Composites Science and Technology 126 (2016) 122e129

Contents lists available at ScienceDirect

Composites Science and Technology

journal homepage: http://www.elsevier.com/locate/compscitech

Fracture toughness of polymeric particle nanocomposites: Evaluation

of models performance using Bayesian method
Khader M. Hamdia c, *, Xiaoying Zhuang e, c, Pengfei He f, Timon Rabczuk a, b, c, d, **
a
Division of Computational Mechanics, Ton Duc Thang University, Ho Chi Minh City, Viet Nam
b
Faculty of Civil Engineering, Ton Duc Thang University, Ho Chi Minh City, Viet Nam
c €t Weimar, 99423 Weimar, Germany
Institute of Structural Mechanics, Bauhaus-Universita
d
School of Civil, Environmental and Architectural Engineering, Korea University, Republic of Korea
e
Department of Geotechnical Engineering, College of Civil Engineering, Tongji University, Shanghai, China
f
School of Aerospace Engineering and Applied Mechanics, Tongji University, Shanghai, China

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

Article history: This study presents a methodology to evaluate the performance of different models used in predicting
Received 27 August 2015 the fracture toughness of polymeric particles nanocomposites. Three analytical models are considered:
Received in revised form the model of Huang and Kinloch, the model of Williams, and the model of Quaresimin et al. The purpose
18 January 2016
behind this study is not to recommend which of the three models to be adopted, but to evaluate their
Accepted 8 February 2016
Available online 15 February 2016
performance with respect to experimental data. The Bayesian method is exploited for this purpose based
on different reference measurements gained from the literature. The models' performance is compared
and evaluated comprehensively accounting for the parameter and model uncertainties. Based on the
Keywords:
Nano particles
approximated optimal parameter sets, the coefficients of variation of the model predictions to the
Fracture toughness measurements are compared for the three models. Finally, the model selection probability is obtained
Modelling with respect to the different reference data.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction have been carried out in order to study the fracture behavior of
polymer/particle nanocomposites ([3e12] among others). On the
Polymeric nanocomposites (PNCs) are commonly formed by an other hand, researchers developed numerical and analytical
epoxy matrix reinforced with a nanosized filler. Due to its inherent methods to get a better understanding of nanocomposite material
characteristic of high crosslink density, an epoxy polymer is known behavior. A close form formula of energy dissipation due to the
to be a relatively brittle material [1]. Nanofillers have shown great interfacial debonding between the particles and matrix was given
improvements in the physical and mechanical properties of epoxy- by Chen et al. [13] considering the effect of particle sizes. Although,
reinforced PNCs. Specifically, they have increased the fracture the increased fracture energy of rubber-toughened epoxy polymers
toughness compared to pristine epoxy. PNCs have numerous ap- was calculated by Huang and Kinloch [14], the model has been
plications in nanotechnology such as: nano-biotechnology, nano- modified for PNCs by Refs. [7,8,10]. The improvement in the fracture
systems, nanoelectronics, and nano-structured materials. Gener- toughness was attributed to two major mechanisms: localized
ally, there are three categories of fillers: nanoparticles, nanoplatelet plastic shear banding and debonding of silica nanoparticles. Further
(layered), and nanofibrous materials. For this scale, the surface area experimental studies also have implied this supposition [15e17].
- to - volume ratio is significantly large. Therefore, the composite According to the assumption of Williams [18], the energy dissipa-
properties are highly modified due to the extreme interfacial area tion is induced by the growth of plastic voids around debonded
between the nanofiller and the matrix [2]. Several experiments particles. The author concluded a large toughness increase for
nanosize particles. Later, his work has been extended to cylindrical
rods and fibres [19,20]. Quaresimin et al. [21] proposed a multiscale
* Corresponding author. approach to predict the overall increase in the fracture toughness
** Corresponding author. Division of Computational Mechanics, Ton Duc Thang taking into account three different damage mechanisms: particle
University, Ho Chi Minh City, Viet Nam
debonding, plastic yielding of nanovoids, and shear banding of the
E-mail addresses: [email protected] (K.M. Hamdia), timon.
[email protected] (T. Rabczuk). polymer. Based on experimental data gathered from the literature,

http://dx.doi.org/10.1016/j.compscitech.2016.02.012
0266-3538/© 2016 Elsevier Ltd. All rights reserved.
 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129 123

