Sie sind auf Seite 1von 45

Semi-analytical Solution for the Generalized Absorbing

Boundary Condition in Molecular Dynamics Simulations

Chung-Shuo Lee, Yan-Yu Chen, Chi-Hua Yu, Yu-Chuan Hsu, Chuin-Shan Chen1

Department of Civil Engineering, National Taiwan University, Taipei 10617, Taiwan

Abstract

We present a semi-analytical solution of a time-history kernel for the generalized

absorbing boundary condition in molecular dynamics (MD) simulations. To facilitate

the kernel derivation, the concept of virtual atoms in real space that can conform with

an arbitrary boundary in an arbitrary lattice is adopted. The generalized Langevin

equation is regularized using eigenvalue decomposition and, consequently, an

analytical expression of an inverse Laplace transform is obtained. With construction of

dynamical matrices in the virtual domain, a semi-analytical form of the time-history

kernel functions for an arbitrary boundary in an arbitrary lattice can be found. The time-

history kernel functions for different crystal lattices are derived to show the generality

of the proposed method. Non-equilibrium MD simulations in a triangular lattice with

and without the absorbing boundary condition are conducted to demonstrate the validity

of the solution.

Keywords: absorbing boundary condition, molecular dynamics simulation, time-

history kernel.

1
Corresponding author, Professor, E-mail: dchen@ntu.edu.tw, Tel: +886233664275, Fax:
+886223631558.
1
1. Introduction

A non-reflecting boundary condition is essential for molecular dynamics (MD)

simulations involving wave transmission [1-4]. Typical boundary conditions adopted in

MD, such as free boundary conditions, fixed boundary conditions, or periodic boundary

conditions (PBC), may not be suitable for non-equilibrium process simulations because

free and fixed boundary conditions lead to wave reflection and PBC leads to different

physical phenomena [5,6].

Several solutions have been proposed to minimize non-physical wave reflection

of a finite domain considered in MD. These include the matching boundary conditions

(MBC) [7], variational boundary condition (VBC) [8-10], perfectly matched multiscale

simulation (PMMS) [11-13], and the generalized Langevin equation (GLE) [2,4,14-22].

MBC takes the form of linear constraints of displacements and velocities from selected

atoms near the boundary and determines the coefficients from matching the dispersion

relation of lattice vibration. VBC is an approximate boundary condition with optimal

coefficients of a finite kernel. Through a series of iterations, the kernel function may

converge to a kernel obtained from the GLE. A PMMS is derived from the discrete form

of a well-known perfectly matched layer for wave propagation in a continuum. It

minimizes reflection coefficients by choosing suitable damping functions. In GLE, a

time-history kernel function is solved by considering the interatomic relation with a

linear assumption. Among these solutions, MBC and GLE are two most promising

approaches to completely eliminate non-physical wave reflection of a finite domain

considered in MD. Performance comparison between MBC and GLE was addressed in

[7].

The GLE is well known to the multiscale community [23]. The GLE can, in

2
principle, completely eliminate wave reflection. It also provides rigorous justification

and a guiding principle for overcoming non-physical wave reflection behavior.

Adelman, Doll, and Myers originally derived the GLE to study the atom/solid-surface

scattering problem in the 1970s [14-16]. In these studies, they considered a gas atom

colliding with a 1-D semi-infinite harmonic mono-atomic chain and derived the

analytical solution for the motion of the boundary atom after collision. This analytical

solution provides the desired displacements of the boundary atom to minimize the

reflection.

An important extension of applying the GLE to multiscale problems was given by

Cai et al. [2]. In this study, the simulation domain was divided into a primary region P

and an outer region Q, for which the P-Q and Q-Q interactions were assumed to be

linear. With a linear assumption for the P-Q and Q-Q interactions, the motion of

boundary atoms can be expressed in terms of the GLE. This extension thus provides a

coupling scheme to apply the GLE to multiscale simulations.

In studies extending the GLE to different 2-D and 3-D crystalline lattices, Liu et

al. made a seminal contribution with their systematic procedure for numerically

computing the time-history kernels for absorbing boundary conditions [17,19,18,20-

22]. By introducing a scheme utilizing the discrete Fourier transform, time-history

kernel functions of any repeating crystal lattices can be solved. The approach was

originally developed for planar boundary conditions (edges in 2-D and surfaces in 3-D)

and later extended to corners using spatial truncations [24]. Recently, Tang et al.

proposed an alternative approach to calculating time-history kernel functions in real

space for a 1-D atomic chain [4] and a 2-D square lattice [25]. The concept of virtual

atoms proposed in these studies motivates us to derive an efficient numerical scheme

in real space for any crystalline lattice.

