See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/308352495

Hybrid Asynchronous Perfectly Matched Layer for seismic wave propagation

in unbounded domains

Article  in  Finite Elements in Analysis and Design · December 2016

DOI: 10.1016/j.finel.2016.07.006

CITATIONS READS

4 477

4 authors, including:

Eliass Zafati Irini Djeran-Maigre

Atomic Energy and Alternative Energies Commission Institut National des Sciences Appliquées de Lyon
9 PUBLICATIONS   19 CITATIONS    87 PUBLICATIONS   963 CITATIONS   

SEE PROFILE SEE PROFILE

Florent Prunier
Institut National des Sciences Appliquées de Lyon
46 PUBLICATIONS   345 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

GECKO Project View project

landslide View project

All content following this page was uploaded by Florent Prunier on 04 December 2017.

The user has requested enhancement of the downloaded file.

 Finite Elements in Analysis and Design 122 (2016) 1–15

Contents lists available at ScienceDirect

Finite Elements in Analysis and Design

journal homepage: www.elsevier.com/locate/finel

Hybrid Asynchronous Perfectly Matched Layer for seismic wave

propagation in unbounded domains
Michael Brun n, Eliass Zafati, Irini Djeran-Maigre, Florent Prunier
Université de Lyon, INSA-Lyon, LGCIE, 34 rue des Arts, F-69621 Villeurbanne Cedex, France

art ic l e i nf o a b s t r a c t

Article history: Perfectly Matched Layer (PML) is recognized as a very effective tool for modeling unbounded domains.
Received 14 March 2016 Nonetheless, the computation time required by the PML may be large, especially when an explicit time
Received in revised form integration scheme is adopted for dealing with the wave propagation problem both in the domain of
6 July 2016
interest and in the PML medium. In this paper, it is proposed to investigate subdomain strategies
Accepted 12 July 2016
Available online 16 September 2016
enabling the appropriate time integration scheme in the PML with its own time step to be chosen,
independently of the choice of the time scheme in the domain of interest. We focus on explicit time
Keywords: integrator in the physical subdomain (Central Difference scheme) associated with a fine time step
Perfect Matched Layers (PML) satisfying the CFL stability criterion. The PML formulation proposed by Basu and Chopra (2004) [1] for 2D
Elastic wave
transient dynamics, has been coupled with the interior physical subdomain using the dual Schur
Transient analysis
approach proposed by Gravouil and Combescure (2001) [2]. Hybrid (implicit time integrator for the PML)
Subdomain coupling
Hybrid Asynchronous PML asynchronous (multi time steps) PMLs have been derived. Their very good accuracy has been shown by
considering the following numerical examples: Lamb's test, loaded rigid strip footing on an half space
and a layered half space.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction
The simulation of wave motion in unbounded media requires
the introduction of artificial boundaries surrounding the bounded
computational domain. Several techniques have been developed to
reproduce an unbounded domain for elastic wave propagation:
the infinite elements (Bettess [3], Su and Wang [4]), appropriate
absorbing boundary conditions (Enquist and Majda [5]), absorbing
layer methods such as the Rayleigh damping layers (Semblat et al.
[6], Rajagopal et al. [7]) or the Perfectly Matched Layers (PML)
(Chew and Liu [8]). The perfectly matched layers (PML) is an
absorbing layer method which surrounds the computational
domain with an uniform thickness layer. The PMLs are character-
ized by their capabilities of providing the same attenuation for all
frequencies and all angles of incidence without any reflection from
the interface.
The PML was originally developed for the electromagnetic
waves by Bérenger [9] using a field-splitting formulation and
n
Corresponding author.
E-mail addresses: michael.brun@insa-lyon.fr (M. Brun),
eliass.zafati@insa-lyon.fr (E. Zafati),
irini.djeran-maigre@insa-lyon.fr (I. Djeran-Maigre),
florent.prunier@insa-lyon.fr (F. Prunier).
became one of the most widely used methods in the simulation of
wave propagation problems in unbounded media. The technique
was then adapted to the elastodynamic equations. Hastings et al.
[10] extended the PML from electromagnetics to elastodynamics
using a formulation in terms of displacement potentials imple-
mented in the finite difference framework. Using also the finite
difference method, Chew and Liu [8] introduced a new split-field
formulation for isotropic media, based on the velocity and stress
fields. Later on, Collino and Tsoga [11] proposed a finite difference
split-field formulation similar to Chew et al. [8], applied to ani-
sotropic media. In [12], Wang et al. developed a new PML for-
mulation, called C-PML based on unsplit-field formulation, using
convolution features adapted to the finite difference method. Next,
Matzen [13] extended the C-PML approach to the finite element
method. In this work, we focus on the unsplit-field formulation in
the framework of the finite element method developed by Basu
and Chopra for applications involving 2D media [14,1]. More
recently, this formulation was extended by Basu to 3D media in
the framework of explicit computations [15] and implemented in
the FE code LS-DYNA [16]. From the frequency-domain equations
of Basu and Chopra, Kucukcoban and Kallivokas derived an unsplit
mixed approach of the PML, by retaining the displacement and
stress fields as unknowns in the time domain [17]. Next, in order
to couple their PML to a displacement-only field formulation in

http://dx.doi.org/10.1016/j.finel.2016.07.006
0168-874X/& 2016 Elsevier B.V. All rights reserved.
2 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15
the physical interior domain, the authors extended their PML to a
mixed hybrid approach [18,19].
Following the same idea, it is of great interest to adopt an
efficient explicit time integration with a fine time step for the
wave propagation into the soil medium without constraining the
choice of the time integrators and time step size for the other
partitions of the complex soil–structure interaction problem.
Indeed, under an earthquake excitation, the problem is multi
physics by nature with phenomena occurring at very different
space and time scales into the soil and solid media. In this paper,
the proposed approach enables the unphysical PML medium to be
integrated in time independently of the physical domain. The main
benefit is a higher versatility and numerical efficiency of the PML
which can be implemented using a more appropriate time inte-
grator associated with a larger time step than the one employed in
the soil medium imposed by the CFL condition for ensuring the
algorithm stability [20]. The proposed PMLs can be viewed as a
hybrid and asynchronous PML version because different time
integrators can be adopted as in the work of Kucukcoban and
Kallivokas [18,19], with the adding desirable properties of dealing
with different time scales in the same dynamic simulation.
Moreover, the proposed approach enables the number of
unknowns to be reduced in comparison to the previous mixed
displacement-stress formulation. More precisely, only displace-
ment quantities are solved in time, at the expense of requiring
more storage and calculations involving strain and stress integrals
in the PML [1].
The subdomain method proposed by Gravouil and Combes-
cure [21] provides the suitable properties for coupling an explicit
time integrator for the subdomain of interest with Newmark
implicit time integrators for the other partitions, including the
PMLs. The method follows a dual Schur approach by ensuring the
velocity continuity at the interface through the use of Lagrange
multipliers. The velocity continuity is considered at the fine time
scale, that is associated with the time step satisfying the CFL
condition. The method is proved to be stable for any Newmark
(explicit and implicit) time integrators [22] using the so-called
energy method (Hughes, [20]). It leads to the first order of
accuracy when coupling second order accurate time integration
schemes due to a slight spurious dissipation at the interface as
soon as different time steps are adopted. When adopting the
same time step, the second order of accuracy is achieved [23].
The GC method was adopted in previous works in order to
design implicit, multi directional, multi time step absorbing
layers, based on increasing Rayleigh damping ratios in the
thickness of the absorbing layers (Zafati et al. [24,25]). Recently,
Brun et al. [26] proposed a general framework to derive a family
of coupling algorithms from the energy method, initially
employed for ODE (Ordinary Differential Equation) and gen-
eralized to DAE (Differential Algebraical Equation) after the
introduction of the Lagrange multipliers. The coupling algo-
rithms are built by ensuring the zero value at the large time scale
of the interface pseudo-energy involved in the generalized
energy method. The derived coupling algorithms can be con-
sidered as Hybrid Asynchronous Time Integrator (HATI) enabling
to couple any Newmark and α schemes, while maintaining the
stability and the second order of accuracy of the coupled time
integrators [27].
In this paper, the GC method is considered because only
Newmark time integrators are investigated. The unsplit field for-
mulation proposed by Basu and Chopra [14,1,28] is adopted and
resumed in the first section. The second section is devoted to the
coupling algorithm, allowing the use of hybrid multi time step
PMLs in explicit computations. The last section concerns numer-
ical examples including Lamb's test and loaded rigid strips lying on
the surface of homogeneous and layered soils. In this last two
examples, the interest of the proposed approach is highlighted by
dealing with three subdomains with different time integrators
associated with their own time step: explicit soil subdomain sur-
rounded by implicit multi time step PMLs, and coupled with
implicit solid subdomain. Efficiency of the proposed approach is
assessed by comparing time histories of displacements and ener-
gies as well as L2 error norms between the numerical results and
the reference results obtained by a monolithic full explicit analysis
using an extended mesh.
2. Perfectly matched layer
2.1. Strong form of the PML in frequency domain
The PML model used in this work has been developed by Basu
and Chopra [1,14]. It is built using the classical elastodynamic
equations by introducing the complex-valued stretching func-
tions λi. The main idea is to replace the real coordinates xi with
the complex ones xi -x~ i : R-C. The complex coordinates are
defined by:
∂x~ i
p
f ðxi Þ
¼ λi ðxi Þ ¼ 1 þ f i ðxi Þ  i i
e
ð1Þ
∂xi bks
where b denotes the characteristic length of the physical pro-
blem, ks ¼ cωs is the wavenumber and cs is the S-wave velocity. The
real-valued positive functions fie and fip vanish at the interface
between the PML and the physical domain so that the unphy-
sical PML perfectly matches the physical domain. The damping
function fip serves to attenuate the propagating waves in the xi
direction, whereas the damping function fie attenuates the eva-
nescent waves. In Eq. (1), the dependence of the complex term
on the factor iω allows for an easy application of the inverse
Fourier transform when expressing the PML in the time domain,
resulting in a PML formulation independent on the frequency. In
other words, all the frequencies are damped out in the
same way.
The PML formulation is obtained by modifying the governing
equations defined in the frequency domain. The classical strong
form of the equation of motion for a homogeneous isotropic
medium under the plane strain assumption is written by sub-
stituting xi by x~ i as follows:
8
> P 1 ∂σ ij
>
> j
> ¼  ω2 ρu j
>
>
<
λj ðxj Þ ∂xj
P
σ ij ¼ k;l C ijkl εij ð2Þ
>
>  
>
> 1 1 ∂u 1 ∂uj
>
: εij ¼ 2 λj ðxj Þ ∂xj þ λi ðxi Þ ∂xi
i
>
where Cijkl are the components of the elastic constitutive tensor.
2.2. Strong form of the PML in time domain
Before writing the governing equations of the PML in time
domain, we introduce the following notations for the PML region:
ΩPML is the region of the PML, bounded by the Γ PML ¼ Γ DPML þ Γ NPML ,
where Γ PML \ Γ PML ¼ ∅, defining decomposition of the boundary
D N
between Dirichlet and Neumann conditions. In addition, g N
denotes the prescribed tractions on ΓPMLN and J ¼ ½0; T is the time
interval of interest. Thanks to the introduction of the stretching
functions expressed in Eq. (1), the inverse Fourier transform can
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15 3

