Binder, Luskin & Ortner (2017)

MULTISCALE MODEL. SIMUL.
c 2017 Society for Industrial and Applied Mathematics

Vol. 15, No. 2, pp. 892919
ANALYSIS OF A PREDICTOR-CORRECTOR METHOD FOR

Downloaded 05/19/17 to 132.239.1.231. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
COMPUTATIONALLY EFFICIENT MODELING OF SURFACE

EFFECTS IN 1D
ANDREW J. BINDER , MITCHELL LUSKIN , AND CHRISTOPH ORTNER
Abstract. The regular CauchyBorn method is a useful and efficient tool for analyzing bulk
properties of materials in the absence of defects. However, the method normally fails to capture
surface effects, which are essential to determining material properties at small length scales. In this
paper, we present a corrector method that improves upon the prediction for material behavior from
the CauchyBorn method over a small boundary layer at the surface of a one-dimensional (1D)
material by capturing the missed surface effects. We justify the separation of the problem into a
bulk response and a localized surface correction by establishing an error estimate, which vanishes in
the long wavelength limit.
Key words. surface, predictor-corrector, CauchyBorn
AMS subject classifications. 74B20, 65Z05
DOI. 10.1137/16M1065884
1. Introduction. Long-range elastic fields dictate the behavior of crystalline

materials in the absence of defects. This elastic response in crystalline systems is of-
ten studied through the use of the CauchyBorn method [1, 18]. In the CauchyBorn
method, the nonlocal interactions in the atomistic system are replaced by a localized
continuum approximation, which may then be further approximated using the finite
element method (FEM). This sequence of approximations allows for a reduction in the
number of degrees of freedom in the model and yields a computationally efficient and
accurate approach to studying the material behavior of perfect crystalline materials.
However, this approach fails in the presence of defects since they create rapid varia-
tions in the atomistic displacement field which can no longer be accurately captured
by the CauchyBorn model. In particular, for the present paper, the CauchyBorn
method often struggles with capturing effects due to the inclusion of surfaces in the
model [17].
Surfaces can be characterized as defects because atoms near a surface experience a
different interaction environment than atoms in an idealized crystal lattice. For large
systems, the surface influence is often negligible since the surface region constitutes
such a small portion of the entire system. The bulkor interiorbehavior dominates.
Surface effects become increasingly significant, though, as the surface area to volume
ratio increases, resulting in a size-dependent response in material behavior, which is
an active area of research [3, 8, 15, 16].
Received by the editors March 15, 2016; accepted for publication (in revised form) February 13,
2017; published electronically May 16, 2017.

http://www.siam.org/journals/mms/15-2/M106588.html
Funding: The work of the first author was supported by the DOD through the NDSEG Fel-
lowship Program and the NSF PIRE Grant OISE-0967140. The work of the second author was
supported in part by NSF PIRE Grant OISE-0967140, NSF grant 1310835, and the Radcliffe Insti-
tute for Advanced Study at Harvard University. The work of the third author was supported by
ERC Starting Grant 335120.
School of Mathematics, University of Minnesota, Minneapolis, MN 55455 (bind0090@umn.edu,
luskin@umn.edu).
Mathematics Institute, University of Warwick, Coventry CV4 7AL, United Kingdom (c.ortner@
warwick.ac.uk).
892
Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

MODELING OF SURFACE EFFECTS 893
While surface effects tend to be most relevant in small systems, these small sys-
tems may still be large enough to present computational challenges for a fully atomistic
numerical simulation. Furthermore, even in bulk crystals, if effects of interest take

place at or near a crystal boundary (e.g., nano-indentation), then accurately capturing
the surface physics remains crucial. Thus, a computationally efficient means to simu-
late surface-dominated systems is desirable. Various methods have been proposed that
have the potential to meet this need. These methods generally entail a modification of
the continuum model [9, 10, 17] or a concurrent coupling of atomistic and continuum
models [4, 5, 6, 7, 12, 13, 19]. Yet, while these methods do result in an approximation
superior to that which comes from the regular CauchyBorn method, the modified
continuum models lack a systematic control over the error in the approximation, while
the atomistic-to-continuum coupling methods may encounter difficulties with surface
geometries that would result in a much reduced computational savings compared to
the savings that might be observed with other defects.
In this paper, we introduce a novel approach to efficiently and accurately captur-
ing surface behavior through the use of a predictor-corrector method that possesses a
more controllable error and an improved approach to handling surface geometries. In
this predictor-corrector method, the CauchyBorn method serves as the initial predic-
tor for material behavior, allowing for the usual tools that study bulk behavior to be
applied to the surface problem too. As the CauchyBorn method works well for the
interior, it should serve well for an approximation of the bulk response of the material.
The solution for the CauchyBorn method is then corrected in the next step to take
into account surface effects by minimizing the energy difference between the atomistic
energy and the CauchyBorn energy about the CauchyBorn solution. This correc-
tion occurs over a boundary layer near the surface at atomistic resolution. The size
of the boundary layer is a controllable parameter in this predictor-corrector method
and allows for a systematic control over the accuracy of the approximation. The
corrected solution represents the approximation of this predictor-corrector method to
the atomistic behavior. Proving the validity of the decomposition of the atomistic
solution into a bulk response and surface correction will be the primary goal of the
analysis in this paper, in addition to assessing the quality of the approximation.
The paper is organized as follows. In section 2, we summarize all main results
and illustrate them via numerical tests. Complete proofs are given in sections 36.
Notation for derivatives. We will employ three types of derivatives. (1) If u :
{0, 1, . . . } is a discrete function, then we define u0` = u`+1 u` . (2) If u : [0, )
is a continuous displacement or deformation function, then u(x) is the standard
pointwise or weak derivative (if it exists). (3) Finally, if V : Rn Rp is a potential
function (or its derivative), then we denote its partial derivatives by i V , while its
Jacobi matrix is denoted by V .
2. Summary of results.
2.1. Atomistic model. We consider a semi-infinite, one-dimensional (1D) chain
of atoms. We index the atoms in the chain by the set of nonnegative integers, Z0 ,
and we denote the individual location of the atom with index ` in the chain by y` R.
The reference position of each atom in the chain is chosen to be y` = `. We denote
the displacement of atom ` from its reference position by u` := y` ` R. The strain
(gradient) in the bond between the atoms indexed by ` and ` + 1 in the chain will be
written as
(1) u0` := u`+1 u` = y`+1 y` 1.

894 A. J. BINDER, M. LUSKIN, AND C. ORTNER
The surface of the chain is located at the atom with index 0.

The atoms in this semi-infinite chain interact according to a nearest-neighbor site
energy (effectively second-neighbor interaction). This is the simplest setting within

which we can still observe the surface effects we are interested in. The energy due to
these interactions is given by

X
(2) E a (y) := V surf (y1 y0 ) + V (y` y`1 , y`+1 y` ),
`=1
where V denotes the site energy for atoms in the interior of the system while V surf
denotes the surface site energy. The surface atom merits a different site energy than
the interior atoms because the 0th atom has only one neighbor while every other
atom has two. We will later assume that V surf is the limit of V as one of the bonds
is stretched to infinity.
Since E a is translation invariant, configurations are not meaningfully distinct if
they differ only up to a translation. Hence, we fix the endpoint so that y0 = 0. Under
this constraint, knowledge of u0 allows us to recover the full displacement y. Hence,
it will often be more convenient to consider the energy in terms of the strain rather
than the deformation:

X
(3) E a (u) := V surf (u00 ) + V (u0`1 , u0` ),
`=1
where we have now also absorbed the reference strain 1 into the definitions of V surf
and V .
We assume throughout that the site energies satisfy the following properties.
Site energy properties:
(i) V C k (R2 ) and V surf C k (R) with k 3.
(ii) V , V surf , and all permissible partial derivatives are bounded.
(iii) V (0, 0) = 0.
(iv) 2 V (0, 0) > 0.
(v) inf {|(r,s)|>} V (r, s) > 0 for any > 0.
(vi) For any s R, limr V (r, s) = V surf (s).
If the semi-infinite chain were extended infinitely in both directions so that it had
no surface and its energy consisted purely of a sum of bulk site energies, then proper-
ties (iii)(v) would imply that the ground state of this bulk model is the configuration
with a uniform strain of 0. Property (iv) would guarantee that the eigenvalues of
the force-constant matrix of such an infinite chain system are positive so that the
infinite crystal is stable. We note that the fixing of the endpoint removes the zero
eigenmode from the system. For the surface model, these properties will allow us to
prove existence of a ground state and to establish properties of the boundary layer.
Property (vi) defines the surface site energy in terms of the limiting behavior of the
bulk site energy, and as a consequence of properties (v) and (vi), we see that
(4) inf V surf (s) > 0.

sR
Property (ii) is one of convenience for simplifying the statement and formulation of
the proofs and is not strictly necessary (see [14]). One way one can weaken this
assumption is by enforcing a minimum distance between atoms. Throughout the
paper, derivatives of V surf with respect to its argument will be denoted by F V surf .

