Lecture Xfem Meshfree

EXtended Finite Element and
Meshfree Methods
Prof. Dr.-Ing. T. Rab zuk
Chair of Me hani s
Bauhaus University Weimar
Version: SS 2009
Contents
1 Introdu tion
13
2 Meshfree methods
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
Completeness, onsisten y and Partition of Unity . .

Conservation, stability and onvergen e . . . . . . .
Continuity . . . . . . . . . . . . . . . . . . . . . . . .
The weighting fun tion . . . . . . . . . . . . . . . . .
2.4.1 Requirements of the weighting fun tion . . .
2.4.2 Spe i weighting fun tions . . . . . . . . . .
2.4.3 Eulerian and Lagrangian kernels . . . . . . .
Spe i meshfree approximations . . . . . . . . . . .
2.5.1 Smooth Parti le Hydrodynami s (SPH) . . .
2.5.2 Corre ted SPH forms . . . . . . . . . . . . .
2.5.3 Element-free Galkerin method (EFG) . . . .
2.5.4 Hp- louds and PUFEM . . . . . . . . . . . .
2.5.5 Generalized Finite Element Method (GFEM)
2.5.6 Examples . . . . . . . . . . . . . . . . . . . .
Governing equations . . . . . . . . . . . . . . . . . .
2.6.1 Stati s . . . . . . . . . . . . . . . . . . . . . .
2.6.2 Dynami s . . . . . . . . . . . . . . . . . . . .
2.6.3 From the strong form to the weak form . . .
Constru tion of spe i meshfree methods . . . . . .
Spatial integration . . . . . . . . . . . . . . . . . . .
2.8.1 Nodal integration . . . . . . . . . . . . . . . .
2.8.2 Stabilized nodal integration . . . . . . . . . .
2.8.3 Stress-point integration . . . . . . . . . . . .
2.8.4 Gauss integration . . . . . . . . . . . . . . . .
Essential boundary onditions . . . . . . . . . . . . .
2.9.1 Lagrange multiplier method . . . . . . . . . .
2.9.2 Penalty method . . . . . . . . . . . . . . . . .
2.9.3 Transformation method . . . . . . . . . . . .
Coupling to nite elements . . . . . . . . . . . . . .
2.10.1 Coupling via ramp fun tions . . . . . . . . .
2.10.2 Coupling with Lagrange multipliers . . . . . .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
17
17
18
19
20
20
21
23
24
25
27
33
40
40
41
50
50
50
51
51
52
52
54
55
56
59
59
60
61
62
62
63
CONTENTS
2.10.3 Bridging domain oupling method . . . . . . . . . . . . .
66
3 Kinemati s of strong and weak dis ontinuities
71
4 Strong dis ontinuities in meshfree methods
75
4.1 A histori al overview . . . . . . . . . . . . . . . . . .

4.1.1 The visibility riterion . . . . . . . . . . . . .
4.1.2 The dira tion riterion . . . . . . . . . . . .
4.1.3 The transparen y riterion . . . . . . . . . .
4.1.4 The see-through and ontinuous line riterion
4.1.5 Mixed riteria . . . . . . . . . . . . . . . . . .
4.2 The on ept of enri hment . . . . . . . . . . . . . . .
4.2.1 Introdu tion into LEFM . . . . . . . . . . . .
4.2.2 Intrinsi enri hment . . . . . . . . . . . . . .
4.2.3 Extrinsi enri hment . . . . . . . . . . . . . .
5 The eXtended nite element method (XFEM)

5.1 Introdu tion . . . . . . . . . . . . . . . . . . . .
5.2 Level sets . . . . . . . . . . . . . . . . . . . . .
5.2.1 Denition . . . . . . . . . . . . . . . . .
5.2.2 Normal and urvature . . . . . . . . . .
5.2.3 Level set for weak form integration . . .
5.2.4 Smoothed Heaviside and Dira fun tion
5.2.5 Signed distan e fun tion . . . . . . . . .
5.2.6 Dis retization of the level set . . . . . .
5.2.7 Updating the level set of the interfa e .
5.2.8 Level set for solid shape des ription . .
5.3 Standard XFEM . . . . . . . . . . . . . . . . .
5.3.1 Appli ation to strong dis ontinuities . .
5.3.2 Appli ation to weak dis ontinuities . . .
5.4 Hansbo-Hansbo XFEM . . . . . . . . . . . . .
5.5 Blending elements . . . . . . . . . . . . . . . .
5.6 Implementation . . . . . . . . . . . . . . . . . .
5.6.1 Shape fun tions . . . . . . . . . . . . . .
5.6.2 The B-Operator . . . . . . . . . . . . .
5.6.3 The element stiness matrix . . . . . . .
5.7 Integration . . . . . . . . . . . . . . . . . . . .
5.8 Appli ation of XFEM to LEFM . . . . . . . . .
5.8.1 Governing equations . . . . . . . . . . .
5.8.2 Weak form . . . . . . . . . . . . . . . .
5.8.3 XFEM approximation for ra ks . . . .
5.8.4 Dis rete equations . . . . . . . . . . . .
5.8.5 Advan ed XFEM in LEFM . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
75
75
75
78
80
81
81
82
83
84
89
89
90
90
90
91
93
93
94
94
95
97
97
99
101
102
107
108
109
111
113
116
116
117
117
121
124
CONTENTS
6 Other Methods
137
6.1 EXtended meshfree methods . . . . . . . . . . . . . . . . . . . . 137

6.2 Embedded elements . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.3 Interfa e separation models . . . . . . . . . . . . . . . . . . . . . 143
Literature
T. Belyts hko, W.K. Liu, B. Moran: Nonlinear Finite Elements for Continua and Stru tures, John Wiley and Sons, 2000
N. Moes, J. Dolbow, T. Belyts hko: A nite element method for ra k
growth without remeshing, International Journal for Numeri al Methods
in Engineering, 1999, 46(1), 133-150
T. Belyts hko, J.S. Chen: Meshfree and Parti le Methods, John Wiley
and Sons, 2007
S. Li, W.K. Liu: Meshfree Parti le Methods, Springer, 2005
T.P. Fries, H. Matthies: Classi ation and overview of meshfree methods,
http://www.digibib.tu-bs.de/?do id=00001418
CONTENTS
Preliminaries and Notations

Tensors will be written in ompa t notation or indi ial notation. In indi ial
notation, the omponents of a tensor are expli itely spe ied, e.g fi for a rst
order tensor, fij for a se ond order tensor. We use Einsein notation and sum
over repeated indi es. Capital indi es denotes a node number, e.g giI is the
rst order tensor for node I . Alternatively, we write the tensor in ompa t
notation, e.g. fi = f or fij = f . The symbol () denotes a tensor ontra tion,
e.g fi gi = r = f g or fijkl gkl = rij is in ompa t notation f : g = r. The symbol
denotes a tensor produ t, e.g. fi gj = rij is in ompa t notation f g = r.
The symbol denotes a ve tor ross produ t, e.g. f g = ijk fi gk where ijk
is the permutation tensor. We will also used Voigt notation where suitable and
write gi = (g1 , g2 , g3 , g12 , g13 , g23 ) instead of gij . When we refer to quantities in
the initial onguration, we use apital letters; if we refer to quantities in the
urrent onguration, small letters are used. Therefore, let us onsider a body
with boundary ; their images in the initial state are 0 and 0 , respe tively1 .
The initial state will also serve as the referen e state. The motion is des ribed
by
x = (X, t),
(1)
where x are the spatial (Eulerian) oordinates and X the material (Lagrangian)
oordinates. The displa ement is then given by
u(X, t) = x X = (X, t) x,
(2)
For example, let us onsider the material time derivatives of the velo ity and
a eleration eld in a Lagrangian des ription2 , then we an write:
v(X, t)
a(X, t)
u(X, t)
= u
t
2
u(X, t)
=u
t2
(3)
where u is the displa ement and v and a the velo ity and a eleration, respe tively. In Eulerian des ription, adve tive terms are present in the total time
1 The subs ript 0 is always used when
2 in terms of material oordinates
we refer to quantities in the initial onguration
CONTENTS
derivatives:
a(X, t)
a(X, t)
v(X, t) vi (x, t) xi (X, t)

+
t
xj
t
v(X, t) vi (x, t)
v
+
t
xj
(4)
A Lagrangian des ription is generally used for solids while an Eulerian des ription of motion is used in uid me hani s. The deformation gradient is dened
by
x
F=
(5)
X
A strain measured that an be derived from the deformation gradient is given
by
u
= IF
=
(6)
X
Another useful strain measure is the rate-of-deformation tensor

D = 0.5 L + LT
(7)
where L = vi,j = F F1 is the velo ity gradient. The Green Lagrange strain
tensor is dened by:

E = 0.5 FT F I
(8)
The work- onjugate stress-tensor of the rate-of-deformation tensor is the Cau hystress tensor and e.g. for E, it is the se ond Piola Kir hho stress tensor. The
dierent stress measures an be transformed into ea h other. For more details,
see Belyts hko et al. [1.
CONTENTS
Symbols
[[()]]
jump
D()
t , ()
material time derivative
()
X , , (),i
nabla operator
Subs ripts/Supers ripts

S
h
u
t
c
P
L
AL
std
enr
blnd
lin
(e)
0
max
min
ext
int
Q
a, b
diag
kin
symmetri part of a tensor or Shepard fun tion or stress point or

ra k surfa e
approximation...
Diri hlet...
von Neumann... or ra k tip...
ra k...
parti le... or penalty...
Lagrange multiplier...
augmented Lagrange...
standard...
enri hed...
blending/transition...
linear...
element...
refer to quantities in the referen e/initial onguration
maximum...
minimum...
external...
internal...
Gauss point...
enri hed node
diagonal...
inertia...
Latin letters
E
G
KI , KII
x, x
X, X
u, u
d
Young's modulus
Shear modulus
Stress intensity fa tor
spatial oordinates
material oordinates
displa ement
displa ement ve tor
10
v, v
a, a
t, t
n
b
p, p
m, m
M, M
w
W
V
A
h
R
f
F
r
P, P
K
N, N
B
C
I
J
e
r, s
S
H
S
CONTENTS
velo ities
a elarations or enri hment
tra tion
normal
body for es or enri hment
basis polynomial or penalty parameter
mass or mass matrix
mass or mass matrix
quadrature weights or meshfree shape fun tion
kernel/window/weighting fun tion
volume or quadrature weight
arrea
dilation parameter
support size
for e
deformation gradient or for e
residuum
rst Piola Kir hho stress
stiness matrix
shape fun tion
B-operator or bran h enri hment
material tangent matrix
identity tensor
Ja obian
error or internal energy
lo al element oordinates
ra k surfa e or step fun tion
Heaviside fun tion
set of nodes
Greek letters
,
,
K
ijk
,
,
Lagrange multiplier or Lagrange multiplier estimate or eigenvalue

Lagrange multiplier estimate or Lagrange multiplier shape fun tion
potential
s aling parameter
in rement...
s aling parameter
ondition number
urvature
permutation tensor
strain
Cau hy stress
hoop stress
CONTENTS
,
,
enri hment fun tion

shape fun tion or smoothing fun tion
level set
meshfree shape fun tion
variation or Krone ker delta or dira delta fun tion
lo al element oordinates
density
domain
boundary
11
12
CONTENTS
Chapter 1
Introdu tion
Finite element methods are a very powerful numeri al tool to study, predi t
and model the behavior of materials, stru tures, uids and events. They were
su essfully applied in many areas in Engineering S ien e. Appli ations range
from the automobile industry, Aerospa e and Aeronauti al Engineering, Environmental Engineering, over Me hani al and Civil Engineering to Geome hani s, Biome hani s and Material s ien e. Even s ientists from dis iplines su h as
applied maths, physi s and omputational hemistry use nite element methods. Finite element methods an be used as design tool. A typi al Engineering
appli ation is the design of ars, ships, air rafts, bridges, dams, armor, new
materials et . Many aspe ts are studied with nite elements in luding uid
me hani s, solid me hani s and uid-uid-intera tion as well as uid-stru ture
intera tion, e.g. the aerodynami behavior, pollution, a ousti emission, safety
and reliability analysis, et . Finite elements an also be used as predi tive tools,
e.g. ra k growths of existing aws in wings of air rafts, avalan hes, land slides,
damage of human tissue in arteries or the pulmonary system, to name a few
appli ations.
However, while standard nite elements have a wide appli ation spe trum,
there are still limitations where standard nite elements do not give a eptable
results. One lass of problems that nite elements are not able to apture su iently are problems that involve strong and weak dis ontinuities, i.e. jump in
the displa ement and strain eld, respe tively. There are two huge appli ation
spe tra that involve strong and weak dis ontintuities:
1. Fra ture and omputational failure me hani s that involves the initiation
and propagation of ra ks, i.e. strong dis ontinuities.
2. So- alled interfa e problems, that an be lassied into solid-solid, soliduid and uid-uid interfa es. While solid-uid interfa e problems are
appli ations with strong dis ontinuities, solid-solid and uid-uid interfa e
problems involve weak dis ontinuities.
13
14
CHAPTER 1. INTRODUCTION
Standard nite elements are not well suited for su h kind of problems due to
their smooth interpolation hara ter. The shape fun tion of standard nite
elements are pie ewise polynomials and strong and weak dis ontinuities an be
in orporated only along element edges. For ra ks for example, the numeri al
results be ome very sensitive with respe t to the orientation of the mesh and
also with respe t to the hosen size and shape of the elements. Su ient results
an only be obtained by omputational expensive remeshing pro edures. The
data from the old mesh has to be mapped to the new mesh that also leads to
loss of a ura y. Similar problems o ur in uid me hani s when the inferfa e
of two dierent uids or a uid and a stru ture has to be aptured.
Within this le ture, extended nite element and meshfree methods will be
studied that an handle arbitrary strong and weak dis ontinuities. These methods modify the approximation spa e of the test and trial fun tion, su h that
arbitrary strong and weak dis ontinuities an be handled. Though it sounds
at rst illogi al, extended nite element methods were born out of meshfree
methods. Hen e, the le ture will rst dis uss meshfree methods and some basi on epts that are inherent of extended nite element methods. We will see
that meshfree methods have ertain advantages over nite element methods
with respe t to ertain appli ations. These aspe ts will be dis ussed within this
le ture.
At the end of the le ture you should know
what is a global and local partition of unity and how it is related to
ompleteness
what is ompleteness and ontinuity
what is a Lagrangian and Eulerian kernel fun tion
what is SPH, RKPM, EFG, PUFEM and how to derive the shape fun tions
of these methods
what is the dieren e between an intrinsi and extrinsi basis
how to derive a weak form from a strong form (and vi e versa)
how to perform a nodal, stress-point and ell-integration in a meshfree
method
how to derive dis rete meshfree equations
why meshfree methods have di ulties in imposing Diri hlet boundary
onditions and how to impose them
what is a strong and weak dis ontinuity
how to in orporate strong dis ontinuities in a meshfree method
what is a level set and a signed distan e fun tion
15
the prin ipal of XFEM
how to in orporate a strong and weak dis ontinuity in XFEM
the implementation pro edure in XFEM
dierent possibilities to lose a ra k at its ra k tip within XFEM
the prin ipal of Embedded elements and Interfa e elements
16
CHAPTER 1. INTRODUCTION
Chapter 2
Meshfree methods
2.1 Completeness, onsisten y and Partition of
Unity
Completeness in nite element and meshfree methods plays the same role as
onsisten y in nite dieren e methods, Belyts hko et al. [1. While for nite
dieren es, onsisten y des ribes how good the dieren e s heme approximates
the dierential operator, ompleteness is expressed in terms of the order of the
polynomial whi h must be represented exa tly. For fun tions depending on the
oordinates X, ompleteness an be stated as follows: a set of approximating
fun tions J (X) reprodu e p(X) with uJ = p(XJ ), when
J (X) uJ = J (X) p(XJ ) = p(X)
(2.1)
If the approximation is able to reprodu e a onstant fun tion, then the approximation is alled zero-order omplete. If the approximation an reprodu e linear
fun tions exa tly, it is alled rst-order omplete or linear omplete and so on.
The term completeness is sometimes referred to as reproducing conditions. In
two dimensions, the onstant and linear reprodu ing onditions are given by
J (X) XJ = X
or in indi ial notation
J (X) = 1
(2.2)
J (X) YJ = Y
(2.3)
J (X) XJi = Xi
(2.4)
If the derivatives of a polynomial eld is reprodu ed orre tly, we all this the
derivative reprodu ing onditions. In two dimensions, the derivative reprodu 17
18
CHAPTER 2. MESHFREE METHODS
ing onditions for a onstant and linear eld are given by

X
X
J,X (X) = 0
J,Y (X) = 0
X
J,X (X) XJ = 1
(2.5)
J,Y (X) XJ = 0
(2.6)
J,X (X) YJ = 0
J,Y (X) YJ = 1
(2.7)
J,i (X) XJj = ij
(2.8)
or in indi ial notation

X
J,i (X) = 0
X
J
If the approximating fun tions J (x) an reprodu e the onstant and assuming that uJ = 1, the approximation should be exa tly unity (see eq. (2.2))
X
J (x) = 1
J
Fun tions that possess this property are alled partition of unities (PU).
Isoparametri nite element shape fun tions meet this ondition. Hen e, isoparametri nite elements an be lassied in the ategory of partition of unity
methods.
2.2 Conservation, stability and onvergen e

Belyts hko et al. [2 have shown that a dis retization has to be zero-order omplete to guarantee onservation of linear momentum and linear omplete to
guarantee onservation of angular momentum.
Conservation of linear momentum requires that the rate of hange of linear
momentum equals the total applied for e su h that the total hange of linear
momentum due to internal for es is zero. Thus, in the absen e of external for es
and body for es, onservation of linear momentum requires that
!
X
D X
mI v I = 0
mI vI =
(2.9)
Dt
IS
IS
where mI are nodal masses and v the velo ity eld. From the linear momentum
equation (without external and body for es) we know that
X
mI v I =
(2.10)
I (XJ ) (XJ ) wJ
JS
where I (XJ ) are shape fun tions and wJ are the quadrature weights. Substituting the RHS of eq. (2.10) into eq. (2.9) gives
X
XX
XX
mI vI =
I (XJ )(XJ ) wJ =
I (XJ )(XJ ) wJ = 0
IS
IS JS
JS IS
(2.11)
19
2.3. CONTINUITY
that requires that
IS
I (XJ ) = 0 for arbitrary stress states that in turn
requires zero-order omplete shape fun tions.

Conservation of angular momentum requires that any hange in angular
momentum is ex lusively due to external for es. Hen e, let us show that the
hange of angular momentum in the absen e of external for es vanishes. The
time rate of hange of angular momentum in the absen e of external for es an
be expressed as
!
X
D X
(2.12)
mI v I XI + vI vI = 0
mI vI XI =
| {z }
Dt
I
=0
Here, denotes the ve tor ross produ t. Substituting eq. (2.10) into eq. (2.12)
leads to
!
!
X
X
D X
ijk
I,m (XJ ) mj (XJ )wJ XIk (2.13)
mI vI XI =
Dt
I
J
I
where ijk is the permutation tensor and XIk refers to the k th omponent of
parti le I . We sum over repeated indi es. Eq. (2.13) an be reformulated to
!
X
X X
mj (XJ )wJ
ijk
I,m (XJ )XIk mj (XJ )wJ = ijk mk
J
{z
mk
X
J
ijm mj (XJ ) wJ = 0 (2.14)

{z
}
|
=0
Note, that we used the linear reprodu ing onditions of the derivatives of the
approximation and the symmetry of the Cau hy stress tensor.
A method is onvergent of order k (k > 0) if
k
max
|{z} |u(Xi ) ui | Ch
(2.15)
where C is a onstant and h is the dilation parameter in meshfree methods1

dened later. A ording to the Lax-Ri htmeyr equivalen e theorem, a method
is onvergent if it is onsistent and stable. Stability guarantees that a small
defe t in the dis retization remains small. As already mentioned, ompleteness
takes the role of onsisten y in a Galerkin methods.
2.3 Continuity
An approximation is onsidered to be Cn if its approximating fun tions are n
times ontinuous dierentiable.
1 h would denote the element size in nite elements
20
R_K
R_I
K
I
Support size of particle I
Figure 2.1: Support size of the kernel fun tion
2.4 The weighting fun tion

2.4.1 Requirements of the weighting fun tion
The weighting fun tion plays an important role in meshfree methods. Other
names for the weighting fun tion are window and kernel fun tion. Typi ally,
three requirements are imposed on the kernel fun tion:
(XI XJ )
(2.16)
(2.17)
W (XI XJ , h0 ) =
lim W (XI XJ , h0 ) =
|{z}
h0 0
W (XI XJ , h0 )d0
XI XJ R
(2.18)
where denotes the Dira delta fun tion, h0 is a shape or dilation parameter
that determines the size of the domain of inuen e of the kernel and R is a
shape parameter often related to h0 , see gure 2.1. The last ondition is alled
ompa t support of the shape fun tion. It an be shown that there is an optimal
value for the ratio between the dilation parameter h0 and the distan e between
parti les x. Figure 2.2 shows that for a xed distribution of parti les, x onstant,
the dilation parameter must be large enough to avoid aliasing (spurious short
waves in the approximated solution). It also shows that an ex essively large
value for h0 will lead to ex essive smoothing. For this reason, it is ommon to
21
2.4. THE WEIGHTING FUNCTION
maintain a onstant ratio between the dilation parameter h0 and the distan e
between parti les x. In most appli ations, the ratio between the support size2
and parti le separation x lies in the range between 2.8 and 3.2. Usually, also
the following symmetry onditions are imposed on the kernel fun tion:
W (XI XJ , h0 ) =
0 W (XI XJ , h0 ) =
W (XJ XI , h0 )
0 W (XJ XI , h0 )
(2.19)
(2.20)
There are basi ally two ways to onstru t kernel fun tions in higher order
dimensions that lead to dierent sizes and shapes of the domain of inuen e.
Either, the kernel fun tion has spheri al support
W (X) = W1D (kXk),
(2.21)
or the kernel fun tion has re tangular support

W (X) = W1D (|X1 |) W1D (|X2 |) W1D (|X3 |)
p
where X = (X1 , X2 , X3 ) and kXk = X12 + X22 + X32 .
(2.22)
2.4.2 Spe i weighting fun tions
A very popular kernel fun tion is the ubi B-spline given by:

C
hD 1 1.5z 2 + 0.75z 3 0 z < 1
3
C
W(r) =
1z2
4 hD (2 z)
0
z>2
where D is the dimension, z = r/h0 and

dimension
2/3
10/(7 )
C=
1/
(2.23)
C is a onstant depending on the

D=1
D=2
D=3
(2.24)
Note, that it is standard to s ale the weighting fun tion with the dilation parameter h0 and express it in non-dimensional form (in terms of z ). The rst
spatial derivatives of the ubi B-spline are obtained by standard dierentiation.
Dening z = ||XI XJ ||, we obtain
W z
W
=
XiJ
z XiJ
with
W
z
3C
2
hD+1 z + 0.75z
2
3C
4 hD+1 (2 z)
(2.25)
0z<1
1z2
z>2
(2.26)
The one dimensional ubi B-spline and its rst spatial derivative is shown in
2 Note
that the support size is proportional to the dilation parameter
22
1.2
1
0.8
1.4
0.6
1.2
0.4
1
0.8
0.2
u(x)
urho(x)
0.6
0.4
0.2
0
3
0.2
1
a) h/x = 1
0.5
0.5
0.5
0.5
(b) h/x = 1
1.2
1
0.8
0.7
0.6
0.6
0.4
0.5
0.4
0.2
u(x)
urho(x)
0.3
0.2
0.1
0
3
0.2
1
) h/x = 2
0.5
(d) h/x = 2
1.2
1
0.8
0.6
0.35
0.3
0.4
0.25
0.2
0.2
u(x)
urho(x)
0.15
0.1
0.05
0
3
e) h/x = 4
0.2
1
0.5
(f) h/x41
Figure 2.2: Meshfree (SPH) approximation fun tions and approximation of

u(x) = 1 x2 with ubi spline window fun tion, distan e between parti les
x = 0.5 and quadrature weights i = x, for h/x = 1, 2, 4, from [3
23
2.4. THE WEIGHTING FUNCTION
Figure 2.3: Typi al kernel fun tion and its rst derivative in one dimension
gure 2.3. Another popular kernel is the quarti spline:

1 6z 2 + 8z 3 3z 4 0 z < 1
W(r) =
0
1z
Other kernels are for example:
kx xI k r
z 2 log z
2
2
W(r) =
ez /c
q
z 2 + R2
linear
thin plate spline
Gaussian
multipolar
(2.27)
(2.28)
where c, R and q are shape parameters.
2.4.3 Eulerian and Lagrangian kernels

When the kernel fun tion is expressed in terms of spatial oordinates, we all it
an Eulerian kernel:
WJ (x) = W (x xJ (t), h(x, t))
(2.29)
The radius h of the support depends on the spatial oordinates and an hange
in time. Simple approa hes to update h are given by:
ht+t =
h =
ht + h t
1/3 v
(2.30)
24
where v indi ates the parti le velo ity and is the density. However, the shape
of the domain of inuen e of a given parti le still remains a sphere. More
realisti tensorial support sizes h an be obtained by means of the deformation
gradient F:
h = h0 F
(2.31)
For radial supported shape fun tions, h spans a sphere in the initial onguration with h0 as radius. In the urrent onguration, the domain of inuen e
be omes an ellipsoid where the eigenve tors of h span the axis of that ellipsoid.
When the kernel fun tion is expressed in terms of material oordinates, it is
alled a Lagrangian kernel:
WJ (X) = W (X XJ , h0 )
(2.32)
For Lagrangian kernels, the neighbors of inuen e do not hange during the
ourse of the simulation but the domain of inuen e in the urrent onguration
hanges with time. For radial kernel fun tions, the domain of inuen e in the
initial onguration is a ir le in two dimensions. In the urrent onguration
the domain of inuen e an be extremely distorted. This is a disadvantage in
simulating uid ow problems and other problems with very large distortions.
On the other hand, it an be shown that many meshfree methods based on
Eulerian kernels introdu e a so alled tensile instability that is not present in
meshfree methods based on Lagrangian kernels.
Note that most meshfree methods that use an Eulerian kernel are based on
a Lagrangian des ription of motion. In other words, the standard derivation of
the a eleration omits the time dependen e of xJ (t). If we let
X
v(x, t) =
(2.33)
W (x xI (t)) vI (t),
IS
then the a elerations are

