Professional Documents
Culture Documents
Ultrasonics: J.R. Pettit, A. Walker, P. Cawley, M.J.S. Lowe
Ultrasonics: J.R. Pettit, A. Walker, P. Cawley, M.J.S. Lowe
Ultrasonics
journal homepage: www.elsevier.com/locate/ultras
a r t i c l e i n f o a b s t r a c t
Article history: Commercially available Finite Element packages are being used increasingly for modelling elastic wave
Received 21 August 2013 propagation problems. Demand for improved capability has resulted in a drive to maximise the efficiency
Received in revised form 25 November 2013 of the solver whilst maintaining a reliable solution. Modelling waves in unbound elastic media to high
Accepted 29 November 2013
levels of accuracy presents a challenge for commercial packages, requiring the removal of unwanted
Available online 12 December 2013
reflections from model boundaries. For time domain explicit solvers, Absorbing Layers by Increasing
Damping (ALID) have proven successful because they offer flexible application to modellers and, unlike
Keywords:
the Perfectly Matched Layers (PMLs) approach, they are readily implemented in most commercial Finite
Absorbing boundaries
Stiffness Reduction Method
Element software without requiring access to the source code. However, despite good overall perfor-
Finite Elements mance, this technique requires the spatial model to extend significantly outside the domain of interest.
Numerical simulations Here, a Stiffness Reduction Method (SRM) has been developed that operates within a significantly
reduced spatial domain. The technique is applied by altering the damping and stiffness matrices of the
system, inducing decay of any incident wave. Absorbing region variables are expressed as a function of
known model constants, helping to apply the technique to generic elastodynamic problems. The SRM
has been shown to perform significantly better than ALID, with results confirmed by both numerical
and analytical means.
Ó 2013 Elsevier B.V. All rights reserved.
0041-624X/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ultras.2013.11.013
J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879 1869
Non-reflecting boundary conditions simulate an unbound med- as increasing the damping, and is thus an extension of the ALID
ium within the same geometric size since they possess no area or concept. Very significant improvements in efficiency have been
volume. To simulate infinite material surrounding the system, the found to be possible. The paper presents the derivation of the
boundary conditions introduce extra variables into the governing method, followed by an analytical assessment of its performance,
equations that aim to prevent reflections occurring. Many different from which its parameters may be optimised. Finally its perfor-
methods have been developed specific to their respective fields mance is illustrated and compared to that of the other methods al-
[12–17]. The first example of a successful implementation is pro- ready discussed.
vided by Givoli and Keller [14]. This provides a frequency domain
solution for a perfect Non-reflecting boundary. This approach has 2. Stiffness Reduction Method
recently been extended by Velichko and Wilcox [15], for NDE
applications where a frequency domain Finite Element model is We start by reviewing the underlying workings of the ALID
used to calculate the complete scattering behaviour of the defect, method, from which we will be able to identify the role of the stiff-
also with boundaries that are perfectly non-reflecting. The use of ness term and thus develop the new stiffness reduction approach.
these perfect non-reflecting boundary conditions allows for only
a single layer of elements surrounding the defect to be used, mak- 2.1. Absorbing layer theory
ing the model spatially computationally efficient. However these
methods require modification of the standard solving procedures, The absorbing layer concept changes the physical properties of
with non-sparse matrices solved only in the frequency domain. the elements within successive layers such that they absorb inci-
Thus, development of specialist FE code is required beyond what dent wave energy. The physical properties must be changed grad-
is possible using commercial packages. ually, but at an optimum rate because of two competing
Absorbing layer techniques append a number of finite sized lay- mechanisms. The first is that reflections are caused by impedance
ers to the boundary of the spatial domain. The layers are made of mismatches between successive absorbing layers. Thus, alterations
the same elements as the model but with gradually varying prop- in the physical properties of adjacent elements cause reflections
erties that absorb incident waves. This technique generates the that return energy back into the Area of Study (AoS) at each inter-
largest relative increase in geometric size, however, it can be di- face. The second is that the incident wave may not be fully
rectly implemented into commercial FE packages. Layer properties absorbed within the total thickness of the absorbing region and
are optimised such that reflected waves are of a magnitude that so may reflect from the model boundary back into the AoS. An effi-
can be considered negligible. The ease of implementation makes cient ALID allows for compensation between these two mecha-
this approach appealing to modellers, provided the additional size nisms by minimising impedance mismatches between successive
of the spatial domain can be minimised. layers in a minimal thickness to produce a net reflection below a
For modelling elastodynamic wave problems two successful pre-defined threshold [11].