a stochastic approach has been presented to predict the fracture to their popularity and their applicability to different experimental
energy of PNCs by Ref. [22]. studies. Moreover, they produce explicit predictions of the
In general, all models inherently underlie an amount of un- enhanced fracture energy of PNCs. Regarding the different theory
certainties which can be related to the model itself and/or its input and mechanism assumed, each of them has its own input param-
parameters. The former might be caused by the simplifications of eters in addition to the joint parameters. Table 1 includes the def-
the physical behavior, while the latter can be related to the number initions of the parameters and their stochastic variation. The
and the stochastic variance of the input parameters. Better pre- uniform distribution was assumed for the parameters uncertainty.
dictions and the subsequent decrease in the model uncertainty are The upper and the lower limits of distributions were mostly pro-
expected by introducing more factors in the model (increasing the posed according to our previous studies [22,28].
model complexity). However, the parameters uncertainties become
more dominant in this case. In light of this, the model with mini-
mum total uncertainty is the most appropriate model, see Fig. 1
[23].
2.1. Huang and Kinloch
In recent years, Bayesian method has been introduced as an
effective tool for evaluating models considering the model and
The model according to Huang and Kinloch [14] was first
parameters uncertainties based on measurements as reference data
developed for the toughening mechanisms of rubber-modified
[24e27].
epoxy polymers and more recently it has been modified for PNCs
This paper is the first attempt to consider the model and pa-
[7,8,10]. The localized plastic shear banding and debonding of
rameters uncertainties in the assessment of the models used for the
nanoparticles which enable plastic void growth of the epoxy matrix
prediction of the fracture energy of PNCs. It aims at presenting a
are the two terms that taking part in the overall enhancement in
methodology to evaluate three different analytical models by using
the fracture toughness of PNCs, while rubber-bridging mechanism
the Bayesian method. In particular, Huang and Kinloch model [14], was disregarded. These two mechanisms are demonstrated in
Williams model [18], and the model according Quaresimin et al.
Fig. 2.
[21] are examined. The purpose of the study is not to give a general The improved fracture energy of PNCs, GIc, is expressed as
recommendation which of the three model to use, but to evaluate
their performance with respect to experimentally tested data se-
GIc ¼ GIm þ DGs þ DGv (1)
ries. The assessment is carried out based on different reference data
(experimental measurements) gathered from the literature [3e12]. where GIm is the fracture energy of the matrix, and DGs and DGv are
Nevertheless, the same methodology can be applied to evaluate the the contribution from the localized shear banding and the plastic
three models based on other measurements. The prior probabilities void growth, respectively.
are first estimated considering the uncertainties in the parameters. The term DGs is given by
Then we find the optimum parameter set which results in best fit of
models prognoses and in consequence the coefficient of variation of

1
the models predictions to the measurements are estimated. Even- DGs ¼ V syc gf F 0 ry (2)
2 f
tually, the model selection probability is calculated.
The remainder of this paper is organized as follows. In Section 2, where Vf is the volume fraction of the nano-filler, gfm is the matrix
the considered models are briefly described. Section 3 presents the shear fracture strain, and syc is the yield stress of the epoxy matrix
method for evaluating the models. Finally, the conclusion of this under compression, which related to the tensile yield stress, sym, by
research is presented in Section 4. Ref. [5].

pffiffiffi !
2. Models for predicting the fracture properties of PNCs
3 þ mm
syc ¼ sym pffiffiffi (3)
Three existing models were chosen to be evaluated; the model 3  mm
of Huang and Kinloch [14], the model of Williams [18], and the
model of Quaresimin et al. [21]. Hereafter, they are abbreviated by mm is a material constant (pressure coefficient).
M1, M2, and M3, respectively. These models have been selected due The parameter F'(ry) is a geometric term given by Ref. [15].
2 !1=3   3 5=2

 4p 8 rn rn rn
F ry ¼ ry 4
0
1 1 
3Vf 5 ry ry ry
3
   
16 rn 7=2 rn 2 165
 2 1 þ (4)
35 ry ry 35

where rn (¼dn/2) is the radius of nanoparticles and ry is the radius of

the plastic zone at the crack tip at fracture in the PNCs

 2
m
ry ¼ 1 þ pmffiffiffi rym Kvm
2
(5)
3
In Eq. (5), rym is radius of the plastic zone of the unmodified
epoxy matrix estimated by Irwin's model [29] and Kvm is the
maximum stress concentration factor of the von Mises stress in the
Fig. 1. Variation in model, parameter, and total uncertainties with respect to the matrix.
number of parameters according to [23]. The term DGv is calculated by
124 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129

Table 1
The definitions of models' parameters.

Parameter Symbol Unit Limits [lower,upper] Modelsa

The volume fraction of the nano-filler Vf % [0.5,10] All

The average diameter of the nano-particles dn nm [10,80] All
The elastic modulus (The Young's modulus) of the matrix Em GPa [2.4,3.6] All
The fracture energy of the matrix GIm J/m2 [40,500] All
The yield strength of the matrix sym MPa [70,120] All
The Poisson's ratio of the matrix nm e [0.33,0.37] All
The shear fracture strain of the matrix gfm e [0.70,0.75] M1,M3
The pressure dependency material constant (pressure coefficient) mm e [0.175,0.225] M1,M3
The matrix maximum stress concentration factor of the von Mises stress Kvm e [2.10,2.25] M1
The average diameter of voids around nanoparticles dv nm [25,120] M1
The percentage of the debonded particles Vdp % [10,18] M1
The interfacial debonding energy Ga J/m2 [0.01,1.5] M2,M3
The shear yielding stress of the matrix tym MPa [40,70] M3
The thickness of the interphase t nm [1,4] M3
The ratio of the shear elastic modulus of the interphase to the shear elastic modulus of the matrix Х e [0.1,2.0] M3
a
M1, M2, and M3 refer to Huang and Kinloch model, Williams model, and Quaresimin et al. model, respectively.

