Download as pdf or txt
Download as pdf or txt
You are on page 1of 91

ADVANCES IN APPLIED MECHANICS.

VOLUME 36

Stochastic Damage Evolution and Failure in


Fiber-Reinforced Composites

W. A . CURTIN*
Engineering Science and Mechunics
Materials Science und Engineering
Virginia folyrechnic Institute cind Stcite University
.
Blackshurg Virginia

I . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
I1. Preliminary Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
A . Fibers, Matrices. and Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . 168
B . Critical Strength and Critical Length Sc . . . . . . . . . . . . . . . . . . . 172
I11. Single-Fiber Composite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
A . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
B . Solutions to the s.f.c. Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
C. Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
D . Comparison to Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
IV . Multifiber Composites: Global Load Sharing . . . . . . . . . . . . . . . . . . . . 185
A . Global Load Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
B. Fiber Pullout and Work of Fracture . . . . . . . . . . . . . . . . . . . . . . . 187
C . Stress-Strain Behavior: Exact and Approximate Results . . . . . . . . . . . . 189
D . In situ Fiber Strength and Fracture Mirrors . . . . . . . . . . . . . . . . . . . 197
E . Localization and Numerical Simulations . . . . . . . . . . . . . . . . . . . . 198
F . Initial Fiber Damage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
G . Comparison to Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
H . Other Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
I . S u m mary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
V . Multifiber Composites: Local Load Sharing . . . . . . . . . . . . . . . . . . . . . 212
A . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
B . Local Load Sharing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
C . Statistical Models of Composite Failure
under Local Load Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
D. Local Load Sharing Model of Zhou and Curtin (1995) . . . . . . . . . . . . . 220
E . Analytic Models and Weak-Link Scaling . . . . . . . . . . . . . . . . . . . . 232
F . Comparison to Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
G . Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

. .
*Present address: Division of Engineering Brown University Providence. RI 029 I2.

163
ADVANCES IN APPLIED MECHANICS VOL. 36 .
Coplright 0 1999 by Academic Press. All n p h b of rrpruduclion in any form reserved.
ISBN 0- I?-(M?O3O-X ISSN 006S-21hS/99 $30.00
164 W A. Curtin

VI. Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244


Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ,248

I. Introduction

A major theme in the field of materials for structural applications is the de-
velopment of the relationship between the material microstructure and its perfor-
mance under various thermomechanical loading conditions. Such a relationship
allows for the tailoring of the microstructure to obtain a desired behavior or for
optimization of a material for multifunctional use. In the engineering design of
structural components, the “performance” includes, among others, material stiff-
ness, inelastic deformation, ultimate failure, reliability, and lifetime. Heterogene-
ity, or nonuniformity, can, however, severely affect these material properties and
complicates the development of microstructure/performance relationships. Het-
erogeneity is used here in the broadest sense, and includes spatial variations in
the elastic, thermal, inelastic, and toughness properties of the constituent materi-
als. Heterogeneity is particularly important in understanding material failure and
reliability because failure is controlled by the development of a “critical” amount
of damage somewhere in the material, damage that becomes unstable to growth at
the failure stress. The damage may be a single crack (preexisting or induced) or a
more diffuse damage zone. In a structural component, there are several sources of
heterogeneity that influence the formation and propagation of damage. First, there
can be a “distributed’ or spatially varying stress field, imposed externally and/or
caused by heterogeneous elastic and thermal properties in the material. Second,
in real materials with flaws and heterogeneities, there are spatially distributed
strength and/or toughness fields. The stress and toughness fields are coupled after
damage occurs: Damage at areas of low strength or low toughness redistributes
stress over length scales comparable to the damaged region and so drives further
damage in the vicinity of the existing damage. Ultimately, the damage becomes
unstable, requiring no further increase in the externally applied field to grow in-
definitely across the entire material.
While conceptually clear, the formal coupling of evolving stress fields in the
presence of damage caused by heterogeneous strength or toughness is a very dif-
ficult problem. One major difficulty stems from the fact that the final failure re-
quires only one critical damage region somewhere in the entire material so that
any type of averaging over the heterogeneity or stress fields is precluded, since
the averaging process can eliminate precisely those fluctuations that drive the
failure. Since the formation of the “critical” crack or damage zone can occur
Fiber-Reinforced Composites 165

anywhere in the material, failure is controlled by such “weak links.” Weak-link


failure is accompanied by a volume scaling of the strength (larger systems have
lower strengths). Hence, models must determine the strength probability distribu-
tions at volumes that are large enough to be in the “weak-link’ scaling regime if
volume effects are to be described properly.
In heterogeneous materials with intrinsically brittle constituents, or materials
where the brittle constituents are the main load-bearing phase, the problem of ac-
curately describing the damage evolution and failure is most acute. Damage is in
the form of cracks with sharp crack tips, and a model cannot eliminate the crack
tips a priori without eliminating the possible source of initiation of the macro-
scopic failure. There are thus at least three (widely differing) scales of importance:
the atomic crack tip scale, the microstructural scale, and the macroscopic scale.
The atomic scale might be replaceable by a larger “near-tip’’ zone but such a scale
is still expected to be much smaller than the microstructural scale. In a homoge-
neous material or weakly heterogeneous material, failure is entirely controlled by
the atomic crack tip and the larger scales are not important. However, failure in a
strongly heterogeneous material can require proper treatment of phenomena over
all the length scales, which span orders of magnitude in physical size.
Traditional analyses of material failure typically involve either (i) direct con-
tinuum finite-element models (FEMs) of stress fields, damage growth, and crack
growth at the microstructural scale; or (ii) continuum damage mechanics (CDM)
wherein a stable damage state is assumed and associated with, or measured by, a
stiffness reduction. In FEM modeling, it is computationally difficult to study large
“representative” regions of a heterogeneous material, large numbers of cracks, or
substantial crack growth. In CDM, there is usually no micromechanics of ac-
tual damage mechanisms and the material is assumed homogeneous on some
largehepresentative length scale; this immediately removes the underlying true
heterogeneous toughness distribution and local stress fluctuations that may be
critical to failure. The material becomes a deterministic nonlinear continuum
medium. The point of failure is then also deterministic and not specifically con-
nected to the microscale damage. Stochastic CDM models can accommodate sta-
tistical variations in the properties of the various “representative elements” but
usually such variations are not derived from the study of known damage modes at
the microstructural scale.
In one class of materials, fiber-reinforced composites (an important class of en-
gineered structural materials), some of the atomickrack tip issues can be elim-
inated in favor of the cumulative damage at a larger scale. The brittle rein-
forcing fibers have intrinsic cracklike flaws and known toughness so that the
strengths of these flaws can be characterized. More importantly, once a flaw be-
comes unstable and fails a single fiber, the crack formed is generally arrested
166 W A. Curtin

at the fibedmatrix interface by matrix plasticity/yielding, interface debonding,


or fibedmatrix sliding at a previously debonded interface. All of these mecha-
nisms prevent the sharp crack from extending up to the surface of the neigh-
boring fibers, which would cause them to fail spontaneously. The neighboring
fibers thus experience a higher load, or stress concentration factor, but not a
singular stress field or stress intensity factor. The fibers can thus be considered
as stochastic load-bearing entities that interact through load transfer from bro-
ken to unbroken fibers. The details of the stress concentration factors or load
sharing among fibers, and its dependence on constituent mechanical properties
and interface properties, is a formidable problem but at a length scale of the
microstructure (fiber diameter or spacing) that is far more tractable than the
crack tip problem. The fiber composite does contain heterogeneity in the fiber
strengths and in the local stress fields at the microstructural scale once some
fiber damage has occurred. So, the connection between the microstructural scale
and the macroscopic scale is still nontrivial and involves mechanics, stochastics,
and volume scaling. Coupling mechanics and stochastics exactly is impossible,
which has driven the development of approximate models wherein the mechan-
ics of the load transfer is approximately described, although in a manner con-
sistent with the physical requirements such as macroscopic mechanical equilib-
rium.
The purpose of this article is to review the accomplishments over the last
10 years in the area of modeling of the mechanical properties of fiber-reinforced
composites, with an emphasis on accurately predicting ultimate tensile strength,
stress-strain behavior, and reliability. A key aspect of the problem is the effect
of heterogeneous or stochastic (nondeterministic) damage evolution that eventu-
ally leads to overall material failure. In particular, we discuss recent efforts to
develop coupled mechanics/stochastic models for fiber composites in a manner
that yields predictive, accurate results for the deformation and strength of real
fiber-reinforced composite systems. These models, although using simple approx-
imations to the complex three-dimensional stress fields, provide the microstruc-
ture/performance relationships in these materials, demonstrating the basic depen-
dencies of macroscopic material performance on underlying constituent proper-
ties. The general problem addressed here can be posed as follows:
Given a statistical distribution of fiber strengths, a matrix strength, prescribed
interfacial fibedmatrix interface behavior, and material thermoelastic constants,
then
Predict the macroscopic stress-strain behavior, tensile strength and failure
strain, work of fracture, reliability (probability of failure versus stress), and any
dependence of these properties on the physical composite volume.
Fiber-Reinforced Composites 167

The success of the resulting predictions is assessed by comparing them to as many


detailed experimental results as possible. This review will attempt to unify much
of the recent work in this field into a common framework and common notation,
with derivations of key concepts and models, presentation of major results, and
extensive comparisons of the theories to experimental data. This review can thus
serve both as a compendium of results for workers in the field and as a primer for
scientists and engineers interested in the subject.
The specific work described below is confined to continuous, unidirectionally
reinforced composites. However, extensions to cross-ply and woven fiber geome-
tries will be discussed. Particulate-reinforced composites are not considered al-
though aligned short-fiber composites are a natural extension of the present work.
We will also consider only fiberlmatrix interfaces having a constant shear yield
strength or interfacial sliding resistance T at the interface; purely elastic inter-
faces, Coulomb friction, and rough asperityxontact interfaces will not be dis-
cussed although work in these areas shows them to be relevant in some materials
(Parthasarathy and Kerans, 1997, and references therein). Lastly, the detailed mi-
cromechanics of the stress concentration factors or load transfer will not be dis-
cussed extensively. Rather, we will focus on two extreme cases that circumscribe
the range of behaviors physically expected. These limitations are not fundamental
ones, however. The major concepts, methods, and general results are anticipated
to survive the relaxation of these restrictions, although not without significant ad-
ditional effort to obtain accurate results.
The remainder of this article is organized as follows. In Section 11, we discuss
the basic model of the unidirectional composite, including fiber strength statistics,
matrix behavior, and interfacial mechanics. In Section 111, we study the single-
fiber composite (s.f.c.), which is the fundamental problem in understanding the
interplay of fiber strength statistics and interface mechanics. The s.f.c. problem
highlights the key aspects relevant for subsequent models for ceramic, metal, and
polymer composites (CMCs, MMCs, and PMCs, respectively). In Section IV, we
introduce the concept of global load sharing and demonstrate how composite fail-
ure properties follow from the s.f.c. problem. Applications to CMCs are made to
demonstrate the capabilities of this model. In Section V, we investigate local load
sharing, which brings many of the important and subtle features of brittle, weak-
link-driven failure into the problem of composite failure. Models for combin-
ing mechanics, stochastics, and size scaling are presented and applied to various
MMCs and PMCs. Finally, in Section VI, we discuss the limitations of the existing
models, and aspects of the existing theories that require improvement or refine-
ment, possible extensions of these models to a host of other problems, and the cou-
pling of these models to macroscale models for predicting structural performance.
168 W A. Curfin

11. Preliminary Issues

A. FIBERS,MATRICES,
AND INTERFACES

The model composite studied throughout this work consists of a volume frac-
tion f of continuous cylindrical fibers of radius r embedded in a matrix material
in a unidirectional (aligned) arrangement. The axial (along the direction of fiber
alignment) Young’s moduli of the fibers, matrix, and composite are E f , E , , and
+
E,, respectively, with the rule of mixtures estimate E, = f E f (1 - f)Em being
highly accurate. Under an applied axial stress cr, and with no interface debonding,
fiber breakage, or inelastic behavior, mechanical equilibrium and strain compati-
bility require that

=f0.f + (1 - f)%, (14


where

are the average axial stresses on the fibers and matrix, respectively. In formulating
simple approximations below, we neglect many complicating features of the true
stress fields, such as the radial variations in axial stresses (Weitsman and Zhu,
1993),but retain the dominant features so that the analysis is tractable and yet the
predictions are accurate.
Fiber composites are designed to take advantage of the strong fibers, which are
thus generally much stronger than the matrix material. We consider three possible
types of matrix behavior: (i) elastic but with Em << E f , as for PMCs; (ii) elas-
ticbrittle where the matrix failure strain is much lower than that of the fibers,
as for CMCs; and (iii) elastic/perfectly plastic where the tensile yield stress of
the matrix, av,is lower than the fiber bundle strength, as for MMCs. In all of
these cases, the matrix then plays only a secondary role in determining the ulti-
mate composite strength. In case (i), the matrix remains elastic but carries negli-
gible stress. In case (ii), after cracking the matrix carries essentially zero stress at
higher applied stresses. In case (iii), after yielding the matrix carries a fixed stress
uY.Therefore, after cracking/yielding, all of the additional applied load is carried
by the fibers. Mechanical equilibrium can then be written as

cr = fcrf + (1 - f>q, (above yield), (2)


where cry M 0 for CMCs and PMCs. The ultimate tensile strength of the compos-
ite, out,,is then controlled by the maximum load-carrying capability of the fibers,
Fiber-Reinforced Composites 169

a;, so that

The major task below is to determine the deformation behavior at versus (T and
the maximum a;.
Our subsequent discussion for obtaining a; and hence outsis not strictly lim-
ited to the three matrix cases above. For work-hardening materials or a fully elas-
tic matrix, the matrix contribution to the stress can be included approximately
by retaining the constitutive dependence of a,,, versus strain with the composite
strain determined self-consistently. Also, in a CMC the matrix cracking itself is a
stochastic process of multiple matrix crack evolution with a range of “strengths”
but typically all cracking is completed well prior to the onset of fiber breakage so
that ( T ~ 0 remains accurate (Ahn and Curtin, 1997; He et nl., 1994a).
The reinforcing fibers are generally brittle materials (ceramic, glass, graphite)
and so are linearly elastic up to the point of failure. The point of failure in any in-
dividual length of fiber is determined by the largest flaw or crack in that particular
fiber. Different pieces of nominally identical fiber have different largest flaws and
hence different strengths. The “strength” of a fiber is thus a stochastic variable
with some probability distribution. The stochastic nature of monolithic ceramics
and glasses is widely recognized and the situation is identical for most reinforcing
fibers. To formalize the discussion, consider a fiber of length L to have a distri-
bution of flaw sizes, with each flaw having a corresponding strength. The total
number of flaws in this fiber that are weaker than a stress (T is @(o,L ) where

Equation (4) assumes a power law distribution of flaw strengths and is a “Weibull
model” for the flaw distribution with a “Weibull modulus” rn (Weibull, 1952);
typical values of m for structural fibers are in the range of 2 5 ni 5 20. The num-
ber of flaws is proportional to L since the flaws are assumed to occur randomly
along the length. 00 and Lo are reference parameters for the fiber strength, with
typically one flaw weaker than strength a()in a length Lo of fiber. In testing of
fibers with a flaw distribution given by eq. (4), failure of a fiber occurs at a stress
(T if there is a flaw of strength (T and no flaws weaker than a. The cumulative

probability of fiber failure in L at stress (T is therefore given by


170 W A. Curtin

which is the well-known Weibull cumulative failure probability distribution. From


eq. (9,it is evident that not only is the fiber strength at fixed length a statistically
distributed quantity but also that the “typical” fiber strength depends on the length
L tested. For a Weibull distribution, the “typical” strength is the characteristic
strength such that Pf(a,L ) = 1 - e-’ = 0.632, or @(u, L ) = 1 in eq. (4).So,
00 is the characteristic strength at length Lo and

is the characteristic strength at any other length L. The strengtMength relation-


ship embodied in eqs. (4)-(6) is very important to understanding composite be-
havior.
Lastly, the underlying assumption that the fibers are governed by eqs. (4) and
(5) is usually assessed by measuring the fiber strength distribution at a length Lo
and fitting the measured distribution to eq. (5j to obtain the parameters no and rn.
The accuracy of eqs. (4)-(6) in describing fiber strengths at other lengths L # Lo
depends critically on the number of tests performed at length Lo. For typical num-
bers of 25-50 tests, the extrapolation of strength using eq. (6) to much smaller or
larger lengths L < Lo/lO or L > lOL0 begins to be inaccurate. So, predictions on
composite properties based on such extrapolations must be regarded cautiously.
The interface between the fibers and matrix occupies a vanishing fraction of the
total composite volume but plays a critical role in determining many composite
properties related to damage and strength. In polymer and metal matrices, the
interface becomes important when fibers break. In ceramic matrices, the interface
is critical first when the matrix cracks and then again when the fibers break. So,
consider a fiber in the matrix that is broken due to failure at some preexisting flaw.
At the fiber cracWmatrix intersection, the stress state is very complex (He et al.,
1994b),but can drive crack extension into the matrix, yielding of the matrix, crack
deflection along the fibedmatrix interface, or yielding along the interface. The
latter two modes are driven by the high shear stresses acting along the interface.
In effective composites, either (i) the interface is engineered to be weak enough
(low toughness) to promote cracking and debonding or (ii) the matrix undergoes
shear yielding along the interface.
After interfacial debonding, there can remain a residual shear sliding resis-
tance across the fibedmatrix interface due to friction. The precise nature of such
“friction” is the subject of considerable study, and the roles of thermal clamping
stresses, interface roughness, Coulomb friction, and Poisson effects are all impor-
tant to some degree. For tractability, however, a common assumption is that there
exists some constant interfacial sliding stress r across the debonded interface.
Fiber-Reinforced Composites 171

FIG. 1 . Schematic of stresses around a broken tiber. and axial fiber stress of vs position :around
a break.

Many, but not all, experimental results on the sliding interface can be interpreted
fairly well with the constant 7 assumption. For the shear-yielding interface in
the absence of work-hardening and multiaxial stress states, the constant t is also
an adequate approximation. So, the constant t model allows for a commonality
among ceramic, metal, and polymer matrix composites. Many of the important
results below will seem to depend explicitly on the constant 7 assumption. But, in
fact, the general concepts developed below can be applied to many more complex
interface models; doing so is simply unwieldy and obscures the main physical
features that are captured clearly within the context of the constant 7 model.
Returning to the broken fiber, we now have a “sliding” or “debond’ or “slip”
zone along the fibedmatrix interface (Figure 1). Neglecting the radial variations
of the axial stress along the fiber, equilibrium between the fiber axial stress and
the interface shear stress is given by, with the break at z = 0 and fiber radius r ,

or, upon integrating,


2tz
= 7
in the slip zone. The axial fiber stress increases linearly in the slip zone. Neglect-
ing the elastic behavior at the end of the slip zone, the slip length 6 is the distance
at which the fiber axial stress attains the far-field value T (which will, due to dam-
age and matrix yielding, differ from the value a E f / E , ) . Setting of ( z = 6) = T
leads to
rT
6 = -.
2s
172 W A. Curtin

Numerous works have shown that the above “shear lag” type of approximation is
quite accurate for the slip length and average axial fiber stresses, particularly in
systems with E f E,,, and “low” t values (He et al., 1994a).

B. CRITICAL STRENGTH a,.AND CRITICAL LENGTH6,.

The fiber strength depends on the gauge length tested. In a single-fiber tension
test, the length can be selected arbitrarily. In a composite, the fiber strengths con-
trol the composite tensile strength. So, is there a particular fiber strength, as well
as an associated length, that is related to the composite strength and deformation?
The answer is yes. To determine the critical strength ac and critical length 6, re-
quires consideration of both the fiber strength statistics and the fiber slip in the
composite. Imagine applying a stress T to the fibers, causing them to break into
fragments of average length (x) = Lo(ao/T)”’.If the spacing (x) is larger than
twice the slip length B = r T / 2 r around each break, then there remain some re-
gions of fiber which experience the far-field load T ; further load could be applied
to the fibers and they could break into smaller pieces. So, the typical maximum
stress that the fibers in the composite can be subjected to is a stress a, for which
the average spacing is exactly twice the slip length at this stress, 6,. = ra,/r.
Since the average fragment length is also related to the applied stress as indi-
cated above, we have 6, = L o ( o ~ ) / a ~as) ~well.
~ ’ Solving these two relationships
simultaneously, we obtain (Curtin, 1991b; Henstenburg and Phoenix, 1989)

Physically, there is typically one flaw of strength a,. in a length 6, of fiber and
6, is twice the fiber slip length at an applied stress of cr,. Equations (9) are the
generalization of the Rosen (1964) and Kelly (1965) critical length and stress
to the case of stochastic fibers. These quantities control several major composite
properties, as we will see below. In particular, we shall find that, within the “global
load sharing” approximation (Section IV),

Tensile strength c( oC,


Fiber pullout a 6,,
Work of fracture 0: a,6,,.
All of the dependencies on the constituent properties no, r, Lo, and t enter only
through a,.and 6,., with only slight additional dependence on the Weibull modulus
m. Even in the more general “local load sharing” approximations (Section V), the
Fiber-Reinforced Composites 173

dominant dependencies remain as indicated above. It is of some interest to note


that the composite tensile strength depends on the interfacial sliding resistance
and fiber radius in a strongly nonlinear manner, dependencies that probably would
never be directly anticipated otherwise.

111. Single-FiberComposite

A. INTRODUCTION

Consider a single fiber embedded in a matrix material (rigid or plastic) and


subjected to uniaxial tension. Under increasing load. the fiber will break at its
weaker flaws and form slip zones around those breaks. Ultimately, at higher loads,
there will be enough fiber breaks so that the entire length of fiber is slipping within
the matrix. The fiber can then be loaded no further and the damage ceases. This
single-fiber composite (s.f.c.) test, though conceptually simple and seemingly far
from the reality of the multifiber composites of practical interest, holds the key
to understanding some fundamental aspects of all multifiber composites (CMCs,
MMCs, and PMCs). In addition, the s.f.c. test can be used to derive information
about the in situ fiber critical length 8, , critical stress ac,and the Weibull modulus
m appropriate at these gauge lengths. These in turn then imply a value for the
interfacial sliding resistance, T = r o c / & , for the particular fibedmatrix interface
studied. The s.f.c. test has a long history of application in the PMC field, primarily
to study the effects of fiber surface treatments on adhesion (a few references are
Fraser et al., 1983; Netravali et al., 1989; Wagner and Eitan, 1990; Rao and Drzal,
1991; Gulino and Phoenix, 1991). More recently, the s.f.c. test has been used to
study the interfacial debonding in Ti-MMCs as well (Majumdar, 1996; Majumdar
et al., 1996a, b contain a number of examples and references).
The questions we wish to answer about the s.f.c. test are as follows. Given the
fiber strength distribution (00, m ) at gauge length Lo and an interfacial T, what
are (i) the number of fiber breaks versus applied stress, (ii) the spatial distribution
of the breaks or, equivalently, the fiber fragment lengths created by the breaks
versus stress, and (iii) the average fragment length distribution at the end of the
test? If these questions can be answered quantitatively, then experimental data
on the fragment distributions and break evolution versus stress can be inverted to
derive the values of a,, &., m , and t. Also, the damage evolution will be used
in the multifiber composite to predict tensile strength, fiber pullout lengths, and
work of fracture, as discussed in Section IV.
The s.f.c. problem is difficult to solve because fiber breaks can only occur where
the stress on the fiber is at the far-field applied value. Flaws in regions of the fiber
174 W A. Curtin

within the slip zone around a previous break experience a lower stress and hence
are strictly excluded from causing another fiber break. Therefore, not all flaws
in the fiber, and not even all of the weakest flaws in the fiber, can actually cause
breaks.
To demonstrate the strict existence of the exclusion zone, consider a section of
fiber that has survived intact up to some stress o ,with a flaw at position z = 0
and of strength (T just about to fail. When this flaw fails, a slip zone is formed in
the region -6 < z c 6, within which the stress is, according to eq. (7b), 2 t z f r
(see Figure 1). The stress everywhere in the slip zone is now lower than it was
just before the break occurred, and will remain lower for all further increases in
the applied stress o. Since the region -6 < z < 6 survived the original stress
(T everywhere except at the one break point, this region will also survive with no

further breaks for all future applied stresses. Regions within a slip length of an
existing break are thus excluded from subsequent failure. The existence of the
excluded regions accounts for the cessation of fiber damage when all regions of
the fiber are within a slip length of some break. The existence of an exclusion zone
does not depend on the constant t assumption (Curtin, 1991a; Hui et al., 1996).
A range of behaviors for the interface shear can be shown to create a rigorous
exclusion zone. A notable exception, however, is the perfectly elastic interface
where all stresses are always proportional to the applied stress.
Figure 2 shows a schematic example of the s.f.c. test for a small section of
fiber with the 12 weakest flaws along the fiber shown explicitly (Curtin, 1991a).
Increasing stress causes flaws to fail, slip/exclusion zones to form or increase in
length, and exclusion zones to increasingly overlap until the entire fiber is sub-
sumed within the exclusion zones and the test “saturates.” In the case shown, the
flaws having strengths os,o7,08, (TI I , ( ~ 1 2 and
, all other stronger flaws not shown
never fail because they become part of an exclusion zone before the applied stress
reaches the value needed to fail those flaws.

