JOURNAL OF

COMPOSITE
Review M AT E R I A L S
Journal of Composite Materials
0(0) 1–17
! The Author(s) 2018
Viscoelastic behavior of an epoxy Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
resin during cure below the glass DOI: 10.1177/0021998318781226
journals.sagepub.com/home/jcm
transition temperature:
Characterization and modeling

Alice Courtois1 , Martin Hirsekorn2, Maria Benavente1,

Agathe Jaillon1, Lionel Marcin3, Edu Ruiz1 and Martin Lévesque1

Abstract
This paper presents a viscoelastic temperature- and degree-of-cure-dependent constitutive model for an epoxy resin.
Multi-temperature relaxation tests on fully and partially cured rectangular epoxy specimens were conducted in a dynamic
mechanical analysis apparatus with a three-point bending clamp. Master curves were constructed from the relaxation test
results based on the time–temperature superposition hypothesis. The influence of the degree of cure was included
through the cure-dependent glass transition temperature which was used as reference temperature for the shift factors.
The model parameters were optimized by minimization of the differences between the model predictions and the
experimental data. The model predictions were successfully validated against an independent creep-like strain history
over which the temperature varied.

Keywords
Stress relaxation, epoxy, cure, viscoelasticity

constitutive models for polymer resins. Most authors

Introduction
Process-induced residual stresses arise in the liquid
molding of composites due to the differential volumet-
ric changes of the constituents and the mechanical
interactions between the mold and the parts. Volume
changes stem from the chemical shrinkage of the resin
as well as from the coefficients of thermal expansion
(CTE) mismatch between the resin and fibers.
Process-induced residual stresses result in unwanted
distortions in composite parts. Numerical tools have
been developed to predict residual stresses in composite
parts made by different manufacturing processes such
as autoclave1 or resin transfer molding (RTM).2 The
main challenges lie in the fact that the resin evolves
from a liquid to a viscoelastic solid state through a
complex time and temperature-dependent process.
Resin’s gel point (gel ) corresponds to a given degree
of cure for which an infinite polymer network is cre-
ated. This is typically associated with a steep buildup of
the resin’s mechanical properties. A number of authors
have therefore developed degree-of-cure-dependent
assumed that before the gel point the resin is so com-
pliant that residual stresses are relaxed immediately.3–7
Bogetti and Gillespie1 proposed a modified degree of
cure rule of mixtures where an epoxy’s Young’s modu-
lus for a given degree of cure () was related to those of
the fully cured and uncured conditions. Ruiz and
Trochu8 proposed a model where the resin’s stiffness
depended on an after-gel-point degree of cure (agp )
after which the resin exhibited a significant rigidity.
The gel point of epoxy resins lies typically at degrees
of cure between 0.6 and 0.7. Khoun and Hubert9 found
an gel of 0.7 for an epoxy resin by taking the degree of
1
Department of Mechanical Engineering, Polytechnique Montréal, Canada
2
ONERA, France
3
Safran Tech, France
Corresponding author:
Martin Lévesque, Polytechnique Montréal Laboratory for Multi-Scale
Mechanics, 2900 Boulevard Edouard Montpetit, Montréal, QC H3C 3A7,
Canada.
Email: martin.levesque@polymtl.ca
2 Journal of Composite Materials 0(0)

