Lattice Models: Rigidity, Randomness, Dynamics, and Optimality

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
4
Lattice Models: Rigidity, Randomness,
Dynamics, and Optimality
There is a crack in everything: that’s how the light gets in.

Leonard Cohen, 1992
This chapter takes lattice models to more general settings. There are four
themes: rigidity, randomness, dynamics, and optimality. The first two of these
involve the introduction of spatial randomness into a lattice as a departure
from an originally periodic geometry of, say, the central-force triangular net-
work of Chapter 3. One path is through a random depletion of bonds, which
leads to a total loss of stiffness, or rigidity, of the lattice. Another way of creat-
ing a random lattice is through a random (instead of a regular) network of, say,
a Poisson–Delaunay variety. The third theme considered here is a generaliza-
tion from statics to dynamics, with nodes acting as quasi-particles—here we
have a coarse scale cousin to molecular dynamics, and, at the same time, an
alternative to finite element methods. Finally, the fourth topic is that of opti-
mal use of material for given loading and support conditions, where a special
case of central-force lattices arises, Michell trusses being the basic paradigm.
4.1 Rigidity of Networks

4.1.1 Structural Topology and Rigidity Percolation
When considering a central force (or truss) network, a question of fundamen-
tal importance is whether such a structure is a sufficiently constrained system
or not. In other words: is it an intrinsically rigid body? This is the subject mat-
ter of a field called structural topology. In the following we provide its basic
concepts.
Any central force network is a set of edges (or bars), and vertices
(or nodes acting as frictionless pivots). We immediately have an edge set
E, and a vertex set V, so that the network is represented by a graph G(V, E).
An edge is an unordered pair of two vertices. Structural rigidity can be based
either on statics or on kinematics, and, as we shall see below, they are, in a
certain sense, equivalent; see Table 4.1.
133
© 2008 by Taylor & Francis Group, LLC

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
134 Microstructural Randomness and Scaling in Mechanics of Materials
TABLE 4.1
A Comparison of Various Terminologies from Structural Mechanics,
Structural Topology, and Physics
Field Terminology Terminology
Structural mechanics Minimally constrained truss Mechanism

Structural topology Isostatic structure Nonisostatic structure
statical approach
Structural topology Infinitesimally rigid structure Nonrigid structure
kinematical approach
Condensed matter physics Rigid network Floppy network
The statical approach involves, in the first place, the concept of an equilib-
rium load. A system of forces assigned to the nodes of a network is said to be
an equilibrium load if and only if (iff ) the sum of the assigned vectors is a zero
vector, and the total moment of those vectors about any one point is zero. A
network resolves an equilibrium load iff there is an assignment of tensions
and compressions to all the bars of E, such that the sum at each node is equal
and opposite to its assigned load. A structure is said to be statically rigid iff
it resolves all equilibrium loads.
The kinematic approach involves the concept of an infinitesimal motion,
which is an assignment of velocities to all the nodes of V, such that the dif-
ference of velocities assigned to the ends of any bar is perpendicular to the
bar itself. This means that the motion does not result in any extension or
compression of the bar. Every connected plane structure has at least three
degrees of freedom (two translations and one rotation), and this is called a
rigid motion. A structure is said to be infinitesimally rigid if and only if all its in-
finitesimal motions are rigid motions. These statical and kinematical pictures
are connected by a theorem due to Crapo and Whiteley (1989): “A structure
is statically rigid iff it is infinitesimally rigid.”
Next, a structure is said to be isostatic iff it is minimally rigid, that is,
when it is infinitesimally rigid but the removal of any bar introduces some in-
finitesimal motion. Clearly, in an isostatic structure all the bars are necessary
to maintain the overall rigidity. In statics this is called a statically determi-
nate structure, as opposed to the indeterminate ones that have more than a
minimally sufficient number of bars for the global rigidity. It is a well-known
result that, in 2D, a determinate structure of |V| nodes has edges numbering
|E| = 2|V| − 3, (4.1)
where || denotes the number of elements in a given set. As an example, let

us consider an incomplete triangular lattice shown in Figure 4.1. Although it
satisfies (4.1) it is not at all clear whether it is isostatic.
This example shows that |E| = 2|V| − 3 is only a necessary but not a
sufficient condition for rigidity. The latter is provided by this theorem (e.g.,
Laman, 1970; Asimov and Roth, 1978): “A planar graph structure is isostatic
if and only if it has 2|V| − 3 bars, and, for every m, 2 ≤ m ≤ |V|, no subset
of m nodes has more than 2m − 3 bars connecting it.” This, effectively, allows

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
Lattice Models: Rigidity, Randomness, Dynamics, and Optimality 135
FIGURE 4.1
A triangular lattice with 71 edges and 37 vertices; it is generically rigid.
one to check whether the edges of the graph are not distributed spatially in
a uniform manner. If they are crowded locally, than the odds are that the
structure is not isostatic.
The isostatic concept so far discussed falls in the category of generic rigid-
ity, where only the topological information on a graph’s connectivity comes
into the picture. However, one may also deal with unexpected infinitesimal
motions when, say, two edges incident onto the same vertex lie on a straight
line. The applicable techniques for analyzing such geometric problems are re-
viewed in Guyon et al. (1990). For a review of problems in 3D see Guest (2000).
When dealing with very large systems—such as encountered in condensed
matter physics—we need to ask the question: what critical fraction, pr , of
edges of E needs to be kept so as to render the structure isostatic? We note
that we would have |E | = pr |E| new edges of thus modified, or depleted,
set |E |. It follows immediately from |E | = 2|V| − 3 that we would have
pr = 2/3. This value is a simple estimate of the so-called rigidity percolation, a
concept also useful in biophysics (Shechao Feng et al., 1985; Boal, 1993; Hansen
et al., 1996). As shown in these references, the actual critical point occurs at
a somewhat different value than 2/3; theoretical methods involved include
effective medium theories and spring network computations. The latter of
these will be demonstrated later on the example of Delaunay networks.
Finally, it is important to keep in mind that the rigidity percolation typi-
cally occurs above the connectivity percolation, that is, pr > pc . For example,
pc = 1/3 in a triangular central-force network, while pr = 2/3. This is demon-
strated in Figure 4.2(a) in terms of the planar bulk and shear moduli in function
of the volume fraction C2 of phase 2, which is 10−5 times softer than phase
1 of the undepleted network, thus simulating zero stiffness bonds. Note that
C2 = 1−C1 , and C1 ≡ p. That figure also shows the bulk and shear moduli of a

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
1
Truss shear modulus µ1/E1
0.9 Beam shear modulus µ2/E1
Truss bulk modulus κ1/E1
0.8 Beam bulk modulus κ2/E1
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
C2
(a)
1
Truss shear modulus µ1/E1
0.9 Beam shear modulus µ2/E1
Truss bulk modulus κ1/E1
0.8 Beam bulk modulus κ2/E1
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
C2
(b)
FIGURE 4.2
Dependence of bulk and shear moduli of triangular truss (central-force) and beam
networks, obtained by computational mechanics on 50 × 50 lattices at contrasts (a) 10−5
and (b) 10−2 .

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
triangular beam network (recall Section 3.4 of Chapter 3) at the same contrast
of phases. Clearly, rigid bonding and beam bending are essential in providing
the rigidity of the network right down to the point of connectivity percola-
tion at C2 = 2/3. The effect of loss of rigidity becomes less dramatic with a
decreasing contrast, see Figure 4.2(b). A much more extensive treatment of
rigidity percolation is in Sahimi (2003).
4.1.2 Application to Cellulose Fiber Networks

4.1.2.1 Rigidity of a Graph of Poisson Line Field Geometry
The planar Poisson line field was introduced in Chapter 1. If one reads the
basic assumptions of the classical article in paper physics/mechanics (Cox,
1952), one arrives at the conclusion that the geometry of a cellulose fiber
network assumed therein must be that of the Poisson line field. Note here
that:
• A homogeneous field of infinite lines cannot be obtained from a

random field of straight and finite segments by extending those
segments to infinity.
• The connectivity of the field of finite fiber segments—which clearly
is a more realistic representation of paper—is lower that that of the
Poisson line field.
• The Cox model also assumed the network of not interacting (!) infi-
nite lines/fibers to be held by a frame so as to make it solid-like.
Clearly, the last of these assumptions is not acceptable, but one way to
perhaps save the Cox model is first to note that under affine motions (ui =
εij0 x j , εij0 = const), straight lines transform into straight lines, and their original
(Lagrangian) points of intersection in plane are preserved. We now identify
the line segments (between any two consecutive intersections) to be edges of
E, and pivots to be vertices of V. Let us recall that the triple-fiber intersections
occur with probability zero for isotropic and anisotropic distributions of lines.
Thus, we typically have vertices of connectivity 4, that is, V4 .
Now, with reference to Figure 4.3(a), which shows a typical realization of
the Poisson line field, we see that there are two types of edges: those in direct
contact with the square-shaped window, and those entirely in the interior.
Clearly, the square window is needed to prevent these boundary layer bonds
from dangling, and this immediately renders the entire network a mechanism.
However, one may argue that the boundary layer of dangling bonds is very
thin relative to the whole field, and ask the question concerning the isostatic
condition for the graph G(V, E) representing the interior network of edges
not directly in contact with the square window boundary; these are shown in
bold in Figure 4.3(a). Here we observe that, while the V4 vertices occur in the
interior of this graph, its boundary involves V2 and V3 vertices. Now, since
there two vertices to every edge, we can calculate the total number of edges