B. SOLUTIONS TO THE S.F.C. TEST

1. nz -+ 00 Limit
In this limit, all flaws have the same strength (TO and when a break occurs the
slip length is always 60 = roof2r. So, no break can occur within a distance 60 of
any other break and breaks continue to appear until there are no break spacings
larger than 260. The s.f.c. problem in this case is identical to the “car-parking”
problem where cars of length 60 are randomly “parked” along a road until there
are no spaces large enough to accommodate another car. The average spacing of
Fiber-Reinforced Composites I75

4. I
q-

0.
a,,

-
f t tt t t t t t t t t
04 4 all % 4
-FIBER
09
AXIS
a,
- %%a, olo

Flci. 2 . Schematic of the evolution of fiber damage and fiber slip (exclusion zones) with iiicreasing
applied stress (bottoni to top) in an s.f’.c. test. The 12 weakest defects are shown explicitly. and are
at strengths rr,, = C J ~ , ! ’ / ’ . At stress “10 the s.f.c. test “saturates” because all regions of the fiber
are within a dip length of an existing break. From Curtin (1991a). Reproduced with permission of‘
Chapman & Hall.
176 W A. Curtin

the breaks at the end of the test has been derived many times in the literature in
different contexts and is ( x r ) = 1.33760 (Widom, 1966). The spacing distribu-
tion, or distribution of fragment lengths x, as a function of the number of breaks
N in a length L , was found by Widom (1966). We denote this family of distri-
butions as Pw(x;q ) where q = N & / L is the fraction of space taken up by the
slip zones (or “cars”). Widom found this distribution to divide naturally into two
components, a distribution for 60 < x < 260 and 260 < x , as

60 < x < 280,

+ ’ e( vx) p [ - ( i - 2 ) + ( q ) ] ,
Pw(x;77) = m 260<x, (llb)

with the function + ( q ) defined by

Widom’s result says nothing, however, about the cases of practical interest, corre-
sponding to 2 < m i20.

2. Curtin ‘s Solution
The present author developed an approach to calculating the fragment length
distribution P ( x , a) for the s.f.c. problem (Curtin, 1991a) that is quite accurate
but is not exact, as thought by many prior to the work of Hui et al. (1995) dis-
cussed below. The gist of the solution is as follows. Consider a fragmented fiber
of some length LT at some stress a with NT breaks in this fiber and a slip length 6
at this stress. There are some fragments, formed earlier in the test, that are smaller
than 6 and the distribution of these will never change since they are too small to
break again and too small to be formed by the breaking of larger fragments. Let
this portion of the distribution be denoted P R ( x , a) (nonzero only for x 5 6) and
contain N R of the breaks occupying a length L R . The remaining portion of the
fragmented fiber then consists of N = NT - N R breaks distributed in a length
L = Lr - L R , with all of the fragment lengths being x > 6 by construction. The
key assumption in (Curtin, 1991a) is then that this latter distribution is identical
to Widom’s distribution Pw(x;q ) with q = N 6 / L (for x > S only). Note that
the stress dependence in PW is implicit in N , L , and 6.
From this starting point, now consider how the two parts PR and PW of the
overall distribution P ( x , a ) change as the stress is increased. Increasing the stress
has two effects: an increase in slip length 6 and an increase in the number of breaks
Fiber-Reirforced Coniposites 177

in the fiber. Fragments just larger than 6 become smaller than the new 6 and so
move from the x > 6 distribution PW to the x < 6 distribution P R , which can be
written as

while the remaining length of fiber containing x > 6 fragments is reduced as

The number of breaks N in the remaining portion x > 6 is decreased by the


fragments lost to the x < 6 distribution but is increased by the occurrence of new
breaks. New breaks can only occur in regions of fiber experiencing the applied
stress, which is thus only in the central x - 26 region of those fragments larger
than 26. So, we have
d-N-
da
d6
- - N P w ( ~ ;q ) -
do
+ dQ,(a,
da
L*)

where the length L* available for new breaks can be written as

and for a Weibull distribution we have from eq. (4)


dQ,(a, L * ) L* CP-’
-
- _ ni -. (16)
da L a(;”
Equations ( 12)-( 15) provide a set of differential equations for determining the
overall fragment distribution as the normalized sum of the two parts PR and PW
versus applied stress:

as a function of the underlying flaw distribution Q, and the slip length S(a).There
is no specijc requirement that be a Weibull distribution or that the i n t e ~ a c e
have a constant r as long as 6 is a proper exclusion zone length.
For the case of a Weibull Q, and constant t such that 6 = r a / 2 t , the above
equations can be simplified by normalizing all lengths by a reference lengths 6~
and all stresses by a reference stress O R , with 6~ = rCfR/2T. The appropriate
choice for these reference quantities comes from simplifying eq. (16) which, upon
178 W A. Curtin

substituting eq. (15), becomes

To make the term in brackets in eq. (1 8) equal to unity then requires the refer-
ence stress and length to be identically OR = o~.and 8~ = &, respectively (see
eqs. (9)). So, all features of the s.8~.test are properly normalized by the critical
stress and length introduced in eqs. (9).
When normalized by a(.and &, the solution to eqs. (1 2)-( 16) leads to a frag-
ment distribution as a function of x E x / 6 , at a normalized stress s = o/a,..
The normalized distributions, the number of breaks versus stress, and the average
fragment lengths then depend only on the Weibull modulus m . Specific results
will be presented below in tandem with the exact results of Hui et al. (1 995).

3. Exact Solution of Hui et al. (I 995)


Recognizing the inadequacy of Curtin's solution, as presaged by some results
on multifiber composites by Neumeister (1993) (see Section V), Hui et al. (1995)
recently formulated an exact set of differential equations for the fragment distri-
bution evolution for Weibull fibers and a constant t interface. They were then able
to obtain analytic closed-form solutions for the fragment distributions in this case.
Hui et al. introduced the same reference parameters n,. and 6,. as above, and
so worked in the normalized length x and stress s, with a normalized slip length
6/6, = s also. For Weibull fibers, the number of breaks per unit length per incre-
ment of stress (the "hazard rate" h ) is

1 d@
h=--
L do

and can be written in dimensionless units as h ( s ) = ms"'-', following from


eq. (16). Hui et al. then considered three different length regimes for the over-
all fragment distribution P ( x , s ) : x < s / 2 ( x < 6 in dimensional units),
s/2 < x < s (6 < x < 26), and s < x (26 < x ) . These are the same regimes of
length that arise in Curtin's solution since the Widom solution divides naturally
into two parts: 6 < x < 26 and 26 < x . For the short fragments x < s / 2 , no new
fragments can occur, so that

d P ( x , s)
=o,
S
x<-
ds 2'
Fiber-Reinforced Composites I79

For fragments in the range s / 2 < x < s , fragments can only form by the breaking
+
of a larger fragment of length x’ > x s / 2 > s, so that

For the largest fragments, new fragments can form by division of larger fragments
but fragments of size x disappear by breaking into smaller fragments. Hence,
dP(x-, s)
= 2h(s) d x ’ P ( x ’ ,s ) - (x - s ) h ( s ) P ( x , s). s < x. (22)
d.S

Continuity at x’ = s / 2 is required, so that eq. ( 2 0 ) is essentially the same as


eq. (12). The differential equations (20)-(22)are quite simple and, in retrospect,
it is surprising that this general form was not found earlier. However, simplicity
is also the hallmark of a seminal result. Note also that the normalized distribution
P ( x . s ) is entirely characterized by the Weibull modulus 111 contained in h ( s ) .Hui
et al. proceeded to demonstrate that eqs. (20)-(22)can be solved analytically, but
the analytic forms still involve definite integrals and are unwieldy to reproduce
here.

C. PREDICTIONS

Of primary experimental interest are the final fragment size distribution (break
spacings), the associated average fragment length ( x f ) at the end of the test, and
the number of breaks versus stress during the test. These major results are easy to
suinmarize because they depend only on m .In other words, the distributions and
averages shown below are “master curves” that are directly usable for comparison
to experimental data. Experimental data are normalized by selecting test values
for a, and 6, and then m is used to best-fit to the master curves. Adjustments of a(
and 6, along with m to obtain an overall best fit to all of the data are then obtained
by iteration. It is important to recognize that all of the measured data, including
the distribution functions, are characterized by only three scalar parameters ( m C ,
6,. and m ) .
Figure 3 shows the predicted average normalized fragment length (sf /6, ) at the
end of the test versus the Weibull modulus m . Also shown are results of a Monte
Car10 simulation study of the fragmentation test by Henstenburg and Phoenix
(1989). All three results are nearly identical. Note also that at small m the average
value significantly exceeds the m = 00 value of 1.337/2 = 0.668 (the factor of
2 arising from the normalization by 6, = 260) and that the convergence to the
n7 = 00 value is quite slow.
180 W A. Curtin

1.1 1 1
D

0 2 4 6 8 10 12 14 16
Weibull modulus m
Frc;. 3 . Predicted final normalized average fragnicnt Icngth (.Y/ )is,.i n the s.1.c. test vs Weibull
modulus i n . Dianionds: Hui ut ti/. (1995) (exact); .sqtt~ircs:Curtin (I99la)result; open triangles: sim-
ulations of Hcnstenburg and Phoenix (1989). The result for 111 = cc is shown as a solid circle at
111 = 16.

Figure 4 shows the predicted cumulative final fragment length distribution, de-
fined as 1;
dx’ P ( x ’ , ,s = co),which is the fraction of fragments smaller than x,
for various values of m. The fragment distributions for small m are substantially
broader than for very large m .The approximate solution agrees fairly well with
the exact results although there are some differences (a few percent in probabil-
ity) at larger fragment sizes and for larger Weibull moduli, and the approximate
model does not obtain the correct saturation (maximum) fragment spacing. The
approximate results were, however, within the statistical error of the Monte Carlo
simulation studies of Henstenburg and Phoenix ( I 989), and are exact, by con-
struction, for the m = 00 limit as well.
Figure 5 shows the evolution of the cumulative fragment length distribution
1; dx’ P ( x ’ , s) for m = 5 and for various stresses corresponding to the break
fractions, as predicted by the Curtin (1991a) theory. Of some interest is that the
smallest fragments must occur early in the test, when the slip lengths are small
and there are few overall breaks, and so this part of the distribution does not
change as more fragments occur at higher stresses. Hence, when reformulated
as a probability distribution, the probability of finding small fragments decreases
with increasing stress.
Fiber-Reinforced Composites 181

0.0 0.5 1 .o 1.5 2.0 2.5 3.0


Normalized Fragment Length
FIG.4. Final normalized fragment length distribution for various Weibull moduli m . Solid lines:
Hui rial. (1995) (exact); dashed lines: Cunin (1991a).

2.0

1.o01
1.

0.0
-2 -1 .o

? -2.0

7 -3.0
C

-4.0

-5.0

-6.0
-1 .o
-1.v -0.5 0.0 0.5 1.o
I n (n
(normalized fragment length)
FIG. 5 . Evolution of normalized fragment length distribution vs applied fiber stress for Weibull
modulus ni = 5 as predicted by Curtin (1991a). Reproduced with the permission of Chapman & Hall.
182 W A. Curtin

D. COMPARISON TO EXPERIMENT

There have been surprisingly few direct comparisons of these new theories for
the s.f.c. test, in light of the widespread use of the fragmentation test over the last
two decades. Here we note three works illustrating the predictive capabilities of
the theory in application to PMCs.
Van der Heuvel et al. (1997) recently studied graphite fibers (Apollo IM 43-750
PAN fibers) in an epoxy matrix (bisphenol A (DGEBA)-type) where the fibers had
varying oxygen plasma surface treatments. Their goal was to assess the interface
properties as a function of this surface treatment. They believed that, for high lev-
els of treatment, the adhesion is excellent and the deformation is controlled by
shear yielding of the matrix. To test this hypothesis, and indirectly the theory of
Curtin (199 la), Van der Heuvel et al. first performed single-fiber tension tests over
a range of gauge lengths to assess the fiber strength statistics ((TO, m ) and verify
the length scaling of eq. (6). They also determined the shear yield stress of the
matrix as 37 MPa at a moderate test rate. Single-fiber composite tests were then
performed and the final fragment lengths were measured, along with various in-
teresting observations of yield propagation and acoustic emission. The measured
and predicted fragment probability distributions (not cumulative distributions) are
shown in Figures 6a and b for the highest surface treatments (200% and 100%).
The agreement is generally very good. The theory predicts slightly longer frag-
ments than observed experimentally but the range and shape of the distributions
are very similar. Tests at a surface treatment of 50% showed poorer agreement, but
this might stem from a fiber strength distribution that was estimated to be rather
different from the strengths obtained at both higher and lower surface treatments
( m = 10.3 at 5096, m = 5.9-6.7 for 0%, lo%, loo%, and 200%). The slight
disagreement in typical fragment sizes shown in Figure 6 could also be due to the
use of a bulk matrix t value; larger t would slowly decrease the fragment lengths
but maintain the same distribution. Overall, the results of Van der Heuvel et al.
confirm the general accuracy of the s.f.c. models that employ constant interfacial
shear stress.
Fukuda and Miyazawa (1994) performed similar s.f.c. tests on a T300 graphite
fiber in Epicote epoxy, and showed results for the evolution of the fragment dis-
tribution at several applied stresses. Converting applied strain to fiber stress via
c f = E f t , their data are replotted in Figure 7a with the length normalized by
the projected final average fragment length of (x,) = 298 wm. Also shown are
the predictions of Curtin’s model where 0, and m have been adjusted to fit all
of the data simultaneously (Curtin and Takeda, 1998b). The agreement is quite
reasonable, although not excellent, which can stem from having only one test
5

v)

m ....-.... Experimental
Experimental
4
t

I
*' :A

Fiber-ReinforcedComposites
3
0

I
-
- --
1-1 d

a Q

2 2
N

1
1 1
-

I
0
1 .o

Y
1 .o

0
0.0 0.5 1.5

c
9
0.0

c
v!
0.5 1.5
Fragment length [mm] Fragment length [mml

FIG. 6. Final fragment length distribution as measured experimentally and as predicted (simulation) by the Cunin theory: (a) 200% 0-plasma treament:
(b) 100% 0-plasma treatment. From Van der Heuvel er 01. ( 1997).Reproduced with permission of Elsevier Science.
184 W A. Curtin
1

0.8
0
E 0.6
3
n
0.4
o.
0.2

0
0.5 1 1.5 2 2.5
Normalized Fragment Length
(a)

0.8
.-a ~

0.6 -

2e 0.4 -
a 0.2
04
0.5 1 1.5 2
Normalized Fragment Length
(b)
FIG. 7. (a) Normalized fragment length distribution vs applied fiber strains of 2.5%, 3.0%. and
3.510 as measured experimentally (Fukuda and Miyazawa, 1994) and predicted by the Curtin theory
(Curtin and Takeda, I998b). (b) Normalized final fragment distribution as measured by Winiolkiatisak
and Bell (1989) and predicted by the Curtin theory (Curtin and Takeda, 1998b).

specimen against which to compare. The derived values of nc = 5800 MPa and
m = 7 were obtained, along with a critical length 6, = 500 pm. For these fibers
of radius r = 3.5 pm, this suggests r = 40 MPa, which is quite close to a Von
Mises estimate r % IT,/& = 45 MPa of the shear yield stress based on the
tensile yield stress CT,, for this epoxy matrix. Fukuda and Miyazawa separately
measured the fiber strength at Lo = 1 in. and found no = 3000 MPa, m = 5.
Using eq. (4) to extrapolate the s.f.c. results up to 1 in. yields 00 = 3100 MPa,
m = 7. The characteristic strengths are quite close. The difference in Weibull
modulus may stem from variations in the single-fiber testing that broaden the
measured distribution (m = 5) relative to the true one (m = 7). Using m = 5
in the size scaling with the measured strength at 1 in. and extrapolating to 6,
would yield a much larger, and erroneous, value for the critical strength. This
Fiber-Reinforced Composites 185

highlights the importance of an accurate measure of the fiber strength at the rel-
evant gauge length 6, , as opposed to the extrapolation from a much larger length
scale. The latter can be appropriate, but must be carried out with care and accu-
racy.
Finally, Figure 7b shows the normalized final fragment distribution for AS-4
graphite fibers (Y = 3.5 p m ) in a PDA matrix, as measured (Wimolkiatisak and
Bell, 1989) and as predicted by the Curtin theory (Curtin and Takeda, 199%).
Single-fiber tension tests were performed at various gauge lengths by Wimol-
kiatisak and Bell to yield DO = 4275 MPa, in = 10.7 at Lo = 12.7 mm, and
were used in making the theoretical predictions. Agreement between theory and
experiment is quite good, with no adjustable parameters. These results will be
used for prediction of the strength of an AS-4Iepoxy PMC in Section V.
Having demonstrated in at least three cases the accuracy and predictive capabil-
ity of the theories for the s.f.c. test, we can now return to the multifiber composites
that are our main interest.

IV. Multifiber Composites: Global Load Sharing

A. GLOBAI.
LOADS H A R I N G

In the real multifiber composites of practical interest, many features of the s.f.c.
are preserved. Namely, the fibers have the same statistical strength distribution
and, once broken, slip with sliding resistance r over a slip length 6 = r a / 2 r
(eq. (7)).The evolution of the fiber fragmentation during loading is, in principle,
different because each individual fiber experiences a nonuniform stress due to the
uniform applied stress plus stresses transferred from other broken fibers in the
composite. The evolution of fiber damage thus depends crucially on the nature of
the load transfer from broken or slipping fibers to unbroken (elastic) fibers.
In this section, we make the assumption of global load sharing (GLS) to deal
with the load transfer (Curtin, 1991b). Global load sharing assumes that the load
lost by a fiber at some axial position z , due to breakage and slippage, is trans-
ferred equally to all unbroken (elastic) fibers in the cross-sectional plane perpen-
dicular to z . For a load drop of Aa(7) along a single broken fiber in a com-
posite of n j fibers, the remaining r z / - 1 each experience an increased stress
of Aa(z)/(n f - 1). For n broken/slipping fibers at position z with stress drops
AD,( z ) ,i = 1, . . . , i i , the remaining n - n fibers experience an increased stress
[ l / ( n / - n ) ] C:’, Aa, ( z ) .In GLS there are no local stress concentrations; dam-
age in any part of the composite at z influences the stress state everywhere else in
the composite at z to an equal extent.
186 W A. Curtin

Is the GLS assumption reasonable? The extent of stress concentrations around


broken fibers depends on many factors: the interface t, fiber and matrix elas-
tic moduli, fiber arrangement, crackeayielded state of the matrix, among others.
However, if t is sufficiently low, then the local stress concentrations cannot be
too large and the stress must be redistributed over some fairly large number of
fibers, approaching the GLS limit. In the limit t += 0, the slip length becomes
infinite and so a broken fiber is decoupled from the remainder of the composite
except through the gripping system and hence GLS becomes exact. In any case,
Section V is devoted to analyzing the effects of local load sharing in considerable
detail.
The assumption of GLS leads to several important simplifications to the mul-
tifiber composite problem that then allow for analytic calculations of the damage
and failure. First, since all unbroken (elastic) fibers experience equal stress in-
creases from all of the brokedslipping fibers, the stress on each of these fibers is
the same, and is given by

where of is the remote applied stress on the fibers. Furthermore, since (i) the
fiber strengths are all selected from the same probability distribution and (ii) these
strengths are statistically uncorrelated along the length of the fiber, every cross-
sectional plane is statistically identical to every other cross-sectional plane in the
composite. In other words, if one analyzed the strength distribution of all the fibers
in arbitrary planes ZI and z 2 , one would find these distributions to be identical,
and also identical to the strength distribution along the length of any one single
fiber in the composite. Therefore, the average amounts of damage n and stress
transfer C:, ACT, ( z ) are independent of location z so that the stresses on the
unbrokenfibers are independent of z as well, T ( z ) = T . These statements hold in
a statistical sense, or for large numbers of fibers in the cross section and, moreover,
for any state of damage at any applied fiber stress. The stress T , being constant
along the unbroken portions of every fiber, can therefore be interpreted as an
effective remote applied fiber stress on each fiber. The value of T is related to both
the true applied fiber stress and the damage (fiber breakage) that is caused by
the stress T itself acting on the fibers.
Now, as the true applied fiber stress of is increased monotonically, T also in-
creases monotonically (increasing damage). Therefore, eachfiber in the mult$ber
composite is, during the loading history of the composite, subjected to an effec-
tive stress T along its length and so undergoes fragmentation that is identical to
the fragmentation of thefiber in a single-fiber composite. To reiterate, every fiber
Fiber-Reinforced Composites 187

behaves as if it were in an s.f.c. under stress T . The relationship between T and


of is as yet unknown but is uniquely determined, as we shall see below. So, at
any applied stress of the fibers are fragmented into pieces having precisely the
statistical distribution P ( x , T(of))where P is the s.f.c. distribution derived in
Section 111.

