INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng. 44, 373391 (1999)

FINITE ELEMENTS FOR MATERIALS WITH
STRAIN GRADIENT EFFECTS
JOHN Y. SHU*, WAYNE E. KING AND NORMAN A. FLECK`
Chemistry and Materials Science Department, awrence ivermore National aboratory, ivermore, CA 94550, .S.A.
`Cambridge niversity Engineering Department, rumpington Street, Cambridge, CB2 1PZ, .K.
SUMMARY
A nite element implementation is reported of the FleckHutchinson phenomenological strain gradient
theory. This theory ts within the ToupinMindlin framework and deals with rst-order strain gradients
and the associated work-conjugate higher-order stresses. In conventional displacement-based approaches,
the interpolation of displacement requires C-continuity in order to ensure convergence of the nite element
procedure for higher-order theories. Mixed-type nite elements are developed herein for the FleckHutchin-
son theory; these elements use standard C"-continuous shape functions and can achieve the same conver-
gence as C elements. These C" elements use displacements and displacement gradients as nodal degrees of
freedom. Kinematic constraints between displacement gradients are enforced via the Lagrange multiplier
method. The elements developed all pass a patch test. The resulting nite element scheme is used to solve
some representative linear elastic boundary value problems and the comparative accuracy of various types
of element is evaluated. Copyright 1999 John Wiley & Sons, Ltd.
KEY WORDS: strain gradient; nite element
INTRODUCTION
Conventional continuum mechanics theories assume that stress at a material point is a function
of state variables, such as strain, at the same point. This local assumption has long been proved
to be adequate when the wavelength of a deformation eld is much larger than the dominant
micro-structural length scale of the material. However, when the two length scales are compara-
ble, the assumption is questionable as the material behaviour at a point is inuenced by the
deformation of neighbouring points. Starting from the pioneering Cosserat couple stress theory,
various non-local or strain gradient continuum theories have been proposed. In the full Cosserat
theory, an independent rotation quantity 0 is dened in addition to the material displacement u;
couple stresses (bending moment per unit area) are introduced as the work conjugate to the
micro-curvature (that is, the spatial gradient of 0). Later, Toupin` and Mindlin` proposed a more
general theory which includes not only micro-curvature, but also gradients of normal strain. Both
* Correspondence to: John Y. Shu, Chemistry and Materials Science Department, Lawrence Livermore National
Laboratory, Livermore, CA 94550, U.S.A. E-mail: [email protected]
Contract/grant sponsor: U.S. Department of Energy
Contract/grant sponsor: Lawrence Livermore National Laboratory; Contract/grant number: W-7405-Eng-48
CCC 00295981/99/03037319$17.50 Received 26 July 1997
Copyright 1999 John Wiley & Sons, Ltd. Revised 24 March 1998
the Cosserat and ToupinMindlin theories were developed for linear elastic materials. After-
wards, non-local theories for plastic materials have been developed by, among others, Aifantis,"
Fleck and Hutchinson.`' FleckHutchinson strain gradient plasticity theory' falls within the
ToupinMindlin framework. Interest in non-local continuum plasticity theories has been rising
recently, due to an increasing number of observed size eects in plasticity phenomena. The reader
is referred to a recent review article' for a summary of the experimental evidence.
The purpose of the current paper is to report the details of a nite element implementation of
the ToupinMindlin formulation of strain gradient theories. The ToupinMindlin formulation
furnishes strain gradients and higher order stresses which enter the principle of virtual work as
work conjugates. The strain gradient quantities are the second-order spatial derivatives of
displacement; therefore, in order to guarantee convergence of a displacement-based nite element
analysis upon mesh renement, the interpolation of displacement should exhibit C-continuity,
i.e. both displacement and its rst-order derivatives are required to be continuous across
inter-element boundaries.` For example, the use of Spechts triangular element` for the special
case of couple stress theory was examined." The element contains displacement derivatives as
extra nodal degrees of freedom (denoted as DOF subsequently), and C-continuity is satised
only in a weak averaged sense along each side of the element; therefore, the element is not a strict
C element. Furthermore, the element fails to deliver an accurate pressure distribution for an
incompressible, non-linear solid." There is a reliable rectangular C element,` but its shape is
obviously a strong limitation.
The lack of robust C-continuous elements drove the development of various C"-continuous
elements for couple stress theory."

