Theoretical and Applied Fracture Mechanics 53 (2010) 194–204

Contents lists available at ScienceDirect

Theoretical and Applied Fracture Mechanics

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

An adaptive dynamic relaxation method for quasi-static simulations using the

peridynamic theory
B. Kilic, E. Madenci *
Department of Aerospace and Mechanical Engineering, The University of Arizona, Tucson, AZ 85721, USA

a r t i c l e i n f o a b s t r a c t

Article history: The peridynamic theory is advantageous for problems involving damage since the peridynamic equation
Available online 11 August 2010 of motion is valid everywhere, regardless of existing discontinuities, and an external criterion is not nec-
essary for predicting damage initiation and propagation. However, the current solution methods for the
Keywords: equations of peridynamics utilize explicit time integration, which poses difficulties in simulations of most
Nonlinear experiments under quasi-static conditions. Thus, there is a need to obtain steady-state solutions in order
Peridynamics to validate peridynamic predictions against experimental measurements. This study presents an exten-
Dynamic relaxation
sion of dynamic relaxation methods for obtaining steady-state solutions of nonlinear peridynamic
Damage
equations.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction Also, in explicit solvers, the nature of explicit time integration

Over past decades, extensive research on computationally ro-
bust techniques has been performed to obtain either transient or
quasi-static nonlinear solutions to structural problems by employ-
ing the finite element method (FEM). Although well-established
and a robust numerical method, the FEM can suffer from computa-
tional difficulties while obtaining solutions to highly nonlinear
problems using implicit solvers. As explained in [1], the basic
Newton–Raphson algorithm might fail when limit or bifurcation
points are encountered along the load–displacement path. Further-
more, implicit solvers might snap through from one stable state to
another, which might conceal important dynamic phenomena.
This difficulty was overcome in [2,3] by introducing the
dynamic relaxation method as a powerful technique for obtaining
solutions to static nonlinear problems based on the fact that the
steady-state part of the transient solution represents the static
conditions. Later, the work in [4,5] independently introduced ficti-
tious density to improve the convergence of the dynamic relaxa-
tion method.
Using the dynamic relaxation method, a nonlinear problem was
first solved in [6] through the introduction of an artificial damping
into the problem. However, it is not always possible to determine
the most effective damping coefficient. Therefore, introduced in
[7,8] is the idea of an adaptive dynamic relaxation method in
which the damping coefficient is changed adaptively in each itera-
tion step.
* Corresponding author. Tel.: +1 520 621 6113; fax: +1 520 621 8191.
E-mail address: madenci@email.arizona.edu (E. Madenci).
requires small time steps and it becomes difficult to obtain solu-
tions under static or quasi-static conditions. The use of the dy-
namic relaxation method becomes very attractive for nonlinear
problems requiring large degrees of freedom as part of the explicit
solvers in which all quantities can be treated as vectors, resulting
in very low storage requirements.
Even with the adaptive dynamic relaxation method, the use of
the finite element method to predict failure can be quite challeng-
ing because remeshing may be required to make an accurate pre-
diction and damage can only propagate in certain directions.
Remeshing can be avoided by employing special elements, such
as cohesive elements. However, these elements require a prior
knowledge of the damage path, which might not be available. Un-
less these elements are correctly placed during model generation,
the damage predictions may be erroneous. The difficulty in pre-
dicting failure using the finite element method comes from the
mathematical form of the classical continuum mechanics equa-
tions. The equations of motion in classical continuum mechanics
are in the form of partial differential equations that involve the
spatial displacement derivatives; however, these derivatives are
undefined when the displacements are discontinuous, such as
across cracks or interfaces. Hence, failure prediction is posterior
and requires special techniques.
Realizing the aforementioned limitation [9], completely refor-
mulated the basic equations of continuum mechanics. The result-
ing approach is known as the peridynamic theory. The main
difference between the peridynamic theory and classical contin-
uum mechanics is that the former is formulated using integral
equations as opposed to derivatives of the displacement compo-
nents. This feature allows damage initiation and propagation at

