simulations. The RANS results show
a broadened or smeared region for the higher temperature, while the LES results show the same type
of jet large-scale structures seen in the experimental
images. Thus, with only a change to LES turbulence
and scalar mixing models that are appropriate for
applications, the simulation results pick up flow
processes that occur in the experiments that were
not previously available in the RANS simulations.
2.3 Expectations of LES for IC engines
In addition to the general expectations of LES listed
above, there are additional expectations related to
IC engine simulations. Generally, these can be
described as the ability to study new physical phenomena in engines and an increased sensitivity to
design changes. These are discussed in more detail
in the following subsection.
2.3.1 Study new phenomena
A very important aspect of using LES for engines is
that it will allow studies of new phenomenon. There
are important aspects of engine flows and combustion that are difficult, if not impossible, to address
with RANS but which are more amenable to LES
approaches. One of the primary features is cycle-tocycle variability. Reynolds Average Navier Stokes
uses models designed to capture the ensemble
averages. This results in higher turbulent viscosity
that almost always removes, or at least smears out,
the variation of in-cylinder flows and combustion
that coincide with cycle-to-cycle variability. Since
LES models are designed to filter out the smaller
scales and retain the larger scales, they are less dissipative. The remaining large scales respond to the
non-linearities inherent in the Navier Stokes equations, and at least some aspects of cycle-to-cycle
variability can occur in the simulations. As discussed in section 4, several research groups are
C J Rutland
00
nT = u
(5)
The terms k and e are interpreted to be the turbulent kinetic energy (TKE) and the turbulent kinetic
(6)
Then,
scale
p k and e provide a turbulent velocity
1.5
00
u e k and a turbulent length scale of ek /e. In
this interpretation, the length scale is thought of as
the integral scale of the turbulence even though the
flow is not homogeneous.
If equation (3) is used for LES models, there are
several approaches for obtaining expressions for nT.
One of the more common models is based on the
ideas of Smagorinsky [18] and results in
nT = CS D2 S~
nT = Cm
energy dissipation rate (or just dissipation). In modern approaches, these terms are obtained from individual transport equations. Thus, the RANS (ke)
model is a two-equation turbulence model.
To provide additional understanding, it is useful
to rewrite the model based on a physical interpretation using a velocity and length scale
(7)
T2 Smagorinsky
Numerical viscosity 0
only
Yes
0
T3 Scale similarity
No
T4 Dynamic Smagorinsky
Yes
T5 k-equation LES
Yes
T7 Dynamic structure
Non-viscosity
Disadvantages
No model required
10
C J Rutland
difference between these additionally filtered quantities and the base quantities calculated on the CFD
grid (see appendix 3). This additional filtering operation is a modest increase in computational cost,
resulting in an increase of ~20 per cent for a simple
turbulent flow.
An interesting variation of the dynamic procedure
was developed by Meneveau et al. [32], in which a
Lagrangian concept was used to develop the model
coefficient. The idea was to average over fluid particle pathlines to improve accuracy. In practice, two
additional transport equations were used to represent the Lagrangian average of terms used to evaluate the dynamic coefficient.
Haworth et al. were also some of the early
explorers in using LES for IC engines [33]. They
mostly used T2 and T4 type models in several different codes. They carried out extensive studies on a
simple, engine type flow with a stationary valve [34].
This configuration, sometimes called the Imperial
College engine, has a large experimental dataset and
is useful for validating valve flows. Haworth et al.
have shown good comparison between ensemble
averaged LES models and experimental data for
both mean and fluctuating velocity profiles at different locations and different crank angles.
The dynamic procedure is very powerful and can
be used in many situations to find modelling coefficients. When used with the Smagorinsky model, the
results are reasonably good for non-reacting flows.
However, dense grids are required and often an
additional averaging must be used to avoid instabilities that arise from negative viscosities. Despite the
improvements found in T4, it still retains the drawbacks of the equation (3) viscosity models discussed
in the previous section. No matter how good a
model is formulated for the turbulent viscosity, the
fundamentals of T2 and T4 are very weak.
T5. The k-equation approach is a practical viscosity-based, one-equation LES model. It was originally
developed for atmospheric flows [35], and is still
common in that field. Some of the first useful
k-equation models for engineering flows were developed by Kim and Menon [36]. This model was still
viscosity based (equation (3)), but now the turbulent
viscosity was formed from the subgrid TKE, ksgs, and
a grid length scale, D, resulting in the following
expression
q
nT = Ck D ksgs
(8)
has several distinct advantages. First, it incorporates more physical processes, such as the convection, production, and dissipation of subgrid kinetic
energy. Second, the subgrid kinetic energy provides
a velocity scaling that can be used in other models,
such as combustion, scalar transport, and sprays.
