You are on page 1of 23

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260972858

A multiscale modeling technique for bridging molecular dynamics with finite


element method

Article  in  Journal of Computational Physics · November 2013


DOI: 10.1016/j.jcp.2013.06.039

CITATIONS READS

12 298

2 authors:

Yongchang Lee Cemal Basaran


University at Buffalo, The State University of New York University at Buffalo, The State University of New York
9 PUBLICATIONS   66 CITATIONS    229 PUBLICATIONS   3,662 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Unified Mechanics Theory: Unification of Newtonian Mechanics and Thermodynamics Laws View project

Thermal Mechanical Behavior of BGA Solder Joints View project

All content following this page was uploaded by Cemal Basaran on 10 December 2018.

The user has requested enhancement of the downloaded file.


Journal of Computational Physics 253 (2013) 64–85

Contents lists available at SciVerse ScienceDirect

Journal of Computational Physics


www.elsevier.com/locate/jcp

A multiscale modeling technique for bridging


molecular dynamics with finite element method
Yongchang Lee ∗ , Cemal Basaran
Electronic Packaging Laboratory, Department of Civil, Structural, and Environmental Engineering, State University of New York at Buffalo,
United States

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

Article history: In computational mechanics, molecular dynamics (MD) and finite element (FE) analysis
Received 1 March 2013 are well developed and most popular on nanoscale and macroscale analysis, respectively.
Accepted 30 June 2013 MD can very well simulate the atomistic behavior, but cannot simulate macroscale
Available online 12 July 2013
length and time due to computational limits. FE can very well simulate continuum
Keywords:
mechanics (CM) problems, but has the limitation of the lack of atomistic level degrees of
Multiscale modeling freedom. Multiscale modeling is an expedient methodology with a potential to connect
Weighted averaging momentum method different levels of modeling such as quantum mechanics, molecular dynamics, and
Molecular dynamics continuum mechanics. This study proposes a new multiscale modeling technique to couple
Wave reflection MD with FE. The proposed method relies on weighted average momentum principle.
A wave propagation example has been used to illustrate the challenges in coupling
MD with FE and to verify the proposed technique. Furthermore, 2-Dimensional problem
has also been used to demonstrate how this method would translate into real world
applications.
© 2013 Elsevier Inc. All rights reserved.

1. Introduction

The insatiate demand for multiscale analysis is not only due to advances in nanotechnology, but also due to experimen-
tal results proving that there is a need for connecting nanoscale physics and macroscale continuum analysis. Significant
advancements in computational power make it feasible to link both powerful methods: molecular dynamics (MD) and finite
element (FE) methods.
MD and FE methods are well suited to a particular level of accuracy on atomistic and continuum simulations, respec-
tively. In general, MD cannot be used for macroscale problems due to the restrictions on the number of atoms that can be
simulated simultaneously, along with the time scale limit. On the other hand, usage of FE method for atomic scale problems
is not accurate for many reasons mainly because continuum mechanics assumes that the substance of body is distributed
continuously throughout the space of body and lacks atomic degrees of freedom. These inherent limitations make connect-
ing these two methods essential but also challenging. Nevertheless, multiscale modeling will allow us to solve complicated
problems with a greater accuracy than ever before. It should be pointed out that MD does not have electronic degrees of
freedom. However, it is expected that methods like the one proposed here will allow us to connect FE, MD, and quantum
mechanics, which has electronic degrees of freedom.

* Corresponding author.
E-mail address: yl83@buffalo.edu (Y. Lee).

0021-9991/$ – see front matter © 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.jcp.2013.06.039
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 65

2. Physical background

2.1. Molecular dynamics

Molecular dynamics (MD) is a statistical method about motion and interaction of atoms. The classical mechanics which
governs the molecular dynamics of a system can be derived from Hamiltonian formulation. The Hamiltonian H is the total
energy of a system that is the sum of kinetic energy and potential energy and is given by:

H MD = K MD + V MD (1)
 p2
i
K MD = (2)
2mi
V MD = Φ(r ) (3)

where K MD is kinetic energy of an atom i with mass mi and momentum p i , V MD is potential energy, and Φ(r ) is the in-
teratomic potential energy of particles with distance, r. The potential energy can be extended depending on the interatomic
potential. The equations of motion in MD are obtained from the Hamiltonian by the following relations:

∂H
= ẋ (4)
∂p
∂H
= − ṗ (5)
∂x
where x is the coordinate and p is the momentum. In these relations, the total energy of a system is conserved by showing
that the time derivative of Hamiltonian is zero: (dH /dt ) = 0. By using Eqs. (4) and (5), the equation of motion is derived
and given by

mi q̈i = f iext − f iint (6)

An atom of mass, mi , moves as a rigid particle with acceleration, q̈i . f iext is the external force. When interatomic potential
is known, the internal force f iint acting on a conservative system can be obtained by f iint = ∇ri V .
In classical molecular dynamics, interactions between neighboring atoms are determined by interatomic potential and are
very crucial for acquiring physically meaningful results. For pair potential, Lennard-Jones potential and Morse potential are
commonly used. Simulations of metals require solving the many-body problems, which lead to development of many-body
potentials. Finnis–Sinclair potential, embedded atom method (EAM), and modified embedded atom method (MEAM) have
been commonly used for many-body potentials.

2.2. Continuum mechanics

Continuum mechanics (CM) deals with the analysis of the kinetics and the behavior of solid or fluid modeled as contin-
uum. The concept of continuum assumes that the substance of body is distributed throughout the space of body and ignores
the fact that the matter consists of atoms, vacancies and atomic degrees of freedom. Due to the assumption of a continuous
and differentiable mass density, a differential equation can be used to solve problems in continuum mechanics. In real life,
objects are very large compared to atoms. Thus, at the macroscopic scale, continuum mechanics provides an appropriate
statistical procedure.
The continuum mechanics is based on two types of equations. One is the laws applied to the entire domain such as
the conservation of energy and mass. Another kind of equation describes the behavior of materials such as constitutive
equations. The equation of motion in continuum mechanics can be derived by the Hamiltonian as:

H CM = V CM + K CM (7)

1
V CM = ε · C · ε dΩ (8)
2
Ω

1
K CM = ρ u̇ 2 dΩ (9)
2
Ω

where ε is the strain tensor, C is the material constitutive tensor, ρ is the material density and u̇ is the nodal velocity. It is
well known that FE can very well simulate the continuum mechanics problems.
66 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

3. Literature review

Pioneering techniques for multiscale methods are the quasi-continuum (QC) method by Tadmor (1996) [1] and macro-
scopic, atomistic, ab initio dynamics (MAAD), or also called as coupling of length scales (CLS), by Abraham et al. (1998) [2].
Based on these techniques, various methods for multiscale modeling have been proposed such as coarse-grained molec-
ular dynamics (CGMD) by Rudd and Broughton [3], bridging domain method (BDM) by Xiao and Belytschko (2004) [4],
and bridging scale method (BSM) by Wagner and Liu (2003) [5] and Park and Liu (2004) [6]. In the following section a
brief review of multiscale modeling techniques is presented. We believe this is necessary to put the proposed method in a
context.

