A meshfree method based on the peridynamic model

of solid mechanics
S.A. Silling
a,
*
, E. Askari
b
a
Department of Computational Physics, MS-0378, Sandia National Laboratories, Albuquerque, NM 87185-0378, United States
b
Math Group, Boeing Phantom Works, P.O. Box 3707, MC 7L-25 Seattle, WA 98124-2207, United States
Accepted 28 November 2004
Available online 21 March 2005
Abstract
An alternative theory of solid mechanics, known as the peridynamic theory, formulates problems in terms of integral
equations rather than partial dierential equations. This theory assumes that particles in a continuum interact with each
other across a nite distance, as in molecular dynamics. Damage is incorporated in the theory at the level of these two-
particle interactions, so localization and fracture occur as a natural outgrowth of the equation of motion and consti-
tutive models. A numerical method for solving dynamic problems within the peridynamic theory is described. Accuracy
and numerical stability are discussed. Examples illustrate the properties of the method for modeling brittle dynamic
crack growth.
2005 Elsevier Ltd. All rights reserved.
1. Introduction
Numerical prediction of crack growth is a long-
standing problem in computational mechanics. The dif-
culty inherent in this problem arises from the basic
incompatibility of cracks with the partial dierential
equations that are used in the classical theory of solid
mechanics. The spatial derivatives needed for these par-
tial dierential equations to make sense do not exist on a
crack tip or surface. Therefore, any numerical method
derived from these equations inherits this diculty in
modeling cracks. In spite of the progress that has been
made in developing meshfree methods aimed at model-
ing fracture, for example [16], such meshfree techniques
generally require some special method of evaluating
the spatial derivatives on each crack surface. They also
require supplemental relations that govern the initiation
of cracks, as well as their growth velocity and direction.
These relations must be applied along each crack tip,
leading to inherent complexity of the method, particu-
larly when multiple cracks occur and interact in three
dimensions. It is also possible to construct constitutive
models that lead asymptotically to localization in a con-
tinuum, for example [7], but these techniques do not en-
tirely avoid the need for special treatment of a crack
after it forms.
As an attempt at improving this situation, a theory of
solid mechanics has been proposed that does not require
spatial derivatives to be evaluated within a body [813].
This theory, known as the peridynamic theory, instead
uses integral equations. The objective is to reformulate
the basic mathematical description of solid mechanics
in such a way that the identical equations hold either
on or o of a discontinuity such as a crack. The pres-
ent paper discusses a numerical approach to solving
0045-7949/$ - see front matter 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2004.11.026
*
Corresponding author. Tel.: +5058443973.
E-mail address: sasilli@sandia.gov (S.A. Silling).
Computers and Structures 83 (2005) 15261535
www.elsevier.com/locate/compstruc
initial-value problems within this theory, with particular
application to fracture and failure.
2. Peridynamic model of a continuum
The peridynamic theory may be thought of as a con-
tinuum version of molecular dynamics. The acceleration
of any particle at x in the reference conguration at time
t is found from
qux; t
_
H
x
fux
0
; t ux; t; x
0
xdV
x
0 bx; t;
1
where H
x
is a neighborhood of x, u is the displacement
vector eld, b is a prescribed body force density eld, q is
mass density in the reference conguration, and f is a
pairwise force function whose value is the force vector
(per unit volume squared) that the particle x
0
exerts on
the particle x. In the following discussion, we denote
the relative position of these two particles in the refer-
ence conguration by n:
n x
0
x; 2
and their relative displacement by g
g ux
0
; t ux; t: 3
Note that g + n represents the current relative position
vector between the particles.
The direct physical interaction (which occurs through
unspecied means) between the particles at x and x
0
will
be called a bond, or, in the special case of an elastic inter-
action to be dened below, a spring. The concept of a
bond that extends over a nite distance is a fundamental
dierence between the peridynamic theory and the clas-
sical theory, which is based on the idea of contact forces
(interactions between particles that are in direct contact
with each other).
It is convenient to assume that for a given material
there is a positive number d, called the horizon, such that
j n j> d )fg; n 0 8g: 4
In other words, the particle x cannot see beyond this
horizon. For the remainder of this discussion, H
x
will
denote the spherical neighborhood of x in R with radius
d (Fig. 1).
The pairwise force function f is required to have the
following properties:
fg; n fg; n 8g; n; 5
which assures conservation of linear momentum, and
n g fg; n 0 8g; n; 6
which assures conservation of angular momentum. The
latter equation means that the force vector between
these particles is parallel to their current relative position
vector.
A material is said to be microelastic if the pairwise
force function is derivable from a scalar micropotential
w:
fg; n
ow
og
g; n 8n; g: 7
The micropotential is the energy in a single bond and
has dimensions of energy per unit volume squared.
The energy per unit volume in the body at a given point
(i.e., the local strain energy density) is therefore found
from
W
1
2
_
Hx
wg; ndV
n
: 8
The factor of 1/2 appears because each endpoint of a
bond owns only half the energy in the bond.
If a body is composed of a microelastic material,
work done on it by external forces is stored in recover-
able form in much the same way as in the classical the-
ory of elasticity. Furthermore, it can be shown that the
micropotential depends on the relative displacement vec-
tor g only through the scalar distance between the de-
formed points. Thus, there is a scalar-valued function
^ w such that
^ wy; n wg; n 8n; g y j g n j : 9
Therefore, the interaction between any two points in a
microelastic material may be thought of as an elastic
(and possibly nonlinear) spring. The spring properties
may depend on the separation vector n in the reference
conguration.
Anisotropy may be included in the microelastic re-
sponse through this dependence on the direction of n.
Fig. 1. Each point x in the body interacts directly with points in
the sphere H
x
through bonds.
S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535 1527
For example, in a model of a unidirectional ber-rein-
forced composite material in which the bers are parallel
to a given unit vector g, one might take
^ wy; n b
2
^ w
fiber
y; j n j 1 b
2
^ w
matrix
y; j n j;
b g n= j n j; 10
which explicitly supplies the dependence of bond energy
on the bond direction.
Combining (7) and (9) and dierentiating the latter
with respect to the components of g leads to
fg; n
n g
j n g j
f j n g j; n 8n; g; 11
where f is the scalar-valued function dened by
f y; n
o^ w
oy
y; n 8y; g: 12
This satises the requirements (5) and (6) provided
^ wy; n ^ wy; n 8y; n; 13
which will henceforth be assumed.
The relation (11), together with the equation of mo-
tion (1), essentially contain the totality of the peridy-
namic model for a nonlinear microelastic material. In
particular, note that the issue of how to treat rigid rota-
tion does not arise in this formulation because y is
invariant under rotation of the body. Similarly, objectiv-
ity of a constitutive model is not an issue in this
approach.
A linearized version of the theory for a microelastic
material takes the form
fg; n Cng 8n; g 14
where C is the materials micromodulus function, whose
value is a second-order tensor given by
Cn
of
og
0; n 8n: 15
This function inherits the following requirement from
(5):
Cn Cn 8n: 16
The properties of C are discussed in detail in [8].
Weckner [13] observed that certain elements of the
linear peridynamic theory just described, particularly
in one dimension, have been reported in earlier inves-
tigations by Kunin [14] and by Rogula [15]. These
methods apparently were proposed as a method for
the study of the properties of linear elastic materials with
microstructure rather than as a more general technique
for modeling deformation and failure of engineering
structures. It should also be noted that the peridynamic
model is a type of nonlocal theory, in the sense that par-
ticles separated by a nite distance can interact directly.
To draw a distinction between this and other nonlocal
theories, note that most nonlocal theories involve a
requirement for averaging some measure of strain within
some neighborhood of a material particle. The present
method diers from such nonlocal theories in that it
does not use strain, which by denition involves partial
derivatives of displacement. A nonlocal theory that has
been successfully applied to crack tip problems was pro-
posed by Eringen [16]. Another nonlocal approach that
has been developed for materials like concrete is de-
scribed by Ozbolt and Bazant [17].
3. Constitutive modeling
With a view toward modeling spontaneous fracture
in a deforming body, a notion of failure will be included
in the nonlinear microelastic material model previously
described. It will be assumed for simplicity that the
scalar bond force f depends only on the bond stretch,
dened by
s
j n g j j n j
j n j

