Artigo Interessante

polymers
Article
Accurate Cure Modeling for Isothermal Processing of
Fast Curing Epoxy Resins
Alexander Bernath 1, *, Luise Kärger 1 and Frank Henning 1,2
1 Karlsruhe Institute of Technology, Institute of Vehicle System Technology,
Chair for Lightweight Technology, Rintheimer Querallee 2, 76131 Karlsruhe, Germany;
[email protected] (L.K.); [email protected] (F.H.)
2 Fraunhofer Institute for Chemical Technology, Joseph-von-Fraunhofer Str. 7, 76327 Pfinztal, Germany
* Correspondence: [email protected]; Tel.: +49-721-608-45895
Academic Editor: Alexander Böker

Received: 9 September 2016; Accepted: 28 October 2016; Published: 3 November 2016
Abstract: In this work a holistic approach for the characterization and mathematical modeling of
the reaction kinetics of a fast epoxy resin is shown. Major composite manufacturing processes
like resin transfer molding involve isothermal curing at temperatures far below the ultimate glass
transition temperature. Hence, premature vitrification occurs during curing and consequently has
to be taken into account by the kinetic model. In order to show the benefit of using a complex
kinetic model, the Kamal-Malkin kinetic model is compared to the Grindling kinetic model in
terms of prediction quality for isothermal processing. From the selected models, only the Grindling
kinetic is capable of taking into account vitrification. Non-isothermal, isothermal and combined
differential scanning calorimetry (DSC) measurements are conducted and processed for subsequent
use for model parametrization. In order to demonstrate which DSC measurements are vital for proper
cure modeling, both models are fitted to varying sets of measurements. Special attention is given to the
evaluation of isothermal DSC measurements which are subject to deviations arising from unrecorded
cross-linking prior to the beginning of the measurement as well as from physical aging effects. It is
found that isothermal measurements are vital for accurate modeling of isothermal cure and cannot
be neglected. Accurate cure predictions are achieved using the Grindling kinetic model.
Keywords: reaction kinetics modeling; differential scanning calorimetry; fast curing resin; epoxy
resin; resin transfer molding
1. Introduction
Processing of highly reactive polymer materials is nowadays a key technology to achieve low
cycle times and subsequently decrease manufacturing costs of both structural and non-structural parts.
However, as processes and part geometries get more and more complex, the commonly adopted
practice, involving many design iterations until a satisfactory result is obtained, becomes far too
expensive. Therefore, it is preferable to experimentally characterize the reaction kinetics to enable
model predictions and eventually optimize the process with the help of numerical simulation methods.
The latter is even more important when fiber reinforced materials are used since the anisotropic
behavior and the strong dependency of material properties on cross-linking makes the resulting part
geometry and tolerances difficult if not impossible to predict.
Relevant effects which have to be considered are thermal expansion or contraction, chemical
shrinkage and the evolution of material parameters as cross-linking advances. Furthermore, the actual
material behavior is also dependent on cure temperature Tc and glass transition temperature Tg .

The ratio of these two temperatures determines whether the material is in glassy Tc ≤ Tg or rubbery

Tc > Tg state. As Tg strongly depends on the degree of cure [1], it is very important to track
Polymers 2016, 8, 390; doi:10.3390/polym8110390 www.mdpi.com/journal/polymers

Polymers 2016, 8, 390 2 of 19
this material property throughout the process. Otherwise it is not possible to determine the correct
material state and consequently inappropriate material properties will be used by the simulation.
This is especially important in processes with isothermal temperature programs, e.g., resin transfer
molding (RTM), where the cure temperature is significantly lower than the ultimate glass transition
temperature Tg,∞ . This results in premature vitrification before the maximum degree of cure αmax has
been reached. In this case, the reaction has to take place in the glassy state where cross-linking either
advances significantly slower than in rubbery state or, as observed in many cases, stops as the reaction
rate completely vanishes [2–4]. Furthermore, important material properties undergo substantial
changes as the resin vitrifies. Lange et al. [4] reported a change in thermal expansion and stiffness
of an epoxy resin by as much as 51% and 260% respectively. In addition, O’Brien et al. [5] found
relaxation to be very sensitive to the degree of cure, especially near gelation. This has huge impact on
the development and relaxation of residual stresses during curing and cooling down which in turn
affects the amount of warpage and the final shape of composite parts. Consequently, a corresponding
process simulation method which aims at predicting this kind of manufacturing influences needs to
take vitrification effects into account.
Various studies have been carried out using rather slow resin systems which have their primary
field of application in small volume manufacturing as it is common e.g., in aerospace industry.
These systems are much easier to analyze compared to highly-reactive resins which are commonly
used by the automotive industry to satisfy requirements of large-batch production, particularly
short cycle times. However, characterization of fast resins is more prone to errors especially under
relevant isothermal conditions as these systems typically require rather high curing temperatures.
This challenges common analyzing methods since the time lag at the beginning of a measurement
needed for heating up to the desired temperature and thermal equalization may mask significant
portions of the already ongoing cure reaction which is not recorded by the differential scanning
calorimetry DSC analyzer [6]. As a result, some authors avoid isothermal measurements and
concentrate solely on non-isothermal data. Spoelstra et al. [7] investigate the chemorheology of
a filled epoxy system and declare isothermal DSC measurements to be inadequate in case of
fast curing thermosets. Furthermore, they claim to be able to predict reaction kinetics for any
temperature-history using their fitted model. Since the post-gelation region is not in their scope,
this statement holds true but it does not necessarily in case of isothermal processes where premature
vitrification cannot be neglected. Hsieh and Su [8] achieve accurate model predictions for isothermal
curing using solely non-isothermal DSC data as long as vitrification does not occur. The same is
reported by González-Romero and Casillas [3] where prediction accuracy impairs dramatically in
case of vitrification. Moreover, they found isothermal measurements to be more usable for parameter
estimation compared to non-isothermal although a higher number of measurements is required.
Atarsia and Boukhili [9] convert non-isothermal cure data to isothermal and vice versa for the purpose
of modeling the cure under non-uniform temperature profiles using an isoconversional representation.
Again, the agreement between isothermal DSC and corresponding model predictions using only
non-isothermal data is good as long as no vitrification occurs.
The above discussion shows that isothermal measurements are vital for proper modeling of
isothermal processes which are affected by vitrification. Therefore, this study aims at demonstrating
which experiments are needed for accurate modeling. Special attention is given to the evaluation of
isothermal DSC measurements which suffer from unrecorded cross-linking at the beginning of the
measurement. The latter was recently addressed by Javdanitehran et al. [10] who developed an iterative
approach for characterization of isothermal curing kinetics. However, they focused on modeling
cross-linking only until the reaction rate became negligibly small due to vitrification. Furthermore,
although their predicted isothermal cure degree in general agrees well with experimental data, the slope
towards the end of the cure curves is not reproduced accurately. As a consequence, incorrect cure
degrees will be obtained for longer periods of time which is a major drawback when utilizing
such predictions for cure cycle optimization. In the present study, this issue is solved by using
Polymers 2016, 8, 390 3 of 19
the complete and uncut time ranges of the isothermal measurements in order to force the applied
kinetic model to accurately reproduce diffusion-controlled reaction kinetics even for lengthy periods
of time. However, achieving this goal is not only a question of having suitable measurements at hand.
Choosing appropriate mathematical models is also an important requirement, though the present
study is not intended to give a comprehensive benchmark on the multitude of available kinetic models.
The comparison is therefore limited to two kinetic models: the Kamal-Malkin kinetic model [11],
which is often applied in commercial process simulation software and the more advanced and complex
Grindling kinetic model [12]. The latter is able to take into account premature vitrification which is
crucial for accurate modeling of isothermal processes like RTM.
2. Experiments
2.1. Material
A commercial epoxy resin system typically used in resin injection processes was chosen for
this study. It consists of two components: resin Biresin CR170 (Bisphenol-A-Epichlorohydrin) and
hardener CH150-3 (3-Aminomethyl-3,5,5-trimethylcyclohexylamin and 2-Piperazin-1-ylethylamin),
kindly provided by Sika Deutschland GmbH (Stuttgart, Germany). The recommended processing
temperatures for this system range between 60 and 120 ◦ C. Batches of 10 g were mixed using a resin to
hardener mixing ration of 100:24 (by weight). From these, the DSC specimens were obtained.
2.2. Characterization
Differential scanning calorimetry (DSC) is by far the most used method for cure kinetics
characterization and thus is also used in the present study. DSC analyzers capture the exothermic
heat flow which is released during the cross-linking process of a reactive polymer in respect of curing
time under isothermal or non-isothermal conditions. Moreover, since the heat capacity of a polymer is
sensitive to temperature and shows significant change during glass transition, the corresponding glass
transition temperature Tg can be measured. All DSC runs are carried out in nitrogen atmosphere and
by using a DSC Q200 analyzer from TA Instruments (New Castle, DE, USA).
As has been implied above, non-isothermal DSC measurements are typically easier to perform
because the reaction rate is virtually zero in the beginning if a sufficiently low start temperature is used.
Therefore, the start temperature was set to −50 ◦ C for all non-isothermal DSC runs. This procedure
of course is not applicable to isothermal measurements which therefore bear the risk of losing
the first seconds or even minutes of cross-linking. Different strategies exist for minimization of
this uncertainty. Slowly curing materials may be heated within the DSC furnace to the desired
curing temperature. However, if the same is applied to fast curing resins, this can lead to a significant
amount of cross-linking already before data acquisition starts since maximal controlled temperature
ramps of typical DSC devices max out at 60 to 120 ◦ C/min. Alternatively, the specimen can be put
into a preheated DSC furnace. In addition, the specimen may also be preheated prior to insertion into
the DSC cell in order to reduce the effective temperature difference. Obviously, this leads to the same
problem as before in case of fast curing resins. Therefore, in this study, the DSC cell was preheated to
the respective isothermal temperature, but the specimens, which were obtained from a freshly mixed
charge at room temperature, were not. In combination with small sample sizes and therefore fast
temperature equalization between DSC cell and specimen, premature cross-linking is avoided as much
as possible. In this study, specimens of about 7 and 10 mg were used. Even though, the assumption of
zero degree of cure at the beginning of isothermal measurements at elevated temperatures is arguable.
2.2.1. Reaction Kinetics

Various measurements were conducted in order to obtain a sufficient amount of data points for
parametrization of the selected kinetic models. Both, isothermal and non-isothermal temperature
programs are investigated in this study. The corresponding temperatures and heat rates are 60, 80, 100
Polymers 2016, 8, 390 4 of 19
and 120 ◦ C and 1, 2.5, 5, 10 and 15 ◦ C/min respectively. Results from the non-isothermal measurements
are used for determination of the maximum total heat of reaction of the resin [13]. This value is a
key quantity on which much of the later shown experimental analysis relies. It is derived by at first
constructing a baseline for the reaction peak as is depicted in Figure 1.
0.7
0.7 1.0
1.0 0.6
0.6
55K/min
K/min(DSC)
(DSC) 55K/min
K/min(DSC)
(DSC) 500
500
0.6
0.6
0.5
0.5 0.8
0.8 0.4
0.4
Specific heat flow q (W/g)
400
400
Reaction enthalpy h (J/g)

