Journal of Computational Physics 231 (2012) 704–716

Contents lists available at SciVerse ScienceDirect

Journal of Computational Physics

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

Analysis and optimization of numerical sponge layers

as a nonreflective boundary treatment
Ali Mani ⇑
Center for Turbulence Research, Stanford University, Stanford, CA 94305, United States

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

Article history: The aim of this work is to provide practical guidelines for designing sponge layers consid-
Received 15 November 2010 ering applications in computational fluid dynamics and computational aeroacoustics. We
Received in revised form 27 September 2011 present the analysis of sponge/flow interactions and provide a characterization of its basic
Accepted 12 October 2011
reflectivity mechanisms. While sponges are perfect absorbers in one-dimensional systems,
Available online 19 October 2011
they can cause their own reflection when encountering oblique sound or oblique vorticity
waves. To minimize this adverse effect, sponge strength and profile need to be selected
Keywords:
optimally. Also for a fixed desired accuracy, sponge length should be above a minimum
Nonreflective boundary condition
Compressible flow
threshold. Our analysis quantifies these requirements for a wide range of conditions in
Sponge/flow interactions terms of inflow/outflow Mach number, incident frequencies, incident angles, and desired
Sponge design accuracy, and covers main concerns with sponges such as sound/sponge interactions and
vortex/sponge interactions. As a test case, we present a nonlinear Euler calculation of a
convecting vortex interacting with sponges with different lengths. We show that sponges
designed by our guidelines achieve accuracies comparable to perfectly matched layers for
the same cost, over moderate to high accuracy demands. The results of the presented anal-
ysis can be used to determine sponge requirements for a wide range of CFD applications. A
summary of these guidelines are listed in the paper.
Ó 2011 Elsevier Inc. All rights reserved.

1. Introduction

External flow simulations are typically computed on truncated domains. In these computations the exact conditions at
the truncation boundary are usually unknown, and the external boundaries are often treated artificially. The aim of these
artificial treatments is to allow for out-going flow features to leave the computational domain without reflecting any
signature inside. These methods involve a variety of techniques [1] including characteristic-based decomposition [2–4],
flow dissipation [5,6], grid-stretching/slow-down operators [7,8], supergrid modeling [9], and perfectly matched layers
[10,11]. While some of these treatments require sophisticated procedures, they still fail to accomplish perfect
nonreflectivity.
One simple approach to treat the external boundaries is to use the sponge terms [5,12]. With this method the compress-
ible Navier–Stokes equations are artificially modified as follows:

⇑ Tel.: +1 650 725 1325.

E-mail address: [email protected]

0021-9991/$ - see front matter Ó 2011 Elsevier Inc. All rights reserved.
doi:10.1016/j.jcp.2011.10.017
 A. Mani / Journal of Computational Physics 231 (2012) 704–716 705

@q @
þ ðquj Þ ¼ rðqref  qÞ;
@t @xj
@ qui @    
þ qui uj þ pdij  sij ¼ r ðqui Þref  qui ; ð1Þ
@t @xj
@E @  
þ ðE þ pÞuj þ qj  uk skj ¼ rðEref  EÞ;
@t @xj
where q, ui, p, E, sij, and qj are the density, velocity, pressure, total energy, viscous stress tensor and heat flux, respectively.
The unphysical terms on the right-hand side are active only near external boundaries where they damp the flow variables to
a known reference solution. The damping coefficient, r, which is an inverse time scale, is the same for all conserved variables
due to stability considerations [12]. One can also show that this form maintains Galilean invariance.
The sponge treatment has been widely adopted owing to its simplicity, robustness, non-stiff nature, and flexibility to han-
dle complex geometries and unstructured grids. Examples of computations with sponge treatment include compressible
mixing layers [13,14], jets [15,16], cavity flows [17,18], airfoils [19], ramps [20], and bluff bodies [21].
Each sponge region can be characterized by its length, lsp, and strength, which is an integral measure of r (defined below).
As we shall see, larger sponges perform better than small sponges with the same strength in the sense that they damp flow
features more quietly. However, larger sponges demand a larger computational domain and are more expensive. This trade-
off determines an optimal sponge design for each simulation.
Existing mathematical analyses of numerical sponge layers establish adequate confidence regarding their usefulness and
concerns such as well-posedness, proof of convergence, and stability [5,12]. However, they do not provide intuition into opti-
mal sponge design considering general purpose CFD applications. Neither does there exist an adequate understanding of
sponge failure; the existing reports only attempt to solve the one-dimensional problem, which as we shall see is irrelevant
in regards with sponge failure. Many previous multi dimensional simulations relied on a combination of trial-and-error
to achieve satisfactory sponge settings, yet cost-optimal sponges have rarely been achieved in practice.
In this paper we elucidate the basic ‘‘physics’’ of sponge/flow interactions beyond the one-dimensional model and provide
a quantitative understanding of its different reflectivity mechanisms. By comparing basic power law profiles we identify the
optimal sponge profile, which we will investigate over a wide range of conditions. We provide results aiming to determine
the correct sponge length and strength for an arbitrarily set performance requirement. Our results cover basic concerns with
sponges such as reflectivity of sound/sponge interactions and vortex/sponge interactions and are presented in terms of
dimensionless parameters which can be estimated a priori. We then present an example of nonlinear Euler calculation of
vortex/sponge interactions and show that a sponge designed by our guidelines achieves a performance comparable with that
by a perfectly matched layer for the same computational cost over moderate sponge lengths.
Many researchers combined sponge terms with grid-stretching and viscous (or higher-order) dissipation [13,22]. Equiv-
alently, coordinate transformation may be used instead of stretching [23]. Grid-stretching allows for larger sponge without
additional computational cost, but needs to be accompanied by high-order dissipation to prevent the reflection of unsup-
ported features. Even though we do not include viscous dissipation in our analysis, the presented results are still useful, since
the sponge requirements are often dictated by the low wavenumber flow content, which is least affected by the high-order
dissipation terms.