y j n j
j n j
: 17
Thus s is positive when the bond is in tension. Note that
such a material is isotropic, since there is no dependence
of f on the direction of n.
The simplest way to introduce failure into a constitu-
tive model is by allowing bonds (springs) to break when
they are stretched beyond a predened limit. After bond
failure, there is no tensile force sustainable in the bond.
Once a bond fails, it is failed forever (there is no provi-
sion for healing of a failed bond); this makes the
model history dependent. For purposes of illustration,
consider the prototype microelastic brittle (PMB) mate-
rial dened by
f yt; n gst; nlt; n; 18
where g is the linear scalar-valued function given by
gs cs 8s; 19
where c is a constant and l is a history-dependent scalar-
valued function that takes on values of either 1 or 0
lt; n
1 if st
0
; n < s
0
for all 0 6 t
0
6 t;
0 otherwise:
_
20
Here, s
0
is the critical stretch for bond failure, which for
the moment will be assumed constant (Fig. 2). Note that
although the PMB material is isotropic in its virgin state,
breakage of bonds oriented in some particular direction
will lead to anisotropy in subsequent response. One
advantage of introducing failure at the bond level is that
it leads to an unambiguous notion of local damage at a
point, which is dened as
ux; t 1
_
H
x
lx; t; n dV
n
_
Hx
dV
n
21
1528 S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535
in which x is now included as an argument of l as a re-
minder that it is a function of position in the body. Note
that 0 6 u 6 1, with 0 representing virgin material, and
1 representing complete disconnection of a point from
all of the points with which it initially interacted.
Because broken bonds are unable to sustain tensile
load, the onset of bond breakage at some point in a
body leads to softening material response. This may
cause an evolution of damage in which broken bonds
coalesce into a surface that becomes a fracture, with
extensive bond breakage across the surface. (The precise
mathematical condition for material instability with re-
gard to damage evolution is an open issue at this point.
Such a criterion appears to be more complex in the pres-
ent setting than loss of ellipticity in PDEs, especially be-
cause of the directional nature of damage due to bond
failure introduced through l(x, t, n).)
Evidently there are only two parameters that charac-
terize the PMB material, a spring constant c and the
critical bond stretch s
0
. The value of c corresponding
to a given bulk modulus k is found as follows. Consider
a large homogeneous body under isotropic extension,
i.e., s is constant for all n, and g = sn. Dening n = jnj
and g = jgj, we have g = sn. Using the denition of the
micropotential (7), since f = cs = cg/n, it follows that
w = cg
2
/2n = cs
2
n/2. Then, applying (8) leads to
W
1
2
_
H
wdV
n