cure when the storage and loss moduli are equal. to obtain relaxation modulus. Adolf et al.23 obtained
Zarrelli et al.10 obtained an gel between 0.64 and the parameters through an extensive experimental cam-
0.66 through isothermal rheometric experiments. paign on four different polymers in the glass transition
Simon et al.4 and Prasatya et al.11 determined the gel region.
point of an epoxy resin theoretically to gel ¼ 0.63. The purpose of this work was to characterize, pre-
The polymer network formation is in fact a continu- dict and validate the relaxation behavior of a viscoelas-
ous process transforming the resin from a viscous liquid tic epoxy resin by accounting for its temperature and
before the gel point to a viscoelastic solid with increas- degree of cure dependencies, based on a thermodynam-
ing viscosity. Therefore, stress relaxation remains ically rigorous approach, which is easily implementable.
important as long as the resin is not fully polymerized The paper is organized as follows: the material studied,
and at temperatures close to the glass transition tem- the specimen manufacturing, the thermo-mechanical
perature Tg . Tg defines the temperature around which testing procedures and the experimental results are first
the polymer evolves from a brittle glassy state to a vis- presented. The mechanical model is then detailed with
cous rubbery state. The resin’s mechanical properties the parameter optimization and the model predictions
drop around that temperature and stress relaxation are finally validated on different load cases.
becomes much more important. Since Tg is usually
exceeded during manufacturing processes, relaxation
will be important during curing and at the first stages
Material and experimental procedures
of the cooling process. The residual stresses are only A commercial DGEBF epoxy resin (DiGlycidyl Ether
really set towards the end of the cooling phase. of Bisphenol F), already prepared, was studied in this
Several authors therefore proposed temperature and work.
cure-dependent viscoelastic models for epoxy resins.
For example, Kim and White12 proposed to model
the -dependency through time shift factors aT .
Samples manufacturing
Simon et al.4 used the cure-dependent glass transition Dedicated silicon molds, shown in Figure 1, were used
temperature Tg as the reference temperature and relied to produce 150 rectangular fully and partially cured
on the DiBenedetto equation13 to relate Tg with the specimens of nominal dimensions of 50 mm
degree of cure. O’Brien et al.14 introduced the 10 mm  2:7 mm. The fully cured specimens were held
-dependency through the stiffness of a linearly visco- at 180 C during 180 min under a pressure of 0:55 MPa
elastic model. These models were all developed for to limit porosities. Partially cured specimens were
epoxy resins and each reproduced experimental data
reasonably well. However, very few authors reported
and discussed the inherent variability of such measure-
ments and validated their model on independent data
sets. The respect of thermodynamics principles was also
rarely accounted for. Caruthers et al.15 and Adolf and
Chambers 16 developed a thermodynamically consistent
viscoelastic approach for thermosets during cure based
on rational thermodynamics, which led to doubly con-
voluted integrals linking strains and stresses histories.
Lévesque et al.17 summarized the viscoelastic constitu-
tive theories developed by Biot18 and Schapery19 based
on the thermodynamics of irreversible processes, which
led to single convoluted integrals. Single convoluted
integrals can be cast under a differential form,20
which delivers highly efficient implementation algo-
rithms, which is a clear advantage of such formulations.
While some authors already attempted to merge
viscoelasticity, temperature, and cure dependen-
cies,4,8,14,21,22 the experimental data sets used to
obtain the constitutive theories parameters were usually
limited to fully cured specimens or to very few partially
cured specimens. O’Brien et al.14 measured viscoelastic
properties on the entire range of degrees of cure Figure 1. Silicon mold used to manufacture epoxy samples for
through creep tests and needed conversion calculations DMA tests.
Courtois et al.
produced by interrupting the cure cycle after 81, 68, 59,
54, 49 and 45 min at 165 C. The samples were then
cooled down by convection outside the mold at room
temperature, which was too low for the studied resin to
continue curing. For each mold and curing cycle, one
sample was broken into smaller parts that served as
differential scanning calorimetry (DSC) samples. The
degrees of cure and the glass transition temperatures
were then measured by DSC on two to four samples
for each mold and curing cycle after the cooling. The
modulated mode was used and a ramp of 2 C/min from
25 C to 290 C was applied. These measurements
showed that the six partially cured specimen batches
were polymerized at 94%, 90%, 86%, 80%, 74% and
66%, at 0:2 %, respectively, with a maximum con-
trast of 0.5% for each silicon mold. The lowest degree
of cure is therefore close to the values for the gel point
found by Simon et al.,4 Prasatya et al.,11Zarrelli et al.10
and Khoun and Hubert.9 In fact, the specimens at 66%
cure were very fragile and demolding and polishing
were particularly challenging, but mechanical testing
was still possible. The gel point of the studied resin
lies therefore not much below 66%, and the tested spe-
cimens cover well the whole range of cure from the gel
point to full polymerization.
Thermo-mechanical testing in the linearly
viscoelastic domain
RTM manufactured composite parts are submitted to
small deformations during the process. For this reason,
Figure 2. Determination of the linearly viscoelastic range during a relaxation experiment at 30 C. 0:1 represents the stress
evolution for a strain level of 0.1%. 0:05 represents the stress evolution for a strain level of 0.05%. 0:052 corresponds to 2  0:05 .
The figure suggests that the material obeys a linearly viscoelastic constitutive theory for a strain range of 0.1% and a temperature of
30 C.
3
this work focused on strains in the order of 0.1%.
Relaxation tests were performed for strain levels of
"1 ¼ 0:05 % and "2 ¼ 2"1 ¼ 0:1 % at 30 C and 150 C
with a three point bending clamp in a TA Instruments
Q800 DMA apparatus on fully cured specimens to
ascertain the linearly viscoelastic domain in this strain
range. Figures 2 and 3 show the stress evolution, at
30 C and 150 C, respectively, for samples tested at
the two different applied strains. The stress responses
to a strain of 0.05%, multiplied by 2, were added and
compared to the stress evolution resulting from an
applied strain of 0.1%. Figures 2 and 3 show that the
stress response due to a strain of 0.1% and twice the
stress response to an applied strain of 0.05% is very
close to each other for both the temperatures. It can
therefore be assumed that the studied resin remains in
the linearly viscoelastic range for strain and tempera-
tures below 0.1% and 150 C (below the glass transition
temperature), respectively, for a fully cured state. It was
assumed that this observation held for the other degrees
of cure (for strain below 0.1% and temperatures below
their glass transition temperature).
Multi-temperature relaxation tests were then carried
out on fully and partially cured specimens. Starting
from 30 C, the specimens were bent to 0.1% of strain
and held at constant strain for 90 min to measure relax-
ation behavior. After each relaxation phase, the bend-
ing load was removed and 10 min of recovery was
allowed. The temperature was then increased by 15 C
over a heating phase of 10 min. The temperature was
held constant at the new level for an additional 5 min
4 Journal of Composite Materials 0(0)

Figure 3. Determination of the linearly viscoelastic range during a relaxation experiment at 150 C. 0:1 represents the stress
evolution for a strain level of 0.1%. 0:05 represents the stress evolution for a strain level of 0.05%. 0:052 corresponds to 2  0:05 .
The figure suggests that the material obeys a linearly viscoelastic constitutive theory for a strain range of 0.1% and a temperature of
150 C.