2. Basic physics of sponge/flow interactions

We present our analysis using the linearized Euler equations which enables capturing the leading order mechanisms of
sponge reflectivity. The linearization is justified since sponges are typically employed in the ‘‘far field’’ where the amplitude
of incident fluctuations is expected to be small.

@u @u @p
þM þ ¼ rðxÞu;
@t @x @x
@v @ v @p
þM þ ¼ rðxÞv ; ð2Þ
@t @x @y
@p @p @u @ v
þM þ þ ¼ rðxÞp;
@t @x @x @y
where u, v, are perturbation velocities in the x- and y-directions, respectively, and p is perturbation in pressure. These per-
turbations are relative to a uniform freestream with Mach number M and are nondimensionalized using free stream speed of
sound, density, and an arbitrary length scale. As shown in Fig. 1(a), the x–y coordinate system is such that the sponge
isocontours are aligned with the y-direction. The background flow is considered to have only an x-component since the
y-component can be eliminated by transforming to a moving frame of reference. Flow features approach the sponge from
x < 0 and interact with it through 0 6 x 6 lsp.
This sponge treatment maintains decoupling of entropy fluctuations from sound and vorticity modes. One can verify that
sponge performs ‘‘ideally’’ on the entropy modes; i.e., entropy fluctuations attenuate exponentially through the sponge as
706 A. Mani / Journal of Computational Physics 231 (2012) 704–716

(a) (b) (c)

ave
vorticity w
incident sound wave

wave
ne
sponge zo

sponge
und
σ=σ (x)

reflected
d s o
cte
y

refle
incident
x

Fig. 1. Schematic of an oblique incident wave approaching a sponge zone (a). An oblique sound wave moving towards a sponge near an inflow boundary
with M = 0.3 (b), and the resulting reflected vorticity wave and sound wave (c) obtained from a nonlinear Euler calculation. The figure have been rotated
slightly in (b) and (c).

they advect, regardless of their scale and alignment. However, the sponge term can cause coupling between incoming and
outgoing acoustic modes as well as coupling between acoustics and vorticity modes. As a demonstrative example, Fig. 1(b)
and c show reflection of a sound wave as well as a vorticity wave due to interactions of an incident sound wave with a
sponge. Before presenting the full analysis of Eq. (2), we present analysis of two simplified cases to gain insight into the basic
physics of sponge/flow interactions.

2.1. One-dimensional limit

In this case Eq. (2) reduces to the following system:

@R @R
þ ðM þ 1Þ ¼ rðxÞR;
@t @x
@v @v
þM ¼ rðxÞv ; ð3Þ
@t @x
@L @L
þ ðM  1Þ ¼ rðxÞL;
@t @x
where R = u + p and L = u  p. One can see that all characteristics remain uncoupled. In other words, sponge does not cause
any reflection itself; it only damps exponentially the reflections that could possibly come off of the end boundary. This can be
simply seen by replacing the left hand side operators in (3) with the time derivatives along the characteristic lines. We would
like to emphasize this important physical observation (empowered by the method of characteristics), since the earlier 1D
analysis [12] (based on incidence/reflection decomposition) fails to distinguish the true source of reflectivity.
If the boundary condition after the sponge at x = lsp remains decoupled (such as the widely used 1D characteristic-based
boundary condition), the sponge plus boundary condition would be perfectly nonreflective. Otherwise, there will be reflec-
tions, but the sponge will exponentially suppress the effects. In this case three general scenarios or their combination can
happen:

1. sound–sound reflectivity: an incident sound wave can reflect in the form of a sound wave,
2. sound–vorticity reflectivity: an incident sound wave can reflect in the form of a vorticity wave for an inflow boundary
with M < 0, and
3. vorticity–sound reflectivity: an incident vorticity wave can reflect in the form of sound wave for an outflow boundary
with M > 0.1

For the sound–sound reflectivity, one can show that the ratio of the reflected to incident sound amplitude is
Z !
lsp
jL j 2
g  reflected ¼ gBC exp  rðxÞdx ; ð4Þ
jRincident j 1  M2 0

where gBC is the (sound–sound) reflection coefficient of the boundary condition at x = lsp. Similar expressions can be obtained
for other mechanisms.2 The exponential term in (4) is the sole contribution of the sponge to attenuate the reflection. This term

