Polymer 51 (2010) 6058e6070

Contents lists available at ScienceDirect

Polymer
journal homepage: www.elsevier.com/locate/polymer

Molecular simulations of crosslinking process of thermosetting polymers

Chunyu Li*, Alejandro Strachan
School of Materials Science and Engineering and Birck Nanotechnology Center, Purdue University, West Lafayette, IN 47906, USA

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

Article history: We use molecular dynamics (MD) with a procedure to describe chemical reactions to predict the atomic
Received 23 July 2010 structure and properties of the thermosetting polymer epoxy EPON-862 and curing agent DETDA. The
Received in revised form DREIDING force field is employed with environment-dependent atomic charges obtained self consis-
13 October 2010
tently during the dynamics. We propose a computationally efficient method to describe charge evolution
Accepted 15 October 2010
based on the observation that atomic charges evolve significantly only during chemical reactions and in
Available online 23 October 2010
a repeatable manner. Two chemistry models with different relative rates for primary and secondary
amine reactions are used to mimic the curing process in two extreme cases of processing conditions. The
Keywords:
Molecular dynamics simulations
simulations show that differences in chemical reaction rates affect properties for intermediate conver-
Thermoset polymer sion degrees (w40e70%) but not for the higher conversion rates of interest in most applications. We also
Polymerization find that performing the polymerization at high temperatures leads to networks with lower internal
strain energy due to increased molecular mobility. The predicted density, coefficient of thermal
expansion, glass transition temperature and elastic constants of the resulting polymers are in excellent
agreement with experiments.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction trial and error leading to lengthy and costly cycles. In this context,
physics-based computational tools capable of predicting the
Fiber-reinforced polymer composite materials have become properties of new formulations would be highly desirable.
increasingly important in aircraft applications because of their Advances in atomistic simulation techniques and the continuing
advantages in weight reduction and energy saving. For example, increase in computer power open the possibility of a molecular-
the Boeing Company has raised the use of composite materials in level description of these materials. Furthermore, being based on
new aircrafts up to 50% of structural weight [1]. Thermosetting first principles these techniques have the potential to predict the
polymers are the matrices of choice for such composites due to performance of new formulations and to provide guidance to
their high stiffness, strength, creep resistance and thermal resis- design and optimization efforts. In this paper we develop a proce-
tance when compared with thermoplastic polymers [2]. These dure based on molecular dynamics (MD) simulations to predict the
desirable properties stem from the three-dimensional (3D) cross- molecular structure and material properties of thermosetting
linked structures of these polymers. Many thermosetting polymers polymers from first principles. The procedure is applied to the
are formed by mixing a resin (epoxy, vinyl ester, or polyester) and epoxy resin EPON-862 with curing agent DETDA. This choice
a curing agent. An irreversible chemical reaction cures the two enables the comparison of our results with available experimental
components into a solid polymer with a 3D network of covalent data and prior simulation results. We characterize the role of the
bonds. polymerization procedure on the internal strain of the generated
The mechanical and physical properties of thermosetting poly- network and the role of the relative reaction rates for the poly-
mers depend on their chemistry, composition, and curing condi- merization chemistry on resulting properties. The predictions (with
tions. Taking into account the various resin and curing agent no adjustable parameters) of glass transition temperature, density,
combinations as well as additives to control processing character- thermal expansion coefficient and elastic constants are in good
istics (e.g. viscosity) and resulting properties, the number of agreement with available experimental values. It should be pointed
possible formulations is enormous. The optimization of thermoset out that the polymerization procedure characterized in this paper is
formulations remains, to a large degree, empirical and based on generally applicable and can be used to predict other thermosetting
polymer systems without modifications.
Initial efforts to study the formation of crosslinked networks
* Corresponding author. with atomic resolution by computer simulations can be traced back
E-mail address: licyemail@yahoo.com (C. Li). to 1980s. A rather thorough review of such works is provided by

0032-3861/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.polymer.2010.10.033
 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070 6059

Doherty et al. [3]. In these initial studies, all the crosslinking reac-
tions were assumed to occur simultaneously without time for the
system to relax as the network is formed; thus, large internal
network strains are expected in the resulting configurations. Since
Doherty et al. [3] performed simulations allowing a progressive
crosslinking and polymerization reaction using MD, several
researchers reported studies of crosslinked polymer networks built
using molecular dynamics simulations. For example, Yarovsky and
Evans [4] developed a methodology for constructing molecular
models of crosslinked polymer networks and applied it to low
molecular weight water soluble epoxy resins cured with different
crosslinking agents. The polymer consistent force field (PCFF) was
used and the volume shrinkage during conversion was predicted.
Heine et al. [5] simulated the structure and elastic moduli of end-
crosslinked PDMS networks. The networks were formed dynami-
cally using MD. The newly formed topology was relaxed using
a modified potential, harmonic at bonding distances, below a pre-
determined cutoff, and linear at larger distances to avoid instabil-
ities when a new bond is formed. Gou et al. [6] built crosslinked
polymer matrix with epoxy resins (EPON-862) and curing agents
(DETDA) to study interfacial load transfer in carbon nanotubes/
polymer composites. Wu and Xu [7] performed crosslinking
simulations for epoxy resin system based on DGEBA (diglycidyl
ether bisphenol A) and IPD (isophorone diamine). They used the
DREIDING force field with charge equilibration to build the struc-
ture but COMPASS force field for property prediction. They found
that COMPASS provides a more accurate description of elastic
properties; however, their DREIDING prediction of Young’s
modulus (about 50 GPa) is not consistent with our experience as
will be described below. Fan and Yuen [8] conducted MD simula-
tions to predict the glass transition temperature (Tg), linear thermal
expansion coefficients (CTE) and Young’s modulus of fully cured
epoxy network composed of EPON-862 epoxy resin and TETA Fig. 1. Molecular structures of (a) EPON-862, (b) activated EPON-862 and (c) DETDA.
(triethylenetetramine) curing agent. Their simulations used the
Material Studio (Accelrys Inc.) with the PCFF force field and their
predictions of Tg, CTE and Young’s modulus were shown to be when chemical reactions occur [10]. Thus, an accurate simulation of
within 10% of the experiments. Lin and Khare [9] presented the crosslinking process should include a procedure for updating
a single-step polymerization method for the creation of atomistic charges to take into account changes in chemistry and topology.
model structures of crosslinked polymers. A simulated annealing Geometry dependent, self-consistent charges can be obtained by
algorithm was used to identify pairs of reacting atoms within minimizing an expression for the total electrostatic energy using
a cutoff distance and all crosslinking bonds were created in a single methods like Charge equilibration (QEq) [12] or electronegativity
step. Recently, Varshney et al. [10] reported a study of molecular equalization method (EEM) [13e15]. These methods are accurate and
modeling of thermosetting polymers with special emphasis on transferable and have been applied to polymerization [7] but they
crosslinking procedure. They describe different approaches to build remain computationally intensive especially for large-scale simula-
highly crosslinked polymer networks and proposed a multistep tions. In this paper, we characterized the charge evolution during
relaxation procedure for relaxing newly formed topology in the polymerization of the EPON-862/DETDA system using explicit EEM.
crosslinking process. Using the consistent valence force field We find that charges evolve significantly only during chemical
(CVFF), they predicted density, glass transition temperature, and reactions and this evolution is predictable from the chemistry of the
thermal expansion coefficient of an epoxy-based thermoset (EPON- reaction and insensitive to the actual atomic configuration of the
862/DETDA); the predicted Tg is lower than the experimental value. process. Based on these observations, we propose a computationally
Komarov et al. [11] reported a new computational method where efficient charge updating approach, denoted electronegative equal-
the polymer network is polymerized at a coarse-grained level and ization-based charge assignment (ECA), which enables large-scale
then mapped into a fully atomistic model. Molecular dynamics are polymerization and crosslinking simulations.
then carried out with the OPLS force field. The predicted Tg is about
20 K lower than experimental results; this underestimation would 2. Computational approach and simulation details
increase once the extremely high cooling rates of MD taken into
account. Our overall procedure to predict the atomic structure and
An aspect that has not been fully explored in these crosslinking thermo-mechanical properties of thermoset polymers consists of
simulations is the evolution of partial atomic charges during curing. three main stages: i) polymerization, where the initial liquid
Electrostatic interactions play an important role in the overall mixture forms a 3D network; ii) annealing, where the polymerized
binding of these systems and an accurate calculation of charge system is cooled down to room temperature and from which we
distribution is essential for the prediction of their structure and obtain glass transition temperature and thermal expansion coeffi-
properties. In most classical molecular dynamics simulations, static cients; and iii) mechanical testing, where we use non-equilibrium
atomic charges are used. However, during polymerization/cross- MD to subject the annealed samples to deformation and from
linking simulations the charge distribution is expected to change which we obtain elastic constants. The following subsections
6060 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070