be easily applied to the previous frequency-domain equations, attenuation function F 1 ðLp Þ in Eq. (9) corresponds to the space
RL
leading to the following time-domain equations [1,28]: integral in the PML of the function fp: F 1 ðLp Þ ¼ s p¼ 0 f p ðsÞds. Sub-
8
> cs μ stituting the expression of the function fp given in Eq. (8), it can be
divðσ F~ þ Σ F~ Þ ¼ ρf m u€ þ ρ f c u_ þ 2 f k u in ΩPML  J
e p
>
> a L
easily shown that: F 1 ðLp Þ ¼ n αþ 1p . Assuming a normal incidence
>
> b
>
> b
> (θ ¼ θs ¼ 0), the coefficient aα can be expressed as a function of the
< σ ¼ C : ε in ΩPML  J
>
ð3Þ reflection coefficient Rpp by:
>
> F eT ϵ_ F e þ F pT ε F e þ F eT ε F p þ F pT E F p ¼ …
>
> cp
>
> ðn þ 1Þ  
>
> 1 1 c 1
: ð∇u_ F þ F ∇u_ Þ þ ð∇u F þ F ∇uÞ in ΩPML  J
> T e eT T p pT
aα ¼ s ln ð10Þ
2 2 2Lp Rpp
Thus, according to the above relationship, the efficiency of the
with the boundary conditions: PML can be controlled by the PML thickness Lp and the degree n of
8 the damping functions. Theoretically, in the continuous formula-
< u ¼ 0 on Γ D
PML tion, the reflected wave can be made arbitrary small but it is not
ð4Þ
: ðσ F~ þ Σ F~ Þn ¼ 0 on Γ N
e p
PML the case anymore after the space and time discretization presented
in the following.
by assuming that traction forces are equal to zero on the Neumann
condition. 2.3. Displacement-based weak form of the PML
Details on the construction of the matrices involved in the
above strong form equation can be found in [1,28]. In the fol- The space discretization proposed by Basu and Chopra [1] is
lowing, we summarize the different matrices expressed as a displacement-based, following a standard finite element for-
function of damping functions f e and f p: mulation. The time discretization employed the classical implicit
" e
# " cs p # Newmark scheme (Constant Average Acceleration scheme, [22]).
e
1 þ f 1 ðx1 Þ 0 p
f ðx Þ
b 1 1
0
F ¼ e ; F ¼ cs p
The space and time discretization is summarized in the following,
0 1 þ f 2 ðx2 Þ 0 f ðx Þ
b 2 2 before presenting the time coupling of an hybrid multi time step
" # " cs p #
e
1 þf 2 ðx2 Þ 0 f ðx Þ 0 PML with the physical domain handled with an explicit time
b 2 2
F~ ¼ ; F~ ¼
e p
e p ð5Þ integrator (Central Difference scheme). It has to be noted that the
0 1 þf 1 ðx1 Þ 0 cs
f ðx Þ
b 1 1
time stepping procedure proposed by Basu and Chopra for tran-
with: sient dynamics is simplified: Newton iterations are not needed
8 e e
anymore.
>
> f ¼ ð1 þ f 1 ðx1 ÞÞð1 þf 2 ðx2 ÞÞ Let v be the test function belonging to an appropriate space, the
< m e p e p
f c ¼ ð1 þ f 1 ðx1 ÞÞf 2 ðx2 Þ þ ð1 þ f 2 ðx2 ÞÞf 1 ðx1 Þ ð6Þ weak formulation is obtained by integrating over the computa-
>
>
: f k ¼ f ðx1 Þf ðx2 Þ
p p tional domain Ω:
1 2
Z Z Z
c μ
Contrary to the expression in [1], it can be noted that the PML is ρf m v  u€ dΩ þ ρ s f c v  u_ dΩ þ f v  u_ dΩ þ …:
2 k
Ω Ω b Ω b
written in this work without material damping, whereas original Z Z Z
ε~ e : σ dΩ þ ε~ p : Σ dΩ ¼ v  ðσ F~ þ Σ F~ Þ:ndΓ
e p
formulation proposed by the authors incorporate Voigt material ð11Þ
Ω Ω Γ
damping. Here, a conservative soil medium is assumed (no
material damping, no Rayleigh viscous matrix), along with a zero In the above equation, we introduce the notation pen þ 1 for the
R R
damping ratio PML so as to ensure the match between the PML internal force given by: pen þ 1 ¼ Ω ε~ e : σ dΩ þ Ω ε~ p : Σ dΩ  Γ ¼ ∂
and the physical domain. Ω is the boundary of Ω, and n the normal vector related to Γ. The
The integral stress and strain tensors Σ and E are related to tensors ε~ e and ε~ p depend on the damping functions of the PML as
stress and strain tensors σ and ε , by the following relationships: follows [1]:
Z t Z t 8
> 1
< ε~ ¼ ððgradvÞF~ þ F~ ðgradvÞ Þ
e eT
Σ¼ σ dt E ¼ ε dt ð7Þ > e T
2
0 0 ð12Þ
> 1
: ε~ p ¼ ððgradvÞF~ þ F~ ðgradvÞT Þ
> p pT
The PML expressed for two-dimensional elastic waves involves 2
time-integral of the stress and strain tensors when unsplit for-
mulation is adopted [14]. The extension to three-dimensional The mass, the damping and the stiffness matrices are given by:
Z Z Z
elastic waves requires an additional tensor defined by the time- c μ
mIJ ¼ ρf m NI NJ dΩI d cIJ ¼ ρf c s NI NJ dΩI d kIJ ¼ f NNd ΩId
integral of the integral stress tensor Σ [15]. The scope of this work Ω Ω b 2 k I J
Ωb
is reduced to the case of two-dimensional elastic waves. ð13Þ
The damping functions fie and fip are written as a polynomial of
where NI is the shape function related to the node I and Id is the
degree n as:
identity matrix. Taking into account Eq. (11), the internal force
 
