J. Fluid Mech. (2008), vol. 596, pp. 387–412.

c 2008 Cambridge University Press 387
doi:10.1017/S0022112007009603 Printed in the United Kingdom

Sound transmission in strongly curved slowly

varying cylindrical ducts with flow
E. J. B R A M B L E Y AND N. P E A K E
Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Centre for
Mathematical Sciences, Wilberforce Road, Cambridge CB3 0WA, United Kingdom

(Received 1 May 2007 and in revised form 3 September 2007)

In this paper we consider the propagation of acoustic waves on top of an inviscid

steady flow along a curved hollow or annular duct with hard or lined walls. The
curvature of the duct centreline (which is not restricted to being small) and the wall
radii vary slowly along the duct, allowing application of an asymptotic multiple-scales
analysis. The modal wavenumbers and mode shapes are determined locally as modes
of a torus with the same local curvature, while the amplitude of the modes evolves
as the mode propagates along the duct. The duct modes are found explicitly at each
axial location using a pseudospectral numerical method.
Unlike the case of a straight duct carrying uniform flow, there is a fundamental
asymmetry between upstream and downstream propagating modes, with some mode
shapes tending to be concentrated on either the inside or outside of the bend
depending on the direction of propagation, curvature and steady-flow Mach number.
The interaction between the presence of wall lining and curvature is also significant;
for instance, in a representative case it is found that the curvature causes the first
few acoustic modes to be more heavily damped by the duct boundary than would be
expected for a straight duct.
Using ray theory, we suggest explanations of these features. For the lowest
azimuthal-order modes, three distinct regimes are found in which the modes are
localized in different parts of the duct cross-section. This phenomenon is explained
by a balance between whispering-gallery effects along the duct and refraction by the
steady flow. At the opposite extreme, strongly spinning modes are investigated, and
are seen to be due to a different whispering-gallery effect across the duct cross-section.

1. Introduction
The propagation of acoustic waves along curved pipes has attracted much attention,
with a wide range of applications. One application, which is the motivation for the
research described in the current paper, is the prediction of unsteady flow along the
sort of convoluted intakes often found on the engines of military aircraft. One issue
here might be the behaviour of sound generated by the fan as it propagates upstream,
or alternatively (but not within the scope of this paper) at very large amplitudes the
propagation of surge events.
A selection of previous work will be mentioned here. Keefe & Benade (1983) used
ideas of impedance matching to study the propagation of very long waves along
a curved pipe. Pagneux and coworkers have developed multimodal techniques to
describe propagation in various sorts of curved ducts with zero mean flow (Pagneux,
Amir & Kergomard 1996a,b; Felix & Pagneux 2001, 2002). Felix & Pagneux (2004)
388 E. J. Brambley and N. Peake

U∞

y (s)

a2(S)

a1(S)

Figure 1. Schematic of the duct geometry.

have also studied sound attenuation round a lined bend, including a ray-tracing
explanation of the effects of curvature, although this was still for zero mean flow.
More analytically based studies have tended to use specific limits, including slender
curved ducts (Ting & Miksis 1983) and weakly curved ducts in two and three
dimensions (Gridin & Craster 2003; Adamou, Gridin & Craster 2005).
In a different direction, including mean flow for a straight duct with a circular cross-
section that varies slowly in the axial direction, Rienstra (1999) derived a multiple-
scales approximation for the unsteady field. This approximation has been validated
by Rienstra & Eversman (2001) by comparison with finite-element computations.
Rienstra’s analysis has been extended in a number of ways; by Rienstra (2003b) to
the case of arbitrary duct cross-section, by Cooper & Peake (2001) to the case of
swirling mean flow and by Ovenden (2005) to a uniformly valid solution that allows
modes to undergo cuton–cutoff transition. However, all of this has been for straight
ducts. The aim of this paper is therefore to investigate both curvature and mean flow
simultaneously.
The paper is organized as follows. In § 2 we derive the steady potential mean flow
through a curved duct. The unsteady linearized flow is described in § 3; as a mode
propagates along the duct it is distorted, and the description of this process involves
first the determination of the local axial wavenumber and mode shape, and second the
determination of the slowly varying amplitude. This local eigenvalue problem must be
solved numerically, and our pseudospectral method for doing this is described in § 4.
Results are presented in § 5, and the possibility of a mode transitioning from cuton
to cutoff is investigated in § 6. Finally, in § 7, ray-tracing theory is used to shed some
light on the structure of the duct modes, and to investigate the intriguing localization
phenomenon discovered in previous sections.

2. Steady mean flow

We consider a duct, as shown in figure 1, whose centreline possesses non-zero
curvature but zero torsion (i.e. the centreline lies in a plane). The duct has a circular
cross-section in planes normal to the centreline, and can be either hollow or annular.
We pick out two length scales for such a duct. The first,
∗ (∗ denotes dimensional
variables), is the length scale associated with the geometry of the duct at a given axial
∗
location, so that the duct radius a1,2 is of order
∗ and the duct curvature κ ∗ is of order
∗
1/
. This means the radius of curvature is on the same scale as the duct radius; were
 Sound in strongly curved lined ducts with flow 389
these two of different scales some asymptotic simplification becomes possible, but
here we deal with the full generality. The cases of a slender curved duct or a straight
duct of varying radius then follow as special cases. The second length scale, L∗ , is the
shortest length scale along the duct centreline over which these parameters vary.
Let us now be more specific. Let s ∗ be the arclength along the duct centreline. Far
upstream (s ∗ → −∞) the duct is assumed to be straight and of uniform outer radius

∗ . (Throughout this paper, the term straight refers to a duct with a straight centreline
(zero curvature), irrespective of variations of the inner and outer radii.) The radii of
the inner (and outer) walls and the centreline curvature then vary along the duct on
the length scale L∗ , so that the inner and outer radii a1,2
∗
and the centreline curvature
∗ ∗ ∗
κ are functions of S ≡ s /L . The requirement of slow variation along the axis is
then equivalent to  ≡ l ∗ /L∗  1. The duct carries a mean flow, which far upstream
∗
has uniform density D∞ , speed U∞∗ and sound speed C∞ ∗
. In what follows speeds
are non-dimensionalized by C∞ , densities by D∞ , distances by
∗ , times by
∗ /C∞
∗ ∗ ∗
,
∗ 2 ∗
and pressures by (C∞ ) D∞ . We introduce the duct-centred coordinate system (s, r, θ),
where r, θ are polar coordinates in planes normal to the duct centreline, and s is
the arclength along the centreline. The duct inner and outer radii are a1,2 (S) and the
centreline curvature is κ(S), where S = εs is the non-dimensionalized slow coordinate
over which the duct geometry varies.
The steady velocity in the duct is written U = U es + V er + W eθ , and it is assumed
that all steady-mean-flow variables are functions of r, θ, S, i.e. vary slowly down
the duct. The geometric factor associated with the curvilinear coordinate s is hs =
1 − κ(S)r cos θ. We assume an inviscid irrotational perfect gas with ratio of specific
heats γ . We apply the steady continuity condition ∇ · (DU) = 0,
ε ∂ 1 ∂ 1 ∂
1 (DU ) + (rhs DV ) + (hs DW ) = 0,
hs ∂S rhs ∂r rhs ∂θ
together with the condition for irrotational mean flow, ∇∧ U = 0,1
1 ∂ 1 ∂V 1 ∂ ε ∂W ε ∂V 1 ∂
(rW ) − = 0, (hs U ) − = 0, − (hs U ) = 0,
r ∂r r ∂θ rhs ∂θ hs ∂S hs ∂S hs ∂r
and the irrotational form of Bernoulli’s equation,
1 2 1
U + D γ −1 = H, (2.1)
2 γ −1
where the enthalpy H = U∞2 /2 + 1/(γ −1) is a constant determined at upstream infinity.
For the steady flow, the duct walls are considered perfectly hard and impenetrable,
with the corresponding boundary conditions
ε daj
V− U =0 at r = aj (S) j = 1, 2. (2.2)
hs dS
We assume no leading-order potential swirl, so that W vanishes to leading order.
Putting all these assumptions together, we find that
D = D0 + O(ε2 ), U = U0 + O(ε 2 ), V = εV1 + O(ε3 ), W = εW1 + O(ε3 ),
where
U† (S)
 1/γ −1
U0 (S, r, θ) = , D0 = (γ − 1) H − 12 U0 2 .
hs (r, θ, S)
390 E. J. Brambley and N. Peake

Intake 0.4
0.45
0.5
0.55
0.6
0.4

0.35

Fan face
0.3

Figure 2. Contours of the axial mean flow Mach number in the RAE 2129 Inlet Diffuser.
Far upstream is a uniform-inlet Mach number of 0.5.