START Controlling parameters : cutoff, crosslinking limit, reacting temperature,

max steps for each stage, charge update method, annealed temperature

Data file: force field and packed monomers

No Yes
EEM charge updating ECA? ECA charge updating

NVT equilibration and EEM STOP NVT equilibration

Yes

NPT equilibration and EEM Crosslinking limit? NPT equilibration

No
Identify potential reactive pairs

No pairs within Yes Yes pairs within No

Create new bonds
cutoff distance? cutoff distance?

No Yes
EEM charge updating ECA? ECA charge updating

Multistep relaxation and EEM Multistep relaxation

Fig. 2. Flowchart of the simulation procedure for crosslinking process.

describe the system of interest and simulation details of the various atomistic systems in terms of: i) bonding interactions including
stages of our approach and Section 3 describes the polymerization bond stretch, angle bending, torsions involving dihedral angles and
procedure. improper torsions and ii) non-bond interactions including elec-
trostatics originating from partial atomic charges and van der
2.1. System of interest Waals interactions to describe short range Pauli repulsion and
dispersion. DREIDING does not specify partial atomic charges and
This paper focuses on the conversion of stoichiometrically we use the EEM method to obtain environment-dependent atomic
perfect mixtures of EPON-862 (diglycidyl ether of bisphenol F) and charges for our systems; such a combination of DREIDING with
curing agent EPI-Cure-W (diethylenetoluenediamine, DETDA). The charge equilibration has been successfully used in the past for soft
molecular structures of the two monomers are shown in Fig. 1. To materials, see for example Ref. [19]. The sensitivity of the DREIDING
simulate the polymerization and crosslinking process, the potential predictions of the liquid density (before polymerization) with
reactive sites in the epoxy resin are activated by hydrating the respect to the choice of atomic charge is included as Supporting
epoxy oxygen atoms at the ends of the molecule, see Fig. 1(b). Each Information; these results show that the choice of partial charges
of the amine groups in DETDA can react with two epoxy molecules: is very important in simulations of these epoxy systems. DREIDING
the initial reaction converts a primary amine to a secondary one van der Waals interactions have been parameterized with either
and the subsequent reaction leads to a tertiary amine. The activa- LennardeJones (LJ) or Buckingham (exponential repulsion and
tion energy for the primary reaction is typically lower than the inverse sixth power attraction, X6) functions. The polymerization
secondary one, a difference of 3.7 kcal/mol was obtained from FTIR procedure is performed with LJ functional with inner and outer
measurements in a similar system [16,17]. Consequently the rate of cutoffs of 8 Å and 12 Å respectively and a switching function that
the primary reaction can be significantly higher than that of the take the potential and force to zero between the inner and outer
secondary, especially at low reaction temperatures. In section 3.3 cutoffs. The same approach is used for the calculation of electro-
below, we discuss how the relative rates of chemical reaction static interactions. For the annealing and mechanical testing
affect the thermo-mechanical response of the polymers. All our simulations from which thermo-mechanical properties are
simulations involve systems with perfect stoichiometries, i.e. there computed, we use X6 van der Waals potentials (the preferred form
are twice as many EPON-862 molecules as DETDA ones. We report for DREIDING) with a cutoff of 12 Å, together with an accurate
results of simulation cells of various size with the following description of the long-range electrostatics interactions using the
numbers of epoxy and curing molecules (16, 8), (64, 32), (128, 64) pppm method as implemented in LAMMPS [20]. Thus, we favor
and (256, 128). In the expression (n, m), that will be used computationally less intensive choices for the polymerization
throughout the paper to identify the various systems, n represents procedure where the key goal is to obtain a well-relaxed topology
the number of epoxy monomers and m the number of curing agent and molecular conformations and perform more accurate simula-
monomers. tions to predict thermo-mechanical properties.
The DREIDING force field uses atom types to distinguish iden-
2.2. Atomic interactions tical atoms in different chemical environments and we use addi-
tional atom types to identify reactive atoms. New types indicate
Atomic interactions are described using the DREIDING force reactive carbon atoms as well as reactive nitrogen atoms in primary
field [18]. DREIDING describes the total potential energy of an or secondary amines.
 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070 6061