α x  x0 n term pen þ 1 can be written as:
f i ¼ aα i α ¼ e; p x0 r xi rx0 þ d ð8Þ Z Z
Lp
B~ σ^ n þ 1 þ B~ Σ^ n þ 1
eT pT
pen þ 1 ¼ ð14Þ
where aα is a positive integer, α being related to evanescent or Ωe Ωe

with both B~ and B~ matrices depending on the shape functions

propagating waves. For a PML of thickness Lp, Basu and Chopra give e p

the reflexion coefficient Rpp for an incident P wave, characterized by and damping functions expressed in terms of their nodal sub-
the angle θ and the reflective angle θs (one way in the PML and the matrices as:
reflected way after being reflected at the end of the PLM, [1]: 2 e 3 2 p 3
  N~ I1 0 N~ I1 0
cos ðθ þ θs Þ cs 6 e 7 6 p 7
Rpp ¼ exp  2 F 1 ðLp Þ cos θ ð9Þ B~ I ¼ 6
e
N~ I2 7 ~p 6 N~ I2 7
cos ðθ  θs Þ cp 4 0 5 BI ¼ 4 0 5 ð15Þ
~ e
~
N I2 N I1
e
~ p
~
N I2 N I1
p
where cp is the P wave velocity and cs is the S wave velocity. The
4 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15

with: 2.4. Time discretization of the PML

N~ Ii ¼ F~ ji NI;j N~ Ii ¼ F~ ji N I;j
e e p p
and ð16Þ In the original formulation of Basu and Chopra [1], the time
The Voigt notation is adopted for the stress and time-integral dicretization employs implicit Newmark time integration schemes
σ  Σ  in a displacement format along with a Newton iteration strategy.
^ ¼ Σ 22 .
11 11
stress tensors, giving the following vectors: σ^ ¼ σσ22
12
and Σ Σ 12 Here, it is proposed a much simpler algorithm without Newton
For the time stepping, Basu and Chopra assume the relation- iteration. The space discrete form of the equation of motion in PML
ship between stress and time-integral stress vectors: given in Eq. (25) is integrated in time by adopting an implicit
Newmark scheme (Constant Average Acceleration scheme). Let us
consider the time step Δt defined by the beginning time tn and the
Σ^ n þ 1 ¼ Σ^ n þ dt σ^ n þ 1 ð17Þ
end time t n þ 1 . The classical approximate Newmark formulas are
It follows: expressed in acceleration format:
Z Z ( nþ1
B~ σ^ n þ 1 þ B~ Σ^ n ¼ U n;p þ βΔt 2 U€ n þ 1
T pT
pen þ 1 ¼ ð18Þ U
Ωe Ωe nþ1 nþ1 ð26Þ
U_ ¼ U_ þ γΔt U€
n;p

where:
β and γ are the classical Newmark parameters and the predictor
B~ ¼ B~ þdt B~
T eT pT
ð19Þ quantities are defined by:
Additional assumptions are required for the time stepping: 8  
> 1 n
>
< U n;p ¼ U n þ Δt U_ þ Δt 2  β U€
n
8
< ε_ n þ 1 ¼ εn þ 1  εn 2 ð27Þ
> n
:U_ ¼ U_ þ Δtð1  γ ÞU€
n;p n
dt ð20Þ >
:E
n þ 1 ¼ E n þ dt εn þ 1

Using the assumptions given in Eq. (20), the third equation of After substituting Eqs. (26) and (27) into the space discrete
the system in Eq. (3) leads to the expression of the strain ε^ n þ 1 at equation of the PML in Eq. (25), the acceleration at the end of the
the end of the time step [1]: time step can be obtained as:
 
1 1 ε Q ~ U€ n þ 1 ¼ F ext  Pðεn ; En ; Σ n Þ ðC þ C~ ÞU_ n;p  ðK þ K~ ÞU n;p
ε^ n þ 1 ¼ Bε U_ n þ 1 þ BQ U n þ 1 þ F^ ε^ n  F^ E^ n ð21Þ M ð28Þ
dt dt
ε Q with the effective stiffness matrix expressed as:
where Bε , BQ, F^ and F^ are matrices depending on the shape
functions and the damping functions (see Appendix A summar- ~ ¼ M þ γΔtðC þ C~ Þ þ βΔt 2 ðK þ K~ Þ
M ð29Þ
izing the above matrices, details can be found in [1]).
As previously noted, Basu and Chopra [1] proposed to integrate
The stress σ^ n þ 1 at the end of the time step is computed from
in time according to the displacement format. A Newton proce-
the strain ε^ n þ 1 thanks to the elastic constitutive matrix C. The
element-wise internal force pen þ 1 can now be written in terms of dure associated with the implicit time integration scheme was
adopted. In this work, the time integration procedure is also
the element vectors of velocity and displacement (U_
e
and U e ) nþ1 nþ1
implicit, but requires no Newton iteration because all the involved
as well as a term, denoted Pðεen ; Een ; Σ n Þ depending only on strain
e
global matrices are independent of displacements. As a con-
and stress tensors known at the beginning of the time step
sequence, the time integration procedure proposed here to
t n ¼ ndt. The element-wise internal force pen þ 1 is written as:
integrate the PML is simplified in comparison to the original
pen þ 1 ¼ c~ e U_ n þ 1 þ k~ U en þ 1 þ Pðεen ; Een ; Σ n Þ
e e e
ð22Þ procedure.
In the following, the coupling strategy based on a dual Schur
where the element matrices c~ e and k~ are defined as:
e
approach is presented in order to derive hybrid asynchronous
Z Z
1 1 PMLs which can be coupled with any Newmark time integration
B~ DBε k~ ¼
e
B~ DBQ
T T
c~ e ¼ ð23Þ
dt Ωe dt Ωe schemes in the domain of interest. Owing to the ability and the
efficiency of the explicit Central Difference scheme to deal with
D is the material constitutive matrix under the plan strain
the wave propagation problem, we focus on this time integrator in
assumption expressed as:
the domain of interest and adopt a Newmark implicit time inte-
2 3
κ þ 43μL κ  23μL 0 grator in the PML region.
6 7
D¼6 4 κ  3μL κ þ 3μL 0 5
2 4 7 ð24Þ
0 0 μL
3. Hybrid multi time step coupling between the physical
where μL is the shear modulus (or Lamé's second coefficient) and κ domain and PML
is the bulk modulus.
Finally, on the whole PML, the space discrete weak form at the Let Ω a bounded domain belonging to R2 with a regular
end time t n þ 1 of the time step h ¼ ½t n t n þ 1  can be expressed as: boundary. J ¼ ½0; T is the time interval of interest. We assume that
the domain Ω is divided into two parts Ω1 and Ω2 as illustrated in
M U€ n þ 1 þðC þ C~ ÞU_ n þ 1 þ ðK þ K~ ÞU n þ 1 þ Pðεn ; En ; Σ n Þ ¼ F ext ð25Þ
Fig. 1 such as: Ω1 \ Ω2 ¼ ∅ and ∂Ω1 \ ∂Ω2 ¼ Γ I . ΓI represents the
in which the global matrices C~ and K~ , related to velocities interface between the two subdomains. The subdomain Ω1 is
and displacements, respectively, are derived from the classical related to a non dissipative linear elastic behavior and the sub-
assembly procedure of the previous elementary matrices c~ e and domain Ω2 is related to the PML region previously presented.
k~ ; Pðεn ; En ; Σ n Þ is the known part of the internal force at the
e
It is assumed that the subdomain Ω1 is linear elastic, char-
beginning of the time step; M, C and K are the global matrices acterized by the density ρ1 and the Lamé coefficients λL1 and μL1 .
resulting from the assembly of the element matrices given E1 and ν1 are the Young modulus and the Poisson ratio. The wave
in Eq. (13). motion in the subdomain Ω1 is governed by the following system
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15 5