absorbing layer techniques are Perfectly Matched Layers (PMLs) The equation of dynamic equilibrium that forms the basis of the
[18–20] and Absorbing Layers by Increasing Damping (ALID) [21– FE discretisation of the wave propagation problem is:
24]. Both techniques append layers to the model boundary with
absorbing properties that gradually increase, thereby attenuating € þ ½Cu_ þ ½Ku ¼ ½F
½Mu ð1Þ
incident waves. For PML the impedance of the absorbing elements
where ½M; ½C and ½K are the respective mass, damping and stiffness
perfectly matches that of the parent material, thus waves enter the
matrices and ½F represents external force [26]. Eq. (1) has a har-
absorbing material without reflection. However for ALID progres-
monic solution expressed in the form uðx; tÞ ¼ eiðkxwtÞ for the one
sion through the absorbing material is accompanied by impedance
dimensional case where k is the wavenumber and x is the angular
mismatches between successive layers, but with the advantage of € , and velocity, u,
_ in terms of
frequency. By expressing acceleration, u
easier implementation. A comparative study between the two car-
displacement, u , Eq. (1) becomes:
ried out by Rajagopal et al. [11], concluded that PML generally
achieves a smaller model size than ALID for comparable perfor- ½Mx2 u ½Cixu þ ½Ku ¼ ½F ð2Þ
mance. However, PML is more suited to frequency domain simula-
tions and requires specific programming and functionality; The damping matrix can be expressed most simply as two separate
currently this is only available in a minority of commercial FE components proportional to the mass and stiffness matrices, consis-
packages. Therefore, ALID is a more accessible tool for practical tent with options available in most commercial FE packages:
use, being directly compatible with explicit time domain solvers ½C ¼ C M ½M þ C K ½K ð3Þ
in commercial FE programs.
A recent development to the ALID approach has been made by Combining Eqs. (2) and (3) produces a new equation for dynamic
Semblat et al. [25] based on a Rayleigh/Caughey damping formula- equilibrium where damping is expressed solely as a function of
tion. Rayleigh/Caughey damping is available in most commercial coefficients C M , mass proportional damping and C K , stiffness pro-
FE packages and the technique developed is thus called Caughey portional damping:
Absorbing Layer Method or CALM. Spatial variation of damping is
CM
controlled by varying the mass and stiffness proportional damping ½M 1 þ i x2 u þ ½Kð1 ixC K Þu ¼ ½F ð4Þ
x
coefficients to give a diagonal damping matrix. Considering the
contribution of stiffness to wave attenuation is comparable to the From the inspection of the coefficients of Eq. (4) it can be shown
approach taken here; however the variation proposed in is this pa- that Eq. (2) can be re-written without the ½Cu_ term, but instead
per, does not exploit stiffness proportional damping, but rather with complex values of the density, q, and Young’s modulus, E:
reduction of the coefficients of the stiffness matrix itself.
The motivation of the work presented in this paper is to develop CM
q!q 1þi ð5Þ
an absorbing layer technique that can improve on the efficiency of x
the ALID technique, whilst remaining accessible to implementation and
using commercial FE programs. The method, which will be pre-
sented in the following sections, exploits benefits that can be E ! Eð1 ixC K Þ ð6Þ
gained by making gradual changes to the material stiffness, as well
1870 J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879
Inside the AoS where damping is normally zero (C M ¼ 0 and C K ¼ 0), was set to be three times the size of the largest wavelength travel-
Eqs. (5) and (6) default to their original form as shown in Eq. (2). In- ling through the system.
side the absorbing region, damping is nonzero, meaning that inci- The development made with SRM is to notice that under these
dent wave energy is absorbed by the imaginary parts of the conditions the wavenumber can be incrementally increased by
complex terms introduced into the solution via q and E. This can using C M ðxÞ but also by decreasing the value of the Young’s modu-
be illustrated by noticing that the wavenumber is related to the lus, E. Therefore within the absorbing region E becomes a function
density and stiffness such that: of x, decreasing gradually across the absorbing region, avoiding any
rffiffiffiffi dramatic changes in material properties.
q
k/ ð7Þ
E 2.2. Explicit Finite Element solution
Inside the absorbing layers k ¼ kreal þ ikimag . Substituting this back
into the original harmonic solution of the one-dimensional problem To understand how best to apply SRM, it is necessary to evalu-
gives: ate the behaviour of the matrices in the explicit time domain FE
solver [26].
