Sie sind auf Seite 1von 41

Journal of Computational Physics 231 (2012) 5571–5611

Contents lists available at SciVerse ScienceDirect

Journal of Computational Physics


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

Three-dimensional boundary conditions for numerical simulations


of reactive compressible flows with complex thermochemistry
Axel Coussement a,b,c,⇑, Olivier Gicquel a,b,c, Jean Caudal b,c, Benoît Fiorina b,c, Gérard Degrez a
a
Université Libre de Bruxelles, Aero-Thermo-Mechanics Departement, Avenue F.D. Roosevelt 51, CP 165/41, 1050 Bruxelles, Belgium
b
Ecole Centrale Paris, Grande Voie des Vignes, 92295 Chatenay-Malabry, France
c
CNRS, UPR 288 Laboratoire dEnergtique molculaire et macroscopique, combustion, Grande Voie des Vignes, 92295 Chatenay-Malabry, France

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

Article history: The Navier–Stokes characteristic boundary conditions (NSCBC) is a very efficient numerical
Received 11 October 2011 strategy to treat boundary conditions in fully compressible solvers. The present work is an
Received in revised form 5 March 2012 extension of the 3D-NSCBC method proposed by Yoo et al. and Lodato et al. in order to
Accepted 7 March 2012
account for multi-component reactive flows with detailed chemistry and complex trans-
Available online 7 April 2012
port. A new approach is proposed for the outflow boundary conditions which enables clean
exit of non-normal flows, and the specific treatment of all kinds of edges and corners is
Keywords:
carefully addressed. The proposed methodology is successfully validated on various chal-
Computational fluid dynamics
Compressible reactive flows
lenging multi-component reactive flow configurations.
Combustion Ó 2012 Elsevier Inc. All rights reserved.
Characteristic boundary conditions
NSCBC

1. Introduction

The tremendous increase of computational power and recent progress in numerical methods enable today the use of com-
putational fluid dynamics (CFD) to conduct numerical experiments of multi-component reactive flows [1–5]. Highly resolved
simulations require not only high order numerical schemes [6] but also a specific treatment of boundaries that avoids spu-
rious oscillation, flow distortion, or numerical instabilities when complicated flow structures enter or exit the domain [7–9].
Various strategies have been proposed to tackle boundary conditions. Among them, techniques based on characteristic
wave decomposition have motivated much effort because of their efficiency. Initially developed for the Euler equations
[10,11], these approaches have been extended to viscous flows with the introduction of the Navier–Stokes characteristic
boundary conditions (NSCBC) by Poinsot and Lele [7]. This technique consists in imposing the boundary conditions in the
form of waves entering or leaving the computational domain with a specific treatment of the viscous terms. Initially devel-
oped for ideal single-component gases, this approach has been extended to more complex situations such as multi-compo-
nent reactive flows [8] or real gas mixtures [9]. Improvement of the method for low frequency oscillating waves or for low
Mach number expansion flows have been also discussed [12].
In these simulations, the flow at the boundary is locally considered as one-dimensional and aligned with the normal at
the boundary. This yields the so-called LODI system that is solved in the characteristic space in order to compute the bound-
ary condition in the physical space. However, in their work Yoo et al. [13] indicate that the consideration of the transverse
term is necessary to obtain a consistent boundary condition. The formulation used by Yoo et al. [13] is therefore no longer

⇑ Corresponding author at: Université Libre de Bruxelles, Aero-Thermo-Mechanics Departement, Avenue F.D. Roosevelt 51, CP 165/41, 1050 Bruxelles,
Belgium. Tel.: +32 2 650 26 71; fax: +32 2 650 27 10.
E-mail address: axcousse@ulb.ac.be (A. Coussement).

0021-9991/$ - see front matter Ó 2012 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.jcp.2012.03.017
5572 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

locally one-dimensional but fully three-dimensional, which gives the so called 3-D-NSCBC. In a further work Yoo and Im [14]
focused on the relaxation coefficient used in the transverse terms in order to achieve a more consistent 3-D-NSCBC formu-
lation at the boundary. Treatment of edges and corners have been assessed by Lodato et al. [15].
Compared with the original 1-D-NSCBC formulation the 3-D extension and the specific treatment of edges and corners
dramatically improve the flow prediction. However, as is demonstrated further in the present work, flow distortions are still
observed when complex patterns exit the domain. Also the 3D-NSCBC formulation must be extended to multicomponent
reactive flows with a detailed treatment of all kinds of edges and corners.
This work follows different objectives: first, an improvement of the outlet boundary condition is proposed along with a
new definition of edges and corners related to the outflows. Second, it proposes a full review (i.e including all kinds of edges
and corners) of the 3-D-NSCBC strategy adapted to high order numerical simulations of reactive flows where detailed chem-
istry effects and complex transport phenomena are accounted for. In Section 2, the 3-D-NSBC is formulated for multi-com-
ponent reactive flows with complex thermochemical properties. For wave decomposition, the temperature is preferred to
the pressure as a non conservative variable because it simplifies reaction rate formulations and wall treatment. A set of
boundary condition treatments for faces, edges and corner are specified in Sections 3–5, respectively. Finally, an implemen-
tation technique of the 3-D-NSCBC using rotations is given in order to drastically simplify the implementation into com-
pressible solvers. Finally the proposed methodology is tested on challenging test cases for multi-component reactive flows.
As shown in Section 6.1.1 this improvement allows a clean exit of flows that are not normal to an outlet, which can be of
great interest in unstructured LES codes.

2. Three-dimensional Navier–Stokes boundary conditions for multicomponent reactive flows with detailed chemistry

2.1. Problem formalism

A multicomponent reactive flow is governed by the Navier–Stokes equations written in the compressible form:
e @e
@U Fi @ Dei
þ þ e ¼0
þX ð1Þ
@t @xi @xi
e is the vector of conservative variables:
where U
e ¼ ½q; qu1 ; qu2 ; qu3 ; qe; qY 1 ; . . . ; qY N T
U ð2Þ
sp

q is the fluid density, u1, u2 and u3 are the velocities in the x, y and z directions, Yk are the species mass fractions and Nsp is the
number of species. The total energy is defined as:
N sp 
X 
1 R
qe ¼ qui ui þ q hk  T Yk ð3Þ
2 k¼1
Wk

where Einstein’s convention is used for the velocities, hk is the enthalpy per unit of mass of species k, R is the perfect gas
constant, Wk is the species k molar mass and T is the temperature.
Using the Kronecker operator di,j, the conservative convective, diffusive and reactive vectors are defined respectively as:
0 1
0 1 0 0 1
qui 0
B C
B qu1 ui þ d1i p C B s1i C B 0 C
B C B C B C
B C B s2i C B C
B qu2 ui þ d2i p C B C B 0 C
B C B C B C
B qu3 ui þ d3i p C B s3i C B 0 C
B C B C B C
e
F ¼ B ðqe þ pÞu C
i B ; e
D ¼ B uj sij þ qi C
i B e B
X¼B C ð4Þ
i C C; 0 C
B C B C B C
B qui Y 1 C C B  @½ Ji;1  C B W x C
B B @xi C B 1 _ 1 C
B C B C B C
B
@
..
.
C
A
B
B
.. C
C
B
@ ... C
A
@ . A
@½J i;Nsp 
qui Y Nsp  @xi W Nsp x _ Nsp

where p is the pressure. The Navier–Stokes equations are completed by state equations. The perfect gas law assumption
leads to:
p ¼ qrT ð5Þ
where r, the mass gas constant is defined as:
R
r¼ ð6Þ
W
and W is the mean molar mass. sij and qi are the stress tensor and heat flux vector respectively defined as:
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5573

  
@ui @uj 1 @uk
sij ¼ l þ  di;j ð7Þ
@xj @xi 3 @xk
  X Nsp
@T
qi ¼ k þ ½J i;k hk ð8Þ
@xi k¼1

The ratio c between the heat capacity at constant pressure, cp, and the heat capacity at constant volume, cv, is defined:
cp
c¼ ð9Þ
cv
The speed of sound, c, is related to r and temperature by:
pffiffiffiffiffiffiffiffi
c¼ crT ð10Þ

2.2. Consistency of the 3-D-NSCBC

A consistent formulation of the 3-D-NSCBC at faces, edges and corners requires an a priori knowledge of the number of
conditions to impose at the boundary. Sutherland and Kennedy [16] went through this analysis on the basis of the work of
Gustafsson and Sundström [17]. In the case of the Navier–Stokes equations and for a subsonic flow, the number of conditions
to impose at a boundary to ensure that the system is well posed is Ndim + 1 + Nsp where Ndim is the number of spatial dimen-
sions. The number of boundary conditions corresponds to the number of balance equations minus one. Indeed, the mass con-
servation equation does not require an explicit Neumann condition.
Neumann boundary conditions can be replaced by Dirichlet conditions, depending on the nature of the boundary condi-
tion. Table 1 gives the number of Dirichlet and Neumann conditions for the most popular boundary conditions. Here only the
subsonic boundary conditions are presented for the 3-D-NSCBC.
The following set of non-conservative variables for rewriting the Navier–Stokes equations is introduced:

U ¼ ½q; u1 ; u2 ; u3 ; T; Y 1 ; . . . ; Y Nsp T ð11Þ


The temperature, T, is preferred to the pressure as a non-conservative thermodynamic variable because it simplifies the
expression of the energy. This formulation introduces additional terms in the wave amplitude equations (see Eq. (14)) with
respect to those expressed with pressure [16]. But using temperature simplifies the treatment of chemical reactions, since
reaction rates appear only in the species equations. Moreover, the use of temperature simplifies the treatment of the isother-
mal wall.
The flux conditions treatment and direction convention are described in the next section.

2.3. Boundary treatment and directions convention

As described above for the 3-D-NSCBC, Dirichlet and diffusive flux boundary conditions must be enforced. The Dirichlet
conditions, called inviscid conditions I , are imposed through the amplitudes of characteristic waves as in the 1D-NSCBC for-
malism [7]. The Neumann conditions, called viscous conditions V, are enforced in the viscous terms of the Navier–Stokes
equations [7]:
e
@U @eFi @Dei
þ þ þ e ¼0
X ð12Þ
@t @xi @xi
|{z} |{z}
I conditions V conditions

As shown in Table 1, the relation I þ V ¼ N dim þ 1 þ N sp is always fulfilled. Details of viscous conditions, along with the non-
conservative variables imposed via the inviscid conditions, are indicated in Table 2. If not specified, the rest of the viscous
terms are computed at the boundary from the flow field inside the domain. The subscript n denotes the direction normal
to the boundary. Table 2 also gives the number of viscous, V and inviscid I , conditions. The adiabatic no-slip wall is a par-
ticular case of the imposed heat flux no-slip wall where the heat flux is imposed to zero: A ¼ 0
As indicated in Appendix A, the chemical source terms x _ k are inserted in the vector X and accounted in the boundary as
in [16]. The conditions are imposed as follow (full details are given in Appendix A):

Table 1
Number of Dirichlet and Neumann conditions for the most common subsonic boundary conditions.

Dirichlet conditions Neumann conditions Total


Subsonic inflow Ndim + 1 + Nsp 0 Ndim + 1 + Nsp
Subsonic outflow 1 Ndim + Nsp Ndim + 1 + Nsp
Imposed heat flux no-slip wall Ndim 1 + Nsp Ndim + 1 + Nsp
Isothermal no-slip wall Ndim + 1 Nsp Ndim + 1 + Nsp
5574 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Table 2
Inviscid and viscous conditions for different types of boundary. Note that k = 1, . . . , Nsp and i = 1, . . . , Ndim.

Type of boundary Viscous condition(s) Inviscid condition(s) V I


Subsonic Inflow – u1, u2, u3, T and Yk 0 Ndim + 1 + Nsp
Subsonic outflow @ sni p Ndim + Nsp 1
@xn ¼ 0 for n – i,
@qn @½J n;k 
@xn ¼ 0, @xn ¼0
Imposed heat flux no-slip wall [Jn,k] = 0, qn ¼ A u1, u2, u3 1 + Nsp Ndim
Isothermal non-slip wall [Jn,k] = 0 u1, u2, u3, T Nsp Ndim + 1

Fig. 1. Illustration of the rotation process for a corner located at x1 = x1max, x2 = x2max, x3 = x3max.

Fig. 2. Algorithm description for the treatment of boundary conditions by the 3-D-NSCBC.