3.1. Macroscopic, atomistic, ab initio dynamics (MAAD)

MAAD by Abraham et al. [2,7] is one of the earlier methods for multiscale simulation. The fundamental idea of this
approach is to make concurrent links between tight-binding (TB) method, molecular dynamics (MD), and finite element
method (FE). In this method, tight-binding method is used for quantum mechanics level degrees of freedom. Molecular dy-
namics is used for the representation of atomistic degrees of freedom. Finite element method is used for the deformation of
continuum mechanics. Here, all three simulations run at the same time, and dynamically communicate required information
between the simulations.
The interactions among three analyses are taken into account by the total Hamiltonian of the system as follows:

H total = H FE (d, ḋ) + H FE/MD (d, ḋ, r , ṙ ) + H MD (r , ṙ ) + H MD/TB (r , ṙ ) + H TB (r , ṙ ) (10)


In this model, ‘handshake’ region was adopted to couple regions with each other in H FE/MD . A very thin handshake region
is used. FE mesh is graded down to the atomic size for the reduction of wave reflection between MD and FE [8]. However,
when connecting molecular dynamics and continuum mechanics by using MD simulation and FE method respectively, this
technique uses the atomic scale mesh size for FE. However, the latter approach leads to two problems: one numerical and
one physical. The numerical issue is that simulation time of FE slows down to picosecond to match the MD time step. Also
FE time step is governed by the smallest element of FE. The physical issue is that atomic scale FE simulation is physically
unreasonable because of the fact that constitutive equation of FE is based on continuum mechanics. Because of the fact that
the time step in FE region depends on the element size, the atomic sized mesh makes the time step too short for realistic
engineering problems. Moreover, although Abraham et al. [2,7] mentioned that there is no visible reflection at the FE–ME
handshake region, they did not discuss the error due to reflection of short wavelength in MD region. The reflection of short
wavelength at the interface between the ME and FE region still exists because atoms on the FE mesh side are stationary
while atoms on the MD side are mobile.

3.2. Coarse-grained molecular dynamics (CGMD)

Rudd and Broughton (1998) developed coarse-grained molecular dynamics (CGMD) approach [3]. CGMD is based on a
statistical coarse graining approximation. This approach removes tight-binding method from MAAD, and links FE and MD.
A key idea of the method is that degrees of freedom are eliminated by using the coarse-graining approximation that con-
verges to the exact atomic energy to reduce the computational cost. The coarse-graining energy for a mono-atomic harmonic
solid of N atoms coarse-grained to N node nodes is stated to be

1
E (uk , u̇k ) = U int + ( M jk u̇ j u̇k + u j K jk uk ) (11)
2
j ,k

Here U int is the internal energy defined as U int = 3( N − N node )kT where k is Boltzmann constant and T is temperature.
The first term of summation M jk u̇ j u̇k is the kinetic energy and the second term of summation u j K jk uk is potential energy.
M jk is mass matrix, K jk is stiffness matrix and u j and u̇ j are the displacement and velocity of node j, respectively. In this
approach, the vital region of simulation is modeled by MD, while peripheral regions are discretized with coarse-graining.
The interface of MD and FE regions is modeled such that CGMD imitates the motion of an FE mesh. Rudd (2001) introduced
the generalized Langevin dynamics into the CGMD formulation. The equation of motion is then given by:

t
−1
M i j ü j = −G ik uk + ηik (t − τ )u̇k (τ ) dτ + F i (t ) = G i j (12)
−∞

where G ik is elastic Green’s function which is defined as u i = G i j f j in static where u i and f j are the CGMD displacement
field and body force respectively, ηik is memory function or a time history, u̇k (τ ) is velocity at node k and F i (t ) is a random
force.
Similar to MAAD, mesh size of CGMD is graded down to atomic scale at the MD region, and coarsened far from the MD
region. Thus, CGMD also experiences the same issues as MAAD such as time step limitation, the wave reflection, and the
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 67

Fig. 1. Illustration of Cauchy–Born rule: (left) an element and atoms that are determined by Cauchy–Born rule, and (right) the deformed element and
homogeneously moved atoms by following the deformed shape of an element.

picosecond total simulation time. In order to reduce spurious wave reflection, the additional terms, the second and third
terms in Eq. (12), are introduced. This technique is very similar to bridging scale method by Liu et al. (2003) which is
also reviewed in the following section. These additional terms lead to additional force calculations in MD simulation which
already suffers from the limited simulation time due to the computational cost. Considering that the most expensive part
of MD simulation is calculations, it is a serious limitation of the method.

3.3. Quasi-continuum (QC) method

Another pioneering approach for multiscale methods is the quasi-continuum (QC) method by Tadmor (1996) [1]. The QC
method is an approach coupling continuum mechanics with atomistic simulation for the mechanical response of polycrys-
talline materials at zero temperature. The QC method is based on an entirely atomistic description of the material domain.
To reduce the computational cost, two assumptions are adapted: one is the reduction of degrees of freedom, and another
is the Cauchy–Born rule: in a crystalline solid subject to a small strain, the positions of the atoms within the crystal lattice
follow the overall strain of the medium as depicted in Fig. 1. The Cauchy–Born rule assumes that the continuum energy
density W can be obtained by using an atomistic potential, with the link to the continuum being the deformation gradient
F given by:
du
F =1+ (13)
dX
where u is the displacement, d X is an undeformed line segment.
By using the Cauchy–Born rule, a continuum stress tensor and tangent stiffness can be acquired from the interatomic
potential W , which allows the usage of nonlinear FE techniques. The continuum stress tensor and tangent stiffness are
given by:
∂W
P= (14)
∂F
∂2W
C= (15)
∂FT∂FT
where P is the first Piola–Kirchoff stress tensor and C is the Lagrangian tangent stiffness.
The particular representation is determined by the local deformation gradient and dictates a small fraction of the atoms
(called representative atoms or “repatoms”). In this approach, the non-local repatoms are used to represent the atomistic
behaviors, and the local repatoms are used to simulate the continuum domain by using the Cauchy–Born rule in the FE
method.
Although QC method suggested a new approach for multiscale modeling, this method suffers from the same issues as
MAAD, that are spurious wave reflection and the total simulation time limit. In this method, even though Cauchy–Born rule
connects atoms in MD region with repatoms in FE, in which the mesh size of FE gradually increases from MD region, the
spurious wave reflection still exists in MD region [9]. The result leads to spurious energy accumulation in MD region, non-
physical heating of the crystal in the MD region, and as a result the solution in MD region becomes unreliable. Moreover,
since this method is implemented in MD and FE regions simultaneously, the time step of MD dominates the total simulation
time, which is very short for any practical engineering problem.

