Backscattering in complex flows: application of the

One-Way Euler equations to Poiseuille flow inside lined

duct
Clément Rudel, Sébastien Pernet, Jean-Philippe Brazier

To cite this version:

Clément Rudel, Sébastien Pernet, Jean-Philippe Brazier. Backscattering in complex flows: application
of the One-Way Euler equations to Poiseuille flow inside lined duct. AIAA AVIATION 2021 FORUM,
Aug 2021, VIRTUAL EVENT, United States. �10.2514/6.2021-2138�. �hal-03526085�

HAL Id: hal-03526085

https://hal.archives-ouvertes.fr/hal-03526085
Submitted on 14 Jan 2022

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Backscattering in complex flows: application of the One-Way
Euler equations to Poiseuille flow inside lined ducts

Clément Rudel∗ , Sébastien Pernet† and Jean-Philippe Brazier‡

ONERA, Université de Toulouse, Toulouse, France, 31400

We present an improved formulation of the numerical One-Way approximation that permits

to diagonalize in a fully numerical way the propagation operator of a general hyperbolic system.
In this paper, it is applied to the linearized Euler equations in the case of a duct with a partially
lined section presenting a Poiseuille flow. Domain decomposition is developed in this paper since
it is needed for the handling of the transverse boundary conditions discontinuity (hard-wall
and impedance condition) which is at the origin of the refraction and reflection of the incident
wave inside the computational domain. The results of the One-Way method are compared to
experimental and numerical results and show a good accuracy for a low amount of computing
resources.

I. Introduction
The reduction of noise produced by an aircraft is, nowadays, an active field of study since it is a major issue
of the development of the commercial aviation worldwide. In particular, we are interested in reducing the sound
emitted by aircraft during take off and approach phases since more and more residential area are built close to airports.
Acoustic liners are already used inside most of the recent airliners to reduce the sound produced by the engines but
their development requires many numerical simulations and experimental setups to be effective. Therefore, computing
quickly and easily numerous configurations of liner is an essential issue for the future of the aviation.
The results obtained through experimental setups give invaluable and accurate insights but are undeniably costly,
difficult to set up and do not allow to get all the wanted data from a case of study.
On the other hand, numerically, the studied setup is classically modeled with a partially lined duct crossed by (in our
case) a subsonic Poiseuille flow. In this configuration, hard wall boundary conditions are imposed at the top all along the
duct while the bottom wall is composed of an impedance condition surrounded by hard walls, creating a discontinuity in
the computational domain. In this kind of simulations, the main challenges are to model accurately the liner behavior
with an impedance condition depending on the frequency and to capture well the reflected and refracted waves coming
from the liner leading and trailing edge. Both issues are to be studied carefully in order to obtain accurate results since
their effects on the propagation of waves is non negligible and the latter is, in particular, the subject of this paper.
Some methods such as the Direct Numerical Simulation (DNS) or Large Eddy Simulation (LES) [1, 2], are able
to give an accurate solution of this wave scattering in all the propagation directions. However, both methods are
particularly costly in computational resources and cannot be easily applied for large models or to quickly compute
several configurations.
The study of the linear phenomena inside ducts has already shown its potential and already gives good results [3]
and takes naturally the reflection and refraction of waves into account but presents some problems. In time domain,
some hypothesis are taken such as a uniform flow inside the duct but the impedance condition applied to model the effect
of the liner is not straightforward to develop and can lead to numerical instabilities [4]. Therefore, the presence of a
shear base flow to linearize the equations can lead to complex developments of impedance and non-reflecting boundary
conditions (NRBC). The simulations can also be carried out inside the frequency domain but the problematic remains
the same with the need to build boundary conditions that allow the waves propagation outside of the computational
domain without reflection. Moreover, inside the frequency domain, this direct resolution method requires to solve a
large linear system that can necessitate a large amount of computational resources and CPU time.
On the other hand, we can decide to solve the system locally by performing a spatial resolution along a privileged
direction which is imposed by the base flow. The construction of One-Way approximations for hyperbolic systems of
equations is precisely based on the uncoupling, in the microlocal sense, between the waves propagating downstream
∗ PhD student, DTIS, clement.rudel@onera.fr
† Dr, DTIS, sebastien.pernet@onera.fr
‡ Dr, DMPE, jean-philippe.brazier@onera.fr

1
 and upstream. This microlocal approach of the One-Way equations has been intensively used in geophysics, acoustics
or electromagnetics for almost fifty years, but they have not been applied to more complex systems yet, like Euler or
Navier-Stokes, since this method requires approximations of pseudo-differential operators that can be hard to obtain in
this context.
Recently, Towne et al. [5] proposed a new formulation which we will call numerical One-Way, in opposition to the
analytical one requiring pseudo-differential operators, that showed convincing results when applied to wave propagation
in inhomogeneous flows. Based on the concept of non-reflecting boundary condition, this approach makes possible to
get rid of the analytical expression of the operators, which permits to apply a One-Way decomposition to more complex
systems. However, it still suffers from a limitation: a slow varying hypothesis for the flow along the privileged direction.
We propose to apply the One-Way method (Sec. II.A) to the case of a lined duct with a sheared flow (Poiseuille).
In order to obtain an accurate result on this configuration presenting discontinuities and, therefore, reflection and
transmission of the incident wave, a new formulation of the numerical One-Way is developed in Sec. II.B. The
computation parameters are presented in Sec. III and the problematic of reflection and transmission of waves is handled
in Sec. IV with the development of a domain decomposition method. Finally, several numerical results are compared to
previous computations and experimental data in Sec. V.

II. Method

A. Linearized Euler equations and numerical One-Way

The development of the One-way equations will be based on the 2D dimensionless Euler equations. Therefore, we
will focus on the formulation of the numerical One-Way as stated by Towne and Colonius [6, 7], but some reviews
are available concerning the more standard One-Way formulation [8, 9]. All the variables and coefficients have been
made dimensionless by reference values taken from the problem and/or the base flow. The Euler equations used for this
method have to be linearized around a base flow and we will use the following decomposition of the variable 𝑞:

q(𝑥, 𝑦, 𝑡) = q̄(𝑥, 𝑦) + q0 (𝑥, 𝑦, 𝑡) , (1)

q0
with q̄ the variable of the steady base flow around which we linearize the equations and the fluctuations of this
variable that we will compute. In our case, we will assume that 𝑝¯ = 𝑝 0 constant, 𝑢¯ = 𝑢(𝑥,
¯ 𝑦) and 𝑣¯ = 0 and denote the
base velocity vector Ū = ( 𝑢¯ 𝑣¯ ) 𝑡 .
Since we are only interested in the stationary behavior of the system, we will consider the Euler equations in a
harmonic regime by decomposing the fluctuations into a summation of frequency modes:
Õ
q0 (𝑥, 𝑦, 𝑡) = q 𝑁 (𝑥, 𝑦)𝑒 −𝑖 𝜔𝑡 ,
e (2)
𝑁

with 𝜔 the wave number.