X
a=
W (x xI (t)) v I + W (x xI (t)) x I vI .
(2.34)
IS
The se ond term is negle ted in most Eulerian kernel based meshfree methods.
This in onsisten y auses severe di ulties when modeling material failure with
Eulerian based meshfree methods.
2.5 Spe i meshfree approximations

A meshfree approximation an be written as the approximation for nite elements:
X
uh (X, t) =
(2.35)
uJ (t) J (X)
JS
where uJ are the nodal parameters, J (X) are the shape fun tions and S is
the set of nodes with J (X) 6= 0. We note one substantial dieren e between
2.5. SPECIFIC MESHFREE APPROXIMATIONS
25
the meshfree and the nite element approximation. While nite elements are
real interpolants, meshfree shape fun tions are approximants, i.e. they do not
go through the data (uh (xI ) 6= uI for meshfree methods). In other words,
meshfree methods do not fulll a ondition often refered to as Krone ker-Delta
property I (XJ ) = IJ where IJ is the Krone ker delta. This ondition is
obviously fullled by a nite element formulation but it is violated by almost
every meshfree approximation. This auses di ulities in imposing Diri hlet
boundary onditions. If the shape fun tions do not vanish along the boundaries,
they do not allow for the spa e H1 to be represented exa tly. This issue will be
addressed in detail in se tion 2.9.
In the following, we will derive all methods in terms of material oordinates (and in terms of a Lagrangian kernel). The equations in terms of spatial
oordinates are obtained in the same manner.
2.5.1 Smooth Parti le Hydrodynami s (SPH)

The ontinuous SPH-form
The ontinuous SPH approximation is given by:
Z
h
u(Y, t) W (X Y, h0 (Y)) dY
u (X, t) =
(2.36)
where 0 is the domain of the problem. The onstant reprodu ing ondition
Z
(2.37)
W (X Y, h0 (Y)) 1 dY = 1
0
is fullled by denition, see eq. (2.16). Linear ompleteness requires that

Z
W (X Y, h0 (Y)) Y dY = X
(2.38)
0
Noting that eq. (2.37) requires that

Z
W (X Y, h0 (Y)) X dY = X
(2.39)
and subtra ting eq. (2.39) from eq. (2.38) leads to

Z
W (X Y, h0 (Y)) (X Y ) dY = 0
(2.40)
This integral is the rst moment of the weighting fun tion. The above equation is met if the weighting fun tion is symmetri about its origin. Hen e, the
ontinuous SPH form is linear omplete.
The spatial partial derivatives of uh (X, t) are given by
Z
h
0 u(Y, t) W (X Y, h0 (Y)) dY
0 u (X, t) =
(2.41)
0
26
Integration by parts gives

0 uh (X, t) =
0 [u(Y, t) W (X Y, h0 (Y))] dY
0 u(Y, t) W (X Y, h0 (Y)) dY
(2.42)
Applying the divergen e theorem on the rst term of the RHS of eq. (2.42), we
obtain
Z
u(Y, t) W (X Y, h0 (Y)) n0 d0
0 uh (X, t) =
0
Z
0 u(Y, t) W (X Y, h0 (Y)) dY
(2.43)
0
In all SPH versions, the surfa e integrals are negle ted so that the nal approximation of the derivatives are
Z
h
0 u(Y, t) W (X Y, h0 (Y)) dY
0 u (X, t) =
(2.44)
0
The dis rete SPH-form

In the dis rete SPH method, the shape fun tions are given by a produ t of the
parti le volume and the weighting fun tion:
J (X) = W (X XJ , h0 ) VJ0
(2.45)
where VJ0 is the volume in the initial onguration3 asso iated with the node J .
In the appli ation to partial dierential equations (PDEs), an approximation of
a fun tion gradient is needed. In SPH, an approximation of the gradient of a
fun tion is given by
X
0 uh (X) =
uJ 0 J (X) with 0 J = 0 W (X XJ , h0 ) VJ0 (2.46)
JS
where the minus sign results from the integration by parts, see eq. (2.42).
It an easily be shown that the dis rete SPH form is unable to reprodu e even
onstant fun tions orre tly even for uniformly spa ed parti le arrangements.
Therefore, onsider a one-dimensional uniform parti le arrangement as shown in
gure 2.4. Parti les are lo ated at the positions 2,3,4,5. The dilation parameter
is hosen to be twi e the parti le distan e, we use the ubi B-spline as kernel
fun tion and try to reprodu e the onstant fun tion of 1. Figure 2.4 shows the
values of every single kernel fun tion and its superposition. We note that the
zero-order ompleteness is not fullled at the boundaries. For an uneven parti le
arrangement, the results will get even worse.
3 Note
that VJ0 is the quadrature weight
27
Figure 2.4: Shape fun tions of the SPH approximation in one dimension
Sin e equation (2.46) does not fulll zero-order ompleteness on the boundary even for a regular parti le onguration, Gingold and Monaghan [4 introdu ed a so alled symmetrization. In this pro edure they assume that
!
X
0
(2.47)
uI 0
0 W (X XJ , h0 ) VJ
JS
although this is only true for a uniform distribution of parti les away from any
boundary. Note that eq. (2.47) is equivalent to assuming zero-order ompleteness (note the quantity inside the parenthesis in (2.47)). Subtra ting equation
(2.46) to equation (2.47) gives
0 uh (X) =
(uJ uI ) 0 W (XI XJ , h0 ) VJ0
(2.48)
JS
A remarkable feature of the symmetrization pro edure is that it yields zeroorder ompleteness for the derivatives of a fun tion for an irregular parti le
arrangement.
2.5.2 Corre ted SPH forms

The Krongauz-Belyts hko orre tion
Johnson and Beissel [5 noted that errors in the extensional strains due to la k
of linear ompleteness ould be orre ted by simple s aling, thus improving
a ura y. Another orre tion that enables the derivatives of onstant or linear
28
elds to be reprodu ed exa tly was developed by Krongauz and Belyts hko [6.
The orre ted derivatives are approximated by
0 uh (X, t) =
or in iditial notation
uh,i (X, t) =
GI (X) uI (t)
(2.49)
IS
GiI (X) uI (t)
(2.50)
IS
where GI is a linear ombination of the Shepard fun tions:

WI (X)
WIS (X) = P
WI (X)
(2.51)
IS
The approximation fun tions for the derivatives GI are dened as linear ombinations of the exa t derivatives by a linear transformation
S
GI (X) = a(X) 0 WIS (X) = aij (X)WjI
(X)
(2.52)
where a(X) are arbitrary parameters. To obtain linear omplete reprodu ing
onditions for the derivatives of the approximation, the following reprodu ing
ondition for the derivative of a linear fun tion must be fullled (see also eq.
(2.8)):
X
IS
(2.53)
GI (X) XI = ij
Let A be the matrix of the ross produ t between the derivatives of the Shepard
fun tion and the Lagrangian oordinate ve tor. Then the parameters a an be
easily determined from
(2.54)
A aT = I
where I is the identity matrix and
A=
S
WI,X
XI
S
WI,X
YI
a=
aXX
aY X
S
WI,Y
XI
S
WI,Y
YI
aXY
aY Y
Finally, we obtain the approximation for the derivatives of a fun tion from (2.52)
and (2.53) as
0 uh (X, t) =
X
IS
a(X) 0 WIS (X) uI (t)
(2.55)
29
Alternative 1
The orre ted derivatives an also be obtained by a orre tion of the form
I
GXI
=
=
(a11 (X) + a12 (X) + a13 (X)) WIS (X)

(a21 (X) + a22 (X) + a23 (X)) WIS (X)
GY I
(a31 (X) + a32 (X) + a33 (X)) WIS (X)
(2.56)
If the oordinate system is shifted to the point of evaluation X, then the oef ients a are obtained as in (2.54) with A given by
1
XI X
YI Y
X
(XI X)2
(XI X)(YI Y )
A = WIS (X) XI X
YI Y (XI X)(YI Y )
(YI Y )2
I
In ontrast to the method above, a 3 3 matrix needs to be inverted. An

advantage of this method is that linear omplete shape fun tions are obtained
in addition to the linear omplete orre ted derivatives.
Alternative 2
Consider shape fun tions and shape fun tion derivatives of the form
I
S
S
a11 (X)WI,X
(X) + a12 (X)WI,Y
(X) + a13 (X)WIS (X)
GXI
S
S
a21 (X)WI,X
(X) + a22 (X)WI,Y
(X) + a23 (X)WIS (X)
GY I
S
S
a31 (X)WI,X
(X) + a32 (X)WI,Y
(X) + a33 (X)WIS (X) (2.57)
to reprodu e linear fun Now, a is obtained by requiring the approximation

tions and the derivatives of the approximation to reprodu e the derivatives of
linear fun tions. To a hieve this, (2.57) is substituted into the onstant and linear reprodu ing onditions, eqs. (2.2) to (2.8). Shifting the oordinate system
to the point of evaluation X, the oe ients a an be obtained by (2.54) with
S
S
WI,X
(X)
WI,Y
(X)
WIS (X)
X
S
S
(X) XI WI,Y
(X) XI WIS (X) XI
A = WI,X
S
S
I
WI,X (X) YI WI,Y (X) YI WIS (X) YI
As in Alternative 1, the omputational ost is higher, but therefore the shape

fun tions themselves are able to reprodu e linear fun tions.
Interpolation estimate
Let us show for the two dimensional ase that for derivative approximations
whi h satisfy linear reprodu ing onditions, the errors in derivative approximants are of order O(h). Therefore, let us develop the fun tion u(X) into a
Taylor series expansion around X:
u(XI ) =
+
+
+
u(X) + u,X (X) (XI X)
u,Y (X) (YI Y ) + 0.5u,XX (X) (XI X)

u,XY (X) (XI X) (YI Y )
0.5u,Y Y (X) (YI Y )2 + O(h3 )
(2.58)
30
Ignoring higher order terms, we an write uh,X (X) u,X as

uh,X (X) u,X
GXI (X) uI u,X
GXI (X) u(XI ) u,X
(2.59)
Substituting (2.58) into (2.59), we obtain

uh,X (X)
u,X
= u(X)
X
I
+ u,Y (X)
GXI (X) + u,X (X)
X
I
GXI (X) (YI Y )
+ 0.5 u,XX (X)

+ u,XY (X)
X
I
+ 0.5 u,Y Y (X)
GXI (X)(XI X) 1
GXI (X)(XI X)2
GXI (X)(YI Y )2
GXI (X)(XI X) (YI Y )

I
(2.60)
P
P
GXI (XI
With the onstant and linear reprodu ing onditions GXI = 0,
I
I
P
X) = 1,
GXI (YI Y ) = 0, eq. (2.60) an be simplied
I
uh,X (X) u,X
0.5 u,XX (X)
u,XY (X)
0.5 u,Y Y (X)
GXI (X)(XI X)2
GXI (X)(YI Y )2
GXI (X)(XI X) (YI Y )

I
(2.61)
From (2.61), we obtain

|uh,X (X) u,X |
0.5 |u,XX (X)| |
|u,XY (X)| |
0.5 |u,Y Y (X)| |
GXI (X)(XI X)2 |
GXI (X)(YI Y )2 |
GXI (X)(XI X) (YI Y )|

I
(2.62)
Sin e the shape fun tions have ompa t support, there exists a onstant d su h
that for any X = (X Y )
|XI X| d,
|YI Y | d
(2.63)
31
Then
|uh,X (X) u,X |
X
|
GXI (X)|
(0.5 |u,XX (X)| + |u,XY (X)| + 0.5|u,Y Y (X)|) d2
(2.64)
If we assume that GXI satises

|GXI |
C1
h0
(2.65)
then if the supports of the shape fun tions are hosen that the size of the support
0 , then
is proportional to the lo al renement parameter h0 , i.e. d = dh
|uh,X (X) u,X |
C
(0.5 |u,XX (X)| + |u,XY (X)| + 0.5|u,Y Y (X)|) h0(2.66)
So the approximation error in the derivative is of order h for a orre ted derivative approximation. Similar observations hold for the Y -derivative.
The Randles-Libersky orre tion

Randles and Libersky [7 developed a similar orre tion whi h they alled normalization. To fulll the rst-order ompleteness, they modied the SPH approximation for the gradient of the fun tion u with a matrix B:
0 uh (X, t) =
with
B(X) =
JS
(uJ (t) uI (t)) 0 W (XJ X, h0 ) VJ0
(XJ X) 0 W (XJ X, h0 ) VJ0
JS
B(X) (2.67)
!1
(2.68)
If the SPH shape fun tions W (X XJ , h) VJ0 are repla ed by the Shepard
fun tions, the expression for B be omes:
!1
X
S
B(X) =
(2.69)
XJ 0 W (XJ X, h0 )
JS
whi h is similar to the Krongauz-Belyts hko orre tion; subtle dieren es arise
from the symmetrization in (2.67) and (2.68). Note, that the SPH shape fun tions in equation (2.67) have to be repla ed by the Shepard fun tions when using
eq. (2.69) for the omputation of B. The approximation for the gradient of the
fun tion u is then formulated in the unsymmetrized form be ause the Shepard
fun tions are zero-order omplete by onstru tion:
!
X
0 uh (X, t) =
(2.70)
uJ (t) 0 W S (XJ X, h0 ) VJ0 B(X)
JS
32
Reprodu ing kernel parti le method (RKPM)

The Reprodu ing kernel parti le method (RKPM) [8 an be onsidered as an
improvement of the ontinuous SPH approximation. In order to in rease the
order of ompleteness of the approximation, a orre tion fun tion C(X, Y) is
introdu ed into the approximation:
Z
h
C(X, Y)W (X Y)u(Y)dY
u (X) =
(2.71)
Y
with K(X, Y) = C(X, Y)W (X Y) and where C(X, Y) is dened su h that

the approximation is n-th order omplete.
u(X) = pT (X)a
(2.72)
p(X)u(X) = p(X)pT (X)a
(2.73)
p(Y)W (X Y)u(Y)dY =
p(Y)pT (Y)W (X Y)dY a
(2.74)
This is a system of equation from whi h a is solved and substituted into the
approximation uh (X) = pT (X)a, it yields
Z
uh (X) = pT (X)
p(Y)pT (Y)W (XY)dY
1 Z
p(Y)w(XY)u(Y)dY
Y
(2.75)
with the orre ted fun tion
C(X, Y)
Z
p (X)
T
p(Y)p (Y)W (X Y)dY
pT (X)[M(X)]1 p(Y)
1
p(Y)
(2.76)
To evaluate this ontinuous expression, numeri al integration must be employed.

This step leads from the reprodu ing kernel method to its dis rete version, the
reprodu ing kernel parti le method [8.
h
u (X)
X
IS
C(X, Y)W (X Y)u(Y)dY
C(X, XI )w(X YI )uI VI0
= pT (X)[M(X)]1
X
IS
p(XI )W (X XI )uI VI0
(2.77)
(2.78)
33
The moment matrix M(X) is also omputed by numeri al integration

Z
M(X) =
IS
p(Y)pT (Y)W (X Y)dY
p(XI )pT (XI )W (X XI )VI0
(2.79)
2.5.3 Element-free Galkerin method (EFG)

The EFG method employs a moving least square approximation that was originally developed for urve tting. We will go one step ba k and rst look at
least-square tting.
Least square approximation

For the sake of simpli ity, we rst present this method in one dimension. The
method is then naturally extended to higher dimensions. Let us onsider the
data tting problem where we want to nd a fun tion uh (x) tting the data
points (xI , uI ) with uI = u(xI ). Assume that the approximation fun tion uh (x)
is a polynomial of order m:
uh (x) = a0 + a1 x + a2 x2 + ... + am xm
(2.80)
that an be written in ompa t form as

(2.81)
uh (x) = pT (x)a
uh(x)
xi
uh(xi)
ui
X
0
xi
Figure 2.5: Data tting using least square method

The parameters a an be determined by minimizing the square of dieren e
between uI and uh (xI )
J=
n
X
I=1
[uh (xI ) uI ]2 =
n
X
I=1
[pT (xI )a uI ]2
(2.82)
34
Taking derivatives with respe t to a leads to the equation

n
X
p(xI )pT (xI )a =
n
X
p(xI )uI
(2.83)
I=1
I=1
whi h allows us to solve for the unknowns a. Then the approximation fun tion
uh (x) is ompletely dened.
Example Let us onsider the data tting example given in table 2.5.3.
xI
uI
1
1
2
1.5
3
4
Let us dene a line approximating these points. In this ase, we have

pT (x) = [1
x] aT = [a0
a1 ]
(2.84)
Eq. (2.83) be omes

3
X
xI
1
u
a=
xI I
xI x2I
I=1

3 6
6.5
a=
6 14
16
3
X
1
I=1
(2.85)
(2.86)
The solution is a0 = 5/6 and a1 = 1.5, hen e the approximating line is given
by
5 3
uh (x) = + x
(2.87)
6 2
It is lear that this way of approximation equals the role of all data points whi h
often gives ina urate results if some points are more important than others.
Moving least square approximation

In the moving least square (MLS) approximation, the unknown a will depend
on the oordinates X or X when we onsider higher order dimensions. The
MLS-approximation fullls the reprodu ing onditions by onstru tion, so no
orre tions are needed. To satisfy the linear reprodu ing onditions, linear base
fun tions p are hosen to be
p(X) =
The MLS approximation is

uh (X, t) =
M
X
1 X
X 2
pI (X) aI (X, t) = pT (Xi ) a(Xi )
(2.88)
(2.89)
I=1
where M determines the size of the polynomial basis and with a hosen to
minimize the quadrati form
35
J(a(Xi ))
N
X
J=1
J , h0 )
W (X X
M
X
I=1
iT
!2
J ) aI (X, t) u(X
J)
pI (X
T
h
i
a(X) u(X)
(2.90)
W(X) P(X)
a(X) u(X)
P(X)
where N denotes the number of neighbors where W (X) 6= 0 and with

= u(X
1 ) u(X
2 ) ... u(X
N)
uT (X)
=
P(X)
W(X) =
1)
p1 (X
2)
p1 (X
..
.
N)
p1 (X
1)
p2 (X
2)
p2 (X
..
.
N)
p2 (X
...
...
..
.
...
1)
W (X X
0
2)
0
W (X X
..
..
.
.
0
1)
pM (X
2)
pM (X
..
.
N)
pM (X
...
...
..
.
...
Minimizing equation (2.90) with respe t to a leads to

J(a(Xi ))
a(Xi )
0
0
0
N)
W (X X
W(X) u(X)
2PT (X)
W(X) P(X)
a(X)
=0
2PT (X)
(2.91)
Bringing the left term on the RHS, we obtain

W(X) u(X)
= PT (X)
W(X) P(X)
a(X)
PT (X)
(2.92)
and solving for a gives nally

a(x)
W(X) PT (X)
PT (X)
W(X) u(X)
PT (X)
|
{z
} |
{z
}
=ARM M
(2.93)
=BRM N
With (2.89), we an write the approximation as
uh (X, t) = pT (X) A1 (X) B(X) u(X)
(2.94)
or in summation form
uh (X, t) =
M X
M X
N
X
J=1 K=1 I=1
pJ (X) A1
JK (X) BKI (X) uI (X)
(2.95)
36
where the meshfree shape fun tions I (X) are

I (X, t) =
M X
M
X
(2.96)
pJ (X) A1
JK (X) BKI (X)
J=1 K=1
Comparing the MLS approximation with the reprodu ing kernel parti le method,
eq. (2.77), we note that if VI0 in eq. (2.77) is equal to one, then both approximations are identi al.
Example Let us illustrate the omputation of the matrix
P11 ... P1N

W1 ...
0
P11 ... PM1
.
.. ..
.. ..
..
..
..
..
=
A(X)
..
.
.
.
. .
. .
.
PM1
...
PMN
...
WN
P1N
...
PMN
for the ase of M = 1, i.e. the polynomial basis p(X) = 1. Then we obtain
1
W1 ...
0

.
.. ..
..
= 1 ... 1
A(X)
..
.
. .
0
... WN
1
and the matrix A turns out to be a s alar value. It is easy to show that the
shape fun tion with the basis p(x) = 1, will lead to the shape fun tion
WI (X)
I (X) = P
WI (X)
(2.97)
IS
that is well known as Shepard fun tion and was dened already in eq. (2.51).
T
Let us onsider the ase4 of M = 3 with p(X) = [1 X Y ] , then the matrix
A is
...
0
1 x1 y1
1
1 ... 1
.
.. ..
..
..
..
= x1 ... xN
A(X)
..
.
. .
.
.
y1 ... yN
0
... WN
1 xN yN
The matrix A for this ase is a 3 3 matrix. The higher the order of ompleteness, the higher will be the order of the matrix A and the higher is the
omputational ost to determine the MLS shape fun tions.
Conditioning of the A-matrix
For the omputation of the MLS-shape fun tion, the matrix A has to be inverted. Sin e the matrix W(X) is non-singular, it an be assumed that A is
non-singular if P is non-singular. A ne essary ondition for a regular matrix A
is that the number of neighbors N is larger than size of the polynomial basis M .
For example, a linear omplete basis in two dimension p(X) = [1 X Y ] requires
at least three neighbor parti les. For a linear omplete polynomial basis and 3
4 This
orresponds to the two dimensional ase with linear ompleteness
37
b)
a)
Figure 2.6: Parti le arrangement for a linear omplete MLS shape fun tion in
2D: a) singular A-matrix, b) regular A-matrix
neighboring parti les in two dimensions (see gure 2.6), it an be shown that
the matrix A be omes singular when the parti les are lying on a line. It an be
shown that the matrix A is regular when the three nodes are lo ated in dierent oordinate dire tions, see gure 2.6. This observation might be ompared
to nite elements where the a ura y of the solution is detoriated for distorted
nite element shapes. The same observation is made in the extension to three
dimensions and for higher order polynomial basis.
A measurement of the onditioning of a matrix is the ondition number
between the maximum and minimum eigenvalue of the matrix
=
max
min
(2.98)
High values of indi ate an ill- onditioned matrix. If , then the matrix
is singular. A shifting and entering of the shape fun tion as will be des ribed
later improves the onditioning of A. Another opportunity to improve the onditioning of the matrix A is by Gram-S hmidt orthogonalization.
Spatial derivatives of the shape fun tions
The rst derivatives of the MLS shape fun tions are obtained by simple dierentiation
(X)
Xi
=
+
pT (X) 1
A1 (X)
A B + pT (X)
B
Xi
Xi
B(X)
pT (X) A1 (X)
Xi
(2.99)
with
B(X)
Xi
P(X)
W(X)
Xi
(2.100)
38
To obtain the spatial partial derivatives of A1 (X), let us onsider

I = A1 (X) A(X)
(2.101)
Taking spatial derivative of eq. (2.101) we obtain

0=
A1 (X)
A(X)
A(X) + A1 (X)
Xi
Xi
(2.102)
Eq. (2.101) an be reformulated to

A1 (X)
Xi
=
=
A(X) 1
A (X)
Xi
W(X) PT (X)
A1 (X)
A1 (X) P(X)
Xi
A1 (X)
(2.103)
The se ond partial derivatives are then

2 (X)
Xi Xj
=
+
+
+
2 pT (X) 1
A (X) B(X)
Xi Xj

B(X)
pT (X) A1 (X)
B(X) + A1 (X)
2
Xi
Xj
Xi

2 1
2
B(X)
A1 (X) B(X)
A (X)
B(X) + A1 (X)
+
pT (X)
Xi Xj
Xi Xj
Xi
Xj

1
A (X) B(X)
(2.104)
pT (X)
Xj
Xi
Fast evaluation of the MLS shape fun tions
A fast evaluation pro edure for the gradient of the shape fun tion J an be
derived by rewriting eq. (2.96):
J (X) = (X) p(XJ ) W (X XJ , h0 )
(2.105)
A(X) (X) = p(XJ )
(2.106)
0 A(X) (X) + A(X) 0 (X) = 0 p(XJ )
(2.107)
with
su h that the oe ients an be obtained by an LU de omposition and ba ksubstitution that requires fewer omputations than inverting the matrix A.
Then the derivatives of the shape fun tions an be written as
Rearranging this equation, 0 (X) is obtained. The evaluation of the derivatives of the shape fun tions requires little extra omputer ost and, moreover,
higher order derivatives an also be omputed repeating the same pro ess.
MLS entered and s aled approa h
For omputational purposes5 , it is usual and preferable to enter in XI and

5 entering
will improve the onditioning of the matrix A
39
s ale with h0 also the polynomials involved in the denition of the meshfree
approximation fun tions. Thus, the EFG shape fun tions will read:

XI X
(X),
I (X) = W (XI , X) PT
(2.108)
h0
whi h is similar to (2.96). Re all also that typi al expressions for the window
fun tion are of the following type: W (Y, X) = W ((Y X)/h0 ). The onsisten y ondition be omes in this ase:

X
XI X
P(0) =
(2.109)
I (X) P
h0
IS
After substitution of (2.108) in (2.109) the linear system of equations that determines (X) is obtained:
(2.110)
A(X) (X) = P(0)
where
A(X) =
W (XJ , X) P
JS
XJ X
h0
XJ X
h0
(2.111)
For a varying dilation parameter h0I , h0I asso iated to parti le XI , is embedded
in the denition of the weighting fun tion:

XI X
W (XI , X) = W
(2.112)
h0I
Note that a onstant h0 is employed in the s aling of the polynomials P. The
onstant value h0 is typi ally hosen as the mean value of all the h0J . The onsisten y ondition in this ase is also (2.109). It also imposes the reprodu ibility
of the polynomials in P.
This entered expression for the EFG shape fun tions an also be obtained
with a dis rete MLS development with the dis rete entered s alar produ t

X
XJ X
XJ X
< f, g >X =
(2.113)
W (XJ , X) f
g
h0
h0
JS
The MLS development in this ase is as follows: for xed X, and for Z near X,
u is approximated as

ZX
h
T
c(X)
u(Z) u (Z, X) = P
(2.114)
h0
where c is obtained, as usual, through a least-squares tting with the dis rete
entered s alar produ t (2.113).
40
a)
(b)
Figure 2.7: Absolute error of the quadrati fun tion F (X, Y ) = X 2 + Y 2 for the
a) SPH approximation (R = 0.8), b) MLS approximation (R = 0.3); R denotes
the support size of the shape fun tion
2.5.4 Hp- louds and PUFEM

While the approximations that were des ribed previously are based on a so
alled intrinsi basis6 , the hp- loud method [9 uses a so alled extrinsi basis
to in rease the order of ompleteness:
!
L
X
X
uh (X) =
(2.115)
J (X) uJ +
pK (X) aJK
K=1
JS
where aJK are additional degrees of freedom introdu ed in the variational formulation. In ontrast to the hp- loud method, the Partition of Unity Finite
Element Method (PUFEM) uses generally nite element shape fun tions.
2.5.5 Generalized Finite Element Method (GFEM)

In GFEM [10 dierent shape fun tions are used for the "usual" part of the
approximation and the extrinsi ally introdu ed additional degrees of freedom:
uh (X) =
JS
J (X) uJ +
JS
J (X)
L
X
pK (X) aJK
(2.116)
K=1
Note that GFEM, PUFEM as well as the hp- loud method employ the partition
of unity on ept. The enri hment in these methods is usually employed in the
entire domain. Therefore, these methods are often referred to as global partition
of unity methods.
6 The
order of ompleteness is in reased by a higher polynomial basis
41
a)
(b)
Figure 2.8: Absolute error of the derivative of the quadrati fun tion F (X, Y ) =
X 2 + Y 2 for the a) SPH approximation (R = 0.8), b) MLS approximation
(R = 0.3); R denotes the support size of the shape fun tion
2.5.6 Examples
In this se tion, we will study the SPH-method, in usual and symmetrized form,
and the linear MLS approximation with respe t to their ability to approximate
some given fun tions. A ubi B-spline with ir ular support is employed as
weighting fun tion. We will look at regular as well as irregular parti le arrangements and restri t our studies to two dimensions.
Fun tion approximation at regular parti le arrangement

The fun tion F (X, Y ) = X 2 + Y 2
Consider a 25 25 regular parti le arrangement with parti le separation of

0.511. Figure 2.7 shows the absolute error of the MLS approximation and the
SPH approximation where a support size R of R = 0.6 and R = 1.6 is hosen
for the MLS and SPH approximation, respe tively. Note that a support size
of R = 0.6 orresponds to the minimum number of neighboring parti les su h
that the matrix A in the MLS approximation just remains regular. Note also
that the a ura y in reases with in reasing dilation parameter. i.e. support
size. So, even with a small dilation parameter, the error does not ex eed 0.05%
for the MLS approximation while the error in reases drasti ally with in reasing
X and Y oordinates for the SPH approximation. The highest errors o ur at
the boundary due to the la k of neighbor parti les and at lo ations with high
gradients.
The spatial derivatives of the fun tion F in x-dire tion is F,X = 2X . The
absolute errors of the numeri al approximation are shown in gure 2.8. The
absolute error for the MLS approximation away from the boundary does not
ex eed 0.005%. Also the error of the "usual" SPH approximation is with a
maximum error of 0.2% away from the boundary a urate. Ina ura ies o ur
42
at the boundaries.

