You are on page 1of 21

Combustion and Flame 195 (2018) 232–252

Contents lists available at ScienceDirect

Combustion and Flame


journal homepage: www.elsevier.com/locate/combustflame

Detailed SGS atomization model and its implementation to two-phase


flow LES
Akira Umemura a,∗, Junji Shinjo b
a
Nagoya University, Nagoya 464-8603, Japan
b
Department of Mechanical, Electrical and Electronic Engineering, Shimane University, 1060 Nishikawatsu, Matsue 690-8504, Japan

a r t i c l e i n f o a b s t r a c t

Article history: A novel turbulent atomization model, which is physically closed itself and free of case-by-case parameter
Received 24 September 2017 tuning using experimental data, has been formulated and demonstrated in the framework of turbulent
Revised 18 November 2017
spray combustion large-eddy simulation (LES). Based on our accumulated research findings that elemen-
Accepted 17 January 2018
tary droplet/ligament generation is a deterministic phenomenon, not something random as considered in
Available online 2 March 2018
the conventional understanding, the model describes two dominant modes of turbulent atomization, i.e.
Keywords: the turbulent resonant mode and the Rayleigh–Taylor (RT) mode, in a physically straightforward manner.
Spray combustion Extending the baseline theory proposed in Umemura (2016), to a hybrid turbulent spray LES formulation
Turbulent atomization which includes both an Eulerian liquid jet core and Lagrangian droplets, the subgrid-scale (SGS) atom-
Subgrid model ization characteristics are completely detailed in this study. Using the LES-resolved turbulent Weber and
Two-phase flow LES simulation Bond numbers on the liquid core surface, the atomization mode and the SGS atomization characteris-
tics such as droplet size, number, ejection velocity and core regression velocity are all identified locally,
and the information is transferred back to the LES code as input information. Test cases of Diesel fuel
jets demonstrate that the present formulation well reproduces the turbulent spray behavior. Thanks to
the obtained detailed data, the spray formation process can be tracked both temporally and spatially,
from the initial head formation with edge atomization to the later core atomization and spray spreading.
It is essentially featured that the present turbulent atomization model works well without ambiguous
user input, contrary to the conventional way of spray simulation. This is a significant breakthrough to
urge paradigm shift in spray simulation, from unclosed/unpredictable to closed/predictable, which en-
ables drastic improvement in the accuracy of spray simulation and may exert a large impact on both
research studies and industrial applications.
© 2018 The Author(s). Published by Elsevier Inc. on behalf of The Combustion Institute.
This is an open access article under the CC BY-NC-ND license.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

1. Introduction Spray combustion is a very complicated combination of multi-


scale and multi-physical phenomena. Conventionally, most spray
Spray combustion is widely used in many engineering applica- simulation cases have been formulated in the framework of
tions such as aircraft and automotive engines and others. Combus- Eulerian–Lagrangian approach, namely the gas phase is solved
tor design is critical for engine performance and emission char- in the Eulerian way and droplets are treated as point parti-
acteristics. Since atomization plays a crucially important role for cles (Paradigm A in Fig. 1). The near nozzle (dense spray) re-
subsequent spray combustion, elucidation of this process is def- gion is omitted and treated as an inlet boundary that is set
initely needed for improving engine performance. Unfortunately, by the user. Even so, such an approach can incorporate near-
accurate prediction of spray is still not possible, because a precise droplet behaviors such as droplet evaporation and even larger-
physical understanding for modeling and prediction is still missing. scale droplet/gas interactions such as vapor mixing and combus-
Currently, the prediction capability of spray simulation is still very tion in the downstream region, by utilizing the fundamental stud-
limited to be fully used in engine design. ies on the evaporation/combustion and hydrodynamic character-
istics of a single droplet for various liquid fuels and ambient
gas conditions [1,2] and collisional droplet dynamics [3] (for the
spray equations) as well as detailed/reduced chemical kinetics and
∗ laminar/turbulent combustion models [4–6] (for the continuous
Corresponding author.
E-mail address: akira@nuae.nagoya-u.ac.jp (A. Umemura).

https://doi.org/10.1016/j.combustflame.2018.01.026
0010-2180/© 2018 The Author(s). Published by Elsevier Inc. on behalf of The Combustion Institute. This is an open access article under the CC BY-NC-ND license.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 233

Nomenclature α wave number of cascaded turbulent eddy =


2π /λ
a orifice or nozzle radius multiplying factor defined by Eq. (38)
= ( 2π ) /K 2
3
B = 5.57 γ surface tension coefficient
Bo Bond number in Eq. (10) L LES grid size
Cs = 0.1677 Smagorinsky constant 0 initial magnitude of SGS interface displacement
d droplet diameter in RT mode
D nozzle diameter δ SGS interface displacement
dN/dt regression speed of LES resolved interface ɛ parameter defined by Eq. (15)
F modification factor for regression speed interfacial LES-resolved shear stress
FL level-set function  integral length scale of nozzle flow turbulence
g normal LES-resolved liquid deceleration defined λ wavelength of cascaded turbulent eddy
by Eq. (9) λm wavelength of most unstable RT mode wave
H parameter defined by Eq. (40) λ∗ wavelength of resonating turbulent eddy
h initial distance of LES-resolved interface from tur- ν kinematic viscosity (liquid: ν L , gas: ν G )
bulent eddy center ρ density (liquid: ρ L , gas: ρ G )
h instantaneous distance of SGS interface from tur- τ Lagrangian time identifying a liquid element
bulent eddy center φ parameter defined in Eq. (36)
hm± upper (+) and lower (−) limit of h χ coefficient specifying, by Eq. (18), a ligament for-
I idem factor matrix (δ ij ) mation onset instant during turnover time for
K = 6.67 coefficient defined in Eq. (8) turbulent resonant eddy
kL liquid turbulent kinetic energy per unit density ψ parameter defined in Eq. (36)
Lb intact core length L interfacial shear rate of LES resolved liquid flow
 reference turbulent eddy size
n droplet ejection number flux defined by Eq. (32)
n
 outward unit vector normal to LES resolved inter- reacting gas field equations). To formulate the inlet spray condi-
face tions, typically, a Lagrangian spray model is extendedly applied
P LES resolved pressure up to the liquid jet injector. One example is implementing certain
Pa ambient gas pressure secondary atomization algorithms that gradually break up an in-
Por probability of eddy orientation for resonant jected large liquid blob (with the same diameter as the injector)
mode (=0.067) into smaller droplets (spray) [7]. The problem in this approach is,
qL turbulent velocity magnitude of cascaded eddy unfortunately, that such an atomization process does not reflect
ReD jet Reynolds number UD/vL the real atomization physics precisely, and therefore case-by-case
r radial distance parameter tuning is inevitably necessary to make the final simu-
s parameter defined by Eq. (37) lated spray characteristics adapted to experimental measurements.

Sd force exerted by mass-point droplets on gas This procedure strongly limits the applicability of spray simulation,
phase which is not a desirable situation. It is a natural desire that the in-
t time put spray droplets can be also predicted numerically in a physically
tRT transient ligament formation period in RT mode straightforward manner (Paradigm shift to B in Fig. 1) since spray
U injection velocity combustion performance crucially depends on the inlet spray char-
U LES resolved velocity acteristics.
Us LES resolved interface velocity This essential question of how the spray droplets are gener-
U s// tangential velocity component of LES resolved in- ated (atomized) has been a difficult question in spray combus-
terface tion. But recently, the understanding of spray physics has been

u fluctuating velocity improved gradually thanks to detailed experiments and numeri-
v y-velocity component cal simulations [8–13]. These achievements should be highly re-
vd SGS droplet velocity in laboratory frame marked, but when conducting practical combustor simulations,
vT turbulent velocity component normal to LES re- DNS (direct numerical simulation) is very limited to the near noz-
solved interface zle region [10–12] and not an appropriate approach due to its high
v∞ SGS droplet ejection speed cost. Therefore, for engine-scale simulation, LES (large eddy simu-
We Weber number in Eq. (10) lation) [14] is best suited for unsteady turbulent conditions [15,16].
WeD liquid jet Weber number = ρ L U2 D/γ Before describing our methodology in this line in detail, we first
We liquid jet Weber number = ρ L U2 /γ review several approaches that have been recently proposed for
X reference Cartesian coordinate used to describe spray simulations.
turbulent eddy flow One approach is to directly extend the DNS approach from the
x Cartesian coordinate along resolved interface or near-nozzle region to the downstream region [17,18]. Generated
Axial coordinate along nozzle jet small droplets are converted into Lagrangian particles. Adaptive
Y reference Cartesian coordinate used to describe mesh refinement (AMR) is typically used to satisfy the local grid
turbulent eddy flow requirement. While the accuracy can be generally high, the com-
y Cartesian coordinate normal to resolved interface putational cost is still relatively high since the primary atomiza-
Z reference Cartesian coordinate used to describe tion region should be fully resolved and numerical complexities
turbulent eddy flow increase if a method like AMR is used. For industrial applications,
z Cartesian coordinate in the third direction currently it is not easy to use this approach directly.
Another approach is to model the primary atomization by a
representative function. The surface density function (surface area
234 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 1. Paradigms of spray simulation. (A) is the conventional one where the droplet generation (primary atomization) is not directly simulated but given by some input
model. (B) is a novel paradigm where the droplet generation is directly considered consistently with the downstream region, typically formulated in a hybrid Eulerian/Eulerian
and Eulerian/Lagrangian manner.

