Science Arts & Métiers (SAM)
is an open access repository that collects the work of Arts et Métiers Institute of
Technology researchers and makes it freely available over the web where possible.
This is an author-deposited version published in:
https://sam.ensam.eu
Handle ID: .
http://hdl.handle.net/10985/7540
To cite this version :
Mohamed JEBAHI, Jean-Luc CHARLES, Frédéric DAU, Lounès ILLOUL, Ivan IORDANOFF - 3D
coupling approach between discrete and continuum models for dynamic simulations
(DEM–CNEM) Computer Methods in Applied Mechanics and Engineering Vol. 255, p.196209
-2013
simulations (DEM-CNEM) MohamedJEBAHI a ,Jean-lu CHARLES b ,Frederi DAU b ,Lounes ILLOUL ,Ivan IORDANOFF b a
Univ. Bordeaux,I2M,UMR5295,F-33400Talen e,Fran e b
ArtsetMetiers ParisTe h,I2M,UMR5295CNRSF-33400,Talen e,Fran e
ArtsetMetiersParisTe h,151,Boulevardde l'Hpital,F-75013 Paris,Fran e
Abstra t
The oupling between two dissimilar numeri al methods presents a major hallenge, espe iallyin ase of
dis rete- ontinuum oupling. The Arlequin approa h provides a exible framework and presents several
advantagesin omparisontoalternativeapproa hes. Manystudieshaveanalyzed,instati s,theingredients
of this approa h in1D ongurations underseveral parti ular onditions. Thepresent study extendsthe
Arlequin parameter studiesto in orporate a dynami behavior using3D models. Based on thesestudies,
a new 3D oupling method adapted for dynami simulations is developed. This method ouples two 3D
odes: DEM-based odeandCNEM-based ode. The3D ouplingmethodwasappliedtoseveralreferen e
dynami stests. Goodresultsareobtainedusingthismethod, omparedwiththeanalyti alandnumeri al
results ofboth DEM andCNEM.
Keywords: multis ale method, oupling method, dis reteelement, natural element,Arlequin approa h,
dynami simulation
1. Introdu tion
Thedis rete element method (DEM)[1 , 2℄ presents an alternative way to study physi al phenomena
requiring a very small s ale analysis or those whi h annot be easily treated by ontinuum me hani s,
su h as wear, fra ture and abrasionproblems. Inthe past de ades,an in reasing interest inthe dis rete
element method has led to the development of many interesting variations of this method. The most
re ent variation involves modeling the intera tion between parti les by ohesive beams [1 ℄. Thismethod
orre tly simulates the 3D linear elasti behavior of the ontinua. However, numeri al simulations are
verytime onsuming(CPU-wise). Furthermore,averygreatnumberofparti lesarerequiredtodis retize
small domains. This methoddoes not onsider large stru ture simulations. However, inmost situations,
theee ts thatmust be aptured byDEM are lo alizedina small portion ofthestudied domain. Thus,
theuseofaspe i multis alemethodto treatthephenomena at ea hs ale appears to be advantageous.
A hallengethatarisesinthemultis ale oupling approa histhatthehighfrequen yportionofwavesare
oftenspuriouslyree tedatthesmall/ oarses aleinterfa e. Thisphenomenonhasalreadybeenaddressed
using theniteelement modelwithdierent element sizes[3℄.
Theimportan eof thismultis aleapproa h hasattra tedmanyresear hers. Therefore,severalpapers
have been published on the subje t, and many oupling methods have been developed. These methods
an bedivided intwo lasses: edge to edgemethods and methods withoverlapping zones( alled overlap
methods). The rst lass [4, 5℄ is mainly applied to stati studies. Indeed, using this method, it is very
di ult to redu e spurious ree tions at the interfa e between models. Therefore, this lass will not be
treatedinthispaper. These ond lassseemstobemoreappli abletodynami studies,whi histhes ope
ofthe present work.
BenDhia [6, 7, 8℄, in a pioneer work, developed the Arlequin approa h asa general framework that
with the mole ular dynami s su h that the two Hamiltonians are averaged in an overlapping zone. A
damping wasusedin the overlapping zoneto redu e the spuriousree tions at theinterfa e between the
two models. Nevertheless, the hoi e ofthedamping oe ient remains di ult.
Smirnovaetal. [11℄developeda ombinedmole ulardynami s(MD)andniteelementmethod(FEM)
model witha transition zonein whi h the FEM nodes oin idewith thepositions of theparti les inthe
MDregion. Theparti lesinthetransitionzoneintera t viatheintera tion potential withtheMDregion.
At the sametime, they experien ethenodalfor es dueto theFEMgrid.
Belyts hkoandXiao[4,12℄havedevelopeda oupling methodfor mole ulardynami s and ontinuum
me hani s models based on a bridging domain method. In this method, the two models are overlaid at
theinterfa e and onstrained withaLagrange multiplier modelintheoverlappingsubdomain.
Ja obetal. [13 ℄formulatedanatomisti - ontinuum ouplingmethodbasedonablendofthe ontinuum
stress andtheatomisti for e. In term ofequations, thismethod isvery similarto theArlequin method.
In an interesting work, Chamoin et al. [14 ℄ have analyzed the main spurious ee ts in the
atomi -to- ontinuum oupling approa hes andtheyproposed a orre tive method basedonthe omputation and
inje tion ofdead for es intheArlequinformulation to osetthese ee ts.
Aubertin et al. [15℄ applied the Arlequin approa h to ouple the extended nite element method
XFEMwiththemole ular dynami s MD to study dynami ra kpropagation.
Baumanetal. [16 ℄developeda3Dmultis alemethod,basedontheArlequinapproa h,betweenhighly
heterogeneous parti le modelsand nonlinear elasti ontinuum models.
Re ently,Combes ure etal. [17 ℄ formulated a 3D oupling method, applied for fasttransient
simula-tions, between the smoothed parti le hydrodynami s SPH and the nite element method. This oupling
methodis, also,basedon theArlequin approa h.
Formoredetails,areviewofthesemethods anbefoundin[18 ℄. A ommonfeatureofoverlap oupling
methods is that a weight fun tion is introdu ed to partition a ertain quantity in the overlapping zone.
Herein, theArlequin approa h [6 , 7,19℄is usedto develop a 3Dmultis alemethod adapted for dynami
simulations between the onstrained natural element method (CNEM) and the dis rete element method
(DEM). TheDEM version, whi h isusedinthis work,isthe most re ent versiondeveloped byAndré[1℄.
TheCNEMisamesh-freemethod,butitisvery losetotheniteelement method. The ouplingmethod
developed here an avoid spurious wave ree tions without any additional ltering or damping. Indeed,
thenes ale solutionisproje tedonto the oarses alesolutionintheoverlappingzoneatea htimestep.
Thus, it ltersthe high frequen ies oming fromthe ne s ale model(dis rete model),whi h aregreater
thanthe utofrequen yofthe oarses alemodel( ontinuummodel). Thispaperisorganizedasfollows:
in Se tion 2, the governing equations of both the DEM and CNEM models are given. Subsequently, we
des ribehowbothmodels are oupledusingthe Arlequinapproa hinthemostgeneral ase. InSe tion3,
several interesting previous studies on the Arlequin parameter are summarized: in luding mathemati al
studiesofBenDhiaetal. [7, 20℄,the studiesofBauman etal. [21 ℄ andthestati 1Dnumeri al studiesof
Guidaultetal. [22 ℄. After,thedierent Arlequinparametersarestudieddynami allyusing3Dmodels. In
Se tion 4,this new oupling method is validated for tensile- ompression, bending and torsional loadings
on beams. Se tion 5 presents the on lusionsand outlooks.