We will use, for the development of this method, the following 2D isentropic compressible linearized Euler equations:
3

 𝐷e𝜌


 +e𝜌 ∇ · Ū + ∇ · ( 𝜌e
¯ u) = 0



 𝐷𝑡
𝐷eu
, (3)
 𝜌¯ + 𝜌¯ (e
u · ∇) Ū + e𝜌 Ū · ∇ Ū + ∇ 𝑝e = 0


 𝐷𝑡
 𝑝e = 𝑎 2 e

𝜌

𝐷
with 𝐷𝑡 the transport operator which can be written in harmonic regime: −𝑖𝜔 + Ū · ∇. We will assume for the rest of
this paper that 𝜌¯ and 𝑎, the speed of sound have a constant value and that we will only work with the primitive variables
𝑝e, e
𝑢 and e𝑣 . Therefore, the variable vector 𝑞e can be written:

𝑝e
© ª
q = ­e
e ­ 𝑢 ®® . (4)
«e𝑣¬
𝜕
Then, by isolating the 𝜕𝑥 part, we can put Equation (3) into the following matrix form:

2
 q
𝜕e q
𝜕e
𝐴(𝑥, 𝑦) = 𝐵(𝑥, 𝑦) + 𝐶 (𝑥, 𝑦)e
q, (5)
𝜕𝑥 𝜕𝑦
with:

𝑢¯ 𝑎 2 𝜌¯ 0 0 0 −𝑎 2 𝜌¯ 𝑖𝜔 0 0
©1 ª © ª © ª
𝐴 = ­ 𝜌¯
­ 𝑢¯ 0® ; 𝐵 = ­ 0
® ­ 0 0 ® ; 𝐶 = ­ 0 𝑖𝜔
® ­ −𝑢¯ 𝑦 ®® , (6)
1
«0 0 𝑢¯ ¬ «− 𝜌¯ 0 0 ¬ «0 0 𝑖𝜔 ¬
with the subscript 𝑦 denoting the derivative in 𝑦 of the base flow variable. This expression of the Linearized Euler
equations is simplified due to the fact that we will study in this case a constant Poiseuille flow presenting variations only
in the 𝑦-direction for the axial velocity. From now on, for the sake of readability, the variable e q will be denoted q.
We proceed to a variable change by diagonalizing the matrix 𝐴 in order to have access to the characteristic variables
 𝑡

𝝓 = 𝜙1 𝜙2 𝜙3 . We introduce the transformation matrix 𝑇 and its inverse 𝑇 −1 to diagonalize the matrix 𝐴 in the
following way:
!
e+ 0
𝐴
−1
𝑇 𝐴𝑇 = 𝐴 e= , (7)
0 𝐴 e−

with 𝐴
e be the diagonal matrix containing all the eigenvalues of 𝐴 which are all real since the system of equation is
hyperbolic. The matrices 𝐴 e+ and 𝐴
e− contain the positive and negative eigenvalues on their diagonal, respectively. In
our particular case, since our base flow velocity is subsonic and positive, we have:
!
𝑢¯ + 𝑎 0 e− = 𝑢¯ − 𝑎 .
𝐴+ =
e and 𝐴 (8)
0 𝑢¯

It can be noticed that, in the case where 𝑢¯ = 0, a singularity can appear in the system, making 𝐴 e non-invertible. This
issue is not within the scope of this paper and the method to deal with it can be find in [10]. The eigenvectors matrices
of 𝐴 can also be easily obtained:
1 1
1 0 1 2 2𝑎 𝜌¯ 0
© ©
0 −𝑎 𝜌¯ ®® and 𝑇 −1 = ­­ 0
ª ª
𝑇 = ­𝑎 𝜌¯
­ 0 1®® . (9)
1
«0 1 0 ¬ «2 − 2𝑎1𝜌¯ 0¬

Since constant values of 𝜌¯ and 𝑎 are assumed, the eigenvectors matrix 𝑇 and its inverse 𝑇 −1 do not depend of 𝑥 or 𝑦.
These matrices are also used to perform the variable change between q and 𝝓 following the relation 𝝓 = T−1 q. Then,
we can put Equation(5) under the following form:
𝜕𝝓
= 𝑀 (𝑥, 𝑦)𝝓 , (10)
𝜕𝑥
e−1𝑇 −1 𝐵(𝑥, 𝑦)𝜕𝑦 + 𝐶 (𝑥, 𝑦) 𝑇. The matrix 𝑀 can be denoted as the propagation operator and its


with 𝑀 (𝑥, 𝑦) = 𝐴
expression is:
¯ 2 𝑢¯ 𝑦
¯
1
0 0 © 𝑖𝜔 − 𝜌𝑎2 𝜕𝑦 − 2𝑎 𝜌¯ 0 ª
© 𝑢+𝑎 ª­ 𝜕
𝑀 (𝑥, 𝑦) = ­­ 0 1 𝜕 ®
(11)
𝑢¯ 0 ®® ­− 𝜌¯𝑦 𝑖𝜔 − 𝜌¯𝑦 ® ,
1 ¯ 2
­ 𝑢¯ 𝑦
®
« 0 0 ¯
𝑢−𝑎 ¬« 0 − 𝜌𝑎
2 𝜕𝑦 + 2𝑎 𝜌¯ 𝑖𝜔 ¬
𝜕
with 𝜕𝑦 be the derivative operator 𝜕𝑦 in 𝑦.
The particularity of the numerical One-Way approach over the microlocal one [11] is that the transverse discretization
of the system appears before the diagonalization of the propagation operator. It will allow us to separate the right-going
waves from the left-going ones without analytically building the dispersion relation, which is difficult to do with a
system like the Euler equations. It can be noted that the transverse boundary conditions like wall or slip conditions,
impedance or non reflective boundary conditions have to be implemented in the system during this step (Sec. III.B). All