Z Z Z
c μ
ρ s f c v 2  u_ 2 dΩ þ
2
f k v 2
 _
u 2
d Ω þ ε~ ðv 2 Þe : σ dΩ
Ω2 b Ω2 b Ω2 2
Z
þ ε~ ðv 2 Þp : Σ dΩ þ ⋯
Ω2 2
Z Z Z
v 1  λ dΓ þ v 2  λ dΓ þ μ :ðu_ 1  u_ 2 ÞdΓ ¼ ⋯
ΓI ΓI ΓI
Z Z 
v 2  σ F~ þ Σ F~ ndΓ
e p
v1  g N
1
dΓ þ ð35Þ
Γ N1 Γ N2 2 2

The coupling between the subdomains Ω1 (physical medium) and

Ω2 (PML medium) is ensured via interface forces corresponding to
Lagrange multipliers according to a dual Schur formulation: the
Lagrange multipliers are introduced in order to ensure the kinematic
continuity at the interface ΓI and provide the interface forces. It is
Fig. 1. Domain Ω divided into two partitions Ω1 and Ω2.
called a dual approach because Lagrange multipliers are the new
unknowns of the problem, in contrast to the hybrid formulation
of equations: proposed by Kucukcoban and Kallivokas [18] based on primal Schur
8
>
> divðσ 1 Þ þ b 1 ¼ ρ1 u€ 1 in Ω1  J approach with the displacement at the interface as unknowns. In the
>
> continuous formulation, imposing any kinematic continuity at the
< σ ¼ C : ε in Ω  J
1 1
1 1 ð30Þ interface leads to the same solution. The situation is different after
>
> h i
>
> 1 discretization (finite element method for the space and Newmark
: ε ¼ gradðu 1 Þ þgradðu 1 ÞT in Ω1  J
1 2 methods for the time) for which one quantity has to be chosen for
gluing the subdomains between each other. On the basis of the FETI
with the boundary conditions:
8 approach for transient dynamics [29], Gravouil and Combescure
on Γ 1  J
D
< u1 ¼ uD
1
proposed a dual Schur method, called GC method, able to couple any
ð31Þ Newmark time integrators with their own time step [2,21]. In this
: ðσ Þn ¼ u N
1
on Γ PML  J
N
1 work, it has been demonstrated that coupling an explicit time inte-
grator, such as the CD scheme, with an implicit time integrator, such
whereas b 1 is the body force per unit volume, u D 1
is the Dirichlet
as the CAA scheme, requires a velocity continuity at the interface,
prescribed displacement and g N 1
the traction force.
otherwise the coupling algorithm becomes unstable. It is important
The weak form of the PML has been given previously in Eq. (11).
to note that the stability demonstration is carried out in the general
The weak form for the whole domain Ω requires the introduction
case of finite element subdomains following the energy argument,
of the test functions in the partitions Ω1 and Ω2, belonging to the
provided by the so-called energy method, proposed by Hughes [20]
appropriate spaces. The coordinates in subdomains Ω1 and Ω2 are
for demonstrating the stability of hybrid explicit/implicit time inte-
denoted by X 1 and X 2 , whereas the coordinates on the interface ΓI
gration procedures. The main advantages of the coupling GC method
are denoted by X I . The test functions v 1 and v 2 are now defined as:
8 n o is to deal with heterogeneous time integrators and multi time steps.
>
< v 1 ðX 1 ; tÞ A V 1  J ; V 1 ¼ v 1 þ regularity : v 1 ¼ 0 on Γ 1
⋆ ⋆ D
As a result, the GC method is a powerful framework to couple phy-
n o ð32Þ sical domains with unphysical ones such as the PMLs. In addition, the
>
2  J ; V 2 ¼ v 2 þ regularity : v 2 ¼ 0 on Γ 2
: v 2 ðX 2 ; tÞ A V ⋆ ⋆ D
multi-time step capability is of great interest for minimizing the
computation time required in the absorbing layer while keeping a
By using the principle of virtual power, we seek to find the solu- target accuracy in the subdomain of interest.
tions u 1 and u 2 belonging to the appropriate spaces: When introducing the spatial discretization using the finite
8 n o element method, the velocity continuity can be written in its
>
< u 1 ðX 1 ; tÞ A V 1  J ; V 1 ¼ u 1 þ regularity : u 1 ¼ u D on Γ 1
D
1 discrete format:
n o ð33Þ
>
: u 2 ðX 2 ; tÞ A V 2  J ; V 2 ¼ u 2 þ regularity : u 2 ¼ u D on Γ 2
D
2 L1 U_1 þ L2 U_2 ¼ 0 ð36Þ

The gluing between the subdomains Ω1 and Ω2 is ensured via
Lagrange multipliers defined on the interface according to a dual
Schur approach. So we introduce the Lagrange multiplier field λ
defined on the interface ΓI and its related test functions μ
belonging to the appropriate spaces:
8   continuity at the interface. The restricted velocity vectors on the
< λ ðX I ; tÞ A Q  J ; Q ¼ λ þ regularity
n o
ð34Þ
: μ ðX I ; tÞ A Q ⋆  J ; Q ⋆ ¼ μ þregularity
All the above considered space time variables are assumed to be
sufficiently smooth and regular.
Now, we can express the principle of virtual power for transient
structural dynamics problems coupling two subdomains with a
dual Schur approach. Find the solution u 1 A V 1  J, u 2 A V 2  J and
λ A Q  J, for which the following weak form is satisfied
where U_1 and U_2 are the nodal velocity vectors related to two
subdomains. L1 and L2 are boolean matrices (containing only 0 and
þ 1 in L1 and 0 and  1 in L2), operating on nodal vectors associated
with the two subdomains Ω1 and Ω2, and picks out the degrees of
freedom lying on the interface ΓI in order to ensure the kinematic
Γ Γ
interface are given by: U_1 ¼ L1 U_1 and U_2 ¼ L2 U_2 . The same
relationships hold for the restricted velocity vectors on the interface
related to the test functions v 1 and v 2 . According to the dual
approach, Lagrange multipliers are introduced so as to ensure the
velocity continuity. The interface terms in Eq. (35) can now be
expressed in a discrete format:
8R
< Γ v 1  λ dΓ ¼ V T1 LT1 λ
R
I
ð37Þ
: Γ v  λ dΓ ¼ V T2 LT2 λ
I 2

1  J, 8 v 2 A V 2  J and 8 μ A Q  J:
⋆
8 v1 A V ⋆ ⋆

Z Z Z Z where V1 and V2 denote the virtual nodal velocity vectors from the
ρ1 v 1  u€ 1 dΩ þ ε ðv 1 Þ : σ dΩ þ
1
v 1  b 1 dΩ þ ρf m v 2  u€2 dΩ þ ⋯ space discretization of the test functions v 1 and v 2 and λ is the
Ω1 Ω1 Ω1 Ω2
Lagrange multiplier vector at the interface. Afterwards, the interface
6 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15

problem can be expressed in a reduced problem at the interface in ~ 2 ¼ M 2 þ β Δt 2 ðK 2 þ K~ 2 Þ þ γ Δt 2 ðC 2 þ C~ 2 Þ

M ð44Þ
2 2 2
which the Lagrange multipliers are the unknowns.
The procedure proposed by Gravouil and Combescure is based
From the weak form of the coupled problem given in Eq. (35)
on splitting the kinematic quantities into two parts: the free and
and using the classical finite element space discretization, the
the linked quantities. The free quantities are obtained by only
standard application of the virtual power leads to the following
considering the internal and external forces, whereas the linked
system of discrete in space equations:
8 quantities are obtained by only taking into account the interface
€ loads defined by the Lagrange multiplier vector λ. If we consider
< M 1 U 1 þ K 1 U 1 ¼ F ext;1  L1 λ
>
>
T