0.4
0.4
Cure degree α (-)
Cure rate (1/s)

0.3
0.3 0.6
0.6 300
300 0.2
0.2
0.2
0.2
0.4
0.4 200
200
0.1
0.1 0.0
0.0
TTg,g,00
0.0
0.0 0.2
0.2 100
100
0.1
0.1 baseline
baseline 0.2
0.2
0.2
0.2 0.0
0.0 00
50
50 00 50
50 100
100 150 150 200
200 00 50
50 100
100 150
150 200
200 00 50
50
Temperature
Temperature ((◦◦C)
C) Temperature
Temperature ((◦◦C)
C)
(a) (b)
Figure 1. Evaluation of DSC data (exemplary for non-isothermal curing). (a) Specific heat flow;
(b) Cure degree.
The area enclosed by the specific heat flow curve q(t) and the baseline corresponds to the specific
reaction enthalpy h (t):
Z t
h (t) = q t̂ dt̂. (1)
t0
In this equation, t0 denotes the time at which the reaction starts. Together with the specific total
heat of reaction ∆h, the temporal evolution of the cure degree α (t) is given by
Z t
1 1
α (t) = · h (t) = · q t̂ dt̂. (2)
∆h ∆h t0
The resulting curves for isothermal and non-isothermal curing are displayed in Figures 2 and 3,
respectively. The evolution of the cure degree α starts at 0 which represents the material in uncured state.
During curing, it can reach a maximum value of 100% or 1, representing fully cured resin.
1.0
1.0 66 1.0
1.0
60
60 ◦◦CC(DSC)
(DSC)
55 80
80 ◦◦CC(DSC)
(DSC)
0.8
0.8 100
100 ◦◦CC(DSC)
(DSC) 0.8
0.8
Cure rate α̇ x10−−33 (1/s)
120
120 ◦◦CC(DSC)
(DSC)
44
Cure degree α (-)
Cure degree α (-)
0.6
0.6 0.6
0.6
33
0.4
0.4 0.4
0.4
22
60
60 ◦◦CC(DSC)
(DSC)
0.2
0.2 80
80 ◦◦CC(DSC)
(DSC) 11 0.2
0.2
100
100 ◦◦CC(DSC)
(DSC)
120
120 ◦◦CC(DSC)
(DSC)
0.0
0.0 00 0.0
0.0
00 20
20 40
40 60
60 80
80 100
100 00 20
20 40
40 60
60 80
80 100
100 00 20
20
Time
Time (min)
(min) Time
Time (min)
(min)
0.9
0.9 (a) 80
80 (b) 160
160
s transition temperature ( ◦◦ C)
0.8
0.8 140
140
60
60
Figure 2. Data for isothermal curing at different temperatures derived from DSC measurements.
0.7
0.7
(a) Cure degree; (b) cure rate.
120
120
Cure degree α (-)
0.6
0.6 40
40 100
100
T-Tg ( ◦◦ C)
0.5
0.5 80
80
20
20
0.4
0.4 60
60
0.3
0.3 00 40
40
0.2
0.2 20
20
Polymers 2016, 8, 390 5 of 19
1.01.0 5 1e
5 1e3 3 1.01.0
1 K/min
1 K/min
(DSC)
(DSC)
2.52.5
K/min
K/min
(DSC)
(DSC)
0.80.8 44 5 K/min
5 K/min
(DSC)
(DSC) 0.80.8
1010
K/min
K/min
(DSC)
(DSC)
d2cure/dt2 (1/s^2)
d2cure/dt2 (1/s^2)
Cure degree α (-)
Cure degree α (-)
Cure rate α̇ (1/s)

Cure rate α̇ (1/s)
1515
K/min
K/min
(DSC)
(DSC)
0.60.6 33 0.60.6
0.40.4 22 0.40.4
1 K/min
1 K/min
(DSC)
(DSC)
2.52.5
K/min
K/min
(DSC)
(DSC)
0.20.2 5 K/min
5 K/min
(DSC)
(DSC) 11 0.20.2
1010
K/min
K/min
(DSC)
(DSC)
1515
K/min
K/min
(DSC)
(DSC)
0.00.0 00 0.00.0
00 5050 100
100 150 150 200 200 250
250 00 5050 100
100 150 150 200 200 250
250 00 5050
Temperature
Temperature( ◦ C)
( ◦ C) Temperature
Temperature( ◦ C)
( ◦ C) T
1.01.0 (a) 140
140 (b) 200
200
Glass transition temperature ( ◦◦ C)

Glass transition temperature ( C)
120
120
0.80.8 3. Data for non-isothermal curing at different
Figure 100
100 temperature ramps derived from DSC 150
150
d3cure/dt3 (1/s^3)
d3cure/dt3 (1/s^3)
measurements. (a) Cure degree; (b) cure rate.

8080
0.60.6 100
100
T-Tg ( ◦◦ C)
T-Tg ( C)
6060
The aforementioned difficulty related to isothermal measurements is visible in Figure 2a. Usually,
4040
0.40.4
higher temperatures lead to higher cure degrees since Tg can reach greater values before vitrification 5050
2020
occurs and the reaction halts. However, the isothermal DSC measurements in Figure 2a do not show
00
such0.2a0.2
behavior. Instead a decrease in final cure degree is noticeable. This is in contradiction to 0 0
2020
the findings of Section 5.1 which prove higher achievable cure degrees for increasing isothermal
0.00.0 4040curing temperatures even for long periods 5050
00
temperatures 5050well 100
as as100 150
150 cure
incomplete 200
200in case
250
250
of lower00 5050 100 100 150 150 200 200 250 250 00 5050
of time. This error Temperature
isTemperature
( C)
therefore (attributed
◦ ◦
C) Temperature
to the amount of cure which Temperature( ◦ C)
is lost ( during
◦
C) specimen T
preparation and measurement initialization. Additionally, as can be seen in Figure 2b, the cure rate
strongly depends on temperature and, more importantly, reaches maximal values at the very beginning
of the measurement. Therefore, even small periods of time without data acquisition in this region can
introduce significant deviations in cure degree. This circumstance has to be addressed during model
parametrization as otherwise poor prediction accuracy is to be expected.
The slope of isothermal cure curves vanishes for long periods of time as is visible in Figure 2a
for all investigated temperatures. The reason for this is the influence of premature vitrification and
the accompanied transition to diffusion-controlled reaction kinetics. Since this phenomenon depends
mainly on the interaction of cure and glass transition temperature, this behavior is also observed
for very low heating rates in non-isothermal DSC (cf. Figure 3a). In this case, the glass transition
temperature increases faster as the cure temperature and finally outruns it, causing a slow down
in reaction rate. As the cure temperature is further increased during non-isothermal DSC, the glass
transition temperature also continues to rise and consequently prohibits the reaction from switching
back to a chemical-controlled state. The effect is indicated by a flattening and almost constant slope of
the curve in the affected cure range. For the studied epoxy resin, this happens during cure at the lowest
heating rate of 1 ◦ C/min and for cure degrees above approximately 88%. However, the influence is
much less pronounced. In particular, it does not show a complete stop in cross-linking as is shown by
isothermal measurements.
The cure rate of non-isothermal experiments shows a dependency on temperature, as is visible
in Figure 3b. However, in case of small heat ramps, most of the reaction takes place at rather low
temperatures since the dwell time in this region is long enough to compensate for low temperatures.
On the contrary, high heat ramps shift the majority of the reaction towards higher temperatures since
these are reached in shorter periods of time. Corresponding curves of the cure degree confirm this
behavior as can be seen in Figure 3a.
Polymers 2016, 8, 390 6 of 19
2.2.2. Glass Transition Temperature

While characterization of reaction kinetics usually involves isothermal and non-isothermal
DSC runs, measurement of Tg at particular cure levels is conducted by combining both temperature
programs. This procedure generally can be used instead of normal isothermal measurements
as was previously shown by Fava [14]. The specimen firstly is hold isothermal using a specific
time-temperature-combination. Afterwards, it is quickly cooled down far below the isothermal
temperature (typically below 0 ◦ C) in order to fully stop cross-linking. Following this, it is heated
up to a high temperature using a constant temperature ramp. Because of this temperature program,
this type of measurement is referred to as cyclic DSC for simplicity in this study.
Determination of Tg from DSC heating scans is generally not recommended since Tg represents the
temperature of transition from liquid/rubber to glass instead of the reverse [15]. However, DSC devices
are usually calibrated on heating and, more importantly, the state of the material is inevitably altered by
first heating the specimen to a temperature markedly above Tg and then cooling down for the purpose
of actual Tg measurement. The reason for incorrect values gathered from heating experiments arises
from commencing the glass transition from a possibly non-equilibrium state. This depends largely on
the thermal history which was previously applied to the specimen. Therefore, reasonable values can
be obtained by starting the heating scan immediately after cooling and by ensuring that curing prior
to cooling down is carried out at a temperature Tc > Tg such that no vitrification occurs [15,16].
Table 1 contains combinations of temperature and hold time used in this study to reach specific
cure states. Afterwards, cooling of the specimen is carried out at the highest rate possible by the
DSC analyzer (about 40 ◦ C/min). Thereby, further cross-linking during cooling down is minimized.
Subsequent heating is realized by using a constant rate of 5 ◦ C/min for all cyclic measurements.
The results are analyzed the same way as has been described above for non-isothermal DSC. However,
since the cyclic measurement starts with an isothermal temperature program, it is subject to the same
problems and errors like pure isothermal DSC runs. Therefore, the degree of cure is calculated reversely
by evaluating the reaction peak of the non-isothermal part. Because of the preceding isothermal stage,
the area of the non-isothermal reaction peak is generally smaller for cyclic DSC. This remaining heat of
reaction is referred to as ∆hR . With this, the degree of cure reached at the end of the isothermal phase
is calculated by
∆h − ∆hR ∆h
α= = 1− R. (3)
∆h ∆h
Table 1. Parameters of cyclic DSC runs.
Curing temperature Tc (◦ C) Hold times th (min)

60 15, 45, 120 *
70 10, 30, 80 *
80 5, 10, 15, 30 *
100 2, 7.5, 20 *
120 1, 5, 10 *
* Vitrification was observed during isothermal curing.
3. Mathematical Modeling
Choosing an appropriate mathematical model for glass transition temperature as well as reaction
kinetics is essential for getting meaningful results and predictions e.g., for process optimization.
However, the field of application needs to be considered since the area of polymer processing includes
a wide range of different temperature boundary conditions. This study concentrates on using the model
for simulation of the resin transfer molding process which typically shows premature vitrification due
to isothermal processing at rather low temperatures, as has been mentioned before.
Polymers 2016, 8, 390 7 of 19
In this section, model equations and corresponding fitting parameters are presented. Identification
of the latter is described in Section 4.
3.1. Glass Transition Temperature