uðx; tÞ ¼ eiðkreal xwtÞ ekimag x ð8Þ Starting from the equation of equilibrium, Eq. (1), solution in
the time domain can be expressed by the central difference finite
kimag x
It is the term e which induces the decay of the waves inside difference equations:
the absorbing region and therefore that must be exploited. Com-
mercial FE packages do not offer a means to alter the wavenumber 1 tDt
€t ¼
u 2
u 2ut þ utþDt ð11Þ
directly. This must be achieved indirectly, by altering values of the Dt
stiffness, damping and mass matrices, to increase the value of kimag
1 tþDt
inside the absorbing region. Consequently, this will also increase u_ t ¼ u utDt ð12Þ
the value of kreal which must be managed carefully in order to min- 2 Dt
imise the reflection between adjacent elements. Substituting Eqs. (3), (11) and (12) into (1) gives:
From Eqs. (5) and (7) it can be shown that the complex wave-
½a0 ½M þ a1 C M ½MutþDt ¼ F t ½½K 2a0 ½Mut
number can be expressed as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ½a0 ½M a1 C M ½MutDt ð13Þ
qð1 þ iCwM Þ where a0 ¼ D1t2 and a1 ¼ 21Dt. The next observation to be made is that
kðkreal ; kimag Þ / ð9Þ
E the dynamic load to the system, F t , is a function
tapplied
of the
The optimal outcome would be a means of increasing the value of stiffness matrix K which can be expressed as F t ð K t Þ . If stability
kimag without increasing kreal . To date such capability for time do- is assumed, utþDt can be solved using an explicit integration method,
main commercial FE packages has not been established, therefore given by Eq. (14).
t t
layer properties must be carefully selected such that successive val- F K ½½K 2a0 ½M
tþDt
ues of kreal result in impedance mismatches that generate reflections u ¼ ut
½a0 ½M þ a1 C M ½M ½a0 ½M þ a1 C M ½M
which can be considered negligible.
½a0 ½M a1 C M ½M tDt
To define an effective absorbing region it is necessary to opti- u ð14Þ
mise variables that the user has control over. Progress made in ½a0 ½M þ a1 C M ½M
ALID by Rajagopal et al. [11], gives a good indication of how best Eq. (14) is thus composed of three terms, each of which is inversely
to optimise initial parameters and forms the basis of the Stiffness affected by the damping coefficient C M . Within the absorbing layer,
Reduction Method (SRM) technique. increasing the value of C M decreases the magnitude of each term,
Firstly, it is noted that introducing damping into the model can resulting in the decay of the incident wave.
decrease the value of the stable time increment within explicit
schemes, thereby reducing computational efficiency by increasing 2.3. SRM mechanisms
computation time. However, a high value of C M causes only a small
decrease in the stable increment as compared to that of C K [11,27]. Eqs. (9), (10) and (14) reveal how best to decay the wave within
It is therefore preferable to avoid using C K to define an absorbing the absorbing region. It is necessary to increase the imaginary com-
boundary within an explicit scheme. Setting C K to zero eliminates ponent of the wavenumber which can be done by increasing the
this issue and allows damping to be controlled solely by C M in Rayleigh damping term C M and decreasing the value of Young’s
which case it is found that the stability is not compromised. modulus, E.
Secondly, damping is defined as an increasing power law set by Changing the variables in this manner also complements
the power p, to gradually alter the value across successive layers. Eq. (14). As well as the decay associated with increasing damping,
the
t first
term is also reduced since the loading component
C M ðxÞ ¼ C Mmax XðxÞp ð10Þ
F ð K t Þ is directly proportional to the Young’s modulus. It is this
C Mmax is a positive real number, XðxÞ ranges from 0 at the AoS mechanism that gives SRM the advantage over conventional ALID.
boundary to 1 at the end of the absorbing region with layer thick- Finally, as in Eq. (10), the thickness of successive layers is equal
ness equal to the element width. Therefore, for a spatial discretised to the element width where x takes discrete values corresponding
FE model, x takes discrete values corresponding to an integer num- to an integer number of element widths. The loading term (and
ber of element widths that increases while moving away from the therefore Young’s modulus) is reduced using an exponential decay
AoS-absorbing layer boundary. function containing an attenuation factor a such that:
There is no established method for directly calculating the ideal
EðxÞ ¼ E0 eaðxÞkinc x ð15Þ
values for the variables in Eq. (10), however there is literature that
suggests how best to begin optimising the technique [11]. The va- where E0 is the Youngs modulus inside the AoS and kinc is the inci-
lue of C Mmax is found using a trial and error method until the dent wavenumber. However, the attenuation factor aðxÞ cannot be a
acceptability criterion of reflection for all modes in the model is constant for each layer since this will induce impedance mis-
validated. In those studies, the thickness of the absorbing region matches between adjacent layers. Therefore it is necessary to
J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879 1871
gradually increase this value across the absorbing region as shown minimise impedance mismatches between successive layers. The
in Eq. (16). larger impedance mismatches that do occur happen deep within
the absorbing region where the incident wave has already been
aðxÞ ¼ amax XðxÞp ð16Þ diminished.