per unit volume) is often used for this purpose since atomiza- that direction. Umemura has recently proposed a novel turbulent
tion is a process to increase the surface area. This approach it- atomization model that is free of case-by-case empirical parame-
self is generally sound, but the modeling of the net produc- ters and includes physical considerations of ligament/droplet gen-
tion term, which represents the surface increase due to atom- eration and thus spray generation [33]. This model is promising
ization (including decrease due to coalescence), is an important in enabling practical engine-scale spray simulation with more en-
key. Jay et al. [19] solved the surface density transport equation hanced accuracy than the conventional approaches. Here, we mean
where the production term is modeled by the growth rate of the a method “free of empirical parameters” as that which is physi-
Kelvin–Helmholtz (KH) instability with a nonlinear saturation ef- cally self-closed and does not need to determine unknown param-
fect. Navarro-Martinez [20] and Lebas et al. [21] similarly used the eters case by case even when the flow conditions are changed. It
surface density transport equation. The surface area is assumed to should be noted that this does not mean it is fully free of models
increase following the relaxation process to an equilibrium value or assumptions, but rather it means that these models are physi-
with some time-scale constant. In these studies, the models de- cally straightforward under the fundamental assumptions adopted
scribe the surface area increase in an overall sense, but details of in LES. By this, we intend to shift spray simulations from adap-
physical paths leading to droplet breakup are all wrapped up in tation (paradigm A in the conventional way) to prediction (new
the production term accompanied by some undetermined model paradigm B).
constants. As a result, this approach still needs parameter tuning Obviously, the use of LES means that, if the subgrid turbulence
case by case. Atomization physics is the very process that deter- could cause SGS (subgrid-scale) atomization, some SGS model
mines the most appropriate path, and it is suggested that some which are compatible with the fundamental assumptions (contin-
more physical understanding is needed to eliminate such model uum requirements, isotropic turbulence in a locally homogenous
constants. shear flow, turbulent inertial subregion) adopted in a standard LES
A different approach focuses on the correlation between turbu- [33] must be implemented to describe the SGS processes. There-
lent eddies and droplet breakup. Such an understanding has been fore, the LES should be able to capture (i) the liquid core behav-
pointed out [8,22] and a simple model based on this has been ior as a continuum, (ii) the liquid shear flow sustaining the liquid
proposed. Saeedipour et al. [23] correlated the droplet size with turbulence, and (iii) various instabilities caused by the presence
the dominant turbulent eddy size and proposed a simple model, of the real liquid jet surface exposed to the outer gas containing
in which droplets are directly ejected (without forming a liga- spray. This could be done if we use interface-capturing two-phase
ment) due to the liquid eddy motion. Very roughly speaking, this flow LES to describe turbulent atomization from the injected liquid
is posed in a proper direction but the specification of the domi- (Paradigm B).
nant turbulent eddy is ad-hoc and surface instability and ligament From this point of view, Umemura made theoretical considera-
formation mechanisms before droplet pinch-off are drastically sim- tions and proposed a turbulent atomization subgrid model for two-
plified. As a consequence, their modeling is still needs unphysical phase flow LES [33]. However, the previous paper focused on theo-
empirical tuning case by case. It will be shown, in this paper, that retical development basing the SGS atomization and did not detail
such a turbulent atomization model might be usable only for an the SGS atomization characteristics in the form which can be di-
immediately-near-nozzle region but cannot simulate the complete rectly implemented to the LES’s. Therefore, in the present paper,
spray jet that is targeted in spray combustion. detailed SGS atomization characteristics are derived and the use-
In order to tackle these difficulties, our recent research has fo- fulness of the model is demonstrated by implementing the model
cused on unveiling the underlying physics of atomization in de- in two-phase flow LES simulating the spray formation process of
tail. Theoretically, Umemura has proposed a role of self-generated turbulent Diesel fuel jets.
capillary waves in droplet breakup and has identified that the
droplet breakup is a deterministic phenomenon [24–31]. Based on 2. Detailed SGS atomization model
this finding, we called for the paradigm shift from A to B in a
plenary lecture of the 9th Asia-Pacific Conference on Combustion The primary atomization of an injected liquid jet usu-
(founded in Japan in 1997 by Prof. C.K. Law as the U.S. member) ally takes place at a relatively low temperature and consid-
in Korea, 2013 [32], and we have a responsibility for efforts toward erably large liquid-to-gas density ratio. For simplicity, neither
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 235

evaporation nor chemical reaction is considered in the present


study. Thus, the governing equations to be solved in the present
LES framework are incompressible filtered Navier–Stokes equations
[33],
∇ • U = 0,
     T 
∂ U 
ρ + U • ∇U  = −∇ P + ρν∇ • ∇ U  + ∇U 
∂t
 
  dN
+∇ • −ρ u   + ρL
u v∞ +γ (∇ • n
) n δ (FL )+Sd (1)
dt
res

where FL is the level-set function (FL = 0 means the LES-resolved


interface, FL > 0 the bulk liquid and FL < 0 the bulk gas) and δ (FL )
is the delta function to identify the interface. The third term on
the right hand side of the momentum equation is the unresolved
Reynolds stress (modeled by the Smagorinsky model), the fourth
term represents the effects of repulsive force in the resonant tur-
bulent atomization mode (later described in 2.7.3) and surface ten-
sion. The last term represents the interaction with the Lagrangian
droplets. The local gas-phase directed interface normal unit vec-
tor can be given by n  = −∇ FL /|∇ FL |. By using relations such as
ρ = ρG + (ρL − ρG )H (FL ), where H(FL ) is the Heaviside function, the
liquid and gas phases are treated in a unified way in the Eulerian
Fig. 2. Physical picture of LES-resolved interface and related atomization dynamics
framework. The level-set function follows on the element.
∂ FL  dN
+ U • ∇ FL = − |∇ FL | (2)
∂t dt
where the right hand side represents the surface regression due to each sub-phenomenon relevant to turbulent atomization is exam-
atomization. ined in detail. The interactions of unresolved interface and turbu-
SGS Lagrangian (point-particle) droplets follow lence inside a cell, which will finally lead to SGS droplet genera-
tion, is modeled. We derive those mathematical expressions that
dxd can be directly used to identify LES-grids involving SGS atomiza-
= vd , (3)
dt tion and determine the resulting atomized droplet characteristics.
dvd f1   
= UG − vd , (4)
dt τd
2.1. Turbulent interface model for LES
where the subscripts d and G refer to droplet and gas phase, re-
spectively, and the drag law for a sphere is used, which is [13,16]
Figure 2 shows a vicinity of each LES-resolved interface element
CD Red frozen at an LES calculation time step. The centered thick red hor-
f1 = , (5)
24 izontal line represents a planar LES-resolved interface element cal-
where Red = |vd − U  G |d/νG is the droplet Reynolds number, τd = culated in the two-phase flow LES with a uniform grid spacing L.
ρL d2 /18(ρG νG ) is the particle response time, and the drag coeffi- An orthogonal LES grid system of the same spacing L could be
cient is [13,16] locally re-constructed as shown in the figure for theoretical consid-

eration here. The LES-resolved interface velocity U  s and gas-phase-
24(1 + Red 2/3 /6 )/Red (Red < 10 0 0 ) oriented normal unit vector n  are calculated from the two-phase
CD = . (6)
0.4392 (Red ≥ 10 0 0 ) flow LES, and they determine the outward normal velocity compo-
nent Us⊥ = U s • n and tangential velocity vector U  s • (I − n
 s// = U n ).
Thus, for non-evaporating droplets, the force source term in The following formulation of SGS processes occurring in the sub-
Eq. (1) is expressed as sequent time refers to a Cartesian coordinate system O-xyz, which

1 md f 1   moves with the LES-resolved interface element and has the x- and
Sd = − ŪG − vd , (7) y-axes parallel to U  s// and n
 , respectively.
V τd
The interface element is tangentially exposed to a locally ho-
where V is the grid cell volume and md = π d3 ρL /6 is the droplet mogenous LES-resolved shear flow of liquid-phase shear rate L .
mass. To describe turbulent atomization caused by the SGS turbulence
It should be noted that GS (grid-scale) turbulence and liquid of LES, we need to know the constant energy flux function
breakup are directly resolved in the LES. Now, it is clear that a value ɛ which specifies the universal turbulent power spectrum
turbulent atomization model should produce SGS atomized droplet E (2π /λ ) ∝ ε 2/3 (2π /λ )−5/3 in the inertial subrange for each in-
information, i.e. (1) ejected droplet number flux n, (2) droplet di- terfacial grid cell. As detailed in [33], this can be done by con-
ameter d, (3) droplet ejection speed v∞ , and (4) liquid core regres- sidering a hypothetical LES-resolved energy-containing turbulent
sion speed dN/dt, combined with when and where in the flow field eddy of size  > L, for which the turbulent energy production
such droplet generation occurs. The proposed model in this study work by the Reynolds stress on the LES-resolved interfacial liq-
gives the information. uid shear flow balances with the molecular dissipation rare ɛ. From
In the previous paper [33], the theoretical aspect of new ap- this consideration,  is determined as  = 2.89 L, which is approx-
proach is discussed in detail. To implement the proposed model imated as  = 3 L in the present study for a reason mentioned in
to an interface-capturing two-phase flow LES code, more detailed Section 3. The interfacial SGS isotropic liquid turbulence produced
characterization of atomized droplets is necessary. In this section, by the LES-resolved interfacial liquid shear flow is found to have
236 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

the intensity from λ/4 to λ/2, respectively. The probability Por that the condi-
1
tion (i) is approximately satisfied may be estimated as Por = 0.067
2kL = L , K = 6.67. (8) [33].
K Consistent with the conventional measurement of turbulent en-
where K (and later introduced B = 5.57) as well as the value 2.89 ergy spectrum, each cascaded eddy is assumed to keep its own
are determined from constants used in the standard turbulence vortical structure and move at its center’s LES-resolved velocity
models. Of course, theses relations hold primarily in a strong tur- during its lifetime of L −1 (eddy turnover time). Then, the evo-
bulence limit, with the LES grid spacing L > > λ. Nevertheless, lution of the SGS interface deformation must be randomly re-
even in a weak turbulence where λ ∼ L, the proposed SGS atom- initialized with a period of L −1 for any SGS cascaded liquid eddy
ization model can also handle this situation, because the SGS tur- of wavelength λ = 2π /α , so as to take into account the cascad-
bulent atomization would not play a large role and LES-resolved ing processes and the random realization of various turbulent eddy
turbulence eddy and atomization (=resolved breakup) would ap- orientation ineffective for spray formation [33]. Thus, the SGS in-
pear instead. Similarly in an extreme limit where L < < λ, the terface element which coincides with the planar LES-resolved in-
present formulation reduces to DNS. In this sense, the present for- terface element at a re-initialized time t = 0 is exposed to the fol-
mulation is perfectly consistent in the LES framework [33]. lowing Y-component of the eddy velocity that tends to deform the
The SGS liquid turbulence determined by Eq. (8) may cause SGS SGS interface element:
atomization from the interface element. Besides, the two-phase
4
flow LES may give the interfacial bulk liquid a local normal de- vT = √ qL cos [α (X − hAB Lt )] sin (α hAB ) cos (α Z )
celeration velocity 3
for 0 < α hAB < π /2
Us • n 
 s • DUs − U

g = − 2 U s • ∇ n
 •U
s (9) and
Us Dt
4
vT = √ qL cos [−π + α (X + hCD Lt )] sin (−α hCD ) cos (α Z )
due to the surface-normal deceleration and the interface curvature. 3
This may also cause SGS atomization from the interface element 4
due to the Rayleigh–Taylor (RT) instability. = √ qL cos [α (X + hCD L t )] sin (α hCD ) cos (α Z )
3
Hence, the purpose of the SGS atomization model is to predict
a spray element for π /2 < α hCD < π , where qL is calculated from the turbulent cas-
produced from the LES-resolved interface element
in terms of 2kL and g, or in their non-dimensional forms, cade relation
 1 / 3
ρL ( 2 k L )  qL 1 λ
Weber number : W e = (10a) = . (11)
γ L  K 

