Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
837
Atmospheric Sound Propagation Over Large-Scale Irregular Terrain
Martin Almquist1 , Ilkka Karasalo
2 and Ken Mattsson
1
1Department of Information Technology, Uppsala University, Lägerhyddsvägen 2, 752 37 Uppsala, Sweden
2KTH Royal Institute of Technology, Stockholm, Sweden
Abstract
A benchmark problem on atmospheric sound propagation over irregular terrain has been solved using a stable
fourth-order accurate finite difference approximation of a high-fidelity acoustic model. A comparison with the
parabolic equation method and ray tracing methods is made. The results show that ray tracing methods can
potentially be unreliable in the presence of irregular terrain.
Keywords
Wave propagation Irregular terrain High-order finite difference methods
1 Introduction
High-fidelity simulations of sound propagation require an accurate treatment of the medium of propagation. In
the atmosphere the speed of sound typically varies in space, which causes refraction of sound rays. As sound
waves hit the surface of the earth they are partly reflected; the degree and direction of reflection depend on the
topography and the type of ground. Measurements have shown that in order to accurately predict sound pressure
levels (SPL), it is important to take the properties of both the atmosphere and the underlying terrain into account
(see for example [23, 35]). In realistic applications, the atmospheric variations can be complex and the terrain is
often irregular, which means both that the topography is non-trivial and that the type of ground can vary.
One high-fidelity model for sound propagation is based on the linearized Euler equations [36], which is most
often solved with the staggered grid finite difference time domain method originally presented in [44]. In [13] a
pseudo-spectral time domain method was used to solve the linearized Euler equations in urban courtyards. This
is an efficient approach when the underlying geometry fits into a block-Cartesian geometry.
Another high-fidelity model is the time-dependent acoustic wave equation. However, the wave equation has
been considered too computationally demanding in realistic large-scale 3-D settings. One way to reduce the
computational cost when the sound source consists of a single frequency is to Fourier transform the wave
equation, which yields the Helmholtz equation. From the Helmholtz equation, it is possible to further reduce the
computational cost via an assumption that reflections at boundaries are negligible. The resulting model is called
the parabolic equation (PE) model, often used in ocean acoustics (e.g. [5, 24, 37]) and aeroacoustics (e.g. [16–
18]). Because of the assumption about reflections, the PE method is only valid for very moderate topographies.
A significant drawback with the Helmholtz equation (and the PE method) is that time-dependent phenomena,
such as a varying atmosphere or sources (for instance wind turbines) can not easily be approximated.
However, despite the possible simplifications, the above models are all generally regarded as too
computationally expensive in realistic 3-D applications. Hence, most numerical methods for sound propagation
are based on much simpler models. A commonly used sound propagation model is NORD2000 [34], based on
ray tracing methods, which combine geometrical ray theory with the theory of diffraction (for examples of
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
838
applications see [7, 22, 42, 43]). Ray tracing methods are computationally cheap, but they incorporate neither a
complex atmosphere nor irregular terrain properly.
Unfortunately, due to a shortage of reliable measurements over large-scale domains, it has been difficult to
validate the sound propagation models currently used in practical applications. An alternative to measurements
is validation through manufactured solutions (where one compares the numerical results with analytical
solutions), but this technique has not yet been widely accepted outside the field of applied mathematics.
In this paper we shall solve a realistic sound propagation problem with the time-dependent acoustic wave
equation as the underlying model. As mentioned above, this model is desirable from an accuracy perspective,
but it is computationally costly. It is therefore imperative to use an efficient numerical method, to minimize the
number of unknowns. It is well known that high-order (higher than second order) spatially accurate finite
difference schemes combined with high-order accurate time marching schemes are very well suited for wave
propagation problems on large domains (see the pioneering paper by Kreiss and Oliger [20]). To guarantee an
accurate approximation, it is necessary that the underlying numerical scheme can be proven stable, which is a
non-trivial task using high-order finite difference methods.
A robust and well-proven high-order finite difference methodology that ensures stability of time-dependent
partial differential equations (PDEs) is the summation-by-parts-simultaneous approximation term (SBP-SAT)
method. The SBP-SAT method combines semi-discrete operators that satisfy a summation-by-parts (SBP)
formula [21], with physical boundary conditions implemented using the simultaneous approximation term
(SAT) method [3].
An added benefit of the SBP-SAT method is that it naturally extends to multi-block geometries while retaining
the essential single-block properties: stability, accuracy, and conservation [4]. Thus, problems involving
complex domains or non-smooth geometries are easily amenable to the approach. References [12, 27, 28,32]
report applications of the SBP-SAT method to problems involving nontrivial geometries. Most of the published
results for the SBP-SAT method are for first order hyperbolic systems (and the Navier–Stokes equations). The
extension of the SBP-SAT technique to the second order wave equation is found in [28–31]. However, those
studies were restricted to 1-D concerning high-order finite difference methods (3-D results for the finite volume
technique were presented). The present study is a direct extension of the SBP-SAT method to the second order
wave equation on multidimensional curvilinear domains, including non-trivial boundary conditions.
In the first part of this paper, the SBP-SAT method is applied to the benchmark problem introduced in [35]. The
SBP-SAT method is here extended to the second order wave equation on a curvilinear 2-D domain with non-
trivial boundary conditions. A fourth-order accurate SBP-SAT approximation is implemented and verified by a
grid-convergence study against manufactured solutions. In the second part of this paper, the benchmark problem
in [35] is used to compare the PE and ray tracing methods with the newly implemented SBP-SAT method.
Conclusions as to the validity of the PE and ray tracing methods in the presence of a complex atmosphere and
irregular terrain are drawn.
The rest of the paper is organized as follows: In Sect. 2 we introduce some definitions and illustrate the SBP-
SAT method by applying it to a 1-D problem. In Sect. 3 we analyze a model problem in 2-D. We then introduce
the benchmark problem in Sect. 4, and describe how the SBP-SAT technique has been adapted to this problem.
The PE method and the ray interpolation methods are introduced in Sects. 5 and 6, respectively. The
implementation of the fourth-order accurate SBP-SAT method is verified in a series of convergence studies in
Sect. 7. In Sect. 8, we present the results from the benchmark problem, comparing the different numerical
methods. Conclusions are presented in Sect. 9. The finite difference (SBP) operators are listed in “Appendix”.
2 The 1-D Problem
In this section we define the SBP-SAT method. To illustrate the power and simplicity of the method we shall
consider the following second-order hyperbolic problem:
autt=(bux)x,L0u=g0(t),L1u=g1(t),u=f1,ut=f2,0≤x≤1,x=0,x=1,0≤x≤1,t≥0,t≥0,t≥0,t=0,
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
839
(1)
where a(x)>0 and b(x)>0, and the boundary conditions are given by
L0uL1u=α0utt+α1ut+α2u+α3ux=g0(t),x=0=β0utt+β1ut+β2u+β3ux=g1(t),x=1.
(2)
In the present study we restrict ourselves to the case α3≠0,β3≠0, referred to as mixed boundary conditions. For
the treatment of Dirichlet conditions we refer to [29].
Before we start employing the SBP-SAT method, some definitions are needed. Let the inner product for real-
valued functions u,v∈L2[0,1] be defined by (u,v)=∫10uvc(x)dx,c(x)>0, and let the corresponding norm
be ∥u∥2c=(u,u). The domain (0≤x≤1) is discretized using the following N+1 equidistant grid points:
xi=ih,i=0,1,…,N,h=1N.
The approximate solution at grid point xi is denoted vi, and the discrete solution vector is v=[v0,v1,…,vN]T.