1
2
_
d
0
cs
2
n
2
_ _
4pn
2
dn
pcs
2
d
4
4
:
22
This is required to equal the strain energy density in the
classical theory of elasticity for the same material and
the same deformation, W = 9ks
2
/2. Combining this with
(22) leads to the spring constant in the PMB material
model [18]
c
18k
pd
4
: 23
The critical stretch for bond failure, s
0
, may also be re-
lated to a measurable quantity by considering a planar
fracture surface in the interior of a large homogeneous
body. To completely separate the two halves of the body
across the fracture surface requires breaking all the
bonds that initially connected points in the opposite
halves. Let w
0
(n) denote the work required to break a
single bond,
w
0
n
_
s
0
0
gsnds; n j n j; 24
where the relation dg = nds has been used. In the present
case of the PMB material, (24) leads to w
0
n cs
2
0
n=2.
The work G
0
required to break all the bonds per unit
fracture area is then found from
G
0

_
d
0
_
2p
0
_
d
z
_
cos
1
z=n
0
cs
2
0
n=2n
2
sin /d/dndhdz:
25
(Refer to Fig. 3 for an explanation of this computation.)
After evaluation of these integrals, the energy per unit
fracture area for complete separation of the two halves
of the body is found to be
G
0

pcs
2
0
d
5
10
: 26
Since the energy release rate is a measurable quantity for
brittle materials, (26) allows s
0
to be evaluated under the
Fig. 2. Bond force as a function of bond stretch in the PMB
material model.
Fig. 3. Evaluation of fracture energy G
0
. For each point A
along the dashed line, 0 6 z 6 d, the work required to break the
bonds connecting A to each point B in the spherical cap is
summed by the integrals in Eq. (25) using a spherical coordinate
system centered at A.
S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535 1529
assumptions of complete separation of the fracture sur-
faces and of the absence of other dissipative mechanisms
near the crack tip. Solving (26) for s
0
and using (23)
leads to
s
0