3
the transversally discretized vectors and matrices valued operators and variables will be denoted by bold characters.
However, even if the discretization of the equation will make us lose the track of the analytical terms during the
diagonalization of 𝑀, the separation of the characteristic variables into two sets depending on the
 sign of 𝐴 is still
e
𝑡
guaranteed and it will be used for the decomposition of the propagation operator. Therefore, 𝝓 = 𝝓+ 𝝓− with the
subscript + denotes the downstream modes while − the upstream ones. The number of discretized points of the vectors
𝝓± is also given by the subscripts: + denotes 𝑁+ points while − designates 𝑁− points. In our case, we can deduce that
𝑁+ = 2𝑁 𝑦 and 𝑁− = 𝑁 𝑦 with 𝑁 𝑦 the number of discretization points in the transverse direction since 𝐴 e+ is a 2 × 2
 𝑡
matrix while 𝐴e− is a scalar function. We can also note that 𝝓+ = 𝝓1 𝝓2 while 𝝓− = 𝝓3 .
The same operation of separation can be applied to the propagation operator M depending on the wave propagation
direction. The equation (10) can be written such as:
! ! !
𝜕 𝝓+ M++ M+− 𝝓+
= , (12)
𝜕𝑥 𝝓− M−+ M−− 𝝓−
with the first subscript denoting the number of rows (𝑁+ or 𝑁− ) and the second, the number of columns of the block
matrix.
The discretized operator M can also be diagonalized following an eigenvalue decomposition with the transformation
V and its inverse U such as:

M = VDU , (13)
with D the diagonal matrix containing all the eigenvalues of M and V the corresponding right eigenvectors matrix or U
the corresponding left eigenvectors matrix.
+𝑁−
𝑁+Í
A new variable change can be performed, using this decomposition, defined by 𝝓 = V𝝍 = 𝝓 = 𝜓 𝑗 v 𝑗 with v 𝑗
𝑗=1
the 𝑗-th eigenvector of V. We keep once again the separation made during the first diagonalization with the + and −
subscript and equation (12) becomes:
! ! ! !
𝜕 𝝍+ D+ 0 𝝍+ 𝜕V 𝝍 +
= +U . (14)
𝜕𝑥 𝝍 − 0 D− 𝝍 − 𝜕𝑥 𝝍 −

Here, the matrix W = U 𝜕V 𝜕𝑥 , is the refraction/reflection matrix and it couples the functions 𝝍 + and 𝝍 − together.
This coupling would make this system ill-posed for a resolution using a spatial marching algorithm in 𝑥. This is why
this term is neglected in the construction of the numerical One-Way to decouple the left and right going information.
Neglecting this term is mandatory to resolve a well-posed system, but it is also at the origin of the slow varying flow
hypothesis [7] along the privileged direction (𝑥 in our case). Equation (14) is now reduced to:
! ! !
𝜕 𝝍+ D+ 0 𝝍+
= . (15)
𝜕𝑥 𝝍 − 0 D− 𝝍 −
It can be noted that the decoupling procedure does not require a complete diagonalization of the matrix M and it
can be realized by a less fine spectral decomposition where only the eigenspaces are identified. In other words, the
construction of the One-Way method can be based on a factorization of 𝑀 where the matrix D e can be a block diagonal
matrix.
The One-Way method is based on the fact that during the computation, the upstream propagating waves are neglected.
We can enforce that by setting the variable 𝝍 − = 0. This condition can be reversed if we want to compute the leftgoing
modes by applying 𝝍 + = 0. Therefore, the One-Way system takes the following form:
𝜕𝝍
 + = D+ 𝝍 +



𝜕𝑥 , (16)
 𝝍 =0

 −

which can be written in term of characteristic variables:

4
 𝜕𝝓
 + = M++ 𝝓+ + M+− 𝝓−



𝜕𝑥 . (17)
 𝝓 = U−1 U−+ 𝝓

 − −− +

The eigenvalues decomposition of the operator M in D, V and U is really costly to perform and it has to be computed
for each station in 𝑥 if the flow is varying along the 𝑥-axis. This is why Equations (14) and (15) are only here to show
the mechanisms of this method and the origin of its limitation. Therefore, as introduced in [7, 12], a non-reflective
boundary condition [13, 14] has been used under the form of a recursive relation in order to approximate the second
equation of system (17) which becomes:
𝜕𝝓+
= M++ 𝝓+ + M+− 𝝓−





 𝜕𝑥
𝛽 −1

 ! 𝑁Ö !
 𝑁𝛽  𝝓
𝝓+
 
𝑗
−1 𝑗 + , (18)
= M − 𝑖𝛽− I M − 𝑖𝛽+ I
 0


 𝑗=0
𝝓 −

 | {z }


 Z𝑟 : Non-reflection matrix
𝑁
with 𝝓+ 𝛽 an auxiliary variable and Z𝑟 the non-reflection matrix for a right One-Way where the right-going waves are
computed while the left-going ones are neglected. Then, the discretized system (18) can be efficiently solved by a spatial
marching algorithm in 𝑥. The coefficient 𝛽± are to be chosen arbitrarily, but with some care. In order to be efficient, the
choice of parameters has to put the 𝛽+ coefficient near the cluster of rightgoing modes of the operator M while the
𝛽− ones are to be chosen near the leftgoing modes cluster. Moreover, the more these two sets of coefficients are far
from each other, the better the convergence is. Their role is to separate the eigenspace generated by the downstream
eigenvectors from the eigenspace generated by the upstream ones. The required insights on the spectrum of M to chose
these sets can be acquired through a simplified local linear stability analysis of the operator.