0167-8442/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tafmec.2010.08.001
 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204
multiple sites, with arbitrary paths inside the material, without
resorting to special crack growth criteria. In the peridynamic the-
ory, internal forces are expressed through nonlocal interactions be-
tween pairs of material points within a continuous body, and
damage is a part of the constitutive model. Interfaces between dis-
similar materials have their own properties, and damage can prop-
agate when and where it is energetically favorable for it to do so.
The peridynamic theory has been utilized successfully for dam-
age prediction of many problems. Considered in [10] is the Kalt-
hoff–Winkler experiment in which a plate having two parallel
notches is hit by an impactor; the peridynamic simulations suc-
cessfully captured the angle of crack growth that was observed
in the experiment. Later, impact damage was predicted using peri-
dynamics in [11,12]. A plate with a center crack was used to show
convergence of their numerical method [12]. The peridynamic the-
ory was applied to damage analysis of plain and reinforced con-
crete structures in [13]. Furthermore, a new constitutive model
was introduced for tearing and stretching of rubbery materials in
[14]. Using this model, they predicted the oscillatory crack path
when a blunt tool is forced through a membrane. The peridynamic
theory was also applied successfully in [15,16] to predict damage
in laminated composites subjected to low-velocity impact and sta-
tic indentation. Moreover, the work in [17,18] also used peridy-
namic simulations to predict damage in laminates subjected to
low-velocity impact and to notched laminated composites under
biaxial loads.
In order to obtain these solutions, the peridynamic equations of
motion are discretized at the collocation points resulting in a sys-
tem of second-order partial differential equations with respect to
time. These discrete equations are then integrated by using a stan-
dard explicit time integrator such as central differencing and Run-
ge–Kutta. However, the nature of explicit time integration
necessitates utilization of small time steps and it becomes difficult
to obtain solutions to the problems subjected to static or quasi-sta-
tic conditions. Therefore, the present study extends the adaptive
dynamic relaxation method introduced in [8] within the realm of
the peridynamic theory. In this study, the damping coefficient in
dynamic relaxation is estimated through Rayleigh’s quotient,
which dampens the system from higher frequency modes to lower
frequency. Also, this approach avoids overshooting, which renders
the dynamic relaxation method advantageous over the implicit
solvers for finding the static solutions of nonlinear problems. The
effectiveness of the present approach is demonstrated by consider-
ing basic structural components subjected to tensile, bending, and
thermal loads.
2. Peridynamic theory
The classical continuum theory is based on the assumption that
a point in a continuum is influenced by the points that are located
infinitesimally small distances away from it. This local nature of
the classical continuum theory necessitates external failure criteria
to predict crack growth since displacement derivatives are un-
bounded at the crack tip. Therefore, introduced is [19–21] are the
nonlocal continuum theory. Because of the nonlocal nature of the
theory, it was shown [22,23] that the stress field ahead of the crack
tip is bounded as the crack tip is approached asymptotically, rather
than unbounded as predicted by the classical continuum theory. It
was suggested [22] that a natural fracture criterion by equating
maximum stress to the cohesive stress that holds the atomic bonds
together. This criterion can be applied everywhere in continuous
media without distinguishing discontinuities. Although their non-
local continuum theory leads to finite stresses at the crack tips, the
derivatives of the displacement field retained in the formulation
have discontinuities arising from the presence of cracks. Another
195
type of nonlocal theory, introduced in [21,24], circumvents this dif-
ficulty because it uses displacement fields rather than their deriv-
atives. However, it is only given for a one-dimensional medium.
Derived in [25] is a three-dimensional nonlocal model by approx-
imating a discrete periodic lattice structure as continuous media.
More recently, Silling [9] independently reintroduced a nonlocal
theory that does not require spatial derivatives—the peridynamic
theory. Compared to the previous nonlocal theory in [21,24], the
peridynamic theory is more general because it considers two-
and three-dimensional in addition to one-dimensional media. Un-
like the nonlocal theory in [25], the peridynamic theory provides
nonlinear material response with respect to displacements. Fur-
thermore, damage is included in the material response in the peri-
dynamic theory.
The peridynamic theory is concerned with the physics of a
material body at a material point that interacts with all points
within its finite distance, as shown in Fig. 1. As in the classical (lo-
cal) continuum theory, the material points of a body are continu-
ous, as opposed to the discrete nature of molecular dynamics.
However, the main distinction between the peridynamic- and con-
tinuum-based methods is that the former is formulated using inte-
gral equations, as opposed to partial differential equations that
include derivatives of the displacement components. Displacement
derivatives do not appear in the peridynamic equations, which al-
low the peridynamic formulation to be valid everywhere whether
or not displacement discontinuities are present. In a Cartesian
coordinate frame, the peridynamic equation of motion at a refer-
ence configuration of position x and time t is given as
Z
@2u
q ¼ dV x0 fðuðx; tÞ; uðx0 ; tÞ; x; x0 ; tÞ þ bðx; tÞ ð1Þ
@t 2 R
in which R is the domain of integration, u is the displacement vec-
tor field, b is a prescribed body-force density field, and q is mass
density. The response function, f is defined as the force vector per
unit volume squared that the material point at x0 exerts on the
material point at x.
The material points interact with each other directly through
the prescribed response function, which contains all of the consti-
tutive behavior of the material. The response function introduced
in [26] can be defined as
nþg 9k 2
fðg; nÞ ¼ lðn; tÞ 4 eðjnj=‘Þ ðs  ahÞ ð2Þ
jn þ gj p‘
in which k is the bulk modulus of the material, a is the coefficient of
thermal expansion of the material, h denotes the change in temper-
ature between material points x0 and x, and ‘ is the internal length,
which is the measure of nonlocal behavior. The locality of interac-
Fig. 1. Pairwise interaction of a material point with its neighboring points.
196 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204

tions depends on the internal length ‘ and interactions become

more local with decreasing internal length, as shown in Fig. 2. The
relative position, n, can be expressed as n = x0  x in the reference
configuration. The relative displacement, g, can be written as
g = u0  u between points x0 and x. The stretch, s, is given as
jn þ gj  jnj
s¼ ð3Þ
jnj
Failure is included in the material response through a history-
dependent scalar-valued function l, which is defined as