The application of exponential decay to the stiffness matrix has a di- The SRM and ALID are directly compared for the one dimen-
rect impact on the dynamic load ½F in the first term of Eq. (14). sional case, with their respective layer properties optimised for
Effectively, ½F has been replaced by ½Feakinc x reducing the magni- maximum absorption. All variables between the two models re-
tude of the contribution of this term to the nodal displacements main constant except for the Young’s modulus in the absorbing re-
within the layer, utþDt , thus decaying the wave. gion of the SRM. Fig. 1a and b show the performance of the two
techniques. The time history of a single excited node is recorded
3. SRM performance
Displacement (m)
been used to optimise and compare model variables. Following
this, the concept is applied to test models set up using the com- 0.25
mercial FE package ABAQUS/Explicit [28] in order to obtain quan-
titative values for the reflection coefficients as functions of incident 0
angle for two-dimensional analysis. Numerical results from these ↑
−0.25 Reflected
simulations are then compared to an analytical model using the Response
Global Matrix method, reported in detail by Lowe [29].
−0.50
showing the incident 5 cycle tone burst and the observed reflection sizable reflections at successive layer boundaries. However, dra-
from the absorbing region later in time. matic changes only occur once the wave has propagated a signifi-
Fig. 1a and b show that the SRM behaves significantly better cant distance into the absorbing region.
than ALID in this example. The maximum amplitude of the re- Here, the thickness of each layer remains constant and is fixed
flected signal for the ALID is 35.2 dB where as the SRM gives to a value of one-element-thickness. Work carried out by Rajagopal
45.0 dB. et al. [11], has stated that it is preferable to minimise the change in
Due to the likened approach to a Rayleigh/Caughey damping any material properties between adjacent layers so that impedance
formulation proposed by Semblat et al. [25], a further comparison changes are gradual and thus the thinner each layer is the better.
is made between SRM and CALM. To incorporate Rayleigh/Caughey However, for one-dimensional FE models, increasing element
damping, the damping matrix is defined by Eq. (3), where C M and thickness could potentially offset impedance mismatches. The
C K are nonzero. Again C M and C K are varied gradually across the stiffness of a single Truss element is the product of the Young’s
absorbing layer, as in Eq. (10), but with the introduction of an addi- modulus multiplied by the element thickness, thereby allowing
tional variable C Kmax , the maximum stiffness proportional damping for another window from which to control material properties.
coefficient. The definition of SRM remains unchanged. This is not however suitable for two and three-dimensional FE
For a fair comparison to be made with SRM, an optimisation models. In this instance a variation in element size must be per-
function is used which calculates the optimal values for C Mmax formed across a rectangular or quadrilateral element, greatly dis-
and C Kmax that will render the lowest possible reflection coefficient torting the structured arrangement of the mesh. Furthermore,
for the CALM absorbing boundary [32]. The SRM and CALM are di- variation of the mesh density will result in scattering from the
rectly compared for the one dimensional case, with their respective mesh itself, producing an additional mechanism whereby un-
layer properties optimised for maximum absorption. All variables wanted reflections are radiated back into the AoS.
between the two models remain constant except for the Young’s
modulus in the absorbing region of the SRM. Like the comparison
made with ALID and SRM in Fig. 1b and c shows the time history 3.2. Two-dimensional Finite Element simulations
of the observed reflection from the absorbing regions.
CALM absorbing boundaries have outperformed the ALID in this Having established a mechanism to generate a successful
example, with a maximum amplitude of the reflected signal being absorbing region, it is now necessary to assess behaviour for the
36.2 dB. This increase in performance over conventional ALID has two-dimensional case in a commercial FE package. The package
been achieved by the inclusion of stiffness proportional damping to ABAQUS/Explicit [28], has been chosen for the study, as it is widely
the damping matrix, with the same ease of implementation as available, developed and supported for wave propagation simula-
mass proportional damping. However, the increase in performance tions. To assess the validity of the numerical model, results are
has come at the expense of a decrease in the stable time increment compared to an analytical approach based upon the Global Matrix
that is associated with having a nonzero C K term. Despite this method [29].
improvement Rayleigh/Caughey damping does not offer the same
performance that has been achieved with SRM.
3.2.1. Numerical model
Fig. 2 shows how the change of matrix variables changes the
A harmonic bulk wave is incident on the absorbing region, as
value of the wavenumber. Both the real and imaginary components
shown in Fig. 3, across a range of angles hI . For each incident angle,
have been plotted. The first observation is that within the SRM,
the model is re-run by adjusting the inclination of the excitation
kimag increases dramatically in the SRM in comparison to the ALID.
plane. The amplitudes of the incident and reflected wave modes
This is as expected and is in agreement with the predictions made
are monitored along their predicted lines of propagation and the
using Eq. (9). It is the increase in this value that produces such suc-
reflection coefficient deduced. Again the FE models are non-dimen-
cessful decay of any incident waves whilst inside each layer. The
sional and the material properties are defined such that the longi-
desired increase in kimag is also accompanied by the undesired
tudinal and shear wavelengths are 1 and 2, respectively; each
increase in kreal . Again the increase observed inside the SRM is con-
incident wave mode is considered separately. The length of the
siderably greater than that in the ALID, which should suggest
absorbing region is equal to 1.5 times the incident wavelength.
