Computational Materials Science 64 (2012) 106–111

Contents lists available at SciVerse ScienceDirect

Computational Materials Science

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

Establishing size effects in discontinuous fibre composites using 2D finite

element analysis
C. Qian, L.T. Harper ⇑, T.A. Turner, S. Li, N.A. Warrior
Division of Mechanics, Materials and Structures, The University of Nottingham, University Park, Nottingham NG7 2RD, United Kingdom

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

Article history: Size effects are observed within discontinuous fibre composites, such that the material properties change
Received 2 November 2011 with the specimen volume. Representative volume elements (RVEs) are commonly used to simulate ran-
Received in revised form 21 May 2012 dom fibre architectures for finite element analysis of discontinuous fibre composites. A series of simpli-
Accepted 31 May 2012
fied 2D RVE models have been created and studied in this paper, in order to determine the relationship
Available online 27 June 2012
between the critical RVE size and fibre length and volume fraction. All models are subjected to periodic
boundary conditions, but average properties are extracted from an inner region offset from the model
Keywords:
boundary by a distance equivalent to two fibre lengths. According to Saint–Venant’s principle, this offset
Carbon fibres
Discontinues fibre composites
removes the uncertainty associated with the approximate boundary conditions. A statistical stopping
Finite element analysis criterion has been adopted to determine the number of realisations required to achieve a representative
RVE set of elastic properties for each fibre architecture. The critical RVE side length is shown to be approxi-
mately four times the fibre length when considering convergence of the tensile and shear stiffnesses
for the range of fibre lengths and volume fractions studied.
Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction volume fraction is reached. The range of achievable fibre volume

Developments in high performance computing have seen an in-
crease in the use of numerical techniques such as Finite Element
Analysis (FEA), to model the response of randomly distributed fibre
composites. This approach is well documented for studying the
transverse response of randomly distributed filaments [1,2] or
spherical inclusions [3] at the micro-scale, but is not so common
for modelling 3D randomly packed fibres at the meso-scale, such
as those found in advanced sheet moulding compounds (SMCs)
[4] and directed fibre preforms (DFPs) [5]. These architectures
are far more challenging to model with many more independent
variables to consider, particularly for high fibre volume fractions
and large aspect ratio fibres [6].
The behaviour of heterogeneous materials is often described
using the concept of a representative volume element (RVE). Accord-
ing to Hill [7], the RVE should be a volume of heterogeneous material
that is sufficiently large to be statistically representative, ensuring a
sample is taken of all micro-structural heterogeneities that occur in
the composite. Random sequential adsorption (RSA) is one of the
commonly used methods for producing numerical RVEs. For RSA, fi-
bres are added to the RVE consecutively and checks are performed to
avoid intersecting fibres. Violating fibres are removed and then
redeposited on the next run of the algorithm until the target fibre
⇑ Corresponding author. Tel.: +44 (0) 115 951 3823; fax: +44 (0) 115 951 3800.
E-mail address: lee.harper@nottingham.ac.uk (L.T. Harper).
fractions is limited due to ‘fibre jamming’ however, which is caused
by the lack of suitable space for accommodating additional incom-
pressible fibres. It is also difficult to define the boundary conditions
for the discontinuous fibre composites, due to the randomness in
fibre architecture. Conventional periodical boundary conditions
cannot be directly applied to the RVE like they are for repeating unit
cell models.
Hill’s [7] conditions is commonly used in the literature as a
stress/strain based criterion to determine critical RVE size, assum-
ing that the effective properties of a heterogeneous elastic material
are energetically and mechanically equivalent. The kinetic vari-
ables (stresses) are compared with the kinematic variables
(strains) on the macroscopic scale, and the RVE is defined as the
minimum material for which the response does not depend on
the volume. A review in [6] indicates that there is no systematic
way to quantify the critical RVE size, but it is important to note
that the geometry of the RVE may be influenced by size effects. Size
effects can be divided into two categories; deterministic size
effects and statistical size effects. It is unlikely that the true
effective mean material properties can be realised if deterministic
size effects are present, as they will always dominate. However,
statistical size effects can be controlled by increasing the number
of realisations to overcome the statistical uncertainty.
This paper presents a novel method for meshing discontinuous
fibre architectures for FEA. Modifications to an RSA model have
been used to eliminate the problem of ‘fibre jamming’ and