B. New diagonalization of the operator

Since our problem presents some variations that are only due to the discontinuity in the transverse boundary
condition, the reflected and transmitted waves will not be computed by taking into account the refraction/reflection
matrix W. Instead, we will use the eigenvectors matrices U and V to decompose the incident wave into its rightgoing
and leftgoing components (Sec. IV). Therefore, if we want to take into account these phenomena, we need to express our
problem through the 𝜓± variables which are not accessible through the current formulation of the numerical One-Way.
In order to do that, the NRBC used for the numerical One-Way will be used again, but this time in the right-going and
left-going directions permitting us to well separate the two different eigenspaces generated. Most of the following
construction will be based on the right One-Way formalism but it is easy to build the same for a left One-Way by
interchanging the + and − parts of the matrices.
The first step is to build our transformation matrices U and V. In order to do that, we will use the same recursive
relation as in the numerical One-Way to form the Z𝑟 matrix. Using the last equation of system (18) to express 𝜙− using
Z𝑟 and 𝜙+ , we obtain:

−1
𝝓− = Z𝑟−− Z𝑟−+ 𝝓+ . (19)
This relation can be compared to the last relation of system (17) to obtain the following equality:

−1
Z𝑟−− Z𝑟−+ ≈ (U−− ) −1 U−+ . (20)
For the rest of the construction, we assume that the choice and the number of 𝛽± coefficient is taken to ensure a good
enough convergence to replace the symbol ≈ by =. The same relation can be obtained using Z𝑙 and the −− and −+ parts
of U.
Since the matrix Z built for the right or the left One-Way is composed by the matrix M, it is easy to see that it has
the same eigenvectors matrices V and U leading to the following relation:
! !
U++ U+− Z++ Z+−
= F (D) U , (21)
U−+ U−− Z−+ Z−−

5
with F(D) the diagonal matrix composed of the reflection coefficient generated by the NRBC applied to the eigenvalues
of M. Formally, these reflection coefficients make F(D+,i ) → ∞ and F(D−,i ) → 0 when using the matrix Z𝑟 and
F(D+,i ) → 0 and F(D−,i ) → ∞ when using the matrix Z𝑙 . The symbol → indicates that 𝑁 𝛽 tends to +∞ and that the
coefficients are well-placed. Then, we can deduce the following relation:

Span U𝑖,∗ ∈ C 𝑁+ +𝑁− | 𝑖 = 𝑁+ + 1, ..., 𝑁+ + 𝑁− ∈ ker Z𝑟 ,


(22)
meaning that the − eigenvectors of U are in the kernel ofZ𝑟 . In the same way, we can easily show that the + eigenvectors
of U are in the kernel of Z𝑙 . Using the converged relation (20) into (21), we can deduce, for the right One-Way:
(
Z𝑟−+ Z𝑟++ + Z𝑟−− Z𝑟−+ = 0
. (23)
Z𝑟−+ Z𝑟+− + Z𝑟−− Z𝑟−− = 0
Once again, these relations show that the vectors inside Z𝑟−+ and Z𝑟−− are also in the kernel of Z𝑟 in the same way as
the − eigenvectors of U are in this kernel. This leads to:

Span Z𝑟𝑖,∗ ∈ C 𝑁+ +𝑁− | 𝑖 = 𝑁+ + 1, ..., 𝑁+ + 𝑁− ∈ ker Zr .


(24)
The same conclusion can be made for Z𝑙 whose kernel contains the vectors generated by Z𝑙++ and Z𝑙+− . By using these
two results, we can build a new matrix U
e composed of the relevant parts of Z𝑟 and Z𝑙 that generates an approximation of
the same eigenspace as U and equally separated in its rightgoing and leftgoing contributions. This matrix has the form:
!
Z l Z l
++ +−
e=
U . (25)
Zr−+ Zr−−
This newly formed matrix of "pseudo eigenvectors" permits a new diagonalization of M following:

UM
e Ve=D
e, (26)
with D
e a diagonal block matrix. From that, we can build an approximation of the One-Way system (16) by replacing D
with its approximate D.
e
The additional computation time induced by the construction of the "pseudo eigenvectors" matrix and the
refraction/reflection term, compared to the standard numerical One-Way, is of around 50 to 60% on this part of the
code. Indeed, since both formulations share the same spatial marching algorithm and the construction of the discretized
system, no extra computational time is induced by the rest of the method.

III. Application to lined ducts

The case of a lined duct presenting a sheared flow has been chosen for this application of the One-Way since we
want to compute a case that is well-documented and has experimental comparison data. Moreover, it is a case where the
variation along 𝑥 does not come from the flow but from the boundary conditions and more particularly, this variation is
a discontinuity inside the discretized propagation operator M. Therefore, it is a case where the information brought
by the One-Way diagonalization is of vital interest since it will be used to decompose the incident wave propagating
through such discontinuity into a reflected and transmitted wave.

A. Configuration of the lined duct

Two different configurations are studied here and they have been chosen in order to apply the One-Way method to a
certain number of different cases. The first one is developed in [15] for the experimental results and in [16] for the
selected numerical ones and it will be referred to as the Configuration A from now on. The second configuration can
be found in [17] for the experimental as well as the numerical results and it will be denoted Configuration B. Both
configurations are very similar and can be described by the Fig. 1. However, some differences can be found in term of
base flow velocity, dimensions and liner characteristics.
Concerning the base flow, the same laminar fully developed Poiseuille flow will be used in both configurations but
with different values of 𝑀max . This flow is modeled by the following equation [17]:
𝑦  𝑦
¯
𝑢(𝑦) = 4𝑀max 𝑎 1 − , (27)
ℎ ℎ

6
 Fig. 1 Configuration of the lined duct

with 𝑀max the maximum value of the Mach number inside the duct, equal to 0.508 for the Configuration A while it is
set to 0.3 in Configuration B.
The physical dimensions of the computational domain are also different due to the fact that we base our computations
on existing experimental installations. Table 1 shows the particular dimensions of each configuration.

Configuration A Configuration B
ℎ 0.051 𝑚 0.0508 𝑚
𝐿 𝑖𝑛 0.203 𝑚 0.210 𝑚
𝐿 𝑙𝑖𝑛𝑒𝑟 0.203 𝑚 0.387 𝑚
𝐿 𝑜𝑢𝑡 0.203 𝑚 0.243 𝑚
Liner CT57 CT73
Table 1 Configuration dimensions

For both configurations, the simulations have been made dimensionless by choosing 𝐿 ref = ℎ, 𝑉ref = 340 m/s and
𝜌ref = 1.23 kg/m3 as reference values.
Finally, the numerical parameters used for all the simulations are more or less the same. In 𝑦, the mesh is composed
of 50 points evenly spaced (unless mentioned otherwise) and 500 points in 𝑥 with an exponential distribution to have a
refined mesh around the discontinuities. This mesh has been chosen to have a well discretized mesh of 60 points per
wavelength on average, in 𝑥, for the case presenting the highest frequency. The transverse discretization is done by a
6-th order compact finite differences scheme [18] while the space marching algorithm is performed using 1st order
Discontinuous Galerkin method, which has the advantage to be unconditionally stable to solve this transport equation
[19] while remaining affordable in term of computational cost.