1
We note that in the most general form the boundary condition at x = lsp could also provide entropy coupling. However, this will not be the case for the
boundary conditions that we study here.
2
Only the factor of 2/(1  M2) will be replaced by the sum of inverse incident and reflected characteristic velocities.
 A. Mani / Journal of Computational Physics 231 (2012) 704–716 707

can be interpreted as a two-passage experience through the sponge with velocities 1 + M and 1  M, and residence times
inversely proportional to the velocities. In this ideal setting sponge provides a targeted exponential damping. We define sponge
target-damping as
" Z !# Z
lsp lsp
2 2 log e
gtarget  20 log exp  2
rðxÞdx ¼ 20 2
rðxÞdx: ð5Þ
1M 0 1M 0

With a good end-boundary-condition (gBC  1) the overall damping, g, will be even better than the target-damping for 1D
systems. In practical situations however, the flow features are often misaligned and the idealized 1D model would be invalid.
In such cases the sponge contributes its own reflection, and the overall damping, g, may be better or worse than gtarget,
R
depending on the degree of sponge failure and the end boundary condition. One can see that the integral, rdx, is a relevant
measure of sponge strength. In this paper we refer to gtarget (i.e., the same integral but properly normalized) as the sponge
strength. As an example, a sponge with 40 dB strength would damp the amplitude of an incident sound wave by a factor of
100 under the one-dimensional condition.

2.2. Oblique waves with no convection

To gain insight into the performance of sponges when encountering oblique waves, we first analyze a simple case in
which M = 0. Eq. (2) can be transformed to an ODE after taking Fourier transforms in the homogeneous directions.
" # 2 3" #
k2
d u^ 0 1 þ ðixyrÞ2 ^
u
¼ ðix  rÞ4 5 ; ð6Þ
dx p
^ ^
p
1 0
where x = 2pf is the angular frequency and ky = 2p sin h/k is the transverse wavenumber with h being the angle of wavefront-
normal measured relative to the x-direction (h and k defined outside of the sponge zone). As mentioned above, a simple
choice for the boundary condition at x = lsp is the 1D characteristics-based (Thompson) condition:

p ^ ðlsp Þ ¼ 0:
^ðlsp Þ  u ð7Þ

We first consider the case of constant r. Under this condition Eq. (6) can be diagonalized to the following system:
" # " #" #
d R c 0 R
¼ ðix  rÞ ; 0 < x < lsp ; ð8Þ
dx L 0 c L
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
where c ¼ 1 þ ky =ðix  rÞ2 . R and L are the right-going and the left-going characteristics, respectively. The eigenvectors
associated with these characteristics can be obtained from the following transformation:
    
^
u c c R
¼ : ð9Þ
^
p 1 1 L

The key observation is that for ky – 0 both eigenvectors are dependent on r; hence they would change across any interface
with a jump in r. As a right-going incident characteristic enters the sponge it can not remain a pure right-going character-
^ and p
istic. This would cause reflection of a portion of an incident wave off the sponge to maintain continuity (for u ^) at the
interface. To see this in the analysis we present the solution to the ODE system of (8):
" # " #"   #" #" #
^ ðlsp Þ
u c c exp cðr  ixÞlsp 0 1
2c
1
2
^ ð0Þ
u
¼   : ð10Þ
^ðlsp Þ
p 1 1 0 exp cðr  ixÞlsp 1
 2c 1
2
^ð0Þ
p
^ ðlsp Þ ¼ p
With the boundary condition u ^ðlsp Þ, one obtains
   
^ ð0Þ cosh cðr  ixÞlsp þ c sinh cðr  ixÞlsp
u
¼   1  : ð11Þ
^ð0Þ cosh cðr  ixÞlsp þ c sinh cðr  ixÞlsp
p

Eq. (11) is the ‘‘apparent’’ effective boundary condition at x = 0 resulting from the system of sponge plus the end boundary.
To obtain the reflectivity of this overall condition, one can use the inverse of Eq. (9) to translate (11) in terms of the ratio of
reflected to incident characteristics (at x = 0).

^
Lð0 Þ cð0 Þ  up^ð0Þ
ð0Þ

g ¼ ; ð12Þ
Rð0 Þ ^ ð0Þ
cð0 Þ þ up^ð0Þ
708 A. Mani / Journal of Computational Physics 231 (2012) 704–716

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
where cð0 Þ ¼ 1  ky =x2 ¼ cos h. Substitution of (11) into (12) results in a closed-form expression for reflectivity in terms
of incidence angle (h), and the products lspf and lspr (nondimensionalized by speed of sound, a, and independent of the ref-
erence length).

2.3. Physical interpretations

Before extending our analysis to a general condition, we present the physical interpretation of results for the case of M = 0
and r = const. The trends that we discuss here are applicable to the general condition and provide useful insight into the
design of sponge zones. We discuss our result in the limit of small incident angle, h ’ ky/x  1. Substitution of (11) into
(12) yields
!
h2 x2 h2 x2
g¼ 1 þ expð2lsp r þ 2ilsp xÞ : ð13Þ
4 ðx þ irÞ2 4 ðx þ irÞ2

The first term in (13) represents reflection by the sponge interface at x = 0, and the second term represents reflection off the
boundary at x = lsp, which is damped by the sponge similar to the 1D case. While a stronger sponge suppresses the reflection
effects of the end boundary, higher r is not always desired as it causes reflection from the sponge itself. Following is a sum-
mary of important observed trends which are also applicable to our extended analysis presented in the next sections.