Vitrification of the resin starts once the glass transition temperature, which persistently rises
during cross-linking, comes close to the current process temperature. Therefore, it is vital to have
precise knowledge of the evolution of Tg during the reaction. A widely used model, which is originally
based on work of DiBenedetto [1,17], is given by
Tg − Tg,0 λα
= . (4)
Tg,∞ − Tg,0 1 − (1 − λ ) α
The glass transition temperature of the material in the uncured and fully cured state, Tg,0 and Tg,∞
respectively, can be determined experimentally through DSC. Adopting the curve to experimental
results is therefore done by adjusting the fitting parameter λ between 0 and 1.
3.2. Reaction Kinetics

Many different models for reaction kinetics exist. A comprehensive overview is given by Halley
and Mackay [13]. From this variety, two models were selected which differ greatly in complexity.
3.2.1. Kamal-Malkin Kinetic Model

As a simple autocatalytic model, the Kamal-Malkin kinetic model [11] is used. It is often applied
and implemented in commercial process simulation software because of its simplicity and the relatively
small number of model parameters which facilitates parametrization. The cure rate α̇ is defined by
dα
α̇ = = ( K1 + K2 · α m ) · ( 1 − α ) n (5)
dt
where reaction rate constants K1 and K2 are given by Arrhenius-type equations

E
K1 = A1 · exp − 1 (6)
R·T

E2
K2 = A2 · exp − . (7)
R·T
Besides the universal gas constant R and the temperature T, this kinetic model contains six fitting
parameters: constants m and n, pre-exponential factors A1 and A2 as well as two activation energies
denoted by E1 and E2 . None of the model equations is able to account for vitrification effects.
In combination with the term (1 − α) on the right hand side of Equation (5), this model presumes
the material to fully cure no matter what temperature is applied. Especially when using lower
temperatures, this can lead to a significant loss in accuracy of the model prediction since the achievable
degree of cure of resins is temperature dependent. However, since the Kamal-Malkin kinetic model is
often used for process simulation, it is considered in this study in order to demonstrate the benefit of
using a more complex model.
3.2.2. Grindling Kinetic Model

In order to overcome the drawback of the Kamal-Malkin kinetic model, Grindling [12] extended
the model towards taking into account diffusion controlled reaction kinetics. It is therefore a suitable
choice for modeling isothermal molding processes and is included in this study in order to show the
benefit from using a more sophisticated model.
Polymers 2016, 8, 390 8 of 19
Since the model was only published in German language, it is presented here in more detail.
In a first step, Equation (5) is modified to
dα
= K1 · (1 − α)n1 + Keff · αm · (1 − α)n2 . (8)
dt
The effective reaction rate constant Keff enables the kinetic model to switch between a
chemical-controlled and a diffusion-controlled state. It is defined by
1 1 1
= + . (9)
Keff K2,diff K2
In these two equations, K1 and K2 represent reaction rate constants which are only limited by
chemistry. Therefore, both are modeled the same way as before (cf. Equations (6) and (7)). On the
contrary, the newly introduced reaction rate constant K2,diff represents a reaction rate that is limited by
the diffusion velocity of the reactants and has therefore a strong dependency on the glass transition

temperature Tg and its distance to the current reaction temperature T − Tg :
T > Tg + ∆Tg :

E2,diff 1 1
K2,diff = K2,diff T = Tg + ∆Tg · exp −

· − (10)
R T Tg + ∆Tg
Tg ≤ T ≤ Tg + ∆Tg :

!
T = Tg c1 · T − Tg
K2,diff = K2,diff · exp (11)
c2 + T − Tg
T < Tg :

T = Tg E1,diff 1 1
K2,diff = K2,diff · exp − · − (12)
R T Tg
Equations (10) and (12) follow an Arrhenius-type approach with pre-exponential factors K2,diff
T = Tg
and K2,diff , as well as activation energies E1,diff and E2,diff . Equation (11) is based on a modified
Williams-Landel-Ferry (WLF) approach and introduces two parameters, c1 and c2 , which enable
the model to adjust the smoothness of the transition between the rubbery and the vitrified state.
The activation energies E1,diff and E2,diff are calculated using both WLF-coefficients of Equation (11),
c1 and c2 :
E1,diff c
= Tg2 · 1 (13)
R c2
2
E2,diff c1 · c2 · Tg + ∆Tg
= 2 (14)
R c2 + ∆Tg
This kinetic model is considerably more complex than the previously shown Kamal-Malkin model.
T = Tg
The eleven fitting parameters are: A1 , E1 , n1 , A2 , E2 , n2 , m, c1 , c2 , K2,diff and ∆Tg . In its original
implementation [12], a fixed value of 100 ◦ C was used in place of the parameter ∆Tg without
detailed explanation. However, ∆Tg has proven to be an important model parameter and therefore
was added to the fitting algorithm in the present study.
4. Parameter Identification
Identifying appropriate model parameters is crucial for precise kinetic modeling. However,
due to the phenomenological character of the used models, it is difficult to estimate even the order of
magnitude of the fitting parameters beforehand. As a consequence, the resulting search space to cover
Polymers 2016, 8, 390 9 of 19
is very large. Therefore, the evolutionary algorithm covariance matrix adaptation evolution strategy
(CMA-ES) introduced by Hansen and Ostermeier [18] is utilized in this study.
The curve fitting algorithm attempts to reproduce the measured values by using the chosen model.
As is shown in the next section, the prediction accuracy of the fitted model depends on the type of
experimental DSC data handed over to the algorithm as target values. The DSC results are categorized
with respect to the type of measurement: non-isothermal, isothermal and cyclic. In order to evaluate
which type of measurement is beneficial for obtaining good fitting results, parameter estimation is
performed by using either solely one of these measurement categories or a combination of them.
The aforementioned problem of non-zero cure degree at the beginning of DSC data acquisition
in case of isothermal measurements is counteracted by calculating the initial amount of cross-linking
iteratively by using the current model parameter set. It is assumed that the missing part of the reaction
can be reconstructed during parameter estimation using the kinetic model. Considering that in this
case the initial cure value of the isothermal curves from both, experiment and model, can vary widely
during fitting, it is necessary to include another measurement type which contributes fixed values
of cure degree, like cyclic or non-isothermal DSC. In other words, it is not advisable to fit a model
by solely using isothermal data since all curves will level out at full cure (α = 1) after fitting due to
the variability in initial cure. Otherwise, fitting to solely isothermal data under negligence of α > 0
at the beginning of the measurement is neither reasonable since then the model needs to handle the
non-physical situation of lower final cure in spite of higher isothermal curing temperature as is visible
in Figure 2.
5. Results and Discussion
5.1. Glass Transition Temperature

The glass transition temperature Tg of the resin at intermediate cure states is evaluated from cyclic
DSC measurements as is outlined in Section 2.2.2. For calculation of the corresponding cure degrees,
the total heat of reaction ∆h is determined from non-isothermal measurements. A mean value of
∆h = 499.8 ± 16.6 J/g is found for the studied epoxy resin. Moreover, a mean glass transition
temperature of the fully uncured resin of Tg,0 = −30.64 ◦ C is determined from non-isothermal DSC
runs as it is indicated in Figure 1b. For the fully cured resin, a glass transition of Tg,∞ = 135.74 ◦ C is
measured by immediately reheating a specimen which was previously cured to α = 100%. During this
second heating, Tg,∞ becomes visible.
Figure 4 shows the dependency of Tg on cure degree as well as the fitted DiBenedetto Tg -model
which is in good agreement with experiments. The corresponding value for the model parameter of
Equation (4) is λ = 0.3895. The experimental values are determined by creating the midpoint between
the glassy and the liquid/rubbery lines and identifying the corresponding temperature. Because of the
different hold times and curing temperatures, the achieved cure degrees differ.
From an experimental point of view, using only a single low isothermal temperature together
with a large number of different hold times for characterization of Tg (α) would be favorable since high
temperatures, which introduce errors in the first seconds, are avoided. However, because of premature
vitrification at low temperatures, this is not possible. Moreover, the material exhibits physical aging
effects when it is hold for long time at a specific temperature after reaching the maximal degree of
cure possible at that temperature [19]. These effects show up during the second stage of cyclic DSC
runs and can make correct determination of Tg and ∆hR difficult if not impossible. Consequently, it is
recommended to use Tg values determined from measurements which do not exhibit vitrification
(valid values) [1]. This is the case for all but the longest times at each temperature (corresponding
times are highlighted in Table 1).
Figure 5 shows the influence of physical aging on the evaluation of data from cyclic DSC
measurements. The specimen which was cured at 60 ◦ C for 15 min (Figure 5a) allows for good
distinction between the glass transition and the onset of the reaction peak. On the contrary, in case of the
Polymers 2016, 8, 390 10 of 19
specimen which was cured at the same temperature but was hold for 120 min (Figure 5b), these events
are superimposed by an endothermic peak which araises from physical aging. Construction of a
baseline and subsequent integration in order to determine α is no longer possible. However, since it is
an advantage for later cure model fitting to have values of α and the corresponding Tg also in vitrified
material state available, an alternative strategy is pursued. In a first step, a Tg -model is fitted using
only valid values of Tg and α. In this case, this is done using the model of DiBenedetto (cf. Section 3.1).
From this relationship, the cure degrees which cannot be determined from cyclic DSC due to physical
aging effects are derived reversely by using associated Tg values. The latter is possible since from the
mentioned three overlapping effects, glass transition is the best-preserved one. Thus determination
of Tg is possible (cf. Figure 5a), although the values might be affected by physical aging. However,
the error of α values gathered that way is compensated to some extent by the non-linear behavior
of Tg (α) towards high cure degrees (cf. Figure 4). Furthermore, because Tg gains in sensitivity with
respect to cross linking in this region, it is even more suited for monitoring cross-linking than α [20].
140
Tg -model
Glass transition temperature Tg ( ◦ C)
120 Tg,0
100 Tg,∞
80 60 ◦ C
70 ◦ C
60 80 ◦ C
40 100 ◦ C
120 ◦ C
20
0
20 DiBenedetto (λ = 0.3895)
40
0.0 0.2 0.4 0.6 0.8 1.0
Cure degree α (-)
Figure 4. Dependency of Tg on cure degree α: experimental values determined from cyclic DSC
according to Table 1, compared to the prediction of the fitted DiBenedetto model.
0.40.4 0.40.4
60 60
◦
C,◦ 15
C, 15
minmin 60 60
◦
C,◦ 120
C, 120
minmin
0.20.2 0.20.2

0.00.0 0.00.0
Tg T=g 6.33
= 6.33CC
◦ ◦
Tg ≈Tg 76.5
≈ 76.5CC
◦ ◦
0.20.2 R h=
∆h∆ R 308.5
= 308.5
J/g J/g 0.20.2
R h=
∆h∆ R ?= ?
0.40.4 0.40.4
endothermic
endothermic
peak
peak
0.60.6 0.60.6
0 0 5050 100100 150150 200200 0 0 5050 100100 150150 200200
Temperature
Temperature
( ◦ C)
( ◦ C) Temperature
Temperature
( ◦ C)
( ◦ C)
(a) (b)
Figure 5. Temperature scanning data after isothermal curing at 60 ◦ C for different periods of time:
influence of physical aging on data evaluation. (a) Cured at 60 ◦ C for 15 min; (b) cured at 60 ◦ C
for 120 min.
5.2. Reaction Kinetics