B. Transverse boundary conditions

The duct has hard walls at the top and the bottom, but a discontinuity of material appears at the bottom with an
acoustic liner that presents an impedance condition. This discontinuity is the origin of the reflection and refraction
phenomena appearing in these simulations.

1. Hard wall
Hard wall boundary conditions are applied at the top and the bottom of the duct and allow the particles to slip
tangentially to the wall while imposing the normal velocity (𝑣) to 0. Moreover, this slip boundary condition models a
infinitely thin acoustic boundary layer [20] and satisfies the following set of equations:

7
  
 𝜕𝑝 𝜕𝑝 2 𝜕𝑢 𝜕𝑣 𝜕𝑝


 + 𝑢¯ + 𝜌𝑎
¯ + + 𝑆𝑎 =0


 𝜕𝑡 𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦

 𝜕𝑢
 𝜕𝑢 1 𝜕 𝑝
+ 𝑢¯ + =0 , (28)


 𝜕𝑡 𝜕𝑥 𝜌¯ 𝜕𝑥


 𝜕𝑣 𝜕𝑣

 + 𝑢¯ =0
 𝜕𝑡 𝜕𝑥
with 𝑆 = 1 or 𝑆 = −1 for the upper or lower wall, respectively. Because we discretize our propagation operator in term
of characteristic variables, we have to express this relation in these variables as well. Since this set of equations does not
change anything to the diagonalization of the matrix 𝐴, the transformation matrices 𝑇 and 𝑇 −1 are also the same as
previously. The same can be said about 𝐴 e and the eigenvalues of 𝐴. Therefore, by applying the same variable change
performed for the Euler equations in this case, we obtain a modified propagation operator 𝑀: e

¯ 2
1 𝜌𝑎
¯ 0 0 ©𝑖𝜔 + 𝑆𝑎 2 𝜕 𝑦 − 2 𝜕𝑦
𝑆𝑎
2 𝜕𝑦 ª
© 𝑢+𝑎 ª­
e (𝑥, 𝑦) = ­ 0 1 (29)
0 0 0
®
𝑀 ­ 𝑢¯
®­
®­ 𝑖𝜔 ® .
1 ¯
𝜌𝑎 2 ®
« 0 0 𝑢−𝑎 𝑆𝑎 𝑆𝑎
¯ ¬ « 2 𝜕𝑦 − 2 𝜕𝑦 𝑖𝜔 + 2 𝜕𝑦 ¬
This new propagation operator is valid and applied only for the 𝜙 variables that are modeling the behavior of the
hard wall boundary condition.

2. Liner
The presence of the acoustic liner is modeled by an impedance condition that couples the pressure fluctuation
with the normal velocity fluctuation since there exists some flow through the liner, contrary to the hard wall boundary
condition. This impedance condition is applied following the classical relation:

𝑝e = 𝑍 (𝜔)e
u·n, (30)
with n the unitary outward pointing normal at the wall and 𝑍 (𝜔) the frequency dependent complex impedance which
means, in our case, that its value has to be modified according to the value of 𝜔 set. For Configuration A, the value of 𝑍
respects the extended Helmholtz resonator relation:
 𝜖
𝑍 (𝜔) = 𝑅 − 𝑖𝑚𝜔 + 𝑖Φ cot 𝜔ℎ + 𝑖 , (31)
2
with 𝑅 the face-sheet resistance, 𝑚𝜔 the face-sheet reactance, Φ a porosity parameter, ℎ the cavity depth and 𝜖 the
damping in the cavity’s fluid. Each of these parameters is imposed by the type of liner used [21] which is here a CT57.
Therefore, for our particular computation, a value of 𝑍 = 5.21 − 1.07𝑖 is chosen at a frequency of 𝑓 = 2000 Hz.
For Configuration B, the impedance values are obtained through a nonlinear least-square fit of the experimental data
and the different values used between 500 and 3000 Hz are available in [22].
Since the discretization of our system is made in term of characteristic variables, we have to express Equation (30)
in this variables, which gives us:

𝜙1 + 𝜙3 = 𝑍 (𝜔)𝜙2 . (32)
This equation is imposed on the second equation which transports 𝜙2 while the other two equations remain
unchanged.

IV. Domain decomposition and transmission of waves

The presence of the liner at the bottom of the duct creates a discontinuity for the One-Way method since it will use a
spatial marching algorithm to solve the equations along 𝑥. Therefore, an incident wave propagating through such a
discrepancy will produce a transmitted wave, propagating in the same direction as before with modified wavelength and
amplitude, and a reflected one, propagating in the opposite direction. Splitting the computational domain into 3 different
subdomains (in our case) that are completely homogeneous in 𝑥 independently leads us to compute the reflected and
transmitted part of an incident wave only at the interface of these subdomains (Sec. IV.B).

8
 However, to take into account these phenomena, solving the One-Way equations in only one direction is not sufficient
to obtain an accurate result since a non negligible part of the waves can propagate in the opposite direction. Hence it
is required to perform a rightgoing One-Way (right OW) to capture the incident wave and its transmission but also a
leftgoing one (left OW) to take into account the reflected waves (Sec. IV.A).

A. Domain decomposition and convergence

In our particular case, the computational domain Ω will be split in three different subdomains Ω = {Ω1 , Ω2 , Ω3 }.
Each of these subdomains is non-overlapping with another but has one or two interfaces positioned exactly at the liner
leading and/or trailing edge. The reflected and transmitted waves will be communicated from a subdomain to another
through these interfaces and their computation will be developed in the next section (Sec. IV.B).

Fig. 2 Decomposition of the computational domain

The layout of Fig 2 is identical for all the simulations performed in this paper, since their configurations are nearly
the same. In each subdomain, a right and left One-Way are computed to capture the rightgoing and leftgoing waves,
respectively. Once this first approximation of the results is obtained, the reflection and refraction of waves is computed
at each interface before updating the initial condition of each One-Way (left and right). A new iteration is then computed,
giving a new result that is closer to the final result since it takes into account the reflected and transmitted waves of the
previous approximation. These iterations can be carried on until a satisfying convergence is reached.

