ACOUSTIC MODELING USING THE ULTRA WEAK VARIATIONAL FORMULATION

Tomi Huttunen, Jari P. Kaipio Peter Monk

University of Kuopio University of Delaware

Department of Applied Physics Department of Mathematical Sciences
P.O.Box 1627 FIN-70211 Kuopio, Finland Newark, DE 19716
tomi.huttunen@uku.fi USA
jari.kaipio@uku.fi monk@math.udel.edu

ABSTRACT

Finite element methods (FEM) are widely used for the modeling of acoustic fields characterized by the Helmholtz
equation. At high frequencies, however, the requirement of a sufficiently large number of elements per wavelength
in standard FEM, may lead to an intolerable computational burden. Recently, a variety of new methods have
been proposed that have the flexibility of FEM for general geometries while relax the need for dense meshes.
One such method is the ultra weak variational formulation (UWVF). For the spatial discretization, the UWVF
uses conventional finite element meshes but instead of polynomials used in FEM, the solution in each element is
approximated using a system of plane waves. We show that a parallel UWVF method on a PC cluster can be used
to simulate 3D acoustic fields that extend over tens of wavelengths.

1. INTRODUCTION

The Helmholtz equation is used to model a single frequency component of a steady-state acoustic pressure field.
Standard tools for the full-wave modeling of the Helmholtz equation have been finite element methods (FEM)
[11] and boundary element methods (BEM) [13]. In audio acoustics, applications of the Helmholtz equation range
from the modeling interior acoustic fields [12] to the modeling of human auditory system [14]. Although FEM and
BEM are well suited for low frequency problems, the requirement of a large number of elements per wavelength
causes a rapid increase in the computational burden as the frequency increses. To extend the methods to higher
frequencies, one is forced to use parallel computing techniques such as proposed in [3].
Recently, a variety of finite element type methods have been developed which try to relax the requirement
for mesh density allowing the size of elements to be several wavelengths. In particular, methods that locally
approximate the wave field using elementary solutions of the underlying Helmholtz problem (most commonly
plane waves), have been a topic of intensive study. A review of this kind of improved methods can be found, for
example, in [10]. An interesting example of these new methods is the ultra weak variational formulation (UWVF)
proposed by Cessenat and Després [4, 5]. The method uses a new variational formulation over the element faces
only (not over element volumes such as in FE methods). In addition, the discrete approximation of the UWVF is
computed using plane wave basis functions in each element.
In this study, we show the feasibility of using the UWVF for audio acoustic simulations. In Section 2, we
briefly outline the UWVF and its numerical approximation. The extension of the UWVF solution to the far-field
is discussed in Section 3. In the numerical examples of Section 4, the UWVF method is applied for scattering of
high frequency audio acoustic fields from a rigid obstacle. Finally, some conclusions are drawn in Section 5.

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-1
 2. ULTRA WEAK VARIATIONAL FORMULATION

In this section we outline the coupled ultra weak variational formulation (UWVF) [5] and perfectly matched layer
(PML) [2] method for the Helmholtz problem in an inhomogeneous medium. For a more detailed presentation of
the method, we refer to [8, 10].

Figure 1: The partition of the computational domain using a tetrahedral finite element mesh. The mesh is used in
the numerical simulations of Section 4.

We seek the solution of the time-harmonic pressure field P of the form

P (r, t) = p(r)e−iωt , (1)

where r = (x, y, z) is the spatial variable, t is time, p is the spatial dependent part of the field and ω = 2πf is the
angular frequency. We assume that the pressure p is the solution of the Helmholtz equation

 
1 κ2
∇· ∇p + p = 0, (2)
ρ ρ

where κ = ω/c + iα is wave number and ρ is density. The complex valued wave number κ at frequency f is
defined by the speed of sound c and absorption coefficient α.
For the UWVF we partition the computational domain Ω using a standard tetrahedral finite element mesh with
elements Kk , k = 1, · · · , N , see Fig.1. We denote the interface between two elements K k and Kj by Σk,j and the
face of element Kk on the exterior boundary of the mesh by Γ k . The outward unit normal for the element K k is
nk .
We assume that the mesh is chosen so that the parameters ρ and κ are constants in each element and we denote
ρk ≡ ρ|Kk and κk ≡ κ|Kk . The Helmholtz problem for the acoustic pressure can represented as a collection of

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-2
coupled problems so that for each element K k , k = 1, · · · , N we have