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
(a) (b)
(c) (d)
FIGURE 4.3
Samples of (a) a planar Poisson line field and (b) a finite fiber field (recall Section 1.4 of Chapter
1), with a 1 = 1 and all other a i = 0. Test windows of size L × L are considered. (c) Deformation
of a network of (b), with 195 fibers with originally straight fibers, with fiber bending present,
subjected to axial strain ε11 = 1%. (d) The same network, with fiber bending almost absent,
subjected to axial strain ε11 = 1%. All displacements in (c) and (d) are magnified by a factor 8 for
clarity. Figure (d) shows large, mechanism-type motions of the network including those of some
fibers which spring outside the original domain of the network.
in the bold drawn graph G(V, E) according to
|E| = |V2 | + 32 |V2 | + 2|V4 |. (4.2)
Evidently, since V = V2 ∪ V3 ∪ V4 , the total number of all the vertices is
|V| = |V2 | + |V3 | + |V4 | (4.3)
so that |E| < 2|V| − 3 and the system is not isostatic, it is underconstrained
(i.e., a mechanism). Given this observation, the Poisson line field of axial force
fiber segments (the so-called Cox model, 1952) is not a valid model of paper,
or any other solid material for that matter. The fact that the Cox model does

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
give finite values for elastic moduli including the shear modulus is easily
explained by the presence of fully stretched fibers, spanning the entire test
window. The situation is analogous to a graph of a square lattice topology,
which, even though it is an obvious mechanism, will give finite axial mod-
uli in two directions, if fully stretched and subjected to kinematic boundary
conditions.
In real networks fibers have finite length, so their ends are loose. When
fiber ends are removed to eliminate their obvious mechanism motions, the
number of vertices in sets V2 and V3 , increases. Consequently, (4.1) is even
further from being satisfied. In order to deal with finite fiber effects, Cox and
others modified the basic model by a so-called shear-lag theory. However,
the latter assumes single-fiber segments to carry axial and shear forces only,
which (see Section 4.2), is not a valid model of a solid element: fiber bending
should also be included.
Paper exhibits finite stiffnesses in 2D as well as in 3D. In the latter case,
condition (4.1) is replaced by an even more stringent one as more constraints
are needed when dealing with the additional degrees of freedom (Asimov
and Roth, 1978).
4.1.2.2 Loss of Rigidity in a Fiber-Beam Network

Besides the foregoing structural topology considerations, there is another fact
that casts doubt on any fiber network model in which fiber segments are
joined by pivots. Namely, any two cellulosic fibers have a finite contact area of
hydrogen bonding (Page et al., 1961), which would be sheared by hinge-type
connections. While it is very difficult to assess experimentally to what extent
this region is deformable, our model will treat it as somewhat deformable in
the sense that bonds are rigid but have no dimension, and fiber segments are
treated as extensible beams from node to node of the graph G(V, E) (Ostoja-
Starzewski et al., 1999).
This modeling of mechanics of fiber networks is similar to that of a cement-
coated wood strands composite (Stahl and Cramer, 1998), as well as to the
one used for highly porous materials (Chung et al., 1996), and is based on the
following assumptions and steps:
1. Generate a system of finite-length straight fibers, such as shown in

Figure 4.3(b) according to specific geometric characteristics: distri-
bution of fiber lengths and widths, distribution of angular orienta-
tions (θ) of fiber chords, etc. The fibers are laid in three dimensions
on top of one another with a possible, slight non-zero out-of-plane
angle. θ ∈ [0, π], that is, the angle a fiber makes with respect to
the x-axis is modeled by the Fourier series-type probability density
function of Chapter 1.
2. Fibers are homogenous, but each fiber may have different dimen-
sions and mechanical properties, all sampled from any prescribed
statistical distribution.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
3. Each fiber is a series of linear elastic 3D extensible Timoshenko beam

elements. Each of these is described by a stiffness matrix written here
in an abbreviated form setup in a corotational coordinate system
(Cook et al., 1989):
 EA 
  0 0 0 0 0  
 F  l  L 

 
  0    

 T 
GJ
0 0 0 0  
 θx 



 
  l  
 


 Ma   l +3g
2
l −6g
2  

 0 0 0 0  θ a
y
=  4a 2a  y
  l +3h l −6h   a  (4.4)
 Mz 
2 2
 θz 
a
   0

0 0 4b
0 2b   

 
   

 My 
b
  0 0 l 2
−6g
0 l 2
+3g
0  θ yh 
 


 b   2a 4a  
 b  
Mz 0 0 0 l −6h
2
0 l +3g
2
θz
2b 4a
where
E Iy E Iz E Iy E Iz
g = 12 h = 12 a= b= .
GA GA l(12g + l 2 ) l(12h + l 2 )
(4.5)
Here F and T are the axial force and the twisting moment, while Mya ,
Mza , Myb , and Mzb are the bending moments around the y and z axes
at the a and b ends, respectively. Also, L, θx , θ ya , θza , θ yb , and θzb
denote axial elongation, angle of twist, and four angles of rotation.
Finally, l, A, J , Ix , and I y are, respectively, the length, cross-sectional
area, cross-sectional polar moment of inertia, and the moments of
inertia with respect to the x- and y-axes. E and G are the Young
modulus and shear modulus of a fiber-beam.
4. After identifying all the fiber–fiber intersection/contact points a
connectivity matrix is set up.
5. Equilibrium is found under the boundary condition ui = εij0 x j .
6. All six effective, in-plane stiffness coefficients are determined from
the postulate of equivalence of strain energy stored in a square-
shaped window of finite thickness with the strain energy of an
equivalent continuum.
The undeformed network, shown in Figure 4.3(b) in its top view, has
the following parameters: window size: 4 × 4 × 0.1 mm, a 1 = 1, and other
coefficients in equation (4.4) are zero; fiber length: 2 mm; fiber width: 0.04 mm;
fiber height: 0.015 mm. As a result of a Boolean process of fiber placement
(Chapter 1), we obtain: 195 fibers with an average of 4.8 bonds per fiber, the
whole system having 859 nodes with six degrees of freedom per node.
The state of deformation corresponding to axial strain ε11 = 8% is shown
in Figure 4.3(c). The analyzed strain is actually 1%, but displacements are
magnified for clarity. Compare this deformed network to that in Figure 4.3(d),
which shows the same network of fibers subjected to the same strain but with

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
the ratio of fiber flexural stiffness to fiber axial stiffness reduced by a factor of
10−4 . Note the following:
1. The sharp kinks we see in both figures are only the artifact of
simple computer graphics—the micromechanical model assumes
fibers deform into differentiable curves. Magnification creates the
appearance of large displacements—actually, an infinitesimal dis-
placement assumption is used in the computational mechanics pro-
gram.
2. The kinks are far more pronounced when fibers have low flexural
stiffness. Portions of the network where connected fibers do not
form triangular pores can generate significant forces in response
to deformation when fibers have high flexural stiffness, but they
cannot do so when fibers rely almost entirely on axial stiffness. These
portions of the network are not stable in the sense of loss of generic
rigidity discussed earlier.
3. We do not study this rigid-floppy transition by turning, in an ad
hoc fashion, all the connections into pivots. Rather, with the model
taking into account all the displacements and rotations of nodes, we
can study it as a continuous function of fiber slenderness; see also
Kuznetsov (1991). Note that this aspect is impossible to
investigate with models based on central-force potentials for single
fiber segments (e.g., Kellomaki et al., 1996). Our model also fills
a gap pointed out in Raisanen et al. (1997) consisting of a need
to set up finite element models of 3D disordered fiber networks,
yet avoids their simplistic mapping into electrical resistor networks
(i.e., second-rank tensor problems) of the same topology.
The fiber network model has also found an application in studies of

mesoscale stiffness and instability (splitting) of fibril networks found in a
bone’s lamellar structure (Jasiuk and Ostoja-Starzewski, 2004), see Figure 7.15
of Chapter 7.
4.2 Spring Network Models for Disordered Topologies