Third, models that use a subgrid k-equation provide
a better model for the subgrid stresses and thus
work better on the coarser grids commonly found in
engine CFD [37, 38].
Menon et al. have applied the T5 turbulence
model to engine flows with good results [39].
Bianchi et al. have performed careful studies of LES
models for engine type flows in simple configurations. For example, they have compared T2 and T5
turbulence models with RANS results for a stationary valve, steady flow bench configuration [28].
The subgrid kinetic energy equation is fairly simple to implement. It requires only one additional
major modelled term, which is for the dissipation of
subgrid kinetic energy. Fortunately, this term plays
its proper role in LES, which at the subgrid scale is
to remove kinetic energy. The dissipation term is
not required to provide the mean value for all scales,
nor is it used to obtain length scales or time scales
as it is in RANS modelling. Thus, dissipation modelling is much less critical, and simple models seem
to work well.
T6. The k-equation LES models have also been
implemented using the dynamic procedure to
obtain a better, local value for the coefficient in
equation (8) [36]. This method is a logical extension
of T5; however, additional implementation details
must be observed to maintain stability. At this time,
it is not clear if this additional complexity beyond
the basic T5 model is useful in engine simulations.
T7. A recent development in LES turbulence models is the dynamic structure approach developed by
Pomraning and Rutland [40] and Chumakov and
Rutland [41]. In this approach, a turbulent viscosity
is not used. Instead, a tensor coefficient is obtained
directly from the dynamic procedure. This tensor
coefficient is multiplied by the TKE that is obtained
from a transport equation (see appendix 4 for more
details). The resulting dynamic structure model is
t ij = Cij ksgs
(9)
11
12
C J Rutland
13
Table 3 List of major combustion modelling approaches that have potential for use in LES.
Original or primary type of combustion for each model is indicated by Mode in column
2: H for homogeneous, P for premixed, D for diffusion
Model type
C1
C2
C3
C4
C5
Mode
Advantages
Disadvantages
Direct Integration
CHEMKIN or other stiff ODE integrators
Blended models
RIF
Time-scale models
(a) Magnusson
(b) CTC
Transport-equation models
(a) Progress variable C
P, D
P, D
P, D
PDF transport
all
14
C J Rutland
successfully in LES applications using the T7 turbulence model for direct injection diesel LTC studies
by Rutland and his group [15, 43, 44].
The C1 approach requires detailed chemical
kinetic mechanisms to be successful in engines, and
this usually results in very large computational run
times. Progress in improving run times is being
achieved by improved load balancing in parallel
computing environments [67]. Additional run-time
improvements are being achieved by applying
advanced numerical techniques such as cell clustering and analytical Jacobians [68] or precomputed,
tabulated results from detailed chemistry calculations [69, 70]. These methods are computationally
efficient, but may require close monitoring of
approximation errors, especially for ignition and
other situations where results are sensitive to kinetic
details.
C2. In an attempt to incorporate more detailed
chemical kinetics but without the computational
penalty, Peters group have developed the representative interactive flamelet (RIF) model [7174].
Individual flamelets that represent the main combustion process are tracked using a Lagrangian
method through the domain. The approach can be
calibrated to work with conventional diesel combustion and provide detailed chemistry for emissions.
However, the approach has difficulty with more
homogeneous flows, wall heat transfer, multiple fuel
injection operation, and spatially non-uniform
mixing that can occur in different regions of the
combustion chamber. Additional flamelets are
sometimes added to help address these issues, and
the method begins to resemble the cell clustering
approach used in C1 models. Combustion is tracked
by the Lagrangian flamelets rather than the processes within each CFD grid cell. The approach is
more of a blending between a CFD flow model and
a system level heat-release model. Since it is not
clear how a representative flamelet concept is consistent with the LES spatial filtering approach, the
RIF approach is not recommended for LES.
C3. For RANS applications, the time-scale
approach was originally developed for spark ignition
engines (Abraham et al. [75]) and later adapted for
diesel engines (Kong and Reitz [76]). The characteristic time-scale (CTC) model is a very practical
approach that can give good results when experimental data are available to adjust coefficients. The
CTC model is an outgrowth of the less commonly
used Magnusson type approaches, but is more
advanced in that CTC drives species concentrations
to a specified value. This specified value is commonly the local equilibrium value. However, in
some models this specified value is obtained from a
15
Fig. 4 Instantaneous temperature fields from an HCCI test engine with a square bowl designed
to increase turbulence levels (from [85], reprinted with permission from SAE paper 200801-1656, 2008, SAE International)
16
C J Rutland
on RANS concepts. The newer one-equation turbulence models (T5, T6, T7) are better at providing
time-scale information because they track the subgrid kinetic energy ksgs using a transport equation.
