Advances in Engineering Software 67 (2014) 1–9

Contents lists available at ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

A local extrapolation method for finite elements

R. Durand ⇑, M.M. Farias
Department of Civil and Environmental Engineering, University of Brasilia, Campus Darcy Ribeiro, Brasilia 70910-900, Brazil

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

Article history: In most numerical analyses using the Finite Element Method, several quantities, such as stresses, strains,
Received 23 April 2013 fluid velocities and gradients, are computed at points in the interior of the solid elements, such as Gauss
Received in revised form 11 July 2013 integration points for instance. Nevertheless, in many applications it is necessary to extrapolate these val-
Accepted 15 July 2013
ues to nodal points. That is the case with most visualization tools and post-processors, also in programs
with auto-adaptive meshes, large deformations schemes such as Arbitrary Lagrangian–Eulerian Methods,
and in programs using the Dynamic Programming Method. A generic methodology to perform this
Keywords:
extrapolation in a precise and efficient way is proposed.
Finite elements
Extrapolation
Ó 2013 Elsevier Ltd. All rights reserved.
Smoothing
Stress recovery
Post-processing

1. Introduction domain [3]. Other application which requires nodal stress values
is related to the Dynamic Programming Method (DPM). DPM
Most boundary value problems in Engineering may be formu- may be used to search for the critical surface in slope stability
lated with the Finite Element Method using an approach similar problems, using the stress field produced (at the Gauss points) in
to the displacement formulation [1]. In these finite elements anal- a previous FE analyses. These values have to be extrapolated to a
yses, primary variables, such as displacements in equilibrium prob- grid of points used in a minimization procedure [4–6].
lems and water head in seepage problems, are computed at the Other numerical methods such as Arbitrary Lagrangian–Euleri-
nodes of the elements. However, secondary or derivate variables an FE (ALE), the Material Point Method (MPM) and its generaliza-
are computed at internal points rather than at the nodal points. tion (GIMP) require computation of derivate variables at nodal
Such variables may be, for instance, the components of stresses points. In ALE, convective terms such as spatial derivatives need
and strain, or the fluid velocities and gradients. to be defined as a continuous field over the domain. For instance,
The internal points at which the secondary variables are com- stresses are computed at the integration points, and the stress field
puted include the center of the element and/or the numerical inte- is generally discontinuous across element edges. In this case the
gration points. The most widely adopted integration points are stresses at the integration points are extrapolated using a least
those given by the Gauss–Legendre quadrature. These points are square approximation to obtain the nodal stresses [7]. The MPM
supposed to give the best sampling position for the secondary vari- uses a background calculation grid where any variable can be ex-
ables [2]. pressed as a piecewise continuous function by using standard fi-
In many applications, however, it is necessary to evaluate the nite element shape functions. In each time step positions,
values of secondary variables at the nodal points. That is the case velocities, stresses and strains are updated to define a new state
with most graphical finite element (FE) post-processors. They use in a deformed background grid. This background grid is then dis-
the nodal values to compute the variables in a grid of points inside carded and a new one is created for the next time step. At the start
the element, using the interpolation functions of that element. This of the next step, the velocities must be calculated by extrapolation
grid is later used to draw iso-lines or contour plots. Another impor- from the material points to the new grid. Since there are always
tant application is related to auto-adaptive meshing. The values more material points than grid vertices, this extrapolation is
computed at the Gauss points of a mesh have to be somehow achieved using a Least Square Method [8]. This extrapolation pro-
transported to the integration points of a new mesh. Nodal stress cedure in MPM requires inversion of a large mass matrix. For a
values are necessary to compute some error measure used to de- more efficient algorithm, a local extrapolation could be used.
fine the degree of refinement necessary in a certain region of the Mathematically, values at the nodal points could be evaluated
directly, i.e., in the same way as those computed at the Gauss
⇑ Corresponding author. points. However, points close the border of a finite element are
E-mail addresses: durand@unb.br (R. Durand), muniz@unb.br (M.M. Farias). the worst sampling points and might produce serious errors, since

0965-9978/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.advengsoft.2013.07.002