2.3. Model setup and simulation details Table 1

Parameters used in multigroup multistep relaxations ðrci ¼ rvdw þ ð2i  1Þ
ðrcutoff  rvdw Þ=6Þ.a
The initial model system is built packing both activated epoxy
and agent monomers into a simulation cell followed by a structural Groups Steps Effective force constants Effective bond lengths
relaxation using the commercial software MAPS [21]. All subse- Group i (i ¼ 1,2,3) 1 k0/5 ric
quent MD simulations are performed by using LAMMPS [22] 2 2k0/5 r0 þ 3(ric  r0)/4
3 3k0/5 r0 þ (ric  r0)/2
a massively parallel simulator from Sandia National Laboratories.
4 4k0/5 r0 þ (ric  r0)/4
In all simulations, 3D periodic boundary conditions are imposed to 5 k0 r0
remove possible surface effects. A NoseeHoover thermostat [23] a
In the equation, rci stands for cutoff distance for group i. rvdw is taken as 3.0 Å and
with 100 fs coupling constant and NoseeHoover barostat [24] rcutoff is 5.64 Å.
with 350 fs coupling constant are used for temperature and pres-
sure control respectively.
3.2. Multi-step bond creation approach
3. Polymerization and crosslinking procedure
When a new bond is formed between a pair of reactive atoms
their separation distance is significantly larger than the equilibrium
The polymerization procedure can be divided into three main
bond length. Thus, in order to avoid large interatomic forces
stages: pre-curing equilibration, curing, annealing of the network
between the newly connected atoms we use a multistep relaxation
polymer. Fig. 2 shows the flowchart of the simulation procedure
procedure similar to the one reported by Varshney et al. [10] where
both for EEM and ECA charge updating methods.
bonds are created in five incremental steps. Furthermore, we divide
the reactive pairs that fall between the cutoff distance, rcutoff, and
3.1. Overall polymerization and crosslinking procedure the distance rvdw, at which the interatomic van der Waals potential
is zero, into three groups:
3.1.1. Step 1: pre-chemistry equilibration  .
After packing the desired number of resin and curing agent 8 0<r
rvdw þ rcutoff rvdw 3
monomers into a 3D periodic simulation cell at low density (0.5 g/ < Region1:  .  .
cm3) the system is equilibrated using isothermal and isochoric Region2: rvdw þ rcutoff rvdw 3<r
rvdw þ2 rcutoff rvdw 3
:  .
(NVT ensemble) MD simulations for 50 ps followed by an Region3:
rvdw þ2 rcutoff rvdw 3<r
rcutoff
isothermal, isobaric (NPT) simulation for 400 ps at atmospheric
pressure that takes the liquid to its equilibrium density.
and pairs in each group are assigned a new bond with equilibrium
distance equal to the mean distance of their region and a force
3.1.2. Step 2: polymerization and crosslinking procedure
constant equal to one-fifth of the regular value for the force field,
After the model system is fully equilibrated in step 1, the poly-
see Table 1.
merization and crosslinking is simulated by periodically perform-
The relaxation is divided into five steps to gradually decrease
ing chemical reactions between reactive atoms during an MD
the equilibrium bond length and increase the force constant of the
simulation. Chemical reactions are performed in a stepwise manner
new bonds to the DREIDING values (r0 ¼ 1.41 Å and k0 ¼ 350 kcal/
using a criterion based on atomic distances and, as will be
mol, respectively, for CeN bonds in this system). Table 1 shows the
described in detail in Section 3.3 below, the type of chemical
bond parameters used in this multigroup multistep relaxation
reaction (primary or secondary amine reactions). The distance
procedure. For each step, a 10 ps MD run is performed to relax the
between all pairs of reactive CeN atoms are computed and new
system. Such 50 ps procedure provides time for the system to
bonds are created between all those that fall within a pre-assigned
adjust to its new topology. Dividing the new bonds into three
cutoff distance; based on prior work [10] we take this distance to be
groups based on their initial bond distance is important to avoid
5.64 Å, four times the equilibrium CeN bond length. After the new
large interatomic forces as new bonds are formed. Fig. 3 shows the
bonds are identified all new additional covalent terms (angles,
volume and instantaneous temperature fluctuations during the
torsions, and improper torsions) are created and hydrogen atoms
process of bond formation for a system containing 16 EPON and 8
from the reactive C and N atoms are removed. As will be described
DETDA molecules. We compare our approach with three groups of
below new bonds are turned on slowly using a multistep relaxation
bonds at each of the five bond formation stages with a case where
procedure to avoid large atomic forces. Once this is accomplished,
a single set of bonds parameters is used at each stage. Fig. 3 shows
or if no reactive pairs are found within the cutoff distance, a 50 ps
that using three groups of bonds together with a multistep
NPT simulation is performed to let the system relax and equilibrate
procedure leads to significantly less fluctuations and has a less
to it new topology before reactive pairs are checked again for
disruptive effect on atomic dynamics. Varshney et al. used
possible new reactions. After each 50 ps equilibration run the
a different bond type for each new bond; this approach is
termination conditions are evaluated: the processes ends if a pre-
expected to lead to even smaller disruptions of the dynamics
determined conversion limit or maximum simulation time is sur-
during bond creation as compared with our three-group method.
passed. If the end conditions are not met, another round of
The advantage of our approach lies in its computational simplicity
chemical reactions begins, see Fig. 2. In practice, the conversion
and scalability. Since reactive pairs are grouped into three cate-
limit is set to 100% and the simulations end based on time.
gories as opposed to creating a different bond type for every new
bond, the number of bond types in our approach is independent of
3.1.3. Step 3: thermalization and annealing
system size.
As will be described below crosslinked system is usually built at
temperatures larger than the glass transition one. An annealing
process is used to cool down the system to room temperature or 3.3. Chemical reaction models
other desired temperature. This annealing process is also used to
predict glass transition temperature and the coefficient of thermal As described above, the activation energy for reactions with
expansion. primary amines is lower than that of secondary ones leading to
6062 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070

Table 2
EEM parameters.