2.2. Williams

The analysis of Williams [18] has referred the energy dissipation

to one basic mechanism; the debonding of nanoparticles which
initiate the plastic void growth. At the first stage the rigid spherical
nanoparticles is bonded to the surrounding matrix, and under
tensile stress, the interfacial stress increases until debonding occur.
This initiates the void growth in the matrix. The critical interfacial
stress (debonding stress), scr, is approximated by
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
4 Em Ga
scr ¼ (9)
1 þ nm rn

where Ga is the interfacial debonding energy, and nm is the Poisson's

ratio of the matrix. If the number of the particles participating in
the process is considered to be more than one, then the fracture
energy of the PNCs is
Fig. 2. Representative diagram for the toughening mechanisms of PNCs according to  
2=3
Huang and Kinloch [14]. GIc ¼ GIm Xt Vf  1:21Vf þ 1 (10)

where Xt is toughening factor that is characterized by a critical

stress ratio factor, x [18].
 
m2  
2 Both factors are given by
DGv ¼ 1  m Vf v  Vfp syc rym Kvm (6)
3 " #
scr 1 þ nm ð1 þ nm Þ2 eðx1Þ 5nm  1
x¼ ; Xt ¼ 
where Vfv and Vfp are the volume fraction of voids and the volume sym 2ð1  nm Þ 2pð1  nm Þ x 2ð1 þ nm Þ
fraction of debonded particles. An expression for (VfvVfp) consid-
ering the average volume of the voids (vv) and the average volume
(11)
of nanoparticles (vp) has been proposed by Refs. [8,10] as

  v 
2.3. Quaresimin et al.
v
Vf v  Vfp ¼  1 Vfp (7)
vp
A multiscale methodology has been adopted by Quaresimin
As reported in the studies of [7,17], not all of the nanoparticles et al. [21] to describe the toughening mechanism of PNCs. The
have been observed to be debonded. Finite element simulations authors have considered the interphase zone surrounding the
suggest that around 14% of the particles are debonded [17]. On this nanoparticle to account for the interactions between the nano-
basis, we include a new factor to quantify the percentage of particles and the matrix. The adjacent polymer chains are disor-
debonded particles (Vdp). It has been assumed to vary from 10% to dered due the addition of the nanofiller, leading to the formation of
18%. The volume fraction of debonded particles, Vfp in Eq. (7), is interphase zones surrounding the nanoparticles with properties
substituted by (Vdp)(Vf) different from that of the bulk matrix. The extent of the impact of
particle/polymer interface is principally influenced by the
  v    manufacturing techniques and the curing processes. The influence
v
Vf v  Vfp ¼  1 Vdp Vf (8) of the interphase was studied experimentally by Refs. [30e33] and
vp
numerically by Refs. [34e37]. Other advanced computational
multiscale methods for crack propagation and material failure were
 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129 125

proposed for instance in Refs. [46-50].

Through its thickness, the interphase layer was assumed to be GIm
GIc ¼   (15)
homogeneous and isotropic [38]. Fig. 3 displays the system 1  Vf Jdp þ Jpy þ JSB
considered at the nanosized scale.
By studying the energy dissipation at the nanoscale, Quaresimin The condition: Vf ðJdp þ Jpy þ JSB Þ < 1, is essential for the
and co-workers indicated that the overall fracture toughness of the applicability of the model in Eq. (15). Other values will produce
nanocomposite is composed of three damaging mechanisms: (i) meaningless results. This limits the applicability of the model in
particle debonding, (ii) plastic yielding of nanovoids, and (iii) shear determining the fracture energy of PNCs reinforced by small frac-
banding of the polymer [38e40]. tions of nanofiller.
The term of toughness improvement due to the debonding of
nanoparticle is expressed by
2.4. Discussion
  !
2Ga 1 þ nc Ec Although, the model of Huang and Kinloch [14] accounts for the
Jdp ¼ (12)
3prn 1  nc s2cr Ch2 main damaging mechanisms, it is based on some simplifying as-
sumptions. The knowledge of the increased volume fraction of
nc and Ec being the Poisson's ratio and the elastic modulus of the voids (VfvVfp) is required to evaluate the energy contribution from
nanocomposite, respectively. In this study, nc was set equal to the void-growth mechanism, DGv (see Eq. (6)). The values for the vol-
matrix Poisson's ratio, nm, while Ec was calculated by the Hashin- ume fraction of voids, Vfv, and the volume fraction of debonded
Shtrikman solution developed in Ref. [41]. The debonding stress, particles, Vfp, can be measured experimentally using for instance
scr, and the reciprocal of the hydrostatic part of the global stress electron micrographs. Instead, based on the expression proposed in
concentration tensor, Ch, depend on the nanoparticle radius, the Refs. [8,10], we introduced a new uncertain parameter to quantify
interphase thickness, and the matrix and interphase elastic prop- the percentage of debonded particles (Vdp). The proposed formula
erties [38]. for (VfvVfp) is shown in Eq. (8).
The part of the fracture toughness enhancement caused by the On the other hand, the model of Williams [18] assumes that the
plastic yielding of nanovoids is [39]. void growth around debonded particles is the only dominant en-
ergy dissipation mechanism. His analysis was based on the energy
    balance concept around a single nanoparticle and assuming the
  s absence of particle-to-particle interaction. In turn, the effect of
3Ch ssym
ya cr 1
4 Ec ð1 þ nc Þð1 þ nm Þ sym a 3 1sym
aggregation was ignored. A similar assumption, i.e. the absence of
Jpy ¼ e
9pCh Em 1  nc scr rn particle-to-particle interaction, was considered in the model of
(13) Quaresimin et al. [21]. However, three damaging mechanisms
through multiscale modelling and the effect of the interphase zone
where sya is the yield stress of the interphase, and a is the external were taken into account.
interphase radius (see Fig. 3). It is widely acknowledged that the nanofillers intend to
The third part, the improvement due to the formation of local- agglomerate in nanocomposites. This may limit the applicability of
ised plastic shear bands, is the models of Williams [18] and Quaresimin et al. [21] to PNCs with
a low volume fraction of fillers. However, since the most important
merit of the PNCs is substantial improvements in the fracture
ISB Ec toughness at low filler content, this is not a short-coming of the
JSB ¼  .pffiffiffi2 G (14)
1  n2c above mentioned models.
4ps2yca 1  mm 3

