Perfectly Matched Layers - An Absorbing Boundary Condition For Elastic Wave Propagation
Perfectly Matched Layers - An Absorbing Boundary Condition For Elastic Wave Propagation
J. Josifovski
University Ss. Cyril and Methodius, Faculty of Civil Engineering- Skopje, R. Macedonia
V. Vitanov
University Ss. Cyril and Methodius, Faculty of Civil Engineering- Skopje, R. Macedonia
O. von Estorff
Hamburg University of Technology, Institute of Modelling and Computation, Hamburg, Germany
ABSTRACT: The solution of the wave equation for unbounded domains is often of interest in various fields of science and engineering. In particular, the solution of the elastodynamic wave equation for an unbounded domain finds application in simulation of wave propagation and soil-structure interaction. The present paper deals with the development of a Perfectly Matched Layer as an absorbing boundary condition for time-harmonic elastodynamic problems. The Perfectly Matched Layer formulation allows a widely used discrete numerical technique such as the finite element method to be employed for investigation of unbounded domains. By enforcing the radiation condition in an unphysical layer positioned adjacent to the bounded domain of interest, one obtains a kind of sponge absorbing the propagating waves of all angles of incidence and frequencies. To demonstrate the applicability and accuracy of the new approach a fairly simple but representative example is evaluated to reveal the special Perfectly Matched Layer features. The solution is compared to other boundary formulations, such as the viscous-damper boundary. Keywords: absorbing boundary condition, perfectly matched layer, propagation of elastic waves.
1. INTRODUCTION In geotechnical engineering the investigation of unbounded domains is often of particular interest. A typical example can be found in structural dynamics, where the elastodynamic wave equation of an unbounded domain needs to be solved in order to describe the dynamic interaction between the structure and its underlying soil. The definition of an unbounded domain requires an enforcement of a radiation condition in any unbounded direction. Irregularities in the geometry of the domain, or in the material, often compel a numerical solution of the problem using different mathematical formulations. According to many researchers and engineers, the finite element method in combination with an efficient artificial boundary, taking care of the absorption of waves, would be the best choice. The artificial boundary presented in this contribution imposes the necessary radiation condition and ensures that no spurious reflections occur into the domain of interest. In the past decade quite a few techniques have been developed which solve the problem of the unbounded domains more or less successfully. One of them is the Perfectly Matched Layer (PML) as an absorbing boundary layer, which had turned out to be very promising since it was reported to be rather efficient and accurate. The Perfectly Matched Layer is working like an absorbing sponge for all propagating waves independent of the incidence angle or frequencies. In general, the PML boundary can be defined with the same material as the one used in the domain of interest but additionally having attenuation characteristics introduced. The general idea is sketched in Figure 1. If the outgoing wave originated in BD , see Figure 1(a), enters the PM so-called Perfectly Matched Medium (PMM) where it is attenuated enough in a finite
distance. Only then, the PM can be truncated by a fixed (Dirichlet) boundary without any significant wave reflection, see Figure 1(b).
x2 x2
Outgoing wave
PMM
Attenuated wave
Outgoing wave
PML
Dirichlet BC
(a)
(b)
Lp
Figure 1 PMM truncation: (a) PMM adjacent to an unbounded domain and (b) PML with a fixed edge
As a consequence, the displacements of the coupled system ( BD PM ) in BD should be almost the same as those of the unbounded elastic domain . Hence, the Perfectly Matched Layer PM , is created from the Perfectly Matched Medium PM , thus BD PM is considered as a replacement of the unbounded elastic domain . 2. GOVERNING EQUATIONS In the present paper the Perfectly Matched Layer has been derived using a second-order displacementbased finite element formulation. A wave propagation analysis is performed, where the near-field (bounded domain) is discretized with standard isoparametric finite elements, surrounded by a dynamic far-field representing an unphysical domain (or layer) where the absorbing boundary condition is enforced through special types of functions, which are able to damp the outgoing as well as the reflected wave within the layer. 2.1. Elastic medium Consider a homogeneous isotropic elastic medium subjected to a time-harmonic excitation. The oscillation of the elastic medium will be in the form u ( x ) exp ( it ) with as circular frequency. The governing equations can be summarized as follows:
x
j
ij
j
= 2 ui (no summation on i)
(1a) (1b)
(1c)
ij = Cijkl kl
kl
ij =
1 ui u j + 2 x j xi
2 Cijkl = ij kl + ( ik jl + il jk ) . 3
(2)
ij and ij are the components of the stress and strain tensors, respectively, and Cijkl are the components of the material stiffness tensor C; is the bulk modulus, the shear modulus, and
the unit mass density of the medium.