The elements developed by Xia and Hutchinson" and by

Shu and Fleck" have a rotation angle 0 as an extra DOF at each node. The elements developed
by Herrmann are based on a mixed formulation which takes the couple stresses as extra nodal
DOF; it is not obvious how Herrmanns elements can be adapted to non-linear problems.
For the more general ToupinMindlin strain gradient theory, the straightforward general-
ization of the elements developed by Xia and Hutchinson" and by Shu and Fleck" would require
roughly 3 times as many nodal DOF as an element for a conventional theory. There is a strong
motivation to devise suitable elements with a smaller increase of computational cost. Six dierent
types of two-dimensional nite elements are developed in this paper. They all use C"-continuous
shape functions, and they all pass a patch test. Computations are performed for two example
problems and the performance of the new elements is evaluated by comparing numerical results
with analytical solutions. Several types of element are capable of producing accurate results, with
only a moderate increase in computational cost compared with elements for conventional
theories. The best-performing element has been used to analyse the shear deformation of
a bicrystal obeying an elasticviscoplastic strain gradient crystal plasticity formulation and is
found to generate satisfactory results.`
In this paper, deformations are assumed to be innitesimally small. A subscript index, unless
stated otherwise, takes a value of 13 and repeated indices implies a summation over 13.
A Cartesian reference-frame is employed throughout the paper, and a bold character represents
a vector or tensor.
REVIEW OF LINEAR ELASTIC STRAIN GRADIENT THEORY
Toupin` and Mindlin` developed a theory of linear elasticity whereby the strain energy density
per unit volume depends upon both strain c
GH
,(u
G H
#u
H G
)/2 and strain gradient p
GHI
,u
I GH
. Here
374 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
u is the displacement eld and a comma represents partial dierentiation with respect to
a Cartesian co-ordinate. Note that c
GH
"c
HG
and p
GHI
"p
HGI
. In addition to Cauchy stress o
GH
, this
theory furnishes higher-order stress t
GHI
("t
HGI
) which is work conjugate to the strain gradient p
GHI
.
Fleck and Hutchinson' extended the ToupinMindlin theory by including plasticity.
The principle of virtual work can be expressed as'

T
[o
GH
oc
GH
#t
GHI
op
GHI
] d"

T
[b
I
ou
I
] d#

Q
[ f
I
ou
I
#r
I
Dou
I
] dS (1)
for an arbitrary displacement increment ou. Here b
I
is the body force per unit volume of the body
while f
I
and r
I
are the truncation and double stress traction per unit area of the surface S,
respectively. They are in equilibrium with the Cauchy stress o
GH
and the higher-order stress
t
GHI
according to
b
I
#(o
GI
!t
HGI H
)
G
"0 (2)
f
I
"n
G
(o
GI
!t
HGI H
)#n
G
n
H
t
GHI
(D
N
n
N
)!D
H
(n
G
t
GHI
) (3)
and
r
I
"n
G
n
H
t
GHI
(4)
In the above equations, D
H
( . )"(o
HI
!n
H
n
I
)c( . ) /cx
I
is a surface-gradient operator and
D( . )"n
I
c( . )/cx
I
is surface normal-gradient operator. n
G
is the ith component of the unit surface
normal vector.
The strain gradient theory dictates that, in order to have a unique solution, six boundary
conditions must be independently prescribed at any point on the surface of the body, i.e. f
I
or u
I
,
and r
I
or Du
I
. The extra boundary conditions on r
I
or Du
I
are characteristic of the strain gradient
theory and, as can be seen for a specic example in the Appendix, they imply strain continuity
across the interface between two dissimilar materials.
The constitutive law governing the stress o
GH
and the higher-order stress t
GHI
for an elastic solid is
derived through o
GH
"cw/cc
GH
and t
GHI
"cw/cp
GHI
where w is the strain energy density per unit
volume. Mindlin` showed that for a general isotropic linear hyper-elastic solid, w can be
expressed as
w"
`
zc
GG
c
HH
#jc
GH
c
GH
#a

p
GHH
p
GII
#a
`
p
GGI
p
IHH
#a
`
p
GGI
p
HHI
#a
"
p
GHI
p
GHI
#a
`
p
GHI
p
HIG
(5)
in terms of the invariants of the second-order strain tensor and the third-order strain gradient
tensor. Here z and j are the standard Lame constants and the ve a
L
are additional constants
of dimension stress times length squared. w can also be rewritten in the following canonical
form:'
w"
`
zc
GG
c
HH
#jc
GH
c
GH
#j(c

p
'''
GHI
p
'''
GHI
#c
`
p
''`'
GHI
p
''`'
GHI
#c
`
p
''`'
GHI
p
''`'
GHI
#c
"
p
'
GHI
p
'
GHI
#c
`
p
'
GHI
p
''`'
HIG
) (6)
where q''G' and q' are the bases of a unique orthogonal decomposition of q introduced by
Smyshlyaev and Fleck.` q' is the hydrostatic part of q while q''G' are the three orthogonal
bases of the deviatoric part q'"q!q'. For an incompressible solid, q',0. The details of
the decomposition are omitted here for the sake of brevity. The constants c
L
have the dimensions
of length squared. In the following context, c
L
will be referred to as constitutive length
scales.
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 375
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
FINITE ELEMENT IMPLEMENTATION
Weak form of the principle of virtual work
Second-order derivatives of displacement appear in the principle of virtual work (equation (1)),
implying that displacement-based elements of C-continuity are generally necessary in a nite
element procedure.` However, no robust C-continuous elements are available in the literature
for application to the strain gradient theory outlined above." In this paper, C"-continuous
elements of mixed type are devised. We will rst derive a weak form of the principle of virtual
work suitable for nite element implementation using C"-shape functions. The strategy is to
introduce additional nodal degrees of freedom and to enforce the kinematic constraints between
displacement and strain by Lagrange multipliers.
We begin by developing a modied virtual work statement involving only rst-order gradients
of kinematic quantities. Consider a strain gradient solid with a stress eld satisfying the
equilibrium relations (2)(4). We introduce a second-order tensor and a related third-order
tensor q such that pL
GHI
is dened by pL
GHI
,(
HI G
#
GI H
)/2. The equilibrium equations (2)(4) may
be re-written in weak form, by making use of Stokes surface divergence theorem, as