Figure 4. Normalized average stress relaxation curves for partially and fully cured samples from  ¼ 0.66 to  ¼ 1.0.

prior to initiating the following relaxation phase in
order to ensure the apparatus and sample thermal equi-
librium. This procedure was repeated up to a tempera-
ture below the glass transition temperature (150 C for
the fully cured specimens for example), because above
Tg the resin becomes too compliant to obtain reliable
experimental data with the used DMA apparatus. The
limitation to temperatures below Tg also ensures that
the degree of cure does not evolve significantly during
the experiments. The entire load history was considered
further in the parameters determination through the
model implementation, as detailed in ‘‘Numerical
implementation’’ section.
Experimental results
Figure 4 shows the relaxation curves for fully and par-
tially cured samples at the tested temperatures. Note
that the stress values were normalized by the actual
stress measured at the beginning of relaxation, on
Courtois et al. 5

fully cured samples and at 30 C, for confidentiality. For an adiabatic and isothermal loading history, equation
clarity, only the average stress evolution is plotted for (1) becomes
each degree of cure. The full experimental data of all  
specimens are shown in ‘‘Comparison between model @ @
i  "_i  _ r ¼ fi "_i þ br _ r  0 ð2Þ
predictions and experimental relaxation data’’ section. @"i @r
The figure shows that the relaxation rate increased as
the temperature raised and the cure decreased. where r are internal variables and fi and br are the
Moreover, at first sight there is no clear hierarchy thermodynamics forces that can be expressed as a func-
between the responses at different degrees of cure at tion of state and internal variables (fi ¼ Fij "_j ,
30 C and 45 C. Since the temperature was limited to br ¼ Brs _ s , with F and B constant matrices) in case of
Tg , the lower was the degree of cure, the lower was the linear viscoelasticity. The free energy can then be
last tested temperature. expanded25

1 1
Mechanical model ðe, vÞ ¼ 0 þ e : L1 : e þ e : L2 : v þ v : L3 : v ð3Þ
2 2
This section recalls the thermodynamics of irreversible
processes framework, developed by Biot18 and where 0 is the free energy in the equilibrium state,
Schapery19 and summarized by Lévesque et al.,17 used
@ @
to write viscoelastic constitutive theories. Then, a spe- L1 ¼ ; L2 ¼ ;
cial form of the viscoelastic model is proposed to @"i @"j @"i @r
  ð4Þ
account for the influence of the degree of cure. The @ L1 L2
L3 ¼ and L ¼
model equations are defined based on the experimental @r @s ðL2 ÞT L3
observations presented previously. The final parts of
this section comprise a description of the numerical L is symmetric and positive semi-definite. As shown
implementation of the model, the parameter optimiza- by Lévesque et al.,17 linearly viscoelastic models can
tion procedure and the comparison between model then be written in the general form
results and experimental data.
rðtÞ ¼ L1 : eðtÞ þ L2 : vðtÞ ð5aÞ
Background :
B : v þL3 : v þ ðL2 ÞT : e ¼ 0 ð5bÞ
Combining the first and second laws of thermo-
dynamics leads to the well-known Clausius–Duhem where
inequality24
@) :
  ¼ B : v ð6Þ
de d dT q  rT @v
r:  þ   0 ð1Þ
dt dt dt T
The solution of equation (5) yields26
where  is Helmoltz’s free energy,  the mass density,   
the entropy density and q the heat flux. Constitutive L2 L2
i ðtÞ ¼ L1ij  ir jr "j
theories must meet this inequality to meet the first L3rr
two principles of thermodynamics. Two main paths Z   ð7Þ
L2ir L2jr t L3 d"j
have been undertaken to derive constitutive theories þ exp  rr ðt  Þ d
L3rr 0 Brr d
from equation (1):25 rational thermodynamics15,16 and
thermodynamics of irreversible processes.18,19 The
rational thermodynamics approach relies on a rigorous which can be cast under the form
Taylor series expansion of the free energy which leads
to doubly convoluted integrals linking stresses and Z tX
N  
t de
strains histories. In the thermodynamics of irreversible rðtÞ ¼ C1 : eðtÞ þ Ck exp  : d ð8Þ
0 k¼1
k d
processes, the free energy is expanded through a mod-
ified second-order Taylor series around a reference
state17 and leads to a single convolution integral relat- C1 is the fully relaxed tensor and the Ck are the
ing the stresses and strains histories, but for which the relaxation tensors associated with the relaxation times
parameters can depend on stress, temperature, etc. In  k. The Ck are positive semi-definite and symmetric. In
addition, this approach expresses the free energy as a one dimension (1D), equation (8) is typically referred to
function of state and hidden variables. When assuming as the generalized Maxwell model.
6 Journal of Composite Materials 0(0)