xi := i ( xi ) dx
0
(3)
This procedure of stretching, in particular, is responsible for the physical mapping of the coordinates in the dynamic wave equation. The coordinate-stretching formally implies
1 = xi i ( xi ) xi
(4)
where xi are real coordinates, xi are complex stretched-coordinates and i ( xi ) are non-zero, continuous, complex-valued coordinate-stretching function. At the same time this procedure creates a complex formulation for inhomogeneous viscoelastic Perfectly Matched Medium (PMM). A plane-strain elastodynamic motion of a PMM is defined by introducing (4) into the governing equations (1) as follows: Equilibrium equation Constitutive relation Kinematic relation
( x ) x = C
j j j
ij ijkl k ,l
ij
j
kl
= 2 ui (no summation on i)
(5a) (5b)
ij =
1 1 ui 1 u j + 2 j ( x j ) x j i ( xi ) xi
(5c)
If such two stretched adjacent media have the same i at their interface, then the matching property will ensure that any propagating wave will pass through the interface without reflection. Consider the x1 x2 plane, with two perfectly matched media defined on the 1) left halfplane ( = ( x1 , x2 ) x1 < 0 ) with i ( xi ) = ilm ( xi ) , and 2) right halfplane ( = ( x1 , x2 ) x1 0 ) with i ( xi ) = irm ( xi ) . The wave propagates from the left medium, upper index (lm), in x1 direction, through the PMM interface into right medium, upper index (rm), as depicted in Figure 2.
surface
lm
interface
rm
wave direction
f1(x1) = 0
f1(x1) = +Re
l1(x1) = 1
WBD
WPM
x2
l1(x1) = Complex
PMM, where the continuous 1 is defined piecewise on the two half-planes. (In this case 2 has not to be considered). The homogeneous isotropic elastic medium governed by the equations (1) is a special case of a PMM where 1lm = 1rm = 1. The same arguments hold for the wave propagation in x2 direction.
with the two bases related by the rotation matrix Q , with components Qij := ei ej . The finite element formulation implementation of the perfectly matched layer condition is related to a rotated coordinate system {ei} through the definition of i ( xi ) as coordinate-stretching function globally on the computational domain. In this way by replacing xi by xi , the Eqs. (5) are expressed in terms of the coordinates xi . Afterwards, the equations with the stretched coordinates are transformed back to the
( ) = ( x ) ( x ) u
2
l2(x2) = 1
l2(x2) = 1
f2(x2) = 0
f2(x2) = 0
x1
(6a) (6b)
(6c)
= C T 1 = ( u T ) + T ( u T ) 2
where i ( xi ) are the coordinate-stretching functions defining the stretch tensors and . The transformation of Eq. (6) to unprimed quantities in the basis {ei } is obtained by
12 = 11 = QQT , = 11 21 22 21
x 12 u1 1 T = Q Q , u = = Qu , = = Q 22 u 2 x2
(7)
corresponding to the primed quantities via the usual change-of-basis rules for vector and tensor components. The weak form of the governing equations Eq. (6) is derived by taking its inner product with an arbitrary weighting function w and integrating the resultant scalar over the entire computational domain using integration-by-parts and the divergence theorem to obtain
: d 2
f m w u d = w n d
(8)
Assuming element-wise interpolations of u and w in terms of shape functions N , imposing Eq. (6b) and (6c) point-wise in Eq. (7), and restricting the integrals to the element domain = e , gives the stiffness and mass matrices for a PML element. In terms of nodal submatrices, with I and J being the node numbers, these are expressed as
e k ij =
B
e
T I DB J
(10) (11)
e m ij = I fm N I N J d
(12)
(1) with nodal shape functions N Ii := ij N I , j and N Ii2 := ij N I , j , described using = Q QT and
( )
( x ) i 1/ ( x ) i := 2 2 and := 1 1 ) 1/ 2 ( x2 1 ( x1 ) i i
(13)
hence diagonal with respect to the basis {ei} , i.e. characteristic basis of the PMM. The Eqs. (8) and (9), the functions i in B , B and in fm are defined globally on the computational domain, not element-wise. Note the evidence of coordinate-stretching in the FE matrices, where the stretch tensors and are incorporated in the nodal compatibility matrices BI and B I , not in the material matrix D . Thus, the system matrices for the are symmetric complex-valued and frequency-dependent
which have to be computed anew for each frequency.The coordinate-stretching procedure performs mapping of a typical element in the layer such that it extends towards infinity. This is obtained in two steps illustratively presented in Figure 3.