1. Scaling with the angle of incidence: For small h the reflection coefficient scales as h2. Therefore, significant improvement
can be achieved by simply aligning the sponge isocontours with the outgoing wavefronts. Fig. 2 shows an example of this
practice.
2. Large wavelength limit: One can verify that in the limit of small x the reflection coefficient is

h2
lim gx!0 ¼ ; ð14Þ
4
which is the same as the reflection coefficient without the sponge. In other words, sponges are ineffective when r is much
larger than the incident frequency.

Fig. 2. Contours of density field computations for flow over cylinder at Re = 10,000 and M = 0.4. The outer boundary and sponge isocontours are circles, but
in (a) they are centered around the cylinder and in (b) they are aligned with the dominant radiating acoustic fronts. The visible spurious sound
contamination is eliminated by sponge alignment and profile optimization.

(a) θ<<1, M=0, σ =const (b) θ<<1, M=0, σ =const

10 10
ηtarget=−20dB ηtarget=−20dB

0 0 η =−40dB
target

−10 ηtarget=−60dB
η /0.25 θ (dB)

−10
−20
2

−20
−30
−30
−40
−40 −50
−50 −60
−2 −1 0 1 2 0 2 4
10 10 10 10 10 10 10 10

Fig. 3. Reflectivity as a function of dimensionless sponge length (a) the upper curve envelop of this profile is shown in (b) and compared for various sponge
strengths.
 A. Mani / Journal of Computational Physics 231 (2012) 704–716 709

(a) (b) 20
(c)
10 20

0 0
0
η /0.25 θ (dB)

−10 −20
−20
2

−20 −40
−40
−30 −60

−40 −60 −80

−50 −80 −100

−1 0 1 −1 0 1 −1 0 1
10 10 10 10 10 10 10 10 10

Fig. 4. Comparison of reflection coefficients for different sponge profiles (h  1, M = 0). r(x) is selected to vary as constant (dotted), linear (dashed),
quadratic (dash-dotted), and cubic (solid) functions. Results are plotted for different sponge strengths indicated by gtarget = 20 dB (a), gtarget = 40 dB (b),
and gtarget = 60 dB (c).

3. Short wavelength limit: In the large x limit, the reflection coefficient becomes

h2
lim gx!1 ¼ expð2lsp rÞ; ð15Þ
4
which is identical to the ideal 1D reflectivity described by Eq. (4), with gBC = h2/4 and M = 0. Achieving this ideal limit (and
simultaneously avoiding the ineffective limit above) requires that x  r. Noting x = 2p/k, this leads to the following
requirement

lsp
 gtarget ; ð16Þ
k
This conclusion is generally applicable to non-constant sponge profiles, but its severity can be significantly relaxed by profile
optimization (see below).
Fig. 3(a) shows reflectivity of a sponge with constant r and gtarget = 20 dB as a function of dimensionless sponge length
(using frequency, f, and speed of sound, a). One can see that for sponge lengths smaller than 30% of the wavelength, the
sponge is ineffective and the reflection coefficient is the same as that without sponge (g ’ 0.25h2). For sponges larger than
10 wavelengths, an almost ideal performance is achievable.
The oscillations in the plot of Fig. 3(a) is due to interference between two reflections at x = 0 and x = lsp represented by two
terms in (13). Fig. 3(b) shows the upper envelope of this curve (i.e., the worst interference scenario) and compares it with
sponges at higher strength. Achieving ideal performance at higher target-damping requires much larger sponges. For
example, a sponge with gtarget = 60 dB needs to be as long as 104 wavelengths to achieve ideal performance. This indicates
that sponges with constant profile are practically ineffective.

2.4. Effect of sponge profile

The stringent requirement on sponge length by (16) can be significantly relaxed if one resorts to non-uniform sponges.
Extension of our analysis to variable r is straightforward. One only needs to discretize the sponge profile and assume con-
stant r in each interval with length Dx (as done in basic integration methods). The ODE system of (6) can be diagonalized for
each interval, and solved analytically resulting in a matrix relation between the two end state-vectors for each interval, sim-
ilar to Eq. (10). Multiplying the series of these matrices for all intervals yields the global solution to the system.
Fig. 4 compares the performance of different sponge profiles and compares them with that of a constant profile. We con-
sider power-law profiles of the form r(x)  xn with n = 0, 1, 2, and 3. The rate of convergence to the ideal performance is
much faster for higher-order sponges than for the constant sponge. This is mainly due to distribution of interfacial reflection
over the entire length of the sponge when r is increased gradually. This treatment reduces the overall reflection in two ways:
first, in deeper regions (higher x) the incident signal is already weakened by the sponge and thus the interfacial reflected
energy will be smaller, and second, there is a possibility of phase cancellation by reflection from a continuous interface.3
Higher-order sponges, however, have a worse low-frequency performance. As shown in the figure, the quadratic sponge
(r  x2) has the best overall performance for gtarget in the range 20 to 60 dB with trends suggesting a cubic sponge for
target-dampings beyond 60 dB. We consider the quadratic sponge as our method of choice and in the following sections
present the general analysis of its performance.