Atom type Electronegativity Hardness h*i Atomic size [26]

c*i (eV) (eV) Ri ¼ 14:4=2h*i
(Å)
H 1.0 17.95 0.4011
C 5.25 9.0 0.8
N 8.8 9.39 0.7668
O 14.72 14.34 0.5021

should be determined self consistently. In our simulations, the

electronegativity equalization method (EEM), a semi-empirical
approach parameterized using density functional theory (DFT) is
adopted. Section 4.1 describes details of our implementation of
EEM. Unfortunately charge equilibration models, including EEM,
remain computationally intensive; this limits their use in large-
scale MD simulations. However, one would expect atomic charges
to be somewhat independent of the details of the environment and
depend mostly on topology and chemistry; consequently signifi-
cant charge evolution may only occur during chemical reactions.
Our EEM-based polymerization simulations confirm this expecta-
tion and based on this observation we propose a new method that
captures the physics of charge transfer in a computationally effi-
cient; Section 4.2 describes this approach.

4.1. EEM

The EEM is based on the electronegativity equalization principle

that when two or more different atoms combine to form a molecule
their electronegativities change to a common intermediate value
[25]. In this study, the formalism of EEM developed by Mortier et al.
[12] is adopted.
Fig. 3. Volume and temperature as a function of time during polymerization. The
The EEM energy for an n-atom system is given by:
multigroup multistep relaxation effectively reduces the fluctuations during the crea-
tion of new bonds. 0 1
X
n
k X qi qj
EðqÞ ¼ @E* þ c* qi þ h* q2 þ A (1)
i i i i 2 jsi Rij
higher chemical reaction rates for the primary reaction. Actually, i¼1

for low temperatures it is possible for the primary reactions to where Ei* ; c*i ¼ ðdE i
Þ and h*i ¼ ðddqEi Þy are the coefficients of the
2

dqi y
evolve to completion with little secondary reactions, which only i
Taylor expansion around the atomic value (asterisks indicate
occur at a significant rate when the reaction temperature is the values are from the Taylor expansion). c*i and h*i represent the
increased [17]. On the other hand, at high temperatures primary electronegativity and hardness of ith atom. The EEM electronega-
and secondary reactions are limited by mass diffusion and exhibit tivity of an atom in the system is then given by the derivative of the
similar probabilities. An intermediate process occurs under normal energy with respect to the effective charge:
production conditions, where the exothermicity of the curing
  X qj
process leads to an increase of the relative rate of secondary dE
reactions. ci ¼ ¼ c*i þ 2h*i qi þ k (2)
dqi y R
jsi ij
In order to investigate how processing conditions affect the
network and resulting properties we perform two types of In Eqs. (1) and (2), qi is the charge on atom i, ci is the equilibrium,
conversion simulations: i) high temperature condition where environment-dependent electronic chemical potential of atom i, Rij
primary and secondary reactions have equal probabilities (bonds stands for the internuclear distance between atoms i and j. The
are created between all reactive pairs within the cutoff distance); constant k is a unit conversion factor, k ¼ 14.4 gives energy in eV
this processes is denoted equal probability chemistry. And ii) low to when distances are expressed in angstroms and charge in electrons.
high temperature condition where for a pre-determined amount of In equilibrium, all atoms should have the same chemical potential
time only reactions involving primary amines are enabled, followed and for an n-atom system there are n þ 1 unknowns (n atomic
by a second stage where all reactive pairs within the cutoff distance charges plus chemical potential) and n equations (Eq. (2)). The
are reacted; this second approach is denoted two-stage chemistry additional equation needed to solve the system of equations is
model. P
qi ¼ 0, which enforces charge neutrality. The most accurate
i
descriptions modify the distance dependent on the electrostatic
4. Partial atomic charges during conversion interaction term in Eqs. (1) and (2) to account for the finite-size of
the charge distribution that leads to shielding for short distance
Coulomb or electrostatic energy, calculated based on partial and deviations from the 1/Rij dependence [12,14].
atomic charges, represents an important fraction of the total energy There have been numerous studies on the calibration of
of many polymers. As described above, we expect charges to evolve parameters c*i and h*i for various elements, usually using large
during the polymerization and crosslinking procedure and they training sets containing small organic molecules and finding a set of
 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070 6063

employed in our simulations. These parameters are from Bultinck

a 0.4 et al. [15], except that we use an expression for the internuclear
H bonded with O distance Rij that
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi takes into account shielding corrections:
0.2 R*ij ¼ q3
R3ij þ 3
ffi Rc . The shield radius is obtained from atomic values
ffiffiffiffiffiffiffiffi
Rc ¼ Ri Rj using an expression from Refs. [12,13], see Table 2.
H bonded with C During the polymerization and crosslinking procedure atomic
Charge(e)

0.0 C bonded with O charges are updated every 10 ps during the 50 ps equilibration
stages and every 1 ps during the bond relaxation procedure.

-0.2 4.2. EEM-based charge assignment (ECA)

C between rings
Charge equilibration is a computationally intensive step in
-0.4
O close to rings force field simulations and since it involves a matrix inversion it
does not, in general, scale linearly with the number of atoms
O at epoxy ends
-0.6 making it especially challenging for large-scale simulations. In
0 300 600 900 1200 1500 1800 this section we show that during polymerization and crosslinking
EEM charges evolve in a relatively simple and predictable manner
Time (ps) amenable to a simple parameterization based on atomic
connectivities. The resulting method denoted EEM-based charge
b 0.0
assignment (ECA) provides, once parameterized, a computation-
C ally efficient approach to describe charge evolution during
-0.2 polymerization.
During MD simulations EEM charges exhibit high-frequency
oscillations around their equilibrium values due to thermal atomic
Charge(e)

-0.4 motion and one would expect large charge variations when
chemical reactions occur. An analysis of a polymerization/cross-
N linking run using EEM for a (16,8) system leads to the following
-0.6 general observations regarding charge distribution:

-0.8 i) Charges on atoms not involved in chemical reaction undergo

small fluctuations around their equilibrium values, as shown
in Fig. 4(a).
-1.0 ii) Charges on equivalent atoms fluctuate around the same
0 300 600 900 1200 1500 1800 average value.
iii) Charges on atoms involved in chemical reactions evolve only
Time (ps)
during the reaction in a predictable manner, as shown in
Fig. 4(b) and (c).
c 0.2
This indicates that it should be possible to develop a scheme to
0.0 assign partial charges and update those for atoms involved in
chemical reactions with minimal computational overhead.
C
-0.2 The first step in our approach is to determine the initial atomic
C
charges for the unreacted epoxy resin and curing agenteliquid
Charge(e)