0927-0256/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.commatsci.2012.05.067
 C. Qian et al. / Computational Materials Science 64 (2012) 106–111
therefore the limitation of volume fraction. Periodic boundary con-
ditions are applied to the FEA model and effective material proper-
ties are calculated based on Saint–Venant’s principle. Various RVE
sizes are studied in order to determine the critical RVE sizes at
different fibre length and volume fractions. A statistical stopping
criterion is also employed to establish the relationship between
the required number of realisations and the effective RVE size,
enabling CPU effort to be optimised.
2. Generation of mesoscopic RVEs
Planar 2D fibre architectures are generated by a modified ran-
dom sequential adsorption scheme programmed in Visual BasicÒ.
Fibre bundles are deposited over a region of interest (ROI), indi-
cated by the outer square in Fig. 1. Random numbers are generated
for the x–y coordinates of the centre of mass for each fibre, and a
third number is used to determine the fibre orientation. This 2D
model allows intersecting fibres, thus removing the fibre volume
fraction limitation. Moreover, this 2D approach is computationally
inexpensive, therefore is practical for simulating larger specimens.
Fibre architectures are meshed suitably for processing by ABA-
QUS/Standard. The matrix material is modelled using structural
2D, plane stress continuum elements (CPS8), since the thickness
of the RVE is typically smaller than the in-plane dimensions. One
dimensional beam elements (B22) with circular cross-section are
used to represent the fibre bundles, where the diameter is assigned
as a function of the filament count and tow volume fraction (Vtow).
The internal structure of each bundle contains resin, and therefore
the volume of deposited bundles (Vdeposited) is adjusted to satisfy
the target volume fraction of the laminate (Vf):
Vf
V deposited ¼
V tow
Beam elements are constrained to the continuum matrix ele-
ments using the ABAQUS built-in function EMBEDDED ELEMENT.
This technique eliminates the extra degree of freedom on the
embedded elements (fibres) compared with the host elements
(matrix) and provides a multi-point constraint. Consequently reg-
ular shaped meshes can be used to model the matrix material,
without node sharing between beam (fibre) and continuum
(matrix) elements. Following [8], the modelling parameters used
for the epoxy resin are 3350 MPa for tensile modulus and 0.38
for Poisson’s ratio. The fibre modulus is 144000 MPa and Poisson’s
ratio is 0.33. All models consist of 24 K fibre tows with an effective
diameter of 1.4 mm, at 60% tow volume fraction. The composite
RVE is modelled to be 3.5 mm thick in all cases, which is a typical
thickness for components manufactured from the range of fibre
Fig. 1. Schematic of random sequential adsorption model. Outer boundary (blue) depicts the edge of the model. Fibres are cropped and boundary conditions are applied at the
region of interest (Red). Elastic constants are extracted from within the inner RVE boundary (black). Black lines represent carbon bundle centre lines. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version of this article.)
107
bundle aspect ratios considered (length/diameter ranges from
1.8 to 7). The influence of RVE thickness on the RVE size has
not been investigated in the current study.
The influence of varying element mesh density on the stress/
strain distribution within the model has been studied, using the
same length for the beam elements and the side length of the con-
tinuum elements. The mesh density is directly dependent on the
number of embedded beam elements within the model, which is
a function of the fibre volume fraction, coupon thickness and fibre
tow size. The mesh needs to be finer to account for a greater num-
ber of bundle ends and crossover points, as more beam elements