 Computation of the wave amplitudes respecting the I conditions and taking into account the transverse terms. As
opposed to the 1D-NSCBC, not only the terms normal to the boundary are taken into account in the wave amplitudes,
but also the terms in the plane of the boundary. For example, if one considers a boundary normal to x1, the wave ampli-
tudes are computed not only using the derivatives along x1, as in the 1-D-NSCBC, but also the derivatives along x2 and
x3.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5575

 Computation of the non-conservative temporal derivatives based on the previously computed wave amplitudes.
 Addition of the source terms.
 Computation of the conservative variable temporal derivatives based on the wave amplitudes and the source terms.
 Inclusion of the viscous terms in the conservative variable temporal derivatives. As indicated in Table 2, V conditions
are imposed through the viscous terms.

Using this algorithm, the I þ V conditions are imposed in the conservative Navier–Stokes equations. In the following sec-
tions, only the wave amplitude treatment for the inviscid conditions and the viscous conditions to be imposed will be de-
tailed as the treatment of the chemical source terms does not change with the boundary type.
For the sake of clarity, the equations are written for inward pointing normal to the faces. Face boundary conditions are
written for a face located at x1 = x1min; the normal is therefore positive and along x1. For the edges, they will be given at
x1 = x1min and x2 = x2min, hence the edge is aligned with x3. Similarly, for a corner the conditions will be given for x1 = x1min,
x2 = x2min, x3 = x3min.
For any other case, a simple rotation of axes prior to the NSCBC treatment allows the application of the expressions pre-
sented here without modifications. Obviously, after the treatment, the inverse rotation must be performed to obtain the set
of conservative equations in the initial frame. For example, for a corner boundary located in x1 = x1max, x2 = x2max, x3 = x3max
which has three negative normals in the reference frame (x1, x2, x3) (see Fig. 1), an axis rotation from x1, x2, x3 to x1 , x2 , x3 must
be applied to recover positives normals at the corner. In that new reference frame the 3-D-NSCBC equations are applied, and
the results are rotated back to the initial frame. All these rotations are detailed in Appendix B.
A schematic view of the full algorithm for the treatment of the 3-D-NSCBC is summarized in Fig. 2.

3. Face boundaries

3.1. General equations

As detailed in Appendix A, the non-conservative Navier–Stokes equations for a boundary located at x1 = x1min can be ex-
pressed with the introduction of the wave amplitudes along x1, Li :

Xsp N
@q q qW
¼ L1 þ L2  L5 þ L5þk  T 1
@t T k¼1
Wk
@u1 c
¼  ðL5  L1 Þ  T 2
@t q
@u2
¼ L3  T 3
@t ð13Þ
@u3
¼ L4  T 4
@t
@T RT
¼ ðL1 þ L5 Þ  L2  T 5
@t qWcv
@Y k
¼ L5þk  T 5þk k ¼ 1; . . . ; Nsp
@t

the wave amplitudes read:


0 !1
XNsp
@q
B ðu1  cÞ 21c @x q @u1 q @T qW @Y k C
B 1
 2c @x 1
þ 2T c @x 1
þ 2W k c @x1 C
B k¼1 C
B ! C
B XN sp C
B RT @q 1 RT @Y k C
B u1  W qc @x1 þ c @x1  @T C
B p W k cp @x1 C
B k¼1 C
B C
B @u
u1 @x12 C
B C
B C
B u1 @u 3 C
L¼B @x1 C ð14Þ
B !C
B XNsp C
B @q q @u1 q @T qW @Y k C
B ðu1 þ cÞ 21c @x þ þ þ C
B 1 2c @x1 2T c @x1 2W k c @x1 C
B k¼1 C
B C
B u1 @Y 1 C
B @x1 C
B C
B .. C
B . C
@ A
@Y Nsp
u1 @x1
5576 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

the transverse terms are noted T i and are given by:


0 1
@q @q
u2 @x þ q @u
@x2
2
þ u3 @x þ q @u
@x3
3

B 2 3
C
B u2 @u 1
þ u3 @u 1 C
B @x2 @x3 C
B C
B X N C
B RT @ q @u2 R @T RT @Y k @u2 C
B þ u2 @x þ þ þ u 3 @x C
B W q @x2 2 W @x 2 W k @x2 3 C
B k¼1 C
B C
B X N sp C
B @u3 RT @ q @u3 R @T RT @Y k C
T ¼B u2 @x2 þ W q @x3 þ u3 @x3 þ W @x3 þ W C ð15Þ
B k @x3 C
B k¼1 C
B RT @u2 RT @u3
C
B þ u 2
@T
þ þ u 3
@T C
B Wcv @x2 @x2 Wcv @x3 @x3 C
B C
B u @Y 1
þ u @Y 1 C
B 2 @x2 3 @x3 C
B C
B .. C
B . C
@ A
@Y N @Y N
u2 @x2sp þ u3 @x3sp
In the following sections, if not specified, the wave amplitudes are computed from Eq. (14). For viscous terms, if not specified
they are computed from the solution inside the domain.

3.2. Subsonic non-reflecting inflow

The procedure of Yoo et al. [13] is applied to close the condition for a non-reflecting inflow. Equations are presented for an
inflow that relaxes towards target temperature, Tt, velocities, u1,t, u2,t, u3,t, and species mass fractions Yk,t which is the most
common case. For a non-reflecting inflow the Li read:
c
L2 ¼ gT ðT  T t Þ  T21 ð16Þ
cLx1
c
L3 ¼ gu2 ðu2  u2;t Þ  T31 ð17Þ
Lx1
c
L4 ¼ gu3 ðu3  u3;t Þ  T41 ð18Þ
Lx1
qð1  M2 Þ
L5 ¼ gu1 ðu1  u1;t Þ  T51 ð19Þ
2Lx1
c
L5þk ¼ gY k ðY k  Y k;t Þ  T15þk k ¼ 1; . . . ; Nsp ð20Þ
Lx1
where Lx1 is the size of the domain in the direction normal to the boundary (i.e x1), M is the Mach number and gn is the
relaxation coefficient of the variable n. All the gn are positive. The order of magnitude of these coefficients depends on
the application, but a value of 0.05 to 0.1 is a good first guess. As in [15] the Tm
k are the characteristic waves perpendicular
to xk and relative to the mth characteristic variable. They are defined as follows:
Xsp N
1 q q qW
T11 ¼ T1 T2þ T5þ T 5þk
2c 2c 2T c k¼1
2W kc

Xsp N
Tðc  1Þ Wcv ðc  1Þ TWðc  1Þ
T21 ¼  T1þ T5 T 5þk
qc Rc k¼1
cW k
T31 ¼T3 ð21Þ
T41 ¼T4
Xsp N
1 q q qW
T51 ¼ T1þ T2þ T5þ T 5þk
2c 2c 2T c k¼1
2W k c
T5þk
1 ¼ T 5þk k ¼ 1; . . . ; Nsp
The Li estimated from Eqs. (16)–(20) are then included in the set of Eq. (13) to compute the non-conservative variables tem-
poral derivatives. L1 is estimated from the inside domain numerical solution. Then the algorithm described in Section 2.3 is
applied to obtain the conservative variables temporal derivatives and to include the viscous and chemical terms. This set of
equations yields Ndim + 1 + Nsp conditions and no viscous condition is enforced.

3.3. Subsonic non-reflecting outflow

The subsonic outflow is derived from the 1-D non-reflecting outflow boundary condition proposed by Rudy and
Strikwerda [18], by adding the transverse relaxation of Yoo et al. [13]. Since Ndim + Nsp viscous conditions are enforced for
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5577

Table 3
Inviscid and viscous conditions for different types of edges. For details see the corresponding section. O = outflow, I = inflow, HFW = imposed heat flux no-slip
wall, IW = isothermal no-slip wall. Note that k = 1, . . . , Nsp and i = 1, . . . , Ndim.

Type of edge Viscous condition(s) Doubled viscous condition(s) (n = 1, 2) Inviscid condition(s) V I


O–O None @ sni
¼ 0 for n – i, @qn
¼ 0,
@½J n;k 
¼0 p Ndim + Nsp 1
@xn @xn @xn
I–O or O–I None None u1, u2, u3, T and Yk 0 Ndim + 1 + Nsp
I–I None None u1, u2, u3, T and Yk 0 Ndim + 1 + Nsp
HFW-HFW None [Jn,k] = 0, qn ¼ A u1, u2, u3 1 + Nsp Ndim
IW-IW or
HFW-IW or None [Jn,k] = 0 u1, u2, u3, T Nsp Ndim + 1
IW-HFW
HFW-I or I-HFW [Jn,k] = 0, qn ¼ A, None u1, u2, u3 1 + Nsp Ndim
for n normal to the wall
IW-I or I-IW [Jn,k] = 0 None u1, u2, u3, T Nsp Ndim + 1
for n normal to the wall
HFW-O or O-HFW [Jn,k] = 0, qn ¼ A, None u1, u2, u3 1 + Nsp Ndim
For n normal to the wall
IW-O or O-IW [Jn,k] = 0 None u1, u2, u3, T Nsp Ndim + 1
For n normal to the wall

Table 4
Inviscid and viscous conditions for different types of corners. O = outflow, I = inflow, HFW = imposed heat flux no-slip wall, IW = isothermal no-slip wall.

Type of corner Viscous condition(s) Tripled viscous condition(s) Inviscid condition(s) V I


(n = 1, 2, 3)
O–O–O None @ sni p Ndim + Nsp 1
@xn ¼ 0 for n – i,
@qn @½J n;k 
@xn ¼ 0, @xn ¼0
I with two O or
two I with one O None None u1, u2, u3, T and Yk 0 Ndim + 1 + Nsp
HFW-HFW-HFW None [Jn,k] = 0, qn ¼ A u1, u2, u3 1 + Nsp Ndim
IW-IW-IW or
IW with one or two HFW None [Jn,k] = 0 u1, u2, u3, T Nsp Ndim + 1
HFW with two I or
Two HFW with one I [Jn,k] = 0, qn ¼ A, None u1, u2, u3 1 + Nsp Ndim
For n normal to the wall (s)
Conditions are doubled
If there are two walls
IW with two I or
Two IW with one I [Jn,k] = 0 None u1, u2, u3, T Nsp Ndim + 1
IW with I and HFW
For n normal to the wall (s)
Conditions are doubled
If there are two walls
HFW with two O or
Two HFW with one O [Jn,k] = 0, qn ¼ A, None u1, u2, u3 1 + Nsp Ndim
for n normal to the wall(s)
conditions are doubled
if there are two walls
IW with two O or
Two IW with one O [Jn,k] = 0 None u1, u2, u3, T Nsp Ndim + 1
IW with O and HFW
For n normal to the wall(s)
Conditions are doubled
If there are two walls

an outflow (Table 2) only one has to been imposed by the Li . Following Yoo et al. [13], in the situation where the boundary is
located at x1 = x1min, the pressure should be imposed by L5 as follows:

ð1  M2 Þ  
L5 ¼ r ðp  pt Þ  T51 þ b T51  T51;ex ð22Þ
2cLx1
where b, 2[0 : 1], is the transverse relaxation parameter, pt is the target pressure value and T51;ex is the exact value of the
transverse terms on the boundary. However, as it is demonstrated in Section 6.1.1, this definition does not allow a clean exit
of vortical structures, if they are not convected perpendicular to the boundary plane.
5578 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Through a low perturbation analysis of 2-D flow, Yoo and Im [14] demonstrated that an optimum value of b is given by
the Mach number based on the normal velocity component at the boundary:

u1
b¼ ¼M ð23Þ
c

To demonstrate this, they assumed the mean flow was normal to the boundary:

u1 mean –0
u2 mean ¼0 ð24Þ
u3 mean ¼0

In the case of non-normal flows u2 mean and u3 mean are non-zero, and so causes inconsistency. This explains why a non-nor-
mal vortex cannot properly exit the domain as shown in Section 6.1.1. Indeed, if the mean flow tangential velocities are non-
zero, additional terms appear in the analysis of Yoo and Im [14].
To be consistent with the analysis of Yoo and Im [14], the following corrected velocities are introduced:

~ 2 ¼ u2  u2mean
u ð25Þ
~ 3 ¼ u3  u3mean
u ð26Þ

Using these new velocities instead of u2 and u3 the additional terms vanishes. Since u2 mean and u3 mean are constant, the
derivatives of the tangential terms do not change:

~2 @u2 @u2mean @u2


@u
¼  ¼ ð27Þ
@xi @xi @xi @xi
~3 @u3 @u3mean @u3
@u
¼  ¼ ð28Þ
@xi @xi @xi @xi
Finally, using these new velocities the same optimum b value is retrieved from the low perturbation analysis.
This demonstrates that if one uses the corrected velocities, and so corrected transverse terms, a clean exit of vortical
structures will be recovered in the case of non-normal flows. Eventually, these corrected velocities are introduced in
the transverse term definitions. Eq. (22), becomes:

ð1  M2 Þ  
L5 ¼ r e5 þ b T
ðp  pt Þ  T f5  T
e5 ð29Þ
1 1 1;ex
2cLx1

f2 and f
Using the corrected tangential velocities u u3 , the corrected transverse terms are defined as follows:

N
f5 ¼ 1 Te þ q Te þ q Te þ X qW Te
sp

T 1 1 2 5 5þk ð30Þ
2c 2c 2T c k¼1
2W k c

with:
@q @u2 @q @u3
Te 1 ¼ u
~2 þq þu ~3 þq
@x2 @x2 @x3 @x3
e @u1 @u1
~
T 2 ¼ u2 þ u3~
@x2 @x3
ð31Þ
RT @u @T RT @u3 @T
Te 5 ¼
2
þu ~2 þ þu~3
Wcv @x2 @x2 Wcv @x3 @x3
@Y @Y
Te 5þk ¼ u
~2 k þ u ~3 k
@x2 @x3
The value, Te 5 , towards which the transverse terms should be relaxed, is in most cases not known a priori. Consequently
1;ex
e 5 ¼ 0 should be imposed, as it will be shown in the test cases presented below. However, if one wants to impose a value
T 1;ex
towards which the transverse terms should relax, the formulation of Te 5 from Yoo et al. [13] must be adapted with respect
1;ex
to the introduction of the corrected velocities:

q^ f q^ f Nsp
X c
q^ W
e5 ¼ 1 f
T Tb 1;ex þ Tb 2;ex þ Tb 5;ex þ
f
Tb ð32Þ
1;ex
2c
^ 2^c 2 Tb c
^ ^ 5þk;ex
2W k c
k¼1

with:
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5579

~^ ~^
f ~^ @ q
Tb 1;ex ¼ u
^
þq
@u
^ 2þu ~^ @ q^
þq^ 3
@u
2 3
@x2 @x2 @x3 @x3
f ~ @ ^1 ~ @ u
u ^1
Tb 2;ex ¼ u^2 þu ^3
@x2 @x3
b @u ð33Þ
f R T ~^ @ Tb R Tb @ u
~^ @ Tb
Tb 5;ex ¼
2
þu ^~ 2 þ
3
þu ^~ 3
cq
W ^ @x2 @x2 W cq ^ @x3 @x3
b b
f ~^ @ Y k þ u
Tb 5þk;ex ¼ u ~^ @ Y k k ¼ 1; . . . ; Nsp
2 3
@x2 @x3
and:

~^ ¼ u
u ^ 2  u2mean
2
ð34Þ
u^~3 ¼ u
^ 3  u3mean

where the hat (^) denotes variables computed with the known solution at the boundary.
According to the work of Rudy and Strikwerda [18], the optimal value of the relaxation coefficient r used in Eq. (29) is
equal to 0.28. Eq. (29) combined with the V viscous conditions indicated in Table 2 provides the Ndim + 1 + Nsp required
conditions.
Finally, in the case where u2mean and u3mean are null, the treatment proposed here is strictly equivalent to the one of Yoo
and Im [14] and Lodato et al. [15].

3.4. Imposed heat flux no-slip wall and isothermal no-slip wall

For an imposed heat flux no-slip wall, the viscous conditions introduce Nsp + 1 constraints as indicated in Table 2. The wall
is assumed normal to x1. The Ndim missing conditions are deduced by introducing the no-slip conditions: u1 = u2 = u3 = 0 on
the face. One can introduce those conditions by prescribing u1 = u2 = u3 = 0 and:

@u1 @u2 @u3


¼ ¼ ¼0 ð35Þ
@t @t @t
This yields the following wave amplitudes:

L5 ¼ L1 ð36Þ
Li ¼ 0; i ¼ 2; 3; 4 ð37Þ
L5þk ¼ 0; k ¼ 1; . . . ; N sp ð38Þ
where L1 is computed from system (14) using the inside domain numerical solution. Doing so, the momentum equations in
the x2 and x3 will degenerate into:

@p @ s2;i
¼ ð39Þ
@x2 @xi
@p @ s3;i
¼ ð40Þ
@x3 @xi
as they should on a no-slip wall. An adiabatic no-slip wall can easily be deduced from the imposed heat flux wall by enforc-
ing the heat flux to zero, that is A ¼ 0 in Table 2. For an imposed heat flux wall, the number of inviscid conditions ðI Þ is Ndim,
along with the 1 + Nsp viscous conditions ðVÞ this yields Ndim + 1 + Nsp conditions.
An isothermal no-slip wall will yield the same wave amplitudes as for the imposed heat flux no-slip wall. The only dif-
ference is in the treatment of the heat flux: this viscous condition is replaced by an inviscid condition: T = Twall on the wall. As
for the velocities the most convenient way to impose the wall temperature is:

T ¼ T wall at t ¼ 0
@T ð41Þ
¼ 0 to be imposed in system ð13Þ
@t
As for the velocities, setting the temporal derivative of the temperature to zero is straightforward (see system (13)).
In this case there are Ndim + 1 inviscid conditions, the velocities and temperature and Nsp viscous conditions, yielding the
Ndim + 1 + Nsp conditions needed.
5580 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

4. Edges boundary conditions

In this section it will be assumed that the edge boundary condition is along x3 and located at x1 = x1min and x2 = x2 min.
Lodato et al. [15], proposed to relate the wave amplitudes Li in direction x1 to the wave amplitudes Mi in direction x2.
The expression of non-conservative variables reads, in terms of the wave amplitudes:

Xsp N Xsp N
@q q qW q qW
¼ L1 þ L2  L5 þ L5þk  M1 þ M3  M5 þ M5þk  T 1
@t T k¼1
W k T k¼1
Wk
@u1 c
¼  ðL5  L1 Þ  M2  T 2
@t q
@u2 c
¼ L3  ðM5  M1 Þ  T 3
@t q ð42Þ
@u3
¼ L4  M4  T 4
@t
@T RT RT
¼ ðL1 þ L5 Þ  L2  ðM1 þ M5 Þ  M3  T 5
@t qWcv qWcv
@Y k
¼ L5þk  M5þk  T 5þk k ¼ 1; . . . ; Nsp
@t
where M and T reads:
0 !1
N sp
X
1 @q
B ðu2  cÞ
B 2c @x2
 2c @x2 þ 2Tqc @x
q @u2
þ@T
2
qW @Y k
2W k c @x2
C
C
B k¼1 C
B C
B u2 @u 1 C
B @x2 C
B ! C
B XN sp C
B RT @q 1 @T RT @Y k C
B u 2  þ  C
B W qcp @x2 c @x2 W k cp @x2 C
B k¼1 C
B C
B u2 @u 3 C
M¼B @x2 C ð43Þ
B !C
B XN sp C
B @q q @u2 q @T qW @Y k C
B ðu2 þ cÞ 21c @x þ þ þ C
B 2 2c @x2 2T c @x 2 2W k c @x2 C
B k¼1 C
B C
B u2 @Y 1 C
B @x2 C
B C
B .. C
B . C
@ A
@Y N
u2 @x2sp

0 @q
1
u3 @x 3
þ q @u
@x3
3

B C
B u3 @u 1 C
B @x3 C
B C
B u3 @u 2 C
B @x3 C
B C
B N sp
X C
B RT @q @Y k C
B þ u3 @u3
þ WR @T
þ RT C
B Wq @x3 @x3 @x3 Wk @x3 C
T ¼B k¼1 C ð44Þ
B C
B RT @u3 @T
þ u3 @x C
B Wcv @x3 3 C
B C
B u3 @Y 1 C
B @x3 C
B C
B .. C
B . C
@ A
@Y Nsp
u3 @x3

Values of Li and Mi are given thereafter and, if not specified, the wave amplitudes are computed from inside the domain
using systems (14) and (43) respectively. The viscous terms, if not specified, are also computed from the computational
solution.

4.1. Outflow/outflow

For outflow/outflow edges, Lodato et al. [15] proposed to solve a system of two equations because of the coupling be-
tween L5 and M5 that appears in their transverse terms. Indeed, M5 appears in the transverse terms to be considered for
L5 and vice versa. However, if one considers the formulation proposed in Section 3.3, this coupling vanishes because the
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5581

corrected waves should be used in the transverse terms. Using this last formulation the wave amplitudes to be imposed for a
double outlet edge are:

ð1  MÞ2  
L5 ¼ r e5 þ b T
ðp  pt Þ  T e5  T
e5
1 1 1;ex
2cLx1
ð45Þ
ð1  MÞ2    
M5 ¼ r p  p0t  Te 5 þ b0 T e5
e5  T
2 2 2;ex
2cLx2
where the transverse terms are:

Xsp N
e5 ¼ q M
T f 2 þ 1 ðM f 5 Þ 1 Te 1 þ q Te 2 þ q Te 5 þ
f1 þ M qW e
T
1
2c 2 2c 2c 2cT k¼1
2 cW k 5þk
ð46Þ
N sp
X
e5 qe 1 e e 5 Þ þ 1 Te 1 þ q Te 3 þ q Te 5 þ qW e
T 2 ¼ L3 þ ðL1 þ L T 5þk
2c 2 2c 2c 2cT k¼1
2cW k

As for the outflow boundary conditions the  refers to waves computed with the corrected velocities; these waves must be
computed from the computational solution.
Two transverse relaxation parameters are introduced in Eq. (45): b and b0 . For the sake of simplicity the same b as for the
corresponding face is used. That is the transverse term’s relaxation set for the outlet located at x1min for b and the one set for
the outlet at x2min for b0 in this case. For consistency purpose, target pressures pt and p0t have to be equal: pt ¼ p0t . Conse-
quently system (45) yields one inviscid condition.
The fully developed expressions are obtained very easily, following the same procedure as for the faces and will not be
given here; briefly, Le i must be computed using u ~ 1 , the M ~ 2 and the Te i using u
f i using u ~ 3 . The corrected velocities are defined
as for the outflow face:
~ i ¼ ui  uimean
u ð47Þ
e 5 and T
with ui mean being the mean flow velocity along xi. If the solution at the edges is known, transverse terms T e 5 can
1;ex 2;ex
be evaluated:   N
X sp c
e5 ¼ q M q^ f q^ f q^ W f
^ f
T c2 þ1 M
f c5 þ 1 f
f
c1 þ M Tb 1 þ Tb 2 þ Tb 5 þ Tb
1;ex ^
2c 2 2c^ 2c^ ^ Tb
2c 2c^W k 5þk
k¼1
ð48Þ
Nsp
X c
e5 ¼ q L q^ f q^ f q^ W
^ e
T b 3 þ 1 ðL
e b 5Þ þ 1 f
e
b1 þ L Tb 1 þ Tb 2 þ Tb 5 þ
f
Tb 5þk
2;ex
2^c 2 2c^ 2^c b
2c T
^ k¼1
2 c
^ W k

e f
b i must be computed using u f
^ 1 , the M ^ 2 and the Te
c i using f f
^ i using u
^ 3 . defined as:
where the L u
~^ ¼ u
u ^ i  uimean ð49Þ
i

As explained in Section 3.3, T51;ex


and T52;ex
are unknown in most practical simulations, therefore it will be assumed that these
quantities are zero.
In the case where u1mean, u2 mean and u3 mean are zero, the proposed boundary condition closure for the wave amplitudes is
equivalent to impose L5 and M5 independently, which differs from the treatment proposed by Lodato et al. [15]. However
this does not involve any stability problems at the edges as it is shown in Section 6. Moreover it is simpler since no system of
equation must be solved.
As discussed previously, the wave amplitudes give only one inviscid condition on the pressure. The V viscous conditions
need to be imposed in the normal direction to the edges. But on an edge, normal directions are ill-defined. To overcome this
problem, the viscous conditions will be imposed along the two normals of the faces surrounding the edges. In this way the
Ndim + Nsp conditions (see Table 3) will be imposed on the edges (how ever the normal is defined). The Ndim + Nsp viscous con-
ditions imposed plus the one from Eq. (45) yield Ndim + 1 + Nsp conditions. For an outlet located at x1 = x1min and x2 = x2min the
viscous conditions read:
@ s1i
¼0 i ¼ 2; 3
@x1
@ s2i
¼0 i ¼ 1; 3
@x2
@q1
¼0
@x1
ð50Þ
@q2
¼0
@x2
@½J 1;k 
¼0 k ¼ 1; . . . ; Nsp
@x1
@½J 2;k 
¼0 k ¼ 1; . . . ; Nsp
@x2
5582 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 3. Pressure isolines, from p = 99,600 Pa to p = 101,325 Pa with a constant step, for different time instants for: left transverse terms from Yoo et al. [13],
right corrected transverse terms.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5583

Fig. 4. x-axis velocity (u1) isolines, from u1 = 20 m/s to u1 = 80 m/s with constant step, for different time instants for: left transverse terms from Yoo et al.
[13], right corrected transverse terms.
5584 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 5. Boundary conditions for the vortex convection case.