∇ · (Ak ∇pk ) + κ2k ηk2 pk = 0 in Kk , (3)

1 1
nk · (Ak ∇pk ) − iςpk = − nj · (Aj ∇pj ) − iςpj on Σk,j , (4)
ρk ρj
1 1
nk · (Ak ∇pk ) + iςpk = − nj · (Aj ∇pj ) + iςpj on Σk,j , (5)
ρk ρj
   
1 1
nk · (Ak ∇pk ) − iςpk = Q − nk · (Ak ∇pk ) − iςpk + g on Γk , (6)
ρk ρk

where pk = p|Kk . The coupled transmission conditions (4) and (5) arise from the continuity of the acoustic
pressure and normal particle velocity across element interfaces [1]. The coupling parameters ς for (4) and (5) on
the interface Σk,j and exterior boundary Γk are
 
1 <(κk ) <(κj ) <(κ)
ς= + and ς = , (7)
2 ρk ρj ρ

respectively, where <(κ) is the real part of κ. On the exterior boundary the parameter Q ∈ C with |Q| ≤ 1 is used
for defining the boundary condition and g is the complex valued source function. Finally, the matrix A and the
parameter η are either A = I and η = 1 if the element is not in the PML or
 
dy dz dx dz dx dy
A = diag , , and η 2 = dx dy dz (8)
dx dy dz

if the element is in the PML. The parameters d x , dy and dz arise from the complex stretching of the spatial variables
so that
∂x0
= dx (9)
∂x
where

x + i R x σ (|x| − x )ν dx, |x| ≥ x0 ,
0 0 0
x = κ x0 (10)
x, |x| < x0 .

Similar expressions are used also for y and z. We also denote r 0 = (x0 , y 0 , z 0 ).
Instead of solving the pressure field p directly, a new function χ k is defined on element faces as follows
 1

χk = − nk · (Ak ∇) − iς pk  , 1 ≤ k ≤ N. (11)
ρk ∂Kk

It is shown in [4, 5] that χk satisfies the variational formulation

N Z
X   N Z
N X
X  
1 1 ∂ 1 1 ∂
χk − − iς vk − χj − iς vk
ς ρk ∂nk Σk,j ς ρk ∂nk
k=1 ∂Kk k=1 j=1
N Z
X   N Z
X  
Q 1 ∂ 1 1 ∂
− χk − iς vk = g − iς vk , (12)
ς ρk ∂nk ς ρk ∂nk
k=1 Γk k=1 Γk

for all piecewise smooth functions v k satisfying the adjoint Helmholtz equation ∇ · (A k ∇v k ) + κ2k η 2 v k = 0
in Kk . Terms of the double summation in the second term on the left hand side are assumed to exist only if

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-3
two elements share a common face Σk,j . Equation (12) is called the ultra weak variational formulation of the
Helmholtz problem.
The discrete UWVF is obtained by approximating the function χ k using complex conjugated plane wave basis
functions
Nk
X  
1
χak = χk,` − nk · (Ak ∇) − iς ϕk,` , (13)
ρk
`=1

where 
exp(iκk dk,` · r 0 ) in Kk
ϕk,` =
0 elsewhere,
In addition, following the Galerkin method developed in [5], we choose v k = ϕk,` . The angularly equidistributed
directions dk,` are obtained by minimizing the maximum distance between N k points on the unit sphere [6]. In
this study, the maximum allowed number of directions per element is limited to 130.
By substituting the plane wave approximation to the equation (12), the discrete UWVF problem can be written
in the form of the matrix equation

(I − D −1 C)X = D −1 b. (14)

The matrix D is a block diagonal matrix arising from the first term of (12). It consists of blocks D k , 1 ≤ k ≤ N
so that the block Dk corresponds to the element Kk . The sparse block matrix C couples the solution in a single
element to adjacent elements or includes terms arising from the boundary conditions. The data vector b corresponds
to the right hand side of (12) and the weights χ k,` form the vector X = (χ1,1 , χ1,2 , ..., χN,Nk ). The parallelized
method for choosing the number of basis functions N k for each element and solving the matrix equation (14) is
discussed in detail in [9, 10].
After the vector X is solved, the approximation of p k in each element Kk can be computed as a direct summa-
tion
Nk
X
pk = χk,` ϕk,` , (15)
`=1

if κk ∈ R. Otherwise, a simple post-processing step is required [8].