1 if sðt 0 ; nÞ  ah < s0 for all 0 < t 0 < t
lðn; tÞ ¼ ð4Þ
0 otherwise

in which s0 is the critical stretch for failure to occur, as shown in

Fig. 3. In the solution phase, the displacement of each material point
Fig. 3. Model for bond failure.
and the stretches between pairs of material points are computed.
When the stretch between two points exceeds the critical stretch,
s0, failure occurs and these two points cease to interact. Thus, dam- into subdomains in order to convert the integral form of the peri-
age in a material is simulated in a much more realistic manner com- dynamic equation of motion into a finite sum as
pared to the classical continuum-based methods. The critical X
N X
Ne
stretch was related to the critical energy release rate in [12] for an- qu€ ðxi ; tÞ ¼ bðxi ; tÞ þ wj fðuðxi ; tÞ; uðx0k ; tÞ; xi ; x0k ; tÞ ð6Þ
other type of response function. Although it is also possible to de- e¼1 j¼1

rive an expression similar to theirs for the present response

function, this method is not followed herein due to the lack of mea-
sured material properties.
The inexplicit nature of local damage at a point, arising from the
introduction of failure in the constitutive model, is removed by
defining the local damage as
R 2
R
dV x0 lðt; nÞeðjnj=lÞ
uðx; tÞ ¼ 1  R 2
ð5Þ
R
dV x0 eðjnj=lÞ
Thus, local damage is the weighted ratio of the number of broken
interactions to the total number of interactions. The material re-
sponse is linear with respect to stretch, as given in (2), but it is also
possible to utilize the bi-linear material response, as described in
[12].
2.1. Steady-state solution to peridynamic equations of motion
It is possible to construct analytical solutions to (1) for some
simple problems; however, most engineering problems concern
complex geometries involving more than one material region.
Therefore, the solution to (1) is constructed by utilizing a numeri-
cal method. The numerical treatment of the equations of motion
involves the discretization of the domain of interest into subdo-
mains. Subsequently, collocation (integration) points are placed
where xi is the position vector located at the ith collocation point, N
is the number of subdomains, and Ne is the number of collocation
points in the eth subdomain. The position vector x0k represents the
jth integration point of the eth subdomain. The parameter wj is
the integration weight associated with point x0k .
In order to obtain the steady-state solution to the system of or-
dinary differential equations given in (6), the dynamic relaxation
method is adopted to simulate the quasi-static loading conditions.
The dynamic relaxation method is based on the fact that the static
solution is the steady-state part of the transient response of the
solution. An artificial damping is introduced to guide the solution
into the steady-state regime as fast as possible. However, it is
not always possible to determine the most effective damping coef-
ficient. Therefore, the steady-state solution of (6) is achieved by
employing the adaptive dynamic relaxation method introduced
in [8] in which the damping coefficient is changed adaptively in
each time step.
By removing the acceleration term and introducing a fictitious
diagonal density and damping matrix that is proportional to the
density matrix, (6) can be cast into set of ordinary differential
equations that become suitable for dynamic relaxation as
€
KUðX; _
tÞ þ cKUðX; tÞ ¼ FðU; U0 ; X; X0 Þ ð7Þ
in which K is the fictitious diagonal density matrix and c is the
damping coefficient whose values are determined through Gres-
chgorin’s theorem [8] and Rayleigh’s quotient, respectively. Fur-
thermore, the vectors X and U represent positions and
displacements at the collocation points, respectively, and they can
be expressed as

XT ¼ fx; x2 ; . . . ; xM g ð8Þ
and

UT ¼ fuðx1 ; tÞ; uðx2 ; tÞ; . . . ; uðxM ; tÞg ð9Þ

where M is the total number of collocation points. Finally, the vector
F is the summation of internal and external forces and its ith com-
ponent can be written as
X
N X
Ne
Fi ¼ bðxi ; tÞ þ wj fðuðxi ; tÞ; uðx0k ; tÞ; xi ; x0k ; tÞ ð10Þ
e¼1 j¼1