The attenuation parameters are defined such that the Young’s
modulus in the final absorbing layer is 1% of the value in the
25 AoS, C Mmax ¼ x and p ¼ 3.
ALID kreal
ALID kimag
20 SRM kreal
SRM kimag
Wavenumber (1/m)
15
10
0
0 0.2 0.4 0.6 0.8 1.0
Distance across absorbing region (m)
Fig. 2. One-dimensional model: Increase of real and imaginary components of the Fig. 3. FE model used to assess performance of SRM in bulk elastic wave
wavenumber with distance. propagation.
J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879 1873
where MfðnÞ and M bðnÞ are the coefficients defined by the material Shear wave incidence angle (degrees)
properties of each layer, n, governed by Hooke’s law and the Helm-
Fig. 4. Analytical and numerical reflection coefficients for a non-dimensional SRM
holtz decomposition linking displacement, u, with the wave poten- for (a) incident longitudinal and (b) incident shear bulk wave.
tials, / and w, at the front boundary, f, and back, b boundary in the
positive, pos, and negative, neg, propagation directions, with the all the layers combined. This is illustrated in Fig. 4b, whereby after
subscript end denoting the final absorbing layer. the critical angle for an incident shear wave, the reflected longitu-
For a given model of known material properties the left hand dinal component is not radiated back into the AoS. Finally, for time
matrix is known. Similarly, for a specified incident wave the initial domain simulation the incident wave packet is composed of waves
h i
amplitude propagating into the absorbing region pos½1 is also with energy over a range of frequencies. The performance of the
known. Matrix inversion is used to calculate the value of the wave absorbing region is a function of the frequency, meaning that indi-
amplitudes within each layer, thereby deducing the amplitude of vidual frequency components of the incident wave packet will all
h i perform differently.
the wave returning into the AoS, neg ½1 . Since the amplitude of A noticeable trade off for reduction in the spatial domain is
the incident wave mode is known, this gives the reflection coeffi- the increase in the critical time step. Optimising the material
cient of both reflected wave modes and for angles of incidence properties of the absorbing region results in the desired increase
within the range 0 to 90 . in wavenumber coinciding with a reduction in wave speed. If
element size remains constant this must result in an increase
3.2.3. Results in the critical time step across successive layers. Since the time
Results for the analytical and numerical models are shown in step must be constant for the entire model and must be stable
Fig. 4a for an incident longitudinal wave and in Fig. 4b for an inci- across all elements within the model, the CFL number is re-
dent shear wave. duced, see Fig. 5.
There is excellent agreement between the numerical and ana- The effect of having a low CFL has been studied by Drozdz [27].
lytical solutions, confirming the reliability of the analyses. Perfor- The findings show that a low CFL number causes an increase in the
mance at low angles of incidence is significantly better than at error of propagation velocity. However this effect can be nullified
higher glancing angles; consistent with other absorbing layer tech- by using a high mesh density to define the system. For this reason
niques. This confirms that the angle of incidence plays a critical such effects are not of concern, as accurate modelling of elastic
role in the selection of SRM variables. Performance at high angles wave scattering problems requires mesh refinement, for example
will always render a less acceptable solution (however it is often to use a minimum of 15 nodes per wavelength, for which errors
the case that these angles are not of concern in simulations). The in propagation velocity are low (less than 1%). Furthermore, inac-
presence of critical angles generates evanescent waves propagating curacies with propagation velocity within the SRM are not of con-
parallel to the back boundary of the SRM. The angle at which this cern. This is only an issue if spurious reflections return into the
occurs is not necessarily obvious since it is due to a net effect from AoS, which has not been observed.
1874 J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879
1.0
will range from 0–1 with 0 being complete absorption of all inci-
ALID
SRM dent energy and 1 being complete reflection.
selection of C Mmax and amax . Initial values for these variables are
0.6 required as a starting point for the optimisation function. C Mmax
can be expressed as a multiple of x and has an approximate start-
ing value of C Mmax ¼ w. Similarly, the starting value of amax should
0.4
be selected to give a Young’s modulus that is 1% of that the AoS.
The advantage to such an approach means that layer characteris-
0.2 tics are a function of the known material properties and the inci-
dent wavelength. This process also prevents the optimisation
function focusing on local minima.
0
0 0.2 0.4 0.6 0.8 1.0
3.3.2. Selecting SRM thickness
Distance across absorbing region (m)
Performance of absorbing layer techniques will always improve
Fig. 5. Reduction in CFL number due to increase in critical time step across with increased thickness. This is due to there being a greater
successive layers. amount of attenuative material within which the wave can decay.