M 2 U€ 2 þ ðC 2 þ C~ 2 ÞU_ 2 þ ðK 2 þ K~ 2 ÞU 2 þ P 2 ¼  LT2 λ ð38Þ the subdomain Ω2, the discrete equation of motion is split into
>
> two parts as follows:
: L U_ þ L U_ ¼ 0
1 1 2 2 8
<M ~ 2 U€ free;m ¼  P 2 ðε0 ; E0 ; Σ 0 Þ  ðC 2 þ C~ 2 ÞU_ 0;p  ðK 2 þ K~ 2 ÞU 0;p
The first equation is the discrete in space equation of motion of 2 2 2
ð45Þ
the subdomain Ω1, the second equation is the discrete in space :M ~ 2 U€ link;m ¼  LT λm
2 2
equation of motion of the subdomain Ω2, corresponding to the
PML medium as given in Eq. (25), and the third equation is the The complete accelerations are obtained by summing the two
m free;m link;m
discrete in space velocity continuity at the interface. Noting that parts as: U€ 2 ¼ U€ 2 þ U€ 2 . The same procedure is applied to
the external forces in the PML are assumed to be equal to zero. the subdomain Ω1 at each time tj. Furthermore, the kinematic
Using the coupling GC method, the partitions can be integrated in quantities of the subdomain 2 at tj are interpolated as well as the
time with any time integrators belonging to the Newmark family. In Lagrange multipliers:
this work, we adopt explicit time integrators in the physical partition 8  
>
> j j
because of its widely use in industrial finite element codes for >
< λ j
¼ 1  λ0 þ λm
m m
simulating the wave propagation (LS-DYNA [16]. The subdomain Ω1   ð46Þ
>
> j j 0 j m
is integrated in time with an explicit time integration >
: W2 ¼ 1m W2 þm W2
scheme (Central Difference scheme), characterized by the para-
meters γ 1 ¼ 0:5 and β1 ¼ 0, whereas the subdomain Ω2 is handled where W j2 denote the free or the linked velocities of the sub-
by an implicit time integration scheme (Constant Average Accelera- domain 2 interpolated at tj. Using the kinematic condition at the
tion scheme), characterized by the parameters γ 2 ¼ 0:5 and interface in Eq. (41) at each time tj, we obtain:
β2 ¼ 0:25. For this purpose, we define the fine time scale Δt 1 for the
L1 U_ 1 þ L2 U_ 2 ¼  L1 U_ 1  L2 U_ 2
link;j link;j free;j free;j
explicit domain and the coarse time scale Δt 2 for the implicit ð47Þ
domain with Δt 2 ¼ mΔt 1 . The discrete in space and time equation of Then, by using the second equation in Eqs. (45) and (46) as well
motion is written for the subdomain Ω2 (PML) at the large time scale as the expression of the linked velocities as a function of linked
tm with Δt 2 ¼ ½t 0 ; t m , while the discrete in space and time equation accelerations, the following interface problem can be derived [30]:
of motion of the subdomain Ω1 is written at the fine time scale tj
H λ ¼ bj
j
(j ¼ 1; 2; …m) with Δt 1 ¼ ½t j  1 ; t j  as follows: ð48Þ
with the interface operator and the second side member vector
 Subdomain 1 on the time step Δt 1 ¼ ½t j  1 ; t j : defined by:
j 8
 LT1 λ^ ~  1 L T þ γ Δt 2 L 2 M
~  1 LT
j
M 1 U€ 1 þ K 1 U j1 ¼ F ext;j
1 ð39Þ < H ¼ γ ΔtL1 M
1 1 1 2 2 2
ð49Þ
: b ¼ L U_ free;j þL U_ free;j
j 1 1 2 2
 Subdomain 2 on the large time step Δt 2 ¼ ½t 0 ; t m :
Finally, once obtained the Lagrange multiplier vector at the end
T ^m
M 2 U€ 2 þ ðC 2 þ C~ 2 ÞU_ 2 þ ðK 2 þ K~ 2 ÞU m
m m
2 þ P 2 ðε0 ; E 0 ; Σ 0 Þ ¼  L2 λ time of the time step, linked accelerations are obtained assuming
ð40Þ only interface forces at the right side of the discrete equation of
motion (second equation of the system in (45)). Linked displace-
ments and velocities are then derived from Newmark formulas
 At the interface, the continuity of velocities is imposed at times (Eqs. (26) and (27)) and added to the free quantities to complete
tj (at the fine time scale) for j ¼ 1; 2; …m as: the fine time step Δt 1 .
L1 U_ 1 þL2 U_ 2 ¼ 0
j j
ð41Þ

4. Numerical examples

From Newmark formulas in Eqs. (26) and (27), we introduce In the following applications, non harmonic waves will be
the predictor quantities related to the two time scales, that is investigated by considering a Ricker incident wave, denoted by Ric,
Δt 1 ¼ ½t j  1 ; t j  for subdomain Ω1 and Δt 2 ¼ ½t 0 ; t m  for subdomain defined as:
Ω2. It leads to the equations of motion for the two subdomains in 2
!
2
!
2 ðt  t s Þ 2 ðt  t s Þ
the form: Ricðt; t p ; t s Þ ¼ A 2π  1 exp  π ð50Þ
8 t 2p t 2p
>
> M~ 1 U€ j ¼ F ext;j  K 1 U j  1;p  LT λj for j ¼ 1; 2; …; m
>
< 1 1 1 1
The Ricker wave is characterized by three parameters: the fun-
M~ 2 U€ m ¼  P 2 ðε0 ; E0 ; Σ 0 Þ  ðC 2 þ C~ 2 ÞU_ 0;p  ðK 2 þ K~ 2 ÞU 0;p  LT λm
> 2 2 2 2 damental period tp, the time shift ts and the amplitude A. The
>
> _j
: L1 U 1 þ L2 U_ 2 ¼ 0 for j ¼ 1; 2; …; m
j
chosen values are: t p ¼ 3s, t s ¼ 3s and A ¼1, as illustrated in Fig. 2.

ð42Þ 4.1. Lamb's test

with the effective stiffness matrices defined for the two sub-
domains by: Lamb's test consists in applying a concentrated load, char-
acterized by the Ricker wave, to the surface of an infinite half
~ 1 ¼ M1 þ β Δ
M t 21 K 1 ð43Þ
1 space medium. In 1904, Lamb [31] analytically calculated the
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15 7

Fig. 2. Waveform and Fourier transform of the Ricker wavelet.

The soil is supposed to be linear elastic with μL1 ¼ 1, ν1 ¼ 0:25

qffiffiffiffiffiffi
μL
ffi
and ρ1 ¼ 1. The shear wave velocity cs ¼ ρ 1 is thus equal to 1.
1
Lamb's test is modeled using the bounded-domain-PML as illu-
strated in Fig. 3, composed of a bounded soil (subdomain 1) with a
size of 43 λS  waves (λS  waves ¼ cs t p being the wavelength of Shear
waves) and a PML (subdomain 2) characterized by a thickness Lp.
The bounded problem is discretized using the four-node bilinear
isoparametric elements.
Fig. 3. Modeling Lamb's test using PMLs. The soil is integrated using explicit time
scheme (Central Difference) while the PML is integrated using implicit time
scheme (Constant Average Acceleration).
displacements at a given point of the surface by assuming
an isotropic linear elastic behavior for the soil. The derived the-
oretical solution involves P, S-waves and Rayleigh waves: Lamb's
test is known to be characterized by a strong Rayleigh wave
excited by the load surface, providing a good test for the per-
formance of PMLs in absorbing surface waves as underlined in
the literature [32].
  The damping functions are chosen as
f i ¼ 0 and f i ¼ ap xi L x0 with ap ¼ 5 and x0 ¼ 4. The characteristic
e p
length of the problem, denoted by b in the definition of the
stretching function in Eq. (1), is taken equal to 1. The length Lp ¼2
is calculated using the relationship in Eq. (10) with R¼ 0.001 for
the reflection coefficient (R ¼ Rpp corresponding to P waves in Eq.
(10)). Finally, a point C located at the distance of 2 from the load
point is chosen in order to record the vertical and the horizontal
displacements.
In this part, the numerical solution is computed using the multi
time step GC method. Two computations are carried out: the full
explicit computation where the whole domain is integrated with
the explicit time integration scheme (CD), considering an exten-
ded mesh (equal to three times the size of the subdomain 1 in
Fig. 3) so as to provide the reference results, and the hybrid multi
time step computation, integrating the subdomain 1 with an
explicit scheme (Central Difference scheme), whereas the sub-
domain 2 (PML) is integrated with an implicit scheme (Constant
Average Acceleration scheme). The effect of the time step ratio
m¼Δ t2
Δt 1 between the two time scales on the result accuracy is
investigated by introducing an error norm between numerical
results. For this purpose, we define the L2 error norm between the
hybrid multi time step PML and the full explicit computation with
an extended mesh. Considering a quantity E over the time interval
½0; T, the L2 norm is defined by:
J EPML  Eref J L2 ð½0;TÞ
err ¼ ð51Þ
J Eref J L2 ð½0;TÞ
8 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15

Fig. 4. Vertical and horizontal displacements at point C using hybrid PMLs (m¼ 1) compared to the reference results using an extended mesh.