3
A perturbation analysis of reflectivity of a general sponge profile for small h reveals that g(lspf/a) is the Fourier transform of a function involving the sponge
profile. Smoother sponges result in a faster drop of the Fourier transform at high frequencies.
710 A. Mani / Journal of Computational Physics 231 (2012) 704–716

η = −20 η = −40 η = −60

θ
70
60
50
− 40
30
20
10
0

θ 60
50

40
−

30

20

10

θ
35
30
25
−

20
15
10
5
0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10

Fig. 5. Sound–sound reflectivity of the quadratic sponge near an inflow boundary. In each case the color represents the ratio of the reflected to incident
sound wave as a function of nondimensional sponge length and angle of incident wavefront normal. Color represents a range from 0 (light) to 140 dB
(dark). The solid line is the isocontour of g = gtarget.

3. General analysis

Extension of this analysis to the general case with nonzero M (non-constant r, and general h) is straightforward. Eq. (2)
can be Fourier transformed in homogeneous directions and rearranged to
2 3 2 31 2 32 3
u^ M 0 1 ix  rðxÞ 0 0
^
u
d 6 7 6 7 6 76 7
4 v^ 5 ¼ 4 0 M 05 4 0 ix  rðxÞ iky 54 v^ 5: ð17Þ
dx
^
p 1 0 M 0 iky ix  rðxÞ ^
p
The sponge profile can be discretized into constant-r intervals, and system (17) can be diagonalized and analytically solved
in each interval to achieve
2 3 2 3
^ ðx þ DxÞ
u u^ðxÞ
6^ 7   1 6 7
4 v ðx þ DxÞ 5 ¼ SðxÞ exp KðxÞ Dx SðxÞ 4 v^ ðxÞ 5; ð18Þ
^ðx þ DxÞ
p ^ðxÞ
p
where S is the local matrix of eigenvectors and K is the diagonal matrix of local eigenvalues. Multiplying the matrices for all
intervals yields a matrix relation between the state-vector at x = 0 and that at x = lsp. Again, a reasonable choice for the
boundary condition at x = lsp is the 1D characteristic-based condition:

^ ðlsp Þ  p
u ^ðlsp Þ ¼ 0;
v^ ðlsp Þ ¼ 0 if M < 0: ð19Þ
The second condition ensures that the incoming 1D vorticity wave will be zero. For M > 0 this condition would be replaced by
another condition set at x = 0. For analysis of an incident sound, the incident vorticity would be set to zero and vice versa.
 A. Mani / Journal of Computational Physics 231 (2012) 704–716 711

η = −20 η = −40 η = −60

θ 100

80

60

40

20

θ 120
100

80

60

40

20

θ 140
120
100
80
60
40
20
0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10

Fig. 6. Sound–sound reflectivity of the quadratic sponge near an outflow boundary. See the caption of Fig. 5.

We note that in the general case with M – 0, the nondimensional sponge length, lspf/a, is not equal to the ratio of the
sponge length to the wavelength. In the physical domain, x 2 (1, 0), the relations between the spatial wave number
and frequency are x = k(1 + Mcosh) and x = kMcos/ for the sound wave and the vorticity wave, respectively, with h and
/ being the angle of the wavefront-normal, measured relative to the x-direction.
We performed numerical calculations of reflectivity for quadratic sponges in a range of inflow and outflow Mach
numbers (M = 0.8, 0.5, 0.2, 0.2, 0.5, 0.8) and for sponge strengths of 20 dB, 40 dB, and 60 dB. We used 200 uniform
intervals to discretize the sponge profile into a piecewise-constant function and verified satisfactory convergence of
the results. The analysis is done for a range of nondimensional frequencies from 0 to 10 with stepping, lspDf/a, equal
to 0.04.

3.1. Reflectivity of the sound–sound mechanism

Figs. 5 and 6 show the sound–sound reflectivity of the quadratic sponge for inflow and outflow boundaries, respectively.
In the limit of small lsp the sponge is ineffective and the results are the same as those without the sponge. But the sponge
effect picks up at lspf/a of order 0.5 to 2, depending on the Mach number and target-damping.
The explored range of incidence angles are from 0 to cos 1(M); the waves outside of this range are irrelevant since they
do not have positive group velocity in the x-direction. As h approaches cos1(M) the waves do not effectively encounter the
sponge since they move almost tangentially.
The solid line in each plot indicates the isocontour of g = gtarget, and one can see that a wide range of frequencies and inci-
dent angles are damped by ratios better than the target level. The results of these plots can guide the design of sponges for
CFD applications. One first needs to estimate the minimum relevant frequency, fmin, in the system to be simulated (for exam-
ple the shedding frequency for flow over a cylinder). The desired simulation accuracy would set the parameter gtarget which
then for the known M narrows the search to a specific sub-plot in Fig. 5 or Fig. 6. The sponge should be designed long enough
so that lspfmin/a lies on the right side of the plotted isocontour for the desired ranges of the incidence angles. Based on the
results of Figs. 5 and 6, a summary of recommended guidelines is presented in Section 4.
712 A. Mani / Journal of Computational Physics 231 (2012) 704–716

η = −20 η = −40 η = −60

θ
70
60
50
40
30
20
10
0

θ 60
50

40

30

20

10

θ 35

30
25
20
15
10
5
0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10

Fig. 7. Sound–vorticity reflectivity of the quadratic sponge near an inflow boundary. In each plot the color represents the ratio of the reflected vorticity
wave amplitude to incident sound wave amplitude with amplitude defined as max{(u2 + v2)1/2}. Color represents a range from 0 (light) to 140 dB (dark).
The solid lines represent isocontours corresponding to 20 dB, 40 dB, and 60 dB.