2. The problem statements
Adomain
Ω
is onsideredwithboundary∂Ω = ∂Ω
u
+ ∂Ω
T
su hthat displa ementsand tra tionsare
pres ribed on
∂Ω
u
and
∂Ω
T
,respe tively. Thisdomainisdividedintotwosubdomains,
Ω
C
andΩ
d
,whi h are modeled using the ontinuum approa h and the dis rete approa h, respe tively. An isotropi linearelasti behaviorand small deformation gradients are assumed for simpli ity. The governing equations of
both the ontinuum and the dis rete subdomains are re alled in Subse tions 2.1 and 2.2, while ignoring
the oupling onditions. These onditions will be introdu ed after detailing the oupling approa h in
2.1. Continuum subdomain
Ω
C
As an isolated system, the governing equations in the ontinuum subdomain
Ω
C
an be written as:∀ x ∈ Ω
C
(t)
andt ∈ [0, T ]
,given theinitial onditions, nd(u, σ) ∈ [H
1
(Ω
C
)]
3
× [L
2
(Ω
C
)]
6
su hthat:
div(σ) + ρf
= ρ¨
u
in
Ω
C
σ
= A : ε(u)
ε(u)
=
1
2
(∇u + ∇
t
u)
u
= u
d
on ∂Ω
u
C
σ.n
= T
d
on ∂Ω
T
C
(1)where
ρ
is the density,u
is the ontinuum displa ement ve tor,σ
is the Cau hy stress tensor,ε
is the straintensor,A
is the stinesstensor,f
isthe bodyfor eve tor,u
d
denesthepres ribed displa ement ve tor on∂Ω
u
C
andT
d
is thepres ribed tra tionve toron∂Ω
T
C
. The asso iated weak formulation an be written as: ndu
∈ U
ad
su h that, given the initial onditions,
∀ δu ∈ U
ad,0
:ˆ
∂Ω
T
C
δ ˙u · T
d
dΓ −
ˆ
Ω
C
ε(δ ˙u) : A : ε(u) dΩ +
ˆ
Ω
C
ρ δ ˙u · f dΩ =
ˆ
Ω
C
ρ δ ˙u · ¨
u
dΩ
(2)with
δ ˙u
asatest fun tion andtheadmissible solution spa es,U
ad
and
U
ad,0
,aredened asfollows:
U
ad
=
u = u(x, t) ∈ [H
1
(Ω
C
)]
3
; u = u
d
on ∂Ω
u
C
; ∀ t ∈ [0, T ]
U
ad,0
=
u = u(x, t) ∈ [H
1
(Ω
C
)]
3
; u = 0 on ∂Ω
u
C
; ∀ t ∈ [0, T ]
2.2. Dis rete subdomain
Ω
d
Inan isolated system of thedis rete domain
Ω
d
whi h is a set of spheri al parti les thatintera t via ohesive beams, the governing equations an be written as: fori = 1..n
p
andt ∈ [0, T ]
, given theinitial onditions,nd(d
i
, θ
i
, f
int
/i
, c
int
/i
) ∈ R
3
× R
3
× R
3
× R
3
su hthat:(
f
ext
/i
+ f
int
/i
= m
i
d
¨
i
c
ext
/i
+ c
int
/i
=
I
i
θ
¨
i
(3)
with
d
i
,θ
i
,m
i
andI
i
representing the displa ement ve tor, therotation ve tor, themass and themass moment of inertia of thei
th
parti le, respe tively.
f
ext
/i
andc
ext
/i
represent the total external for es and the total external torques applied on thei
th
parti le, respe tively.
f
int
/i
andc
int
/i
are the total internal for es andthetotal internal torquesapplied byotherparti lesvia the ohesive beams on thei
th
parti le, respe tively.
f
int
i
=
nnp
X
j=0
f
ij
=
nnp
X
j=0
(E
µ
S
µ
∆l
l
µ
µ
x
−
6E
l
µ
2
I
µ
µ
((θ
jz
+ θ
iz
)y + (θ
jy
+ θ
iy
)z))
c
int
i
=
nnp
X
j=0
c
ij
=
P
nnp
j=0
(
G
µ
I
Oµ
l
µ
(θ
jx
− θ
ix
)x −
2E
µ
I
µ
l
2
µ
((θ
jy
+ 2 θ
iy
)y + (θ
jz
+ 2 θ
iz
)z))
(4) With:• nnp
is total numberofneighbor parti lesof thei
th
parti le
• f
ij
andc
ij
arebeamrea tionfor esandtorquesa tingonthei
th
parti lebythe
j
th
one,respe tively.
• (O
i
, x, y, z)
is lo alframe asso iatedto thebeam onne tingi
th
and
j
th
parti les.
• θ
i
(θ
ix
, θ
iy
, θ
iz
)
andθ
j
(θ
jx
, θ
jy
, θ
jz
)
are the rotations of beam ross se tions expressed in the beam lo alframe.PSfragrepla ements
Ω
Ω
C
Ω
O
Ω
d
Figure1: Globaldomainde omposition
• l
µ
,S
µ
,I
Oµ
andI
µ
arethebeamlength,beam rossse tionarea,polarmomentofinertiaandmoment of inertiaalongy
andz
.• E
µ
andGµ
arethe beam Young and shearmodulus.As in the ontinuum, the asso iated weak formulation an be dened as follows: nd
(d, θ, f , c) ∈
D
ad
× O
ad
× F
ad
× C
ad
su h that, given theinitial onditions,∀ (δ ˙d, δ ˙θ) ∈ ˙
D
ad,0
× ˙
O
ad,0
:n
p
X
i=1
f
ext
/i
· δ ˙d
i
+
n
p
X
i=1
f
int
/i
· δ ˙d
i
+
n
p
X
i=1
c
ext
/i
· δ ˙θ
i
+
n
p
X
i=1
c
int
/i
· δ ˙θ
i
=
n
p
X
i=1
m
i
d
¨
i
· δ ˙d
i
+
n
p
X
i=1
I
i
θ
¨
i
· δ ˙θ
i
(5) with:D
ad
= {d = d
i
(t) i = [1..n
p
] ∀t ∈ [0, T ]}
O
ad
= {θ = θ
i
(t) i = [1..n
p
] ∀t ∈ [0, T ]}
F
ad
= {f = f
int
/i
(t) i = [1..n
p
] ∀t ∈ [0, T ]}
C
ad
= {c = c
int
/i
(t) i = [1..n
p
] ∀t ∈ [0, T ]}
n
p
: total numberof DEM parti les. 2.3. Couplingapproa hAs mentioned in the previous se tions, the oupling approa h used here is based on the Arlequin
approa h [6, 7,8℄. Thisapproa h onsistsof:
1. Asuperpositionofme hani al statesinthegiven subdomains
Ω
C
andΩ
d
withanoverlappingzoneΩ
O
(Fig. 1).2. A weak oupling (based on theweakformulation):
(a) Denition ofthegluing zone
Ω
G
:Inthis study,thegluing zone
Ω
G
isthe same astheoverlappingzoneΩ
O
. Hereafter,the term overlapping zone will beusedto design theoverlappingzone orthegluing zone.(b) Mediator spa e
M
:To ensure the orre t dialogue between the models, the ontrol quantities in the overlapping
zonemust be hosen arefully. Here, the velo ity oupling, ina weak sense in
Ω
O
), is hosen. Fromanalgorithmi pointofview,thevelo ity ouplingiseasierthanthedispla ement oupling(Remark2). The mediatorspa edenotedby
M
isdenedasthespa eof thevelo itiesdened inΩ
O
.a) Discontinuous f unctions
b) Linear continuous f unctions
c) Continuous dif f erential f unctions
f (x)
f (x)
f (x)
¯
f (x)
¯
f (x)
f (x)
¯
Ω
C
Ω
C
Ω
O
Ω
d
Ω
O
Ω
d
Ω
C
Ω
O
Ω
d
1
1
1
0
0
0
Figure2: Examplesofweight fun tions
( ) Proje tion operator andjun tion model:
Theproje tionoperator
Π
proje tsthe ontinuumanddis retevelo itiesonthemediatorspa e. The jun tion model denes the linking onditions between thetwo models in theoverlappingzone. To proje t the velo ities on
M
, an interpolation, whose shape fun tions will be dened later,is used. The jun tionmodelusedinthis workis theH
1
(Ω
O
)
s alarprodu tdened by:< λ, q >
H
1
(Ω
O
)
=
ˆ
Ω
O
λ
· (Π ˙
u
− Π ˙
d) + l
2
ε(λ) : ε(Π ˙