3.4. Bridging domain method

Xiao and Belytschko (2004) [4] have developed a coupling method for molecular dynamics and continuum mechanics
based on a bridging domain method. In this approach, the system consists of three domains: ΩMD (molecular dynamics),
ΩCM (continuum mechanics), and ΩHS (handshake region). The main idea of the model is using a linear combination of
Hamiltonian on the handshake region, ΩHS . Hamiltonian is defined by:

H = (1 − α ) H MD + α H CM
  p iM · p iM  p iC · p iC
= 1 − α( Xi ) + (1 − α ) V MD + α( Xi ) + α V CM (16)
2mi 2M i
i i
68 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

where p iM is momentum of MD region, p iC is momentum of CM region, parameter α = [0, 1] (linear) in ΩHS , α = 0 in


ΩMD − ΩHS , and α = 1 in ΩCM − ΩHS . Lagrange multipliers method is applied to enforce displacement compatibility in the
handshake region between the molecular and continuum regions by imposing the following equation

 
g I = u i ( X I ) − di I = N J ( X I )u i J − d i I =0 (17)
J

i.e. the atomic displacements are required to conform to the continuum displacements at the handshake region. The con-
straints are applied to all components of the displacements. In the Lagrange multiplier method, the total Hamiltonian and
the equation of motion are written as:

H L = H + λT g = H + λTI g I (18)
I
 ∂g J
mi q̈i = f iext − f iint − f iL = f iext − f iint − λTJ (19)
∂ qi
J

where λ is a vector of Lagrange multipliers whose components correspond to the components of the displacement q i of
atom i. f iL is the force due to the constraints enforced by the Lagrange multipliers. ∂ g J /∂ qi in Eq. (19) is introduced by
substituting the Hamiltonian in Eq. (18) into the relations in Eqs. (4) and (5).
As shown above in Eq. (16), the energy within the handshake region goes from entirely atomistic at MD boundary
to entirely continuum at FE boundary. The effect of this energy transition is that short wavelength atomic scale energy
is filtered. The idea of spatial filtering is proven by the numerical examples in [4]. The example shows that a minimum
handshake distance is required for the method to eliminate wave reflection effectively. The minimum handshake distance in
this method is relatively long, as a result increasing the computational cost and decreasing the size of MD zone.

3.5. Bridging scale method

The fundamental idea of bridging scale developed by Wagner and Liu (2003) [5], and Park and Liu (2004) [6] is to resolve
the total displacement u (x) in terms of course scale u (x) and fine scale u  (x) at the position x. The coarse scale is governed
by the continuum mechanics and simulates the entire field, while the fine scale is used to simulate the region of high
interest.
The total displacement u (x) is resolved into the course scale u (x) and fine scale u  (x) as follows:

u (x) = u (x) + u  (x) (20)

The coarse scale and the fine scale are defined as, respectively

u( Xα ) = Nα
I dI (21)
I

u = u − P u (22)

where N αI = N I ( X a ) is the shape function on atomic position X a , d I is the nodal displacement of FE, and P is the projection
matrix which is determined by minimizing the mass-weighted square of the fine scale. This method assumes that the result,
q, of any atomistic level simulation could be used to generate an exact solution. Thus, the total displacement is given by:

u = Nd + q − P q (23)

Hamiltonian formulation generates the coupled multiscale equations of motion as follows:

M d̈ = N T f (u ) (24)
ma q̈ = f (25)

where M is the mass matrix of FE, and ma is the atomic mass matrix of MD.
Bridging scale method starts from an entire molecular system. To save computational time, in this approach, the system
area of MD is reduced from the entire region to a small area of interest. An entire molecular system can be changed into
the reduced MD system along with external forces that act on the boundaries of the reduced lattice. The latter represents
the combined effects of all the atomistic degrees of freedom accounted for by using the generalized Langevin equation
(GLE). The effect of using the GLE in conjunction with FE is the dissipation of small wavelengths which FE cannot capture
because FE can only capture longer wavelengths that are on the order of the FE mesh spacing or larger. The application of
GLE generates the external force on the equation of motion for MD simulation. The additional force f imp is applied at the
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 69

Fig. 2. An example of 1-D longitudinal wave; the arrow represents the direction of displacement of particles, and the amplitude of wave represents the
displacement.

boundary of MD region, defined by:


ncrit t
imp  
f m (t ) = θm−m (t − τ ) q0,m (τ ) − u 0,m (τ ) − R 0,m (τ ) dτ (26)
m =−ncrit 0

where θm−m (t − τ ) is the time history kernel (or memory kernel) function that describes renormalization of the atomic
interaction along the boundary of MD domain, ncrit refers to the maximum number of atomic neighbors with MD bound-
ary atoms, and R 0,m (τ ) is the stochastic displacement that accounts for external force in FE domain. Thus, the coupled
multiscale equations of motion are given by:

ma q̈ = f (q) + f imp (t ) + R f (t ) (27)


T
M d̈ = N f (u ) (28)
where R f (t ) is the random force that accounts for thermal effect in FE region.
This approach does not scale down the mesh size of FE to the atomic size, as a result provides different simulation time
scales on both MD and FE sides. It yields excellent wave reflection results because the time history kernel generated by
the generalized Langevin equation (GLE) leads to reduced elastic wave reflection. In comparing bridging scale method to the
other multiscale methods (CGMD, MAAD), one clear advantage of this approach is that FEM simulates the entire system, and
is not graded down to the atomic scale. The result of this is that long time step used in FE region is not restricted by the
atomic sized elements in the mesh. This allows the staggered time integration algorithm. Thus, the coarse scale variables
can evolve on an appropriate time scale, while the fine scale variables can evolve (appropriately) on a much smaller time
scale.
However, although the generalized Langevin equation is a mathematically exact representation of the MD degrees of
freedom, in multiple dimensions, the time history integral is hard to compute and brings additional computational cost in
MD simulation [5], because, in multiple dimensions, the calculation of time history kernel becomes more complex and has
to use numerical techniques increasing the computational cost. The most expensive part of this method is the computational
cost which is computing forces between atoms. When the computational limitation of MD is considered, the calculation of
additional force can be a significant disadvantage on multiscale modeling.

4. Challenges of bridging molecular dynamics with finite element analysis

The main problem on a concurrent multiscale modeling is the spurious wave reflection which is generated in the hand-
shake region between MD and FE [3–6,8–25]. Two kinds of wave reflection can occur: one is due to the mesh size in FE
region where representation of short wavelength is not possible [3–6,8–25], and another is due to the different wave dis-
persion speed in each domain [24]. Wave reflection causes serious accuracy problems mainly in the MD region because
high frequency waves cannot transfer into FE region. They are reflected back into MD region. In order to illustrate the wave
reflection problem, an example of 1-Dimensional longitudinal wave is used, Fig. 2. The waves used for the example are
illustrated in Fig. 3, and are applied to an atom of which the initial position is zero. Fig. 3(a) is for long wavelength and
low frequency, Fig. 3(b) is for short wavelength and high frequency, and Fig. 3(c) is the sum of these two waves shown in
Fig. 3(a) and (b).