·10−2

4
𝐿 2 norm

1
0 2 4 6 8 10 12 14 16 18 20
Iteration

Fig. 3 𝐿 2 norm of the left and right One-Way in each domain depending on the iteration number. is for the
right One-Way, is for left One-Way. are for Ω1 , are for Ω2 and are for Ω3

We can see in Fig. 3 the 𝐿 2 norm of each iteration of the left and right One-Way in each subdomain for the case at
𝑓 = 500 Hz in Configuration B (Sect. III.A). This case has been initialized with a local mode of a hard-walled duct

9
(plane wave) to obtain a pure rightgoing wave at the inlet. At iteration 0, only the right OW in Ω1 is visible since it is
the only one that has a non-null initialization vector (incident rightgoing wave). In the same way, it is expected for
its norm to stay the same throughout the iterations since this initialization is not modified. The same analysis can be
made about the left OW inside Ω3 because of its initialization with a null vector meaning that no leftgoing wave goes
inside our computational domain from the outlet. The norm of the rightgoing OW in Ω2 and leftgoing OW in Ω1 first
appears at iteration 1 due to the fact that the incident wave at the inlet has to travel through Ω1 at iteration 0 before being
transmitted and reflected. Similarly, the norm of the rightgoing OW in Ω3 and the leftgoing OW in Ω2 take a different
value from zero at iteration 2 for the exact same reason. Finally, it can be noticed that the norm of each One-Way in each
domain is stabilized to its final value after 4 or 5 iterations, which are necessary for the reflected waves to travel in all
subdomains, meaning that our transmission condition is a good approximation of the exact operator.

B. Transmission of waves
Before communicating the reflected and transmitted waves to other subdomains, we first need to compute which part
of the incident wave leaving the current subdomain is reflected back in it and which part actually leaves it before being
transmitted to the next domain. To ensure the continuity of our final result even through a discontinuity, we want to
impose the following relation at the interface:

q1 = q2 , (33)
 𝑡  𝑡
with q1 = 𝑝 1 𝑢 1 𝑣 1 the discretized variables in Ω1 at the interface and q2 = 𝑝 2 𝑢 2 𝑣 2 the same variables at
the interface but in Ω2 . For the rest of this development, the under-script 𝑖 means that the variable belongs to Ω𝑖 at the
interface with Ω𝑖+1 or Ω𝑖−1 . The example of the transmission is taken between Ω1 and Ω2 here but it is also relevant
for any other domain having a transmission interface. This condition can present some convergence issues for wave
transmission problem and more complex conditions are usually employed for this kind of concern [23]. However, the
fact that, thanks to the numerical diagonalization of the propagation operator, we are able to decompose a wave into
its rightgoing and leftgoing components ensure us a well-posed condition. To decompose these waves and make our
transmission problem accurate, we need to express Equation (33) in term of the One-Way variables 𝝍 1 and 𝝍 2 . To do
that, we use first the characteristic variables:

T1 𝝓1 = T2 𝝓2 . (34)
Since, in our case, the switch in boundary conditions does not lead to a discontinuity in T𝑖 (i.e. T1 = T2 ), we can
simplify Equation 34 into:

𝝓1 = 𝝓2 . (35)
This leaves us, in term of One-Way variables:

V1 𝝍 1 = V2 𝝍 2 . (36)
By decomposing V𝑖 in its + and − components, we obtain:
(
V++ + +− − ++ + +− −
1 𝝍 1 + V1 𝝍 1 = V2 𝝍 2 + V2 𝝍 2
−− − . (37)
V−+ + −− − −+ +
1 𝝍 1 + V1 𝝍 1 = V2 𝝍 2 + V2 𝝍 2

We now have an expression of our original relation in Equation (33) in term of leftgoing and rightgoing components.
This decomposition allows us to impose only rightgoing information as the transmitted wave and leftgoing information
as the reflected wave. The last step is to express directly the wanted terms. In Fig. 4, we can see a diagram of the
situation we are in for this refraction/reflection between Ω1 and Ω2 . After performing a right One-Way in Ω1 , we get an
incident rightgoing wave 𝝍 +1 while the left One-Way in domain Ω2 gives us 𝝍 −2 . These variables are approximations
of the final incident waves but are enough to make 𝝍 +1 and 𝝍 −2 known values leaving us with two equations for two
unknowns.
We can easily express the missing values 𝝍 +2 and 𝝍 −1 from 𝝍 +1 and 𝝍 −2 , which gives:

10
 Fig. 4 Transmission of wave through a discontinuity

  −1     
 𝝍 +2 = V++ +− −−
−1 −+ ++ +− −−
−1 −+ + +− +− −−
−1 −−
𝝍 −2


 2 − V 1 V 1 V 2 V1 − V1 V 1 V 1 𝝍 1 − V2 − V1 V 1 V 2
. (38)
 𝝍 − = V−−
−1 V−+ 𝝍 + + V−− 𝝍 − − V−+ 𝝍 +


 1 1 2 2 2 2 1 1
Then, we can set the newly obtained values as initialization for the left One-Way in Ω1 and for the right One-Way in
Ω2 . By applying this relation at each iteration and at each interface, we are able to compute more and more accurate
approximations of the reflected and transmitted waves, converging towards the exact result.

V. Numerical results
The numerical results presented here have been computed on the two different configurations presented above (Sec.
III.A). To compare the obtained results with the One-Way with experimental data and other numerical results, we have
to introduce the sound pressure level (SPL):
 
| 𝑝|
𝑆𝑃𝐿 = 20 log , (39)
𝑝 ref
with 𝑝 ref = 2 × 10−5 𝑃𝑎 the reference sound pressure. The SPL is expressed in decibels (dB).

132

130
SPL (dB)

128

126

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

x (m)

Fig. 5 Sound Pressure Level along the upper wall for Configuration A. Comparison between experimental data
( ) from [15], time domain simulation ( ) from [16] and the One-Way simulations with domain decomposition
( )

In both cases, the One-Way simulations have been initialized by a plane wave mode that has been scaled to correspond
to the same SPL as the numerical data. These plane waves have been obtained through a Local Stability Analysis

11
performed at the inlet of the computational domain and by selecting the eigenvectors corresponding to the plane wave
mode of the operator spectrum.

f = 3.0 kHz
130

120

110

100

90

130 f = 2.5 kHz

120
SPL (dB)

130 f = 2.0 kHz

130 f = 1.5 kHz

120

130 f = 1.0 kHz

120

110

100

90

80

130 f = 0.5 kHz

120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
x (m)

Fig. 6 Sound Pressure Level along the upper wall for Configuration B. Comparison between experimental
data ( ), multifrequency ( ) and single frequency ( ) from [17] and the One-Way simulations with
domain decomposition ( )

12
Configuration A For the first configuration, we will refer to numerical results that can be found in [16] and
experimental data obtained in [15]. The numerical simulation has been carried out globally and in the time domain,
contrary to our local solution computed in the frequency domain.
We can see in Fig. 5 the SPL obtained at the upper wall all along the computational domain. It is clear that the
One-Way results are quite close to the other numerical results. The fact that these SPL are really close to each other
means that the reflection and refraction of wave at the discontinuity (i.e. at the leading and trailing edge of the liner) are
accurately captured and propagated since the global computation takes naturally into account these phenomena. In term
of computational cost, this One-Way simulation was performed on only two 2.60 GHz Xeon processors with a total
CPU time of 90 seconds while only 4 iterations of One-Way were needed for the domain decomposition to converge.