4.2. Inflow/outflow and outflow/inflow edges

This boundary condition introduces one characteristic condition for the outflow and Ndim + 1 + Nsp for the inflow. In that
form, this boundary condition is equivalent to simultaneously imposing the pressure, velocity and temperature at the same
location. This leads to physical inconsistency issues because the problem is over-constrained. Indeed, the pressure on a reg-
ular inflow boundary condition is implicitly given by the velocity at inlet, the outlet target pressure and the pressure losses
predicted in the computational domain. In this context, enforcing a target pressure at an inflow/outflow edge would over-
constrain the problem. The solution proposed here is to assume a perfectly non-reflecting outflow at the edge. This enables
the edge pressure to reach a value consistent with the flow solution.
The wave amplitude condition leads to Ndim + 1 + Nsp conditions, so that there should be no viscous conditions to en-
force. 1
In summary the wave amplitude condition reads for an inflow/outflow edge:
c
L2 ¼ gT ðT  T t Þ  M3  T21
cLx1
c c
L3 ¼ gu2 ðu2  u2;t Þ  ðM5  M1 Þ  T31
Lx1 q
c
L4 ¼ gu3 ðu3  u3;t Þ  M4  T41
Lx1 ð51Þ
qð1  M2 Þ q 1
L5 ¼ gu1 ðu1  u1;t Þ  M2  ðM1 þ M5 Þ  T51
2Lx1 2c 2
c 5þk
L5þk ¼ gY k ðY k  Y k;t Þ  M5þk  T1 k ¼ 1; . . . ; Nsp
L x1
M5 ¼ 0
and for outflow/inflow edges:
c c
M2 ¼ gu1 ðu1  u1;t Þ  ðM5  M1 Þ  T22
Lx2 q
c
M3 ¼ gT ðT  T t Þ  L2  T32
cLx2
c
M4 ¼ gu3 ðu3  u3;t Þ  L4  T42
Lx2 ð52Þ
qð1  M2 Þ q 1
M5 ¼ gu2 ðu2  u2;t Þ  L3  ðL1 þ L5 Þ  T52
2Lx2 2c 2
c
M5þk ¼ gkþ5 ðY k  Y k;t Þ  L5þk  T5þk
2 k ¼ 1; . . . ; Nsp
Lx2
L5 ¼ 0

1
However, an outlet boundary condition should require Ndim + Nsp viscous conditions. But, as the velocities, temperature and species are set on the edge, all
the viscous fluxes in the direction normal to the outflow are consequently imposed.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5585

Fig. 6. Pressure isolines, from p = 99,600 Pa to p = 101,325 Pa with a constant step, for different time instants for a vortex convected through a
outflow/outflow edge. Corrected transverse terms.

with the Tik defined as:

Xsp N
Tðc  1Þ Wcv ðc  1Þ TWðc  1Þ
T21 ¼  T1þ T5 T 5þk
qc Rc k¼1
cW k
T31 ¼T3
T41 ¼T4 ð53Þ
Nsp
X
1 q q qW
T51 ¼ T1þ T2þ T5þ T 5þk
2c 2c 2T c k¼1
2W kc

T15þk ¼ T 5þk k ¼ 1; . . . ; Nsp

T22 ¼ T 2
Xsp N
Tðc  1Þ Wcv ðc  1Þ TWðc  1Þ
T32 ¼  T1þ T5 T 5þk
qc Rc k¼1
cW k
T42 ¼T4 ð54Þ
Nsp
X
1 q q qW
T52 ¼ T1þ T3þ T5þ T 5þk
2c 2c 2T c k¼1
2W kc

T25þk ¼ T 5þk k ¼ 1; . . . ; Nsp

4.3. Inflow/inflow edges

Inflow velocities, temperature and species mass fractions target values are equal on both edges sides. The incoming waves
amplitudes Li are then computed in x1 directions, and the amplitudes Mi relative to x2 are set to zero. This method, although
5586 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

not theoretically rigorous (see [19] for details), allows good numerical stability on the edges. It results in specifying only
Ndim + 1 + Nsp conditions, and no viscous conditions have to be added. The system of equations reads:

c
L2 ¼ gT ðT  T t Þ  M3  T21
cLx1
c c
L3 ¼ gu2 ðu2  u2;t Þ  ðM5  M1 Þ  T31
Lx1 q
c
L4 ¼ gu3 ðu3  u3;t Þ  M4  T41
Lx1 ð55Þ
qð1  M2 Þ
L5 ¼ gu1 ðu1  u1;t Þ  T51
2Lx1
c
L5þk ¼ gY k ðY k  Y k;t Þ  M5þk  T15þk k ¼ 1; . . . ; Nsp
L x1
Mi ¼ 0; i ¼ 2; . . . ; 5 þ Nsp

where Ti1 are defined by system (53). Note that the boundary conditions could also be imposed on the Mi waves, in this case
it is the Li waves that are set to zero.

4.4. No-slip wall/wall edges

This section will cover the treatment of the edges for the two types of walls covered in this work. For the no-slip walls, as
for face boundary, the three constraints, u1 = u2 = u3 = 0, have to be included in the wave amplitude. The wave amplitudes to
be imposed are:

Li ¼ 0; i ¼ 2; 3; 4
L5 ¼ L1
L5þk ¼ 0; k ¼ 1; . . . ; Nsp
ð56Þ
Mi ¼ 0; i ¼ 2; 3; 4
M5 ¼ M1
M5þk ¼ 0; k ¼ 1; . . . ; Nsp

L1 and M1 are computed from system (14) and (43) using the computational solution. In practice it is more convenient to
impose:

u1 ¼ u2 ¼ u3 ¼ 0 at t ¼ 0
@u1 @u2 @u3 ð57Þ
¼ ¼ ¼ 0 to be imposed in system ð42Þ
@t @t @t

For the isothermal wall/wall, as for the isothermal face boundary condition, the same inviscid condition for the temperature
is added on the two walls surrounding the edge:

T ¼ T wall at t ¼ 0
@T ð58Þ
¼ 0 to be imposed in system ð42Þ
@t
These Ndim or Ndim + 1 inviscid conditions for, respectively, a no-slip imposed heat flux or isothermal wall are completed by
viscous conditions detailed in Table 3. As for outlet/outlet the ill-definition of the normal requires some viscous conditions to
be imposed along the two normals, which are also indicated in Table 3. For an edge composed of a no-slip isothermal wall
and an imposed heat flux wall, the choice has been made to replace the viscous heat flux condition by the temperature invis-
cid condition on the edge, so that it degenerates in a double isothermal wall edge.

4.5. Imposed heat flux or isothermal wall with an outlet or an inlet

In order to ensure a zero wall velocity one must verify system (56). Any combination of an imposed heat flux wall bound-
ary condition with an inflow or an outflow will degenerate into an imposed heat flux wall/wall boundary condition for the
inviscid conditions. If one considers an isothermal wall the boundary condition degenerates into an isothermal wall/wall
boundary condition for the inviscid conditions. The same procedure is also followed for the viscous conditions as indicated
in Table 3.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5587

The viscous conditions should only be enforced on the wall side of the edges yielding Ndim + 1 + Nsp conditions in any
cases.

5. Corner boundary conditions

The boundary conditions will be given for a corner located at x1 = x1min, x2 = x2min, x3 = x3min. To formulate the 3-D-NSCBC
at a corner, the characteristic wave equations are formulated along the three directions. The temporal derivatives of the con-
servative variables computed using the waves along the three directions reads:

Xsp N
@q q qW
¼ ðL1 þ L5 Þ  ðM1 þ M5 Þ  ðN 1 þ N 5 Þ þ ðL2 þ M3 þ N 4 Þ þ ðL5þk þ M5þk þ N 5þk Þ
@t T k¼1
Wk
@u1 c
¼  ðL5  L1 Þ  M2  N 2
@t q
@u2 c
¼ L3  ðM5  M1 Þ  N 3
@t q ð59Þ
@u3 c
¼ L4  M4  T 4  ðN 5  N 1 Þ
@t q
@T RT RT RT
¼ ðL1 þ L5 Þ  L2  ðM1 þ M5 Þ  M3  ðN 1 þ N 5 Þ  N 4
@t qWcv qWcv qWcv
@Y k
¼ L5þk  M5þk  N 5þk k ¼ 1; . . . ; Nsp
@t

with the wave amplitudes along x3 being define as:


0 N sp
!1
@q q @u q @T
X qW @Y k
B ðu3  cÞ 21c @x3  2c @x33 þ 2T c @x þ 2W k c @x3 C
B 3 C
B k¼1 C
B C
B u3 @u 1 C
B @x3 C
B C
B C
B u3 @u 2
C
B @x3 C
B ! C
B N sp
X C
B RT @q 1 @T RT @Y k C
B u3  W qc @x3 þ c @x3  C
B p W k cp @x3 C
N ¼B k¼1 C ð60Þ
B !C
B XN sp C
B @q q @u3 C
B ðu3 þ cÞ 21c @x
B 3
þ 2c @x3
þ 2Tqc @x
@T
3
þ qW
2W k c
@Y k
@x3
C
C
B k¼1 C
B C
B C
B u3 @Y 1
C
B @x3 C
B C
B .. C
B . C
@ A
@Y N
u3 @x3sp

Since the waves are computed in the three directions there is no transverse term any more. All types of corner boundary
conditions are presented in the next sections.
If not specified, the waves amplitudes Li , Mi and N i are computed using systems (14), (43) and (60) respectively using
the computational solution. The viscous terms are also computed from inside the domain if not specified.

5.1. Outflow/outflow/outflow corner

Following the process described in Section 4.1 the system of wave amplitudes to consider is:

ð1  M2 Þ  
L5 ¼ r e5 þ b T
ðp  pt Þ  T e5  T
e5 ð61Þ
1 1 1;ex
2cLx1
ð1  M2 Þ  
M5 ¼ r e 5 þ b0 T
ðp  p0t Þ  T e5
e5  T ð62Þ
2 2 2;ex
2cLx2
ð1  M2 Þ    
N5 ¼ r e 5 þ b00 T
p  p00t  T e5  T
e5 ð63Þ
3 3 3;ex
2cLx3
5588 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 7. Pressure isolines, from p = 99,600 Pa to p = 101,325 Pa with a constant step, for different time instants for a vortex convected through a outflow/
outflow edge. Formulation from Yoo et al. [13] and Lodato et al. [15].

with:

e 5 ¼ 1 ðM f5 þ f q f qf
T 1
f1 þ M N1 þfN 5Þ þ M 2 þ N2
2 2c 2c
e 5 ¼ 1 ðL e5 þ f qe qf
T 2
e1 þ L N1 þfN 5Þ þ L 3 þ N3 ð64Þ
2 2c 2c
e 5 ¼ 1 ðL
T e5 þ M
e1 þ L f 5Þ þ q L
f1 þ M e4 þ q M f4
3
2 2c 2c

and:

e 5 ¼ 1 ðM
f c5 þ f
f c f q^ f
c2 þ q f
^ c
T 1;ex
c1 þ M N1þcN 5Þ þ M N2
2 2^c 2^c
e 5 ¼ 1 ðL
e b5 þ f
e c f q eb
^ qf
^
T 2;ex
b1 þ L N1þcN 5Þ þ L3 þ c
N3 ð65Þ
2 2^c 2^c
c 5Þ þ q L b4 þ q M
^ e ^ f
e 5 ¼ 1 ðL
T
e e f
b5 þ M
b1 þ L f
c1 þ M c4
3;ex
2 2^c 2^c

As for the double outlet edge, the same target pressure has to be retained, yielding one inviscid condition and three trans-
e i and Le
b i must be computed using u ~ 1 defined in Eq.
verse term relaxation parameters, one for each wave. Furthermore, the L
f f
f i and M
(47), M c i must be computed with u ~ 2 (Eq. (25)) and f
N i and c
N i must be computed with u ~ 3 (Eq. (26)).
This formulation is not equivalent to the one proposed by by Lodato et al. [15] in the case where u1 mean = u2mean = u3 mean = 0,
but, as for the double outflow edges, it displays remarkable stability.
For the viscous conditions, the normal is ill-defined as for the double outlet edge. This imposes the enforcement of the
viscous conditions on the three components of the normal at the corners. This gives the Ndim + Nsp physical conditions needed
(see Table 4). The total number of conditions is therefore Ndim + 1 + Nsp.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5589

Fig. 8. x-axis velocity (u1) isolines, from u1 = 20 m/s to u1 = 80 m/s with constant step, for different time instants for a vortex convected through a outflow/
outflow edge.Corrected transverse terms.