Following the same thermodynamically rigorous evolution of aT with temperature changes around the
approach, Schapery27 has introduced nonlinear consti- glass transition temperature Tg . Below Tg , it is well
tutive theories through scalar nonlinearizing functions. described by an Arrhenius relationship31,32
The expansion of the free energy becomes  
Ea 1 1
1 1 log aT ¼  ð14Þ
ðe, vÞ ¼ 0 þ e : L1 : e þ p3 e : L2 : v þ p2 v : L3 : v ln 10R T Tref
2 2
ð9Þ while above Tg , the well-known Williams–Landel–
Ferry equation (WLF)33
and
M1 ðT  Tref Þ
log aT ¼ ð15Þ
@) : M2 þ T  Tref
¼ p1 B : v ð10Þ
@v
is the most appropriate. In these equations, Ea is an
where p1, p2 and p3 are scalar functions that can depend activation energy (J/mol), R is the universal gas con-
on state variables such as strain eðtÞ, temperature T, or stant (8.314 J/(Kmol)), M1 and M2 are adjustable
degree of cure . Included into the formulation recalled material constants, and Tref is the reference
by Lévesque et al.17, equation (5) yields (see Crochon temperature.
et al.20)
  The proposed viscoelastic model
@ @p3
rðtÞ ¼ ðtÞ þ  L2 : eðtÞ þ p3 L2 : vðtÞ ð11aÞ
@e @e In the studied strain range relevant for the formation of
residual stresses, the viscoelastic behavior of the studied
:
p1 B : v þp2 L3 : v þ p3 ðL2 ÞT : e ¼ 0 ð11bÞ resin is linear with strain. The nonlinearizing functions
in equations (11) and (12) are therefore supposed to
where  is a scalar function. Solution of equation (11) depend only on temperature and degree of cure.
yields Introducing  ¼ 12 e : C1 : e, p1 ¼ pðÞaT ðT, Þ, and
  Zt p2 ¼ p3 ¼ pðÞ, equation (12) becomes
@ @p3
rðtÞ ¼ þ  e þ p3 I : CððtÞ  ðÞÞ Z t X
N
@e @e 0
  rðtÞ ¼ C1 : e þ pðÞ  Ck exp
d p3 ðÞ 0 k¼1
 eðÞ d  Z t Z 
d p2 ðÞ 1 1 1 de
ð12Þ  d  d  : d
k 0 aT ðT, Þ 0 aT ðT, Þ d
P
N   ð16Þ
where CðtÞ ¼ Ck exp  tk and
k¼1 In one dimension, equation (16) reduces to
Z t Z t
p2 ðÞ 1
ðtÞ ¼ d ¼ d ð13Þ Z t X
N
0 p1 ðÞ 0 aT ðÞ
ðtÞ ¼ C1 : " þ pðÞ  Ck exp
0 k¼1
is a reduced time. The time () dependence of p2 and p3  Z Z 
t 
in equation (12) is interpreted through the time depend- 1 1 1 d"
 d  d  : d
ence of the state variables, on which p2 and p3 depend. k 0 aT ðT, Þ 0 aT ðT, Þ d
The factor aT in equation (13) is called a shift factor. ð17Þ
When p2 ¼ p3 ¼ 1 and p1 ¼ aT ðTÞ, equation (13)
reduces to the so-called time–temperature superpos- The model obeys the TTS principle through the
ition principle (TTSP28). The TTSP can be used to con- cure- and temperature-dependent shift factors
struct master curves by shifting relaxation or creep aT ðT, Þ. The relaxation modulus is written in a Prony
curves obtained at different temperatures on the loga- series as in equation (12). The cure dependence of the
rithmic time scale by a temperature-dependent shift relaxation modulus is described by the function pðÞ.
factor. A good superposition of the relaxation or
creep curves has to be observed. The relaxation/creep
Model identification
time can be extended to ensure a true superposition or
tests at closer temperatures can be carried out. It was Shift factors. According to the TTS principle, each indi-
found by different authors29,30 that for epoxy resins the vidual isothermal relaxation curve was shifted along the
Courtois et al. 7

logarithmic time scale to construct master curves for
every degree of cure, with 30 C taken as the reference
temperature. The shift factors aT were determined auto-
matically with a Matlab program developed at Onera
that seeks the shift along the logarithmic time scale
minimizing the average distance of the stress values
with respect to the relaxation curve at the next lower
temperature. In this way, an independent shift factor
was determined for each temperature. One shift factor
set for each degree of cure is plotted for clarity in
Figure 5. It can be seen that the temperature depend-
ence of the shift factors is well approximated by the
Arrhenius relationship given by equation (14).
However, the slope of the curves decreases with the
degree of cure, indicating that the activation energy
Ea depends on cure.
The activation energy determined by fitting the
Arrhenius model individually to the shift factors
obtained for each specimen is plotted as a function of
the degree of cure in Figure 6. A decreasing linear rela-
tionship between the degree of cure and the activation
energy was observed, suggesting that the relaxation
mechanisms of partially cured specimens were more
sensitive to temperature. We therefore introduce the

Figure 5. Shift factors with respect to Tref ¼ 30 C as a function of the temperature (T 5 Tg ) obtained from the master curve
construction, from  ¼ 0:66 to  ¼ 1, together with the Arrhenius model identified individually for each degree of cure.