4.2.1 Granular-Type Media
4.2.1.1 Load Transfer Mechanisms
The spring network models are most natural when applied to systems that
have the same topology as the underlying lattice. One such example has
been discussed above: a cellulose fiber network. Another one is offered by a
granular medium. Here, the principal method of computational mechanics
analyses, dating back to Cundall and Strack (1979), is the so-called discrete
element (DE) model.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
2 V2
1 V1
3 V3
(a) (b)
V2 V2
kn
ka ka
V1 V1
kn
V3 V3
(d) (c)
FIGURE 4.4
(a) A cluster of three grains (1, 2, 3) showing the three lines of interactions; (b) a discrete element
model showing the normal force, the shear force and the moment exerted by grains 2 and 3
onto grain 1; (c) a most general model showing the same grain-grain interactions as before but
augmented by an internal, angular spring constant k a ; a simplified model showing only normal
(k n ) and angular (k a ) effects.
Let us employ a graph representation of the planar granular medium: a

graph G(V, E), whereby vertices of the set V signify grain centers and edges
of the set E represent the existing grain–grain interactions, Figure 4.4(a). We
fix an r − θ polar coordinate system at a grain center. There are several types
of the DE model that one may consider:
• Central interactions: the total energy is a sum total of central inter-

actions of all the edges
U = U cental , (4.6)
and this model is a generalization of the basic model of Section 3.2.3
in Chapter 3.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
• Central and angular interactions: the total energy is modified to
U = U cental + U angular , (4.7)
and this model is a generalization of what we saw in Section 3.2.4 in

Chapter 3. Continuum mechanics tells us that εθ θ equals r −1 ∂uθ /∂θ +
ur /r . This shows that the angular changes φ/φ between two adja-
cent edges V1 − V2 and V1 − V3 in Figure 4.4 correspond to r −1 ∂uθ /∂θ
in the r − θ polar coordinate system fixed at a grain center. This term
does not show up in two other expressions for εrr and εr θ . However,
ur /r is due to a radial displacement, and so φ/φ does not exactly
equal εθθ , which leads us to call it a “θθ -type strain.” We now adopt
the Kirkwood model (recall Chapter 3) to account for φ/φ in addi-
tion to the normal grain–grain interactions, and introduce angular
springs of constant k a acting between the edges V1 − V2 and V1 − V3
incident onto the node V1 , Figure 4.4(d); the edges remain straight
throughout deformation.
• Central, shear, and bending interactions: the total energy is
U = U cental + U shear + U moment , (4.8)
and this model is a generalization of what we saw in Sections 3.4.1

to 3.4.2 of Chapter 3. This is a typical DE model, which, of course,
may be termed a “locally inhomogeneous micropolar continuum,”
with inhomogeneity varying on the scale of grains; see the section
below.
• Central, shear, bending, and angular interactions: the total energy is
U = U cental + U shear + U moment + U angular . (4.9)
One may argue that the three-point interaction should be introduced in

the DE models so as to better represent the micromechanics, and to make,
in accordance with Figure 4.4(c), the strain energy stored in a single Voronoi
cell equal to (4.9). However, there exist successful DE models that account for
normal and shear forces only (Bathurst and Rothenburg, 1988a,b); this neglect
of the contact moment is justified by the fact that only small numerical errors
are thus caused in problems of interest in granular materials. In the case of a
regular, triangular array of disks, this model is equivalent to a classical Born
model of crystal lattices, which is known to lack the rotational invariance
(Jagota and Bennison, 1994).
4.2.1.2 Graph Models

Let us pursue the planar graph representation of granular media in some more
detail (Satake, 1976, 1978). First, we list in Table 4.2 a correspondence between
a system of round grains and its graph model. Besides the vertex (V) and edge
(E) sets introduced earlier, we also have a loop set L. With this geometric

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
TABLE 4.2
Granular Structure and Graph Terminology
Assembly of grains Graph Index No. of elements
Grain Vertex v |V|

Contacting point Edge e |E|
Void (in 2D) Loop l |L|
Note: After Satake (1978).
reference, we can then set up an assignment of load quantities—forces on the

left and corresponding kinematic measures on the right—in Table 4.3.
The connectivity of the graph is described by the incidence matrix Dve .
Let us write down a total of 3|V| scalar equilibrium equations, each one with
respect to a typical grain of radius r v and volume V v
e
F Bv
D ve
+V v
= 0, (4.10)
Me Nv
where

0 Dve
ve
D = v ve e . (4.11)
r D n × Dve
Here ne is the unit vector of edge e in the nonoriented graph. The operator
Dve (and its dual Dev mapping from vertices into edges) also plays a key role
in the kinematics of all the edges:

ue uv
= −D ev
= 0, (4.12)
we wv
where

Dev −ne × Devr v
ev =
D . (4.13)
0 Dev
The kinematics is subject to 3|L| compatibility constraints written for all the
loops, where we make a reference to Satake’s work.
TABLE 4.3
Load Versus Kinematic Quantities
Load quantity Notation Number of elements Notation Kinematic quantity
Body force Bv |V| uv Grain displacement

Body couple Nv |V| wv Grain rotation
Contact force Fe |E| ue Relative displacement
Contact couple Me |E| we Relative rotation
Note: After Satake (1978).

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
The above should be augmented by 3|E| constitutive equations connecting

the contact force Fe and contact moment Me with the relative displacement
ue and relative rotation we . Given three global equilibrium conditions, we
have a total of
3(|V| − 1 + |E| + |L|) = 6|E| (4.14)
equations. Taking note of the Euler formula |V| − |E| + |L| = 1 in 2D (as
opposed to that in 3D in Section 1.4 of Chapter 1), we see that this budget of
equations agrees with the total of 6|E| unknowns: Fe , Me , ue , and we , all
defined on edges of the set E.
Finally, we note a formal analogy of (4.10) and (4.12) to the equilibrium
and strain-displacement equations of Cosserat continua (Chapter 6)

σ b γ u
div + =0 = grad . (4.15)
µ m κ w
The similarity of compatibility relations for graph and continuum descrip-

tions has also been shown by Satake (1976, 1978). See Goddard (2006) for an
in-depth analysis of these issues.
Applications of graph models to cell biomechanics have been explored
in Hansen et al. (1996, 1997). While this section focuses on classical contin-
uum modeling, multifield models for granular-type materials are analyzed
in Trovalusci and Masiani (2005).
4.2.1.3 Periodic Graphs with Disorder

Randomness may be introduced into the periodic networks in various ways.
Figure 4.5 displays two basic possibilities: substitutional disorder and topolog-
ical disorder. The first of these connotes a variability in properties per vertex
(or node), and the second consists in a departure from the periodic topol-
ogy. There is also a third case, of much more interest in solid-state physics:
geometric disorder, which involves the variability in the geometry of a net-
work’s structure—like uneven lengths of various bonds—but preserving a
(a) (b)
FIGURE 4.5
Substitutional (a) versus topological disorder (b) of a hard-core Delaunay network.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
topological periodic structure (Ziman, 1979). The topological disorder is typ-

ically caused by an incompatibility of crystal-like domains in a granular ma-
terial. For example, the material may consist of equisized disks, which are
organized into regular, periodic arrays, but the fact that they happen to be
differently oriented in space causes an irregular structure and network of
domain boundaries.
As observed earlier in connection with the DE model carrying torques,
the topological disorder leads to a locally inhomogeneous polar, or micropo-
lar, continuum—depending on the type of vertex–vertex interactions—with
inhomogeneity varying on the scale of grains. Such a continuum model con-
tains a lot of information, but, in the first place, one wants to establish the
eff
effective, in the macroscopic sense, first-order moduli Ceff (≡ Cijkl ) of the mate-
rial. These are obtained from the so-called periodic boundary conditions on an
L × L square B:
ui (x + L) = ui (x) + εij0 L j ti (x) = −ti (x + L) ∀x ∈ ∂ B. (4.16)
Here εij0 is the macroscopic strain, ti is the traction on the boundary ∂ B of B,