The two initial SGS interface locations can be consolidated by a


1 (ρL − ρG )g2
and Bond number : Bo = , (10b) single random variable α h which ranges over 0 < α h < π and rep-
( 2π ) 2 γ resents a random realization of initial interface location AB or CD
where ρ L and ρ G are the liquid and gas densities and γ is the sur- for 0 < α h ≤ π /2 or π /2 ≤ α h < π , respectively, and we express
face tension coefficient. Expected operations are; When We or Bo 4
is in a range for which SGS atomization takes place, (i) the sta- vT = √ qL cos {α (X ∓ hLt )} sin (α h ) cos (α Z ) (12)
3
tistical characteristics of locally atomized droplets are determined
on the basis of the We or Bo values and (ii) input into the La- A linear stability analysis to the effects of vT and g [33] shows
grangian spray tracking scheme. At the same time, (iii) the local that the subsequent SGS interface deformation behavior is classi-
LES-resolved interface element is regressed by the total volume of fied as follows:
the atomized droplets so that the entire liquid mass might be con- 2 1 ρL +ρG (α√
h )2 We
(i) 0 < [( λl ) − 2 ]( λ )
Bo l
= 2 [ ρL ] generating a propa-
served. Later, at the end of this section, Fig. 14 will again summa- 2B
rize the numerical procedures in the present formulation. gating capillary wave,
(ii)
 2   
2.2. Atomizing grid cell l Bo l 1 ρL + ρG (α h )2W e
− = √ (13)
λ∗ 2 λ∗ 2 ρL 2B
An analysis of dynamic equilibrium between liquid- and gas-
phase turbulences at the LES-resolved interface reveals that tur-
bulent atomization is characterized by interfacial liquid-phase tur- causing a diverging interface deformation resonant with a par-
bulence [33]. The interfacial grid cell contains SGS liquid turbu- ticular cascaded eddy of wavelength λ = λ∗ ,
lent eddies which are cascaded from an energy-containing eddy of (iii) ( λl )2 − Bo
2 < 0 causing a diverging interface deformation due to
wavelength  = 3 L. Referring to a local Cartesian coordinate sys- RT instability for any α h.
tem O-XYZ, these cascaded eddies can be represented as Fourier
components with various wavelength λ < . For isotropic turbu- The excitation of capillary waves in (i) plays an important role
lence, the Y-axis may orient to an arbitrary direction. However, SGS to describe the formation of self-sustained unstable convective
atomization effective for spray formation occurs only when (i) the waves and turbulent transition for laminar issued jets [24–28]. In
Y-axis is normal and X- and Z-axes are parallel to the planar LES- the present model, we are mainly interested in (ii) and (iii); A SGS
resolved interface element (i.e., the X-,Y- and Z-axes are parallel to spray element is produced from liquid ligaments that develop out-
x-,y- and z-axes), and (ii) as shown in Fig. 3, the converging liquid ward due to (a) capillary wave resonance with a particular cas-
eddy flow impinges on the interface element for a representative caded liquid turbulent eddy or (b) RT instability associated with
Fourier component of wavelength λ [33]. There are two possible outward inertial force g acting on the interfacial bulk liquid. We
interface locations relative to the eddy center as indicated by hAB will analyze (ii) in more detail and explore its relation with (iii).
and hCD , which are random variables ranging from 0 to λ/4 and The solution of Eq. (13) is given as
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 237

Fig. 3. LES-resolved interface positions relative to the center of a cascaded liquid eddy of wavelength λ.

⎧ 1/3 
⎨ ρL + ρ G (α h√)2 We √ 1 / 3  √ 1 / 3  region enclosed by these two √ curves and the curve of Eq. (16). In
 ρL 1− 1−ε + 1+ 1−ε : ε≤1 ρ
= 4 2B (14) the figure, (α h )ε=1 = [2 ρ +Lρ 2We2B ]1/2 ( Bo
6 )
3/4 is also plotted for ref-
λ ∗
⎩ Bo 1 √  L G
2 6 cos 3 tan
−1
ε−1 : ε>1 erence.

where
 √ 2  3 2.4. We-Bo map for SGS atomization
ρL 2 2B Bo
ε= 2 (15)
ρL + ρG (α h ) We 2 6
The critical curves (α h)c1 and (α
√h)c2 intersect with the curve of
Eq. (16) at α h = (α h )c1 = sin−1 ( 3K/16/32/3 ) = 0.353 and α h =
#
Therefore, λ∗ takes a random value corresponding to an
α h−realization for given We and Bo. (α h )c2 # = π − 0.353. Using these α h values, the following two
curves C1 and C2 can be drawn on the W e − Bo plane as shown
2.3. Conditions for SGS atomization onset in Fig. 6.
 √
2.3.1. SGS condition 3 2B(18 − Bo)
C1 : 0.353 = → W e = 189.6(18 − Bo),
Turbulent eddies targeted by the SGS model have a wavelength We
λ shorter than the LES grid cell size L, i.e., λ/ L = (/ L )(λ/ ) = Bo = 18 − 0.005274 W e (19a)
3(λ/ ) < 1. The critical condition /λ = 3 is substituted into Eq.
(13) to yield
  √

ρL 3 2B(18 − Bo) 3 2B(18 − Bo)
(α h )λ= L = . (16) C2 : π − 0.353 = → W e = 3.039(18 − Bo),
ρL + ρG We We
Bo = 18 − 0.3291 W e (19b)
2.3.2. Atomizing liquid ligament formation condition
For a resonant liquid eddy, the subsequent dominant interface Resonant atomization takes place from LES-resolved interface
deformation is brought about by the same eddy flow. Hence, its elements whose We and Bo values are located on the right-hand
amplitude δ is proportional to the elapsed time t as side of curve C2 . On the right-hand side of curve C1 , any α h re-
alization between (α h)c1 and (α h)c2 occurs at an equal probability
1 ∗ of 1/2π , while the lower α h bound is determined by Eq. (16) be-
δ ( X = h L t ) = λ A L α h L t with AL
2 tween curves C2 and C1 . Other curves C3 ∼ C5 in the figure will be

4 ρL sin (α h )
  2 explained later.
= √ 3. (17)
3K ρL + ρG αh λ∗ 2.5. SGS interface regression process in an α h-realization
Once δ exceeds λ∗ /8, the restoring surface tension force can
no longer confine the bulk liquid [29–31,33]. A free liquid liga- For a given atomizing α h−realization, we consider the defor-
ment disintegrating into droplets is formed by the inertial liquid mation of an SGS interface element which coincides with the lo-
eddy flow into the ligament (see Fig. 4). Putting t = χ L −1 in Eq. cal LES-resolved interface element at a re-initialization time (=an
(17), the condition for this free liquid ligament formation to occur LES calculation time) t = 0. For the Fourier component eddies un-
within the lifetime of the resonant eddy is expressed as der consideration, it is enough to consider the SGS interface be-
√   −2/3 havior within a representative eddy of wavelength λ∗ as shown in
3K ρL + ρG 1
χ= < 1. (18) Fig. 4 for the case of 0 < α h ≤ π /2. The eddy flow deforms the SGS
16 ρL sin (α h ) λL ∗ interface. Once a free liquid ligament begins to be formed, the sub-
Substituting Eq. (14) into Eq. (18) determines the lower and sequent ligament development is achieved by inflow of the iner-
upper limits of α h, i.e., (α h)c1 for 0 ≤ α h ≤ π /2 and (α h)c2 for tial liquid eddy flow illustrated by arrowed solid streamlines in
π /2 ≤ α h ≤ π as a function of We and Bo. In the following, (ρL + Fig. 4. Correspondingly, the reference plane of the instantaneous
ρG )/ρL ∼
= 1 is assumed to simplify numerical presentation. SGS interface (defined in Fig. 4) decreases (for h ≤ π /2) or increases
Figure 5 shows the (α h)c1 and (α h)c2 curves on the α h − Bo (for h ≤ π /2) its distance from the eddy center, h (t). The instanta-
plane for W e = 10 0 0. SGS turbulent atomization may occur in the neous liquid velocity entering the developing ligament is expressed
238 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 4. Formation of a free liquid ligament.

Fig. 5. α h realization range relevant to SGS atomization. Fig. 6. SGS atomization regimes in We-Bo space.

by the positive part of Eq. (12) with h replaced by h (t). Since


  lifetime, h has a lower limit hm− or an upper limit hm+ for each
∗ 2 dh
 λ∗
4
λ∗
4 4λ∗
2   atomizing α h- realization. Putting τ = (1 − χ )L −1 (0 < χ < 1) in
∓λ = dx dzvT = √ qL sin α h
dt ∗
− λ4 − λ4

π2 3 Eq. (22) and applying Eq. (11), we obtain
(− : h ≤ π /2, + : h ≥ π /2), (20)    2 / 3
1 ± cos (α h ) 16 1
we have the following ordinary differential equation for h . η≡ exp (1 − χ ) √ (24)
1 ∓ cos (α h ) π 3K λ∗
dh 4  
∓ = √ q sin α h . L (21)  η − 1
dt π2 3 α hm± = cos−1 ± (25)
This equation is integrated subject to the initial condition h =
η+1
h, yielding The reference interface regression distance is given by
  |h − h m± |, whose dependence on α h is illustrated in Fig. 7 together
1 ± cos α h 1 ± cos (α h ) 8 1
η≡ = exp [2ωτ ] , ω = √ qL ∗ with the variation of resonant eddy wavelength λ∗ .
1 ∓ cos (α h ) 1 ∓ cos (α h ) π 3 λ
(22)
2.6. Breakup characteristics intrinsic to an unsteadily stretching
 η − 1 liquid ligament
α h = cos−1 ± (23)
η+1 As shown in Supplementary material A, the central velocity of
where τ denotes the time measured from the ligament formation vT in Fig. 4 plays the most important role on the resulting axi-
onset instant (t = χ L −1 ). Since τ is upper-limited by the eddy symmetric liquid ligament shape r(t, y). Utilizing a 1D simulator
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 239

Fig. 7. Interface regression distance and resonant eddy wavelength vs α h.