Figure 6. Master curve data for the Ea -dependency, from  ¼ 0.66 to  ¼ 1.0, together with the postulated linear model.
8 Journal of Composite Materials 0(0)

cure dependence of the shift factors into the Arrhenius
relationship (equation (14)) through an activation
energy in the form
Ea ðÞ ¼ a þ b
where a and b are the adjustable parameters.
Master curves. The master curves were constructed using
the method described in the previous paragraph with
the independent shift factors for each temperature.
Figure 7 shows the relaxation data of a partially
cured specimen for each temperature along a
logarithmic time scale and the resulting master curve,
with a good overlap for each step. This observation
supports the relevance of using the TTSP for the stu-
died epoxy resin. The master curves for each degree of
ð18Þ cure are shown in Figure 8 as a function of the reduced
time . One curve for each degree of cure is plotted for
clarity. It can be seen that the master curves for higher
degrees of cure are stiffer that those for lower .
Moreover, the relaxation modulus decrease was
delayed for higher degrees of cure. The same lack of
hierarchy noticed in the experimental relaxation results
can be observed at the beginning of the master curves,
corresponding to the first temperature steps.

Figure 7. Master curve construction for a partially cured sample at Tref ¼ 30 C.

Figure 8. Master curves for partially and fully cured samples at Tref ¼ 30 C, from  ¼ 0.66 to  ¼ 1.0.
Courtois et al. 9

Every specimen master curve was independently Ck can therefore be interpreted as a sum of a Gaussian
fitted with Prony series in order to obtain a first and a sigmoid distribution.
approximation of the Ek ¼ pðÞCk . The relaxation For illustration purposes, Figure 10 shows the dis-
times  k were set to one per decade and every Ek was crete Ek extracted from a fully cured specimen master
associated with a  k such that Ek ðk Þ. Similar distribu- curve, the fitted Ek as well as the sigmoid and Gaussian
tions for Ek ¼ pðÞCk ðk Þ were obtained for each degree parts. The advantage of using this continuous function
of cure, as can be seen in Figure 9. These distributions to approximate the Ek is that it reduces greatly the
exhibit a plateau of approximately constant Ek at short number of independent parameters of the model
relaxation times, followed by a peak that can be well (from 15 (one Ek per relaxation time) to 4), thus facil-
fitted by a Gaussian distribution at longer relaxation itating the optimization process required for obtaining
times. Therefore, the distribution of the Ck ðk Þ was the model parameters.
assumed to be of the shape
" "  #
log k  log peak 2 Influence of the degree of cure
Ck ðk Þ ¼ exp 
lpeak Figure 9 shows that the peak position log peak increases
   monotonically with the degree of cure. Since both
log k  log peak ð19Þ
þ 1  erf asymptotes of the continuous function towards short
lpeak
relaxation times (the plateau) and long relaxation
log kþ1  log k1 times (zero) are horizontal, instead of introducing a
 ðC0  C1 Þ
2 cure dependent peak position, an -dependency can
where C0 is the instantaneous relaxation tensor be introduced into the shift factors aT . The approach
of Simon et al.4 was adopted to use the cure-dependent
X
k
C0 ¼ C1 þ p Ck ð20Þ glass transition temperature Tg as reference tempera-
k¼1
ture for the shift factors.
The evolution of Tg with the degree of cure was
and , lpeak and log peak are adjustable parameters determined from DSC measurements. It is well
related to the plateau height, the peak width and the approximated by the DiBenedetto equation13
position of the peak on the relaxation time axis,
respectively. erf is defined as Tg  Tg0 
¼ ð22Þ
Z Tg1  Tg0 1  ð1  Þ
x
2
erfðxÞ ¼ pffiffiffi expðt2 Þdt ð21Þ
0 where Tg0 and Tg1 are the glass transition temperatures
of the uncured and the fully cured resin, respectively,

Figure 9. Ek distributions with Tref ¼ 30 C, from  ¼ 0.66 to  ¼ 1.0, and best fitting continuous functions (equation (19)) computed
individually for each degree of cure.
10
Figure 10. Discrete Ek values for  ¼ 1:0 and corresponding analytical relationship.
Figure 11. Comparison between DSC data and the DiBenedetto model.
and is a material parameter. Both the experimental
DSC results and the DiBenedetto model are plotted in
Figure 11.
Figure 12 reports the same plots as that of Figure 9,
but for each degree of cure Tref ¼ Tg ðÞ was taken as
the reference temperature. The figure shows that the
curves almost collapse into a single distribution,
which confirms that the approach of Simon et al.4 is
appropriate for the studied resin. However, the peak
areas vary slightly with the degree of cure. Its values
that yield the curves shown in Figure 12 are plotted
in Figure 13 as a function of the degree of cure.
Journal of Composite Materials 0(0)
The heights or widths variations lead to similar effects.
The peaks width was therefore fixed and the cure
dependence focused on the peak heights was account-
ing for through parameter pðÞ in Ek ¼ pðÞCk . It can
be deduced that p ¼ lpeak pffiffi decreases with the degree
area
of cure, the same way as the peaks area shown in
Figure 13. Therefore, pðÞ was set to
pðÞ ¼ K1  þ K2 ð23Þ
where Ki are constants.
Courtois et al. 11

Figure 12. Ek distributions with Tref ¼ Tg ðÞ, from  ¼ 0.66 to  ¼ 1.0.

Figure 13. Discrete peak area values extracted from each master curve fit and a linear approximation from  ¼ 0.66 to  ¼ 1.0.