and L = Lei , with ei being the unit base vector. The periodicity means that
the network topology is modified so as to repeat itself with some periodicity
L in x1 and x2 directions, whereby L is usually taken much larger than the
typical vertex-vertex spacing (or edge length).
Now, the periodic conditions (4.16) require that a periodic network be
set up, and this, in turn requires a periodic Poisson point field on the L × L
square; recall Chapter 1. Topologically, our square turns into a torus, but, as
Figure 4.6 shows, there are also three other possibilities.
B B A A
Cylinder
Torus
A A A A
(a) (c)
B A A A
Möbius band
Klein bottle
A B A A
(b) (d)
FIGURE 4.6
Mapping of edges of a a square-shaped domain resulting in (a) a cylinder, (b) a Möbius band,
(c) a torus and (d) a Klein bottle. The torus corresponds to periodic boundary conditions.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
FIGURE 4.7
A periodic Poisson–Delaunay network with 200 vertices.
A typical realization B(ω) of a periodic Poisson–Delaunay network, num-

bering 200 vertices, is shown in Figure 4.7. The set B ={B(ω); ω ∈ } forms
the random medium; a single ω indicates one realization of the Poisson point
field and a chosen assignment of spring constants. In actual simulations only
a minute subset of the entire sample space can be investigated, but by the
standard Monte Carlo and ergodicity arguments, this subset is representative
of the whole system. Thus, already the response of a single network much
larger than the grain size is sufficient to gain a good estimate of Ceff . The
ensemble average of that tensor is isotropic for a microstructure of space-
homogeneous and isotropic statistics, but, with the number of vertices large,
even one realization of the network should be close to isotropic.
Using the formula for the strain energy of a 2D elastic continuum of volume
V = L 2:

V 1 0 0
U= κεii εjj + 2 εij εij − εii εjj ,
0 0 0 0
(4.17)
2 2
leads, with reference to Chapter 5, to planar bulk and shear moduli:
2U(1) 2U(2)
κ= µ= 0 2 . (4.18)
0 2
V ε11
0
+ ε22 V ε12
0
Here U(1) and U(2) denote energies under dilatational (ε11 = ε22
0
) and shear
0
(ε12 ) strains, respectively. The moduli computed this way were compared to
various models in Ostoja-Starzewski et al. (1995):

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
• A symmetric self-consistent approximation for elliptical disks with

perfectly bonded interfaces, which treats the medium as a mixture
of N phases without distinguishing any one as a matrix.
• An asymmetric self-consistent approximation for elliptical disks
with perfectly bonded interfaces, which treats the medium as a mix-
ture of N − 1 phases embedded in a matrix.
• An asymmetric self-consistent approximation for circular disks with
perfectly bonded interfaces.
• An asymmetric self-consistent approximation for circular disks with
springy interfaces.
• Voigt and Reuss bounds for circular disks with springy interfaces.
Comparisons to such continuum models are possible providing one inter-

prets a vertex and all the edges incident onto it as a “spider,” which allows
interpretation of that object as a continuum-type inclusion. As a result we
have a “spider-inclusion” analogy of Figure 4.8.
4.2.2 Solutions of Truss Models

A specialized method of determination of apparent moduli applies in the case
of truss-type microstructures. Thus, let us focus on the response of a window,
arbitrarily larger than the average edge length, subjected to displacement
boundary conditions, Figure 4.9 (Ostoja-Starzewski and Wang, 1989). First,
a window is cut out of the network as indicated by a square with dashed
lines in Figure 4.9(a), and treated as an independent body (b)—in this case,
given the assumption of central forces in all the Delaunay edges, a planar
truss. At this stage, all the boundary points (on four sides) of the window
are pinned, and this is where the displacement condition is applied—see
equation (4.23) below. For example, a state of uniaxial extension and shear is
depicted in Figure 4.9(c). The actual mechanical response of this planar truss
is now solved by a structural mechanics method.
FIGURE 4.8
A periodic Poisson–Delaunay network with 200 vertices, showing a “spider” of edges incident
onto a vertex in the figure on the right.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
(a) (b) (c)
FIGURE 4.9
A finite window (b) cut out of a Delaunay network (a), and subjected to uniform displacement
boundary conditions in (c).
We introduce a vector notation, so that the effective mechanical response

of the window is described by
    0 
 σ 11 
  C11 C12 C13   ε11 

 
σ 22 ≡ {σ } = [C]{ε } ≡  C21 C22 C23  ε22
0 0
. (4.19)

 
  0 
 
σ 12 C31 C32 C33 ε12
The structural mechanics of the truss is governed by
[K ]{u} = {F }, (4.20)
where [K ], {u}, and {F } are the stiffness matrix of the system and the dis-
placement and force vectors at the nodes, respectively. Noting that the set of
global equilibrium equations governing all the elements may be partitioned
into those corresponding to the degrees of freedom at the window boundaries
(b) and the ones in the interior (i), we have

K (ii) K (ib) u(i) F (i)
[K ] = , {u} = , {F } = . (4.21)
K (bi) K (bb) u(b) F (b)
The net force on any interior node (and, thus, any interior degree of freedom)
must be zero, F (b) = 0, so that the static condensation gives
[K ]{u(b) } = {F (b) } [K ] = [K (bb) ] − [K (bi) ][K (ii) ]−1 [K (ib) ]. (4.22)
For any boundary node the displacement can be written as

 0 
 ε11 
u1(b) x1 0 x2 /2  0 
= [A]{ε} ≡ ε22 , (4.23)
u2(b) 0 y2 x1 /2  0  
ε12

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
while for an interior node we have
[u(i) ] = [K (ii) ]−1 [K (ib) ][A]{ε 0 }. (4.24)
Now, noting that the strain energy of the window is

1 (b) T (b) 1
U= {u } {F } = {u(b) }T [K ]{u(b) }, (4.25)
2 2
we obtain, by equivalence with the continuum model, the effective moduli
1
[C] = [A]T [K ][A], (4.26)
V
where the strain-displacement matrix [ A] is obtained through the assembly
of the nodal matrices for all the boundary nodes.
In the case of traction boundary conditions
{F (b) } = [H]{σ }, (4.27)
a force-based homogenization technique has been proposed (Huyse and Maes,
2001). Here [H] is a matrix that depends on the actual boundary’s microge-
ometry. The structural mechanics of the truss is now rewritten as

f (ii) f (ib)
[ f ]{F } = {u} [ f ] = , (4.28)
f (bi) f (bb)
where [ f ] is the flexibility matrix and the strain energy of the window is
1 (b) T (b) 1
U= {u } {F } = {F (b) }T [ f (b) ]{F (b) }. (4.29)
2 2
From the above equations we obtain, by equivalence of the discrete with the
continuum model, the apparent compliances
1
[S] = [H]T [ f (bb) ][H], (4.30)
V
where [ f (bb) ] is obtained through an inversion.
4.2.3 Mesoscale Elasticity of Paper

4.2.3.1 Dilemma of Special In-Plane Orthotropy of Paper
Paper is one of the most challenging engineered materials. The difficulties in
understanding it are due to its complex multiscale structure—this necessitates
random fields of Chapters 1 and/or 2, random network models of this chapter,
(non)classical elasticity of Chapters 5 and 6, as well as micromechanics models
discussed here and also later. Things get still more complicated when the
inelastic behavior is considered, see Chapter 9.
For the sake of reference, because we are dealing with the machine-made
(i.e., oriented) paper, for the linear elastic tensile range, an orthotropic model

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
is adopted (e.g., Uesaka et al., 1979):

E1 E 2 ν12
σ11 = ε11 + ε22
(1 − ν12 ν21 ) (1 − ν12 ν21 )
E 1 ν21 E2
σ22 = ε11 + ε22
(1 − ν12 ν21 ) (1 − ν12 ν21 )
τ12 = Gγ12 = 2Gε12 (4.31)
where the engineering constants are:
E 1 = Young’s modulus in x1 -direction (MD)

E 2 = Young’s modulus in x2 -direction (CD)
−ε22
ν12 = is the Poisson ratio for strain in x2 -direction (CD) when
ε11
paper is stressed in x1 -direction (MD) only
−ε11
ν21 = is the Poisson ratio for strain in x1 -direction (MD) when
ε22
paper is stressed in x2 -direction (CD) only
G = shear modulus in x1 x2 -plane
Here MD (x1 ) and CD (x2 ) are the so-called machine- and cross-directions,
respectively. To ensure that the stiffness matrix is orthotropic, we require
E 1 ν21 = E 2 ν12 . (4.32)
The shear modulus G is very difficult to measure by quasi-static experiments,

but ultrasonic experiments on many paper materials indicate that the follow-
ing relation is approximately satisfied (Horio and Onogi, 1951; Campbell,
1961):
E1 E2
G= . (4.33)
E 2 + E 1 + E 1 ν21 + E 2 ν12
Thus, paper has a special in-plane orthotropy property where only three elas-
tic constants are independent. This “special orthotropy” means that the shear
modulus is (approximately) invariant with respect to rotations of the coordi-
nate system (Horio and Onogi, 1951; Campbell, 1961). Another way to express
this invariance, fully equivalent to (4.33) above, is in terms of the in-plane
compliances
4S1212 = S1111 + S2222 − S1122 . (4.34)
Note here that “approximately” reflects the fact that the compliance S is an
apparent (i.e., mesoscale) property, typically measured on scales of centi-
meters, and being actually random since we are below the RVE.
Note: Given the multiscale structure of paper shown in Figure 4.10, there
is no clear length scale at which homogenization to a perfect RVE can be