Here the goal is to achieve a tolerable level of reflection from the
layer within a minimal thickness, thereby offering the maximum
3.3. Optimised SRM variables efficiency possible to the FE solver.
The addition of absorbing layers will inevitably increase the
For a given model, the two-dimensional analytical model number of degrees of freedom of the system. A generic expression
should be used to calculate values of C Mmax and amax that will min- for the percentage increase in this value can be expressed for rect-
imise reflections across all incident angles within the smallest angular domains containing structured quadrilateral elements.
absorbing region thickness possible. Each variable has a combined Fig. 6 shows such a system with length, a, height, b, absorbing layer
contribution to the overall performance of the layers, therefore, in thickness, nk containing structured rectangular elements of size, d.
order to have a simple scalar value to assess performance the opti- Eq. (25) provides an expression for the number of degrees of
misation function described in Section 3.1 is used to minimise the freedom that are to be solved inside the AoS, DoF 2DAoS .
Total Reflected Energy (TRE) from the SRM:
ab
RE DoF 2D
AoS ¼ ð25Þ
TRE ¼ ð19Þ d
2
IE
The addition of absorbing layers with thickness nk increases the
where RE is the reflected energy and IE is the incident energy.
number of degrees of freedom, DoF 2D
total , as given by Eq. (26).
RE can be expressed as a sum over all incident angles for both
the longitudinal, RLE and shear, RSE , reflected wave modes.
ab þ 2nkða þ bÞ þ ð2nkÞ2
Z 2p Z 2p DoF 2D
total ¼ 2
ð26Þ
RE ¼ RLE ðhÞdh þ RSE ðhÞdh ð20Þ d
0 0
From Eqs. (25) and (26), the percentage increase in degrees of free-
where RLE ðhÞ and RSE ðhÞ are proportional to the amplitudes of the dom, DoF 2D% , due to the addition of absorbing layers can be calcu-
measured mode displacements squared, such that lated, Eq. (27).
which gives wider bandwidths would need to consider the performance of the
SRM across a frequency range rather than a central maximum. This
2nk½ða þ bÞ þ 2nk
DoF 2D
% ¼ 100 ð28Þ property partly explains the good performance of PML, since it is
ab normally used in the frequency domain and therefore only has to
Eq. (28) can be readily extended to a three-dimensional problem by be optimised for the single frequency of excitation.
the inclusion of breadth, c, containing structured quadrilateral ele- From the observations made in this study, a suitable SRM thick-
ments. In this case the percentage increase in degrees of freedom, ness of 1:5kinc is capable of achieving the tolerable levels of reflec-
DoF 3D
% is given by Eq. (29).
tions required for all possible incident angles across a sufficiently
h i broad frequency spectrum.
2nk ð2nkÞða þ b þ cÞ þ ðab þ ac þ bcÞ þ ð2nkÞ2
DoF 3D
% ¼ 100
abc 3.4. Comparison of SRM, ALID and PML
ð29Þ
Comparison of SRM, ALID and PML absorbing layer techniques
For absorbing regions operating within significantly reduced spatial
can be made by examining results published by Rajagopal et al.
domains, performance is more susceptible to changes in the fre-
[11], with those in this study. Reflection coefficients for the three
quency of the incident wave. This point is illustrated in Fig. 7a
methods are calculated using analytical models applied to the
and b. Here, the one-dimensional model described in Section 3.1
same non-dimensional case. The layer thickness for all three cases
is used to evaluate the performance of the SRM in the frequency do-
is the same as the wavelength of the incident wave, and the opti-
main for increasing layer thickness expressed as a function of the
misation function is used to select the model variables that will
wavelength of the incident wave. The frequency spectrum of the
produce the lowest TRE value for each case.
incident 5 cycle tone burst is compared with the frequency spec-
The results for the incident compression wave are shown in
trum of the reflected signals.
Fig. 8a and b and for an incident shear wave in Fig. 9a and b.
Fig. 7a and b show that thicker absorbing layers are less suscep-
The results show that PML generally outperforms both ALID and
tible to wider bandwidth signals. If the optimisation function is
SRM for almost all angles of incidence across both modes. How-
used to minimise layer thickness, it should be done with care, since
ever, SRM shows a significant improvement over the ALID. The
the high performance of the absorbing region may only be achiev-
presence of critical angles is also evident, maintaining that the
able within a narrow frequency range. For this reason, sources with
(a) (a) 0
Lwave Reflection SRM
1 Incident Wave Lwave Reflection ALID
SRM Thickness = λ/4 Lwave Reflection PML
0.9
SRM Thickness = λ/2
Reflection coefficient (dB)
−20
0.8 SRM Thickness = 3λ/4
SRM Thickness = λ
Amplitude (no unit)
0.7
0.6 −40
0.5
0.4
−60
0.3
0.2
0.1 −80
0 0 15 30 45 60 75 90
0 0.5 1 1.5 2 2.5
Longitudinal wave incidence (degrees)
Frequency (Hz)
(b) 0 (b) 0
Swave Reflection SRM
Swave Reflection ALID
−10 Swave Reflection PML
Reflection coefficient (dB)
Reflection coefficient (dB)
−20
−20
−30 −40
−40
−60
−50
SRM Thickness = λ/4
SRM Thickness = λ/2
−60
SRM Thickness = 3λ/4
SRM Thickness = λ −80
−70 0 15 30 45 60 75 90
0 0.5 1 1.5 2 2.5
Frequency (Hz) Longitudinal wave incidence (degrees)
Fig. 7. Figure showing (a) the frequency response from the incident and reflected Fig. 8. Reflection coefficients for SRM, ALID and PML for incident longitudinal bulk
signals for increasing layer thickness and (b) the reflection coefficient as a function wave: (a) longitudinal analytical reflection coefficients and (b) shear analytical
of frequency for each layer. reflection coefficients [11].