3. FAR-FIELD PATTERN

In many applications it is desirable to compute an approximation to the pressure p at a large distance from the sound
source. To do this, we discuss briefly the computation of the far-field pattern from the UWVF approximation.
At a point r = (R, θ, φ) (given in spherical coordinates) at a large distance (i.e. R >> 1) from the scatterer,
the field p can be decomposed into angular and radial parts so that

eiκR 1
p(r) = p∞ (θ, φ) + O( 2 ), (16)
R R
The angular part of the field is called the far-field pattern and it can be represented as [11]
Z  
1 ∂e−iκr0 ·rS ∂p(rS ) −iκr0 ·rS
p∞ (θ, φ) = p(rS ) − e dS, (17)
4π S ∂n ∂n
r
where the surface S with outward normal n surrounds the scatterer, r S is a point on S and r0 = |r| .

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-4
 The computation of the far-field pattern from the UWVF approximation relies on the following identity
     
1 ∂p 1 ∂v 1 ∂v 1 ∂p
− − − iσp − iσv + − − iσv − iσp
ρ ∂n ρ ∂n ρ ∂n ρ ∂n
 
σ ∂v ∂p
= 2i p − v . (18)
ρ ∂n ∂n

Let us assume next that the surface S is constructed of the collection of the internal faces of the mesh, so that
S = ∪N n n
n=1 Σk,j . Then, each face Σk,j , 1 ≤ n ≤ NS is shared by two elements K k and Kj . Let us also assume
S

that element Kk has outward normal nk = n. Consequently, for element Kj we get nj = −n. By defining
v = exp(−iκr0 · rS ), using the definition (11) for χk and χj ; and with the help of the identity (18), the far-field
pattern from the UWVF approximation can be computed via
NS Z     
ρ X 1 ∂v 1 ∂v
p∞ (θ, φ) = χj − − iσv − χk − iσv , (19)
8πiσ Σn ρ ∂nk ρ ∂nk
n=1k,j

where the integration is over all faces Σ nk,j , 1 ≤ n ≤ NS that form the surface S. The discrete form of (19) is
obtained by replacing χk and χj with their discrete approximations (13).

4. ACOUSTIC SCATTERING PROBLEM

As a numerical example we investigate the scattering of an acoustic pressure field from a perfectly rigid sphere
(radius a = 10 cm) that is surrounded by an infinite, homogeneous medium. Rather than trying to imitate any
particular application, this problem is chosen due to the existence of the analytical Fourier series solution [11]
which allows us to evaluate the accuracy of the UWVF approximation. Similarly, it is also possible to derive
an analytical solution for the far-field pattern. For this problem, the total field p can be divided into the sum of
scattered ps and incident fields pi so that p = ps + pi . In this study, the incident field pi (r) = A exp(iκd · r) is a
plane wave propagating in the direction of positive x-axis, i.e. d = (1, 0, 0). The normal particle velocity is zero
on the surface of the object Γobs which yields the boundary condition

∂ps ∂pi
=− on Γobs . (20)
∂n ∂n
Since we are seeking the UWVF approximation for the scattered field p s , the exterior boundary condition (6) for
the surface Γobs corresponds to (20) if we choose Q = 1 and g = ρ2 ∂p i
∂n .
3
The sphere is embedded in air so we have ρ = 1.2 kg/m and c = 340 m/s. Results are computed using the
frequency f = 20 000 Hz which corresponds to the wavelength λ = 1.7 cm. Hence, we have κa ≈ 37.0. The
computational domain in which the obstacles are enclosed is 40 × 40 × 40 cm cube surrounded by a 5 cm thick
PML. The decay parameters in (10) are set to σ 0 = 30 m−1 and ν = 0. On the exterior boundary of the PML we
set Q = 0 and g = 0.
The computer code used for the simulations is coded in Fortran90 and parallelized using MPI (Message Passing
Interface). The computations are done with a Beowulf PC cluster consisting of 24 2.6 GHz Pentium 4 processors
and having 48.0 GB total RAM. The processors are connected with a 1GB ethernet switch.
In Fig. 2 we show the UWVF approximations of scattered and total fields. A comparison with the Fourier
series solution gives a relative error of 2.7% for the scattered field. The difference between the Fourier series
solution and the UWVF approximation is shown in Fig. 3. The figure reveals that error is mainly originating from
the surface of the scatterer. In fact, it is possible to improve the accuracy by refining the discretization near the
sphere or using curved elements on the surface of scatterer. The latter approach has been used in 2D simulations
in [7].