-0.4 mixture. These are obtained from EEM charges during a pre-
crosslinking run. The charge on each individual atom is time-
averaged and as mentioned before atoms with equivalent chemical
-0.6 N
environments exhibit very similar charges and those are further
averaged together. The initial charges for EPON-862 and DETDA
-0.8 based on our method are listed in the Appendix.
The second step involves updating atomic charges of the atoms
-1.0 involved in chemical reactions. Each reactive carbon in EPON-862
0 300 600 900 1200 1500 1800 can react with a nitrogen atom in DETDA and each nitrogen atom
Time (ps) in DETDA can form two new bonds with two EPON-862 carbon
atoms; these reactions with nitrogen can happen simultaneously
Fig. 4. Time evolution of partial charges on various atoms before and after reaction (a) or not. Fig. 4(b) shows the temporal evolution of the charge of
Charges on atoms not involved in chemical reactions; (b) Charges on carbon and a reactive C/N pair and Fig. 4(c) shows a nitrogen atom that reacts
nitrogen atoms undergoing a primary reaction; (c) Charges on carbon and nitrogen with two carbon atoms at two different times. When a nitrogen
atoms with primary and secondary reactions.
atom bonds with one carbon atom, as shown in Fig. 5(a) and (b),
its charge increases by 0.195 e, from 0.755 e to 0.56 e and the
parameters that can give the best fit between the EEM and DFT charge on the carbon atom changes by þ0.2185 e from 0.3328 e
charges. Although all parameterizations are applicable to our to 0.1143 e. Note that the change in total charge is zero consid-
polymers, the parameters are often quite different from one set to ering the charges on two hydrogen atoms that are removed during
another due to differences in training sets and optimization tech- the chemical reaction, one from the nitrogen and another from the
niques used. Therefore, we tested various parameterizations on carbon. When a nitrogen atom forms two new bonds with carbon
a set of small molecules; the parameters listed in Table 2 are atoms, as shown in Fig. 5(c) and (d), its charge changes
6064 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070

Fig. 5. Charges on atoms before and after reaction (a) and (b): DETDA reacting with one EPON; (c) and (d): DETDA reacting with two EPONs.

þ2  0.195 e from 0.755 e to 0.365 e and the charge on each of are obtained from small model molecules in the gas phase (the
the two carbon atoms increases by þ0.2185 e. The net charge actual method to obtain those charges was not described in
change is also zero considering the charges on four removed detail, the authors point out that they use Insight II, a commercial
hydrogen atoms. In summary, when a reaction occurs, the charge molecular modeling software). Our detailed analysis of the
on a nitrogen atom increases 0.195 e whereas the charge on charge evolution during polymerization in the condensed phase
a carbon atom increases 0.2185 e. The change in atomic charge of lends credence to the charge assignments in Ref. [10] and in our
other atoms during the reaction is negligible. Thus, ECA uses paper.
simple rules to describe the charge evolution during chemical
reactions providing a computationally efficient alternative to EEM.
As will be shown below ECA leads to polymerization/crosslinking 5. Polymerization and crosslinking results
results very similar to EEM both in terms of volume contraction
and total energy evolution. A single polymerization simulation 5.1. Role of temperature and charge update on polymerization/
using EEM is required to parameterize ECA for a given set of crosslinking
monomer and curing agent and this can performed on a relatively
small system; once parameterized ECA can be applied to much We performed polymerization/crosslinking simulations using
larger simulations cells. the equal probability chemistry approach for small model systems
It should be pointed out that ECA is similar to the approach (16, 8) using both EEM and ECA charge updates and under various
reported in the paper of Varshney et al. [10] An important temperatures from 300 K to 600 K. The purpose of these simula-
difference is that the ECA approach is parameterized based on tions is to verify the feasibility of ECA method and characterize the
self-consistent charge calculations during the actual process of role of simulation temperature on the network structure produced
polymerization in the condensed phase while charges in Ref. [10] during polymerization and crosslinking.
 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070 6065

a 5000 250
ECA 600K EEM
EEM 600K bond
4000 200 angl
Total energy (kcal/mol)

dihs

Energy change (%)

imps
3000 150
vdws
coul
2000 100 ECA(open)

1000 50
ECA 300K
EEM 300K
0
0
0 20 40 60 80 100
300 400 500 600 700 800
Conversion (%)
Temperature (K)
b 1.5
Fig. 7. Increase in energy components during polymerization as a function of
EEM ECA
temperature. ECA results show the change in energy from 0% to 87.5% conversion; EEM
600K 300K
1.4 500K 400K results show the change from 0% to 78.125%.
Specific volume (A /g)

400K 500K
3

300K 600K
1.3
sizes we see large fluctuations but volume shrinkage is clear as the
crosslinking progresses. The shrinkage rate is in the range of 8e12%
1.2 for 80e90% crosslinked polymers, which are all achieved with the
cutoff distance predefined to 4r0 (r0 ¼ 1.41 Å). The EEM and ECA
1.1 methods give essentially identical results on volume shrinkage.
In order to better understand how conversion proceeds during
1.0 the simulation and optimize our approach we now focus on char-
acterizing how the internal energy of the system evolves during
conversion. As mentioned earlier, care should be taken when
0.9 comparing energies of systems at different conversion degrees.
0 20 40 60 80 100 DREIDING describes the total energy of the system as a sum of
Conversion (%) covalent terms (including bond stretch, angles, dihedral and
improper torsions), van der Waals and electrostatics. Fig. 7 shows
Fig. 6. Comparison of EEM and ECA for (16, 8) systems using the equal probability the percentage increase of the various energy terms during poly-
chemistry method and two different temperatures: (a) Total energy variations as
a function of conversion; (b) Volume reduction as a function of conversion.
merization/crosslinking for various simulation temperatures. The
bond, angle, dihedral and improper energies are each normalized
by the actual number of terms for each conversion degree (i.e. bond
energy per bond, angle energy per covalent angle, etc.). The van der
Fig. 6(a) shows the total energy as a function of conversion degree Waals and electrostatics energy terms are normalized by the cor-
for two temperatures using both EEM and ECA. The conversion responding volume of unit cell. We find that all energy components
degree is defined as the number of new bonds formed out of increase during polymerization/crosslinking but the dihedral
maximum number of pairs of reactive sites. We find an increase in (torsions) energy plays a dominant role increasing over 100% of its
total energy with conversion degree in an approximately linear original value. Furthermore, the increase becomes more important
fashion. Both EEM and ECA exhibit the same trends and very similar as the simulation temperature is decreased. These results show that
increases in total energy during reaction. The total energy from ECA dihedral energy dominates the total increase in internal network
is slightly higher than that of EEM since EEM is able to adjust charges strain during the formation of the 3D crosslinked polymer indi-
in response to the instantaneous configuration. An increase in total cating that these relatively floppy modes are forced away from their
energy during polymerization and crosslinking may seem counter equilibrium value as the network forms and molecular groups loose
intuitive; the reason for this behavior is that the force field we use mobility.
does not take into account the decrease in energy due to the To confirm the trends observed in Fig. 7 regarding the role of
formation of new bonds. Thus, since the number of covalent terms temperature in the internal strain of the network and obtain a more
change with conversion, comparing total energies for different direct measure of the degree of relaxation of structures built using
conversion degrees is meaningless. However, EEM and ECA energies different procedures, we compare the energy of a structure poly-
for the same conversion degree can be compared and Fig. 6(a) shows merized at T ¼ 300 K (denoted low-T sample) and one polymerized
that ECA leads to similar quantitative trends as the more computa- at T ¼ 600 K and then annealed to T ¼ 300 K using a cooling rate of
tionally intensive EEM. Another important quantity to monitor 10 K/200 ps (high-T sample). Both samples have the same degree of
during curing is the specific volume. At each step of the reaction we conversion (87.5%) and energies are averaged over 200 ps at
average the volume after the multistep relaxation and after the T ¼ 300 K. The total energy of the high-T sample is 2200 kcal/mol,
subsequent equilibration run; these average values are plotted along 10% lower that that of the low-T sample, confirming that high
with the conversion degree in Fig. 6(b). As expected for small system temperature polymerization leads to a decrease in internal network
6066 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070