Fig. 8. Evolution of ligament and droplet breakup. Here, r and v denote the liquid surface radius and the axial liquid velocity, respectively.

developed in [24], we calculated a liquid plug


flow∗ jet emanat- [34] which is established in ultrasonic atomization studies may be
ing at a temporarily varying velocity v = √4 2kL ( λ )1/3 sin(α h ) adopted to correlate the average atomized droplet diameter with
3
from a circular orifice of radius a = λ∗ /4. To mimic the liquid liga- the resonant eddy wavelength. In fact, the diameter of the droplet
ment evolution in an α h realization, the orifice initially located at produced by the shortwave breakup in this calculation is estimated
y = 0 should move back with time t by a distance of |h − h |. How- to be d/2 = 0.6(λ∗ /4 ) = 0.15λ∗ , which is slightly smaller than the
ever, |h − h| could be at most a = λ∗ /4, whereas a first jet disin- Lang equation d = 0.34λ∗ . The local velocity of the thin liquid col-
tegration occurs at a distance of O(10a) from the orifice. Therefore, umn increases with the distance from the orifice and determines
the orifice was fixed at the same location y = 0. The thick lines in respective atomized droplet velocities.
Fig. 8 show the calculation results for the case of We = 500 and It is also
 notable that the first droplet breakup takes place
γ
α h = π /2. The simulator is available until a first jet disintegration around t
a ρL a = 5. The ratio of this elapsed time t to the resonant
takes place somewhere along the jet. In the figure, the jet shape √
5X 23/4 KB
and axial velocity distribution, predicted from inertial analyses in eddy lifetime L −1 is t L = 8 (α h )
= 4.08 for α h = π /2.
Supplementary material A, are also drawn with the thin lines for
comparison.
It is found that the jet becomes a thin liquid column and dis- 2.7. Resonant atomization characteristics
integrates to droplets. The jet tip contracts by a surface tension
effect and generates an upstream propagating capillary wave. As a The proposed SGS model characterizes the locally atomized
result, the jet tip liquid bulb is formed and broken off at its up- droplets as quantities averaged over the α h realization ensemble
stream necked portion, resulting in a considerably smaller diame- on the atomization map of Fig. 6. Therefore, the atomization char-
ter of droplet at a shorter distance from the orifice, compared to acteristics are expressed as functions of We and Bo. In the previ-
the well-known constant velocity jet case. The same short-wave ous paper [33], the strong turbulence limit was considered, and
breakup mechanism [24] operates repeatedly for the newly pro- it was assumed for simplicity that hm− = 0 and hm+ = π . In this
duced jet tip. This feature is similar to the disintegration of a liquid section, complete formula necessary to calculate the local atom-
ligament produced by RT instability [29–31], and the Lang equation ization characteristics are presented for the case of ρ L > > ρ G . The
derived atomized droplet characteristics give the input data to the
240 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 9. Ensemble-averaged resonant eddy wavelength for droplet diameter estima- Fig. 10. Ensemble averaged regression distance for regression speed estimation.
tion.

Lagrangian spray tracking starting at each atomizing LES-resolved


interface element.

2.7.1. Atomized droplet diameter


As explained in Section 2.6, the liquid ligament disintegration
in turbulent atomization is different from the Rayleigh breakup
[35] adopted in conventional spray simulators. Instead, the aver-
aged atomized droplet diameter is given by the Lang equation.

(26)

(27)

Figure 9 shows the dependences of λ∗ / on We and Bo.

2.7.2. LES-resolved interface regression speed


The average regression speed of reference interface in an α h re-
alization is derived by dividing |h − h m± | by L −1 . For an interfa-
cial grid cell (given We and Bo), its ensemble average yields the
LES-resolved interface regression speed as

Fig. 11 Schematic of ligament evolution and droplet generation. Here, τ denotes


Lagrangian time identifying an ejected liquid element.

(29)

The solid lines in Fig. 10 show the dependences of



|α h − α h m± | λ  on We and Bo.

(28) 2.7.3. Average ejection velocity v∞ of atomized droplet and droplet


ejection conditions
As illustrated in Fig. 11, a yellow atomized droplet mass could
where be traced back as a yellow ligament element mass and a yellow
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 241

eddy element mass in a Lagrangian √ description (see Supplemen-


tary material A). Therefore, (4/ 3 )qL sin(α h ) represents an out-
ward average velocity of the disintegrated droplets from the free
liquid ligament. However, it is important to recall that the reso-
nant atomization occurs at large We and only for the particular
turbulent eddy orientation as accounted by the probability Por in
Eq. (29). We have to consider what happens in other turbulent
eddy orientations that are not effective for spray formation. Those
atomized droplets that are not effective for spray formation should
be excluded to avoid unnecessary complicating calculations in a
two-phase flow LES simulator as described below.
In an α h-realization, the stretching liquid ligament formed
within the eddy lifetime is disintegrated to a√series of droplets
that√ have an outward velocity between (4/ 3 )qL sin(α h ) and
(4/ 3)qL sin(α hm± ) relative to the LES-resolved interface. As men-
tioned in Section 2.6, this ligament disintegration may take place
after the eddy lifetime, i.e. in another realization excluded in the
model. The atomized droplets that have small outward velocities
stay near the LES-resolved interface and they are likely to coalesce
with the bulk liquid when the turbulent eddy orientation changes.
In addition, small cascaded eddies are convected in random direc-
tions by larger eddies. The convection of resonant eddies parallel Fig. 12. We-dependence of droplet ejection speed.
to the LES-resolved interface does not cause any significant effect
on turbulent atomization on average because the normally extend-
ing liquid ligaments also move with the convected eddies. How-
ever, the convection of resonant eddies normal to the LES-resolved
interface adds a random translational velocity to √ the atomizing liq-
uid ligament. Therefore, simply calculating (4/ 3 )qL sin(α h ) is
not appropriate for characterizing the ejection speed of atomized
droplets forming a spray.
We have to consider the atomized droplet behaviors suffered
from random turbulent perturbations as follows. The largest con-
vective velocity relevant to the SGS turbulence is represented by
the turbulent intensity 2kL . Thus, each outward droplet (liquid
ligament element) velocity is superimposed by a completely ran-
dom velocity V whose amplitude is equal to 2kL . If the resulting
outward droplet (liquid ligament element) velocity relative to the
LES-resolved interface becomes negative, the droplet will collide
and coalesce with the resolved interface in another turbulent real-
ization excluded in the model. As a result, only those droplets that
have positive values contribute to the spray formation. In Supple-
mentary material B, it is explained how to calculate the ensemble-
averaged outward droplet velocity v∞ accounting for such turbu-
lent perturbations, and we have Fig. 12 for the case of Bo = 0 to
express v∞ as a function of We. In the present model, hence, we
Fig. 13. We-dependence of modification factor F.
simply assume that all locally atomized droplets effective for spray
formation have the same ejection velocity

uid entering the ligament at small velocities is considerably small


compared to that at large velocities.
(30) Figure 11 allows us to describe an atomizing liquid ligament as
a series of atomized droplets ejected from the LES-resolved inter-
for 54.8 < We < 10 0 0 of practical interest. Then, serious numeri- face element. In the laboratory system, each atomized droplet of
cal problems arising droplet collisions near the interface can be diameter d has the initial velocity
avoided.
The liquid volume that is not effective for spray forma-
tion should be subtracted from the interface regression volume
λ∗ 2 |h − hm± | in each α h-realization. Therefore, |h − hm± |in Eq. . (31)
(28) should be replaced with f |h − hm± |, 0 < f ≤ 1, and
 f |h − hm± | = F |h − hm± |. In Supplementary material B, it is ex- The number n of such droplets ejected from a unit area of the
plained how f is estimated to modify dN/dt by F. Figure 13 shows LES-resolved interface element per unit time is calculated from
the dependence of F on We for the case of Bo = 0. It is found that F
may be approximated as F = 1 in the range of We∼O(100) of inter-
est where the resonant atomization takes place near the injector
exit. This is because qL does not change significantly from 2kL
with decreasing λ∗ from  at such We, and the volume of the liq- (32)
242 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Since the isotropic turbulence assumed in LES has no LES- to ψ during the transitional period. Then, defining
resolved momentum orientation, the repulsive force must be ap-    
λ ρ ( L  ) 2  γ λ∗
s = L =
plied as a pressure of acing on the LES- g ρ γ g2
√  √ 
resolved interface element upon this resonant atomization for the K We λ∗ W e λ∗
momentum conservation in the LES simulation (see Eq. (1)). This = √ = 1.06 √ , (37)
(2π ) Bo  Bo 
point is different from RT atomization considered next.
in Eq. (36) is expressed as
⎧1 1 − χ s
⎨ : ≥1
2.7.4. Effect of positive Bo after onset of free liquid ligament = s 2 1 − χ . (38)
formation ⎩1 − 1 s s
1 −χ
:0≤ ≤1
Once a free liquid ligament is established, the restoring sur- 2 1−χ
face tension force acting on the resonant liquid eddy is at the hol- The differential Eq. (37) can be solved analytically. hm± can be
low surface portion alone. This implies that the factor Bo/2 in Eq. calculated from the following equations.
(13) should be changed to Bo and that the inertial force g > 0 act-  2
ing on the bulk liquid may enhance the liquid flow into the lig- α hm± 1 + H∓ φ φ
ament and thereby the interface recession speed when Bo/2 < tan = − 12 −
√ √ 2 1 − H∓ ψ ψ
/λ∗ < Bo. This consideration leads to /λ∗ = Bo (instead of
/λ = Bo/2) for the critical resonant eddy wavelength. Then, Eq.
∗ φ
for >1 (39)
(13) yields ψ
where
 
Bo √

(α h )2W e 
φ 2
Bo −
2
Bo = √ , (33) tan α2h + φψ − ψ − 12
2 2B H∓ =  
φ 2
and the condition (18) is rewritten as tan α2h + φψ + ψ − 12
√ ⎛  ⎞
3K 1
  2 / 3  2 √ √ 
< = Bo1/3 φ ( 1 − χ )0.225 3(2π )2 Bo 
16 sin (α h ) λ∗ × exp ⎝∓ −1 √ ⎠.
ψ K We λ
or
√  (40)
−1 3K −1/3
(α h )3 ≡ sin Bo < αh and

16   2
√  α hm± φ
ψ
φ tan α2h + 1
< π − sin
−1 3K −1/3
≡ (α h )4 tan = 1− tan ⎣tan−1 
Bo 2 ψ 
φ 2
16 −1
ψ
for Bo > 9. (34)  ⎤
2   2 / 3
φ 8 (1 − χ )
⎦− φ .
Substituting (α h)3 and (α h)4 into Eq. (33), we obtain ∓ −1 √
ψ π 3K λ ψ

2B
C3a : W e =  √
3/2
2 Bo , φ
for <1 (41)
sin
−1 3K
16
Bo−1/3 ψ
√ As α h approaches 0 or π where the resonant eddy contribution
2B 3/2 vanishes, the RT contribution becomes dominant. The α h can cross
C3b : W e =  √  2 Bo . (35)
−1 3 K 0 or π and the reference interface plane can penetrate into the
π − sin Bo−1/3 neighboring liquid turbulent eddy region until t = (1 − χ )L −1 ,
16
and |α h − α hm± | is significantly expanded depending on the φ / ψ
which are drawn as curves C3a and C3b in Fig. 6. Hence, we need value. As a result, dN/dt calculated from Eq. (28) may be increased
to modify Eq. (28) for the We and Bo values below curve C3b in to a great degree at smaller We as shown by the dashed lines in
Fig. 6. Fig. 10. Correspondingly, v∞ should be replaced by v∞ = 2kL +

In this case, similarly to the RT instability treated in the next 2 × 0.225 3 λ∗ g in Eqs. (30) and (31). However, the dN/dt value
subsection, g > 0 would cause the steady reference interface reces-
√ that is not modified by the above-mentioned Bo effect should be
sion speed 0.225 3 λ∗ g in the absence of turbulent eddies [31]. used to calculate the repulsive pressure −ρL (dN/dt ) 2kL on the
Thus, using this expression, Eq. (21) should be changed to LES-resolved interface element.

dh 4   √  
∓ = √ qL sin α h + 0.225 3 λ∗ g ≡ φ sin α h + ψ , 2.8. RT atomization characteristics
dt π 3
2
√   √  
φ 2 W e  1/6 W e  1/6 Because of capillary waves propagating from various portions,
= √ = 0.0957 √ . (36) the instantaneous SGS interface of an LES-resolved interface el-
ψ 3X 0.225π 3
Bo λ ∗
Bo λ∗

ement may have various wavelength deformation components,
among which RT instability may be excited for /λ < Bo/2
It takes a period of tRT = λ∗ /g for the steady reference inter-
when the interfacial bulk liquid is exposed to g > 0. As We de-
face recession to be established in the resonant eddy region [31].
creases to a value below C3b in Fig. 6, there emerges a possibil-
The multiplying factor is introduced to account for the transi-
ity that the RT instability is predominantly excited from the ini-
tional contribution. We assume that the reference interface reces-
tial free liquid ligament formation stage according to = 0 eβ t
sion speed increases linearly from zero (at the re-initiation time)
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 243

where 0 is the initial wave amplitude of a candidate wave- v∞ is given by


length of /λ < Bo/2 and the growth rate is given by β =
 
γ 2
ρL  (2π ) 2
 
λ [Bo − 2 ( λ ) ]. The dominant RT instability
1 3/2 1/4

. (46)
wave has the following wavelength λm and maximum growth rate
β max depending on Bo.
Eqs. (26), (31) and (32) can be used to calculate d, vd and n by
replacing λ∗  with λm .