4.1. Wave reflection due to the different wave dispersion speeds

Wave propagation speed is a crucial quantity to be able to build a suitable handshake region between the continuum
and the atomistic description regions. In order to demonstrate the effect of coupling domains with different wave dispersion
speeds, only a low frequency wave shown in Fig. 3(a) is applied in the 1-D example. High frequency wave can cause wave
reflection also due to element size. Therefore, it is not used for this example. The details of the wave reflection due to high
frequency wave will be discussed in the following section.
On the one hand, when the wave propagation has the same dispersion speed on both regions, no wave reflection happens
as shown in Fig. 4. On the other hand, coupling two domains with different wave dispersion speeds leads to wave reflection
70 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 3. Applied displacement time-history at x = 0; (a) long wavelength, (b) short wavelength, and (c) the wave that is a summation of (a) and (b).

Fig. 4. Wave with long wavelength travels from left to right with no reflection when the wave propagation speed is same on both sides.

Fig. 5. Wave with long wavelength travels from a faster wave speed region (bold line) to a slower wave speed region.

at the interface as shown in Figs. 5 and 6. Fig. 5 shows the result when wave travels from faster speed region to slower
speed region. In Fig. 5, in the slower wave speed region, amplitude of displacement decreases and the wavelength is shorter.
The faster wave speed region in Fig. 5 exhibits a concave wave reflection. Fig. 6 shows the result for when the wave travels
from slower region to faster region. In Fig. 6, in the fast wave speed region, amplitude of displacement increases and the
wavelength is bigger. The slower wave speed region in Fig. 6 exhibits a convex wave reflection. These wave reflections can
generate serious error in MD region. Accordingly, we have to check if the momentum is conserved when discretization scale
changes in a multiscale analysis. Because of these reflections, conservation of momentum will not be satisfied in the MD
region.

4.2. Wave reflection due to the steep change in mesh size between MD and FE

The mesh size for FE is much larger than the interatomic distance used in MD. This different mesh size leads to wave
reflection in bridging molecular dynamics with finite element method. For FE scale mesh size, representation of short
wavelength at the atomic scale is not possible, since we need at least two discretization points per wave length.
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 71

Fig. 6. Wave with long wavelength travels from a slower wave speed region (bold line) to a faster wave speed region.

Fig. 7. Full MD simulation (entire domain is modeled with MD).

Fig. 8. Equilibrium element size in FE is 4 times larger than equilibrium atomic distance of MD. Straight line is MD region (red on the web version of this
article), and circle is FE region (blue on the web version of this article).

Fig. 9. Equilibrium element size in FE is 10 times larger than equilibrium atomic distance of MD. Straight line is MD region (red on the web version of this
article), and circle is FE region (blue on the web version of this article).

In order to demonstrate the wave reflection due to relatively large element size in FE region, short wavelength shown
in Fig. 3(b) is used. In order to compare the wave transfer between MD and FE, Fig. 7 shows the results when full MD
simulation is done. Figs. 8 and 9 show wave propagation and reflection using a different mesh size in FE domain. In Figs. 8
and 9, MD domain is connected with FE domain without using a handshake region. In all cases, the applied wave is moving
from left to right. When nodal distance in the FE region is 4 times the equilibrium atomic distance, results are shown in
Fig. 8, where the wave with short wavelength is coarsely transferred into FE domain. In Fig. 8, some part of wave is reflected
at the border between MD and FE domains, and dispersion speed of the transferred wave decreases in FE domain. Fig. 9
shows the result in which the nodal distance in FE region is 10 times the equilibrium atomic distance. In Fig. 9, the wave
with short wavelength is perfectly reflected at the border between MD and FE. This reflection happens because the element
size in FE region is too large to represent the short wavelength of the applied wave.
72 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 10. Full MD simulation (entire domain is modeled with MD).

Fig. 11. Element size in FE is 4 times the equilibrium atomic distance. Straight line is MD region (red on the web version of this article), and circle is FE
region (blue on the web version of this article).

Fig. 12. Element size in FE is 10 times the equilibrium atomic distance. Straight line is MD region (red on the web version of this article), and circle is FE
region (blue on the web version of this article).

4.3. Wave reflection for long and short wavelengths

In the above example, we can observe two kinds of wave reflection: one is due to large element size in FE region, and the
other is due to different wave dispersion speed between the two regions. In this section, a wave with both long and short
wavelengths shown in Fig. 3(c) is applied at an atom, located at x = 0. In this example both domains have the same wave
dispersion speed, so we can see the effect of the applied wave with respect to the element size. Fig. 10 shows the results of
full MD simulation where both sides of the border are discretized with MD. Figs. 11 and 12 show the results of bridging MD
with FE without using handshake region. The results are similar to the earlier example for short wavelength propagation.
The wave with long wavelength is successfully transferred into FE domain from MD region. The short wavelength cannot be
transferred into FE region and is reflected at the border of MD and FE regions (Fig. 12). However, Fig. 11 shows that when
element size in FE region is 4 times the equilibrium atomic distance of MD, additional error is introduced at the border
of MD and FE regions. In Fig. 12, when element size in FE region is 10 times the equilibrium atomic distance, the error
disappears. It becomes more clear when the reader compares the displacement amplitude at the border among Figs. 10, 11,
and 12.

5. New multiscale modeling approach, WAMM

5.1. Introduction

The atomic motion in MD simulation contains short wavelength that FE region cannot represent. It means the total
momentum in MD domain cannot be transferred into FE domain in a concurrent multiscale analysis. The undelivered mo-
mentum generates noise (spurious additional energy) in the MD region. In this work, to solve this problem, we couple the
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 73

Fig. 13. Schematic description for 1-D coupling with handshake region.

momentum of MD and FE in the handshake region ΩHS . The coupled momentum p coupled is given by:

p coupled = α p MD + (1 − α ) p FE (29)
where p MD and p FE are the momentum of MD and FE region respectively. α is an averaging parameter which is zero in
FE domain ΩFE and at the border near FE region in ΩHS , and 1 in MD domain ΩMD and at the border near MD region in
ΩHS . The schematic description of each domain and parameter α for 1-D case is illustrated in Fig. 13. By using this coupled
momentum, total Hamiltonian can simply be represented as:

H total = K coupled + V MD + V FE (30)



N 2
( p coupled )
K coupled = (31)
2mi
i
N
atoms
ext
V MD = V (r i j ) + V MD (32)
i

1 ext
V FE = σi j εi j dΩFE + V FE (33)
2
ΩFE

where N and N atoms are the number of total particles (including atoms and finite element nodes) and atoms, respectively.
MD potential V MD includes atoms not only in MD region, but also the atoms in the handshake region. V MD can be extended
depending on the interatomic potential used for MD simulation. Superscript ext is used for the external effect. In the
potential energy of FE region, V FE , σi j and εi j are stress tensor and strain tensor, respectively.
In the handshake region, atoms of MD and nodes of FE are overlapped and nodal positions of overlapped elements
are determined by the positions of overlapped atoms. The coupling details are shown in Section 5.3. A weighted average
momentum combination is enforced in the handshake region ΩHS as shown in Eq. (29).
In order to determine the averaging parameter α , a linearly decreasing kinetic energy of MD, K MD , is considered in the
handshake region first.
N N
atoms
p 2MD,i atoms
(βi0.5 p MD,i )2
β K MD = β = (34)
2mi 2mi
i i

where the parameter βi varies linearly from zero to one in handshake region ΩHS , p MD,i is the momentum of atom i, and
mi is the mass of atom i. In the above form, the decreasing momentum βi0.5 p MD,i can be used for a momentum averaging
equation. Introducing (29) into (34) yields the following relationship:

αi = βi0.5 (35)
In the following section, we prove that this linear averaging scheme reduces the elastic wave reflection. However, some
insignificant wave reflection still remains. To reduce the wave reflection even more, generalized equation (36) is introduced
using parameter p instead of constant 0.5 that is used in Eq. (35), and is given by:

αi = βip (36)
The effects of weighted averaging of momentum and parameter p are discussed below.