potentially share the same host elements. The error between an
element length of 0.05 mm and 0.0125 mm is less than 1% for all
of the fibre volume fractions (10%, 30%, 50%) studied at a thickness
of 3.5 mm. An element length of 0.05 mm has been used for all
subsequent models.
3. Boundary conditions
Boundary conditions are difficult to derive for random hetero-
geneous fibre composites and also have a large impact on the
RVE generation. It is common to see periodic conditions imposed
[9,10], where the inhomogeneous material is approximated by an
infinitely extended model material with a periodic phase arrange-
ment. However, enforcing periodicity for long slender fibres with
larger aspect ratios influences the local fibre volume fraction distri-
bution around the edge of the model, which can affect the critical
size of the RVE [3].
In the present paper, all models are subjected to periodic
boundary conditions, assuming translational symmetries in the x
and y directions. However, it is incorrect to assume periodicity as
ð1Þ the material is non-repeating and fibres bridging the RVE cause
local discontinuities at the boundary. However, according to
Saint–Venant’s principle [1], the effects of incorrectly prescribed
boundary conditions affect only a limited zone next to the bound-
ary. An approximation to the exact solution can therefore be found
if a sub-domain is extracted from the model at some critical ‘decay
length’ away from the boundary [1], as shown in Fig. 1.
The RVE boundary length was set to at least two times the fibre
length in all cases, as suggested in [11]. Fig. 2 shows an example of
how the predicted in-plane stiffnesses vary with increasing decay
length (‘d’) for a model consisting of 10 mm long fibres, randomly
distributed at a fibre volume fraction of 10% (results are norma-
lised wrt fibre length ‘l’). All three stiffness values converge after
a decay length of two times the fibre length (20 mm). All three
curves plateau at d/l = 2 and the error between this point and d/
l = 5 is less than 1% for all three stiffness components. Von Mises
108 C. Qian et al. / Computational Materials Science 64 (2012) 106–111
Fig. 2. Effect of decay length on the calculated in-plane stiffnesses for an RVE
containing 10 mm long fibres at 10% vf. (decay length is plotted as a function of
fibre length).
(a)
(b)
(c)
Fig. 3. Von Mises stress (MPa) contours extracted from the inner RVE boundary for
three different decay length/fibre length ratios: (a) 0.5, (b) 2, and (c) 5. RVE
consisted of 10 mm long fibres at 10% vf. Loading was in shear in the 1–2 plane.
stress contours plotted in Fig. 3 show that a decay length ratio of d/
l = 2 also gives a good approximation of the stress field within the
inner RVE boundary. All three plots share similar internal features,
but the magnitude of the stresses around the boundary of d/l = 0.5
are generally lower than d/l = 2 and d/l = 5. The stress at an arbi-
trary nodal point (node 519581) has been compared for the decay
lengths and it is seen that the error is less than 2% between d/l = 2
and d/l = 5 (110.9 MPa for d/l = 0.5, 125.3 MPa for d/l = 2 and 123.3
for d/l = 5). The displaced shape for d/l = 2 is also very similar to d/
l = 5. The edges are not straight and therefore conventional dis-
placement boundary conditions for this RVE model would indeed
be inappropriate, which further supports the use of Saint–Venant’s
principle for heterogeneous materials in this application.
The critical decay length has been established for a range of
models containing different fibre lengths (2.5 mm, 5 mm, 10 mm),
volume fractions (10%, 30%, 50%) and orientation distributions (ran-
dom, aligned). The error between the elastic properties at d/l = 2
and d/l = 5 is less than 2% in all cases, independent of fibre length
and fibre volume fraction. The critical decay length is therefore
found to be two times the length of the fibres in all cases. This