u
− Π ˙
d) dΩ
(6)where
(Π ˙
u
− Π ˙
d)
is the dieren e between the proje ted ontinuum and dis rete velo ities onΩ
O
,λ
is the Lagrange multiplier eld andl
, the jun tion parameter, is anH
1
oupling
parameter.This parameter has the unit of a length, it is added to ensure the homogeneity of
theintegral termsof the
H
1
oupling model. Inthiswork,
l
is onsideredavariableparameter and will be studied in Se tion 3. Ifl = 0
, theH
1
(Ω
O
)
s alar produ t be omes equivalent to theL
2
(Ω
O
)
s alarprodu t(7)knownasthe Lagrange multipliermodel.< λ, u >
L
2
(Ω
O
)
=
ˆ
Ω
O
λ
· (Π ˙
u
− Π ˙
d) dΩ
(7)The displa ement and velo ityelds in
Ω
C
andΩ
d
do not have thesame nature. Indeed,Ω
C
is a ontinuumwhereas theΩ
d
is a dis rete subdomain. The dis rete eld asso iated withΩ
d
is dened only at the parti le positions. To be able to ompute the jun tion models (6) and(7),anintepolationisdenedontheDEMparti lesin
Ω
O
usingshapefun tions,whi hwillbe dened later.3. Theenergy partition between the ontinuum anddis rete media intheoverlapping zone:
Asshown inFigure 1, the two models oexist in
Ω
O
. Therefore, the energies inthis zonemust be weighted, and,a kindof partitionof unityintermsofenergy isperformed. Three weight fun tions,α(x)
,β(x)
andγ(x)
, are introdu ed for the internal energy, the kineti energy and the external work ofthe ontinuum subdomain,respe tively. Allof thefun tionsverifythefollowing:f (x) : Ω → [0, 1]
x
→
1
in Ω
C
\Ω
O
[0, 1] in Ω
C
∩ Ω
O
0
otherwise
(8)In a omplementary manner, the internal energy, the kineti energy and the external work of the
dis rete subdomain are weighted by
α(x) = 1 − α(x)
¯
,¯
β(x) = 1 − β(x)
andγ(x) = 1 − γ(x)
¯
, respe tively. Figure2presentsexamples ofweight fun tions.This oupling approa h is applied to ouple the ontinuum and thedis rete models dened on
Ω
C
andΩ
d
,respe tively. By introdu ing theweight fun tions, in(2)and (5), and the oupling ondition (6),the global weighted weakformulationbe omes: nd(u, d, θ, λ) ∈ U
ad
× D
ad
× O
ad
× M
su hthat, giventhe
initial onditions,
∀ (δ ˙u, δ ˙d, δ ˙θ, δ ˙λ) ∈ U
ad,0
× D
ad,0
× O
ad,0
× M
´
Ω
C
β ρ δ ˙u · ¨
u
dΩ +
´
Ω
C
α ε(δ ˙u) : A : ε(u) dΩ −
´
∂Ω
T
C
γ δ ˙u · T
d
dΓ −
´
Ω
C
γ ρ δ ˙u · f dΩ
+
n
p
X
i=1
¯
β
i
m
i
d
¨
i
· δ ˙d
i
+
n
p
X
i=1
¯
β
i
I
i
θ
¨
i
· δ ˙θ
i
−
n
p
X
i=1
(¯
γ
i
f
ext
/i
+ ¯
α
i
f
int
/i
) · δ ˙d
i
−
n
p
X
i=1
(¯
γ
i
c
ext
/i
+ ¯
α
i
c
int
/i
) · δ ˙θ
i
+ δ
ˆ
Ω
O
λ
· (Π ˙u − Π ˙d) + l
2
ε(λ) : ε(Π ˙u − Π ˙d) dΩ
= 0
(9)2.4. Spatial dis retizationand integration issues
Intheprevious subse tions, the global weak formulation (9)is presentedin a ontinuous form. Now,
thespatial dis retization is introdu ed. In the literature, thereare many interesting ontinuum methods
used for solving partial dierential equations, su h as SPH [23℄, NEM [24 ℄ and FEM[25℄. Ea h method
is distinguished byits apability to spatially dis retizethe studied model. Among them,the onstrained
natural element method CNEM[26, 27,28 ℄,whi h isanextension of thenatural element method(NEM)
[29 ℄ to non- onvex domains, is hosen in this study. This method has pra ti ally all of the advantages
of the FEM approa h, and it ir umvents the major drawba ks related to the meshing. Indeed, using
the FEMapproa h, the approximation is dependent on the mesh quality. In ontrast, using the CNEM
approa h, the approximation is dependent only on the relative position of the nodes [30 ℄. Unlike the
othermesh-freeapproa hes: (i)the supportsof onstrained naturalneighbor (CNN)shape fun tionsused
in CNEM approa h are automati ally dened, (ii) the values of CNN shape fun tions asso iated with
internal nodesarenullonthe borderofthedomain. Thislastpropertyis parti ularlyinteresting be ause
itallowsadire timpositionoftheboundary onditions,exa tlyasintheniteelementsframework. Given
thebroadsimilarity between the CNEMand FEMapproa hes, theCNEM-DEM oupling havethe same
performan esastheFEM-DEM oupling withbetterappli abilityon omplex domainsand/orbehaviors.
Therefore,the ontinuumsubdomainin
Ω
C
isdis retizedwiththeCNEMapproa h. Consequently,Ω
C
is approximated byaset ofnodesinwhi h onne tivityis not ne essary[26℄.Toobtain a ontinuouseldfromthe dis retequantities dened attheDEM parti lepositionsin
Ω
O
, a onstrainednaturalneighbor (CNN)interpolationisintrodu ed inΩ
d|Ω
O
. Thus,theparti lesasso iated withthis subdomain are also onsideredCNEM nodes. TheCNN interpolationis only applied inΩ
d|Ω
O
, whi h is assumed to be far from the ne s ale ee ts. The mediator spa e is also dis retized with theCNEMapproa h. Wedenoteby
M
h
C
,M
h
d
andM
h
O
thedis retizedspa esofU
ad
,
D
ad
and
M
,respe tively. The asso iated dis retized subdomains are designedΩ
h
C
,Ω
h
d
andΩ
h
O
,respe tively. The dis rete domainΩ
d
is a set of parti les, thenit is naturally dis retized andΩ
h
d
= Ω
d
. A ording to the ongurations of thedis retized spa es inthe overlapping zone,four ases an bedistinguished (Fig. 3).Inthisstudyand ontrarytoprevious studieson ontinuum/dis rete oupling approa hes,no oin iden e
onditionsareimposedonthe oexistingdis retizedsubdomainsin
Ω
O
. Therefore,thefourth onguration is studied here (Fig. 3-d)as thegeneral onguration thatin ludes the three other ongurations. Thissimplies theuseofthis method in3D omplexdomains. Indeed,inthis ase,it issu ient to dis retize
thesubdomainsindependentlyandmountthemasindi atedinFigure3-d. Infa t,usingthis onguration,
itisverydi ulttoprovemathemati allythe existen eanduniquenessofthesolution. Alsoand ontrary
to thethreeother ongurations,there arenonumeri al works,inliterature, studying this onguration.