ISB is referring to the stress concentration around the nano-

particles, syca being the interphase yielding stress under compres- 3. Assessment of PNCs fracture models using Bayesian
sion, and G is quantifying the energy produced at the nanoscale. method
The shear yielding stress of the matrix, tym, in addition to Vf, gfm, a,
and rn are required to calculate G, while ISB is a function of nc, mm, Ch, The Model selection refers to the problem of selecting one
and HvM (the deviatoric component of the global stress concen- model from a list of candidate models based on available data. The
tration tensor) [40]. Bayes' rule of statistics has motivated [42] to develop the Bayesian
Eventually, the total fracture energy of PNCs is calculated by approach for model selection by incorporating the different sources
of uncertainties based on response measurements (reference data),
D. The model selection probability is represented by the conditional
probability of the model Mi given the reference data D. It can be
calculated by

PðDjMi ÞPðMi Þ
PðMi jDÞ ¼ P (16)
i PðDjMi ÞPðMi Þ

where P(Mi) is the prior probability of Mi which is based on the

user's judgment on the initial plausibility of the models. The data-
dependent term P(DjMi) is the evidence of Mi. It defines the prob-
ability that the measurements of reference data D being repre-
sented by the predictions of the model Mi. Making use of the
Fig. 3. 2D representation of the nanoscale system considered according to Quaresimin theorem of total probability, the evidence can be calculated by Ref.
et al. [21]. [42].
126 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129