The quantity U† may be found in terms of U∞ by applying conservation of mass at

different axial locations, to yield the implicit equation
 2π a2
U0 D0 r dr dθ = πU∞ (1 − a1 (−∞))2 , (2.3)
0 a1

which can easily be solved numerically. Note that the mean velocity components in
the plane of cross-section arise, in irrotational flow, from the slow variation of the
duct, and are therefore O(). In what follows it turns out that the value of the radial
velocity V1 is only required on the walls, and can be found simply from the O()
terms in the boundary condition (2.2). The value of the O(ε) tangential velocity W1
will not be required at all in our final answer for the unsteady flow.
One duct geometry we shall consider in particular is the RAE 2129 Inlet Diffuser
duct (as considered by Menzies 2002), which is a much studied reference duct
geometry. A cross-section along the RAE 2129 duct centreline is shown in figure 2,
along with the mean flow for a uniform-inlet Mach number U∞ = 0.5. The duct
geometry is defined in terms of the lateral offset of the centreline, y ∗ , from its position
at the intake s ∗ = 0 (as shown in figure 1), with
  ∗ 
h∗ πs
y ∗ (s ∗ ) = − 1 − cos , (2.4)
2 L∗
The lateral offset at the downstream exit (s ∗ = L∗ ) is then −h∗ . The duct itself
is hollow, with outer radius varying quartically between upstream (radius
∗ ) and
downstream (radius af∗ ) as
 4  3
a ∗ (s ∗ ) −
∗ s∗ s∗
=3 1− ∗ −4 1− ∗ + 1.
af∗ −
∗ L L

For the RAE 2129 duct, L∗ /
∗ = 7.1, h∗ /L∗ = 0.3 and (af∗ /
∗ )2 = 1.4. This leads to a
value of ε based on the duct length of ε = 1/7.1, for which it is reasonable to suppose
that the small-ε asymptotics of this paper will work well.
 Sound in strongly curved lined ducts with flow 391
3. Unsteady flow
Consider a small time-dependent perturbation (u, ρ, p) with time dependence
exp{iωt} to the mean flow (U, D, P ). Introducing a scalar potential u = ∇φ, and
neglecting vortical and entropic perturbations, the equations for the unsteady
linearized flow due to Goldstein (1978) reduce to
 
D 1 Dφ 1 Dφ
2
− ∇ · (D∇φ) = 0, p = C 2 ρ = −D , (3.1)
Dt C Dt D Dt
where D/Dt = iω + U · ∇ is the convective derivative with respect to the mean flow,
and C 2 = D γ − 1 is the square of the wave speed. (A similar, although less general,
equation describing sound propagation through a non-uniform medium was originally
given by Blokhintzev 1946.) Equation (3.1) is to be solved subject to the usual Myers
(1980) boundary condition for a lined duct, namely
iωn · ∇φ = {iω + U · ∇ − [(n · ∇)U] · n}(p/Zj ) on r = aj (S) for j = 1, 2, (3.2)
where Z1,2 (S) are the wall impedances and n is the corresponding wall normal pointing
out of the fluid. (Note that, had an exp{−iωt} convention been adopted, Z1,2 would
be the complex conjugate of what it is here.)
In order to account for the slowly varying duct geometry and mean flow we follow
Rienstra (1999) and pose the multiple-scales WKB ansatz (see for instance Hinch
1991, chapter 7)
 
i S
φ = [A0 (S, r, θ) + εA1 (S, r, θ) + O(ε2 )] exp iωt − k(S  ) dS  .
ε 0
The details of what follows are given in the Appendix. In short, substituting this
into (3.1) and taking just the O(1) terms gives
     2 
1 1 ∂ ∂A0 1 ∂ ∂A0 Λ k2
rhs D0 + 2 hs D0 + − A0 = 0, (3.3)
hs D0 r ∂r ∂r r ∂θ ∂θ C0 2 hs 2
where Λ = ω − kU0 / hs and C0 2 = D0 γ−1 . The O(1) boundary condition from (3.2) is
∂A0 D0 Λ2 A0
=± on r = aj (S) for j = 1, 2, (3.4)
∂r iωZj
where ± refers to the outer and the inner walls respectively. For hard walls (3.4)
becomes simply ∂A0 /∂r = 0. One crucial difference here from the case of a straight
circular duct is the highly non-trivial dependence of A0 on θ. When κ = 0, (3.3)
and (3.4) can be solved using separation of variables to yield a well-known solution
proportional to exp{−imθ } for integer m and depending on Bessel functions in the
radial direction. But when κ = 0, (3.3) is no longer separable. Equations (3.3) and (3.4)
must therefore be solved numerically to determine the axial wavenumber k(S) and
the corresponding wave function A0 (S, r, θ). Details are given in the next section. For
definiteness, we normalize the solution so that
 2π a2
D0 U0 ω 2
A0 (S, r, θ) = N (S)Â0 (S, r, θ), Â0 r dr dθ = 1. (3.5)
0 a1 C0 2
The reason for this strange-looking normalization is that it becomes much more
natural when we consider turning point in § 6. The unknown amplitude N(S) must be
determined from the solvability condition obtained using the O(ε) terms from (3.1)
and (3.2), following the procedure given by Rienstra (1999). The details of this are
392 E. J. Brambley and N. Peake
given in appendix A. In short, by multiplying the O(ε) part of (3.1) by hs D0 A0 ,
integrating across the duct cross-section, and applying the mean flow equations of
motion to eliminate V1 and W1 , we arrive at the requirement that the quantity
{F (S) + I1 (S) + I2 (S)}N(S)2 (3.6)
is independent of S, where
 2π  
a2
ωU0 k U0 2
F (S) = D0 Â0 2
+ 1− 2 r dr dθ,
0 a1 C0 2 hs C0
 2π
ΛU0 D0 2 r 2
Ij (S) = Â0 dθ .
0 iωZj r=aj

2
Notice here that this condition involves Â0 , rather than |Â0 |2 , which has arisen from
the non-self-adjoint nature of the k eigenvalue problem (in fact the adjoint solution is
A∗0 ). In the case of rigid walls (3.6) reduces to the condition that F (S)N (S)2 is constant
along the duct, which for cuton modes can be interpreted as conservation of energy.
For finite impedance, acoustic energy from the flow is dissipated by the walls, and this
effect manifests itself in both the fact that the axial eigenvalue k(S) will be complex
and by the presence of the surface integrals I1,2 (S) in (3.6). Note that I1,2 are both
additive in (3.6), corresponding to the fact that energy is dissipated at both walls.
Putting all this together, we now have the leading-order solution (3.5) for the
unsteady flow, in which the local axial wavenumber and mode shape are determined
by numerical solution of (3.3) and (3.4) and the slowly varying amplitude is then
given by (3.6).