T
[o: oc#t oqL#Vot : (o!Voou) ] d"

T
[b ) ou] d#

Q
[f ) ou#nr : o] dS
#

Q
(n ) t!nr) : (o!Voou) dS (7)
for arbitrary variations of ou and o. Here Vo is the standard forward gradient operator. If is
subjected strictly to the constraint of
o"Voou (8)
everywhere, then (7) degenerates to (1) and q becomes identically the strain gradient q as dened
in the previous section. However, as previously discussed, the strict enforcement of this constraint
will demand C-continuous elements. Therefore, to facilitate the use of convenient C"-continuous
elements, the constraint (8) is enforced in the following weighted residual manner:

T
(
HI
!u
I H
) ot
GHI G
d"0 (no sum on j and k) (9)
for an arbitrary variation of the Lagrange multipliers ot
GHI G
. Satisfaction of (8) in the volume
average sense by application of (9) ensures that the third term on the right-hand side of (7) is
negligible. Finally, by denoting the Lagrange multipliers Vo ) t as p the modied virtual work
statement (7) becomes

T
[o: oc#t oq #p : (o!Voou) ] d+

T
[b ) ou] d#

Q
[f ) ou#y` : o] dS (10)
where the second-order tensor r` denotes nr for later convenience and (9) becomes

T
(
HI
!u
I H
) oj
HI
d"0 (no sum on j and k) (11)
376 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Henceforth, we shall refer to as the relaxed displacement gradient in contrast to the true
displacement gradient Vou and similarly, c "(#') /2 as the relaxed strain in contrast to the
true strain c. Then, the Lagrange multiplier technique (9) ensures that is a good approximation
to Vou. Consequently, the relaxed strain c is a good approximation to c, and the relaxed strain
gradient q is a good approximation to the true strain gradient q. Both and u are treated as
independent nodal degrees of freedom. Only rst-order derivatives of the DOF are involved in the
weak form of the principle of virtual work equation (10), therefore C"-continuous interpolation of
both u and are sucient to ensure convergence of the nite element procedure. The virtual
work statement (10) is suitable for a compressible solid. In the incompressible limit, the mean
stress o
K
"o
II
/3 is excluded from o and treated as an additional Lagrange multiplier to enforce
c
II
"0, in a well-documented way for solids without strain gradient eects.`
As stated in the previous section, to obtain a unique solution for a boundary value problem, six
independent boundary conditions should be prescribed on the surface for a three-dimensional
problem (and four independent boundary conditions for a two-dimensional problem). While the
boundary conditions associated with nodal DOF of displacement are easily set, the way to
prescribe boundary conditions associated with nodal DOF needs to be specially addressed. Let
us designate a prescribed quantity on the surface with a superscript*. Then, for the case of pure
traction-loading, stress-tractions f * and double stress tractions r* are applied, giving
f"f * and r`"nr* (12)
For displacementloading, displacement u* and displacement gradient Du* are prescribed as
follows. Denote t and s as two orthogonal surface tangential unit vectors which form a triad with
the normal unit vector n. The boundary conditions are specied as
u"u* and "nDu*#tD
R
u*#sD
Q
u* (13)
where D
R
( . )"t ) Vo( . ) and D
Q
( . )"s ) Vo( . ) are the surface tangential gradient operators. D
R
u* are
D
Q
u* are known if the surface displacement u* is prescribed. With this boundary condition, there
will be no error introduced in (10) when the third term on the right-hand side of (7) is dropped.
Now, consider mixed boundary conditions. For the case of f * and Du* given, we specify f"f *
and
n ) "Du*, s ) r`"0 and t ) r`"0 (14)
Alternatively, if u* and r* are prescribed, then we enforce u"u* and
n ) r`"r*, s ) "D
Q
u* and t ) "D
R
u* (15)
Stipulation that s ) "D
Q
u* and t ) "D
R
u* reduces the error in the approximation equation
(10), associated with dropping the third term on the right-hand side of (7).
An iterative solution procedure" is adopted to solve the system of linear equations for nodal
DOF involving zero diagonal terms.
Finite elements
In this section we introduce six types of isoparametric element for the strain gradient solid; they
are sketched in Figure 1. We restrict attention to two-dimensional deformation in the (x

, x
`
)
plane, and assume that subscript indices range from 1 to 2.
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 377
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 1. Sketch of six types of element
First, we explain the code used to label each type of element, for example TU24L12. Triangular
elements are designated by a leading letter , and quadrilateral elements by a leading letter Q.
The nodal degrees of freedom are of displacement-type (u or ), and the total number of degrees
of freedom n per element is written as n. Finally, the number m of Lagrange multipliers (p) is
designated Lm. Hence, element TU24L12 is a triangular element, with 24 nodal degrees of
freedom at 12 Lagrange multipliers j
GH
. The six types of element described below fall within three
classes:
(i) triangular elements where the number of DOF is dierent for corner nodes and for
mid-side nodes;
(ii) quadrilateral elements where the number of DOF is dierent for corner nodes and for
mid-side nodes;
(iii) triangular and quadrilateral elements which possess the same number of DOF for all
nodes.
378 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
(i) TU2412 and 244 elements
Both elements TU24L12 and TU24L4 are six-noded isoparametric triangular elements. The
location of each node in terms of the area co-ordinates for a triangle is standard. It has six degrees
of freedom, u