1876 J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879
(a) 0 effect further, since the wave speed reduction is even more
pronounced.
The reducing wave speed and increasing path length effect is not
apparent across all incident angles. As is consistent with all other
Reflection coefficient (dB)
−20
absorbing boundary methods, high incident angles produce high
reflection coefficients due to glancing effects. As a result, a trade
off is observed between the two. For incident angles approaching
−40 40 the increase in path length and reduction in wave speed is dom-
inating wave attenuation since the incident angle is not so large as
to introduce reflections from the higher glancing angles.
−60 It is important to note that the PML analytical solution can have
a further improved performance. PML variables have been capped
Lwave Reflection SRM to give a 60 dB reflection coefficient at 0 since it is observed that
Lwave Reflection ALID beyond this point the analytical model breaks down and does not
−80 Lwave Reflection PML coincide with what is observed in numerical models. Work by
0 15 30 45 60 75 90 Rajagopal et al. [11], discusses the effects of numerical reflections
Shear wave incidence (degrees) caused by element boundaries as a cause for this disagreement.
As a result it is recommended that optimal values of 3kinc and kinc
(b) 0 be used for ALID and PML thickness respectively. From the obser-
vations made thus far and the discussion that follow, proposed
optimal SRM thickness is 1:5kinc .
Reflection coefficient (dB)
−20
4. Application
↑
Incident S ↑
−0.5
Wave Reflected S
Wave
−1.0
0 4 8 12 16
Time (μs)
6. Discussion
(a) 1.00
The SRM offers a means to apply the absorbing region technique
to reduced spatial domains for use in commercial explicit FE solv-
0.95
ers. Previously, ALID had been chosen by the authors in such in-
stances over the more spatially efficient PML due the flexibility
R (no unit)
1. Assign an SRM thickness of 1:5kinc . It may be necessary to [1] W.D. Smith, The application of finite element analysis to body wave
increase this thickness if narrow-band signals are being used. propagation problems, Geophys. J.R. Astron. Soc. 42 (1975) 747–768.
[2] R. Ludwig, W. Lord, A finite element formulation for the study of ultrasonic
A 1:5kinc SRM thickness has been successful for a 5 cycle Han-
NDT systems, IEEE Trans. Ultrason. Ferroelectr. Freq. Contr. 35 (1988) 809–820.
ning window signal. [3] P. Rajagopal, M.J.S. Lowe, Scattering of fundamental shear horizontal guided
2. Assign layers to the SRM thickness each with a thickness of one- wave by a part-thickness crack in an isotropic plate, J. Acoust. Soc. Am. 124
(2008) 2895–2904.
element-width.
[4] J. Zhang, B.W. Drinkwater, P.D. Wilcox, Longitudinal wave scattering from
3. Select a maximum damping value C Mmax that is equal to the rough crack-like defects, IEEE Trans. Ultrason. Ferroelectr. Freq. Contr. 58
angular frequency, w, of the incident wave. (2011) 2171–2180.
4. Vary the mass proportional damping of the elements within [5] L. Moreau, A. Velichko, P.D. Wilcox, Accurate finite element modelling of
guided wave scattering from irregular defects, NDT&E Int. 45 (2012) 46–54.
each layer by C M ðxÞ ¼ C Mmax XðxÞp , see Eq. (10). [6] P. Rajagopal, E.A. Skelton, W. Choi, M.J.S. Lowe, R.V. Craster, A generic hybrid
5. Calculate amax which can be done by re-arranging Eq. (16) to model for bulk elastodynamics, with application to ultrasonic nondestructive
EðxÞ
lnð E Þ evaluation, IEEE Trans. Ultrason. Ferroelectr. Freq. Contr. 59 (2012) 1239–
give amax ¼ 0
kinc x
. In the final absorbing layer at x ¼ 1:5kinc let 1252.