Similar to the continuous inner product, we define an inner product for discrete real-valued vector
functions u,v∈RN+1 by (u,v)Hc=uTHCv, where H is diagonal and positive definite andC is the projection
of c(x) onto the diagonal. The corresponding norm is ∥v∥2Hc=vTHCv.
Remark 1
The matrix product HC defines a norm if and only if HC is symmetric and positive definite. This can only be
guaranteed if H is a diagonal matrix (see [39] for a detailed study on this).
The following vectors will be frequently used:
e0=[1,0,…,0]T,eN=[0,…,0,1]T.
(3)
2.1 The SBP-SAT Method
SBP operators are essentially central finite difference stencils, closed at the boundaries with carefully chosen
one-sided difference stencils which mimic the underlying integration-by-parts formula in a discrete norm. In the
present paper we address the SBP operators by the accuracy of the central scheme and the type of norm which
they are based on. A 2pth order diagonal norm SBP operator is closed with pth order accurate one sided stencils
(see [30]). For first order hyperbolic problems, this implies that the convergence rate (i.e., global convergence)
drops to (p+1)th order when using a 2pth order diagonal norm SBP operator. For strongly parabolic problems
and second order hyperbolic problems the convergence rate instead drops to (p+2)th order (see [9, 40] for more
information on the accuracy of finite difference approximations).
To define the SBP-SAT method, we present Definitions 1–2 (first stated in [33] and [28]). We here say that a
scheme is explicit if no linear system of equations needs to be solved to compute the difference approximation.
Definition 1
A difference operator D1=H−1Q=H−1(Q¯−12e0eT0+12eNeTN) approximating ∂/∂x, using a 2pth-order
accurate stencil of minimal width, is said to be a 2pth-order accurate narrow-diagonal first-derivative SBP
operator if H is diagonal and positive definite and Q¯+Q¯T=0.
Definition 2
Let D(b)2=H−1(−M(b)−e0b0S0+eNbNSN) approximate ∂/∂x(b∂/∂x), where b(x)>0, using a 2pth-order accurate
stencil of minimal width. D(b)2 is said to be a 2pth-order accurate narrow-diagonal second-derivative SBP
operator, if H is diagonal and positive definite, M(b) is symmetric and positive semi-definite
and S0 and SN approximate the first-derivative operator at the boundaries.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
840
The superscript (b) emphasizes that M(b) and D(b)2 depend on b(x). The explicit dependence can be found
in [26]. For completeness we have included the fourth order SBP operators (used in the present study) in
“Appendix”.
The following definition introduced in [26] is also central in this paper:
Definition 3
Let D1 and D(b)2 be 2pth-order accurate narrow-diagonal first- and second-derivative SBP operators.
If M(b)=DT1HBD1+R(b), and the remainder R(b) is positive semi-definite, D1 and D(b)2 are calledcompatible.
2.1.1 Continuous Analysis
Multiplying the first equation in (1) by ut and integrating by parts (referred to as “the energy method”) leads to
ddt(∥ut∥2a+∥ux∥2b)=2(buxut)x=1−2(buxut)x=0,
(4)
where (buxut)x=1 means (buxut) evaluated at x=1. We also identify
E=∥ut∥2a+∥ux∥2b,
(5)
as the total energy (kinetic and potential).
Multiplying the first equation in (1) by ut, integrating by parts and imposing the boundary conditions (2) leads
to
ddtE¯=BTx=0+BTx=1,
(6)
where
E¯=∥ut∥2a+∥ux∥2b+b(0)α3(α0u2t+α2u2)x=0−b(1)β3(β0u2t+β2u2)x=0,
(7)
and
BTx=0BTx=1=+2b(0)α1α3(ut−g02α1)2x=0−g20b(0)2α1α3=−2b(1)β1β3(ut−g12β1)2x=1+g21b(1)2β1β3.
(8)
Here we assume that α3≠0 and β3≠0. The following Lemma is central in the present study,
Lemma 1
Equation (1) with boundary conditions (2) has a bounded energy in terms of initial and boundary data
if α3≠0,β3≠0,α0α3≥0,α2α3≥0,β0β3≤0,β2β3≤0,α1α3<0 and β1β3>0 hold.
Proof
E¯ is non-negative and well defined if α3≠0,β3≠0,α0α3≥0,α2α3≥0,β0β3≤0,β2β3≤0 hold. By integrating (6) in
time, we obtain
E¯(t)+∫0t⎛⎝2b(1)β1β3(ut−g1(τ)2β1)2x=1−2b(0)α1α3(ut−g0(τ)2α1)2x=0⎞⎠dτ=E¯(0)+∫0t(g21(τ)b(1)2β1β3−g2
0(τ)b(0)2α1α3)dτ.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
841
If α1α3<0 and β1β3>0 hold we have a strong estimate of E¯(t) in terms of initial and boundary data.
Remark 2
The implication of Lemma 1 is that we can have at most linear time-growth of ∥u∥. Linear growth (in t) does not
violate well-posedness (see [10]). However, linear time-growth of ∥u∥ can only occur if we have f2≠0combined
with pure Neumann boundary conditions, i.e, α0=α1=β0=β1=0, so that we have a zero eigenvalue in the
spectrum [25].
2.1.2 Semi-Discrete Analysis
The discrete approximation of (1) using the SBP-SAT method is
Avtt=D(b)2v+τ0H−1e0(L0v−g0)+τ1H−1eN(L1v−g1),
(9)
where e0 and eN are defined in (3). (We assume the same initial conditions v=f1,vt=f2 as in the continuous
case). The matrix A has the values of a(x) injected on the diagonal. The semi-discrete boundary operators that
mimic (2) are given by
L0vL1v=α0(vtt)0+α1(vt)0+α2v0+α3S0v=β0(vtt)N+β1(vt)N+β2vN+β3SNv.
(10)
Applying the energy method by multiplying (9) by vTtH and adding the transpose leads to
ddtEH=−2b0(vt)0S0v+2bN(vt)NSNv+2τ0(α0(vt)0(vtt)0+α1(v2t)0)+2τ0(α2(vt)0v0+α3(vt)0S0v−(vt)0g0)+2τ1(β0(v
t)N(vtt)N+β1(v2t)N)+2τ1(β2(vt)NvN+β3(vt)NSNv−(vt)Ng1),
(11)
where
EH=∥vt∥2Ha+vTM(b)v.
Lemma 2
Equation (9) with boundary operators (10) exactly mimics the continuous energy estimate (6)
if τ0=b0α3,τ1=−bNβ3, and is thus stable if the conditions in Lemma 1 hold.
Proof
insert τ0=b0α3,τ1=−bNβ3 in (11) to obtain
ddtE¯H=BT0+BTN,
(12)
where
E¯H=∥vt∥2Ha+vTM(b)v+b0α3(α0(v2t)0+α2(v2)0)−bNβ3(β0(v2t)N+β2(v2)N),
and
BT0BTN=+2b0α1α3((vt)0−g02α1)2−g20b02α1α3=−2bNβ1β3((vt)N−g12β1)2+g21bN2β1β3.
Equation (12) is a semi-discrete analogue to (6), and stability follows if the conditions in Lemma 1 hold.
3 Analysis in 2D
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
842
In this section we analyze the scalar 2-D wave equation with mixed boundary conditions. To allow for complex
domains, we transform the equation given on a curvilinear domain to an equation on the unit square. We then
derive an energy estimate for the continuous case. After discretizing the model in space with the SBP-SAT
method, we prove stability by exactly mimicking the continuous energy estimate.