, u
`
,

,
`
,
`
and
``
at each corner node and two degrees of freedom, u

and u
`
at each mid-side node. The total number of degrees of freedom per element is 24. Dis-
placements u
G
are interpolated using standard quadratic shape functions in terms of the area
co-ordinates. Relaxed displacement gradients
GH
are interpolated linearly. The dierent interpo-
lation is motivated by the consideration that the true displacement gradients will depend linearly
on the co-ordinates if all sides of the element are straight.
In a TU24L12 element, four nodal values of Lagrange multiplier ( j

, j
`
, j
`
and j
``
) exist at
each quadrature point of the three-point Laursen and Gellert` integration rule. The Lagrange
multipliers are interpolated linearly between these three points. Since this triangular element has
12 Lagrange multipliers j
GH
and 24 DOF it is designated the code TU24L12. An alternative
element (TU24L4) has also been constructed with 4 constant Lagrange multipliers within the
element.
For both types of elements, the stiness matrix is calculated as follows. The deviatoric Cauchy
stress terms and the higher-order stress terms in the principle of virtual work are integrated by the
three-point integration scheme.` For a compressible solid, the pressure term is also integrated by
this scheme. For an incompressible solid, the pressure is treated as an additional Lagrange
multiplier, is assigned a constant value throughout the element` and is integrated by one-point
quadrature. The Lagrange multiplier j
GH
terms are integrated by full three-point integration.
It is known that for a nite element mesh, the total number of independent nodal DOF
(denoted as n

) should exceed the total number (denoted as n

H
) of all kinematic constraints
(including incompressibility) in order to generate numerically stable solutions.` Although the
element TU24L12 performs satisfactorily in some of the numerical tests conducted, it often
violates the numerical stability condition for a mesh with a large number of elements. On the
other hand, element TU24L4 generally satises the stability condition. Numerical experiments
have shown that TU24L4 used alone or in a proper combination with TU24L12 satises the
stability condition and can generate reasonably accurate results.
(ii) QU34L16 and QU34L4 elements
These are nine-noded isoparametric quadrilateral elements, as shown in Figure 1. The location
of each node in terms of the local co-ordinates (

,
`
) (!1)
G
)1) is standard. There are six
DOF, u

, u
`
,

,
`
,
`
and
``
at each corner node, and 2 DOF, u

and u
`
at each of the
remaining ve nodes. The total number of nodal DOF is 34. The two displacements are
interpolated by the conventional biquadratic Lagrangian shape functions and the relaxed
displacement gradients
GH
are interpolated bilinearly.
In a QU34L16 element, the four Lagrange multipliers j
GH
are interpolated bilinearly between
nodal values given at the four quadrature points of a 2;2 Gauss integration scheme. There are
a total of 16 nodal values of Lagrange multipliers j
GH
and therefore the element is denoted by
QU34L16. A slightly modied element (QU34L4) uses four Lagrange multipliers j
GH
, assumed
constant throughout the element.
The deviatoric stress terms, the pressure terms for a compressible solid and the higher-order
stress terms are all integrated using full integration of 3;3 Gauss quadrature. For the case of an
incompressible solid, pressure forms an additional Lagrange multiplier, and is interpolated
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 379
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
bilinearly between four nodal values at the Gauss points of a 2;2 quadrature scheme.` 2;2
Gauss quadrature is used to integrate all terms involving any Lagrange multiplier.
Meshes consisting solely of QU34L16 elements often violate the numerical stability condition,
for both compressible and incompressible solids. On the other hand, the QU34L4 elements satisfy
the stability condition and have been found to work well for both compressible and incompress-
ible solids. A combination of QU34L16 and QU34L4 also perform satisfactorily.
(iii) QU54L16 and TU36L12 elements
Element QU54L16 is identical to QU34L16 except that now there are six DOF,
u