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
ZD
(a)
CD MD
(b)
(c)
P S1 S2 S3
35A (d)
100A
(e) H
CH2OH H1 OH
H H
CH2OH H OH
H
O O O O
H OH H H OH H
OH H H OH H H
O H H O O H H O O
H OH CH2OH H OH CH2OH
(f )
FIGURE 4.10
A hierarchy of scales in paper: (a) roll of paper on a paper machine (up to 10 m wide, thousands
of kilometers long) with a possible presence of streaks; (b) paper sheet (scale of centimeters);
(c) random fiber network (scale of millimeters); (d) cellulose fiber with its layers P, S1, S2, S3
(all made of fibrils) and the lumen (microns); (e) fibrils; (f) molecular chains.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
carried out; see the discussion of RVE in Chapters 7 to 10. Although the con-
stitutive law above applies to length scales of centimeters, specimens display
some scatter there. Going to scales of tens of centimeters, one begins to see
streaks, while on the scales of meters there is a strong dependence on the
cross-direction of the paper machine. There are also complex fluctuations in
the machine direction, which do not qualify as a WSS random field.
Staying on the scale of centimeters, the question becomes: Why does the
relationship (4.34), in fact hardly found in other materials, occur in paper?
4.2.3.2 Explanation via Random Fiber Network

The above question is addressed by Ostoja-Starzewski and Stahl (2000) with
the help of models introduced earlier. First, we employ the germ-grain model
of Chapter 1, constructed so as to obtain nonuniform fiber flocculation, com-
monly seen as spatially nonuniform grayscale effects in a typical sheet of
paper held against light. Next, we take the computational mechanics model
of fiber networks of Chapter 3. In brief, fibers are placed in 3D with possible
non-zero angles to control out-of-plane orientation of the fiber axis and the
“roll” of the fiber about its own axis. Each fiber, depending on its contacts
with other fibers, is a series of linear elastic three-dimensional Timoshenko
beam elements e, also with torsional response included, between bonds with
other fibers. In the analysis of each and every body Bδ (ω), having solved for
displacements and rotations of all the bonds under uniform kinematic bound-
ary conditions ui = εij0 x j , we establish the network’s effective stiffness tensor
Cijkm (ω) from a postulate of equivalence of the total strain energy of all the
networks’ elements e, Utot , with that of an effective continuum (V = L x L y t)
V
Utot = εij Cijkm (ω)εkm = [Ueaxial + Ueshear + Uemoment + Uetorsion ]. (4.35)
2 e∈E
It may be argued, by virtue of a qualitative analogy to in-plane conductiv-

ity of networks of identical geometry, that the kinematic boundary condition
allows a much faster asymptotic scaling to the RVE than the uniform traction
boundary condition ti = σji0 n j . The point is that fiber networks are somewhat
analogous to plates with holes, reported in Section 3.4.3 of Chapter 3, recall
Figure 3.11 there.
Let us now go back to Figure 1.11 of Chapter 1. It displays two networks
differing only by the degree of flocculation. That is, parameter b equals 2.0
in (a) and 0.4 in (b), so that we have a highly versus a weakly flocculated
network. Their responses under overall ε11 0
-strain are shown in the middle
figures, while the figures on the right show the difference between the result-
ing displacement of each node and what the displacements would be if the
strain field in the interior of the network had been uniform. In the latter case,
the figures would consist simply of dots; thus, the lines represent deviation
from uniformity. Let us first note that the deviation from uniform displace-
ments is apparent in both networks, but it is certainly higher in the network
with b = 2.0 (strong flocculation) than with b = 0.4. (weak flocculation).

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
This is so because fibers belonging to a given floc move together, and this
motion, being different from the affine one, resembles a swirl.
Now, it turns out that the special orthotropy relation (4.34) is satisfied
for a network with b between the values employed here. Various parameters
of the network may be altered, but some degree of flocculation—neither too
much nor too little—is always necessary. [In our original paper the factor “4”
in equation (4.34) was inadvertently missed, but this had no effect on the
results since we worked with G in computational mechanics.] To sum up,
this analysis shows that random network geometry involving two scales—
(1) random fiber placement within a floc with random angular orientation
and (2) flocculation modeled by the germ-grain process—together with a
network model possessing generic rigidity (thanks to beam-type fibers) offers
an explanation of the peculiar property of many papers.
4.2.4 Damage Patterns and Maps of Disordered Composites

As mentioned in Chapter 3, the lattice method can also be used to simulate
damage of heterogeneous materials. This works particularly well in the case of
elastic-brittle failure of composites, where one uses a mesh (much) finer than
the typical size of the microstructure. In principle, one needs to determine
which lattice spacing ensures mesh-independence, or nearly so. Such a study
has been conducted for a thin aluminum polycrystalline sheet discussed in
Section 3.3 of Chapter 3.
Following Ostoja-Starzewski et al. (1997) and Alzebdeh et al. (1998), we
now focus on two-phase composites in antiplane shear, under periodic bound-
ary conditions and (necessarily) periodic geometries. Now, since both phases
(inclusion i and matrix m) are isotropic and elastic-brittle (Figure 4.11a), the
σ C i/C m
Stiff Stiff
i
C weak strong
Cm
εicr /εm
cr
εicr εm ε Stiff Soft

cr
weak strong
(a) (b)
FIGURE 4.11
(a) Elastic-brittle stress-strain curves for matrix and inclusion phases; (b) sketch of the damage
plane.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
10.0
C i/ C m
1.0
0.1
0.1 1.0 10.0

i m
ε/ε
FIGURE 4.12
Crack patterns in the damage plane on a scale 4.5 times larger than that of the inclusion diameter.
The center figure of a homogeneous body is not shown as it corresponds to all the bonds failing
simultaneously.
composite can be characterized by two dimensionless parameters
εcr
i
/εcr
i
C i /C m , (4.36)
where εcri
is the strain-to-failure of either phase, and C its stiffness. This leads
to the concept of a damage plane (Figure 4.11b), where we see various combi-
nations of strengths and stiffnesses. While the response in the first and third
quarters of damage plane is quite intuitive, this is not so for the second and
fourth quarters. In those two quarters there is a competition of either high
stiffness with low strength of the inclusions with the reverse properties of the
matrix, or the opposite of that. The damage plane is useful for displaying ef-
fective damage patterns of any particular geometric realization of the random

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
σ/σm
σ/σm
σ/σcr
cr
cr
10.0
ε/εm
cr ε/εm
cr ε/εm
cr
m
σ/σm
σ/σcr
cr
m
1.0
C /C
i
m m
ε/εcr ε/εcr
m
σ/σm
σ/σm
σ/σcr
cr
0.1 cr
m m m
ε/εcr ε/εcr ε/εcr
0.1 1.0 10.0

i m
ε /ε
FIGURE 4.13
Damage maps of statistics of constitutive responses for twenty realizations of the random com-
posite, such as that in Figure 4.12.
composite while varying its physical properties (Figure 4.12) as well as other
characteristics, say, statistics of response in the ensemble sense (Figure 4.13).
A number of other issues are studied in the referenced papers:
• Stress and strain concentrations

• Finite size scaling of response
• Function fitting of statistics (where it turns out that the beta proba-
bility distribution offers a more universal fit than either Weibull or
Gumbel)
• Effects of disorder versus periodicity
Also, see Ostoja-Starzewski and Lee (1996) for a similar study under in-plane
loading; computer movies of evolving damage are at http://www.mechse.
uiuc.edu/research/martinos.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
4.3 Particle Models