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
2 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9

equilibrium is satisfied only in an average sense when using weak nodes, an efficient scheme should be used, since the problem is
formulations (displacement approach). The existence of optimal underdetermined. This problem is illustrated in Fig. 1.
sampling points has been suggested and investigated numerically Therefore, the main objective of this paper is to present a local
and mathematically by many authors [9–11]. These points might extrapolation methodology, which might be generically applied to
coincide with the ones given by Gauss–Legendre quadrature in quantities computed at any number of internal points of finite ele-
some elements, but this is not always the case. If these optimal ments with different shapes. The proposed scheme shows accu-
points or other points with good overall accuracy are known, then racy, efficiency and it can be easily implemented in any
it is intuitive that some extrapolation from these values to the no- computer program.
dal points, plus some smoothing procedure such as averaging,
might give more accurate solutions than those obtained directly 2. Extrapolation methodology
from the displacement formulation. This idea produced several no-
dal stress recovery algorithms such as the original work proposed In some cases, as depicted in Fig. 1, there might be infinite
by Hinton and Campbell [12]. More details may be found in [13]. combinations of nodal values that produce, upon interpolation,
Stress recovery procedures may be applied in a local (element) the desired set of computed values at Gauss integration points.
area or globally, i.e., over the whole finite element mesh [14,15]. Therefore, the extrapolation of Gauss point values to nodal val-
Global methods are considerably more costly and most efforts ues may be underdetermined. Yet, it is possible to devise extrap-
have been concentrated in refining local procedures. Among olation schemes, which fit the sampling points and keep a
these, perhaps the most popular is the Superconvergent Patch smooth trend over the whole element domain. For the develop-
Recovery (SPR) method introduced by Zienkiewicz and Zhu ment of such a scheme the following requirements should be
[16,17]. In this technique, the smoothed stresses corresponding fulfilled:
to a fixed nodal point are obtained from a polynomial expansion
defined over a ‘‘patch of elements’’ sharing the node. This polyno- i. Extrapolation should be independent of element type or
mial is determined from a minimum square fitting using the cal- stress state, i.e., it should be applicable to any type of ele-
culated stresses at the surrounding superconvergent points. ment either in two or three dimensions (2D or 3D).
Usually, the superconvergent points are the same as the Gauss ii. The scheme should be valid for any number of nodes and
integration points and the polynomial expansion has the same any number of internal sampling points (i.e. number of inte-
complete order as the shape functions. Contrary to the original lo- gration points). The number of nodes could be greater, equal
cal methods focused in an element rather than a patch of ele- or less than sampling points.
ments, the SPR method requires the extra identification of the iii. The scheme should minimize the error when the extrapo-
elements and the corresponding superconvengent points sur- lated nodal point values are interpolated back to the sam-
rounding the particular nodal point for which the stresses are de- pling points.
sired. Moreover, a system of equations must be solved for each iv. The procedure should keep a smooth variation inside the
nodal point individually. Therefore, in the present paper, the element and not produce spurious nodal values.
authors propose a method that enhances and generalizes the local
recovery process over an element. The proposed method can be In the following let m denote the number of sampling (Gauss)
eventually modified to work over a patch of elements although points and n denote the number of nodal points. By the way, the
this extension is not presented here. sampling points may also lie along the sides of the elements. Con-
Local recovery procedures, such as proposed by Hinton and dition (ii) will be divided in two general cases: (1) m P n; or (2)
Campbell [12], work only for certain element types and particular m < n. These cases are treated separately in the next sections.
cases, as for instance, when the number of Gauss points is greater In the derivation below, it is used the concept of generalized in-
or equal to the number of nodal points. On the other hand, it is not verse matrix attributed to Moore and Penrose, see e.g. [18,19]. This
uncommon the use of reduced integration in many engineering concept indicates that for a given non-square matrix A, its general-
applications. In fact, eight node quadrilateral elements (Q8 Seren- ized inverse A+ can be calculated by the following equation, when
dipity type element), with 2  2 Gaussian integration (four points), the number of lines is greater than the number of columns:
are perhaps the most widely used in two-dimensional problems. 1
Also it is common to use brick elements of 20 nodes with Aþ ¼ ðAT AÞ AT ; ð1Þ
2  2  2 integration in three-dimensional (3D) finite element
and by the following expression when the number of columns is
applications. If someone wishes to extrapolate the stress compo-
greater than the number of rows:
nents computed at the four Gauss points of a Q8 element to its
1
Aþ ¼ AT ðAAT Þ : ð2Þ
In addition, the sample points values are denoted by wi and the
resultant nodal values are denoted by vi.