10G
0
pcd
5
_

5G
0
9kd
_
: 27
This determines s
0
for the PMB material under the
assumptions stated above.
For some real materials, it is an oversimplication to
assume that the critical stretch for failure of a given
bond is completely independent of conditions in all
other bonds. For example, better agreement with exper-
imental results on fracture in glass is obtained when the
critical stretch s
0
in (20) is allowed to depend on s
min
, the
current minimum stretch among all bonds connected to
a given material point
s
0
t s
00
as
min
t; s
min
t min
n
yt j n j
j n j
_ _
;
28
where s
00
and a are constants, with a typically on the or-
der of 1/4. Thus, compressive (negative) strains trans-
verse to the direction of maximum stretch increase the
critical stretch for bond failure. Other dependencies of
s
0
could be introduced to model various eects that
inuence fracture behavior, such as environmental ef-
fects and manufacturing defects. Time-dependent behav-
ior of s
0
can be introduced to model such eects as aging
and fatigue.
Interfaces between materials are characterized
through the properties of bonds that connect points in
the dierent materials. The properties of these interface
bonds can be chosen independently of the bonds within
the individual materials. For example, a weak interface
would be modeled by reducing s
0
for the bonds across
the interface.
The model described above allows only one elastic
constant, called c in the case of the PMB material,
whereas linear elastic materials in the classical theory
are characterized by two such constants. This dierence
occurs because an elastic solid that involves only two-
particle interactions (a Cauchy crystal) always has a
Poisson ration of 1/4. A renement of the peridynamic
theory that allows dependence of strain energy density
on local volume change in addition to two-particle inter-
actions removes this restriction [8] but has not yet been
implemented in the numerical method described in the
remainder of this paper. Also under development are
material models that incorporate permanent deforma-
tion at the bond level to model plastic response.
Another aspect of constitutive modeling in the peri-
dynamic theory that has no analogue in the classical the-
ory is the dependence of f on n. For the virgin PMB
material, it was assumed for convenience that f depends
only on s, which implies a particular dependence on n.
However, many alternatives are possible, and it is shown
in [8] that the nature of this dependence aects the dis-
persion relation of the material for linear waves. This
can potentially be exploited in nanoscale modeling of
materials.
4. Numerical method
The region is discretized into nodes, each with a
known volume in the reference conguration. Taken to-
gether, the nodes form a grid. The method is meshfree in
the sense that there are no elements or other geometrical
connections between the nodes. For simplicity, the de-
tails of the method will be discussed below using the lin-
earized version of the peridynamic theory applied to a
homogeneous body. The discretized form of the equa-
tion of motion (1) replaces the integral by a nite sum:
qu
n
i

p
fu
n
p
u
n
i
; x
p
x
i
V
p
b
n
i
; 29
where f is supplied by (11), n is the time step number,
and subscripts denote the node number, so that
u
n
i
ux
i
; t
n
: 30
V
p
is the volume of node p. To preserve the dimensions
of quantities used in the threee-dimensional expression
(1), it is helpful to assume that the one-dimensional
problem refers to a bar of constant cross-sectional area
A; thus V
p
= ADx, where Dx is the grid spacing, which
is assumed to be constant. The sum in (29) is taken over
all nodes p such that jx
p
x
i
j 6 d.
The discretized form of the linearized peridynamic
model is
qu
n
i

p
Cx
p
x
i
u
n
p
u
n
i
V
p
b
n
i
; 31
where C is dened in (15).
An explicit central dierence formula is used for
acceleration in either (29) or (31):
u
n
i

u
n1
i
2u
n
i
u
n1
i
Dt
2
; 32
where a constant time step size Dt is assumed.
Truncation error associated with the discretized
model (31) can be analyzed in one dimension, assuming
a homogeneous body:
qu
n
i

p
Cx
p
x
i
u
n
p
u
n
i
V
p
b
n
i
; 33
where the micromodulus function is now a scalar-valued
function C.
Under these assumptions, to derive the truncation er-
ror associated with (33), note that the equation of
motion
1530 S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535
qux
i