This can be combined with length scales (gradients
or filter length scales) to provide time scales to combustion models.
Multimode combustion. In some engine applications, combustion does not easily fall into the
traditional classifications of premixed mode or
non-premixed mode. Or combustion may occur in
multiple modes within a cycle. Examples are direct
injection gasoline technologies and some of the
newer LTC technologies such as partially PCCI.
These types of combustion processes are probably
best described by combinations of direct integration
for ignition (C1), premixed and partially premixed
combustion for early, more highly mixed processes
(C3, C4ac), and mixing controlled combustion for
later processes (C4de). These multimode operations can occur in a time sequence, or simultaneously, but in different regions in the combustion
chamber, or some combination of these two situations. A combination C4a and C4d model for nonengine LES was reported by Ihme and Pitsch [120].
For engine applications, hybrid approaches have
been explored for RANS diesel applications [107]
and premixed/diffusion combustion in the ECFM3Z
model [95]. More recent work has demonstrated
LES simulations of diesel engine simulations using a
combination of C1, C3b, and C4d combustion models with a T7 turbulence model [15, 44].
The difficulty with multimode approaches is designating and accurately evaluating the best parameters for switching between the modes.
Commonly, these parameters measure a mixing
state (for example scalar dissipation rate), relative
timescales (Damkohler or Karlovitz numbers), or
reaction progress (for example, reaction products or
normalized temperature). Currently, there is not a
good theoretical framework for determining the
switching parameters, so they are often developed
based on physical arguments. In addition, the
switching procedure and the value at which
the switch occurs may have a greater impact on the
results than the details of the individual combustion
models. Clearly, much more work needs to be done
in this area for both RANS and LES modelling.
17
Transport equations
Advantages
Disadvantages
SC1
None
Poor results
SC2
Viscosity based
(a) Simple turbulent viscosity
SC3
Self-similarity
SC4
SC5
Dynamic structure
SC6
many
18
C J Rutland
energy (equation (8)) and combined with a turbulent Schmidt or Prandtl number. This approach is
used with LES in IC engine applications [15]. It provides good results at reasonable computational
expense. This is primarily because it is combined
with advanced turbulence models T5T7.
SC3. Scale-similarity models are based on the
same concepts that underlie many of the turbulence
models (T3, T4, T6, T7, and appendix 3). Thus, this
approach has the potential to be very accurate. This
has been demonstrated by Moin and others [103,
123] from comparison with DNS results and for
dump-combustor
simulations.
However,
the
approach does not appear to have been used for
LES engine simulations yet. In addition, since scalesimilarity models allow backscatter, which can
result in unrealizable scalar values, SC3 models
would require additional dissipation.
SC4. The transport equation model approach uses
a traditional transport equation for the subgrid scalar flux or the subgrid scalar fluctuations. This is a
logical evolution of LES scalar models (see Jimenez
et al. [19] for discussion), but has only been used in
LES modelling in the context of the dynamic structure approach (SC5). Probably the primary reason
this approach has not been used is that every scalar
(e.g. species and energy) requires an additional
transport equation, and this can become very
expensive. This is true, especially if simple turbulence models are used and high grid resolution is
required. In addition, each transport equation
requires a model for the subgrid scalar dissipation.
This approach could be used in C4 type combustion
models, especially the mixture fraction models
where the only scalar transport required is for the
mixture fraction variable.
SC5. The dynamic structure approach of T7 was
extended for scalar transport modelling by
Chumakov and Rutland [124]. This involves a
transport equation for the subgrid scalar fluctuations that is then used with the dynamic structure
approach to model the scalar flux. This can work
well, but can be expensive because, just as with
SC4, an additional transport equation for the fluctuating component of each species is required.
Chumakov is continuing work on similar advanced
scalar flux modelling for LES, and the results look
promising [125]. However, testing on engine applications is still required.