5.2. Equations of motion

In this work, a new total Hamiltonian is introduced by bridging momenta. Substituting the Hamiltonian in Eq. (30) into
the relations in Eqs. (4) and (5) gives:
∂H ∂ K coupled p coupled
= = = ẋ (37)
∂ p coupled ∂ p coupled m
∂H ∂ V MD ∂ V FE
= + = − ṗ coupled (38)
∂x ∂x ∂x
74 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 14. Application of Cauchy–Born rule for 1-D coupling example; dots are atoms, and circles are finite element nodes.

Fig. 15. The force contributions of atoms in ΩHS and virtual atoms in ΩCM .

When the above equations are considered for each domain, the equations of motion can be given by:

p MD ∂ V MD ∂ V MD
= q̇, = − ṗ MD → ma q̈ = − = f MD in ΩMD (39)
ma ∂q ∂q
p coupled ∂ V MD ∂ V MD
= q̇hs , = − ṗ coupled → ma q̈hs = − = f hs in ΩHS (40)
ma ∂ qhs ∂ qhs
p FE ∂ V FE ∂ V FE
= ḋ, = − ṗ FE → M d̈ = − = f FE in ΩFE (41)
M ∂d ∂d
where ma and M are corresponding mass of atoms and the nodes. f MD and f FE are corresponding force acting on atoms
and nodes. q and d are the displacement of atoms and nodes respectively. qhs and f hs are the displacement and interatomic
force obtained by using coupled momentum p coupled in handshake region ΩHS .

5.3. Cauchy–Born rule for MD/FE coupling

We now present the details of Cauchy–Born rule to couple MD with FE. In order to exploit the benefit of using an
independent mesh size in MD and FE regions, the nodal space in FE region is not scaled down to interatomic spacing.
However, to couple MD with FE using a new multiscale method, FE nodes’ degrees of freedom in handshake region must be
the same as MD’s degrees of freedom. Accordingly, Cauchy–Born rule described in QC method [1] is applied in this study.
The different degrees of freedom are illustrated in Fig. 14. In handshake region ΩHS , MD has 21 atoms, and FE has
3 nodes. In order to have the same degrees of freedom in both regions, Cauchy–Born rule is applied to nodes in FE region.
This application generates 18 virtual atoms in handshake region ΩHS : 9 virtual atoms per an element. Through Cauchy–Born
rule, the number of total particles of FE in handshake region ΩHS becomes 21 that consists of 3 nodes and 18 virtual atoms.
The virtual atoms outside of handshake region ΩHS , which are also called ghost atoms or pad atoms, are generated for
the force calculation of MD simulation. If there are no virtual atoms outside of handshake region ΩHS , an atom at the
boundary of MD region has only the force from the handshake region and as a result cannot be in equilibrium. Fig. 15
shows the atomistic 1-D chain with a cutoff radius of 4r0 where r0 is equilibrium interatomic distance. The number of
virtual atoms outside of handshake region ΩHS is decided by cutoff radius in force calculation of MD. We must have virtual
atoms in the FE region to satisfy the equilibrium of forces at zero atom location. When we do not consider virtual atoms,
the force acting on atom 0, f 0 , is given by:
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 75

−1

f0 = fi (42)
i =−4

where f i is the force contribution of atom i to atom 0. The consequence is imbalance in the forces acting on the atom 0.
By introducing virtual atoms in FE region ΩFE , the atoms 1, 2, 3, and 4 can exert force on atom 0. When the distance
between atoms is the same, the equilibrium of forces at atom 0 is satisfied. The force equilibrium at atom 0 is given by:


4
f0 = fi (43)
i =−4, i =0

5.4. Time integration algorithm

The equations of motion for all domains are derived in the previous section. The fundamental time step algorithm
in handshake region is the same as the classical MD simulation. In this work, velocity Verlet algorithm is used for MD
simulations because of its efficiency and accuracy. Through the relation of momentum, p = mv, the coupled momentum
equation in Eq. (29) can be represented by

mcoupled v coupled = αmMD v MD + (1 − α )mFE v FE (44)


where mcoupled , mMD , and mFE are coupled mass of atoms, mass of atoms and mass of virtual atoms in handshake region re-
spectively. v coupled , v MD , and v FE are coupled velocity of atoms, velocity of atoms and velocity of virtual atoms in handshake
region respectively. If the atomic masses, mcoupled , mMD and mFE , are same, the above equation can be rewritten as:

v coupled = α v MD + (1 − α ) v FE in ΩHS (45)


Utilizing the Cauchy–Born rule at FE mesh in the handshake region, the following relationship can be used to obtain the
virtual atom’s displacement dhs from the nodal displacement d in handshake region:

dhs = Nd (46)
where N is a shape function. Similarly, the virtual atom’s velocity, ḋhs , and the virtual atom’s acceleration, d̈hs can be
calculated by the equation which is given by:

ḋhs = N ḋ (47)
d̈hs = N d̈ (48)
where ḋ is the nodal velocity and d̈ is the nodal acceleration in handshake region.
In order to utilize the benefit of independent time steps, staggered time integration algorithm is used in the proposed
multiscale modeling approach. The time step is defined by t = mtm , t is used in FE region, and tm is used in MD and
handshake regions. [ j] will be shorthand for the fractional time step n + ( j /m). Superscripts are used to denote the time
step. In the time integration algorithms, it is assumed that the initial conditions are known.
In MD domain ΩMD , displacement q, velocity q̇, and acceleration q̈ of atoms are updated via velocity Verlet algorithm as
follows:
1 [ j] 2
q[ j +1] = q[ j ] + q̇[ j ] tm + q̈ tm (49)
2
 
q̈[ j +1] = ma−1 f MD q [ j +1 ]
(50)
1 
q̇[ j +1] = q̇[ j ] + q̈[ j ] + q̈ [ j +1 ]
tm (51)
2
where ma is mass of atoms, and tm is time step of MD.
Similarly, in handshake region ΩHS , displacement qhs , velocity q̇hs , and acceleration q̈hs of atoms are updated as follows:
[ j +1 ] [ j]
ḋhs = ḋhs + d̈nhs tm (52)
[ j +1 ] [ j] [ j] 2 1 [ j]
qhs = qhs + q̇hs tm + q̈hs tm (53)
2
[ j +1 ]  [ j +1 ] 
q̈hs = ma−1 f MD qhs (54)
[ j +1 ] [ j] 1  [ j] [ j +1 ] 
q̇trial = q̇hs + q̈hs + q̈hs tm (55)
2
[ j +1 ] [ j +1 ] [ j +1 ]
q̇hs = αq̇trial + (1 − α )ḋhs (56)
where q̇trial is a trial velocity.
76 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 16. Displacement time-history at x = 0.