_
d
d
Cnux
i
n ux
i
Adn bx
i
34
may be written without approximation in the form
qux
i
A

p
_
Dx=2
Dx=2
Cx
p
x
i
nux
p
n
ux
i
dn bx
i
: 35
(The argument t is dropped from u here to shorten the
notation.) Under the assumption that the functions C
and u are twice continuously dierentiable, we can ex-
pand C near x
p
x
i
and u near x
p
using the rst few
terms of a Taylor series:
Cx
p
x
i
n Cx
p
x
i
C
0
x
p
x
i
n On
2
;
36
ux
p
n ux
p
u
0
x
p
n On
2
: 37
Combining (35)(37), and integrating leads to the
conclusion
qux
i

p
Cx
p
x
i
ux
p
ux
i
V
p
bx
i
ODx
2
;
38
where use has been made of the fact that terms in the
integrand that are linear in n integrate to 0. The error
in the time dierence formula (32) is well known to be
O(Dt
2
), so the conclusion is that the overall error in
(33) is O(Dx
2
) + O(Dt
2
).
However, when discontinuities in C or u are present,
the assumption made prior to (36) that these functions
are twice continuously dierentiable no longer holds.
In fact, the error changes under these conditions to
O(Dx) + O(Dt
2
). The convergence study given in Section
5 illustrates the resulting linear rate of convergence in a
crack growth problem, which necessarily contains dis-
continuities in u.
A stability condition for the numerical scheme (33)
will now be derived. Applying the standard assumption
for a von Neumann stability analysis [19], let
u
n
i
f
n
exp ji

1
p
_ _
; 39
where j is a positive real number and f is a complex num-
ber. The objective of the analysis is to determine a condi-
tion on Dt such that jfj 6 1 for all j. This condition is
necessary for waves not to grow unboundedly over time.
Using the notation q = p i and C
q
= C(x
p
x
i
), the
above assumptions when applied to (33) imply
q
Dt
2
f 2 f
1

1
q1
AC
q
exp jq

1
p
_ _
1
_ _
Dx

1
q1
2AC
q
cos jq 1Dx; 40
where the fact that C is an even function has been used
in the last step. Now dene
M
j

1
q1
AC
q
1 cos jqDx: 41
Then, using (40) and (41), and the quadratic formula to
solve for f
f 1
M
j
Dt
2
q

1
M
j
Dt
2
q
_ _
2
1

42
From this it is follows that the requirement jfj 6 1
implies:
Dt <

2q=M
j
_
8j: 43
To ensure that (43) holds for all wavenumbers j, note
that from (41)
M
j
6 2ADx