x2 x2
Outgoing wave
PMM
Attenuated wave
Outgoing wave
PML
Dirichlet BC
Lp
(a)
(b)
The PML element is first stretched to discretize unbounded domain as in Figure 3(a), then it is truncated (with Dirichlet condition) to a Perfectly Matched Layer element with a depth Lp, see Figure 3(b). These elements are positioned adjacent to those discretizing the bounded domain.
3. INVESTIGATIONS
To estimate the performance of the proposed PML formulation it will be compared to other methods by means of a foundation vibration analysis. Consider a rigid surface massless strip-foundation over a half-space, excited by a vertical harmonic displacement with unit amplitude u0 = 1 and excitation frequency , see Figure 4.
u(0,t)=u0 exp(iwt)
B = 2b x2
x1
r,m,n,x
Figure 4 Surface strip-foundation over a half-space excited by vertical oscillations
half-space
The foundation dimensions are: width B = 3 m and height h f = 1.2 m . The soil half-space is defined as linear isotropic viscoelastic material with a shear wave velocity of cs = 92.2m / s , a unit mass density of = 2000 kg m3 , a Poissons ratio of = 0.3 and damping coefficient of = 0.05 . In the present analyses, two parallel finite element models are discretized, one using viscous-damper as boundary (VDB) and the other a PML as absorbing condition, both are shown in Figure 5. The VDB finite element model depicted in Figure 5(a) is discretized by 56x46 isoparametric quadrilateral elements (with dimensions a = b = B / 6 = 0.25 m.
Lx + Lp = 14b + bp
L = 28b
B/2 = 3b
h = 2a
L = 14b B/2 = 3b
Lp = bp
H = 46a
H + Lp = 19a + ap
H = 19 a
h = 2a
WBD WPM
W WBD BD
WPM
Lp = ap
WPM
(a)
(b)
Figure 5 Finite element model of strip-foundation over elastic half-space with (a) VDB and (b) PML condition
Hence, the viscous-damper boundary condition is enforced at the edge of the model. The PML finite element model differs conceptually because it defines a bounded domain BD discretized by 28 x 19 isoparametric quadrilateral elements, adjacent to which an absorbing layer PM is positioned as shown in Figure 5(b). This PML with depth L p = 3.5 m enlarges mesh to a 30x20 with additional n p = 68 elements. It should be noted both finite element models discretized an identical region which positions the boundary condition at distance L = 7 m from the source of excitation. The analysis quantitatively compares both methodologies by looking at the dimensionless response functions related through a compliance matrix F ( a0 ) as follows:
u = F ( a0 ) P
u2 F22 ( a0 ) 0 0 P2 F11 ( a0 ) F1r12 ( a0 ) P u1 = 0 1 M b bu F1r12 ( a0 ) Fr12 ( a0 ) r12 0 r12
(14)
(15)
The conversion is done by normalization of the physical parameters the transformation from dimensional variables to their dimensionless equivalent variables ^ through the relation: = F F ii ii (16)
For pure translation degrees (i=1,2) all dependent from the dimensionless frequency a0 = ( b ) cs with b (half of the foundation width) as a characteristic dimension. The complex compliance (a ) = F Re ( a ) + iF Im ( a ) of the foundation vertical degree with real ( Re ) and imaginary ( Im ) F 22 0 22 0 22 0 part is presented in Figure 6, respectively.