Z [3e12] have been utilized as reference data. For each, Table 2 shows
PðDjMi Þ ¼ PðDjXi ; Mi ÞPðXi jMi ÞdXi (17) the values of the calculated optimal parameter set.
Interestingly, the incorporation of the parameter Vdp in M1 has
enhanced the model predictions to fit the measurements. By the
where PðDjXi ; Mi Þ is the likelihood function and PðXi jMi Þ is the prior finite-element analysis of [45], the value of the maximum stress
probability of the input parameters. concentration for the von Mises stresses around a void, Kvm, was
The likelihood is the joint conditional probability of the refer- estimated to be 2.22 for a matrix of elastic modulus equal 3.2 GPa
ence data, D, given the input parameters, Xi. It measures how the which agrees well with the optimal values obtained in this study.
model fit the data. A higher likelihood factor corresponds to better The interfacial debonding energy, Ga increases as the diameter of
fit of Mi to D. The prior probability of the input parameters char- the nanofiller increases. Its optimal values were in the range of
acterizes what is known about the parameters before any actual [0.184,1.360] and [0.010,0.046] for M2 and M3, respectively. Similar
observation or modelling being considered. In the presence of values of Ga for M2 were reported in Refs. [18] and [20]. The high
measurements and model predictions, the prior probability is value of these results may be explained by assuming that the
updated to posterior probability [43]. optimal values of Ga were reduplicated since the total energy
Assuming that the posterior probability of the parameters is dissipation in M2 was attributed only to one mechanism. Based on
approximated by a Gaussian distribution, the Laplace's method for this, the probability distribution of Ga can be updated to a uniform
asymptotic approximation can be applied to estimate the evidence distribution in the range of [0.1,1.5] for M2 and [0.01,0.1] for M3. The
as [44]. elastic property of the interphase was softer than that of the matrix
   h
 i1 in the measurements; D10 and D11 (c¼1.162 and 1.231, respectively),
P ðDjMi Þ ¼ P D X^i ; Mi P X^i Mi H X^i 2p
2
(18) whereas the matrix showed stiffer elasticity in the remaining
measurements.
where Xbi is the optimal parameter set that maximize the posterior Exploiting the optimal parameter sets, the models predictions
probability and Hð Xbi Þ is the Hessian matrix of versus the nanofiller volume fraction are depicted in Fig. 4. Obvi-
ln½PðDjXi ; Mi ÞPðXi jMi Þ with respect to Xi calculated at Xbi . The ously, M2 and M3 mostly have a similar ascending trend but it dif-
models are compared according to their model selection proba- fers slightly from M1.
bility calculated in Eq. (16). The model with the largest probability The model uncertainty can be demonstrated by the differences
is the optimum one. between the predictions and the measurements. This uncertainty is
In the present work, the models of predicting the fracture en- measured by the coefficient of variation (CV).
ergy of PNCs were evaluated. We considered the model of Huang
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
and Kinloch [14] (M1), the model of Williams [18] (M2), and the u
1u X Nj
1
model of Quaresimin et al. [21] (M3), which were described in CVij ¼ t ðDm  Yim Þ2 (19)
Section 2. The prior probabilities of these models were assumed to Dj Nj  1 m¼1
be equal, i.e. P(M1) ¼ P(M2) ¼ P(M3) ¼ 1/3.
Thanks to the uniform distribution assumed for the input pa- where Dj and Nj are the mean value and the number of the indi-
rameters, the prior probabilities of the model parameters, PðXi jMi Þ, vidual experiments of the j reference data, Dm is the measured
are constant disregarding the value of the parameter. The input value, and Yim is the corresponding predicted value of the model Mi.
parameters Vf ; dn ; Em ; and GIm were fixed as deterministic param- The CV values for M1, M2, and M3 are shown in Fig. 5. Except of
eters, while we calculated the most probable value (optimal the measurements: D8, D11, D12, and D13, M1 shows better perfor-
parameter value), which realized the best fit of the model pre- mance compared to M2 and M3, where its CV values are the least.
dictions to the measurements, for the remaining parameters. The predictions of M2 have the lowest discrepancies from the
Different experimental measurements gathered from the literature measurements of D11, D12, and D13. M3 produces the best fit

Table 2
The values of the input parameters used in the assessment of the models.

Reference All M1b M2b M3b

modelsa

Data dn Em sym nm gfm mm Kvm dv Vdp sym nm Ga sym nm Ga gfm mm tym t c

nm GPa MPa e e e e nm % MPa e J/m2 MPa e J/m2 e e MPa nm e

D1 [3] 20 3.20 71.6 0.35 0.732 0.194 2.227 37.7 14.1 72.3 0.36 0.287 85.1 0.36 0.015 0.722 0.184 66.0 3.00 0.882
D2 [3] 20 3.20 115.8 0.36 0.740 0.203 2.216 43.5 11.4 80.0 0.36 0.309 77.6 0.35 0.011 0.746 0.197 52.9 2.26 0.788
D3 [4] 12 3.53 80.5 0.33 0.728 0.206 2.236 25.5 10.4 76.9 0.36 0.184 81.4 0.34 0.013 0.730 0.204 67.3 2.24 0.644
D4 [4] 20 3.53 107.1 0.34 0.730 0.192 2.180 25.5 10.3 86.5 0.36 0.289 110.4 0.34 0.016 0.747 0.184 69.6 1.95 0.670
D5 [4] 40 3.53 118.4 0.34 0.704 0.177 2.108 45.0 13.2 82.7 0.36 0.473 97.0 0.36 0.015 0.727 0.214 59.0 2.90 0.375
D6 [5] 23 3.50 71.6 0.34 0.715 0.210 2.244 25.1 10.8 75.8 0.36 0.250 117.5 0.36 0.010 0.727 0.189 65.1 2.7 0.742
D7 [5] 74 3.50 71.6 0.37 0.745 0.222 2.236 91.2 15.2 75.8 0.35 0.914 110.1 0.34 0.046 0.727 0.188 63.4 2.18 0.447
D8 [6] 20 2.86 118.6 0.34 0.704 0.191 2.110 47.4 16.4 81.7 0.35 0.374 87.2 0.33 0.018 0.713 0.184 68.5 2.97 0.710
D9 [7] 20 2.96 70.5 0.36 0.737 0.205 2.190 25.7 12.5 80.2 0.35 0.310 112.5 0.36 0.011 0.745 0.192 62.6 1.08 0.758
D10 [8] 20 2.41 70.3 0.35 0.747 0.208 2.224 42.2 14.1 73.5 0.36 0.363 94.8 0.35 0.011 0.710 0.187 67.3 1.32 1.162
D11 [8] 80 2.41 70.7 0.36 0.730 0.223 2.239 118.8 17.9 71.5 0.37 1.360 88.5 0.34 0.046 0.716 0.184 63.8 1.20 1.231
D12 [9] 25 3.02 118.4 0.34 0.704 0.177 2.108 29.8 13.3 81.8 0.34 0.340 109.6 0.34 0.012 0.734 0.203 48.3 3.91 0.676
D13 [9] 25 2.78 116.7 0.36 0.704 0.181 2.112 56.6 16.7 72.7 0.34 0.383 87.2 0.34 0.010 0.709 0.220 60.2 3.78 0.656
D14 [10] 20 2.96 71.5 0.37 0.741 0.222 2.229 42.0 13.7 77.4 0.33 0.385 79.0 0.36 0.011 0.708 0.185 50.8 3.43 0.649
D15 [11] 13 2.60 80.6 0.36 0.705 0.181 2.221 35.3 11.1 71.1 0.35 0.224 97.1 0.36 0.010 0.749 0.176 67.4 2.83 0.770
D16 [12] 25 3.27 117.1 0.33 0.708 0.185 2.204 36.4 14.9 83.5 0.36 0.322 110.0 0.36 0.011 0.731 0.215 46.4 3.89 0.830
a
The values of dn and Em are obtained from the corresponding references.
b
These are the optimal values approximated in the current study.
 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129 127