Fig. 5. Kinetic and internal energies computed using hybrid PMLs (m¼ 1) compared to the reference results using an extended mesh.

where EPML is obtained using hybrid multi time step PML and Eref is Table 1
obtained using a full explicit computation with an extended mesh. L2 error norms in energies and displacements obtained from hybrid multi time step
PMLs for different time step ratios m (PML Δt 2 , with Δt 2 ¼ mΔt 1 ).
In the case of an homogeneous time step in both subdomains,
Figs. 4 and 5 display the time-histories of displacements and Time step Kinetic Internal Horizontal Vertical
energies obtained by the hybrid PML in comparison to the refer- ratio energy (%) energy (%) displacement (%) displacement (%)
ence results, highlighting the excellent behavior of the hybrid PML
for Lamb's test. It can be seen that high accuracy is achieved. It is m¼ 1 0.0002 0.003 0.001 0.001
m¼ 3 0.3 0.1 1.8 1.3
confirmed by the computation of the L2 norm summarized in m¼ 5 0.7 0.3 3.6 2.6
Table 1: an error of 0:001% is achieved in displacements and less
than 0:0003% in energies.
Figs. 6 and 7 compare the results obtained using the hybrid be explained by the following points. First, the time step in the
explicit/implicit multi time step computation for different values
implicit PML increases, leading to an increase of numerical errors
of the time step ratio m with the reference results. One can remark
in PML. Moreover, the GC coupling algorithm is known to be dis-
that the different quantities (displacements and energies) remain
sipative as soon as different time steps are adopted in the sub-
in a very good agreement with respect to reference results even if
small discrepancies can be observed in the case of m ¼ 5. Table 1 domains. As a consequence, the incident wave is altered by the
summarizes the errors computed for each quantity by varying the interface and spurious energies can be generated, coming pollute
ratio m from 1 to 5. It can be observed that the errors remain small the solution in the physical subdomain. Finally, the discrepancies
and do not exceed 3.6% in displacements and 0.7% in energies. The are also due to the discrete PML formulation itself: indeed, it is
observed decrease of accuracy as the time step ratio increases can well known that the perfect PML in the continuous problem
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15 9

Fig. 6. Vertical and horizontal displacements plotted at point C using hybrid multi time step PMLs for different values of the time step ratio m compared to the reference
results using a full explicit integration and an extended mesh.

Fig. 7. Kinetic and internal energies computed using hybrid multi time step PMLs for different values of time step ratio m compared to the reference results using a full
explicit integration and an extended mesh.

becomes not perfect as soon as spatial and time discretizations are

employed.

4.2. Rigid strip on a homogeneous isotropic elastic half-plane

In Fig. 8, the classical problem of the rigid strip-footing (strip

Fig. 8. Rigid strip on a homogeneous soil. The soil is integrated using explicit time
scheme (Central Difference) while the PML and the rigid strip are integrated using
implicit time scheme (Constant Average Acceleration).
Young's modulus very important with respect to the soil Young's
modulus) on a half plane is considered. In this application, the
strip represents the third subdomain Ω3, integrated in time using
an implicit time scheme with the same time step as the hybrid
multi time step PML. In soil structure interaction problem, it
represents a very nice feature for the proposed hybrid multi time
step approach because the most appropriate time integration
method can be selected for the PML and the structure, indepen-
dently of the soil which is dealt with the same explicit scheme and
time step as in previous Lamb's test. The mesh of the third
10 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15
Fig. 9. Vertical and horizontal displacements plotted at point C using the hybrid PMLs (m¼ 1) compared to the reference results using an extended mesh.
subdomain Ω3 matches the mesh of the soil subdomain Ω2 as
displayed in Fig. 8.
Three different subdomains are considered: the solid medium
corresponding to the rigid strip (subdomain 3), the soil medium
(subdomain 1) and the PML medium (subdomain 2). The size of
the soil medium is 4 in the horizontal and depth directions,
whereas the PML layer depth Lp is set to 2 as in previous Lamb's
test. A recording point, located at the soil surface (C point located
at 2 m from the symmetry vertical axis as shown in Fig. 8), is
considered in order to assess the efficiency of the PML layers to
model an infinite half-space medium. The reference results are
obtained from the full explicit computation with an extended
mesh equal to three times the size of the soil subdomain 1. The
rigid strip on the soil is characterized by a thickness of 0.5, cor-
responding to the finite element size, and Young's modulus equal
to 100 times the one of soil medium. The load is defined by a
Ricker wave with the same parameters as in Lamb's test and is
applied uniformly to the strip as shown in Fig. 8. The same
properties for the soil and the PML given in previous Lamb's test
are adopted. The comparison is carried out in terms of displace-
ments computed at point C and energies by summing the energies
in the soil (subdomain 1) and the rigid strip (subdomain 3).
Fig. 9 compares the displacements at C obtained using the
hybrid PML and the extended mesh (reference results). It can be
observed that the results closely match the reference results with
an error, given in Table 2, less than 0.2% when homogeneous time
step is employed in the three subdomains.
Figs. 10 and 11 display the displacements and energies obtained
using the proposed strategy for an increasing time step ratio m in
comparison to the reference results. Although the observed accu-
racy decreases as the time step ratio increases, it can be noted that
the different curves are quite close with an error less than 5% in
displacements and 2.4% in energies (Table 2). Furthermore, an
important reduction in computation time can be highlighted in
comparison to the full computation with the extended mesh:
computation time reduction factors of 6, 16 and 21 are achieved
for m ¼1, m ¼ 3 and m ¼5, respectively.
4.3. Rigid strip on a layered isotropic elastic half-plane
The case of a rigid foundation on a layered medium is described
in Fig. 12. Again, three different subdomains are considered: the
solid medium corresponding to the foundation (subdomain 3), the
Table 2
L2 error norms in energies and displacements computed for the rigid strip footing
on an elastic homogeneous soil medium and for different time step ratios m (PML
Δt 2 , rigid strip Δt 3 , with Δt 2 ¼ Δt 3 ¼ mΔt 1 ).
Time step Kinetic Internal Horizontal Vertical
ratio energy (%) energy (%) displacement (%) displacement (%)
m¼ 1 0.2 0.1 0.8 0.6
m¼ 3 1.3 0.9 3 2
m¼ 5 2.4 1.8 5 3.7
soil medium (subdomain 1) and the PML medium (subdomain 2).
The size of the soil medium is 4 according to the horizontal and
depth directions, whereas the PML layer depth Lp is taken equal to
2. A recording point, located at the surface of the soil (C point at
2 m from the symmetry vertical axis as shown in Fig. 12), is con-
sidered in order to assess the efficiency of the PML layers to model
an infinite half-space medium. As in previous example, Young's
modulus of the foundation is equal to 100 times the one of the soil
medium. As a result, it is of great interest to adopt implicit time
integration schemes (Constant Average Acceleration scheme) for
the solid medium and explicit time integration scheme (Central
Difference scheme) for the wave propagation into the soil medium.
The third subdomain is the PML integrated in time with an implicit
time integration scheme (CAA). The main difference in this tests
with respect to the previous one lies in the validation of the
implicit multi time step PML approach in the case of a layered
medium.
The soil medium is composed of two layers, the second layer
being characterized by a shear modulus five times the one of the
surface layer. The stiffness contrast between the two layers leads
to a series of wave reflexions and refractions at the interface of the
two layers: it entails a dispersive propagation wave problem
similar to a guided wave propagation problem.
The mechanical properties of the surface soil are denoted by:
μL1 ¼ 1, ν1 ¼ 0:25, ρ1 ¼ 1 for the shear modulus, the Poisson coefficient
and the density, respectively. The mechanical parameters of the
underneath layer are given by: μL2 ¼ 5μL1 , ν2 ¼ ν1 , ρ2 ¼ ρ1 . The
interface between the layers has to be taken into account into the PML,
leading to two PMLs with different mechanical properties to match the
physical layers. The efficiency of the hybrid multi time step PML
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15 11

Fig. 10. Vertical and horizontal displacements plotted at point C using hybrid multi time step PMLs for different time step ratios m compared to the reference results using a
full explicit integration and an extended mesh.