F (X, Y ) = sin X 2 + Y 2
Let us onsider a more omplex fun tion F within the interval 0 X 2 ,
0 Y 2 . The partial derivatives7 of the fun tion F in x-dire tion is shown
p
in gure 2.9a. The parti le separation is set to /300 and the support size R
is set as in the last example. We will look only at the approximation of the rst
derivatives sin e the derivatives are ru ial in the appli atiion to partial dierential equations (PDE). The absolute error of the numeri al approximation is
shown in gure 2.9b, ,d. The global absolute error of the MLS approximation
is 0.01 while the error of the standard SPH approximation is 0.42. A symmetrization is able to de rease the error of the SPH approximation to a value
of 0.12.
Finally, it an be on luded that the MLS approximation produ es signi antly better results than the SPH approximation. However, for omplex fun tions, even the MLS approximation looses a ura y. The symmetrized SPH
version is able to redu e the error signi antly.
The fun tion
Error evolution for the SPH method in two dimensions

Let us onsider the error evolution of the partial derivatives of a quadrati fun tion in x-dire tion. Therefore, we will onsider a regular parti le arrangement.
The parti le volume is assumed to be V = d2 where d is the parti le separation. Furthermore, h is hosen su h that only parti le 2,4,6 and 8 are within
the domain of inuen e
of parti le 5 (gure 2.10), the parti le under onsider
ation: d < h < 2d. The standard SPH approximation of the derivatives in
x-dire tion of a fun tion is given by:
u,X (X) =
N
X
VJ WJ,X (X)uJ
(2.117)
J=1
where the minus in front of the sum results from the integration by parts that
is standard in SPH pro edures. For parti le 5, we obtain

(25)
(45)
(55)
(65)
(85)
u,X (X(5) ) = VJ W,X u2 + W,X u4 + W,X u5 + W,X u6 + W,X u8
(2.118)
(25)
(55)
Sin e the x- oordinate of parti le 2,5 and 8 are identi al, W,X = W,X =
(85)
W,X = 0. Due to the symmetry ondition of the kernel, the indi es in W IJ

an be ex hanged that results in a hange in the sign:

(56)
(54)
u,X (X(5) ) = VJ W,X u4 + W,X u6
(2.119)
Error evolution for quadrati fun tion
Let us onsider the fun tion f (X) = aX 2 + bX + c. Sin e Y is onstant for

parti le 4 and 6, the argument in y -dire tion an be omitted. In gure 2.11, the
7F
,X
= 2X cos X 2 + Y 2
43
a)
b)
d)

Figure 2.9: a) The derivative of the fun tion F (X, Y ) = sin X 2 + Y 2 . b)
Absolute error of the derivative of the sinus fun tion F (X, Y ) = sin X 2 + Y 2
for the MLS approximation (R = 0.6); b)
Absolute error of the derivative of the
sinus fun tion F (X, Y ) = sin X 2 + Y 2 for the SPH approximation (R = 1.6);
Absolute error of the derivative of the sinus fun tion F (X, Y ) = sin X 2 + Y 2
for the symmetrized SPH approximation (R = 1.6)
44
Figure 2.10: Regular parti le arrangement
Figure 2.11: Graphi al representation of the fun tion f (X) = aX 2 + bX + c and

the value of the shape fun tion of the parti les 4 and 6 for this approximation
45
quadrati fun tion f is shown in blue while the red urve shows the derivatives
of the ubi B-spline of parti le 5 in x-dire tion. The value w of the derivatives
of the spline at the lo ation of parti le 4 and 6 are illustrated in green and
yellow, respe tively. Note that w is assumed to be always positive. With eq.
(2.119), we an write now

(54)
(56)
f,X (X(5) ) = VJ W,X f4 + W,X f6
h

i
= VJ w a (x(4) )2 + bx(4) + c + w a (x(6) )2 + bx(6) + c

i
h
= VJ w a (x(5) + d)2 (x(5) d)2 + b (x(5) + d) (x(5) d)

= VJ w 4 a d x(5) + 2 b d

= 2 VJ w d 2 a x(5) + b
(2.120)
With the parti le volume VJ = d2 , we obtain at least a linear fun tion that
should be obtained by dierentiation of the fun tion f :

f,X (X(I) ) = 2 w d3 2 a x(I) + b
(2.121)
While the parameters a and b are given by the fun tion f , the parameter d and
h an be varied8 . The absolute error of the approximation an be given in terms
of d, h and x(I)

errabs (d, h, x(I) ) = 2 a x(I) + b 2 w(d, h) d3 2 a x(I) + b

=
2 a x(I) + b 1 2 w(d, h)d3
(2.122)
The relative error is obtained by s aling with the analyti al solution:

2 a x(I) + b 1 2 w(d, h)d3
(I)
errrel (d, h, x ) =
2 a x(I) + b
= 1 2 w(d, h)d3
(2.123)
To remove the error, the following ondition holds:

1 2 w(d, h)d3 0 w(d, h) d3 = 0.5
(2.124)
the
It is easy to show that there does not exist any pair of (d, h) that fullls
above ondition. However, the best approximation is obtained for d/h = 2.
The relative error is 35%. In reasing the dilation parameter signi antly improves the results. With 28 neighoring parti les, the error an be de reased to
0.2%. However, we omit an analyti al onsideration for that ase.
Irregular parti le arrangement

8 Note
that w in eq. (2.121) depends on d and h
46
Figure 2.12: Regular and irregular parti le arrangements around parti le 407
47
Figure 2.13: Relative error of the gradient of the quadrati fun tion F (X, Y ) =
X 2 + Y 2 for the SPH method for regular and irregular parti le arrangements
(R = 1.6)
To realize an irregular parti le arrangement, the boundary parti les are kept
on their original lo ations. The parti les inside the domain were relo ated randomly with a deviation of 5% and 10% from their original position. The out ome
for one parti le, that has the number 407, is shown in gure 2.12-the parti les
within its domain of inuen e are illustrated. For the 10% deviation, an additional parti le (number 333) is within the domain of inuen e of parti le 407.
We will now onsider again the quadrati fun tion of se tion 2.5.6 and look at
the error in the approximation of it's derivatives.
The error of the usual SPH approximation is shown in gure 2.13 for two
layers of parti les, from parti le number 401 to 425 (rst layer) and 426 to 450
(se ond layer). Re all that we used a 25 25 = 625 parti le dis retization that
is numbered from 1 to 625. We note that the error is in reased signi antly
with the irregularity of the parti le dis retization. One reason might be the insu ient assumption for the quadrature weights VJ = d2 . The error introdu ed
by this assumption an be maximal 21% sin e Vnew,J = (1.1d)2 = 1.21 Vold,J .
However, sin e the approximation is based on a summation where the dierent
values an have dierent signs (due to the value of the kernel fun tion), it is di ult to study the error of the approximation of the partial derivatives in general.
Therefore, let us onsider one spe i parti le, number 407 (gure 2.12), that
has a large relative error of around 70% (gure 2.13). Under the assumption of
equivalent parti le volumes we an reformulate the derivative approximation by
X
0 u(X(407) )
(407)
(X) uJ
0 WJ
VI
JS
(2.125)
48
Figure 2.14: Volume-free relative error and parti le deviations (deviation is 10%)
from their original position
The "volume-free" relative error is then al ulated by
(407)
(X) uJ
0 WJ
P
(407)
0 WK (X) uK
(2.126)
KS
(407)
(X)
J
In gure 2.14, all parti les with the relative error
uJ 6= 0 are shown
X
for the volume-free relative error for the ase of a 10% parti le lo ation deviation
from the original onguration. The deviation from the original parti le position
is illustrated as well. As an be seen from that gure, the parti le position
deviation plays a minor role. The largest error o urs for parti le number 432
with a relatively small deviation. Hen e, the kernel approximation itself is
mainly responsible for the ina urate approximation. Note that the relative
position of the neighboring parti les with respe t to the entral parti le, the
parti le of interest, is ru ial in the kernel approximation.
The error of the symmetrized SPH approximation is shown in gure 2.15
for the 5% and 10% deviation. We note that the symmetrization drasti ally
de reases the relative error. The relative error of the MLS approximation is
by far smaller ompared to the SPH approximation for an irregular parti le
arrangement, see gure 2.16.
49
X 2 + Y 2 for the symmetrized SPH method for regular and irregular parti le
arrangements (R = 1.6)
X 2 + Y 2 for the MLS method for an irregular parti le arrangements (R = 0.6)
50
2.6 Governing equations

2.6.1 Stati s
In stati s, the governing equation is the equation of equilibrium given in the
total Lagrangian des ription for small strains by
(2.127)
0 P b = X 0
where P is the nominal stress tensor (note that P is the transpose of the rst
Piola Kir hho (1PK) tensor), b are the body for es, X are the material oordinates, 0 is the gradient operator with respe t to the material oordinates
and 0 is the domain of the body in the initial onguration. The boundary
onditions are
(X, t) on u0
u(X, t) = u
(2.128)
n0 P(X, t) = t0 (X, t)
on
t0
(2.129)
and t0 are T
the pres ribed displa ements and tra tions, respe tively and
where
S u
u0 t0 = 0 , (u0 t0 ) = .
2.6.2 Dynami s
Negle ting thermome hani al and fri tional for es, the onservation equations
in the total Lagrangian formulation are given by:
J = 0 J0
=
u
(2.130)
1
0 P + b on 0
0
(2.131)
1
F : PT
0
(2.132)
e =
where J and J0 are the Ja obian determinant and initial Ja obian determinant,
respe tively, u is the displa ement ve tor, 0 is the initial density, P are the
nominal stresses, b are the body for es per mass unit, e is the internal energy,
F = u+I denotes the deformation gradient where I is the se ond order identity
tensor and a superimposed dot denotes material time derivatives. Note that the
mass onservation equation is written in algebrai form sin e it is integrable for
a Lagrangian des ription.
The boundary onditions are:
u(X, t)
n0 P(X, t)
(X, t) on u0
u
= t0 (X, t) on t0
=
(2.133)
(2.134)
and t0 are the pres ribed displa ement and tra tion, respe tively, n0
where u
is the outward normal to the domain and u0 t0 = 0 , (u0 t0 ) = .
51
2.7. CONSTRUCTION OF SPECIFIC MESHFREE METHODS
2.6.3 From the strong form to the weak form

Let us onsider at least C0 test and trial fun tions:
X
uh (X) =
J (X) uJ
(2.135)
JS
uh (X)
(2.136)
J (X) uJ
JS
The test and trial fun tions belong to the following spa es:

(t) on u0 ,
V =
u(, t)|u(, t) H1 , u(, t) = u

V0 =
u|u H1 , u = 0 on u0 ,
(2.137)
The weak from of the linear momentum equation is obtained by multiplying

(2.131) with test fun tions:
Z
Z
) u d0 = 0
(2.138)
0 (b u
0 P u d0 +
0
The rst term on the RHS of the momentum equation an be transformed by

integration by parts
Z
Z
Z
T
(0 u) : P d0 (2.139)
0 (P u) d0
0 P u d0 =
0
The Gauss theorem applied on the rst term of the right hand side of (2.139)
gives:
Z
Z
(2.140)
n0 P u d0
0 (P u) d0 =
t0
With the relation between stress ve tor and stress tensor t = n0 P we obtain
Z
Z
(2.141)
0 (P u) d0 =
t u d0
t0
so that we nally obtain

Z
Z
(0 u)T : P d0
0
0 b u d0 +
+
t0
t0 u d0
d0 = 0
0 u u
(2.142)
2.7 Constru tion of spe i meshfree methods

Let us onsider the weak form of the linear momentum equation9 , eq. (2.142).
Substituting the test and trial fun tions into the linear momentum equation
9 The
equilibrium equation ould be onsidered as well, for example
52
leads nally to the system of equations that we need to solve10 :

X
J = fIext fIint ,
mIJ u
(2.143)
where fIext and fIint are the external and the internal for es, given by
Z
Z
I t d0
0 I b d0 +
fIext =
t0
fIint
and
mIJ =
XZ
IS
0 I P d0
0 I (X) J (X) d0 .
(2.144)
(2.145)
(2.146)
The above mass matrix is the onsistent mass matrix. To obtain the nal dis rete equations, the integrals above have to be evaluated by numeri al integration, that will be dis ussed in the following se tions. Depending on the hoi e
of the shape fun tion for the test and trial fun tion, a large variety of meshfree
methods an be onstru ted. For example, if we hoose J (X) = (X XI ),
then we will obtain a ollo ation method that is losely related to a Galerkin
method based on nodal integration. Dieren es o ur only for the boundary
integrals. If we hoose J (X) = J (X), then we obtain a Bubnov Galerkin
method. If J (X) 6= J (X), we will obtain a Petrov Galkerin method.
2.8 Spatial integration

2.8.1 Nodal integration
An e ient dis retization in a Galerkin method is by nodal integration:
Z
X
(2.147)
f (X) d0 =
f (XJ ) VJ0
0
JS
where the weights VJ0 represent tributary volumes asso iated with parti le J .
Note that we used nodal integration previously to obtain the dis rete SPH and
RKPM approximations in se tion 2.5.1 and 2.5.2. Applying this te hnique e.g.
to the omputation of the internal for es (2.145) gives:
X
fIint =
(2.148)
VJ0 0 I (XJ ) PJ
JS
The nodal masses mI are usually obtained by a Voronoi pro edure, i.e. after
pla ement of the nodes, a triangulation is performed and the interse tions of the
10 For
more details, see e.g. Belyts hko et al. [1
53
2.8. SPATIAL INTEGRATION
perpendi ular bise tors of the side on the triangles form the Voronoi ell11 , see
also gure 2.19. The parti le masses mI are then omputed just by multiplying
the volume of the Voronoi ell VJ0 with the density. Note that in general the
physi al mass mI varies from the numeri al mass mIJ . However, it an be
shown that the physi al mass is re overed by diagonalizing the onsistent mass
matrix mIJ with a standard row-sum te hnique:
mI
XZ
mIJ =
JS
JS
0 I (X) J (X) d0
0
0 I (X)
J (X)
JS
(2.149)
d0
With an at least zero-order omplete approximation, eq. (2.149) nally be omes

mI =
0 I (X) d0
(2.150)
J I (X) VJ0
(2.151)
After nodal integration we obtain

mI =
JS
When the test fun tions are also at least zero-order omplete, it is easy to verify
that after adding up all lumped masses, the total physi al mass M of the body
is retained:
N
tot
X
mI =
I=1
N
tot
X
I=1 JS
J I (X) VJ0 =
JS
N
tot
X
I=1
I (X) VJ0 =
JS
J VJ0 = M
(2.152)
Cru ial is also how to onsider and dis retize the 'real' geometri volume of a
body. It an be dis retized with parti les arranged as in gure 2.17a where
the parti les are inside the volume. No parti les lie on the boundaries of the
body. The body an also be dis retized as illustrated on gure 2.17b where
parti les are pla ed exa tly on the boundaries. For gure 2.17a, all masses are
equal. For the dis retization in gure 2.17b, the boundary parti les have only
half of the mass of the interior parti les for a Voronoi method. The masses
of the parti les at the orners are only a quarter of the masses of the interior
parti les. The boundary integrals for natural boundary onditions (external
for es) (2.144) dier for the two models as shown in gure 2.17. When the
parti les are lo ated dire tly on the boundary, the value of the shape fun tion
diers from when the parti les are at a ertain distan e from the boundaries
(see gure 2.17).
11 Note
that the volume of the Voronoi ell is the quadrature weight
54
a)
b)
Figure 2.17: Value of the shape fun tion for boundary parti les lying dire tly
on the boundaries or inside the boundaries
2.8.2 Stabilized nodal integration

Nodal integration often leads to instabilities due to rank de ien y and to low
onvergen e rates. Chen et al. [11 showed that the vanishing derivatives of
the meshfree shape fun tions at the parti les are the ause of the instabilities.
They noti e that linear omplete meshfree methods have to satisfy the following
integration onstraint
Z
Z
(2.153)
n0 I (X) d0
I (X) d0 =
0
This integration onstraint omes from the equilibrium of the internal and external for es of the linear omplete Galerkin approximation and is similar to the
linear ompleteness in the onstant stress pat h test in nite elements. Chen
et al. [11 propose a stabilized onforming nodal integration using strain smoothing. In the strain smoothing pro edure, the nodal strains are omputed as the
divergen e of a spatial average of the strain eld. The strain smoothing avoids
evaluating derivatives of the shape fun tions at the nodes and hen e eliminates
defe tive modes. The smooth strain eld at a material point XM an be
expressed as
Z
(X XM ) d0
(XM ) =
(2.154)
0
where is the strain from the ompatibility ondition ( = 0.5(ui,j + uj,i )), 0
is the domain of the Voronoi ell and (X XM ) is the smoothing fun tion
that has to fulll the following requirements:
Z
(X XM ) 0
(X XM ) d0 = 1
(2.155)
55
particle
particle
particle
particle
stress point
stress point
stress point
particle
particle
particle
particle
Figure 2.18: Arrangement of the stresspoints in two dimensions

Chen et al. [11 hose
(X XM ) =
1
AM
XM 0 , otherwise (X XM ) = 0
(2.156)
where AM is the area of the smoothing (=Voronoi) ell. Substituting eq. (2.156)
into (2.154), we obtain
Z
1
(XM ) =
(ui,j + uj,i ) d0
2AM 0
Z
1
(2.157)
(ui nj + uj ni ) d0
=
2AM 0
As an be seen, integration is performed along a surfa e (the surfa e of the

Voronoi ell) instead of a volume. It should be noted that the smoothed strain
eld does not satisfy the ompatibility relations with the displa ement eld at all
points in the dis retized domain. However, the smooth, non-lo al strain and the
lo al strain an be onsidered as two independent elds- the non-lo al strain eld
an be viewed as an assumed strain eld. Thus, a two-eld variational prin iple
is suitable for this approximation.
2.8.3 Stress-point integration

Stress point integration was proposed by Dyka and Ingel [12 in one dimension
to stabilize the SPH method. Randles and Libersky [13 extended stress point
integration to higher dimensions to stabilize the normalized form of SPH. Stress
point integration eliminates instabilities due to rank de ien y.
In stress point integration methods, stress points are interspersed between the
parti les and the ontributions of the stresses are added to the integration of
(2.147):
Z
X
X
(2.158)
f (X) d0 =
fJP VJ0P +
fJS VJ0S
0
JNP
JNS
56
master particles
slave particles
Figure 2.19: S heme of stress point integration

where NP and NS indi ate the supporting original parti les and stress point
nodes, respe tively, to the original parti le XP
I . There are dierent ways for
arranging the stress points between the original parti les. Some of them are
shown in gure 2.18. Note that all kinemati values su h as displa ements and
velo ities are obtained via interpolation from the original parti les:
uSI =
J (XSI ) uP
J,
vIS =
JS
J (XSI ) vJP
(2.159)
JS
where the supers ripts S indi ate stress points and the P the original parti les.
J (XSI ) is the shape fun tion of the supporting master node J at XSI . The
internal for es for examples are al ulated by
X
X
P
VJ0S 0 I (XSJ ) PSJ
VJ0P 0 I (XP
fIint =
(2.160)
J ) PJ +
JNS
JNP
The volumes VJ0P and VJ0S are omputed from the Voronoi diagramm (see gure
2.19) so that their sums results in the total geometri initial volume:
V0 =
JNP
VJ0P +
VJ0S
(2.161)
JNS
2.8.4 Gauss integration

There are two ways to onstru t a ba kground mesh for the Gauss integration.
The ba kground mesh an be aligned with the parti les as shown in gure 2.20a
or the ba kground mesh is not aligned with the parti les as illustrated in gure
2.20b. In ea h integration ell, Gauss quadrature is performed. Note that in
Gauss quadrature a polynomial of degree 2nQ 1 an be exa tly integrated with
57
nQ quadrature points12 . However, exa t integration is di ult to perform in

meshfree methods sin e meshfree shape fun tions are often not polynomial. It
is ommon to in rease the number of quadrature points to redu e integration
errors. The number nQ of Gauss points depends on the number of nodes in
a ell. The number of Gauss points an be given by numeri al experiments
a ording to
(2.162)
nQ = m + 2
where m denotes the number of nodes in one ell.

The integral of a fun tion is given by
Z
f (X) d0 =
Z+1Z+1
m
X
wJ f (J ) det J (J )
f (, ) det J (, ) dd =
J=1
1 1
(2.163)
where = (, ) are s aled lo al oordinates, m is the total number of quadrature points, wJ = w(J ) w(J ) are the quadrature weights whi h is the produ t
of the weight at the orresponding Gauss point in - and -dire tion and det J
is the Ja obian determinant given by
X
J =
(2.164)
The internal for es are then

fint =
m
X
J=1
wJ detJ ( J ) 0 (X( J ) XP ) P( J )
(2.165)
where the supers ript P indi ates the parti le position.
Example: The EFG-method for elastostati s

Let us onsider a Bubnov-Galerkin method in a stati setting with EFG-shape
fun tions. Furthermore, we assume linear elasti material behavior, the hypothesis of small strains and no geometri non-linearities. Then we an write the
equilibrium equation in variational formulation: nd u V su h that
W = Wint Wext = 0 u H1
where
Wint =
( u) : P d0
(2.166)
(2.167)
12 Example: For a 2D quadrilateral nite element with bilinear shape fun tions, we have a
maximum of quadrati terms for the shape fun tions plus linear terms for the Ja obian, i.e.
maximum ubi terms; so we need nQ = 2 quadrature points. An 8-node quadrilateral will
require nQ = 3 quadrature points sin e the highest polynomial is of order 5 (order 3 for the
shape fun tion and order 2 for the Ja obian)
58
b)
a)
Figure 2.20: Dierent ways of ell integration
Wext =
where
V =
V0
0 u b d0 +
t0
u t0 d0

(t) on u0 ,
u(, t)|u(, t) H1 , u(, t) = u

u|u H1 , u = 0 on u0 ,
(2.168)
(2.169)
After substituting the test and trial fun tions into eq. (2.166), the system of
equations reads:
K u = f ext
(2.170)
where K is the stiness matrix
KIJ =
BI Ct BJ d0
(2.171)
with the following B-matrix (in two dimensions)
I,X
0
BI = 0
I,Y
I,Y I,X
The external and body for es are given by
Z
Z
I (X) t0 d0 +
fIext =
t0
I (X) b d0
(2.172)
(2.173)
Note that the test fun tions do not vanish at the boundaries and hen e do not
allow for H1 to be represented exa tly. Hen e, spe ial te hniques need to be
introdu ed to handle essential boundary onditions.
59
2.9. ESSENTIAL BOUNDARY CONDITIONS
2.9 Essential boundary onditions

2.9.1 Lagrange multiplier method
Essential boundary onditions an be imposed by means of Lagrange multiplier
method. Therefore, we modify the variational formulation given in the last
se tion: nd u V su h that
W = Wint Wext Wu = 0 u H1
where
Wint =
Wext =
where
V =
V0
0 \c0
0 \c0
( u) : P d0
0 u b d0 +
t0
u t0 d0

(t) on u0 ,
u(, t)|u(, t) H1 , u(, t) = u

u|u H1 , u = 0 on u0 ,
(2.174)
(2.175)
(2.176)
(2.177)
The term Wu in eq. (2.174) is introdu ed for the imposition of Diri hlet boundary onditions. With the Lagrange multiplier method, Wu reads:
Z
Z
) d0 +
u d0
(u u
Wu =
(2.178)
0u
where are the Lagrange multipliers that an be onsidered as tra tions a ross
the boundaries. As for the approximation of the test and trial fun tions, there
are several ways to dis retize the Lagrange multiplier eld:
X
=
(2.179)
L
J (X) J
JS
where L
J (X) are the approximation fun tions for the Lagrange multipliers that
an be nite element shape fun tions, meshfree shape fun tions or Dira Delta
fun tions. The resulting dis rete form reads then

ext
K G
u
f
=
(2.180)
G 0
q
where K = KIJ is the stiness matrix given above,
Z
I (X) L
GIK =
K (X) S d
(2.181)
and
qK =
d
L
K (X) S u
(2.182)
where S is a 2 2 matrix with Sij equal to 1 for j = i and 0 otherwise. The

disadvantages of using Lagrange multipliers are:
60
The dimension of the resulting system of equations is in reased.