Fig. 4. Predictions of the models using the optimal parameter set for the different reference data.

Fig. 5. The coefficient of variation for the different references data.

predictions in the measurements of D8 only. of M2 and M3 have steeper posterior probabilities. Significant
When considering both the model and parameters uncertainties changes in their prognoses are expected due to slight variations in
in the evaluation, M1 outperforms M2 and M3 for all the different the parameters values. One possible explanation is that the natural
measurements. It has significantly higher model selection proba- exponential relation in M2 and in M3 results in high values of the
bility, PðMi jDÞ (See Table 3). It can be concluded that the parameters determinant of their Hessian matrices.
128 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129

Table 3 Mech. 97 (2013) 193e206.

The models selection probability values for the different reference data. [5] P. Dittanet, R.A. Pearson, Effect of silica nanoparticle size on toughening
mechanisms of filled epoxy, Polymer 53 (9) (2012) 1890e1905.
Reference P(MijD)a [6] H.Y. Liu, G.T. Wang, Y.W. Mai, Y. Zeng, On fracture toughness of nano-particle
modified epoxy, Compos. Part B Eng. 42 (8) (2011) 2170e2175.
Data M1 M2 M3
[7] T. Hsieh, A. Kinloch, K. Masania, J.S. Lee, A. Taylor, S. Sprenger, The toughness
D1 [3] 0.988 0.000 0.012 of epoxy polymers and fibre composites modified with rubber microparticles
D2 [3] 1.000 0.000 0.000 and silica nanoparticles, J. Mater. Sci. 45 (5) (2010a) 1193e1210.
D3 [4] 1.000 0.000 0.000 [8] Y. Liang, R. Pearson, Toughening mechanisms in epoxyesilica nanocomposites
D4 [4] 1.000 0.000 0.000 (ESNs), Polymer 50 (20) (2009) 4895e4905.
[9] H. Zhang, L.C. Tang, Z. Zhang, K. Friedrich, S. Sprenger, Fracture behaviours of
D5 [4] 0.805 0.000 0.195
in situ silica nanoparticle-filled epoxy at different temperatures, Polymer 49
D6 [5] 0.999 0.000 0.001
(17) (2008) 3816e3825.
D7 [5] 0.998 0.000 0.002
[10] B. Johnsen, A. Kinloch, R. Mohammed, A. Taylor, S. Sprenger, Toughening
D8 [6] 0.998 0.000 0.002 mechanisms of nanoparticle-modified epoxy polymers, Polymer 48 (2) (2007)
D9 [7] 1.000 0.000 0.000 530e541.
D10 [8] 0.995 0.000 0.005 [11] B. Wetzel, P. Rosso, F. Haupert, K. Friedrich, Epoxy nanocompositesefracture
D11 [8] 0.744 0.001 0.255 and toughening mechanisms, Eng. Fract. Mech. 73 (16) (2006) 2375e2398.
D12 [9] 1.000 0.000 0.000 [12] H. Zhang, Z. Zhang, K. Friedrich, C. Eger, Property improvements of in situ
D13 [9] 1.000 0.000 0.000 epoxy nanocomposites with reduced interparticle distance at high nanosilica
D14 [10] 1.000 0.000 0.000 content, Acta Mater. 54 (7) (2006) 1833e1842.
D15 [11] 1.000 0.000 0.000 [13] J.k. Chen, Z.P. Huang, J. Zhu, Size effect of particles on the damage dissipation
D16 [12] 0.997 0.000 0.003 in nanocomposites, Compos. Sci. Technol. 67 (14) (2007) 2990e2996.
[14] Y. Huang, A. Kinloch, Modelling of the toughening mechanisms in rubber-
a
The probability of selecting the model Mi given the different reference data modified epoxy polymers. part ii a quantitative description of the
calculated by Eq. (16). microstructure-fracture property relationships, J. Mater. Sci. 27 (10) (1992a)
2763e2769.
[15] T. Hsieh, A. Kinloch, K. Masania, A. Taylor, S. Sprenger, The mechanisms and
mechanics of the toughening of epoxy polymers modified with silica nano-
4. Conclusion particles, Polymer 51 (26) (2010b) 6284e6294.
[16] P. Dittanet, R.A. Pearson, Effect of bimodal particle size distributions on the
Three existing models used for the prediction of the fracture toughening mechanisms in silica nanoparticle filled epoxy resin, Polymer 54
(7) (2013) 1832e1845.
toughness of PNCs were evaluated. The Bayesian method was [17] D.J. Bray, P. Dittanet, F.J. Guild, A.J. Kinloch, K. Masania, R.A. Pearson,
employed to quantify the model selection probabilities of Huang A.C. Taylor, The modelling of the toughening of epoxy polymers via silica
and Kinloch [14] model, Williams [18] model, and Quaresimin et al. nanoparticles: the effects of volume fraction and particle size, Polymer 54 (26)
(2013) 7022e7032.
[21] model. The model and parameters uncertainties were
[18] J. Williams, Particle toughening of polymers by plastic void growth, Compos.
considered in the assessment based on the experimental mea- Sci. Technol. 70 (6) (2010) 885e891.
surements of [3e12]. The optimal models predictions with respect [19] J. Williams, B. Blackman, H. Steininger, K. Zuo, Toughening by plastic cavita-