By utilizing the central-difference explicit integration, displace-

Fig. 2. Effect of internal length. ments and velocities for next iteration step can be written as
 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204
h i
U_ nþ1=2 _ n1=2 þ 2DtK1 Fn =½2 þ cDt
¼ ð2  cDtÞU
Unþ1 ¼ Un þ Dt U_ nþ1=2
in which n indicates the nth iteration. However, the integration
algorithm given by (11) and (12) cannot be used to start the integra-
tion due to the unknown velocity field at t1/2. By assuming that
U0 – 0 and U_ 0 ¼ 0, integration can be started by using
U_ 1=2 ¼ DtK1 F0 =2
The physical term in this algorithm is only the summation of inter-
nal and external forces, F. The density matrix K, damping coeffi-
cient c, and time step size Dt do not have to be physically
meaningful quantities, as mentioned in [8]. Thus, they can be cho-
sen to obtain faster convergence.
In dynamic relaxation, a time step size of 1 (Dt = 1) is a conve-
nient choice. There are many different methods for choosing the
density matrix but one of the most common method, given in
[8], is based on Greschgorin’s theorem, which can be written as
1 2X
kii  Dt jK ij j
4 j
in which Kij is the stiffness matrix of the equation system under
consideration. The inequality sign ensures stability of the central-
difference explicit integration, and derivation of this stability condi-
tion is given in [8]. Although this approach achieves near optimal
density, the density is coordinate frame dependent because it de-
pends on absolute values of the global stiffness matrix, as stated
in the context of the finite element method of [27]. Therefore, an
alternative way was suggested [28] to choose the mass based on
the minimum element dimension to make the mass frame invari-
ant. This method seems to reduce overshooting as compared to
the mass allocation based on Greschgorin’s theorem. Therefore,
the present solutions of peridynamic equations also utilize a frame
invariant density matrix. However, the construction of the stiffness
matrix requires knowledge of the response function’s derivative
with respect to the relative displacement g. Since the response
function given in (2) is a nonlinear function of g, it is not always
possible to obtain its derivative. However, by considering the line-
arized version of the response function for small displacements,
an absolute row sum of the stiffness matrix can be written as
X X
N X
Ne
jn  ej 9k ðjnj=‘Þ2 1
jK ij j ¼ wj e ð15Þ
j e¼1 j¼1
jnj p‘4 jnj
in which e is a unit vector along x-, y-, or z-direction. The summa-
tion given in Eq. (15) can be employed for the density matrix but it
is not sufficient to make the density frame invariant. Therefore, the
maximum value of |n  e| is used for all directions.
As described in [8], the damping coefficient can be determined
by using the lowest frequency of the system. Therefore, the lowest
frequency is determined by making use of Rayleigh’s quotient,
which is given as
rffiffiffiffiffiffiffiffiffiffiffiffi
uT Ku
x¼
uT Ku
However, the density described by (14) might have large numerical
values, which makes the denominator of (16) numerically difficult
to compute. In order to avoid this problem, (7) can be rewritten
at nth iteration as
U _ n ðX; tn Þ ¼ K1 Fn ðUn ; U0n ; X; X0 Þ
€ n ðX; t n Þ þ cU
The damping coefficient for (17) can be expressed by using (16) for
each iteration step as
197
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð11Þ cn ¼ 2 ððUn ÞT 1 Kn Un Þ=ððUn ÞT Un Þ ð18Þ
in which 1Kn is the diagonal ‘‘local” stiffness matrix, which is given
ð12Þ
as
1
K nii ¼ ðF ni =kii  F n1
i =kii Þ=ðDtu_ n1=2
i Þ ð19Þ
Since the local stiffness matrix calculation involves division by
the velocity component, it is important to start the solution with
ð13Þ a nonzero displacement field to avoid division by zero. Further-
more, U0 – 0 is not sufficient to eliminate division by zero because
there is no guarantee that the velocity field at a point will take non-
zero values during the solution procedure. Therefore, velocities are
monitored and local stiffness is set to zero if a zero velocity is
encountered, which results in removal of the U0 – 0 requirement.
2.2. Boundary conditions
Since peridynamics is a nonlocal theory and its equations of
motion utilize integral equations as opposed to partial differential
equations in the case of the classical continuum theory, the appli-
ð14Þ cation of boundary conditions is also very different from that of the
classical continuum theory. The difference can be illustrated by
considering a body that is subjected to mechanical loads. If this
body is fictitiously divided into two domains X and X+ as shown
in Fig. 4a, there must be a net force, F+, which is exerted to domain
X+ by domain X so that force equilibrium is satisfied. According
to the classical continuum mechanics, force F+ can be determined
by integrating surface tractions on the cross-section area of do-
mains X and X+ as
I
Fþ ¼ dAT ð20Þ
@X
in which T is the surface tractions, as shown in Fig. 4b. In the case of
the peridynamic theory, the material points located in domain X+
interact with the points in domain X, as show in Fig. 4c. Hence,
the force densities, L, acting on points in domain X+ must be deter-
mined by integrating the response function over domain X as
Z
0
LðxÞ ¼ dV fðx; x0 Þ ð21Þ
X
Finally, the force, F+, can be computed by volume integration of
these force densities (Fig. 4d) over domain X+ as
Z Z Z
Fþ ¼ dVLðxÞ ¼ dV
0
dV fðx; x0 Þ ð22Þ
Xþ Xþ X
Hence, the tractions or point forces cannot be applied as boundary
conditions since their volume integrations result in a zero value.
Therefore, the boundary conditions are applied over the volumes
as body forces, displacements, and velocities.
2.3. Modeling of material
According to the response function given in (2), a material point
interacts with all other points within the problem domain. How-