0.40 0.30
^( ) F22 (Im)
0.5
(a)
1.5
2.5
(b)
Although an exceptional matching of the curves can be observed in both parts of the complex function, there is also a small deviation especially for lower frequencies. The PML as a rigorous condition is considered to produce exact solutions, in contrast to the VDB which is an approximate condition. To complete the picture, a relative amplitude difference with respect to the PML method is illustratively presented in Figure 7.
20% 18% 16% 14% e 12% c n e r 10% e f f i 8% d . l e r 6% 4% 2% 0% 0.5 1 1.5 ao 2 2.5 3
3.5% 15.0%
% =
100
Fzz(Re) Fzz(Im)
10.2%
6.2%
2.9%
2.3%
2.0%
1.6%
1.2%
Figure 7 Relative differences in compliance functions for a rigid surface foundation on half-space
From the cumulative diagram it is obvious that at the lower excitation frequencies the relative difference is more pronounced decreasing with the increasing frequency. The qualitative side of the analysis shows that the VDB model produces 15% error in lower frequency range for a0 0.5 and around 10% at a0 = 1.0 . This could be argued with the fact that at lower frequencies the viscousdamper boundary condition has restrictions with respect to the angle of wave incidence which may lead to some stability problems. Thus, the viscous-damper condition should be positioned at distances greater than L = 7 m from the source of the oscillation to obtain the same accuracy as in the case of the PML. Another aspect of the comparative analyses is presented in Table 1. The ratio t/N of the elapsed calculation time t and the related number of elements N defines the computation time per element.
Table 1 Comparison of PML and VDB with respect to computation time per element Method PML VDB t (min) 13.4 24.8 N (/) 600 2576 t/N (s) 1.34 0.58
Since the PML attenuation is defined on the global computational domain, not element-wise, the model can be discretized with fewer elements which reduce the computational time significantly. In this analysis the spatial discretization in the VDB model needs 4.3 times more elements than the PML. In fact, although PML is computationally more expensive due to its complex formulation, it still remains more efficient because fewer elements are needed for the same level of accuracy. In the current study the time gain was 46 percent. Finally, it can be summarized that the Perfectly Matched Layer formulation not only takes less computation time but also produces better results with fewer elements making it an efficient method and preferable as a boundary condition in solution of elastodynamic problems.
4. CONCLUSIONS
The formulation and application of a Perfectly Matched Layer in definition of unbounded domains for elastodynamics has been presented. Its adsorption capability was found to be very convincing, even in the case of a rather small PML thickness and even for relatively low frequencies. In fact, the results have shown that the PML absorption remains equally efficient at wavelengths far larger than the PML thickness. As a consequence, the PML thickness can be kept minimal even for studies involving low frequencies, and no rescaling of the model size is required. The recommended value of the PML thickness is defined through the ratio max L p 10 which will ensure an accurate solution. The performance of the PML condition has been compared with another dynamic boundary method. The results show that the PML does not only take less computational time but also produces better accuracy with fewer elements. This suggests that the bounded domain may be restricted to the region of interest in order to lower the computational cost. This constitutes the major advantage of the PML with respect to the other methods classifying it as very efficient and preferable choice in the solution of elastodynamic problems. The current contribution brings a formulation of the PML condition derived in the frequency domain. It is recommended that for further research the PML condition should be derived directly in the time domain, which will eventually enable the investigation of a nonlinear behaviour inside the bounded soil region. Moreover, the PML offers an efficient modelling alternative for the simulation of wave propagation in unbounded domains not only for elastodynamic problems but also in other fields of engineering, such as, for instance, acoustics.
ACKNOWLEDGMENT
The financial support of the first author by the DAAD (Deutscher Akademischer Austausch Dienst) through the SEEFORM (South Eastern European Ph.D. Formation in Engineering) program is gratefully acknowledged.
REFERENCES Brenger, J.-P. (1994). A perfectly matched layer for the absorption of electromagnetic waves, J. Comput. Phys. 114, 185200. Basu, U. and Chopra, A.K. (2003). Perfectly matched layers for time-harmonic elastodynamics of unbounded domains. Theory and finite-element implementation, Comput. Methods Appl. Mech. Engrg. 192, 1337 1375. Harari, I. and Albocher, U. (2006). Studies of FE/PML for exterior problems of time-harmonic elastic waves, Comput. Methods Appl. Mech. Engrg. 195, 38543879.