4.3.1 Governing Equations
4.3.1.1 Basic Concepts
Particle models are a generalization of lattice models to include dynamic
effects, and can also be viewed as an offshoot of molecular dynamics (MD).
The latter field has developed over the past few decades in parallel with the
growth of computers and computational techniques. Its objective has been to
simulate many interacting atoms or molecules in order to derive macroscopic
properties of liquid or solid materials (Greenspan, 1997, 2002; Hockney and
Eastwood, 1999). The governing Hamiltonian differential equations of motion
need to be integrated over long time intervals so as to extract the relevant
statistical information about the system from the computed trajectories.
Techniques of that type have been adapted over the past two decades
to simulate materials at larger length scales, whereby the role of a particle
is played by a larger-than-molecular piece of material, a so-called particle or
quasi-particle. The need to reduce the number of degrees of freedom in complex
systems has also driven models of galaxies as systems of quasi-particles, each
representing lumps of large numbers of stars. In all these so-called particle
models (PM), the material is discretized into particles arranged in a periodic
lattice, just like in the spring network models studied in earlier sections, yet
interacting through nonlinear potentials, and accounting for inertial effects,
that is, full dynamics. As shown in Figure 4.14, the lattice may be in 2D or
in 3D.
Note that, by comparison with finite elements (FE), which indeed also
involve a quite artificial spatial partitioning, PMs are naturally suited to in-
volve interparticle potentials of the same functional form as the interatomic
potentials, providing one uses the same type of lattice. The PM can therefore
take advantage of the same numerical techniques as those of MD, and rather
easily deal with various highly complex motions. Thus, the key issue is how
to pass from a given molecular potential in MD to an interparticle potential
in PM. In the case when the molecular interactions are not well known, the
PM may still turn out to be superior relative to the FE. Among others, this
indeed is the case with comminution of minerals where scales up to meters
are involved (Wang and Ostoja-Starzewski, 2006; Wang et al., 2006).
In MD, the motion of a system of atoms or molecules is governed by clas-
sical molecular potentials and Newtonian mechanics. As an example, let us
consider copper. Following Greenspan (1997), its 6–12 Lennard-Jones poten-
tial is
1.398068 −10 1.55104 −8

φ(r ) = − 10 + 10 erg. (4.37)
r6 r 12

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
(a)
Z
Y
X
–4
Z (cm)
0 –2
m)
0
X (c
2
–6 –4 –2 0 2 4 4
6
Y (cm)
(b)
FIGURE 4.14
Particle models and intermediate stages of fracture in (a) 2D and (b) 3D. (After Wang et al. [2006]).
Here r is measured in Å. It follows that the interaction force between two
copper atoms is
dφ(r ) 8.388408 −2 18.61248
F (r ) = − =− 10 + dyn. (4.38)
dr r7 r 13
In (4.38) F (r ) = 0 occurs at r0 = 2.46Å, and φ then attains the minimum:
φ(r0 ) = 3.15045 · 10−13 erg.
Let us next recall Ashby and Jones’s (1980) simple method to evaluate
Young’s modulus E of the material from φ(r ): compute
S0 d 2 φ(r )
E= where S0 = |r . (4.39)
r0 dr 2 0
With this method, we obtain Young’s modulus of copper as 152.942 GPa,
a number that closely matches the physical property of copper and copper
alloys valued at 120∼150 GPa. Ashby and Jones also defined the continuum-
type tensile stress
σ (r ) = NF (r ), (4.40)

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
where N is the number of bonds/unit area, equal to 1/r02 . Tensile strength,

σTS , results when d F (r )/dr = 0, that is, at rd = 2.73Å (bond damage spacing),
and yields
σTS = NF (rd ) = 462.84 MN/m2 . (4.41)
This value is quite consistent with data for the actual copper and copper-based
alloys reported at 250∼1000 MPa.
In the PM, the interaction force is also considered only between nearest-
neighbor (quasi-)particles and assumed to be of the same form as in MD
G H
φ(r ) = − p
+ q. (4.42)
r r
Here G, H, p, and q , all positive constants, are yet to be determined, and this
will be done below. Inequality q > p must hold so as to obtain the repulsive
effect that is necessarily (much) stronger than the attractive one. Three
examples of interaction force for three pairs of p and q are displayed in
Figure 4.15(a). The dependence of Young’s modulus for a wide range of p
and q is shown in Figure 4.15(b).
The conventional approach in PM, just as in MD, is to take the equation
of motion for each particle Pi of the system as

d 2 ri Gi Hi rji
mi 2 = α − p + q i=
j, (4.43)
dt j
rij rij rij
where mi is the mass of Pi and rji is the vector from P j to Pi ; summation is

taken over all the neighbors of Pi . Also, α is a normalizing constant obtained
by requiring that the force between two particles must be small in the presence
of gravity:

Gi Hi

α − p + q < 0.001 · 980mi . (4.44)
D D
Here D is the distance of local interaction (1.7r0 cm in this particular example),

where r0 is the equilibrium spacing of the particle structure. The reason for
introducing the parameter α by Greenspan (1997) was to define the interaction
force between two particles as local in the presence of gravity. However, since
setting α according to (4.44) would result in a “pseudo-dynamic” solution,
we set α = 1.
According to equation (4.42), different ( p, q ) pairs result in different
continuum-type material properties, such as Young’s modulus E. Clearly,
changing r0 and volume of the simulated material V (= A × B × C) will
additionally influence Young’s modulus. Therefore, in general, we have some
functional dependence
E = E( p, q , r0 , V). (4.45)

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
20
10
Interaction force (KN)
–10
(p, q) = 3, 5
(p, q) = 5, 10
–20 (p, q) = 7, 14
–30
–40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Equilibrium position r0 (cm)
(a)
15
14
13
12
11
10
9
q
8
7
6
5
4
3
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14
p
Young’s modulus (GPa)
14 37 59 82 104 127 149 172 195 217 240 262 285 307 330
(b)
FIGURE 4.15
(a) The interaction force for pairs of ( p, q ) exponents, at r0 = 0.2 cm. (b) The variability of Young’s
modulus in the ( p, q )-plane.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
In Wang and Ostoja-Starzewski (2005) we formulated four conditions to deter-

mine continuum-level Young’s modulus and tensile strength, while maintain-
ing the conservation of mass and energy of the particle system and satisfying
the interaction laws between all the particles in the PM model for a given MD
model.
4.3.1.2 Leapfrog Method

Just like in MD, there are two commonly used numerical schemes in parti-
cle modeling: the completely conservative method and the leapfrog method. The
first scheme is exact in that it perfectly conserves energy, linear, and angular
momentum, but requires a very costly solution of a large algebraic prob-
lem. The second scheme is approximate. Since in most problems one needs
large numbers of particles to adequately represent a simulated body, the com-
pletely conservative method is unwieldy and, therefore, usually abandoned in
favor of the leapfrog method (Hockney and Eastwood, 1999). That method
is derived by considering Taylor expansions of positions ri,k+1 and ri,k of the
particle Pi (i = 1, 2, . . . , N) at times tk = kt and tk+1 = (k +1)t, respectively,
about time tk+1/2 = (k + 1/2)t (with t being the time step):
t t 2 t 3
ri,k = ri,k+1/2 − vi,k+1/2 + ai,k+1/2 − ȧi,k+1/2 + O(t 4 )
2 8 48
t t 2 t 3
ri,k+1 = ri,k+1/2 + vi,k+1/2 + ai,k+1/2 + ȧi,k+1/2 + O(t 4 ). (4.46)
2 8 48
Here vi and ai denote velocity and acceleration. Upon addition and subtrac-
tion of these we get the new position and velocity
t 2
ri,k+1 = 2ri,k+1/2 − ri,k + ai,k + O(t 4 )
4
vi,k+1/2 = (ri,k+1 − ri,k )/t + O(t 2 ), (4.47)
which shows that the position calculation is two orders of magnitude more
accurate than the velocity calculation. However, the error in computation
of velocity accumulates only as fast as that in position because it is really
being calculated from positions. It is easy to see that the leapfrog method
is more accurate than the conventional Euler integration based on vi,k+1 =
vi,k + (t)ai,k and ri,k+1 = ri,k + (t)vi,k .
Often the leapfrog formulas relating position ri , velocity vi , and accelera-
tion ai for all the particles Pi (i = 1, 2, . . . , N) are written as
t
vi,1/2 = vi,0 + ai,0 (starter formula)
2
vi,k+1/2 = vi,k−1/2 + (t)ai,k k = 0, 1, 2, . . . (4.48)
ri,k+1 = ri,k + (t)vi,k+1/2 k = 0, 1, 2, . . . .