0.30 100
EEM 600K
ECA 600K 90
0.25
Dihedral Energy (kcal/mol)

80
0.20

Conversion (%)
70

0.15
60
Model system
0.10 50 (16,8)
(32,16)
40
0.05 EEM 300K
(64,32)
ECA 300K
30 (128,64)
0.00
(256,128)
0 20 40 60 80 100 20
Conversion (%) 0 500 1000 1500 2000 2500 3000
Time (ps)
Fig. 8. Dihedral energy variations as a function of conversion.

Fig. 10. Conversion degree vs. time for different sizes of model systems.

strain energy. The main terms contributing to this energy differ-

ence are bond angles (w20% difference) and dihedrals (w10% role of system size in the process of polymerization/crosslinking.
difference). Fig. 9 shows the specific volume as a function of reaction
Fig. 8 shows the average dihedral energy per mode (covalent percentage for two different sizes; as expected, specific volume is
dihedral angle terms in the Hamiltonian) as a function of conver- not strongly dependent on size. The volume shrinkage for both
sion degree for T ¼ 300 K and 600 K both for EEM and ECA simu- sizes is about 10%, in good agreement with the simulation
lations. An interesting observation is that the rates of increase of observation of Yarovsky and Evans [4] that a volume reduction of
internal network strain energy with conversion degree grow as the 5e12% occurs for various systems they simulated. Also, it is
reaction progress. This implies that pulling molecules together as comparable to the 7% volume reduction reported by Varshney
new bonds are formed becomes increasingly difficult as the et al. [10] for larger systems. It should be pointed out that Var-
network develops leading to increasing internal strain in the shney et al. did not observe volume shrinkage in small systems
material. but our simulations show significant volume shrinkage even for
small systems and the rate of volume shrinkage is largely inde-
pendent of system size and temperature. Fig. 10 shows conversion
5.2. Polymerization/crosslinking as a function of system size
degree as a function of time for systems of various sizes, from
1000 atoms for the (16, 8) system to 16 000 atoms for (256,128).
We have shown that the ECA method is not only computa-
As expected, conversion proceeds at a fast rate initially and this
tionally efficient but also captures the role of charge evolution
rate is essentially independent of systems size. The rate of
during reactions in the energetics and density of the polymers.
conversion slows down significantly at 60e70% conversion as the
Furthermore, we found that higher simulation temperatures are
mean separation between reactive sites increases and molecular
desirable since they lead to lower internal strain energies. Thus,
mobility decreases due to the formation of the network. This
now we focus on ECA simulations at 500 K and characterize the
range of conversion degrees where we observe conversion rate
slowdown coincides with a steep increase in the molecular
weight of the systems as described in Ref. [10].
1.25

(256,128)
5.3. Role of chemical kinetics
(128, 64)
Specific volume (A /g)

1.20
3

Fig. 11 shows the time evolution of the number of primary,

secondary, and tertiary amines for the two chemistry models dis-
cussed in section 3.3. The equal probability chemistry model, Fig. 11
1.15 (a), shows secondary reactions starting at early times, as soon as
secondary amines become available to react. On the other hand,
Fig. 11(b) shows results from the two-stage model where secondary
reactions are not allowed until primary reactions are almost fully
1.10
completed (time 1.5 ns). Comparing these two extreme reaction
models at an overall conversion percent of w85%, we find for the
equal probability model (corresponding to high curing tempera-
1.05 ture) 6.25% primary amines (unreacted) and 15.5% secondary
0 20 40 60 80 100 amines (those that have undergone one reaction). The two-stage
Conversion (%) models lead to a lower number of primary amines (2.75%) with
23.5% of secondary amines. Interestingly, the number of tertiary
Fig. 9. Specific volume as a function of conversion for (128, 64) and (256, 128) systems. (i.e. fully reacted) amines is similar in both cases: w78% for the
 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070 6067

1.15
a 100
Cooling:10K/200ps

80
1.10 Tg=420~430K
Conversion (%)

Density (g/cm )
3
60
Primary%
1.05
Secondary%
40
Tertiary%

20 1.00

0
0 1000 2000 3000 4000 5000 6000 0.95
300 350 400 450 500 550 600 650
Time (ps)
Temperature (K)
b 100
Fig. 12. Density as a function of temperature for X6 van der Waals interactions and for
the two chemistry models: equal probability (red) and two-stages (blue) (for inter-
pretation of the references to colour in this figure legend, the reader is referred to the
80 web version of this article).
Conversion(%)

60
Primary%
involving a primary amine are allowed, every reaction reduces the
Secondary%
40 number of chain ends by two. This leads, in average, to a larger
Tertiary% density increase than when a secondary amine reacts reducing the
number of chain ends by one.
20