Fig. 9. x-axis velocity (u1) isolines, from u1 = 20 m/s to u1 = 80 m/s with constant step, for different time instants for a vortex convected through a outflow/
outflow edge. Formulation from Yoo et al. [13] and Lodato et al. [15].
5590 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

5.2. Inflow/outflow/outflow corner

The strategy followed to handle inflow/outflow edges is retained for this kind of corner. Target velocities, T and Yk are only
specified at the inlet. Assuming inflow in the x1 direction, wave amplitudes are governed by:
c
L2 ¼ gT ðT  T t Þ  M3  N 4
cLx1
c c
L3 ¼ gu2 ðu2  u2;t Þ  ðM5  M1 Þ  N 3
Lx1 q
c c
L4 ¼ gu3 ðu3  u3;t Þ  M4  ðN 5  N 1 Þ
Lx1 q
qð1  M2 Þ 1 q ð66Þ
L5 ¼ gu1 ðu1  u1;t Þ  ðM1 þ M5 þ N 1 þ N 5 Þ  ðM2 þ N 2 Þ
2Lx1 2 2c
c
L5þk ¼ gY k ðY k  Y k;t Þ  Mkþ5  N 5þk k ¼ 1; . . . ; Nsp
L x1
M5 ¼ 0
N5 ¼ 0

or if one supposes the inflow in the x2 direction:

c c
M2 ¼ gu1 ðu1  u1;t Þ  ðL5  L1 Þ  N 2
Lx2 q
c
M3 ¼ gT ðT  T t Þ  L2  N 4
cLx2
c c
M4 ¼ gu3 ðu3  u3;t Þ  L4  ðN 5  N 1 Þ
Lx2 q
qð1  M2 Þ 1 q ð67Þ
M5 ¼ gu2 ðu2  u2;t Þ  ðL1 þ L5 þ N 1 þ N 5 Þ  ðL3 þ N 3 Þ
2Lx2 2 2c
c
M5þk ¼ gY k ðY k  Y k;t Þ  Lkþ5  N 5þk k ¼ 1; . . . ; Nsp
Lx2
L5 ¼ 0
N5 ¼ 0

and finally for an inflow in the x3 direction:

c c
N 2 ¼ gu1 ðu1  u1;t Þ  ðL5  L1 Þ  M2
L x3 q
c c
N 3 ¼ gu2 ðu2  u2;t Þ  L3  ðM5  M1 Þ
L x3 q
c
N 4 ¼ gT ðT  T t Þ  L2  M3
cLx3
qð1  M2 Þ 1 q ð68Þ
N 5 ¼ gu3 ðu3  u3;t Þ  ðL1 þ L5 þ N 1 þ N 5 Þ  ðL3 þ M3 Þ
2Lx3 2 2c
c
M5þk ¼ gY k ðY k  Y k;t Þ  Lkþ5  M5þk k ¼ 1; . . . ; Nsp
Lx3
L5 ¼ 0
M5 ¼ 0

Along with those inviscid conditions on the wave amplitudes, no viscous conditions are required, yielding the Ndim + 1 + Nsp
conditions required.

5.3. Inflow/inflow/outflow corner

Methods proposed for inflow/inflow edges and the inflow/outflow edges are retained for this type of corner. Inflow waves
amplitudes are specified whereas wave amplitudes associated with the remaining inflow and outflow are zero. In the situ-
ation where the outflow follows the x1 direction, the system reads:
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5591

Fig. 10. Iso-contour of the pressure at p = 101,320 Pa for different time instants.

c c
M2 ¼ gu1 ðu1  u1;t Þ  ðL5  L1 Þ  N 2
Lx2 q
c
M3 ¼ gT ðT  T t Þ  L2  N 4
cLx2
c c
M4 ¼ gu3 ðu3  u3;t Þ  L4  ðN 5  N 1 Þ
Lx2 q
qð1  M2 Þ 1 q ð69Þ
M5 ¼ gu2 ðu2  u2;t Þ  ðL1 þ L5 þ N 1 þ N 5 Þ  ðL3 þ N 3 Þ
2Lx2 2 2c
c
M5þk ¼ gY k ðY k  Y k;t Þ  Lkþ5  N 5þk k ¼ 1; . . . ; Nsp
L x2
L5 ¼ 0
N i ¼ 0; i ¼ 2; . . . ; 5 þ Nsp
If the outflow is in the x2 or x3 direction the inflow conditions can be imposed along x1:
c
L2 ¼ gT ðT  T t Þ  M3  N 4
cLx1
c c
L3 ¼ gu2 ðu2  u2;t Þ  ðM5  M1 Þ  N 3
Lx1 q
c c
L4 ¼ gu3 ðu3  u3;t Þ  M4  ðN 5  N 1 Þ
Lx1 q
qð1  M2 Þ 1 q ð70Þ
L5 ¼ gu1 ðu1  u1;t Þ  ðM1 þ M5 þ N 1 þ N 5 Þ  ðM2 þ N 2 Þ
2Lx1 2 2c
c
L5þk ¼ gY k ðY k  Y k;t Þ  Mkþ5  N 5þk k ¼ 1; . . . ; Nsp
L x1
M5 ¼ 0
N i ¼ 0; i ¼ 2; . . . ; 5 þ Nsp
5592 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 11. Pressure profile along a diagonal of the domain for different time instants.

Ndim + 1 + Nsp conditions are recovered without any viscous conditions. As for the inflow/inflow edges, the wave amplitudes
can be imposed on either side of the two inflows.

5.4. Inflow/inflow/inflow corner

The triple inflow wave amplitudes are obtained by imposing the inflow boundary condition on one of the three directions
and enforcing zero wave amplitude on the two last directions:
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5593

c
L2 ¼ gT ðT  T t Þ  M3  N 4
cLx1
c c
L3 ¼ gu2 ðu2  u2;t Þ  ðM5  M1 Þ  N 3
Lx1 q
c c
L4 ¼ gu3 ðu3  u3;t Þ  M4  ðN 5  N 1 Þ
Lx1 q
qð1  M2 Þ 1 q ð71Þ
L5 ¼ gu1 ðu1  u1;t Þ  ðM1 þ M5 þ N 1 þ N 5 Þ  ðM2 þ N 2 Þ
2Lx1 2 2c
c
L5þk ¼ gY k ðY k  Y k;t Þ  Mkþ5  N 5þk k ¼ 1; . . . ; Nsp
L x1
Mi ¼ 0; i ¼ 2; . . . ; 5 þ Nsp
N i ¼ 0; i ¼ 2; . . . ; 5 þ Nsp
As for the Inflow/outflow/outflow, no viscous conditions are imposed because of the presence of inlets. Ndim + 1 + Nsp condi-
tions are obtained.

5.5. Wall/wall/wall corner

An imposed heat flux wall/wall/wall corner is prescribed by ensuring zero temporal derivatives of the velocities. This
gives the following equations for the wave amplitudes:

Li ¼ 0; i ¼ 2; 3; 4
L5 ¼ L1
L5þk ¼ 0; k ¼ 1; . . . ; Nsp
Mi ¼ 0; i ¼ 2; 3; 4
M5 ¼ M1 ð72Þ
M5þk ¼ 0; k ¼ 1; . . . ; Nsp
N i ¼ 0; i ¼ 2; 3; 4
N5 ¼ N1
N 5þk ¼ 0; k ¼ 1; . . . ; Nsp
Including these constraints in the non-conservative equations, the temporal derivatives of the velocity components in the
three directions vanish. As for the double imposed heat flux wall edge, the viscous conditions on the three sides of the corner
are enforced because the normal is ill-defined. Doing this, the Ndim + 1 + Nsp physical conditions are recovered. If there is at
least one isothermal wall on a side of the corner, the same assumption as for edges is made: the corner degenerates into a
triple isothermal wall condition with the inviscid conditions:

T ¼ T wall at t ¼ 0
@T ð73Þ
¼ 0 to be imposed in system ð59Þ
@t
For consistency purposes the same temperature is imposed on the three corner sides if there is at least one isothermal wall
composing the corner. And again a complementary condition must be introduced on the heat flux normal to the wall, which
is not considered as a viscous condition.

5.6. Corners with a wall on at least one side

Since the velocity in the three directions remains zero, these boundary conditions always degenerate in a triple wall
boundary condition. Wave amplitudes are given by system (72).
Regarding the viscous conditions, as for the edges with a wall on one side, the viscous conditions of a wall are applied on
the wall side of the corner if there is only one wall in the corner, and on the two wall sides if there are two walls at the corner.
If there is at least one isothermal wall, the viscous conditions on the heat flux is replaced by the inviscid condition on the
temperature as for the edges, or the triple wall corner. All the viscous conditions are summarized in Table 4.

6. Test cases

The next sections present different test cases which validate the multi-species reactive 3-D-NSCBC and the formulation of
edges and corners. The advantages of the 3-D-NSCBC against 1-D-NSCBC will not be demonstrated here since this analysis
could be found in [12,13,15]. This section is divided into three parts: first, the new outflow formalism will be compared to
the one of Lodato et al. [15], especially in the case of a diagonally convected vortex. The new edges and corners formalism for
5594 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 12. Pressure profile along a diagonal of the domain at t = 235.2 ls for the initial domain (solid black) and a domain twice as large (grey dots).

Fig. 13. Boundary conditions for the burning vortex ring.

double outflow and triple outflow respectively will also be tested using the diagonally convected vortex and the 3-D pressure
wave case of Lodato et al. [15]. The ability of the outflow boundary condition to cleanly evacuate a flame front will then be
accessed on the basis of a vortex flame interaction.
In the second part, the stability of the edges and corners boundary conditions introduced here will be demonstrated using
an inlet/outlet cube and a stagnation point flow test case.
Finally in the last part, the interest of 3-D NSCBC in DNS (or LES) will be demonstrated on the basis of a turbulent hydro-
gen jet.
The boundary conditions treatment has been implemented in the 3-D compressible Navier–Stokes solver YWC developed
at the EM2C laboratory. The spatial derivatives are computed using finite differences with either a 6th order compact scheme
[6] or with a 4th, 6th or 8th centered scheme [20]; only a 6th order compact scheme is used here. An 8th order filtering [21]
is used for stability purpose. The code is explicit in time using a 4th order Runge–Kutta method. The species diffusion fluxes,
Ji,k, molar reaction rates x _ k , and the thermo-chemical quantities are estimated using the CHEMKIN package formalism [22].
The mathematical formalism of the 3-D-NSCBC is indicated in Appendix A.

6.1. New formalism validation

6.1.1. Vortex convection


The first test is a 2-D compressible vortex convected diagonally through a non-reflective outflow. The flow is initialized
by:
!
s ðx2 þ y2 Þ
u1 ðx; yÞ ¼ u10 þ y exp  ð74Þ
R2c 2R2c
!
s ðx2 þ y2 Þ
u2 ðx; yÞ ¼ u20  x exp  ð75Þ
R2c 2R2c
!!
c  s 2 ðx2 þ y2 Þ
pðx; yÞ ¼ p0 exp  exp  ð76Þ
2 cRc R2c
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5595

Fig. 14. Comparison of the YH field at different time instants for the burning vortex ring, on the short and long domain.

with a constant temperature T0 = 300 K. s is the vortex intensity equal to 1, Rc = 20  103 m is the vortex radius, u10 = 50 m/s
is the convection speed along x1, u2 0 = 50 m/s is the convection speed along x2, and p0 = 101,325 Pa is the reference pressure.
The flow is composed of oxygen and nitrogen: Y O2 ¼ 0:233, YN2 = 0.767. The grid, cartesian and uniform, is made of 161  161
nodes, with a grid spacing equal to 103 m. Periodic boundary conditions are applied on the y-normal boundaries, an inlet is
applied on x = xmin imposing flow conditions at infinity (u1mean = 50 m/s, u2 mean = 30 m/s, T = 300 K, Y O2 ¼ 0:233 and
Y N2 ¼ 0:767). At the outlet, at x = xmax, a target pressure of pt = 101,325 Pa is specified. As in [15], the coefficient b, needed
to close Eq. (29) is assumed to be equal to the Mach number defined from the mean flow properties at the outlet:
u1mean
b ¼ pffiffiffiffiffiffiffiffi ð77Þ
crT
5596 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 15. Profiles of YH along y at x = 1.9  103 m. The solid gray line (Long) refers to the long domain, the black dots (Short) to the small domain.
 