ð16Þ
ever, it is computationally expensive for a large number of colloca-
tion points. Fortunately, the magnitude of the response function
decreases rapidly when the distance between the points increases,
as shown in Fig. 2. Thus, the cutoff radius, rc, is introduced to re-
duce the computation time by limiting the interaction range, as de-
scribed in [26]. On the other hand, the internal length, ‘, cannot be
ð17Þ chosen arbitrarily small since the accuracy of the volume integrals
in (1) is dependent on the variation of the response function within
the subdomains. A numerical investigation of the effect of the
internal length and cutoff radius on numerical predictions is pre-
sented in [26]. Based on this investigation, the internal length, ‘,
198 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204
Fig. 4. Boundary conditions: (a) domain of interest; (b) tractions in classical continuum mechanics; (c) interaction of a point in domain X+ with domain X and (d) force
densities acting on domain X+ due to domain X.
is chosen to be approximately the edge length of the hexahedral
subdomains and the cutoff radius, rc, is then set to 2.5 times the
internal length, ‘.
Due to pairwise interactions, Poisson’s ratio is restricted to be 1/
4 in three-dimensions and 1/3 in two-dimensions in the peridy-
namics theory as explained in [13]. Hence, the effective Poisson’s
ratio is influenced by the geometric dimension, which is on the
same order of the internal length as investigated in [26]. If the
dimension is smaller than the internal length in one direction,
Poisson’s ratio approaches to 1/3 since most of the interactions oc-
cur in a plane. On the other hand, Poisson’s ratio approaches its
theoretical value of 1/4 with an increase in the dimension. Because
the numerical solution method assumes the Poisson’s ratio to be 1/
4, small deviations are observed in the numerical results as com-
pared to the finite element method.
Because the peridynamic theory is nonlocal, material points
interact across the interfaces. Hence, the response function needs
to be specified for the interface in addition to the response func-
tions for individual materials. Therefore, the peridynamic theory
is capable of modeling different adhesion strengths of the materi-
als. If the domain consists of two dissimilar materials as shown in
Fig. 5, three different interactions need to be specified. Two of
these interactions occur between material points having the same
material, labeled 1 and 2 in Fig. 5, and the material properties used
by the response function are trivially chosen to be that of the mate-
rial point. In the case of interactions across the interface, labeled 3
in Fig. 5, the numerical experimentations revealed that displace-
ment predictions are insensitive to properties of the interface
material if the numbers of interactions across the interfaces are
much smaller than those for material points having the same
material. When the numbers of interactions across the interfaces
are comparable to those between the material points having the
same material, the study in [29] showed that the use of the smaller
of the bulk moduli of the two materials gives better results. Hence,
the response function for interactions across the interface utilizes
Fig. 5. Interactions among material points.
the material properties of the region having the smaller bulk
modulus.
2.4. Numerical solution of equations
In an effort to reduce the peridynamic equation of motion into
its discrete form, the solution domain shown in Fig. 6a is discret-
ized into subdomains, as shown in Fig. 6b. The numerical method
is mesh-free in the sense that problem domains can be discretized
without having any connectivity among the subdomains, as shown
in Fig. 6b. After discretization, Gaussian integration points are
placed into each subdomain, as illustrated in Fig. 6b. The subdo-
mains with different shapes can be utilized during discretizations.
In this study, volume integration is performed using hexahedral
subdomains. In each subdomain, one or eight integration points
are utilized. This type of discretization usually leads to a large
number of collocation points. Therefore, parallel processing using
OpenMP is also employed to reduce computation time while utiliz-
ing the uniform grids as arrays of linked lists as described in [26]. A
binary space decomposition, introduced in [26,30], is used to ob-
tain good load balancing on each processor. The decomposition in-
volves division of the space into two subunits, with an
approximately equal number of collocation points. Each subunit
is then continuously divided into two subunits until the numbers
of subunits and processors become equal. Division is performed
only on the x, y, or z plane, with the constraint of minimum divi-
sion surface to reduce the communication cost among the
processors.
3. Numerical results
The effectiveness of the present approach is demonstrated by
considering basic structural components subjected to tensile and
thermal loads. The basic structural components considered are a
plate with a hole and a bimaterial strip. The plate with a hole is
subjected to uniform tension, and the bimaterial strip is under uni-
form temperature change.
3.1. Plate with a hole under tension
The geometry of a plate with a hole is described in Fig. 7.The ori-
gin of the coordinate system is attached to plate’s geometric cen-
ter. As shown in Fig. 7, the plate geometry is defined by a length,
a, of 228.6 mm, a width, b, of 76.2 mm, and thickness, t, of
5.08 mm. The circular hole with a radius, r, of 12.7 mm is located
at the center of the plate. The material is assumed to be isotropic
with elastic modulus and Poisson’s ratio of E = 68.947 GPa and
t = 0.25, respectively.
 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204 199