1
q1
C
q
44
hence the inequality (43) is satised whenever
Dt <

q
ADx

1
q1
C
q

: 45
A more useful form of (45) is
Dt <

2q

p
V
p
C
ip

; 46
where C
ip
= C(x
p
x
i
). This form of the stability condi-
tion does not require the nodes to be numbered sequen-
tially and is applicable to any number of dimensions by
setting C
ip
= jC(x
p
x
i
)j. Finally, for nonlinear material
models, the appropriate form of the stability condition is
(46) with C
ip
= jof/ogj. In this case it is important to ap-
ply a safety factor less than one to the stable time step
to account for possible nonlinear material response that
would make the estimated stable time step based on the
linear analysis above too large.
Ainteresting feature of (46) is that for a givend, the sta-
ble Dt is only weakly dependent on Dx. This is in contrast
to the famous CourantFriedrichsLewy condition that
controls stability in conventional wavecodes, which re-
quires that Dt = O(Dx) as Dx !0. However, it turns out
that the dependence of Con d implies that as d is reduced,
Dt = O(d). So, the maximum stable time step tends to be
limited by the horizon rather than the grid spacing.
The value of d may depend on the physical nature of
the application being modeled. For example, at the
nanoscale, d would be determined by the distance over
which physical interactions between atoms or molecules
occur. However, in macroscale calculations, d can be
chosen according to convenience, since as shown
previously, the parameters such as c that determine the
bulk elastic properties of the material can be tted to
S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535 1531
experimental data for any value of d. (See, for example,
(23).) In practice for macroscale modeling, the value
d = 3Dx usually works well. Values much smaller than
this typically result in undesirable grid eects (cracks
grow along the rows or columns of the grid). Values
much larger than this may result in excessive wave dis-
persion and require very large computer run times.
Boundary conditions are somewhat dierent in the
peridynamic method than in the classical theory. As
shown in [8], the variational formulation of the peridy-
namic equations does not lead to natural boundary con-
ditions, which in the classical theory are traction
boundary conditions. Instead, forces at the surface of
a body must be applied as body forces b within some
layer under the surface. In practice, this layer usually
has thickness d. Similarly, displacement boundary con-
ditions must be prescribed within a layer of nite thick-
ness under the surface.
Contact may be treated using short-range repulsive
forces that prevent nodes from getting too close. These
short-range forces are independent of the positions of
the nodes in the reference conguration, so that contact
is treated in a consistent way even under large deforma-
tion. This approach avoids the need for dening contact
surfaces, as is required by many nite-element codes.
The discretized equation of motion (29) includes non-
linearity and damage at the bond level, since these are in-
cluded in the constitutive model f. Recall that the history
dependent variable lin(20) is denedfor eachbond rather
than merely for each node. Since the number of bonds is
much larger than the number of nodes, computer memory
requirements may be minimized by using a bit variable to
store values of l. Cracks initially present in a body are
modeled by breaking each bond (the line segment con-
necting a pair of nodes) that intersects the surface of the
crack. Here, breaking a bond means setting the corre-
sponding bit variable l to 0 rather than 1 initially.
Although the above analysis assumes a regular lattice
of nodes, the method is not restricted to such structured
grids. Irregular grids may be used, the only complication
being the need to dene a volume associated with each
node. This may be accomplished using computational
geometry techniques such as Delaunay triangulation or
Voronoi diagrams [20]. A less accurate but much simpler
approach is to set the nodal volume based on the dis-
tances of a given node to some small set of its nearest
neighbors. Using irregular grids, complex geometries
may be modeled, and the full advantages of a meshfree
method are realized.
5. Example problem: Convergence in a fracture
calculation
In this section, the convergence of the method as grid
spacing is reduced is demonstrated for a nonlinear
problem that includes damage and fracture. Crack initi-
ation, growth, and arrest are all involved in this prob-
lem, providing insight into convergence of the
numerical method for practical applications involving
dynamic fracture.
Consider a thick square plate (Fig. 4) of side length
50 mm, containing a center crack of initial length
2a
0
= 10 mm. The material density is 8000 kg/m
3
, and
Youngs modulus is 192 GPa. A constant horizon
d = 1.6 mm is assumed, which is slightly greater than
3Dx for the case of the coarsest grid studied. The
PMB material model is used with a constant value of
the critical stretch for bond failure, s
0
= 0.02. Velocity
boundary conditions are applied as shown in Fig. 5
consisting of a tensile square pulse followed by a
compressive square pulse. The velocities are applied
symmetrically on the horizontal edges of the plate within
strips of width d. The rst pulse creates tensile stress
waves that move toward the crack, creating mode-I
loading at the crack tips. The duration of the tensile
pulse is 5.0 ls, during which interval the crack grows.