Derivatives of V with respect to its first and second arguments will be denoted by
1 V and 2 V , respectively. Higher-order derivatives for the two site energies will be
denoted similarly.
We are concerned with finding the energy-minimizing configuration of the chain
in the presence of external forces. To consider the situation where bulk and surface
effects are roughly of the same order, we consider only finite-energy displacements,
i.e., displacements from the space
(5) U := {u : Z0 R | u(0) = 0 and u0 `2 (Z0 )}.
When U is equipped with the H 1 -seminorm |u|H 1 = ku0 k`2 (Z0 ) , it becomes a Hilbert
space due to the constraint u(0) = 0. It is easy to see [12] that compact displacements
are dense in U.
Applied forces take the form of a lattice function f : Z0 R, which must be an
element of U . We say that f U if and only if there exists a constant kf kU such
that
(6) hf, ui kf kU ku0 k`2 (Z0 ) for all u U with supp(u) compact,
P
where hf, ui := `=0 f` u` . For arbitrary u U, hf, ui is defined by continuity.
Given such a force f U , we seek a minimizer
(7) ua arg min{E a (u) hf, uiZ0 | u U}.
This problem may have several or no solutions. We consider the existence of solutions
to this problem as well as their decay in the following section.
We conclude the description of the model by mentioning that properties (i) and
(ii) guarantee that the first and second variations are globally Lipschitz continuous.
We refer the reader to [12] for a proof.
Remark. It is easy to see that f U if and only if there exists g `2 (Z0 )
such that f = g 0 by considering a discrete summation by parts of hf, ui. From this
summation by parts, it can also be shown that

X
(8) gn = f` < ;
`=n+1
see section 4 for a proof that this sum is well defined. Therefore, hf, ui = hg, u0 i
kgk`2 ku0 k`2 , and, in fact, it is clear that kf kU = kgk`2 .
2.1.1. Example: EAM model. To provide a concrete example of the types of
systems encompassed by the model described above, we show how the semi-infinite
chain satisfying the embedded atom model (EAM) [2, 11] with only nearest-neighbor
interactions is described in this framework. The energy for such a system is written
in terms of the strain as
(9)

X
0
a
E (u ) = (1 + u00 ) + ((1 + u00 )) + 0 0 0
(1 + u` ) + (1 + u` ) + (1 + u`1 ) ,
`=1
where is a nearest-neighbor pair potential, is an embedding energy function, and

is an electron density function. In order to write this energy in the form of (3), we

define the bulk and surface site energies, respectively, as

1 1
V (u0`1 , u0` ) = (1 + u0`1 ) + (1 + u0` ) + (1 + u0`1 ) + (1 + u0` ) and

2 2
1
V surf (u00 ) = (1 + u00 ) + (1 + u00 ) .

2
For the analysis, we will consider the generalized framework described by (3), but we
will return to the EAM system for the numerical results in section 2.6.
2.1.2. Example: Next-nearest neighbor model. This framework also allows
for next-nearest neighbor type interactions. For example, consider the following next-
nearest neighbor system energy for the semi-infinite chain:

X
(10) E a (u0 ) = (1 + u00 ) + (1 + u0` ) + (2 + u0`1 + u0` ) ,
`=1
where is some potential. To translate this into the form of (3), we define the bulk
and surface site energies, respectively, as
1 1
V (u0`1 , u0` ) =(1 + u0`1 ) + (1 + u0` ) + (2 + u0`1 + u0` ) and
2 2
surf 0 1 0
V (u0 ) = (1 + u0 ).
2
2.2. Existence, decay, and stability of atomistic solutions. Unlike in a
bulk model without surfaces, we do not expect that an energy-minimizing configu-
ration for the semi-infinite chain is homogeneous. Therefore, we have to be satisfied
with weaker existence proofs and establishing general properties of a minimizer.
Theorem 2.1. There exists a minimizer of E a : U R {+}.
Proof. See section 3.
While for many systems it is reasonable (and natural) to expect a unique ground
state, our assumptions do not preclude the existence of multiple states that achieve
the same minimal energy. In the following, uagr may refer to any ground state.
The key property of uagr that motivates our subsequent developments is that the
surface effects are highly localized. This is established next.
Theorem 2.2. Let uagr be a critical point of E a . Then, there exists 0 a < 1
such that
(11) |(uagr )0` | . `a for all ` Z0 .

The exponential decay of the strain due to surface effects normally justifies ig-
noring surface effects in large systems. However, we will see that when surface effects
are the focus of interest, this boundary layer cannot be ignored.
We now proceed to incorporating external forces into the analysis. To that end, it
is convenient to assume that a ground state uagr is strongly stable; that is, we suppose
that there exists an atomistic stability constant ca > 0 such that
(12) h 2 E a (uagr )v, vi ca kv 0 k2`2 (Z0 ) for all v U.

This stability assumption on uagr enables us to prove the existence of nearby strongly
stable local minimizers of the atomistic problem from (7) with small external forces.
For future reference, an element ua U is a strongly stable solution to the atomistic
problem if and only if it satisfies the EulerLagrange equation
(13) hE a (ua ), vi = hf, vi for all v U
as well as the stability condition
(14) h 2 E a (ua )v, vi ckv 0 k2`2 (Z0 ) for all v U,
for some constant c > 0. The exact forms of the first and second variations of E a are
provided in Propositions 7.1 and 7.2.
Corollary 2.3. There exist , C > 0 such that, for all f U with kf kU < ,
the atomistic problem (7) has a unique, strongly stable solution with k(ua uagr )0 k`2
Ckf kU .
Proof. This is an immediate consequence of the inverse function theorem (Theo-
rem 7.3).
2.3. CauchyBorn model. A common approach to determining the approxi-
mate bulk behavior of perfect crystalline materials utilizes the CauchyBorn model
of atomistic interactions and approximates the nonlocal atomistic interactions with
a local approximation. A limiting procedure then turns the discrete problem into a
continuum one [1]. The CauchyBorn energy for the semi-infinite chain model is
Z
(15) E cb (u) := W (u(x)) dx for u U cb ,
0
where W (F ) := V (F, F ) is the CauchyBorn energy density function and
U cb := u Hloc 1
(0, ) | u L2 (0, ) and u(0) = 0 .

(16)
The CauchyBorn energy density function W inherits the smoothness of V bulk , i.e.,
W C 3 (R). Derivatives of W with respect to its argument (as opposed to x) will be
indicated by F W . We also note that the space U may be considered a subspace of U cb
if we identify the lattice functions u U with their piecewise continuous interpolants.
For atomistic systems without defects, the CauchyBorn approximation and the
exact model agree under homogeneous deformation. However, in systems containing
defects, such as surfaces, this property is lost. In (15), this discrepancy arises as the
CauchyBorn model treats every point in the chain as an interior point. The absence
of a surface component to the energy results in an O(1) error in the consistency
estimate for the CauchyBorn model as compared to the atomistic system, which we
can demonstrate when f = 0.
Proposition 2.4. The unique minimizer of E cb in U cb is ucb = 0. Its atomistic
residual is bounded by
(17) sup |hE a (0), vi| = kE a (0)kU = |F V surf (0)|.

vU ,kv 0 k`2 (Z ) =1
0
In particular, k(uagr ucb )0 k`2 M 1 |F V surf (0)|, where M is the global Lipschitz
constant of E a .


In general, we expect that |F V surf (0)| 6= 0 so that the CauchyBorn and atom-
istic energy-minimizing configurations are not in agreement because of surface effects.

This lack of consistency at the surface indicates that the CauchyBorn model
alone will not serve as an accurate approximation of the original system for the purpose
of studying surface effects. However, as we have shown in Theorem 2.2, surface effects
are a local phenomenon. The CauchyBorn model may then still serve as an efficient
means to computing the behavior of the system in the interior.
We will show that the bulk and surface effects decouple, which means that a
complex concurrent coupling scheme is not required. Instead, we can add the surface
effects in a predictor-corrector type approach, where the predictor is the CauchyBorn
solution.
We therefore consider the general CauchyBorn problem
(18) ucb arg min{E cb (u) hf, uiR+ | u U cb },
where we identify Rthe lattice function f with its continuous piecewise affine interpolant
and hf, uiR+ := 0 f u dx. It is easy to see that f U implies that hf, iR+ (U cb ) ;
see section 4.
Analogously to the atomistic problem, we say ucb is a strongly stable solution to
(18) if it satisfies the first-order and strong second-order optimality conditions,
hE cb (ucb ), vi = hf, viR+ for all v U cb and

2 cb cb
(19) h E (u )v, vi ccb kvk2L2 ([0,)) for all v U cb ,
for some constant ccb > 0. The exact forms of the first and second variations of the
CauchyBorn energy may be found in Propositions 7.1 and 7.2.
For small enough external forces, we can guarantee the existence of strongly stable
solutions to the CauchyBorn problem and deduce some additional facts concerning
their regularity.
Theorem 2.5. There exists cb > 0 such that for all f U with kf kU < cb , a
strongly stable solution ucb U cb of the CauchyBorn problem (18) exists.
Moreover, ucb Hloc3
, and it satisfies the bounds
(20)
|ucb (0)| . kf kU , |2 ucb (x)| . |f (x)|, and |3 ucb (x)| . |f (x)| + |f (x)|2 .