Once quantities in the MD and handshake regions are obtained using the above algorithm at time n + 1, the nodal
displacement d, the nodal velocity ḋ, and the nodal acceleration d̈ are updated from time step n to time step n + 1. These
updates use a central difference scheme:

n +1 dn + ḋn t + 12 d̈n t 2 in ΩFE


d = (57)
( N T N )−1 N T qnhs+1 in ΩHS
 
d̈n+1 = M −1 f FE d n +1
(58)
1 
ḋn+1 = ḋn + d̈n + d̈ n +1
t (59)
2
If we assume that no external force acts upon the system, f FE = f FE
int
. The internal force in FE region is computed by the
following equation:
int
f FE = Kd (60)
where K is the stiffness matrix, and d is the nodal displacement vector. In the handshake region of Eq. (57) we transfer
atomic positions calculated in MD to nodal positions by assuming that the MD simulation results in the exact displacement
solution to the problem.

5.5. Conservation of energy

So far we described the detailed methodology of a weighted averaging momentum method (WAMM) to couple molecular
dynamics with finite element method. We now consider the energy conservation of system in proposed method, WAMM.
Suppose that we have the material at certain temperature T without motion over time. In this material, although kinetic
energy at continuum scale is zero, the kinetic energy at atomistic scale has a certain value corresponding to the tempera-
ture T . Therefore, we need to use thermal energy in the continuum region to explain the energy conservation that nodal
displacement cannot represent in finite element analysis. Total energy in FE region will be subdivided by potential energy
V FE , kinetic energy K FE , and thermal energy T FE . Note that kinetic energy and potential energy partly overlap with thermal
energy. Thus, total energy E total of whole system is given by:

E total = E MD + E FE = ( K MD + V MD ) + ( K FE + V FE + T FE − E overlap ) (61)


Here, E overlap is the overlapping energy. The energy term, T FE − E overlap , is for the part that nodal displacement in continuum
mechanics cannot describe but atomistic displacement in molecular dynamics can.
When the wave in Fig. 3(c) is applied to molecular dynamics region in multiscale model, total energy in molecular
dynamics region, E MD , can be subdivided by transportable energy E transportable and reflected energy E reflected [24]:

E MD = E transportable + E reflected (62)


Here, E transportable and E reflected are corresponding to K FE + V FE and T FE − E overlap in Eq. (61) respectively. Multiscale modeling
techniques, BDM, BSM, and WAMM, remove the reflected energy at the interface during elimination of the wave reflection.
In order to conserve the total energy of system in multiscale model, the reflected energy in molecular dynamics region
needs to be transferred into thermal energy in the handshake region. Otherwise the temperature in the MD region increases
continuously.

6. 1-D wave propagation example

6.1. Details

In a 1-D wave propagation example, displacement is propagated from MD to FE zone. The prescribed displacement used
in the analysis is plotted in Fig. 16.
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 77

Fig. 17. Schematic details of 1-D wave propagation example.

 
e −((t −7)/xa ) − u c
2

u (t , x = 0) = A 1 + b cos (t − 7) (63)
1 − uc H

where u c = e −( L c /xa ) , A = ha = 0.005, xa = 500ha , L a = 25xa , b = 0.1, and H = xa /4. Applied units are nanometer (nm) for
displacement and picosecond (ps) for time. As shown in Fig. 17, the total number of atoms used for MD simulation is 181,
the number of (overlapped) atoms in the handshake region is 20, the number of (overlapped) element in the handshake
region is 2, and the total number of elements for FE is 100. The region designated for MD is 0  x  22.23 nm. The
handshake region is 22.23 nm < x  24.7 nm. The time step in MD and handshake region is 0.002 ps and the time step in
FE is 0.02 ps.
Lennard-Jones (LJ) potential, ΦLJ , is used for the interatomic potential which is given by:
 12  6
σ σ
ΦLJ (ri j ) = 4ε − (64)
ri j ri j
Here, r i j is interatomic distance between atoms i and j, ε = 0.2 J and σ = 0.11 nm and the equilibrium bond length is
r0 = 21/6 σ . The cutoff radius rc is 2.5r0 . The interatomic force in MD simulation can be obtained by following equation:
 12  6
∂Φ(ri j ) 48ε σ 1 σ
f (r i j ) = − = − (65)
∂ ri j ri j ri j 2 ri j

In continuum mechanics, the energy density W is obtained by using LJ 6-12 potential. As following the stiffness defini-
tion in Cauchy–Born rule in Eq. (15), the stiffness k0 between two nodes in 1-D is defined as:

∂ 2 W r0 624ε 168ε
k0 = = − (66)
∂ F T ∂ F T r =r0 l0 214/6 σ 2 28/6 σ 2
where l0 is the initial nodal distance.

6.2. Results

Weighted averaging momentum in the handshake region yields the different levels of reduction in the high frequency
wave reflection depending on the value of parameter p. To optimize the value of parameter p, a parametric study for
different p values was conducted. Fig. 18 shows the displacement at 17 ps with two overlapped elements in the handshake
region. Normalized kinetic energy in MD region at 17 ps is plotted in Fig. 19. MD region kinetic energy is normalized with
respect to the maximum value. As shown in Figs. 18, 19, and 20, the wave reflection and normalized kinetic energy have
the minimum value for p = 0.01. Parameter p below and above 0.01 has a larger wave reflection and larger kinetic energy.
To optimize the length of handshake region, various lengths of handshake region are studied and presented in Figs. 21,
22, and 23. Figs. 21 and 22 show that using only one overlapped element in handshake region shows the worst kinetic
energy transfer. Increasing the number of overlapped elements in handshake region shows excellent kinetic energy transfer.
Among them, the case of two overlapped elements in the handshake region has the minimum wave reflection with the
benefit of shortest handshake region, Fig. 23.
The amplitude of the input low frequency wave is A = 0.005 nm, and amplitude of input high frequency is 2 Ab =
0.001 nm. In Fig. 24, influence of a weighted averaging momentum in handshake region is compared with and without
handshake region. When handshake region is not used, the maximum displacement in FE is 5.024 × 10−3 nm, and the
maximum displacement in MD is 0.9479 × 10−3 nm. It is obvious that the deformation and associated energy of short
wavelength is not transferred into the FE region. When handshake region is introduced, for the case of p = 0.01 which has
the minimum wave reflection, the maximum displacement in FE is 5.13 × 10−3 nm, and the maximum displacement of MD
is 7.266 × 10−5 nm. When handshake region is used, the maximum displacement in FE with handshake region increases
78 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 18. Parametric study for different p values at 17 ps with two overlapped elements in handshake region.