The global matrix is no longer positive denite even if K is symmetri and
semi-positive denite.
The nal dis rete system and the weak form problem indu e a saddle point
problem whi h pre ludes an arbitrary hoi e of the interpolation/approximation spa e for u and . The resolution of the multiplier eld must be
ne enough in order to obtain an a eptable solution, but the system of
equations will be singular if the resolution of Lagrange multipliers eld
is too ne. In fa t, the interpolation spa es for the Lagrange multiplier
and the prin ipal unknown u must verify an inf sup ondtion, known
as Babuska-Brezzi stability ondition, in order to ensure the onvergen e
of the approximation. While in nite elements, it is trivial to hose the
approximation for the Lagrange multiplier to verify the Babuska-Brezzi
ondition, this hoi e is not trivial in meshfree methods.
Example
For simpli ity, let us onsider the following problem: nd the minimum of the
fun tion with onstrained onditions:
with onstraint
= 2a21 2a1 a2 + a22 + 18a1 + 6a2
(2.183)
a1 = a2
(2.184)
This problem is solved by introdu ing a Lagrange multiplier as new unknown

= + (a1 a2 )
= 2a21 2a1 a2 + a22 + 18a1 + 6a2 + (a1 a2 )
(2.185)
Dierentiation with respe t to three unknowns ai and gives:
=0
=0
=0
a1
a2
The solution of the above equations gives us values for a1 , a2 and .
(2.186)
a1 = a2 = 12 = 6
2.9.2 Penalty method

In the penalty method, a dierent Wu is used as ompared to the previous
se tion:
Z
k2 d0
Wu = 0.5p
(2.187)
ku u
0u
where p is the penalty parameter that has to be spe ied by the user. A high
penalty number will lead to a more a urate solution sin e in the penalty method
61
2.9. ESSENTIAL BOUNDARY CONDITIONS
the imposed boundary ondition is not exa t. However, a high penalty number also leads to an ill onditioned global stiness matrix. The advantage of
the penalty method is that the dimension of the resulting system of equations
remains un hanged.
Another alternative is the augmented Lagrange method that ombines the
Lagrange multiplier and penalty method and will be dis ussed in se tion 2.10.3.
2.9.3 Transformation method

Let us onsider the approximation:
uh (x) =
N
X
(2.188)
I (xJ )
uI
I=1
where uh (x) is a tually displa ement at parti le xJ whereas uI are titious

values sin e the meshfree shape fun tions do not satisfy the Krone ker delta
property. With N parti les in the entire domain, we have N equations that an
be given in matrix form by
(2.189)
u = D
u
Noting that the size of matrix D is N N . Sin e the essential boundary onditions are imposed over the true displa ement not over the titious parameters
, it is natural to eliminate them by the relation u
= D1 u. Hen e, we have
u
the transformed approximation written in terms of true nodal displa ement
uh (x) =
n
X
(2.190)
I (x)D1
IJ uI
I=1
Now, the essential boundary onditions an be applied dire tly. However, this
method requires the inversion of a matrix of size N N whi h is omputationally
too expensive. Hen e, let us onsider separately the boundary nodes on u from
the interior nodes. The number of interior parti les is denoted by N and the
number of parti les on the essential boundary u as Nu . We an write the
approximation as follows
uh (x) =
N
X
I (xJ )uI +
I=1
N
u
X
I (xJ )uI
(2.191)
I=1
Along u , the pres ribed displa ement is imposed by

u(xJ ) = g(xJ ),
(2.192)
J = 1, ..., Nu
whi h is rewritten in matrix form

}
|D {zu
(Nu N )(N 1)
D
}
| {zu
(Nu Nu )(Nu 1)
g
|{z}
(Nu 1)
(2.193)
62

u
Eliminating u
from the above equation, we have
h u i1
u
u
= D
(g D u
)
(2.194)
and get the transformed meshless approximation

uh (x) =
N
X
I (xJ )uI +
I=1
N
u
X
I=1

u
I (xJ ) [DIJ ]1 (gI D

u
)
IJ J
After rearranging, the nal form of the new approximation reads

N
N
u
X
X
u
h
u 1
u (x) =
I (x) I (x)[DIJ ] DIJ uI +
I (x)[DIJ ]1 gJ
i=I
(2.195)
(2.196)
I=1
2.10 Coupling to nite elements

2.10.1 Coupling via ramp fun tions
particle domain
particle boundary
blending region
element boundary
F E
PSfrag repla ements

element domain
F E
FE node
particle
Figure 2.21: Coupling of Finite Elements and Parti les via ramp fun tions
Consider a domain of problem with a hybrid dis retization between nite
elements and parti les as illustrated in gure 2.21. The transition region is
designates by B , P denotes the parti le domain and F E is the element
domain. The element and parti le boundary is F E and P , respe tively. In
the interfa e region, an approximation is given by:

uh = uF E (X) + R(X) uP (X) uF E (X) , X B
(2.197)
2.10. COUPLING TO FINITE ELEMENTS
63
where uF E and uP are the nite element and parti le approximations for u in
the transition region and R(X) is a ramp fun tion, so that R(X) = 1, X P
and R(X) = 0, X F E . It is onstru ted with the use of a linear ramp fun tion
along the interfa e element boundaries so that ontinuity is ensured:
with
R(X) = 3 r2 (X) 2 r3 (X)
(2.198)
(2.199)
r(X) =
NJ (X)
JSP
where SP is the set of nodes on P . Substituting the FE approximations and

the meshfree approximation into eq. (2.199) the approximation in the transition
region is obtained:
X
I (X)uI , XI B
uh (X) =
(2.200)
N
I
with the interfa e shape fun tion
I (X) = (1 R(X)) NI ((X)) + R(X) NI (X) X B

N
(2.201)
I (X) = R(X) NI (X) X

N
/ B
(2.202)
I (X) = NI (X) X B on F E
N
(2.203)
I (X) = 0 X
N
/ B on F E
(2.204)
I (X) = N (X) X
N
/ B on P
(2.205)
Linear ompleteness is preserved in the entire domain. If the integrals for the
parti les are evaluated by a nodal integration with stress points, the shape
fun tions in the blending domain have only to be evaluated at the parti le
boundary P and element boundary F E and are redu ed to:
sin e R(X) = 1 on and R(X) = 0 on . The approximation of the test

fun tions in the blending region have the same stru ture.
P
FE
2.10.2 Coupling with Lagrange multipliers

The parti le and nite element domain an also be oupled by use of Lagrange
multipliers. For the stati ase, the potential to be minimized is
W = W int W ext + T g
(2.206)
where W int is the internal and W ext is the external energy. The last term on
the RHS are the onstraints. The Lagrange multipliers are denoted by and
64
PSfrag repla ements
0F E
0
0P
0F E
0P
Figure 2.22: Coupling of parti le and nite elements

g = uF E uP is the gap of the parti le and the nite element domain along
the ommon boundary as illustrated in gure 2.22. The Lagrange multipliers
are lo ated at the parti le positions and are
N
X
gh =
J=1
E
NJF E (X, t) uF
NJP (X, t) uP
J
(2.207)
JS
The Lagrange multiplier estimates are pla ed at the parti le position and nite
element shape fun tions are used to dis retize the Lagrange multiplier eld :
P
h (X, t) =
N
X
NJF E (X, t) J (t)
(2.208)
J=1
Note that for the interpolation in eq. (2.208), the position of the Lagrange
multipliers in the lo al element oordinate system13 has to be known. If the
global position XL of the Lagrange multiplier is known then the lo al position
an simply be obtained by solving XL = I ()XI with respe t to the lo al
oordinates . The test and trial fun tions are
uh (X, t) =
N
X
E
NJF E (X, t) uF
J (t) +
J=1
uh (X, t) =
N
X
J=1
and
NJP (X, t) uP
J (t)
(2.209)
NJP (X, t) uP
J (t)
(2.210)
JS
E
NJF E (X, t) uF
J (t) +
JS
N F E (X, t) = 0 X P
0
E
N P (X, t) = 0 X F
0
(2.211)
where S is the set of nodes in the parti le model. Minimizing eq. (2.206) with
respe t to u and leads to the following equations:
W
W int
W ext
g
g
=
+
= f int f ext +
=0
u
u
u
u
u
W
=g=0
13 on the boundary
(2.212)
65
The derivatives of W int and W ext with respe t to u are the internal and external
for es, respe tively:
Z
(0 u)T : P d0
f int =
(2.213)
FE
P
0 0
ext
FE
P
0 0
u b d0 +
F E,t
P,t
0 0
u t0 d0
(2.214)
g
are linear ombinations of the Lagrange multipliers.
The additional for es u
To obtain the dis rete system of nonlinear equations we will do a linearization.
Therefore, we take a Taylor series expansion of eq. (2.212) negle ting any higher
order terms:
0 =
f int f ext +
0 =
u+
f ext
g
2g
g f int
+
u
u +
+
u
u
u
u
u
uu
g
u
u
(2.215)
Substituting the test and trial fun tions into (2.215) we nally obtain the following system of equations:
T
2 g
KF E + uu
0
KF EF E
T
2g
KF EP
0
KP + uu

T
KF EF E
KF EP
0
ext,F E
T
f
f int,F E KF EF E
= f ext,P f int,P T KF EP
g
E
uF
J
uP
J
(2.216)
where KF EF E and KF EP denotes the derivatives of g with respe t to u for

the nite element (uF E ) and parti le domain (uP ), respe tively and KF E and
KP are the derivatives of the internal and external for es with respe t to u. If
neither b nor the tra tion boundary onditions t depend on u, then we an give
the matri es:
Z
T
F EF E
K
=
NF E NF E d0
KF EP
KP
=
=
P
0
KF E
NF E
E
F
0
BP
T
BF E
T
NP d0
C BP d0
T
C BF E d0
(2.217)
66
and the ve tors for internal and external for es

Z
Z
T
f ext,F E =
NF E
b d0 +
E
F
0
f ext,P
NP
P
0
f int,F E
T
BF E
E
F
0
int,P
P
0
BP
T
b d0 +
T
P,t
0
P d0
NF E
E,t
F
0
NP
T
T
t0 d0
t0 d0
(2.218)
P d0
For a linear problem in elastostati s, we have a linear stiness matrix K, the

2g
starting value is zero, i.e. u = u and = and the term uu
vanishes
for linear onstraint. Hen e we obtain the well-known dis rete equations for a
linear problem in elastostati s:
T
FE
KF E
0
KF EF E
uJ
uP
0
KP
KF EP
J

T
KF EF E
KF EP
0
ext,F E
f
= f ext,P
(2.219)
g
2.10.3 Bridging domain oupling method
We denote the omplete domain in the initial onguration by 0 and its boundaries by 0 ; 0 onsists of tra tion boundaries t0 and the displa ement boundaries u0 . The domains are subdivided into the subdomains treated by nite
E
P
element methods, F
0 , and that treated by parti le methods, 0 ; the latter
is the domain en ompassed by the parti les of the model. The interse tion of
these two domains is denoted by int
in the initial onguration, int in the
0
int
urrent onguration; is often alled the overlapping subdomain (or bridging domain);
0 denotes the edge of the nite element domain; an example of
a model is shown in gure 2.23.
In expressing the total internal potential energy of the system, we employ a
s aling parameter in the overlapping subdomain. The parameter is dened
as = l(X)
l0 where l(X) is the least square of the proje tion of X onto 0 as
shown in gure 2.23. The s aling parameter is unity at the edge of the nite
element domain and vanishes at the other edge of int
0 ; it is important that
int
in ludes
the
last
line
of
parti les.
In
the
absen e
of
heat transmission, the
0
onservation of energy of governing equations in the entire domain is:
Z
Z
E
P FT PdP
F E FT PdF
+
W int =
(2.220)
0
0
E
F
0
P
0
67
FE
0
int
finite element node

particle
=1
=0
Figure 2.23: Finite element model oupled with parti le method

where the s aling multiplier eld
F E (X) =
P (X) =
is dened as
0
in P
0
1 in int
0
E
1
in F
int
0
0
E
0 in F
0
int
in 0
int
1 in P
0 0
(2.221)
(2.222)
The external energy is:

W ext
E
F
0
E
F
0
E
F E 0 b udF
+
0
E
F E t udF
+
0
P
0
P
0
P 0 b udP
0
P t udP
0
(2.223)
the displa ements an be approximated in the terms of nite element

In int
0
shape fun tions NI (X) or meshfree fun tions wI (X), respe tively, by:
X
E
uF E (X, t) =
(2.224)
NI (X)uF
I (t)
I
u (X, t) =
wI (X)uP
I (t)
(2.225)
Therefore, the onstraints ondition in int

at the dis rete position of parti les
0
are
(
)
X
X

FE
P
FE
P
gI = {giI } = uiI uiI =
(2.226)
NJI uiJ
wKI uiK
J
68
The Lagrange multiplier eld is also expressed in terms of shape fun tions denoted by I (X):
X
i (X, t) =
(2.227)
I (X)iI (t)
I
Generally, the shape fun tions for the Lagrange multiplier eld I (X) will dier
from that for the displa ement, NI (X) or wI (X), and they must satisfy the
Babuska Brezzi onditions. The Lagrange multiplier eld is usually represented
by inserting nite elements in the interse tion domain and the nite element
approximation is applied on the parti les of the overlapping subdomain. To
distinguish the Lagrange multiplier eld i in eq. (2.227), iI is denoted as the
unknown Lagrange multiplier at the Lagrange multiplier nodes.
Dis rete Equations
The energy fun tion for the augmented Lagrangian method is

1
WAL = W int W ext + T g + pgT g
2
(2.228)
where p is the penalty parameter. If p = 0, eq. (2.228) will be identi al to

the expression of the energy fun tion for the Lagrange multiplier method. The
dis rete equations are then obtained by setting the derivatives of WAL with
respe t to uI and I to zero. This gives
#
!
"
X
X
WAL
int
ext
= (FiI FiI ) +
KL K NIL
E
uF
iI
L
K
"
#
!
X
X
X
E
+ p
NKL uF
wKL uP
NIL = 0 (2.229)
iK
iK
L
WAL
uP
iI
int
(fiI
WAL
iI
"
IL
ext
fiI
)
X
K
X
L
"
E
NKL uF
iK
"
X
E
NKL uF
iK
KL K
wKL uP
iK
wKL uP
iK
where
NKI = NK (XI )
wIL
KI = K (XI )
wIL = 0 (2.230)
=0
(2.231)
(2.232)
Fint and Fext are internal and external for e in the nite element subdomain
E
F
and they are expressed as:
0
Z
E
F E NI,j (X)Pji (X)dF
Fint
=
(2.233)
0
iI
E
F
0
Fext
iI
E
F
0
E
F E NI (X)0 bi dF
+
0
t0
F E NI (X)ti dt0 (2.234)
69
and f int and f ext are internal and external for e in the parti le subdomain P
0
given by
int
fiI
ext
fiI
P
0
(2.235)
P wI,j (X)Pji (X)dP

0
P
0
wI (X)0 bi dP
0
t0
(2.236)
P wI (X)ti dt0
Letting d denote the array of u, the in rements in the internal nodal for e an
be approximated in terms of in rements in the nodal displa ement by stiness
matri es:
Fint
I
E
FE
KF
IJ uJ
(2.237)
or Fint = KF E dF E
fIint
P
KP
IJ uJ
(2.238)
or f int = KP dP
where KF E and KP are tangent stiness matri es given by
KF E
KP
FE
FE
d1
E
dF
2
=
..
FE
dn
E
KF
11
E
KF
21
=
KP
11
KP
21
=
E
dF
I
E
KF
12
FE
K22
KP
12
KP
22
A11
A21
LF E
A12
A22
LP
..
.
E
KF
nn
KP
mm
E
uF
xI
FE
uyI
The system an then be written as
..
E
KF
IJ =
KP
IJ =
P
d1
dP
2
dP =
..
P
dm
Fint
I
E
uF
J
fIint
uP
J
dP
I

T
LF E
dF E rF E
T
dP
rP
=
LP
(2.239)
(2.240)
uP
xI
uP
yI
(2.241)
(2.242)
If we let di denote ukP and dj denote ulQ , the ingredients of eq. (2.242) an be
70
expressed as:
rF E
rP
=
=
Fint Fext + T GF E + pgT GF E

f int f ext + T GP + pgT GP
(
)
X
{giI } =
IK giK
(2.243)
(2.244)
(2.245)
A11
A12
KF E + pGF E GF E
FET
pG
PT
pG
A22
iI
A21
(2.246)
(2.247)
(2.248)
FE
PT
(2.249)
K + pG G
X
K (XI )iK
(2.250)
FE
KP
FE
=
=
LP
=
=
#
" int # "Z
FiI
Fint
FE
FE
=
=
NI,j C jilk NQ,k d0 (2.251)
E
E
dF E
uF
F
lQ
0
#
int " int # "Z
f
fiI
P
P
=
wI,j C jilk wQ,k d0
(2.252)
=
dP
uP
P
lQ
0

"
X
L
gL
IL F E
di
"
X
gjL
IL F E
di
#
gL
IL NP I jk
(2.253)
IL F E =
ukP
L
L
# "
# "
#
"
X
X
X
gjL
gL
gL
IL P =
IL P
IL P =
di
di
ukP
L
L
L
#
"
X
IL wP I jk
(2.254)
"
"
X
FE
GP

gjI
gI
=
= [NP I jk ]
E
E
dF
uF
i
kP

gjI
gI
=
= [wP I jk ]
dP
uP
i
kP
(2.255)
Chapter 3
Kinemati s of strong and

weak dis ontinuities
A jump in the displa ement eld is referred to as strong dis ontinuity. A typi al
strong dis ontinuity is a ra k. Let us onsider an innite bar with a ra k in
one dimension, gure 3.1. A linear displa ement eld adja ent to the ra k is
assumed. Hen e, the strain eld will be onstant.
Let us onsider now the two dimensional ase as shown in gure 3.2a. The
strong dis ontinuity is denoted by c and shown in the initial onguration. The
dis ontinuity in the displa ement is often onsidered in the initial onguration
where the two opposite ra k surfa es oin ide in one line. In the urrent onguration, the two ra k surfa es won't oin ide. The displa ement eld of the
one side of the ra k an be ompletely dierent as the displa ement eld on
the opposite side of the ra k, even in slope and orientation. The kinemati s of
the strong dis ontinuity is given by
]](x, t)
+ HS [[u
u(x,
t) = u
(3.1)
where u is the displa ement eld, the superimposed dot denotes time derivative,
spatial derivatives with respe t to spatial oordinates x, t is the time, r and
s are lo al oordinates, HS is the Heaviside fun tion a ting on the dis ontinuity interfa e and S is the Dira delta fun tion also a ting on S . With the
ompatibility ondition and eq. (3.1), the strain rate eld an be expressed by

S
+ HS S [[u
]] + S [[u
]] n
(x,
t) = S u = S u
(3.2)
]](s, t)
+ Hh (r, t)[[u
u(x,
t) = u
(3.3)
A weak dis ontinuity is a jump in the strain eld as shown in gure 3.2b. It
o urs e.g. at material interfa es. It is not well suited to des ribe the kinemati s
of a ra k. The domain is split by a dis ontinuity interfa e S into + and
. The kinemati s of the weak dis ontinuity an then be des ribed by
71
72CHAPTER 3.
KINEMATICS OF STRONG AND WEAK DISCONTINUITIES
Figure 3.1: Cra king of a one-dimensional bar with orresponding ra k kinemati s
Figure 3.2: Kinemati s for a) weak dis ontinuity, b) strong dis ontinuity
73
where Hh denotes the ramp fun tion dened by
h
0 x \
+
h
1 x \
Hh =
ss
h
s+ s x
(3.4)
Formal dierentiation gives:
]]
]] + Hh [[u
+ Hh S [[u
(x,
t) = S u = S u
(3.5)
where denotes spatial derivatives with respe t to the lo al oordinate s. The

derivatives of Hh with respe t to the lo al oordinate system (in s-dire tion)
are:
n
Hh =
(3.6)
h(r)
where h(r) is the thi kness of the lo alization band and n is the (unit) normal to
the dis ontinuity interfa e; note that h = s+ s is the width of the lo alization
band. Dening the ollo ation fun tion a that is 1 if X h and 0 otherwise,
we obtain:
S
a
]] n
]] +
+ Hh S [[u
(3.7)
[[u
(x, t) = S u = S u
h
Note, that the kinemati s des ribed by eqs. (3.3) and (3.1) are independent of
the dis retization.
74CHAPTER 3.
KINEMATICS OF STRONG AND WEAK DISCONTINUITIES
Chapter 4
Strong dis ontinuities in

meshfree methods
4.1 A histori al overview
4.1.1 The visibility riterion
The visibility method was the rst approa h that introdu ed a strong dis ontinuity into a meshfree dis retization. In the visibility method, the ra k boundary
is onsidered to be opaque. Thus, the displa ement dis ontinuity is modeled by
ex luding the parti les on the opposite side of the ra k in the approximation
of the displa ement eld, see gure 4.1.
Di ulties arise for parti les lose to the ra k tip sin e undesired interior
dis ontinuities o ur (gure 4.2a) sin e the shape fun tion is ut abruptly, see
gure 4.1. Nevertheless, Krysl and Belyts hko [14 showed onvergen e for the
visibility method. For the pra ti al important ase of linear omplete EFG
shape fun tions, they even showed that the onvergen e rate is not ae ted by
the dis ontinuity.
It should also be noted that the visibility riterion leads to dis ontinuities
in shape fun tions near non- onvex boundaries su h as kinks, ra k edges and
holes, as shown in gure 4.2b in two dimensions.
4.1.2 The dira tion riterion

The dira tion method is an improvement of the visibility method. It removes
the undesired interior dis ontinuities, see gure 4.3 (see also gure 4.1) and
the dira tion method is also suitable for non- onvex ra k boundaries. The
method is motivated by the way light dira ts around a sharp orner but the
equations used in onstru ting the domain of inuen e and the weight fun tion
bear almost no relationship to the equation of dira tion. The method is only
appli able to radial kernel fun tions with a single parameter hI0 .
75
76CHAPTER
4. STRONG DISCONTINUITIES IN MESHFREE METHODS
CRACK
Crack line
Visibility criterion
CRACK
Crack line
Diffraction criterion
CRACK
Crack line
PSfrag repla ements
Transparency criterion
Figure 4.1: Prin iple of the visibilty, dira tion and transparen y method with
orresponding shape fun tions, from Belyts hko et al. [2
77
4.1. A HISTORICAL OVERVIEW
Domain of influence
inter
discontinuities
crack
crack
a)
(b)
Figure 4.2: a) Undesired introdu ed dis ontinuities by the visibility method, b)

Di ulties with the visibility method for on ave boundaries and kinks
xI
s1
PSfrag repla ements
s0 (x)
xc
s2 (x)
crack
crack
Figure 4.3: The dira tion method
78CHAPTER
The idea of the dira tion method is to treat the ra k as opaque but to
evaluate the length of the ray h0 by a path whi h passes around the orner of
the dis ontinuity. This removes the abrupt ut of the shape fun tion to zero
a ross the undesired interior dis ontinuity. A typi al weight fun tion is shown
in gure 4.1. The weight parameter hI0 is omputed by
hI0 (X) =
s1 + s2 (X)
s0 (X)
(4.1)
s0 (X)
where
s0 (X)
s1
s2 (X)
= kX XI k
= kXc XI k
(4.2)
= kX Xc k
The parameter is usually set between one and two and adjusts the distan e of
the support on the opposite side of the ra k. For a better illustration, see also
gure 4.3. It should be noted that the shape fun tion of the dira tion method
is quite omplex with several areas of rapidly varying derivatives that ompli ates quadrature of the dis rete Galerkin form. Moreover, the extension of the
dira tion method into three dimensions is omplex. The orre ted shape fun tions are obtained by substituting the modied dilation parameter hI0 into the
orresponding kernel fun tion. For the use in a Galerkin formulation, the spatial
derivatives of the shape fun tions are needed. This requires the derivatives of
the kernel fun tion:
W h0I
W
=
Xi
h0I Xi
(4.3)
The rst term on the right hand side is un hanged and the se ond term is
h0I
=
Xi
s1 + s2 (X)
s0 (X)
1
s2
+ (1 )
Xi
s1 + s2 (X)
s0 (X)
s0
Xi
(4.4)
with
s2
X Xc
=
X
s2 (X)
s0
X XI
=
X
s0 (X)
(4.5)
4.1.3 The transparen y riterion

The transparen y method was developed as an alternative to the dira tion
method by Organ et al. [15. The transparen y method is easier extendable into
three dimensions than the dira tion method. In the transparen y method, the
ra k is made transparent near the ra k tip. The degree of transparen y is
79
4.1. A HISTORICAL OVERVIEW
PSfrag repla ements
crack
Figure 4.4: The transparen y method
80CHAPTER
related to the distan e from the ra k tip to the point of interse tion, see gure
4.4.
Consider a ray from the evaluation point X to the node XI , gure 4.4. The
dilation parameter h0I is modied as follows when the ray interse ts the ra k:
h0I = s0 (X) + hmI
sc (X)
sc
(4.6)
where s0 (X) is dened in eq. (4.2), hmI is the radius of the nodal support SI
and sc (X) is the distan e from the ra k tip to the interse tion point, gure
4.4. The parameter sc = h sets the interse tion distan e at whi h the ra k
segment is ompletely opaque where is used to vary the opa ity and h is a
measure of nodal spa ing. Note that the additional term in eq. (4.6) is at least
quadrati so that the weight fun tion derivatives will be ontinuous.
An additional requirement is usually imposed for parti les lose to the ra k.
Sin e the angle between the ra k and the ray from the node to the ra k tip is
small, a sharp gradient in the weight fun tion a ross the line ahead of the ra k
is introdu ed. In order to redu e this ee t, Organ et al. [15 imposed that all
nodes have a minimum distan e from the ra k surfa e. The spatial derivatives
are obtained by formal dierentiation:
h0I
s1 sc
s0
=
+ hmI c
X
X
sc X
(4.7)
where we an write
s0
X
sc
X1
sc
X2
X XI
s0 (X)
Xb Xc
sc (X)
Yb Yc
= sin() =
s2 (X)
= cos() =
(4.8)
where is the angle between the ra k and the x-axis. The weight fun tion of
the transparen y method looks similar to the one of the dira tion method and
is illustrated in gure 4.1.
4.1.4 The see-through and ontinuous line riterion

The "see-through" method was proposed by Terry [16 for onstru ting ontinuous approximations near non- onvex boundaries. Therefore, the boundary was
onsidered as ompletely transparent su h that the dis ontinuity is removed.
Though the "see-through" method works well for apturing features su h as
interior holes, it is not well suited to model strong dis ontinuities.
In the ontinuous line method from Krysl and Belyts hko [14 and Duarte
and Oden [9, the ra k is ompletely transparent at the ra k tip. In other
words, parti les whose domain of inuen e are partially ut by the ra k, an see
4.2. THE CONCEPT OF ENRICHMENT
81
na
nB
PSfrag repla ements
Figure 4.5: Domain of inuen e near a wedge-shaped non- onvex boundary.