3
The aim of this study is to develop an approach to calculating time-history kernels

in real space for absorbing boundary conditions that can be applied in any lattice and

on any geometric boundary. The concept of virtual atoms is extended to facilitate kernel

derivation. The GLE is regularized using eigenvalue decomposition and, consequently,

an analytical expression of an inverse Laplace transform is obtained. A semi-analytical

form of the time-history kernel function can then be found.

The remainder of this paper is organized as follows. In Section 2, the proposed

approach is derived. In Section 3, a simple 1-D mono-atomic chain example is used to

demonstrate the main characteristics of the proposed approach. In Section 4, we provide

some time-history kernel function results for different lattices (a 2-D square lattice and

a 2-D triangular lattice) and different geometry boundaries (edge and corner). In

addition, non-equilibrium MD simulations in a triangular lattice with and without the

absorbing boundary condition are conducted to demonstrate the validity of the solution.

2. Methodology

We first give a derivation of the time-history kernel function with the concept of

virtual atoms. We then introduce the layer-wise concept to systemically solve the time-

history kernel function for different lattices.

2.1. Virtual Domain and Analytical Expression

Considering a full atomic system, the equations of atomic motion can be described

by Newtons law:

mu f f ext , (1)
where u is the displacement of atoms, m is the mass of atoms, and f and f ext are

the internal and external forces acting on the atoms, respectively.

We consider an atomic system divided into two parts: a real domain R and a
4
virtual domain V , as illustrated in Fig. 1. Using Eq. (1), the equations of motion for

the virtual domain are:

muV f V f VR , (2)
where the superscripts V and R denote the virtual domain and the real domain,

respectively, and f VR is an external force acting on the virtual domain, resulting from

the interaction between real and virtual domains.

We furthermore make an assumption that the R V and V V

interactions are linear. Based on this assumption, the internal forces and the systems

potential energy U have the following relation:

2U
f u. (3)
uu
The above relation can be simplified by a dynamical matrix D :

2U
D m 1 . (4)
uu
Applying this relation to Eq. (2), the equations of motion for the virtual domain can be

expressed as:

uV DV uV DVRu R . (5)

Applying the eigenvalue decomposition on the dynamical matrix DV , we obtain:

DV XXT , (6)
where is a diagonal matrix with all the eigenvalues of DV , and X is a matrix of

the corresponding eigenvectors. Equation (5) can be rewritten as:

uV XXT uV DVRu R (7)


or:

XT uV XT uV XT DVRu R . (8)

Performing the Laplace transform on Eq. (8) with uV t 0 0 and uV t 0 0 ,

we obtain:

5
s I X u
2 T V
XT DVRu R , (9)

where s 2u s u t and u s u t . Thus, u V can be related to u R as

follows:

u V X s 2 I XT DVR u R .
1
(10)

The displacements of virtual atoms can then be related to the atoms in the real domain

by taking the inverse Laplace transform of Eq. (10):

uV 1
u X sin t
V 1
XT DVR u R u R , (11)

where is a time-history kernel function, is a convolution operator, and are

frequencies defined as:

i i . (12)

We note that Eq. (11) indicates that the displacements of the virtual atoms can be

determined from the displacements of the real atoms if we can obtain the time-history

kernel function a priori. The novelty of eigenvalue decomposition opens an avenue for

an analytical inverse Laplace transform in Eq. (10). Practically, this eliminates the

computationally expensive part of solving the inverse Laplace transform numerically.

2.2. Unit Cell Interaction and Dynamical Matrix Construction

For a real lattice, we need a systematic way to construct the dynamical matrix. The

first step is to choose the smallest repeating unit cell conforming to the boundaries of a

simulation box. To this end, a dynamical matrix assembly is needed to construct the

interaction between the unit cells.