to these measurements were obtained using the optimal parameter tion around cylindrical particles and fibres, Compos. Sci. Technol. 103 (2014)
119e126.
sets. In contradiction to the references data of D8, D11, D12, and D13, [20] K. Zuo, B. Blackman, J. Williams, H. Steininger, The mechanical behaviour of
the optimal predictions of Huang and Kinloch model showed better ZnO nano-particle modified styrene acrylonitrile copolymers, Compos. Sci.
performance compared the other two models. However, for all the Technol. 113 (2015) 9e18.
[21] M. Quaresimin, M. Salviato, M. Zappalorto, A multi-scale and multi-
reference measurements, the model of Huang and Kinloch showed mechanism approach for the fracture toughness assessment of polymer
a distinctly higher model selection probability. On this base, we can nanocomposites, Compos. Sci. Technol. 91 (2014) 16e21.
conclude that it is the most robust model with regard to the applied [22] K.M. Hamdia, T. Lahmer, T. Nguyen-Thoi, T. Rabczuk, Predicting the fracture
toughness of pncs: a stochastic approach based on ANN and ANFIS, Comput.
reference measurements. Mater. Sci. 102 (2015a) 304e313.
[23] B. Fisher, M. Ireland, D. Boyland, S. Critten, Why use one model? an approach
for encompassing model uncertainty and improving best practice, Environ.
Acknowledgement Model. Assess. 7 (4) (2002) 291e299.
[24] T. Most, Assessment of structural simulation models by estimating un-
The authors gratefully acknowledge the support for this certainties due to model selection and model simplification, Comput. Struct.
89 (17) (2011) 1664e1672.
research provided by the IRSES-MULTIFRAC and the Alexander von
[25] H. Keitel, A. Dimmig-Osburg, L. Vandewalle, L. Schueremans, Selecting creep
Humboldt Foundation in the framework of the Sofja Kovalevskaja models using bayesian methods, Mater. Struct. 45 (10) (2012) 1513e1533.
Award endowed by the Federal Ministry of Education and Research, [26] K. Farrell, J.T. Oden, Calibration and validation of coarse-grained models of
atomic systems: application to semiconductor manufacturing, Comput. Mech.
Germany. Dr. Xiaoying Zhuang thanks the National High-end
54 (1) (2014) 3e19.
Foreign Experts by State Administration of Foreign Experts Af- [27] E. Prudencio, P. Bauman, D. Faghihi, K. Ravi-Chandar, J. Oden, A computational
fairs, PRC of Tongji University. framework for dynamic data-driven material damage control, based on
bayesian inference and model selection, Int. J. Numer. Methods Eng. 102 (3e4)
(2015) 379e403.
Appendix A. Supplementary data [28] K.M. Hamdia, M.A. Msekh, M. Silani, N. Vu-Bac, X. Zhuang, T. Nguyen-Thoi,
T. Rabczuk, Uncertainty quantification of the fracture properties of polymeric
nanocomposites based on phase field modeling, Compos. Struct. 133 (2015b)
Supplementary data related to this article can be found at http:// 1177e1190.
dx.doi.org/10.1016/j.compscitech.2016.02.012. [29] G. Irwin, Analysis of stresses and strains near the end of a crack traversing a
plate, J. Appl. Mech. 24 (1957) 361e364.
[30] J. Berriot, F. Lequeux, L. Monnerie, H. Montes, D. Long, P. Sotta, Filler-
References eelastomer interaction in model filled rubbers, a 1 h nmr study, J. Non Cryst.
Solids 307 (2002) 719e724.
[1] A. Argon, R. Cohen, Toughenability of polymers, Polymer 44 (19) (2003) [31] J. Berriot, F. Martin, H. Montes, L. Monnerie, P. Sotta, Reinforcement of model
6013e6032. filled elastomers: characterization of the cross-linking density at the filler-
[2] E.T. Thostenson, C. Li, T.W. Chou, Nanocomposites in context, Compos. Sci. eelastomer interface by 1 h nmr measurements, Polymer 44 (5) (2003)
Technol. 65 (3) (2005) 491e516, http://dx.doi.org/10.1016/ 1437e1447.
j.compscitech.2004.11.003. [32] A. Bansal, H. Yang, C. Li, K. Cho, B.C. Benicewicz, S.K. Kumar, L.S. Schadler,
[3] M. Zappalorto, A. Pontefisso, A. Fabrizi, M. Quaresimin, Mechanical behaviour Quantitative equivalence between polymer nanocomposites and thin polymer
of epoxy/silica nanocomposites: experiments and modelling, Compos. Part A films, Nat. Mater. 4 (9) (2005) 693e698.
Appl. Sci. Manuf. 72 (2015) 58e64. [33] S. Watcharotone, C.D. Wood, R. Friedrich, X. Chen, R. Qiao, K. Putz, L.C. Brinson,
[4] M. Zamanian, M. Mortezaei, B. Salehnia, J. Jam, Fracture toughness of epoxy Interfacial and substrate effects on local elastic properties of polymers using
polymer modified with nanosilica particles: particle size effect, Eng. Fract. coupled experiments and modeling of nanoindentation, Adv. Eng. Mater. 13
 K.M. Hamdia et al. / Composites Science and Technology 126 (2016) 122e129 129