B. FIBER
PULLOUT A N D WORK OF FRACTURE

Putting aside the T (of) dependence temporarily, imagine that the composite
has been tested to infinite strain or stress so that of + 00, T + 00. Then, each
fiber is fully fragmented with the final fragment spacing distribution P ( x , 00)
(see Figure 4 ) . The composite will separate into two pieces along some arbitrary
plane (recall all cross sections are statistically identical). The fibers will pull out
of the matrix around this fracture plane, with the shorter of the two embedded
lengths (one on each side of the plane) being pulled out because the force to pull
out an embedded length h against the interfacial shear is ( 2 n r ) s h .
Knowing the distribution of fragment lengths P ( x , oo),the distribution of pull-
out lengths Pp,(h) is obtained as follows (Curtin, 1991b). The probability of a
plane intersecting a fiber fragment of length x is simply x P ( x , o o ) / ( x f )where
( x f )is the final average fragment length as before. The pullout length distribution
p t ( h )caused by such a size-x fragment has equal probability between 0 and x / 2 ,
i.e.,
2 X
p , ( h ) = -, 0 < h < -. (24)
X 2
The probability of obtaining a pullout length h , Pp,(h), considering all possible
fragment lengths x is then

The normalized distribution Ppo(h)6, versus h/6, is shown in Figure 8 for various
rn values. Since most fragments are longer than 6,/2 (see Figure 4 ) , the pullout
length distribution is nearly constant for lengths h < 6,./4.
The average pullout length is ( h ) = 1; d h h Pp,(h), but is easier to understand
as follows. For an intersected fragment of length x , the average resulting pullout
is x / 4 since it is equally likely to intersect the fragment anywhere along the length
and the shorter piece pulls out. So,
188 W A. Curtin

1.o
Pullout Length h/Sc
F I G .8. Normalized liber pullout length distribution lor various Weibull moduli. Reproduced from
Curtin ( I99 I b) with permission of the ./o~trucr/q f r h Anrericm Cerrrmic Sociefv.

where (x; ) = dx x ” P ( x , 00) is the nth moment of the final fragment length
distribution. Introducing the reference length S,, we thus find

1
( h ) = - h(m)S,. (271
4

Here, h ( m ) = ((xf /S02)/(xf /ac) is a pure number that vanes slowly with rn and
is reasonably approximated (within 5%) by

0.664
W)
= ,0.6 +0.716, m 1.

Fiber pullout thus scales directly with the characteristic length 6,. . Measurement
on ( h )can thus be used to estimate 6,. even if in is not known. Measuring the full
distribution P,,(h) can be used to determine the value of m as well.
The work to pull out the broken fibers is the major contribution to the work
of fracture per unit area in many composites. The work to pull out an embedded
length h is ( 2 x r ) r ( h 2 / 2 )Therefore,
. the average work to pull out a fragment of
Fiber-Reinforced Composites 189

length x is
nrr
) -x .
d h ( n r r h z ) p , y ( h=
12
The work per unit area of composite to pull out all of the fibers, W,, is then
(Curtin, 1991b)

upon introduction of a,.and 8,. Here, y ( m ) = ( ( x j /S,.)')/(xj /ac) is another pure


number that is well approximated by
1.87
v ( m > = ,0.75 +0.50, m 2 1.

The work of fracture per unit area is determined only by the material parameters
through the product of the critical stress and critical length, 0~8,.
It should be noted that several workers preceding Curtin (199 1b) derived results
that were subsequently shown to reduce to the same forms as eqs. (27) and (30)
but with different statistical prefactors. Specifically, Thouless and Evans ( 1988)
studied the case of a ceramic composite with a single matrix crack, an impor-
tant limiting case, and found the pullout and work of fracture to be somewhat
smaller (smaller coefficients h and y ) than for the multiple matrix cracking case in
eqs. (27) and (30). Sutcu (1989) considered multiple matrix craclung with an ap-
proximate statistical analysis and obtained results fairly close to those of eqs. (27)
and (30) (see Curtin, 1991b).

c. STRESS-STRAIN BEHAVIOR: RESULTS


EXACTAND APPROXIMATE

To obtain the fiber portion of the composite stress-strain relation, we first rec-
ognize that the composite strain is controlled, at all cross sections, by the elastic
stretching of the unbroken fibers. Since these fibers have a stress T , the composite
strain is precisely E = T / E f . Therefore, the T (a/ ) relationship is the straidfiber
stress relationship for the composite.
The relationship between T and af is now found as follows. Static mechanical
equilibrium requires that the average stress borne by all the fibers (unbroken, bro-
ken, sliding) in any cross section must equal the remote applied fiber stress af.
Considering an arbitrary cross section, it is clear that there is some load-bearing
capacity on all fibers, including the slipping fibers, except for those fibers broken
exactly at this cross section. Because the fibers are statistically identical and under
190 W A. Curtin

an identical effective applied stress T , the average of the fiber stresses across the
cross section is identical to the average stress along the length of any one fiber.
Therefore, we must have

where L , is the length of the composite and a(z) is the stress distribution along
the length of an arbitrary fiber. Since the fiber is broken into a set of fragment
lengths x according to P ( x , T ) and since the stress in a fragment of length x is
known, the integral in eq. (32) can be rewritten as a sum over all fragment lengths
of the integrated stress over a fragment of length x:

where o,( z ) is the stress along a fragment of length x and N is the total number of
fragments (breaks) in the length L ; . With an “applied” stress T and with reference
to Figures 1 and 2, it is trivial to find the integrated stress over a fragment length
x as

r
Substituting these results into eq. (33) yields

= pT
lh X2
dx P ( x , T ) -
46
+ pT dx P ( x , T ) ( x -a),

with p = N / L ; the number of breaks per unit length. Using the identi-
(35)

ties Jg dx ~ ( x T, ) = I - Jo2h d x ~ ( x T, I , J20$5 d x x ~ ( x T, ) = (x) -


Jib d x x P ( x , T ) , and ( x ) = L , / N = p - ’ allows eq. (35) to be rewritten in
the convenient and physical form

The first term on the r.h.s. of eq. (36) is the linear relationship that exists in the
absence of damage. The second term on the r.h.s. is primarily due to the reduced
load-carrying capability in the slip regions at the ends of the longer fragments
(x > 26) while the third term on the r.h.s. accounts for the reduced load-carrying
capability in regions of overlapping slip lengths between two nearby breaks. The
damage parameter p includes the total number of breaks, independent of relative
Fiber-Reinforced Composites 191

positions or fragment lengths, and thus makes the second and third terms on the
r.h.s. ofeq. (36) not completely due to long and short fragments, respectively.
Since P ( x , T ) is the s.f.c. fragment distribution at stress T and p is the fraction
of breaks per unit length in the s.f.c. problem, eq. (36) is a unique relationship
between the applied fiber stress a, and the effective stress T , which is also pro-
portional to the composite strain. Both P ( x , T ) and p are really only functions of
the normalized length and stress variables x’ = x / S , and F = T / o ( .so that, upon
introduction of ac and 6,. and the normalized quantities iif= a f /ac,5 = p6,
i i
(the number of breaks in length a(.), and 6 / & = T / o C= T , eq. (36) becomes

Equation (37)is the nonnalized stress-strain curve for the j b e r portion of the
composite, and in this form depends only on the Weibull modulus in. It is the
major resultfor the GLS theory of comnposite defonnation and failure. The leading
nonlinear term is explicitly linear in the damage parameter and quadratic in the
4
strain, as expected physically. The factor of in the second term on the r.h.s. of
eq. (37) expresses the fact that, on average, sections of slipping fiber that intersect
4
the cross section of interest carry - of the stress camed by an elastic fiber. The
broken but slipping fibers can thus make a substantial contribution to the load-
bearing capacity of the composite.
The ultimate tensile strength (u.t.s.) a,,,for
, the composite corresponds to the
maximum of a/.versus T or, in dimensionless form, 6, versus f. Exact and
various approximate solutions to eq. (371, discussed below, all lead to the general
form for the maximum fiber stress as 6; = cp(rn) or 01 = cp(m)o,, where cp(m)
depends rather weakly on m . Substitution into eq. (3) then yields the composite
strength.
The present author obtained a simple analytic form for af versus T by assum-
ing that the damage at the maximum stress was small, in some sense (Curtin,
1991b). Then, with few breaks, it is unlikely to find small fragments x < 26 so
that the third term on the r.h.s. of eq. (37) due to such short fragments can be ne-
glected. Furthermore, the occurrence of fiber breaks is not substantially affected
by the small fraction of fiber length excluded by slip, so that

5 @(6(., T ) = f”’. (38)

With these assumptions, eq. (37) becomes simply


192 W A. Curtin

I
I
0'9

0.8
c I . i

0 1 2 3 4 5 6 7 8 9 10
Weibull modulus m
FIG. 9. Normalized fiber bundle strength 6* vs Weibull modulus m as predicted by various analy-
t
ses. Squares: eq. (40a) (Curtin, 1991b); triangles: eq. (48) (Curtin and Zhou, 1995); crosses: following
from eq. ( 5 2 ) (Neumeister, 1993); diamonds: Hui rr al. (1995) (exact); circles: eq. (54) (Curtin e t a / . ,
1998) (identical to Neumeister for larger m ) .

The maximum of eq. (39) is at

so that the tensile strength of the fibers is

These results for 6; and f* versus m are shown in Figures 9 and 10.
Returning to dimensional units and using eq. (3), the composite u.t.s. is

the failure strain (uncorrected for thermal residual strains) is


Fiber-Reinforced Composites 193
1.6

1.4

'5C 1.2
5
e
.--m 1
U

0.8

0.6
0 1 2 3 4 5 6 7 8 9 10
Weibull modulus m
FIG. 10. Normalized tiher bundle failure strain f* vs Weihull modulus ~n as predicted by various
analyses. Squares: eq. (JOb) (Curtin, I991 h): triangles: eq. (49) (Curtin and Zhou. 1995): crosses:
following from eq. (52) (Neumeister. 1993): diamonds: Hui er ril. ( 1995) (exact); circles: eq. ( 5 5 )
(Curtin rt d..1998).

the stress-strain curve after yielding or matrix cracking is

and the damage per length 8,. at failure is

These results were anticipated to be accurate for ,Z << 1, which is valid for m ? 3
or so.
An indication that eqs. (39)-(44) become inaccurate at small m is that the pre-
dicted 6;drops below the average fiber pullout stress upo,which is not physically
possible. Recalling that the force to pull out the average embedded fiber of length
( x l ) is 2 r r r s ( x f ) ,the stress up" to pull the composite apart against the sliding
resistance after all of the fibers have broken is, using eq. ( 2 6 ) ,

op0= 51 h(m)a, or
- 1
ap0= - k ( m ) .
2
(45)
194 W A. Curtin
For m < 2, 8ppo exceeds the 8; of eq. (40); eq. (45) must be a lower bound for the
true 5;.
To improve the tensile strength and failure strain predictions at lower m and to
include the possibility of initial fiber damage due to processing, Curtin and Zhou
(1995) accounted approximately for the random occurrence of close fiber breaks
(small fragments) as follows. If N breaks are placed randomly in a fiber of length
L , with no regard for the exclusion zones, the fragment distribution is exponential

P ( x , T ) x pePPX, (46)
with p = N / L z as before. Some break spacing will be smaller than 26 merely due
to the random placement. Using eq. (46) in eq. (37), the integrations are easily
performed to yield, after a bit of algebra,
- 1
fff = z ( 1 - e-Pf), (47)
P

as the dimensionless stress-strain curve. Assuming ,ij = Fnl again, an approxi-


mate analytic form for the maximum of eq. (47) was derived as

with the damage parameter at failure fi* = (2/m)m/(mf')and the failure strain
given by

The predictions of eqs. (48) and (49) are shown in Figures 9 and 10, and the
predictions are improved significantly at lower m values.
Neumeister (1993) followed the work of the present author by using the Curtin
(199 1a) solution for P (x , T ) to obtain the stress-strain relationship more accu-
rately. Neumeister (1993) nicely showed that the integrals required, such as in
eqs. (36) and (37), could be performed analytically so that the resulting equa-
tions involved only simple functions of the break density q that enters the Widom
fragment distribution (see Section 111). Moreover, Neumeister then showed that
the Curtin solution provides no actual maximum in the stress-strain curve for
m 5 1.6, although down to m = 1.6 the results looked very reasonable. Thus,
there was something amiss with the Curtin solution, although the pullout stress
a,,. (eq. (45)) does exist below m = 1.6 and so provides the maximum tensile
stress that would be obtained asymptotically at infinite strain.
Fiber-Reinforced Composites 195

In light of the limitation of the Curtin theory and its unwieldy implementa-
tion, Neumeister (1993) then developed an excellent analytic model for the stress-
strain behavior by better accounting for the exclusion zones formed by slip and
for the stresses carried by the short fragments x < 26. His result cannot be ob-
tained directly from eq. (37), however, and so we merely quote the analytic result
of

The damage parameter, including the effect of the exclusion zones in preventing
some breaks, could be written as

leading to

The ultimate strengths and strains predicted by maximizing eq. (52) are shown
in Figures 9 and 10, respectively. The predictions for m > 1.6 are essentially
identical to the values obtained from using Curtin’s theory and so the latter result
is not shown separately. The Neumeister expression of eq. (52) does not have a
maximum for m < 1.2, which has not been previously recognized in the literature;
quotes of Neumeister’s result for smaller m do appear in the literature but may
have been obtained through an assumption that the maximum strain is given by
eq. (49). Jansson and Kedward (1996) found an alternative analytic form that
reproduces the tensile strength results of Neumeister (1993) quite closely but this
result does not have an associated stress-strain curve with a maximum at the
analytic strength.
Hui et al. (1995) used their exact solution for P ( x , T ) to numerically calculate
3f versus T , using the concepts leading to eq. (37), and numerically evaluated
the ultimate stress and failure strain. Exact analytic results were not possible. The
exact stress-strain curves for various m are shown in Figure 11 , and the predicted
3; and F* are shown in Figures 9 and 10. Hui et al. proposed approximate ana-
lytic relationships for the stress-strain curve and failure stress for various regimes
of m, but these results are no more accurate than the Neumeister results form > 1
and so are not reproduced here. Hui et al. did demonstrate, however, that the fail-
ure strain estimate of eq. (49), also derived from separate analyses of Ibnabdeljalil
and Phoenix (1995b), was quite accurate for m > 1.
196 W A. Curtin

: 4 0.8 -
a
2
*
2- 0.6-
0
3
0.4 -
0
LL
0.2-

0.0
0.0 0.5 1.o 1.5 2.0 2.5
Normalized Strain T
Fici.1 I . Exact normalized fiber bundle stress-strain curve ( f i t vs f*)for various Weihull moduli
117. Reproduced from Hui ei ol. (1995) with permission of Elsevier Science.

Finally, Curtin et al. (1998) recently derived an approximate result by alter-


native means that is surprisingly simple and accurate. The analysis is based on
considering more completely the presence of discrete matrix cracks in CMCs, but
in the limit of zero matrix crack spacing such an analysis leads to a new result for
the present problem. Here, we quote the results only. The stress-strain curve was
found to be

with a tensile strength of

and a failure strain of

These predictions are numerically almost identical to those of Neumeister (shown


in Figures 9 andlo), but retain a maximum for all m. Figures 9 and 10 show
the predictions of eqs. (54) and (55) at rn = 1, where the Neumeister results do
not exist. The failure strain of eq. (55) is smaller by a factor of ((in l)/(m + +
2))”(’”+’) than eq. (49), which is fairly negligible for m 3 3. In CMCs with
finite matrix crack spacings, Curtin et al. (1998) showed that the tensile strength
Fiber-Reinforced Composites 197

(eq. (54)) is only modestly affected whereas the failure strain (eq. (55)) can be
substantially affected by the degree of matrix cracking.
Several other works are worthy of note. Sutcu (1989) and Schwietert and Steif
(1990) derived models for tensile strength that, when analyzed carefully, embody
the same mechanics and approximations as found in Curtin (1991b). Thouless
and Evans (1988) also analyzed the composite strength in the presence of only
one matrix crack to obtain an upper bound to the strength with multiple cracks.
Phoenix and Raj (1992) developed strict, tight, upper and lower bounds on the
tensile strength but their results were superseded by the more precise results of
Neumeister (1993) and of Hui et al. (1995).
In summary, a number of results have been derived for the stress-strain curve,
tensile strength, and failure strain. It is evident that all of these results are nearly
identical for m 2 3. This similarity arises from the fact that, at higher m , the
amount of damage at the peak applied stress is fairly small when measured as
the fraction of damage per length 6,. Hence, the amount of overlap in slip zones
and exclusion of regions from breaking is small enough to be almost neglected
entirely. While the deviations for m < 3 are important in some systems, the
deviations in tensile strength are still relatively small. We shall compare these
predictions to detailed experiments in Section 1V.F.

D. INSlTU FIBERSTRENGTH AND FRACTURE


MIRRORS

The strengths of the individual fibers in situ, i.e., after full processing of the
composite, are often different from the values measured on pristine fibers. The
observed degradation is not surprising since the brittle ceramic fibers are exposed
to high temperatures and possible abrasion during composite fabrication. Two op-
tions for assessing the in situ strength have been used. The first method is dissolu-
tion of the matrix, using reagents that do not affect the fibers, followed by direct
single-fiber tension tests on the extracted fibers. The second method involves in-
terrogating the fracture surfaces of the pulled-out fibers on the fracture surface
of a tested composite. Many ceramic fibers, particularly the widely used Nicalon
CG fibers, show “fracture mirrors” on the fracture surfaces. A mirror is indicative
of an underlying critical flaw that caused the fiber failure. Empirical relationships
between the mirror radius a , and the fiber strength S due to failure at the flaw
indicate that
198 W A. Curtin

1.2 -

f
+.L 1 -
I

!i
0.9 -

0.8 -
5
0.7 -

0.6
0 5 10 15
Weibull rn
F I G . 12. Fracture mirror parameters S* and in* vs true Weibull modulus. Squares: S*/o;.; di-
amonds: m * / m . Reproduced from Curtin (1991b) with permission of the Journol o f t h e Americrrri
Cerortiic. Society.

where K is a fracture-toughness-like parameter. The value of K can be obtained


by comparing measured S and values for single fibers tested in tension ex siru.
The measurement of many mirror radii ( a w zon } the composite fracture surface
allows for the creation of an in situ strength distribution ( S ] . Often, the distri-
bution can be adequately characterized by a Weibull distribution with Weibull
modulus m* and characteristic strength S*. However, S* and m* are not identical
to the needed quantities 0;.and m. Nonetheless, since within the GLS approxima-
tion the fibers break as if in an s.f.c. test, there is a unique relationship between
( S * , m*) and (ac,m ) (Curtin, 1991b). The distribution ( S } is precisely the num-
ber of breaks versus stress in the s.f.c. test and does indeed follow an approximate
Weibull form over the major portion of the probability range. The relationship
between ( S * , m*) and (ac,m ) is expressed as ratios of S*/uCand m * / m ver-
sus the true Weibull modulus m , and is shown in Figure 12. The differences are
small (less than 10%) for m 3 3 but become significant for smaller m. Even for
larger m , quantitative agreement with experiments requires proper conversion of
(S*, m*) to (cC, m).

E. LOCALIZATION
A N D NUMERICAL
SIMULATIONS

Prior to the availability of the exact results of Hui el al. (1993, several work-
ers devised numerical simulation methods to check the accuracy of approximate
Fiber-Reinforced Composites 199

analytic results within the GLS theory. In addition, the issue of strain and damage
localization is important, in principle. Since the macroscopic stress-strain behav-
ior has a peak at 67 followed by a softening regime (see Figure 1l), localization
of damage is expected to occur somewhere in the material once 6 ; is reached. The
localization then interrupts the overall progression of fiber damage envisioned in
the analytic models: i.e., each fiber is no longer undergoing an s.f.c. test. The lo-
cal stress T in one region of width % 26, along the length becomes larger while
damage elsewhere ceases since the applied stress is not increasing. Since the total
damage 5" at the maximum stress (e.g., eq. (44) or prior to eq. (49)) is small,
much of the observed fiber damage on the fracture surface occurs after the local-
ization. The localization thus formally invalidates the predictions of pullout, work
of fracture, and fracture mirrors using the s.f.c. theory. Although the stress and
length can continue to be referenced by cr( and 6, , the values of the coefficients
h ( m ) ,y ( m ) ,and in Figure 12 are not necessarily correct. The simulation methods
are thus valuable to obtain accurate coefficients for these important quantities.
The simulation methods have been adequately described (Curtin, 1993a-c; Ibn-
abdeljalil and Phoenix, 1995a; Zhou and Curtin, 1995; Iyengar and Curtin, 1997a,
b), but various small errors, such as incorrect pullout planes and possible effects
due to the small lengths of composite simulated, raise some doubts as to the valid-
ity of some of the results for pullout properties. We have thus used the simulation
models most recently described by Iyengar and Curtin (1997a, b) to revisit the
issue of pullout and to analyze the fracture mirror statistics, which have not pre-
viously been studied numerically. We performed five simulated tests at various
Weibull moduli, on composites composed of 1000 fibers and a length 46,. The
longer length was used to avoid possible boundary effects in the determination of
the pullout. The coefficients h ( m ) ,y(m), S*/a,, and m " / m so derived are shown
in Table 1, labeled as Sim., along with the predictions of the analytic models from
eqs. (28) and (31) and Figure 12, labeled as Anal. The agreement is remarkably

TAIjLk I
COMPARISON O F ANALYTIC A N D SIMULATION-DERIVED VALllES FOR PULLOLIT, WORK O F
FRACTURE,A N D F R A C T I I RME I K K O KPARAMETERS, FOR SEVERAL WE1BUI.L MODULIm

Pullout h(,rr) Work y(m) Minors S*/rr, Minor5 ~JI*/JII

m Anal. Sirn. Anal. Sim. Anal. Sim. Anal. Sim.


1 1.36 1.38 2.24 2.62 0.65 0.7 1 1.1s 1.21
2 1.18 1.19 1.62 1.95 0.89 0.86 I .07 I .08
s 0.97 1.01 1 .05 1.16 I .o I 1.01 0.98 0.98
10 0.86 0.94 0.81 0.96 I .03 I .01 0.94 0.98
200 W A. Curtin
31

I /’
I I l l , I /

-6 -5 -4 -3 -2 -1 0 1 2 3
In (stress)
FIG. 13. Simulated fracture mirror strength distribution for various Weibull moduli, plotted in
Weibull form. Note the rollover in the distributions at higher stresses. Derived parameters S* and m *
are shown in Table 1.

good. The fiber fracture strength probability distributions, or mirror distributions,


IS} obtained in these simulations are shown in Figure 13. These distributions
show a characteristic rollover at higher probabilities such that the m* is best de-
termined by only considering the probability range below about 0.7 or so (below
In(- ln(0.3)) = 0.2 on a Weibull plot, as in Figure 13). Otherwise, the simulated
results are quite consistent with the analytic models.