4. Numerical solution
Our task is to solve the leading-order eigenvalue problem (3.3) and (3.4) so as
to determine the local axial wavenumber k and corresponding eigenfunctions as
functions of the slow arclength S. The leading-order equation for A0 and k is recast,
by introducing B0 = kA0 , as the generalized eigenvalue problem
⎛  ⎞
   2ωU0 1 U0 2  
L 0 A0 ⎜ 2 2
1 − 2 ⎟ A0
= k ⎝ hs C0 hs C0 ⎠ , (4.1)
0 1 B0 B0
1 0
where
   
1 1 ∂ ∂A0 1 ∂ ∂A0 ω2
LA0 = rhs D0 + 2 hs D0 + 2 A0 ,
hs D0 r ∂r ∂r r ∂θ ∂θ C0
subject to the boundary conditions
 
∂A0 iωD0 2iU0 D0 iD0 U0 2
± + A0 = k A0 − 2 B0 ,
∂r Z1,2 hs Z1,2 hs ωZ1,2
where the negative sign is taken for the inner boundary (if one is present), and the
positive sign for the outer boundary.
We use a pseudospectral method with Chebyshev polynomials as the radial
basis (see for example Khorrami, Malik & Ash 1989; Boyd 2001, chapter 18.6,
p. 391) and trigonometric polynomials in the azimuthal direction. The number of
collocation points in the radial and azimuthal directions are denoted by nr and nθ
respectively (note that nθ must be odd, since all trigonometric polynomials have an
 Sound in strongly curved lined ducts with flow 393
odd number of degrees of freedom). For an annular duct, the collocation points are
given by
    
a1 + a2 a1 − a2 jπ 2
π j = 0, . . . , nr − 1,
(rj , θ
) = + cos , ,
2 2 nr − 1 nθ
= 0, . . . , nθ − 1.
For a hollow duct, a1 is replaced by a small non-zero constant, typically of the
order of the spacing between neighbouring collocation points at the centre, a1 ≈
a2 π2 /(2nr − 2)2 . While this is ugly, it allows both annular and hollow cases to be
calculated using the same numerical code, and offers the potential to treat a hollow-
to-annular transition (as pointed out by Rienstra 1999). For alternative ways to
discretize a hollow duct, see Boyd (2001, chapter 18.5, pp. 386–391).
Our system is now discretized for an annular duct by requiring the boundary
conditions to be satisfied at collocation points (0,
) and (nr −1,
) for
= 0, . . . , nθ −1,
and (4.1) to be satisfied at collocation points (j,
) for
= 0, . . . , nθ − 1 and
j = 1, . . . , nr − 2. For a hollow duct, (4.1) is also required to be satisfied at
collocation points (0,
) for
= 0, . . . , nθ − 1, and the inner boundary condition is
dropped. After a series of manipulations, which include representation of the r and θ
derivatives using standard spectral differentiation matrices, we arrive at a generalized
eigenvalue problem which is 2nθ nr square. This was solved using the QZ algorithm,
as implemented in the LAPACK library routine ZGGEV (Anderson et al. 1999).
In order to avoid spurious eigenvalues, two filtering processes were used. In the first,
based loosely on the description by Boyd (2001, pp. 137–139), eigenvalues which vary
significantly with small changes in the discretization are discarded. Second, we discard
eigenvectors for which nr or nθ are not large enough to resolve properly the true
eigenfunction. This is done by decomposing the numerical eigenvector into its spectral
representation in both r and θ directions, and then ensuring that the outlying spectral
coefficients are sufficiently small. Typically two thirds of the eigenvalues computed
failed these tests, although this fraction is strongly dependent on the parameters used.
We validated our numerical calculations by comparison with known analytic results
for the straight hard-walled ducts, and with numerical results for a lined straight
pipe. An exponential decrease in error with increasing (nr , nθ ) is obtained, as is
expected from a pseudospectral method. The results presented in the next section
were computed using typically nr × nθ = 31 × 61. For hollow ducts, a value of
a1 = 10−3 a2 was found to give good convergence.

5. Results
We first consider a hypothetical duct, with upstream conditions U∞ = 0.5, a1 = 0.4,
a2 = 1.0, γ = 1.4 and ω = 10. The curvature of the duct was considered to vary slowly
from κ = 0 upstream to κ = 0.1, at which point the spectrum is calculated. The
numerical eigenvalues for this duct with hard walls are shown in figure 3(a, b).
Figure 3(c, d) shows examples of cross-sectional modal shapes. Both of these are
upstream propagating modes, and the fundamental mode (the curved-duct equivalent
of a plane-wave mode) in figure 3(c) is seen to be localized in the inside of the bend.
The downstream propagating modes have similar shapes, but are localized on the
outside of the bend. Figure 3(e, f ) shows typical higher-order modes, both of which
are cutoff. These modal shapes are termed bouncing-ball and whispering-gallery type
modes respectively, see Babic & Buldyrev (1991).
One interesting feature of the spectrum is the diagonal vanes of eigenvalues
occurring periodically across the usual vertical line of cutoff modes in figure 3(a)
394 E. J. Brambley and N. Peake
(a) (b)
40 40

30 30

20 20

10 10

Im(k)
Im(k)

0 0

–10 –10

–20 –20

–30 –30

–40 –40
–25 –20 –15 –10 –5 0 5 10 –7.2 –7.1 –7.0 –6.9 –6.8 –6.7 –6.6 –6.5 –6.4
Re(k) Re(k)

(c) (d)
0.2 0.4
0.1 0.3 0.5
0.6
0.7
0.8
0.9
r
θ

(e) (f )

Figure 3. Upstream propagating modes in a hard-walled annular duct. (a) eigenvalues; (b)
cutoff eigenvalues; (c) fundamental mode; (d ) low-order mode; (e) bouncing-ball mode; and
(f ) whispering-gallery mode. Dashed lines indicate negative values. The axis orientation is
shown in (d ), giving the inside of the bend on the right and the mean flow into the page.
U∞ = 0.5, ω = 10, a1 /a2 = 0.4, and κ = 0.1.
 Sound in strongly curved lined ducts with flow 395
and in close-up in figure 3(b). This feature only seems to appear in the presence of
both non-zero mean flow and non-zero curvature, and can perhaps be associated with
the asymmetric mean flow and asymmetric mode shapes, leading to slightly different
Doppler shifts experienced by modes localized in different parts of the cross-section.
Similar results can also be seen for hollow rather than annular duct cross-sections.
Figure 4(a) shows eigenvalues for a curved (κ = 0.1), lined (Z = 2 − i) hollow duct
with mean flow (U∞ = 0.5). The additional series of eigenvalues in the lower-half
k-plane correspond to surface modes, exactly as for a straight duct (Rienstra 2003a;
Brambley & Peake 2006). Figure 4(d ) shows how such modes are strongly localized
near the boundary, while figure 4(c) shows the upstream-propagating acoustic mode
of the same order. This latter mode is a whispering-gallery mode, and while still being
localized close to the outer boundary is noticeably more pervasive into the duct than
the surface mode. (The true difference is that the mode in figure 4c decays algebraically
away from the boundary, while that in figure 4d decays exponentially.) At the other
extreme is the bouncing-ball mode shown in figure 4(b). These modes are similar to the
high-order modes in a hard-wall duct, except that for the lined duct there is very little
oscillation at the duct wall. Figure 4(e, f ) shows the fundamental duct modes, and
illustrate the dramatic asymmetry between upstream- and downstream-propagating
modes. The upstream-propagating mode is removed from the boundary, similarly to a
mode with a pressure-release boundary condition, while the curvature biases the mode
slightly to the inside of the bend (the right-hand side). The downstream-propagating
mode, in contrast, is strongly localized on the outside of the bend, and is oscillating
significantly on the duct boundary; it is very similar in form to a hard-wall duct mode.
Figure 5 shows how the axial wavenumbers k vary with the curvature κ. Note that
owing to the left–right symmetry (in the sense of the cross-sections shown in figures 3
and 4) for κ = 0, two distinct modes with κ = 0 may merge into a double mode with
κ = 0. As the curvature increases from zero, the first few downstream modes (on the
right of figure 5b) become more damped. In contrast, the surface modes (i.e. the lower
branch in the right half-plane) for k . 5 become less damped, while most of the well
cutoff acoustic modes (i.e. the ones in the line parallel to the vertical in figure 4a)
maintain the same rate of decay, although their phase speed shifts slightly towards
upstream.
Turning now to the RAE 2129 (hard-walled) duct described in § 2, the cuton
eigenvalues (i.e. those with real axial eigenvalue k) are plotted against the position
along the duct centreline in figure 6, for a realistic aeroengine rotor-alone frequency.
As can be seen, many modes which are cuton at the fan face (s = 7.1) will propagate
all the way to the intake (s = 0). In figure 7 we plot the sound pressure level (SPL)
of one such cuton mode (in the initial straight portion of the duct this corresponds to
the m = 24 first radial order mode, a typical aeroengine rotor-alone mode), and in this
case the amplitude varies rather little along the duct and the mode is concentrated
close to the duct wall all the way along. The curvature does give an amplification of
4 dB to the mode, occurring on the inside of the left-most bend. For these parameters,
there are also several duct modes which transition from cuton to cutoff within the
duct, i.e. the sequence of real wavenumbers starting at the fan reaches a minimum
value of s before turning round and moving back towards the fan (e.g. in figure 6 there
are four modes which turn around near s ≈ 2, etc.). These transitions correspond to
wave reflection by the changing geometry and flow, and will be described in detail in
the next section. The cutoff transition confines these modes to the fan end of the duct.
The amplitude of one such mode is shown in figure 8 (in the initial straight portion
of the duct corresponding to the m = 24 second radial order mode, a harmonic of
396 E. J. Brambley and N. Peake
(a) (b)
40