2.9. Implementation in two-phase flow LES

All ensemble-averaged SGS atomization characteristics are ex-


pressed as functions of We and Bo. In two-phase flow LES in which
both We and Bo change continuously in space and time, neigh-
(42a) boring interfacial grid cells must be in a similar state. The local
turbulent eddy lifetime may be longer than the calculation time
step determined by the CFL condition and each SGS interface de-
formation process occurring during the eddy lifetime may span
over neighboring interfacial grid cells. Nevertheless, the ensemble-
averaged SGS atomization characteristics, which are also time-
averaged quantities over the eddy lifetime, can be treated as in-
stantaneous local quantities in two-phase flow LES. This is because
(42b) these quantities are partitioned evenly among the neighboring grid
cells and time steps involved in a SGS atomization process for
given We and Bo. Figure 6 and the SGS atomization characteris-
The liquid ligament formation onset time tRT is determined
tics can be tabulated and incorporated in a numerical simulation
from the condition 0 eβmax tRT = λm /8, yielding
code. It is not difficult to find their analytical correlations for the
⎧   ρ 
⎨ln 18 λ∗3 Bo −3/2 −1/4 respective regions classified by C1 ∼ C5 in Fig. 6.
γ ( 2π )
2 
L 2 √ 1 : 18 ≤ Bo ≤ 54
tRT = √ 0
 3[Bo−18]
(43) Figure 14 summarizes the flow of numerical procedures at each
⎩ln 3 λ∗  ρL  (2π )−3/2 2−1/2 33/4 1 : Bo ≥ 54
8 0 γ Bo3/4 calculation time step t = tn in the present LES framework. First,
√ √ √ the local interfacial We and Bo are determined by the Eulerian
3 L 6 3 3 L 2 3 3 L λ∗ 3 λ∗
where use is made of 8 0 Bo = 8 0 Bo = 8 0  = 8 0 . This We- part. At each LES-resolved interface element that has a velocity U s
independent ligament formation onset time should be compared to the lig- and outward normal unit vector n, the tangential shear stress

ament formation onset time for the resonant eddy at the same Bo value,  s • (I − n
√ acting in the direction of U n ) can be calculated. This de-
i.e., χ L −1 = 16
3K 1
( λ∗ )2/3 1 . For the realization α h = π /2 which min-
sin(α h ) L termines the magnitude of liquid-phase shear rate L from the
χ L −1 , the critical condition (tRT = χ L −1 ) is expressed as
imizes relation | | = ρL (Cs L )2 L 2 , provided that molecular viscosity is
#  $ negligibly smaller than eddy viscosity. Thus, we can calculate the
−2 1 λ∗ Bo value of We according to Eq. (10a). We can also determine the
We = 19.6ln [Bo − 18]Bo−2/3 (18 ≤ Bo ≤ 54 ),
8 0 3 2
value of g according to Eq. (9), and the value of Bo according to
√ ∗  Eq. (10b). These local We and Bo values specify a point on the
−2 3 λ
We = 1.77ln Bo5/6 , (Bo ≥ 54 ) (44) map of Fig. 6. When the point is located within the SGS atom-
8 0
ization region, the SGS atomization model equations relevant to
Considering the possible 0 value 
in the non-resonant case, it the point are used to produce necessary input information for nu-

merical atomization procedures and subsequent tracking of atom-
is reasonable to assume that ln( 18 λ∗3 3 λ∗
2 ) ≈ ln ( 8 ) = 1. Then,
Bo
0 0 ized droplets in the Lagrangian solver, namely, droplet diameter d,
Eq. (44) is expressible as curve C5 in Fig. 6, below which the RT droplet ejection speed v∞ , surface regression speed dN/dt, droplet
instability appears dominantly.
√ velocity in the laboratory frame vd = U  s + [v∞ − g(t − tn )]n̄ and the
If the resulting steady interface recession speed 0.225 3 λm g
ejected droplet number flux n. The intermittent nature of the mod-
is larger than the turbulent intensity 2kL , the outward liquid eled turbulent atomization can be mimicked by ejecting new at-
ligament development and the interface regression are not inter- omized droplets generated by the above input information, at an
rupted by re-initialization nor random change to unfavorable tur- interval of turnover time L −1 or RT characteristic time λm /g.
bulent eddy orientations. This critical condition is expressed by The Eulerian–Eulerian part would be a central part of the
curve C4 in Fig.√ 6, which consists of W e = 1.996Bo (18 ≤ Bo ≤ 54) solver. To solve a flow field including a liquid/gas interface, several
and W e = 14.7 Bo (Bo ≥ 54). Since C4 is located above curve C5 , methods have been proposed [10–13], and any of these interface-
the RT atomization in regions RT(1) and RT(2) can be excited re- capturing method can be a baseline method for implementing the
gardless of turbulent eddy orientation. Thus, at C5 , the atomization present model. The SGS atomization model and a Lagrangian solver
characteristics become discontinuous. On average, turbulent ran- could be readily added to such a baseline code.
dom fluctuations in instantaneous SGS interface recession speed
are cancelled out. The LES-resolved interface recession speed coin-
√ 3. Demonstration
cides with the steady interface recession speed 0.225 3 λm g and
is expressed as We implemented the proposed SGS atomization model to sim-
ulate the process of spray formation of turbulent Diesel fuel
jets. For a baseline code in this study, the TATM-MEX (Turbu-
lent AToMization -MicroEXplosion) code developed by the authors
is used, which is based on the CIP (Cubic Interpolated Pseudo-
(45) particle or Constrained Interpolation Profile) formulation [10,36].
244 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 14. Flow chart of numerical procedures in the present LES code.

Table 1
Flow conditions. The upper half shows common values for three cases.

Ambient pressure Gas density ρ G [kg/m3 ] Liquid viscosity ν L [Pa s] Gas viscosity μG [Pa s] Gas velocity UG
Pa [MPa] [m/s]

3 34.5 790 × 10−6 18.9 × 10−6 0

Case Nozzle Liquid density Liquid velocity Liquid Surface tension Bulk liquid Bulk liquid
diameter D ρ L [kg/m3 ] UL = U [m/s] kinematic γ [N/m] Reynolds Weber number
viscosity ν L W eD = ρL Uγ D
2
[mm] number
[m2 /s] ReD = UD
νL

1 0.3 848 200 0.93 × 10−6 30.0 × 10−3 6.4 × 104 3.4 × 105
2 3.6 10 0 0 100 0.79 × 10−6 70.0 × 10−3 4.6 × 105 5.1 × 105
3 0.1 848 100 0.93 × 10−6 30.0 × 10−3 1.1 × 104 2.8 × 104

Interface-capturing is based on the level-set method (Eq. (2)) and fluctuations are imposed whose mean turbulent intensity is about
combined with a VOF (volume of fluid) method to assure vol- 5% of the mean flow velocity U using the digital filtering (DGF)
ume conservation. The effectiveness of the TATM-MEX code has method for a pipe flow [42]. High pressure effects, such as su-
been demonstrated, in the studies of turbulent atomization [10], percritical processes are not considered since they are beyond the
droplet pinch-off [10], droplet heating [37], emulsion droplet puff- scope of this study. For a very short nozzle, the complete spray
ing/microexplosion [38] and droplet vapor mixing and combustion jet is never realized [8,27,43]. In an appropriately long nozzle, the
[39,40], and further validation of interface capturing capability will cavitation phenomenon occurring in the nozzle inlet may serve as
not be repeated here. The SGS atomization model and a Lagrangian a similar role to a tripping wire that greatly shortens the turbulent
solver have been added to the above baseline code, and the total transition distance in the strong shear flow which the emanating
capability will be validated hereafter. liquid has. Thus, such a turbulent jet can be also treated exactly
Importantly, the present SGS atomization model predicts the at- in the same code settings, provided that freely suspended cavities
omized droplet diameters independent of the choice of L [33], only play a minor role in the memory-free turbulent flow struc-
except the maximum SGS droplet diameter determined by the L ture established by the strong shear flow between the liquid and
value. It is, however, known that too large L may hinder resolving gas phases as expected at moderately high pressures under con-
the liquid core accurately and cause a numerical instability. After sideration. In this sense, the code is free of case-by-case empirical
conducting grid resolution tests, L = D/12 is used. This is reason- parameter tuning for turbulent jets, and solving the liquid core ex-
able because the integral length scale of turbulent nozzle flow is hibits an advantage over the conventional codes in paradigm A.
known to be  = D/8[41]. To avoid the empty range of droplet size In the present simulations, the following algorithms were in-
(0.34 L < d < 0.34) between LES-resolved and SGS droplet sizes, tentionally excluded to investigate the intrinsic features caused by
atomized droplets associated with the interface deformation wave- SGS atomization.
length λ <  = 3 L were treated as the SGS atomization in the
present simulation. (1) Secondary atomization for droplets atomized from the liquid
Table 1 shows the flow conditions and material properties for core surface, except the one that is resolved by the two-phase
two main cases and one reference case considered here. Case 1 flow LES grid.
is a baseline case mimicking a Diesel jet. Case 2 is a jet simi- (2) Coalescence/breakup of colliding droplets.
lar to the experiment in [22]. Case 3 is a low Reynolds number (3) Random displacement of atomized droplets, although the
case which is not suited for spray combustion, but solved for ref- droplets are in SGS sizes and may be exposed to gaseous SGS
erence. For these jets, experimental data are available for compar- turbulence.
ison, although they are limited. At time t > 0, the completely tur- (4) Droplet evaporation.
bulent nozzle flow was assumed to emanate from a circular nozzle
Therefore, in each case, the spray only consists of droplets at-
into an otherwise quiescent air at temperature 300 K and pressure
omized from the liquid core surface.
3 MPa. The mean injection velocity profile is given by the 1/7th
The used grid system had an equal spacing in the cuboid cal-
law, where a thin shear layer exists outside of the rather flat ve-
culation domain (0 < x < 82D, −5.75D < y,z < 5.75D for case 1 and
locity area in the central region. Upon this, LES-resolved unsteady
0 < x < 47D, −5D < y,z < 5D for cases 2 and 3), namely in total with
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 245

Fig. 15. Temporal evolution of spray (case 1). (a) Initial head formation and RT atomization from the head edge. The instantaneous Bo number distribution is also shown.
(b) Evolution of spray and liquid core. The color on the droplets indicates the axial velocity. The color on the liquid core indicates the liquid regression speed. (c) Temporal
history of spray jet front. The thick lines indicate the present result. The circles are experimental data [43].