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
Clearly, the name of the method comes from taking velocities at intermediate
time steps relative to positions and accelerations; it is also known as a Verlet
algorithm.
It can be shown that the global (cumulative) error in position going from
ri,k to ri,k+n (i.e., over T = nt) of Pi is
error(ri,k+n − ri,k ) = O(t 2 ). (4.49)
By the argument following (4.46) above, the global error in velocity is also
O(t 2 ).
Stability is concerned with the propagation of errors. Even if the trunca-
tion and round-off errors are very small, a scheme would be of little value if
the effects of small errors were to grow rapidly with time. Thus, instability
arises from the nonphysical solution of the discretized equations. If the dis-
crete equations have solutions that grow much more rapidly than the correct
solution of the differential equations, then even a very small round-off error
is certain to eventually seed that solution and render the numerical results
meaningless. By the root locus method for an atomistic unit of time, the safe
time-step used in the leapfrog method meeting this requirement is
1/2
1 dF
t 2 = . (4.50)
m dr max
We see that, as r → 0, d F /dr → ∞, which results in t → 0. Because this
may well cause problems in computation, we introduce the smallest distance
between two particles according to these conditions:
1. For a stretching problem of a plate/beam, take (dF/dr) max dF/

dr|r =r0 , which with (4.50) dictates t 10−7 − 10−6 s.
2. For an impact problem, one often needs to set up a minimum dis-
tance limiting the spacing between two nearest particles, e.g., rmin =
0.1r0 . It is easy to see from Figure 4.15(a) that, in this case, this suit-
able time increment is greatly reduced because of a rapid increase
in . This leads to t 10−8 s.
Following the MD methodology (Napier-Munn√ et al., 1999), we can also

set up a criterion for convergence: t < 2 m/k, where m is the smallest mass
to be considered, k is the same stiffness as S0 in (4.39)2 . An examination of
these two criteria shows there is not much quantitative difference between
them in the case of elastic or elastic-brittle, but not plastic, materials.
4.3.2 Examples
The maximum entropy formalism of Section 2.6 in Chapter 2 is much more
suited to deal with quasi-static rather than dynamic fracture. The dynamic
character of fracture in these experiments, combined with the presence of

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
multiple incipient spots, was also a big challenge in several computational me-
chanics models reviewed in Al-Ostaz and Jasiuk (1997) employing
commercial finite element programs, as well as in an independent study using
a meshless element program (Belytschko et al., 1995). Upon trying various fail-
ure criteria, several numerical methods, and even being forced to initialize the
cracking process by a subjective choice in the meshless model, the modelers
(including this author) have run into uncertainty as to which modeling as-
pect is more critical, and whether there is a way to clarify it. The recent study
by Ostoja-Starzewski and Wang (2006) was motivated by this outstanding
challenge, and offered a way to test the PM vis-à-vis experiments.
In Section 2.6 of Chapter 2 we considered crack patterns in epoxy plates
perforated by holes. Two analyses—one based on the minimum potential en-
ergy formulation and one based on the maximum entropy method—relied on
the assumption of quasi-static response. Strictly speaking, although the load-
ing was static, the fragmentation event was dynamic. Clearly, the preparation
of mineral specimens—involving measurement of highly heterogeneous and
multiphase microstructures—for a direct comparison with the model predic-
tion is very hard. Thus, we can apply the model to the experimentally tested
plate with 31 holes and follow this strategy:
1. Decrease the lattice spacing until we attain mesh-independent crack
patterns.
2. Find out whether the lattice of (1) will also result in the most domi-
nant crack pattern of Figure 4.16. Indeed, the crack patterns “stabi-
lize” as we refine the mesh.
3. Assuming the answer to (2) is positive, introduce weak perturba-
tions in the material properties—either stiffness or strength—to de-
termine which one of these has a stronger effect on the deviation
away from the dominant crack pattern, that is, on the scatter in
Figure 4.16.
4.3.2.1 Other Models

The PM is but one of the variations on the theme of MD. Here are some other
possibilities:
• Molecular statics (MS)—by disregarding the inertia forces, it in-
volves a static solution of the system of atoms (Chen, 1995; Vino-
gradov, 2006). Although the MD allows simulations of large systems
with a constraint to very short time scales (transient phenomena of
the order of nanoseconds), the MS allow large (macroscopic type)
time scales albeit with a limitation by the size of a (nonlinear) alge-
braic system one is able to solve and a restriction to 0 K.
• Derivation of a continuum model from a microscopic model based
on the assumption that the displacements on the macroscopic level
are the same as those on the molecular level (Blanc et al., 2002).

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
FIGURE 4.16
Final crack patterns for four mesh configurations at ever finer lattice spacings: (a) r0 = 0.1 cm;
(b) r0 = 0.05 cm; (c) r0 = 0.02 cm; (d) r0 = 0.01 cm. (From Ostoja-Starzewski and Wang, 2006.
With permission.)
• Introduction of a finite extension and spin for continuum-type par-

ticles (Yserentant, 1997).
• Direct incorporation of interatomic potentials into a continuum anal-
ysis on the atomic scale (Zhang et al., 2002).
• Computational mechanics of granular media (Hermann and
Luding, 1998).

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
4.4 Michell Trusses: Optimal Use of Material

4.4.1 Study via Hyperbolic System
4.4.1.1 Governing Equations
The classical problem of an optimal truss concerns a minimum-weight design
of a planar truss T that transmits a given load to a given rigid foundation
with the requirement that the axial stresses in the bars of the truss stay within
an allowable range σ0 ≤ σ ≤ σ0 (Michell, 1904). This forms the basis for a
problem of layout of a truss whose locally orthogonal members are of a rigid-
perfectly plastic material with tensile and compressive stresses ±σ0 , (Saveè
and Prager, 1985; Rozvany et al., 1995, and references therein). The solution
to this problem is provided by a so-called Michell truss-like continuum, whose
members form a dense structure having the geometry of an orthogonal net of
characteristics. That is, as the mesh spacing becomes infinitesimally fine, the
volume (and hence the weight) of the material reaches a minimum.
Allowable stresses in all the truss members are in the range σ0 ≤ σ ≤ σ0 .
Given a modulus E, we have a range for strains −k ≤ ε ≤ k, where k = σ0 /E.
As pointed out by Rozvany et al. (1995), equal permissible stresses in tension
and compression are necessary for the Michell (1904) criteria to hold. With
length li and cross-sectional area Ai of bar i, the design variables are the yield
forces Yi = σ0 Ai , and the design objective is the minimization of the cost

= Yi li , (4.51)
i
which is proportional to the total volume of the bars V = i Ai li .

As an example let us now consider a problem of optimal layout of a truss
set up on a rigid circular foundation F , which can support a force P acting
at a point A, Figure 4.17. The solution is provided by a so-called field of type
T (Save and Prager, 1985) for which the principal strains have equal absolute
values k, but carry opposite signs, and the principal lines are logarithmic
spirals. Indeed, plots (a)–(d) of this figure show a sequence of four ever-finer
trusses providing supports under the same global conditions (the foundation
F and the force P). These trusses are based on, respectively, 2n + 1 (n =
2, 3, 4, and 5) boundary nodes. With n growing, their geometries tend to
an optimal truss-like continuum whose principal strain/stress directions are
mutually orthogonal characteristics of a quasi-linear hyperbolic system—this
is discussed below. This convergence of trusses (of volume V ∗ = V(n), n
finite) to the optimal truss-like continuum (of volume V = V(n)|n→∞ ) can be
quantified by the efficiency
E f f = V/V ∗ . (4.52)
We indicate Eff in all four cases of Figure 4.17.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
P P
A A
F F
(a) Eff = 0.939 (b) Eff = 0.965
P P
A A
F F
(c) Eff = 0.981 (d) Eff = 0.991
FIGURE 4.17
Successive approximations to the Michell truss, all governed by (3.6) for a homogeneous material,
according to meshes based on, respectively, 2n + 1(n = 2, 3, 4 and 5) boundary points on the rigid
foundation F.
Our presentation of the governing equations follows Hegemier and Prager

(1969). First, we let u and v be the displacement components with respect to
the x, y coordinates in the plane of the truss. Then

∂u ∂v ∂v ∂u ∂v ∂u
εx = εy = γ = + /2 W= − /2 (4.53)
∂x ∂y ∂x ∂y ∂x ∂y
are the strain components and the rotation. If by θ we denote the angle be-
tween the negative y-direction and arbitrarily assigned positive direction
along the line with the unit extension k, then
εx = −k cos 2θ ε y = k cos 2θ γ = −k sin 2θ. (4.54)

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
It now follows from the above that