Configuration B The second configuration has for particularity to be set at 𝑀max = 0.3, which is slower than in the
previous case where 𝑀max = 0.5. Moreover, the One-Way simulations have been carried out at different frequencies,
from 𝑓 = 500 Hz to 𝑓 = 3000 Hz, every 500 Hz to explore the validity domain of such method in this type of application.
All the numerical and experimental comparison data are taken from [17]. Once again, these results have been obtained
through a global time domain computation with a Fast Fourier Transform (FFT) applied to the solution in order to get
the SPLs at the upper wall.
In Fig. 6, the SPL along the upper wall is shown for every frequency studied here. It is directly visible that the
One-Way results are quite close to either the multifrequency source simulation or the single frequency source one in
most of the cases. In particular, at 𝑓 = 1500 Hz, where a deviation between the numerical result and the experimental
data was noticed [17] for an unknown reason, the same deviation with the One-Way seems to appear. On the contrary,
this very same difference at 𝑓 = 3000 Hz is less important in the case of our OW simulation. The same analysis can be
made for Configuration A and B about the reflected and refracted waves that are well captured since both the rightgoing
waves at the outlet and the leftgoing waves at the inlet gives an accurate result that is in accordance with the global
simulation. The CPU time required for these simulations is the same as for the previous configuration with 90 seconds of
total CPU time. However, the case at 𝑓 = 1000 Hz required the implementation of a compact 4th order finite differences
scheme in the transverse direction for a non uniform mesh [18], since oscillations in the SPL appeared after the liner
with a constant step grid. Therefore, this discretization scheme allows a more refined mesh at the hard wall and liner
boundary conditions, making these oscillations disappear with only 100 points in 𝑦 and 600 seconds of total CPU time.
The reason behind the difference of results for the multi and single frequency source simulations might be the fact that
with an incident wave presenting a SPL of around 130 dB, some areas inside the simulations might have a higher SPL
that could lead to a nonlinear behavior of the computed waves.

130

120

110
SPL (dB)

100

90

80

70
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
x (m)

Fig. 7 Sound Pressure Level along the upper wall for Configuration B at 𝑓 = 3000 Hz. Decomposition of the
final result ( ) into its rightgoing ( ) and its leftgoing ( ) components

13
 One of the particular feature of the One-Way method is that it is possible to decompose our result into its rightgoing
and leftgoing contributions. In the same way as previously, Fig. 7 shows the SPL at the upper wall for the case
𝑓 = 3000 Hz along with the SPL of the leftgoing and rightgoing components. We can see that the reflected waves,
with a maximum SPL of 120 dB are not negligible compared to the maximum SPL of the rightgoing waves that is of
133 dB, even if we speak here of a logarithmic scale. Moreover, the oscillations present on the liner and before it are
entirely the results of these reflected waves since they can not be observed on the contribution propagating to the right.
In the end, the SPL of the wave propagating downstream is of 103 dB which is lower than the SPL of the outgoing wave
propagating upstream with a value of 104 dB. It can be noticed that the SPL after the liner is null for the leftgoing waves,
as expected, since nothing is imposed at the outlet and no reflection can occur after the second discontinuity.

Fig. 8 Decomposition of the Sound Pressure Level into the rightoing and leftgoing components and total results
for Configuration B at 𝑓 = 3000 Hz

The SPL on the whole computational domain is presented in Fig. 8. As expected, the SPL of the rightgoing part is
homogeneous at a fixed value before the liner before decreasing, with the lowest values at the bottom (where the liner is)
before stabilizing after the liner. For the leftgoing part, the SPL is null on the right side of the liner before presenting
some strong variations at the bottom of the duct and decreasing in value due to the effect of the acoustic liner. These
waves come only from the reflection of the rightgoing part at the liner trailing edge. Between the inlet and the beginning
of the liner, the SPL values are bigger due to the fact that the rightgoing part is at around 130 dB which is more than the
104 dB, on average, present for the second discontinuity. As in Fig. 7, we can see the drop in SPL on the second plot on
the upper wall, along with some other in the middle and the bottom of the duct. Discontinuities can be seen in both the
downstream and upstream propagating components since only the summation of them guarantees a smooth transition.

VI. Conclusion
We presented a diagonalization method based on a non-reflective boundary condition that can be applied to a generic
hyperbolic system since this decomposition of the propagation operator is fully numerical and does not require heavy
analytical developments. In particular, we applied this method to the linearized Euler equations to compute a case of
partially lined duct presenting a shear flow. One of the particularities of this kind of configuration is the presence of
discontinuities in the boundary conditions along the propagation axis and each of them has been processed to avoid any
ill-posedness.
This discontinuity issue has been handled by the projection of waves, leading to their decomposition in rightgoing

14
and leftgoing components, which is now possible thanks to the numerical diagonalization of the operator. This feature
has been used here to compute the reflected and refracted part of a wave across a discontinuity but it can be used in a
more wider range of application, such as domain decomposition in heterogeneous media or preconditioning.
Finally, some numerical results were presented with comparison to existing numerical and experimental data. It
seems to give good results, in accordance to the previous existing numerical data for a wide range of velocity and
frequency and for a low computational cost. This type of method also allows a decomposition of the final results into
rightgoing and leftgoing components, giving more insights on the phenomena appearing in this kind of configurations.

Acknowledgments

This work was partly supported by the french "Programme d’Investissements d’avenir" ANR-17-EURE-0005
conducted by ANR

References
[1] Colonius, T., and Lele, S., “Computational Aeroacoustics: Progress on Nonlinear Problems of Sound Generation,” Progress in
Aerospace Sciences, Vol. 40, 2004, pp. 345–416.