The boundaryis enfor ed if nA nB
through the ra k. This drasti ally shortens the ra k. If no spe ial te hniques
are introdu ed, the ra k also does not lose at the ra k tip that leads to
ina urate solutions.
4.1.5 Mixed riteria

Belyts hko and Fleming [17 suggested to ombine dierent methods depending
on the onvexity of the ra k boundary e.g. to use the visibility for onvex
boundaries and other methods for non- onvex ra k boundaries. They suggest
a riterion based on the angle of the wedge that an be written in terms of the
surfa e normal, see gure 4.5. When nA nB with = 0o as uto value,
the boundary an be onsidered as onvex, otherwise non- onvex. The uto
value of = 0o orresponds to the wedge angle of = 90o in gure 4.5.
4.2 The on ept of enri hment

There are basi ally two ways of enri hing an approximation: an intrinsi enri hment and an extrinsi enri hment. The prin ipal is basi ally equivalent to
the prin ipal of in reasing the order of ompleteness that an be a hieved also
intrinsi ally or extrinsi ally. However, the enrichment is aimed to in rease the
a ura y of the approximation by in luding information of the analyti al solution. The on ept of enri hment will be outlined for a strong dis ontinuity in
two dimensions, i.e. for a ra k in linear elasti fra ture me hani s (LEFM).
82CHAPTER
crack
Figure 4.6: Denitions for the ra k tip enri hment

Therefore, we will briey des ribe the analyti al solution of the near ra k tip
eld in LEFM in the next se tion. In the following se tion, we will dis uss the
dierent enri hments.
4.2.1 Introdu tion into LEFM

In LEFM, the asymptoti near ra k tip displa ement eld in two dimensions
is given by
r

r
1
KI Q1I () + KII Q1II ()
u1 =
G 2G
r

1
r
(4.9)
u2 =
KI Q2I () + KII Q2II ()
G 2G
where G is the shear modulus, r and are explained in gure 4.6 and
Q1I ()
Q2I ()
Q1II ()
Q2II ()
= cos + sin sin

2
2
= + sin + sin cos

2
2
= + sin + sin cos

2
2
= cos sin sin

2
2
(4.10)
are the angular fun tions for LEFM, KI and KII are the mode-I and mode-II
stress intensity fa tors (SIF) where = (3 )/(1 + ) for plane stress and
= (3 4) for plane strain is the kolosov onstant. Using trigonometri
83
identities, it an be shown that the basis, given by

r sin(/2), r cos(/2), r sin(/2)sin(), r cos(/2)sin()

(4.11)
spans the LEFM ra k-tip displa ement eld in eq. (4.9). The angular fun tions
p = [B1 , B2 , B3 , B4 ] are illustrated in gure 4.7
pT (X) =
B function
B2 function
2.5
1
1.5
0
1
0.5
2
3
10
0
10
10
10
4
0
(a) B1
(b) B2
B function
B function
2
2
1
0
1
0
2
3
10
1
10
10
8
10
8
4
0
2
0
( ) B3
4
0
2
0
(d) B4
Figure 4.7: Bran h enri hment fun tions
4.2.2 Intrinsi enri hment

Let us onsider the EFG method with MLS basis fun tion p. The intrinsi
enri hment is realized by introdu ing the asymptoti near-tip displa ement eld
of the Westergaard solution into p:

p (X) = 1, X, Y, r sin , r cos , r sin sin(), r cos sin()

2
2
2
2
T
(4.12)
84CHAPTER
where r is the radial distan e to the ra k tip and the angle to the ra k (or its
virtual extension), see gure 4.6. Note that the linear terms are not related to
the asymptoti near ra k tip displa ement eld and an be repla ed by higher
order polynomials.
Re alling the EFG shape fun tions
J (X) = p(X)T A(X)1 pJ (X) W (X XJ , h)
A(X) =
JS
pJ (X) pTJ (X) W (X XJ , h)
the additional omputational ost introdu ed by the intrinsi enri hment be omes obvious. Moreover, the moment matrix A be omes ill onditioned. The
size of the domain of inuen e has to be enlarged to guarantee the regularity
of A. Though this ill- onditioning does not ae t the nal solution, it is quite
troublesome. By diagonalizing A with a Gram-S hmidt orthogonalization, the
regularity of the moment matrix an be drasti ally improved but leads to more
omplex shape fun tions. Moreover, interior dis ontinuities will be introdu ed
if the intrinsi basis is not employed in the entire body of onsideration. Therefore, Fleming et al. [18 proposed a pro edure to blend nodes with dierent basis
in a ertain transition region, see gure 4.8. The approximation is written by
uh (X) = R uenr (X) + (1 R) ulin (X)
(4.13)
where u (X) is the enri hed approximation, u is the linear approximation

and R is a ramp fun tion that is 1 on the enri hed boundary of the oupling
domain and 0 on the linear boundary of the oupling domain. This is the only
requirement imposed on R. Belyts hko and Fleming [17 suggested to hoose
R = 1 or R = 1 10 3 + 15 4 6 5 with = (r r1 )(r2 r1 ) (see gure
4.8). The nal approximation then reads
X
J (X)
uh (X, t) =
(4.14)
uJ (t)
enr
JS
with
J (X) = R enr (X) + (1 R)lin (X)
J
J
(4.15)
where the shape fun tions

and
are formed by the enri hed and
linear basis, respe tively. This blending ensures ontinuity in the displa ement
eld. However, in ase for linear blending, i.e. R = 1 , dis ontinuities in the
strain eld o ur.
enr
J (X)
lin
J (X)
4.2.3 Extrinsi enri hment

Extrinsi MLS-enri hment
In the MLS extrinsi enri hment, the near ra k tip asymptoti eld of LEFM
is as follows:
nc
X
X

K
K
uh (X, t) =
(4.16)
p(XJ )T a(X, t) +
kIK QK
I (XI ) + kII QII (XI )
JS
K=1
85
r2
r1
crack
Enriched
Transition
PSfrag repla ements

Linear
Figure 4.8: Coupling between enri hed and linear approximations
86CHAPTER
where nc is the number of ra ks in the model, uh is the approximation of u,

p is the usual polynomial basis of n th order ompleteness and kI and kII
are additional degrees of freedom asso iated with mode-I fra ture and mode-II
fra ture. Note that kI and kII physi ally indi ate the stress intensity fa tors
(SIF). They an be used as approximate SIFs that an be obtained by dire tly
solving the system of equations without onsidering the J -integral.
As for the standard MLS approximation, the oe ients a are determined
by minimizing the weighted dis rete L2 error norm
X1
J=
2
p(XJ ) a(X, t) +
JS
nc
X
kIK QK
I
K K
kII
QII
K=1
uJ (t)
!2
W (XXJ , h0 )
that leads with the stationarity of J to

A(X)a(X) =
JS
with
A(X) =
PJ (X) uJ
X
JS
and
nc
X
kIK QK
I
K K
kII
QII
K=1
!
p(XJ ) pT (XJ ) W (X XJ , h0 )
PJ (X) = [W (X X1 , h0 )p(X1 ), ..., W (X Xn , h0 )p(Xn )]
with n numbers of nodes. Solving eq. (4.18) with respe t to a gives

!
nc
X
X
K K

1
K K
a(X) =
A (X)PJ (X) uJ
kI QI + kII QII
(4.17)
(4.18)
(4.19)
(4.20)
(4.21)
K=1
JS
After some algebra, the nal approximation in terms of the nodal parameters is
obtained:
!
nc
X
X

K K
h
T
1
K K
u (X) =
p (X)A (X)PJ (X) uJ
kI QI + kII QII
+
JS
nc
X
K=1
K=1

K K
K K
kI QI + kII
QII
(4.22)
with the shape fun tions
J (X) = pT (X)A1 (X)PJ (X)
(4.23)
that an be written as
uh (X)
JS
J +
J (X) u
nc
X
K K

K K
kI QI + kII
QII
K=1
(4.24)
87
with
J
u
uJ
nc
X
K K
kIK QK
I + kII QII
K=1
(4.25)
Note that the parameters kI and kII are global parameters and no spatial derivatives are taken with respe t to them.
Extrinsi PU enri hment

In the PU-extrinsi enri hment, the asymptoti near-tip displa ement eld of
the Westergaard solution, eq. (4.12) is added externally to the approximation:
!
X
X
h
u (X) =
(4.26)
bIJ pJ (X)
I (X) uI +
IS
JSc
where bIJ are additional unknowns introdu ed into the variational formulation
and Sc is the set of nodes whose domain of inuen e is ut by the ra k tip. The
approximation (4.26) is learly a partition of unity. Note the similarity between
this approximation and the approximations des ribed in se tions 2.5.4 to 2.5.5.
As might be noted, a ompatible approximation is even guaranteed by a
local partition of unity, meaning only nodes lose to the ra k tip are enri hed.
This is a huge advantage from a omputational point of view.
88CHAPTER
Chapter 5
The eXtended nite element

method (XFEM)
5.1 Introdu tion
"Standard" nite element approximations are approximations with pie ewise
dierentiable polynomials that are obviously not well suited for problems with
strong and weak dis ontinuities. The only opportunity to a urately model
these kind of dis ontinuities is to onform the nite element mesh with the line
of dis ontinuity. This be omes a major di ulty for moving dis ontinuities1 as
already mentioned at the very beginning of this manus ript.
The XFEM is a nite element approximation that is able to handle arbitrary strong and weak dis ontinuities. XFEM is based on a lo al2 extrinsi PU
enri hment that idea was used before for strong dis ontinuities, i.e. ra ks, in
meshfree methods, see se tion 4.2.3. Hen e, it is not surprising that the rst
XFEM approximations were developed for ra ks. The XFEM on ept was later
extended to interfa e problems and weak dis ontinuities. The basi idea is the
same as des ribed in se tion 4.2.3: To enri h the approximation spa e su h that
it is able to reprodu e ertain features of the problem of interest, e.g. ra ks or
interfa es.
One important aspe t of problems with moving interfa es is the tra king of
these interfa es. A powerful tool for tra king interfa es is the level set method
that will be explained in the next se tion. Though it is not mandatory to use
level sets in XFEM, most XFEM formulations take advantage of the level set
method.
1 That an o ur for moving interfa es in two phase-ow problems or evolving ra ks for
exemaple
2 lo al means that the enri hment is only employed in a ertain local subdomain. Note that
the methods in se tions 2.5.4 to 2.5.5 were proposed as global PU
89
90CHAPTER
5. THE EXTENDED FINITE ELEMENT METHOD (XFEM)
5.2 Level sets

5.2.1 Denition
Let us onsider a domain that is divided into 2 domains A and B . The
interfa e between these two domains is denoted by , see g. 5.1.
= A B
A
A B =
~n
>0
:
B
surfa e of dis ontinuity (or interfa e)

between
and
<0
=0
Figure 5.1: Des ription of a two-domain problem by using level sets

Let us dene the level set fun tion (x) as
(x) > 0 x A
(x) < 0 x B
(x) = 0 x
(5.1)
The interfa e is the zero iso ontour of (x). Note that the position of the
interfa e an hange with time. Hen e, (x, t) will be time dependent as well.
5.2.2 Normal and urvature

The normal n to the interfa e at the point x is dened by
n=
k k
(5.2)
If k k= 1, then n = . The normal n is oriented from B to A if B is

the -negative domain and A is the -positve domain.
The urvature of at a point x is dened by
K = ni,i
(5.3)
K = ni,i = ,ii
(5.4)
If k k= 1 then
91
5.2. LEVEL SETS
5.2.3 Level set for weak form integration

Integration of a fun tion over
Let us suppose that we need to integrate a fun tion f (x) in the domain . This
integral an be de omposed into 2 terms on A and B :
Z
Z
Z
(5.5)
f (x)
f (x) +
f (x) =
With the Heaviside fun tion H()

H() =
1 > 0
0 < 0
(5.6)
the 2 domains A and B an be dened by

A = {x /H((x)) = 1}
B = {x /H((x)) = 1}
The previous integral an obviously be re-written as

Z
Z
Z
f (x) =
f (x)H((x)) +
f (x)H((x))
Integration of the derivative of a fun tion over
Let us onsider the integral of the derivative of a fun tion in A :

Z
Z
f,i (x)H((x))
f,i (x) =
A
(5.7)
(5.8)
The domain integral on the LHS an be transformed into a surfa e integral by

the divergen e theorem:
Z
Z
(5.9)
f (x)ni
f,i (x) =
A
where ni is the exterior normal to A . Integration by parts on the RHS of eq.

(5.46) gives
Z h
Z
i
(f (x)H((x))) ,i f (x) (H((x))),i
f,i (x) =
(5.10)
A
In order to ompute the right hand side, the derivative of H((x) is needed:

H((x)) = ,i (x)H,i ((x)) = ,i (x)((x))
(5.11)
,i
92CHAPTER
PSfrag repla ements

A
A
B
int A =
ext A =
ext A =
Case 1: A = ext A int A
Case 2: A =
Case 3: A = ext A
Figure 5.2: Des ription of a two-domain problem by using level sets

where () is the dira delta fun tion. By assuming that the norm of the gradient
of is 1, it has been shown that ,i (x) is then equal to the normal nBA to
the interfa e if evaluating on the interfa e. Finally,

H((x))
= nBA
on
i
,i
(5.12)
0 otherwise.
The right hand side of the proposed expression leads to:

Z
Z
Z

f,i (x)H((x)) =
f (x)H((x))
f (x) H((x))
,i
,i
Z
Z
=
f (x)H((x))ni f (x)nBA
i
Z
Z
(5.13)
f (x)H((x))ni + f (x)nAB
=
i
From the last expression, 3 ases an be examined (g. 5.2):

Case 1:
f,i (x)H((x))
|{z}
= ext A
f (x) H((x)) ni +
| {z }
=1
f (x)ni
|{z}
f (x)nAB
i
= int A
(5.14)
Case 2:
f,i (x)H((x))
Z
Z
f (x) H((x)) ni +
| {z }
=0
f (x)ni
|{z}
f (x)nAB
i
=A
(5.15)
93
5.2. LEVEL SETS
Case 3:
f,i (x)H((x))
f (x)
H((x))
| {z }
ni +
=1 onlyif xA
f (x)nAB
i
f (x)ni
(5.16)
5.2.4 Smoothed Heaviside and Dira fun tion

It an be ne essary to use smoothed fun tions to avoid numeri al problems
su h as instabilities in the numeri al version. For instan e, one an use for the
Heaviside fun tion:
for <
0
1
1
H() =
+
+
sin
for < <
2 2 2
1
for <
or
0
for <

3
1
1
for < <
H() =
2 + 8 9 5( )
1
for <
Note that shall be smaller than the element size. A possible smoothed Dira
delta fun tion is
for <
0
1
1
() =
+ sin
for < <
2 2
0
for <
5.2.5 Signed distan e fun tion
A usual level set fun tion is the signed distan e fun tion (g. 5.3). This fun tion
rea hes the property of beeing zero on the surfa e of dis ontinuity. The distan e
d from a point x to the interfa e is
d =k x x k
(5.17)
where x is the normal proje tion of x on . The level-set fun tion (x) is set
to
(x) = d in A ,
(x) = d in B .
This an be written as

k sign n (x x
)
(x) = |{z}
min k x x
(5.18)
Note that the signed distan e fun tion has the following property:
k k= 1
(5.19)
94CHAPTER
~n
<0
>0
=0
Figure 5.3: Signed distan e fun tion
5.2.6 Dis retization of the level set

In the general ase, the signed distan e to the interfa e is not known analyti ally. The level set is usually dis retized. The dis retization is based on a nite
element mesh where NI (x) is the shape fun tion asso iated to the node I . The
set of nodes whi h belong to the mesh is denoted by S. Note that the same
nite element shape fun tions as for the me hani al properties an be used that
fa ilitates implementation though this is not mandatory, ? . The dis retized
level set is
X
(x) =
(5.20)
NI (x)I
IS
where I is the value of the level set on node I . This dis retization be omes
usefull when the value of the level set is needed at the element level. It an
be evaluated by interpolation. Moreover, the derivative of the level set involves
only the well known derivatives of the shape fun tions:
X
(x),i =
(5.21)
NI,i (x)I
IS
5.2.7 Updating the level set of the interfa e

When the dis ontinuity evolves in time, the level set has to be updated. Note
that there are subtle dieren es between ra ks and phase-interfa es. While
phase-interfa es generally evolve perpendi ular to the interfa e and the phaseinterfa es are lose urves, ra ks are open urves and the ra k propagates
perpendi ular to its normal from its ra k tip. Sin e a ra k is an open urve,
one level set is not su ient to des ribe the ra k. Another level set at the
ra k tip, generally perpendi ular to the level set that des ribes the ra k, has to
be introdu ed. Hen e, after updating the level set of the moving dis ontinuity,
i.e. the ra k, the level set at the ra k tip has to be updated as well. Usually, it
is required that these two level sets fulll the orthogonality ondition ,i ,i = 0.
95
5.2. LEVEL SETS
Updating of the interfa e in its normal dire tion

In general, the position of the dis ontinuity is not known a priori. The level set
has to be zero on the interfa e. Hen e, the material time derivative of has to
be zero:
D(x, t)
(5.22)
=0
Dt
By assuming that the dis ontinuity moves with a velo ity eld v, the previous
equation an be written as
(x, t)
+ (x, t) v(x, t) = 0
t
(5.23)
+ ,i vi = 0
(5.24)
or
This equation is known as the Hamilton-Ja obi equation. Using a rst-order

time s heme leads to
n+1 n
(5.25)
= n,i vin
t
and to
n+1 = n t n,i vin
(5.26)
where t is the time step. For a phase-interfa e problem su h as in bi-phasi

uid ow or in uid-stru ture intera tion (FSI) problems, the velo ity eld is
given by the ow or by the stru ture that moves in the uid eld, respe tively.
However, in ase of a propagating ra k, the ra k speed is generally unknown.
Hen e, to update (meaning propagate) the ra k surfa e, other riteria have to
be found.
Reinitialization
After evolving under a general speed vi , it generally does not remain a signed
distan e fun tion. Therefore, an be reinitialized by nding a new with the
same zero level set but with || = 1. If the position of the dis ontinuity is
known at ea h time step, the signed distan e fun tion an be re- omputed in
order to build the level-set.
5.2.8 Level set for solid shape des ription

Computational model
Let us onsider a body whi h is initially in the domain 0 with boundary 0 .
We will rst des ribe the pro edure for the ase when the obje t is en losed by
a single surfa e. It may be onvex or on ave. The pro edure onsists of the
following steps:
an impli it fun tion is onstru ted su h that (X) = 0 denes the surfa e
of the body and (X) > 0 inside the body.
96CHAPTER
voxels (background mesh)
f(X)>0
CD
active
particle
PSfrag repla ements
f(X)<0
f(X)=0
f(X)=0
Figure 5.4: Computational s heme: the impli it fun tion (X), parti le arrangement and pixels, whi h also serve as ba kground quadrature ells
a set of pixels that en ompass the entire domain for whi h (X) 0 is
onstru ted; at the orners of the pixels, nodes XI are dened. They do
not ne essarily need to be onne ted to a nite element dis retization.
The pro edure is also appli able to meshfree methods.
a set of a tive nodes is dened by I Nact if (XI ) 0 or if (XI ) = 0
for any pixel ontaining XI .
We start with the set of points that dene the obje t. The points an be
taken either dire tly from the obje t by a s anner when an existing omponent
is to be analyzed, or they may be based on a CAD model. The rst step is
to onstru t the impli it fun tion des ription of the body. The surfa es are
I , I =1
des ribed by tting an approximant to a set of points on the surfa e X
to nsp . A set of o-surfa e points XI , I =1 to nip is also needed; these are

generated by moving away from the surfa e by an approximate surfa e normal.
The surfa e of the body is des ribed by
(X) = 0
(5.27)
i.e. the surfa e is the zero isobar of the fun tion (X). The expression for (X)
in terms of the approximating fun tions NI (X) is
X
(X) =
(5.28)
NI (X) XI
IS
The out ome of these steps are illustrated in gure 5.4.
97
5.3. STANDARD XFEM
Ba kground mesh (voxels) and a tive pixels

The impli it fun tion is en losed by a bounding box B that in ludes all points X
su h that (X) > 0; this bounding box is alled the omputational domain CD .
The bounding box refers to the referen e domain (i.e. the initial onguration
of the body), so its size does not hange as the body deforms. Note that the
bounding box does not ne essarily have to be re tangular. The bounding box
ould also onsist of unions of re tangular boxes.
An array of pixels is then onstru ted. The sele tion of the pixel size is a
ru ial step in the pro edure, sin e it is ne essary to determine a pixel size that
leads to reasonably e onomi al omputations, yet orresponds losely enough
with the details of the obje t so that the boundaries of the obje t in ea h pixel
are simple. Note that in the initial sele tion of pixel size, resolution of the
geometry is more important than a ura y of the solution. The sele tion of
a tive pixels is then made by the riterion (XI ) hp , where hp is the pixel
dimension and is a onstant between zero and one.
5.3 Standard XFEM

XFEM is based on a lo al partition of unity. It uses an extrinsi enri hment
to model the weak or strong dis ontinuity within a nite element. The XFEM
approximation an be de omposed into a usual part and into an enri hed part
X
X X
JK (X) K (X) aK
N
uh (X) =
(5.29)
NJ (X) uJ +
J
JS
KE JSc
where the rst term on the RHS of eq. (5.29) is the usual approximation and
the se ond term on the RHS of eq. (5.29) is the enri hment. S is the set of
nodes in the entire dis retization and Sc is the set of enri hed nodes that are
J are the shape fun tions3 . (X) is an
inuen ed by the interfa e; NJ and N
enri hment fun tion that is hosen a ording to the problem of interest4 and aJ
are additional degrees of freedom introdu ed into the variational formulation.
The set E is the number of interfa es5 . If we have only a single interfa e, then
the rst sum of the se ond term of the right hand side of equation (5.29) an
be omitted (as well as the superimposed K ).
5.3.1 Appli ation to strong dis ontinuities

J (X) to be able to model
Let us use the XFEM enri hment with NJ (X) = N
one single strong dis ontinuity in one dimension. Therefore, the enri hment
fun tion is hosen to be the step fun tion S :

1 > 0
S() =
(5.30)
1 < 0
3 Note that generally dierent shape fun tions an be used for the standard part and for
the enri hment
4 (X) will dier for a strong and weak dis ontinuity
5 e.g. ra ks
98CHAPTER
<0
=0
crack 3
>0
N3 (X)
N2 (X)
Shifting
crack
N2 (X) (H(f (X)) H(f (X2 )))

N2 (X)H(f (X))
PSfrag repla ements

N3 (X) (H(f (X)) H(f (X3 )))
N3 (X)H(f (X))
Figure 5.5: Prin iple of XFEM in 1D for a strong dis ontinuity