In the next section, we will use the CauchyBorn solution to approximate the
atomistic solution. To that end, we introduce a projection operator a : U cb U via
Z `+1
(21) (a u)0 = 0 and (a u)0` = u(s) ds for ` Z0 , u U cb .
`
2.4. Predictor-corrector method. Proposition 2.4 demonstrates that the

CauchyBorn method commits an O(1) consistency error at a material surface. How-
ever, the CauchyBorn method is an excellent method for approximating the bulk
response of materials, and Theorem 2.2 indicates that the error due to surface effects
may be quite localized. We therefore propose a mechanism to correct the Cauchy
Born solution to obtain an approximation of the form
(22) ua a ucb + qL ,

where the parameter L for the corrector qL determines the quality of the approxima-
tion.
Given a predictor ucb solving the CauchyBorn problem (18), we define the cor-
rector problem via the minimization of a corrector energy, which is given by
(23) E (q; F0 ) = V surf (F0 + q00 ) W (F0 ) q00 F W (F0 )

X
0
, F0 + qj0 ) W (F0 ) qj0 F W (F0 ) ,

+ V (F0 + qj1
j=1
where we will normally take F0 := ucb (0).

The idea of the corrector problem is that it should depend only in a local way
on the elastic field ucb . In this case, this dependence is only on ucb (0). This
choice was made deliberately so that the corrector problem can be understood as a
cell problem on a surface element when the CauchyBorn model is discretized using
finite elements.
For L N {}, a corrector strain on the interval [0, L] is found by solving
(24) qL arg min{E (q; F0 ) | q QL },
where QL describes the boundary layer over which we correct the CauchyBorn so-
lution and is defined to be
QL := {q U | q`0 = 0 for all ` L}. In particular, Q = U.
The choice of L affects the computational expense of solving the corrector problem as
well as the accuracy of the approximation of the atomistic systems behavior. Note
that the external force f does not enter directly into the corrector problem. The
CauchyBorn method accounts for the external forces on its own, but the influence
of the external force is felt through the dependence on F0 .
Theorem 2.6. There exists > 0 such that, for all F0 R with |F0 | < , the
corrector problem (24) with L = has a solution q Q . For all v Q , this
solution q satisfies
ca 0 2
(25) h 2 E (q ; F0 )v, vi kv k`2 ,
2
where ca is the atomistic stability constant for uagr in (12). In addition, there exists a
constant 0 q < 1 such that
(26) |(q )0` | . `q for all ` Z0 .

We can now consider the existence of a solution to the corrector problem for a
finite boundary layer as an approximation of the infinite case.
Proposition 2.7. Under the conditions of Theorem 2.6, there exists L0 > 0 such
that the corrector problem (24) with L L0 has a solution qL QL satisfying
0 0
(27) kq qL k`2 . L
q.

If kf kU is sufficiently small, then Theorem 2.5 guarantees the existence of a

solution ucb U cb with |ucb (0)| . Thus, if we choose a sufficiently large
boundary layer for the corrector problem (24), Theorem 2.6 and Proposition 2.7 ensure
that the predictor-corrector approximation
(28) upc
L := a u
cb
+ qL
is well defined. Next, to determine the accuracy of this approximation, we show that
a solution to the atomistic problem (7) exists in a neighborhood of upc L , which we
quantify.
Theorem 2.8. There exists an > 0 such that, for all f U with kf kU < ,
there exists an atomistic solution ua U to (7) satisfying
(29) (ua )0 (upc 0 L 2 cb 2 cb 2 3 cb

L ) `2 . q + | u (0)| + k u kL4 + k u kL2 + kf kL2 .
Proof. This result is a consequence of the inverse function theorem (Theorem 7.3)
and Theorems 6.2 (consistency) and 6.3 (stability).
Remark. The term |2 ucb (0)| is an error due to the fact that, in the corrector
problem, we replaced the nonlinear elastic field ucb (x) with a homogeneous field
ucb (0). Identifying this term is the main result of our analysis.
The term L q , which arises due to the approximation of the corrector problem,
may now be balanced against the (uncontrollable) |2 ucb (0)| contribution by choosing
L log |2 ucb (0)|. The remaining terms are the typical error committed by the
CauchyBorn model and the error committed in the approximation of the external
forces, which are represented by k2 ucb k2L4 + k3 ucb kL2 and kf kL2 , respectively.
In preparation for the next section, we observe that the error estimate can be
rewritten, using Theorem 2.5, in terms of f only:
(u ) (upc )0 2 . L
a 0 2
(30) L ` q + |f (0)| + kf kL4 + kf kL2 .
2.5. Rescaling of the external forces. We now consider a standard scaling of

the external force typically employed in the analysis of the CauchyBorn model [12,14].
For the sake of simplicity, we discuss this only formally.
Let 1 denote a length-scale over which we expect elastic strains to vary. This
suggests that after a rescaling of space through x x and u u, the variation will
be on an O(1) scale. The corresponding dual scaling for the force is f (x) f (x),
which motivates us to consider an external force f f () given by
()
(31) f` := f(`),
where f = g for some g H 2 (0, ) with kgkH 2 sufficiently small. We then obtain
that
|f () (0)| = |f(0)|, kf () k2L4 = 3/2 kfk2L4 , and kf () kL2 = 3/2 kfkL2 .
We remark that, in the following, we do not use rescaled x and u, but we only
mentioned this rescaling to motivate our chosen scaling of the external force.
For kgkH 2 sufficiently small and L sufficiently large, we can deduce the existence
of the predictor ucb , the corrector qL , and the corresponding atomistic solution ua .
We can then arrive at the resulting error estimate
(u ) (upc )0 2 . L
a 0 3/2
(32) L ` q ++ ,

where we kept the O(3/2 ) term to emphasize that the CauchyBorn contribution to
the error is negligible compared to the surface contribution. We may now balance the
error again by choosing L = log / log q to finally obtain
(u ) (upc )0 2 . + 3/2 .
a 0
(33) L `
Remark. If we take f as a smooth function rather than a piecewise affine inter-

polant, then the CauchyBorn solution is rescaled due to the force rescaling according
to ucb cb cb
(x) = u (x). In particular, u (0) would then be independent of , and
thus the corrector problem is independent of as well.
In the setting of our analysis, our choice of interpolating f means this observation
is still approximately true. In particular, it is fairly straightforward to extend our
analysis rigorously to the setting of section 2.5.
2.6. Numerical results. In this section, we provide numerical demonstrations
of the predictor-corrector method as well as corroboration of the error estimates we
established.
For the atomistic model, we use the EAM site energies described in (9) with , ,
and given by

r r
(r) = e exp 1 , (r) = fe exp 1 ,
re re
/ /

() = Ec 1 log e ,
e e e
where e = 10.6, fe = 1, Ec = 3.54, = 21, = 6, e = 2, re = 1, and = 8.

These parameters and potentials are taken from [11] and represent an EAM poten-
tial describing a system composed of copper atoms. We note that re represents the
equilibrium distance in the infinite atomistic model without surfaces. Therefore, the
equilibrium spacing in the infinite model is simply 1 and corresponds to a strain of
0 as our reference spacing is also 1. We also note that property (ii) does not hold
for this potential, but as mentioned above, the results remain the same without this
assumption.
For the numerical implementation, we seek energy-minimizing configurations of
the semi-infinite atomistic chain model in the space UN := QN , instead of U. Pro-
vided N is sufficiently large so that supp(f ) [0, N/2], say, it can be proven using
the techniques we employed in our analysis that the additional error committed is ex-
ponentially small in N . Due to the low computational cost in 1D experiments, we can
choose N = 1000 throughout, which guarantees that the additional error committed
from replacing U with UN is negligible.
We find approximate solutions of the CauchyBorn problem (18) by seeking min-
imizers of a discretized CauchyBorn energy given by

X 1 1
cb
W (u0` ),

(34) E (u) := where W (F ) = (1+F )+ (1+F )+ 2(1+F ) .
2 2
`=0
Proposition 4.2 can be employed to prove that this discretization does not introduce
new terms to the error bound (29).
The energy-minimizing configurations for the above models are found using the
steepest-descent method with a backtracking algorithm.

Comparison of Ground States 1

Error Due to Increasing Boundary Layer
0.03 10
Atomistic
CauchyBorn
0.02 PredictorCorrector: L = 2
PredictorCorrector: L = 5 2
10
0.01
L ) k`2
0
3
10
k(ua )0 (upc
0
Strain
0.01 4
10
0.02
5
10
0.03
6
0.04 10
0 2 4 6 8 10 12 14 16 18 20 0 5 10 15
Atom Index Size of Boundary Layer L
(a) (b)
Fig. 1. Ground states for the atomistic, CauchyBorn, and predictor-corrector methods along
with a demonstration of the exponential convergence of the boundary layer error.
2.6.1. Test 1: Ground state. The ground states (f = 0) of the atomistic,

pure CauchyBorn, and predictor-corrector methods are shown in Figure 1a. The
atomistic ground state exhibits the predicted exponential decay, but we also observe
an alternating sign behavior common to certain metals such as copper.
The pure CauchyBorn solution clearly does not capture any surface effects. For
the predictor-corrector method, we see that upc 2 already yields good accuracy, while
upc
5 is visually indistinguishable from ua
gr . In Figure 1b, we quantify this by numeri-
cally demonstrating the exponential convergence of qL = upc L to the ground state.
2.6.2. Test 2: Long wavelength limit. We next consider the error under
the long wavelength limit and rescaling described in section 2.5. We define f(x) =
cos(3x/8)[0,4) (x), where A (x) denotes the characteristic function, and f` = f(`).
For each , we compute the solution to the predictor-corrector method with the size of
the boundary layer set to L = 3+dlog10/9 (1 )e. This choice of L is motivated by the
discussion of balancing the boundary layer contribution to the error in section 2.5.
By choosing a logarithm with a base less than 1 q in the choice of L, the surface
contribution to the error becomes , where > 1 is a constant. Figure 2 shows the
first-order convergence that results when the scaling factor is taken to zero. The
predictor-corrector method behaves precisely as predicted in section 2.5.
2.6.3. Test 3: Error with external forces. We wish to highlight the fact that
the error made by the predictor-corrector methods approximation will not vanish
with increasing boundary layer in the case of fixed nonzero external forces. To that
end, we consider f` = f(`), where f is given in section 2.6.2. In Figure 3a, we
display the energy-minimizing configurations for the atomistic solution, the Cauchy
Born solution, and the predictor-corrector method with L = 5. We observe that
a small error in the boundary layer still persists. Indeed, since the error depends
on the magnitude and regularity of f as demonstrated in (30), it cannot be driven
to zero. This is numerically demonstrated by Figure 3b. We emphasize, however,
that despite our extremely concentrated and nonsmooth choice of f , the predictor-
corrector method is able to reduce the error by a factor of 30, which is an excellent
improvement given the simplicity of the predictor-corrector scheme.
2.7. Conclusions. We produced a detailed analysis of an atomistic surface in a
1D model problem. Based on this analysis, we showed that, while a regular Cauchy