2.1. Case 1: number of sampling points greater than or equal to the

number of nodes (m P n)

For the sake of generality, let w denote the value of any quantity
evaluated at any point with natural coordinates (n, g, f) inside a fi-
nite element. This value can be approximated by a weighted aver-
age of nodal values vi, using the interpolation functions of the
element. This approximation is denoted by w  and is expressed in
the following equation:
X
n
 g; fÞ ¼
wðn; Ni ðn; g; fÞv i ð3Þ
i¼1
Fig. 1. Extrapolation of internal quantities to nodes in a Q8 element.

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9 3

! P
X @ ni¼1 Np1i vi
where n is the number of nodes, Ni is the shape function of the node n
@E p1
i evaluated at point (n, g, f), and vi are the nodal values, yet to be ¼  2 w1  Ni v i
@v k i¼1
@ v k
determined, for i = 1, . . . , n. ! P
Xn n p2
Therefore the values at the m sampling points may be expressed @ i¼1 i v i
N
 2 w2  Np2 v i
as a function of the unknown nodal values vi according the follow- i¼1
i
@v k
ing expression: ! P
Xn
@ ni¼1 Np3i vi ð9Þ
 2 w3  Np3
i vi
8 9 2 p1 38 9 @v k
> 1 >
w N1 Np1
2 ... Np1
n > v1 > i¼1
>
> >
> 6 p2 7>> > >
> 2 > 7>< > ..
Np2 Np2 7 v2
<w = 6 N1 ... = .
6 2 n

> .. > ¼ 6
6 .. .. .. 7
.. 7> .. > ð4Þ ! P
Xn
@ ni¼1 Npmi vi
>
> . >
> 4 . . . . 5> > . > >
>
: >
; >
: > ;  2 wm  pm
Ni v i ¼0
m
w Npm 1 Npm
2 . . . Nn pm vn i¼1
@ v k

in which w  1; . . . ; w
 m are the calculated values at the sampling Evaluating the derivatives in Eq. (3) and substituting into Eq. (9)
points, N ji is the value of the shape function of node i evaluated at and then rearranging, one gets:
the location of sampling point j, and v1, . . . , vn are the nodal values.
X
n X
n X
n X
n
Eq. (4) can be written using matrices (denoted by bold face letters) Np1
k Np1 p2
i v i þ Nk Np2 p3
i v i þ Nk Np3 pm
i v i þ    þ Nk Npm
i vi
in a condensed notation, as: i¼1 i¼1 i¼1 i¼1