The growth is straight because of the symmetry of the
problem. The tensile pulse is followed by a compressive
pulse that stops the crack growth. Fig. 6 shows contours
of elastic energy density at time 6.7 ls, as the crack is
growing. As would be expected, the highest elastic
energy density is found ahead of the crack tips.
The calculation was carried out with all parameters
held constant, but four dierent values of grid spacing,
Dx = 0.10 mm, 0.25 mm, 0.35 mm, and 0.50 mm. The
grid in all cases was a square lattice with uniform spac-
Fig. 4. Square plate containing an initial center crack.
1532 S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535
ing throughout. The resulting predictions for crack
growth distance Da as a function of time are shown in
Fig. 7. The crack tip position at any time is dened to
be the maximum value of x
1
among all nodes i such that
u
i
> u
0
, where u
i
is the nodal value of damage dened in
(21), and u
0
= 0.3. (This value of u
0
, although slightly
arbitrary, was chosen because u usually reaches a max-
imum of about 0.4 along the crack faces.) The stairstep
appearance of the curves for the coarser grids occurs be-
cause, using this denition, the crack tip position jumps
discontinuously by Dx as the crack grows over time.
The maximum predicted crack growth speed (ob-
tained from the slope of the curves in Fig. 7) is about
1300 m/s, which is less than the Rayleigh wave speed
(2800 m/s). The predicted crack growth speed is there-
fore well within the theoretical limit for steady mode-I
fracture [21].
Fig. 7 shows that variations in the predicted crack
initiation time, arrest time, total growth distance, and
velocity, are on the order of the variations in grid spac-
ing. This demonstrates in a practical sense the linear
convergence of the predicted dynamic fracture behavior
as Dx is reduced.
6. Example problem: impact of a sphere on a brittle target
Consider the impact of a rigid sphere against a homo-
geneous block of brittle material. The sphere has mass
4.16 gm, diameter 10 mm, and impact velocity 35 m/s di-
rected normal to the target surface. The target material
has density q = 2200 kg/m
3
. The PMB material model
is used for the target with bulk modulus k = 14.9 GPa
and a critical bond stretch for failure given by (28) with
s
00
= 0.0005 and a = 0.25. A three-dimensional cubic grid
with spacing Dx = 0.5 mm and horizon d = 1.50 mm is
used. The target is a cylinder with diameter 34 mm and
height 25 mm. The target grid contains approximately
185,000 nodes and has free surfaces at the boundaries.
Fig. 8 shows two views of the predicted crack. The plot
onthe left shows a cross-section of the target after the pro-
jectile has rebounded, showing damage contours in the
target material, where damage is the quantity u dened
in (21). The damage is localized in a cone-shaped surface
which is known as a Hertzian crack. A three-dimensional
view of this crack is shown on the right. The curvature
near the circular cracktipis presumably due tointeraction
with the free cylindrical surface. The formation of conical
Hertzian cracks in glass under impact by spheres is ob-
served experimentally, for example by Ball [22].
Predictions for a similar problem, but with impact
velocity 100 m/s and a target with diameter 74 mm and
thickness 2.5 mm, are shown in Fig. 9. The upper plot
shows the classic debris cloud of high-velocity impact
on a thin target. The lower plot omits the small debris
and shows the predicted fragmentation of the entire tar-
get (displacements are exaggerated by a factor of 5).
This illustrates the ability of the method to model com-
plex patterns of crack growth and mutual interaction.
Fig. 5. Velocity applied to plate horizontal edges.
Fig. 6. Contours of elastic energy density as the crack grows.
Fig. 7. Crack tip growth distance Da(t) predicted with four
dierent grid spacings.
S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535 1533
An additional example concerned with dynamic frac-
ture in a high-strength steel plate may be found in [10].
Other examples may be found in [11,12].
7. Discussion
The primary advantage of the peridynamic approach
to fracture modeling is that it does not require any sup-
plemental relation that dictates when a crack should
grow, its velocity, direction, branching and instability,
and arrest. All of these phenomena emerge as conse-
quences of the equation of motion and the constitutive
model, which incorporates damage at the bond level.
In particular, it is not necessary to evaluate a stress
intensity factor at a crack tip for use in such a supple-
mental relation, or even to keep track of the crack tips
position. Because the same underlying equations apply
either on or o of a crack surface, it is unnecessary to
treat opposite sides of a crack separately. For these rea-
sons, any number of cracks can be treated, including all
their mutual interactions.
The question of the validity of the underlying integral
equations with respect to experimental data is beyond
the scope of this paper. However, it is worth noting that
these equations have been shown [9] to converge to the
classical PDEs, at least in the limited case of a one-
dimensional linear elastic bar, as the horizon shrinks,
d !0. Also, any microelastic peridynamic constitutive
model has an equivalent in the classical theory, in the
sense of having identical response under homogeneous
deformation [8]. The ability to prescribe, at least approx-
imately, an energy release rate (27) makes a connection