2
Error Due to Bulk Scaling
10
3
10
L ) k`2
0
k(ua )0 (upc
4
10 C
5
10
6
10 2 1 0
10 10 10

Fig. 2. Rate of convergence to atomistic solution in the long-wavelength limit.
Comparison of Minimizers in Presence of Forces 1

Error Due to Increasing Boundary Layer
0.02 10
Atomistic
CauchyBorn
PredictorCorrector
0.01
L ) k`2
0
0
k(ua )0 (upc
Strain
2
0.01 10
0.02
0.03
3
0.04 10
0 2 4 6 8 10 12 14 16 18 20 0 5 10 15
Atom Index Size of Boundary Layer L
(a) (b)
Fig. 3. (a) Energy-minimizing configuration for atomistic and predictor-corrector systems in

the presence of external forces. Here, the size of the boundary layer for the predictor-corrector
problem is L = 5. (b) Demonstration of nonvanishing error in the presence of nonzero external
forces and increasing boundary layer.
Born method fails to accurately capture the surface effects, it can be corrected (post-
processed) using a computationally cheap surface cell problem. We gave a sharp error
analysis of this new predictor-corrector scheme and, in particular, showed that its
error relative to the fully atomistic solution tends to zero in a natural scaling limit.
Our numerical results show promising quantitative behavior of the proposed scheme.
While the analysis is 1D, we anticipate that it can shed some light even on the
three-dimensional case, when surface behavior dominates edge or corner effects. An
important new effect that will have to be taken into account in two and three dimen-
sions is surface stresses that act tangentially. Despite the presence of this additional
stress, we still expect surface effects to be quite localized for certain higher-dimensional
systems as numerical experiments in [10] and analysis in [20] have indicated exponen-
tially decaying responses in two dimensions; thus, we are optimistic regarding the
validity of an extension of the proposed surface cell problem in this paper to higher
dimensions.

3. Proofs: Atomistic model.

Proof of Theorem 2.1. We employ the direct method of the calculus of variations.
Property (iii) on the site potential energies and the fact that |V surf (0)| < imply that
E a (u) 0. We may therefore consider an energy minimizing sequence {un } n=1 U
such that
(35) E a (un ) inf E a (u) and E a (un ) < inf E a (u) + 1.

uU uU
We wish to show that |un |H 1 = ku0n k`2 (Z0 ) is uniformly bounded. To that end,
we will consider separately the contributions to the norm from the small and large
strains. Let > 0. For each n N, we denote the set of indices that indicate the
locations of -defects in the chain to be Dn :
(36) Dn := {` Z0 | |u0n,` | > or |u0n,`+1 | > }.
Property (v) implies that there exists a > 0 such that
(37) V (u0n,` , u0n,`+1 ) for all ` Dn .
Since V 0, we obtain
X
(38) E a (un ) V (u0` , u0`+1 ) #Dn .
`Dn
Hence, the cardinality of Dn is uniformly bounded; that is,
(39) max #Dn < .

n
We now analyze small strains. By Taylors theorem and property (iv), we may
show that for small enough there exists a constant C > 0 such that
V (u0n,` , u0n,`+1 ) C |u0n,` |2 + |u0n,`+1 |2 / Dn .

(40) for all `
With this inequality and the positivity of the site energies, we obtain
X
E a (un ) = V surf (u0n,0 ) + V (u0n,` , u0n,`+1 )
`Z0
X X
= V surf (u0n,0 ) + V (u0n,` , u0n,`+1 ) + V (u0n,` , u0n,`+1 )
`Dn / n
`D
X
C (|u0n,` |2 + |u0n,`+1 |2 ).
/ n
`D
From (35), we deduce the bound

X
(41) |u0n,` |2 C1 inf E a (u) + 1 < .
uU
/ n
`D
Note that the upper bound (41) excludes only a finite number of strains on the
chain (the -defects). If we can show that there exists a positive constant C > 0 such
that
(42) |u0n,` | C for all n N, ` Z0 ,

then we obtain a uniform bound on ku0n k`2 .

Suppose, for contradiction, that (42) fails. Then, there exists a subsequence
{unk }
k=1 and a sequence of indices {`k }k=1 such that
(43) lim |u0nk ,`k | .

k
Since maxn #Dn is finite, we may add the condition on the indices `k that there exists
a constant S > 0 such that
(44) |u0nk ,`k +1 | S for all k.
We now split the chain into two components,

k 1
`X
E a (unk ) = V surf (u0nk ,0 ) + V (u0nk ,` , u0nk ,`+1 )
`=0

X
+ V (u0nk ,`k , u0nk ,`k +1 ) + V (u0nk ,` , u0nk ,`+1 )
`=`k +1
=: E1 + E2 .
Since V 0, we obtain
k 1
`X
(45) E1 = V surf (u0nk ,0 ) + V (u0nk ,` , u0nk ,`+1 ) inf
0
V surf (s0 ) > 0.
s R
`=0
To bound E2 , we observe that this group represents a semi-infinite chain that has,
up to a small error, decoupled from the first group. Therefore, it is bounded below
by the infimum of the energy. To make this precise, let 1 := 21 inf sR V surf (s) > 0.
By properties (vi) and (i), there exists R > 0 such that if |r| > R, then |V (r, s)
V surf (s)| < 1 for all |s| S. From (43), there exists a K > 0 such that |u0nk ,`k | > R
for all k K. Therefore, we can conclude that, for k K,

X
E2 = V (u0nk ,`k , u0nk ,`k +1 ) + V (u0nk ,` , u0nk ,`+1 )
`=`k +1

X
V surf (u0nk ,`k +1 ) 1 + V (u0nk ,` , u0nk ,`+1 )
`=`k +1
inf E a (u) 1 .
uU
In summary, we have shown that, for k K,
(46) E a (unk ) = E1 + E2 1
inf V surf (s)
2 sR + inf E a (u).
uU
Since 12 inf sR V surf (s) > 0, this contradicts the fact that {unk } is a minimizing se-
quence.
Therefore, (42) holds and together with (39) and (41), and we deduce that ku0n k`2
is uniformly bounded. Upon passing to a subsequence (not relabeled), we may assume
that there exists u0 `2 (Z0 ) such that u0n * u0 weakly in `2 . In particular, this
implies that
(47) u0n,` u0,` for all ` Z0 .

We may now prove that u is a minimizer by employing Fatous lemma to estimate
inf E a (u) = lim inf E a (un )

uU n

X
lim inf V surf (u0n,0 ) + lim inf V (u0n,` , u0n,`+1 ) = E a (u ).
n n
`=0
Thus, u is an energy minimizing configuration of the atomistic energy E a in U.