3.1 Definitions
To make the notation more compact we introduce the Kronecker product:
C⊗D=⎡⎣⎢⎢c0,0D⋮cp−1,0D⋯⋯c0,q−1D⋮cp−1,q−1D⎤⎦⎥⎥,
(13)
where C is a p×q matrix and D is an m×n matrix. We also let IN be the (N+1)×(N+1) identity matrix.
If the problem is given on a curvilinear domain Ω (referred to as the physical domain) we transform it to the unit
square, Ω′. The unit square is discretized using the (Nξ+1)(Nη+1) grid points:
(ξi,ηj)=(iNξ,jNη),i=0,1,…,Nξ,j=0,1,…,Nη.
The boundaries of Ω′ are denoted by W (west), N (north), E (east) and S (south), respectively, as shown in Fig. 1.
The approximate solution at a grid point (ξi,ηj) is denoted by vij, and the discrete solution vector
is v=[v00,…,v0Nη,v10,…,vNξNη]T. The matrix RW is defined so that RWv is a vector with the same length
as v and the same elements on the positions corresponding to the west boundary, but zeros everywhere else. The
matrices RN,RE and RS are defined similarly for the north, east and south boundaries, respectively.
Fig. 1
The mapping between cartesian (left) and curvilinear (right) coordinates
By D1ξ we denote the 2-D version of the narrow-stencil first-derivative SBP operator D1, approximating ∂∂ξ.
Similarly, D(b)2ξ approximates ∂∂ξ(b∂∂ξ). In the same manner, we let Hξ denote the 2-D version of the
diagonal matrix H, applied in the ξ-direction. D1η,D(b)2η and Hη are defined similarly for the η-direction.
To simplify the notation (without any restriction) we here assume Nξ=Nη=N. The 2-D operators can be neatly
expressed in terms of the 1-D operators using the Kronecker product:
D1ξHξEWRWEERE=D1⊗IN,D1η=IN⊗D1=H⊗IN,Hη=IN⊗H=e0⊗IN,ES=IN⊗e0=EWETW,RS=ESETS=eN
⊗IN,EN=IN⊗eN,=EEETE,RN=ENETN,
(14)
where the vectors e0 and eN are defined in (3). Assuming that the coefficient b is constant, we can also write
D(b)2ξ=D(b)2⊗IN,D(b)2η=IN⊗D(b)2.
(15)
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
843
In the case of a variable coefficient b, however, Eq. (15) does not hold. To cover also that case, we introduce the
notation
b(ξ)i(η)=b(ξi,η),b(η)i(ξ)=b(ξ,ηi),i=0,1,…,N.
(16)
We also define
d(b)ij=⎡⎣⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢(D(b(η)0)2)i,j(D(b(η)1)2)i,j⋱(D(b(η)N)2)i,j⎤⎦⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥.
(17)
We then have
D(b)2ξ=⎡⎣⎢⎢⎢⎢⎢⎢⎢⎢d(b)11d(b)21⋮d(b)(N+1)1d(b)12d(b)22⋮d(b)(N+1)2⋯⋯⋱⋯d(b)1(N+1)d(b)2(N+1)⋮d(b)(N+1)(N
+1)⎤⎦⎥⎥⎥⎥⎥⎥⎥⎥
(18)
and
D(b)2η=⎡⎣⎢⎢⎢⎢⎢⎢⎢⎢⎢D(b(ξ)0)2D(b(ξ)1)2⋱D(b(ξ)N)2⎤⎦⎥⎥⎥⎥⎥⎥⎥⎥⎥.
(19)
3.2 The Continuous Problem
We consider the following problem:
utt=bΔuγ1u+γ2∇u⋅n+γ3ut=0u=f1,ut=f2,(x,y)∈Ω,(x,y)∈∂Ω,(x,y)∈Ω,t≥0t≥0t=0,
(20)
where b(x,y)>0. We have chosen homogeneous boundary conditions to avoid unnecessary notation in the
analysis. Similarly to the 1-D analysis in Sect. 2.1, the analysis holds for inhomogeneous conditions as well. We
also limit our present study to the case γ2≠0, which includes the important case of Neumann
conditions (γ1=0,γ2=1,γ3=0).
We can add dissipation to (20) by adding a term b∇⋅(σ∇ut),σ(x,y)≥0 to the right hand side of the PDE. The added
dissipation term will be used to create absorbing layers at artificial boundaries in Sect. 4.2. Including the
dissipation term, the problem reads
utt=bΔu+b∇⋅(σ∇ut)γ1u+γ2∇u⋅n+γ3ut=0u=f1,ut=f2,(x,y)∈Ω,(x,y)∈∂Ω,(x,y)∈Ω,t≥0t≥0t=0.
(21)
We now transform the problem to the unit square. Assume that there is a smooth one-to-one mapping
{x=x(ξ,η)y=y(ξ,η),
from Ω′ to Ω. The Jacobian J of the transformation is
J=xξyη−xηyξ.
The scale factors η1 and η2 of the transformation are defined as
η1=x2ξ+y2ξ−−−−−−√,η2=x2η+y2η−−−−−−√.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
844
(22)
Since the mapping is one-to-one, the Jacobian is everywhere non-zero. By the chain rule, we have
{uξ=uxxξ+uyyξuη=uxxη+uyyη,
which is equivalent to
⎧⎩⎨⎪⎪⎪⎪ux=1J(uξyη−uηyξ)=1J((uyη)ξ−(uyξ)η)uy=1J(uηxξ−uξxη)=1J((uxξ)η−(uxη)ξ).
(23)
Replacing u with ux and uy in (23) yields
uxxuyy=1J(1J(uξyη−uηyξ)yη)ξ−1J(1J(uξyη−uηyξ)yξ)η=1J(1J(uξxη−uηxξ)xη)ξ−1J(1J(uξxη−uηxξ)xξ)η.
(24)
By adding uxx and uyy and rearranging terms, the first equation in (21) can be written as
J~utt=Δ~u+Δ~σut,(ξ,η)∈Ω′,
(25)
where we have defined
Δ~uΔ~σuα1=(α1uξ)ξ+(βuξ)η+(βuη)ξ+(α2uη)η,=(σα1uξ)ξ+(σβuξ)η+(σβuη)ξ+(σα2uη)η,=1J(x2η+y2η),β=−1J(xηx
ξ+yηyξ),α2=1J(x2ξ+y2ξ),
and
J~=Jb.
Using Eq. (23) to transform ∇u⋅n in the second equation in (21) yields the transformed boundary conditions:
⎧⎩⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪γ1η2u−γ2(α1uξ+βuη)+γ3η2ut=0,γ1η2u+γ2(α1uξ+βuη)+γ3η2ut=0,γ1η1u−γ2(α2uη+βuξ)+γ3
η1ut=0,γ1η1u+γ2(α2uη+βuξ)+γ3η1ut=0,(ξ,η)∈W(ξ,η)∈E(ξ,η)∈S(ξ,η)∈N.
(26)
The complete transformed problem is given by (25), (26) and the initial conditions stated in (21). Applying the
energy method (here assuming that also the time derivative of the boundary condition (26) holds) leads to
ddtE=−∫Wγ3+σγ1γ2η2u2tdη−∫Eγ3+σγ1γ2η2u2tdη−∫Nγ3+σγ1γ2η1u2tdξ−∫Sγ3+σγ1γ2η1u2tdξ−∫Ω′[utξutη]T[σα1σβ
σβσα2][utξutη]dΩ′,
(27)
where
E=12⎛⎝⎜∫Ω′J~u2tdΩ′+∫Ω′[uξuη]T[α1ββα2][uξuη]dΩ′+BT⎞⎠⎟,
(28)
and
BT=∫Wγ1γ2η2u2dη+∫Eγ1γ2η2u2dη+∫Nγ1γ2η1u2dξ+∫Sγ1γ2η1u2dξ+∫Wσγ3γ2η2u2tdη+∫Eσγ3γ2η2u2tdη+∫Nσγ3γ2η1
u2tdξ+∫Sσγ3γ2η1u2tdξ.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
845
(29)
The matrix [α1ββα2] is positive definite since α1>0 and α1α2−β2=(xξyη−xηyξ)2=J2>0. Thus, the problem has
an energy estimate if the relations
γ1γ2≥0,γ3γ2≥0
(30)
hold. The last term in (27) implies damping of the energy for σ>0.