F. INITIAL
FIBERDAMAGE

Another important issue in practical composites is that of actual fiber break-


age during composite processing. Preexisting fiber damage modifies the over-
all stress-strain behavior and reduces the composite tensile strength and failure
strain. Experimental work on Ti-MMCs by Groves et al. (1994) has clearly identi-
fied various causes of fiber damage during processing. Since early analytic models
were based on the assumption of limited fiber damage, such concepts were antici-
pated to fail in applications to predamaged composites at moderate damage levels.
Fiber damage is, however, easy to incorporate into the exact and approximate
GLS models derived in Section 1V.C. The total possible fiber damage (preexisting
plus stress-induced) is simply represented by a new “flaw” population which is
not of the Weibull form. Various workers have investigated the case where the
number of flaws per unit length weaker than the stress CT is described by
Fiber- Reinforced Composites 20 1
0.8 .-- - ~

0.75
0.7
5
C
0.65
2 0.6
3
Q
0.55
7 0.5
0.45
I-
0.4
0.35
0.3 I 1
0 1 2 3
Damage parameter

FIG. 11. Predicted normalized fiher hundle strength Ci* vs diinen\ionless initial damage parameter
t
pg& (number of initial break5 per length J,.) for several Weibull moduli as obtained in the model of
Curtin and Zhou ( 199s). Note the convergence at high darnage.

rather than by eq. (4), such that there is a certain preexisting density of breaks po
that adds to the additional damage from the “usual” fiber flaws. This assumption is
probably the most severe case, since it assumes that the preexisting damage does
not deplete the remaining available flaws in the fiber. Curtin and Zhou (1995) and
Duva et al. (1995) used eq. (47) along the dimensionless damage parameter 6
modified to accommodate eq. (57) as

6 = p0Sc + f’”, (58)


which shows that the relevant scale for preexisting damage relative to induced
damage is the amount of damage per length &.. Curtin and Zhou ( 1995) and Ibn-
abedeljalil and Phoenix (1995a) used the general result of eq. (57) in simulation
models, and they, Duva ef al. ( 1995), and Hui et al. ( 1995) used eq. (57) in analytic
models based on Curtin’s or the exact s.f.c. solution.
Figure 14 shows typical results for the tensile strength versus initial damage
po& for various Weibull moduli, from the model of Curtin and Zhou (1995); other
workers show essentially identical results. Figure 14 shows that the decrease in
tensile strength is nearly linear with increasing damage, with significant decreases
occurring only for ,006, > 0.25 or so. Since 6, is typically on the order of a few
millimeters or smaller, as we shall see below, the amount of damage required to
degrade the tensile strength is on the order of a few breaks per inch in every fiber.
This is a fairly large amount of damage; the ability of the composite to maintain
strength at such damage levels is due to the fact that it is the damage over the crit-
202 W A. Curtin

ical length 6, that is important. These results also suggest that aligned, short-fiber
composites with lengths only somewhat longer than 6, should perform nearly
as well as continuous-fiber composites. Elzey et al. (1994) used these numerical
studies in tandem with models for initial damage formation in Ti-MMCs to de-
velop comprehensive time-temperature-pressure-damage-strength maps relating
the detailed processing to the predicted final composite tensile performance.

G. COMPARISON TO EXPERIMENTS

The GLS theory has been used to predict and derive properties of a large num-
ber of CMCs and Ti-MMCs. Here we describe applications to CMCs because it is
anticipated that local load sharing provides an even better description of MMCs
and so we defer most of the discussion of MMCs until Section V. The interested
reader should see Curtin (1993~)for the initial applications of the GLS model
to MMCs. Majumdar (1996) and the references therein show limitations of the
GLS model in application to Ti-MMCs. In CMCs, the fibers used to date have
been almost exclusively the Nicalon CG fibers. However, fiber degradation dur-
ing processing makes composites with different interface coatings and matrices
somewhat different. Here we present all of the complete applications of the the-
ory known at present.
Inputs to the theory are the in situ fiber strength parameters ao, rn at length
Lo, fiber radius r , interfacial sliding resistance t,fiber volume fraction f , and
fiber elastic modulus E f . For CMCs, a, = 0 is assumed after matrix cracking.
Predictions of the theory are then outs,the stress-strain curve, the pullout length
distribution and ( h ) ,and the work of fracture W,?.When fracture mirrors are used
for the in situ strengths, the inputs are S* and rn*, which lead directly to a, and
rn via Figure 12, f , and E f . Values of t can then be derived from the measured
pullout length ( h )using eq. (26).
Below, we use the results of eqs. (48) and (49) from the model of Curtin and
Zhou (1995) because of their reasonable accuracy for both strength and strain and
their relative simplicity. Differences with the exact results are generally negligible.
Prewo (1986) extracted Nicalon fibers ( r = 7.5 pm, E f = 200 GPa) from
a unidirectional LAS-I1 glass matrix CMC and measured a 0 and rn at Lo =
25.4 mm. Pushout tests showed r = 2-3 MPa so we use 2.5 MPa here. Inter-
face debonding is accomplished due to a very thin carbon layer formed in situ in
this material. The strength and t values lead to a, and 6, as shown in Table 2
along with f . The predicted aUt,and E / compare very well with the measured
values obtained by Prewo, also shown in Table 2. The predicted pullout length
is ( h ) = 1.7 111111, which is in the range estimated by Prewo for these systems.
Fiber-Reinforced Composites 203
TABLE 2
PROPERTY DATAA N D M E A S I J R LAI )N D PREDICTED STRENGTH A N D FAILURE
CONS'I'ITUTIVE
STRAIN FOR LAS-II/NICAI.ON
(PKEWO. 1986), LAS-V/SCS-6 (JAKMONA N D P R E W O , 1986).
A N D PYKEX/NICALON(TsuD.4 ('t d . . 1996) COMPOSITES

LAS-I1 0.46 1730 3.8 2.5 2411 7.23 758 756 0.97 I.05
LAS-I1 0.46 1740 2.7 2.5 2657 7.97 664 796 0.86 I.22
LAS-11 0.44 1615 3.9 2.5 2257 6.77 670 678 0.90 0.95
LAS-I1 0.44 1632 3.1 2.5 2129 7.39 680 697 I .03 I .09
LAS-V 0.20 3500 8 10 3986 2X.3 557 605 - 0.85
Pyrex 0.48 2900 3.0 4.5 4069 6.78 71 1 I360 I.03 I .75
Pyrex' 1.0 2993 22.5 98 I I.29

Stresses in MPa, lengths in mm. strains in (2: Lo = 25.3 iiiiii in all cases. A sample calculation for
the NicalonPyrex system for a reduced r is indicated by the asterisk and is for illustration purposes.

Table 2 contains four slightly separate systems with marginally different matrix
structure, fiber strengths, f,tensile strengths, and failure strains. The theory pre-
dictions agree exceptionally well with three of the four and are within 10% for
the fourth system, and so seem to capture even subtle changes in material proper-
ties.
Jarmon and Prewo (1986) investigated the properties of LAS-V glass reinforced
with Textron SCS-6 S i c fibers ( r = 71 pm, E f = 400 GPa). The fiber strength
statistics at 1 in., the estimated r . and derived CT~ and 6, are shown in Table 2. The
ultimate strength and failure strain, as predicted and measured, are also shown in
Table 2 and very good agreement is again obtained.
Tsuda et al. (1996) have used a similar approach for the system of Nicalon
fibers coated with 140 nm of pyrolitic carbon and embedded in a Pyrex ma-
trix. The in situ fiber strengths measured after matrix dissolution were quite high,
showing almost no degradation relative to the pristine fibers, as shown in Table 2.
The interface r = 4.5 MPa was quoted based on pushout tests, leading to a, and
6, as shown in Table 2. The predicted tensile strength and failure strain are also
shown in Table 2 along with the experimental values. The theory greatly overpre-
dicts the strength and failure strain in this material. The origin of this discrepancy
is unknown at present. Corroborating data on fiber pullout and fiber fracture mir-
rors for in situ strengths were not presented. It is possible that the interface r is
lower than quoted; Table 2 shows predicted results using a much lower hypothet-
ical r = 1 MPa for illustration. Analysis of pullout would help confirm the value
of r for these materials. Also, stress concentrations at the fiber surface may play
a role as well, although in all other applications such an effect has been neglected
and good results have been obtained. This system clearly requires further study
204 W A. Curtin
TARLC 3
CONSTITUTIVE PKOI’EKTY DATAANI) DliRlVED QUANTITIES I:OR VARIOUS NICALONFIBER
C M C SYSTEMS

Matrix Arch. f s* )?I* (!I) uc rti 6,. 1


-
Carbon Woven 0.22 2200 4.5 410 2200 4.5 I640 I0
Alumina Woven 0.20 1875 5.0 305 1875 5.0 1258 I I (25)
CAS Uni 0.37 2000 I .8 300 2250 1.7 I016 17(17-20)
Blackglas x-ply 0.20 1704 2.1 256 1982 2.0 853 17.5
in-Cord. Uni 0.37 1618 7.0 1620 1585 7.3 7200 I .7
in-Cord. x-ply 0.19 2409 1.8 482 2890 1.6 1517 14
LAS-111 Uni 0.45 2470 2. I - 2800 2.0
Soda-lime Uni 0.43 1380 3.1 - 1440 3.0

Stresses are in MPa, lengths in pin. strains in ‘%.


Other values for r are shown in parentheses. See text
for references.

to elucidate any new mechanisms that may drive composite failure much earlier
than predicted by the present GLS models.
In a large number of other CMCs, fracture mirrors have been used to assess
a,.and m for Nicalon fibers. These systems include a carbon matrix with no fiber
coating (Heredia et al., 1992), an alumina matrix with dual BN/SiC fiber coating
(Heredia et al., 1995), a CAS glass matrix with no explicit fiber coating (Bey-
erle er al., 1992), a Blackglas matrix with pyrolitic carbon fiber coating (Stawovy
e f al., 1997),a modified-cordierite matrix with pyrolitic carbon fiber coating (Sta-
wovy et al., 1997), an LAS-I11 glass matrix with no explicit coating (Jansson and
Leckie, 1992), and a soda-lime glass matrix also with no explicit coating (Cao
et al., 1990). Table 3 shows the measured constitutive properties of these vari-
ous systems and the derived values of oc,m , a,., and r . Values of r measured
by other methods are noted for comparison. Note that these are not all unidirec-
tional materials, and in applications to cross-ply and woven fiber geometries f
refers to the fiber volume fraction in the loading direction only (generally $ of the
total).
Table 4 shows the predicted tensile strength and failure strain for all of the
systems mentioned above. Agreement between theory and experiment is excellent
in almost all cases, particularly for the tensile strength. The failure strain can be
affected by three factors. First, the stress-strain curve approaches zero tangent
modulus at failure so that even small differences in tensile strength can have much
larger effects on the failure strain. Second, it has recently been shown that the
matrix crack spacing, usually neglected, can have a marked effect on the failure
strain but not the tensile strength (Curtin et al., 1998); this effect is neglected here.
Fiber-Reinforced Composites 205
TAHLE 4
TENSILE STKENGTH A N D FAILURE
PREDICTED AN11 M E A S U R E D STRAIN
FOR VARIOUS N I C A L O NFIBERCMC SYSTEMS (SEE T ~ 8 i . e3 )

Ten5ile strength qII\ Failure strain E ,

Matrix Expt Theory Expt. Theory


Carbon 300 330 0.62 0.95
Alumina 270 265 0.83 0.79
CAS 460 520 1.01 1.19
Blackglas 244 26 I 0.78 0.99
in-Cord uni 454 474 0.64 0.60
m-Cord x-ply 348 334 0.98 I .49
LAS-111 790 857 - 1.40
Soda-lime 348 417 - 0.65

Third, there is a relief of residual thermal strains upon matrix cracking so that an
additional strain of

(59)

must be added, where Aa is the thermal expansion mismatch and AT is the cool-
ing range during processing. This correction has been added to the modified-
cordierite materials and is not needed for the Blackglas materials, which are
cracked extensively upon cooling, or the LAS materials, for which the resid-
ual stresses are negligible. The predicted ultimate tensile strength is generally
within 5-10% of the measured value, with the one exception being the soda-
lime system where the difference is 20%. This system is also one in which the
fibers experience the most severe degradation during processing. Overall, note
that there are no adjustable parameters in this theory-all quantities input to the
theory are measured-and hence the general level of agreement found here is
impressive.
The Blackglas material cracks extensively upon cooling. Therefore, the entire
stress-strain curve is dominated by the response of the fibers and, as noted above,
thermal strains are relieved at the outset. The predicted and measured stress-strain
behaviors are shown in Figure 15a, and excellent agreement is evident. The entire
nonlinearity in the deformation is controlled by the statistical evolution of the fiber
damage and provides strong validation of the general phenomenon of cumulative
fiber damage in CMCs. Figure 15b shows the stress-strain curve for an modified-
cordierite cross-ply material and, above the matrix cracking regime, the predicted
behavior again agrees well with experiment.
206 W A. Curtin

400 -/

500
400

0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014


Strain
FIG. 1 5 . Measured and predicted stress-strain curves for (a) Nicalon/Blackglas CMC and
(b) Nicalon/modified-cordierite glass cross-ply CMC. The dashed portion in (b) shows the fiber con-
tribution prior to saturation of the matrix cracking. Reproduced from Stawovy et cil. (1997) with per-
mission of Elsevier Science.

The modified-cordierite unidirectional and cross-ply materials are an interest-


ing pair of materials. Although processed similarly, and with fairly small residual
stresses, there is a factor of 10 difference in the interface t.This is also accompa-
nied by a large difference in critical fiber strength. The origin of the t difference
is not understood. But the fiber strength difference is, at least in part, due to the
strengtMength scaling (eq. (6)) and the very different critical lengths 6, due to the
different t values (demonstrated by the very different pullout lengths). The the-
oretical analysis thus helps show how macroscopic measured properties are con-
nected to underlying constituent properties in each specific system under study.
Note that the theory works equally well for unidirectional, cross-ply, and woven
fiber geometries. The effects of local fiber volume fraction and weave geometry
apparently play a very secondary role in these materials. Although weave geom-
etry is often cited as a source of material variation in PMCs, in CMCs it appears
that, since the critical length 6,. is much smaller than the weave length scale, the
weaving has a negligible effect once in situ strength differences (if any) are ac-
counted for properly.
Fiber-Reirforced Composites 207

40
35
30

10
5
0
0 100 200 300 400 500 600 700 800 900 100
0
Pullout length (microns)
FIG. 16. Measured (squares) and predicted (solid line) pullout length distributions for Nicalonl
CAS. Predictions use !)I = 2 and 6,. = 1000 pni.

There is one system, the S i c matrix deposited by chemical vapor infiltration


onto Nicalon fibers, for which the GLS theory has not apparently been success-
ful and strongly overpredicts the strength and failure strain. Unfortunately, this
is one of the more important materials for current applications. One difficulty in
applying the theory is that the in situ strengths have been more difficult to mea-
sure accurately. In addition, the interfacial t in these materials may be rather
higher than in the CMCs discussed here so that local load sharing may be an
important additional factor. Finally, the influence of incomplete matrix cracking
may also play some role in reducing the failure strain. Recent work based on the
GLS concept has, however, taken into account matrix crackmg and has been quite
successful in predicting the stress-strain behavior, strength, and failure strain of
single-tow “minicomposites” of CVI-SiC/Nicalon material (Curtin et al., 1998).
Future work on full-scale composite coupons using similar approaches may thus
rectify the theory and experiment in this particular system.
To complete our comparisons of theory and experiment, we consider the fiber
pullout length distributions. The measured distribution for CAS glass/Nicalon ma-
terials (Beyerle et ul., 1992) is shown in Figure 16 along with the predicted value
using the Curtin s.f.c. theory with m = 2 and 6, = 1000 pm. There are some
differences at small lengths, which may be associated with the difficulty of seeing
small pullout lengths in a forest of much longer lengths or due to factors not yet
understood and missing in the theory. However, the overall shape of the distribu-
tion is generally well predicted by the theory. Such a measurement helps confirm
the value of the Weibull modulus determined through fracture mirrors.
208 W A. Curtin

H. OTHERAPPLICATIONS

The GLS model shows that the composite strength and deformation are con-
trolled almost entirely by the characteristic strength a,..Changes in the material
parameters DO, m , and t induced by fatigue, temperature, corrosion, and creep
can thus be directly translated into changes in composite tensile strength. Fatigue,
fiber strength degradation, and creep have all been studied carefully to date.
The GLS model shows that strength is proportional to t ' / ( ' ' +(see ' ) eq. (49)
coupled with the definition in eq. (9)). The physical reason for this is that t es-
tablishes the critical gauge length and the critical fiber strength is the strength at
this gauge length. Under cyclic fatigue loading above matrix cracking, the relative
fibedmatrix interface sliding leads to wear of the interface, among other possible
effects. The wearing phenomenon decreases the interfacial t with increasing cy-
cles N , which can be measured directly during testing through the unloadreload
hysteresis. Assuming no other material changes with cyclic loading, strength de-
creases as t (N)l/'"+l'as the critical gauge length increases. When the strength
decreases to the maximum applied stress S in the fatigue test, the material fails.
The fatigue life "S-N" curve can thus be mapped out knowing only t ( N ) and m.
Specifically, the peak stress S at which failure occurs in N cycles satisfies, with
nutsthe unfatigued strength,

If the wear leads to some nonzero asymptotic value t ( N + oo),then the fatigue
threshold stress Sth (life is infinite for stresses below the threshold) is

+
Because of the (m l)-I power in eq. (61), even a reasonably large decrease in
t can yield a high fatigue threshold.
The above fatigue model has been applied to several CMC systems (Rouby and
Reynaud, 1993; Evans et al., 1995). Figure 17 shows t versus N for CAS/Nicalon
(Evans er al., 1995), which was found to be largely independent of the maximum
applied load S, and t decreases by about a factor of 4 in less than 100 cycles.
Using this t ( N ) in eq. (60) with rn = 3, Evans el al. predicted the S-N curve
shown in Figure 18, along with the measured behavior. The agreement is quite
good, with a fatigue threshold at Stl, = 325 MPa based on outs= 460 MPa. For
m = 2 (see Table 3), the threshold would decrease slightly to 290 MPa, which
Fiber-Reinforced Coinposites
25

20
2
I
-J 15
d
QI
0 10
f
).

C
- 5

0
0 1 2 3 4 5
loglO(Cycles)

FIG. 17. Interfacial sliding resistance T vs fatigue cycles N for N i c a l d C A S glass. After Ev;m
PI c d . (19%) with permission of Elsevier Science.

still agrees reasonably well with the data and shows that the sensitivity to m is not
too strong for moderate m values.
Interfacial shear stresses can also decrease under constant load by creep. Du
and McMeeking (1993, Fabeny and Curtin (1996), and Ohno et al. (1997,1998)
have developed creep models for t ( t ) that are driven by the creep relaxation of the
matrix in shear, after the tensile stresses have been relaxed by cracking or tensile

450 I_._--

250
200 I
0 1 2 3
,A
4 5
logl O(Cycles to failure)

FIG. 18. Measured and predicted S-N curve foi- NicaloidCAS glass. Solid symbols denote ex-
perimental failure; open symbols denote test slopped (“run-out”). Predictions are shown for Weibull
moduli rzi = 3 (solid line) and m = 2 (dashed line). After Evans cr t i / . (1995) with permission of
Elsevier Science.
210 W A. Curtin

creep. Inserting t ( t ) into a,.then leads to a time-dependent composite tensile


strength auts(t).
Similar models for PMCs were developed earlier by Mason et al.
(1992). For a power law shear creep relation given by

where f J n is the shear strain rate, t is the shear stress, and n and B are the matrix
power law creep exponent and prefactor for tensile creep, Ohno et af. (1997,1998)
have derived a form for t ( t ) that is particularly convenient

Here,

is a characteristic relaxation time for the shear creep in the composite, and w is
the fiber spacing. Combining eqs. (60) and (62b) and solving for the failure time
yields

Predictions and data for time to failure in a Ti-MMC using this approach are
shown in Figure 19 and the theory clearly shows the trend exhibited by the data.
The absolute time scale is in poor agreement, however. While this may be due to
uncertainty in some of the constitutive property data used, other physical features
may be missing from the simple model. In any case, the failure time is sufficiently
long at these low stresses that the tensile creep has effectively driven the matrix
contribution 0)to 0 well before t f is reached; hence, the analogy with fatigue in
the CMCs is quite appropriate. Du and McMeeking (1993, Ohno et al. (1997,
1998), Fabeny and Curtin (1996), and Weber et al. (1996) have all addressed the
issue of failure at higher loads and shorter times during the period of effectively
decreasing 0).Predictions of another Ti-MMC by Fabeny and Curtin (1996) in
this regime are depicted in Figure 20, and show generally good agreement with
the experimental failure times.
Ibnabdeljalil and Phoenix ( I995b) and lyengar and Curtin (1997b) have also
investigated composite strength versus time due to time-dependent fiber strength
loss. However, experimental data on such phenomena is limited at present and so
these theories will not be discussed here. Iyengar and Curtin (1997~)have also re-
Fiber-Reinforced Composites 21 1

1400

800 -1 I
-1 0 1 2 3 4 5 6 7
loglO(fai1ure time, hrs)

F I G . 19, Applied stress vs hilure time for a Ti-MMC. Solid symbols: measured failure; open
symbols: test stopped ("run-out"): solid line: prediction due to shear creep relaxation. After Ohno PI d.
(1998).

cently theoretically investigated the strength versus time when both creep/fatigue
and fiber strength loss operate simultaneously; their results show a marked non-
linear coupling of these two mechanisms which sharply reduces the composite
lifetime relative to either mechanism acting separately. This may have strong im-
plications in the interpretation of real experimental data.

1,100
1,100

-a"
ao=l.47GPa

1,000 5
I
Y

v)

$
(13
-oO=1.29G
U
Q,
a.
8 800 -

'
0
I 1
700
-4 -2 0 2 4 6
Base10 Logarithm of Time-lo-Failure in Hours
FIG. 20. Applied stress vs failure time for a Ti-MMC. Symbols: measured failure: solid lines:
predicted due to matrix tensile creep lor two different initial fiher strength values. Reproduced from
Fabeny and Curtin ( 1996) with permission of Elsevier Science.
212 M? A. Curtin

I. S U M M A R Y

The GLS model of composite failure is based on the assumption that fiber
breaks do not cause local stress concentrations. From that one assumption flows a
complete theory for the stress-strain behavior, tensile strength and failure strain,
fiber pullout, and work of fracture as functions of the constituent material prop-
erties. The deformation and failure are controlled by the stochastic evolution of
the fiber damage, with many key composite properties determined by the critical
stress rr, and critical length 6, . The success of this theory in applications to CMCs
appears quite good, and with no adjustable parameters. This success has motivated
extensions to time-dependent and cycle-dependent phenomena, and these models
appear to capture the major factors controlling lifetime.

V. Multifiber Composites: Local Load Sharing

A. INTRODUCTION

When a single fiber fails in a composite, its stress is certainly not transferred
to fibers infinitely far away except in the limit r -+ 0. Some sort of localized
stress transfer must exist in the real composite, a situation we generally refer to
as “local load sharing” (LLS). Although GLS is a good approximation, as seen
in Section IV, the existence of LLS to any degree has some important conse-
quences for the nature of the composite failure. First, fiber damage in some local
region increases the stresses around that region and drives further damage lo-
cally. Composite failure is then caused by the formation, at some place in the
material, of a “critical cluster” of damaged fibers that grows/propagates in an
unstable manner across the remainder of the composite. This situation is thus
similar to failure of a monolithic ceramic or the individual fibers themselves,
where one critical flaw causes failure. As a result, the composite failure becomes
statistical, with some probability distribution and some limited reliability. Fur-
thermore, the composite strength depends on the sample volume, decreasing as
the volume increases because the larger volume provides more possible loca-
tions for the “critical cluster” to develop. Finally, the composite is sensitive to
local stress concentrations, stress gradients, or localized induced damage due to
notches, holes, or impact events. None of these features is present within the GLS
approximation, yet all are critically important to the engineering application of
composites.
There is considerable experimental evidence in PMC materials that the failure
strength is both statistically distributed and volume dependent. Although difficult
Fiber-Reinforced Composites 213

to determine and quantitatively assess because of competing failure modes (e.g.,


compression and tension in bend tests), the volume scaling has been presumed to
be due to the localized stress transfer in PMCs. We reference some recent work
along these lines by Wisnom (Wisnom, 1991a, b) and, for ceramics, the work of
McNulty and Zok (1997).
To understand how LLS influences composite behavior, one must address
(i) how to determine the nature of the load sharing in any particular system for
arbitrary spatial locations of breaks and (ii) how to use such information to assess
the damage evolution and failure in large composites. In this section, we review
progress on both of these issues and present in detail one recent model for LLS
that appears to capture features observed in experiments on MMCs and possibly
PMCs.