Proof of Theorem 2.2. Critical points of E a satisfy the EulerLagrange equation
(48) hE a (u0 ), v 0 i = 0 for all v U.
Expanding the first variation E a (u0 ) about u0 = 0 yields
(49) hE a (0) + 2 E a (0)u0 + (u0 ), v 0 i = 0,
where (u0 ) is the remainder term from the expansion which can be readily shown to
satisfy the bounds
(50) |0 | . |u00 |2 + |u01 |2 and |` | . |u0`1 |2 + |u0` |2 + |u0`+1 |2 for ` 1.
We may explicitly compute the remaining terms in the Taylor series expansion keeping
in mind that V (F, F ) achieves its minimum at (0, 0):
n o
hE a (0), v 0 i = v00 F V surf (0) and
n o n o
h 2 E a (0)u0 , v 0 i = v00 u00 F2 V surf (0) + 11
2
V (0, 0) + v00 u01 12
2
V (0, 0)
X n o n o
+ 0
v` u0`1 12 2
V (0, 0) + u0` 22 2 2
V (0, 0) + 11 V (0, 0)
`=1
n o
+u0`+1 2
12 V (0, 0) .
Let
(51)
2 2 2
b := 12 V (0, 0), a := 11 V (0, 0)+22 V (0, 0), and as := F2 V surf (0)+11
2
V (0, 0);
then we see that u0 solves the system

(52) as u00 + bu01 = 0 F V surf (0) =: 0 ,
(53) bu0`+1 + au0` + bu0`1 = ` for ` 1.
We now suppose that b 6= 0. The case when b = 0 is analogous. Solving for the
homogeneous solution to this system of equations gives

a a2 4b2
(54) u0` = c+ `+ + c ` , where = .
2b
Positive definiteness of 2 V (0, 0) from property (iv) implies that a 2b > 0 and
a + 2b > 0. Hence, the discriminant is always positive, a2 4b2 > 0, and, as a result,
+ 6= . The symmetry of (53) implies that + = 1/ . This relation combined
with the fact that the discriminant is never 0 implies that 6= 1. Without loss of

generality, then, we must have that 0 < |+ | < 1 and | | > 1. As solutions to the
atomistic problem must belong to U, we have that u0` 0 as ` . This boundary
condition at infinity implies that c = 0 in order to prevent exponential growth in the

strain. Thus, letting := + , we have that solutions of the homogeneous equation
are of the form u0` = u00 ` . A discrete Greens function argument provides the solution
for the inhomogeneous equation:

X
(55) u0` = CBC ` + D1 |`k| k ,
k=0

where D := a2 4b2 and CBC can be determined from u00 . Taking the absolute
value of both sides, using the triangle inequality, and applying (50), we can estimate

X
X
|u0` | CBC ||` + D1 |||`k| |k | . CBC ||` + D1 |||`k| |u0k |2 .
k=0 k=0
Let be a constant such that || < < (1 + ||)/2. Then,

X
X X
(56) |`m| |u0m | . CBC |`m| ||m + D1 |`m| |||mk| |u0k |2 .
m=0 m=0 m,k0
Using the observation that

X
(57) |`m| |||mk| . |`k| ,
m=0
we arrive at

X
X
(58) |`m| |u0m | . CBC ` + D1 |`k| |u0k |2 .
m=0 k=0
Ignoring the prefactor D1 , the second term on the right-hand side of (58) can be
bounded by
(59)

X X X X
|`k| |u0k |2 = |`k| |u0k |2 + |`k| |u0k |2 . ` + sup |u0k | |`k| |u0k |.
kk0
k=0 kk0 kk0 kk0
For any > 0, we can choose k0 sufficiently large so that supkk0 |u0k | . Hence, we
arrive at

X X
(60) D1 |`k| |u0k |2 . ` + |`k| |u0k |.
k=0 kk0
Substituting this bound into (58) yields

X
(61) (1 )|u0` | (1 ) |`k| |u0k | . ` .
k=0

4. Proofs: CauchyBorn model.

Proof of Proposition 2.4. Let u = 0. Then, properties (iii)(v) imply that E a (u)
is given by
X
hE a (u), vi = v00 F V surf (u00 ) + 1 V (u00 , u01 ) + v`0 2 V (u0`1 , u0` ) + 1 V (u0` , u0`+1 )
`=1
= v00 F V surf (0).
This is maximized by taking v00 = sign(F V surf (0)) and v`0 = 0 for ` > 0, thus proving
the first statement.
To deduce the lower bound on the error, we simply observe that
(62) |F V surf (0)| = kE a (0)kU = kE a (0) E a (uagr )kU M k(ucb uagr )0 k`2 .
Before we state the next result, recall that f U if and only if there exists
g `2 (Z0 ) such that hf, ui = hg, u0 i. Summation by parts, with the convention
g1 = 0, may be used to show that
(63) f` = g` g`1 .
Conversely, if we are given f , we may recover g via

X K
X
(64) g` := fk := lim fk .
K
k=`+1 k=`+1
In the CauchyBorn model, we identify f with its piecewise affine interpolant.

The continuous analogue of (64) is
Z Z K
(65) g(x) := f (s) ds := lim f (s) ds.
x K x

Lemma 4.1. Let f U . Then, g and g are well defined and satisfy the estimates
g` g(` + 1/2) 1 |f`0 |

(66) 8 for all ` Z0 and
`+1
X
(67) |g(x)| C |gm | for all x [`, ` + 1].
m=`1
Proof. Since f U , there exists g `2 such that (63) holds. Let M > ` + 1.
Then,
M
X M
X
(68) fm = (g`1 g` ) = g` gM g` as M .
m=`+1 m=`+1
Hence, (64) is well defined.

Similarly, let M Z and M > ` + 1. Then,
Z M M
X Z `+1 Z M M
X
f (s) ds fm = f (s) ds + f (s) ds fm
`+1/2 m=`+1 `+1/2 `+1 m=`+1
M
X 1 h i
1 1
+ f`+1 ) + 12 f`+1 + 1

= 2 4
(f` 2
(fm + fm+1 ) fm fM
m=`+1
1 1 1
= 8
(f` f`+1 ) f
2 M
8
(f` f`+1 ) as M .

Thus, we can conclude that g(x) is well defined for all x, as well as the stated estimate
(66).
To prove the second estimate, we simply note that, for x [`, ` + 1],
Z x

(69) |g(x) g(` + 1/2)| =
f (s) ds . |f` | + |f`+1 |,
`+1/2
and then apply (63).

Next, we prove Theorem 2.5. In addition to the bounds stated therein, we will
also establish that
(70) |ucb (x)| . |g(x)|.
Proof of Theorem 2.5 and of (70). The EulerLagrange equation of the Cauchy
Born problem (18) is

(71) F W (u) = f.
Formally integrating over (x, ) and using the fact that F W (0) = 0, we obtain
Z
(72) F W (u) = f (s)ds = g(x).
x
3
We will now produce a function u Hloc
satisfying this equation.
From properties (iv) and (i), there exists G > 0 such that F2 W (F ) > 21 F2 W (0) >
0 on [G, G]. Thus, F W (F ) is strictly monotone on this interval. The bound (67)
implies that, for cb sufficiently small, g([0, )) F W ([G, G]). Hence, we can
define
(73) u(x) = (F W )1 (g(x)),
where (F W )1 is defined to have range in [G, G]. Due to the restriction on the
range of the inverse function, u(x) [G, G]; hence it follows that the solution
must be stable with a stability constant ccb 21 F2 W (0).
It is furthermore easy to show with this that (F W )1 is twice continuously
differentiable and, in particular, Lipschitz. Noting that (F W )1 (0) = 0, Lipschitz
continuity then yields the bound
(74) |ucb (x)| = |(F W )1 (g(x)) (F W )1 (0)| . |g(x)|,
where we have ignored the Lipschitz constant for the inverse function.
The remaining estimates are consequences of the fact that g C 1,1 (R0 ) with
g = f and 2 g = f along with an elementary computation.
Next, we recall an auxiliary result that allows us to reduce the continuous Cauchy
Born model to a discrete model. To that end, we recall that we can identify discrete
test functions v U with continuous test functions v U cb via piecewise affine
interpolation. Through the same identification, we can also admit u U as arguments
for E cb .
Proposition 4.2. Under the conditions of Theorem 2.5, we have

cb cb
E (u ) E cb (a ucb ), v . k2 ucb k2 4 kvkL2

(75) L for all v U.
Proof. This result is a simplified variant of [12, Lemma 5.2].

5. Proofs: Corrector problem.

Proof of Theorem 2.6. Consistency. It is reasonable to expect that, for |F0 | small,
the corrector solution q is close to uagr , so we may find q by applying the inverse
function theorem in a neighborhood of uagr . From Proposition 7.1, the first variation
of the corrector energy evaluated at q = uagr is

hE (q; F0 ), vi = v00 F V surf (F0 + q00 ) + 1 V (F0 + q00 , F0 + q10 ) F W (F0 )

X
+ v`0 2 V (F0 + 0
q`1 , F0 + q`0 ) + 1 V (F0 + q`0 , F0 + 0
q`+1 ) F W (F0 ) ,
`=1
where v Q . Using the fact that E a (q) = 0 and F W (0) = 0, we obtain
hE (q; F0 ), vi = hE (q; F0 ) E a (q), vi

= v00 F V surf (F0 + q00 ) + 1 V (F0 + q00 , F0 + q10 ) F W (F0 )

F V surf (q00 ) 1 V (q00 , q10 ) + F W (0)

X
+ 0
v`0 2 V (F0 + q`1 , F0 + q`0 ) + 1 V (F0 + q`0 , F0 + q`+1
0
) F W (F0 )
`=1

0
2 V (q`1 , q`0 ) 1 V (q`0 , q`+1
0
) + F W (0)

X
=: v`0 A` .
`=0
Using Lipschitz continuity of all potential derivatives, we easily obtain that
|A0 | . |F0 |.
To estimate A` for ` 1, we proceed more carefully. Expanding with respect to qj

for j = ` 1, `, ` + 1 and employing the identity F W (F ) = 1 V (F, F ) + 2 V (F, F )
and the global Lipschitz continuity of 2 V , we obtain
n o n o

|A` | = 2 V (F0 , F0 ) + 1 V (F0 , F0 ) F W (F0 ) 2 V (0, 0) + 1 V (0, 0) F W (0)
Z 1 n o q 0
0
+ 2 V (F0 + tq`1 , F0 + tq`0 ) 2 V (tq`1
0
, tq`0 ) `1 dt
0 q`0
Z 1n
q 0
o
1 V (F0 + tq`0 , F0 + tq`+1
0
) 1 V (tq`0 , tq`+1
0