SC6. The linear eddy model (LEM) approach uses
a very different concept to model subgrid scalar
mixing. In LEM, a one-dimensional (1D) unsteady
equation, which can contain mixing, diffusion, and
reaction is solved in each CFD computational grid
cell. The emphasis of the LEM approach is on the
mixing term that is modelled using a triplet mapping. This is a simple rearrangement of values based
on length and time scales that mimic turbulent
eddies. The original concept was developed by
Kerstein [126] for turbulence. However, it is
Menons group that has carried out a lot of work
extending LEM for reacting flows [39, 60]. This
approach can be expensive because a large number
of grid points are used within each CFD grid cell for
the 1D equations. Partly from the increased CPU
resources and improved algorithms, the approach
gives nice results [38, 127, 128]. The LEM can be
parallelized, but it is not practical for applications at
this time and has not been used in engine LES.
3.3.1 Scalars: additional considerations
Scalar dissipation rate. A scalar dissipation rate term
can arise in several different ways when modelling
scalars and combustion. As noted in the previous
section, it may be used as a time scale in combustion modelling. For scalars, it often occurs in prescribed PDF combustion models, in which a mean
and variance of some scalar is required. Commonly,
the scalar is the mixture fraction, and transport
equations for both the mean and variance are used.
Within the transport equation for the variance, the
scalar dissipation rate occurs.
The most common approach for scalar dissipation rate modelling in LES is one that is based on
RANS modelling and uses the turbulence time scale
[19]. However, this does not make sufficient use of
LES concepts, and has been shown to give poor
results in a priori engine studies [63]. More promising approaches based on scale-similarity ideas have
been developed by Chumakov [129] for general
flows and by Zhang et al. [63] for engine LES.
Scalar variance. Another approach to modelling
the scalar variance is to avoid using a transport
equation and use an algebraic closure. This
approach has been developed for use in LES using
scale-similarity concepts by Cook and Riley [130],
Cook [131], and Jimenez et al. [19]. This avoids the
need for a scalar dissipation rate model in the transport equation. In addition, the approach seems to
provide a sufficiently accurate mixture fraction scalar variance for use in C4d prescribed PDF combustion models. The approach has been used in LES
diesel engine modelling as part of a multimode
combustion model, and gives good results when
compared to experimental data [44]. However, there
is strong evidence from Colin and Benkenida [132]
that an algebraic closure is insufficient in flows
with sprays, indicating that additional research is
needed.
19
20
C J Rutland
Advantages
Disadvantages
S1
Laminar correlations
S2
RANS correlations
S3
LES modifications
S4
21
Fig. 5 Droplet radius (rd(cm)) and vapour mass fraction (Yf) for 110 MPa injection pressure into a
spray vessel with a density of 30.2 kg/m3: (a)
RANS, (b) LES with no spray source model, and
(c) LES with the spray source model [146]
22
C J Rutland
Fig. 6 Cyclic variation demonstrated by velocity magnitudes in a centre cut-plane of a PFI engine at
235CA BTDC (from [161], reprinted with permission from SAE paper 2007-01-0151, 2007,
SAE International)
23
Inc., the National Science Foundation [grant number 0500056], and the Air Force Office of Scientific
Research [grant number F49620-02-1-0348].
Authors 2011
FUNDING
The author acknowledges support over the years in
the broad area of LES for engineering applications from a number of sources, including the
US Department of Energy [grant numbers DE-FC0402AL67612, DE-FC26-06NT42628, and DE-EE000
0202], General Motors Research primarily through
the GM-UW Cooperative Research Lab, Caterpillar
24
C J Rutland
67
68
69
70
71
72
73
74
75
76
77
78
79
80
25
26
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
C J Rutland
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
like environment by the conditional moment closure model. SAE paper 2000-01-0200, 2000.
De Paola, G., Mastorakos, E., Wright, Y. M., and
Boulouchos, K. Diesel engine simulations with
multi-dimensional conditional moment closure.
Combust. Sci. Technol., 2008, 180(5), 883899.
Steiner, H. and Bushe, W. K. Large eddy simulation of a turbulent reacting jet with conditional
source-term estimation. Phys. Fluids, 2001, 13(3),
754769.
Triantafyllidis, A. and Mastorakos, E. Implementation issues of the conditional moment closure
model in large eddy simulations. Flow, Turbul.
Combust., 2010, 84(3), 481512.
Pope, S. B. PDF methods for turbulent reactive
flows. Prog. Energy Combust. Sci., 1985, 11(2),
119192.
Haworth, D. C. Progress in probability density
function methods for turbulent reacting flows.
Prog. Energy Combust. Sci., 2010, 36(2), 168
259.
Subramaniam, S. and Haworth, D. C. A probability density function method for turbulent mixing
and combustion on three dimensional unstructured deforming grids. Int. J. Engine Res., 2000,
1(2), 171190.
Kung, E. and Haworth, D. C. Transported probability density function (tPDF) modeling for
direct-injection internal combustion engines. SAE
paper 2008-01-0969, 2008.