To establish the interaction between the unit cells, we start by considering the

interaction between atoms and rewriting Eq. (3) using the stiffness matrix K :

2U
f u =Ku . (13)
uu

6
The dynamical matrix is then:

2U
D m 1 m 1K . (14)
uu
Once we have the stiffness matrix K atom or dynamical matrix Datom to describe the

interaction between atoms, we can assemble a set of Datom matrices to construct the

dynamical matrix of a cell, Dcell . For example, the dynamical matrix of a cell with two

atoms can be constructed as:

D1,1
atom atom
D1,2
D cell
atom atom
, (15)
D2,1 D2,2

where the lower index indicates the interaction between different atoms.

The next step is to choose a repeating layer, which is a set of unit cells

corresponding to the geometry of a given boundary, as a high-level unit to construct the

dynamical matrix DVR or DV . It should be noted that the dynamical matrix DVR

connects only the atoms near the interface with the linear assumption, whereas other

non-nearby atoms in the real domain make no contribution. We denote the

displacements of the layer in the real domain as u , and those of the layers with the
0

same topology as u
0
in the virtual domain as u1 , u 2 , u3 , etc. (Fig. 2). We can

assemble a set of Dcell values to construct the dynamical matrix of layer, Dlayer . For

example, the dynamical matrix of a layer with M cells can be constructed as:

D1,1
cell cell
D1,2 D1,cellM
cell
D Dcell
D layer
2,1 2,2 (16)

cell
DM ,1 Dcell
M ,M

where the lower index indicates the interaction between different cells.

The final step is to assemble a set of Dlayer matrices to construct the dynamical

matrix DV or DVR . Since there is no interaction between non-nearby layers, the

dynamical matrices DV or DVR are banded and are given by:


7
D1,1
layer layer
D1,2 D1,0
layer
0
layer layer
D D
D 2,1 , DVR 0 0 0 ,
V 2,2
(17)
Dlayer
N 1,N
0 0

Dlayer
N , N 1 Dlayer
N ,N

where the subscripts indicate the interaction between different layers. Furthermore, for

an identical linear interaction between atoms, Eq. (17) can be simplified as:

P L LT 0
LT
P , DVR 0 0 0 ,
D
V
(18)
L 0 0

LT P

where:

P Dlayer
i ,i , i 1, 2, ,N

L D j , j 1 , j 1, 2, , N 1 . (19)
layer

LT Dlayer , j 1, 2, , N 1
j 1, j

From Eqs. (13) to (18), we can construct the dynamical matrix systematically via

a series of matrix assembly. Substituting Eq. (18) into Eq. (11), we obtain the semi-

analytical time-history kernel function:

t X sin t 1XT DVR , (20)

where X and are respectively the eigenvectors and eigenvalues of the matrix

DV .

The procedure for implementing the generalized absorbing boundary condition

proposed herein is outlined in Fig. 3. As mentioned above, the core concept is to choose

a suitable virtual domain to construct the matrix D and to compute the corresponding

time-history kernel function . In this procedure, eigenvalue decomposition leads to

the semi-analytical form of the time-history kernel function, which significantly

reduces the computational cost of solving the inverse Laplace transform. A systematic

way to construct the dynamical matrices from a prescribed virtual domain is vital to

8
calculate the time-history kernel function in different lattices or different geometric

boundaries, as will be demonstrated in a later section.

3. 1-D Atomic Chain

In this section, we use a 1-D atomic chain as shown in Fig. 4 to demonstrate how

to obtain the time-history kernel function following the procedure outlined in Fig. 3.

Additionally, a 1-D atomic chain is the simplest example and we can look closely at the

properties of the time-history kernel function obtained from the proposed method.

Finally, we will connect the solution obtained herein with the analytical solution.
The interatomic forces of the 1-D harmonic lattice are assumed to be linear with

the displacements:

mui k ui 1 2ui ui 1 . (21)

Without loss of generality, we can further simplify Eq. (21) to the rescaled form

k m 1 :

ui ui 1 2ui ui 1 . (22)

The unit cell of the 1-D atomic chain contains only an atom, and each layer only

contains a unit cell, i.e., P 2 and L 1 . Substituting the components into Eq.

(18), we can construct the dynamical matrices:

2 1 1 0
1 2 1 0 0 0

DV 1 2 , DVR 0 0 . (23)


1 2 N N
N N

Here, the virtual domain contains N atoms. We can then obtain the absorbing

boundary condition via Eq. (11):

9
u1 1,1 u0 , (24)

where u1 is only related to u0 .