Thus, thewell posedness oftheglobal problemwill beanalyzed numeri ally inthis paper.
UsingtheCNNinterpolationonthedierentdis retizedsubdomains,
Ω
h
C
,Ω
h
O
andΩ
h
d|Ω
O
,thedispla e-ment elds
u
andd
and theLagrange multiplier unknownsλ
areapproximated by:u
h
(x) =
n
C
X
i=1
PSfragrepla ements a)
Ω
h
C|Ω
O
= Ω
h
O
= Ω
h
d|Ω
O
b)Ω
h
C|Ω
O
⊂
Ω
h
O
= Ω
h
d|Ω
O
)Ω
h
C|Ω
O
= Ω
h
O
⊂
Ω
h
d|Ω
O
d)Ω
h
C|Ω
O
6= Ω
h
O
6= Ω
h
d|Ω
O
Nodesofthe ontinuumsubdomain
Ω
h
C
NodesoftheoverlappingsubdomainΩ
h
O
Parti lesofthedis retesubdomainΩ
h
d
Figure3: Thedierent ongurationsofthedis retizedsubdomains
d
h
(x) =
n
pO
X
i=1
N
i
d
(x)d
i
(11)λ
h
(x) =
n
O
X
i=1
N
i
O
(x)λ
i
(12)where
n
C
andn
O
aretotal number of nodes lo ated inΩ
h
C
andΩ
h
O
,respe tively.n
pO
istotal number of parti leslo atedinΩ
h
d|Ω
O
.
u
i
isthe nodaldispla ements,d
i
aretheparti le displa ementsandλ
i
arethe nodalLagrange multipliers.N
C
i
,N
O
i
andN
d
i
arethe CNN shape fun tions onstru ted onΩ
h
C
,Ω
h
O
andΩ
h
d|Ω
O
,respe tively.For the remainderof thispaper,thesupers ript
h
will be omittedfromtheapproximated quantities for larity. Be ause the global weightedweak formulation (9) is truefor anysmall arbitrary variations of˙u
,˙d
,˙θ
andλ
,it an be reformulated asfollows: nd(u, d, θ, λ) ∈ U
ad
× D
ad
× O
ad
× M
su h that, given
theinitial onditions,
∀ (δ ˙u, δ ˙d, δ ˙θ, δ ˙λ) ∈ U
ad,0
× D
ad,0
× O
ad,0
× M
:
´
Ω
C
β ρ δ ˙u · ¨
u
dΩ −
´
∂Ω
T
C
γ δ ˙u · T dΓ +
´
Ω
C
α ε(δ ˙u) : A : ε(u) dΩ −
´
Ω
C
γ ρ δ ˙u · f dΩ
+
´
Ω
O
λ
· δΠ ˙u + l
2
ε(λ) : ε(δΠ ˙u) dΩ = 0
(13)n
p
X
i=1
¯
β
i
m
i
d
¨
i
· δ ˙d
i
−
n
p
X
i=1
(¯
γ
i
f
ext
/i
+ ¯
α
i
f
/i
int
) · δ ˙d
i
−
´
Ω
c
λ.δΠ ˙d + l
2
ε(λ) : ε(δΠ ˙d) dΩ = 0
(14)n
p
X
i=1
¯
β
i
I
i
θ
¨
i
· δ ˙θ
i
−
n
p
X
i=1
(¯
γ
i
c
ext
/i
+ ¯
α
i
c
int
/i
) · δ ˙θ
i
= 0
(15)´
Ω
O
δλ · (Π ˙u − Π ˙d) + l
2
ε(δλ) : ε(Π ˙u − Π ˙d) dΩ = 0
(16)The integralterms will be omputed numeri ally using an integration te hnique. Integrationbya Gauss
to ompute the ontinuum termsusing theVoronoï ells astheba kground of theintegration. However,
on erning the oupling terms, this te hnique annot be applied dire tly. Indeed, the integrands in lude
variables dened on dierent Voronoï diagrams. The issue here is how to hoose the ba kground of
the integration. In this work, the Voronoï ells asso iated with the mediator spa e
M
h
O
are hosen as ba kground of integration in the overlapping zoneΩ
O
. All of thevariables that are not dened onM
h
O
areproje ted on thisspa e.
By repla ing
u
¨
,δ ˙u
,˙d
,δ ˙d
,λ
andδλ
by their approximated expressions in (13), (14), (15) and (16), the dis retized equations an be writtenas:•
DEM equations:[m
β
]
n ¨do = f
int
α
+ f
ext
γ
+ {f
c
}
[I
β
]
n¨θo
=
c
int
α
+ c
ext
γ
(17) where
(m
β
)
ij
= ¯
β
i
δ
ij
· m
i
,(I
β
)
ij
= ¯
β
i
δ
ij
I
i
,(c
int
α
)
i
= ¯
α
i
c
int
/i
,(c
ext
γ
)
i
= ¯
γ
i
c
ext
/i
,(f
int
α
)
i
= ¯
α
i
f
int
/i
,(f
ext
γ
)
i
= ¯
γ
i
f
ext
/i
and{f
c
} = [c
d
] {λ} = ([c
L
2
d
] + l
2
[c
H
1
d
]) {λ}
representsthetotal oupling for e.c
L
d
2
≈
n
O
X
K=1
V
O
I
h
N
d
(x
I
)
i
T
N
O
(x
I
) dΩ
Andc
H
1
d
≈
n
O
X
K=1
V
O
I
h ˜
B
d
(x
I
)
i
T
h ˜
B
O
(x
I
)
i
Where
x
I
arethe oordinatesoftheI
th
nodeof
Ω
h
O
,V
O
I
isthevolume oftheVoronoï ellasso iated withI
th
node ofΩ
h
O
,N
d
andN
O
are the interpolation matri es asso iated with
Ω
h
d
andΩ
h
O
, respe tively.h ˜
B
d
i
andh ˜
B
O
i
are the smoothed gradient matri es [31 , 32℄ asso iated with
Ω
h
d
andΩ
h
O
,respe tively.•
CNEMequations[M
β
] {¨
u} = −
F
int
α
+ F
ext
γ
− {F
c
}
(18) where(M
β
)
ij
= δ
ij
β(x
i
) M
i
,M
i
isthelumpedmassofthei
th
nodelo atedat
x
i
position,F
int
α
=
[K
α
] {u}
,[K
α
]
isthe weightedstinessmatrix and{F
c
} = [C
C
] {λ} = ([C
L
2
C
] + l
2
[C
H
1
C
]) {λ}
repre-sentsthetotal oupling for e.C
C
L
2
≈
n
O
X
K=1
V
O
I
N
C
(x
I
)
T
N
O
(x
I
) dΩ
AndC
H
1
C
≈
n
O
X
K=1
V
O
I
h ˜
B
C
(x
I
)
i
T
h ˜
B
O
(x
I
)
i
WhereN
C
andh ˜
B
C
i
are,respe tively,theinterpolationandsmoothedgradientmatri esasso iated
with
Ω
h
C
.•
Interfa e equations: Equation16 leadsto:[C
O
]
.