3.2. Reflectivity of the sound–vorticity mechanism

For an inflow condition with M < 0, it is possible that a portion of the energy of an incident sound wave reflects back into
the physical domain in terms of a vorticity wave. Fig. 7 shows the results of the reflectivity analysis of the sound–vorticity
mechanism. Again, the contours in the zero incident frequency limit indicate reflectivity without the sponge. As frequency
increases the sponge effect picks up very rapidly and the reflectivity of the sound–vorticity mechanism is much smaller than
that of the sound–sound mechanism over the entire range of parameters. Therefore, this mechanism is typically of lesser
concern when designing sponges.

3.3. Reflectivity of the vorticity–sound mechanism

For an outflow condition with M > 0 it is possible that a portion of the energy of an incident vorticity wave reflects back
into the physical domain in terms of a sound wave. Fig. 8 shows the results of the reflectivity analysis of the vorticity–sound
mechanism.
We note that only the waves with jsin/j < M create propagating acoustic reflection; the /’s outside of this range would
cause reflection in the form of evanescent waves and are excluded from our analysis. Again, sponges perform better at higher
frequencies and lower incidence angle.
In physical scenarios, vorticity is less likely to be encountered in the form of ‘‘waves’’. Instead compact vortices are likely to be
found in the domain. The analysis presented in Fig. 8 needs to be further processed to obtain estimates of reflectivity due to vor-
tex/sponge interactions and hence provide insights for sponge design when compact vortices must leave the domain.

3.4. Vortex/sponge interactions

Consider a single vortex convecting towards an outflow boundary with a sponge. The vortex can be characterized by its
diameter and by the shape of its velocity profile; the amplitude does not affect the reflection coefficient in the linear analysis.
 A. Mani / Journal of Computational Physics 231 (2012) 704–716 713

η = −20 η = −40 η = −60

φ
10

φ 30
25

20

15

10

φ
50

40

30

20

10

0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10

Fig. 8. Vorticity–sound reflectivity of the quadratic sponge near an outflow boundary. In each plot the color represents the ratio of the reflected sound wave
amplitude to incident vorticity wave amplitude with amplitude defined as max{(u2 + v2)1/2}. k is the vorticity wavelength and / is the angle of incident
vorticity wavefront normal. Color represents a range from 0 (light) to 140 dB (dark). The solid lines represent isocontours corresponding to 20 dB,
40 dB, and 60 dB.

While the profile shape affects the quantitative details of the sponge/vortex system, we expect that the trends to be inde-
pendent of the profile and the vortex diameter to be the most important parameter characterizing the behavior. We assume
the following velocity profile for the vortex:
 
r 1 r2
v h ðrÞ ¼ v max exp 1 2 ; ð20Þ
b 2 b
where vmax is the maximum circumferential velocity, and b is the measure of radius of the vortex with vh(b) = vmax. This
velocity field can be decomposed into Fourier vorticity modes. With the analysis that was presented in the previous section,

−2
10
(a) (b) (c)
−3
10
energy ratio

−4
10

−5
10

−6
10

−7
10
0 1 0 1 0 1
10 10 10 10 10 10

Fig. 9. ratio of the total reflected sound to the kinetic energy of an incident vortex as a function of sponge length over vortex diameter. Results are compared
for different sponge strengths: gtarget = 20 dB (solid), 40 dB (dashed), and 60 dB (dash-dotted); and different outflow Mach numbers: M = 0.2 (a),
M = 0.5 (b), and M = 0.8 (c).
714 A. Mani / Journal of Computational Physics 231 (2012) 704–716

one could compute the amplitude of the reflected sound wave for each mode. Combining these reflected modes together pro-
vides a quantitative understanding of the reflected sound field due to an incident vortex.
To make the analysis useful from a practical point of view we assign a simple measure of reflectivity to each vortex/
sponge interaction event. The natural choice would be to report the ratio of the energy of the vortex which has been reflected
as sound.4
Fig. 9 shows results of the reflectivity analysis of vortex/sponge interaction over a range of incident Mach numbers. Each
plot shows the energy portion of an incident vortex reflected into the physical domain as a function of dimensionless sponge-
length. For sponge lengths larger than 0.5–2 vortex diameters (depending on the outflow Mach number) the reflectivity
drops rapidly when increasing the sponge size. A power law fit reveals that reflectivity drops as ln with 3 < n < 4. In the very
large lsp limit the reflectivity asymptotes to a constant number since the sponge reaches its ‘‘ideal’’ limit similar to what pre-
viously observed in Figs. 3 and 4 for the sound–sound mechanism. Stronger sponges provide more damping in the asymp-
totic limit, but have poorer performance in the smaller lsp ranges.