For this simple 1-D case, we can obtain the eigenvalues and eigenvectors

analytically [22]:

i 2sin i / 2
2 , (25)
X i, j sin i j
N

where . The time-history kernel function can then be rewritten as:
N 1

N
2 sin 2t sin i / 2
1,1 t sin 2 i . (26)
i 1 N 2sin i / 2

We remark that it is interesting to take N at the semi-analytical form to

reveal its identity with the analytical solution. To this end, trigonometric function

substitutions in Eq. (26) are performed:


N
8 N 1 2
1,1 t sin i cos i sin 2t sin i . (27)
i 1 N 2 2 2 2
Next, we take N , which leads to 2 0 and N . Letting 2 , Eq.

(27) can then be rewritten into an integral form:



8
1,1 t sin cos 2 sin 2t sin d .

2
(28)
0

Converting all the cosine functions to sine functions in Eq. (28), we obtain:


1,1 t
8

sin sin sin 2t sin d .
0
2 3
(29)

Applying the triple-angle formulae, Eq. (29) becomes:



2
1,1 t sin sin 3 sin 2t sin d .

2
(30)
0

The above formula can be rewritten by the Bessel function as:

1,1 t J1 2t J 3 2t , (31)

10
where J is the Bessel function of the first kind. Finally, we obtain:
2 J 2 2t
1,1 t , (32)
t
which is the analytical solution for a 1-D atomic chain [4,15,17,23].

We now further explore long-tail and convergence characteristics of space and

time on computing 1,1 . For the spatial truncation, we choose different sizes of the

virtual domain N 20, 30, 40, 50 , and obtain 1,1 for each, as shown in Fig. 5. It

clearly shows that a larger N results in a longer valid cutoff time tc . The cutoff time

tc , defined as the maximum time for the valid time interval, is proportional to the length

of the virtual domain N, as indicated in [2]. The relationship between the cutoff time

tc and the length of the virtual domain N for this 1-D atomic chain is plotted in Fig. 6.

4. Results and Discussion

In this section, we demonstrate the procedure to obtain the time-history kernel

functions for 2-D square and triangular lattices. Moreover, we discuss MD simulation

in 2-D triangular lattice with and without the absorbing boundary conditions.

4.1. 2-D square lattice at an edge

The setup to obtain the time-history kernel function of a square lattice at an edge

interface is shown in Fig. 7. In this simple lattice, we treat an atom as a unit cell, and

each layer is arranged parallel to the edge interface. The rescaled interaction between

each atom in the out-of-plane direction is:

ui , j ui 1, j ui 1, j ui , j 1 ui , j 1 4ui , j . (33)

From the interaction between cells in the same layer, the cell dynamical matrix is:

11
4 , for i j
,j
Dicell . (34)
1 , for i j
Consider a virtual domain being divided into N layers where each layer contains

M cells. The layer dynamical matrix system becomes:

4 1 1
1 4 1
P , L= , (35)
1

1 4 M M 1 M M

in which the dimensions for the P and L matrices are M M and the dimensions

for the dynamical matrices DV or DVR are M N M N . Substituting Eq.

(35) into Eqs. (18) and (20), we obtain the semi-analytical time-history kernel

functions. Following Eq. (11), the displacements of atoms in the first layer are

controlled by the displacements of atoms in the zeroth layer:


N
ui i , j u j .
1 0
(36)
j 1

Furthermore, the atom arrangement in the plane interface gives us the symmetric

property:

i , i j i ,i j (37)

and only half of the kernel functions are needed for each convolution process at each

atom. Meanwhile, considering the virtual domain in a semi-infinite domain

( M , N ), we obtain:

0sq edge i ,i
sq edge , for any i, j . (38)
j i ,i j

Applying the properties into Eq. (36), the displacements of atoms in the first layer (that

is, the absorbing boundary condition) of a square lattice with an edge interface become:


M 2
1
ui sq edge
ui jsq edge ui j ui j .
0 0 0
0 (39)
j 1

12
The time-history kernel functions 0sq edge with different virtual domain sizes are

plotted in Fig. 8. Similar to the case of the 1-D atomic chain, we observe convergence

as the size of the virtual domain increases. The value of the converged time-history

kernel function also quickly decays to zero, as observed in the case of the 1-D atomic

chain, and tc 50 is sufficient to capture the behavior of the kernel. Figure 8 plots the

time-history kernel functions 0sq edge , 2sq edge , and 4sq edge for M N 41 and

tc 50 . The results agree very well with the results reported by Karpov et al. [18].
In practice, the contribution of unit cells far from a cutoff distance can be neglected.