18.7 million and 9 million grid points, respectively. The so-called 3.1. Comparison with available experimental data for validation
free inflow and outflow conditions were imposed on the side and
downstream boundaries of the calculation domain. Obviously, it is First of all, we emphasize that the droplet ejection condition of
not practical to perform a DNS which resolves the minimum sized the resonant mode, explored in Section 2, gives a physical reason
droplet formation process in the same target since the necessary to Faeth et al. experimental observations [22] that turbulent atom-
computational resources could be huge. The wall clock time for the ized droplets have the outward normal velocity equal to the turbu-
present simulation was 2–5 days using a small desktop PC cluster lent intensity and the tangential velocity equal to the local liquid
composed of 7 CPUs, which is beneficial for industrial use. surface velocity in the near-nozzle region. This strongly supports
Figures 15 and 16 show the overall temporal evolution of the our modeling from the physical point of view in Section 2.
simulated spray formation process for cases 1 and 2. Overall, both Unfortunately, other experimental data are limited due to the
cases exhibit similar behavior of initial head formation, core thin- difficulty in measurement of the liquid core behaviors. Still, valida-
ning, development of a dense droplet layer over the liquid core, tion can be possible in some aspects. Here, we compare four repre-
gradual spray spreading and core head vanishment. Before going sentative diesel spray characteristics, and their underlying physics
into the spray mechanisms, first, the simulation results are com- will be discussed in the subsequent subsections.
pared with available experimental data for a validation purpose.
246 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 16. Temporal evolution of spray and liquid core (case 2) at Ut/D = 65.

Fig. 17. Comparison of intact core length. The original figure was reproduced from
[44 and 45].

3.1.1. Penetration length


Fig. 18. Comparison of spray spreading angle. Experimental data are reproduced
In Fig. 15(c), experimental validation of the simulated liquid from [46–49].
core tip history is made in comparison with that of similar injec-
tion conditions but at different gas pressures Pa = 0.1, 2 and 4 MPa
[43]. The nozzle diameter was D = 0.24 mm in the experiments. that the spray angle established within the intact core length takes
The jet issue speed U = 219 m/s was estimated from the slope of a little bit smaller value because the liquid core flow drags the sur-
a straight line passing the origin and the three black solid circles rounding spray flow.
for Pa = 0.1 MPa. It is known that the core tip history is influenced
by the ambient pressure and the behavior in the present simula- 3.1.4. Droplet size distribution
tion is well consistent with the experimental observations. In our simulation, all SGS atomized droplet information is
stored. Figure 19 shows details of the SGS droplet size distribu-
3.1.2. Intact core length tion generated from the stored data for case 1. The corresponding
The two black solid circles in Fig. 17 show the intact core length normalized distribution cannot be measured in experiments and
(or core breakup length) Lb /D for cases 1 and 2 comparing with direct comparison is impossible. Most experimental data present
Hiroyasu’s data (Pa = 3.0 MPa) [44,45]. This comparison was made the droplet size distribution of a spray element located near the
only after the calculations, i.e. this result was not tuned. The com- spray boundary or unspecified portions. As a result, significant dif-
plete spray jet (with atomization onset immediately downstream ferences among reported droplet size distributions are found in
of the nozzle and an invisible liquid core) realizes only on the experimental and numerical literature. For the validation of the
right-hand side of the red curve. The three yellow bars will be ex- present simulation, we are interested in the variation of droplet
plained later. The present results predict good agreement and show size distribution along the liquid core.
that the dimensionless intact core length of complete spray jet is The figure indicates that the RT mode on the core surface be-
almost independent of the injection velocity. comes dominant as the jet penetrates. In order to discuss the
shape of droplet size distribution, some typical experimental data
3.1.3. Spray angle is taken. As will be explained in Section 3.3, the RT√atomized
Figure 18 shows the comparison of spray angle. It is known that droplet diameter is scaled according to d ∝ (ρL /ρG )1/4 D/U. Us-
the liquid/gas density ratio significantly affects the spray angle. Al- ing this scaling, the red droplet size distribution (normalized in
though the experimental data scatter with some range [46–49], the the same way as ours) in Fig. 19 was reproduced from the mea-
present results are well within the range of experimental observa- surement data (having a peak droplet size of 12.5μm) at the in-
tions. Most experiment data are supposed to be obtained for the stant of jet penetration length 6 mm (Ut/D = 17.1) for U = 100 m/s,
spray flow downstream of the intact core length. It is reasonable D = 0.35 mm and Pa = 1 MPa [50]. Consistency of the peak droplet
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 247

Fig. 19. Atomized droplet size distribution for case 1.

a log–normal distribution in the final spray. However, this is not


a critical issue since the important point in the Lagrangian track-
ing of SGS droplets is that the dominant size of locally atomized
droplets is correctly captured. Of course, the proposed SGS model
without ensemble-averaging can be used as a stochastic model
with a much simpler algorithm, but this will be a costly simula-
tion. Therefore, with the above considerations, the present aver-
aged modeling is an appropriate approach for practical-scale spray
simulation.
Similar to the simulation results, the experimental data also in-
dicated the presence of a small fraction of large droplets sepa-
rated from the main log–normal-like distribution. However, it is
more noticeable in the simulation results. This is a result of the
intended neglect of droplet collision and secondary atomization.
Such large droplets are never broken in the present uniform grid
system which cannot resolve the thin shear flow developed on the
atomized droplet surface. Again, this is not a critical issue, but
if further sophistication is needed, a secondary atomization algo-
rithm may be applied.
Further experimental data are not readily available. But, from
the above shown comparisons, it can be said that the present
Fig. 20. Probability density function of droplet diameter atomized in random α h- method can predict the spray behavior well. It should be noted
realization. This defines the ensemble-averaged atomized droplet diameter as a that the above agreement has been made for sprays under differ-
delta function indicated by a vertical line for given We and Bo. ent conditions, using exactly the same code without changing any
model constants case by case.

size is fairly good under the above-mentioned experimental un-


certainties. The two profile shapes are slightly different, i.e., the 3.2. Atomization from head umbrella
simulation result is relatively peakier, but this is intended by the
proposed SGS model for computational cost reduction as explained The liquid core head collides with the stagnant gas and initially
next. deforms to an umbrella-like shape (Fig. 15(a)), which has been also
In the model, the atomized droplet diameter (expressed as a confirmed in our previous DNS study [10]. Vigorous RT atomization
function of We and Bo) is averaged using those droplet size distri- occurs at the umbrella edge due to its curvature. This is indicated
butions that are illustrated in Fig. 20 for the resonant mode. Simi- by the Bo number distribution shown in Fig. 15(a) and the atom-
lar profiles are obtainable for the RT mode, as known in ultrasonic ization mode shown in Fig. 19(a). All the injected liquid reaching
atomization studies. The abscissa of Fig. 20 corresponds to each the liquid core head is atomized. A similar feature continues until
α h-realization allowable for given We and Bo. Each profile is of the liquid core head becomes small and finally vanishes, as seen in
a log–normal-like distribution form. The proposed SGS model ap- Fig. 15(b).
proximates this as a delta function that has a spike at the averaged It can be observed in Fig. 19 that the dominant mode is shifting
droplet diameter, and intends to realize an equivalent drag force from the head RT mode to the core RT mode, along with some con-
effect on the gas flow by Eq. (7). Therefore, if we recalculate the tribution from the resonant mode in the near-nozzle region. Simi-
droplet size distribution using such a probability density function lar behavior has been also observed in case 2. Therefore, it is sug-
for each atomized droplet, we could reproduce a more similar pro- gested that some complicated physical processes occur inside the
file to the red less-peakier profile, because, in turbulent jets, the spray. In experiments, it is difficult to observe what are happening
local We and Bo values randomly change on the liquid core sur- within the spray jet. The SGS atomization not only deforms the
face both spatially and temporarily, and would eventually result in LES-resolved liquid core shape but also changes its own type lo-
248 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 21. Liquid core dynamics.

cally through two-way interactions with the LES-resolved gas flow (c). This atomization mode transition can be also confirmed by the
containing atomized droplets. It is expected that these complicated change in the slope of the liquid core’s equivalent diameter (Fig.
features are successfully captured by the present simulation. The 21(d)). It is important, in causality, to note that the resonant mode
following are physical insights from the present simulation results, in the near-nozzle region is indispensable to trigger the subsequent
mainly using case 1 as a baseline case. RT mode in the turbulent atomizing jet. The LES-resolved interface
regression rate caused by the RT atomization is much higher and
the surface tension effect on the LES-resolved interface is negligi-
3.3. Determination of intact core length ble. Accordingly, the liquid core faster reduces its equivalent diam-
eter of irregularly deformed cross-section along the liquid core and
Figures 15(b) and 16 show that the core thins due to atomiza- results in Fig. 21(d).
tion as the penetration progresses. Near the nozzle exit, the liquid The RT atomization in this process produces a relatively large
jet has thin turbulent liquid- and gas-phase boundary layers on its diameter of atomized droplets, which can penetrate through the
surface. The large liquid shear rate causes large We but small Bo spray layer produced by the upstream resonant atomization. The
(see Fig. 21(a) and (b)) and sets off the resonant atomization. As thinning liquid core reduces its head umbrella size accordingly.
a result, a large number of small droplets are produced and flow Therefore, the thickened spray layer with a large axial velocity
down as a spray layer surrounding the liquid core. As the bound- comes to overleap the head umbrella and collides with the stag-
ary layers grow downstream, We decreases initially according to nant air ahead, as can be seen in Figs. 15(b) and 23(a). This is the
an expected power law (x/D )−2/3 and the atomized droplets be- stage where the liquid core head is disappearing, and determines
come larger. The ejected droplets tend to further reduce the inter- the intact core length Lb . The intact core length obtained in the
facial gas shear rate and, thereby, liquid shear rate L , as seen in present simulation does not change significantly in the subsequent
Fig. 21(a). Then, the resonant atomization no longer takes place time, because the upstream spray flow has already entrained air
dominantly. Instead, the SGS RT atomization is locally excited due from outside and settled down to a quasi-steady-state spray flow
to the LES-resolved liquid turbulence or core deformation driven with a certain spray angle.
by the Kelvin-Helmholtz instability as found from Fig. 21(b) and
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 249

Fig. 22. Comparison of Sauter mean diameter (SMD). The original figure was reproduced from [22].