∂u ∂u
= −k cos 2θ = −k(2w + sin 2θ)
∂x ∂y
∂v ∂v (4.55)
= k(2w − sin 2θ ) = k cos 2θ,
∂x ∂y
where w = W/k.
Eliminating u and v from (4.55) by cross-differentiation, we obtain
∂w ∂θ ∂θ
k + k cos 2θ + k sin 2θ = 0,
∂x ∂x ∂y
(4.56)
∂w ∂θ ∂θ
k + k sin 2θ − k cos 2θ = 0.
∂y ∂x ∂y
and, setting θ = π/2, we find equations
d d
(w − θ ) = 0 (w + θ ) = 0, (4.57)
ds1 ds2
which hold along two characteristics s1 and s2 , at angles specified, respectively,
by
θ = α + π/4 θ = α + 3π/4. (4.58)
α is defined as the angle formed by the positive direction along the foundation
F with the positive x-direction. On this boundary, w = −1.
w = −1. (4.59)
4.4.1.2 Example of an Optimal Layout

In practice, as illustrated by the example of Figure 4.17, the stochastic quasi-
linear hyperbolic system governing the field is solved by finite differences.
At the typical point Q of the foundation F , let the positive direction along
this boundary form the angle α with the positive x-direction. As the rigid
foundation is inextensible, its tangent and normal at Q bisect the right angles
formed by the principal axes of strain at Q. As u vanishes along F , cos α ∂∂ux +
sin α ∂∂uy = 0 at Q. From this, along with (4.55) and (4.58) there follows
w = ∓1 (4.60)
along F ; in Figures 4.1 and 4.3 the upper sign in the above is appropriate.
Thus, we have an inverse Cauchy problem: “find the net of characteristics
supporting the given load P at point A, that emanates from the foundation F
with conditions (4.58) and (4.59) specified on it.” Figures 4.17(a)–(d) display
four deterministic solutions, all governed by (4.57) according to meshes based
on, respectively, 2n +1 (n = 2, 3, 4 and 5) boundary points. In the limit n → ∞,
we arrive at a truss-like continuum with Eff = 1.

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
In case one wants to make (i.e., manufacture) the Michell truss from a metal
plate—such as a polycrystal—one encounters the effect of random material
microstructure. In effect, the denser the truss—or the finer the mesh spacing—
the more significant is the effect of microstructural fluctuations on the plastic
limit. This leads to a question we address in Chapter 8: Can one truly reach
the limit of truss-like continuum as the mesh is refined ad infinitum?
4.4.2 Study via Elliptic System

The foregoing approach applied when one assumed a fixed load and searches
for a structure of optimal (i.e., minimal) weight. Another approach, much
more in line with the conventional methodology of shape optimization of
engineering structures (Bendsøe and Kikuchi, 1988), is to seek the shape of
a structure with minimum compliance (maximum stiffness), that possesses a
prescribed weight. The minimum compliance problem for a planar (respec-
tively, spatial) body B of volume V in R2 (R3 ) subjected to body forces f and
tractions t takes the form:
min L(u)
(4.61)
Cijkl ∈ Uad
subject to a C (u, v) = L(v), all v ∈ U, design constraints, where

a C (u, v) = ε(u)C(x)ε(v)dx (4.62)
B
is the energy bilinear form, and

L(u) = fvdx+ tvdx (4.63)
B ∂ Bt
is the load linear form. That is, we seek the optimal choice of stiffness tensor
C in some given set of admissible tensors Uad . C are generally fields over R2 ,
so that Uad ∈ (L ∞ (V)) 6 , corresponding to the six independent elements of in-
plane stiffness tensor. By the “design constraints” we understand constraints
on stresses, strains, displacements, etc., while sizing constraints, volume con-
straints, etc., are accounted for in the choice of Uad . Finally, U is the space of
kinematically admissible displacement fields.
In the case of optimal shape design, elements C(≡ Cijkl ) of Uad take on the
form
Cijkl (x) = χ (x)C ijkl , (4.64)
where C ijkl is the constant stiffness tensor for the material employed for the
construction of the mechanical element, and χ (x) is the indicator function.
The discretized formulation of the topology optimization problem can then

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
be stated as follows:
min f (ρ)
s.t. V = ρ j νi ≤ V ∗ (4.65)
η ≤ ρi ≤ 1, i = 1, . . . , N,
where f represents the objective function, ρi and νi are element densities
and volumes, respectively, V ∗ is the target volume, N is the total number
of elements and η is the small number that prevents stiffness matrix from
being ill-conditioned. The common objective function is the weighted sum of
compliances under all load cases. Note that the problem in (4.65) is a relaxation
formulation of the topology problem, where the density should only take
value 0 or 1. To force the design to be close to a 0/1 solution, a penalty is
introduced to reduce the efficiency of intermediate density elements, namely,
by a following power law formulation
Ki∗ (ρ) = ρi Ki
p
(4.66)
where Ki∗ and Ki represent the penalized and the real stiffness matrix of the ith
element, respectively, and p is the penalization factor that is bigger than 1. For
a survey of historical development and a summary of theory and techniques
of topology optimization, see Rozvany et al. (1995), while for the penalty
formulation see Allaire and Kohn (1993).
In Chapter 8 we return to Michell trusses in the setting of materials pos-
sessing microstructural randomness.
Problems
1. The extension of the condition (4.1) in 3D is the so-called Maxwell’s
rule:
|E| = 6|V| − 6.
Justify this equation.

2. Verify that the network of Figure 4.1 is generically rigid.
3. Using the graph representation, formulate the virtual work prin-
ciple in granular media relating the work of forces and moments
of grain-grain interactions with the work of forces and moments
acting at the boundary of the granular medium.
4. In the case of iron, the interatomic potential is

φ(r ) = D e −2a (r −r0 ) − 2e −a (r −r0 )
where the binding energy D = 0.4174eV, a = 1.3885( Å) −1 and

ϒ0 = 2.845 Å (Milstein, 1971). Compute E, c and ωmax . To see the

P1: BINAYA DASH
June 1, 2007 12:16 C4174 C4174˙C004
applicability of linear elastodynamics, compare the latter quantity

with the highest frequencies of about 106 /sec excited in transducers.
Translate these results into wavelengths.
5. Modify the 1D chain of particles of Section 3.1.1 by adding a cubic
term in the equation (3.1)1 :
1 1
U= K (ui+1 − ui ) 2 + C(ui+1 − ui ) 3 .
i
2 3
Derive the resulting dynamical equation generalizing (3.2). Then,

using a Taylor expansion with terms up to the fourth derivative,
obtain the continuum equation

s2
E u, xx + u, xxxx + a u, x u, xx = ρ ü.
12
Identify E, α and ρ.
6. Consider two particles (of mass mi , i = 1, 2 located at ri , respec-
tively, separated by a distance rij = rji , and interacting via a central-
force potential φ (rij ). In the completely conservative method of
particle modeling (PM) one replaces their Newtonian equations of
motion (no summation convention and no tensor index notation)
d 2 ri ∂φ ri − r j
mi =− , i, j = 1, 2
dt 2 ∂ri j ri j
by two first order difference equations

ri,n+1 − ri,n Vi,n+1 − Vi,n
=
t 2
vi,n+1 − vi,n φ(ri j,n+1 ) − φ(ri j,n ) ri,n+1 + ri,n − r j,n+1 − r j,n
mi =− ,
t i, j =r1, 2 − ri j,n
i j,n+1 ri j,n+1 + ri j,n
Where the first subscript specifies the particle number while the
second subscript indicates the time t or t+t. These equations can be
solved by the Newton’s method. Show that the numerical solution
conserves the total energy, the linear and the angular momentum.
7. Consider the system of three particles interacting via central forces.
Write their Netwonian equations of motion, and then the corre-
sponding first order difference equations.
8. Show that the numerical method discussed in Problem 5 is invariant
with respect to rotations of the frame of coordinates.
9. Verify (4.56), and derive a more general version of the equations
governing w and θ for the case of k being smooth in x and y.

Lattice Models: Rigidity, Randomness, Dynamics, and Optimality

Uploaded by

Copyright:

Available Formats

Lattice Models: Rigidity, Randomness, Dynamics, and Optimality

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lattice Models: Rigidity, Randomness, Dynamics, and Optimality

Uploaded by

Copyright:

Available Formats

P1: BINAYA DASH

June 1, 2007 12:16 C4174 C4174˙C004

There is a crack in everything: that’s how the light gets in.

4.1 Rigidity of Networks

© 2008 by Taylor & Francis Group, LLC

134 Microstructural Randomness and Scaling in Mechanics of Materials

Structural mechanics Minimally constrained truss Mechanism

|E| = 2|V| − 3, (4.1)

where || denotes the number of elements in a given set. As an example, let

© 2008 by Taylor & Francis Group, LLC