4. Sponge design guidelines

The following presents a summary of our findings:

1. Sponge performs ideally for entropy modes, but creates coupling between sound and vorticity modes.
2. For target-dampings in the range 20–60 dB, the quadratic sponge provides reasonably ‘‘optimal’’ performance (assuming
the use of Thompson boundary condition at the sponge end). For target-dampings beyond 60 dB, cubic and higher-order
polynomials need to be used for the sponge profile.
3. Sponges perform best when the incident features are aligned with the sponge. In the small angle limit, reflectivity is pro-
portional to the square of the incidence angle.
4. For a given problem and fixed sponge length, there exists an optimal sponge strength leading to the lowest reflectivity; a
strong sponge causes its own reflection, whereas a very weak sponge does not prevent reflections by the end boundary
condition.
5. Conversely, for a given problem and a fixed target-damping (translating into fixed desired accuracy), there is a minimum
sponge length requirement. This length is the larger of the two lengths imposed due to incident sound and vorticity (dis-
cussed below).
6. For effective damping of incident sound waves, minimum dimensionless sponge lengths from 0.5 to 2 are recom-
mended. The length is normalized by minimum relevant incident frequency and speed of sound. For higher target-
dampings, longer sponges are recommended. This recommendation covers a wide range of incidence angles, with details
presented in Figs. 5–7.
7. For effective damping of incident vortices, dimensionless sponge lengths above 2 are recommended. This length is nor-
malized by the maximum relevant vortex diameter. Reflectivity of the sponge/vortex mechanism improves rapidly as the
length increases faster than the third power of the sponge length. More details are presented in Fig. 9.
8. In low-Mach-number CAA applications, reflection of a small portion of an out-going vorticity mode can create significant
acoustic errors. Therefore, strong sponges with lengths of 10 vortex diameters or more are recommended. This higher
(vortex-based) length demand is also consistent with sound–sound requirements as acoustic wavelengths become much
larger than vortex size in the limit of low Mach number. However, the sound–sound mechanism is of lesser concern since
the sound–sound mechanism can generally be tackled by improving alignment.

5. Comparison with the perfectly matched layer

In this section we present results from computation of nonlinear Euler equations applied to a single isentropic vortex
interacting with a sponge near an outflow boundary. We compare the performance of sponge with that of a perfectly
matched layer (PML) using the data published by Hu et al. [10]. In our computation we replicate the same scenario as that
considered by Hu et al. [10] in all physical aspects including domain size, vortex size and profile (same as in Eq. (20)), and
flow parameters.
We considered a vortex with radius b = 0.2 in a physical domain bounded by 1 < x < 1 and 1 < y < 1. The vortex has a
maximum rotational Mach number of 0.25 and convects in a uniform flow with M = 0.5. We used a sixth-order compact
finite difference scheme on a staggered mesh [24] to discretize the nonlinear Euler equations in space. The mesh size was
selected to be 1/30 in each direction. The fourth-order Runge–Kutta scheme with acoustic CFL number of 0.2 was used
for time stepping. Similar to the approach of Hu et al., a computation on a larger physical domain (by a factor of 2 in the
y- and a factor of 4 in the x-direction) was used as a reference solution to compute the errors. Sponges with different lengths
were used outside of the physical domain to damp the vortex. Guided by the results in Fig. 9(b), we selected sponge strength
of 20 dB for our tests.

4
Note that the ratio of the reflected to incident energy flux for each mode involves the amplitude ratio as well as the ratio of group velocities in the x-
direction. Also the reflected acoustic wave involves both pressure and velocity modes which equally contribute to energy.
 A. Mani / Journal of Computational Physics 231 (2012) 704–716 715

−1
10
sponge
PML

−2
10

error
−3
10

−4
10
0 1
10 10
cost

Fig. 10. Maximum reflection error for v velocity along the line x = 0.9 near the outflow boundary of the vortex/sponge problem. The cost for sponge is lsp/
Dvort and for PML is 14lsp/5Dvort due to additional transport equations. The PML data are from Hu et al. [10].

Fig. 10 compares the reflectivity performance of the two methods. We note that, for the same length, a PML calculation is
generally more expensive than a sponge treatment. This is because PML requires computation of auxiliary transport equa-
tions in addition to the primary quantities. For a single boundary (i.e., only in the x-direction), PML requires solving for 9
additional quantities and thus is more expensive by a factor of 14/5. Considering this factor, one can see from Fig. 10 that
the sponge has comparable performance to PML for moderate sponge lengths (5 to 10 vortex diameters). The trends sug-
gest that as the length becomes higher (leading to higher accuracy), the sponge becomes more effective than the PML with
the same cost.

6. Discussions and summary

In our analysis we have ignored discretization errors. This is a reasonable assumption, as long as the sponge profile is well
resolved and the mesh size does not vary rapidly. The presence of highly stretched meshes, which may be used for cost sav-
ing in the sponge region, can cause new sources of reflection. As practiced by many researchers, one remedy would be to
combine dissipative numerics with mesh-stretching in the sponge zone to avoid high wave number reflections while at
the same time saving computational cost.
Also we note that by searching over non-power-law profiles, it is possible to further optimize sponges. Through try-
and-error we have been able to gain 3 dB additional improvement in reflection coefficients over the quadratic profile. This
translates into 25% cost saving, which is an incremental improvement compared to the savings achieved within the power-
law optimization. Therefore, we limited our search to power-law sponges here since they achieve practically efficient and
simple designs.
In summary, we identified major reflectivity mechanisms for sponge/flow interactions and presented a quantitative anal-
ysis of these mechanisms using linearized Euler equations. We showed that sponges perform ideally (exponential damping)
for both one-dimensional systems and high-frequency features. The choice of sponge profile was shown to have a significant
influence on its performance. To achieve damping in the range 20–60 dB the quadratic sponge was found to be a practical
choice. The system of a sponge plus 1D characteristics-based end boundary condition was extensively analyzed for different
reflectivity mechanisms over a wide range of parameters. These sponge performance studies for different inflow/outflow
conditions and target-dampings can be used to estimate sponge requirements for practical CFD purposes as guidelined in
Section 4. Our computation of vortex/sponge interactions indicates that accuracies comparable to PML can be achieved with
sponges at the same cost. The simple implementation of the sponge treatment and its flexibility in handling complex geom-
etries and unstructured meshes makes it an attractive method to treat the external boundaries in general CFD applications.