u − [c
o
]
n
.
d
o
= 0
where:[C
O
] =
h
C
L
2
O
i
+l
2
h
C
H
1
O
i
=
h
C
L
2
C
i
T
+ l
2
h
C
H
1
C
i
T
and[c
o
] =
h
c
L
2
o
i
+l
2
h
c
H
1
o
i
=
h
c
L
2
d
i
T
+ l
2
h
c
H
1
d
i
T
.The numeri al time integration is based on an expli it integration s heme that is well adapted for
dynami omputations. Many expli it s hemes an be used, su h as the Runge-Kutta, position Verlet
and velo ity Verlet s hemes. A omparison between these s hemes an be found in [33℄. A ording to
this referen e [33 ℄, the velo ity Verlet s heme provides good results and is also easy to implement. For
these reasons, this s heme is usedin this paperto solve theglobal dynami problem. Thiss heme gives
an
O(h
3
)
approximation for both velo ities and displa ements. Thus, the velo ity oupling, usedin this
work,doesnot ae t the oupling approa h a ura y ompared to thedispla ement oupling.
2.5.1. DEM algorithm (DEM ode)
•
Initializationn ¨do
n
,n
.
d
o
n
,{d}
n
,n ¨θo
n
,n
.
θ
o
n
and{θ}
n
: Theinitial onditionsortheinterfa eresults.•
Computationof{d}
n+1
and{θ}
n+1
:{d}
n+1
= {d}
n
+ ∆t
n
d
.
o
n
+
∆t
2
2
n ¨do
n
{θ}
n+1
= {θ}
n
+ ∆t
n
θ
.
o
n
+
∆t
2
2
n ¨θo
n
(19)•
Computationoff
int
α
n+1
,f
ext
γ
n+1
,c
int
α
n+1
etc
ext
γ
n+1
•
Computationof the predi tive lineara elerationsn ¨do
∗
n+1
(omittingthe oupling for es
{f
c
}
from Equation17).n ¨do
∗
n+1
= [m
β
]
−1
(f
int
α
n+1
+
f
ext
γ
n+1
)
(20)•
Computationof the angular a elerationsn ¨θo
n+1
:n ¨θo
n+1
= [I
β
]
−1
(c
int
α
n+1
+
c
ext
γ
n+1
)
(21)•
Computationof the predi tive linearvelo itiesn
.
d
o
∗
n+1
:n
.
d
o
∗
n+1
=
n
.
d
o
n
+
∆t
2
(n ¨do
n
+
n ¨do
∗
n+1
)
•
Computationof the angular velo itiesn
.
θ
o
n+1
:n
.
θ
o
n+1
=
n
.
θ
o
n
+
∆t
2
(n ¨θo
n
+
n ¨θo
n+1
)
•
Transfer of the predi tive linearvelo ities and a elerations to theinterfa e:n
.
d
o
∗
n+1
andn ¨do
∗
n+1
2.5.2. CNEM algorithm (CNEM ode)
•
Initialization{¨
u}
n
,.
u
n
et{u}
n
: The initial onditionsor theinterfa e results.•
Computationof{u}
n+1
:{u}
n+1
= {u}
n
+ ∆t
.
u
n
+
∆t
2
2
{¨
u}
n
(22)•
Computationofthe predi tivelineara elerations{¨
u}
∗
n+1
: (omittingthe ouplingfor es{F
c
}
from Equation18).{¨
u}
∗
n+1
= [M
β
]
−1
(−
F
int
α
n+1
+
F
ext
γ
n+1
)
(23)•
Computationof the predi tive linearvelo ities.
u
∗
n+1
:.
u
∗
n+1
=
.
u
n
+
∆t
2
({¨
u}
n
+ {¨
u}
∗
n+1
)
•
Transfer of the predi tive linearvelo ities and a elerations to theinterfa e:.
u
∗
n+1
and{¨
u}
∗
n+1
•
Re overy of the predi tive linear velo ities from both the CNEM and DEM odes:.
u
∗
n+1
andn
.
d
o
∗
n+1
•
Computationof{λ}
n+1
.
u
n+1
=
.
u
∗
n+1
−
∆t
2
[M
β
]
−1
{F
c
}
n+1
n
.
d
o
n+1
=
n
.
d
o
∗
n
+
∆t
2
[m
β
]
−1
{f
c
}
n+1
(24){F
c
}
n+1
= [C
C
] {λ}
n+1
{f
c
}
n+1
= [c
d
] {λ}
n+1
(25)[C
O
]
.
u
n+1
− [c
o
]
n
.
d
o
n+1
= 0
(26)By introdu ingEquations 24 and25 into Equation26,the interfa esystemof equations an be
written as:
[A] {λ}
n+1
= {b}
n+1
(27)where the oupling matrix
[A]
and{b}
n+1
are dened,respe tively,as:[A] =
∆t
2
([C
O
] [M
β
]
−1
[C
C
] + [c
o
] [m
β
]
−1
[c
d
])
(28){b}
n+1
= [C
O
]
.
u
∗
n+1
− [c
o
]
n
.
d
o
∗
n+1
(29) By solvingEquation27,{λ}
n+1
an beobtained.•
Computationof{F
c
}
n+1
and{f
c
}
n+1
using Equation25.•
Computationof the linearvelo ities.
u
n+1
andn
.
d
o
n+1
usingEquation 24.•
Thelinear a eleration orre tions:{¨
u}
n+1
andn ¨do
n+1
:{¨
u}
n+1
= {¨
u}
∗
n+1
− [M
β
]
−1
{F
c
}
n+1
n ¨do
n+1
=
n ¨do
∗
n+1
+ [m
β
]
−1
{f
c
}
n+1
(30)•
Transfer of{¨
u}
n+1
and.
u
n+1
to theCNEM pro ess andn ¨do
n+1
and
n
.
d
o
n+1
theDEM pro ess.
Remark1:. The system (27) is solved using the well-known
LU
de omposition method [34℄. Con erning thesystem (28), sin e the massmatri es ofboth ontinuum and dis rete models arediagonals, itis easyto derive their inverse matri esand ompute the oupling matrix
A
.Remark2:. From an algorithmi point of view,the velo ity oupling used inthis work is easier than the
displa ement oupling. This is be ause thedispla ement oupling requires, inaddition to thepredi tive
a elerations and velo ities, the omputation of the predi tive displa ements whi h must be sent to the
Interfa e odefor orre tion. Therefore,inthe aseofdispla ement oupling,additionalstepsarene essary
to omputeand orre tthepredi tivedispla ements. Whereas,inthe aseofvelo ity oupling,the orre t
of the parti les in the overlapping zone. These quantities are orre ted impli itly. Indeed, the internal
for es are omputed a ounting for parti le displa ements and rotations. These for es are later used to
ompute thenew displa ementswhi h are orre tedusing the oupling ondition(16).