With these specifications, the model (equation (17)) and log peak . C1 was obtained from equation (20)
reduces to and involved parameters C0 and Ck. The shift factors
aT were computed from equation (14) and involved par-
Z t X
N ameters Ea and Tref ¼ Tg . Ea was obtained from equa-
ðtÞ ¼ C1 : " þ pðÞ  Ck exp tion (18) and involved constants a and b, as well as Tg
0 k¼1
 Z t from equation (22), which involved fitting .
1 1 ð24Þ
 d
k 0 aT ðT, Tg ðÞ, Þ
Z  Numerical implementation
1 d"
 d  : d The model (equation (24)) was implemented in Matlab
0 a T ðT, T g ðÞ, Þ d
using a recursive strategy.34,35 The time scale was
p was obtained from equation (23) and involved divided into several small increments over which the
fitting parameters K1 and K2. The Ck were obtained strain evolution was assumed as linear as a function
from equation (19) and involved parameters , lpeak of time. The temperature and degree of cure were
12 Journal of Composite Materials 0(0)

assumed constant over each increment. Under these t

where  ¼ ð27bÞ
assumptions, equation (24) can be integrated and writ- aT
ten considering the internal variables and the instant-
aneous relaxation tensor (equation (20)) The strain state (total strain and values of the inter-
nal variables k) at the end of each recovery phase was
X
N taken as the starting state of the following loading step,
 ¼ C0 "  p Ck k ð25aÞ because 10 min of recovery plus heating was not suffi-
k¼1
cient to let the specimens fully recover. The heating
Z t   phases were modeled as recovery phases with zero
t   d"
where k ¼ 1  exp  d ð25bÞ applied stress and increasing temperature.
0 aT k d

Parameter optimization
leading to the incremental form
As stated previously, each specimen master curve was
X
N first independently fitted with Prony series, yielding dis-
 ¼ C0 "  p Ck k ð26Þ crete values for the shift factors and relaxation moduli.
k¼1
On the basis of these discrete values, continuous func-
tions were defined in order to reduce the number of
C1 , p and Ck are the material constants, and k was independent parameters of the model. In the final
developed by Machado et al.36 model, these continuous functions are used to calculate
    the relaxation moduli Ek for each discrete relaxation
k   time  k (one per decade). The parameters of the con-
k ¼ nþ1
k   n
k ¼ þ exp  1 "
 k k tinuous functions obtained from the independent
   master curve fits are used as initial parameter set

þ 1  exp ð"n  nk Þ ð27aÞ for an optimization of the model parameters based on
k
a simulation of the whole experimental dataset

Table 1. Optimized model parameters.

a ðkJ=molÞ b ðkJ=mol Þ K1 K2 lpeak log peak C1 ðMPaÞ

–62.18 256.81 363 697 0.087 1.28 0.57 0.27 6.27

Figure 14. Ek values after parameter optimization with Tref ¼ Tg .

Courtois et al. 13

(a) (e)

(b)
(f)

(c)
(g)

(d)

Figure 15. Model predictions compared to the normalized experimental stress relaxation for the fully cured specimens (a) and the
specimens cured at 94% (b), 90% (c), 86% (d), 80% (e), 74% (f), and 66% (g).

(all relaxation and recovery curves, for all tested tem-
peratures and degrees of cure). Only the DiBenedetto
parameters, Tg0 , Tg1 and , were identified separately
from the others to fit the DSC measurements. The
relaxed modulus C1 was also fixed separately to the
horizontal asymptote of the relaxation curves of the
fully cured samples at the highest temperature. The
remaining seven unknown model parameters (a and b
from equation (18), , lpeak and log peak from equation
(19), K1 and K2 from equation (23)) were obtained by
minimizing the cost function
P
ðmðti Þ  d ðti ÞÞ2
r¼ P ð28Þ
½d ðti Þ 2
with Matlab’s lsqnonlin algorithm. During the relax-
ation and recovery phases, mðti Þ are the predicted and
d ðti Þ the experimentally measured stress or strain,
respectively. The residue of one multi-temperature
relaxation test was defined as the sum of the residues
14
Figure 16. Temperature ( C) and load (N) applied on fully cured specimens.
of the different relaxation and recovery phases. The
parameter set that minimized the discrepancies between
the model predictions and the experimental data is
listed in Table 1. The Ek calculated with these param-
eters are shown in Figure 14 for each tested degree of
cure together with the corresponding continuous func-
tions of equation (19).
Comparison between model predictions and
experimental relaxation data
The numerically calculated stress curves of the multi-
temperature relaxation tests are shown in Figure 15 and
compared to the experimental data of all tested speci-
mens. The stress evolution predicted by the model glo-
bally lies within the experimental data range for the
tested temperatures and degrees of cure, which shows
that the stress evolution in time during relaxation,
and its dependence on temperature and cure is well
taken into account in the model. Due to the large vari-
ability in the experimental data observed for certain
degrees of cure, there may be important differences
between the model predictions and an individual test.
It is therefore important that several tests are carried
out for a given degree of cure, in order to evaluate these
variabilities.
Model validation: Comparison with
independent experimental creep data
Creep experiments were carried out in order to test the
robustness of the proposed model. Fully cured epoxy
specimens were manufactured according to the proto-
col detailed previously. The same DMA and clamp as
Journal of Composite Materials 0(0)
described in the ‘‘Thermo-mechanical testing’’ section
were used. These tests were deemed as a challenging
validation case since model parameters were extracted
from relaxation curves. Moreover, the samples were
tested under non-isothermal conditions. The thermal
expansion was then taken into account, considering a
CTE of 67:106 ðm=ðmKÞÞ.
Temperature and force ramps were successively
applied onto the specimens, as shown in Figure 16.
The temperature ramps were of 0.5 C/min to ensure
the apparatus and sample thermal equilibrium and
the force ramp of 0.1 N/min. After a stabilization at
30 C, a temperature ramp was applied to 45 C and a
force ramp until 1 N followed. Then, temperature and
force ramps were alternated successively until 80 C,
1.5 N, 120 C, and 2 N. Finally, a temperature ramp
until 150 C was applied.
The experimentally measured strain was compared
with the model predictions, considering the optimized
parameters a and b from equation (18), K1 and K2 from
equation (23), , lpeak and log peak from equation (19),
from equation (22) and C1 from equation (20), gath-
ered in Table 1. Figure 17 shows a good agreement
between the predictions and the experimental data in
the range of temperatures for which the model was
identified.
Discussions
It should be noted that we were expecting an increase
of the instantaneous elastic stiffness with an
increasing degree of cure. However, it can be seen
from Figure 13 that p decreases with cure, which
means that the instantaneous elastic stiffness also
Courtois et al. 15