The approximation (5.29) then reads:
X
X
uh (X) =
NJ (X) S((X)) aJ
NJ (X) uJ +
JS
(5.31)
JSc
Moreover, we assume linear shape fun tions N1 = 0.5(1 r), N2 = 0.5(1 + r),
where r is a lo al s aled element oordinate. Let us onsider three nite elements
as shown in gure 5.5. The shape fun tions N2 (X) of node 2 and N3 (X) of node
3 are shown in gure 5.5 as well. The nodes are numbered in as ending order
from 1 to 4. The element in the middle supposed to have a strong dis ontinuity,
i.e. a ra k, at an arbitrary lo ation Xc between node 2 and node 3. Therefore,
the nodes 2 and 3 will be enri hed. The nodes 1 and 4 will not be enri hed sin e
they are not inuen ed by the ra k.
The ra k is dened with the level set (Xc ) = 0 where Xc is the global
position of the ra k and (X) < 0 when X < Xc and (X) > 0 when X > Xc .
Sin e for the oordinate of node 2, X2 < Xc , it follows that S((X2 )) = 1.
A ordingly, S((X3 )) = 1 sin e it lies on the opposite side of the ra k with
X3 > Xc . The resulting enri hed shape fun tions NJ (X) S(X) for nodes 2 and
3 are shown on the LHS of gure 5.5 as well. From (5.31), it is obvious that the
value of u(X) on an enri hed node K Sc is
u(XK ) = uK + S((XK )) aJ
(5.32)
Thus, the nodal parameter uK is not the real displa ement value on the node.
99
5.3. STANDARD XFEM
In order to satisfy this relation, the enri hed shape fun tions are shifted around
the node of interest:
X
X
uh (X) =
(5.33)
NJ (X) (S((X)) S((XJ ))) aJ
NJ (X) uJ +
JS
JSc
su h that u(XK ) = uK . The result of this shifting is illustrated on the RHS of

gure 5.5. We note that the enri hed region is getting narrower. This shifting
is standard in XFEM. The jump in the displa ement eld is given by
[[uh (X)]] =
=
u(X + ) u(X )
X
X

NJ (X + ) S((X + )) aJ
NJ (X + ) uJ +
JS
JSc
NJ (X ) uJ +
JS
JSc
JSc

NJ (X ) S((X )) aJ

NJ (X) S((X + )) S((X )) aJ
NJ (X) aJ
(5.34)
JSc
Here, we have used the ontinuity of the shape fun tions a ross the ra k, i.e.
NJ (X ) = NJ (X + ). Several authors prefer to use the Heaviside fun tion instead of the step fun tion to model the jump in the displa ement eld. In that
ase, eq. (5.34) be omes:
X

NJ (X) H((X + )) H((X )) aJ
[[uh (X)]] =
JSc
(5.35)
NJ (X) aJ
JSc
and hen e the jump is
NJ (X) aJ instead of 2
JSc
NJ (X) aJ . Note that
JSc
the hoi e of the jump in the enri hment fun tion does not matter. The nodal
parameters will adjust automati ally.
5.3.2 Appli ation to weak dis ontinuities

For a weak dis ontinuity, the enri hment fun tion is hosen to be the signed
distan e fun tion :
J (x, t) = |(x, t)| |(xJ , t)|
The approximation (5.29) then reads:

X
X
NJ (x) J ((x), t) aJ (t)
v h (x) =
NJ (x) vJ (t) +
JS
JSc
(5.36)
(5.37)
100CHAPTER 5.
<0
THE EXTENDED FINITE ELEMENT METHOD (XFEM)
=0
2 interface 3
>0
N3 (X)
N2 (X)
2 (X)
N2 (X) (H(f (X)) H(f (X2 )))
3 (X)
N3 (X) (H(f (X)) H(f (X3 )))
PSfrag repla ements

2 (X)
Figure 5.6: Prin iple of XFEM in 1D for a weak dis ontinuity
101
5.4. HANSBO-HANSBO XFEM
where Sc denotes the inferfa en. Note that we have written the approximation
in terms of spatial oordinates sin e weak dis ontinuities are often asso iated in
uid me hani s (two-phase ow). Hen e, we have also used the letter v instead
of u that should indi ate velo ity instead of displa ement. We onsider again 3
elements in one dimension that are numbered from 1 to 4 in as ending order and
with the interfa e between node 2 and node 3, see gure 5.6. The enri hment
fun tion for node 2 and node 3 is illustrated as well. As an be seen, a kink
is introdu ed that will ause the jump in the gradient of the fun tion sin e the
jump o urs in the derivatives of , see the bottom on the LHS of gure 5.6.
The resulting shape fun tion N2 (x, t) 2 (x, t) for node 2 and N3 (x, t) 3 (x, t)
for node 3 are shown in gure 5.6, too. The velo ity gradient is obtained by
formal dierentiation of eq. (5.37):
X
v h (x) =
NJ (x)vJ (t)
JS
JSc
with
(NJ (x) J ((x), t) + NJ (x) J ((x), t)) aJ (t)(5.38)
J (x, t) = sign() = sign()nint
(5.39)
where nint denotes the normal to the interfa e. The only term that an ause
the jump is J (x, t). Similar to the jump in the displa ement eld, we obtain
the jump in the strain eld
X
NJ (X) aJ nint
[[v h (X)]] = 2
(5.40)
JSc
or the jump normal to the inferfa e:
[[v h (X)nint ]] = 2
NJ (X) aJ
(5.41)
JSc
where the fa tor 2 results from the step size from 1 to 1, see gure 5.6.
5.4 Hansbo-Hansbo XFEM

An alternative to the standard XFEM for strong dis ontinuities was proposed
by Hansbo and Hansbo [19. They do not model the ra k kinemati s with
additional degrees of freedom but by overlapping elements, see gure 5.7. The
Hansbo and Hansbo [19 XFEM-version an be derived from standard XFEM.
Therefore, let us onsider the approximation of the dis ontinuous displa ement
eld for a linear nite element in one dimension:
uh (X) =
2
X
I=1
NI (X) [uI + aI (H(X Xc ) H(XI Xc ))]
= u1 N1 + u2 N2 + a1 N1 H(X Xc )
+ a2 N2 [H(X Xc ) 1]
(5.42)
102CHAPTER 5.
where H is the Heaviside fun tion. With NI = NI H(XXc)+NI (1 H(X Xc )),

I = 1, 2, we an rewrite eq. (5.42)
uh (X) =
+
(u1 + a1 ) N1 H(X Xc ) + u1 N1 (1 H(X Xc ))

(u2 a2 ) N2 (1 H(X Xc )) + u2 N2 H(X Xc ) (5.43)
Let us dene
element1
element2
u11 = u1
= u 2 a2
(5.44)
u21 = u1 + a1
u22 = u2
(5.45)
u12
where supers ripts and subs ripts denote the element and node numbers, respe tively. Eq. (5.43) an then be rewritten as
uh (X) = u11 N1 (1 H(X Xc )) + u12 N2 (1 H(X Xc ))
+ u21 N1 H(X Xc ) + u22 N2 H(X Xc )
(5.46)
Thus, we an onsider the displa ement eld to onsist of the displa ement
elds of two elements: element 1, whi h is only a tive for X < Xc due to
(1 H(X Xc )) and element 2, whi h is only a tive for X > Xc due to H(X
Xc ). The displa ement jump a ross the ra k is then:
[[uh (X)]]X=Xc
= |{z}
lim [u(X + ) u(X )]X=Xc
0

= N1 (Xc ) u21 u11 + N2 (Xc ) u22 u12
= a1 N1 (Xc ) + a2 N2 (Xc )
(5.47)
From eq. (5.46), we an see that the dis ontinuous eld an be onstru ted by
adding an extra element, element 2 in that ase, as shown in gure 5.7. Then,
two additional nodes are added (u12 and u21 ). The two parts of the model are
ompletely disjoint.
5.5 Blending elements

Around the enri hed domain PU whi h is the support of the partition of unity,
there is a partial enri hed area where the elements do not have all their nodes
enri hed. These elements are alled blending (or partial enri hed) elements (g.
5.8). The approximation in a 4-node element whi h has less than 4 enri hed
nodes, 3 for instan e, is
uhi (x) =
4
X
I=1
NI (x)uIi +
3
X
J=1
NJ (x)(x)aJi
(5.48)
103
5.5. BLENDING ELEMENTS
crack
<0
XC
3
>0
2
<0
N1 (X) (H(X Xc ) H(X1 Xc ))
XC
N1 (X)
4 >0
crack
N4 (X)
1
PSfrag repla ements
N2 (X) (H(X Xc ) H(X2 Xc ))
N2 (X)
N1 (X)
u+
u+
NI u I
NI u I
[[u]]
1
[[u]]
u
4
Figure 5.7: Standard XFEM vs. Hansbo XFEM

In this element the fun tion (x) an not be re overed by taking uIi = 0 and
aJi = 1 be ause (N1 , N2 , N3 ) is not anymore a partition of unity, in other terms,
3
X
J=1
NJ (x) 6= 1.
(5.49)
The fa t that the additional fun tion an not be re overed in these elements is
not important sin e these elements do not ontain the dis ontinuity. The main
point is that it may introdu e spurious terms in the approximation whi h produ e an error in the solution. The spurious terms an be automati ally orre ted
by the standard part of the approximation if the order of the standard part is
more or equal than the order of the partition of unity times the enri hment.
Table 5.1 shows a few possible ombinations.
The spurious terms an be orre ted by an assumed strain method [20. The
nite element shape fun tions form a partition of unity
X
(5.50)
NI (x) = 1
IN
It follows from the above that for an arbitrary fun tion (x), the following
satises
X
(5.51)
NI (x)(x) = (x)
IN
104CHAPTER 5.
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
11
00
00
11
00
11
00
11
0011
11
0011
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
0011
11
00
0011
11
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
0011
11
0011
11
0011
00
00
11
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
00
11
00
11
00
11
00
11
00
11
00
11
0011
11
00
11
00
11
00
11
00
00
00
00
00
00
00
0011
11
0011
11
0011
11
0011
11
0011
11
0011
11
0011
11
0011
11
00
PU , enri hed elements
Set Senr of enri hed nodes
p.e. , partial enri hed or blending
11
00
00
11
00
11
Figure 5.8:
elements
Partial enri hed around the enri hed area.
Standard shape
fun tions
NI (x)
Partition
unity
fi (x)
of
4-node element
1st order
4-node element
1st order
9-node element
2st order
4-node element
1st order
4-node element
1st order
4-node element
1st order
Enri hment
(x)
Heaviside
zero order
Ramp
1st order
Ramp
1st order
order
Spurious
of
terms
fi (x)
(x)
1
No
Yes
No
Tableau 5.1: A few standard ombinations for shape fun tions, partition of unity
and enri hment
Therefore any fun tion an be reprodu ed by a set of fun tions NI . This
is the key property of enri hed nite element methods based on a partition
of unity. Although one ould enri h the entire domain, only a sub-domain is
usually enri hed sin e the features need to be modeled lo al for instan e, a
ra k ompared to the plate ontaining it. Moreover, keeping enri hment lo al
permits keeping the matrix banded. This is why X-FEM an be onsidered as
a lo al partition of unity enri hed nite element method. A partitioning of a
typi al domain into its non-enri hed sub-domains and enri hed sub-domains is
shown in Figure 5.9. In this lo al enri hment s heme, three types of elements
are distinguished. The rst types are the lassi al nite elements, those in
whi h none of its nodes are enri hed, these elements are grouped in std . The
se ond type are fully enri hed elements, i.e. all of its nodes are enri hed. These
105
5.5. BLENDING ELEMENTS
elements are denoted as enr . The third type of elements, alled partiallyenri hed elements, are those for whi h only some but not all of the nodes are
enri hed. These elements form the blending sub-domain blnd .
blnd
std
111111
000
000111
000
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
111
000
111
000
111
000
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
000111
111
000111
000
111
enr
PSfrag repla ements
Figure 5.9: Typi al dis retization illustrating enri hed domain enr , transition
domain blnd and standard domain std ; lled nodes are enri hed
Let uI = 0 and aJ = 1 in the enri hed
we have
eJ (x)(x) =
N
X
h
e (x)(x) 6=
u (x) =
N
eJ
JN enr N
J (x)(x) =
nite element approximation, then

(x) x enr
(x) x blnd
0 x std
(5.52)
Therefore the approximation an reprodu e the enri hment in enr and it vanishes in std . However, in the blending domain, it onsists of the produ t of a
eJ and the enri hment fun tion so this
subset of the enri hed shape fun tions N
enri hment fun tion annot be reprodu ed. The blending elements or transition
elements lead to a lower onvergen e rate for enri hed nite element methods
ompared to standard nite element methods. The following example, extra ted
from [20, shows the reason. Consider a one dimensional mesh as illustrated in
Figure 5.10 with a dis ontinuity in the derivative in element 0. The enri hment
fun tion is the ramp fun tion
(x) = xH(x)
(5.53)
106CHAPTER 5.
Figure 5.10: A 1D example of how a lo ally enri hed nite element method fails
to be able to reprodu e a linear eld. The desired pie ewise linear eld is shown
in (b) and the enri hed part (dotted line ), standard part (dashed line ), and the
total (solid line ) approximation are shown in ( ). The dis retization is shown in
(a), where the enri hed nodes denoted by lled ir les, the two blending elements
are lled and the fully enri hed element is hashed [20.
where H is the Heaviside step fun tion. This enri hment adds a dis ontinuity
in the gradient of the approximation at x = 0. Linear shape fun tions are used
for both the standard approximation and the partition of unity. Let element
0 be the fully enri hed element and element 1 be the blending element to the
right. The approximation of element 1 is given by
uh (x) =
2
X
I=1
NI (x) + N1 (x)(xH(x) x1 H(x1 ))a1
uh () = u1 (1 ) + u2 + a1 h(1 )
(5.54)
(5.55)
where
=
x x1
h
(5.56)
and h is the length of element 1. Let the nite element interpolation to the
solution be given by uh and denote the error in the interpolation by e, we have
e u uint
(5.57)
The maximum error o urs at the point x where

e,x|x
d
e(x) = 0
dx
(5.58)
107
5.6. IMPLEMENTATION
Then a Taylor expansion about x gives
or
1
e(x) = e(x) + e,x|x(x x) + e,xx|x(x x)2 + O(h3 )
2
(5.59)
1
e(x) = e(x) + e,xx|x (x x)2
(5.60)
2
If we let x = x1 , then e(x1 ) = 0 sin e uh is the nite element interpolation of
u, i.e. uh (xI ) = u(xI ). Therefore, we obtain
1
e(x) = e,xx|x(x x)2
2
(5.61)
Sin e
2a1
h
(5.62)
1
1
(x x1 )2 h2
2
8
(5.63)
e(x) = u,xx +
and sin e
it follows that
2a1
1 2
h max(u,xx +
)
(5.64)
8
h
The last term, 2a1 /h, does not appear for standard nite elements. It in reases
the interpolation error in the blending elements from order h2 to h. Although
this o urs only in few elements, it redu es the rate of onvergen e of the entire
approximation. The reason for this is that the partition of unity property ( ompleteness) of the approximation is not veried in the whole domain. Therefore,
the theoreti al rate of onvergen e annot be attained. When the enri hment
is a polynomial of order n, i.e. n , then for n > 1 the interpolation error in the
blending elements is in reased even further. If we go through the same steps as
before, we nd that
e(x)
1 2
2a1
(5.65)
h max(u,xx + n )
8
h
At this time, we an understand why in the quadrati XFEM formulation,
the partition of unity shape fun tions have been hosen as linear. To get improved rate of onvergen e, Chessa proposed the enhan ed strain formulation
for the blending elements. By properly hoosing an enhan ed strain eld, the
undesired terms in the enri hed approximation an be eliminated.
e(x)
5.6 Implementation
The implementation of XFEM will be explained for a 4-node quadrilateral element with linear shape fun tions, see gure 5.11.
108CHAPTER 5.
5.6.1 Shape fun tions

s=1
node4
: (x4 , y4 )
node3
(1, 1)
: (x3 , y3 )
(1, 1)
r=1
r = 1
r
r
y
node1
(1, 1)
: (x1 , y1 )
node2
s = 1
: (x2 , y2 )
Physi al element
Parent element
Figure 5.11: Physi al and parent 4-node elements

The shape fun tions NI , I = 1...4 are given by:
1
(1 r)(1 s)
4
1
N2 (r, s) = (1 + r)(1 s)
4
1
N3 (r, s) = (1 + r)(1 + s)
4
1
N4 (r, s) = (1 r)(1 + s)
4
N1 (r, s) =
where r and s are s aled oordinates in the parent element, see gure 5.11. The
standard nite element approximation of the displa ement is
2
ue (M ) =
ux
uy
N1
0
N2
0
N3
0
N4
0
0
N1
0
N2
0
N3
6
6
6
6
6
0
6
N4 6
6
6
6
4
ux1
ux2
ux3
ux4
uy1
uy2
uy3
uy4
7
7
7
7
7
7 = Nestd (M ) qe
7
7
7
7
5
(1, 1)
109
5.6. IMPLEMENTATION
The enri hed nite element approximation of the displa ement is

ue (M )
ux
uy
N1
0
N2
0
N3
0
N4
0
0
N1
0
N2
0
N3
0
...
N4
2
...
ue (M )
N1 1
0
N2 2
0
N3 3
0
N4 4
0
0
N1 1
0
N2 2
= [ Nestd (M ) Neenr (M ) ] qe
0
N3 3
0
N4 4
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
4
ux1
ux2
ux3
ux4
uy1
uy2
uy3
uy4
a x1
a x2
a x3
a x4
ay 1
ay 2
ay 3
ay 4
3
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
5
In other terms, this leads to

ue (M ) = Ne (M ) qe
where Ne (M ) = [Nestd (M ) Neenr (M )]. In this last expression, it is assumed

that the 4 nodes of the element are enri hed by a single additional fun tion
(x) where I denotes the shifted fun tion:
I (x) = (x) (xI )
5.6.2 The B-Operator

The strain tensor omponents in Voigt notation are
xx
= yy = Due (M )
2xy
with
x
0
D=
By repla ing ue (M ) by its approximated form, it gives

= DNe (M ) qe = Be (M ) qe
110CHAPTER 5.
where Be (M ) is the dis retized gradient operator. It ontains both the standard
part and the enri hed part and it an be written as
Be (M ) = [Bestd (M ) Beenr (M )]
In the last expression, the dis retized gradient operator Bestd (M ) is equal to:
N1,x N2,x N3,x N4,x

0
0
0
0
0
0
0
N1,y N2,y N3,y N4,y
Bestd = 0
N1,y N2,y N3,y N4,y N1,x N2,x N3,x N4,x
and the enri hed dis retized gradient operator Beenr (M ) is equal to
Beenr
(N1 1 ),x
0
=4
(N1 1 ),y
(N2 2 ),x
0
(N2 2 ),y
(N3 3 ),x
0
(N3 3 ),y
(N4 4 ),x
0
(N4 4 ),y
0
(N1 1 ),y
(N1 1 ),x
0
(N2 2 ),y
(N2 2 ),x
0
(N3 3 ),y
(N3 3 ),x
In the ase of an Heaviside enri hment, the derivative of the approximation

an be written as
X
X
uhi,j =
NJ,i (x) ujJ +
(NJ (x)H((x))) ,i ajJ
IS
IS
NJ,i (x) ujJ +
IS
(NJ,i (x)H((x)) + NJ (x)H,i ((x))) ajJ
IS
The derivatives of the Heaviside is the dira delta fun tion H,i ((x)) = .
That means that H,i = 1 at the ra k interfa e and H,i = 0 otherwise. If we
assume the ra k to be tra tion-free, we an omit the derivatives of the Heaviside
fun tion and an give the enri hed B-operator as
Beenr
N1,x 1
0
=4
N1,x 1
N2,x 2
0
N2,x 2
N3,x 3
0
N3,x 3
N4,x 4
0
N4,x 4
N1,y 1
N1,y 1
N2,y 2
N2,y 2
N3,y 3
N3,y 3
N4,y 4 5
N4,y 4
In the ase of a ramp enri hment, (x) = |(x)|, the omputation of the
derivative of (x) is needed:

(x) = sign((x)) ,i (x)
,i
The derivative of (x) is then needed. Sin e
(x) = [ N1 N2 N3
the derivative a ording to x is
1
2
N4 ]
3
4
(x),x = [ N1 ,x N2 ,x N3 ,x N4 ,x
1
2
]
3
4
3
0
(N4 4 ),y 5
(N4 4 ),x
111
5.6. IMPLEMENTATION
and the derivative a ording to y is
(x),y = [ N1 ,y N2 ,y N3 ,y N4 ,y
5.6.3 The element stiness matrix
1
2
]
3
4
The previous omputations involve the derivative of the shape fun tions in terms
of the physi al oordinates. The relations between the derivative in the parent
and in the physi al oordinates are
NI r
NI s
NI
=
+
x
r x
s x
NI
NI r
NI s
=
+
y
r y
s y
whi h an be written, for ea h fun tion NI :
N,x
N,y
N,r
N,s
r r

x y
s s
x y
|
{z
1
= J
where J is the Ja obian. The derivatives of the fun tions NI (r, s) in terms of
the parent oordinates r and s are:
1
N1,r = (1 s)
4
1
N2,r = (1 s)
4
1
N3,r = (1 + s)
4
1
N4,r = (1 + s)
4
The expression of the Ja obian is
with
x
r
J=
y
r
1
N1,s = (1 r)
4
1
N2,s = (1 + r)
4
1
N3,s = (1 + r)
4
1
N4,s = (1 r)
4
x
s
y
s
112CHAPTER 5.
x=
4
X
I=1
I=1
I=1
x X NI
x X NI
NI xI ,
=
xI ,
=
xI
r
r
s
s
whi h is also
x
= N1,r
r
x
= N1,s
s
N2,r
N3,r
N4,r
N2,s
N3,s
N4,s
and
y=
4
X
NI yI
I=1
x1
x2
x3
x4
x1
x2
x3
x4
y X NI
=
yI
r
r
I=1
y X NI
=
yI
s
s
I=1
whi h is also
y
= N1,r
r
y
= N1,s
s
N2,r
N3,r
N4,r
N2,s
N3,s
N4,s
y1
y2
y3
y4
y1
y2
y3
y4
The expression of the stiness matrix for an enri hed element is

Z 1Z 1
Z
T
T
Be (M ) Ce Be (M ) d =
Be (r, s) Ce Be (r, s) det J dr ds
Ke =
e
where C is the onstitutive tangent operator. The stiness matrix an be

de omposed into four 8 8 blo ks:
Z
Z
T
eT
e e
Bstd (M )C Bstd (M )
Bestd (M )Ce Beenr (M )
e
Kel = Z
eT
e e
eT
e e
Benr (M )C Benr (M )
Benr (M )C Bstd (M )
e
The up-left blo k orresponds to the standard stiness matrix. The elemental
enri hed stiness matrix has a 16 16 size.
113
5.7. INTEGRATION
crack
2
5
3
9
Crack path produced

by level set
background cell
Crack path recognized by the code
crack
2
3
7
9
4
6
5
11
10
background cell
Figure 5.12: Sub-triangulation of nite elements
5.7 Integration
Numeri al integration is the last step to obtain the dis rete equations. In FEM,
Gauss quadrature is usually applied. However, standard Gauss quadrature annot be used if the element is rossed by a dis ontinuity. One integration strategy
is based on sub-dividing an existing elements into several smaller triangular elements as shown in gure 5.12 for the 2D ase. Di ulties o ur for high
urvatures of interfa e . For example: Consider linear nite elements. If an
interfa e severely urves as shown for the two dimensional ase in gure 5.12
and the interfa e is dis retized with level sets where for the dis retization of
the level set, the linear FE shape fun tions are used, then the level set annot
apture the urvature of the ra k orre tly. Therefore, an additional node is
introdu ed to maintain the a ura y of the integration. Adding one more point,
the error is redu ed se ond order small.
After the element with the interfa e is sub-triangulated, the integration of a
fun tion F an be done as follows:
Z
Z
F(X)d
F(X)d +
F =
+
Z
Z
F(X()) detJ+ () d (5.66)
F(X()) detJ ()d +
=
114CHAPTER 5.
Figure 5.13: Transformation of an integration method on a square into an integration method on a triangle for ra k tip fun tions
onto . With the mapping

where the Ja obian maps the domains
, we nally obtain:
Z
()d
F(X()()) detJ (()) detJ
F =
Z
+ ()d
F(X()()) detJ+ (()) detJ
+
(5.67)
+
and in its dis rete form using Gauss quadrature:

n+
GP
n
GP
F=
()wI +
F( I ) detJ () detJ
I=1
+ ()wI (5.68)
F( I ) detJ+ () detJ
I=1
+
where n
GP and nGP are the Gauss points in and , respe tively, and are
the lo al oordinates of the Gauss points and wI are their quadrature weights.
An alternative approa h is the modi ation of the quadrature weights, that
is illustrated in gure 5.14. In that method, the quadrature weights rossed by
the ra k are omputed a ording to their areas A+ and A :
A+
I
AI
A
wI = w I
AI
wI+ = w
(5.69)
The method requires subdivision of the element; we note that a ertain number
of Gauss points are needed in order to obtain a urate results.
In ra k problems, it is preferable to have an a urate stress eld around
the ra k tip in order to model the propagation of the ra k as a urate as
possible. In a sub-triangulation pro edure, state variables have to be mapped
from the original Gauss points to the new Gauss points generated by the subtriangulation pro edure at the time the ra k enters the element. Therefore, it
is preferable to use the se ond approa h with higher numbers of Gauss points
adja ent to ra k-tip elements. The mapping is then done bef ore the ra k
approa hes the element that will be ra ked.
115
5.7. INTEGRATION
Delaunay triangulation
A1
A2
A4
A5
A7
A8
Voronoi cells
A3
A6
A9
Crack
Gauss point
Node
PSfrag repla ements
A+
i
A
i
Figure 5.14: Integration with modied quadrature weights
116CHAPTER 5.
crack
crack tip
crack
crack tip
enriched nodes
not enriched nodes
Figure 5.15: One opportunity to lose the ra k in XFEM
Remark: In LEFM, the elements that ontain the ra k tip do not only have
a dis ontinuity but also a singularity at the ra k tip. Hen e, a sub-triangulation
pro edure might not be a urate enough if Gauss points of the sub-triangles are
lose to the stress singularity. This drawba k an be ir umvented by expressing
the elementary integrals of the type
Z
(5.70)
(Fj i ) (Fl k ) dx
in polar oordinates that will remove the r0.5 singularity of Fi . The geometri
transformation

x
xy
G:
(5.71)
y
y
maps the unit square onto a triangle, see gure 5.13. With this transformation,
it is possible to build a quadrature rule on the triangle from a quadrature rule on
are obtained
the unit square. The new integration points and their weights w
from those of the original quadrature rule with
= G() , w
= w det(G)
(5.72)
5.8 Appli ation of XFEM to LEFM