observation agrees well with findings in [1], which shows that the
critical decay length is two times the length scale of the reinforcing
inclusion. A decay length of d/l = 2 is used for all subsequent
analyses.
4. Determining the critical RVE size
Statistical measures are commonly used to establish the critical
RVE size by determining the condition of statistical homogeneity.
Convergence of the effective homogenised properties is often stud-
ied to determine the characteristic RVE size. It is difficult to estab-
lish exactly when the effective size has been reached, and common
statistical methods are reviewed in [12]. Kanit et al. [13] relate RVE
size to number of realisations and estimation accuracy for linear-
elastic properties. In their approach a relative error is calculated
for the effective properties of models of increasing volume, which
is reverse engineered to determine the final RVE size for a given
precision. The accuracy of this model strongly depends on the
number of realisations performed and can require a large number
of FE computations (typically 2500). The use of smaller RVE sizes
has to be compensated by a larger number of realisations.
Gitman et al. [3] purposed a stopping criterion for determining
non-linear material properties, using a statistical analysis on the
average stress from numerical computations. Gitman’s criterion
is less time consuming than Kanit’s, because only five repeats are
used for each volume size. However, such a small number of real-
isations cannot be sufficient unless the volume of material is large,
as reported in [13].
Considering the higher accuracy, the stopping criterion devel-
oped by Kanit et al. [13] has been adopted in the current paper
to determine how many realisations are required to achieve a rep-
resentative volume, based on the stabilisation of the estimated
mean and variance. The size of the RVE can be directly related to
the mean value of the effective property Z, of the domain size V,
for different independent realisations n. The relative error can be
calculated as follows:
eabs 2DZ ðVÞ
erel ¼ ¼ pffiffiffi ð2Þ
Z Z n
where Dz(V) is the standard deviation of the effective property Z for
a domain size V. The relative error has been calculated for all
scenarios and has been presented as a function of number of
realisations. A relative error threshold of 5% has been used as the
stopping criterion for all subsequent simulations, which therefore
represents a confidence of 95% in the mean value of the effective
property Z.
 C. Qian et al. / Computational Materials Science 64 (2012) 106–111
5. Results
Effective material properties have been calculated in order to
determine the critical RVE size. In-plane tensile stiffness values
in the longitudinal direction (E1) are presented in Fig. 4 as a func-
tion of edge length normalised with respect to fibre length (L/a).
The tensile stiffness values in the transverse direction (E2) are also
calculated but the graph is not included, as the error between E1
and E2 is less than 5%, which is smaller than the relative error va-
lue selected for the stopping criterion.
Fig. 4 shows an increase in the average tensile stiffness for
increasing fibre length and fibre volume fraction for any given
RVE size. The average tensile stiffness decreases for increasing
RVE size, before reaching a plateau value, for all of the material
architectures in Fig. 4. The average stiffnesses for materials with
a volume fraction of 30% are 4683 MPa, 7880 MPa and
12227 MPa for fibre lengths of 2.5 mm, 5 mm and 10 mm respec-
tively. The highest average stiffness is for the 5 mm fibre length
with a 50% volume fraction; 13576 MPa. The convergence point
of the effective stiffness occurs when the RVE edge length is four
times the fibre length in all cases (L/a = 4). There is less than 1% er-
ror between L/a = 4 and the next available data point for each
material.
Similarly, for the shear stiffness G12 (Fig. 5), convergence also