Fig. 8. Nonuniform discretization for Gaussian integration for only a quarter of the
plate with a hole.

Fig. 6. Discretization of problem domain: (a) problem domain and (b) its
discretization.

Fig. 7. Geometry of a plate with a hole.

The plate is discretized into hexahedrons, as shown in Fig. 8,

and the hexahedral volumes are approximately 2.048 mm3. Eight
integration (collocation) points, based on a Gaussian integration
scheme, are placed in the subdomains to reduce the equation of
motion, (6) to its discrete form. The parameter defining the inter-
nal length, ‘, and the cutoff radius, rc, are chosen to be 1.27 mm
and 3.175 mm, respectively. Furthermore, the plate is free of any
displacement constraints and subjected to uniaxial tension of
26.689 kN. The tensile loads are applied along the vertical edges
of the plate as body forces in the rectangular regions over a length
of l = 1.27 mm, as illustrated in Fig. 7. The corresponding value of
Fig. 9. Displacement variations at two specific points for an increasing number of
the applied uniform body-force density is 54.289 N/mm3. On the iterations: (a) x-direction and (b) y-direction.
mid-plane, z = 0 along the symmetry line, y = 0, the change in dis-
placement at two specific points, x = 12.7 mm and x = 114.3 mm
are shown in Fig. 9a as the number of iterations increases. Simi- at points y = 12.7 mm and y = 38.1 mm are shown in Fig. 9b. The
larly, along the symmetry line, x = 0, the change in displacement changes in velocities at these points are shown in Fig. 10. As shown
200 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204
Fig. 10. Velocity variations at two specific points for an increasing number of
iterations: (a) x-direction and (b) y-direction.
in Figs. 9a and 10a, the longitudinal displacements reach their stea-
dy-state values in approximately 800 iterations. However, it takes
approximately 1000 iterations for the convergence of transverse
displacements, as presented in Figs. 9b and 10b. The reason for
the slow convergence in transverse displacements can be attrib-
uted to the existence of small in-plane bending due to the presence
of a hole. Nevertheless, these results indicate that nonuniform
discretization does not produce any ill effects for achieving
convergence.
The validity of the peridynamic predictions is established by
comparing against the finite element method using ANSYS. In the
case of the finite element model, discretization is the same as that
for the peridynamic model, as shown in Fig. 8. Furthermore, the
SOLID45 brick element of ANSYS is utilized to approximate the dis-
placements. Figs. 11 and 12 show horizontal and vertical displace-
ments, respectively, along the bottom line of the plate. The
comparison of horizontal displacements indicates close agreement,
as shown in Fig. 11. Although there is also good correlation be-
tween vertical displacements of the present and the finite element
method, as shown in Fig. 12, there is a larger deviation for vertical
displacements than for horizontal displacements near the hole be-
cause the Poisson’s ratio of 1/4 is not exactly satisfied in the peri-
dynamic theory. Since the plate thickness is comparable with the
value of the cutoff radius, there are more interactions in the x–y
Fig. 11. Horizontal displacements along the bottom line.
Fig. 12. Vertical displacements along the bottom line.
plane resulting in a Poisson’s ratio between 1/4 and 1/3, as numer-
ically investigated in [26].
Unlike the classical continuum methods, as mentioned previ-
ously, the peridynamic equation of motion is valid everywhere
regardless of existing discontinuities in displacement fields. There-
fore, the peridynamic theory is very suitable for predicting damage
initiation and growth because the damage model is built into the
material response. However, the damage prediction can still be dif-
ficult because damage happens progressively and its path is depen-
dent on its history. Therefore, damage problems are very nonlinear,
even if the elastic material response is linear.
After validating against the finite element method, the plate
with a hole is further investigated to determine how damage initi-
ates and grows under quasi-static conditions. In order to simulate
quasi-static loading, the tension loading is applied as a displace-
ment, incremented linearly with time, at the ends of the plate.
The displacements are applied in a finite volume over the length
l of 5.08 mm, as shown in Fig. 7. In this study, if the magnitudes
of traveling waves are numerically insignificant, the solution repre-
sents the quasi-static conditions. In order to determine the highest
loading rate for quasi-static conditions, the tensile load is applied
 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204 201

Fig. 15. Variation of ultimate failure force with increasing critical stretch.
Fig. 13. Force due to applied displacements at the cross-section.

failure pattern in a plate with a hole, as illustrated in Fig. 14, in

by incrementing the displacement at each iteration of adaptive dy-
the form of local damage. As expected, the crack emanates from
namic relaxation with a constant value.
the hole boundary at which high displacement gradients exist, as
The force required to sustain the applied displacements is com-
shown in Fig. 14a, then the crack propagates toward the edges of
puted by considering a cross-section located at w = 25.4 mm away
the plate, as shown in Fig. 14b. The effect of critical stretch is fur-
from the plate center as shown in Fig. 7. In peridynamics, the force
ther investigated by fixing the loading rate to 1.27E6 in/iteration.
that is applied to domain X+ by X can then be determined by
Fig. 15 shows the change in ultimate force with increasing critical
using (22). Fig. 13 shows the variation of the cross-sectional force
stretch. It is observed in Fig. 15 that the ultimate failure force is lin-
for varying loading rates for a critical stretch, s0, of 0.0025. For high
early dependent on the critical stretch. Thus, it may be possible to
loading rates, the force does not increase linearly due to dynamic
determine the critical stretch for the peridynamic analysis by mea-
effects. Furthermore, the force at ultimate failure is higher for high
suring the ultimate force experimentally. By knowing the ultimate
loading rates than for small loading rates, which is consistent with
failure force, the experimental setup can be simulated for two dif-
the experimental results. As observed in Fig. 13, the quasi-static
ferent critical stretch values to determine the ultimate failure
conditions apply for the loading rates smaller than 6.35E6 in/iter-
force. Finally, the critical stretch of a specific material can be calcu-
ation. Furthermore, the peridynamic theory captures the typical
lated by linear interpolation.

3.2. Bimaterial strip under uniform temperature change

A thin bimaterial strip is considered in order to examine the

effectiveness of adaptive dynamic relaxation in the presence of
uniform temperature loading. The geometry of a bimaterial strip
is described in Fig. 16. The length of the strip is specified as
l = 30 mm. The height of each material strip is denoted by
h1 = 3 mm and h2 = 1 mm. Their thickness has a value of
t = 0.1 mm. The strip is free of any displacement constraints and
is subjected to a uniform temperature of h = 50 °C. Both materials
are isotropic; their material constants are given in Table 1.
The three-dimensional peridynamic model is constructed by
discretizing the problem domain with cubic regions having an edge
length of D = 0.1 mm, as shown in Fig. 16. Moreover, eight integra-

Fig. 14. Contour of local damage: (a) crack initiation and (b) ultimate failure. Fig. 16. Dimensions and discretization for bimaterial strip.
202 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204

Table 1
Material properties for bimaterial strip.