between the present theory and established principles
of fracture mechanics.
As noted in the Introduction, several meshfree meth-
ods such as SPH have been proposed as a means to
model fracture. There are fundamental distinctions in
both theory and practice between these methods and
the one proposed here. The main theoretical dierence
is that the peridynamic approach solves a dierent set
of continuum equations. The integral equations of the
peridynamic model remain directly applicable when a
discontinuity appears, while the PDEs of the classical
theory do not. The practical dierence is that cracks
emerge spontaneously in a peridynamic model as a nat-
ural consequence of the equation of motion and consti-
tutive models. Another practical dierence is that the
peridynamic model is based on interactions between
pairs of nodes, and therefore does not require a statisti-
cally large population of neighboring nodes to determine
spatial derivatives. One consequence of this pairwise ap-
proach is that the peridynamic method does not suer
from the well-known tensile instability in SPH.
A promising alternative numerical model based on
the peridynamic theory, with particular application to
fracture and failure of reinforced concrete, is under
development by Gerstle [23]. Another numerical ap-
proach, which may result in improved accuracy, is being
investigated by Emmrich and Weckner [24].
Acknowledgement
Sandia is a multiprogram laboratory operated by
Sandia Corporation for the United States Department
Fig. 8. Hertzian cracks due to impact on a brittle solid: (left)
Cross-section showing damage contours and (right) conical
crack.
Fig. 9. Debris cloud and fragmentation due to impact on a thin
brittle target.
1534 S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535
of Energy under Contract DE-AC04-94AL85000. This
work was supported by the Joint DOE/DOD Munitions
Technology Program, by the U.S. Nuclear Regulatory
Commission, by the US Department of Energy, and
by The Boeing Company. Discussions with John Chum-
ley of Techneglas, Inc. regarding Hertzian cracking are
gratefully acknowledged. The authors thank Professor
Walter Gerstle for helpful discussions and for providing
a correction to Eq. (23).
References
[1] Randles PW, Libersky LD. Smoothed particle hydrody-
namics: some recent improvements and applications.
Comput Methods Appl Mech Eng 1996;139:375408.
[2] Krysl P, Belytschko T. The element free Galerkin method
for dynamic propagation of arbitrary 3-D cracks. Int J
Numer Methods Eng 1999;44:767800.
[3] Carpinteri A, Ferro G, Ventura G. An augmented
Lagrangian element-free (ALEF) approach for crack
discontinuities. Comput Methods Appl Mech Eng
2001;191:94157.
[4] Li S, Liu W-K, Qian D, Guduru PR, Rosakis AJ. Dynamic
shear band propagation and micro-structure of adiabatic
shear band. Comput Methods Appl Mech Eng
2001;191:7392.
[5] Chen YP, Lee JD, Eskandarian A. Dynamic meshless
method applied to nonlocal crack problems. Theoret Appl
Fract Mech 2002;38:293300.
[6] Rao BN, Rahman S. Mesh-free analysis of cracks in
isotropic functionally graded materials. Eng Fract Mech
2003;70:127.
[7] Oliver J, Huespe AE. Continuum approach to material
failure in strong discontinuity settings. Comput Methods
Appl Mech Eng 2004;193:3195220.
[8] Silling SA. Reformulation of elasticity theory for discon-
tinuities and long-range forces. J Mech Phys Solids
2000;48:175209.
[9] Silling SA, Zimmermann M, Abeyaratne R. Deformation
of a peridynamic bar. J Elast 2003;73:17390.
[10] Silling SA. Dynamic fracture modeling with a meshfree
peridynamic code. In: Bathe KJ, editor. Computat Fluid
Solid Mech. Amsterdam: Elsevier; 2003. p. 6414.
[11] Silling SA, Askari E. Peridynamic modeling of impact
damage. In: Moody FJ, editor. PVP-Vol 489. New
York: American Society of Mechanical Engineers; 2004.
p. 197205.
[12] Silling SA, Bobaru F. Peridynamic modeling of membranes
and bers. Int J Non-Linear Mech 2005;40:395409.
[13] Weckner O, Abeyaratne R. The eect of long-range forces
on the dynamics of a bar. J Mech Phys Solids
2005;53:70528.
[14] Kunin IA. Elastic media with microstructure I: one-
dimensional models. Berlin: Springer-Verlag; 1982. p. 27.
[15] Rogula D. Nonlocal Theory of Material Media. Ber-
lin: Springer-Verlag; 1982. p. 13749, 24378.
[16] Eringen AC, Speziale CG, Kim BS. Crack-tip problem in
non-local elasticity. J Mech Phys Solids 1977;25:
339355.
[17] Ozbolt J, Bazant ZP. Numerical smeared fracture analysis:
nonlocal microcrack interaction approach. Int J Numer
Methods Eng 1996;39:63561.
[18] The version of this equation that appears in [11] contains
an error.
[19] Lapidus L, Pinder GF. Numerical solution of partial
dierential equations in science and engineering. New
York: Wiley; 2003. p. 171.
[20] Okabe A, Boots B, Sugihara K, Chiu SN. Spatial tessela-
tions: concepts and applications of Voronoi dia-
grams. New York: Wiley; 2000.
[21] Yoe EH. The moving Grith crack. Philos Mag
1951;42:73950.
[22] Ball A. On the bifurcation of cone cracks in glass plates.
Philos Mag A 1996;73:1093103.
[23] Gerstle W, Sau N, Peridynamic modeling of concrete
structures. In: Li L, editor. Proceedings of the Fifth
International Conference on Fracture Mechanics of
Concrete Structures, Ia-FRAMCOS, vol. 2; 2004.
p. 94956.
[24] Emmrich E, Weckner O, Analysis and numerical approx-
imation of an integro-dierential equation modelling non-
local eects in linear elasticity [to appear].
S.A. Silling, E. Askari / Computers and Structures 83 (2005) 15261535 1535