3.3 The Semi-Discrete Problem
In the semi-discrete setting we use the following notation for the matrices corresponding to the continuous
variable coefficients, for readability purposes: If λ denotes a variable coefficient in the continuous setting, we
here denote the matrix with the values of λ(ξ,η) at the grid points injected on the diagonal by λ. There is no risk
of confusion since it will always be clear from context whether we are in a continuous or semi-discrete setting.
The semi-discrete version of (26) is given by
⎧⎩⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪LWv=RW{γ1η2v−γ2(α1Sξv+βD1ηv)+γ3η2vt}=0LEv=RE{γ1η2v+γ2(α1Sξv+βD1ηv)+γ3η2v
t}=0LSv=RS{γ1η1v−γ2(α2Sηv+βD1ξv)+γ3η1vt}=0LNv=RN{γ1η1v+γ2(α2Sηv+βD1ξv)+γ3η1vt}=0.
(31)
In the numerical scheme we also impose the time-derivative of the boundary conditions when σ>0. For instance,
if we have the boundary condition Lv=f, we impose both Lv=f and Lvt=ft using the SAT technique.
The semi-discrete approximation of (25) and (26) using the SBP-SAT method is
J~vtt=D(α1)2ξv+D1ξβD1ηv+D1ηβD1ξv+D(α2)2ηv+D(σα1)2ξvt+D1ξσβD1ηvt+D1ησβD1ξvt+D(σα2)2ηvt+τ1H−
1ξLWv+τ1H−1ξLEv+τ2H−1ηLSv+τ2H−1ηLNv+τ3σH−1ξLWvt+τ3σH−1ξLEvt+τ4σH−1ηLSvt+τ4σH−1ηLNvt.
(32)
The first main result of the present study is stated in the following theorem:
Theorem 1
The scheme (32) is stable if τ1=τ2=τ3=τ4=−1γ2 and (30) holds.
Proof
Applying the energy method by multiplying (32) by vTtHξHη and adding the transpose leads to
ddtEH=vTt(1+τ1γ2)Hηα1(−RW+RE)Sξv+vTt(1+τ2γ2)Hξα2(−RS+RN)Sηv+vTt(1+τ3γ2)Hησα1(−RW+RE)Sξvt+vT
t(1+τ4γ2)Hξσα2(−RS+RN)Sηvt+vTt(1+τ1γ2)Hη(−RW+RE)βD1ηv+vTt(1+τ2γ2)Hξ(−RS+RN)βD1ξv+vTt(1+τ3γ2)
Hη(−RW+RE)σβD1ηvt+vTt(1+τ4γ2)Hξ(−RS+RN)σβD1ξvt+vTt(τ1γ3+τ3σγ1)Hηη2(RW+RE)vt+vTt(τ2γ3+τ4σγ1)H
ξη1(RS+RN)vt+vTtHηM(σα1)ξvt−vTtHξM(σα2)ηvt+2(D1ξvt)TσβHξHη(D1ηvt),
where
EH=12vTtHξHηJ~vt+12(vTHηM(α1)ξv+vTHξM(α2)ηv+2(D1ξv)TβHξHη(D1ηv))+12(−vTτ1γ1Hηη2(RW+RE)v−v
Tτ2γ1Hξη1(RS+RN)v)+12(−vTtτ3σγ3Hηη2(RW+RE)vt−vTtτ4σγ3Hξη1(RS+RN)vt).
By choosing τ1=τ2=τ3=τ4=−1γ2 we obtain an energy estimate completely analogous to (27). If (30) holds, we
have a non-growing energy. □
4 Model Problem
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
846
We begin this section by stating the problem, which can be solved by any appropriate method, in Sect. 4.1. In
Sect. 4.2 we derive a time-domain model of the problem. Section 4.3 is devoted to the details of applying the
SBP-SAT method to this model.
4.1 Problem Description
The benchmark problem introduced in [35] consists of the following components:
A 3-D topography, which is assumed to be axisymmetric. Figure 2 shows the vertical cross section of
the topography, which is available for download at [41].
A point source emitting spherical waves with a frequency of 50 Hz, located at range r=0 m,
height z=10 m.
Constant sound speed c=340 m/s.
The boundary condition at the ground, which in frequency domain is described by Eq. 15 in [35]. In
time domain, it is described by (37)–(39).
The objective is to compute the propagation loss P1(r) 1 meter above the ground, for range r< 2,000 m. Let A be
the amplitude of the point source and let u1(r) be the amplitude of the sound pressure field 1 m above ground at
range r. The definition of propagation loss yields
P1(r)=−20log(u1(r)A).
(33)
The propagation losses obtained with the PE approach in [35] and with the SBP-SAT method in the present
paper are available for download at [41].
Fig. 2
The topography
4.2 The SBP-SAT Model
When deriving the SBP-SAT model, we assume that propagation of sound waves is governed by the acoustic
wave equation
utt=bΔu,
(34)
where u is the acoustic pressure and b is the square of the wave velocity. As in Sect. 3.2, we introduce
dissipation,
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
847
utt=bΔu+b∇⋅(σ∇ut).
(35)
Expressing Eq. (35) in cylindrical coordinates (r,ϕ,z) and assuming symmetry in the azimuthal direction (the ϕ-
direction) results in the axisymmetric two-dimensional restriction of (35),
1butt=1r(rur)r+uzz+1r(σrur)r+(σuz)z.
(36)
We construct the physical domain by introducing artificial boundaries in the manner shown in Fig. 3. In the
simulations, the west boundary was placed at r=1 m and the north boundary at heights ranging
from z=300 to z=750 m. The dot at the z-axis marks the location of the source, just outside the west boundary of
the domain.
Fig. 3
Qualitative description of the physical domain. The dot at the z-axis represents the point source
The boundary condition at the south boundary is a locally-reacting impedance condition given by (see [35])
(pω0c−χ2)u+∇u⋅n+qcut=0,
(37)
where c is the wave speed, n is the unit outward normal, ω0 is the angular frequency of the source, χ is the
curvature and p and q are real numbers that satisfy the relation
p+qi=iZ^,
(38)
where Z^ is the normalized sound impedance, Z^=18.3+17.5i. The curvature χ is defined as
χ(r)=−Hrr(r)(1+H2r(r))3/2,
(39)
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
848
where H(r) denotes the height of the ground at horizontal position r.
The north and east boundaries are artificial boundaries that are introduced to truncate the unbounded domain.
We need to avoid reflections at these boundaries. Absorption of waves at artificial boundaries is an important
numerical issue. One possible approach is to apply an absorbing boundary condition (ABC), for example the
first order Engquist Majda ABC [8],
⎧⎩⎨⎪⎪⎪⎪(ut+b√∇u⋅n)|z=zmax=0,(ut+b√∇u⋅n+b√2ru)|r=rmax=0.