occurs at L/a = 4 and the effective shear stiffness is generally higher
for smaller models below this L/a threshold, except for 10 mm fibre
length. The average shear stiffness for materials with a volume
fraction of 30% is 1719 MPa, 2828 MPa and 4431 MPa for fibre
Fig. 4. In-plane tensile modulus E1 as a function of RVE edge length normalised by
fibre length. All x-axis values are integers; however some data points have been
offset ±0.05 along the x-axis to provide clarity.
Fig. 5. In-plane shear modulus G12 as a function of RVE edge length normalised by
fibre length. All x-axis values should be integers; however some data points have
been offset ±0.05 to provide further clarity.
109
Fig. 6. Effect of increasing RVE size on Poisson’s ratio. Example curve for the 10 mm
fibre length/30% vf case. Linear equation for trend line is included. v12 data series
has been offset by 0.02 on the x-axis from v21 to provide clarity between standard
deviation bars.
lengths of 2.5 mm, 5 mm and 10 mm respectively. The highest
average stiffness is for the 5 mm fibre length with a 50% vf;
4910 MPa. However, the average Poisson’s ratio (Fig. 6) is indepen-
dent of increasing RVE size. The average error between v12 and
v21 is less than 3% in all cases; and there is no distinguishable
trend for the magnitude of Poisson’s ratio as a function of fibre
length or fibre volume fraction. This can be attributed to the
dimensionality of the elements used to model the fibre bundles.
The 1D beams do not exhibit any Poisson effects, therefore the
Poisson ratio of the composite is dominated by the elastic proper-
ties of the matrix material. The average Poisson’s ratio for the
2.5 mm, 5 mm and 10 mm materials with 30% vf are approximately
the same; 0.367, 0.370, 0.360 respectively. The average Poisson’s
ratio for the 2.5 mm, 50% vf material is 0.360.
The curves presented in Figs. 4 and 5 can be split into two dis-
tinct zones based on the convergence of the stiffness at L/a = 4. A
representative average of the effective stiffness cannot be achieved
for models where L/a < 4 because deterministic size effects are
present, caused by bridging fibres across the RVE and also the
material scale. The effective stiffness is constant for L/a > 4, but
the variance continues to decrease for larger models. Statistical
size effects dominate the properties of larger models, presenting
a trade-off between the size of the RVE and the number of realisa-
tions required to achieve a representative average. The fundamen-
tal question is whether it is more efficient to test many smaller
models or fewer larger ones, in order to achieve the effective mean
properties at the desired confidence level of 95% (erel = 5%).
The standard deviation of the effective properties is found to
follow a power law relationship
DZ ðVÞ ¼ KLc ð3Þ
where L is the RVE edge length. The parameters in Eq. (3) are sum-
marised in Table 2 for each material. This relationship can be used
to calculate the number of realisations required to achieve a relative
error of 5% for larger RVE sizes. The number of realisations for each
RVE size has been multiplied by the number of degrees of freedom
in each model. This provides a measure of CPU effort required to
achieve a representative set of elastic constants as a function of
RVE size. Fig. 7 shows CPU effort as a function of RVE edge length
for E1. Similar trends were also observed for G12 and v12 and a
summary of the critical RVE sizes are presented in Table 1. Filled
points in Fig. 7 represent the data from the simulations and lines
of best fit have been calculated using the power law relationship
of the standard deviation from Eq. (3) and Table 2. A local minimum
occurs for each material architecture, indicating the optimum RVE
size in terms of CPU effort required to achieve a representative
110 C. Qian et al. / Computational Materials Science 64 (2012) 106–111