One major goal of this study is to find out which measurements need to be conducted in order
to achieve good prediction accuracy. In particular, the benefit of incorporating data from cyclic
Polymers 2016, 8, 390 11 of 19
DSC measurements, which are necessary anyhow for characterization of the relationship between the
glass transition temperature Tg and the cure degree α, is demonstrated by firstly discussing results
fitted without this data. Afterwards, the same is done under consideration of cyclic data in order to
compare both cases.
For reasons of clarity and comprehensibility, the layout of the presentation of fitting results,
consisting of three graphs placed side by side, is kept constant throughout the discussion. On the
left graph, the predicted isothermal curing is plotted against time. The plot in the middle shows
the model-prediction for non-isothermal temperature profiles and on the right, the consistence with
cyclic DSC measurements is evaluated. The latter plot contains additional black lines which show the
discrepancy between the prediction and the experimental data point.
While each figure contains plots for all three measurement types, the types that were actually
used for curve fitting vary (cf. Section 4). This way, it is possible to visualize how good a kinetic model
is able to reproduce each of the measurement types although not all of them may have been actually
considered during fitting. However, one exception is formed by isothermal DSC data. Since the initial
degree of cure cannot be predicted when isothermal measurements are not used for fitting of the model,
it is not feasible to plot all of the corresponding experimental curves in such a way that it is possible to
evaluate the prediction quality. Hence, only the experimental data of the lowest temperature (60 ◦ C) is
plotted because in this case, the amount of cure which is initially lost is insignificant.
5.2.1. Prediction Accuracy without Using Cyclic DSC Measurements

Figures 6 and 7 show the goodness of fit for both kinetic models in case only non-isothermal DSC
data is used. As can be seen from the non-isothermal plots, the Grindling model is able to reproduce
the experimental data very closely whereas the Kamal-Malkin model shows a significant deviation.
From the plot on the right (Figure 7c) it can be seen that the Grindling model shows good accuracy
also for the isothermal conditions during cyclic DSC. However, the slope at the end of the curves on
the isothermal plot (Figure 7a) is overestimated and does not exhibit slowing down of cross-linking
due to vitrification (cf. Figure 2).
1e
2.5 1e
2 2 1e 3
1.0 1.0 2.51.0 1.0 51.0 1.0
0.8 0.8 2.00.82.0 1 ◦ C/min DSC 0.8 40.8 0.8
d2cure/dt2 (1/s^2)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
2.5 ◦ C/min DSC

rate (1/s)
Cure rate (1/s)
Cure rate (1/s)
0.6 0.6 1.50.61.5 5 ◦ C/min DSC 0.6 30.6 0.6

Curedegree
10 ◦ C/min DSC
15 ◦ C/min DSC
0.4 0.4 60 ◦ C60DSC
◦
C DSC 1.00.41.0 1 ◦ C/min 0.4 20.4 0.4
Cure
60 ◦ C60 ◦ C 2.5 ◦ C/min

0.2 0.2 80 ◦ C80 ◦ C 0.50.20.5 5 ◦ C/min 0.2 10.2 0.2
100 ◦100
C ◦C 10 ◦ C/min DSC DSC
120 ◦120
C ◦C 15 ◦ C/min modelmodel
prediction
prediction
0.0 0.0 0.0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0.000.0 0 0 205020 4010040 6015060 ◦8020080 100
250100 0 0 0 20 502040 100
4060 6015080 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
(min)(min) Temperature
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1201.0
120 160
140160 200
Glass transition temperature ( ◦ C)
(1/s^3) ( ◦ C)
( ◦ C)
(a) 100 100 (b) 140

120140 (c)
temperature
Glass transition temperature
0.8 0.8 800.880 120

100120 150
Cure degree α (-)
Cure degree α (-)
Figure 6. Cure prediction of Kamal-Malkin model fitted solely to non-isothermal 10080100 data. (a) Isothermal;
0.6 0.6 600.660 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
(b) non-isothermal; (c) cyclic. 806080

40 40
d3cure/dt3
0.4 0.4 6040 60

200.420 50
Glass transition
402040
0.2In
0.2Section 4 it was concluded that 00.2 fitting
0 to only isothermal data20is020not advisable. Figures 8 and 9 0
20 20 020 0
show the result of the fitting algorithm
0.0 0.0 for this situation for Kamal-Malkin and Grindling kinetic model,50
-3 -3 -2 -2 -1 -1 0 0 1 1 2 2 40 0.0
0 40 250100 2040
0 0 205020 4010040 6015060 ◦8020080 100 0 0200 205020 4010040 6015060 ◦ 8020080 100
250100 0 50 100 150
10 10 10 10 10 10 10 10 10 10 10 10
respectively. Obviously,
TimeTime
(min)(min)prediction accuracy ofTemperature
TimeTime
both ( C)
(min)(min)
models is poor which Temperature
TimeTime
is especially (evident
(min)(min)C) from Temperature
the comparison with cyclic DSC data. The algorithm overpredicts the initial cure degree such that all
isothermal curves eventually end up at full cure (α = 1). As a consequence, the reaction never has to
take place under diffusion-controlled conditions. This leads to very similar cure predictions since the
Grindling model is an extension of the Kamal-Malkin model and behaves similar in the absence of
vitrification (cf. Section 3.2). Therefore, as has been mentioned before, using only isothermal data for
kinetic model fitting is not sufficient.
Polymers 2016, 8, 390 12 of 19
1e
1.0 1e
2 2 1e 3
1.0 1.0 1.01.0 1.0 51.0 1.0
0.8 0.8 0.80.80.8 1 ◦ C/min DSC 0.8 40.8 0.8
d2cure/dt2 (1/s^2)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
2.5 ◦ C/min DSC
rate (1/s)
Cure rate (1/s)
Cure rate (1/s)
0.6 0.6 0.60.60.6 5 ◦ C/min DSC 0.6 30.6 0.6
Curedegree
10 ◦ C/min DSC
15 ◦ C/min DSC
0.4 0.4 60 ◦ C60DSC
◦
C DSC 0.40.40.4 1 ◦ C/min 0.4 20.4 0.4
Cure
60 ◦ C60 ◦ C 2.5 ◦ C/min
0.2 0.2 80 ◦ C80 ◦ C 0.20.20.2 5 ◦ C/min 0.2 10.2 0.2
100 ◦100
120 ◦120
prediction
prediction
0.0 0.0 0.00.00.0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 60150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1401.0
140 200
140200 200

(1/s^3) ( ◦ C)
( ◦ C)
(a) 120 120 (b) 120 (c)
150 150
temperature
Cure degree α (-)Glass transition temperature

0.8 0.8 1000.8
100 100 150
Cure degree α (-)
Cure degree α (-)
Figure 7. Cure prediction of Grindling 80 80 model fitted solely to non-isothermal 10080100 data. (a) Isothermal;
0.6 0.6 600.660 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
50 50
d3cure/dt3
0.4 0.4 400.440 40 50

α (-)Glass transition
20 20 020 0
6
0 1e 6
0 1e
3 3 5 1e 3 1.0
1.0 0.2
1.0 1.0 1.0 0
0.2 0.2 50 1.0
50 0
205 205 20
0.8 0.0
0.0 0.8 0.840
400.0 4
0.8 100
0.8 0.8
100 10040 250100 500
1 C/min DSC ◦
d2cure/dt2 (1/s^2)
10-3 10-310-2 10-210-1 10-1 100 100 101 101 102 102 40 040 205020 4010040 6015060 ◦802002.5 80C/min
100
250 0 0 0 205020 4010040 6015060 ◦ 8020080 100 50 100 150
Cure degree α (-)
Cure degree α (-)
α (-)
DSC
rate (1/s)
rate (1/s)
Cure rate (1/s)
0.6 0.6 TimeTime

(min)(min) 60
◦
C
60DSC
◦
C DSC 0.6 Temperature
TimeTime
(min) (
(min)C)5 C/min DSC 0.6 30.6
◦ Temperature
TimeTime
(min) (
(min) C) 0.6 Temperature
Curedegree
Curedegree
80 C80DSCC DSC 10 C/min DSC
3 3
◦ ◦ ◦
100 100
◦
C DSCC DSC
◦
15 C/min DSC ◦
0.4 0.4 120 120

◦
C DSCC DSC
◦ 0.4 1 C/min 0.4 20.4
◦ 0.4
2 2
Cure
Cure
60 C60 C
◦ ◦
2.5 C/min ◦
0.2 0.2 80 C80 C

◦ ◦
10.2 1
5 C/min 0.2 10.2
◦
0.2
100 100
◦
C C ◦
10 C/min ◦
DSC DSC
120 ◦120
prediction
prediction
0.0 0.0 00.0 0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 60150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1401.0140 150
140150 200
(1/s^3) ( ◦ C)
( ◦ C)
1.0 1.0 (a) 120 120 (b) 120 (c)
temperature
100 0.8
100 100
100100 150
0.8Figure
0.8
Cure degree α (-)
Cure degree α (-)
8. Cure prediction of Kamal-Malkin model fitted solely to isothermal

80 data. (a) Isothermal;
800.680 506050 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
(b)
0.6 0.6 non-isothermal; (c) cyclic. 60 60
d3cure/dt3
40
0 0
400.440 50
0.4 0.4 20
1e 1e
2060.2 3 3 1e 3
1.0 1.0 1.0206 1.0 5
50 01.0
50 1.0
0
0.2 0.2 0 0 20
0.8 0.8 50.8 5 40.8
0.840100 0.8
200.0 20 1 C/min DSC ◦
100 250100 500
d2cure/dt2 (1/s^2)
100 100 101 101 102 102 40 040 205020 4010040 6015060 ◦802002.5 80C/min
100
250100 0 0 0 205020 4010040 6015060 ◦ 8020080 100 50 100 150
Cure degree α (-)
Cure degree α (-)
α (-)
DSC
rate (1/s)
rate (1/s)
Cure rate (1/s)
0.6 0.6 TimeTime

(min)(min) 60 C
60DSC
◦
C DSC ◦
0.6 Temperature
TimeTime
(min) (
(min)C)5 C/min DSC ◦
0.6 30.6 Temperature
TimeTime
(min) (
Curedegree
Curedegree

3 3
◦ ◦ ◦
100 100
C DSCC DSC
◦ ◦
15 C/min DSC ◦
0.4 0.4 120 120

C DSCC DSC
◦ ◦0.4 1 C/min ◦ 0.4 20.4 0.4
2 2
Cure
Cure
60 C60 C
◦ ◦
2.5 C/min ◦
0.2 0.2 80 C80 C