Material number Elastic modulus (GPa) Poisson’s ratio Critical stretch (s0) Thermal expansion coeff. (ppm/°C)
1 5.1 0.25 0.004 50.0
2 128 0.25 0.005 16.6

tion (collocation) points are placed in cubic regions to reduce the

equation of motion, (6), to its discrete form. The parameter defin-
ing the characteristic internal length, ‘, is chosen to be 0.1 mm.
Furthermore, a cutoff radius, rc, of 0.25 mm is used to reduce the
computation time.
The changes in displacements with increasing number of itera-
tions are presented at two specific points along the interface in
Fig. 17. These points are defined by x = 0.0 mm and x = 15.0 mm.
The displacement variation in the x-direction is only given at the
point located at x = 15.0 mm because the displacement at
x = 0.0 mm is zero due to symmetry. The velocity variations of
these points are given in Fig. 18. As shown in Figs. 17 and 18, the
bimaterial strip reaches its steady-state regime in approximately
10,000 iterations. As expected, the convergence is slower than that
of the plate with a hole under tensile loading since the bimaterial
strip is subjected to bending forces due to the presence of dissim-
ilar materials with different coefficients of thermal expansion. Fur-
thermore, displacements in the y-direction converges their steady-

Fig. 18. Velocity variation at the interface for an increasing number of iterations:
(a) x-direction and (b) y-direction.