Target transverse terms are set to zero T51;ex ¼ 0 . Finally u2 mean is set to 30 m/s. To compare the outlet condition proposed
by Yoo et al. [13] to the one introduced here, two computations where carried out: with and without the proposed transverse
term correction (see Section 3.3). Pressure and x-velocity fields are displayed in Figs. 3 and 4 respectively at four different
instants for both computations.
From Figs. 3 and 4, one can conclude that the use of corrected transverse terms allows a clean exit of flows that are not
normal to the outlet.
To further demonstrate the enhancement provided by the corrected transverse terms, a second test case based on a vortex
convection is computed. The same vortex intensity and radius are used for the initial flow, the following inflow conditions
are imposed: u1 = 50 m/s, u2 = 60 m/s, T = 300 K, Y O2 ¼ 0:233 and Y N2 ¼ 0:767. This configuration allows the vortex center to
exit near the double outflow edges. The remaining boundary conditions are summarized in Fig. 5 and the mesh stays the
same with respect to the previous case. For the outflow/outflow edge, as explained in Section 4.1, two different transverse
relaxation parameters are used. Each of them equals the one defined for the corresponding face. The corrected velocities used
for the transverse terms are the mean tangential velocities on the faces.
The results are shown for the x-velocity and pressure field in Figs. 6 and 8. Those figures can be compared to Figs. 7 and 9
which show the same field computed using the conditions from Yoo et al. [13] and the edge treatment of Lodato et al. [15].
No significant deformation can be seen, which demonstrate the ability of the corrected transverse terms to evacuate vor-
tical structures in very challenging configurations. Furthermore, the outlet/outlet edge formulation proposed here does not
induce any flow distortion and displays good numerical stability.

6.1.2. 3-D pressure wave


A very challenging configuration, initially proposed by Lodato et al. [15], for outlet edges and corners is a spherical pres-
sure wave propagating out of a cubic domain. The initial pressure profile is given by:
!!
x2 þ y 2 þ z 2
pðx; y; zÞ ¼ pt 1 þ d exp  ð78Þ
2R2p
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5597

Fig. 16. Profiles of YOH along y at x = 1.9  103 m. The solid gray line (Long) refers to the long domain, the black dots (Short) to the small domain.

Fig. 17. Boundary conditions for the inlet–outlet cube.

where d = 106, Rp = 20  103 m and the initial temperature is 300 K. Boundaries located at the cube faces are outlets with
pt = 101,325 Pa, r = 0.28 and b = 0.5, as proposed by Lodato et al. [15]. The reference velocities used to correct the transverse
terms are set to zero. The grid is made of 161  161  161 nodes, with a grid spacing equal to 103 m.
Fig. 10 shows the iso-pressure for different time instants and Fig. 11 gives the profile of pressure along a diagonal of the
domain for the same time instants. Both figures show that the faces, edges and corners treatments proposed here for the
outflows enable a proper wave exit, as did the formulation of Lodato et al. [15]. To further verify the soundness of the bound-
ary conditions, the same computation was launched on a domain twice as large in each direction. The comparison between
the two solutions is shown in Fig. 12 where the profile of pressure was extracted along a diagonal of the domain at
t = 87.2 ls. This last figure confirms the good behavior of the outflow boundary conditions proposed here.
5598 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 18. Profile of u1, u2, u3 and Y H2 O  10 extracted along the diagonal of the cube binding the triple inlet corner to the triple outlet corner.

Fig. 19. Boundary conditions for the stagnation point flow.

Fig. 20. Streamlines of velocity. The streamlines start on a line linking two opposite corners of the inlet.

6.1.3. 2D – Burning vortex ring


To demonstrate the ability of the temperature formulated 3D-NSCBC to cleanly evacuate a flame front, premixed hydro-
gen flame-vortex interaction is computed. This test case is very challenging. Indeed, in most of the cases the computation of
such configurations must be stopped before the flame passes through the boundary [23]. The goal of this test case is to verify
that the outlet boundary condition is able to evacuate the flame front without inducing spurious oscillations in the compu-
tational domain. Moreover this test case differs from those of Sutherland and Kennedy [16] and Yoo and Im [14], in the sense
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5599

Fig. 21. Field of the planar velocity (Vp) for a slice normal to the x-axis at x = 25  103 m.

Fig. 22. Profiles of planar velocity (Vp) along the diagonal (from (y = 0, z = 0) to (y = 25  103, z = 25  103)) for different planes.

Fig. 23. Maximum planar velocity (Vp) in different x-normal planes (circle) compared with the analytic profile away from boundary layer, solid black line
[25].

that there is a convection speed and the flow structure is much more complicated. Indeed, the cases proposed by Sutherland
and Kennedy [16] and Yoo and Im [14] are flame front propagations in a flow at rest.
5600 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

Fig. 24. Mesh of the reference domain. The gray line indicates the location of the outlet boundary for the small domain.

The test case configuration is presented in Fig. 13. The grid consists of 161  161 nodes with a constant grid spacing in
both directions Dx = Dy = 12.5  106 m which results in a box of 2  103 m by 2  103 m. The premix hydrogen flame
chemistry is described using the mechanism of Katta et al. [24]. The b parameter is computed using Eq. (77) with the refer-
ence velocity based on the post-flame mean velocity, and the reference speed to correct the transverse terms is set to zero.
As advised by Renard et al. [23] the flow is initialized by a one dimensional premixed flame extended along the y-axis on
which two counter-rotating vortices are superposed in the unburnt gases. The vortices are initialized using Eqs. (74)–(76)
with Rc = 0.2  103 m and s = 20  103 and the distance between the vortices centers is equal to Rc. This results in a maxi-
mum x-axis velocity of around 122 m/s and so a convection speed of the ring of 61 m/s.
To further check the good behavior of the outlet boundary condition, a second computation is made on a domain twice as
long, that is, on a grid using 321  161 nodes. This last case was initialized with the same flow field as the one used on the
shorter domain. Fig. 14 shows a comparison of the YH field during the vortex passage on the two cases.
YH and YOH profiles during the vortex passage are plotted in Figs. 15 and 16. The profiles are extracted at
x = 1.9  103 m. The two sets of profiles display very good agreement, even though slight errors can be observed. Theses
errors can be explained by the hypothesis made on the diffusion fluxes and heat fluxes at the boundary, as pointed out by
Sutherland and Kennedy [16]. Indeed, on the boundary, the heat fluxes and species diffusion flux derivatives normal to the
boundary are set to zero as recommended by Poinsot and Lele [7] which, according to Sutherland and Kennedy [16], is
more stable but yet physically not coherent. Because a detailed thermo-chemistry model is used here on an hydrogen
flame, the species diffusion effects are significant in the flame front. This, in combination with the hypothesis made on
the derivatives of the fluxes, can explain the errors observed.

Fig. 25. Instantaneous Y H2 O fields for the reference domain (A) and the small domain (B). Note the outlet boundary for the small domain can be seen.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5601

Fig. 26. Profiles of Y H2 along y at x = 2  103 m (A), x = 3  103 m (B), x = 4  103 m (C) and x = 5  103 m (D). The solid black line (Long) refers to the
reference domain, the dashed gray line (Short) to the small domain.

6.2. Edges and corners stability

6.2.1. Inlet–outlet cube


This test case and the following are conducted to verify that the multiple boundary conditions formulations for the edges
and corners are consistent. The first test case is composed of three inlets and three outlets, the inlets are normal to the
minimum coordinates along x, y and z and the outlets are located at the maximum coordinates. The boundary conditions
used are summarized in Fig. 17. The grid, composed of 50  50  50 nodes, is regular with Dx = Dy = Dz = 106 m. The initial
flow is at rest, composed of pure oxygen at a temperature of 300 K. Relaxation coefficients are 0.5, 0.1 and 1 for velocities,
temperature and species respectively. At t = 0 s the inlets are activated, and thus H2O is injected in the domain. The simu-
lation is stopped when Y O2 < 1% at the outlets. The transverse term relaxation parameters are set using Eq. (77) with refer-
ence velocities corresponding to the normal velocities at the outlet when the flow has converged. The velocities used to
correct the transverse terms are the mean tangential velocities at convergence on the faces, edges or corner.
The velocities and mass fraction of Y H2 O profiles along a diagonal of the computational domain at the end of the simula-
tion are shown in Fig. 18. These profiles do not display any distortion, which means that the triple inlet corner behaves as
expected.

6.2.2. Stagnation point flow


In this test case, a uniform flow enters from one side of a cube and impacts on a solid adiabatic wall on the opposite side.
This is a severe test case for the boundary conditions on edges and corners, especially on the inlet–outlet edges and the cor-
ner on their sides. Indeed, part of the flow is directly leaving the domain via those edges and corners.
The computational domain is cubic and made of 50  50  50 nodes with a mesh size Dx = Dy = Dz = 103 m. The inlet is
located at x = xmin = 0 m, the wall at x = xmax = 49  103 m. The others boundaries are outlets. The boundary conditions are
5602 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

summarized in Fig. 19 and velocities are imposed with g = 2.5. Those target values for the velocities at the inlet are set
accordingly to the three dimensional stagnation point flow presented in [25]. Finally, the b parameter is computed from
Eq. (77), using the mean normal velocity at the outlet. The reference velocities used to correct the transverse terms are
set to zero.
Streamlines, computed from an inlet diagonal are presented at Fig. 20. The flow impacts the wall and deviates towards the
outlets, without showing any visible unphysical distortion even close to the edges and corners.
Fig. 21 shows the planar velocity V p ¼ u22 þ u23 , at x = 25  103 m. Despite the non axi-symmetry of the computational do-
main, the flow conserves its axi-symmetric properties. This is confirmed by the planar velocities profiles extracted along the
diagonal (from (y = 0, z = 0) to (y = 25  103, z = 25  103)) of different planes perpendicular to the x direction: x = 5  103 m,
x = 15  103 m, x = 25  103 m and x = 35  103 m. According to [25], those profiles should be linear away from the boundary
layer, which is verified by Fig. 22. Moreover, the maximum velocity of each profile should also display a linear decrease with
x. Fig. 23 shows the theoretical profile compared with the maximum Vp in the different x-normal planes, which are located
on the outflow/outflow edges. Good agreement is found, despite a small error for the first point.

6.3. Hydrogen jet

The efficiency of the 3-D-NSCBC with respect to the classical 1D-NSCBC has already been demonstrated in complex con-
figurations. More specifically, Lodato et al. [15] have shown that the 3-D-NSCBC prevents unphysical jet deviation observed
with the 1D-NSCBC. Those unphysical jet deviations are caused by the complex vortical structures exiting the domain and, in
general, require long domain or increased dissipation near the exit of the domain, to prevent them from disturbing the flow
in the zone of interest.
Here, it will also be proven that the 3-D-NSCBC, by correctly modeling the flow at an outlet, can lead to a drastic reduction
of the spatial extent of the computational domain. To show this, the same direct numerical simulation will be undertaken
with a full spatial extent of a computational domain, the reference domain, and with a smaller domain that includes only
one quarter of the points of the reference domain, the small domain. The jet penetration in those domains will then be
compared.
A 2-D computation of an hydrogen jet in an air co-flow is proposed. This setup is derived from the direct numerical sim-
ulation done by Mizobuschi et al. [3,4]. The reference domain consists of a rectangle whose dimensions are close to those
used by Mizobuschi et al. [3,4] (see Fig. 24). The grid is made of 2000  500 nodes, the grid spacing in the x and y direction
is given by:
(
Dx ¼ 50  106 m if x < 95  103 m
ð79Þ
Dxiþ1 ¼ 1:03 Dxi otherwise

(
Dy ¼ 50  106 m if  8:75  103 m < y < 8:75  103 m
ð80Þ
Dyiþ1 ¼ 1:05Dyi otherwise
The small domain is a subset of the original mesh that includes only the first 500 points in the x direction. The gray line in
Fig. 24 indicates the limit of the small domain. The flow set up consists of a hydrogen jet with a 103 m diameter centered on
y = 0 which is injected at 600 m/s yielding a Reynolds number of Re = 5260. The co-flow of air is injected at a constant speed
of 10 m/s. The injection temperature of both flows is 300 K. The jet velocity profile is parabolic to mimic a boundary layer,
and relaxation coefficients at the inlet are set to 0.1 except for u, for which g = 10.
At the outlet, the target pressure is set to 101,325 Pa and b is computed from a reference velocity of 10 m/s. Two adiabatic
wall boundary conditions are used at y = ymax and y = ymin. The two computations start with a flow of air, and a convection
speed along x of 10 m/s.
For the small domain, a lot of strong vortical structures are leaving the domain through the outlet. Some of those generate
flow inversions at the outlet. This flow inversion problem can be tackled in general by imposing a perfectly non reflecting
outlet on the boundary where the flow is reversed (i.e. setting L2 to L4 and L5þk to zero) like in Lodato et al. [15]. However,
if vortical structures are passing through the boundary with small convective velocities or stop on the boundary before
slightly reentering the domain, this technique tends to impose a non-physical temperature at the outlet.
To tackle this problem, when the flow is reverted, an inlet condition is imposed at the exit only where the flow is reverted,
with a temperature of 300 K, tangential velocity of 0 m/s and species mass fraction corresponding to air. Those inlet condi-
tions are imposed using the wave amplitudes given in Eqs. (16), (17) and (18), (20) for L2 to L4 and L5þk . L5 stays unmodified.
When the flow is again exiting the domain the classical outlet boundary condition is imposed. To avoid perturbations of the
flow, these conditions are imposed with small relaxation coefficients ( 0.01).
The resulting turbulent flows for the two domains are shown in Fig. 25. As can be seen, the two flows display similar tur-
bulent structures despite the fact that the domain is a lot smaller for the second case.
The comparison of these two turbulent flows can only be made on the basis of means and the jet penetration is of interest,
so that Y H2 , the Reynolds average of Y H2 , was computed. Y H2 profiles are plotted in Fig. 26 at different axial positions corre-
sponding to x = 2  103 m, x = 3  103 m, x = 4  103 m and x = 5  103 m.
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5603

The profiles are almost identical for the short and long domain, which demonstrate the ability of the present 3-D-NSCBC
formulation to correctly model the behavior of turbulent flow without inducing unphysical distortion in the zone of interest.
Moreover, it proves that by using the 3-D-NSCBC, an increase in the dissipation or in the size of the domain is not needed to
cleanly exit turbulent structures, which can drastically reduce computational time, especially in DNS.