B. LOCALLOADSHARING
MODELS

The existence of localized stress transfer has long been recognized in work on
predicting the strength of PMCs. The earliest model was a fully elastic model
developed by Hedgepeth and Van Dyke (1967). Hedgepeth and Van Dyke mod-
eled the fibers as one-dimensional elastic elements aligned in a regular array and
surrounded by a matrix capable of carrying only shear stresses. They consid-
ered shear coupling between near-neighbor fibers only and were able to derive a
second-order differentialldifference equation for the coupled axial displacements
{ u ( z ) } of the elastic fibers, with the axial stress following from equilibrium as
p ( z ) a duldz. For fibers in a regular array, they showed that (i) the differen-
tial/difference equations in the presence of a single broken fiber could be solved
exactly to obtain the stress transfer, or “break influence function,” and (ii) the
induced stress transfer in the presence of multiple breaks could be obtained by
solving a matrix equation involving only the fiber displacements at the break lo-
cations followed by simple matrix multiplication. Specifically, for a square lattice
of fibers denoted, Hedgepeth and Van Dyke showed that the in-plane stress on
fibers (tr,l , ) due to a single break at the origin (0, 0) under stress c could be
expressed as K c , , e,u with K c , , y , = -ky,, e , lko, 0 and

-
kL6 - -1. ln 1’
do dg2cos(t,,g2)cos(tr8)

(v))
7r2

1 /2

x (1 + sin2 - sin2 ,
214 W A. Curtin

When multiple breaks exist at locations (t.r,t ,,} in the plane, one then solves

It:. :0 broken)

for the displacements uy,. y , of all the ( t r ,t , }interacting broken fibers under
unit compressive loads (the source of the -1 on the r.h.s. of eq. (66)) and then
the stresses Kp, , p, (r transferred to the remaining unbroken fibers (l:, ti 1 are
obtained by solving

( V , , Y , broken)

The matrix size in the set of eqs. (66) is equal to the number of actual breaks, not
the total number of fibers, and is therefore optimally efficient. Similar results are
obtained for other lattices, and the generalization to out-of-plane stresses straight-
forward.
Hedgepeth and Van Dyke (1967) utilized the above model to analyze stress
concentration factors for various compact clusters of in-plane broken fibers for
linear (two-dimensional), square, and hexagonal fiber arrays. The in-plane stress
concentrations for single breaks in the two types of arrays are shown in Figure 21;
clearly, the load transfer is quite localized, with about 60% of the stress from a
single fiber transferred to the nearest neighbor fibers. In spite of the great power
of this approach, in which only the interactions between the broken fibers need be
determined, it was not utilized to study composite failure. A number of subsequent
workers analyzed the same model problem and presented further results, both nu-
merical and analytical. Suemasu (1984) included some analysis of the probability
of damage around preexisting clusters of breaks and concluded that the enhanced
stresses did not translate into significant increases in failure probability. However,
the extension to full predictions of composite failure with stochastic fibers was
not accomplished.
Sastry and Phoenix (1993, 1995) and Beyerlein and Phoenix (1996~1, b,
1997a, b) have revived the Hedgepeth-Van Dyke (HVD) model. Sastry and
Phoenix (1993, 1995) investigated interactions between out-of-plane breaks to
demonstrate the full capability of the model but did not pursue problems in three
dimensions. Beyerlein and Phoenix (1996b) showed that, for a large linear ar-
ray of aligned breaks, the stresses predicted by the HVD shear lag model agree
very well with linear elastic fracture mechanics for orthotropic materials down to
the scale of one fiber diameter. The stresses in this case were also investigated
previously by Hikami and Chou (1989), who obtained an analytic form for the
stress concentration at the crack tip. Beyerlein and Phoenix (1996a) also devel-
Fiber-Reinforced Composites 215

u (3
(3 @@@

/I@@@@
@Be@@@
\ 0.004

@@@a@
(@@@a@@
@a@@@ (b)
FIG 2 1 In-plane load transfer for a single broken fiber under unit apphed load. a5 calculated by
the Hedgepeth and Van Dyke ( 1967) method (a, square lattice, (b) hexagonal lattlce

oped a method for incorporating matrix yielding, beyond what was done in the
HVD model, and showed the strong decrease in stress concentration factors near
fiber breaks when the yielding is included. This detail is accompanied, however,
by a marked increase in the computational requirements since the matrix regions
must also be treated as non-linear elements. Beyerlein and Phoenix (1997a, b)
also analyzed the failure of a two-dimensional composite with an initial crack and
stochastic fiber strengths, for in-plane breaking only, and showed how the statisti-
cal variability in fiber strength could give rise to toughening or “resistance-curve’’
behavior even in the absence of fiber pullout toughening (whlch dominates in real
216 W A. Curtin

composites). Lastly, Beyerlein and Phoenix ( I 998) have extended this approach
to consider a viscoelastic matrix, and hence the evolution of stress concentrations
as a function of time. All of these efforts put the HVD technique on the verge of
exploring full composite failure but the final step has not yet been taken.
Several other workers have recently analyzed the load transfer around broken
fibers using finite-element modeling aimed at CMCs, MMCs, or PMCs. He et al.
(1992) investigated the stresses around an elastic fiber with a break and inter-
face debonding embedded in an elastic axisymmetric geometry. A ring of elastic
fiber material represented the neighboring fibers, and the stresses on the neigh-
boring fibers were studied as a function of the applied fiber stress C T ~ interfacial
,
t,elastic moduli, and volume fraction. He ef al. found nonconstant stress con-
centrations that were very roughly linear in t/of for small t/af < 0.1 and
attained values approaching the HVD result for hexagonal geometry for large
E f I E , , , and t/a/% 0.2. Since the matrix remained elastic throughout and was
uncracked, the He et al. results showed significant reductions in stress concentra-
tion for Em M E f . Du and McMeeking (1993) analyzed stress concentrations in
a similar axisymmetric model but with an elastic/plastic matrix and a thin elas-
tic/plastic fiber coating. They found the stress concentration on the neighboring
fibers was a function of the applied fiber stress af divided by the matrix yield
stress no, in a manner similar to Beyerlein and Phoenix (1996a), and a function of
the coating shear yield stress to. For equal matrix and coating shear yield stresses
(to = no/&), the stress concentration typically increased rapidly beyond full
matrix yielding, reached a peak of 543% at around af/oo = 5-10, and then de-
creased rapidly with further increases in applied fiber stress. For coatings with a
fairly reduced shear yield stress (TO 5 O.loo), the peak stress concentration factor
typically decreased slightly, occurred at small applied fiber stress, and decreased
rapidly at higher applied stresses. An example of this behavior is shown in Fig-
ure 22 for f = 0.35, E I / E , ~= 3, and varying ro. Load transfer to the level
found in the HVD model was never achieved. Caliskan (1996) used a different
numerical method for the same axisymmetric geometry and elastic fiber, matrix,
and interface to investigate the very small E,/Ef regime. He found that the av-
erage stress concentration on the neighboring fibers was almost identical to that
predicted by the HVD model for both 1 broken fiber, 7 broken fibers (central fiber
and the near-neighbor ring of fibers), and 19 broken fibers (central fiber and two
nearest-neighbor rings of fibers). The in-plane load transfer was also independent
of the matrix modulus when small, as in the HVD model. This work confirmed the
accuracy of the HVD shear lag result in the limit of elastic interfaces and very low
matrix modulus, cases of possible practical applicability in PMCs. Finally, Nedele
and Wisnom (1994) performed full three-dimensional finite-element models (al-
Fiber-Reinforced Composites 217

1.1

;1.02
1
0 2 4 6 8 10
Fiber stress normalized by matrix yield strength. o,/q

FIG.22. Load transferred to near-neighbor fibers (stress concentration factor) for a single broken
fiber in an elastic-plastic matrix (yield s i r e s U O ) with an elastic-plastic fiber coating vs fiber stress
0, and for various normalized coating shear yield stresses rv(,,/cr0as indicated. Reproduced from Du
and McMeeking (1993) with permission of Kluwer Academic. Dordrecht.

though using a cylinder of effective composite material beyond the neighboring


fibers) for one specific case of a PMC (low E , , , / E f , elastic behavior throughout).
They found a stress concentration of only about 6% in the plane of the break,
which is rather smaller than that obtained by Hedgepeth and Van Dyke. Such
results should be revisited again, and in more depth, to fully map out the stress
concentrations predicted in three-dimensional models.
None of the above works included matrix cracking, as relevant to CMCs, and
only Du and McMeeking (1993) permitted full matrix yielding, as expected in
MMCs prior to failure. In total, the above results show that some local load trans-
fer does exist and that the HVD results appear to be an upper bound. However, a
full correlation of the load transfer and the material properties E f , E,,, f ,r , and
matrix yielding does not yet exist in the literature although the Du and McMeek-
ing results are a fairly comprehensive set of data. And, most unfortunately, none
of these detailed numerical techniques can easily accommodate multiple-fiber
breaks in a realistic manner including out-of-plane breaks.

c. STATISTICAL MODELSOF COMPOSITE FAILURE


UNDER LOCALLOADSHARING

With the difficulty of obtaining approximate load transfer values for single bro-
ken fibers in composites, workers interested in the general effects of LLS on com-
posite failure have developed models using the HVD values or other approximate
218 W A. Curtin

rules for the load transfer which are tractable for use in composite theories. An
early model is that of Zweben (1968) and Zweben and Rosen (1970), who con-
sidered the development of successively larger compact clusters of broken fibers
when there is LLS and the fibers have a statistical strength distribution. Exact
enumeration of the statistics became difficult, however, beyond clusters consist-
ing of more than just a few fiber breaks so that Zweben and Rosen could only
provide some general insight and approximate guidelines for composite strength.
Such models do contain the important feature of size scaling, but could not con-
sider all of the likely “critical” damage clusters causing failure. Around the same
time, Scop and Argon (1967a, b) developed similar models for two-dimensional
structures. They recognized that a composite of length L; could be broken into a
series arrangement of L;/S smaller composites, where the slip length 6 was also
stress dependent, and then weak-link scaling could be used to relate the length-
S composite strength to the length-l, composite strength. The use of weak-link
ideas, discussed further in Section V.E, was also introduced earlier by Gucer and
Gurland (1962). The full composite problem then reduces to finding the strength
statistics of a length4 composite. Like Zweben and Rosen, Scop and Argon pre-
sented a conceptual approach to obtaining the strength statistics but could only op-
erationally carry out the calculations for fairly small numbers of broken fibers in a
two-dimensional linear arrangement. They did propose asymptotic forms for the
strength for large numbers of fibers in the cross section but the accuracy of these
forms remained untested. A decade later, Phoenix and co-workers revisited the
“chain of bundles” (chain of length4 fiber bundles) and performed very impres-
sive analyses of failure. Harlow and Phoenix (1981) studied linear arrays of fibers
under the assumption that all of the stress from a broken fiber is transferred onto
the two nearest-neighbor fibers (one on each side). With such a simple load trans-
fer, Harlow and Phoenix found exact recursion relations for the composite failure
probability for increasing composite size. They then showed that, since failure in
very large systems is controlled by small amounts of local damage at low overall
applied stresses, asymptotic methods could be used to accurately relate failure in
a small system to larger systems. These works clearly demonstrated the effects
of local damage-driven failure and stochastic/size-dependent composite strengths
that are the hallmark of LLS. However, the simplified load sharing and geometry
prevented direct predictions of strength in realistic three-dimensional fiber com-
posites. Smith et al. (1983) and Pitt and Phoenix (1983) followed this effort with
work on three-dimensional composites and with “tapered” local load sharing, re-
spectively, but did not make any applications to specific systems.
Various other models for composite failure have been developed since the work
of Harlow and Phoenix (1981), including numerical simulation models. Most re-
Fiber-Reiiqorced Composites 219

cently, Leath and Duxbury ( 1994) have revisited the problem studied by Harlow
and Phoenix and obtained new exact recursion relations. None of these models
has demonstrated broad prediction capabilities and so they have not obtained
widespread use to date. One model, developed by Batdorf ( 1982), captures much
of the dominant features of the Harlow-Phoenix, Zweben-Rosen, and Scop-
Argon models, however, and is fairly easy to apply. Batdorf (1982) considered
the probability of forming Q I isolated breaks in a composite, and then deter-
mined the probability that the neighbors of these breaks would also break under
the stress concentration (‘1 to form Q2 pairs of breaks. The number of triplets
(three breaks) emanating from the pairs was then determined, and so on. Batdorf
thus obtained a recursive relationship between the number Qi of i-clusters and
Qi+l in terms of the number of neighbors ni around an i-cluster and the stress
concentration ci on those neighbors. An underlying assumption, when applied to
three-dimensional fiber arrangements, was the neglect of specific different shapes
and neighbor stress concentrations of the i-clusters. The n ; and c; were consid-
ered averages over the different types of i-clusters. The Batdorf model can be
rewritten to read as (Foster et ul., 1998)

for a composite of n / fibers in length L ; . The composite strength in this model


is defined as the stress at which the largest cluster (size i * for which Qi* = 1) is
unstable to growth because Q;*+l 1. That is, once there exists one cluster of
size i*, it will grow to larger sizes with no further increase in the applied stress.
Batdorf showed that this model produces results very close to those of Harlow
and Phoenix (1981) for the linear array. Batdorf and Ghaffarian (1982) applied
the model to polymer composites with some success using compact clusters and
the HVD results to obtain the requisite ti; and c i . In principle, the Batdorf model
can also incorporate load transfer rules that are a function of both local damage
cluster size and applied stress, although these aspects have not been included to
date.
Two limitations of many of the above models are as follows. First, there is the
lack of spatial variability of the fiber breaks within the length 6 ; this eliminates the
load carried by slipping fibers, which was shown to be one-half the full load within
the GLS model. Second, there is the assumption that the clusters of fiber breaks
are compact and near neighbors; more-diffuse clusters of breaks and long-ranged
interactions among damage are not included in these models. As remarked earlier,
however, the general three-dimensional version of the HVD model, as presented
220 W A. Curtin

by Sastry and Phoenix (1993, 1995) for elastic matrices and by Beyerlein and
Phoenix (1996a) for elastic-plastic matrices, is ideally suited toward pursuing
numerical studies of composite failure. Such studies remain to be carried out,
however.
As a final note, several workers attempted to use the GLS models to approxi-
mately investigate the effects of LLS. Specifically, Curtin (1993b), Phoenix and
Raj (1992), and Phoenix et al. (1997) envisioned that failure due to some critical
cluster occurring in LLS could be approximated by failure in GLS in a bundle
of some effective number of fibers n. Since a finite-size bundle of fibers in GLS
shows statistical variations, as known since the early work of Daniels (1945), the
statistical variations in GLS at size n were thought to somehow reflect those of
LLS. The strengths of large systems could then be obtained by weak-link scaling
in a manner similar to that shown much earlier by Gucer and Gurland (1962).
However, none of these workers could show a definite connection between LLS
and GLS at some particular size n so that while the concept was interesting, spe-
cific predictions could not be developed for specific systems. This work presaged
some very recent work (Ibnabdeljalil and Curtin, 1997a) in which a specific con-
nection between LLS and GLS has been shown to exist (see below).

D. LOCALLOADSHARING MODELOF ZHOU A N D CURTIN (1995)

1. Model Development
Zhou and Curtin (1995) developed a model for studying composite failure that
retained all of the features of the GLS model but included local load sharing
within any plane z of the composite. The introduction of LLS was accomplished
through a discrete model very similar to the continuum model of Hedgepeth and
Van Dyke. However, because of the discretization, Zhou and Curtin (ZC) were
able to induce load sharing that varied from GLS to that of Hedgepeth and Van
Dyke with the tuning of one single parameter in the model. With a tunable load
sharing, this method also is suitable for incorporating results from much more
detailed determinations of stress transfer (Section V.B) with relative ease. The re-
striction to in-plane load transfer also allows the full three-dimensional composite
to be considered as a coupled set of two-dimensional (in-plane) problems, thereby
yielding considerable computational efficiency.
In the ZC model, each individual fiber is modeled as a linear set of springs in
series, with each spring representing a small section of the linearly elastic fiber.
These springs are then arranged in a regular array (e.g., square lattice) and are
coupled through the introduction of pure shear (leaf) springs between the nodes
Fiber-Reinforced Composites 22 1

of nearest-neighbor fiber springs. A schematic of this setup of springs and nodes


is shown for the two-dimensional case in Figure 23. The fiber springs have a
spring constant k, and the “matrix” shear springs have spring constant k, related
to the matrix shear modulus G,fl.As in the HVD model, the “matrix” carries
no tensile load. The entire composite consists of n f fibers labeled by index n
(1 5 n 5 n t ) of length L ; , each of which is divided into N z elements of length
8 = L , / N , labeled by index m (1 5 m 5 N 2 ) . A remote applied stress o f
per fiber is applied to the composite which, in the absence of any fiber damage,
induces a stress (force) o,],= o f in the nth fiber at axial location m .
The key issue is the treatment of fiber breaks within the above spring model for
the composite. Suppose spring element ( n , m ) is broken (i.e., k, for this element
is set to 0). The induced slip along this nth fiber must be put in “by hand,” since
the matrix is purely elastic, as follows. According to eq. (7b), the stresdforce is
linear with distance from the break and takes on the values

(69)

This slip occurs over those fiber elements m’ for which the slip-controlled stress is
less than the total “far-field” applied stress o,]. which includes both the applied
nlj,

fiber stress and stress transferred to element ( n , m’) from other broken fibers. The
slipping fiber elements carry the stress p l l , independent of the displacements of
the nodes connected to spring ( n , m’) and so are “perfectly plastic” springs with
zero spring constant k, = 0. Operationally, the slipping fiber elements ( n , m’) are
then treated as breaks, k, = 0, but with an effective applied stress of - p I I .
The broken and slipping regions along the discretized fiber n must also transfer
the loads af - p n , to the neighboring fibers n’ # n . In this model, we assume that
the transferred load at - p N nl . in plane m is transferred only to otherfibers in the
same axial plane m, as if there were no other damage in any other plane. Equi-
librium requires that this be true on average; i.e., the axial stress carried across
any plane m must be n f o f , independent of any details of the load transfer. How-
ever, the spatial distribution of the loads is, in principle, affected by the damage
or slipping that occurs in other planes (as evident in the three-dimensional HVD
results of Sastry and Phoenix (1993), for example). With this one approxima-
tion, the full three-dimensional problem of determining the stress fields arl, on
every unbroken, nonslipping fiber element, given the applied load of plus the
transferred loads (of - P , ~,,?, ) from the brokedslipping fibers, is reduced to a se-
quence of N z highly coupled planar (two-dimensional) problems. The coupling
in the two-dimensional problems occurs since an actual break in element ( n , m )
affects, through the slip pI1, the state of stress in planes m’ # m for those m’
,,lj,

within a slip length of plane m . Thus, the induced stress state around a broken
222 W A. Curtin

n.m+l

n.m+ I

nm

n.m- I

F I G . 23. Schematic of spring/node model for unidirectional composites with load applied in the
vertical ( z ) direction. (a) Labeling of fibers and nodes by spatial position. (b) Schematic after some
fiber damage (denoted by -x-) -m-).
with slip represented by “plastic” springs (denoted by
(c) Representation of an individual layer with plastic springs replaced by broken springs and applied
closing forces acting against the remote applied forces.
Fiber-Reinforced Composites 223

fiber is three-dimensional but not the exact solution of the full three-dimensional
problem.
Consider now the planar problem of a set of broken fibers ( k , = 0 ) in some
plane in with applied stresses a/ - p,, ,,! on each broken fiber and applied stresses
o f on each unbroken fiber. We wish to obtain the stresses (forces) on all of the
unbroken fiber elements (springs) in this plane m ,assuming no other damage in
the material. This problem is essentially identical to the HVD problem for in-
plane breaks, but for a discrete lattice rather than a continuum medium. We solve
this general problem using a Green’s function technique that is also the underlying
basis for the HVD “break influence function” approach.

2 . Lattice Green’s Function Method

Green’s function methods are widely used in applied mechanics and in physics.
The discrete models used here are carried over from their applications in atomic
deformations in solids and are well developed in the solid-state physics literature
(Tewary, 1973; Hsieh and Thomson, 1973; Thomson et al., 1987). Because of its
origins outside of applied mechanics, we discuss the method in some detail here
to provide a clear reference for future applications.
Consider the lattice of springs and nodes shown in Figure 23. Denote the posi-
+
tion of the node located at physical position t , ; t!, $ t!,? by t! and the axial +
displacement of that node by ug. To hold the nodes at specified positions u = ( u t )
requires the application of forces F = (Ft ] of

F = $u or Ft = @ ~ , p u ~ ~ , 6‘0)

where @ = ( @ i , ~ is ~ ) the force constant matrix with entry @[,!’ equal to the
spring constant connecting sites f! and t’.The repeated index t!’ in eq. (70) implies
summation over that index, as usual. The form of $ for the perfect lattice (no
broken springs) is straightforward to determine by writing out eq. (70) for some
particular site t!. For a square lattice, this becomes

Fu,.e,. PI = k t [ u L , t , . / ; + l + U I , . t v ,(:-I - 2ut.,,i,.,t,]

+ k s [ ~ ( . , + I . t , . i+, u t , - l , t , . f : +uf,,t,+l,t:

+ u t , . t , - l ~ t :-4Uf,,t,,e:]? (71)

so that $ t . y = -2k, - 4k,y,$(, = k , if e, t’ are in-plane neighbors, IlrU, y~ = k ,


1c

if t!, e’ are axial neighbors, and @,, = 0 otherwise.