30

20

10
Im(k)

–10

–20

–30

–40
–20 –15 –10 –5 0 5
Re(k)
(c) (d )

(e) (f )
0.7 0.8 0.9
0.6
0.5
0.4
0.3
0.2
0.1

Figure 4. Results for a lined curved duct: (a) eigenvalues; (b) a bouncing-ball mode; (c) a
whispering-gallery mode; (d ) a surface mode; (e) the upstream fundamental mode; and (f ) the
downstream fundamental mode. Dashed lines indicate negative values. The axes are oriented
as in figure 3. U∞ = 0.5, κ = 0.1, Z = 2 − i, and ω = 10.
 Sound in strongly curved lined ducts with flow 397
(a) (b)
0
20
–0.5

10
–1

0 –1.5

–2
–10

–2.5
–20
–3
–25 –20 –15 –10 –5 0 5 10 0 2 4 6 8 10

Figure 5. Graph showing the motion of eigenvalues in the k-plane due to varying the curvature
from κ = 0.0 (+) to κ = 0.4 (×). Z = 2 − i, U∞ = 0.5, and ω = 10.0. (b) A close up of the nearly
cuton modes in (a).

30

20

10
Axial wavenumber, k

–10

–20

–30

–40

–50

–60
0 1 2 3 4 5 6 7
Distance along duct centreline, s

Figure 6. Real eigenvalues in the (hard-walled) RAE 2129 Inlet Diffuser shown in figure 2,
for ω = 26.2, giving a2∗ ω∗ /C∞
∗
= 31.0 at the fan face (s = 7.1). The mean flow has Mach number
0.5 at the intake (s = 0).

a typical aeroengine rotor-alone mode), indicating the standing-wave pattern formed

by the mode and its reflection between the fan and the transition point. This suggests
the possibility of acoustic resonance, in which acoustic modes are trapped upstream
of the fan by the cutoff transition and are prevented from propagating downstream
by the swirl in the rotor–stator gap (see Cooper & Peake 2000).

6. Turning points and wave reflection

For a hard-walled duct the secularity condition (3.6) represents conservation of axial
energy flux. Since a cuton mode has a non-zero energy flux and a cutoff mode has
398 E. J. Brambley and N. Peake

Intake

–20 dB
4 dB
–40 dB
–80 dB
0 dB

Fan face
–80 dB

–40 dB
–20 dB

Figure 7. SPL for the first radial order, 24th azimuthal order mode propagating from right
to left, normalized to give a maximum fan-face wall pressure of 0 dB. The axial wavenumber
k for this mode is shown in figure 6 as the middle of the three solid lines.

(a)
(b)

Fan face
Fan face

Figure 8. The second radial order, 24th azimuthal order mode propagating from the fan face
(on the right) towards the intake (on the left), before being reflected by the duct geometry and
propagating back towards the fan face: (a) Re(φ), (b) φ. The axial wavenumbers k for these
modes are shown as the upper solid line in figure 6. The incoming wave has unit maximum
amplitude. Contours of Re(φ) in (a) are at 0.5, 1.0, 1.5, and 2.0, with dashed lines indicating
negative φ. Contours of |φ| in (b) are at 1.0, 2.0, 3.0, and 4.0.

zero energy flux, it is to be expected that the secularity condition breaks down in the
neighbourhood of a cuton–cutoff transition. The secularity condition (3.6) becomes
singular when F (S) → 0, a so-called turning point. This is well understood in straight
ducts (see, for example, Rienstra 2003b), and here we now derive the turning-point
behaviour incorporating curvature.
 Sound in strongly curved lined ducts with flow 399
0.3
Real part
0.2
Imaginary part

F(S)2 0.1

–0.1
1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4
Distance along duct centreline, s

Figure 9. Graph showing the variation of F (S)2 over a cutoff to cuton transition in the
RAE 2129 Inlet Diffuser.
Define
 2π a2  
D0 2 U0 2
G(S) = Â0 1 − 2 r dr dθ,
0 a1 hs C0
so that k = (F (S) − 1)/G(S). From previous studies on a straight duct (Rienstra 1999),
we can expect that this transition occurs over a portion of the S-axis of length O(ε2/3 ),
and since G(S) will vary little over this interval it follows that the change in k from
real (cuton) to complex (cutoff) is associated with a change of sign of F (S)2 . Figure 9
shows the variation in F (S)2 for a mode in the RAE 2129 duct that undergoes a
cuton–cutoff transition; in the neighbourhood of the cutoff region, F (S)2 is seen to
be a linear function of S, and goes through zero as the mode transitions from cuton
to cutoff.
The singularity in the secularity condition (3.6) is artificial and is due to the neglect
of a term occurring at a higher power of ε. Including this term, the secularity condition
becomes
d d2 N
(F N 2 ) + iεGN 2 = 0. (6.1)
dS dS
√
Let the turning point be at S0 , so that F (S0 + S) = a S for some a. In what
follows an inbound downstream propagating mode is considered (exactly the same
analysis can be applied to an upstream propagating mode, for example as in the
RAE 2129 duct), so that a is negative. By introducing the inner variable x, (6.1)
becomes
√ iN
N  + 2i xN  + √ = 0, where x = Sε−2/3 G−2/3 a 1/3 ,
2 x
with branch cuts taken along the positive imaginary axis. This has leading-order
solution

2i
N = (AAi(−x) + BBi(−x)) exp − x 3/2 .
3
Using asymptotic expansions of the Airy function for large |x| (see Abramowitz &
Stegun 1964, pp. 448–449), N has the large-x behaviour

eiπ/4 4i
N ∼ √ 1/4 (A − iB) exp − x 3/2 − i(A + iB) as x → ∞,
2 πx 3

1 4 3/2
N∼ √ A + 2B exp |x| as x → −∞.
2 π|x|1/4 3
400 E. J. Brambley and N. Peake
The outer solution to (3.6) about the singular point S0 is
⎧
⎪ N0
⎪
⎪ for S < S0
⎪ (a S)1/4
⎨ iπ/4
N = T N0 e for S > S0 .
⎪
⎪
⎪ |a S|
1/4
⎪
⎩

Expanding in terms of the inner variable x and matching with the inner solution gives
B = 0 and T = 1. However, there is still an unmatched term as x → ∞. This term is
   S
Aeiπ/4 4i 3/2 iN0 2i F (S  ) 
√ 1/4 exp − x ∼√ exp dS .
2 πx 3 F (S) ε S0 G(S  )
The inbound downstream propagating mode has k = (F (S) − 1)/G(S). The corres-
ponding upstream propagating mode would have k̃ = (−F (S) − 1)/G = k − 2F (S)/
G(S). Hence, the downstream propagating mode from the singular point would have
the form
 
RN0 i S
√ exp − k̃(S  ) dS 
F (S) ε S0
    S
i S   RN0 2i F (S  ) 
= exp − k(S ) dS √ exp dS .
ε S0 F (S) ε S0 G(S  )
This is exactly the solution needed to match with the extra term in the inner solution,
provided R = i.
This conclusion is exactly the same as was reached for a turning point in a straight
duct by Rienstra (2003b). This is perhaps surprising, since within the inner region s
varies by O(ε−1/3 ), which is long compared with the local radius of curvature. The
inner region is therefore strongly curved, albeit with a constant curvature, and so one
might have expected a curvature-dependent reflection and transmission coefficient.
An example of the results of this matching behaviour is shown in figure 8, as
previously discussed. Note that only the outer solution for φ is plotted in figure 8,
and therefore a singularity at the reflection point is shown. This singularity is not
physical, and is smoothed over by the inner Airy function solution. A method for
obtaining a uniformly valid solution, incorporating both the inner and outer solutions,
is given by Ovenden (2005), motivated by investigating cuton–cutoff transition near
the ends of the duct (especially for civilian aeroengines, where the ducts are very
short). This was not investigated further here, as the reflection points for a curved
duct tend to occur in the centre of the duct, and away from this reflection point
the accuracy of the outer solution is unaffected by the singularity. In particular, the
reflection and transmission coefficients derived above, which are arguably the most
important result of this section, are correct without needing to resort to a uniformly
valid solution.