Remark4:. To be able to ompute the predi tive a elerations (Eqs. 20, 21 and 23), the lumped mass
matri es must be invertible. Thus, the weight fun tions
β
and¯
β
must be stri tly positive inΩ
O
and at the border∂Ω
O
. Then, a smallε
will be used insteadof zero at the nodes assigned to∂Ω
O
. Therefore, thedenitionof this weight fun tionβ
(8) isslightly modied as:β(x) : Ω → [0, 1]
x
→
1
in Ω
C
\Ω
O
[ε, 1 − ε] in Ω
C
∩ Ω
O
0
otherwise
(31)where
ε
isa small stri tlypositive realnumberto be hosen.2.5.4. Implementation
TheDEM al ulusisa hievedusingtheGranOOworkben h(GranularObje tOriented). Granoowas
developed at theMe hani s institute of Bordeaux (I2M)by J.L. Charles etal. The ode provides C++
librariesthatimplement lassesusefultodes ribeandsolvedynami me hani al problemsusingDEMand
expli ittemporalintegration s hemes. TheCNEM al ulusisa hieved usingaCNEM-based ode,whi h
was developed at the PIMM laboratory by G. Cognal et al. It provides C++ libraries that interfa e
with Python modules. The oupling between DEM and CNEM, des ribed in the previous se tions, is
performedbyan interfa e written inthe Python language. This interfa e ommuni ates dire tlywith
theCNEM-based pro ess using Python lasses. The Inter Pro ess Communi ation (IPC) toolis usedto
ensure asyn hronized ommuni ation between theGranOO pro ess and theinterfa e pro ess.
3. Parametri study of the oupling parameters
Several workshavestudied mathemati ally theArlequin method forboth ontinuum- ontinuum oupling
[7 ,19,20℄and ontinuum-dis rete oupling[21℄. Themainresults on erningthewell-posednessofthe
ou-pling problemarere alledinthis paper. Theweight fun tion
α
mustbestri tly positive inΩ
O
. Without this onditionthe oer ivityof theinternal energy annotbe veried. Another signi ant resulton ern-ing the oupling jun tion models is that for the dis retized problem, ontrary to the
H
1
oupling whi h
yields a well-posed problem, the
L
2
oupling model an lead to an ill- onditioned system of equations,
espe ially in the ase of very small mesh size. In this ontext, Bauman et al. [21℄ have studied another
oupling model, the
H
1
seminorm, inwhi hthe rst term ofthe
H
1
modelisremoved. Thismodel leads
to a well-posed problem, but it does not onstrain enough the ontinuum and dis rete displa ements in
the overlapping zone. Other works [22, 35 , 21℄ have studied numeri ally the ingredients of theArlequin
method using
1D
models. Guidault et al. [22, 35 ℄ noted that, for theL
2
oupling model, the weight
fun tion
α
mustbe ontinuousattheboundaryofthegluingzone∂Ω
O
. Indeed,theuseofadis ontinuous weight fun tion an ause undesirable free onditions at∂Ω
O
.Con erning the hoi e of the mediator spa e, Ben Dhia [7 , 20 ℄ mentioned thatin the ase of ontinuous
domains, itis onvenient to hoose
M = H
1
(Ω
O
)
;however,itis verydi ultto hoosethenite approx-imation spa eM
h
O
. To address this di ulty, several works[22 , 35,21 ℄ proposed a 1Dnumeri al study ofM
h
O
. The dierent ongurations thatwerestudied arepresentedina,band of Figure3. Thestati studiesof Guidaultetal. [22℄ showthat: (i)inthe aseofa nemultiplier spa e(Fig. 3-b), theresponseof the stru ture do not depend on the weight fun tions and a lo king phenomenon takes pla e, i.e, the
E
µ
= 265 GP a
ν
µ
= 0.3
r
˜
µ
= 0.71
Table1: Themi ropropertiesofthe ohesivebeambondsintheDEMsubdomain
˜
r
µ
isanadimensional ohesivebeamradius,denedastheratiobetweenthebeamradiusandthemeanparti leradius;E
µ
andν
µ
are themi roYoungmodulusandthePoissonratioofthebeams,respe tively.multiplier spa e, the weight fun tions has an inuen e on the solutions su h that the larger the weight
fun tion onthene mesh,the smaller be omesthe maximumjumpbetween thetwo meshes.
This work proposes a 3D numeri al dynami study using the general onguration given in Figure 3-d.
It will be demonstrated that some of the results proven in stati using 1D models are not valid in 3D
dynami simulations.
Assumingthegeneral aseoftheapproximatedLagrangemultiplierspa e,thevarious ouplingparameters
studied are:
•
Thejun tion modelparameterl
,•
Theweight fun tionsα
,β
andγ
,•
Thewidth ofthe overlappingdomainL
O
,•
Thedis retization ofthe approximated Lagrangemultipliers spa eM
h
O
.A 3D beam model is used for the dynami study (Fig. 4), in whi h the length and the diameter are
L = 20 mm
andD = 2 mm
,respe tively. The modelis dividedinto two subdomainswithan overlapping zone. TheleftsubdomainismodeledbytheCNEMapproa husing626nodes(theasso iated hara teristilength is about
l
c
= 0.47 mm
) and xedat the leftend (x = 0
). Theright subdomain is modeledbythe DEM approa h using20 000
spheri al parti les havingr
c
= 0.05 mm
as mean radius. Based on the hara teristi length of DEM and CNEM dis retization (l
c
andr
c
), the uto frequen ies of the two models an be determined:f
CN EM
c
= 1.9 M Hz
andf
DEM
c
= 18.2 M Hz
. To ontrol thehigh frequen y wave reexion at the CNEM-DEM interfa e, thefree end (x = L
) is submitted to a tensile loading with a very steep slope (Fig. 5). As shown in lower viewgraph of Figure 5, the Fourier spe trum ontainspowerful high frequen y waves (greater than
f
CN EM
c
). The material of the beam is the sili a: Young's modulusE = 72 GP a
, Poisson's ratioν = 0.17
and densityρ = 2200 Kg/m
3
. The orresponding mi ro
propertiesof the ohesivebeambondsintheDEM approa h aregivenin[1℄ and presentedinTable1.