, u
`
,

,
`
,
`
and
``
at every node and all DOF are interpolated using standard
biquadratic Lagrange shape functions. The element is denoted as QU54L16 because the total
number of DOF per element is 54. Regardless of the degree of compressibility, meshes of this
element satisfy the numerical stability condition. The same integration scheme is used as that
described for the element QU34L16. Though the element performs well in the numerical tests
conducted, it suers from the drawback that, for a mesh with a signicant number of such
elements, there are roughly twice as many nodal DOF as exist in a mesh with the same number of
QU34L4 elements.
A six-noded triangular element TU36L12 was also developed. It employs quadratic interpola-
tion for u, and linear interpolation for p. A mesh consisting solely of these elements marginally
satises the stability condition when the material is compressible but fails to satisfy the stability
condition for an incompressible solid. This element performed poorly during a 2Dtest calculation
(see section below), and although it is able to give adequate results in a 1D simulation, it suers
from the same drawback as that of QU54L16 as discussed in the previous paragraph. Hence the
use of this element alone is not recommended.
NUMERICAL TESTS
Three types of numerical test were performed: (1) patch tests for a 2-D mesh of elements under
uniform loading; (2) 1-D boundary layer analysis and (3) 2-D stress analysis for a plate contain-
ing a hole. In all cases, plane-strain deformations were assumed and body forces were absent. In
the presentation of results, nodal values of stress and strain were extrapolated from Gauss point
values by a smoothing scheme described in Appendix I.
Patch test
All types of element were subjected to patch tests as shown in Figure 2. In each case the mesh
was loaded in uniaxial tension by end stress-tractions, as indicated in the gure. The material was
discretized into a mesh of n;n identical quadrilateral elements or 2;n;n identical triangular
elements, with n ranging from 1 to 5. Triangular elements were obtained by sub-dividing each
square along a diagonal (dashed line).
The elements TU24L12 and QU34L16 passed all patch tests for any choice of constitutive
parameters, including the case of all length scales c
G
being identically zero.
The remaining elements QU34L4, TU24L4, QU54L16 and TU36L12 give the correct stress
and strain states when at least one of the material length scales c
G
is nite. For the limiting case of
all c
G
"0 (i.e. the material has degenerated to a conventional solid), these four elements have
380 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 2. Mesh and the boundary conditions for the patch test
Figure 3. (a) Notation and geometry of a bimaterial under uniform shear; (b) sketch of the mesh
zero-energy modes. This is because the number of nodal relaxed displacement gradients
GH
ex-
ceeds that of the linearly independent kinematic constraints and the vanishing length scales
render the derivatives of
GH
indeterminate. For the particular choice c

"0, with at least one of

the remaining c
G
taking on a nite value, the solution for the relaxed displacement gradient

GH
depends upon the value chosen for the accelerator parameter of iterative solution procedure
laid out in Reference 13.
We conclude that all six types of element pass the patch tests, if at least one c
G
O0.
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 381
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Boundary layer analysis
Higher-order gradient theories predict the existence of boundary layers adjacent to in-
homogeneities such as interfaces. Consider, for example, a bimaterial composed of two perfectly
bonded half planes of elastic strain gradient solids, subjected to a remote shear stress o`
`
as
shown in Figure 3(a). An analytical solution is presented in the Appendix II: for the strain
gradient solid specied by equations (1)(6) the shear strain c
`
has a continuous but non-
uniform distribution within a boundary layer adjacent to the interface. A similar solution was
obtained for a couple stress theory solid.` In a specic quantitative example, we shall make the
following arbitrary choice of constitutive parameters. The shear modulus j

of material 1 is taken
Figure 4. Accuracy study of various elements in the boundary layer analysis: (a) strain c
`
by QU34L16; (b) strain c
`
by
QU34L4; (c) relaxed strain c
`
by QU34L4
382 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 4. (Continued)
to be twice that of material 2. For both materials, the constants c

to c
"
(as dened in equation (6))
are equal to l` while c
`
"0. We shall examine the ability of the six types of element described
above to predict this boundary layer.
The nite element mesh, sketched in Figure 3(b), is xed at a width of l in the x

-direction, and
a length of 50l in the x
`
-direction on each side of the interface. A central region of length 10l on
each side of the interface is discretized into N/2 identical quadrilateral elements. The two
remaining end regions of the mesh are each divided into 10 quadrilateral elements. (Triangular
elements are obtained by sub-dividing the quadrilaterals along a diagonal.) The following
periodic boundary conditions are applied along the left- and right-hand sides of the mesh to
enforce uniformity of deformation in the x

-direction. For any pair of nodes P and P' which are

equally distanced from the interface and do not lie on the top or bottom surface, each nodal DOF
at P is equal to its counterpart at P' while the corresponding work-conjugate nodal force or
double force at P equals that at P' in magnitude but has an opposite sign. For any surface node,
the double forces vanish while the nodal forces are constructed from the prescribed surface
traction. For this one-dimensional problem, all types of element gave converged solutions. We
evaluate below the relative performance of the elements by examining the accuracy of solution as
a function of N.
The shear strain obtained using QU34L16 elements is plotted in Figure 4(a) after normaliz-
ation by the average shear strain of c
`
"o`
`
(j