(40)
However, a first order Engquist Majda ABC is perfectly absorbing only at normal incidence. At 45∘ incidence
on a straight boundary, 17 % of the incoming wave is reflected, and close to glancing the reflection coefficient
tends to unity. Another, more efficient approach is to introduce an absorbing layer (AL) close to the artificial
boundary. This can be done by smoothly increasing the dissipation coefficient σ from zero to some fixed value.
The dissipation must increase quickly enough that the waves are damped efficiently inside the layer, but it must
also increase smoothly enough that we avoid reflections at the interface between the AL and the interior domain.
The wider one can afford to make the layer, the easier it is to find a function σ that fulfills both criteria. This AL
is a special case of the method presented by Appelö and Colonius [1]. Appelö and Colonius also slowed down
waves inside the layer by stretching the grid at the boundaries and included higher-order dissipation operators
for better performance. Since the focus in the present study is not on optimal absorbing layers, we here settle for
a simple version of their AL. We will verify that the truncation of the domain does not affect the solution by
placing the artificial boundaries at different locations in the simulations.
At the west boundary, the boundary condition is determined by the source. Consider a point source with
amplitude A and frequency f. At a distance r~ from the source, the acoustic pressure is given by
u(s)(r~,t)=Ar~sin(2πf(t−r~c)).
(41)
Now let the source be located at (r,z)=(0,z0). The distance r~ from the source to a point (r,z) is given by
r~=r2+(z−z0)2−−−−−−−−−−√.
(42)
Combining (41) and (42) yields
u(s)(r,z,t)=Ar2+(z−z0)2−−−−−−−−−−√sin⎛⎝⎜2πf⎛⎝⎜t−r2+(z−z0)2−−−−−−−−−−√c⎞⎠⎟⎞⎠⎟.
(43)
We will impose this boundary data on the west boundary. This can be done using a Dirichlet boundary condition
by extending the SAT technique developed in [29]. However, in the present study we instead impose the data
using a mixed boundary condition (which allows for a stronger energy estimate in terms of boundary data),
ut+b√∇u⋅n+b√2ru=u(s)t+b√∇u(s)⋅n+b√2ru(s),r=rmin.
(44)
We now introduce logical coordinates (ξ,η) and perform a transformation from the unit
square (0≤ξ≤1,0≤η≤1) onto the physical domain (r(ξ,η),z(ξ,η))∈Ω. Equation (36) transforms into
J~utt=Δ~u+Δ~σut,(ξ,η)∈Ω′,
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
849
(45)
where we have defined
J~Δ~uΔ~σu=rJb,=(α1uξ)ξ+(βuξ)η+(βuη)ξ+(α2uη)η,=(σα1uξ)ξ+(σβuξ)η+(σβuη)ξ+(σα2uη)η,
and
α1=rJ(r2η+z2η),β=−rJ(rξrη+zξzη),α2=rJ(r2ξ+z2ξ).
To summarize, the model that we solve with the SBP-SAT method is the Eq. (45) with the boundary conditions
(37) and (44) at the south and west boundaries, and either the ABC (40) or the AL at the north and east
boundaries. In the ABC approach, σ is identically zero. In the AL approach, σ is non-zero close to the north and
east boundaries. The Eq. (45) has the same form as (25), and all the boundary conditions (37), (44) and (40) are
of the mixed type analyzed in Sect. 3. Thus, the analysis performed in Sect. 3, proving well-posedness for the
continuous problem and stability for the discrete scheme, holds for this model too.
4.3 Implementation Details
We have implemented a fourth-order SBP-SAT method of the model problem. The spatial discretization is thus
fourth-order accurate in the interior scheme and second-order accurate in the boundary closures. For
completeness we have included the operators (first presented in [26]) in “Appendix”. The classical fourth-order
accurate Runge–Kutta method was used for discretization in time.
In order to apply the solver to the model problem, a computational grid must be constructed in the physical
domain. Generating a good grid on a complex domain is not a trivial task. If the grid is not smooth enough, the
convergence rate will decrease. In this case, the terrain profile has only two continuous derivatives, and hence
we suspect that the grid will not support high-order accuracy. We have used Pointwise, a commercial software
for creating grids. Figure 4 shows an example of a coarse grid generated in Pointwise. Table 1 lists the number
of gridpoints required when using 6, 9 and 12 points per acoustic wavelength, with the north boundary at a
height of 500 m.
Fig. 4
Coarse example grid
Table 1
Number of grid points corresponding to different resolutions. North boundary at z=500 m
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
850
Resolution (points per wavelength) Nr Nz Nr×Nz
6 1,801 451 0.81×106
9 2,701 676 1.83×106
12 3,601 901 3.24×106
In the simulations, we time-advanced until the solution became periodic in time and then computed the
amplitude of the sound waves by measuring |v|max, the maximum absolute value of the solution during one
period, one meter above ground. The propagation loss P (measured in dB) was computed as
P=−20log(|v|maxA),
(46)
where A is the amplitude of the point source.
5 The Parabolic Equation Method
The PE method used in the model problem of Sect. 8 is described briefly below. More details can be found
in [19].
As an initial step, a smooth approximation h(r) of the ground height as function of range is computed from the
data, using an interpolating or a variance-reducing B-spline expansion [6, Ch. XI], the choice depending on the
smoothness of the data. The geometry is then mapped from the physical (r,z) domain to a rectangle in
the (ξ,η) plane by an orthogonal curvilinear transform
r=r(ξ,η),z=z(ξ,η),rξrη+zξzη=0,
(47)
such that η is constant along the boundaries of the computational domain. Assuming cylindrical symmetry, the
Helmholtz equation for the complex pressure u(ξ,η) is
f(f−1uξ)ξ+f(g−1uη)η+k2a2u=0,
(48)
where, using the unit dB/wavelength for the attenuation α,
fa=ρa/rb,g=ρb/ra,=(r2ξ+z2ξ)1/2,b=(r2η+z2η)1/2,k=ωc(1+iαlog(10)40π).
The PE approximations are derived by writing (48) in the form
T2u=(1−L)u−k−20Ru,
(49)
where k0 is a reference wavenumber, T and L are differential operators, and R a function:
Tu=−ik−10f1/2(f−1/2u)ξ,
(50)
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
851
Lu=−k−20[f(g−1(uη))η+k2a2u]+u,
(51)
R=(3f2ξ−2ffξξ)/(4f2).
(52)
Discarding the term Ru, which is small since R∼ξ−2 as ξ grows, Eq. (49) is simplified to
T2u=(1−L)u.
(53)
The PE schemes compute one-way solutions to (53) by solving
Tu=1−L−−−−√
(54)
with the pseudo differential operator 1−L−−−−√ replaced by a rational function of L,
Tu=Pm(L)Qn(L),
(55)
analogously with [2]. Pm and Qn are polynomials of degrees m and n in the Padé approximation
1−x−−−−√=Pm(x)/Qn(x)+O(xm+n+1),x→0.
(56)
Thus the JEPE PE-approximations are
u=Qn(L)v,
(57)
TQn(L)v=Pm(L)v,m=max(n,1),n=0,1,2,…
(58)
Increasing the Padé order n reduces the phase error as function of elevation angle, but also increases the
computational work. In practice, Padé orders n=0,1,2 are the most frequently used and correspond to the narrow
angle (15∘), the wide angle (35∘) and the very wide angle (55∘) approximations, cf. [15, Sec. 6.2.4].
Equation (58) with initial conditions at ξ=ξ0 and boundary conditions at η=0 (the upper boundary)
and η=−H (the ground) is solved using the method of lines. Thus, u,T,L and the boundary conditions are
discretized vertically using a centered second-order finite difference scheme [38, Sec. 9]. The vertically
discretized form of Eq. (58) is a system of ODEs (omitting the indices m and n)
ddξDQ(L)w=ik0D[P(L)−Q(L)]ww(ξ0)=w0
(59)
for the scaled and wavenumber-shifted complex pressure
w(ξ)=e−ik0ξ1/2v(ξ).