To obtain the forces in the axial springs, we must solve for the displacements
u = ( u t } given the applied nodal forces F = (Fp].Then, the spring force is
224 W A. Curtin

simply kt(ue,.e , , rz - uy,, (J, !.-I) for the spring with upper axial node l . The
displacements are obtained from the inversion of eq. (70):

u =$-IF = GF, (72)


where G = $I-’ is the Green’s function matrix. The quantity Gel, is the displace-
ment induced at site [’ for a unit force applied at site l . The matrix G is not sparse
like its inverse $ but is the quantity of interest for calculating the spring forces.
In the perfect lattice (no broken springs), the lattice retains its translational
symmetry and hence both $ and G are functions only of the relative displace-
ment l - [’ between two nodes. The perfect lattice Green’s function Go can be
obtained from eq. (70) by using Fourier transforms. Defining the Fourier trans-
form relationship between u y and up in the infinite discrete system through

and similarly for Fr and $[-if, substituting these into eq. (70), and solving for
the Fourier coefficients u q , F,, and GI: = u q / F y , we obtain the perfect lattice
Green’s function as
0 - 1
G, - 2 ( k , (cos(q1) + cos(q, ) - 2 ) + k , (cos(q,) - 1)). (74)

G: in real space is then obtained using the analog of eq. (73), and is the Green’s
function connecting the node at the origin to the node at C in the perfect, undam-
aged lattice.
Now consider the damaged system in which some fibers are broken or, within
the spring model, some springs are missing ( k , = 0). The force constant matrix
can then be written as

$ = $(I - S$, (75)


where S$ is the portion of @ due to the missing springs, or “defects,” in the lattice.
For a defect between neighboring vertical nodes [ and l - ? , the perturbation in S$
has just four entries, S @ t . e = 8$?-:- f - f = -k, and 6@e,c-2 = 6$c-t, = k,.
Hence, in general, S$ has a number of nonzero entries equal to only twice the
number of nodes associated with the defects in the system. The Green’s function
G = @-’ for the lattice with defects can then be written as the solution of
G = G” + G”S$G. (76)
The node displacements ( m e } under the forces { Fp} are then
0
u t = G,_,, F ~ I + G ~ - , , , s ~elup.
~, (77)
Fiber-Reinforced Composites 225

Writing a similar expression for node u f -:, i.e., the node vertically below node
C , and subtracting from eq. (77) leads to the displacement Aut across the spring
with upper node e as

Aut = u i - ~ t - i = ( G ~ - Y , - G ~ - 2 - ~ , ) F t ~ + ( G ~ p i , , - S+tii,
G~-~ - ~ , ,(78)
e’uv‘. )

Now since 81) only connects pairs of vertical nodes e, e-2, the product S $ p l , I ‘ u p
depends only on Auttl, the displacement across the broken spring having top node
l” (a node in the defect space). Applying equal and opposite forces to all pairs
of nodes, F,-; = - F t , and noting that, by symmetry, G:,(-: = GP-:.. (, and
GY = G:-:, y - 2 , we obtain our main result

kl A u ~= gFpv,Fti + @-i,,kl Au,~! (79a)


e, l” E (defect space), t ‘ E (all space),
+
kf A U Y= G0~ p t ~ F y ~0 Gf-ltjkt AUC“ (79b)
t’’ E (defect space), t $ (defect space), e‘ E (all space),

where the sums are only over the top nodes of each spring. In eqs. (79) we have
introduced the dimensionless dipole Green’s function

G:-(, = 2k,(G:-Yr - G:pz-Y,). (80)

Equation (79a) is a matrix equation for the unknown displacements Auy of the
broken springs only (in the defect space), in the presence of arbitrary applied
forces {Fe} on all nodes in [he system. Equation (79b) then determines the dis-
placements of all of the unbroken springs (outside the defect space) in terms of
the displacements of the broken springs obtained in eq. (79a). The main problem
is thus the solution of eq. (79a), which involves matrices of size only equal to the
number of broken springs in the entire system. Equation (79b) involves matrix
multiplication only, and is straightforward. In this formulation of the problem, as
in the HVD approach, the size ofthe matrix inversion is only the size of the number
of defects in the system and so this is the minimum possible size and an optimally
efJicient method. Solving eq. (79a) formally and substituting into eq. (79b), we
find that the forces of interest, kl Aue on the unbroken springs, are given by

k, Au, = G:-rrFti + - G:,,pI,,,)- 1G 0l l l l p u I F ~ t . (81)

where e”, e”’ E (defect space), e $ (defect space), and e’ E (all space), and
repeated indices imply summation within the restricted space.
A few simple results stemming from eqs. (79) and (81) are in order. First, con-
sider a single broken spring at site e’’ with applied force F at this site only. Then
226 W A. Curtin

the forces on all other sites follow from eq. (81) as

or

The quantity G:-t,,/(l - G): is the fractional force transferred to site l from site
l", and so is the load transfer function. Summing up the load transferred to all of
the sites shows that

=1
Y

is required to satisfy equilibrium. Next consider a single break at site l" but with
a uniform applied force F on all sites in the system. Equation (81) then yields,
using the sum rule of eq. (84),

so that the load transfer is the same as if only the broken site were loaded, as
required by superposition.
The above analysis shows that only the dimensionless dipole Green's functions
are needed to determine the spring forces in the presence of arbitrary applied
forces. In applying this approach to the ZC model for composites, we only con-
sider individual layers of springs (fixed layer rn, or fixed axial top node coordinate
l : ) .So, the dipole Green's functions of interest are those which couple the pair
of nodes at lri + + +
+ l , j l,i and l,x^ l , j ( l , - 1 ) i to another (top) node at
l',; + l : j + l,i, i.e., to nodes having the same axial coordinate l , . Since the per-
fect lattice Green's functions only depend on the difference in node positions, we
therefore need the Green's function connecting the nodes at the origin 0 and -?
to the node l , i + lv.$for all integers l , , l , . The dimensionless dipole Green's
functions needed are then
Fiber-Reinforced Composites 227

Hsieh and Thomson ( 1973) showed that the q: integral in eq. (86) could be per-
formed analytically to yield

Jsin’(q,/2) + sin’(qy/2) exp(iq,l.,) exp(iq,.e,.)


. .

X (87)
J I + ~:(sin’(q.,/2) + sin’(qy/2))

where we have introduced the quantity R = m. For a finite system with peri-
odic boundary conditions in A- and y, the integrals over the wavevectors q.y,qy in
eq. (87) are replaced by discrete sums over the Brillouin zone, i.e., q,, = n,,r / N ,
for - N , /2 5 n , 5 N , /2 for a system of N., nodes in the x direction, and simi-
larly for the y direction (Thomson et al., 1987).
Note that the dimensionless dipole Green’s function depends only on the quan-
tity 52, which is the anisotropy ratio of the shear and tensile springs, and the lattice
symmetry. The parameter R thus controls the load transfer function (e.g., eq. (83))
and is an adjustable parameter in the model. There are two limits of particular in-
terest. For R +, 00, the square-root integrands in eq. (87) cancel out and the
stress transfer defined by eq. (83) becomes independent of position and vanish-
ingly small but such that the sum rule of eq. (84) is still satisfied. This case thus
corresponds precisely to GLS conditions. For R + 0, the stress transfer defined
by eq. (83) remains well defined and the resulting integrals become identical to
the HVD results, after some algebra. To relate the discrete problem to the con-
tinuum problem, Zhou and Curtin showed that, for fibers separated by a 2r, a +
single spring representing a physical length 8 of fiber has
Elxr’
k, = -,
s
while the shear springs have

Thus, k, and k , depend on continuum properties and on the axial discretization


length 8. From eqs. (88) and (89),it is clear that R 8. In the continuum limit
6 + 0, we have Q + 0 and hence the expected correspondence with the HVD
results. In the ZC model, however, only in-plane load transfer is considered and
hence the axial and transverse stress variations are explicitly decoupled. This al-
lows any value for R to be selected to control the in-plane load transfer and,
independently, maintain any value of interfacial r for the axial slip. Although the
in-plane load transfer and interface r are coupled in reality, as shown by the vari-
228 W A. Curtin

ous finite-element models discussed in Section V.B, it is nevertheless convenient


to have a general method which allows both features to be controlled separately
rather than fixed at the values required by the (elastic) HVD problem.
Figure 2 1 has already shown examples of in-plane stress concentrations around
a single break in the HVD limit. Figure 24 shows the in-plane ( z = 0 plane) stress
redistributed to surrounding fibers from a cluster of four broken fibers in the z = 0
plane for the square lattice, where each of the broken fibers is loaded with a unit
load ( F = 1). Three different cases of 52 are shown: The C2 = 0.001 case is
essentially identical to HVD; the 52 = 10.0 case is clearly approaching the GLS
limit of equal loads on all remaining fibers; the case !2 = 1.0 is an intermediate
result, showing moderate stress concentrations around the broken cluster that are
not as large as in the HVD limit.

3. Algorithm and Qualitative Results


To employ the overall method described above, the algorithm is as follows.
First, select the number of fibers n 1 , the overall composite length L , > 2&, and
the desired fiber geometry. Second, select the load-sharing parameter R and calcu-
late the dimensionless dipole Green’s functions G: from eq. (87). Third, discretize
the axial length into units 8 << 6,. Fourth, assign each spring ( n , m ) a strength
. T , ~ , ~=~ ~q l n ( l / ( l - R ) ) where C T ~= a[6,/6]1/”1
is the Weibull characteristic
strength at length 6 and R is a random number in the interval (0, 1). These steps
complete the initial setup of the composite to be simulated. A tensile test is then
performed by the following algorithm:
1. Apply a uniform stress (force) at to all the springs.
2. Break the springs that are overloaded, otr, 2 s j j ,
3. Introduce an axial slip along the broken fiber(s) by setting kt = 0 and
assigning the residual plastic loads p j j ,m .
4. Use the Green’s functions and eq. (8 1) to calculate the new stresses (forces)
on all of the remaining elastic springs in each layer m .
5. Under the new local stresses, determine if any fibers are overloaded by the
transferred stress. If yes, go to step 2. If no, go to step 6.
6. Increase the applied stress by a small increment and recalculate the local
stresses on the elastic springs in each layer by solving eq. (81). Return to
step 2.
Mechanical equilibrium at any applied stress is obtained when all unbroken
springs carry stresses smaller than their strengths, which occurs on proceeding to
Fiber-Reinforced Composites 229

(c) (2@@@@@
4

FIG. 24. Load transfer around four in-plane breaks to surrounding in-plane fibers (broken fibers
have unit applied load) for various load transl‘er parameters R. (a) R = 0.001 (HVD lirnir): (b) R =
1.0; (c) R = 10.0 (approaching global load sharing). After Zhou and Curtin (1995) with permission
of Elsevier Science.
a

00000000
000000000

00000000000000000000
0000000000000
000000000000

(b)
F I G . 25. Critical clusters of damage at onset of macroscopic failure instability and in the eventual
plane of composite separation. as simulated in the LLS model of Zhou and Curtin. Fihers broken in
the plane are shown in black, fibers broken in a different plane hut slipping in the plane of failure
are shown in gray. and unbroken tibers are shown in white. (a) Square lattice, m = 10: (b) square
lattice, 111 = 5 ; (c) hexagonal lattice. rn = 10: (d) hexagonal lattice, In = 5 . The critical clusters are
schematically identified hy the outlined regions.
Fiber-Reinforced Composites 23 1

3
0

step 6 above. The simulation is complete when, after step 3 , all springs in any one
or more layers are broken or sliding. At this point, the material cannot support the
applied load and must pull apart.
Failure under LLS occurs by the formation of some localized cluster of fiber
damage that causes stress concentrations that, in turn, drive continued fiber dam-
age with no increase in applied stress. The stress at which failure occurs, and
the configuration of fiber breaks, are statistically distributed. Hence, the material
shows features similar to the failure of monolithic ceramics. Here we show some
of these basic features and defer detailed analysis to the next section. For the re-
mainder of this section, the load transfer studied is identical to the HVD limit
(52 + 0 ) ;this is the most-local load sharing obtainable within the present model
and so represents the case that differs most from the GLS model of Section IV.
Critical clusters Figure 25 shows the fiber damage in the eventual plane of
failure for several different composites just at the point of failure, i.e., where a
tiny increase in stress will cause unstable breaking of the remaining fibers. Most
232 W A. Curtin

of the damage is due to fibers failing above or below the plane of separation, and
hence the “critical cluster” is a complex three-dimensional object. The critical
cluster size increases with decreasing fiber Weibull modulus, and, in general, a
fairly large number of fibers must be damaged locally to precipitate failure. There
are several features that show a distinct contrast with earlier approximate models
of composite failure, such as those of Zweben and Rosen (1970), Scop and Argon
(1967a, b), and Batdorf (1982). First, the critical cluster sizes are much larger than
a few broken fibers, particularly for smaller m . Second, the critical clusters are not
necessarily compact, near-neighbor clusters of breaks but are more diffuse. Third,
the breaks are not all coplanar so that the broken fibers retain some load-carrying
capability in any one plane.
Strength variations Performing many tensile tests on nominally identical ma-
terials (only differing in the specific strength values of (sn,m ) in the fibers) leads
to a distribution of composite strengths. The cumulative probability of failure for
volume n f L , , H n f ~ z ( ~ atj )applied
, fiber stress af is shown in Figure 26 for
various-sized composites. The data are plotted such that a Weibull distribution
would appear as a straight line. Clearly, there is variability in the strength at fixed
size.
Size scaling Figure 26 also shows that the strength distribution depends on
the physical volume of the composite (nf fibers, length L z ) . The characteristic
strength (where In(- ln(1 - H ) ) = 0) decreases steadily with increasing size n f
andor L,.

E. ANALYTIC
MODELSAND WEAK-LINK
SCALING

Materials that fail by unstable propagation of localized damage must, for suf-
ficiently large sizes, have strength distributions that obey weak-link scaling. That
is, a large system can be formally considered as composed of a collection of in-
dependent subsystems coupled in series so that failure in the weakest subsystem
causes failure across the entire system. So, the strength distributions at different
sizes must be related as follows. Let &(of) and H r l f ( o fbe ) the cumulative fail-
ure probability distributions (f.p.d.) at fiber stress C T ~ for systems of size n and n’,
respectively. Considering the size n system as composed of n/n’ subsystems of
size n’ n, the probability of survival at size n , 1 - & ( o f ) , is then simply the
product of the probabilities of survival 1 - H , l j ( a f )for the n/n’ subsystems. The
f.p.d. Hn/(aj) at size n can thus be related to that of size n’ by

HfI(Uf=
) 1 - (1 - Hflr(cTf))fl’n’.
233

-6

-8

FIG. 26. Cumulative probability of tiber bundle failure vs stress as obtained from the LLS simula-
tion model. for ni = 5, length Li = 2&, and n = 100, 196,400, 576. and 900 fibers. Distribution is
plotted such that a Weibull distribution appears as a straight line. Reproduced from Ibnabdeljalil and
Curtin (1997a) with perniission of the Interncrtionnl Joitnial of Structures and Solids.

Ibnabdeljalil and Curtin (1997a) showed that the data in Figure 26 at different
sizes are related by the above weak-link scaling relationship.
Moreover, Ibnabdeljalil and Curtin (1997a) postulated that, at some small crit-
ical “link” size, the failure probability under LLS is identical to that obtained
under GLS for the same size. In other words, failure in a large system follow-
ing LLS is controlled by failure in a small subregion whose failure probability is
the same as the failure probability in GLS. Thus, LLS and GLS are postulated
to give identical statistics at some small size. Phoenix er al. (1997) had previ-
ously derived the strength probability distributions for finite-sized GLS systems
and found that the link length was 0.46, in the axial direction. They also found
that the failure probability distribution for a bundle of n fibers of length 0.46,. is
a Gaussian distribution with cumulant @((of - p:)/y,:*), having mean pL,*and
standard deviation y:*. Ibnabdeljalil and Curtin (1997a) thus argued that there
exists a critical link of size n1 fibers of length 0.46, for which the LLS strength
distribution is Hn,0.46,.(~f) = @((of - p.f,)/y,*,*). This correspondence was ar-
234 W A. Curtin

gued physically by noting that the large amount of spatially distributed damage in
the critical cluster just prior to failure tends to homogenize the local stress fields
and so approaches the spatially constant value prevailing in GLS. With this cor-
respondence at some (as-yet unknown) size n / , Ibnabedeljalil and Curtin claimed
that, by use of eq. (90). the strength distribution in LLS at any size n f and length
L z could be expressed as
,I/ L-/O4b,ll,
&,L,(q) = 1 - (1 - @ ( ( O f - PL,T/)/Y,TI*)) (91)
To show that their postulated correspondence between LLS and GLS held, and
to find the critical size n / , Ibnabdeljalil and Curtin proceeded as follows. They
used the simulation data, as in Figure 26, and the known Gaussian distribution for
GLS to determine what size nl would make eq. (91) true for all of the simulation
data at any fixed fiber Weibull modulus. Inverting eq. (91) to read as
0 4&11//11/ L -
@((Of - &)/Vn:*) = 1 - (1 -fL/l (Of>) (92)
shows that one must find a value for n / such that the r.h.s. of eq. (92) becomes
a cumulative Gaussian with mean j ~ ; ,and standard deviation y,?. Such a corre-
spondence is highly nontrivial.
However, the correspondence does exist: Figure 27 shows how the simulated
data from Figure 26 for rn = 5 at n f = 576 and 900 and L , = 26, transform
under eq. (92) to a Gaussian for n/ = 54, length 0.46,, with mean and standard
deviation identical to those for the GLS case for nr = 54 fibers. Similar correspon-
dences were obtained form = 10 with n / = 21 and m = 2 with n [ = 165. These
results show that the failure of the large bundle of fibers in LLS is controlled en-
tirely by the failure of a GLS bundle of size n l , length 0.46,, via weak-link scaling.
Observation of the critical damage clusters formed (Figure 25) shows that these
clusters (composed of broken, sliding, and unbroken fibers and roughly shown as
the outlined regions) are very close in size to the n/ values obtained purely nu-
merically through the GLSLLS. Based on the analysis of their simulation data,
Ibnabdeljalil and Curtin proposed the empirical relationship

n/ = 403mF' '* (93)


to relate the fiber Weibull modulus m to the critical link size n / . The associated
values of the mean j~r*;,and standard deviation y,T,* for each rn were then tabulated,
and are shown in Table 5.
The detailed origins of the very close relationship between LLS and GLS at
size n / remain unknown. However, the result appears to be quite general. Recent
work has shown that the strength distribution for hexagonal fiber arrangements
is nearly identical to that for square lattices (Curtin and Takeda, 1998a), and the
Fiber-Reinforced Composites 235

i
.---
-1
:s-
\

f -2
tj
W

-3

1 1 , , 1 ,
-5 ' ' ' I " " " " " '
0.62 0.64 0.66 0.68 0.70 0.72
CT
F I G . 27. Small points: cumulative probahility of libcr bundle failure vs stress as oh-
tained from the LLS simulation model after weak-link scaling to critical siie n / = 53.
length 0.48,. (parameters )ti = 5. L: = 28,. and ? I / = 576. and 900). Solid
line: GLS strength distribution for i f / = 54. length 0.4&. Data are plotted so that
a cuniulative Gaussian distribution appears as a straight line. Reproduced from Ihnabdel-
jalil and Cui-tin ( 1997a) with permis5ion of' the /ntenw/rouci/ J o w t i d of Stnrctirws ofid
sol;t/.v.

critical sizes are quite similar (compare Figures 25). Simulations in the absence
of pullout (dry bundles, or r + 0) show the same correspondence between GLS
and LLS but the critical size nl is different from the case studied here (Curtin,
1998).
The relationship between LLS and GLS embodied in eqs. (91) and (92) also
leads to powerful analytic results. Specifically, using the analytic asymptotic be-
havior of the cumulative Gaussian @, Smith and Phoenix (1981) and Phoenix
et al. (1997) have found accurate approximations to the weak-link scaling form
of eq. (9 1) for large sizes. Of interest here is the result from Phoenix et al. ( 1 997),
which shows that the strength distribution for a composite size n and length L ,
can be written as
236 W A. Curtin

with
+
ln(ln(n)) ln(4n)
4 ln(n)
1. (94b)
a*
ii = +5GGj,
Y nI
and where n is the dimensionless composite size

n = -.nfLz (944
0.48,nl
In other words, the strength distribution for nf fibers of length Lz under LLS is
accurately described by a Weibull distribution with characteristic strength 6; and
Weibull modulus ii. Equations (94) also show that the scaling of strength depends
predominantly on a. Along with the tabulated data in Table 5 for the nl, p:, ,
and y:* versus m , eqs. (94) provide an analytic formula for the tensile strength
distribution for arbitrary fibers and composite sizes. Comparisons of this result
with experiment are discussed in the next section.

TABLE5
GLS MEANS T R E N G T H @,; A N D S T A N D A R D DEVIATION
y;, , NORMALIZED B Y n,, FOR THE CRITICAL SIZEn / I N
T H E L L s FAILURE PROBLEM, FOR V A R I O U S FIBER
W E I B U L L MODULIm . THESENUMBERS AREUSED IN
THE ANALYTIC THEORY OF EQS. (94)

2.0 I66 0.6869 0.023I


3.0 99 0.6996 0.0257
4.0 68 0.7256 0.0278
4.5 59 0.7415 0.0285
5.0 51 0.7442 0.0293
5.5 45 0.7516 0.0300
6.0 41 0.7595 0.0303
6.5 37 0.7676 0.0307
7.0 33 0.7760 0.03 10
7.5 31 0.7832 0.0314
8.0 28 0.7909 0.0319
8.5 26 0.7979 0.0321
9.0 24 0.8047 0.0322
9.5 23 0.8106 0.0324
10.0 21 0.8173 0.0327
Fiber-Reinforced Composites 237

F. COMPARISON TO EXPERIMENT

The LLS model summarized by eqs. (94) and Table 5 and/or the numerical
simulation results themselves are specifically for the most-local load-sharing case
possible, corresponding to the HVD case. Hence, the predictions of these models
are expected to best apply to materials with higher sliding resistance or shear yield
strengths T . Comparisons to date have been made on Ti-MMCs, Al-MMCs, and
graphite/epoxy PMCs.
Inputs to the theory are few: As in the GLS case, we require fiber parameters 00,
m , LO,r , and f ;interfacial r ; matrix yield stress o v ;and now for LLS the com-
posite size (number of fibers n f and length L z ) .The reference parameters orand
6 , follow directly from eqs. (9). The critical link size nl is obtained from eq. (93),
the dimensionless fiber bundle strength 5; is then calculated from eqs. (94) us-
ing the requisite data in Table 5, and the composite strength is calculated using
eq. (3).
Ibnabdeljalil and Curtin (1997a) applied the analytic model to a Ti-24Al-
11Nb MMC reinforced with SCS-6 S i c fibers carefully characterized by McKay
et al. (1994). McKay et al. quoted values of a0 = 4577 MPa, m = 8.6 at
LO = 12.7 mm, with r = 56 MPa, crl = 546 MPa, and f = 0.26. The coupons
had dimensions of Lz = 25.4 111111,and cross-sectional area A = 8.26 mm2.The
measured tensile strength was found to be 1251 f 93 MPa. Application of the
LLS analytic model, with 6, = 6.29 mm, n, = 26, and nfL,/0.46,n1 = 53,
leads to the predicted strength of 1338 f 71 MPa. The agreement is within 7%.
The standard deviations are also very comparable, with the experiments broader
in part due to an uncertainty of f0.028 in the fiber volume fraction. This indicates
that the LLS model accounts reasonably well for the reliability of the strength as
well as the average strength.
Foster et al. (1998) have used the simulation model to predict the strengths of
Ti-1 100 MMCs reinforced with Textron SCS-6 S i c fibers, a system studied in de-
tail over a range of fiber volume fractions by Gundel and Wawner (1997). Gundel
and Wawner determined the postprocessed fiber strengths 00,m at Lo = 12.7 mm
by matrix dissolution and single-fiber tension tests. Fiber pushout tests deter-
mined the T values for the debonded, sliding interface. The matrix yield strength
was a, = 950 MPa. The constitutive information and calculated 0, and 6, are
shown in Table 6. Because of the small size of the composites (tape geometry
of 4 x 26 fibers), Foster et al. (1998) performed direct numerical simulations of
the entire %in. composite gauge length rather than using the analytic model of
eqs. (94) valid for large composites. Foster et al. also developed methods to in-
clude free edges along the tape but found results nearly identical to those obtained
238 W A. Curtin
TABLF,
6
MATEKIAI. PARAMETERS FOR Ti- 1 100 REINFORCEI) W I T H SCS-6
Sic F I H E R (CUNDEL
S ANI) 1997)
WAWNER,

Sample f' 00 111 r a,. 4


B 0.15 3930 10.I 188 5082 1.89
C 0.18 4310 13.9 190 5191 1.91
D 0.20 2890 5.8 190 4608 1.70
F 0.26 4270 12.3 65 4856 5.23
G 0.28 4640 12.6 65 5229 5.63
H 0.30 3330 6.8 65 4280 4.61
I 0.35 4410 11.6 81 5126 4.43
-
Stresscs in MPa. lengths in mni

using periodic boundary conditions. The measured and simulated LLS strengths
are shown in Table 7, and very good agreement is found, typically within less
than 10% over the full range of samples. The GLS results are also shown in Ta-
ble 7 and are consistently higher in strength, but not by very much because the
overall composite size (n f L,/0.46,nl FZ 30) is fairly small. Predictions of the
Batdorf model described earlier (eq. (68)) are also shown, and fare slightly bet-
ter than the LLS model. However, Foster et al. showed that the Batdorf model is
actually an approximation to the LLS model so that improved agreement is due
to somewhat fortuitous cancellations of various effects. They also showed that
the Batdorf model predictions differ significantly from the LLS model for smaller
Weibull moduli m 5 5 and do not show the correct trends for notched compos-
ites.