#j
`
) /(2j

j
`
). The plot of the relaxed shear
strain, c
`
"(
`
#
`
) /2, is indistinguishable from that of the true shear strain and is not
explicitly shown. It is clear from the gure that the nite element solution converges quickly to
the exact solution with increasing renement of mesh, N. It is interesting to note that the strain
calculated at the interface is accurate even for a very coarse mesh (N"4). The element TU24L12
gives comparable performance to QU34L16 for this 1-D problem.
Next, consider the element QU34L4. The shear strain and the relaxed shear strain are plotted
in Figure 4(b) and 4(c), respectively. It can be seen that c
`
and c
`
converge to the exact solution
at a rate marginally slower than that demonstrated by element QU34L16. The accuracy is
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 383
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 5. Notation and geometry of an innite plane subjected to a remotely uniform tension
noticeably poorer only at the interface (x
`
"0). Similar performance tests were also carried out
on elements TU24L4, TU36L12 and QU54L16; they showed similar convergence properties to
those of QU34L4.
A combination of QU34L16 and QU34L4 elements was also used: the N elements adjacent to
the interface were chosen to be QU34L16 while the remaining were QU34L4 elements. The shear
strain distribution is indistinguishable from that shown in Figure 4(a).
Stress concentration due to a hole
In order to examine the convergence properties of the various nite elements for a 2-D
problem, we consider an innite solid containing a circular cylindrical hole. The solid is subjected
to remote uniform tension, as shown in Figure 5. An analytical solution was obtained' for this
problem for the special case of a couple stress solid, i.e. the strain energy density per unit volume
w depends upon strain and upon that part of strain gradients which can be expressed in terms of
curvature ,
GH
by
w"
`
zc
GG
c
HH
#jc
GH
c
GH
#2jl`,
GH
,
GH
(16)
Here, ,
GH
is related to the rotation of material elements 0 by ,
GH
"0
G H
"
`
e
GNI
u
I NH
. The problem is
re-analysed using the elements developed here. We note that the couple stress constitutive law is
recovered by setting the ve constitutive length scales in equation (6) to be c

"0, c
`
"3l`/4,
c
`
"5l`/8, c
"
"l`/8 and c
`
"!l`.
The symmetry of geometry and boundary conditions imply that a quarter mesh is adequate. All
computations were carried out using the mesh shown in Figure 6 which has 720 quadrilateral
elements and 2983 nodes. Triangular elements were generated by sub-dividing a quadrilateral
element along a diagonal. The boundary conditions prescribed are illustrated in Figure 6. Results
384 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 6. (a) Global view of a mesh used for the stress concentration problem as depicted in Figure 5. The mesh has 720
quadrilateral elements and 2983 nodes. Boundary conditions are indicated; (b) the close-up view of the ne mesh zone.
Boundary conditions on the hole surface are included
are presented in the form of the stress concentration factor at the edge of the hole; values as
a function of normalized hole radius a/l are given in Table I for Poissons ratio v"0 and in Table
II for Poissons ratio v"0)5.
The elements QU34L4 and QU54L16 both satisfy the stability condition (n

/n
H
'1) for both
compressible and incompressible solids, and give accurate solutions to the problem. Of the two,
a mesh of QU54L16 elements has roughly twice as many DOF as a mesh of QU34L4 elements,
leading to a signicantly higher computational cost for no signicant increase in accuracy. In
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 385
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Table I. Stress concentration factor (v"0)
QU34L4#
a/l Analytical* QU34L4 QU54L16 QU34L16 TU24L4
10` 3)000 3)006 3)010 3)009 3)003
10 2)878 2)894 2)888 2)888 2)902
8 2)824 2)843 2)834 2)834 2)849
6 2)729 2)750 2)733 2)738 2)758
4 2)545 2)568 2)549 2)552 2)577
3 2)389 2)412 2)397 2)396 2)422
2 2)169 2)188 2)176 2)176 2)201
1 1)889 1)897 1)893 1)895 1)912
*Analytical solution obtained in Reference 16
Table II. Stress concentration factor (v"0)5)
QU34L4#
a/l Analytical* QU34L4 QU54L16 QU34L16 TU24L4
10` 3)000 3)006 3)005 3)024 2)975
10 2)937 2)948 2)947 2)965 2)908
8 2)908 2)920 2)918 2)935 2)876
6 2)855 2)869 2)865 2)879 2)814
4 2)743 2)760 2)752 2)761 2)685
3 2)639 2)657 2)648 2)653 2)557
2 2)476 2)492 2)483 2)486 2)351
1 2)231 2)243 2)236 2)251 1)927
*Analytical solution obtained in Reference 16
addition, we found that for element QU34L4 (and similarly for TU24L4), contours of stress and
strain based on nodal values calculated by linear local smoothing are distorted around a central
part of the hole surface, but constant local smoothing (see Appendix I) damps out the zig-zag of
the contours, as shown in Figure 7(a) and 7(b). The contours of strain c
``
and the relaxed strain
c
``
are mostly indistinguishable.
Element TU24L4 is adequate for a compressible solid, but for an incompressible solid it
converges to an inaccurate distribution of pressure, particularly for l"a. The relatively poor
performance of TU24L4 in this case could be due to the assumed uniformity of pressure within an
element.
A mesh consisting only of elements QU34L16 or TU24L12 fails to converge for any v and nite
a/l, as the stability criterion (n

/n
H
'1) is violated. The element TU36L12 is a borderline case: for
the compressible solid (n

/n
H
+1)02) the solution converges slowly; for the incompressible case
(n

/n
H
+0)92) the solution does not converge.
We conclude that for a mesh made solely of one type of element, the optimal element is
QU34L4. This conclusion is moderated by the fact that elements can be used in combination to
give satisfactory results. For example, a combination of QU34L16 and QU34L4 was tried by
386 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 7. (a) Contours of normalized strain and relaxed strain around the hole with l/a"1. The material is incompress-
ible. Nodal values of the strain are calculated with bilinear local smoothing. The mesh consists of QU34L4 elements;
(b) Contours of normalized strain and relaxed strain around the hole with l/a"1. The material is incompressible. Nodal
values of the strain are calculated with constant local smoothing and the surface values are extrapolated from inner nodes.
The mesh consists of QU34L4 elements
taking the rst four layers of elements of the mesh along the hole surface (see Figure 6(b)) as
QU34L16 and the rest as QU34L4. This combination has n