◦ ◦
10.2 1
5 C/min ◦
0.2 10.2 0.2
100 ◦100
120 ◦120
prediction
prediction
0.0 0.0 00.0 0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100
250100 0 0 0 20 502040 100
4060 60150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1401.0
140 150
140150 200
(1/s^3) ( ◦ C)
( ◦ C)
1.0 1.0 (a) 120 120 (b) 120 (c)

temperature
T-Tgtemperature
1000.8
100 100
100100 150
0.8Figure
0.8
Cure degree α (-)
Cure degree α (-)
9. Cure prediction of Grindling model fitted solely to isothermal 80 data. (a) Isothermal;
800.680 506050 100
T-Tg ( ◦ C)
( ◦ C)
T-Tg ( ◦ C)
0.6(b)
0.6 non-isothermal; (c) cyclic. 60 60
d3cure/dt3
40
0 0
400.440 50
Glass transition
Glass transition
0.4 0.4 20
By using non-isothermal as well0.2 20 20
as isothermal data as the fitting50target, 050 it is possible to iteratively 0
0.2 0.2 0 0 20
calculate the initial cure degree for 200.0the experimental isothermal10040 curves. When comparing the 50
100 100 101 10in
1
102 10 0 20
102 and 0 0 205020 4010040 6015060 ◦8020080 100
250100 0100
0 0 205020 4010040 6015060 ◦ 8020080 100
250100 0 50 100 150
corresponding plots
TimeTime
(min)(min)Figures 11, Temperature
the TimeTime
(min)
advantage (
(min)C) of diffusion-controlled Temperature
TimeTime
(min) ( C)
(min)
kinetics in case Temperature
of the Grindling kinetic model becomes obvious. The Kamal-Malkin model is not able to reproduce
both isothermal and non-isothermal data in a sufficient way. From Figure 10a it is visible that the
fitting algorithm chooses the initial cure degrees such that all curves reach a cure level of 100%.
This was to be expected since the model is not capable of modeling vitrification effects. This model
actually delivered much better prediction accuracy for non-isothermal and isothermal processing
when only non-isothermal data was used for fitting (cf. Figure 6). The Grindling model shows very
good prediction accuracy for both temperature programs due to its ability to adjust the cure rate not
only to cure temperature but also to the current glass transition temperature.
Polymers 2016, 8, 390 13 of 19
1e6 1e
3 3 1e 3
1.0 1.0 61.0 1.0 51.0 1.0
0.8 0.8 50.8 5 1 ◦ C/min DSC 0.8 40.8 0.8
d2cure/dt2 (1/s^2)
α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
4 4 2.5 ◦ C/min DSC
rate (1/s)
rate (1/s)
Cure rate (1/s)
0.6 0.6 60 C60DSCC DSC
◦ ◦
0.6 5 ◦ C/min DSC 0.6 30.6 0.6
Curedegree
Curedegree
80 ◦ C80DSCC DSC 10 ◦ C/min DSC
3 3
◦
100 ◦100
C DSC◦
C DSC 15 ◦ C/min DSC
0.4 0.4 120 ◦120
C DSC◦
C DSC 0.4 1 ◦ C/min 0.4 20.4 0.4
2 2
Cure
Cure
60 C60 C
◦ ◦
2.5 ◦ C/min
0.2 0.2 80 ◦ C80 ◦ C
10.2 1
5 ◦ C/min 0.2 10.2 0.2
100 ◦100
120 ◦120
prediction
prediction
0.0 0.0 00.0 0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 60150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1401.0
140 150
140150 200

(1/s^3) ( ◦ C)
( ◦ C)
1.0 1.0 (a) 120 120 (b) 120 (c)
temperature

1000.8
100 100
100100 150
0.8Figure
0.8
Cure degree α (-)
Cure degree α (-)
10. Cure prediction of Kamal-Malkin model fitted to isothermal 80 and non-isothermal data.
800.680 506050 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
0.6(a)
0.6Isothermal; (b) non-isothermal;60(c)60cyclic.
d3cure/dt3
400.440 040 0 50
0.4 0.4 20
20
6 1e20
6 1e
3 3 5 1e 3 1.0
1.0 1.0 1.0
0.2 1.0 0
50 1.0
50 0
0.2 0.2 0 0 20
0.8 0.8 5 5
0.820 4
0.8 100
0.8 0.8
200.0 1 C/min DSC
100 10040
◦
250100 500
d2cure/dt2 (1/s^2)
100 100 101 101 102 102 40 040 205020 4010040 6015060 ◦80200
2.5 80C/min
100
250 0 0 0 205020 4010040 6015060 ◦ 8020080 100 50 100 150
Cure degree α (-)
Cure degree α (-)
α (-)
DSC
rate (1/s)
rate (1/s)
Cure rate (1/s)
0.6 0.6 Time

Time
(min)
(min) 60 C
60DSC
◦
C DSC ◦
0.6 Temperature
TimeTime
(min) (
◦ Temperature
TimeTime
(min) (
Curedegree
Curedegree
C DSCC DSC 3 3
◦ ◦ ◦
100 100 ◦ ◦
15 C/min DSC ◦
0.4 0.4 120 120

C DSCC DSC
◦ 0.4
◦
1 C/min 0.4 20.4 ◦ 0.4
2 2
Cure
Cure
60 C60 C
◦ ◦
2.5 C/min ◦
0.2 0.2 80 C80 C

◦ ◦
10.2 1
5 C/min 0.2 10.2 ◦
0.2
100 100
C C ◦ ◦
10 C/min ◦
DSC DSC
120 ◦120
prediction
prediction
0.0 0.0 00.0 0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100
250100 0 0 0 20 502040 100
4060 60 150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1401.0
140 120 120
140 200
(1/s^3) ( ◦ C)
( ◦ C)
(a) 120 120 (b) 120
100 100 (c)
temperature

0.8 0.8 1000.8
100 10080
80 150
Cure degree α (-)
Cure degree α (-)
Figure 11. Cure prediction of Grindling 80 80 model fitted to isothermal 80 and non-isothermal data.
0.6 0.6 0.6 60 60 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
(a) Isothermal; (b) non-isothermal;60(c)60cyclic. 60
40 40
d3cure/dt3
0.4 0.4 400.440 40 50

Glass transition
20 20 202020
5.2.2. Prediction Accuracy Using Cyclic
0.2 0.2 00.2 0 DSC Measurements 0 0 0 0
20 20 202020
0.0Up0.0 to0 this point, kinetic model 400.040
fitting was carried out 250 100 4040
without 0 40
0 0using
205020 data from cyclic DSC 500
10 100 101 101 102 102 0 0 0 205020 4010040 6015060 ◦8020080 100 4010040 6015060 ◦ 80200
80 100250100 50 100 150
measurements. Time Time Temperature
TimeTime
(min) (
(min)C)
(min)(min) when using diffusion-controlled kinetic models, these measurements need
However, Temperature
TimeTime
(min) (
(min) C) Temperature
to be performed anyway in order to characterize the dependency of Tg on cross-linking. It may
therefore be beneficial to also use them for fitting of kinetic models.
By comparing Figures 6 and 12, it becomes evident that by adding cyclic data to the fitting target,
the prediction accuracy of the Kamal-Malkin kinetic model for isothermal processing is improved
considerably. This is especially evident when comparing the match with cyclic DSC measurements
(cf. Figures 6c and 12c). However, the slope at the end of isothermal curves still remains too steep.
The reason for that is the type of data which is gathered from cyclic DSC runs. In contrast to pure
isothermal measurements, cyclic DSC measurements as they are evaluated in the present paper,
only provide data at distinct points. More specifically, there is no information about how a specific
cure degree was reached or for how long a possible vitrification might have been present. As can be
seen in Figures 7 and 13, the Grindling model shows slightly better accuracy after adding cyclic data
to the fitting target. Moreover, a sudden transition from chemically-controlled to diffusion-controlled
reaction is indicated by a kink in the curve for isothermal cure at 60 ◦ C in Figure 13a.
In general, it is observed that isothermal cure prediction gains in accuracy by adding cyclic
data to the fitting target, which apart from that only consists of non-isothermal data. At the same
time, prediction quality for non-isothermal conditions decreases when applying the Grindling model
(cf. Figures 7b and 13b). This was to be expected since the model now has to not only reproduce
non-isothermal curing, but also, at least to some extent, isothermal processing in the form of cyclic DSC.
The opposite is observed in case of the Kamal-Malkin model, which shows a slight improvement in
non-isothermal cure predictions, particularly for low heat rates (cf. Figures 6b and 12b).
Polymers 2016, 8, 390 14 of 19
1e1.6 1e
2 2 1e 3
1.0 1.0 1.61.0 1.0 51.0 1.0
1.4 1.4
0.8 0.8 1.20.81.2 1 C/min DSC 0.8 40.8
◦ 0.8
d2cure/dt2 (1/s^2)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
2.5 C/min DSC
rate (1/s)
Cure rate (1/s)
Cure rate (1/s)
◦
0.6 0.6 1.00.61.0 5 C/min DSC 0.6 30.6

◦
0.6
Curedegree
10 C/min DSC ◦
0.8 0.8 15 C/min DSC ◦
0.4 0.4 60 C60DSCC DSC 0.60.40.6

◦ ◦
1 C/min 0.4 20.4
◦ 0.4
Cure
60 C60 C
◦ ◦
2.5 C/min ◦
0.2 0.2 80 C
80 C 0.40.20.4 5 C/min 0.2 10.2 0.2
◦ ◦ ◦
100 100
C C
◦ ◦
0.2 0.2 10 C/min ◦
DSC DSC
120 120
C C
◦ ◦
15 C/min ◦
modelmodel
prediction
prediction
0.0 0.0 0.0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0.000.0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 6015080 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1401.0
140 200
140200 200
Cure rate (1/s)Glass transition temperature ( C)

(1/s^3) ( ◦ C)
( ◦ C)
(a) 120 120 (b) 120 (c)
◦
150 150
temperature

0.8 0.8 1000.8
100 100 150
Cure degree α (-)
Cure degree α (-)
Figure 12. Cure prediction of 80 Kamal-Malkin

80 model fitted to non-isothermal10080100 and cyclic data.
0.6 0.6 0.6 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
50 50
d3cure/dt3
0.4 0.4 400.440 40 50

Cure degree α (-)Glass transition
20 20 020 0
1.6
0 1e1.6
0 1e
2 2 5 1e 3 1.0
1.0 0.2
1.0 1.0 1.0 0
0.2 0.2 50 1.050 0
20 1.4
1.4 20 20
0.8 0.0
0.0 0.8 0.81.2
400.0 40 4
0.8 100
0.8 0.8
100 10040 250100 500
1.2 1 C/min DSC ◦
d2cure/dt2 (1/s^2)
10-3 10-310-2 10-210-1 10-1 100 100 101 101 102 102 0 0 0 205020 4010040 6015060 ◦80200 2.5 80C/min
100
250 0 0 0 205020 4010040 6015060 ◦ 8020080 100 50 100 150
Cure degree α (-)
Cure degree α (-)
α (-)
DSC
rate (1/s)
Cure rate (1/s)
0.6 0.6 TimeTime

(min)(min) 1.00.61.0 Temperature
TimeTime
(min) (
◦ Temperature
TimeTime
(min) (
Curedegree
10 C/min DSC ◦
0.8 0.8 15 C/min DSC ◦
0.4 0.4 60 C60DSCC DSC 0.60.40.6

◦ ◦
1 C/min 0.4 20.4
◦ 0.4
Cure
60 C60 C
◦ ◦
2.5 C/min ◦
0.2 0.2 80 C80 C 0.40.20.4 5 C/min 0.2 10.2 0.2