TABLE7
COMPOSITE TENSILE
STRENGTHS ( I N MPA) FOK Ti- I 100 MMCS AS M E A S ~ J K E(CUNDEL
D
A N D WAWNEK, 1997) AN11 A S PREDICTEII H Y THE L L s MODEL(PERIODIC A N D FREE-EDGE).
THE GLS MODEL,THE R U L E OF MIXTUKES, A N D THE BATDORF MODEL.ALSO S H O W N I S THE
Fll3EK PULLOUT LENG'I'H( I N M M ) O B T A I N E D F K O M THE L L s SIMULATIONS.

Sample Measured LLS (per.) LLS (free) GLS ROM Baldorf Pullout
~~ ~ ~ ~ ~~ ~

B I252 1341 1348 I392 1384 1334 0.2 I


C 1300 1470 I474 1531 1543 1464 0.22
D 1230 I353 I367 1414 1326 1319 0.19
F 1496 1630 1643 I708 1802 1631 0.60
G I724 1768 1789 I856 1972 1766 0.65
H I317 I535 IS46 1605 1654 1502 0.55
I 1716 I929 I938 204 I 2151 1900 0.5 I
Fiber-Reinforced Composites 239
2500 1
I

i
-
2000

'E
u)
1500

s! 1000
Q

500

0
0.0 0.2 0.4 0.6 0.8 1.0 1.2
% Strain
FIG. 28. Measured (symbols) and predicted (solid lines). by GLS and LLS models, stress-strain
curves for Ti-I 100 MMC (Sample I). including thermal residual stress corrections. Adapted from
Gundel and Wawner (1997) with permission of Elsevier Science.

Gundel and Wawner used the GLS model, and included thermal residual
stresses, to predict the stress-strain behavior in this Ti-MMC and found excel-
lent agreement up to failure, aside from the over prediction of the failure point.
Their results for two particular samples are shown in Figure 28 along with the
GLS and LLS predictions of Foster et al. The LLS model reduces both the tensile
strength and the strain but otherwise the deformation follows the GLS prediction,
as shown previously by Zhou and Curtin (1995). Foster er ul. also analyzed the
predicted fiber pullout for these materials, and found values of about 200 p m for
materials B-D (high r ) and about 550 p m for materials F-I (low T). These values
are about one-half of those estimated from the GLS model. The measured pull-
out was about 200 p m for all the samples studied by Gundel and Wawner. The
predicted pullout is thus in the right range for samples B-D, but is too large for
samples F-I. The latter disagreement may be due to effects of dynamic fracture
during final composite failure, which are not in the LLS model, or due to incor-
rect assumptions about the axial load transfer in the LLS model itself. This issue
requires further study.
Foster (1998) has recently simulated the behavior of Ti-6-4 MMCs reinforced
with BP Sigma fibers, a system studied by Weber et al. (1996) and Ramamurty
et al. (unpublished). These fibers have a radius r = 50 p m and a postprocessed
strength of 1470 MPa with m = 5.3 at to = 1 m. The composite consisted of a
6 x 40 array of fibers with gauge length 3 in. and .f = 0.326, and other properties
are c,,= 820 MPa and T = 130 MPa. The measured tensile strength was found
240 W A. Curtin

*
to be 1482 14 MPa. Using the LLS simulation model, with a,.= 4820 MPa and
6, = 1.85 mm from eqs. (9), the predicted tensile strength is 1589f 14 MPa. The
predicted strength is again in good agreement with the measured value (7%)and
the reliability (standard deviation) is predicted very well in this system.
Ramamurty et al. (1997) have also used the LLS theory to predict the strengths
of Al-MMCs reinforced with Nextel alumina fibers tested in tension, 4-point
bending, and 3-point bending, and have compared these predictions to their ex-
perimental results. This work provides the first explicit test of the predicted size
scaling of tensile strength. It is well known that, for a linearly elastic brittle mate-
rial with a Weibull strength distribution characterized by Weibull modulus &, the
tension, 4-point, and 3-point, strengths are related due to the volume scaling of the
strength and the different effective volumes sampled in these test configurations.
The effective sample volumes are in the ratio of

m ++21)’
v , : ( 4(rn v4) : ( +
2(rn 1 1)’ v3) 3 (95)

when the overall nominal tested specimen volumes experiencing any amount of
tension are V,, V4, and V3, for tension, 4-point, and 3-point bending, respectively.
The tensile strengths are then in a ratio of the effective volumes of eq. (95) raised
to the ( - l / g ) power, according to the Weibull scaling of strength with size. The
AI-MMCs are essentially elastic to failure, due to a very low matrix (T, , so that
eq. (95) and the associated strength ratios are expected to be valid. Ramamurty et
al. tested the Al-MMC materials with physical volumes of V, = 1.08 x m3,
V4 = 4.74 x m3, and V3 = 2.42 x lop7 m3 and obtained characteristic
strengths of 1821 MPa, 2051 MPa, and 2171 MPa, respectively, clearly show-
ing the volume scaling of strength in the experiments. Comparing the measured
tension and 4-point bending strengths, the strength ratio following from eq. (95)
indicates 6 = 5 I z!c 10, while comparing the tension and 3-point bending strengths
indicates 61= 59 f 6. To apply the LLS theory to this system, Ramamurty et al.
considered the 4-point and 3-point bend tests to be identical to tension tests at
the effective volumes given in eq. (95) with & = 55 and calcuIated the tensile
strengths at these effective volumes using the LLS theory. Such an approach is
valid as long as the effective volumes contain a large number of fibers in both
area and through-thickness of the test specimens, which is the case here.
The Nextel 610 fiber properties are (TO = 2060 MPa, m = 9 at Lo = 1 m,
*
r = 5.5 vm, E j = 380 GPa, and f = 0.652 0.022. The matrix 0.2% off-
set yield stress is = 100 MPa with E,,, = 60 GPa, and so with no debond-
ing at the interface the appropriate r is estimated to be the matrix shear stress
t, = CT, /a = 60 MPa. According to the results of Du and McMeeking (1993)
5 2000
b
v)
1800 ~

1600 --t- I---r ~ 1

(Figure 22), such materials should exhibit LLS approaching the HVD results. The
reference parameters are then a,.= 4856 MPa, 6,. = 445 pm,with t t / = 24. The
predicted composite characteristic strengths versus dimensionless composite size
follow directly from eqs. (94) and are shown in Figure 29 along with the experi-
mental results and the GLS results, which are volume independent. The scaling of
the strength is predicted fairly well by the LLS model, and the predicted strength
is only about 10%too large. In contrast, the GLS result is 35% larger at the largest
volume (tension test). The LLS-predicted trend of strength with length is not per-
fect; the experimental strengths appear to decrease slightly faster than predicted.
However, the general level of agreement does demonstrate the capability of the
LLS model to capture the major size-scaling effects quantitatively with no ad-
justable parameters.
Ramamurty et 41. (1997) also measured the strength distributions at fixed coin-
posite size and estimated much lower composite Weibull moduli of 15-20. This
was attributable primarily to the variations in fiber volume fraction from sample
to sample, variations that do not depend on sample volume and hence do not affect
the volume scaling of the strength. This observation highlights the care that must
be taken in using the strength distributions at fixed size to predict the strength
versus volume via eq. ( 5 ) .
As a final application of the LLS theory, we consider a graphitekpoxy PMC
system studied by Madhukar and Drzal (1991). The fibers are surface-treated
242 W A. Curtin

AS-4 ( E f = 234 GPa) and analysis of s.f.c. tests on these same fibers (Figure 7b)
leads directly to a, = 5783 MPa and 6, = 501 p m (Curtin and Takeda, 1998b).
These fibers were incorporated into an Epon 828 matrix cured with mPDA that is
very similar to the matrix used in the s.f.c. tests. Uniaxial tension tests were per-
formed on specimens of length 152 mm, width 12.5 mm, and thickness 1.8 mm.
The fiber volume fraction was measured as f = 0.677 but the composite Young’s
modulus E , is much more consistent with an effective volume fraction of 0.59. To
avoid this issue (see Curtin and Takeda, 1998b, for discussion), we consider the
tensile strength divided by the composite modulus, nuts/ E, = auts/ f E 1 , which
is independent of the fiber volume fraction. Using n[ = 20, p:, = 0.825, and
y,T,* = 0.0328 for rn = 10.7, eqs. (94) and eq. (3) (with n) = 0 for the matrix)
yields

autS/E,= 0.0161, (96)


while the experimental value is

~uts/E<= 0.0137. (97)


The predicted strength is 18% larger than experiment. For reference, the GLS pre-
diction is 0.0195, which is 43% larger than experiment and again demonstrates the
importance of the volume scaling of the strength for larger specimen sizes when
LLS is expected to be important. The difference between LLS theory and experi-
ment is rather larger than found in all of the previous applications to MMC materi-
als but, as a first application to PMCs, it is reasonable. Similar levels of agreement
have been obtained in other graphite/epoxy systems (Curtin and Takeda, 1998b).
One issue in PMCs that might contribute to the larger discrepancy between
theory and experiment is the role of the loads carried by broken fibers. A constant
t model may not be appropriate here or the deformations around the broken fiber
may be more complex than envisioned by the simple models. The effect of the
load-carrying capability of the broken fibers can be assessed using recent work
(Curtin, 1998) in which, effectively, 5 = 0 over the length 0.46, with the axial
stress abruptly increasing to the far-field value. In such a situation, the composite
strength does decrease below the LLS model of eqs. (94) because the broken
fibers do not carry load near the fiber breaks. For the graphite/epoxy system of
Madhukar and Drzal(1991), the predicted strength becomes

auts/E, = 0.0134, (98)


which is slightly below the measured value. Therefore, the precise nature of the
deformation around the fibedmatrix interface and ‘‘pullout” load of the broken
fibers may become more important to obtain quantitative agreement in PMC sys-
tems.
Fiber-Reinforced Composites 243

A second issue involves the fiber strength distribution for carbon and graphite
fibers. Tensile strength data on these fibers often show a deviation from the
strength-length relationship of eqs. (4)-(6). Specifically, the strength distribu-
tion at fixed size and the scaling of average strength with changing length are
not related simply through the Weibull modulus rn. Thus, some additional work
is necessary to determine how to incorporate the different strength variations into
an appropriate model.

G. SUMMARY A N D DISCUSSION

We have reviewed the development of tensile strength models incorporating lo-


cal load-sharing and fiber strength statistics. The model of Zhou and Curtin ( 1995)
and Ibnabdeljalil and Curtin (1997a) has been reviewed in the greatest detail due
to its computational tractability, inclusion of many physical features, flexible load-
sharing rule, and accurate predictions in several different MMC and PMC sys-
tems. However, most of the composite strength models mentioned in Section V.C
contain the same common physics and mechanics. Hence, even the most-recent
models could, in principle, have been developed and brought to fruition almost
30 years ago except for the limitations of computational power. The reasonably
good success of the recent models in applications to MMCs and, to a lesser extent,
PMCs suggests that the major features of the failure have been captured in these
models.
More sophisticated and complete models are being developed at present. The
work of Sastry and Phoenix (1993,1995) and Beyerlein and Phoenix ( 1996a) does
not make assumptions regarding the three-dimensional nature of the load transfer
and is the faithful extension of the HVD approach to three-dimensions and yield-
ing matrices. McGlockton and McMeeking (unpublished) have extended a finite-
element model due to Cox et al. ( 1994) and Xu e f al. ( I 995), which bears some
resemblance to the HVD model, to include interface slip. They find strengths that
are closer to GLS in simulations on small composites because the load sharing
depends on both r and the applied stress (see Figure 22). The detailed results that
emerge from these models may show some important differences with the LLS
model of Zhou and Curtin and Ibnabdeljalil and Curtin. However, these more-
complete models are also computationally demanding. Since in any numerical
model many simulations must be performed on moderate-sized systems to cap-
ture the necessary statistics for performing weak-link scaling, comprehensive new
results from these models may not be forthcoming in the near future.
Another possibility for future directions exists if models such as those of Mc-
Glockton and McMeeking (unpublished) or Beyerlein and Phoenix (1 996a) can
demonstrate that the planar decoupling approach used by Curtin and co-workers
244 W A. Currin

is reasonably accurate. Then, there exists the prospect of combining the very flex-
ible LLS model described in detail here with load-sharing rules (possibly stress-
dependent) obtained from detailed finite-element models of load transfer around
various arrays of fiber breaks, as in the works of Du and McMeeking (1993),
Nedele and Wisnom (1994), Caliskan ( 1996), and others.

VI. Future Directions

After 10 years of active research, considerable progress has been made in pre-
dicting the strength and deformation in tiber composites, including the very im-
portant aspect of stochastic fiber damage.
The GLS model, stemming from the underlying s.f.c. problem, provides the
basis for understanding the relevant strength and length scales in the composite
and shows that many features of composite tensile failure (strength, stress-strain
behavior, fiber pullout, fracture mirrors, work of fracture) are related and depend
almost entirely on the key parameters mc, 6,., and m . Predictions of CMC proper-
ties using the GLS model are in very good agreement with available experimental
data, and provide a first approximation to MMCs as well. The GLS model also
provides a framework for trying to understand what physical features may exist
in some materials that cause them not to agree with the theory. Furthermore, the
theory provides a guide for material optimization and for the analysis of the trade-
offs expected in performanc- as constituent properties are varied. The success of
this basic model has generated a host of extensions to related time-dependent fail-
ure problems, such as fatigue, creep, and fiber degradation. The extended GLS
models then provide guidelines for quantitative prediction of composite lifetime
under various modes of degradation.
The LLS models recently developed have built upon the GLS model but in-
corporate the nonglobal nature of the load sharing expected to be prevalent in
many real materials. The LLS models then give rise to finite reliability, size scal-
ing of strength, and sensitivity to local stresses and/or damage. All of these issues
are of considerable importance in taking composite materials from the laboratory
coupon scale into realistic engineering applications. Preliminary applications of
one LLS model to several MMC systems show the predictive capabilities of these
models and provide some new concepts for dealing with the problem. However,
the field of “LLS models” for quasistatic tensile strength is still evolving due to
the complexity of dealing with both heterogeneous stresses and fiber strengths in
large systems. Some important questions still to be resolved are as follows. What
is the precise dependence of “load sharing” on the underlying material parame-
ters 5 , fly,E,,,, E t , and j ’ , and on the applied stress and state of damage (multiple
Fiber-Reinforced Coniposites 245

fiber breaks in three-dimensional, matrix cracking in CMCs)? What are the ap-
propriate models for the interface shear stress, and, in pai-ticular, how do features
of Coulomb friction and asperity-controlled friction enter into the problem? How
effective are broken fibers in carrying load in the “slip regions” in the presence
of multiple breaks, and does this vary from material to material? Do specific de-
formations around fiber breaks, such as the shear banding in Ti-MMCs observed
by Majumdar et al. (1996a, b), drive local fiber failure in a manner different from
that envisioned in all current LLS models’! Is it accurate to neglect the radial vari-
ations in the fiber axial stress by using the shear-lag models? Recent work by
Weitsman and Zhu (1993) shows a stress concentration at the tibedmatrix inter-
face that could drive fiber failure to occur at lower applied stresses than in models
that neglect these variations. Will further careful applications of the existing LLS
models to PMC systems lead to satisfactory agreement or will new features such
those noted above be necessary for obtaining quantitative agreement? Can these
models predict notch strengths and strengths in the presence of stress gradients in
accord with experimental data? These and other issues are all important in bring-
ing any LLS model to a state of widespread acceptance within the composites
community.
While questions remain about the LLS models in general, there are also many
new problems to address in extending the existing LLS models (with their in-
herit assumptions) to other problems of importance in composite applications.
For instance, what is the dependence of the tensile strength on initial fiber dam-
age that might occur upon processing (Groves rt a/., 1994)? How does damage
progress around a preexisting crack in the composite (see Ibnabdeljalil and Curtin,
1997b. for initial progress on this issue)? What is the predicted time- or cycle-
dependent strength degradation due to creep and/or fatigue (see Beyerlein and
Phoenix, 1998, for an extension of the three-dimensional HVD models to matrix
creep)? What is the stress rupture life under conditions where the fiber strengths
degrade in time? How do material strengths degrade when multiple degradation
phenomena (fatigue, fiber degradation) are occurring simultaneously? These are
issues that have been addressed within the GLS model but may show some sig-
nificant changes under LLS; in particular, the composite lifetime will show some
size dependence.
There are also some basic issues within the LLS quasistatic models that war-
rant further study. In particular, how does the connection between LLS and GLS
at some critical size F I ~arise? Can such a relationship be demonstrated fundamen-
tally rather than simply as a correspondence found through analysis of simulation
data? Why is the predicted tensile strength so weakly dependent on the fiber ar-
rangement (Curtin and Takeda, 1998a)? How does the tensile strength prediction
246 W A. Curtin

change for “random” arrangements of fibers (see Foster, 1998, for some prelimi-
nary results)?
For engineering design of composite structures, the micromechanical results
obtained within GLS and LLS models must be useable in larger-scale design
codes based on continuum methods such as finite-element models. It is compu-
tationally impossible to directly simulate large composite structures at the scale
of the individual fibers. Thus, coupled micromechanics/macromechanics methods
must also be developed.
One strategy is simply to use the nonlinear GLS constitutive relation for el-
ements within an finite-element model (although the constitutive relations for
shear deformations and transverse tension must also be incorporated). The fail-
ure strength of such elements must be dealt with very carefully, however. First,
the element strengths must be assigned stochastically based on the actual element
volume according to the LLS size scaling of strength. Second, the element sizes
cannot be made too small. Elements smaller than the critical size nl and length
0.46, are not appropriate since damage in the fibers is closely correlated over
this size scale. Mesh refinement around macroscopic stress concentrators such as
holes and notches must therefore be done carefully. Third, after “failure” of an
element, the propagation of the localized damage must also be incorporated; this
may involve the introduction of interface elements to represent a crack bridged
by broken fibers. In any case, it is simply inappropriate to assume either (i) com-
ponent failure or (ii) element failure, which is insufficient in the treatment of the
very localized stress transfer at the real crack tip in the material.
Another promising strategy is to intimately combine component simulations at
both micromechanical and macromechanical levels. Specifically, each individual
finite element in a macro model could obtain state information from an associated
fiber-level LLS model, although the LLS model size must be tractable. The LLS
model could be invoked when stresses attain a preset minimum to avoid exces-
sive computations for largely undamaged material. Around stress concentrators
or induced damage, the LLS models could faithfully represent the entire element.
Failure of an element would be obtained naturally in the underlying LLS sim-
ulation, along with the formation of a bridged crack or fiber damage zone. The
propagation of this damage zone could then be followed by remeshing to larger
sizes or by appropriate modification of the “applied stress” in the neighboring el-
ement LLS simulations to include the effect of the localized damage zones. Such
a “direct” combined simulation approach might also be useful for the study of
various time-dependent degradation problems as well.
Finally, we return to the very general issue of failure in heterogeneous mate-
rials discussed in Section I. The fundamental progress made in fiber-reinforced
Fiber-Reinforced Composites 247

composites is due to the fact that stresses from fiber breaks are transferred over
the scale of the fibers themselves. In polycrystalline materials or particulate-
reinforced composites, such a “blunting” of the microscopic crack tip does not
exist generally. Therefore, the direct transfer of results from the fiber-reinforced
systems to other materials is not possible, even if the appropriate strength statis-
tics for the polycrystalline grain boundaries or particle/matrix interfaces, etc. are
supplied by some means. Careful analysis of mechanisms to inhibit crack prop-
agation is required in these other systems, and such mechanisms are probably
very material specific. Heterogeneity alone also may not increase toughness. Re-
cent work on the propagation of a large crack through a material with discrete
heterogeneous toughnesses, using the Green’s function technique, indicates little
toughness improvement over a homogeneous material (Curtin, 1997). Therefore,
crack bridging and microcracking, features that may be an indirect consequence
of the heterogeneity, appear needed to provide enhanced toughness. In systems
for which crack tip stress intensities can be neglected, such as heterogeneous ma-
terials with very weak interfaces failing in shear (Lawn et ul., 1994). materials
with a very broad distribution of residual stress (Padture et ul., 1991), or systems
with intrinsic mechanical decoupling at some length scale (possibly earthquake
fault systems, see Ben-Zion and Rice, 1993), the damage evolution may be ap-
proachable using models similar to the LLS models. Interpretation of results must
be approached carefully, though, since Jagota and Bennison ( 1995) have shown
that unphysical or anomalous results, such as stress distributions dependent on
the mesh shape, can arise in discrete models, especially when treating fracture.
Nonetheless, several important concepts developed within the LLS model (see
also Harlow and Phoenix, 1981) for fiber composites will be important in other
heterogeneous systems. Specifically, the concept of weak-link scaling coupled
with the existence of a critical damage size that controls the macroscopic failure
may be a universal feature of failure in many heterogeneous materials. The re-
sulting size scaling of the strength predicted by eqs. (94) may also be a general
form for the failure strength scaling in heterogeneous materials. Thus, although
considerable effort must be expended to make significant and practical progress
in this broader area of failure in heterogeneous materials, the seeds of some key
concepts exist within the work reported here on fiber composites.

Acknowledgments

I gratefully thank the Air Force Office of Scientific Research (Grant F49620-
95-1-0158) and the National Science Foundation (Grant DMR-942083 1) for fi-
nancial support since 1994, and BP Research for support in prior years. I also
M? A. Curtin
thank my colleagues S. J. Zhou, M. Ibnabdeljalil, H. Scher, and S. L. Phoenix
for their very valuable explicit and implicit contributions to much of the work
reported herein; H. D. Wagner for introducing me to the s.f.c. problem; R. Thom-
son for guidance in the Green's function method; and A. G . Evans for particularly
stimulating suggestions and conversations over the last 10 years.