7. Conclusions

A numerical strategy adapted to high order numerical simulations of reactive flows with detailed chemistry effects and
complex transport phenomena has been proposed. The temperature was preferred to the pressure as a non conservative var-
iable for wave decomposition, because it simplifies reaction rates formulations and wall treatment. A new formulation for
the transverse terms relative to the outlet boundary condition has been proposed and successfully validated against previous
formulations. Edges and corner conditions were also carefully addressed. The proposed methodology has been successfully
validated on various test cases for multi-component reactive flows, such as a vortex crossing the outlet non-normally, or a
vortex distorted flame crossing the boundary.

Acknowledgement

The first author was supported by a fellowship from the Fonds National de la Recherche Scientifique, FRS-FNRS (Commu-
nauté Française de Belgique). Computing resources were provided by the IDRIS under the allocation 2009-i2009020164
made by GENCI (Grand Equipement National de Calcul Intensif).

Appendix A. Navier–Stokes equation in characteristic form

The Navier–Stokes equations can be rewritten in the following form:


e @e
@U Fi @ Dei
þ þ e ¼0
þX ðA:1Þ
@t @xi @xi
with:
e ¼ ½q; qu; qv ; qw; qe; qY 1 ; . . . ; qY N T
U ðA:2Þ
sp

N sp 
X 
1 R
qe ¼ qui ui þ q hk  T Yk ðA:3Þ
2 k¼1
Wk

and
0 1
0 1 0 0 1
qui 0
B C
B quui þ d1i p C B 2lA1i C B 0 C
B C B C B C
B C B 2lA2i C B C
B qv ui þ d2i p C B C B 0 C
B C B C B C
B qwui þ d3i p C B 2lA3i C B 0 C
B C B C B C
e
F ¼ B ðqe þ pÞu C
i B ; e
D ¼ B 2luj Aij þ qi C
i B e B
X¼B C ðA:4Þ
i C C; 0 C
B C B @½J i;1  C B C
B qui Y 1 C B  @xi C B W x C
B C B C B 1 _ 1 C
B . C B C B . C
B . C B .. C B .. C
@ . A B . C @ A
@ A
@½J i;Nsp 
qui Y Nsp  @x W Nsp x _ Nsp
i

To formulate the boundary conditions in the form of characteristic waves, one must first rewrite the equations in non-con-
servative variables and then perform a characteristic analysis of those equations. The non-conservative equations read:
@U @U
þ Ai þDþX¼0 ðA:5Þ
@t @xi
with:
U ¼ ½q; u; v ; w; T; Y 1 ; . . . ; Y Nsp T ðA:6Þ
The viscous terms in the conservative equation, D, will be neglected in the characteristics analysis, however as in the work of
Sutherland and Kennedy [16] the source terms will be included in the analysis. To compute the source term vector and the
Jacobian matrices in non-conservative variables one must first define:
e
@U
P¼ ðA:7Þ
@U
In the framework of this paper P and its inverse P1, which allow to rewrite the equations in non-conservative variables (and
the other way around) read:
5604 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

0 1
1 0 0 0 0 0  0
B C
B u1 q 0 0 0 0  0 C
B C
B u2 0 q 0 0 0  0 C
B C
B C
B u3
B
0 0 q 0 0  0 C
   C
P¼B
B e qu1 qu2 R
qu3 qcv q h1  W 1 T    q hN  W Nsp T C
R
C ðA:8Þ
B C
B C
B Y1
B
0 0 0 0 q  0 C
C
B . .. .. .. .. .. .. .. C
B .. . . . . . . . C
@ A
Y Nsp 0 0 0 0 0  q
and
0 1
1 0 0 0 0 0  0
B  u1 1
0 0 0 0  0 C
B q q C
B C
B  u2 0 1
0 0 0  0 C
B q q C
B C
B  u3 0 0 1
0 0  0 C
B q q C
B C
P1 ¼B
B 1 u21 þu22 þu23 h1 þWR T hNsp þW R T C
N sp C ðA:9Þ
B 2 qc v  quc1v  quc2v  quc3v 1
qcv qcv
1
... qcv C
B C
B Y1
0 0 0 0 1
 0 C
B q q C
B C
B .. .. .. .. .. .. .. .. C
B . . . . . . . . C
@ A
Y Nsp 1
q 0 0 0 0 0  q

From this, one can write the Jacobian matrices as:

@e
Fi
Ai ¼ P1 P ðA:10Þ
e
@U
or in a fully developed from:
0 1
ui d1i q d2i q d3i q 0 0  0
B d RT ui 0 0 d1i WR d1i WRT1    d1i WRTN C
B 1i W q C
B sp C
B RT RT C
B d2i 0 ui 0 d2i WR d2i WRT1    d2i W N C
B Wq sp C
B C
B d3i RT 0 0 ui d3i WR d3i WRT1    d3i WRTN C
B sp C
Ai ¼ B W q C ðA:11Þ
B 0 d1i Wc RT RT
d2i Wc RT
d3i Wc ui 0  0 C
B C
B v v v C
B 0 0 0 0 0 ui  0 C
B C
B . .. .. .. .. .. .. .. C
B . C
@ . . . . . . . . A
0 0 0 0 0 0  ui
One can also write the source term vector as:
e
X ¼ P1 X ðA:12Þ
or:
0 1
0
B 0 C
B C
B C
B 0 C
B C
B C
B 0 C
B C
B 0 C
X¼B
B
C
C ðA:13Þ
B 0 C
B C
B 1 C
q W 1 x1
B _ C
B C
B .. C
B . C
@ A
1
q W Nsp xNsp
_

The Ai can be diagonalized using the usual transformation:


A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5605

R i Ki L i ¼ Ai ðA:14Þ
with:
0 1
ui  c
B C
B ui C
B C
B C
B ui C
B C
B C
B ui C
B C
i B
K ¼B C ðA:15Þ
u þ c C
B i C
B C
B u C
B i C
B C
B . C
B .. C
@ A
ui
0 1
1 d1i qT d2i qT d3i qT 1  qWW1     WqNW
B sp C
B C
B d c 1  d 0 0 d1i q c
0  0 C
B 1i q 1i C
B C
B c c C
B d2i q 0 1  d2i 0 d2i q 0  0 C
B C
B C
B d3i c
0 0 1  d3i d3i q c
0  0 C
B q C
Ri ¼ B C ðA:16Þ
B RT RT C
B qWc d1i d2i d3i 0  0 C
B v qWcv C
B C
B 0 0 0 0 0 1  0 C
B C
B C
B .. .. .. .. .. .. .. .. C
B . . . . . . . . C
@ A
0 0 0 0 0 0  1
0 1
1 q q q q qW qW
2c
d1i 2c d2i 2c d3i 2c 2T c 2W 1 c
 2W Nsp c
B C
B C
B d1i RT 1  d1i 0 0 d1i 1c d1i WRT1 cp    d1i W NRT cp C
B W qc p sp C
B C
B d RT 0 1  d2i 0 d2i 1c d2i WRT1 cp    d2i W NRT cp C
B 2i W qcp C
B sp C
B C
B d3i RT 0 0 1  d3i d3i 1c d3i WRT1 cp RT C
   d3i W N cp C
L ¼B
i
B
W qc p sp
C ðA:17Þ
B 1 q q q q Wq Wq C
B d1i 2c d2i 2c d3i 2c  C
B 2c 2T c 2W 1 c 2W Nsp c C
B C
B 0 0 0 0 0 1  0 C
B C
B .. .. .. .. .. .. .. .. C
B C
@ . . . . . . . . A
0 0 0 0 0 0  1

A.1. Faces

The general implementation strategy of the 3-D-NSCBC is similar to the 1D-NSCBC; first, one applies the boundary con-
dition on the characteristic waves, then the temporal derivatives of the non-conservative variables are computed, source
terms are added. Finally the temporal derivative of the conservative variables are computed and viscous terms, modified
by the boundary conditions if needed, are added. This process is described with the relevant equation below.
If one supposes that the face is orthogonal to x1, the characteristic waves will travel through the boundary along this
direction. This means that only A1 needs to be diagonalized to obtain the characteristic wave system. The non conservative
system thus reads:
@U @U @U @U
þ R 1 K1 L 1 þ A2 þ A3 þDþX¼0 ðA:18Þ
@t @x1 @x2 @x3
If one follows the classical approach for the NSCBC proposed by Thompson [10] the vector L is defined as:
@U
L ¼ K1 L1 ðA:19Þ
@x1
whose components are the amplitude of the wave traveling through the boundary:
5606 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

0 !1
N sp
X
1 @q
B ðu1  cÞ
B 2c @x1
 2c @x1 þ 2Tqc @x
q @u1 @T
þ 1
qW @Y k
2W k c @x1
C
C
B k¼1 C
B ! C
B XN sp C
B @q RT @Y k C
B u1  WRT þ 1 @T
 C
B qcp @x1 c @x1 W k cp @x1 C
B k¼1 C
B C
B u1 @u 2 C
B @x1 C
B C
B @u
u1 @x13 C
L¼B C ðA:20Þ
B !C
B XN sp C
B @q q @u1 q @T qW @Y k C
B ðu1 þ cÞ 21c @x þ þ þ C
B 1 2c @x1 2T c @x1 2W k c @x1 C
B k¼1 C
B C
B u1 @Y 1 C
B @x1 C
B C
B .. C
B . C
@ A
@Y Nsp
u1 @x1

Those characteristic waves will be used to compute the waves exiting the domain through the boundary and the waves
entering the domain will be imposed with respect to the desired target values. Once the L vector is computed, one must re-
cover the time derivative of the conservative variable to integrate the 3-D-NSCBC conditions into the computational algo-
rithm. To do so, the non-conservative equation will first be recovered using Eq. (A.18). The transverse terms will be
defined as:
@U @U
T ¼ A2 þ A3 ðA:21Þ
@x2 @x3
or:
0 @q @q
1
u2 @x þ q @u
@x2
2
þ u3 @x þ q @u
@x3
3

B 2 3
C
B u2 @u 1
þ u3 @u 1 C
B @x2 @x3 C
B C
B X N C
B RT @ q @u2 R RT @Y k @u2 C
B þ u2 @x þ
@T
þ þ u 3 @x C
B Wq 2
@x 2 W @x 2 W k @x2 3 C
B k¼1 C
B C
B X N sp C
B @u3 RT @ q @u3 R @T RT @Y k C
T ¼B u þ þ u3 @x3 þ W @x3 þ W k @x3 C ðA:22Þ
B 2 @x2 W q @x3 C
B k¼1 C
B @u @u C
B RT 2
þ u2 @x @T
þ WcRT 3
þ u3 @x @T C
B Wcv @x2 2 v @x 3 3 C
B C
B u @Y 1
þ u @Y 1 C
B 2 @x
2
3 @x
3 C
B C
B .. C
B . C
@ A
@Y Nsp @Y Nsp
u2 @x2 þ u3 @x3

If one neglects, in a first stage, the viscous and source terms, the non conservative system reads:
@U
¼ R1 L  T ðA:23Þ
@t
or:
8
> XNsp
>
> @q
>
>
> @t
¼ L1 þ qT L2  L5 þ qW
L T1
W k 5þk
>
>
>
>
k¼1
>
>
> @u@t1 ¼  qc ðL5  L1 Þ  T 2
>
<
@u2
¼ L3  T 3 ðA:24Þ
>
>
@t
>
> @u3
>
> ¼ L4  T 4
>
>
@t
> @T ¼  RT ðL1 þ L5 Þ  L2  T 5
>
>
> @t qWcv
>
> @Y
:
@t
k
¼ L 5þk  T 5þk k ¼ 1; . . . ; Nsp