7. Ray theory for a hard-walled duct

We now investigate the high-frequency limit using ray theory. This will turn out
not only to provide an asymptotic approximation for certain types of modes, but
will also provide a more intuitive insight into why modes have the characteristics
discovered above. We consider two different limiting cases. First, we are concerned
with the fundamental modes, i.e. those which are very close to being plane waves.
 Sound in strongly curved lined ducts with flow 401
For a straight duct these modes would have azimuthal order m = 0, and would be
uniform across the duct cross-section. With wall curvature (and non-uniform mean
flow), we see that the fundamental modes can be localized on one side of the bend
or the other. In contrast, we consider second the limit of very high azimuthal order,
and apply a known result of Babic & Buldyrev (1991) to determine the whispering
gallery modes.
In what follows, S will only occur as a parameter specifying which slice of the
duct is being considered, and dependence on S will not be explicitly mentioned in the
following equations. The gradient ∇⊥ is used for the derivative in the cross-section of
the duct, i.e. the r and θ directions only. We start off by introducing the ray ansatz
for A0 (r, θ) by writing
A0 (r, θ)
A0 (r, θ) = √ exp(−iωψ(r, θ)). (7.1)
hs (r, θ)D0 (r, θ)
The reason for normalizing by the square root is to transform (3.3) into a form suitable
for ray tracing (see (A 6) in appendix A). We suppose that ω, the dimensionless
frequency, is large, so that (7.1) corresponds to highly oscillatory modes, while ψ is
O(1). The corresponding unknown axial eigenvalue is O(ω), and so we write k = µω
with µ = O(1). Substituting (7.1) into (3.3) (or equivalently into (A 6)) and taking
just the leading terms in ω, i.e. O(ω2 ), we get
 
1 2µU0 µ2 U0 2
2 2
(∇⊥ ψ) = α , α = 2−
2
− 1− 2 , (7.2)
C0 hs C0 2 hs 2 C0
which is the standard ray-tracing result for propagation through a medium with
non-uniform wave speed 1/α (this is a standard ray-tracing derivation; see Babic &
Buldyrev 1991, chapters 1 and 4 for details). Substituting (7.1) into (3.4), taking the
hard-wall limit Zj → ∞, and taking just the leading terms in ω, i.e. O(ω), gives the
boundary condition ∂ψ/∂r = 0 at r = aj . This gives simply normal reflection of a ray
by the boundary (again, see Babic & Buldyrev 1991, chapter 4).
This result is quite important because it shows that, for large dimensionless
frequencies ω, our problem for the curved duct with variable mean flow can be
replaced by ray tracing in two dimensions within a circular boundary with spatially
varying sound speed 1/α. Note that the duct curvature and the duct radii affect α 2
through the mean flow terms (U0 , C0 and D0 ), as well as the curvature appearing in
the metric factor hs . In fact, it may be seen that α 2 only depends on x = r cos θ, the
transverse position towards the inside or outside of the curve of the duct, and the
position along the duct labelled by the slow axial coordinate S. The duct modes may
therefore be thought of as bouncing around inside the cross-section of the duct, being
reflected normally by the boundary, subject to a variable wave speed 1/α that varies
horizontally from the inside to the outside of the bend, but not vertically.
7.1. Plane-wave localization
The plane-wave fundamental modes for a straight duct are uniform across the duct
cross-section. For a curved duct, some fundamental modes are shown in figures 3(c)
and 4(e, f ). As mentioned above, for these parameters the upstream-propagating
modes are localized on the inside of the bend, while the downstream-propagating
modes are localized on the outside. However, this is not always the case: figure 10
shows that, for κ = 0.2, U∞ = 0.5, and ω = 31 there are two upstream-propagating
fundamental modes, one localized on the inside of the bend and one on the
outside, with the downstream-propagating mode still localized on the outside. The
402 E. J. Brambley and N. Peake
(a) (b) (c)

Figure 10. Fundamental modes for a hard-wall duct with κ = 0.2, U∞ = 0.5, and ω = 31. The
inside of the bend is on the right. (a) and (b) Upstream-propagating modes with k = −66.4
and −62.1 respectively, while (c) is a downstream-propagating mode with k = 25.1.

0.8

0.4

0 Not
Resolved
–0.4 35
30
–0.8 25
20
–60
15 m
–40
–20 10
k 0 5
20
0

Figure 11. The wavenumbers k for propagating modes against their approximate azimuthal
order m, for a hard-wall duct with κ = 0.2, U∞ = 0.5, and ω = 31. The vertical axis plots
Wx , with modes localized on the inside and outside that have stalks extending upwards and
downwards respectively. Only numerically resolved modes are shown (nr = 23, nθ = 131); the
curve denotes the unresolved region.

wavenumbers for these modes are shown in figure 11, together with a left–right
weighting
 2π  a2
r cos(θ)|A0 (r, θ)|2 rdrdθ
Wx = 0  2π
a1
 a2 .
|A0 (r, θ)| r dr dθ
2
0 a1

Modes which are highly localized on the left/right of the duct (i.e. on the
outside/inside of the bend) have a value of Wx close to ∓1 respectively. The presence
of a couple of upstream modes localized on the outside is clear in figure 11. The
question is, therefore, what parameters influence the localization of duct modes?
In three dimensions, the fundamental modes may be thought of as travelling nearly
axially down the duct, reflecting occasionally from the duct boundary. In order to do
this, the flow and the geometry must be such that a ray, having just reflected from the
boundary, is driven back towards the boundary, as shown in figure 12(a). Figure 12(b)
 Sound in strongly curved lined ducts with flow 403
(b)

(a)

Figure 12. Schematic of localized fundamental rays. The dashed lines are where α 2 = 0. (a) A
projection of the three-dimensional rays onto a vertical plane, and (b) the corresponding pro-
jection onto the cross-sectional plane.

shows the corresponding projection of the rays onto the duct cross-section, as derived
above. The dashed lines are where α 2 becomes negative, and the rays are restricted
to regions where α 2 is positive.
A ray travelling purely axially at the local speed of sound along the duct at a fixed
horizontal offset x from the duct centreline would give the ray-tracing parameter
µ± (x) = hs /(U0 ± C0 ), with + for downstream and − for upstream. Substituting µ± (x)
into (7.2) shows that α 2 = 0 for each x, corresponding to the fact that such a ray is
travelling completely in the axial direction, and therefore has no motion in the duct
cross-section. Here, we are concerned with modes localized on the inside or outside
boundary, and so (normalizing such that the duct radius is a2 = 1) we are concerned
only with the value of µ± (x) at x = 1 (for the inside of the bend) and x = −1 (for the
outside of the bend). Figure 13 shows the variation of α 2 across the duct for the four
cases in which µ takes one of the values µ+ (±1), µ− (±1). Figure 13(b) corresponds
to figure 3, while figure 13(d ) corresponds to figure 10. By perturbing the value of µ
slightly from the values µ± (±1), it is possible that a small pocket of positive α 2 might
be created close to the duct wall, and thus (provided the frequency is high enough) a
localized mode on that boundary is possible. In order for a small perturbation to µ to
lead to a localized mode on the boundary, α 2 (x) must decrease away from that bound-
ary. This is also the requirement that a ray having just reflected from the boundary is
driven back towards the boundary (it is a standard ray-tracing result that rays bend
towards regions with larger α, as may be seen by taking ∇⊥ of the left-hand side of
(7.2)). A change in the number of localized fundamental modes is therefore seen when
the derivative of α 2 at x = ±1 changes sign. As an example, figure 13(a) demonstrates
the possibility of a downstream fundamental mode localized on the outside of the
bend and an upstream fundamental mode localized on the inside. Figures 13(b) and
13(d) both demonstrate the possibility of a downstream fundamental mode localized
on the outside, and two upstream fundamental modes, one localized on the inside and
one on the outside. Figure 13(c) demonstrates the possibility of both the downstream
and upstream fundamental modes being localized on the outside of the bend.
In order to investigate which values of U∞ and κ give rise to which types of
localization behaviour, we will now look for a change in derivative of α 2 at x = ±1.
Differentiating α 2 with respect to x gives
 