Figure 17. Model predictions compared to the experimental creep data on a fully cured specimen.

Figure 18. Experimental relaxation stress at 30 C, gathering the data in two groups, with confidence intervals of 95%.

decreases with the degree of cure. A statistical analysis
was carried out to confirm this trend.
The experimental data set was split into two groups:
group A contained the experimental data obtained for
100% and 94% cured specimens, while group B con-
tained the experimental data for 80%, 74% and 66%
cured specimens. Figure 18 shows the mean relaxation
stress curves obtained at 30 C for groups A and B,
along with a 95% confidence interval on this mean
value. At the lowest tested temperature, the stress relax-
ation is the weakest and thus the contribution of the
instantaneous elastic stiffness is the most important.
The figure shows that none of the confidence intervals
overlap, which evidences statistically different responses
between group A and group B. This observation sup-
ports the fact that the instantaneous elastic stiffness
decreases with an increase in degree of cure. The che-
mico-physical reasons for this observation are still an
open question to be investigated in the future.
Conclusions
This work proposed a thermodynamically consistent
viscoelastic model to predict the coupled degree of
cure- and temperature-dependent response of an
epoxy resin. The proposed model was able to reproduce
16
the experimental data used for identifying its param-
eters within their variability. Its predictions are also
in good agreement with the strain evolution observed
in an independent non-isothermal multi-level creep
experiment used for validating the model.
The observation of the experimental data revealed
that there is a significant scatter in the results.
Moreover, the partially cured samples were quite brit-
tle, which complicated their mechanical testing. Further
tests would be required to obtain more confidence in
the mean responses. Moreover, it was found that par-
tially cured samples were stiffer at low temperatures
than fully cured samples, while an opposite tendency
would be expected. The chemico-physical reasons for
this behavior should be investigated further.
Since for the identification of the proposed viscoelas-
tic model only unidirectional experimental data were
available, its formulation in this article is limited to
1D. The model can easily be extended into 3D, but
for a proper formulation of the 3D behavior, multi-
axial experimental data would be needed, in particular
on the evolution of the Poisson ratio during relaxation.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.
Funding
The author(s) disclosed receipt of the following financial sup-
port for the research, authorship, and/or publication of this
article: This work was funded by the Research Chair held by
Pr. Ruiz, gathering the Natural Sciences and Engineering
Research Council of Canada (NSERC) and Safran group.
ORCID iD
Alice Courtois http://orcid.org/0000-0002-3669-8602
References
1. Bogetti TA and Gillespie JW. Process-induced stress and
deformation in thick-section thermoset composite lamin-
ates. J Compos Mater 1992; 26: 626–660.
2. Ruiz E and Trochu F. Multi-criteria thermal optimization
in liquid composite molding to reduce processing stresses
and cycle time. Compos Part A 2006; 37: 913–924.
3. Adolf D and Chambers R. Verification of the capability
for quantitative stress prediction during epoxy cure.
Polymer 1997; 38: 5481–5490.
4. Simon SL, Mckenna GB and Sindt O. Modeling the evo-
lution of the dynamic mechanical properties of a commer-
cial epoxy during cure after gelation. J Appl Polym Sci
2000; 76: 495–508.
5. Khoun L, Centea T and Hubert P. Characterization meth-
odology of thermoset resins for the processing of compos-
ite materials-case study: Cycom 890rtm epoxy resin.
J Compos Mater 2010; 44: 1397–1415.
Journal of Composite Materials 0(0)
6. Msallem YA, Jacquemin F, Boyard N, et al. Material
characterization and residual stresses simulation during
the manufacturing process of epoxy matrix composites.
Compos Part A 2010; 41: 108–115.
7. Khoun L and Hubert P. Cure shrinkage characterization
of an epoxy resin system by two in situ measurement
methods. Polym Compos 2010; 31: 1603–1610.
8. Ruiz E and Trochu F. Thermomechanical properties
during cure of glass-polyester RTM composites: elastic
and viscoelastic modeling. J Compos Mater 2005; 39:
881–916.
9. Khoun L and Hubert P. Processing characterization of a
RTM carbon epoxy system for aeronautical applications.