/n
H
+1)24 for v"0)5 and 2)05 for
v(0)5. Tables I and II indicate that the combination is able to produce the stress concentration
factor with accuracy generally comparable to that by QU54L16. Echoing the nding in the
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 387
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
boundary layer analysis in the previous section, we nd for the hole problem that a small number
of QU34L16 elements around a free surface gives slightly more accurate results at the surface.
A word of caution: in practical applications, choosing the optimal small number of QU34L16
elements is not an easy task and is problem-dependent.
Our numerical tests indicate that for a 2-D problem, if a mesh fails to satisfy the numerical
stability condition (n
B
/n
H
'1), then the iterative equation solution procedure" also fails to reach
a converged solution. Furthermore, the closer n

/n
H
is to unity, the more dicult it is to obtain
a converged solution.
CONCLUDING REMARKS
A total of six types of triangular element and quadrilateral element have been developed for the
ToupinMindlin framework of strain gradient theory. Displacement gradients are introduced as
extra nodal degrees of freedom, and are termed the relaxed displacement gradients. The kinematic
constraints between the relaxed displacement gradients and true gradients of displacement are
enforced via Lagrange multipliers. For an incompressible solid, the pressure is treated as an
additional Lagrange multiplier. This mixed formulation allows all the nodal degrees of freedom
to be interpolated using standard C"-continuous shape functions. All six types of element passed
a patch test. Boundary value problems were solved using the various elements in order to
evaluate their relative performance. Elements QU54L16 and QU34L4 yield the most accurate
results with the latter one being computationally less expensive. TU24L4 is less accurate than
QU34L4 for an incompressible solid. Element TU36L12 is not recommended for practical
applications. Both QU34L16 and TU24L12 perform well in 1-D problems, but for general 2-D
problems they can only be used in combination with QU34L4 and TU24L4. We conclude by
recommending QU34L4 elements for practical applications.
APPENDIX I
Smoothing scheme
In this paper, nodal values of stress, strain and double stress were calculated from Gauss point
values by element-wise (local) smoothing, followed by averaging over all elements adjoining
a given node. Linear local smoothing` was used in all elements except for QU34L4 and TU24L4.
For these two elements, uctuations in stress and strain occurred within elements adjacent to
a free surface and were not damped-out by the linear smoothing. This can be attributed to the fact
that the kinematic constraint equation (8) is enforced only at the geometrical center of an element
and a free surface provides no additional constraints. For elements QU34L4 and TU24L4
smooth distributions of stress and strain were obtained by the following two-step scheme. First,
constant local smoothing (taking the average over all Gauss points within an element and
assigning the average onto all nodes of the element) was performed on the whole mesh, including
surface elements. The values at any interior node were obtained by averaging the contributions
from all elements sharing that node. Surface nodal values calculated by this step alone obviously
would lose some accuracy, and so a second stage of smoothing was performed on these surface
nodes. Linear extrapolation of values from two inner nodes was used to estimate surface nodal
values, as sketched in Figure 8. For example, consider nodes at one end of a strip mesh sketched
in Figure 8(a). A nodal value, denoted by P

, at node 1 was calculated from the values at inner

388 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
Figure 8. Illustrations for extrapolations from inner nodes onto surface nodes. The dashed lines represent a side of
a triangular element: (a) nodes at the ends of a strip mesh used for the boundary layer analysis; (b) nodes around a corner
of the mesh used for the hole problem.
nodes 3 and 9 by P

"2
*
P
"
!P
`
. Similarly, P
"
"2
*
P
"
!P
`
and P
`
"2
*
P
"
!P
'
. For nodes
around a corner of a mesh illustrated in Figure 8(b), the surface nodal values were obtained by
P

"2
*
P
"
!P
`
, P
`
"2
*
P
"
!P
'
, P
"
"2
*
P
`
!P
`
, P
`
"2
*
P
"
!P
`
and nally P
`
"2
*
P
'
!P
`
.
The scheme has been found to work well in the numerical tests.
APPENDIX II
A bimaterial under remote shear
An analytical solution is presented for a bimaterial, consisting of two perfectly bonded half
planes of dissimilar linear elastic strain gradient solids. The bimaterial is subjected to a remote
uniform shear stress o`
`
, as shown in Figure 3(a). Here, we assume that material 1 laying below
the interface has a shear modulus j

, and an internal length scale l

by taking
c

"c
`
"c
`
"c
"
"l

and c
`
"0. Material 2, laying above the interface, has a shear modulus j
`
,
and an internal length scale l
`
by taking c

"c
`
"c
`
"c
"
"l
`
and c
`
"0.
For this bimaterial system, conventional elasticity theory dictates that the shear stress is
uniform and the shear strain jumps in magnitude at the interface from c
`
"o`
`
/2j

in material
1 to c
`
"o`
`
/2j
`
in material 2.
By including strain gradient eects, a continuously distributed shear strain can be obtained. In
this problem, the only non-zero displacement, strain, stress and higher-order stress are
u