+ ) 0` dt
0 q`+1
0
| + |q`0 | + |q`+1
0
. |F0 | |q`1 | .
An application of the CauchySchwarz inequality yields
(76) hE (q; F0 ), vi kAk`2 kv 0 k`2 . |F0 |(1 + kq 0 k`2 )kv 0 k`2 .
Stability. Recall that q = uagr is strongly stable in the atomistic model (12) with
stability constant ca > 0. To prove stability of 2 E (q; F0 ), we simply bound the error

in the Hessians:
n o

2 a 2 0 2 2 surf 0 2 surf 0
E (q) E (q; F0 ) v, v = |v0 | F V (q0 ) F V (F0 + q0 )
n o
0 0 0 0 0 0
X 2 2 2
+ |v` + v`+1 | 12 V (q` , q`+1 ) 12 V (F0 + q` , F0 + q`+1 )
`=0
n o
0 2
2 0 0 2 0 0 2 0 0 2 0 0
11 V (q` , q`+1 ) 12 V (q` , q`+1 ) 11 V (F0 + q` , F0 + q`+1 ) + 12 V (F0 + q` , F0 + q`+1 )
+ |v` |
n o
0 2 2 0 0 2 0 0 2 0 0 2 0 0
+ |v`+1 | 22 V (q` , q`+1 ) 12 V (q` , q`+1 ) 22 V (F0 + q` , F0 + q`+1 ) + 12 V (F0 + q` , F0 + q`+1 ) .
Employing the Lipschitz continuity of F2 V surf and 2 V , we can immediately deduce

that
E (q) 2 E (q; F0 ) v, v C|F0 |kv 0 k22 ,

2 a
`
where the constant C is the upper bound on the Lipschitz constants involved in the
estimate multiplied by a simple factor. Thus, we obtain, for |F0 | 14 ca /C,
3ca 0 2
(77) h 2 E (q; F0 )v, vi kv k`2 for all v Q .
4
The consistency estimate (76), the stability estimate (77), the Lipschitz continuity
of q 7 2 E (q; F0 ), and an application of Theorem 7.3 imply the existence of an
equilibrium q with kq 0 q 0
k`2 . |F0 |. Choosing F0 sufficiently small and once
again using the Lipschitz continuity of 2 E implies (25).
The exponential decay in (26) follows from a straightforward modification of The-
orem 2.2.
To analyze the projection of the corrector problem from Q to QL , we introduce
a projection operator L : Q QL via
0
0 q` , ` = 0, . . . , L,
(78) L q` :=
0, ` > L.
Lemma 5.1. Let q Q satisfy |q`0 | . ` for some [0, 1). Then,
(79) kq 0 L q 0 k`2 . L .
Proof. First, note that

0 0, 0 ` L 1,
(80) q`0 L q =
` q`0 , ` > L.
Thus,

X X 2L
(81) kq 0 L q 0 k2`2 = |q`0 |2 . 2` = .
1 2
`=L `=L
Proof of Proposition 2.7. The result is proven by an application of the quantita-

tive inverse function theorem, Theorem 7.3, using the projected solution L q as an
approximate solution. Observe that
(82) hE (q ; F0 ), vi = 0 for all v QL Q ,

where q solves the infinite corrector problem. By the Lipschitz continuity of E ,

(83)
0 0
|hE (L q ; F0 ), vi| = |hE (L q ; F0 ) E (q ; F0 ), vi| . kq k`2 kv 0 k`2 .
L q
According to Theorem 2.6, there exists 0 q < 1 such that |(q )0` | . `q . Hence,
Lemma 5.1 may be applied to show that
(84) kE (L q ; F0 )kQL = sup |hE (L q ; F0 ), vi| . L

q.
vQL ,kv 0 k`2 =1
Stability of L q follows from the stability of q (Theorem 2.6), the Lipschitz

continuity of 2 E , and Lemma 5.1,

hE (L q ; F0 )v, vi = hE (q ; F0 )v, vi hE (L q ; F0 )v, vi hE (q ; F0 )v, vi
0 0
hE (q ; F0 )v, vi M kq L q k`2 kv 0 k2`2
c
a 0
CM L q kv k`2 ,
2
where M is the Lipschitz constant for E and C is the unlisted constant in (84).
For L sufficiently large, all assumptions of Theorem 7.3 are met, and its applica-
tion completes the proof.
6. Proofs: Predictor-corrector method. The centerpiece of our analysis of
the predictor-corrector method is the following consistency error estimate. In its
statement, we employ the notation
(85) u00` := u0`+1 u0` and u000 0 0 0

` := u`1 2u` + u` .
Theorem 6.1. Let w, q U, u := w + q, and F0 R. Then, for all v U,

"
|hE (u) E (w) E (q; F0 ), vi| . kw00 k2`4 (Z0 ) + kw000 k`2 (Z>0 ) + kq 00 w00 k`2 (Z0 )
a cb

!1/2 #
X
+ |w10 F0 | + 0
(|q`1 | + |q`0 | + |q`+1
0
|)2 |w`0 F0 |2 kv 0 k`2 .
`=1
Proof. The difference in the first variations is given by
hE a (u) E cb (w) E (q; F0 ), vi

n o n o
= v0 F V surf (u00 ) + 1 V (u00 , u01 ) F W (F0 )
0
n o
surf 0 0 0
F V (F0 + q0 ) + 1 V (F0 + q0 , F0 + q1 ) F W (F0 )

X n o n o
+ v`0 2 V (u0`1 , u0` ) + 1 V (u0` , u0`+1 ) F W (w`0 )
`=1
n o
0 0 0 0
2 V (F0 + q`1 , F0 + q` ) + 1 V (F0 + q` , F0 + q`+1 ) F W (F0 )
=: S + B,
where S, B denote, respectively, the surface and bulk contributions to the consistency
error.

Surface term. Using the Lipschitz continuity of the bulk site energy, we can show
that
hn o n o
|S| = v00 F V surf (u00 ) + 1 V (u00 , u01 ) F W (F0 )

n oi
F V surf (F0 + q00 ) + 1 V (F0 + q00 , F0 + q10 ) F W (F0 )

= |v00 | 1 V (F0 + q00 , w10 + q10 ) 1 V (F0 + q00 , F0 + q10 )

(86) . |v00 | |w10 F0 |.

P
Bulk sum. We write B = `=1 v`0 B` , where
n o n o
B` = 2 V (u0`1 , u0` ) + 1 V (u0` , u0`+1 ) F W (w`0 )
n o
0
2 V (F0 + q`1 , F0 + q`0 ) + 1 V (F0 + q`0 , F0 + q`+1
0
) F W (F0 ) .
Using the identity F W (F ) = 1 V (F, F ) + 2 V (F, F ) and adding 0, we can split B`

into
h i h i
B` = 2 V (u0`1 , u0` ) + 1 V (u0` , u0`+1 ) 2 V (w`0 + q`1
0
, w`0 + q`0 ) + 1 V (w`0 + q`0 , w`0 + q`+1
0
)
h i h i
0 0 0 0 0 0 0 0 0 0 0 0
+ 2 V (w` + q`1 , w` + q` ) + 1 V (w` + q` , w` + q`+1 ) 2 V (w` , w` ) + 1 V (w` , w` )

0
2 V (F0 + q`1 , F0 + q`0 ) + 1 V (F0 + q`0 , F0 + q`+1
0
) 2 V (F0 , F0 ) 1 V (F0 , F0 )
(1) (2) (3)
=: B` + B` B` .
(1)
Estimate for B` . Expanding V , we obtain, using u = w + q,
u0`1 w`0 q`10

= 2 V (w`0 + q`1
0
, w`0 + q`0 )
(1)
B` 0 0 0
u` w` q`
u0` w`0 q`0

+ 1 V (w`0 + q`0 , w`0 + q`+1
0
) 00
+ O(|w`1 |2 + |w`00 |2 )
u0`+1 w`0 q`+1
0
= 12 V (w`0 + q`0 , w`0 + q`+1

0
) w`00 12 V (w`0 + q`1
0
, w`0 + q`0 ) w`1
00 00
+ O(|w`1 |2 + |w`00 |2 ).
Expanding the result again yields

(87)
(1)
B` = 12 V (w`0 +q`0 , w`0 +q`0 ) (w`00 w`1
00
)+O |q`00 ||w`1
00
|+|q`00 ||w`00 |+|w`1
00
|2 +|w`00 |2 .