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
852
(60)
D and L are diagonal and tri-diagonal matrix-valued functions of ξ, respectively, with D real and the imaginary
part of L diagonal and non-positive. The initial profile w0 is computed from the given source data (height and
vertical directivity), by low-pass filtering w.r.t. vertical wave number to the validity interval of the PE scheme.
Equation 59 is then solved by a two-step fourth order A-stable second derivative method by Jeltsch (method J4
in [14]).
6 The Ray Interpolation Method
In the ray interpolation method the ground height is described by a smooth function of range identical to that in
the PE method described in Sect. 5. A ray trajectory (r(s),z(s)) where s is arc length, is a solution to the ODE
system [15, Sec. 3.2.1]
dr/dsdz/dsdϕ/ds=cos(ϕ)=sin(ϕ)={sin(ϕ)∂c/∂r−cos(ϕ)∂c/∂z}/c.
(61)
c=c(r,z) is the sound speed and ϕ=ϕ(s) the elevation angle of the ray. In the high-frequency limit, the ray
trajectories are streamlines of the acoustic intensity field i.e., propagation paths of acoustic energy. The
wavefield at a point (r,z) is then a sum of contributions from all rays passing through (r,z)—all eigenraysfrom
the source to (r,z)). Each eigenray contributes to the sum with the field inside an infinitesimal tube surrounding
the ray. With a mono-frequency monopole source with amplitude P0 at 1 m range, the value of the ray-tube field
along an eigenray (r(s),z(s)) is
P(f,s)=P0α(s)ei2πfτ(s)eiπnc(s)/2Πnb(s)j=1γj
(62)
where
ffrequency
(63)
α(s)=cosϕ0rA(s)ray tube area factor
(64)
ϕ0=ϕ(0)launch angle
(65)
A(s)=−sin(ϕ)∂r/∂ϕ0+cos(ϕ)∂z/∂ϕ0
(66)
τ(s)travel time along ray
(67)
nc(s)number of caustic points along ray
(68)
nb(s)number of ground reflections along ray
(69)
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
853
γjreflection coefficient for j′th ground reflection
(70)
The wavefield P(f,r) at points (r,h(r))0<r<R on the ground from a source at (0,zs) is computed in two steps.
First, a pre-defined number K of ray paths (rj(s),zj(s)) are computed by solving the ray ODEs with a fourth order
Runge–Kutta method [11, p. 178] with variable stepsize and local error control. The rays start from the
source (rj(0),zj(0))=(0,zs) with uniformly distributed vertical launch angles Φj,j=1,…,K, and the ODE system
(61) is augmented with one equation each for the travel time τ(s) along the ray and the partial
derivatives ∂r(s)/∂ϕ0,∂z(s)/∂ϕ0,∂ϕ(s)/∂ϕ0 with respect to launch angle ϕ0. Rays are reflected at the ground and
terminated at the maximal range r=R or at the upper boundary z=Z of the computational region. The number and
the locations of caustic points and ground reflections along each ray are determined.
Then, for each point (rj,h(rj)) on a receiver grid, the field P(f,r) is obtained as a sum of contributions of the
form (62) approximating the eigenrays by cubic interpolation to appropriate ray subsets.
7 Convergence Results
In this section we verify the implementation of the fourth-order SBP-SAT method and investigate the quality of
the grid in a series of convergence studies. We will calculate the convergence rate q as
q=log10(∥vref−v(N2)∥h∥vref−v(N1)∥h)/log10(N1N2)1/d,
(71)
where d is the dimension (d=2 here), vref is a reference solution, v(N) is the corresponding numerical solution
with N grid points and ∥vref−v(N)∥h is the discrete l2 norm of the error.
7.1 Convergence Study Without Absorbing Layer
To separate the effects of the grid from the numerical method, we here present a convergence study on a smooth
curvilinear grid. We use the analytical solution
u(a)(r,z,t)=Ar2+(z−z0)2−−−−−−−−−−√sin⎛⎝⎜2πf⎛⎝⎜t−r2+(z−z0)2−−−−−−−−−−√c⎞⎠⎟⎞⎠⎟
(72)
which is the pressure field created by a point source with frequency f and amplitude A, located at (r,z)=(0,z0).
We here set A=1,f=12.5 Hz and z0=10 m. In the computations we use the same setup as for the benchmark
problem, except that we here set the dissipation coefficient σ to zero everywhere (in order to have an analytical
solution) and impose the analytical solution u(a) as initial and boundary data. The setup with the smooth domain
and the initial condition is shown in Fig. 5. The convergence results are presented in Table 2. We note that we
obtain the expected fourth order convergence rate.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
854
Fig. 5
The analytical solution at time t=0 on a smooth grid
Table 2
log(l2−errors) and convergence rates on a smooth grid, without absorbing layer
Nr×Nz logel2 q
81×21 -0.26 0.00
161×41 -1.30 3.46
321×81 -2.62 4.45
641×161 -3.73 3.64
1,281 × 321 -4.88 3.80
Next, we investigate the quality of the grid generated (using the commercial grid-generator Pointwise) for the
benchmark problem by running a convergence study with exactly the same setup on that grid. The results are
presented in Table 3. We note that we obtain approximately third order convergence on this grid and draw the
conclusion that the grid, as expected, is not smooth enough to support high-order accuracy (higher than third
order). When solving the benchmark problem, we will thus have to make do with third order convergence.
Table 3
log(l2−errors) and convergence rates on the grid generated for the benchmark problem, without absorbing layer
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
855
Nr×Nz logel2 q
226×56 0.33 0.00
451×111 -0.53 2.87
901×221 -1.37 2.77
1,801 × 441 -2.19 2.72
7.2 Convergence Study with Absorbing Layer
To verify also the implementation of the AL, we here perform a convergence study with a non-zero dissipation
coefficient σ on the benchmark grid. The solution obtained with 3,601 × 881 grid points was used as a reference
solution. The results are presented in Table 4. Similar to Table 3, we obtain slightly less than third order
convergence on this grid.
Table 4
log(l2−errors) and convergence rates on the grid generated for the benchmark problem, with absorbing layer
Nr×Nz logel2 q
226×56 0.46 0.00
451×111 -0.32 2.59
901×221 -1.09 2.56
1,801 × 441 -1.84 2.48
8 Computations
We have solved the benchmark problem described in Sect. 4 for two different sound speed profiles:
profile 1: Constant profile, c=340 m/s
profile 2: Linear profile, c=c0+kz with c0=340 m/s and k=0.1 s−1.
To guarantee a correct solution it is important to verify: (1) that the SPL at ground level is grid-converged, and
(2) that reflections at artificial boundaries are negligible. We begin this section by investigating the effects of the
artificial boundary treatment, and then perform a grid-convergence study. The grid-converged results obtained
with the SBP-SAT method are then compared with the results obtained with the PE and ray tracing methods.
8.1 Domain Truncation
To investigate the effect of the artificial boundaries, we computed the propagation loss for different locations of
the north boundary, for case 1. In Fig. 6 we compare the first order Engquist Majda ABC with the AL approach
to truncate the domain at the north boundary. The effects of the reflections using the first order Engquist Majda
ABC decrease as we move the north boundary higher, but even with the north boundary at a height of 2,000 m
the reflections interfere with the interior waves and cause rapid oscillations in the SPL at ground level. The
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
856
spurious reflections when using the AL are much smaller. After 500 m, the results do not change visibly. In the
remaining computations we place the north boundary at z=500 m and employ the AL approach to truncate the
domain.