To ontrol thewave propagation in themodel, four he k points are pla ed along this beam (Fig. 4) as
follows:
•
CnemChe kPoint: at the middleoftheCNEMsubdomainwhere the ontrolled quantitiesare om-putedusing the CNEMnodesinthiszone•
OverlapCnemChe kPoint: atthemiddleofthe overlappingzonewherethe ontrolledquantitiesare omputed using onlytheCNEM nodesinthis zone.•
OverlapDemChe kPoint: at the middle of the overlapping zonewhere the ontrolled quantities are omputed using onlytheDEM parti lesinthis zone.•
DemChe kPoint: atthemiddleoftheDEMsubdomainwherethe ontrolledquantitiesare omputed usingthe DEMparti les inthiszone.Figure 6 presents the referen e results obtained by DEM and CNEM separately. Table 2 presents the
meandispla ement ofthe right end andtherstthreenatural frequen ies. It anbeseenthattheresults
are in good agreement, and they are also in agreement with the beam theory results. This ensures the
0.0
0.1
0.2
0.3
0.4
0.5
T ime
(µs)
0
200
400
600
800
1000
1200
1400
F
o
r
ce
(N
)
(a) F orce vs. T ime
20
40
60
80
100
F requency
(M Hz))
50
55
60
65
70
P
o
w
e
r
(d
B
)
(b) F ourier spectrum
Figure5: Tensileloadingofstudymodelandtheasso ietedspe tralanalysis( omputedfromFFT)
U
mean
(mm)
f
0
(Hz)
f
1
(Hz)
f
2
(Hz)
Theory
0.087
71 757
215 272
358 787
DEM
0.083
72 408
217 246
362 072
CNEM
0.088
71 359
214 023
356 491
( omputedfromFFT)
Remark5:. Inthisstudy,forthesakeofsimpli ity,theweight fun tionsare hosenasfollows:
α = β = γ
.3.1. Inuen e of the jun tion parameter
l
Theparameter
l
ismainlyemployedto omputethe ouplingmatrixA
(28). Theparameter'sinuen e on the onditioning ofA
(Cond = kAk.kA
−1
k
) is analyzed. Figure 7 shows the onditioning of
A
with respe ttol
obtainedforL
O
= 6 mm
usinga oarse multiplierspa eand ontinuousweight fun tionswithε = 0.005
(ε
is dened in (31)). The onditioning de reases withl
and rea hes a minimum at a smalll = l
opt
. Beyond this value, the onditioning in reases exponentially asl
in reases. This is true for any hoi es ofL
O
,theweight fun tionsandM
h
O
. Be ausel
de reasesthe onditioning,H
1
oupling (6)for a
small valueof
l
,isbetter thanL
2
oupling (7). However, ontrary to whatis presentedintheliterature,
H
1
oupling be omes worse if
l
ex eeds some small value. In pra ti e, this parameter an be hosen as the hara teristi length oftheoverlapping zonedis retizationl
Ω
O
c
(l
opt
≈ l
Ω
O
c
). Intheremainder of thisse tion we will usetheH
1
oupling with
l = l
opt
.3.2. Inuen e of the weight fun tions
Inthissubse tion,a nedis retization oftheapproximated multiplierspa e
M
h
O
is hosen, i.e,atthe same s ale asM
h
d
. The widthoftheoverlapping zoneL
O
isxed at2 mm
.3.2.1. Constant weight fun tions
α
CN EM
= α
DEM
= 0.5
Themeandispla ementobtainedwiththe ouplingmethodis
0.081 mm
. Thisisinagreementwiththe referen emean displa ements(Tab. 2). However, the temporal urve (Fig. 8) presents several deviationswithregardtothereferen e urves. Figure9presentsthevelo itiesinthedierent he kpoints(Figure4)
for therstround tripof the wave propagation. It an beseen thatthemajor partof thehighfrequen y
waves (HFW) are ree ted without entering the overlapping zone. Indeed, the HFW initially aptured
intheDemChe kPoint didnot appearinOverlapDemChe kPoint or OverlapCnemChe kPoint. This
explainsthedeviationinthetemporaldispla ementea htimetheglobalwave rossestheoverlappingzone.
Thus, onstant weight fun tions are not a good hoi e for dynami simulations. Indeed, the proje tion
me hanism, whi h o urs in
Ω
O
, annot dampen the HFW, and an additional ltering is required. In ontrast,the stati studiesofGuidaultetal. [22 ℄showed that onstant weight fun tions anbe usedwithFigure7: Conditioningof
A
withrespe ttol
L
O
= 6 mm
, oarseMultiplierspa e, ontinuousweightfun tionswithε
= 0.005
Figure8:Thefree-enddispla ementsobtainedusingDEMandCNEMseparatelyandthe ouplingmethod,andtheasso iated
spe tralanalyses( omputedfromFFT)
L
O
= 2 mm
,neMultiplierspa e, onstantweight fun tionsα
CN EM
= 0.5
3.2.2. Constant weight fun tions
α
CN EM
6= 0.5
Thissub-subse tion analyzestheinuen e ofthe weight onstant onthewavepropagation. Two ases
are studied here; the rst ase uses
α
CN EM
= 0.3
(then,α
DEM
= ¯
α
CN EM
= 0.7
), and the se ond ase usesα
CN EM
= 0.8
(then,α
DEM
= 0.2
). Theasso iated resultsare presentedinFigure 10.Figure 10: The free-end displa ements obtained using DEM and CNEM separately and the oupling method, and the
asso iatedspe tralanalyses( omputedfromFFT)fordierentweight onstants
L
O
= 2 mm
,neMultiplierspa e, onstantweightfun tionsα
CN EM
= 0.3
,α
CN EM
= 0.5
andα
CN EM
= 0.8
A largedieren e between theresults isobserved. Inthe rst ase(
α
CN EM
= 0.3
), themagnitudeofthe free-end displa ement is greater than that obtained usingα
CN EM
= 0.5
. However, it is smaller for the ase ofα
CN EM
= 0.8
. To provide an explanation for these results, the temporal velo ities at the he k pointsarepresentedinFigure11.It anbeseenthatfor
α
CN EM
= 0.8
,aportionoftheprin iplewaveispositively ree tedattheinterfa e between the two models, or more pre isely, without entering the overlapping zone. Furthermore, only at
num
DEM −CN EM
=
magnitude of the transmitted wave
magnitude of the incident wave
=
9,90
22,01
= 0.44
andr
num
DEM −CN EM
= 0.56
Byanalogywiththewavepropagationbetweenmediawithdierenta ousti impedan es,thetransmission
and ree tion oe ients antheoreti ally be dened as:
t
th
DEM −CN EM
=
α
DEM
2 α
+α
DEM
CN EM
and
r
th
DEM −CN EM
=
α
α
CN EM
DEM
+α
−α
CN EM
DEM
Then, it an be veried thatt
num
DEM −CN EM
andr
num
DEM −CN EM
are of the same order of magnitude ast
th
DEM −CN EM
= 0.4
andr
th
DEM −CN EM
= 0.6
,respe tively. Forα
CN EM
= 0.3
,thesame ree tion me ha-nism takespla e but witha negative oe ient. Indeed, the velo itymagnitude ofthe transmitted wave(measured at CnemChe kPoint) is greater than the velo ity magnitude of the forward wave (initially
measuredat DemChe kPoint).
t
num
Dem−Cnem
=
31.49
21.86
= 1.44
r
num
Dem−Cnem
= 1 −
31.49
21.86
= −0.44
andt
th
Dem−Cnem
=
2×0.7
1
= 1.4
t
num
Dem−Cnem
= 1 − 1.4 = −0.4
Then,for the aseofa onstantweighting,
α
CN EM
= α
DEM
= 0.5
mustbeused. Otherwise,therewillbe aree tionof apartofthe prin ipalforward wave. Thisresultprovesthatthe1Dstati studiesavailablein literature annot be used to perform dynami oupling. In ee t, Guidault et al. [22℄ noted that, in
stati s andusing a nemultiplier spa e, thesolutions do not depend on theweight fun tions.