References

Ahn. B. K., and Curtin, W. A. (1997). Strain and hysteresis by stochastic inatrix cracking in ceramic
matrix composiles. J. Mrch. Phys. Solid\ 45. 177-209.
Batdorf, S. B. (1982). Tensile strength of unidirectional reinforced composites. I. J. Reinforced Plrrstic
Cor,rpositc,s 1, 153- 164.
Baldorf. S. B.. and Ghaffrlrian. R. (1982).Tensile strength of unidirectional reinforced composites. 11.
J. Keir!forced Plrrsric Coniposi/es 1. 165-1 76.
Ben-Zion, Y., and Rice. J. R . ( 1993). Earthquake failure sequences along a cellular fault zone in a 3d
elastic solid containing asperity and non-asperity regions. J. Genphys. Re.r. 98, I4 109.
Beyerle, D. S . . Spearing, S. M.. Zok. F. W., and Evans. A. G. (1992). Damage. degradation. and failure
in a unidirectional ceramic-matrix composite. J. Anirr: Crmrir. Soc. 75, 271 9-2725; correction
( 19Y3). J. Amrr: Cermi. So<. 76, 560.
Beyerlein. I. J.. and Phoenix, S. L. (19963). Stress concentralions around multiple tiber breaks in an
elastic matrix with local yielding or debonding using quadratic influence superposition. .I Mrclr.
.
Phy.s. Solid7 44, 1997-2039.
Beyerlein. I. J., and Phoenix. S. L. (1996b). Comparison of shear-lag theory and continuum fracture
mechanics for modeling fiber and matrix stresses i n an elasticcracked composite lamina. Irrrerncit.
J. Solids Stnicturus 33. 2543-2574.
Beycrlein. 1. J., and Phoenix. S. L. (1997a). Statistics of fracture for an elastic notched composite lam-
ina containing Weibull fibers. I. Features from Monie-Carlo simulation. Dzgrg. Frcrcrirre Mech.
57.241-265.
Beyerlein. I. J.. and Phoenix, S. L. (1997b). Statistics of fracture for an elastic notched composite
lamina containing Weibull libcrs. I. Probability models of crack growth. Erigrg. Frtrcture Mech.
57, 267-299.
Beyerlein. 1. J . . Phoenix, S. L.. and Ra.1, R. (1998). Time evolution of stress redistribution around
multiplc fiber breaks i n a composite wilh viscous and viscoelastic matrices. hternnr. J. Solirls
Structirres. To appear.
Caliskan. A. G. (1996).Microniechanics-based approach to predict strength and stiffness of composite
materials. M.S. thcsia. Virginia Polytechnic Institute and State University, Blacksburg, VA.
Cao. H. C.. Sbaizero. 0..Ruhle, M., Evans. A. G.. Marshall. D. B.. and Brennan, J. J. (1990). Effect
0 1 intcrfaces on the propeltie5 of fiber-reinforced ceramics. .J. Arne,: Cemni. Soc. 73, 1691-1699.
Cox. B. N.. Carter. W. C.. and Fleck. N. A. (1994). A binary model of tcxtile composites. I. Formula-
tion. Acrci Mefull. Mote,: 42. 3463-3479.
Curtin. W. A. (1991a). Exact theory of fibrc fragmentation in a single-filament composite. J. Mute,:
Sci. 26. 5239-5253.
Curtin, W. A. (1991b). Theory of mechanical propertics of ceramic-matrix composites. J. Ames Ce-
r m i . Soc. 74. 2837-2845.
Cuitin. W. A. (19931). Fiber pullout and strain localiration i n ceramic matrix composites. J. Mech.
Phyr. Solids 41, 35.53.
Fiber-Roinjorred Cornyosites 249

Cui-tin. W. A . ( I993h). The "tough" to hrittlc tKuiaitioii in brittle matrix coniporitc\. J. Mrch. Plivs.
S(J/lt/.Y41, 2 17-245.
Curtin. W. A. i 1993~).Ultimate strength?. of fibre-reinforced ceraniics and iiietals. C'orripsirc.s 24,
'2-102.
Curtin. W. A . (1997). Toughening in disonlcrcd hrittls iii~itei-i~ils. Nix\. Kei: H 55. I 1270-1 1276.
Curtin. W. A. (199x1. Size scnling o f btrength i n Iieterogeneous ni;iterials. PI1j.c. KO\: h r r . 80. 1445-
1448.
Cui-tin. W. A,. iiiid Zhou. S. J . ( 19%). Iiitlueiice (11' proce5\ing dnmage on perforiiiance of liber-
reinfoi-ced coiiip~sites.J. M r d i . Phi,\. .Si~/i(/\13.343-363.
Curtin. W. A,. aiitl Takeda. N.i I99Xa).Teiisile strength of tiller-rciiili,i-cctl coiiipo\itch. 11. Model and
effect\ of local geometry. Couipo,sir~~\. To appear.
Curtin. W. A,. and Tiikcda. N. t 1998h).Tcn,ilc strength of fiber-reinforced composites. 11. Application
to polyiiier iiiatris cnnipnsites. Coirip~.\irc~.\. To appear.
Cui-tin. W. A.. Ahn. B. K.. and TaLeda. N.( 199x1. Modeling brittle and tough stress-sti-aiii behavior
iii unidircctionnl cci-amic iiiiitrix coiiipo\itc\. ,4t,rrr Morcv: 46. 34i)9-3420.
Donirls. H. E. (1945). The w t i w c a l theory of the \trength 0 1 huiirllea o f threads. PUJC.K o v . & I ( , .
L ~ i i h ScJr:
i A 183. 405435.
Du. 7,.%.. and McMeeking. R. M . (1093).Control ol'strcngtli aniatropy of metal matrix fiber compos-
hlrrri,r: I k . \ i , q i i 1. 243-264.
ites. J. ~(Jrrr/'riIrr-Ail/(,(/
Du. Z. Z.. and McMeeking. R. M. ( l995j Creep iiiodel\ for metal iiiiitrix conipositcs \bith long hi-ittle
tiherc. J. McL,/I.P/ry.s. Solids 43. 701 -726.
Duva. J. M., Curtin. W. A , . and Wadley. H. N.G. (19951.Aii ultiniatc teiirile \trength dependence on
prncessing for c o n d i d a ~ c diiictiil iiiatrix compo\ites. clc,r(i Mcmll. Mtrr
Elrcy. D.. Duva. I . M.. and Wadley. H. N.G. ( l W 4 l . LJltimatc teii\ile stre
ships for iiictal niiiti-ix coiiipnsites. h r . Ciir!\.'K r ( . c w A(liwric.cc iri Tir(rriiw1i M w i \ C'(mpo,\irc\
(F. H. !'roes and J . Storer. eds.). TMS. Wari-endale. PA.
Evans. A. G.. Zok. F. W.. and McMeeking. R. M ( 19%). Fatigue ol'cci-miic matrix ctmiposites. i \ c ~ t r
Muttill. Mtrtur: 43. 859-875.
Faheny. B.. and Curtin. W. A. i 1996). Dama~c-~nlianc~.d creep arid rupture iii fiber-reinforced c o n -
pwites. A m Mrrrcv: 44. 3439-34.5 I
Foster. G. C. ( 1998). Tensile and flexure strciigtli 0 1 lihrr-reinftmcd composites: direct ~ i u i i i e r i c ~ i l
simulation and analytic nietliodq. M.S. thesib. Vii-ginia Polytechnic Institute and State University.
BlncL.;burg. VA.
Foster. G. C.. IhiiahdeI,jalil. M . . and Curtin. W. A . ( 1998). Tensile strength of ti ti i i i i i i i n iiiatrix compos-
ites: direct numei-ical simulation\ and analytic models. Iritr~riiot.J. Solids Srrlrmrc.r. 35. 2523-
2536.
Fraser. W. A,. Anckci-. F. H.,DiBenecletto. A . T.. and Elhirli. B. i 1983). Eialuation of surlace treat-
ment\ for fibers iii compo\itc ni;iteriaI\. Po/wier- Corirpi'.\irr.\ 4. 238-249.
Fuhuda. H.. and Miyamwa. T. [ lY94). Micromechanical :iliprnacli to tlie tenbile \trength of unidirec-
tional composites. I. Monotilament strength and liher/m;ttrix interaction. A h ! ~ ~ I / ? J / J ( J \ ~Mrrrrr: IC
4. 101-1 10.
Grove\. J. F,. Eliey. D. M., and W;idlry. El. I. fracture during tlie coiisnlitlation metal

matrix composite\. A m Mrroll. M"r(


Gucc~;D. E.. and Gurlantl. J . (1962). Comparisoii o f the statistics of t w o fracture moder. J. Mrc.h.
PhY.\. & J / i t / \ 10. 365-373.
Gulino. R.. and Phoenix. S. L. i 199 I ). Weibull strength \tiiti&b for graphite tihers tncaaured I'rum tlie
breah progression in ii model graphite/glas/cpoxy microcomporite. J. Mrircv: &I. 26. 3 107-3 1 I X .
Gundel, D. B.. and Wauner. F. E. i 1997). Experimental and theoretical a i s e ~ \ m e n tof the longitu-
iiiiil ten4le \treiigth of unidircctioiial SiC-liIierltitanium-matrix compohites. Coiiip~.s;roSci. f i ~ h -
IICJ/. 57. 47 1 4 8 I .
W A. Curtin

Harlow, D. G., and Phoenix. S. L. (1981). Probability distributions for the strength of composite ma-
terials 11. A convergent sequence of tight bounds. lnrernat J. Fracture 17,601-629.
He, M. Y., Evans, A. G., and Curtin, W. A. (1992). The ultimate tensile strength of metal and ceramic-
matrix composites. Acru Metall. Muter: 41. 871-878.
He, M. Y., Wu. B. X.. Evans, A. G.. and Hutchinson, J. W. (1994a). Inelastic strains due to matrix
cracking in unidirectional fiber-reinforced composites. Mech. Mate,: 18, 2 13-229.
He, M. Y.,Evans, A. G., and Hutchinson, J . W. (1994b). Crack deflection at an interface between
dissimilar elastic materials: role of residual stresses. Internat. J. Solids Strucrures 31, 3443-3455.
Hedgepeth. J. M. (1961). Stress concentrations in filamentary structures. NASA Technical report
D-882.
Hedgepeth. J . M.. and van Dyke, P. J. (1967). Local stress concentrations in imperfect filamentary
composite materials. J. Composite Mate% I, 294-309.
Henstenburg. R. B., and Phoenix, S. L. (1989). Interfacial shear strength studies using the single
filament composite test. 11. A probability model and Monte Carlo simulation. Polymer Compos-
ites 10, 389408.
Heredia, F. E., Spearing. S. M.. Evans, A. G., Mosher, P., and Curtin, W. A. (1992). Mechanical prop-
erties of continuous-fiber-reinforced carbon matrix composites and relationships to constituent
properties. J. Amer: Crrclni. Soc. 75, 3017-3025.
Heredia. F. E., Evans, A. G.. and Andersson, C. A. (1995). The tensile and shear properties of con-
tinuous fiber-reinforced SiC/A1203 composites processed by melt oxidation. J. Amer: Ceram.
Soc. 78, 2790-2800.
Hikami. F.. and Chou, T.-W. (1989). Explicit crack problem solutions of unidirectional composites:
elastic stress concentrations. AIAA J., March, 499-505.
Hsieh. C., and Thomson, R. ( 1973). Lattice theory of fracture and crack creep. J. Appl. P h p . 44,
205 1-2063.
Hui. C.-Y.. Phoenix, S. L., Ihnabdeljalil, M., and Smith, R. L. (1995). An exact closed form solution
for fragmentation of Weibull fibers in a single filament composite with applications to fiber-
reinforced ceramics. J. Mech. Phys. Solids 43, 155 1-1585.
Hui. C. Y.. Phoenix, S. L., and Kogan, L. (1996). Analysis of the fragmentation in the single filament
composite: roles of fiber strength distributions and exclusion zone models. J. Mech. Phys. Solids
44, 1715-1737.
Ibnabdeljalil, M.. and Phoenix, S. L. (199%). Scalings in the statistical failure of brittle matrix com-
posites with discontinuous fibers. I. Analysis and Monte Carlo simulations. Acra Metall. Mater:
43.2975-2983.
Ibnabdeljalil, M., and Phoenix, S. L. (1995b). Creep rupture of brittle matrix composites reinforced
with time-dependent fibers: scalings and Monte Carlo simulations. J. Mech. Phys. Solids 43,
897-93 I .
Ihnabdeljalil. M., and Curtin, W. A. (l997a). Strength and reliability of fiber-reinforced composites:
localized load sharing and associated size effects. Internat. J. Solids Srructures 34, 2649-2668.
Ibnabdeljalil, M., and Curtin, W. A. (1997b). Strength and reliability of notched fiber-reinforced com-
posites. Actti Mate,: 45. 364 1-3652.
Iyengar, N., and Curtin, W. A. (1997a). Time-dependent failure i n fiber-reinforced composites by fiber
degradation. Acta Murer: 45, 1489-1502.
Iyengar, N.. and Curtin, W. A. ( l997b). Time-dependent failure in fiber-reinforced composites by
matrix and interface shear creep. Acta Muter: 45, 3419-3429.
Iyengar. N.. and Curtin, W. A. (1997~).Time-dependent failure in ceramic composites by fiber degra-
dation and interface creep. Proc. F i f h Internat. Conf: on Brittle Matrix Composites, pp. 497-506.
Woodhead Publishing, Warsaw.
Jagota. A,. and Bennison, S. J . (1995). Element breaking rules in computational models for brittle
fracture. Mod. Simid. Mute,: Sci. Engrg. 3, 485-501.
Fiber-Reinforced Composites 25 1

Jansson. S.. and Kedward. K. (1996). Ultimate tensile strength of composites exhibiting fihcr frag-
mentation. Coniposite Sci. Techno/. 56. 3 1-35.
Jansson. S.. and Leckie. F. A. (1992). The mechanics of failure i n silicon carbide fiber-reinforced
glass-matrix composites. A m Metrill. Muter: 40. 2967-2978.
Jarmon. D. C.. and Prewo. K. (1986). Final report on Contract NOOO14-81-C-0571. Office of Naval
Research.
Kelly. A,. and Tyson. W. R. ( 1965). J. Mecti. Phys. Solids 13. 329.
Lawn. B. R.. Padture, N. P.. Cai. H.. and Guiherteau. F. (1994). Making ceramics "ductile." Science
263, 11 1 4 - 1 116.
Leath. P. L.. and Duxbury. P. M. ( 1 994). Fracture of heterogeneous materials with continuous distri-
butions of local breaking strengths. Phys. Reit B 49. 14905-14917.
MacKay. R.. Draper, S. L.. Ritter, A . M., and Siemers. P. A. (1994). A comparison of the mechanical
properties and microstructures of intermetallic matrix composites fabricated by two different
methods. Mernll. Morel: Trnns. A 25A. 1443-1455.
Madhukar, M. S.. and Drzal, L. T. (1991). Fiber-matrix adhesion and its effect on composite mechan-
ical properties. 11. Longitudinal and transverse tensile and flexural behavior in graphitekpoxy
composites. J. Composite Mutel: 25.958-99 1.
Majumdar. B. S. (1996). Interfaces in metal matrix composites. In Titunium Murrix Coniposite.s (Mall
and Nicholas. eds.). Technomic Publications.
Majumdar, B. S.. and Miracle. D. B. (1996a). Interface measurenients and applications in fiber-
reinforced MMCs. J. Key. Engrg. Muterids 116117. 153-172.
Majumdar, B. S., Matikas. T., and Karpur. P. (1996h). Fiber fragmentation in single and multi-fiher
composites. Proc. ICCE-3. New Orleans.
Mason, D. D., Hui. C. Y., and Phoenix, S. L. (1992). Stress profiles around a fiber break in a composite
with a nonlinear, power law creeping matrix. Internut. J. Solitls Structures 29. 2829-2854.
McNulty. J., and Zok, E W. ( 1997). Application of weakest-link fracture statistics to fiber-reinforced
ceramic-matrix composites. J. Amer: Cercrm. Soc. 80, 1535-1 543.
Nedele, M. R.. and Wisnom. M. R. (1994). Three-dimensional finite element analysis of the stress
concentration at a single fibre break. Conipyire Sci. Teckno/. 51, 517-524.
Netravali, A. N.. Henstenburg. R. B.. Phoenix. S. L.. and Schwartz, P. (1989). Interfacial shear strength
studies using the single-filament-composite test. 1. Experiments on graphite fibers in epoxy. Polv-
rner Cotnposites 10, 226-24 1
Neumeister. J. M. (1993). A constitutive law for continuous fiber reinforced brittle matrix compositea
with fiher fragmentation and stress recovery. J. Mech. Plrys. Solids 41, 1383-1404.
Ochiai. S., Schulte. K.. and Peters, P. W. M. (1991). Strain concentration factors for fibers and matrix
in unidirectional composites. Cornposire Sci. Techno/. 41, 237-256.
Ohno. N., Kawabe, H.. and Miyake. T. (1997). Effcct of matrix viscosity on stress distribution in
broken fibers in unidirectional composites: analysis based on energy balance. P roc. Inrernrrt.
Cot$ Mech. Muter. Tokyo.
Ohno. N., Kawabe. H.. and Miyake. T. ( 1998).Stress relaxation in broken fibers in unidirectional com-
posites: modeling and applications to creep rupture analysis. J. Mecli. Phys. Solids. Submitted for
pu bl ication.
Padture. N. P.. Bennison, S. J.. Runyan. J . L.. Rodel, J., Chan. H.. and Lawn. B. R. (1991). Flaw
tolerant A1203-AI2Ti05 composites. Cerom. Trms. 19. 715-72 I .
Parthasarathy, T. A., and Kerans, R. J. (1997). Predicted effects of interfacial roughness on thc behavior
of selected ceramic composites. J. Amel: Crrnm. Soc. 80. 2043-2055.
Phoenix. S. L.. and Raj, R. (1992).Scalings in fracture probabilities fora brittle matrix fiber composite.
Acto Met~dl.Mute:: 40, 2813-2828.
Phoenix, S. L.. Ibnahde1,jalil. M.. and Hui. C.-Y. (1997). Size effects in the distribution for strength of
brittle matrix fibrous composites. Inrernnt. J . solid.^ Structures 34, 545-568.
W A. Curtin

Pitt, R. E.. and Phoenix, S. L. ( 1983). Probability distributions for the strength of composite materials.
1V. Lmxlizcd load sharing with tapering. Intrrritrt .J. Frtrctrrrr 21, 243-276.
Prcwo. K. M. ( 1986). Tension and flexural strength of silicon carbide fibre-reinforced glass ceramics.
.I. Matcr: sci. 21, 359(3-3600.
Ramainurty, U.. Zok. F. W.. Leckie. F. A., and Devc. H. E. (1997). Strength variability in alumina
fibei--reinforced aluminum matrix composites. A m r MuteI: 45. 4 6 0 3 4 6 13.
Rao, V.. and Drzal. L. T. ( I99 I).Thc depcndencc of interfacial shear strength on matrix and interphase
properties. Polyrwrr Coriiposi/e.s12, 48-56.
Kosen. B. W. (1964).Tcnsilc failure oftibrow composites. AIAA J. 2, 1985-1991.
Rouby, D.. and Rcynaud. P. (1993). Fatigue behaviour relatcd to interface modification during load
cycling in ceramic-matrix tibre composites. Composite Sci. Tech~id.,109-1 18.
Sastry, A. M.. and Phoenix, S. L. (1993). Load redistribution ncar non-aligned fibre breaks i n a two-
dimcnsional unidirectional composite using br intiuencc supcrposition. J. Murpr: Sci. Lett. 12.
I 596- 1599.
Sastry. A. M.. and Phoenix. S . L. (1995). Shielding and magnification of loads in elastic. unidirectional
coinposites. SAMPE J. 30, 61-67.
Schwictcrt. H. R.. and Steif. P. S. ( 1990).A theory for thc ultimate strength of brittle matrix compos-
ites. J. Mrch. P l r y . Solids 38. 325-343.
Scop. P. M.. and Argon, A. ( 1967a). Statistical theory of strength of laminated composites. J. Coni-
posite Muter: 1. 92-99.
Scop. P. M., and Argon. A. (1967b). Statistical theory of strength of laminated composites. 11. J. Coni-
/JO.Si/eMtI/rK 3. 3 0 4 7 .
Smith, R. L.. and Phoenix, S. L. ( 1981j. Asymptotic dihbutions for the failure of fibrous matcrials
under series-parallel structure and cclual load-sharing. J. A ~ J J M
/ . d i . 103. 75-82.
Smith. R. L., Phoenix. S. L.. Grcenficld. M. R., Henstenburg. R. B., and Pitt. R. E. (1983). Lower-
tail approximations for the probability of failure of three-dimensional fibrous composites with
hexagonal geometry. Proc. Roy. Soc. Lorirlorr Srr: A 388. 353-39 I .
Stawovy. R. H.. Kainpc. S. L., and Curtin. W. A. (1997). Mechanical behavior of glass and Blackglas
ccraniic matrix composites. Acrtr Mriter: 45, 53 17-5325.
Suemasti. H. (1984). Prohabilistic aspects of strength of unidirectional fibre-reinforced composites
with matrix failurc. J. Mmrr: Sci. 19. 574-584.
Sutcu. M. (1989). Weihull statistics applicd to liber failure i n ceramic composites and work o f fracture.
Acto Metoll. 37. 65 1-66 1 .
Tewary. V. K. (1973). Grccn-function method for lattice statics. Ad,: Phy.s. 22. 757-810.
Thoinson, R., Tewary. V. K.. and Masuda-Jindo. K. ( 1987). Theory of chcinically induced kink for-
mation on cracks in silica. 1. 3-d crack Green's functions. J. Marcr: Rrs. 2, 619-630.
Thouless. M.. and Evans. A. G. (1988). Effccts of pulloLit 011 the mechanical properties of ceramic
matrix compositcs. Actn Mrtc7ll. 36. 5 17-52?,
Thouless, M.. Sbaizero. 0..Sigl. L. S.. and Evans, A. G. (1989). Effect of intertice mechanical prop-
erties on pullout in a Sic-fiber-reinforccd lithium aluininosilicate glass matrix. J. Airier: Crmni.
Soc 72, 525-532.
Tsuda. H., Takahashi. J . . Kcmniochi. K., and Hayashi. K. (1996). Fracture process of silicon carbide
fiber-reinforced glasses. ./. Anwr: Crrtori. Soc. 79. 2293-2299.
Van der Heuvcl, P. W. J.. Hogeweg, B., and Peijs. T. (1997). An experimental and numerical in-
vestigation into the single-Abcr fragmentation test: stress transfer by a locally yielding matrix.
Coiirpo,sitrs Prirr A 28A. 237-249.
Wagner. H. D., and Eitan, A. (1990). Interpretation of fragmentation phenomenon in single-filament
composite experiments. Appl. Plrys. /A/. 56, 1965- 1967.
Weher. C. H.. Du, Z. Z., and Zok, F. W. (1996). High temperature deformation and fracture o f a fiber
reinforced titanium matrix composite. Acrtr MNIPI:44, 683-695.
Fiber-Reinforced Composites 253

Weibull. W. (1952). A statistical distribution function of widc applicability. ASME J.. 293-297.
Weitsman. Y.. and Zhu. H. (1993). Multi-fracture of ceramic compositcs. J. Much. P/i\,.s.%/ids 41.
351-388.
Widoni. B. (1966). Random sequential addition of hard sphercs to a voluiiie. J. Clirni. Ph,.v. 44, 3888-
3893.
Wimolkiatisak. A. S., and Bell. J . P.(1989). Interfacial shear strength and failure modes of interphase-
modified graphite+poxy composites. Po/ynrer Corizposirrs 10. 162-1 72.
Wisnoni. M. R. (199la). Relationship bctwcen qtrength variability and size effect in unidirectional
carbon fibrekpoxy. Corirpmitrv 22. 47-52.
Wisnom. M. R. (1991b). The effect of specimen sizc on the bending strength 0 1 unidircctional carbon
tibre-epoxy. Contpsite Sfrrrcrrtrrs 18. 47-63.
Xu. J..Cox. B. N.. McGlockton. M. A,. and Carter, W. C. (1995).A binary model oftextilr composites.
11. The elastic rcgime. Actcr Mrtrr//. Mrifcr: 43. 351 1-3524.
Zhou. S. J., and Curtin. W. A. (1995).Failurc in fihcr composites: a lattice Grccn function model. Acto
Mum//. Mtrtur: 43. 3093-3 103.
Zweben, C. (I968). Tensile failure analysis of librous compositcs. AIAA J. 6. 2325-233 I .
Zweben, C.. and Rosen, B. W. ( 1970). A statistical theory of strength with application to composite
materials. ./. Mech. Phys. Solidc 18. 189-206.

You might also like