Fig. 6
Propagation loss measured 1 m above ground for different locations of the north boundary, using a first order
Engquist Majda ABC and b an AL
Remark 3
We also extended the domain and moved the east boundary further to the right, but the location of the east
boundary turned out to have no impact on the SPL at ground level.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
857
8.2 Grid-Convergence
To verify that the discretization errors are negligible, we study how the computed propagation loss varies with
grid refinement. The results are shown in Fig. 7. The curve obtained using 4.5 grid points per wavelength
deviates significantly from the others, while the curves corresponding to 6 and 12 grid points per wavelength are
almost identical, i.e., indicating grid convergence. In the remaining simulations, grids with 12 points per
wavelength were used.
Fig. 7
Convergence study. The graphs show the propagation loss 1 m above the ground for different levels of grid
refinement
8.3 Comparison of Models
Figure 8 shows the propagation loss, obtained with the SBP-SAT method, in the entire domain up to a height of
100 m. The effects of the refraction that occurs with profile 2 is most apparent far away from the source.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
858
Fig. 8
Propagation loss (dB) for different sound speed profiles, a profile 1 and b profile 2
In Fig. 9 we compare the result obtained with the SBP-SAT method with the result published in [35] and the
result obtained with the PE method described in Sect. 5, for profile 1. We observe that the results are in fairly
good agreement. The maximum difference between the SBP-SAT and the PE methods in Sect. 5 and in [35] are
8 and 4 dB, respectively.
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
859
Fig. 9
Propagation loss 1 m above ground for SBP-SAT and PE, with profile 1
In Fig. 10 we compare the ray tracing methods with the SBP-SAT method, for profile 1. The difference between
the computed SPL using the SBP-SAT method and the most accurate ray tracing method is always greater than
15 dB beyond 1,100 m. We also note that the results obtained with the different ray tracing methods differ
significantly from one another, and they all under-predict the SPL.
Fig. 10
Propagation loss 1 m above ground for the SBP-SAT method and the ray tracing methods, with profile 1
In Fig. 11 we compare the SBP-SAT method with the PE and ray tracing methods, with profile 2. The PE
method again shows reasonable agreement with the SBP-SAT method. The ray tracing methods are here in
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
860
better agreement with the SBP-SAT method than they were in Fig. 10 (profile 1), but again they under-predict
the SPL, except ”Ray interpolation” that now over-predicts (except at the very end of the domain).
Fig. 11
Propagation loss 1 m above ground for the SBP-SAT method, the PE and ray tracing methods, with profile 2
9 Conclusion
The theory surrounding the SBP-SAT technique has been extended with a result that proves the stability of the
SBP-SAT method for the second order wave equation on a curvilinear 2-D domain with mixed boundary
conditions. A fourth-order accurate SBP-SAT method has been applied to the benchmark problem on
atmospheric sound propagation introduced in [35]. Since the SBP-SAT method is here applied to the full wave
equation model, it can be used as a reference against which simpler (and computationally cheaper) methods can
be validated.
The present study has shown that, when applying the SBP-SAT method to sound propagation problems, the
following should be considered:
The introduction of artificial boundaries must not affect the solution. One way to achieve this is with
carefully constructed absorbing layers.
A grid generated for a realistic topography might not support high-order accuracy.
The SPL must be grid-converged.
The results presented in Figs. 10 and 11 show that ray tracing methods are not reliable for prediction of SPL in
the case of irregular terrain. The PE methods show reasonable agreement with the SBP-SAT method, both with
constant speed of sound and with a linear sound speed profile, which is expected since the topography in this
problem is rather gentle. One would expect the PE methods, and the ray tracing methods in particular, to be
more unreliable in the case of more pronounced topography. This is something we hope to address in a coming
study.
Appendix: Finite Difference Operators
For completeness we present the fourth order SBP operators. Here h denotes the grid-spacing. The interior
stencils (in D1 and M(b)) are the standard central 4th order accurate finite difference stencils. At the boundaries
we use one-sided stencils that are formally second order accurate. The discrete norm H is defined:
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
861
The first derivative SBP operator is given by,
The third-order accurate boundary derivative operator S0 is given by,
S0=1h[116−332−1300…]
The interior stencil of −hM(b) at row i is given by (i=7…N−6):
mi,i−2mi,i−1mi,imi,i+1mi,i+2=16bi−1−18bi−2−18bi=16bi−2+16bi+1+12bi−1+12bi=−124bi−2−56bi−1−56bi+1
−124bi+2−34bi=16bi−1+16bi+2+12bi+12bi+1=16bi+1−18bi−18bi+2.
The left boundary closure of −hM(b) (given by a 6×6 matrix) is given by
m1,1m1,2m1,3m1,4m1,5m1,6m2,2m2,3m2,4m2,5m2,6m3,3m3,4m3,5m3,6m4,4m4,5m4,6m5,5m5,6m6,6=1217b1
+59192b2+27010400129345067064608b3+694623760312070402387648b4=−5968b1−6025413881211265549
76b3−5374166637042184992b4=217b1−59192b2+21331800516049630912b4+20839385998024815456b3=36
8b1−124472400121126554976b3+75280666721126554976b4=4957908710149031312b3−4957908710149031
312b4=−1784b4+1784b3=34813264b1+92582828316238757669235228057664b3+2360243299962031278205
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
862
871342944b4=−59408b1−2929461579460729725717938208b3−294467388102329725717938208b4=−591088
b1+2602973192328912556411742685888b3−608341868138411278205871342944b4=−132818869266337594
290333616b3+132818869266337594290333616b4=−86732904112b3+86732904112b4=151b1+59192b2+1377
705022330059726218083221499456b4+56446113384296b5+3782888823025465122092707643413496776874
56b3=1136b1−125059743572b5−48363400904421872275525802884687299744b3−172204932779818917715
3814624b4=−1053241207733542840005263888b4+16139767610328843057963657098519931984b3+564461
4461432b5=−9601191280713392b4−33916692148b5+3323505419126452850508784b3=31088b1+507284006
600757858213475219048083107777984b3+18691032230716b5+124b6+195006219843699738346176140288
32b4=−495927181498464461320965546238960637264b3−16b6−1599871490964937594290333616b4−37517
7743572b5=−3683952230716b5+752806667539854092016b3+10636498712336b4+18b6=8386761355510099
813128413970713633903242b3+22247172617734372763180339520776b4+56b6+124b7+280535371786b5=−
35039615213452232b4−16b7−1309181092513226425254392b3−11187492230716b5−12b6=32906368004458
7b4+55801816692148b5+56b7+124b8+66020484313226425254392b3+34b6
The corresponding right boundary closure is obtained by replacing bi→bN+1−i for i=1,…,8 followed by a
permutation of both rows and columns. Let mi,j be the entry at row i and column j in M(b). The matrix M(b) is
symmetric, which means that it is completely defined by the entries on and above the main diagonal,
i.e., mj,i=mi,j,i=1,…,N,j=i,…,N.
References
1.Appelö, D., Colonius, T.: A high-order super-grid-scale absorbing layer and its application to linear
hyberbolic systems. J. Comput. Phys. 228, 4200–4217 (2009)
2.Bamberger, A., Engquist, B., Halpern, L., Joly, P.: Higher order paraxial wave equation approximations in
heterogeneous media. SIAM J. Appl. Math. 48, 129–154 (1988)
3.Carpenter, M.H., Gottlieb, D., Abarbanel, S.: Time-stable boundary conditions for finite-difference schemes
solving hyperbolic systems: methodology and application to high-order compact schemes. J. Comput.