Ihme, M. and Pitsch, H. Prediction of extinction
and reignition in nonpremixed turbulent flames
using a flamelet/progress variable model 2. Application in LES of Sandia flames D and E. Combust.
Flame, 2008, 155(12), 90107.
Tamagna, D., Ra, Y., and Reitz, R. D. Multidimensional simulation of PCCI combustion using
gasoline and dual-fuel direct injection with
detailed chemical kinetics. SAE paper 2007-010190, 2007.
Tamagna, D., Gentili, R., Ra, Y., and Reitz, R. D.
Multidimensional simulation of the influence of
fuel mixture composition and injection timing in
gasoline-diesel dual-fuel applications. SAE paper
2008-01-0031, 2008.
Moin, P., Squires, K. D., Cabot, W. H., and
Lee, S. A dynamic subgrid-scale model for compressible turbulence and scalar transport. Phys.
Fluids A, 1991, 3(11), 27462757.
Chumakov, S. G. and Rutland, C. J. Dynamic
structure models for scalar flux and dissipation
in large eddy simulation. AIAA J., 2004, 42(6),
11321139.
Chumakov, S. G. A priori study of subgrid-scale
flux of a passive scalar in isotropic homogeneous
turbulence. Phys. Rev. E, 2008, 78(3), 111.
Kerstein, A. Linear-eddy modelling of turbulent
transport. Part 3. Mixing and differential molecular diffusion in round jets. J. Fluid Mech., 1990,
216, 411435.
Goldin, G. M. and Menon, S. A scalar pdf construction model for turbulent non-premixed
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
27
28
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
C J Rutland
160
161
162
163
164
165
APPENDIX
Notation
k
ksgs
p
ui
Cij
Ck
Cm
G
Gij
Lij
Sij
Tij
aij
bij
e
n
nT
f
fj
r
t ij
t rij
D
G ij
~
00
^
kinematic viscosity
turbulent kinematic viscosity
generic scalar
generic scalar subgrid flux
density
subgrid stress tensor
residual subgrid stress tensor,
solenoidal
filter length scale; grid cell size scale
viscous stress tensor in the momentum
equation
averaged (RANS) or filtered (LES)
quantity
Favre averaged (RANS) or Favre filtered
(LES) quantity
fluctuating (RANS) or subgrid (LES)
quantity
test filtered quantity
Abbreviations
ATDC
BTDC
CA
CERFACS
CTC
EGR
GM TCC
IFP
LIF
ODE
PDF
PFI
PIV
APPENDIX
u(x)i =
29
G(x, y)u(y)i dy
(8)
rui
r
(9)
The total or instantaneous velocity can be decomposed into the resolved term and a reminder indicated by a double prime
ui = u~i + u00i
(10)
g
ui 99 6 0
(11)
Applying the filtering operation to the incompressible Navier Stokes momentum equation, and assuming that the filtering operation commutes with the
differential operators, results in the LES equation
ru~i u~j
rt ij
ru~i
p G ij
+
=
+
xi xj
t
xj
xj
(12)
(13)
1
dij t kk
3
(14)
30
C J Rutland
(15)
GT (x, y)u(x)i dy
(16)
(17)
(18)
(19)
(20)
^^ 1
Tij 2C2S 2D2 S
Sij + dij Tkk
3
(21)
(22)
(23)
APPENDIX
(24)
where
ksgs =
1
ui ui ui ui
2
(25)
The subgrid kinetic energy is obtained from a transport equation. The tensor coefficient is obtained
from the dynamic approach. Assuming the coefficient can be removed from the filtering integral
gives an algebraic form of the model that is simple
to implement
Cij = 2
Lij
Lkk
(26)
where Lij is the modified Leonard term from equation (19). Thus, the subgrid stresses obtain their tensor structure from the normalized modified Leonard
term via the dynamic procedure and their magnitude from the subgrid kinetic energy.
As explained here, the dynamic structure model
can be derived using the dynamic procedure.
However, it can also be viewed as a particular form
of a scale-similarity model [30]. The original scalesimilarity model was simply t ij= Lij [14]. However,
this approach did not work well and required an
additional viscosity term in what is commonly called
a mixed model. In comparison to the scale-similarity
model, the dynamic model uses the subgrid kinetic
energy to determine the magnitude of the term.
More recent work has resulted in a family of
dynamic structure models (Lu et al. [12, 42]). The
primary insight was to start with equation (24) and
31
Gij
Gkk
where; Gij =
ui uj
xk xk
(27)