∂α 2 (γ − 1)κU0 2 2κU0 H (γ − 1)µ(2hs − µU0 ) 2κµ2 U0 2
= − − 1− 2 , (7.3)
∂x C0 4 hs C0 4 hs 3 hs 3 C0
404 E. J. Brambley and N. Peake
(a) (c)
0.8 0.4

0.4 0.2

0
0

–0.2
–0.4
–0.4
–0.8
–1.0 –0.5 0 0.5 1.0 –1.0 –0.5 0 0.5 1.0

(c) (d )

0.2 0.4

0 0
α2
–0.2 –0.4

–0.4 –0.8

–1.0 –0.5 0 0.5 1.0 –1.0 –0.5 0 0.5 1.0

x x

Figure 13. Graphs showing the variation in α 2 (x) across the duct cross-section, for potential
upstream- and downstream-propagating modes localized on the inside and outside of the
duct bend: (a) U∞ = 0.6, κ = 0.1; (b) U∞ = 0.5, κ = 0.1; (c) U∞ = 0.3, κ = 0.1; (d)
U∞ = 0.5, κ = 0.2. The solid line is the upstream outside mode (i.e. µ = µ− (−1)), the long
dashed line the upstream inside (i.e. µ = µ− (1)), the short dashed line the downstream inside
(i.e. µ = µ+ (−1)), and the dash-dot line the downstream outside (i.e. µ = µ+ (1)).

and substituting µ = µ± (x) = hs /(U0 ± C0 ) gives

∂α 2 κ U0 2 U0
= (γ − 1) ∓ 4 −2 . (7.4)
∂x hs (U0 ± C0 )2
C0 2
C0
A sign change of ∂α 2 /∂x for a ray propagating axially down the duct is therefore
given by a zero of the square brackets in (7.4). After some algebraic manipulation,
this gives
  
2
U0 = H 1 ±
2
. (7.5)
γ +1
Note that the + solution of this (corresponding to a downstream-propagating mode)
is very close to U0 2 = 2H , at which point the mean density on the inside of the bend
becomes zero.
Equation (7.5) may be evaluated to give κ in terms of U∞ , or U∞ in terms of κ,
for the critical parameters for which an upstream or downstream propagating mode
(− or + in (7.5)) may be localized on the inside or outside of a bend (evaluating
(7.5) at x = 1 or x = −1). Unfortunately, since U† depends on both κ and U∞ and is
calculated numerically, (7.5) must in general be solved numerically. However, if the
duct cross-sectional area is the same as far upstream, then it can be shown from (2.3)
 Sound in strongly curved lined ducts with flow 405
1.0

Zero density
0.8

0.6

Cho
Curvature, k

ked
0.4

0.2

0.0
0 0.2 0.4 0.6 0.8 1.0
Up stream Machnumber (U∞)

Figure 14. The different localization behaviours of plane waves for different upstream Mach
numbers (U∞ ) and curvatures (κ). The sketches show whether upstream and downstream
modes localized on the inside and outside of the intake are possible, with downstream being
the right-facing arrow. Solid lines are numerical results from (7.5), and dashed lines are
asymptotic results from (7.6). The two dash–dot lines (one numerical and one asymptotic) are
where U0 2 = 2H , for which the density on the inside of the bend becomes zero.

that U† = U∞ + O(κ 2 ). Hence, if the cross-sectional area is the same as far upstream
and κ is small, (7.5) gives
⎛    −1/2 ⎞
1⎝ 2 ⎠,
κ= 1 − U∞ H 1 ± (7.6)
x γ +1

with x being either +1 or −1 for the inside or outside of the bend. Here H is given
by (2.1) upstream.
Figure 14 plots the small-curvature asymptotics given by (7.6) and the numerically
calculated solutions of (7.5). The numerically generated solutions stop around
U∞ = 0.65, since for these parameters the duct is choked; i.e. there is no solution
to (2.3) for U† that gives the required mass flow rate.
Interestingly, the boundaries between the different behaviours of the upstream
fundamental mode intersect at κ = 0, as shown in figure 14. At this point, the small-
curvature asymptotics give the exact answer (assuming the duct cross-sectional area
is the same as far upstream), and rearranging (7.6) gives the upstream Mach number
for which this occurs as
 
2 γ +1
U∞ = −1 .
γ −1 2

For γ = 1.4, as used for all the examples presented here, this gives U∞ ≈ 0.477. This
means that, for U∞ close to this value (and U∞ = 0.5 has been used for most examples
presented here) for all but very small curvatures it is possible for both the upstream
406 E. J. Brambley and N. Peake
–3 dB
–10 dB
–20 dB
–30 dB
Intake

–30 dB

–20 dB

–10 dB

Fan face
–3 dB

Figure 15. SPL for the plane-wave mode propagating from right to left, normalized to give a
maximum fan-face wall pressure of 0 dB. The axial wavenumber k for this mode is shown in
figure 6 as the bottom of the three solid lines.

localized modes to be present. Indeed, this can be seen in figure 10, where both inside-
and outside-localized upstream modes are shown.
For low Mach number flows, the geometry keeps both upstream and downstream
modes localized on the outside of the bend. This is exactly the result seen by Felix &
Pagneux (2004). However, different behaviour is seen for larger Mach number flows.
For large Mach number flows, the mean flow is fastest on the inside of the bend and
slowest on the outside, giving a refraction effect which curves upstream-propagating
rays towards the inside of the bend and downstream-propagating rays towards the
outside. Hence, as the Mach number is increased from zero, the upstream mode
first becomes present on the inner wall, and then disappears from the outside as
the Mach number is increased further. Increasing the curvature makes this effect
more pronounced. This case is similar to the unsteady flow through a loaded cascade
considered by Atassi, Fang & Patrick (1993), in which upstream propagating waves
can propagate along the suction surface of the blade (corresponding to the inside of
our bend) and become blocked if the mean flow becomes locally close to sonic, giving
increased local amplitudes on the suction surface.
Figure 15 gives an example of an upstream-propagating plane wave in the RAE 2129
duct, in a similar manner to figure 7. In this case, the mode is always localized on
the outside of the bends, and becomes a straight-duct plane wave at the midpoint
of the intake where the curvature passes through zero. On the right (fan-end) of the
duct only this localization is possible. On the left (intake-end) a mode localized on
the inside of the bend is also possible, but is not excited.
The validity of the results of this section requires the frequency ω to be high enough.
For the relatively low frequency ω = 10, as in figures 3 and 4, no upstream modes
localized on the outside can be found. However, for the high frequency ω = 31, the
ray-theory prediction of inside and outside upstream fundamental modes presented
in figure 14 is confirmed in figure 11.
 Sound in strongly curved lined ducts with flow 407
While the results presented here have been for a duct with the same cross-section
as far upstream, (7.5) is still valid even if this is not the case. The small-κ asymptotics
of (7.6) do require this restriction, however.
7.2. Whispering-gallery modes
Keller & Rubinow (1960) used ray theory to construct a method for determining
the eigenvalues of the Helmholtz equation in certain closed domains containing a
uniform acoustic medium. Their procedure was adapted by Babic & Buldyrev (1991)
to domains with a varying sound speed. We now make use of Babic & Buldyrev’s
procedure to obtain analytically based approximations for the eigenvalues of our
curved duct at high frequency and high azimuthal order for the case of hard duct
walls.
One key step in the Keller & Rubinow procedure is the determination of the shape
of a caustic surface inside the domain, which provides an envelope for all possible
ray directions. In the case of a circular domain with a uniform medium these caustics
are simply concentric circles, whose radii are related to the allowed eigenvalues of
the problem. However, for our equivalent variable sound speed it does not appear
possible to determine the shape of this caustic in general, but analytical progress can
be made for so-called whispering-gallery modes. Whispering-gallery modes consist of
rays running close to the boundary, bouncing a large number of times at very short
intervals. A typical modal shape is shown in figure 4(c).
Babic & Buldyrev (1991) determined an asymptotic expression for the eigenvalues
of whispering-gallery modes of a circle with non-uniform sound speed. These modes
are parameterized by two integers: m  1, the azimuthal order, and j = 1, 2, . . . ,
the radial order. For the case of a curved duct, with the effective sound speed given
by (7.2), their analysis gives (see Babic & Buldyrev 1991, § 5.3)
 
1/3
πµ 9m
kj m = 2m + I2 (a2 ) (j − 3/4)2 , (7.7)
I1 (a2 ) 4I1 (a2 )

where
 2π  2π  2/3
1 1 ∂α 2 ∂α 2 ∂α 2
I1 (r) = α r dθ, I2 (r) = α 1/3 + 2 r dθ, = cos θ ,
0 0 r 2α ∂r ∂r ∂x
2
with ∂α /∂x given in (7.3). This expression is valid for both annular and hollow
ducts; because the rays are bouncing around the outer boundary, the inner boundary
plays no part. Note that (7.7) gives kj m (µ) implicitly as a function of µ ≡ kj m /ω,
and an iterative method was therefore needed to find the axial wavenumber kj m for
a specified value of ω. Figure 16 shows the results of the ray-tracing asymptotics
against numerically calculated eigenvalues. A frequency of ω = 40 was used for the
comparison, so as to allow high azimuthal-order modes to be cuton. The results are
plotted against the azimuthal order m so the individual modes can be distinguished.
The agreement is seen to be reasonable, especially for nearly cutoff large-m modes,
as is to be expected from using large m asymptotics.