Fig. 10 plots several peak values of the time-history kernels and confirms the values

become negligible as M increases. This allows us to perform a further truncation in

space for Eq. (39):


Mc
ui 0sq edge ui jsq edge ui j ui j ,
1 0 0 0
(40)
j 1

in which M c is the number of truncated cells and M c M.

4.2. 2-D square lattice at a corner

The procedure for computing the absorbing boundary condition at the corner is the

same as that for the edge. However, the corner geometry leads to a different virtual

domain as shown in Fig. 11 and a more complex dynamical matrix D . In this case, the

equations of motion and the cell dynamical matrix are identical to Eqs. (33) and (34),

respectively. The only difference is the layer topology. We remark that the flexibility to

choose different layer topologies conforming to the desirable topology of the boundary

is indeed the key to enabling our method to be applied to any lattice and any geometric

boundary.

Considering the layer interaction, the corresponding dynamic matrix components

are:

13
4 1 0
1 4 1 1 0



P 1 4 1 , L 1 0 1 .


1 4 1 0 1
1 4 M M 0 M M

(41)

Comparing Eq. (41) with Eq. (35), all the diagonal terms of L derived for the corner

boundary condition are zero. This is because there is no interaction between ui , j and

ui 1, j 1 . In addition, the non-zero terms of L derived for the corner boundary condition

are consequences of the interaction between ui , j and ui , j 1 or between ui , j and

ui 1, j .

Next, we look for geometric symmetry in the model to reduce the convolution

computation. For the corner, this is:

isq, j corner sqi ,corner


j . (42)

Thus, for the displacements of the corner atoms u0 , we obtain:


1


Mc
u0 0sq corner u0 jsq corner u j u j .
1 0 0 0
(43)
j 1

In general, the displacements of atoms in the first layer are:

ui1 isq, j corner u j0 (44)


j

and the summation for the corner boundary needs two parameters to characterize the

spatial cutoff of the interacting cells:

Mc2
ui1
j M c1
isq, j edge u j0 , (45)

14
in which M c1 and M c 2 are the indicial parameters characterizing the number of

truncated cells. The time-history kernel functions obtained herein are plotted in Fig. 12

and the results agree very well with those reported by Pang and Tang [25].

4.3. 2-D triangular lattice at the edge and corner

In this section, we derive the time-history kernel functions at the edge and corner

for in-plane motion in a 2-D triangular lattice. The resulting absorbing boundary

conditions will be used in the MD simulation in a later section.

In a 2-D triangular lattice, each unit cell contains two atoms, as shown in Fig. 13.

Following similar procedures to those for the 2-D square lattice, the P and L

matrices corresponding to the edge and corner are:

Edge:

Dcell
0,0
cell
D0,1 Dcell
1,0 Dcell
1,1

cell cell cell cell
D D , L D1, 1 D
P 0,1 0,0 1,0 (46)
cell
D0,1 Dcell
1,1

cell cell
Dcell
0, 1 D0,0 Dcell
1, 1 D1,0

Corner:

15
Dcell
0,0 Dcell
1,0

cell cell
D1,0 D 0,0
Dcell
1,0


P cell
D1,0 cell
D0,0 cell
D0,1 ,
Dcell
0, 1

Dcell
cell
0,0 D0,1
Dcell cell
D0,0
0, 1

Dcell
1, 1

cell
D
0,1
D1,cell1


L D1,cell1 D0,cell1 Dcell Dcell Dcell
1, 1 1,0 1,1


cell
D1,1

1,0
Dcell
, (47)
Dcell
1, 1

in which the dynamical matrices between the unit cells are:

3 0 1
4
3 4 0 0 0 0
0
3 3 0 0
, D0,cell1 Dcell
0,1 ,
0 3 0
D0,0 4

cell 4 cell T
, D0,1
1 4 3
4 3 0 1 4 3
4 0 0
3 3
3 3

4 4 0 3 4 4 0 0
1 0 0 0
0 0
1,0 D1,0 ,
0 0
D1,0
cell
, Dcell cell T

1 4 3
4 1 0
3 3

4 4 0 0
0 0 0 0
0 0
1, 1 D1,1 ,
0 0
cell
D1,1 , Dcell cell T

1 4 3
4 0 0
3 3

4 4 0 0
0 0 0 0
0 0 0 0
D1,1 1,1 D1, 1 .
cell cell T
, Dcell
0 0 0 0
(48)
0 0 0 0
We notice that the interaction is in-plane. Thus, any pair of atoms will have four kernel

functions: ixx, j , ixy, j , iyx, j , and iyy, j . Some of these kernel functions are plotted in Fig.

16
14.

4.4. MD simulation

In this section, we impose the absorbing boundary condition obtained from the

previous section to study wave transmission and reflection in an MD simulation. All

the values reported herein are in Lennard-Jones units. We consider three different

models with a 2-D triangular lattice: a small MD model ( 30a1 20a2 , where a1 1 6
2

and a2 3a1 ) with the absorbing boundary condition, a mid-sized MD model

( 50a1 44a2 ) with PBCs, and a large MD model. A Lennard-Jones interatomic potential

with 1 26 and 1 6
2 is adopted and a single wave source with a prescribed

displacement in the y-direction is applied at the center:

u y 0.05 1 cos 3 0.2t , t 0,10 . (49)

We note that the large MD model is designed such that the wave will not reach its

boundaries within the time of interest. For the small MD model, 7a1 4a2 corner

boundary atoms are imposed with the corner interface, and the other boundary atoms

are imposed with the edge interface, as illustrated in Fig. 15. Each time-history kernel

function is truncated with tc 30 .

From the snapshots shown in Fig. 16, we observe that the wave initially propagates

outward from the center atom for all three models as expected. In the mid-sized PBCs

MD model, after the wave reaches the boundary, it reflects backward toward the system.

In contrast, little reflection is observed in the small MD model with the absorbing

boundary condition.

To further quantify the error due to spurious wave reflection, we consider the

potential energy and temperature in the observed regions (Fig. 16). The quantities

initially increase monotonically and reach a steady state after the prescribed

17
displacement forms a complete wave at t 10 . When the wave reaches the boundary

of the observed region, the energy and temperature decrease to a small value, indicating

the wave has passed through the observed regions boundary.

After the energy and temperature reach the minimum, we observe that the

quantities increase again in the mid-sized MD model with PBCs. This is clearly a non-

physical effect due to spurious wave reflection. For the small MD model with absorbing

boundary conditions, the results agree well with those of the large MD model.

5. Conclusions

In this paper, we developed an approach to calculate time-history kernels in real

space for absorbing boundary conditions. The flexibility to choose different virtual

domains with layer topologies conforming to the desirable topology of the boundary

enables us to apply absorbing boundaries in any lattice and on any geometric boundary.

While computing the corresponding time-history kernel function, eigenvalue

decomposition leads to a semi-analytical form of the kernel function, which

significantly reduces the computational cost of solving the inverse Laplace transform.

We used a simple 1-D mono-atomic chain example to demonstrate the main

characteristics of the proposed approach. In Section 4, we provided some time-history

kernel function results in different lattices (a 2-D square lattice and a 2-D triangular

lattice) and different geometric boundaries (edge and corner). In addition, non-

equilibrium molecular dynamics simulations in a triangular lattice with and without the

absorbing boundary conditions were conducted to demonstrate the validity of the

solution.

We finally remark that the approach developed herein is limited to absorbing

boundary conditions for semi-infinite problems. Nevertheless, as our approach is based

18
on the concept of the virtual domain in real space, it is possible to extend the

formulation to include multiple combinations of real and virtual domains that will

eliminate the limitation posed herein. We are currently working on this topic and will

report our results in the future.

Acknowledgements

This research was supported by the Ministry of Science and Technology, Taiwan,

under grant no. 105-2221-E-002-025-MY3 and by National Taiwan University under

grant no. 103R891805 (The NTU Excellence in Research Program). We are grateful to

the National Center for High-Performance Computing and National Taiwan University

for providing the computational resources.