Phys. 111(2), 220–236 (1994)
4.Carpenter, M.H., Nordström, J., Gottlieb., D.: A stable and conservative interface treatment of arbitrary spatial
accuracy. J. Comput. Phys. 148, 341–365 (1999)
5.Collins, M.D.: A higher-order parabolic equation for wave propagation in an ocean overlying an elastic
bottom. J. Acoust. Soc. Am. 86(4), 1459–1464 (1989)
6.de Boor, C.: A Practical Guide to Splines. Springer, New York (1978)
7.Embrechts, J.J.: Broad spectrum diffusion model for room acoustics ray-tracing algorithms. J. Acoust. Soc.
Am. 107(4), 2068–2081 (2000). doi:
8.Engquist, B., Majda, A.: Absorbing boundary conditions for the numerical simulation of waves. Math.
Comput. 31, 629–651 (1977)
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
863
9.Gustafsson, B.: The convergence rate for difference approximations to general mixed initial boundary value
problems. SIAM J. Numer. Anal.18(2), 179–190 (1981)
10.Gustafsson, B., Kreiss, H.O., Oliger, J.: Time Dependent Problems and Difference Methods. Wiley, New
York (1995)
11.Hairer, E., Nørsett, S., Wanner, G.: Solving Ordinary Differential Equations I. Springer, Berlin (1993)
12.Hicken, J.E., Zingg, D.W.: Aerodynamic optimization algorithm with integrated geometry parameterization
and mesh movement. AIAA J. 48(2), 400–413 (2010). doi:
13.Hornikx, M., Forsse’n, J.: Modelling of sound propagation to three-dimensional urban courtyards using the
extended Fourier PSTD method. Appl. Acust. 72, 665–676 (2011)
14.Jeltsch, R.: Multistep methods using higher derivatives and damping at infinity. Math. Comput. 31, 124–138
(1977)
15.Jensen, F., Kuperman, W., Porter, M., Schmidt, H.: Computational Ocean Acoustics. AIP Press, New York
(1994)
16.Kampanis, N.A.: Numerical simulation of low-frequency aeroacoustics over irregular terrain using a finite
element discretization of the parabolic equation. J. Comput. Acoust. 10(01), 97–111 (2002)
17.Kampanis, N.A.: A finite element method for the parabolic equation in aeroacoustics coupled with a nonlocal
boundary condition for an inhomogeneous atmosphere. J. Comput. Acoust. 13(04), 569–584 (2005)
18.Kampanis, N.A., Ekaterinaris, J.A.: Numerical Prediction of Far-field wind turbine noise over a terrain of
moderate complexity. Syst. Anal. Model. Simul. 41, 107–121 (2001)
19.Karasalo, I., Sundström, A.: JEPE—a high-order PE-model for range-dependent fluid media. In: Proceedings
of the 3rd European Conference on Underwater Acoustics, pp. 189–194. Heraklion, Crete, Greece (1996)
20.Kreiss, H.O., Oliger, J.: Comparison of accurate methods for the integration of hyperbolic equations.
Tellus XXIV(3), 199–215 (1972)
21.Kreiss, H.O., Scherer, G.: Finite Element and Finite Difference Methods for Hyperbolic Partial Differential
Equations. Mathematical Aspects of Finite Elements in Partial Differential Equations. Academic Press Inc
(1974)
22.Laine, S., Siltanen, S., Lokki, T., Savioja, L.: Accelerated beam tracing algorithm. Appl. Acoust. 70(1), 172–
181 (2009). doi: 23.Larsson, C.: Weather effects on outdoor sound propagation. Int. J. Acoust. Vib. 5, 33–36
(2000)
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
864
24.Larsson, E., Abrahamsson, L.: Helmholtz and parabolic equation solutions to a benchmark problem in ocean
acoustics. J. Acoust. Soc. Am.113, 2446–2454 (2003)
25.Mattsson, K.: Boundary procedures for summation-by-parts operators. J. Sci. Comput. 18, 133–153 (2003)
26.Mattsson, K.: Summation by parts operators for finite difference approximations of second-derivatives with
variable coefficients. J. Sci. Comput.51, 650–682 (2012)
27.Mattsson, K., Almquist, M.: A solution to the stability issues with block norm summation by parts operators.
J. Comput. Phys. 253, 418–442 (2013)
28.Mattsson, K., Ham, F., Iaccarino, G.: Stable and accurate wave propagation in discontinuous media. J.
Comput. Phys. 227, 8753–8767 (2008)
29.Mattsson, K., Ham, F., Iaccarino, G.: Stable boundary treatment for the wave equation on second-order form.
J. Sci. Comput. 41, 366–383 (2009)
30.Mattsson, K., Nordström, J.: Summation by parts operators for finite difference approximations of second
derivatives. J. Comput. Phys. 199(2), 503–540 (2004)
31.Mattsson, K., Parisi, F.: Stable and accurate second-order formulation of the shifted wave equation.
Commun. Comput. Phys. 7, 103–137 (2010)
32.Mattsson, K., Svärd, M., Carpenter, M., Nordström, J.: High-order accurate computations for unsteady
aerodynamics. Comput. Fluids 36, 636–649 (2006)
33.Mattsson, K., Svärd, M., Shoeybi, M.: Stable and accurate schemes for the compressible Navier–Stokes
equations. J. Comput. Phys. 227(4), 2293–2316 (2008)
34.Nord 2000. comprehensive outdoor sound propagation model. part 2: Propagation in an atmosphere with
refraction. Technical Report 1851/00, Nordic Noise Group & Nordic Road Administration (2006)
35.Parakkal, S., Gilbert, K., Xiao, D., Bass, H.: A generalized polar coordinate method for sound propagation
over large-scale irregular terrain. J. Acoust. Soc. Am. 128(5), 2573–2580 (2010)
36.Renterghem, T.V., Botteldooren, D.: Prediction-step staggered-in-time FDTD: an efficient numerical scheme
to solve the linearised equations of fluid dynamics in outdoor sound propagation. Appl. Acust. 68, 201–216
(2007)
37.Senne, J., Song, A., Badiey, M., Smith, K.B.: Parabolic equation modeling of high frequency acoustic
transmission with an evolving sea surface. J. Acoust. Soc. Am 132(3), 1311–1318 (2012).
International Journal of Mathematics and Computer Sciences (IJMCS) ISSN: 2305-7661 Vol.31 July 2014 www.scholarism.net
865
doi:10.1121/1.474272038.Sundström, A.: Energy-Conserving Parabolic Wave Equations. FOA Report C20895–
2.7. National Defence Research Establishment, Stockholm (1992)
39.Svärd, M.: On coordinate transformation for summation-by-parts operators. J. Sci. Comput. 20(1), 29–42
(2004)
40.Svärd, M., Nordström, J.: On the order of accuracy for difference approximations of initial-boundary value
problems. J. Comput. Phys. 218, 333–352 (2006)
41.“Wavetools”. 42.Weinberg, H., Burridge, R.: Horizontal ray theory for ocean acoustics. J. Acoust. Soc.
Am. 55(1), 63–79 (1974). doi:
43.Yang, C.F., Wu, B.C., Ko, C.J.: A ray-tracing method for modeling indoor wave propagation and
penetration. IEEE Trans. Antennas Propag.46(6), 907–919 (1998)
44.Yee, K.: Numerical solution of initial boundary value problems involving Maxwell’s equations in isotropic
media. IEEE Trans. Antennas Propag.14, 302–307 (1966)