(5) (2011) 400e404. validation processes for predictive stochastic models, Int. J. Numer. Methods
[34] G. Odegard, T. Clancy, T. Gates, Modeling of the mechanical properties of Eng. 87 (1e5) (2011) 262e272.
nanoparticle/polymer composites, Polymer 46 (2) (2005) 553e562. [44] J.L. Beck, K.V. Yuen, Model selection using response measurements: Bayesian
[35] R. Qiao, L.C. Brinson, Simulation of interphase percolation and gradients in probabilistic approach, J. Eng. Mech. (2004).
polymer nanocomposites, Compos. Sci. Technol. 69 (3) (2009) 491e499. [45] Y. Huang, A. Kinloch, Modelling of the toughening mechanisms in rubber-
[36] S. Yu, S. Yang, M. Cho, Multi-scale modeling of cross-linked epoxy nano- modified epoxy polymers. part i finite element analysis studies, J. Mater.
composites, Polymer 50 (3) (2009) 945e952. Sci. 27 (10) (1992b) 2753e2762.
[37] A. Pontefisso, M. Zappalorto, M. Quaresimin, An efficient RVE formulation for [46] H. Talebi, M. Silani, T. Rabczuk, Concurrent multiscale modelling of three
the analysis of the elastic properties of spherical nanoparticle reinforced dimensional crack and dislocation propagation, Adv. Eng. Softw. 80 (2015)
polymers, Comput. Mater. Sci. 96 (2015) 319e326. 82e92.
[38] M. Zappalorto, M. Salviato, M. Quaresimin, Influence of the interphase zone on [47] H. Talebi, M. Silani, S. Bordas, P. Kerfriden, T. Rabczuk, A computational library
the nanoparticle debonding stress, Compos. Sci. Technol. 72 (1) (2011) 49e55. for multiscale modelling of material failure,, Comput. Mech. 53 (5) (2014)
[39] M. Zappalorto, M. Salviato, M. Quaresimin, A multiscale model to describe 1047e1071.
nanocomposite fracture toughness enhancement by the plastic yielding of [48] H. Talebi, M. Silani, S.P.A. Bordas, P. Kerfriden, T. Rabczuk, Molecular dy-
nanovoids, Compos. Sci. Technol. 72 (14) (2012) 1683e1691. namics/XFEM coupling by a three-dimensional extended bridging domain
[40] M. Salviato, M. Zappalorto, M. Quaresimin, Plastic shear bands and fracture with applications to dynamic brittle fracture, Int. J. Multiscale Comput. Eng. 11
toughness improvements of nanoparticle filled polymers: a multiscale (6) (2013) 527e541.
analytical model, Compos. Part A Appl. Sci. Manuf. 48 (2013) 144e152. [49] P. Budarapu, R. Gracie, S. Bordas, T. Rabczuk, An adaptive multiscale method
[41] A. Pontefisso, M. Zappalorto, M. Quaresimin, Influence of interphase and filler for quasi-static crack growth, Comput. Mech. 53 (6) (2014) 1129e1148.
distribution on the elastic properties of nanoparticle filled polymers, Mech. [50] P. Budarapu, R. Gracie, Y. Shih-Wei, X. Zhuang, T. Rabczuk, Efficient coarse
Res. Commun. 52 (2013) 92e94. graining in multiscale modeling of fracture, Theor. Appl. Fract. Mech. 69
[42] D.J. MacKay, Bayesian interpolation, Neural Comput. 4 (3) (1992) 415e447. (2014) 126e143.
[43] J.T. Oden, S. Prudhomme, Control of modeling error in calibration and