5.8.1 Governing equations
The governing equation is the equation of equilibrium:
0 P b = X 0 \ c0
(5.73)
The boundary onditions are
(X, t)
u(X, t) = u
on
u0
(5.74)
117
5.8. APPLICATION OF XFEM TO LEFM
n0 P(X, t) = t0 (X, t)
n0 P(X, t) = 0
on
on
t0
c0
(5.75)
(5.76)
and t0 are the pres ribed

and
respe tively
and
where u
S displa ements
S
Ttra tions,
S
T
S
T
c
0 is the ra k surfa e with u0 t0 c0 = 0 , (u0 t0 ) (t0 c0 ) (u0 c0 ) =
. We assume that the ra k surfa es are tra tion-free, see eq. (5.76).
5.8.2 Weak form

The weak form of the equilibrium equation is given by: nd u V su h that
W = Wint Wext = 0 u
where
Wint =
Wext =
(5.77)
( u)T : P d0
(5.78)
(5.79)
u b d0 +
t0
u t0 d0
where

(t) on u0 , u discontinuous on c0
u(, t)|u(, t) H1 , u(, t) = u

(5.80)
=
u|u H1 , u = 0 on u0 , u discontinuous on c0
V =
V0
In the stati ase, the solution of a boundary value problem (BVP) with internal
boundaries is to be sought in the so- alled Space of Bounded Def ormations.
5.8.3 XFEM approximation for ra ks

The appli ation of XFEM to ra ks poses additional hallenges onto the method:
It has to be ensured that the ra k loses at its ra k tip.
During a loading y le, a ra k an open and lose. It has to be ensured
that the two ra k surfa es do not overlap.
While in uid-uid or uid-stru ture intera tion problems, the interfa e
is given at the beginning of the omputation, ra ks an be initiated at
any time during a loading y le. Hen e, we need also a riterion that tells
us when a ra k is initiated, how the ra k will be oriented and how the
shape of the ra k looks like, e.g. how long is the ra k?
Cra k path tra king algorithms are more omplex than interfa e tra king
algorithms in two-phase ow or uid-stru ture intera tion problems.
118CHAPTER 5.
Figure 5.16: Example of enri hed nodes when ra k losure is realized by losing
the ra k tip at the element sides
We will dis uss these aspe ts only for a single ra k. Moreover, we onsider
two dimensional triangular and quadrilateral nite elements with linear shape
fun tions. The ra k line is approximated with level sets where the level set is
dis retized with the same shape fun tions as the me hani al properties. Hen e,
the urvature of the ra k an be maximal bi-linear for the quadrilateral element and linear for the triangular element. We don't onsider geometri al and
material non-linearities. For now, we assume that the ra k does not propagate
and study only the ra k kinemati s. The subsequent issues will be dis ussed
in the following se tions.
The Heaviside enri hment is able to apture the jump in the displa ement
eld. However, it does not ne essarily guarantee the losure of the ra k at its
ra k tip. The simplest possibility to ensure ra k losure at the ra k tip is to
lo ate the ra k tip at one of the element edges and not to enri h the adja ent
nodes, see gure 5.15. Two potential ra k pathes in luding the enri hed nodes
are illustrated in gure 5.16. The bla k dots denote the enri hed nodes. The
standard nodes are not shown. This kind of ra k losure has the drawba ks
that the ra k length is governed by the element size sin e it is required to
grow the ra k through the entire element. Another opportunity is to lose
the ra k within an element. However, the enri hment pro edure has to be
modied. Therefore, let us onsider a triangular element as shown on the RHS
in gure 5.17. The ra k is assumed to pass through the side 23 and interse t
side 12. Other relationships an be obtained by permuting the node numbers.
Due to ompatibility, the enri hment has to vanish on the sides 12 and 13 and
is ontinuous a ross 23 with the eld in the adja ent element. To meet this
ondition, only node 3 is enri hed and the dis ontinuous displa ement eld in
the tip element is
udisc = 3 3 ( ) a3
(5.81)
where = [1 2 3 ] are the parent oordinates of the sub-triangle 23P in

gure 5.17. The shaded parent area oordinates are related by 3 = 1 1 2
and 3 ( ) = sign (( )) sign (3 ). The relation between and is given
119
P
PSfrag repla ements
1
N3 () = 1 1 2
N1 () = 1
N2 () = 2
Figure 5.17: XFEM enri hment for a ra k tip that ends in an element
by
1 =
1
, 2 = 2
1P
(5.82)
where 1P is the area oordinate of point P . When the dire tion of the ra k
interse ts side 31, see the LHS of gure 5.17, then the dis ontinuous part of the
displa ement eld is
udisc = 2 2 ( ) a2
(5.83)
with
1 = 1
1P
2
2 , 2 =
2P
2P
(5.84)
and 2 ( ) = sign (( )) sign (2 ) and a3 = aP = 0. The enri hment an be

implemented easier by letting
X
udisc =
(5.85)
I I ( ) aI
I
and onstru ting aI to vanish for the nodes on the edge towards whi h the
ra k is heading. These enri hment displa ements udisc vanish on the boundary
of enr . Hen e, only the elements in enr need a spe ial treatment of the
same type. This is a partition of unity in enr and there is no inner blending
between dierent enri hments. Moreover, sin e the enri hment vanishes on the
boundary of enr , blending outside the enri hed subdomain does not o ur.
Thus, although it is a lo al partition of unity, it is indistinguishable from a
global partition of unity.
Another opportunity to lose the ra k is by use of bran h fun tions B.
In fa t, the bran h fun tions were already introdu ed in se tion 4.2 where the
120CHAPTER 5.
crack tip enrichment

Heaviside enrichment
Figure 5.18: Cra k with enri hed elements

on ept of enri hment was dis ussed:
B
= [B1 B2 B3 B4 ]

r sin , r cos , r sin sin(), r cos sin()

=
2
2
2
2
(5.86)
Here, we will denote the bran h fun tions6 with B a ording to the LEFM literature. It is obvious that for r = 0, the ra k loses at the ra k tip. Moreover,
the solution will be more a urate sin e we put the information of the analyti al
solution into our approximation. The approximation of the displa ement eld
is now given by:
X
X
I (X)) aI
NI (X) H(f
uh (X) =
NI (X) uI +
IS
ISt (X)
ISc (X)
NI (X)
K (X) bKI
B
(5.87)
where St are the set of nodes that are inuen ed by the ra k tip and the
superimposed bar denotes a shifting as des ribed in se tion 5.3.1. The rst
term on the RHS of eq. (5.87) is the standard approximation, the se ond term
and the third term is the enri hment, see also gure 5.18. Generally, only
the element ontaining the ra k tip is enri hed with the bran h fun tions B.
However, this is not mandatory. It is also possible to enri h adja ent elements
as shown on the RHS of gure 5.18.
Finally, we would like to mention the spe ial ase, where a ra k is lo ated
lose to a node7 as shown in gure 5.19. Therefore, let us onsider the element
spanned by the nodes a, b, c, d. The support of the node is dened by its
adja ent neighbor elements. Node a for example has four neighboring elements.
6 In se tion 4.2, the bran h fun tions were denoted with p
7 The node annot be enri hed with a ra k tip enri hment
121
111111
000000
000000
000000111111
111111
000000
111111
000000111111
111111
000000
000000
111111
000000
000000111111
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+
000000
111111
000000
111111
A
000000
111111
000000
000000111111
111111
000000
111111
000000111111
111111
000000
000000111111
111111
000000
000000111111
111111
000000
000000
111111
000000
a
000000111111
111111
000000
111111
000000
111111
000000
000000111111
111111
000000
111111
000000111111
111111
000000
crack
a
PSfrag repla ements
Figure 5.19: Enri hment riteria for ra ks that are lose to a node
Hen e, the support size is equal to the area of these four neighboring nite
elements. Let us dene the parameters r+ and r between the area on both
sides of the ra k and the total area of the support:
r+ =
A+
A
, r = +
+A
A + A
A+
(5.88)
If one of these ratios is smaller than a given threshold, the support node is no
longer enri hed. In the ase of our element a, b, c, d, only nodes a and b will be
enri hed.
5.8.4 Dis rete equations

To obtain the dis rete equations, the test and trial fun tions have to be substituted into the weak form of the equilibrium equation (5.77). We will onsider
the approximation with ra k tip enri hment as dened in eq. (5.87). The
approximation of the test fun tions look identi al. Then we obtain the nal
system of equations:

Kuu
Kua
Kub
uJ fIext
IJ
IJ
IJK
Kau
aJ
f ext
=
Kaa
Kab
(5.89)
IJ
IJ
IJK
Iext
bb
ba
bu
bJK
fIK
KIJK KIJK KIJK
or
K d = f ext
(5.90)
where K is the stiness matrix, d = {u a b}T is the ve tor with the nodal pa
T

rameters, f ext = f u f a f b is the external for e ve tor with f b = f b1 f b2 f b3 f b4
and
Z
Z
u
NI t d
NI b d +
fI =
(5.91)
fIa =
NI (H((X)) H((XI ))) b d+
NI (H((X)) H((XI ))) t d
(5.92)
122CHAPTER 5.
fIbl =

NI BIl (X) BIl (XI ) b d+
The stiness matrix is
K=
T C B
d
B
(5.94)
is the B-operator dened by:

where B
NI,X
BuI = 0
NI,Y

NI BIl (X) BIl (XI ) t d (5.93)
NI,X (H((X)) H((XI )))

0
BaI =
NI,Y (H((X)) H((XI )))
NI,Y
NI,X
NI,Y
NI,X
(5.95)
0
(H((X)) H((XI )))
(H((X)) H((XI )))
(5.96)
0
l
l
BK
(X) BK
(XI ) ,Y
Bbl
I |l=1,2,3,4 =

l
l
BK
(X) BK
(XI ) ,X
NI
(5.97)
In eq. (5.96), we have already omitted the derivatives of the Heaviside fun tion,
see se tion 5.6.2. The partial derivatives for the bran h fun tion are

l
l
l
(X) + NI BK
(X),i
NI BK
(X) ,i = NI,i BK
(5.98)

NI

l
l
BK
(X) BK
(XI ) ,X
0

l
l
BK (X) BK
(XI ) ,Y

N
I
NI
To obtain the derivatives of the bran h fun tions, let us dene a lo al ra k

oordinate system, see gure 5.20. The angle denotes the slope of the lo al
ra k oordinate system with respe t to the global oordinate system. The
derivatives of the bran h fun tions in the lo al ra k oordinate system are
l
l
B,li = B,r
r,i + B,
,i
(5.99)
where and r are dened in gure 4.6 and the subs ript ", i" denotes derivatives
l
l
and B,
are
with respe t to the lo al ra k oordinate system. The terms B,r
obtained by formal dierentiation:
sin(/2)
2cos(/2)
1
1
B,r =
B, =
2
2 2
2sin(/2)
cos(/2)
2
2
B,r
=
B,
=
2
2 2

cos(/2)
sin()
sin(/2)
sin()
3
3
+ sin(/2) cos()
B, = r
B,r =
2
2 2

sin(/2) sin()
cos(/2) sin()
4
4
B, = r
B,r =
+ cos(/2) cos() (5.100)
2
2 2
123
Y
X
Figure 5.20: Denition of a lo al ra k oordinate system

The derivatives of r and with respe t to the lo al ra k oordinate system are
r,X = cos() ,X = sin/r
r,Y = sin() ,Y = cos/r
(5.101)
With (5.100) and (5.101) in (5.99), we have the derivatives of the bran h fun tions in the lo al ra k oordinate system:
sin(/2)
cos(/2)
B,1Y =
2 2
2 2
sin(/2)
cos(/2)
B,2Y =
B,2X =
2 2
2 2
sin(/2) + sin(3/2) cos()
sin(3/2) sin()
B,3Y =
B,3X =
2 2
2 2
cos(3/2) sin()
cos(/2) + cos(3/2) cos()
B,4X =
B,4Y =
2 2
2 2
B,1X =
(5.102)
Finally, the derivatives in the global oordinate system are obtained by:
B,X
B,X cos() + B,Y sin()
B,Y
B,X sin() + B,Y cos()
(5.103)
where is the in lination angle of the lo al ra k oordinate system versus the

global oordinate system.
124CHAPTER 5.
2 (x) = 0
2 (x) = 0
1 (x) = 0
1 (x) = 0
PSfrag repla ements

Intersecting discontinuity
Branching discontinuity
Figure 5.21: XFEM element with interse ting dis ontinuities and bran hing
dis ontinuities
5.8.5 Advan ed XFEM in LEFM

Cra k bran hing and ra k jun tion
Let us onsider the ase of ra k bran hing and ra k interse tion in a single
nite element as shown in gure 5.21. Let S1c be the set of nodes whose domain
of inuen e is ompletely ut by the dis ontinuity des ribed with the signed
distan eTfun tion 1 (X) = 0 and S2c the orresponding set for 2 (X) = 0;
S3c = S1c S2c . The same applies a ordingly for nodes whose domain of inuen e
is ut by the ra k tip enri hment. We will denote this set of nodes with S1t and
S2t . Then the approximation of the displa ement may be given by
X
X
(1)
NI (X) H(1 (X)) aI
NI (X) uI +
uh (X) =
IS1c (X)
IS(X)
(2)
NI (X) H(2 (X)) aI
IS2c (X)
(3)
NI (X) H(1 (X)) H(2 (X)) aI
IS3c (X)
NI (X)
IS1t (X)
IS2t (X)
(1)
(1)
(2)
(2)
BK (X) bKI
NI (X)
BK (X) bKI
(5.104)
Note, that ra k bran hing requires the introdu tion of another level set. Cra k
jun tion an be treated similarly. A omputationally more e ient approa h
was proposed by Zi et al. [21 by modifying the signed distan e fun tions so that
no ross terms are needed for jun tion or bran h problems.
125
Figure 5.22: Sign fun tions for ra k jun tion

When two ra ks are joining, the ra k tip enri hment is removed. By using
the signed distan e fun tions of the pre-existing and approa hing ra k, the
signed distan e fun tion of the approa hing ra k is modied, see gure 5.22.
Three dierent subdomains have to be onsidered: (1 < 0, 2 < 0), (1 >
0, 2 > 0), (1 > 0, 2 < 0) as in gure 5.22b or (1 > 0, 2 < 0), (1 > 0, 2 >
0), (1 < 0, 2 < 0) as in gure 5.22d. The signed distan e fun tion of ra k 1
of a point X is then obtained by:
0
1 (X), if 02 (X1 ) 02 (X) > 0
1 (X) =
(5.105)
02 (X), if 02 (X1 ) 02 (X) < 0
where the superimposed 0 denotes the sign distan e fun tion before ra k jun tion. Therefore, the nal approximation without the ross term reads:
uh (X)
NI (X) uI +
m=1 ISt (X)
(n)
(n)
NI (X) H(I (X)) aI
n=1 ISc (X)
IS(X)
mt
X
nc
X
NI (X)
(m)
(m)
BK (X) bKI
(5.106)
where nc and mt are the number of ra ks that ompletely ross the element or
ontain the ra k tip, respe tively.
126CHAPTER 5.
Cra k opening and ra k losure

The governing equations in se tion 5.8 were formulated for a tra tion-free ra k
problem. To be more pre ise, the problem was stated only for the ase of ra k
opening. The ase of ra k losure that an arise for example under y li
loading onditions was not onsidered in that formulation. To generalize the
ra k problem in elastostati s, let us rephrase the problem:
(5.107)
0 P b = X 0 \ c0
with boundary onditions

(X, t)
u(X, t) = u
n0 P(X, t) = t0 (X, t)
n0 P(X, t) = 0
t+
0t = t0t = 0,
on
t+
0N = t0N
(5.108)
u0
on
on
(5.109)
t0
c0
if not in contact
(5.110)
on
c0
(5.111)
[[uN ]] 0
[[n P]] = 0
on
on
if in contact
(5.112)
c0
(5.113)
c0
where eqs. (5.107) to (5.110) were already formulated in se tion 5.8 and eqs.
(5.111) to (5.113) guarantee no interpenetration where t0N = n P n is the
normal tra tion and t0t is the tra tion a ting in tangential dire tion of the ra k.
The inequality (5.112) with [[uN ]] = u+ n+ = u n 0 guarantees that the
ra k surfa es do not interpenetrate in the ase of ra k losure and eq. (5.113)
ensures tra tion ontinuity. The superimposed plus and minus sign indi ates the
dierent sides of the ra k surfa es. Note that it is often assumed that n+ = n
though this assumption is not mandatory.
If the Lagrange multiplier method is used to enfor e no-interpenetration
onditions of the adja ent ra k surfa es, the weak form of the equilibrium
equation reads:
Z
( u)T : P d0
u b d0
t0
u t0 d0 +
c0
[[uN ]] d0 0
(5.114)
where it is su ient to hose C1 approximation fun tions for the Lagrange
multiplier eld if the approximating fun tions for the test and trial fun tion are
C0 .
127
Cra king riteria in LEFM

Let us onsider a spe imen that has a pre- ra k or a not h and we are interested
if the pre- ra k will propagate (or not) under a ertain stress state. There are
basi ally four major ra king or ra k propagation8 riteria in LEFM:
Maximum hoop stress riterion or maximum prin ipal stress riterion
Minimum strain energy density riterion, Shih [22
Maximum energy release rate riterion, Wu [23
The zero KII riterion (Vanishing in-plane SIF (KII ) in shear mode for
innitesimally small ra k extension), Goldstein and Salganik [24
The rst two riteria predi t the dire tion of the ra k traje tory from the
stress state prior to the ra k extension. The last two riteria require stress analysis for virtually extended ra ks in various dire tions to nd the appropriate
ra k-growth dire tion. Note, that these riteria will give only the orientation
of the ra k but not its length. To determine the ra k length, these riteria
have to be he ked in dierent distan es around the ra k tip. Often a onstant
ra k propagation speed is assumed. In omputational LEFM, the rst of the
above mentioned riteria is mostly used. The ra k is propagated in an angle of
c from the ra k tip.
In the maximum hoop stress or maximum prin ipal stress riterion, a ra k
is oriented perpendi ular to the dire tion of the maximum prin ipal stress. The
maximum ir umferential stress , often alled hoop stress, in the polar oordinate system around the ra k tip orresponds to the maximum prin ipal
stress and is given for a ra k propagating with onstant velo ity vc by:
KI I
KII II
=
fh (, vc ) +
fh (, vc )
2r
2r
(5.115)
where the fun tions fhI and fhII represent the angular variation of stress for
dierent values of ra k-tip speed vc . When the maximum hoop stress is larger
c
, then the ra k is propagated in the dire tion
equal a riti al hoop stress
c
perpendi ular to the maximum hoop stress. For pure mode I fra ture,
is
given by
Kc
c
= I
(5.116)
2r
with the fra ture toughness KIc that is obtained from experiments. The lo al
dire tion of the ra k growth is determined by the ondition that the lo al shear
stress is zero that leads to the ondition:
KI sinc + KII (3cosc 1) = 0
8 Note
that ra k initiation annot be modeled with the riteria mentioned here
(5.117)
128CHAPTER 5.
that results in the ra k propagation angle

!
p
2
KI KI2 + 8KII
c = 2arctan
4KII
(5.118)
Diagonalized mass matrix

Methods that are based on an expli it time integration generally use a diagonalized mass matrix that is obtained by a simple row-sum te hnique. A diagonalized mass matrix fa ilitates the solution of the system of equations enormously.
However, if additional degrees of freedom are introdu ed as in XFEM, it an
be shown that a standard row-sum te hnique will lead to an erroneous lumped
mass matrix9 . A row-sum te hnique for standard XFEM an be obtained based
on the assumption that for a rigid body motion, the dis rete kineti energy has
to be exa t. A lumped mass matrix an then be obtained by:
Z
1
m
2 del
mdiag =
(5.119)
nnodes mes(el ) el
where el is the element being onsidered, m is its mass, mes()el its length,
nnodes the number of nodes in and is the enri hment fun tion. If the
step fun tion is used, the mass matrix an be diagonalized by one of these two
pro edures:
X
Mlumped
=
Mconsistent
, or
IJ
II
J
Mconsistent
Mlumped
= m P IIconsistent
II
MIJ
(5.120)
One di ulty with a diagonalized mass matrix in an XFEM formulation is that
the riti al time step t tc = 2/max is redu ed drasti ally when a ra k
is lo ated lose to a node10 . Within the approa h proposed above, the riti al
time step is not diminished so drasti ally when the dis ontinuity is lose to a
node; drasti ally means by a fa tor around 2 ( ompared to the CFL riterion of
the element without dis ontinuity). Note that the lumping pro edure is slightly
dierent when the Heaviside fun tion is used as enri hment.
Example Let us onsider a one-dimensional element with two nodes. The
approximation is given by
uh (X) = N1 u1 + N1 1 a1 + N2 u2 + N2 2 a2
and the lumped mass matrix of the element by
m1 0
0
0 m2 0
Mlumped =
0
0 m3
0
0
0
9 Note
10 Note
0
0
0
m4
(5.121)
(5.122)
that for the Hansbo-Hansbo XFEM, a standard row sum te hnique is su ient
that max is the largest solution of det (K M) where K is the stiness matrix
and M the mass matrix
129
h
The oe ients mRi have to be determined su h that Ekin
= 0.5u T Mlumped u
equals Ekin = 0.5 el v2 d. Let us onsider that the element moves with a
in the same dire tion. Hen e, we set a equal zero and obtain
onstant velo ity u

2
h
(m1 + m2 )
Ekin
= 0.5 m1 u 21 + m2 u 22 = 0.5 u
(5.123)
2
h
= Ekin
that m1 = m2 = 0.5 m where m is
and obtain with Ekin = 0.5 m u
the mass of the element. Next let us onsider the separation of the element into
= a
1 (x). Thus, we set u equal zero and obtain
two parts, i.e. u

h
2 (m3 + m4 )
Ekin
= 0.5 m3 a 21 + m4 a 22 = 0.5 a
(5.124)
and
2
Ekin = 0.5 a
el
so that nally the mass m3 and m4 are

m
m3 = m4 =
2 mes()el
12 del
(5.125)
(5.126)
el
12 del
Let us now determine the minimal riti al time step for the one-dimensional
XFEM element and ompare it to the riti al time step of a standard element.
If the length of the element is denoted by l, the linear shape fun tion an be
given by
x
N1 (x) = 1
l
x
(5.127)
N2 (x) =
l
The onsistent mass matrix and the stiness matrix of the standard element is
given by

EA
1 1
1/3 1/6
MF E = A l
(5.128)
, KF E =
1 1
1/6 1/3
l
where E is the Young's modulus and A the ross se tion. The riti al time step
is easily omputed by
r
2
=l
tc,F E =
(5.129)
max
3E
With the lumped mass matrix

(5.130)
= 3tc,F E
E
(5.131)
Mlumped
=Al
FE
1/2 0
0 1/2
the riti al time step is

tlumped
c,F E
=l
130CHAPTER 5.
Now, let us study the riti al time step for the XFEM approximation. The
dis ontinuity is lo ated at position s and with the generalized step fun tion
entered in s, the approximation of the displa ement eld is
uh (x)
= N1 (x) u1 + N1 (x) S(x s) a1
+ N2 (x) u2 + N2 (x) S(x s)a2
(5.132)
and the onsistent mass matrix and stiness matrix
MXF EM
...
KXF EM
1/3
1/6
A l
2s2 2s + 1/3 2/3s3
1/6 s2 + 2/3s3
2s2 2s + 1/3 2/3s3
1/6 s2 + 2/3s3
1/3
1 2s
1/6
1
1
E A
1
1
=
l 1 2s 2s 1
2s 1 1 2s
1/6
1/3
...
1/6 s2 + 2/3s3
1/3 21
2
1/6 s + 2/3s3
1/3 2/3s3
(5.133)
1/6
2s 1
1/3
1 2s 2s 1
2s 1 1 2s
1
1
1
1
The lumped mass matrix for the XFEM approximation is
1 0 0 0
0 1 0 0
Mlumped
XF EM = 0.5 A l 0 0 1 0
0 0 0 1
(5.134)
(5.135)
The riti al time step for the XFEM-approximation depends on the lo ation
of the dis ontinuity s in the element. The smallest riti al time step is obtained when the dis ontinuity is lo ated at x = 0 or x = l. For the onsistent
XFEM mass matrix, the riti al time step goes to zero when the dis ontinuity approa hes 0 or l while the riti al time step at x = 0 and x = l for the
1 tlumped . Hen e, even when
lumped XFEM mass matrix is tlumped
c,F E
c,XF EM =
2
the dis ontinuity is lo ated very lose to a node, the riti al time step is not
destroyed. In ontrast, for the Hansbo-Hansbo approa h, a standard row sum
te hnique an be employed. However, the riti al time step will tend to zero
when the dis ontinuity approa hes a node. In pra ti e, a ertain minimum mass
is assigned to a node su h that the omputation an be pro eeded.
Limitations
For parti ular dis retizations and ra k ongurations, the XFEM approximation annot a urately represent the dis ontinuity in the near-tip displa ement
131
crack
crack
a)
b)
crack
c)
effective crack length
d)
Figure 5.23: a),b) Cra k length that approa h the lo al element size annot
be a urately represented by the standard XFEM approximation. Dots denote
single enri hed nodes and squares denote double (in our ase, the node will
ontain the enri hment of two ra k tips) enri hed nodes; ) the dashed line
shows the ee tive ra k length; d) even if no ra k tip enri hment is used, in
order to lose the ra k within a single element, no nodes have to be enri hed
with a step fun tion
132CHAPTER 5.
elds. Di ulties o ur for the non-physi al ase of too lose parallel ra ks
(within a single element) and when the extent of the ra k approa hes the support size of the nodal shape fun tions. Then, the asymptoti bran h fun tions
for ea h tip may extend beyond the length of the ra k, resulting in a non onforming approximation. Although the XFEM approximation is apable of
representing ra k geometries that are independent of element boundaries, it
also relies on the intera tion between the mesh and the ra k geometry to determine the sets of enri hed nodes. This leads to parti ular ra k ongurations
that annot be a urately be represented by eq. (5.87). Su h ases are shown
in gure 5.23. As the ra k size approa hes the lo al nodal spa ing, the set Sc
of nodes for the Heaviside or step enri hment is empty, gure 5.23b. Moreover,
node 1 for the ra king ase in gure 5.23a or nodes 1 to 4 for the ra king
ase in gure 5.23b, respe tively, ontain two bran h enri hments. Thus, the
standard approximationgets di ulties with this ra k onguration sin e the
dis ontinuous fun tion 2sin(/2) extends too far. This problem always arises
when one or more nodal supports ontain the entire ra k geometry. Usually,
this kind of problem arises whenever a ra k nu leates. Similar di ulties o ur
for approximations without any ra k tip enri hment, see gure 5.23d. In order
to lose the ra k within a single element, the set Sc is empty as well. One
solution is to rene the mesh lo ally su h that the hara teristi element size
falls below that of the ra k. An admissible ra k onguration is shown in
gure 5.24d.
Tra king the ra k path

There are numerous ways to tra k the ra k path and to represent the ra k
surfa e. The ra k surfa e an be represented expli itly, meaning by introdu ing another mesh for the ra k surfa e. Most ommonly, the ra k surfa e is
des ribed by pie ewise linear ra k segments though it is also possible to des ribe the ra k surfa e smoothly, e.g. with B-splines or NURBS (Non uniform
rational B-Spline). The ra k surfa e an also be des ribed impli itly with the
help of level sets or signed distan e fun tions11 . Cra k path tra king algorithms
an be lassied into three lasses: global methods, lo al methods and the level
set method.
The basi idea of the global ra k tra king pro edure is to dene a linear
thermal problem to be solved ea h time step of the original me hani al problem.
Therefore, two ve tor elds a and b are introdu ed that have to fulll the
following ondition:
a n0 = b n0 = 0
(5.136)
where n0 is the ra k normal in the initial onguration. The family of surfa es,
enveloping both ve tor elds a and b an be des ribed by a temperature-like
11 Note that if the same approximating fun tions are used for the level set and the physi al
properties, meshfree methods are well suited for des ribing urved ra ks due to their higher
ontinuous shape fun tions
133
crack
Figure 5.24: Admissible ra k representation

fun tion T(X) if
a 0 T = 0 T a = 0 in 0
b 0 T = 0 T b = 0 in 0
(5.137)
holds. Eq. (5.137) an be rephrased as an anisotropi heat ondu tion problem.