[2] Tam, C., Computational Aeroacoustics: A Wave Number Approach, Cambridge University Press, 2012.

[3] Richter, C., Hay, J. A., Panek, ł., Schönwald, N., Busse, S., and Thiele, F., “A Review of Time-Domain Impedance Modelling and
Applications,” Journal of Sound and Vibration, Vol. 330, No. 16, 2011, pp. 3859–3873. https://doi.org/10.1016/j.jsv.2011.04.013.

[4] Brambley, E. J., “Fundamental problems with the model of uniform flow over acoustic linings,” Journal of Sound and Vibration,
Vol. 322, No. 4, 2009, pp. 1026–1037. https://doi.org/10.1016/j.jsv.2008.11.021.

[5] Towne, A., and Colonius, T., “One-Way Spatial Integration of Hyperbolic Equations,” Journal of Computational Physics, Vol.
300, 2015, pp. 844–861. https://doi.org/10.1016/j.jcp.2015.08.015.

[6] Towne, A., and Colonius, T., “Improved Parabolization of the Euler Equations,” 19th AIAA/CEAS Aeroacoustics Conference,
AIAA Paper 2013-2171, American Institute of Aeronautics and Astronautics, 05/27-29/2013. https://doi.org/10.2514/6.2013-
2171.

[7] Towne, A. S., “Advancements in Jet Turbulence and Noise Modeling: Accurate One-Way Solutions and Empirical Evaluation
of the Nonlinear Forcing of Wavepackets,” Ph.D. thesis, California Institute of Technology, 2016.

[8] Halpern, L., and Trefethen, L. N., “Wide-angle One-way Wave Equations,” The Journal of the Acoustical Society of America,
Vol. 84, No. 4, 1988, pp. 1397–1404. https://doi.org/10.1121/1.396586.

[9] Angus, D. A., “The One-Way Wave Equation: A Full-Waveform Tool for Modeling Seismic Body Wave Phenomena,” Surveys
in Geophysics, Vol. 35, No. 2, 2014, pp. 359–393. https://doi.org/10.1007/s10712-013-9250-2.

[10] Majda, A., and Osher, S., “Initial-boundary value problems for hyperbolic equations with uniformly characteristic boundary,”
Communications on Pure and Applied Mathematics, Vol. 28, No. 5, 1975, pp. 607–675. https://doi.org/10.1002/cpa.3160280504.

[11] Halpern, L., and Trefethen, L. N., “Wide-angle one-way wave equations,” The Journal of the Acoustical Society of America,
Vol. 84, No. 4, 1988, pp. 1397–1404. https://doi.org/10.1121/1.396586, URL https://asa.scitation.org/doi/10.1121/1.396586,
publisher: Acoustical Society of America.

[12] Towne, A., and Colonius, T., “Continued Development of the One-Way Euler Equations: Application to Jets,” 20th AIAA/CEAS
Aeroacoustics Conference, AIAA Paper 2014-2903, American Institute of Aeronautics and Astronautics, 06/16-20/2014.
https://doi.org/10.2514/6.2014-2903.

[13] Higdon, R. L., “Numerical Absorbing Boundary Conditions for the Wave Equation,” Mathematics of Computation, Vol. 49, No.
179, 1987, pp. 65–90. https://doi.org/10.2307/2008250.

[14] Hagstrom, T., and Warburton, T., “A New Auxiliary Variable Formulation of High-Order Local Radiation Boundary Conditions:
Corner Compatibility Conditions and Extensions to First-Order Systems,” Wave Motion, Vol. 39, No. 4, 2004, pp. 327–338.
https://doi.org/10.1016/j.wavemoti.2003.12.007.

15
[15] Jones, M., Watson, W., and Parrott, T., “Benchmark Data for Evaluation of Aeroacoustic Propagation Codes with Grazing Flow,”
11th AIAA/CEAS Aeroacoustics Conference, Aeroacoustics Conferences, American Institute of Aeronautics and Astronautics,
2005. https://doi.org/10.2514/6.2005-2853.

[16] Pascal, L., Piot, E., and Casalis, G., “A New Implementation of the Extended Helmholtz Resonator Acoustic Liner
Impedance Model in Time Domain CAA,” Journal of Computational Acoustics, Vol. 24, No. 01, 2015, p. 1550015.
https://doi.org/10.1142/S0218396X15500150.

[17] Özyörük, Y., and Long, L. N., “Time-Domain Calculation of Sound Propagation in Lined Ducts with Sheared Flows,” AIAA
Journal, Vol. 38, No. 5, 2000, pp. 768–773. https://doi.org/10.2514/2.1056.

[18] Gamet, L., Ducros, F., Nicoud, F., and Poinsot, T., “Compact Finite Difference Schemes on Non-Uniform Meshes. Application
to Direct Numerical Simulations of Compressible Flows,” International Journal for Numerical Methods in Fluids, Vol. 29,
No. 2, 1999, pp. 159–191. https://doi.org/10.1002/(SICI)1097-0363(19990130)29:2<159::AID-FLD781>3.0.CO;2-9.

[19] Pietro, D. A. D., and Ern, A., Mathematical Aspects of Discontinuous Galerkin Methods, Mathématiques et Applications,
Springer-Verlag, Berlin Heidelberg, 2012. https://doi.org/10.1007/978-3-642-22980-0.

[20] Nayfeh, A. H., Kaiser, J. E., and Telionis, D. P., “Acoustics of Aircraft Engine-Duct Systems,” AIAA Journal, Vol. 13, No. 2,
1975, pp. 130–153. https://doi.org/10.2514/3.49654.

[21] Richter, C., Thiele, F. H., Li, X. D., and Zhuang, M., “Comparison of Time-Domain Impedance Boundary Conditions for Lined
Duct Flows,” AIAA Journal, Vol. 45, No. 6, 2007, pp. 1333–1345. https://doi.org/10.2514/1.24945.

[22] Özyörük, Y., Long, L. N., and Jones, M. G., “Time-Domain Numerical Simulation of a Flow-Impedance Tube,” Journal of
Computational Physics, Vol. 146, No. 1, 1998, pp. 29–57. https://doi.org/10.1006/jcph.1998.5919.

[23] Collino, F., Ghanemi, S., and Joly, P., “Domain Decomposition Method for Harmonic Wave Propagation : A General
Presentation,” Research Report RR-3473, INRIA, 1998.

16