◦ ◦ ◦
100 100
C C
◦ ◦
0.2 0.2 10 C/min ◦
DSC DSC
120 120
C C
◦ ◦
15 C/min ◦
modelmodel
prediction
prediction
0.0 0.0 0.0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0.000.0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 6015080 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1401.0
140 200
140200 200
(1/s^3) ( ◦ C)
( ◦ C)
(a) 120 120 (b) 120 (c)
temperature

0.8 0.8 1000.8
100 150
100150 150
Cure degree α (-)
Cure degree α (-)
Figure 13. Cure prediction of Grindling 80 80 model fitted to non-isothermal and 80 cyclic data. (a) Isothermal;
0.6 0.6 600.660 10060100 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
(b) non-isothermal; (c) cyclic.
d3cure/dt3
0.4 0.4 40 0.440 40

50 50 50
Glass transition
20 20 20
0.2When
0.2 fitting to solely isothermal 00.2 data,
0 the match between experimental 00 0 and numerical curves 0
20 20 20
improves significantly by additionally
0.0 0.0-3 -3 -2 -2 -1 -1 0 0 1 1 2 2 400.0 incorporating cyclic data into the fitting target. This is visible 50
10 10 10 10 10 10 10 10 10 10 10 10 0 40
0 0 205020 4010040 6015060 ◦8020080 100 250100 5040 0 0500 205020 4010040 6015060 ◦ 8020080 100
250100 0 50 100 150
in Figures 8 and 14
TimeTime for
(min)(min)the Kamal-Malkin kinetic Temperature
TimeTime
model ( C) and in Figures 9 andTemperature
(min)(min) Time
15Time
for ( C) Grindling
(min)(min)
the Temperature
kinetic model. However, the gain is more pronounced in case of the latter model.
The Kamal-Malkin model compensates for premature vitrification under isothermal conditions
by significantly slowing down the reaction after reaching medium high cure levels (cf. Figure 14a).
The match with cyclic DSC data is therefore good but even though the reaction is slowed down,
it will reach full cure for longer periods of time. Moreover, vitrification occurs for distinct isothermal
temperatures at different cure levels which the Kamal-Malkin kinetic model is not able to account for.
This results in a too early slow down of the reaction in case of the lowest as well as a too late slow
down for the highest isothermal temperature. Hence, the glass transition temperature which is directly
related to the cure degree is underestimated, or in the latter case, overrated. This can have a significant
impact when using such predictions for process simulation since many material properties depend on
the glass transition temperature and whether the material is in glassy or rubbery state.
Although none of the non-isothermal experimental data is used for fitting both models give good
predictions for non-isothermal curing. The accuracy is best for the lowest heating rate which is due
to the fact that the isothermal measurements provide many data points for widespread values of
cure degree within a temperature range of 60 to 120 ◦ C. Most of the reaction of specimens which are
cured at a heating rate of 1 ◦ C/min, happens within this temperature range (cf. Figures 14c and 15c)
and therefore the model is able to accurately predict the curing process. The agreement between
model prediction and experiment declines as the heating rate is increased because the reaction is
shifted towards higher temperatures which require the kinetic models to extrapolate the temperature
dependency of the cure rate into this region. This introduces deviations which are most evident in case
of the highest heating rate of 15 ◦ C/min (cf. Figures 14b and 15b).
Polymers 2016, 8, 390 15 of 19
1e
1.61.01.6 1e
2 2 1e 3
1.0 1.0 1.0 61.0 1.0
1.4 1.4 5
0.8 0.8 1.20.81.2 1 C/min DSC 0.8 0.8 ◦ 0.8
d2cure/dt2 (1/s^2)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
2.5 C/min DSC 4
rate (1/s)
Cure rate (1/s)
Cure rate (1/s)
◦
0.6 0.6 60 C60DSCC DSC 1.00.61.0

◦ ◦
5 C/min DSC 0.6 0.6
◦
0.6
Curedegree
80 C80DSCC DSC 10 C/min DSC 3
◦
C DSCC DSC 0.8 0.8

◦ ◦
100 100 ◦ ◦
15 C/min DSC ◦
0.4 0.4 C DSCC DSC 0.60.40.6

120 120 1 C/min 0.4 0.4 ◦ 0.4
2
◦ ◦
Cure
60 C60 C◦ ◦
2.5 C/min ◦
0.2 0.2 80 C
80 C 0.4 0.4
0.2 5 C/min 0.2 10.2 0.2
◦ ◦ ◦
100 100
C C ◦
0.2 0.2
◦
10 C/min ◦
DSC DSC
120 120
C C ◦ ◦
15 C/min ◦
modelmodel
prediction
prediction
0.0 0.0 0.0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0.000.0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 60 150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1201.0
120 140 140
140 200
Cure rate (1/s)Glass transition temperature ( C)

(1/s^3) ( ◦ C)
temperature ( ◦ C)
(a) (b) 120 120
120 (c)
◦
100 100
temperature
0.8 0.8 800.880 100 100
100 150
Cure degree α (-)
Cure degree α (-)
Figure 14. Cure prediction of Kamal-Malkin model fitted solely 80 to8080

isothermal and cyclic data.
0.6 0.6 600.660 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
d3cure/dt3
0.4 0.4 404040

200.420 50

202020
0
1.4 1e 0
1.4 1e
2 2 5 1e 3 1.0
1.0 1.0
0.2 0.2 1.0
0.2 0
1.0 0 0
1.0 0
20 1.2
1.2 20 202020
0.8 0.0
0.0 0.8 0.840
400.0 1 C/min DSC 4
0.8 0.8 0.8
100 40400 40 250100 500
◦
d2cure/dt2 (1/s^2)
100 100 101 101 102 1021.001.00 0 205020 4010040 6015060 ◦80200
2.5 80C/min
100
250 0 0 205020 4010040 6015060 ◦ 8020080 100 50 100 150
Cure degree α (-)
Cure degree α (-)
α (-)
DSC
rate (1/s)
Cure rate (1/s)
0.6 0.6 TimeTime

(min)
(min) 60 C60DSC
◦
C DSC ◦
0.6 Temperature
TimeTime
(min) (
(min)C)5 C/min DSC 3
0.6 0.6
◦ Temperature
TimeTime
(min) (
80 C80DSCC DSC 0.8 0.8
Curedegree
◦ ◦
10 C/min DSC ◦
100 ◦100
C DSC
0.60.40.6
◦
C DSC 15 C/min DSC ◦
0.4 0.4 120 ◦120

C DSC◦
C DSC 1 C/min 0.4 20.4 ◦ 0.4
Cure
60 C60 C
0.4 0.4
◦ ◦
2.5 C/min ◦
0.2 0.2 0.2

80 ◦ C80 ◦ C 5 C/min 0.2 10.2 ◦
0.2
0.2 0.2
100 ◦100
C ◦C 10 C/min ◦
DSC DSC
120 ◦120
C ◦C 15 C/min ◦
modelmodel
prediction
prediction
0.0 0.0 0.00.00.0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100 250100 0 0 0 20 502040 100
4060 60150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.0 1.0 1401.0
140 140
140140 200
(1/s^3) ( ◦ C)
( ◦ C)
(a) 120 120 (b) 120120
120 (c)
temperature

0.8 0.8 1000.8
100 100100
100 150
Cure degree α (-)
Cure degree α (-)
Figure 15. Cure prediction of Grindling 80 80 model fitted to isothermal and 80 cyclic data. (a) Isothermal;
0.6 0.6 80 80
600.660 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
60 60
d3cure/dt3
0.4 0.4 400.440 40 50

Glass transition
20 20 402040
0.2Figures
0.2 10 and 16 show the difference 00.2 0 20 020
between fitting to non-isothermal as well as isothermal 0
20 20 020 0
and0.0doing
0.0 so with additionally using cyclic data in case of the Kamal-Malkin kinetic model.50
0 0 1 1 2 2 400.0
0 40
0 0 205020 40
10040 60
150 60 80
20080 100250
100 2040
0 20
0 0 205020 4010040 6015060 ◦ 8020080 100
250100 0 50 100 150
10 10 10 10 10 10
An improvement is(min)
TimeTimevisible
(min) for each measurement Temperature
TimeTime
(min)
type. ( C)However, isothermalTemperature
(min)◦
Time
cureTime
(min) ( C)
(min)
predictions still Temperature
end up at too high values of cure degree, which again arises from the fact, that the model is not able
to render vitrification effects. The slow down in cure rate, which was observed when fitting against
isothermal and cyclic data (cf. Figure 14a), is less pronounced in this case. This is due to a change
in weighting of each of the measurement types. If fitting is carried out under consideration of only
isothermal and cyclic data, vitrification is present in all of the experimental data. However, this is
not the case for non-isothermal data and by adding this type of measurement to the fitting process,
data with vitrification loses in weight.
1e6 1e
3 3 1e 3
1.0 1.0 61.0 1.0 51.0 1.0
0.8 0.8 50.8 5 1 ◦ C/min DSC 0.8 40.8 0.8

d2cure/dt2 (1/s^2)
α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
4 4 2.5 ◦ C/min DSC

rate (1/s)
rate (1/s)
Cure rate (1/s)
0.6 0.6 60 C60DSCC DSC

◦ ◦
0.6 5 ◦ C/min DSC 0.6 30.6 0.6
Curedegree
Curedegree
80 ◦ C80DSC C DSC 10 ◦ C/min DSC

3 3
◦
100 ◦100
C DSC◦
0.4 0.4 120 120
◦
C DSC◦
C DSC 0.4 1 ◦ C/min 0.4 20.4 0.4
2 2
Cure
Cure
60 ◦ C60 ◦ C 2.5 ◦ C/min

0.2 0.2 80 ◦ C80 ◦ C
10.2 1
5 ◦ C/min 0.2 10.2 0.2
100 ◦100
120 ◦120
prediction
prediction
0.0 0.0 00.0 0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100
250100 0 0 0 20 502040 100
4060 60150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.2 1.2 2001.0
200 150
140150 200
(1/s^3) ( ◦ C)
( ◦ C)
(a) (b) 120 (c)

1.0 1.0
temperature
T-Tgtemperature
1500.8
150 100
100100 150
Cure degree α (-)
Cure degree α (-)
0.8Figure
0.8 16. Cure prediction of Kamal-Malkin model fitted to isothermal, non-isothermal 80 and cyclic data.
1000.6
100 506050 100
T-Tg ( ◦ C)
( ◦ C)
T-Tg ( ◦ C)
0.6(a)
0.6Isothermal; (b) non-isothermal; (c) cyclic.
d3cure/dt3
500.450 40
0 0 50
Glass transition
Glass transition
0.4 0.4 20
0.2 the contrary, by comparing00.2
0.2On the0 results of the Grindling model 50 050 in Figures 11 and 17, a very 0
20
good match between experiment and
0.0 0.0 500.0 model is observed regardless
50 10040of whether cyclic data is used 50
100 100 101 101 102 102 0 0 0 205020 4010040 6015060 ◦8020080 100250100 0100
0 0 205020 4010040 6015060 ◦ 8020080 100
250100 0 50 100 150
TimeTime
TimeTime
(min) (
(min)C) Temperature
TimeTime ( C)
Polymers 2016, 8, 390 16 of 19
for fitting or not. The slight deviations may also arise from the fitting algorithm itself. However,
by comparing Figures 15b and 17b, it is observed that by adding non-isothermal data to the fitting
target, the prediction accuracy for this type of processing is improved whereas at the same time,
the match with cyclic data loses in quality. This is again because of the fact that the model now has to
satisfy a more comprehensive situation.
1e6 1e
3 3 1e 3
1.0 1.0 61.0 1.0 51.0 1.0
0.8 0.8 50.8 5 1 ◦ C/min DSC 0.8 40.8 0.8
d2cure/dt2 (1/s^2)
α (-)
Cure degree α (-)
Cure degree α (-)
Cure degree α (-)