The drawba k of this method is that the heat ondu tion problem has to be
solved at every time step, that makes the method omputationally expensive.
With lo al ra k tra king algorithms, the alignment of the ra k surfa es
is enfor ed with respe t to its neighborhood. Lo al ra k tra king algorithms
are usually hara terized by re ursively " utting" elements and are espe ially
ee tive in three dimensions.
As mentioned earlier, the original level set method is not well suited to tra k
the ra k path. First, the zero level set must be updated behind the tip to take
into a ount the fa t that on e a material point is ra ked, it remains ra ked.
Se ond, the level set fun tions are not updated with the speed of an interfa e
in the dire tion normal to itself but with the speed at the ra k fronts. Hen e,
the equations for updating the interfa e
+ v = 0
t
(5.138)
annot be adopted to model ra k propagation. And third, the ra k is an open

surfa e that grows from its ra k front. Therefore, an additional level set has to
be introdu ed to be able to ompletely des ribe the ra k surfa e. This additional level set fun tion (at the ra k tip) is perpendi ular to the original level
set fun tion (and perpendi ular to the urrent ra k front) and hen e has to
be updated as well. The a ura y of the ra k surfa e representation depends
134CHAPTER 5.
on the dis retization (shape fun tions) of the level set fun tions. Usually, the
same dis retization is employed for the me hani al properties and the level set
fun tion that makes the method very attra tive and elegant from an implementational point of view. If highly urved ra ks o ur, linear nite element shape
fun tions will fail to represent the ra k surfa e exa tly. Note, that a reinitialization is ne essary after updating the level set fun tions to ensure that the zero
level set fun tion remains the signed distan e fun tion.
Cra k propagation with level sets an be modelled by dierent te hniques
that an be lassied into four groups. In the rst group, the level set is updated by the solution of dierential equations, similar to eq. (5.138) where the
level set fun tions are the unknowns. These methods require the dis retization
of the level set fun tions. The se ond group is dened on algebrai relations
between the oordinates of a given point, the oordinates of the ra k front and
the ra k advan e ve tor. The Ve tor level set method is dened in terms of
geometri transformations. In the ve tor level set method, the distan e to the
ra k surfa e is stored in addition to the signed distan e fun tion. This fa ilitates implementation sin e there is no need to solve a PDE to update the level
set. The last lass of methods are based on algebrai and trigonometri equations involving the initial value of the level set fun tions and the ra k advan e
ve tor. Some of them also require the des ription of the ra k front.
The level set te hnique ouples well for methods that use a ra k tip enri hment sin e
p they an be exploited to ompute the distan e to the ra k
front r = 2 + 2 where is the level set perpendi ular to and the angle = arctan(/), see gure 4.6. The level sets are espe ially useful for the
approximation of sin e they guarantee that = on the ra k surfa e sin e
= 0.
Non-mat hing meshes

As stated before, one big advantage of tra king the ra k path with level sets
is that no expli it representation of the ra k geometry is needed. The level set
an be approximated with the same shape fun tions as the me hani al properties. However, in ertain ases, this approximation might not be su ient. For
example, if the "usual" mesh is des ribed with linear nite elements12 , then the
urvature of the ra k might be not represented a urately enough for ertain
appli ations. In this ase, the level set is approximated with dierent13 shape
fun tions and a dierent mesh that is non- onforming with the "usual" nite
element mesh.
Then, the me hani al model will provide the ra k orientation and ra k
length. Update, reinitialization and orthogonalization an be done on the
"usual" mesh. The ra k update is made by solving the Hamilton-Ja obi equation at ea h point of the "level-set mesh". On e the new level set fun tions are
determined, a proje tion is needed to give their value on the "usual" mesh. In
12 i.e. with triangular elements
13 in that ase higher order shape
fun tions
135
order to keep proje tion errors small, the size of the level-set mesh should be of
the order of the "usual" mesh.
136CHAPTER 5.
Chapter 6
Other Methods
6.1 EXtended meshfree methods
EXtended meshfree methods basi ally have the same stru ture as the XFEM.
One major di ulty in extended meshfree methods is the losure of the ra k
at its ra k tip due to the strong overlapping of the meshfree shape fun tions,
see gure 6.1. The simplest way to ensure that the ra k loses at its tip is
with the use of bran h fun tions. Another opportunity is to modify the support
size of the nodes lose to the ra k tip su h that its domain of inuen e will
be ut ompletely as shown in gure 6.2. The domain of inuen e of node 1 is
ompletely ut by the ra k and it is enri hed by the sign fun tion. Cra k tip P
is lo ated inside the domain of inuen e of node 2 whi h is partially ut. The
shape fun tion I for the dis ontinuous displa ement is s aled down so that the
ra k tip is positioned at its edge as shown in the gure, i.e.
J (X) =
A (X) =
pT (X) A (X)1 D (XJ )

X
p(XJ ) pT (XJ ) W (
rJ ; h )
(6.1)
(6.2)
D (XJ ) =
p(XJ ) W (
rJ ; h )
(6.3)
in whi h the asterisk denotes the modi ation for the ra k tip and h is the
modied size of the domain of inuen e. Note that the shape fun tion for the
ontinuous displa ement remains un hanged.
The domain of inuen e of node 3 is also partially ut and the shape fun tion
may be shrunk, too. However, the node is not enri hed sin e it is very lose to
the ra k tip and the shape fun tion be omes very small ompared to others
after it is modied and the approximation for the dis ontinuous displa ement
eld be omes bumpy. Therefore, when the domain of inuen e of a node is
partially ut, we enri h the node if the support of the shape fun tion in ludes
at least one enri hed node after it is modied. Node 4 is not enri hed be ause
the shape fun tion be omes not ut by the ra k as it is modied.
137
138
CHAPTER 6. OTHER METHODS
Figure 6.1: a) XFEM enri hment su h that the ra k loses at its ra k tip and
b) the problemati of ra k losure in meshfree methods
One drawba k of the method is that the ra k appears to be shorter for
parti les lose to the ra k and the ra k tip. Instead of modifying the shape
fun tion of the node of whi h domain of inuen e is partially ut, we may
onsider the use of the Lagrange multiplier method. If only the sign fun tion
enri hment is used, there is the extension of dis ontinuity c,ext beyond the
ra k tip; see Fig. 6.3. To model the ra k, the dis ontinuity on c,ext should
vanish. Be ause the ondition should be satised along a line, not at a point,
the Lagrange multiplier must be dis retized, too. To avoid introdu ing another
nodes for the dis retization, the same shape fun tions as those partially ut by
the ra k an be used.
6.2 Embedded elements

In 1987, Ortiz et al. [25 modied the approximation of the strain eld to apture weak dis ontinuities in nite elements to improve the resolution of shear
bands. Therefore, they enri hed the strain eld to obtain the kinemati relations shown in gure 6.4a. Based on this idea, Belyts hko et al. [26 allowed
two parallel weak dis ontinuity lines in a single element, gure 6.4b, so that
the element was able to ontain a band of lo alized strain. Dvorkin et al. [27
were the rst who developed a method that was able to deal su essfully with
strong dis ontinuities in nite elements. The lass of embedded elements was
born. The name omes from the fa t that the lo alization zone is embedded in a
single element, see gure 6.4 . This way, ra k growth an be modeled without
remeshing. This lass of methods are mu h more exible than s hemes that
allow dis ontinuities only at element interfa es, and it eliminates the need for
ontinuous remeshing. As in XFEM, embedded elements introdu e additional
6.2. EMBEDDED ELEMENTS
139
Figure 6.2: De reasing the support size to lose the ra k at its ra k tip
unknowns into the variational formulation. However, the enri hment is on the
element level. This has the advantage, that the additional unknowns an be
ondensed on the element level, so that dis ontinuities an be aptured only
with very small hanges of the existing ode. The drawba k is that embedded
elements are less exible than XFEM. The ra k an only open pie ewise onstant. Moreover, the ra k an be propagated one element at a time. A typi al
ra k representation for embedded elements are shown in gure 6.5. Note, that
in embedded elements, a tip enri hment annot be employed as is XFEM1 sin e
the ra k opens pie ewise onstant.
The rst version of embedded elements is often alled stati al optimal symmetri (SOS) sin e tra tion ontinuity is fullled but it is not possible to apture
the orre t ra k kinemati s. It has been shown, that SOS formulations lead
to stress lo king, i.e. stresses are transmitted a ross the ra k even if the ra k
is wide open. The kinemati al optimal symmetri (KOS) version by Lofti and
Shing [28 ensures the orre t ra k kinemati s but violates the tra tion ontinuity ondition. Consequently, the riteria for the onset of lo alization written
in terms of stresses in the bulk are no longer equivalent to the same riteria
1 due
to the la k of a ra k tip
140
Figure 6.3: The dis ontinuity c,ext beyond ra k tip P when nodes are enri hed
by using only the sign fun tion.
written in terms of the tra tions on the dis ontinuity area; e.g. for the Rankine
riterion, the normal tra tion at the onset of lo alization should be equal to the
tensile strength and the shear tra tion should be zero. This annot be properly reprodu ed by the KOS formulation. The kinemati al and stati al optimal
non-symmetri (KSON) version of embedded elements [27 guarantees tra tion
ontinuity and the appropriate ra k kinemati s but leads to a non-symmetri
stiness matrix with all its disadvantage with respe t to solving the linearized
system of equations.
In embedded elements, the approximation of the displa ement eld is given
a)
b)
c)
Figure 6.4: Element with a) one weak dis ontinuity, b) two weak dis ontinuities,
) one strong dis ontinuity
141
6.2. EMBEDDED ELEMENTS
b)
a)
Figure 6.5: a) Pie ewise onstant ra k opening in embedded elements and b)

linear ra k opening for linear XFEM
by:
uh (X) =
(e)
J + Ms(e) (X) [[
NI (X) u
uI (X)]]
(6.4)
IS
is the usual approximation, u

is the enri hment and the superimposed
where u
"(e)" indi ates the enri hment on element level. In other words, only one additional degree of freedom in every oordinate dire tion is introdu ed per element.
Let us onsider a linear quadrilateral element for example. In XFEM, this element would have 4 additional degrees of freedom in every oordinate dire tion,
i.e. 8 additional degrees of freedom while in embedded elements only one additional degree of freedom2 in every oordinate dire tion is introdu ed. The
(e)
fun tion Ms is given by
0
(e)
/S
(e)
(e)
Hs
(e) S
(e)
Ms (X) =
(6.5)
+
N
e
(e) =
NI+ (X)
I=1
where Hs is the step fun tion a ting on the ra k line S and Ne+ is the number of
nodes of element (e) that belong to the domain +
0 , see gure 6.6. By standard
dierentiation, the dis ontinuous strain eld is obtained:
h (X) =
X
IS
S (e)

S
s
(e)
(e)
I )S (e) [[
(0 NI (X) u
uI (X)]] +
[[
uI (X)]] n
k
where the superimposed S denotes the symmetri part and the term
2 This
enri hment is inherent of the element
(6.6)
is a
(e)
s /k
142

(e)
regularized Dira delta fun tion and s is a ollo ation fun tion dened as

1 X Sek
(e)
s =
(6.7)
0 X
/ Sek
with the thi kness k of the lo alization band. Considering the equilibrium equation in elastostati s and with the trial and test fun tion of the stru ture of (6.4),
the dis rete equations in matrix form an be written as
#
"
ext
(e)
(e)
(e)
u
FI
Kuu Kuu
=
(6.8)
(e)
(e)
(e)
0
[[
uI ]]
Kuu Kuu
with
(e)
Kuu
(e)
Kuu
(e)
Kuu
(e)
Kuu
BT C B d0
d0
BT C B
=
=
T C B d0
B
T C B
d0
B
where C is the elasti ity tensor, B is the B-operator dened earlier,
(e)
0
x
(e)
(e) =
y
0(e)
(e)
n(e)
nx
= 0
ny
(6.9)
(6.10)
0
ny
nx
(6.11)
is the B-operator of the enri hment that depends on the embedded eleand B
6= B
,
ment formulation (SOS, KOS or KSON). For the KSON formulation, B
that will result in a non-symmetri stiness matrix. For the SOS and KOS
(e)
= B
. The elemental enri hment [[
formulation, B
uI ]] an be ondensed on
the element level:
i1
h
(e)
(e)
(e) (e)
[[
uI ]] = Kuu
(6.12)
Kuu u
(e)
uI ]] to the system of equations of the type

that leads with the expression for [[
u=f
K
(6.13)
= Kuu Kuu K1 Kuu

K
u
u
(6.14)
with
143
6.3. INTERFACE SEPARATION MODELS
PSfrag repla ements

S
Figure 6.6: Embedded element ut by a dis ontinuity S

The embedded element formulation is proposed in dierent format, starting from
the extended prin iple of virtual work over Hellinger Reissner to Hu-Washizu
variational prin iple, using an enhan ed assumed strain (EAS) or B-bar format.
6.3 Interfa e separation models

Standard nite elements will generally have di ulties to apture the ra k
kinemati s sin e they use ontinuous trial fun tions that are not parti ularly well
adapted for solutions with dis ontinuous displa ement elds. Therefore, it is not
surprising that the rst models apable of des ribing the ra k are the so- alled
interelement separation methods. In these methods, ra ks are only allowed
to develop along existing interelement edges. This endows the method with
omparative simpli ity, but an result in an overestimate of the fra ture energy
when the a tual ra k paths are not oin ident with element edges. The results
depend severely not only on the mesh size (and form of the hosen element) but
also the mesh bias that an be ompensated only by omputational expensive
remeshing. Furthermore, it has been noted that the solutions sometimes depend
signi antly on mesh renement. This sensitivity has been mollied by adding
randomness to the strength, as in Zhou and Molinari [29 and Espinosa et al. [30,
though su h orre tions are not yet understood. Nevertheless, many interesting
problems have been studied by this method.
There are basi ally two types of interlement separation models. The Needlemangroup introdu es (initially ina tive) ohesive surfa es at the beginning of the
omputation. This method is from the implementational point very simple but
suers from all the drawba ks mentioned above. The interelement separation
models of Ortiz use remeshing and adaptive insertion of ohesive segments in
the ourse of the simulation. This te hnique is more a urate but omputa-
144
tionally expensive (be ause of the remeshing) and from an implementational

point of view more burdensome. Moreover, errors are introdu ed be ause of the
mapping that is needed during remeshing.
Bibliography
[1 T. Belyts hko, W. K. Liu, and B. Moran. Nonlinear Finite Elements for
Continua and Stru tures. John Wiley and Sons, Chi hester, 2000.
[2 T. Belyts hko, Y. Krongauz, D. Organ, M. Fleming, and P. Krysl. Meshless
methods: An overview and re ent developments. Computer Methods in
Applied Me hani s and Engineering, 139:347, 1996.
[3 A. Huerta, Belyts hko T., Fernandez-Mendez S., and Rab zuk T. En y lopedia of Computational Me hani s, hapter Meshfree Methods. John Wiley
and Sons, 2004.
[4 R.A. Gingold and J.J. Monaghan. Smoothed parti le hydrodynami s:theory and appli ations to non-spheri al stars. Mon. Not. R. astr. So .,
181:375389, 1977.
[5 G.R. Johnson and S.R. Beissel. Normalized smoothing fun tions for sph
impa t omputations. International Journal for Numeri al Methods in Engineering, 39:27252741, 1996.
[6 Y. Krongauz and T. Belyts hko. Consistent pseudo derivatives in meshless
methods. Computer Methods in Applied Me hani s and Engineering, 146:
371386, 1997.
[7 P.W. Randles and L.D. Libersky. Re ent improvements in sph modeling
of hypervelo ity impa t. International Journal of Impa t Engineering, 20:
525532, 1997.
[8 W. K. Liu, S. Jun, and Y. F. Zhang. Reprodu ing kernel parti le methods.
20:10811106, 1995.
[9 C.A.M Duarte and J.T. Oden. An h-p adaptive method using louds.
Computer Methods in Applied Me hani s and Engineering, 1996.
[10 T. Strouboulis, K. Copps, and I. Babuka. The generalized nite element
method: An example of its implementation and illustration of its performan e. International Journal for Numeri al Methods in Engineering, 47
(8):14011417, 2000.
145
146
BIBLIOGRAPHY
[11 J.S. Chen, C.T. Wu, S. Yoon, and Y. You. A stabilized onforming nodal
integration for galerkin meshfree-methods. International Journal for Numeri al Methods in Engineering, 50:435466, 2001.
[12 C.T. Dyka and R.P. Ingel. An approa h for tensile instability in smoothed
parti le hydrodynami s. Computers and Stru tures, 57:573580, 1995.
[13 P.W. Randles and L.D. Libersky. Normalized sph with stress points. International Journal for Numeri al Methods in Engineering, 48:14451462,
2000.
[14 P. Krysl and T. Belyts hko. Element-free galerkin method: Convergen e
of the ontinuous and dis ontinuous shape fun tions. Computer Methods
in Applied Me hani s and Engineering, 148:257277, 1997.
[15 D. Organ, M. Fleming, T. Terry, and T. Belyts hko. Continuous meshless approximations for non onvex bodies by dira tion and transparen y.
Computational Me hani s, 18:225235, 1996.
[16 T.G. Terry. Fatigue ra k propagation modeling using the element free
galerkin method. Master thesis, Northwestern University, 1994.
[17 T. Belyts hko and M. Fleming. Smoothing, enri hment and onta t in
the element-free galerkin method. Computers and Stru tures, 71:173195,
1999.
[18 M. Fleming, Y.A. Chu, B. Moran, and T. Belyts hko. Enri hed element-free
galerkin methods for ra k tip elds. International Journal for Numeri al
Methods in Engineering, 1997.
[19 A. Hansbo and P. Hansbo. A nite element method for the simulation of
strong and weak dis ontinuities in solid me hani s. Computer Methods in
Applied Me hani s and Engineering, 193(33-35):35233540, 2004.
[20 J. Chessa, H. Wang, and T. Belyts hko. On the onstru tion of blending
elements for lo al partition of unity enri hed nite elements. International
Journal for Numeri al Methods in Engineering, 57(7):10151038, 2003.
[21 G. Zi, J.-H. Song, E. Budyn, S.-H. Lee, and T. Belyts hko. A method
for grawing multiple ra ks without remeshing and its appli ation to fatigue ra k growth. Modelling and Simulation in Materials S ien e and
Engineering, 12(1):901915, 2004.
[22 G.C. Shih. Strain energy density fa tor applied to mixed-mode ra k problems. International Journal of Fra ture, 10:305321, 1974.
[23 C-H Wu. Fra ture under ombined loads by maximum energy release rate
riterion. Journal of Applied Me hani s, 45:553558, 1978.
[24 R.V. Goldstein and R.L. Salganik. Brittle fra ture of solids with arbitrary
ra ks. International Journal of Fra ture, 10:507527, 1974.
BIBLIOGRAPHY
147
[25 M. Ortiz, Y. Leroy, and A. Needleman. Finite element method for lo alized
failure analysis. Computer Methods in Applied Me hani s and Engineering,
61(2):189214, 1987.
[26 T. Belyts hko, J. Fish, and B. Englemann. A nite element method with
embedded lo alization zones. Computer Methods in Applied Me hani s and
Engineering, 70:5989, 1988.
[27 E. N. Dvorkin, A. M. Cuitino, and G. Gioia. Finite elements with displa ement interpolated embedded lo alization lines insensitive to mesh size and
distortions. International Journal for Numeri al Methods in Engineering,
30:541564, 1990.
[28 H.R. Lofti and P.B. Shing. Embedded representation of fra ture in on rete
with mixed nite elements. International Journal for Numeri al Methods
in Engineering, 38:13071325, 1995.
[29 F. Zhou and J.F. Molinari. Dynami ra k propagation with ohesive elements: a methodolgy to address mesh dependen e. International Journal
for Numeri al Methods in Engineering, 59(1):124, 2004.
[30 H.D. Espinosa, P.D. Zavattieri, and S. Dwidivedi. Adaptive fem omputation of geometri and material nonlinearities with appli ation to brittle
failure. Spe ial issue of Me hani s and Materials, 29(3-4):275305, 1998.

Lecture Xfem Meshfree

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Lecture Xfem Meshfree

Загружено:

Авторское право:

Доступные форматы

EXtended Finite Element and

Prof. Dr.-Ing. T. Rab zuk

Completeness, onsisten y and Partition of Unity . .

2.10.3 Bridging domain oupling method . . . . . . . . . . . . .

3 Kinemati s of strong and weak dis ontinuities

4 Strong dis ontinuities in meshfree methods

4.1 A histori al overview . . . . . . . . . . . . . . . . . .

5 The eXtended nite element method (XFEM)

6.1 EXtended meshfree methods . . . . . . . . . . . . . . . . . . . . 137

Preliminaries and Notations

we refer to quantities in the initial onguration

v(X, t) vi (x, t) xi (X, t)

material time derivative

Subs ripts/Supers ripts

symmetri part of a tensor or Shepard fun tion or stress point or

Lagrange multiplier or Lagrange multiplier estimate or eigenvalue

enri hment fun tion

or in indi ial notation

CHAPTER 2. MESHFREE METHODS

ing onditions for a onstant and linear eld are given by

J,i (X) XJj = ij

or in indi ial notation

2.2 Conservation, stability and onvergen e

that requires that

I (XJ ) = 0 for arbitrary stress states that in turn

requires zero-order omplete shape fun tions.

ijm mj (XJ ) wJ = 0 (2.14)

where C is a onstant and h is the dilation parameter in meshfree methods1

CHAPTER 2. MESHFREE METHODS

Support size of particle I

Figure 2.1: Support size of the kernel fun tion

2.4 The weighting fun tion

2.4. THE WEIGHTING FUNCTION

or the kernel fun tion has re tangular support

2.4.2 Spe i weighting fun tions

where D is the dimension, z = r/h0 and

C is a onstant depending on the

that the support size is proportional to the dilation parameter

CHAPTER 2. MESHFREE METHODS

Figure 2.2: Meshfree (SPH) approximation fun tions and approximation of

2.4. THE WEIGHTING FUNCTION

where c, R and q are shape parameters.

2.4.3 Eulerian and Lagrangian kernels

CHAPTER 2. MESHFREE METHODS

then the a elerations are

2.5 Spe i meshfree approximations

2.5. SPECIFIC MESHFREE APPROXIMATIONS

2.5.1 Smooth Parti le Hydrodynami s (SPH)

is fullled by denition, see eq. (2.16). Linear ompleteness requires that

Noting that eq. (2.37) requires that

and subtra ting eq. (2.39) from eq. (2.38) leads to

CHAPTER 2. MESHFREE METHODS

Integration by parts gives

The dis rete SPH-form

that VJ0 is the quadrature weight

2.5. SPECIFIC MESHFREE APPROXIMATIONS

(uJ uI ) 0 W (XI XJ , h0 ) VJ0

2.5.2 Corre ted SPH forms

CHAPTER 2. MESHFREE METHODS

GiI (X) uI (t)

where GI is a linear ombination of the Shepard fun tions:

where I is the identity matrix and

a(X) 0 WIS (X) uI (t)

2.5. SPECIFIC MESHFREE APPROXIMATIONS

5 The eXtended nite element method (XFEM)

we refer to quantities in the initial onguration

ing onditions for a onstant and linear eld are given by

2.4.2 Spe i weighting fun tions

2.5 Spe i meshfree approximations

is fullled by denition, see eq. (2.16). Linear ompleteness requires that

If we assume that GXI satises

with K(X, Y) = C(X, Y)W (X Y) and where C(X, Y) is dened su h that

Figure 2.5: Data tting using least square method

Let us dene a line approximating these points. In this ase, we have

and solving for a gives nally