Note that Fig. 21 shows the simulation results at an instant of on U should be weak for the boundary-layer type flow with a
Ut/D = 110 or 65. The flow field is unsteady and the plots fluctu- small spray angle. Furthermore, Fig. 21(b) indicates that fn is an in-
ate temporally about the presented ones even taking a short time creasing function of x/D. Here, we consider the average value C =
%
D Lb /D √
interval about the instant. In addition, the local We and Bo values
L 0
fN d ( Dx ) and use the expression g = C ρG /ρL (U 2 /D ). Then,
b
vary in the circumferential direction as well. This is found from the we have, from Eq. (10)b,
 3 2  ρ ρ U 2 D  3 2  ρ
differences in the We and Bo distributions along the two different
generation lines. Since √the present simulations cause the resonant C g L C g
Bo = = We (47)
atomization for W e = 2 2B/π 2 = 1.36 in Fig. 21(a), Figs. 21(a) and ( 2 π )2 i ρL γ (2π )2 i ρL D
(d) indicate that the contribution of resonant atomization may be
neglected for the determination of intact core length Lb . The fol- where i = D/ L. Naturally, the Bo value depends on the grid size
lowing is a simple analysis based on this fact. This analysis was by definition. The Bo value of Eq. (47) is drawn in Fig. 21(b)
conducted to interpret and validate Fig. 21(b) and (c) physically with the red horizontal line for i = 12 and C = 1 for reference.
by considering the Bo distribution as a function of x, which is ob- Since C is expected be of O(1), the figure implies that g(x ) =

tained by averaging over time and in the circumferential direction. fn (x/D ) ρG /ρL (U 2 /D ) correlates the envelop curve of the simu-
As found from Figs. 15(c) and 23(c), the local axial velocity lated Bo(> 0) values well and that the simulation captures the ex-
of the liquid core maintains a value close to the injection veloc- pected physics successfully.
ity U over the intact core length, and the strong KH instability is When Eq. (47) is substituted into d/ = 0.34 6/Bo, we obtain
excited by the interaction between the liquid core flow and the    2 
spray-laden gas shear flow. Therefore, the magnitude of local g in- d d  3 6 3 6(2π )2 i ρL 1
= = 0.34 = 0.34
duced by this KH instability may be estimated through the follow- D D i Bo i C 3 ρG W e D
ing three steps of analysis. (I) The local shear layer thickness and   
the outward liquid surface velocity v are scaled by D and U, re- 6 ρL 1 / 4 1
spectively. Therefore, the magnitude of the gas-phase shear rate = 2π X 0.34 (48)
C ρG W eD
G is proportional to U/D, which leads to the magnitude of the

liquid-phase shear rate L ∝ ρG /ρL (U/D ) through the interfacial This indicates that the size of SGS droplets produced by the RT
turbulent stress balance. (II) The absolute value of Bo < 0 takes a atomization does not depend on the choice of the LES grid spacing
similar profile to those in Fig. 21(a). Therefore, the LES-resolved L = D/i since i is canceled out in Eq. (48).
interface is unsteadily deformed in a sinusoidal-like form in both Figure 22 shows the Sauter mean diameter (SMD) of the
axial and circumferential directions. In Fig. 21(c), the axial wave- present results and experimental data from [22]. The original fig-
length of the local interface deformation is plotted against the lo- ure is reproduced from Fig. 19 in [22], which has collected and
cal Bo value. This indicates that the LES-resolved RT instability is correlated, by the thick black straight line, the SMD data ob-
excited by the KH instability. (III) The temporal change in v, ob- tained at the atomization onset location of a quasi-steady liq-
served in the frame of reference moving with the liquid core ele- uid jet emanating from a nozzle into the atmosphere. It is
ment, should be caused by the transverse motion of the KH insta- known that the droplet size is initially small in the near-nozzle
bility flow, and it is expressible in the form of v∝Usin (L t). There- region and gradually increases as the spray flow field devel-
fore, the envelope curve of positive g = −dv/dt ∝ U L is expressed ops toward the downstream region [22]. The shaded area in

as g(x ) = fn (x/D ) ρG /ρL (U 2 /D ). Obviously, the dependence of fn Fig. 22 shows the estimated SMD range covering the variation
250 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Fig. 23. Radial profiles of liquid, gas and droplet velocities at Ut/D = 110 for case 1. All the velocities are averaged in the circumferential direction.

from the near-nozzle region to the downstream region. The two resonant atomization algorism yields the expected SMD consis-
marked stars (left: Pa = 1 MPa, D = 3.6 mm, U = 35 m/s, SMD = 71 tent with the correlation line. The present results are shown as
μm and right:Pa = 1 MPa, D = 36 mm, U = 35 m/s, SMD = 150 μm) square marks in the figure. The white squares indicate the SMD
denote the results of our initial simulations using a short length of the initial stage, and the gray squares include droplets from the
of calculation domain, which were conducted to confirm that the RT mode. For case 1 (We = WeD /8), the calculated SMD of the
A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252 251

spray was 2.8 μm (resonant mode only) and 6.8μm at Ut/D = 40 droplets drag the gas where the gas velocity comes to decrease sig-
(lower squares connected with a dashed line), and it gradually nificantly. The radial gas velocity there is negative, which indicates
increased up to 18 μm (the uppermost square) until the liquid air entrainment. This air entrainment is caused to satisfy the conti-
core head disappears at Ut/D = 110, because larger droplets atom- nuity equation when the inner high-speed gas flow exerts a shear
ized by the core RT mode are accumulated in the stored atomized stress to increase the outer axial gas velocity. The increase in axial
droplet information. This variation is related to the mode shift as gas velocity, caused by the dragging force of atomized droplets, en-
seen in Fig. 21. It should be noted that the white squares (res- hances this air entrainment. Therefore, it is important to note that
onant mode droplets only) well agree with the near-nozzle ex- the enhanced air entrainment tends to decelerate those droplets
perimental data, where it can be speculated that resonant mode which are ejected from the core surface or driven by the LES vorti-
droplets are actually counted in the experiments. Case 2 shows cal flow and move outward by their inertia. This tendency is clearly
a similar trend. These results show good agreement both in the observed in Fig. 23(d) since the radial droplet velocity is rapidly
qualitative trend of size increase and in the quantitative range of decreased to a negative value as the droplets go outward. As a re-
size. For reference, the d/values estimated using Eq. (50) with sult, the radial distance that atomized droplets can reach at each
C = 1 for Cases 1 and 2 are marked with the two red crosses on a axial station is upper limited. This implies that the spray with sig-
line d/ = 2π X 0.34 6X 8/C (ρL /ρG )1/4 / W eL . Since SMD takes a nificantly large droplet number density (droplet layer) is confined
larger value than the mean diameter d, it can be said that the red in a region bounded by a spray angle in the enhanced air entrain-
line well predicts the SMD of the spray produced in the simula- ment region. Essentially, the spray angle is established by the pres-
tions. ence of a locally stable location r = rstab (x) in the radial movement
Using Eq. (47), it is easy to show that the intact core length of spray droplets, i.e., vd (r = rstab ) = 0.
should increase with ReD in Fig. 17 if the circumferential core Due to the difference in inertia (characterized by the Stokes
surface deformation were ignored. Therefore, it is found that the number), it is natural that larger droplets are observed at the edge
circumferential core surface deformation whose degree increases of the spray layer as shown in Fig. 23(b). Figures 23(b) and (d)
with ReD is indispensable to realize the almost ReD -independent indicate that larger droplets from the core and the head can pene-
intact core length in the complete spray jets. This is because the trate longer and reach the periphery while smaller droplets cannot.
net RT atomization per unit axial length is enhanced by the cir- Similar observations have been made experimentally.
cumferential core surface deformation effect. As seen above, the motion of droplets is strongly influenced by
The two vertical lines in Fig. 23 respectively express the lower the interaction with the gas flow. The turbulent vortex ring formed
limit of the complete spray jet region at Pa = 3 and 0.1 MPa, which by the collision of the spray flow on the stagnant air entrains air
were calculated from Fig. 17 for a water jet into air. The com- effectively and enlarges the spray head size gradually with decreas-
plete spray jet is realized on the right-hand side of the respec- ing moving speed, forming a shape with a rather flattened head
tive vertical lines. It is obvious that most jets in the original fig- connected with the cone-shaped spray region.
ure (Pa = 0.1 MPa) are laminarlized at a location far apart from the
nozzle and disintegrate into large droplets due to an air-assisted
Plateau–Rayleigh instability even if turbulent atomization occurs in
the near-nozzle region where the jet has a thin turbulent liquid 4. Conclusion
shear layer on its surface [27,33]. A similar behavior was observed
in the present simulation for case 3 (ReD = 1.1 × 104 ) in Fig. 17, In this paper, the detailed SGS atomization model was derived,
where the three yellow bars denote the calculation domain length which can describe the SGS turbulent atomization from the LES-
used in the three simulations. Case 3 exhibits the resonant mode resolved interface of a bulk liquid phase and couple the interface-
near the nozzle. But the liquid core flow is laminarlized down- tracking Eulerian two-phase flow LES with Lagrangian spray track-
stream and extends longer than the calculation domain. A much ing scheme. With detailed physical considerations, the proposed
longer calculation domain with a refined grid would be needed to turbulent atomization model is free of case-by-case parameter tun-
fully capture the laminarlized jet breakup behavior. Basically, the ing, which gives a substantial advantage over conventional spray
present formulation can even treat such transitional cases as well, simulation models.
although such cases are not interesting for spray combustion. As an example for demonstrating the effectiveness of the pro-
posed SGS atomization model, the spray formation process of tur-
3.4. Determination of spray angle within the intact core length bulent Diesel fuel jets was simulated. The results were very good
in predicting the liquid core behavior, atomized droplet character-
An LES-resolved large vortical structure is gradually established istics, and spray flow behavior in comparison with experimental
along the jet in the gas phase. This implies that the spray droplets data. Since the model distinguishes the type of SGS atomization
may move at velocities associated with the vortical gas velocity corresponding to the local state of each interfacial grid cell, two-
near the outer boundary of the spray region. Conventionally, spray way interactions among the bulk liquid flow, the droplets-laden
angle is determined experimentally from a shadow graph of the gas flow and SGS atomization itself are captured. The example
spray jet downstream of the intact core length. Figure 23(a) shows cases clearly demonstrated the importance of these two-way inter-
an instantaneous shadow graph calculated at Ut/D = 110 when the actions to simulate the complete spray jet structure and the advan-
core head umbrella disappears. It suggests the establishment of tage of the proposed turbulent atomization model free of tuning
a certain spray angle in the region x/D < 50, where an influence parameters. It is a straightforward extension to implement droplet
of the recirculation zone which moves with the spray jet head is evaporation and chemical kinetics in the downstream region, as
weakened (it still remains as the increasing droplet size portion done in conventional LES spray combustion simulations.
r/(D/2) > 7 in Fig. 23(b), which is caused by a small number of large
droplets atomized from the head umbrella edge and not of inter-
est in the following). The mechanism of this spray angle establish-
ment is understood from the radial distributions of axial and ra- Acknowledgments
dial velocity components averaged in the circumferential direction
(see Figs. 23(c) and (d)). The axial gas and droplet velocities de- This work was supported by JSPS KAKENHI Grant Number
crease from Us// to zero as r increases. Their difference is small. The JP17H03481.
252 A. Umemura, J. Shinjo / Combustion and Flame 195 (2018) 232–252