6. Thermo-mechanical properties of the resulting networks

0
0 1000 2000 3000 4000 5000 6000 In this section we characterize the thermo-mechanical proper-
Time (ps) ties of the resulting 3D network polymers. All these simulations are
performed with the (256, 128) systems containing 15,198 atoms
c 1.00 with a conversion degree of w78%.
Equal probability
Primary first 6.1. Glass transition temperature
0.95
After polymerization and crosslinking, the samples are cooled
Density (g/A )
3

down to room temperature and the glass transition temperature

0.90 and thermal expansion coefficient are obtained from these
simulations. As described earlier these simulations are carried
out using the Buckingham (X6) form for van der Waals
0.85 interactions.
We cool down the system from 600 K to 300 K at a rate of 10 K/
200 ps under atmospheric pressure. The density as a function of
0.80 temperature is plotted in Fig. 12 for both chemistry models. The
0 20 40 60 80 100 predicted density at T ¼ 300 K (w1.12 g/cm3) is in good agreement
Conversion (%) with the experimental value of (w1.17 g/cm3) [27]. The change in
slope of the densityetemperature curves denotes the glass tran-
Fig. 11. Percentage of primary, secondary and tertiary amines as a function of time for sition. To obtain the transition temperature and the error in this
the (a) equal probability and (b) two-stage chemistry models. c) Density as a function calculation, linear fits were performed to the MD densi-
of conversion degree for the two chemistry models.
tyetemperature data below and above Tg in various temperature
ranges. This procedure leads to a predicted Tg in the range
420e430 K for both chemistry models. These results show that the
equal probability approach and w74% for the two-stage model. order in which chemical reactions occur has little effect on the
Fig. 11(c) shows the density evolution during the process of room temperature density and the glass transition temperature for
conversion for the two chemical models. These simulations indi- the high conversion degree tested here. For a slightly higher
cate that the density for the two-stage model is slightly higher than degree of conversion, 86%, the equal probability chemistry model
the equal probability approach for intermediate conversion leads to a slightly higher glass transition temperature of about
degrees, in the range of 40e80%. For higher conversion degrees the 450 K.
two chemistry models lead to essentially identical densities. The The reported experimental Tg values of EPON-862/DETDA are
likely reason for this observation is that when only reactions 413 K [28], and 417e432 K [29]. The actual conversion degree of the
6068 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070

0.040 polymer tested in these experiments was not reported but typical
values lie between approximately in the range 80e95% [17]. Tg is
0.035 a kinetic quantity and depends on cooling/heating rate; an increase

/K
-6
10
of 3 K is expected per order of magnitude increase in rate [30]. Thus,

2x
0.030
the Tg predicted from our MD (with a rate of 5  1010 K/s) should be

17
(L-L0)/L0(at 300K)

α=
0.025 about 30 K higher than the experimental value, i.e. the expected Tg
for MD rates is 445e460 K. Our MD predictions, between 420 and
0.020 450 K, (for conversion degrees between 78 and 86%) using
DREIDING are very close to this experimental range. In comparison,
0.015 the Tg value given by Varshney et al. [10] using the CVFF force field
-6 /K is significantly lower (378 K).
0.010 10
7x
α =8
0.005 6.2. Coefficient of thermal expansion

0.000
The coefficient of thermal expansion (CTE) can then be calcu-
300 350 400 450 500 550 600 650
lated from the simulation cell lengths during our isobaric cooling
Temperature (K) simulations described in Section 6.1, see Fig. 13. The simulation cell
is maintained cubic throughout the simulation and an equal
Fig. 13. Cell length normalized with T ¼ 300 K value as a function of temperature for
probability model is shown. The coefficient of linear thermal
polymer with w78% degree of conversion. Coefficients of thermal expansion before
and after Tg are shown. expansion is defined as:

 
a 250
a ¼
1 vL
(3)
E=3.5 GPa L0 vT P
E=3.2 GPa where L0 is the box size at 300 K and ðvL=vTÞP is the slope of linear
200
part of the box size change vs. temperature. Our prediction for CTE
using the Buckingham potential is 87  106 K1 for temperatures
below Tg and we find a ¼ 172  106 K1 above Tg. Our calculations
Stress (MPa)

150
are in good agreement with experiments: 61  106 K1 below and
195  106 K1 above Tg [8]. Note that the accuracy of the MD
predictions is higher above Tg where the polymer exhibits shorter
100
relaxation times. Prior MD simulations on similar epoxy systems
Temperature: 300K showed similar level of agreement with experiments, Ref. [10]
predicts 40  106 K1 at room temperature, and Ref. [8]
50 Conversion: ~78% 55  106 K1 below Tg and 184  106 K1 above Tg.
8 -1
Strain rate: 5x10 s
0 6.3. Elastic constants

0 5 10 15 20 We also calculated the elastic constants for the system described

Strain (%) in the previous subsection (built using the equal probability
chemistry model) via non-equilibrium MD simulations. We uni-
b 10
Temperature: 300K
axially strain the system while maintaining atmospheric pressure
in the transverse directions and obtain the Young’s modulus and
Conversion: ~78% Poisson ratio by fitting the MD data to linear functions up to strains
8 8 -1
Strain rate: 5x10 s within the linear response region of the system. Fig. 14(a) shows the
MD engineering stressestrain curve for a strain rate of 5  108 s1
up to 20% longitudinal strain and Fig. 14(b) shows transverse vs.
ε y (%)

6 longitudinal strains. As compared with experimental results for

epoxy systems well below their glass transition temperatures the
υ=0.43 MD data in Fig. 14 exhibits a higher strength and lack of brittle
Strain -

4 failure. These are general features of all MD simulations of polymer

systems cause by two key factors: i) the extremely large strain rates
υ=0.47 in MD (due to the limitations in total simulation time), and ii) the
2 inability of the material to localize strain because of small sample
sizes.
From the data shown in Fig. 14 we obtain a Young’s modulus in
the range 3.2e3.5 GPa and a Poisson ratio of 0.43e0.47. These
0
ranges are obtained by fitting the MD data up to a maximum of
0 5 10 15 20
various maximum strain values between 2% and 4%. The MD
Strain ε z (%) predictions are close to the experimental values reported for this
system: Ref. [26] reports a Young’s modulus of 2.76 GPa while
Fig. 14. Predicted stressestrain curve (a) and transverse strain vs. longitudinal strain
Ref. [31] obtained an average Young’s modulus of 2.65 GPa and
(b) together with resulting Young’s modulus and Poisson’s ratio. A running average of
the MD data with a 2 ps averaging period (strain of 0.095%) is shown to reduce the a Poisson ratio of 0.35. As mentioned above, Wu and Xu [7] found
thermal fluctuations. extremely large elastic constants (w50 GPa) for DGEBA/IPD using
 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070 6069

DREIDING with self-consistent charges, this result is inconsistent

with ours results on EPON-862/DETDA and other polymers and
might be erroneous.