Fig. 11. Kinetic and internal energies computed using hybrid multi time step PMLs for different time step ratios m compared to the reference results using a full explicit
integration and an extended mesh.

approach is assessed with respect to a reference problem associated

with an extended mesh (3 times in horizontal and vertical directions
the soil medium subdomain shown in Fig. 12) and a time step satisfying
the CFL condition imposed by the mechanical properties of the rigid
foundation, that is Δt 3 ¼ 0:003 s.
In the subdomain approach, the soil medium is integrated
using the CD scheme with a time step Δt 1 ¼ 0:025 s, whereas the
rigid foundation and the PML are integrated using the CAA
scheme, with a time step Δt 3 ¼ Δt 2 ¼ mΔt 1 , m denoting the time
step ratio. The damping functions for propagative waves in the
PML are defined by: f i ¼ ap ðxi  x0 Þ
p
Lp , index i being related to the X or
Y direction. The slope coefficient of the damping function in the
PML depth is taken as ap ¼ 2:5.
Considering first a time step ratio m ¼ 1, the displacement and
velocity time histories recorded at point C are compared in Fig. 13
with the reference results obtained with an extended mesh. Here,
the coupling is only between different time integrators because Fig. 12. Rigid strip on a layered soil. The soil is integrated using explicit time
the same time step size is adopted in all the subdomains. From scheme (CD) while the rigid strip and PMLs are integrated using implicit time
Fig. 13, it can be seen that excellent agreement is achieved in scheme (CAA).
12 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15

Fig. 13. Displacements at point C with a time step ratio m¼ 1 compared to the reference results using a full explicit computation and an extended mesh.

Fig. 14. Displacements at point C for different time step ratios m compared to the reference results using a full explicit computation and an extended mesh.

Fig. 15. Kinetic and internal energies computed using time heterogeneous implicit PMLs for different values of time step ratios m compared to the reference results using a
full explicit integration with an extended mesh.
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15 13

comparison to the reference results from the extended mesh. It is
important to note that the time steps between the two compared
simulations are not the same: Δt ¼ 0:003 s for the full explicit
computation (reference results) and Δt ¼ 0:025 s for the hybrid
explicit/implicit simulation. As a result, an error is expected only
due to the difference in the time step size.
The time history displacements are also plotted in Fig. 14 for an
increasing time step ratio ranging from m¼1 to m¼ 5. The predic-
tion using the hybrid multi time step PML approach remains in a
good agreement with respect to the reference results for these time
step ratios. However, discrepancies can be observed for the largest
time step ratio, indicating numerical errors coming from the fol-
lowing two points: first, the increase of the time step size Δt 2 in
PML leading to increased spurious reflections at the interface
between the soil and the PML, and second, the classical drawback of
the GC coupling algorithm which is known to be dissipative at the
interface as soon as different time steps are considered. This energy
dissipative feature has been demonstrated on the basis of the so-
called energy method in [2] and has been highlighted in numerical
examples (Brun et al. [33,30,34]).
The time history energies contained in subdomains 1 (soil
medium) and 3 (rigid strip) are displayed in Fig. 15 for an
increasing time step ratio ranging from m ¼1 to m ¼5. It can be
seen that all the incident energy coming from the loading on the
strip is damped out thanks to the hybrid multi time step PML. The
discrepancies in terms of energies and displacements at the point
C can be summarized in Table 3 according to the L2 ð½0; TÞ norm.
Again, the accuracy of the proposed approach can be highlighted
when keeping low time step ratio between the large time step
(implicit subdomains: strip and PML) and the fine time step
(explicit subdomain: soil).
Two computations are added in Table 3 with a finer time step
size Δt 1 ¼ 0:003 for the soil medium, corresponding to the time step
of the full explicit computation (reference results). In other word,
the time step in the soil is taken as less than the CFL critical time
step in the rigid strip. Again, we adopt for the time step sizes in the
strip and PML: Δt 3 ¼ Δt 2 ¼ mΔt 1 . The excellent agreement of the
PML with respect to the full explicit results has to be highlighted,
achieving an error of 0.005% in energies and 0.05% in displacements.
The case of m¼5 is again considered. The very weak errors tends to
demonstrate that result accuracy mainly depends on the time step
in the PMLs: the larger is the time step in PML, the lower is the
accuracy; moreover, the time step ratio m seems to have a minor
effect if the time step in the PML remains sufficiently low.
Finally, long-time simulations are added to illustrate the sta-
bility of the multi time step hybrid approach. For a time step
Δt 1 ¼ 0:003 in the soil medium and a time step ratio m equal to 5,
we run the simulation for a total duration of 120 s, corresponding
to 40,000 micro time steps. In addition, a long-time with a time
step Δt 1 ¼ 0:025 in the soil medium and a time step ratio m ¼1 is
carried out. The responses in terms of vertical displacements at the
recording point C are compared in Fig. 16 to the reference results
obtained from a full explicit computation. It can be observed that
the motion died out after approximately 40 s. In Fig. 16, the decay
of the mechanical energy (sum of the kinetic and internal ener-
gies) is plotted in a logarithmic scale. It can be highlighted that no
numerical instability is observed during the simulation time for
homogeneous and heterogeneous time steps.

Table 3
L2 error norms related to the energies (kinetic and internal) and displacements
(X and Y directions at point C) for a time step ratio m ranging from 1 to 5 and two
5. Conclusion
values of the time step Δt 1 in the soil medium (PML Δt 2 , rigid strip Δt 3 , with
Δt 2 ¼ Δt 3 ¼ mΔt 1 ). An hybrid asynchronous approach in 2D transient dynamics
has been presented in order to couple physical media with
Time step ratio Kinetic Internal Horizontal Vertical
unphysical ones such as the Perfectly Matched Layers whose
energy (%) energy (%) displacement (%) displacement (%)
purpose is to damp out propagating waves at the frontiers of a
m ¼ 1; Δt 1 ¼ 0:025 0.2 0.1 0.6 0.3 truncated finite element mesh so as to mimic infinite media. The
m ¼ 3; Δt 1 ¼ 0:025 1.2 0.8 2.1 1.1 PML formulation adopted in this work was the one proposed by
m ¼ 5; Δt 1 ¼ 0:025 2.2 1.6 3.6 1.8
Basu and Chopra, corresponding to an unsplit formulation, dis-
m ¼ 1; Δt 1 ¼ 0:003 0.005 0.005 0.02 0.05
m ¼ 5; Δt 1 ¼ 0:003 0.1 0.1 0.3 0.2 cretized in space by the finite element method, written in the time
domain and integrated via implicit Newmark time integrators [1]