Fig. 19. Normalized kinetic energy in MD region versus parameter p values at 17 ps with two overlapped elements in handshake region.
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 79

Fig. 20. Time history of normalized kinetic energy in MD region for different p values.

by 2.11%, while the high frequency wave’s amplitude in MD region is reduced by 92.73%. It implies that when WAMM
technique is used, the most of the high frequency wave kinetic energy is dissipated.
After 18 ps, kinetic energy in MD region with p = 0.01 is almost zero, while that of MD without handshake region is
still large. This result indicates that kinetic energy of high frequency motion is almost dissipated in the handshake region,
while kinetic energy of high frequency wave in MD without handshake region is reflected at the boundary of MD region
and remains in MD region until the end of simulation. The spurious kinetic energy in MD region, without handshake
region, leads to noise in the computation of MD region. In contrast, the values of kinetic energy in FE region with and
without handshake region are almost same because high frequency energy cannot be transferred to the FE region because
of element size. In Fig. 25, the curve of normalized kinetic energy of FE using a weighted averaging momentum shows a
parallel shift because the wave arrives later in FE region due to the handshake region. In Fig. 25 after 18 ps, kinetic energy in
MD region when there is no handshake region periodically fluctuates near a certain large energy value due to the reflected
high frequency energy.

6.3. Comparison with other multiscale modeling techniques

MAAD [2] and QC method [1] have two main issues: one is the short wavelength reflection at the interface and another
is the time step limitation that the time step of FE must be the same scale as that in MD region. BDM [4], BSM [5,6],
and the technique proposed in this paper, weighted averaging momentum method (WAMM), solve the short wavelength
reflection problem and show similarly excellent results in eliminating the wave reflection. The direct comparison of 1-D
wave propagation example is shown in Figs. 26 and 27 for BDM and BSM respectively. In Fig. 26, the results of BSM and
the proposed technique WAMM are almost same. The results in Fig. 27 are very similar but have a small difference in
continuum region. It may be because BSM uses meshfree method for the analysis in continuum region while we use finite
element methods for WAMM. The details of 1-D example are in Refs. [4,5].
Among all the multiscale techniques discussed above only, BDM, BSM, and the proposed technique WAMM use separate
time steps in FE and MD regions. In spite of these independent time steps, computational cost of the multiscale method is
still the most important criterion in choosing a multiscale approach. To estimate the increasing computing load due to the
multiscale modeling, we consider a typical case with nhs atoms in handshake region and n B atoms at MD boundary. In one
MD time step tm , BDM increases computational cost on the order of O(nhs ), BSM increases computational cost on the order
of O((2ncrit + 1)n H n B ) where the period of time history kernel is n H tm [22] and ncrit is referred in Eq. (26), and WAMM
proposed here increases computational cost on the order of O(nhs ). When these values are normalized by n B , the normalized
value, nhs /n B , of BDM is about 64 [4], the normalized value of BSM is (2ncrit + 1)n H [22], and the normalized value of
WAMM is 20. In this comparison BSM could theoretically have a smaller number of normalized value of computations by
using smaller ncrit . However, it has to be emphasized that the calculations for each atom in WAMM are much simpler than
BDM and BSM. While BDM and BSM have to execute complex force calculations, WAMM has a simple weighted averaging
velocity calculation. For this reason, we believe that the proposed technique WAMM has a big advantage, over BSM and
BDM, in reducing the computational cost of multiscale analysis.
80 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 21. Time history of normalized kinetic energy in MD region along the number of elements in handshake region using p = 0.01.

Fig. 22. Normalized kinetic energy in MD region versus the number of overlapped elements in handshake region with p = 0.01 at 17 ps.

Fig. 23. Amplitude of the first reflected waves depending on the number of overlapped elements in handshake region.
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 81

Fig. 24. Effect of using a weighted averaging momentum method.

Fig. 25. Energy transfer from MD to FE for two overlapped elements and p = 0.01.

Fig. 26. Comparison of the proposed weighted averaging momentum method (WAMM) with BDM [4] for 1-D wave propagation example [4].

7. 2-D wave propagation example

In a 2-D wave propagation example, displacement is propagated from MD to FE region. The displacement used in the
analysis is applied to one atom at the center of model in x direction, and is given by:
 
e −((t −7)/xa ) − u c
2

u (t , x = 0, y = 0) = A 1 + b cos (t − 7) (67)
1 − uc H
82 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 27. Comparison of the proposed weighted averaging momentum method (WAMM) with BSM [5,6] for 1-D wave propagation example [6].

Fig. 28. Schematic details of the model used for 2-D wave propagation example.

where u c = e −( L c /xa ) , A = ha = 0.005, xa = 500ha , L a = 25xa , b = 0.3, and H = xa /4. Applied units are nanometer (nm) for
displacement and picosecond (ps) for time.
As shown in Fig. 28, the total number of atoms used for 2-D simulation is 154 241, the number of (overlapped) atoms in
the handshake region is 23 512, the number of (overlapped) element in√the handshake region is 112, and the total number
of elements for FE is 32 000. The region designated for MD is |x|  64 √ 3r0 and | y |  216r
√ 0 . The equilibrium bond length
r0 is same as in 1-D example. The region designated for handshake is 64 3r0 < |x|  80 3r0 and 216r0 < | y |  240r0 . The
time step in MD and handshake region is 0.002 ps and the time step in FE is 0.02 ps. Parameter p is 0.01.
In molecular dynamics and handshake region, Lennard-Jones (LJ) potential ΦLJ is used for the interatomic potential given
by Eq. (64) and same parameters and cutoff radius as in 1-D example are used. The interatomic force in MD simulation can
be obtained by Eq. (65). In continuum mechanics, the constitutive relationship and constitutive matrix in 2-D are defined
as:
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 83

Fig. 29. A four-noded element and 6 atoms of distance r0 with the atom 0 of which interatomic stiffness is k.

Fig. 30. The cases of force generation for an atom in x and y directions: (a) force generation at the surface in x direction by missing atoms 4 and 5, (b) force
generation in y direction at the surface by missing atoms 2, 3, and 4, and (c) force generation in y direction at the surface by missing atom 3.

σi j = [C ]εi j (68)
⎡ ⎤
c 11 c 12 c 13
[C ] = ⎣ c 21 c 22 c 23 ⎦ (69)
c 31 c 32 c 33
To obtain the components of constitutive matrix, following Cauchy–Born rule, atomic positions in an element are il-
lustrated in Fig. 29. Because the interatomic distance r0 dominates the force contribution, the three cases in Fig. 30 are
considered. We assume that the set of atoms are connected with a spring of stiffness k. The force acting on an atom from
the surface by strain εi j can be calculated and be translated to the stress σi j . The constitutive components c i j are obtained
by using the constitutive relationship given by Eq. (68).