8. Concluding remarks
In this paper we have considered the propagation of acoustic waves along a
curved duct carrying mean potential flow. The wall radii, impedance of the acoustic
lining and curvature of the duct centreline may vary slowly along the duct, allowing
408 E. J. Brambley and N. Peake
45

40
Azimuthal order, m

35

30

25

20

15
–80 –70 –60 –50 –40 –30
Axial wavenumber, k

Figure 16. Comparison of ray-tracing asymptotics (×) with numerical results (+).
U∞ = 0.5, κ = 0.1 and ω = 40.

application of an asymptotic multiple-scales analysis. Our general results have not

relied on any assumption of weak curvature. We have seen that the spectrum of
local eigenmodes is more complicated than in the straight-duct case, due to the
fundamental asymmetry between upstream and downstream propagating modes. For
the plane-wave fundamental modes in a hard-wall duct, this asymmetry has been
classified into different regions of parameter space, depending on the local curvature
and the upstream Mach number. This asymmetry is even more pronounced when an
acoustic lining is present. In contrast, our closed-form expression for the variation
of the slowly-varying amplitude is very similar in form to the straight-duct case,
possibly due to its connection with the energy flux along the duct. This similarity
even extends to cuton–cutoff turning points, for which the straight-duct solution is
recovered despite the duct being significantly curved (albeit with constant curvature)
on the inner asymptotic region.
Throughout this paper we have assumed an inviscid mean flow, and indeed in
most circumstances this is a valid assumption. In some cases, however, particularly
for strongly curved ducts and high Mach number mean flows, it is possible for the
viscous boundary layer along the duct wall to separate. This possibility has been
neglected here. However, assuming the boundary layer remains attached to the duct
walls, it is expected that the inviscid assumption will be acceptable.
The pseudospectral eigenvalue solver is sufficiently general to allow straightforward
extension of our solution to more complicated geometries; for example, to the curved-
duct version of Rienstra’s (2003b) solution for a duct of arbitrary slowly varying cross-
section. Another possible extension of this work would be to relax the restriction that
the duct centreline is planar. This means that as well as a non-zero slowly varying
curvature, the duct centreline would possess a slowly varying torsion τ (S). Germano
(1982) showed how the coordinate system can be modified to account for this effect,
simply by replacing the cross-sectional polar angle θ by

1 S 
θ+ τ (S ) dS  ,
ε 0
 Sound in strongly curved lined ducts with flow 409
which yields an orthogonal coordinate system which effectively twists with the duct
centreline. For zero mean flow the effect of torsion can therefore be included as a very
simple modification of the results given in § 3. For non-zero mean flow, however, the
cross-sectional components V , W become O(1) when τ = 0, rather than O(ε) when
τ = 0, and this would lead to significant, but perhaps not intractable, complication.

This work has been supported by an EPSRC CASE award from Rolls-Royce.
Helpful discussions with Dr S. W. Rienstra are gratefully acknowledged.

Appendix. Details of the multiple-scales derivation

We consider a small perturbation (u, ρ, p) with time dependence exp{iωt} to the
steady mean flow (U, D, P ). We neglect vortical and entropic perturbations, and
introduce a scalar potential u = ∇φ. Our starting point is (3.1), the equation for the
unsteady linearized flow due to Goldstein (1978)
 
D 1 Dφ 1 Dφ
2
− ∇ · (D∇φ) = 0, p = C 2 ρ = −D , (A 1)
Dt C Dt D Dt
where D/Dt = iω + U · ∇ and C 2 = D γ − 1 , and (3.2), the Myers (1980) boundary
condition for a lined duct,
iωn · ∇φ = {iω + U · ∇ − [(n · ∇)U] · n}(p/Zj ) on r = aj (S) for j = 1, 2, (A 2)
where Z1,2 (S) are the wall impedances and n is the corresponding wall normal pointing
out of the fluid.
We pose the multiple-scales WKB ansatz (see for instance Hinch 1991, chapter 7)
 
i S
φ = A(S, r, θ) exp iωt − k(S  ) dS  . (A 3)
ε 0
For brevity, define the linear operators
kU0 U0 ∂ ∂ 1 ∂
Λ = ω− D= + V 1 + W1 ,
hs hs ∂S ∂r r ∂θ
so that D/Dt = iΛ + εD + O(ε 2 ). Substituting (A 3) into (A 1) gives
     2 
1 1 ∂ ∂A 1 ∂ ∂A Λ k2
rhs D0 + 2 hs D0 + − A
hs D0 r ∂r ∂r r ∂θ ∂θ C0 2 hs 2
   
2Λ Λ 2k ∂A A ∂ D0 k
= iε DA + AD + 2 + + O(ε2 ), (A 4)
C0 2 C0 2 hs ∂S hs D0 ∂S hs
subject to the boundary conditions (from (A 2))
 
∂A Λ2 D0 ik daj 1 ΛD0 A2
± − A= − ε ± 2 A+ D
∂r iωZj h dS A ωZj
  s 
D0 Λ daj ∂ U0 ∂V1
+ − A
Zj ω dS ∂r hs ∂r
+O(ε 2 ) (A 5)
at r = aj , where the positive sign is taken for j = 2 and the negative sign for j = 1.
410 E. J. Brambley and N. Peake
For ray-tracing applications, (A 4) may be rearranged to give
   
1 ∂ ∂A 1 ∂ 2A κ U0 2 ∂A 1 ∂A
r + 2 2 − 1+ 2 cos θ − sin θ
r ∂r ∂r r ∂θ hs C0 ∂r r ∂θ
 
ω2 2ωU0 k k2 U0 2
+ − − 1 − A
C0 2 hs C0 2 hs 2 C0 2
 2   
iε AΛ ∂ D0 kA2
= hs D0 D + ,
hs D0 A C0 2 ∂S hs
√
and then, by making the substitution ψ = hs D0 A,
 
1 ∂ ∂ψ 1 ∂ 2ψ
r + 2 2
r ∂r ∂r r ∂r
   
ω 2
2ωU0 k k2 U0 2 κ2 U0 2 U0 4
+ − − 1 − + 1 + 4 + (2γ − 3) ψ
C0 2 hs C0 2 hs 2 C0 2 4hs 2 C0 2 C0 4
   
iε ψ 2Λ ∂ kψ 2
= h s D0 D + . (A 6)
ψ hs D0 C0 2 ∂S hs 2
Returning to (A 4) and expanding A = A0 + εA1 + O(ε2 ) gives, to leading order,
     2 
1 1 ∂ ∂A0 1 ∂ ∂A0 Λ k2
LA0 ≡ rhs D0 + 2 hs D0 + − A0 = 0,
hs D0 r ∂r ∂r r ∂θ ∂θ C0 2 hs 2
(A 7)
subject to the boundary conditions (from (A 5))
∂A0 Λ2 D0
=± A0 at r = aj , j = 1, 2. (A 8)
∂r iωZj
Equations (A 7) and (A 8) are exactly the leading order equations (3.3) and (3.4).
At O(ε), (A 4) gives
 2   
i A0 Λ ∂ D0 kA0 2
LA1 = hs D0 D + , (A 9)
hs D0 A0 C0 2 ∂S hs
where L is the operator defined in (A 7), subject to the boundary conditions (from
(A 5))
∂A1 Λ2 D0 ik daj
± − A1 = ∓ 2 A0
∂r iωZj hs dS
     