3.2.3. Continuous weight fun tions
Asexplained in Remark 4,the weight fun tions must not vanish at the boundary of theoverlapping
zone, and a small value
ε
must be adopted rather than0
at∂Ω
O
. Prior to studying theinuen e of the ontinuous weight fun tions, the inuen e ofε
is studied. Figure 12 presents the free-end displa ement using ontinuous weight fun tions forε = 0.05
,ε = 0.005
andε = 0.0005
and the same onditions forL
O
andM
h
O
. The parameterε
, whenlessthan0.05
,hasno pra ti al inuen e on theresults,but a very smallε
an lead to instability problems. Indeed, as shown inTable3, the smallertheε
,the greaterthe onditioning of the oupling matrixA
be omes.Figure12: Thefree-enddispla ementsobtainedusingthe ouplingmethod,andtheasso iatedspe tralanalyses( omputed
fromFFT)fordierentvaluesof
ε
ε
0.05
0.005
0.0005
Cond[A(l
opt
)]
2.53e4
8.93e4
5.67e5
Table 3: Conditioningof
A
withrespe ttoε
L
O
= 2 mm
,neMultiplierspa e, ontinuousweightfun tionsIn the remainder of this paper,
ε = 0.05
will be hosen ea h time a ontinuous weight fun tion is used. Figure 14 shows the free-end displa ement for the ase of a ontinuous weight fun tion (Fig. 2-b). Nohigh frequen y waves (HFW) are ree ted at the interfa e between the two models. Using ontinuous
weight fun tions, the HFW enter the overlapping zone, and then, they are dampened by the proje tion
ontothe oarse spa e. Thelowersubplot inFigure14eviden es thatwitha nemultiplier spa e,a small
overlappingzoneissu ient to an elout alloftheHFW.AsshowninFigure13,theuseofa ontinuous
weight fun tion signi antly improves the results. However, a small deviation from the referen e results
still persists and be omes greater ea h time the wave travels ba k (CNEM-DEM dire tion). Be ause of
thevery ne dis retizationof theDEM subdomain,theweight of theparti lesin
Ω
O
de reases smoothly when approa hing the CNEM subdomain. Therefore, the forward wave orre tly rosses the interfa ebetween the pure DEM (
Ω
d
\Ω
O
) and the overlapping zoneΩ
O
. Thus, byexamining the rst round-trip (Fig. 13),it isapparent thatno deviation fromthe referen esis notedwhen wave travels fromthe DEMsubdomaintotheCNEMsubdomain. IntheCNEMsubdomain,a oarsedis retizationisused. Thejump
betweentheweightsoftwoadja ent nodesisrelatively large. Thedis reteweightfun tionsoftheCNEM
subdomain aredis ontinuous stair asefun tionswithlarge jumps. Thus, thesameree tion me hanism,
observedpreviously,o urswhenthewavetravelsba k(CNEM-DEMdire tion). Toredu ethedeviation,
thewidthoftheoverlapping zonemustbe in reasedto redu etheslope oftheweight fun tions. Another
solution onsists of using ontinuous dierentiable weight fun tions (Fig. 2- ) to redu e the weighting
jump in the vi inity of the overlapping zone boundary
∂Ω
O
. Figures 19, 20 and 21 present the results using the twosolutions. Thewave orre tly rossesΩ
O
withoutanydeviation.Figure 13: The free-end displa ements obtained using DEM and CNEM separately and the oupling method, and the
asso iatedspe tralanalyses( omputedfromFFT)
Figure14: Thelinearvelo ities(a)atthedierent he kpointsforthe aseof ontinuous
α
and(b)atthe"DemChe kPoint" forthe aseof ontinuousand onstantα
L
O
= 2 mm
,neMultiplierspa e, ontinuousweightfun tions,ε
= 0.05
3.3. Inuen e of the approximated multiplierspa e
M
h
O
In the previous subse tion, a ne multiplier spa e was used. In this ase, the velo ity in
Ω
d|Ω
O
is pra ti allylo kedatavalueequaltothevelo ityintheΩ
C|Ω
O
,asshownintheuppersubplotofFigure15. Indeed,thevelo ity urveatOverlapDemChe kPoint oin ides withthatatOverlapCnemChe kPoint.The same lo king phenomenon is noted when the se ond onguration (Fig. 3-b) is used [22℄. Now, to
study theinuen e of
M
h
O
onthe results,a oarse multiplierspa e isused, i.e,at thesames ale asM
h
C
. Asshown in thebottom subplot of Figure 15,equality of thevelo ities inΩ
O
issatised only in a weak sense and not in ea h multiplier spa e node. This allows the ne model (DEM model) to orre tly a tin
Ω
O
. However, in this ase a small overlapping zone is insu ient to orre tly transmit the prin ipal tensile waveand an el the highfrequen y waves.3.4. Inuen e of the width of the overlapping zone
L
O
Itisapparent thatforthe aseofnemultiplierspa e,thelo kingphenomenono ursandasmall
L
O
issu ientto an elthehighfrequen ywaves(HFW).Be ausetheDEMparti lesarestrongly onstrainedin
Ω
O
,theuseofalargeoverlappingzone an slightlydampen theglobalfree-end displa ement (Fig. 16). Forthe aseofa oarsemultiplierspa e,the DEMparti lesareableto orre tlya tintheoverlappingzone. Then, even for the aseofalarge
Ω
O
,the global results willnot bedampened. AsshowninFigure 17,the largertheoverlapping zone,the better the resultsbe ome. Indeed,theuse of alargeΩ
O
redu es theHFWree tion andallows abetter transferoftheforward wave.3.5. How to hoose the oupling parameters in a general ase?
Ina general ase, there is not an obvious method to determine, ina single way, thevarious oupling
parameters to avoid wave reexion. This subse tion gives several re ommendations and trends to hoose
orre tly these parameters. The weight fun tions must be ontinuous. Indeed, with onstant weight
fun tions,thehighfrequen ywaves(HFW)areree tedwithoutenteringtheoverlappingzoneand annot
bedampened bytheproje tion me hanism. The hoi es ofthewidthoftheoverlappingzonedependson
the hara teristi dimension of the dis retization inthis zone. In the aseof ne dis retization, a narrow
L
O
= 2 mm
, ontinuousweightfun tion,ε
= 0.05
Figure16: Thefree-enddispla ementsobtainedusingthe ouplingmethodfor
L
O
= 2 mm
,L
O
= 4 mm
andL
O
= 6 mm
. FineMultiplierspa e, ontinuousweightfun tions,ε
= 0.05
Figure17: Thefree-enddispla ementsobtainedusingthe ouplingmethodfor
L
O
= 2 mm
,L
O
= 4 mm
andL
O
= 6 mm
. CoarseMultiplierspa e, ontinuousweightfun tions,ε
= 0.05
Dynami simulation
Couplingtype
H
1
oupling
Jun tion parameter
l
l = l
opt
Approximated Multiplierspa e Coarse Multiplierspa e
Widthofthe overlappingzone
L
O
As largeaspossibleWeight fun tions Continuous,
ε = 0.05
Table4: Re ommendedArlequinparametersfordynami simulations
in
Ω
O
. On the ontrary, oarse dis retization requires a largeΩ
O
. To minimize the onditioning of the oupling matrixA
,theH
1
oupling with
l = l
opt
isre ommended.l
opt
an be hosenasthe hara teristi length of theoverlapping zone dis retizationl
Ω
O
c
(l
opt
≈ l
Ω
O
c
). Table 4 presents the onvenient Arlequin parameters to orre tlyperform the oupling.4. Validation
The previous parametri -based study on tensile loading has allowed us to retain the onvenient
pa-rameters to perform a orre t oupling. In this se tion, the results of this study areused to validatethe
oupling between CNEM and DEM in a general 3D ase. Contrary to the tensile ase, in bending and
torsion, thedeformationsinthe rossse tions aresigni ant. To a ount for theseee ts, newgeometri
hara teristi s of the 3Dmodelare used:
L = 100 mm
andD = 20 mm
(L/D = 5
). The DEM method is applied only for the portion lo ated20 mm
from the right end (the se tionlo ated atx = L
) and the remainder ofthemodelismodeled usingtheCNEM method (Fig. 18).ThefollowingArlequin's parameters are hosen:
L
O
= 10 mm
, ontinuousdierentiableweight fun tions,ε = 0.05
,l = l
opt
and oarse multiplier spa e.Figures 19 and 20 present the temporal free-end displa ements with respe t to the x-axis and y-axis,
respe tively, using tensile and bending loading. The deviation from the referen e, as observed in the
previous simulations when the wave rosses the
Ω
O
, disappeared in the present results. Then, for this model,L
O
= 10 mm
is su ient to orre tly transmit the wave between the dis rete and ontinuum models.Figure19: Test1:Thefree-enddispla ements
U
x mean
obtainedbyDEM,CNEMandthe ouplingmethodL
O
= 10 mm
, oarseMultiplierspa e, ontinuousdierentiableweighting,ε
= 0.05
Figure20: Test2: Thefree-enddispla ements