, c
`
, o
`
and t
``
, respectively, and they are functions of the co-ordinate x
`
only. From the
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 389
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
constitutive equation (6), it follows that
o
`
"2j
G
c
`
and t
``
"2jl`
G
p
``
"4jl`
G
cc
`
cx
`
(17)
in material i. Substitution of the above relations into the equilibrium equation (2) leads to
cc
`
cx
`
!lK `
G
c`c
`
cx`
`
"0 (18)
where lK
G
"(2l
G
. The general solution to the above ordinary dierential equation is
c
`
"d

#d
`
eV`
lK
#d
`
eV`
lK
for x
`
(0 (19)
and
c
`
"d
"
#d
`
eV`
lK
`#d
'
eV`
lK
` for x
`
'0 (20)
Here d

to d
'
are 6 constants yet to be determined. The general solution is subject to the following
boundary conditions:
(i) c
`
Po`
`
/2j

and x
`
P!Rand c
`
Po`
`
/2j
`
as x
`
PR
and at the interface;
(ii) continuity of traction: (o
`
!t
`` `
)
V`"

"(o
`
!t
`` `
)
V`"
>
;
(iii) continuity of double stress traction: t
``

V`"

"t
``

V`"
>
;
(iv) continuity of strain: c
`

V`"

"c
`

V`"
>
.
The particular solution satisfying all these conditions is
c
`
"
o`
`
2j

1#
j

!j
`
j
`
j
`
lK
`
j

lK

#j
`
lK
`
eV`
lK

for x
`
(0 (21)
and
c
`
"
o`
`
2j
`

1#
j
`
!j

lK

lK

#j
`
lK
`
eV`
lK
`

for x
`
'0 (22)
ACKNOWLEDGEMENTS
This work is supported by the U.S. Department of Energy and the Lawrence Livermore National
Laboratory under contract W-7405-Eng-48. The authors acknowledge many helpful comments
by a reviewer.
REFERENCES
1. E. Cosserat and F. Cosserat, heorie des Corps Deformables, Hermann et Fils, Paris, 1909.
2. R. A. Toupin, Elastic materials with couple stresses, Arch. Rational Mech. Anal., 11, 385414 (1962).
3. R. D. Mindlin, Micro-structure in linear elasticity, Arch. Rational Mech. Anal., 16, 5178 (1964).
4. E. C. Aifantis, On the microstructural origin of certain inelastic models, rans. ASME J. Engng. Mater. ech., 106,
326330 (1984).
5. N. A. Fleck and J. W. Hutchinson, A phenomenological theory for strain gradient eects in plasticity, J. Mech. Phys.
Solids, 41, 18251857 (1993).
6. N. A. Fleck and J. W. Hutchinson, Strain gradient plasticity, Adv. Appl. Mech., 33, 295361 (1997).
390 J. Y. SHU, W. E. KING AND N. A. FLECK
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)
7. O. C. Zienkiewicz and R. L. Taylor, he Finite Element, 4th edn., McGraw-Hill, New York, 1994.
8. B. Specht, Modied shape functions for the three node plate bending element passing the patch test, Int. J. Numer.
Meth. Engng., 26, 705715 (1988).
9. Z. C. Xia and J. W. Hutchinson, Crack tip elds in strain gradient plasticity, J. Mech. Phys. Solids, 44, 16211648
(1996).
10. J. Y. Shu and N. A. Fleck, Prediction of a size eect in micro indentation, Int. J. Solids Struct., 35, 13631383 (1998).
11. L. R. Herrmann, Mixed nite elements for couple-stress analysis, in S. N. Atluri, R. H. Gallagher and O. C.
Zienkiewicz (eds.), Hybrid and Mixed Finite Element Methods, Wiley, New York, 1983.
12. J. Y. Shu and N. A. Fleck, Strain gradient crystal plasticity: size-dependent deformation of a bicrystal, J. Mech. Phys.
Solids (1998), in press.
13. V. P. Smyshlyaev and N. A. Fleck, The role of strain gradients in the grain size eect for polycrystals, J. Mech. Phys.
Solids, 44, 465495 (1996).
14. O. C. Zienkiewicz, J. P. Vilotte and S. Toyoshima, Iterative method for constrained and mixed approximationan
inexpensive improvement of f.e.m. performance, Comput. Meth. Appl. Mech. Engng., 51, 329 (1985).
15. M. E. Laursen and M. Gellert, Some criteria for numerical integrated matrices and quadrature formulas for triangles,
Int. J. Numer. Meth. Engng., 12, 6776 (1978).
16. R. D. Mindlin, Inuence of couple-stresses on stress concentrations, Exp. Mech., 3, 17 (1963).
17. E. Hinton, F. C. Scott and R. E. Ricketts, Local least squares stress smoothing for parabolic isoparametric elements,
Int. J. Numer. Meth. Engng., 9, 235256 (1975).
FINITE ELEMENTS FOR MATERIALS WITH STRAIN GRADIENT EFFECTS 391
Copyright 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 44, 373391 (1999)