1 ΛD0 A0 2 D0 Λ daj ∂ U0 ∂V1
− D − − A0 (A 10)
A0 ωZj Zj ω dS ∂r hs ∂r
at r = aj , with ± being + at r = a2 and − at r = a1 (this convention will be assumed
from here on). Note that, by using integration by parts to move derivatives from A1
onto A0 and the fact that LA0 = 0,
 2π a2  2π   a2
∂A1 ∂A0
A0 L(A1 )hs D0 r dr dθ = A0 − A1 rhs D0 dθ ,
0 a1 0 ∂r ∂r a1

where the [ · · · ]2π

0 terms that would have appeared above are zero owing to 2π
periodicity. Multiplying (A 9) by A0 hs D0 r and integrating over r and θ, and then
 Sound in strongly curved lined ducts with flow 411
substituting for ∂A0 /∂r and ∂A1 /∂r using the boundary conditions (A 8) and (A 10)
therefore gives
 2π        a2
k daj 2 ΛD0 A0 2 D0 ΛA0 2 daj ∂ U0 ∂V1
−i A 0 ± D ± − rhs D 0 dθ
0 hs 2 dS iωZj iωZj dS ∂r hs ∂r a1
 2π a2  2   
A0 Λ ∂ D0 kA0 2
=i rhs D0 D +r dr dθ. (A 11)
0 a1 C0 2 ∂S hs
Note that, since ∇ · (DU) = 0, for any function f (S, r, θ), ∇ · (DUf ) = εDDf , and
hence
∂ ∂ ∂
rhs D0 Df = r (D0 U0 f ) + (rhs D0 V1 f ) + (hs D0 W1 f ).
∂S ∂r ∂θ
Using this to eliminate the D operator from (A 11), substituting V1 = (daj /dS)(U0 / hs )
from the mean flow boundary conditions at r = aj and rearranging gives
 2π a2  
∂ D0 kA0 2 A0 2 D0 U0 Λ
Q1 + Q2 + + r dr dθ
0 a1 ∂S hs C0 2
 2π   a2
daj D0 kA0 2 A0 2 D0 U0 Λ
+ r + dθ = 0, (A 12)
0 dS hs C0 2 a1

where, setting fj = ΛD0 2 A0 2 r/(iωZj ),

 2π     
∂ ∂ daj ∂ U0 ∂V1 
Qj = (U0 fj ) + (hs V1 fj ) + hs fj − dθ  ,
0 ∂S ∂r dS ∂r hs ∂r r=aj

where a term involving W1 integrates to give zero owing to 2π periodicity. Expanding

the ∂/∂r terms and again using V1 = (daj /dS)(U0 / hs ) gives
 2π 
∂ daj ∂ 
Qj = (U0 fj ) + (U0 fj ) dθ 
0 ∂S dS ∂r r=aj
  
d 2π 2
U0 ΛD0 A0 2 
= rdθ  . (A 13)
dS 0 iωZj r=aj
 ! "
Ij (S)

Hence, (A 12) may be rearranged, moving the ∂/∂S derivative to the other side of the
integrals as we just have done for Qj above, to get finally
  2π a2    
d ωU0 k U0 2
I1 + I2 + D0 A0 2
+ 1 − 2 rdrdθ = 0,
dS 0 a1 C0 2 hs C0
 ! "
F (S)

with I1,2 defined in (A 13). Setting A0 (S, r, θ) = Â0 (S, r, θ)N(S) gives the secularity
condition as written in (3.6)

REFERENCES
Abramowitz, M. & Stegun, I. A. 1964 Handbook of Mathematical Functions, 9th edn. Dover.
Adamou, A. T. I., Gridin, D. & Craster, R. V. 2005 Acoustic quasi-modes in slowly varying
cylindrical tubes. Q. J. Mech. Appl. Maths 58, 419–438.
412 E. J. Brambley and N. Peake
Anderson, E., Bai, Z., Bischof, C., Blackford, S., Demmel, J., Dongarra, J., Du Croz, J.,
Greenbaum, A., Hammarling, S., McKenney, A. & Sorensen, D. 1999 LAPACK Users’
Guide. Society for Industrial and Applied Mathematics.
Atassi, H. M., Fang, J. & Patrick, S. 1993 Direct calculation of sound radiated from bodies in
nonuniform flow. Trans. ASME: J. Fluids Engng 115, 573–579.
Babic, V. M. & Buldyrev, V. S. 1991 Short-Wavelength Diffraction Theory. Springer.
Blokhintzev, D. 1946 The propagation of sound in an inhomogeneous and moving medium I.
J. Acoust. Soc. Am. 18, 322–328.
Boyd, J. P. 2001 Chebyshev and Fourier Spectral Methods, 2nd edn. Dover.
Brambley, E. J. & Peake, N. 2006 Classification of aeroacoustically relevant surface modes in
cylindrical lined ducts. Wave Motion 43, 301–310.
Cooper, A. J. & Peake, N. 2000 Trapped acoustic modes in aeroengine intakes with swirling flow.
J. Fluid Mech. 419, 151–175.
Cooper, A. J. & Peake, N. 2001 Propagation of unsteady disturbances in a slowly varying duct
with mean swirling flow. J. Fluid Mech. 445, 207–234.
Felix, S. & Pagneux, V. 2001 Sound propagation in rigid bends: A multimodal approach. J. Acoust.
Soc. Am. 110, 1329–1337.
Felix, S. & Pagneux, V. 2002 Mulitmodal analysis of acoustic propagation in three-dimensional
bends. Wave Motion 36, 157–168.
Felix, S. & Pagneux, V. 2004 Sound attenuation in lined bends. J. Acoust. Soc. Am. 116, 1921–1931.
Germano, M. 1982 On the effect of torsion on a helical pipe flow. J. Fluid Mech. 125, 1–8.
Goldstein, M. E. 1978 Unsteady vortical and entropic distortions of potential flows round arbitrary
obstacles. J. Fluid Mech. 89, 433–468.
Gridin, D. & Craster, R. V. 2003 Quasi-modes of a weakly curved waveguide. Proc. R. Soc.
Lond. A 459, 2909–2931.
Hinch, E. J. 1991 Perurbation Methods. Cambridge.
Keefe, D. H. & Benade, A. H. 1983 Wave propagation in strongly curved ducts. J. Acoust. Soc.
Am. 74, 320–332.
Keller, J. B. & Rubinow, S. I. 1960 Asymptotic solution of eigenvalue problems. Ann. Phys. 9,
24–75.
Khorrami, M. R., Malik, M. R. & Ash, R. L. 1989 Application of spectral collocation techniques
to the stability of swirling flows. J. Comput. Phys. 81, 206–229.
Menzies, R. D. D. 2002 Investigation of S-shaped intake aerodynamics using computational fluid
dynamics. PhD thesis, University of Glasgow.
Myers, M. K. 1980 On the acoustic boundary condition in the presence of flow. J. Sound Vib. 71,
429–434.
Ovenden, N. C. 2005 A uniformly valid multiple-scales solution for cut-on cut-off transition of
sound in flow ducts. J. Sound Vib. 286, 403–416.
Pagneux, V., Amir, N. & Kergomard, J. 1996a A study of wave propagation in varying cross-
section waveguides by modal decomposition. Part I. Theory and validation. J. Acoust. Soc.
Am. 100, 2034–2048.
Pagneux, V., Amir, N. & Kergomard, J. 1996b A study of wave propagation in varying cross-section
waveguides by modal decomposition. Part II. Results. J. Acoust. Soc. Am. 101, 2504–2517.
Rienstra, S. W. 1999 Sound transmission in slowly varying circular and annular lined ducts with
flow. J. Fluid Mech. 380, 279–296.
Rienstra, S. W. 2003a A classification of duct modes based on surface waves. Wave Motion 37,
119–135.
Rienstra, S. W. 2003b Sound propagation in slowly varying lined flow ducts of arbitrary cross-
section. J. Fluid Mech. 495, 157–173.
Rienstra, S. W. & Eversman, W. 2001 A numerical comparison between the multiple-scales and
finite-element solution for sound propagation in lined flow ducts. J. Fluid Mech. 437, 367–384.
Ting, L. & Miksis, M. J. 1983 Wave propagation through a slender curved tube. J. Acoust. Soc.
Am. 74, 631–639.