Then one can add the source terms and isolate the temporal derivatives:
@U @U
¼ þX ðA:25Þ
@t @t
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5607

Finally the temporal derivatives of the non-conservative variables are multiplied by P and the viscous terms are added to
recover the temporal derivatives of the conservative variables for integration into the computational algorithm:
e
@U @U @ Dei
¼P þ ðA:26Þ
@t @t @xi

A.2. Edges

The edges follow a quite similar process: computation of the characteristic waves, computation of the temporal derivative
of the conservative variable and finally insertion to the computational algorithm via the temporal derivative of the conser-
vative variables. If one suppose an edge located along x3, two sets of characteristic waves must be computed:
@U @U @U @U
þ R 1 K1 L 1 þ R 2 K2 L 2 þ A3 þDþX¼0 ðA:27Þ
@t @x1 @x2 @x3
and:
@U
L ¼ K1 L1 ðA:28Þ
@x1
@U
M ¼ K2 L2 ðA:29Þ
@x2
The L vector stays the same as for the face (see Eq. (A.20)). The characteristic waves along x2 are defined as:
0 !1
XN sp
@q
B ðu2  cÞ 21c @x q @u2 q @T qW @Y k C
B 2
 2c @x 2
þ 2T c @x 2
þ 2W k c @x2 C
B k¼1 C
B C
B u2 @u 1 C
B @x2 C
B ! C
B XN sp C
B RT @q 1 RT @Y k C
B u2  W qc @x2 þ c @x2  @T C
B p W k cp @x2 C
B k¼1 C
B C
B u2 @u 3 C
M¼B @x2 C ðA:30Þ
B !C
B XN sp C
B @q q @u2 q @T qW @Y k C
B ðu2 þ cÞ 21c @x þ þ þ C
B 2 2c @x2 2T c @x2 2W k c @x2 C
B k¼1 C
B C
B u2 @Y 1 C
B @x2 C
B C
B .. C
B . C
@ A
@Y Nsp
u2 @x2

The transverse terms vector contains only the derivatives along x3 for the edge:
@U
T ¼ A3 ðA:31Þ
@x3
or:
0 @q
1
u3 @x 3
þ q @u
@x3
3

B C
B u3 @u 1 C
B @x3 C
B C
B u3 @u 2 C
B @x3 C
B C
B N sp
X C
B RT @q @Y k C
B þ u3 @u 3
þ WR @T
þ RT C
B Wq @x3 @x3 @x3 Wk @x3 C
T ¼B k¼1 C ðA:32Þ
B C
B RT @u3 @T
þ u3 @x C
B Wcv @x3 3 C
B C
B u3 @Y 1 C
B @x3 C
B C
B .. C
B . C
@ A
@Y Nsp
u3 @x3

The non-conservative equations without the viscous and sources terms reads:
@U
¼ R1 L  R2 M  T ðA:33Þ
@t
5608 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

which yields:

8
N sp
X
>
>
>
> @@tq ¼ ðL1 þ L5 Þ  ðM1 þ M5 Þ þ qT ðL2 þ M3 Þ þ qW
ðL5þk þ M5þk Þ  T 1
>
> Wk
>
> k¼1
>
>
>
>
>
> @u@t1 ¼  qc ðL5  L1 Þ  M2  T 2
>
>
>
<
@u2
@t
¼ L3  qc ðM5  M1 Þ  T 3 ðA:34Þ
>
>
>
> @u3 ¼ L  M  T
>
>
>
> @t 4 4 4
>
>
>
> @T RT
¼  qWc ðL1 þ L5 Þ  L2  qWcRT
ðM1 þ M5 Þ  M3  T 5
>
> @t
>
>
v v
> @Y
:
@t
k
¼ L5þk  M5þk  T 5þk k ¼ 1; . . . ; Nsp

The procedure to recover the temporal derivative of the conservative variables stays the same as for the face boundary.

A.3. Corners

Like for the edges boundary the procedure stays similar as for the faces, except that three direction should be considered
and so the transverse terms vanish.The non-conservative set of equations is:

@U @U @U @U
þ R 1 K1 L 1 þ R 2 K2 L 2 þ R 3 K3 L 3 þDþX¼0 ðA:35Þ
@t @x1 @x2 @x3

and:

@U
L ¼ K1 L1 ðA:36Þ
@x1
@U
M ¼ K2 L2 ðA:37Þ
@x2
@U
N ¼ K3 L 3 ðA:38Þ
@x3

L and M are the same as for the edges, the N vector reads:
0 !1
N sp
X
1 @q
B ðu3  cÞ
B 2c @x3
 2c @x3 þ 2Tqc @x
q @u3 @T
þ 3
qW @Y k
2W k c @x3
C
C
B k¼1 C
B C
B @u1 C
B u3 @x3 C
B C
B C
B u3 @u 2 C
B @x3 C
B C
B ! C
B XN sp C
B RT @q 1 @T RT @Y k C
B u3  W qc @x3 þ c @x3  W k cp @x3 C
B p C
N ¼B k¼1 C ðA:39Þ
B !C
B XN sp C
B C
B ðu3 þ cÞ 1 @ q þ q @u3 þ q @T þ qW @Y k C
B 2c @x3 2c @x3 2T c @x3 2W k c @x3 C
B k¼1 C
B C
B C
B u @Y 1
3 @x C
B 3 C
B C
B . C
B .. C
B C
@ A
@Y N
u3 @x3sp

If one rewrites the non-conservative equations without the viscous and source terms, this yields:

@U
¼ R1 L  R2 M  R3 N ðA:40Þ
@t
A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5609

or:
8 Nsp
>
> @q q
X qW
>
> @t ¼ ðL1 þ L5 Þ  ðM1 þ M5 Þ  ðN 1 þ N 5 Þ þ T ðL2 þ M3 þ N 4 Þ þ ðL5þk þ M5þk þ N 5þk Þ
>
> Wk
>
> k¼1
>
>
>
>
>
> @u1
¼  qc ðL5  L1 Þ  M2  N 2
>
> @t
>
>
>
>
< @u2
@t
¼ L3  qc ðM5  M1 Þ  N 3 ðA:41Þ
>
>
>
>
> @u3 c
> @t ¼ L4  M4  T 4  q ðN 5  N 1 Þ
>
>
>
>
>
>
>
> @T RT
¼  qWc RT
ðL1 þ L5 Þ  L2  qWc RT
ðM1 þ M5 Þ  M3  qWc ðN 1 þ N 5 Þ  N 4
>
>
> @t
>
v v v
>
>
: @Y k
@t
¼ L5þk  M5þk  N 5þk k ¼ 1; . . . ; Nsp
The rest of the procedure to recover the temporal derivatives of the conservatives variables is again similar to the one for the
faces.

Appendix B. Rotation

As explained above, all the boundary conditions presented are written for a face located at x1 = x 1max, for an edges located
at x1 = x1min and x2 = x2min and for a corner located at x3 = x3min. A rotation has to be made to make a boundary at any location
correspond to the one presented. Such rotations are based on the tensor and vector rotation formulae, with the upper case
letter representing the new frame and the lower caser letter the old one:
xP ¼ aPi xi ðB:1Þ
T PQ ¼ aPi aQj T ij ðB:2Þ
with aPi being the cosine of the angle between axis Q, in the new frame, and i in the old one:
aPi ¼ cosðxP ; xi Þ ðB:3Þ
The following tables will give the aPi used to change the frame from the general coordinates (x, y, z) to the local boundary
coordinates (x1, x2, x3). Note that an inverse rotation must be applied to recover the temporal derivatives of the conservative
equations in the general flow frame. Also, not only the velocity components have to be rotated but also the derivative and the
target value for the velocities. All this is done by applying the tensor rotation formula (Eq. (B.2))

B.1. Faces rotations


5610 A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611

B.2. Edges rotations

B.3. Corners rotations


A. Coussement et al. / Journal of Computational Physics 231 (2012) 5571–5611 5611

References

[1] E. Hawkes, J. Chen, Direct numerical simulation of hydrogen-enriched lean premixed methane-air flames, Combustion and Flame 138 (2004) 242–258.
[2] T. Echekki, J. Chen, Direct numerical simulation of autoignition in non-homogeneous hydrogen-air mixtures, Combustion and Flame 134 (2003) 169–
191.
[3] Y. Mizobuchi, S. Tachibana, J. Shinio, S. Ogawa, T. Takeno, A numerical analysis of the structure of a turbulent hydrogen jet lifted flame, Proceedings of
the Combustion Institute 29 (2002) 2009–2015.
[4] Y. Mizobuchi, J. Shinjo, S. Ogawa, T. Takeno, A numerical study on the formation of diffusion flame islands in a turbulent hydrogen jet lifted flame,
Proceedings of the Combustion Institute 30 (2005) 611–619.
[5] T. Colonius, S. Lele, Computational aeroacoustics: progress on nonlinear problems of sound generation, Progress in Aerospace sciences 40 (2004) 345–
416.
[6] S. Lele, Compact finite difference schemes with spectral-like resolution, Journal of Computational Physics 103 (1992) 16–42.
[7] T.J. Poinsot, S.K. Lele, Boundary-conditions for direct simulations of compressible viscous flows, Journal of Computational Physics 101 (1992) 104–129.
[8] M. Baum, T. Poinsot, D. Thevenin, Accurate boundary-conditions for multicomponent reactive flows, Journal of Computational Physics 116 (1995) 247–
261.
[9] N. Okong’o, J. Bellan, Consistent boundary conditions for multicomponent real gas mixtures based on characteristic waves, Journal of Computational
Physics 176 (2002) 330–344.
[10] K.W. Thompson, Time dependent boundary conditions for hyperbolic systems, Journal of Computational Physics 68 (1987) 1–24.
[11] K.W. Thompson, Time-dependent boundary conditions for hyperbolic systems, ii, Journal of Computational Physics 89 (1990) 439–461.
[12] W. Polifke, C. Wall, P. Moin, Partially reflecting and non-reflecting boundary conditions for simulation of compressible viscous flow, Journal of
Computational Physics 213 (2006) 437–449.
[13] C.S. Yoo, Y. Wang, A. Trouve, H.G. Im, Characteristic boundary conditions for direct simulations of turbulent counterflow flames, Combustion Theory
and Modelling 9 (2005) 617–646.
[14] C.S. Yoo, H.G. Im, Characteristic boundary conditions for simulations of compressible reacting flows with multi-dimensional, viscous and reaction
effects, Combustion Theory and Modelling 11 (2007) 259–286.
[15] G. Lodato, P. Domingo, L. Vervisch, Three-dimensional boundary conditions for direct and large-eddy simulation of compressible viscous flows, Journal
of Computational Physics 227 (2008) 5105–5143.
[16] J.C. Sutherland, C.A. Kennedy, Improved boundary conditions for viscous, reacting, compressible flows, Journal of Computational Physics 191 (2003)
502–524.
[17] B. Gustafsson, A. Sundström, Incompletely parabolic problems in fluid dynamics, SIAM Journal on Applied Mathematics 35 (1978) 343–357.
[18] D. Rudy, J. Strikwerda, A nonreflecting outflow boundary condition for subsonic Navier–Stokes calculations, Journal of Computational Physics 36
(1980) 55–70.
[19] R. Higdon, Initial-boundary value problems for linear hyperbolic systems, SIAM Review (1986) 177–217.
[20] C. Kennedy, M. Carpenter, Several new numerical methods for compressible shear-layer simulations, Applied Numerical Mathematics 14 (1994) 397–
433.
[21] M. Visbal, D. Gaitonde, On the use of higher-order finite-difference schemes on curvilinear and deforming meshes, Journal of Computational Physics
181 (2002) 155–185.
[22] R. Kee, F. Rupley, E. Meeks, J. Miller, CHEMKIN-III: a FORTRAN chemical kinetics package for the analysis of gas-phase chemical and plasma kinetics,
Sandia National Laboratories, Report SAND96-8216, 1996.
[23] P. Renard, D. Thevenin, J. Rolon, S. Candel, Dynamics of flame/vortex interactions, Progress in Energy and Combustion Science 26 (2000) 225–282.
[24] V. Katta, L. Goss, W. Roquemore, Effect of nonunity Lewis number and finite-rate chemistry on the dynamics of a hydrogen-air jet diffusion flame,
Combustion and Flame 96 (1994) 60–74.
[25] H. Schlichting, K. Gersten, K. Gersten, Boundary-Layer Theory, Springer Verlag, 2000.

Das könnte Ihnen auch gefallen