In: CANCOM, Winnipeg, 14–17 August 2007.
10. Zarrelli M, Skordos A and Partridge I. Investigation of
cure induced shrinkage in unreinforced epoxy resin. Plast
Rubber Compos 2002; 31: 377–384.
11. Prasatya P, McKenna GB and Simon SL. A viscoelastic
model for predicting isotropic residual stresses in thermo-
setting materials: effects of processing parameters.
J Compos Mater 2001; 35: 826–848.
12. Kim YK and White SR. Stress relaxation behavior of
3501-6 epoxy resin during cure. Polym Eng Sci 1996;
36: 2852–2862.
13. DiBenedetto A. Prediction of the glass transition tem-
perature of polymers: a model based on the principle of
corresponding states. J Polym Sci Part B 1987; 25:
1949–1969.
14. O’Brien DJ, Mather PT and White SR. Viscoelastic prop-
erties of an epoxy resin during cure. J Compos Mater
2001; 35: 883–904.
15. Caruthers JM, Adolf DB, Chambers RS, et al. A thermo-
dynamically consistent, nonlinear viscoelastic approach
for modeling glassy polymers. Polymer 2004; 45:
4577–4597.
16. Adolf DB and Chambers RS. A thermodynamically con-
sistent, nonlinear viscoelastic approach for modeling
thermosets during cure. J Rheol 2007; 51: 23–50.
17. Lévesque M, Derrien K, Baptiste D, et al. On the devel-
opment and parameter identification of schapery-type
constitutive theories. Mech Time Depend Mater 2008;
12: 95–127.
18. Biot M. Theory of stress-strain relations in anisotropic
viscoelasticity and relaxation phenomena. J Appl Phys
1954; 25: 1385–1391.
19. Schapery RA. Application of thermodynamics to ther-
momechanical, fracture, and birefringent phenomena in
viscoelastic media. J Appl Phys 1964; 35: 1451–1465.
20. Crochon T, Schönherr T, Li C, et al. On finite-element
implementation strategies of schapery-type constitutive
theories. Mech Time Depend Mater 2010; 14: 359–387.
21. Zarrelli M, Skordos AA and Partridge IK. Toward a
constitutive model for cure-dependent modulus of a
high temperature epoxy during the cure. Eur Polym J
2010; 46: 1705–1712.
22. Mahnken R. Thermodynamic consistent modeling of
polymer curing coupled to visco–elasticity at large
strains. Int J Solids Struct 2013; 50: 2003–2021.
23. Adolf DB, Chambers RS and Caruthers JM. Extensive
validation of a thermodynamically consistent, nonlinear
Courtois et al.
viscoelastic model for glassy polymers. Polymer 2004; 45:
4599–4621.
24. Schapery R. On a thermodynamic constitutive theory and
its application to various nonlinear materials. In: Boley
BA (ed) Thermoinelasticity. Vienna: Springer, 1970,
pp.259–285.
25. Crochon T. Modeling of the viscoelastic behavior of a
polyimide matrix at elevated temperature. PhD Thesis,
École Polytechnique de Montréal, Montreal, Canada,
2014.
26. Luk-Cyr J, Crochon T, Li C, et al. Interconversion of
linearly viscoelastic material functions expressed as
prony series: a closure. Mech Time Depend Mater 2013;
17: 53–82.
27. Schapery RA. On the characterization of nonlinear visco-
elastic materials. Polym Eng Sci 1969; 9: 295–310.
28. Ferry JD. Viscoelastic properties of polymers. New-York:
John Wiley & Sons, 1980.
29. Crowson RJ and Arridge RG. Linear viscoelastic proper-
ties of epoxy resin polymers in dilatation and shear in the
glass transition region. 1. Time-temperature superpos-
ition of creep data. Polymer 1979; 20: 737–746.
30. Sadeghinia M, Jansen K and Ernst L. Characterization
and modeling the thermo-mechanical cure-dependent
17
properties of epoxy molding compound. Int J Adhes
Adhes 2012; 32: 82–88.
31. Hojjati M, Johnston A, Hoa SV, et al. Viscoelastic behav-
ior of Cytec FM73 adhesive during cure. J Appl Polym
Sci 2004; 91: 2548–2557.
32. Miyano Y, Nakada M and Cai H. Characterization of
time-temperature dependent static and fatigue behavior
of unidirectional CFRP. In: 16th international conference
on composite materials, Kyoto, Japan, 8–13 July 2007.
33. Williams ML, Landel RF and Ferry JD. The temperature
dependence of relaxation mechanisms in amorphous
polymers and other glass-forming liquids. J Am Chem
Soc 1955; 77: 3701–3707.
34. Taylor RL, Pister KS and Goudreau GL.
Thermomechanical analysis of viscoelastic solids. Int J
Numer Meth Eng 1970; 2: 45–59.
35. Muliana A, Rajagopal K, Tscharnuter D, et al. A non-
linear viscoelastic constitutive model for polymeric solids
based on multiple natural configuration theory. Int J
Solids Struct 2016; 100: 95–110.
36. Machado M, Cakmak UD, Kallai I, et al.
Thermomechanical viscoelastic analysis of woven-rein-
forced thermoplastic-matrix composites. Compos Struct
2016; 157: 256–264.