Table 1
Critical RVE size as a function of fibre length and fibre volume fraction. Results are presented for two methods (1) convergence of effective properties, and (2) convergence of CPU
effort. L is RVE edge length and a is fibre length.

Architecture Effective properties CPU effort (decrees of Freedom  realisations)

Fibre length Vf E1 G12 v12 E1 G12 v12
(mm) (%) L (mm) L/a L (mm) L/a L (mm) L/a L (mm) L/a L (mm) L/a L (mm) L/a
2.5 30 10 4 10 4 – – 24 9.6 14 5.6 20 8
5 30 20 4 20 4 – – 68 13.6 59 11.8 81 16.2
10 30 40 4 40 4 – – 94 9.4 110 11 120 12
5 30 20 4 20 4 – – 82 16.4 62 12.4 92 18.4

Table 2
Constants for Eq. (3) for coarseness, departure from isotropy and a range of elastic constants as a function of fibre architecture. R2 values indicate the goodness of fit.

Architecture E1 G12 v12

K c R2 K c R2 K c R2
2.5 mm, 30% 2472 0.945 0.936 850 1.127 0.936 0.229 1.110 0963
5 mm, 30% 20006 1.075 0.996 3904 0.935 0.980 0.567 0.945 0994
10 mm, 30% 62277 1.137 0.932 7454 0.909 0.964 1.360 1.058 0972
50 mm, 30% – – – – – – – – –
5 mm, 50% 42355 1.105 0.992 6515 0.962 0.980 0.654 0.955 0995

Fig. 7. Effect of RVE size on CPU effort to determine E1 (expressed in terms of degrees of freedom multiplied by the number of realisations) at a target relative error of 5%.
Data points represent simulation averages and lines are power-law fits.

value for E1. This local minimum occurs when the number of real-
isations required to achieve a relative error of 5% converges to unity.
Whilst the number of realisations has converged by this point, total
CPU effort begins to increase as the number of degrees of freedom
continues to follow a squared relationship. Local minima occur at
L = 24 mm, 68 mm and 94 mm for the 2.5 mm, 5 mm, and 10 mm
fibre length materials with a 30% volume fraction, indicating that
the RVE size is fibre length dependent.
The convergence of the CPU effort appears to be dominated by
the number of realisations, rather than the total number of degrees
of freedom for each RVE size. Convergence typically occurs when
n = 1, which implies that it is more computationally efficient to test
one large model rather than several smaller ones. In practice how-
ever, this is not always feasible depending on the scale of the prob-
lem and the computational resources available.
composites. Saint–Venant’s Principle has been demonstrated for
discontinuous random fibre composites to enable the use of
periodic boundary conditions, and the critical decay length is
two times the fibre length using the current modelling
approach.
Both deterministic and statistic size effects have been found
for effective material properties. For tensile and shear modulus,
convergence starts as the statistical size effects become domi-
nate; however, Poisson’s ratio is independent of the RVE size.
The critical RVE size for the convergence of effective properties
is four times the fibre length using the current modelling
approach.
References

6. Conclusions
A random sequential adsorption model has been used to
generate architectures for meso-scale discontinuous fibre
[1] A. Wongsto, S. Li, Composites Part A: Applied Science and Manufacturing 36 (9)
(2005) 1246–1266.
[2] D. Trias, J. Costa, A. Turon, J.E. Hurtado, Acta Materialia 54 (13) (2006) 3471–
3484.
[3] I.M. Gitman, H. Askes, L.J. Sluys, Engineering Fracture Mechanics 74 (16) (2007)
2518–2534.
 C. Qian et al. / Computational Materials Science 64 (2012) 106–111
[4] P. Feraboli, E. Peitso, F. Deleo, T. Cleveland, P.B. Stickler, Composite, Materials
Technology (2008) 289–299.
[5] L.T. Harper, T.A. Turner, N.A. Warrior, J.S. Dahl, C.D. Rudd, Composites Part A:
Applied Science and Manufacturing 37 (11) (2006) 2136–2147.
[6] Y. Pan, L. Iorga, A. Pelegri, Composites Science and Technology 68 (13) (2008)
2792–2798.
[7] R. Hill, Journal of the Mechanics and Physics of Solids 11 (1963) 357.
[8] R. Luchoo, L.T. Harper, M.D. Bond, N.A. Warrior, A. Dodworth, Rubber and
Composites 39 (2010) 216–231.
[9] N. Pan, Journal of Composite Materials 28 (16) (1994) 1500–1531.
111
[10] S. Kari, Micromechanical Modelling and Numerical Homogenization of Fibre
and Particle Reinforced Composites, Institute of Mechanics Department, Otto-
von-Guericke University Magdeburg, Magdeburg, 2005.
[11] L. Iorga, Y. Pan, A.A. Pelegri, Journal of Mechanics of Materials and Structures 3
(7) (2008) 1279–1298.
[12] C. Pelissou, J. Baccou, Y. Monerie, F. Perales, International Journal of Solids and
Structures 46 (14–15) (2009) 2842–2855.
[13] T. Kanit, S. Forest, I. Galliet, V. Mounoury, D. Jeulin, International Journal of
Solids and Structures 40 (2003) 3647–3679.