19
6. References

1. Moseler M, Nordiek J, Haberland H (1997) Reduction of the reflected pressure


wave in the molecular-dynamics simulation of energetic particle-solid collisions. Phys
Rev B 56 (23):15439-15445. doi:10.1103/PhysRevB.56.15439
2. Cai W, de Koning M, Bulatov VV, Yip S (2000) Minimizing boundary
reflections in coupled-domain simulations. Phys Rev Lett 85 (15):3213-3216.
doi:10.1103/PhysRevLett.85.3213
3. Jones RE, Kimmer CJ (2010) Efficient non-reflecting boundary condition
constructed via optimization of damped layers. Phys Rev B 81 (9):094301.
doi:10.1103/PhysRevB.81.094301
4. Tang S (2010) A Two-Way Interfacial Condition for Lattice Simulations. Adv
Appl Math Mech 2 (1):45-55. doi:10.4208/aamm.09-m0944
5. Kantorovich L (2008) Generalized Langevin equation for solids. I. Rigorous
derivation and main properties. Phys Rev B 78 (9):094304.
doi:10.1103/PhysRevB.78.094304
6. Kantorovich L, Rompotis N (2008) Generalized Langevin equation for solids.
II. Stochastic boundary conditions for nonequilibrium molecular dynamics simulations.
Phys Rev B 78 (9):094305. doi:10.1103/PhysRevB.78.094305
7. Wang X, Tang S (2013) Matching boundary conditions for lattice dynamics.
International Journal for Numerical Methods in Engineering 93 (12):1255-1285.
doi:10.1002/nme.4426
8. E W, Huang Z (2001) Matching conditions in atomistic-continuum modeling of
materials. Phys Rev Lett 87 (13):135501. doi:10.1103/PhysRevLett.87.135501
9. Li X, E W (2006) Variational boundary conditions for molecular dynamics
simulations of solids at low temperature. Commun Comput Phys 1 (1):135-175
10. Li X, E W (2007) Variational boundary conditions for molecular dynamics
simulations of crystalline solids at finite temperature: Treatment of the thermal bath.
Phys Rev B 76 (10). doi:10.1103/PhysRevB.76.104107
11. To AC, Li S (2005) Perfectly matched multiscale simulations. Phys Rev B 72
(3). doi:10.1103/PhysRevB.72.035414
12. Li S, Liu X, Agrawal A, To AC (2006) Perfectly matched multiscale
simulations for discrete lattice systems: Extension to multiple dimensions. Phys Rev B
74 (4). doi:10.1103/PhysRevB.74.045418
13. Guddati MN, Thirunavukkarasu S (2009) Phonon absorbing boundary
conditions for molecular dynamics. J Comput Phys 228 (21):8112-8134.
doi:10.1016/j.jcp.2009.07.033
14. Adelman SA, Doll JD (1974) Generalized Langevin equation approach for
atomsolid-surface scattering: Collinear atomharmonic chain model. J Chem Phys
20
61 (10):4242. doi:10.1063/1.1681723
15. Doll JD, Myers LE, Adelman SA (1975) Generalized Langevin equation
approach for atom/solid-surface scattering: Inelastic studies. J Chem Phys 63 (11):4908.
doi:10.1063/1.431234
16. Adelman SA, Doll JD (1976) Generalized Langevin equation approach for
atom/solid-surface scattering: General formulation for classical scattering off harmonic
solids. J Chem Phys 64 (6):2375. doi:10.1063/1.432526
17. Wagner GJ, Liu WK (2003) Coupling of atomistic and continuum simulations
using a bridging scale decomposition. J Comput Phys 190 (1):249-274.
doi:10.1016/s0021-9991(03)00273-0
18. Karpov EG, Wagner GJ, Liu WK (2005) A Green's function approach to
deriving non-reflecting boundary conditions in molecular dynamics simulations. Int J
Numer Meth Eng 62 (9):1250-1262. doi:10.1002/nme.1234
19. Wagner GJ, Karpov EG, Liu WK (2004) Molecular dynamics boundary
conditions for regular crystal lattices. Comput Method Appl M 193 (17-20):1579-1601.
doi:10.1016/j.cma.2003.12.012
20. Park HS, Karpov EG, Liu WK, Klein PA (2005) The bridging scale for two-
dimensional atomistic/continuum coupling. Philosophical Magazine 85 (1):79-113.
doi:10.1080/14786430412331300163
21. Karpov EG, Yu H, Park HS, Liu WK, Wang QJ, Qian D (2006) Multiscale
boundary conditions in crystalline solids: Theory and application to nanoindentation.
International Journal of Solids and Structures 43 (21):6359-6379.
doi:10.1016/j.ijsolstr.2005.10.003
22. Tang S, Hou TY, Liu WK (2006) A pseudo-spectral multiscale method:
Interfacial conditions and coarse grid equations. J Comput Phys 213 (1):57-85.
doi:10.1016/j.jcp.2005.08.001
23. Liu WK, Karpov EG, Park HS (2006) Nano mechanics and materials : theory,
multiscale methods and applications. John Wiley, Chichester, England ; Hoboken, NJ.
doi:9780470018514
24. Medyanik SN, Karpov EG, Liu WK (2006) Domain reduction method for
atomistic simulations. Journal of Computational Physics 218 (2):836-859.
doi:10.1016/j.jcp.2006.03.008
25. Pang G, Tang S (2011) Time history kernel functions for square lattice. Comput
Mech 48 (6):699-711. doi:10.1007/s00466-011-0615-4