Eð1:5kinc Þ
[7] S. Mahaut, N. Leymarie, C. Poidevin, T. Fouquet, O. Dupond, Study of complex
E0
= 0.01. ultrasonic NDT cases using hybrid simulation method and experimental
6. Calculate the values of the function aðxÞ for each layer by validations, Insight 53 (2011) 664–667.
[8] P.D. Wilcox, A. Velichko, Efficient frequency-domain finite element modelling
aðxÞ ¼ amax XðxÞp , see Eq. (15). of two-dimensional elastodynamic scattering, J. Acoust. Soc. Am. 127 (2010)
7. Vary the Young’s modulus of the elements within each layer by 155–165.
EðxÞ ¼ E0 eaðxÞkinc x , see Eq. (15). [9] P. Bettess, Infinite Elements, Penshaw Press, 1992.
[10] G.R. Liu, S.S. Quek Jerry, A non-reflecting boundary for analysing wave
propagation using the finite element method, Finite Elem. Anal. Des. 39 (2003)
To make small adjustments to the performance of the layers 403–417.
vary the values of C Mmax and amax (by altering Eð1:5k
E0
inc Þ
). Generally [11] P. Rajagopal, M. Drozdz, E.A. Skelton, M.J.S. Lowe, R.V. Craster, On the use of
absorbing layers to simulate the propagation of elastic waves in unbounded
speaking, performance can always be improved by increasing the isotropic media using commercially available finite element packages, NDT&E
total thickness of the absorbing boundary. Int. 51 (2012) 30–40.
J.R. Pettit et al. / Ultrasonics 54 (2014) 1868–1879 1879
[12] B. Engquist, A. Majda, Radiation boundary conditions for acoustic and elastic [23] M. Castaings, C. Bacon, B. Hosten, M.V. Predoi, Finite element modeling of
wave calculations, Commun. Pur. Appl. Math. 32 (1979) 313–357. torsional wave modes along pipes with absorbing materials, J. Acoust. Soc. Am.
[13] A. Bayliss, E. Turkel, Radiation boundary conditions for wave-like equations, 119 (2006) 3741–3751.
Commun. Pur. Appl. Math. 33 (1980) 707–725. [24] W. Ke, M. Castaings, C. Bacon, 3D finite element simulations of an air-coupled
[14] D. Givoli, J.B. Keller, Non-reflecting boundary conditions for elastic waves, ultrasonic NDT system, NDT&E Int. 42 (2009) 524–533.
Wave Motion 12 (1990) 261–279. [25] J.F. Semblat, L. Lenti, A. Gandomzadeh, A simple multi-directional absorbing
[15] A. Velichko, P.D. Wilcox, Efficient finite element modelling of elastodynamic layer method to simulate elastic wave propagation in unbounded domains, Int.
scattering with non-reflecting boundary conditions, in: Rev. Prog. QNDE, J. Numer. Meth. Eng. 85 (2011) 1543–1563.
volume 1430 of AIP Conference Proceedings, pp. 142–149. [26] K.J. Bathe, Finite element procedures, Prentice Hall, 1996.
[16] J.M. Carcione, Boundary conditions for wave propagation problems, Finite [27] M. Drozdz, Efficient Finite Element Modelling of Ultrasound Waves in Elastic
Elem. Anal. Des. 16 (1994) 317–327. Media, Ph.D. thesis, Imperial College of Science and Technology and Medicine,
[17] D. Givoli, High-order local non-reflecting boundary conditions: a review, Wave 2008.
Motion 39 (2004) 319–326. [28] ABAQUS v 6.9, http://www.abaqus.cc.ic.ac.uk/v6.9/, Version Abaqus 6.9.1,
[18] J.P. Berenger, A perfectly matched layer for the absorption of electromagnetic 2006.
waves, J. Comp. Phys. 114 (1994) 185–200. [29] M.J.S. Lowe, Matrix techniques for modelling ultrasonic waves in multilayered
[19] E.A. Skelton, S.M.D. Adams, R.V. Craster, Guided elastic waves and perfectly media, IEEE Trans. Ultrason. Ferroelectr. Freq. Contr. 42 (1995) 525–542.
matched layers, Wave Motion 44 (2007) 573–592. [30] R. Courant, K. Friedrichs, H. Lewy, Uber die partiellen differenzengleichungen
[20] F.D. Hastings, J.B. Schneider, S.L. Broschat, Application of the perfectly matched der mathematischen physik, Math. Ann. 100 (1) (1928) 32–74.
layer (PML) absorbing boundary conditions to elastic wave propagation, J. [31] R. Courant, K. Friedrichs, H. Lewy, On the partial difference equations of
Acoust. Soc. Am. 100 (1996) 3061–3069. mathematical physics, IBM J. Res. Dev. 11 (2) (1967) 215–234.
[21] M. Israeli, S.A. Orszag, Approximation of radiation boundary conditions, J. [32] MATLAB R2007a, The MathWorks Inc., United Kingdom, 2012.
Comp. Phys. 41 (1981) 115–135.
[22] M. Castaings, C. Bacon, B. Hosten, M.V. Predoi, Finite element predictions for
the dynamic response of thermo-viscoelastic material structures, J. Acoust.
Soc. Am. 115 (2004) 1125–1133.