Acknowledgments

This work was supported by the United States Department of Energy under the Predictive Science Academic Alliance Pro-
gram (PSAAP) at Stanford University. The author gratefully acknowledges comments and fruitful discussions with Dr. J. Nic-
hols and Dr. G. Lodato on a draft of this manuscript.

References

[1] T. Colonius, Modeling artificial boundary conditions for compressible flow, Ann. Rev. Fluid. Mech. 36 (2004) 315–345.
[2] M.B. Giles, Nonreflecting boundary conditions for Euler equation calculations, AIAA J. 28 (1990) 2050–2058.
[3] T.J. Poinsot, S.K. Lele, Boundary conditions for direct simulations of compressible viscous flows, J. Comput. Phys. 101 (1992) 104–129.
[4] C.K.W. Tam, Z. Dong, Radiation and outflow boundary conditions for direct computation of acoustic and flow disturbances in a nonuniform mean flow,
J. Comput. Acoust. 4 (1996) 175–201.
[5] M. Israeli, S.A. Orszag, Approximation of radiation boundary conditions, J. Comput. Phys. 41 (1981) 115–135.
[6] J.B. Freund, Proposed inflow/outflow boundary condition for direct computation of aerodynamic sound, AIAA J. 35 (1997) 740–742.
[7] T. Colonius, S.K. Lele, P. Moin, Boundary conditions for direct computation of aerodynamic sound generation, AIAA J. 31 (1993) 1574–1582.
[8] S. Karni, Far-field filtering operators for suppression of reflections from artificial boundaries, SIAM J. Numer. Anal. 33 (1996) 1014–1047.
716 A. Mani / Journal of Computational Physics 231 (2012) 704–716

[9] T. Colonius, H. Ran, A super-grid-scale model for simulating compressible flow on unbounded domains, J. Comput. Phys. 182 (2002) 191–212.
[10] F.Q. Hu, X.D. Li, D.K. Lin, Absorbing boundary conditions for nonlinear Euler and Navier–Stokes equations based on the perfectly matched layer
technique, J. Comput. Phys. 227 (2008) 4398–4424.
[11] F.Q. Hu, Development of PML absorbing boundary conditions for computational aeroacoustics: a progress review, Comput. Fluids 37 (2008) 336–348.
[12] D.J. Bodony, Analysis of sponge zone for computational fluid mechanics, J. Comput. Phys. 212 (2006) 681–702.
[13] C. Bogey, C. Bailly, D. Juve, Numerical simulation of sound generated by vortex pairing in a mixing layer, AIAA J. 38 (2000) 2210–2218.
[14] M.F. Barone, S.K. Lele, Receptivity of the compressible mixing layer, J. Fluid Mech. 540 (2005) 301–335.
[15] D.J. Bodony, S.K. Lele, Low-frequency sound sources in high-speed turbulent jets, J. Fluid Mech. 617 (2008) 231–253.
[16] M. Shoeybi, M. Svard, F.E. Ham, P. Moin, An adaptive implicit–explicit scheme for the dns and les of compressible flows on unstructured grids, J.
Comput. Phys. 229 (2010) 5944–5965.
[17] X. Gloerfelt, C. Bailly, D. Juve, Direct computation of the noise radiated by a subsonic cavity flow and application of integral methods, J. Sound Vibration
266 (2003) 119–146.
[18] J. Larsson, L. Davidson, M. Olsson, L. Eriksson, Aeroacoustic investigation of an open cavity at low Mach number, AIAA J. 42 (2004) 2462–2473.
[19] D.J. Bodony, Scattering of an entropy disturbance into sound by a symmetric thin body, Phys. Fluids, 21.
[20] N.A. Adams, Direct numerical simulation of turbulent compression ramp flow, J. Comput. Phys. 212 (2006) 681–702.
[21] A. Mani, P. Moin, M. Wang, Computational study of optical distortions by separated shear layers and turbulent wakes, J. Fluid Mech. 625 (2009) 273–
293.
[22] C. Bogey, C. Bailly, D. Juve, Noise investigation of a high subsonic, moderate Reynolds number jet using a compressible large eddy simulation, Theoret.
Comput. Fluid Dynamics 16 (2003) 273–297.
[23] D. Appelo, T. Colonius, A high-order super-grid-scale absorbing layer and its application to linear hyperbolic systems, J. Comput. Phys. 228 (2009)
4200–4217.
[24] S.K. Lele, Compact finite difference schemes with spectral-like resolution, J. Comput. Phys. 103 (1992) 16–42.