7. Conclusions

In summary we proposed a method to mimic the process of

polymerization and crosslinking in thermoset polymers based on
MD simulations and characterized it via its application to the EPON-
862/DETDA system. Our approach builds on previous efforts and
provides justification to performing charge updates only during
chemical reactions. We explored in detail, for the first time, how
differences in chemical reactions related to processing conditions
affect the resulting structures and thermal properties of the
network. Our simulations show that whether primary reactions
occur first followed by secondary ones or both reactions exhibit
identical reaction rates from the onset does not affect the proper-
ties of the resulting network for conversion degrees higher than
80%.
Due to the limitation in timescales accessible to MD, the total
simulation time for the polymerization/crosslinking processes is
only a few nanoseconds, several orders of magnitude shorter than
in experiments. The high rates of chemical reactions could lead to
poor relaxation and larger internal strains in the network produced.
Our results indicate that performing the simulations at high
temperatures to increase molecular mobility mitigates this
problem.
Using the general-purpose molecular force field DREIDING
with Buckingham van der Waals interactions and EEM-based
charges, our approach leads to predictions of density, glass tran-
sition temperature, thermal expansion coefficient, and elastic
constants in good agreement with experiments. The predicted
density is only slightly lower than the experimental value, this
could be due to the high reaction rates used in the simulations; we
are currently investigating how timescales affect the predicted Fig. A1. Initial partial atomic charges on an EPON-862 monomer.

properties. The glass transition temperature obtained from our

MD simulations is between 420 and 430 K for a conversion degree
of 78% and w450 K for 86%. Once the effects of cooling rate are
taken into account these results are within the experimental
range. As noted in the original DREIDING paper, Buckingham van
der Waals potentials lead to more accurate predictions than
LenardeJones, which underestimates the density of the cross-
linked polymer by about 10%.
Finally, the proposed approach is generally applicable to any
thermoset system and we are currently applying it to other
systems. Obtaining realistic atomistic structures of thermoset
polymers is a key step towards the development of physics-
based models that can be used to quantitatively predict and,
eventually, optimize the properties of this important class of
materials.
Fig. A2. Initial partial atomic charges on a DETDA monomer.

Acknowledgements References

This work was supported by a grant with The Boeing Company [1] Griffiths B. High Performance Composites. 2005;5:1e4.
[2] Pascault JP, Sautereau H, Verdu J, Williams RJJ. Thermosetting polymers. New
and the US National Science Foundation (NSF) under contract York: CRC Press; 2002.
CMMI- 0826356. Useful discussions with G. Medvedev, J. A. Car- [3] Doherty DC, Holmes BN, Leung P, Ross RB. Comp Theor Polym Sci
uthers, R. B. Pipes, S. Christensen and J. Goose and computational 1998;8:169e78.
[4] Yarovsky I, Evans E. Polymer 2002;43:963e9.
resources of nanoHUB.org are gratefully acknowledged.
[5] Heine DR, Grest GS, Lorenz CD, Tsige M, Stevens MJ. Macromolecules
2004;37:3857e64.
[6] Gou J, Minaie B, Wang B, Liang ZY, Zhang C. Comp Mater Sci 2004;31
Appendix (3e4):225e36.
[7] Wu CF, Xu WJ. Polymer 2006;47:6004e9.
[8] Fan HB, Yuen MMF. Polymer 2007;48:2174e8.
Initial charges on atoms of EPON-862 and DETDA. [9] Lin PH, Khare R. Macromolecules 2009;42:4319e27.
6070 C. Li, A. Strachan / Polymer 51 (2010) 6058e6070

[10] Varshney V, Patnaik SS, Roy AK, Farmer BL. Macromolecules [22] LAMMPS (Large-scale Atomic/Molecular Massively Parallel Simulator), open
2008;41:6837e42. source code, http://www.cs.sandia.gov/wsjplimp/lammps.html.
[11] Komarov PV, Chiu YT, Chen SM, Khalatur PG, Reineker P. Macromolecules [23] Hoover WG. Phys Rev A 1985;31:1695.
2007;40:8104e13. [24] Hoover WG. Phys Rev A 1986;34:2499.
[12] Rappe AK, Goddard III WA. J Phys Chem A 1991;95:3358e63. [25] Sanderson RT. Chemical bonds and bond energy. New York: Academic Press; 1976.
[13] Mortier WJ, Genechten KV, Gasteiger JJ. Am Chem Soc 1985;107:829e35. [26] Note: In using LAMMPS, the parameters for calculating atomic charge are
[14] Njo SL, Fan JF, van de Graaf BJ. Mol Catal A Chem 1998;134:79e88. denoted as and, where gi ¼ 1/Ri.
[15] Bultinck P, Langenaeker W, Lahorte P, De Proft F, Geerlings P, Waroquier M, [27] http://www.resins.com/Products/TechnicalDataSheet.aspx?id¼3950.
et al. Phys Chem A 2002;106:7887e94. [28] Sun L, Warren GL, O’Reilly JY, Everett WN, Lee SMD, Lagoudas D, et al. Carbon
[16] Wang XR, Gillham JK. J Appl Poly Sci 1991;43:2267e77. 2008;46:320e8.
[17] Schiering DW, Katon JE. J Appl Poly Sci 1987;34:2367e75. [29] Tao K, Yang SY, Grunlan JC, Kim YS, Dang B, Deng Y, et al. J Appl Poly Sci
[18] Mayo SL, Olafson BD, Goddard III WA. J Phys Chem 1990;94:8897e909. 2006;102:5248e54.
[19] Jang SS, Goddard WA. J Chem Phys C 2007;111:2759e69. [30] Ferry JD. Viscoelastic properties of polymers. New York: John Wiley & Sons
[20] Hockney RW, Eastwood JW. Computer simulation using particles. NY: Adam Inc.; 1980.
Hilger; 1989. [31] Tack JL. Thermodynamic and mechanical properties of EPON 862 with curing
[21] MAPS (The Materials And Processes Simulations platform). Scienomics Inc.; agent DETDA by molecular simulation. Master’s thesis, Texas A&M University,
2008. 2006.