∂ 2 Φ(ri j ) 624εσ 12 168εσ 6 72ε
k= = − = 1/3 2 (70)
∂ ri2j r i j =r0 =2 σ
1/ 6 r 14
0 r 8
0
2 σ
√ √
3 3 3
c 11 = c 22 = k, c 21 = c 12 = c 33 = k, c 13 = c 31 = c 23 = c 32 = 0 (71)
4 4

The calculated constitutive matrix shows isotropy and plane stress conditions (elastic modulus E = (2/ 3)k, Poisson’s ratio
ν = 1/3). For a 2-D linear elastic system, the stiffness matrix can be defined as:

K= [ B ]T [C ][ B ] dV (72)

In this 2-D example, a four-noded rectangular finite element is used with the size, l x = 8 3r0 , l y = 12r0 .
Fig. 31 shows the contour of the result from 2-D wave propagation at 16 ps. As Fig. 31 shows, when a √ handshake region
is not used, while long wavelength is transferred from MD to FE, short wavelength is reflected at |x| = 80 3r0 ≈ 17.11 nm
84 Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85

Fig. 31. Contour of 2-D wave propagation example at 16 ps: (left) direct coupling without the handshake region, and (right) the proposed weighted
averaging momentum method (WAMM).

that is the border of MD and FE regions. The reflected wave travels in MD region resulting in spurious energy accumulation
in MD region. However, when WAMM is utilized, the short-wavelength wave reflection is successfully eliminated during
wave propagation. This result indicates that the parameter p and handshake length, which are determined in 1-D example,
work successfully at 2-D example as well.
Comparison with other methods at 2-D is not possible, because there is no published data of 2-D wave reflection ex-
amples in the literature. Programming other multiscale methods for a 2-D example is almost impossible due to the lack of
details in the published papers.

8. Conclusions

In this study, a weighted averaging momentum method, WAMM is introduced for multiscale modeling and shows excel-
lent coupling results in 1-D and 2-D examples through the transfer of the energy and displacement. This method does not
require scaling down FE mesh size to MD atomic resolution. As a result MD region and FE region have independent time
steps. We use staggered time integration algorithms. Moreover, because this approach has a short handshake region, it has
the benefit of reducing computational cost significantly. The biggest advantage of WAMM is the fact that it is a simple way
to link MD with FE. On handshake region, FE nodes coarsely represent the MD energy, and MD velocity is averaged with FE
velocity. By using this approach, the spurious wave reflection at the interface is dramatically reduced.

Acknowledgements

The research project has been sponsored by US Navy Office of Naval Research (ONR) Advanced Electrical Power Systems
program under the direction of program director Dr. Peter Cho.

Appendix A. Supplementary material

Supplementary material related to this article can be found online at http://dx.doi.org/10.1016/j.jcp.2013.06.039.

References

[1] E.B. Tadmor, M. Ortiz, R. Phillips, Quasicontinuum analysis of defects in solids, Philos. Mag. A 73 (1996) 1529–1563.
[2] F.F. Abraham, J.Q. Broughton, N. Bernstein, E. Kaxiras, Spanning the continuum to quantum length scales in a dynamic simulation of brittle fracture,
Europhys. Lett. 44 (1998) 783–787.
[3] R.E. Rudd, J.Q. Broughton, Coarse-grained molecular dynamics and the atomic limit of finite elements, Phys. Rev. B 58 (1998) R5893–R5896.
[4] S.P. Xiao, T. Belytschko, A bridging domain method for coupling continua with molecular dynamics, Comput. Method. Appl. M. 193 (2004) 1645–1669.
[5] G.J. Wagner, W.K. Liu, Coupling of atomistic and continuum simulations using a bridging scale decomposition, J. Comput. Phys. 190 (2003) 249–274.
[6] H.S. Park, W.K. Liu, An introduction and tutorial on multiple-scale analysis in solids, Comput. Method. Appl. M. 193 (2004) 1733–1772.
Y. Lee, C. Basaran / Journal of Computational Physics 253 (2013) 64–85 85

[7] J.Q. Broughton, F.F. Abraham, N. Bernstein, E. Kaxiras, Concurrent coupling of length scales: Methodology and application, Phys. Rev. B 60 (1999)
2391–2403.
[8] L.E. Shilkrot, R.E. Miller, W.A. Curtin, Multiscale plasticity modeling: coupled atomistics and discrete dislocation mechanics, J. Mech. Phys. Solids 52
(2004) 755–787.
[9] R.E. Miller, E.B. Tadmor, The quasicontinuum method: Overview, applications and current directions, J. Comput.-Aided Mater. 9 (2002) 203–239.
[10] W. Cai, M. de Koning, V.V. Bulatov, S. Yip, Minimizing boundary reflections in coupled-domain simulations, Phys. Rev. Lett. 85 (2000) 3213–3216.
[11] J. Knap, M. Ortiz, An analysis of the quasicontinuum method, J. Mech. Phys. Solids 49 (2001) 1899–1923.
[12] E. Lidorikis, M.E. Bachlechner, R.K. Kalia, A. Nakano, P. Vashishta, G.Z. Voyiadjis, Coupling length scales for multiscale atomistics–continuum simulations:
Atomistically induced stress distributions in Si/Si3N4 nanopixels, Phys. Rev. Lett. 87 (2001).
[13] W.N. E, Z.Y. Huang, A dynamic atomistic–continuum method for the simulation of crystalline materials, J. Comput. Phys. 182 (2002) 234–261.
[14] W.A. Curtin, R.E. Miller, Atomistic/continuum coupling in computational materials science, Model. Simul. Mater. Sc. 11 (2003) R33–R68.
[15] N.M. Ghoniem, E.P. Busso, N. Kioussis, H.C. Huang, Multiscale modelling of nanomechanics and micromechanics: an overview, Philos. Mag. 83 (2003)
3475–3528.
[16] B. Liu, Y. Huang, H. Jiang, S. Qu, K.C. Hwang, The atomic-scale finite element method, Comput. Method. Appl. M. 193 (2004) 1849–1864.
[17] D.D. Vvedensky, Multiscale modelling of nanostructures, J. Phys.-Condens. Mat. 16 (2004) R1537–R1576.
[18] H. Kadowaki, W.K. Liu, A multiscale approach for the micropolar continuum model, CMES-Comp. Model. Eng. 7 (2005) 269–282.
[19] X.T. Li, E. Weinan, Multiscale modeling of the dynamics of solids at finite temperature, J. Mech. Phys. Solids 53 (2005) 1650–1685.
[20] A.C. To, S.F. Li, Perfectly matched multiscale simulations, Phys. Rev. B 72 (2005).
[21] W.K. Liu, E.G. Karpov, H.S. Park, Nano Mechanics and Materials: Theory, Multiscale Methods and Applications, John Wiley, Chichester, England and
Hoboken, NJ, 2006.
[22] S.Q. Tang, T.Y. Hou, W.K. Liu, A pseudo-spectral multiscale method: Interfacial conditions and coarse grid equations, J. Comput. Phys. 213 (2006) 57–85.
[23] S.Q. Tang, A finite difference approach with velocity interfacial conditions for multiscale computations of crystalline solids, J. Comput. Phys. 227 (2008)
4038–4062.
[24] K. Fackeldey, The weak coupling method for coupling continuum mechanics with molecular dynamics, Dissertation, Universität Bonn, 2009.
[25] J.D. Lee, X.Q. Wang, Y.P. Chen, Multiscale computation for nano/micromaterials, J. Eng. Mech.-ASCE 135 (2009) 192–202.

View publication stats

You might also like