α (-)
4 4 2.5 ◦ C/min DSC
rate (1/s)
rate (1/s)
Cure rate (1/s)
0.6 0.6 60 C60DSCC DSC
◦ ◦
0.6 5 ◦ C/min DSC 0.6 30.6 0.6
Curedegree
Curedegree
80 ◦ C80DSC C DSC 10 ◦ C/min DSC
3 3
◦
100 ◦100
C DSC◦
0.4 0.4 120 120
◦
C DSC◦
C DSC 0.4 1 ◦ C/min 0.4 20.4 0.4
Cure 2 2
Cure
60 C60 C
◦ ◦
2.5 ◦ C/min
0.2 0.2 80 ◦ C80 ◦ C
10.2 1
5 ◦ C/min 0.2 10.2 0.2
100 ◦100
120 ◦120
prediction
prediction
0.0 0.0 00.0 0 0.0 00.0
0 0 20 20 40 40 60 60 80 80 100 100 0 0 0 205020 4010040 6015060 ◦8020080 100
250100 0 0 0 20 502040 100
4060 60 150
80 ◦80200 250120 0.00
100 100120 50 100 150
TimeTime
TimeTime ( C)
TimeTime ( C)
1.2 1.2 1501.0
150 120 120
140 200
(1/s^3) ( ◦ C)
( ◦ C)
(a) (b) 120
100 100 (c)
1.0 1.0
temperature

1000.8
100 10080
80 150
Cure degree α (-)
Cure degree α (-)
0.8Figure
0.8 17. Cure prediction of Grindling model fitted to isothermal, non-isothermal 80 and cyclic data.
0.6 60 60 100
T-Tg ( ◦ C)
T-Tg ( ◦ C)
T-Tg ( ◦ C)
60
0.6(a)
0.6Isothermal; (b) non-isothermal;50(c)50cyclic. 40 40
d3cure/dt3
0.4 40 50
Glass transition
0.4 0.4 202020

00.2 0 of the different fitting 0results, 00 the goodness of fit is 0
0.2In0.2 order to simplify the comparison 202020
quantified
0.0 0.0 0 0 by standard errors, as summarized
500.050 in Table 2. In this table, non-isothermal and isothermal 50
10 10 101are 102 102 0as
101 abbreviated 250100 4040
0 0 205020 4010040 6015060 ◦8020080 100 0 40
0 0 205020 4010040 6015060 ◦ 8020080 100
250100 0 50 100 150
DSC measurements TimeTime
(min)(min) dyn Temperature
andTimeTime
iso, ( C)
(min)(min)
respectively. Lower Temperature
TimeTime
values of ( C)
(min)(min)
standard error Temperature
denote a better accordance between measurement and model prediction. However, it is important
to mention that the prediction quality for isothermal conditions is overrated when the initially lost
amount of cross-linking is overestimated by the fitting algorithm. This is observed for both models
when fitted against solely isothermal data, and in case of the Kamal-Malkin model when fitted to
isothermal and non-isothermal data. On the contrary, the prediction accuracy for isothermal curing
is underrated when isothermal measurements are not used for model fitting, since the initial cure
degrees cannot be estimated under this circumstances. In this case, the standard error is calculated
with the assumption of zero initial cure (cf. Figure 2a). Corresponding values of standard error are
marked accordingly in Table 2. Due to the mentioned restrictions, the validity and comparability of
the shown standard errors is limited. Interpretation of the data in Table 2 should therefore always be
accompanied by visual comparison of corresponding cure predictions.
Table 2. Prediction accuracy (standard error) of the kinetic models parametrized using different fitting
targets (lower values denote better accuracy).
Fitting target
Kinetic model Prediction case
dyn * iso dyn-iso dyn-cyclic iso-cyclic dyn-iso-cyclic
isothermal 0.1634 ‡ 0.0230 † 0.0393 † 0.1614 ‡ 0.0642 0.0177
non-isothermal 0.0337 0.0734 0.0693 0.0311 0.0655 0.0430
Kamal-Malkin
cyclic 0.1521 0.1061 0.1193 0.0580 0.0119 0.0684
mean error 0.1164 ‡ 0.0675 † 0.0760 † 0.0835 ‡ 0.0472 0.0430

isothermal 0.1360 ‡ 0.0230 † 0.0087 0.1496 ‡ 0.0519 0.0089
non-isothermal 0.0040 0.0734 0.0258 0.0194 0.0520 0.0258
Grindling
cyclic 0.0874 0.1061 0.0457 0.0541 0.0038 0.0524
mean error 0.0758 ‡ 0.0675 † 0.0267 0.0744 ‡ 0.0359 0.0290
* dyn = non-isothermal DSC (dynamic temperature scanning); initial cure degree is overestimated †
(no premature vitrification during isothermal curing); ‡ initial cure degrees assumed to be zero because
isothermal DSC data is not used for fitting (cf. Figure 2a).
Since it is favorable to reduce the needed amount of measurements, Figure 18 shows the fitting
quality of the Grindling kinetic model in case cyclic data and only a reduced set of isothermal and
non-isothermal measurements are used for model parametrization. The isothermal data is limited to
Polymers 2016, 8, 390 17 of 19
the temperatures 60, 80 and 100 ◦ C. In case of non-isothermal DSC, only results for heating rates of
1, 5 and 15 ◦ C/min are used.
By comparing Figures 17a and 18a only minor differences can be identified. This suggests that
isothermal measurements at high temperatures close to the ultimate glass transition temperature
Tg,∞ can be omitted. This may be due to the fact that at this elevated temperatures, the influence of
vitrification is much smaller compared to lower temperatures. Since DSC data of isothermal curing at
120 ◦ C was not used for fitting, a corresponding initial degree of cure was not predicted. Therefore,
the experimental curve in Figure 18a shows far too low cure degrees.
Prediction quality for non-isothermal processing (cf. Figures 17b and 18b) loses in accuracy.
However, this is acceptable since for the majority of composite manufacturing processes, i.e., RTM,
good and accurate results for isothermal temperature programs is much more important. Furthermore,
the match with cyclic data is improved (cf. Figures 17c and 18c) which again is advantageous for
modeling of isothermal processes.
1.0 1.0 0.101.0

0.10 1e 3
1.051.0 1.0
0.8 0.8 0.080.8
0.08 1 ◦ C/min DSC 0.840.8 0.8
d2cure/dt2 (1/s^2)
Cure degree α (-)
Cure degree α (-)

α (-)
Cure degree α (-)
α (-)
2.5 ◦ C/min DSC
Cure rate (1/s)
rate (1/s)
rate (1/s)
0.6 0.6 60 C60DSCC DSC

◦ ◦
0.060.6
0.06 5 ◦ C/min DSC 0.630.6 0.6
Curedegree
Curedegree
80 ◦ C80DSC
◦
100 ◦100
C DSC
◦
0.4 0.4 120 120
◦
C DSC
◦
C DSC 0.040.4
0.04 1 ◦ C/min 0.420.4 0.4
Cure
Cure
60 C60 C
◦ ◦ 2.5 ◦ C/min
0.2 0.2 80 ◦ C80 ◦ C 0.020.2
0.02 5 ◦ C/min 0.210.2 0.2
100 ◦100
C ◦C 10 ◦ C/min DSCDSC
120 ◦120
C ◦C 15 ◦ C/min model prediction
model prediction
0.0 0.0 0.0 0.000.0
0 0 20 20 40 40 60 60 80 80 100100 0.000.00 0 0 0 205020 40
10040 60 15060 ◦ 80
20080 100
250100 00 0 205020 40 100 15080 ◦80200
40 60 60 120120 0.00
100100250 50 100 150
TimeTime
TimeTime
(min) ( C)
(min) Temperature
Time (min)
Time ( C)
(min) Temperature
1.0 1.0 1401.0
140 200
140200 200
(1/s^3) ( ◦ C)
( ◦ C)
(a) 120120 (b) 120 (c)

temperature
T-Tgtemperature
0.8 0.8 1000.8100 100150

150 150
Cure degree α (-)
Cure degree α (-)
Figure 18. Cure prediction of Grindling 80 80 model fitted to cyclic data and a80reduced set of isothermal and
0.6 0.6 0.660 100100 100
T-Tg ( ◦ C)
( ◦ C)
T-Tg ( ◦ C)
non-isothermal experiments: 60, 8060and 100 ◦ C (isothermal) and 1, 5 and6015 ◦ C/min (non-isothermal).
d3cure/dt3
0.4Isothermal; (b) non-isothermal;40