state values at a slower rate than those in the x-direction, as in the

Fig. 17. Displacement variation at the interface for an increasing number of
iterations: (a) x-direction and (b) y-direction.
case of plate with a hole under tension.
In order to establish the validity of the peridynamic predictions,
the displacement components along the interface are compared
against the finite element method, using SOLID45 elements of
ANSYS, a commercially available program. The same discretization
of the peridynamic model as shown in Fig. 16 is also utilized to
construct the finite element mesh. As shown in Fig. 19, the com-
parisons of displacements along the interface indicate very close
agreement. Furthermore, the displacements in the horizontal and
vertical directions are shown in the form of contour plots in
Fig. 20. The coefficient of thermal expansion for material 1 is
approximately three times that for material 2. Therefore, it can
be observed in Figs. 19b and 20b that the bimaterial strip bends
downward.
 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204
Fig. 19. Comparison against FEM: (a) horizontal and (b) vertical displacements
along interface.
Fig. 20. Peridynamic predictions of displacements: (a) horizontal and (b) vertical
components.
In the prediction of damage, the temperature of the strip is in-
creased 5  104 °C at each iteration step instead of applying a
constant uniform temperature. Fig. 21 shows the contour plots of
local damage given in (5) at different iterations. As expected, two
203
Fig. 21. Progression of damage for bimaterial strip subjected to incremental
thermal load.
cracks initiate at the interface and grow along the interface toward
the middle of the strip, as shown in Fig. 21.
4. Conclusions
This study presents an adaptive dynamic relaxation method to
solve governing equations of the peridynamic theory that are non-
linear with respect to displacements. Hence, the explicit form of
the stiffness matrix is not known. Therefore, the row summation
of the stiffness matrix is determined by an approximate approach,
and the numerical results indicate very good convergence. The re-
sults show that the problems involving bending converge slower
than those subjected to tension. This can be attributed to the deter-
mination of the critical damping coefficient by using Rayleigh’s
quotient. Rayleigh’s quotient estimates the lowest extensional fre-
quency accurately and the lowest extensional frequency is greater
than the lowest active bending frequency. Hence, the system is
heavily overdampened, resulting in slow convergence. Further-
more, the results show that the nonuniform discretization does
not reduce the computational performance significantly. It is also
illustrated that adaptive dynamic relaxation method can be uti-
lized while predicting damage by using the peridynamic theory.
References
[1] R.C. Yu, G. Ruiz, Explicit finite element modeling of static crack propagation in
reinforced concrete, International Journal of Fracture 141 (2006) 357–372,
doi:10.1007/s10704-006-9002-0.
[2] A.S. Day, An introduction to dynamic relaxation, The Engineer 219 (1965) 218–
221.
[3] J.R.H. Otter, Computations for prestressed concrete reactor pressure vessels
using dynamic relaxation, Nuclear Structural Engineering 1 (1965) 61–75.
[4] A.K. Welsh, Discussion on dynamic relaxation, Proceedings of the Institution of
Civil Engineers 37 (1967) 723–735.
[5] A.C. Cassell, P.J. Kinsey, D.J. Sefton, Cylindrical shell analysis by dynamic
relaxation, Proceedings of the Institution of Civil Engineers 39 (1968) 75–84.
[6] K.R. Rushton, Large deflection of variable-thickness plates, International
Journal of Mechanical Sciences 10 (1968) 723–735, doi:10.1016/0020-
7403(68)90086-6.
[7] M. Papadrakakis, A method for the automatic evaluation of the dynamic
relaxation parameters, Computer Methods in Applied Mechanics and
Engineering 25 (1981) 35–48, doi:10.1016/0045-7825(81)90066-9.
[8] P. Underwood, Dynamic relaxation, Computational Methods for Transient
Analysis 1 (1983) 245–265.
[9] S.A. Silling, Reformulation of elasticity theory for discontinuities and long-
range forces, Journal of Mechanics and Physics of Solids 48 (2000) 175–209,
doi:10.1016/S0022-5096(99)00029-0.
[10] S.A. Silling, Dynamic fracture modeling with a meshfree peridynamic code, in:
K.J. Bathe (Ed.), Second MIT Conference on Computational Fluid and Solid
Mechanics, Elsevier, Amsterdam, 2003, pp. 641–644.
[11] S.A. Silling, E. Askari, Peridynamic modeling of impact damage, in: F.J. Moody
(Ed.), 2004 ASME Pressure Vessels and Piping Conference, ASME, New York,
2004, pp. 197–205 (PVP 489).
[12] S.A. Silling, E. Askari, A meshfree method based on the peridynamic model of
solid mechanics, Computers and Structures 83 (2005) 1526–1535,
doi:10.1016/j.compstruc.2004.11.026.
204 B. Kilic, E. Madenci / Theoretical and Applied Fracture Mechanics 53 (2010) 194–204
[13] W.H. Gerstle, N. Sau, Peridynamic modeling of concrete structures, in: V.C. Li,
C.K.Y. Leung, K.J. William, S.L. Billington (Eds.), Fracture Mechanics of Concrete
Structures: Proceedings of the Fifth International Conference on Fracture
Mechanics of Concrete Structures, Ia-FraMCoS, 2004, pp. 949–956.
[14] S.A. Silling, F. Bobaru, Peridynamic modeling of membranes and fibers,
International Journal of Non-linear Mechanics 40 (2005) 395–409,
doi:10.1016/j.ijnonlinmec.2004.08.004.
[15] K.W. Colavito, B. Kilic, E. Celik, E. Madenci, E. Askari, S. Silling, Effects of
nanoparticles on stiffness and impact strength of composites, in: 48th AIAA/
ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, Honolulu, HI, USA, 2007 (AIAA 2007-2021).
[16] K.W. Colavito, B. Kilic, E. Celik, E. Madenci, E. Askari, S. Silling, Effect of void
content on stiffness and strength of composites by a peridynamic analysis and
static indentation test, in: 48th AIAA/ASME/ASCE/AHS/ASC Structures,
Structural Dynamics, and Materials Conference, Honolulu, HI, USA, 2007
(AIAA 2007-2257).
[17] E. Askari, J. Xu, S. Silling, Peridynamic analysis of damage and failure in
composites, in: 44th AIAA/ASME/ASCE/AHS/ASC Aerospace Sciences Meeting
and Exhibit, Reno, NV, USA, 2006 (AIAA 2006-88).
[18] J. Xu, A. Askari, O. Weckner, H. Razi, S. Silling, Damage and failure analysis of
composite laminates under biaxial loads, in: 48th AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics, and Materials Conference, Honolulu, HI, USA,
2007 (AIAA 2007-2315).
[19] A.C. Eringen, D.G.B. Edelen, On nonlocal elasticity, International Journal of
Engineering Science 10 (1972) 233–248, doi:10.1016/0020-7225(72)90039-0.
[20] E. Kroner, Elasticity theory of materials with long range cohesive forces,
International Journal of Solids and Structures 3 (1967) 731–742, doi:10.1016/
0020-7683(67)90049-2.
[21] I.A. Kunin, Elastic Media with Microstructure I: One Dimensional Models,
Springer-Verlag, Berlin, 1982.
[22] A.C. Eringen, B.S. Kim, Stress concentration at the tip of crack, Mechanics
Research Communications 1 (1974) 233–237, doi:10.1016/0093-
6413(74)90070-6.
[23] A.C. Eringen, B.S. Kim, On the problem of crack tip in nonlocal elasticity, in: P.
Thoft-Christensen (Ed.), Proceedings, Continuum Mechanics Aspects of
Geodynamics and Rock Fracture Mechanics, D. Reidel Publishing Co.,
Dordrecht, Holland, 1974, pp. 107–111.
[24] D. Rogula, Nonlocal Theory of Material Media, Springer-Verlag, Berlin, 1982.
pp. 137–149, 243–278.
[25] B. Kilic, Peridynamic Theory for Progressive Failure Prediction in Composite
Materials, Ph.D. Dissertation, Department of Aerospace and Mechanical
Engineering, The University of Arizona, Tucson, AZ, 2008.
[26] I.A. Kunin, Elastic Media with Microstructure II: Three-dimensional Models,
Springer-Verlag, Berlin, 1983.
[27] T.G. Lovie, D.R. Metzger, Lumped mass tensor formulation for dynamic
relaxation, in: G.M. Hulbert (Ed.), Computer Technology, ASME, New York,
1999, pp. 255–260 (PVP 385).
[28] M.J. Berger, S.H. Bokhari, A partitioning strategy for nonuniform problems on
multiprocessors, IEEE Transactions on Computers C-36 (1987) 570–580.
[29] R.G. Sauve, D.R. Metzger, Advances in dynamic relaxation techniques for
nonlinear finite element analysis, Journal of Pressure Vessel Technology 117
(1997) 170–176.
[30] B. Kilic, A. Agwai, E. Madenci, Damage prediction in notched composites using
peridynamic theory, in: 49th AIAA/ASME/ASCE/AHS/ASC Structures, Structural
Dynamics, and Materials Conference, Schaumburg, IL, USA, 2008 (AIAA-2008-
2000).