Fig. 16. (a) Displacements at point C for different time step ratios m compared to the reference results (long-time simulation; reference solution truncated at 12 s).
(b) Mechanical energies computed using time homogeneous and heterogeneous implicit PMLs compared to the reference results.
14
and implemented in an industrial finite element code in its 3D
explicit version [15,16]. The hybrid asynchronous approach is
based on the dual Schur approach, more precisely along the lines
of the coupling method proposed by Gravouil and Combescure
[21,2]. The derived hybrid asynchronous PLMs are integrated in
time according to an implicit time integration (Constant Average
Acceleration scheme) with potentially large time steps, while the
physical medium is handled by an explicit time integrator, the
widely used Central Difference scheme, suitable for simulation of
wave propagation problems. First, the adopted PML formulation
has been summarized and the time stepping procedure has been
simplified, avoiding Newton iterations as proposed by Basu and
Chopra. Then the coupling framework is presented in its weak
form, followed by the finite element discretization and the hybrid
asynchronous time stepping procedure for coupling in time the
physical subdomain with the PML region. Different time inte-
grators with their own time step can be adopted depending on the
considered subdomain, making the proposed framework very
useful to couple complex PML formulations while conserving
classical time integrators in other subdomains. In order to high-
light the accuracy of the hybrid asynchronous PMLs, numerical
applications have been investigated such as Lamb's test, rigid
footing on a half space and on a layered half space. Excellent
agreement has been achieved in comparison to the full explicit
computation with an extended mesh. The result accuracy
decreases when large time steps in the PML are considered. Other
PML formulations could be employed using the presented frame-
work. Weak coupling in time could also be investigated on the
basis of the recent works [26,27], enabling to adopt more
advanced time integrators as α methods for damping out the high
frequencies due to the finite element discretization.
Appendix A.
ϵ Q
The matrices Bϵ , BQ, F^ and F^ in Eq. (21) are given by Basu and
Chopra [1]. Here, we remind the following definitions:
 1
Fe
Fϵ ¼ FeFl
Fl ¼ Fp þ FQ ¼ FpFl
dt
The matrix Bϵ is given in terms of the submatrix for the node I
by:
2 3
F ϵ11 N lI1 F ϵ21 N lI1
6 7
BϵI ¼ 6
4 F ϵ12 N lI2 F ϵ22 N lI2
F ϵ11 N lI2 þ F ϵ12 NlI1 F ϵ21 N lI12 þF ϵ22 N lI1
with:
N lIi ¼ F lij N I;j
The matrix BQ is defined, similarly, by replacing F ϵ with FQ.
ϵ
Finally, the matrix F^ is given by:
M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15
References
[1] U. Basu, A.K. Chopra, Perfectly matched layers for transient elastodynamics of
unbounded domains, Int. J. Numer. Methods Eng. 59 (2004) 1039–1074.
[2] A. Gravouil, A. Combescure, A multi-time-step explicit-implicit method for
non-linear structural dynamics, Int. J. Numer. Methods Eng. 50 (2001)
199–225.
[3] P. Bettess, Infinite elements, Int. J. Numer. Methods Eng. 11 (1977) 53–64.
[4] J. Su, Y. Wang, Equivalent dynamic infinite element for soil–structure inter-
action, Finite Elem. Anal. Des. 63 (2013) 1–7.
[5] B. Enquist, A. Majda, Absorbing boundary conditions for the numerical
simulation of waves, Math. Comput. 31 (1977) 629–665.
[6] J.F. Semblat, L. Lenti, A. Gandomzadeh, A simple multi-directional absorbing
layer method to simulate elastic wave propagation in unbounded domains,
Int. J. Numer. Methods Eng. 85 (2011) 1543–1563.
[7] P. Rajagopal, M. Drozdz, E.A. Skelton, M.J.S. Lowe, R.V. Craster, On the use of
the absorbing layers to simulate the propagation of elastic waves in
unbounded isotropic media using commercially available finite element
packages, NDT&E Int. 51 (2012) 30–40.
[8] W.C. Chew, Q.H. Liu, Perfectly matched layers for elastodynamics: a new
absorbing boundary condition, J. Comput. Acoust. 4 (1996) 341–359.
[9] J. Bérenger, A perfectly matched layer for the absorption of electromagnetic
waves, J. Comput. Phys. 114 (1994) 185–200.
[10] F. Hastings, J.B. Schneider, S.L. Broschat, Application of the perfectly matched
layer (pml) absorbing boundary condition to elastic wave propagation, J.
Acoust. Soc. Am. 100 (1996) 3061–3069.
[11] F. Collino, C. Tsoga, Application of the perfectly matched layer model to the
linear elastodynamic problem in anisotropic heterogeneous media, Geo-
physics 66 (2001) 294–307.
[12] T. Wang, X. Tang, Finite-difference modeling of elastic wave propagation: a
nonsplitting perfectly matched approach, Geophysics 68 (2003) 1749–1755.
[13] R. Matzen, An efficient finite element time-domain formulation for the elastic
second-order wave equation: a non split complex frequency shifted con-
volutional pml, Int. J. Numer. Methods Eng. 88 (2011) 951–973.
[14] U. Basu, A.K. Chopra, Perfectly matched layers for time-harmonic elastody-
namics of unbounded domains: theory and finite-element implementation,
Comput. Methods Appl. Mech. Eng. 192 (2003) 1337–1375.
[15] U. Basu, Explicit finite element perfectly matched layer for transient three-
dimensional elastic waves, Int. J. Numer. Methods Eng. 77 (2009) 151–176.
[16] LS-Dyna, Recent Developments in LS-DYNA, Version 971-R4 developments,
Livermore Software Technology Corporation, LSTC.
[17] S. Kucukcoban, L.F. Kallivokas, Mixed perfectly-matched-layers for direct
transient analysis in 2d elastic heterogeneous media, Comput. Methods Appl.
Mech. Eng. 200 (2011) 57–76.
[18] S. Kucukcoban, L.F. Kallivokas, A symmetric hybrid formulation for transient
wave simulations in pml-truncated heterogeneous media, Wave Motion 50
(2013) 57–79.
[19] A. Fathi, B. Poursartip, L.K. Kallivokas, Time-domain hybrid formulations for
wave simulations in three-dimensional PML-truncated heterogeneous media,
Int. J. Numer. Methods Eng. 101 (2015) 165–198.
[20] T.J.R. Hughes, The Finite Element Method: Linear Static and Dynamic Finite
Element Analysis, Prentice-Hall, Englewood Cliffs, NJ, 1987.
ð52Þ [21] A. Combescure, A. Gravouil, A numerical scheme to couple subdomains with
different time-steps for predominantly linear transient analysis, Comput.
Methods Appl. Mech. Eng. 191 (2002) 1129–1157.
[22] N.M. Newmark, A method of computation for structural dynamics, J. Eng.
Mech. Div., 85, (1959) 67–94.
[23] N. Mahjoubi, A. Gravouil, A. Combescure, Coupling subdomains with hetero-
geneous time integrators and incompatible time steps, Comput. Mech. 44,
2009, 825-843, http://dx.doi.org/10.1007/s00466-009-0413-4.
7
5 ð53Þ [24] E. Zafati, M. Brun, I. Djeran-Maigre, F. Prunier, Multi-directional and multi-
time step absorbing layer for unbounded domain, Comptes Rendus Mec. 342
(2014) 539–557.
[25] E. Zafati, M. Brun, I. Djeran-Maigre, F. Prunier, Design of an efficient multi-
directional explicit/implicit Rayleigh absorbing layer for seismic wave pro-
pagation in unbounded domain using a strong form formulation, Int. J. Numer.
Methods Eng., Published online in Wiley Online Library (wileyonlinelibrary.
ð54Þ com), http://dx.doi.org/10.1002/nme.5002, 2015.
[26] M. Brun, A. Gravouil, A. Combescure, A. Limam, Two FETI-based heterogeneous
time step coupling methods for Newmark and α-schemes derived from the
energy method, Comput. Methods Appl. Mech. Eng. 283 (2015) 130–176.
[27] A. Gravouil, A. Combescure, M. Brun, Heterogeneous asynchronous time
2  ϵ 2  ϵ 2 3 integrators for computational structural dynamics, Int. J. Numer. Methods Eng.
F 11 F 21 F ϵ11 F ϵ21 102 (2015) 202–232.
ϵ 6  2  ϵ 2 7
F^ ¼ 6 7
ϵ [28] U. Basu, A.K. Chopra, Erratum to: perfectly matched layers for transient elas-
4 F 12 F 22 F ϵ12 F ϵ22 5 ð55Þ todynamics of unbounded domains, Int. J. Numer. Methods Eng. 61 (2004)
ϵ ϵ
2F 11 F 12 2F ϵ21 F ϵ22 F ϵ11 F ϵ22 þ F ϵ12 F ϵ21 156–157.
[29] C. Farhat, L. Crivelli, F.X. Roux, Transient FETI methodology for large-scale
Q parallel implicit computations in structural mechanics, Int. J. Numer. Methods
Likewise, the matrix F^ is obtained by replacing F ϵ with FQ. Eng. 37 (1994) 1945–1975.
 M. Brun et al. / Finite Elements in Analysis and Design 122 (2016) 1–15
[30] M. Brun, A. Batti, A. Limam, A. Combescure, Implicit/explicit multi-time step
co-computations for predicting reinforced concrete structure response under
earthquake loading, Soil Dyn. Earthq. Eng. 33 (2012) 19–37.
[31] H. Lamb, On the propagation of tremors over the surface of an elastic solid,
Proc. R. Soc. Lond. 72 (1903) 128–130.
[32] S. Ma, P. Liu, Modeling of the perfectly matched layer absorbing boundaries
and intrinsic attenuation in explicit finite-element methods, Bull. Seismol. Soc.
Am. 96 (2006) 1779–1794.
View publication stats
15
[33] M. Brun, A. Batti, A. Combescure, A. Gravouil, External coupling software based
on macro- and micro-time scales for explicit/implicit multi-time-step co-
computations in structural dynamics, Finite Elem. Anal. Des. 86 (2014)
101–119.
[34] M. Brun, A. Batti, A. Limam, A. Gravouil, Explicit/implicit multi-time step co-
computations for blast analyses on a reinforced concrete frame structure,
Finite Elem. Anal. Des. 52 (2012) 41–59.