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-5
 | ps | | ps + pi |
1 1.5
150 150
0.8
100 100
1
50 0.6 50
z (mm)

z (mm)
0 0
−50 0.4 −50
0.5
−100 −100
0.2
−150 −150
0 0
−100 0 100 −100 0 100
x (mm) x (mm)

Figure 2: Scattered and total field for the scattering from the sphere problem. The error of the near field solution
is 2.7 %

Error in | p |
s

150 0.025
100
0.02
50
z (mm)

0 0.015

−50 0.01
−100
0.005
−150

−100 0 100
x (mm)

Figure 3: The difference field indicates that the major source of the error is the insufficently large triangulation of
the spherical surface.

The far-field pattern of the scattered field is shown in Fig. 4. Due to the relatively small error of the UWVF
approximation, the lines for the exact solution and the UWVF approximation are hardly distinguisable from each
other.

5. CONCLUSIONS

The aim of this study was to show that the parallel UWVF method can be used for simulating 3D audio acoustic
Helmholtz problems in which the field extends over tens of wavelengths. Although the numerical example of this
study is a field in a homogeneous medium, previous studies [7, 8, 10] show that the UWVF is also feasible for
problems in inhomogeneous media. Therefore, the UWVF may offer potential advantages over boundary element
methods whose complexity increases with the number of inhomogeneities within the computational domain. A
benefit of the UWVF approach over standard finite element methods are much coarser meshes which reduce the
computational burden.

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-6
 log (|p |)
10 ∞

90
3
120 60

150 30

180 0

210 330

240 300

270

UWVF approximation
Exact solution

Figure 4: Polar plot of the far-field pattern for the scattered field from the sphere. The error of the UWVF approx-
imation is 0.9 %, so the lines for exact and UWVF solutions can not be distinguished.

6. REFERENCES

[1] J.D. Benamou and B. Després. A domain decomposition method for the Helmholtz equation and related
optimal control problems. Journal of Computational Physics, 136:68–82, 1997.

[2] J. Bérenger. A perfectly matched layer for the absorption of electromagnetic waves. Journal of Computational
Physics, 114:185–200, 1994.

[3] P. Geng, J.T. Oden and R.A. van der Geijn. Massively parallel computation for acoustical scattering problem
using boundary element method. Journal of Sound and Vibration, 191(1):145–165, 1996.

[4] O. Cessenat. Application d’une nouvelle formulation variationelle des equations d’ondes harmonigues, Prob-
lemes de Helmholtz 2D et de Maxwell 3D. PhD thesis, Paris IX Dauphine, 1996.

[5] O. Cessenat and B. Després. Application of an ultra weak variational formulation of elliptic PDEs to the
two-dimensional Helmholtz problem. SIAM Journal of Numerical Analysis, 35(1):255–299, 1998.

[6] R.H. Hardin, N.J.A. Sloane and W. D. Smith. Spherical coverings, May 1997. Available on:
http://www.research.att.com/∼njas/coverings/index.html.

[7] T. Huttunen, P. Monk, and J.P. Kaipio. Computational aspects of the ultra-weak variational formulation.
Journal of Computational Physics, 182:27–46, 2002.

[8] T. Huttunen, J.P. Kaipio, and P. Monk. The perfectly matched layer for the ultra weak variational formulation
of the 3D Helmholtz problem. International Journal for Numerical Methods in Engineering, 2004 Accepted.

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-7
 [9] T. Huttunen, P. Monk and J.P. Kaipio. Parallelized UWVF method for 3D Helmholtz problems Proceeding
of ECCOMAS, July 24-28. 2004 Jyväskylä, Finland, Submitted.

[10] T. Huttunen. The ultra weak variational formulation for ultrasound transmission problems. PhD thesis,
University of Kuopio, 2004. Available on: http://physics.uku.fi/∼huttunen

[11] F. Ihlenburg. Finite Element Analysis of Acoustic Scattering. Springer, 1998.

[12] S. Kopuz and N. Lalor. Analysis of interior acoustic fields using the finite element method and the boundary
element method. Applied Acoustics, 45: 193-210, 1995

[13] P.K. Kythe. An Introduction to Boundary Element Methods. CRC Press, 1995.

[14] T. Walsh, L. Demkowicz and R. Charles. Boundary element modeling of the external human auditory system.
The Journal of the Acoustical Society of America, 115(3):1033-1043, 2004

Joint Baltic-Nordic Acoustics Meeting 2004, 8-10 June 2004, Mariehamn, Åland BNAM2004-8