Thus, we obtain
(1)
(88) |B` | . |w`000 | + |w`1
00
|2 + |w`00 |2 + |q`00 | |w`1
00
| + |q`00 | |w`00 |.
(2) (3) (2) (3)

Estimate for B` B` . The two terms B` and B` are of similar structure

(2) (3)
and will need to be treated together. First, we rewrite B` and B` in the form
Z 1 0
q
(2)
B` = 2 V (w`0 + tq`1
0
, w`0 + tq`0 ) `1 dt
0 q`0
Z 1 0
q
+ 1 V (w`0 + tq`0 , w`0 + tq`+1
0
) 0 ` dt and
0 q `+1
Z 1 0
(3) 0 q
B` = 2 V (F0 + tq`1 , F0 + tq`0 ) `1 dt
0 q`0
Z 1 0
q
+ 1 V (F0 + tq`0 , F0 + tq`+1
0
) 0 ` dt.
0 q`+1
Subtracting and rearranging terms yields

Z 1 0
q
2 V (w`0 + tq`1
0
, w`0 + tq`0 ) 2 V (F0 + tq`1
0
, F0 + tq`0 ) `1
(2) (3)
B` B` = dt
0 q`0
Z 1 0
q
+ 1 V (w`0 + tq`1
0
, w`0 + tq`0 ) 1 V (F0 + tq`0 , F0 + tq`+1
0
) 0 ` dt.
0 q `+1
Making use of the Lipschitz continuity of 2 V once again, we deduce that

Z 1 0 0 Z 1 0 0
(2)
B B (3) .
|w` F0 | q`1 |w` F0 | q
` ` 0 0 dt + 0 0 ` dt
0 |w` F0 | q ` 0 |w` F 0 | q `+1
0
. (|q`1 | + |q`0 | + |q`+1
0
|)|w`0 F0 |.
Combining this estimate with (88), summing `, and applying the CauchySchwarz
inequality yields the interior contribution of the stated result.
Theorem 6.2 (consistency). Let kf kU be sufficiently small and L be sufficiently
large so that the conditions of Theorems 2.5 and 2.6 with F0 = ucb (0) and of Propo-
sition 2.7 are satisfied. Let ucb and qL denote the corresponding CauchyBorn and
corrector solutions. Then,
(89) kE a (a ucb +qL )f kU . L 2 cb 2 cb 2 3 cb
q +| u (0)|+k u kL4 +k u kL2 +kf kL2 .
Proof. As ucb and q are solutions to their respective problems, hE cb (ucb ), vi =

hg, viR0 for all v U cb with g defined as in (65) and hE (q ; F0 ), vi = 0 for all
v Q . We recall that U can be seen as a subspace of U cb and that Q = U.
Let u := a ucb + qL and F0 = ucb (0), then using the fact that ucb solves (18),
we can split the consistency error into
hE a (u) f, vi = hext + cb + + pc , vi, where
ext := hf, iR0 hf, iZ0 ,
cb := E cb (a ucb ) E cb (ucb ),
:= E a (a ucb + qL ) E a (a ucb + q ), and
pc := E a (a ucb + q ) E cb (a ucb ) E (q ; F0 ).
The term ext represents the error in the action of the external forces. In the
atomistic model, the external forces can be written as hf, viZ 0 = hg, v 0 iZ0 , where g
is defined as in (64). Using (66), we get that
(90) hext , vi = |hg, viR0 hg, v 0 iZ0 | . kf kL2 kv 0 k`2 for all v U.

The term cb represents the error associated with the discretization of the Cauchy
Born problem and was estimated in Theorem 4.2 to be
(91) kcb kU . k2 ucb k2L4 .
The term represents the error associated with the Galerkin projection of the
corrector problem and was estimated in Theorem 2.7:
(92) k kU . L
q.
Finally, pc , the error due to the predictor-corrector method is estimated in The-

orem 6.1:
kpc kU . k(a ucb )00 k2`4 + k(a ucb )000 k`2 + kq 00 (a ucb )00 k`2

!1/2
X 2 2
+ (a ucb )00 F0 + 0
|q`0 | 0
(a ucb )0` F0

|q`1 | + + |q`+1 |
`=1
=: P1 + P2 + P3 + P4 + P5 ,
where we have written q q for the sake of simplicity of notation.

To proceed, we need to relate discrete and continuous derivatives. For example,
`+1
Z
(a ucb )00` = (a ucb )0`+1 (a ucb )0` = cb cb

u (x + 1) u (x) dx
`
Z `+1 Z 1 Z `+2
2 ucb (x + t) dt dx |2 ucb (x)| dx.

=
` 0 `
Analogously, we can prove

Z `+2
(a ucb )000 |3 ucb (x)| dx

` and

`1
Z `+1
(a ucb )0` F0 |2 ucb (x)| dx.

(93)
0
Using these estimates, it is straightforward to establish that
P1 = k(a ucb )00 k2`4 . k2 ucb k2L4 and P2 = k(a ucb )000 k`2 . k3 ucb kL2 .
The remaining three terms have a similar structure. Estimating |q`00 | . `q = e`

for some > 0, we deduce
Z
(94) P32 = kq 00 (a ucb )00 k2`2 . e2x |2 ucb (x)|2 dx.
0
Applying (93), and using the fact that ex is bounded below on [0, 1], we have
2
Z 1
P42 = (a ucb )00 F0 . e2x |2 ucb (x)|2 dx.

(95)
0

Finally, applying (93) again, together with |q`0 | . `q = e` , we can estimate

X
0
2 2
P52 = |q`1 | + |q`0 | + |q`+1
0
| (a ucb )0` F0
`=1

X Z `+1 2 Z Z t 2
. 2`
q |2 ucb (x)| dx . e2t |2 ucb (x)| dx dt
`=1 0 0 0
Z Z t Z Z
. e2t t |2 ucb (x)|2 dx dt = |2 ucb (x)|2 e2t t dt dx
Z0 0 0 x
2x 2 cb 2
. (1 + x)e | u (x)| dx.
0
Thus, we can combine

Z
(96) P32 + P42 + P52 . (1 + x)e2x |2 ucb (x)|2 dx.
0
To finalize, we use the same argument as in the estimate for P5 to deduce that
Z
P32 + P42 + P52 . |2 ucb (0)|2 + (1 + x)e2x |2 ucb (x) 2 ucb (0)|2 dx
Z0
2 cb
. | u (0)| + 2
(1 + x)2 e2x |3 ucb (x)|2 dx
0
. |2 ucb (0)|2 + k3 ucb k2L2 .
Combining the estimates for ext , cb , and those for Pj , j = 1, . . . , 5, we arrive
at the stated result.
Theorem 6.3 (stability). There exist , L0 > 0 such that for all f U with
kf kU < and all L L0 , we have
ca
(97) h 2 E a (a ucb + qL )v, vi kv 0 k2`2 ,
2
a
where ca denotes the stability constant for ugr .
Proof. First, let f be small enough and L be large enough so that Theorems
2.5 and 2.6 and Proposition 2.7 hold. The second variation of the atomistic energy
evaluated at the predictor-corrector solution can be written as
h 2 E a (a ucb + qL )v, vi = h 2 E a (uagr )v, vi h 2 E a (uagr )v, vi h 2 E a (q )v, vi

h 2 E a (q )v, vi h 2 E a (qL )v, vi h 2 E a (qL )v, vi h 2 E a (a ucb + qL )v, vi .

The highlighted difference terms can all be bound using the Lipschitz continuity of
the second variation of the atomistic energy:
(98) |h 2 E a (uagr )v, vi h 2 E a (q )v, vi| . k(uagr )0 (q )0 k`2 kv 0 k2`2 ,
(99) |h 2 E a (q )v, vi h 2 E a (qL )v, vi| . k(q )0 (qL )0 k`2 kv 0 k2`2 ,
0
(100) |h 2 E a (qL )v, vi h 2 E a (a ucb + qL )v, vi| . k a ucb k`2 kv 0 k2`2 .
Each of these bounds goes to zero either for large L or for small kf kU . The bound
in (98) goes to zero as kf kU 0 by Theorem 2.6, the bound in (99) goes to zero as
L by Proposition 2.7, and the bound in (100) goes to zero as f 0 by Theorem
2.5. Finally, h 2 E a (uagr )v, vi ca kv 0 k2`2 by definition. Thus, there exist an L0 > 0 and
an > 0 such that the theorem statement holds.

7. Appendix.
Proposition 7.1 (first variations). Let u, v U. Then,
X
hE a (u), vi = v00 1 V surf (u00 ) + 1 V (u00 , u01 ) + v`0 2 V (u0`1 , u0` ) + 1 V (u0` , u0`+1 ) ,
`=1

cb
hE (q; u ), vi = v00 1 V surf (F0 + q00 ) + 1 V (F0 + q00 , F0 + q10 ) F W (F0 )

X
+ v`0 2 V (F0 + q`1
0
, F0 + q`0 ) + 1 V (F0 + q`0 , F0 + q`+1
0
) F W (F0 ) .
`=1
cb
Let u, v U . Then,
Z
cb
(101) hE (u), vi = F W (u) vdx.
0
Proposition 7.2 (second variations). Let u, v, w U. Then,

n o n o
h 2 E a (u)w, vi = v00 F0 F2 V surf (u00 ) + 11
2
V (u00 , u01 ) + v00 w10 12 2
V (u00 , u01 )
X n o n o
0 0
+ v` w`1 2
12 V (u0`1 , u0` ) + w`0 222
V (u0`1 , u0` ) + 11
2
V (u0` , u0`+1 )
`=1
n o
0 2 0 0
+w`+1 12 V (u` , u`+1 ) ,
n o
h 2 E (q; ucb )w, vi = v00 F0 F2 V surf (F0 + q00 ) + 112
V (F0 + q00 , F0 + q10 )
n o
+ v00 w10 122
V (F0 + q00 , F0 + q10 )
X n o
+ v`0 w`1
0 2
12 0
V (F0 + q`1 , F0 + q`0 )
`=1
n o
+w`0 22
2 0
V (F0 + q`1 , F0 q`0 ) + 11
2
V (F0 + q`0 , F0 + q`+1
0
)
n o
0
+w`+1 2
12 V (F0 + q`0 , F0 + q`+10
) .
Let u, v, w U cb . Then,
Z
2 cb
(102) h E (u)w, vi = F2 W (u) w vdx.
0
7.1. Inverse function theorem.