Supplementary materials [24] A. Umemura, Self-destabilizing mechanism of a laminar inviscid liquid jet is-
suing from a circular nozzle, Phys. Rev. E 83 (2011) 046307.
[25] A. Umemura, S. Kawanabe, S. Suzuki, J. Osaka, Two-valued breakup length of a
Supplementary material associated with this article can be water jet issuing from a finite-length nozzle under normal gravity, Phys. Rev.
found, in the online version, at 10.1016/j.combustflame.2018.01.026. E 84 (2011) 036309.
[26] A. Umemura, J. Osaka, Self-destabilizing loop observed in a jetting-to-dripping
References transition, J. Fluid Mech. 752 (2014) 184–218.
[27] A. Umemura, Model for the initiation of atomization in a high-speed laminar
liquid jet, J. Fluid Mech. 757 (2014) 665–700.
[1] C.K. Law, Combustion physics, Cambridge University Press, 2010.
[28] A. Umemura, Self-destabilising loop of a low-speed water jet emanating from
[2] W.A. Sirignano, Fluid dyanamics and transport of droplets and sprays, second
an orifice in microgravity, J. Fluid Mech. 797 (2016) 146–180.
ed., Cambridge University Press, 2010.
[29] Y. Li, A. Umemura, Two-dimensional numerical investigation on the dynamics
[3] Y.J. Jiang, A. Umemura, C.K. Law, An experimental investigation on the collision
of ligament formation by Faraday instability, Int. J. Multiph. Flow 60 (2014)
behavior of hydrocarbon droplets, J. Fluid Mech. 234 (1992) 171–190.
64–75.
[4] N. Peters, Turbulent combustion, Cambridge University Press, 20 0 0.
[30] Y. Li, A. Umemura, Threshold condition for spray formation by Faraday insta-
[5] D. Veynante, L. Vervisch, Turbulent combustion modeling, Prog. Energy Com-
bility, J. Fluid Mech. 759 (2014) 73–103.
bust. Sci 28 (2002) 193–266.
[31] Y. Li, A. Umemura, Numerical study on the jet formation due to Rayleigh–Tay-
[6] D.C. Haworth, Progress in probability density function methods for turbulent
lor instability, Jpn. J. Appl. Phys. 53 (2014) 110302.
reacting flows, Prog. Energy Combust. Sci. 36 (2010) 168–259.
[32] A. Umemura, J. Shinjo, Toward a new paradigm of spray combustion research,
[7] J.C. Beale, R.D. Reitz, Modeling spray atomization with the Kelv-
9th Asia-Pacific Conference on Combustion, Gyeongju, Korea (2013), pp. 7–10.
in–Helmholtz/Rayleigh–Taylor hybrid model, Atom. Sprays 9 (1999) 623–650.
[33] A. Umemura, Turbulent atomization subgrid model for two-phase flow
[8] P.K. Wu, R.F. Miranda, G.M. Faeth, Effects of initial flow conditions on pri-
large eddy simulation (theoretical development), Combust. Flame 165 (2016)
mary breakup of nonturbulent and turbulent round liquid jets, Atom. Sprays
154–176.
5 (1995) 175–196.
[34] R.J. Lang, Ultrasonic atomization of liquids, J. Acoust. Soc. Am. 34 (1962) 6–9.
[9] M. Linne, Imaging in the optically dense regions of a spray: a review of devel-
[35] L. Rayleigh, On the instability of jets, Proc. Lond. Math. Soc. 10 (1878) 4–13.
oping techniques, Prog. Energy Combust. Sci. 39 (2013) 403–440.
[36] T. Yabe, F. Xiao, T. Utsumi, The constrained interpolation profile method for
[10] J. Shinjo, A. Umemura, Simulation of liquid jet primary breakup: dynamics of
multiphase analysis, J. Comput. Phys. 169 (2001) 556–593.
ligament and droplet formation, Int. J. Multiph. Flow 36 (2010) 513–532.
[37] J. Shinjo, J. Xia, A. Megaritis, L.C. Ganippa, R.F. Cracknell, Modeling temperature
[11] O. Desjardins, V. Moureau, H. Pitsch, An accurate conservative LES ghost fluid
distribution inside an emulsion fuel droplet under convective heating: a key to
method for simulating turbulent atomization, J. Comput. Phys. 227 (2008)
predicting microexplosion and puffing, Atom. Sprays 26 (2016) 551–583.
8395–8416.
[38] J. Shinjo, J. Xia, L.C. Ganippa, A. Megaritis, Physics of puffing and microexplo-
[12] M. Herrmann, Detailed numerical simulations of the primary atomization of a
sion of emulsion fuel droplets, Phys. Fluids 26 (2014) 103302.
turbulent liquid jet in crossflow, J. Eng. Gas Turbines Power 132 (2010) 061506.
[39] J. Shinjo, J. Xia, L.C. Ganippa, A. Megaritis, Puffing-enhanced fuel/air mixing of
[13] X. Jiang, G.A. Siamas, K. Jagus, T.G. Karayiannis, Physical modelling and ad-
an evaporating n-decane/ethanol emulsion droplet and a droplet group under
vanced simulations of gas-liquid two-phase jet flows in atomization and
convective heating, J. Fluid Mech. 793 (2016) 444–476.
sprays, Prog. Energy Combust. Sci. 36 (2010) 131–167.
[40] J. Shinjo, J. Xia, Combustion characteristics of a single decane/ethanol emulsion
[14] D.K. Lilly, The representation of small-scale turbulence in numerical simulation
droplet and a droplet group under puffing conditions, Proc. Combust. Inst. 36
experiments, IBM Scientific Computing Symposium on Environmental Sciences,
(2017) 2513–2521.
Yorktown Heights, N.Y. (1967), pp. 195–210.
[41] J.O. Hinze, Turbulence, McGraw-Hill, New York, USA, 1875.
[15] F. Jaegle, J.M. Senoner, M. Gareia, F. Bismes, R. Lecourt, B. Cuenot, T. Poinsot,
[42] M. Klein, A. Sadiki, J. Janicka, A digital filter based generation of inflow data
Eulerian and Lagrangian spray simulations of an aeronautiacal multipoint in-
for spatially developing direct numerical or large eddy simulations, J. Comput.
tector, Proc. Combust. Inst. 33 (2011) 2099–2107.
Phys. 186 (2003) 652–665.
[16] S. Tachibana, K. Saito, T. Tamamoto, M. Makida, T. Kitano, R. Kurose, Experi-
[43] W. Li, T. Suzuki, Y. Ochiai, T. Oda, S. Tanabe, Initial development of diesel
mental and numerical investigation of thermo-acoustic instability in a liquid–
sprays and spatial distribution of fuel drops in the outer layer, Atomization
fuel aero-engine combustor at elevated pressure: validity of large-eddy simu-
6 (1997) 140–149.
lation of spray combustion, Combust. Flame 162 (2015) 2621–2637.
[44] H. Hiroyasu, M. Arai, M. Shimizu, Break-up length of a liquid jet and internal
[17] M. Herrmann, A parallel Eulerian interface tracking/Lagrangian point particle
flow in a nozzle, LCLASS-91, Gaithersburg, MD, USA, 1991, pp. 275–282.
multi-scale coupling procedure, J. Comput. Phys. 229 (2010) 745–759.
[45] M. Tabata, M. Arai, H. Hiroyasu, Sauter mean diameter of a diesel spray in-
[18] Y. Ling, S. Zaleski, R. Scardovelli, Multiscale simulation of atomization with
jected into an environment of high pressure, Trans. Jpn. Soc. Mech. Engr. B 55
small droplets represented by a Lagrangian point-particle model, Int. J. Mul-
(1989) 3239–3245.
tiph. Flow 76 (2015) 122–143.
[46] Y. Wakuri, M. Fujii, T. Amitani, R. Tsuneya, Studies on the penetration of fuel
[19] S. Jay, F. Lacas, S. Candel, Combined surface density concepts for dense spray
spray of diesel engine, Trans. Jpn. Soc. Mech. Engr. B 25 (1959) 820–826.
combustion, Combust. Flame 144 (2006) 558–577.
[47] N. Ishikawa, K. Tujimura, Measurement of diesel spray angle near the nozzle
[20] S. Navarro-Martinez, Large eddy simulation of spray atomization with a prob-
exit, Atomization 8 (1999) 51–58.
ability density function method, Int. J. Multiph. Flow 63 (2014) 11–22.
[48] X. Wang, Z. Huang, O.A. Kuti, W. Zhang, K. Nishida, Experimental and analytical
[21] R. Lebas, T. Menard, P.A. Beau, A. Berlemont, F.X. Demoulin, Numerical simu-
study on biodiesel and diesel spray characteristics under ultra-high injection
lation of primary break-up and atomization: DNS and modelling study, Int. J.
pressure, Int. J. Heat Fluid Flow 31 (2010) 659–666.
Multiph. Flow 35 (2009) 247–260.
[49] K.-J. Wu, R.D. Reitz, F.V. Bracco, Measurements of drop size at the spray edge
[22] G.M. Faeth, L.-P. Hsiang, P.-K. Wu, Structure and breakup properties of sprays,
near the nozzle in atomizing liquid jets, Phys. Fluids 29 (1986) 941–951.
Int. J. Multiph. Flow 21 (1995) 99–127.
[50] T. Ishiyama, K. Miwa, M. Kamogawa, Y. Liu, S. Miyashiro, Observations of mi-
[23] M. Saeedipour, S. Pirker, S. Bozorgi, S. Schneiderbauer, An Eulerian–Lagrangian
croscopic structures of diesel sprays with a nano-spark light source, Trans. Jpn.
hybrid model for the coarse-grid simulation of turbulent liquid jet breakup,
Mech. Engr. B 60 (1994) 715–721.
Int. J. Multiph. Flow 82 (2016) 17–26.

You might also like