21
7. Figures

Fig. 1: An atomic region is divided into a real domain R and a virtual domain V .

An important assumption is that R V and V V interactions are linear, but

no assumptions are made on the R R interaction.

22
N

u
0
u u
2
u
1 3

R V
Fig. 2: In this illustration, the virtual domain is divided into N layers. Each layer

contains M cells, and each cell contains two atoms.


23
Fig. 3: The procedure for calculating the time-history kernel function and absorbing

boundary condition.
24
u0 u1 u2
R V

Fig. 4: The model for a 1-D atomic chain.

25
Fig. 5: Comparison of the time-history kernel functions 1,1 t with different sizes

N 20, 30, 40, 50 of the virtual domain with the analytical solution as N .

26
Fig. 6: The linear relationship between the cutoff time tc and the length of the virtual

domain N.

27
N

u
0
u u
1 2

R V
Fig. 7: The virtual domain for a 2-D square lattice with the edge boundary in which the

virtual domain contains N layers and each layer contains M cells.

28
Fig. 8: The time-history kernel function 0 t of a 2-D square lattice with the edge

boundary for different sizes of the virtual domain N 11, 21, 31, 41 .

29
Fig. 9: Comparison of the time-history kernel functions 0 t , 2 t , and 4 t for

a 2-D square lattice at the edge boundary calculated using our approach with those

reported by Karpov et al. [18].

30
Fig. 10: Maximum peak values of the time-history kernel functions for a 2-D square

lattice at the edge boundary.

31
M

u
2

u
1

u
0
R V

Fig. 11: The virtual domain for a 2-D square lattice at the corner boundary.

32
(a)

(b)

33
Fig. 12: Comparison of the time-history kernel functions (a) 2,0 t for a 2-D square

lattice at the corner boundary and (b) 1,2 t for a 2-D square lattice at the edge

boundary calculated using our approach with those reported by Pang and Tang [25].

34
(a)

0 1 2
u u u
R V
35
(b)

f d

e
2
u

1
u

u
0

R V
36
Fig. 13: The virtual domains for a 2-D triangular lattice at (a) the edge boundary and

(b) the corner boundary.

37
(a)

(b)

38
Fig. 14: The time-history kernel functions for a 2-D triangular lattice at (a) the edge

boundary and (b) the corner boundary.

39
(a)

(b)

Fig. 15: MD models in a 2-D triangular lattice: (a) a small MD model ( 30a1 20a2 )

with absorbing boundary conditions, where yellow indicates an edge and green

indicates a corner; and (b) a mid-sized MD model ( 50a1 44a2 ) with PBCs (black line),

where the region enclosed by the red solid lines is the observed region.

40
(a) t = 10

(b) t = 20

(c) t = 30

41
(d) t = 40

(e) t = 50

42
(f) t = 60

Fig. 16: MD simulation snapshots from t 10 to t 60 . In the small MD model, the

wave transmits through the boundary with negligible wave reflection. In the mid-sized

model, the wave hits the PBCs and reflects backward into the system.

43
(a)

(b)

44
Fig. 17: (a) Potential energy and (b) temperature in three different models.

45

Das könnte Ihnen auch gefallen