0.4(a) 0.440cyclic. 504050 50
Glass transition
Glass transition
(c)
20 20 20
0.2 0.2 0
0.2 0 0
0 0 0
20 20 20
It should be clear from the
0.0 0.0-3 -3 -2 -2 -1 -1 0 0 1 1 2 2 400.0
above that the often applied approach of only using non-isothermal
400 5020 10040 15060 20080 250100 50400500 5020 10040 15060 20080 250100 500
10 10 10 10 10 10 10
DSC measurements for 10 0 0 parametrization
10 10 10model
10kinetic 20 40 60 ◦ 80 100
is not suitable 0
for20prediction
40 60 ◦ 80 100
of
50 100 150
Time
Time (min)(min) Temperature
TimeTime
(min) ( C)
(min) Temperature
Time
Time ( C)isothermal
(min)
(min) Temperature
curing at temperatures below the ultimate glass transition temperature. Instead, the results of the
present study indicate that the non-isothermal measurements are actually not as important as the
isothermal ones. This is due to the fact that isothermal measurements contain information of both,
cross-linking with and without influence of vitrification. Non-isothermal DSC shows this effect only
for very low heating rates, for which the glass transition temperature is able to catch up with the
reaction temperature. This results in a temporarily decrease in reaction rate. However, since the
temperature is continuously increased, non-isothermal DSC data does not contain information about a
complete stop in reaction as is observed under isothermal conditions. Using isothermal measurements
for model parametrization, however, requires the fitting strategy to account for the initially lost
amount of cure which cannot be avoided when conducting DSC measurements of fast resins at
elevated isothermal temperatures. By iteratively approximating the initial cure degrees during fitting,
as presented in this study, this major drawback of isothermal DSC runs was successfully eliminated.
6. Conclusions
In this study, the prediction accuracy of a simple (Kamal-Malkin) and a complex (Grindling)
kinetic model was investigated and compared to each other for non-isothermal, isothermal and
cyclic temperature programs. The Grindling model was chosen because of its ability to account for
premature vitrification, which typically occurs during isothermal processes like RTM. In this context,
the cure dependency of the glass transition temperature was characterized and a non-linear behavior
of the investigated epoxy resin towards high cure degrees was observed. Furthermore, since some
cyclic DSC measurements showed enthalpy relaxation due to physical aging, a strategy was presented
Polymers 2016, 8, 390 18 of 19
which enables derivation of reliable cure degree values from this data with the aid of a parametrized
Tg -model. This additional values proved to be beneficial for subsequent fitting of the kinetic models.
Comparison of the goodness of fit of both kinetic models was carried out for parameter sets
which were fitted using various combinations of the three measurement types. It was shown that
for obtaining accurate and reliable cure predictions for non-isothermal or isothermal processing, it is
generally necessary to include corresponding measurements in the fitting process. It is especially
not sufficient to use only non-isothermal data for fitting of a model which is intended to predict
isothermal curing at temperatures below the ultimate glass transition temperature. Instead, it is vital
to include measurements into the fitting process which contain information concerning relevant effects
like vitrification. However, using solely isothermal data also led to poor results regardless of the
applied kinetic model due to fundamental problems closely related to specimen preparation and DSC
data acquisition in general. A strategy was presented in order to use this type of measurement for
model fitting despite of the mentioned problems by iteratively approximating the initial amount of
cure which was not covered by the DSC analyzer. However, as has been shown, this approach is only
of advantage if additional data from non-isothermal or cyclic DSC runs is available during fitting.
Consequently, very good accuracy for isothermal processing was achieved when fitting the complex
Grindling model against isothermal and cyclic data. The simpler Kamal-Malkin model also benefits
from this approach. However, prediction quality for isothermal curing is inferior to the results of the
Grindling model. Non-isothermal curing is predicted by both models with good results although
none of the corresponding measurement data was used for fitting. Compared to fitting by using solely
non-isothermal data, the mean error for both models is reduced by about 60%. An equal reduction
in prediction error is achieved when using isothermal and non-isothermal data, although in this case
good results were only achieved by the Grindling kinetic model.
Predictions of the Kamal-Malkin kinetic model are generally inferior to predictions of the
Grindling model. When using the Grindling model, the standard error can be reduced by a maximum
of 65% compared to the Kamal-Malkin model.
In relation to the question which measurements are required for accurate cure modeling in
isothermal processes, it is important to make clear that this can only be accomplished by a suitable
kinetic model. As has been shown in this study, the simpler Kamal-Malkin kinetic model is not able to
account for vitrification effects which are important in isothermal processing. As a result, it generally
fails in accurately predicting the evolution of the cure degree under isothermal conditions. The more
complex Grindling model achieved significantly better results but necessitates the characterization of
the evolution of the glass transition temperature during cross-linking. However, since cure modeling is
often done in the context of process simulation, this relationship needs to be determined anyway, e.g.,
for the purpose of modeling the evolution of mechanical parameters during cross-linking. Furthermore,
results of this study indicate that having isothermal and cyclic data available obviates the need for
conducting non-isothermal DSC runs. However, the total heat of reaction ∆h, which is needed for
evaluation of the DSC data, as well as the ultimate glass transition temperature Tg,∞ were determined in
this study from non-isothermal data. It is therefore not possible to completely disregard non-isothermal
DSC runs but a decrease in number is feasible. Moreover, results of this study indicate that isothermal
measurements at temperatures near the ultimate glass transition temperature can be omitted. This way,
very good prediction accuracy is achieved by at the same time reducing the experimental effort.
Acknowledgments: The present study was financially supported by Deutsche Forschungsgemeinschaft (DFG)
within the Priority Program (SPP) 1712, which is gratefully acknowledged. Open Access publishing was kindly
supported by Deutsche Forschungsgemeinschaft and Open Access Publishing Fund of Karlsruhe Institute
of Technology. Furthermore, the authors want to thank the staff of the Institute for Technical Chemistry and
Polymer Chemistry—Polymeric Materials (ITCP) for their support regarding DSC measurements, and the Sika
Deutschland GmbH (Stuttgart, Germany) for providing the resin.
Author Contributions: Alexander Bernath conceived and designed the experiments, analyzed the data and wrote
the paper; Luise Kärger and Frank Henning revised the paper and supervised the research.
Polymers 2016, 8, 390 19 of 19
Conflicts of Interest: The authors declare no conflict of interest. The founding sponsors had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the
decision to publish the results.
Abbreviations
The following abbreviations are used in this manuscript:
DSC Differential Scanning Calorimetry
RTM Resin Transfer Molding
CMA-ES Covariance Matrix Adaptation Evolution Strategy
References
1. Pascault, J.P.; Williams, R.J.J. Glass transition temperature versus conversion relationships for thermosetting
polymers. J. Polym. Sci. Part B 1990, 28, 85–95.
2. Montserrat, S. Calorimetric measurement of the maximum glass transition temperature in a thermosetting
resin. J. Therm. Anal. 1993, 40, 553–563.
3. Gonzalez-Romero, V.M.; Casillas, N. Isothermal and temperature programmed kinetic studies of thermosets.
Polym. Eng. Sci. 1989, 29, 295–301.
4. Lange, J.; Altmann, N.; Kelly, C.; Halley, P. Understanding vitrification during cure of epoxy resins using
dynamic scanning calorimetry and rheological techniques. Polymer 2000, 41, 5949–5955.
5. O’Brien, D.J.; Mather, P.T.; White, S.R. Viscoelastic Properties of an Epoxy Resin during Cure. J. Compos. Mater.
2001, 35, 883–904.
6. Henne, M.; Breyer, C.; Niedermeier, M.; Ermanni, P. A new kinetic and viscosity model for liquid composite
molding simulations in an industrial environment. Polym. Compos. 2004, 25, 255–269.
7. Spoelstra, A.B.; Peters, G.W.M.; Meijer, H.E.H. Chemorheology of a highly filled epoxy compound.
Polym. Eng. Sci. 1996, 36, 2153–2162.
8. Hsieh, T.H.; Su, A.C. Cure behavior of an epoxy–novolac molding compound. J. Appl. Polym. Sci. 1992, 44,
165–172.
9. Atarsia, A.; Boukhili, R. Relationship between isothermal and dynamic cure of thermosets via the
isoconversion representation. Polym. Eng. Sci. 2000, 40, 607–620.
10. Javdanitehran, M.; Berg, D.C.; Duemichen, E.; Ziegmann, G. An iterative approach for isothermal curing
kinetics modelling of an epoxy resin system. Thermochim. Acta 2016, 623, 72–79.
11. Kamal, M.R.; Sourour, S.; Ryan, M.E. Integrated thermo-rheological analysis of the cure of thermosets.
SPE Tech. Pap. 1973, 19, 187.
12. Grindling, J. Simulation zur Verarbeitung von Reaktiven Non-Post-Cure-Epoxidharz-Systemen im
Druckgelieren und Konventionellen Vergiessen. Ph.D. Thesis, University of Paderborn, Paderborn, Germany, 2006.
13. Halley, P.J.; Mackay, M.E. Chemorheology of thermosets—An overview. Polym. Eng. Sci. 1996, 36, 593–609.
14. Fava, R. Differential scanning calorimetry of epoxy resins. Polymer 1968, 9, 137–151.
15. Hutchinson, J.M. Determination of the glass transition temperature: Methods correlation and structural
heterogeneity. J. Therm. Anal. Calorim. 2009, 98, 579–589.
16. Plazek, D.J.; Frund, Z.N. Epoxy resins (DGEBA): The curing and physical aging process. J. Polym. Sci. Part B
1990, 28, 431–448.
17. DiBenedetto, A.T. Prediction of the glass transition temperature of polymers: A model based on the principle
of corresponding states. J. Polym. Sci. Part B 1987, 25, 1949–1969.
18. Hansen, N.; Ostermeier, A. Completely Derandomized Self-Adaptation in Evolution Strategies.
Evolut. Comput. 2001, 9, 159–195.
19. Lin, Y.G.; Sautereau, H.; Pascault, J.P. Epoxy network structure effect on physical aging behavior. J. Appl.
Polym. Sci. 1986, 32, 4595–4605.
20. Pang, K.P.; Gillham, J.K. Competition between cure and thermal degradation in a high Tg epoxy system:
Effect of time and temperature of isothermal cure on the glass transition temperature. J. Appl. Polym. Sci.
1990, 39, 909–933.
c 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright of Polymers (20734360) is the property of MDPI Publishing and its content may
not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.

Artigo Interessante

Uploaded by

Copyright:

Available Formats

Artigo Interessante

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Artigo Interessante

Uploaded by

Copyright:

Available Formats

polymers

Academic Editor: Alexander Böker

Polymers 2016, 8, 390; doi:10.3390/polym8110390 www.mdpi.com/journal/polymers

2.2.1. Reaction Kinetics

Reaction enthalpy h (J/g)

Cure degree α (-)

Cure rate (1/s)

Cure degree α (-)

Cure rate α̇ (1/s)

Glass transition temperature ( ◦◦ C)

measurements. (a) Cure degree; (b) cure rate.

2.2.2. Glass Transition Temperature

Table 1. Parameters of cyclic DSC runs.

Curing temperature Tc (◦ C) Hold times th (min)

3.1. Glass Transition Temperature

3.2. Reaction Kinetics

3.2.1. Kamal-Malkin Kinetic Model

3.2.2. Grindling Kinetic Model

5. Results and Discussion

5.1. Glass Transition Temperature

Specific heat flow q (W/g)

5.2. Reaction Kinetics

5.2.1. Prediction Accuracy without Using Cyclic DSC Measurements

0.8 0.8 2.00.82.0 1 ◦ C/min DSC 0.8 40.8 0.8

Cure degree α (-)

2.5 ◦ C/min DSC

0.6 0.6 1.50.61.5 5 ◦ C/min DSC 0.6 30.6 0.6

60 ◦ C60 ◦ C 2.5 ◦ C/min

(a) 100 100 (b) 140

Glass transition temperature

0.8 0.8 800.880 120

(b) non-isothermal; (c) cyclic. 806080

0.4 0.4 6040 60

0.8 0.8 0.80.80.8 1 ◦ C/min DSC 0.8 40.8 0.8

Cure degree α (-)

Glass transition temperature ( ◦ C)

Cure degree α (-)Glass transition temperature

0.4 0.4 400.440 40 50

0.6 0.6 TimeTime

0.4 0.4 120 120

0.2 0.2 80 C80 C

Cure degree α (-)Glass transition temperature

8. Cure prediction of Kamal-Malkin model fitted solely to isothermal

0.6 0.6 TimeTime

80 C80DSCC DSC 10 C/min DSC

0.4 0.4 120 120

0.2 0.2 80 C80 C

1.0 1.0 (a) 120 120 (b) 120 (c)

0.8 0.8 50.8 5 1 ◦ C/min DSC 0.8 40.8 0.8

Cure degree α (-)

Glass transition temperature ( ◦ C)

Cure degree α (-)Glass transition temperature

0.6 0.6 Time

0.4 0.4 120 120

0.2 0.2 80 C80 C

Glass transition temperature

0.4 0.4 400.440 40 50

Cure degree α (-)

0.6 0.6 1.00.61.0 5 C/min DSC 0.6 30.6