Theorem 7.3. Let H be a Hilbert space equipped with norm k kH , and let G
C 1 (H, H ) with Lipschitz-continuous derivative G:
(103) kG(u) G(v)kL M ku vkH for all u, v H,
where k kL denotes the L(H, H )-operator norm.

Let u H satisfy
kG(u)kH ,
hG(u)v, vi kvk2H for all v H,

such that M , , and satisfy the relation

2M
(104) < 1.
2
Then, there exists a (locally unique) u H such that G(u) = 0,

ku ukH 2 , and

2M
hG(u)v, vi 1 2 kvk2H for all v H.

REFERENCES
[1] X. Blanc, C. Le Bris, and P.-L. Lions, From molecular models to continuum me-
chanics, Arch. Ration. Mech. Anal., 164 (2002), pp. 341381, https://doi.org/10.1007/
s00205-002-0218-5.
[2] M. S. Daw and M. I. Baskes, Embedded-atom method: Derivation and application to im-
purities, surfaces, and other defects in metals, Phys. Rev. B, 29 (1984), pp. 64436453,
https://doi.org/10.1103/PhysRevB.29.6443.
[3] J. Diao, K. Gall, and M. L. Dunn, Surface-stress-induced phase transformation in metal
nanowires, Nature Materials, 2 (2003), pp. 656660, https://doi.org/10.1038/nmat977.
[4] M. Dobson and M. Luskin, An analysis of the effect of ghost force oscillation on qua-
sicontinuum error, M2AN Math. Model. Numer. Anal., 43 (2009), pp. 591604, https:
//doi.org/10.1051/m2an/2009007.
[5] M. Dobson, M. Luskin, and C. Ortner, Accuracy of quasicontinuum approximations near
instabilities, J. Mech. Phys. Solids, 58 (2010), pp. 17411757, https://doi.org/10.1016/j.
jmps.2010.06.011.
[6] M. Dobson, M. Luskin, and C. Ortner, Sharp stability estimates for the force-based quasi-
continuum approximation of homogeneous tensile deformation, Multiscale Model. Simul.,
8 (2010), pp. 782802, https://doi.org/10.1137/090767005.
[7] M. Dobson, M. Luskin, and C. Ortner, Stability, instability, and error of the force-based
quasicontinuum approximation, Arch. Ration. Mech. Anal., 197 (2010), pp. 179202, https:
//doi.org/10.1007/s00205-009-0276-z.
[8] K. Gall, J. Diao, and M. L. Dunn, The strength of gold nanowires, Nano Lett., 4 (2004),
pp. 24312436, https://doi.org/10.1021/nl048456s.
[9] M. E. Gurtin and A. I. Murdoch, A continuum theory of elastic material surfaces, Arch.
Ration. Mech. Anal., 57 (1975), pp. 291323, https://doi.org/10.1007/BF00261375.
[10] K. Jayawardana, C. Mordacq, C. Ortner, and H. S. Park, An analysis of the boundary
layer in the 1D surface CauchyBorn model, ESAIM Math. Model. Numer. Anal., 47
(2013), pp. 109123, https://doi.org/10.1051/m2an/2012021.
[11] R. A. Johnson, Analytic nearest-neighbor model for fcc metals, Phys. Rev. B, 37 (1988),
pp. 39243931, https://doi.org/10.1103/PhysRevB.37.3924.
[12] M. Luskin and C. Ortner, Atomistic-to-continuum coupling, Acta Numer., 22 (2013),
pp. 397508, https://doi.org/10.1017/S0962492913000068.
[13] C. Ortner, A priori and a posteriori analysis of the quasinonlocal quasicontinuum
method in 1D, Math. Comp., 80 (2011), pp. 12651285, https://doi.org/10.1090/
S0025-5718-2010-02453-6.
[14] C. Ortner and F. Theil, Justification of the Cauchy-Born approximation of elastodynam-
ics, Arch. Ration. Mech. Anal., 207 (2013), pp. 10251073, https://doi.org/10.1007/
s00205-012-0592-6.
[15] H. S. Park, K. Gall, and J. A. Zimmerman, Shape memory and pseudoelasticity in metal
nanowires, Phys. Rev. Lett., 95 (2005), 255504, https://doi.org/10.1103/PhysRevLett.95.
255504.
[16] H. S. Park and P. A. Klein, A surface Cauchy-Born model for silicon nanostructures, Com-
put. Methods Appl. Mech. Engrg., 197 (2008), pp. 32493260, https://doi.org/10.1016/j.
cma.2007.12.004.
[17] H. S. Park, P. A. Klein, and G. J. Wagner, A surface Cauchy-Born model for nanoscale
materials, Internat. J. Numer. Methods Engrg., 68 (2006), pp. 10721095, https://doi.org/
10.1002/nme.1754.

[18] E. B. Tadmor and R. E. Miller, Modeling Materials: Continuum, Atomistic, and Multiscale
Techniques, Cambridge University Press, Cambridge, UK, 2011.
[19] E. B. Tadmor, M. Ortiz, and R. Phillips, Quasicontinuum analysis of defects in solids,
Philos. Mag. A, 73 (1996), pp. 15291563, https://doi.org/10.1080/01418619608243000.

[20] F. Theil, Surface energies in a two-dimensional mass-spring model for crystals, ESAIM Math.
Model. Numer. Anal., 45 (2011), pp. 873899, https://doi.org/10.1051/m2an/2010106.

Binder, Luskin &amp; Ortner (2017)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Binder, Luskin &amp; Ortner (2017)

Uploaded by

Copyright:

Available Formats

MULTISCALE MODEL. SIMUL.

c 2017 Society for Industrial and Applied Mathematics

ANALYSIS OF A PREDICTOR-CORRECTOR METHOD FOR

COMPUTATIONALLY EFFICIENT MODELING OF SURFACE

Key words. surface, predictor-corrector, CauchyBorn

AMS subject classifications. 74B20, 65Z05

1. Introduction. Long-range elastic fields dictate the behavior of crystalline

2017; published electronically May 16, 2017.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

numerical simulation. Furthermore, even in bulk crystals, if effects of interest take

(1) u0` := u`+1 u` = y`+1 y` 1.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

The surface of the chain is located at the atom with index 0.

energy (effectively second-neighbor interaction). This is the simplest setting within

(4) inf V surf (s) > 0.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

(5) U := {u : Z0 R | u(0) = 0 and u0 `2 (Z0 )}.

(7) ua arg min{E a (u) hf, uiZ0 | u U}.

where is a nearest-neighbor pair potential, is an embedding energy function, and

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

define the bulk and surface site energies, respectively, as

V (u0`1 , u0` ) = (1 + u0`1 ) + (1 + u0` ) + (1 + u0`1 ) + (1 + u0` ) and

(11) |(uagr )0` | . `a for all ` Z0 .

Proof. See section 3.

(12) h 2 E a (uagr )v, vi ca kv 0 k2`2 (Z0 ) for all v U.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

problem if and only if it satisfies the EulerLagrange equation

(13) hE a (ua ), vi = hf, vi for all v U

as well as the stability condition

(14) h 2 E a (ua )v, vi ckv 0 k2`2 (Z0 ) for all v U,

where W (F ) := V (F, F ) is the CauchyBorn energy density function and

(17) sup |hE a (0), vi| = kE a (0)kU = |F V surf (0)|.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

Proof. See section 4.

istic energy-minimizing configurations are not in agreement because of surface effects.

(18) ucb arg min{E cb (u) hf, uiR+ | u U cb },

hE cb (ucb ), vi = hf, viR+ for all v U cb and

Proof. See section 4.

2.4. Predictor-corrector method. Proposition 2.4 demonstrates that the

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

rector problem via the minimization of a corrector energy, which is given by

(23) E (q; F0 ) = V surf (F0 + q00 ) W (F0 ) q00 F W (F0 )

where we will normally take F0 := ucb (0).

(24) qL arg min{E (q; F0 ) | q QL },

QL := {q U | q`0 = 0 for all ` L}. In particular, Q = U.

(26) |(q )0` | . `q for all ` Z0 .

Proof. See section 5.

Proof. See section 5.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

If kf kU is sufficiently small, then Theorem 2.5 guarantees the existence of a

(29) (ua )0 (upc 0 L 2 cb 2 cb 2 3 cb

2.5. Rescaling of the external forces. We now consider a standard scaling of

|f () (0)| = |f(0)|, kf () k2L4 = 3/2 kfk2L4 , and kf () kL2 = 3/2 kfkL2 .

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

error again by choosing L = log / log q to finally obtain

Remark. If we take f as a smooth function rather than a piecewise affine inter-

where e = 10.6, fe = 1, Ec = 3.54, = 21, = 6, e = 2, re = 1, and = 8.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

Comparison of Ground States 1

2.6.1. Test 1: Ground state. The ground states (f = 0) of the atomistic,

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

Fig. 2. Rate of convergence to atomistic solution in the long-wavelength limit.

Comparison of Minimizers in Presence of Forces 1

Fig. 3. (a) Energy-minimizing configuration for atomistic and predictor-corrector systems in

Binder, Luskin & Ortner (2017)

Binder, Luskin & Ortner (2017)