 ¼ Nv
w ð5Þ ¼ w1 Np1 p2 p3 pm
k þ w2 N k þ w3 N k þ    þ wm N k ð10Þ

where w is a vector containing the values computed at the sampling Reorganizing properly and writing the n equations in (10) using
points, N is a (m  n) matrix with the values of the shape functions matrices, the following equation is obtained:
evaluated at sampling points and v is a vector containing the un-
known nodal values. NT Nv ¼ NT w ð11Þ
The values ðw  j Þ at the sampling points, computed according to
where matrix N was previously defined in Eq. (5).
the interpolation of Eq. (5), may in principle differ from the original
The term NTN in the left hand side of Eq. (11) produces an
input values (wj). Therefore an approximation error can be defined
(n  n) square matrix M, which may be inverted to finally obtain
as follows:
the desired vector v of nodal values according to:
X
m
E¼  j  wj Þ2
ðw ð6Þ v ¼ ðNT NÞ
1
NT w ð12Þ
j¼1
or simply:
In order to satisfy condition (iii), the error defined in Eq. (6) must
be minimized. This can be accomplished by using the Least v ¼ Nþ w ð13Þ
Square Method as suggested by [2]. Thus, taking the partial deriv- T 1 T
The term (N N) N in Eq. (12) is a (n  m) matrix, which can be
atives of the error function in Eq. (6) with respect to each com-
applied to any vector with the m given values of a quantity at the
ponent vi of the nodal value vector and forcing them to be
sampling points in order to produce the desired values at the n
zero, one gets:
nodes of the element. Note that the matrix N does not necessarily
Pm need to be a square matrix. If m = n, the problem has a unique solu-
@E @   wj Þ2
j¼1 ðwj
¼ ¼0 ð7Þ tion and the sampling point values are reproduced exactly when
@v i @v i evaluated from the interpolation of the computed nodal values.
Evaluating the derivatives of Eq. (5) and substituting into Eq. The error in Eq. (6) will be zero. In this case (m = n), matrix N itself
(7), the following system with n equations and n unknowns is invertible (NTN = I) and the solution could be obtained directly
(v1, . . . , vn) is obtained: from the inversion of Eq. (5). Accordingly, in this situation, the sys-
tem in Eq. (12) reduces to:
@E 1
@w @w2 !1
¼ 2ðw1  w  1Þ  2ðw2  w  2Þ
@v 1 @v 1 @v 1 v ¼ T
N |{z} N NT |{z}
w
|{z} |{z} |{z}
3
@w m
@w n1 nn nn nn n1
 3Þ
 2ðw3  w     2ðwm  w  mÞ ¼0 |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl}
@v 1 @v 1 nn
@E 1
@w @w2 !1 !
¼ 2ðw1  w  1Þ  2ðw2  w  2Þ
@v 2 @v 2 @v 2 ¼ T
N |{z} N T
N 1
N |{z}
N N1 |{z}
w ¼ |{z} w ð14Þ
|{z} |{z} |{z} |{z}
3
@w m
@w nn nn nn nn nn nn n1
 3Þ
 2ðw3  w     2ðwm  w  mÞ ¼0 ð8Þ |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} n1
@v 2 @v 2 nn
..
. If m > n, the system in Eq. (12) is over-determined. The solution
@E 1
@w @w2 obtained for the nodal values is just an approximation. However, it
¼ 2ðw1  w  1Þ  2ðw2  w  2Þ will ensure the best fit through the given values at the sampling
@v n @v n @v n
3 m points, minimizing the error of Eq. (6).
@w @w
 3Þ
 2ðw3  w     2ðwm  w  mÞ ¼0 Finally, if m < n the problem is undetermined, and a singular
@v n @v n
matrix is obtained when trying to compute the inverse of NTN. This
For each equation k (with k = 1, . . . , n) of system (8), substituting happens because the square matrix NTN will result in a higher or-
the values w  j at the samplings points (with j = 1, . . . , m), der (n  n), but is generated from a single smaller matrix N (m  n).
estimated according to Eq. (5), the following expression is Thus the additional (n  m) lines will be linear combinations of the
obtained: original ones, and the determinant will be zero. To solve this

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
4 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9

problem, additional constraints are necessary. These are described It should be emphasized here that the number of equations in
in the next section. system (18) is the same as the number of unknowns a1, . . . , am
and equal to the number of sampling points m. Substituting the
2.2. Case 2: number of sampling points less than the number of nodes values of the estimates at the sample points, w  1; . . . ; w
 m , given by
(m < n) Eq. (16), into the system of Eq. (18), each equation k of this system
takes the following form:
In this case, system (11) will be undetermined and a different ! ! !
@E X n X m
@ X n Xm
approach or additional equations are necessary. Here a new ap- ¼ 2 w1  pj
aj N 1 N i p1
 aj N 1 Np1
pj
i
proach is proposed. Eq. (4) is generally regarded as an operation @ak i¼1 j¼1
@ak i¼1 j¼1
that allows expressing the values at the sampling points w j ! ! !
X n X m
@ X n Xm
(j = 1, . . . , m) as a linear combination of the nodal values vi  2 w2  aj Npj
2 Np2
i  aj Npj
2 Np2i ...
(i = 1, . . . , n) in which the shape functions N ji are the coefficients i¼1 j¼1
@ak i¼1 j¼1
! ! !
of this linear combination. However, the same expression can be X n X m
@ X n Xm
pj pm
viewed as a linear combination of the shape functions values N ji ,  2 wm  aj Nm Ni  aj Nm Npm
pj
i ¼0
i¼1 j¼1
@ak i¼1 j¼1
in which the ‘‘coefficients’’ are the nodal values. Therefore, it is also
reasonable to suppose that the nodal values may be expressed as a ð19Þ
linear combination of the shape functions evaluated at the sam-
Proceeding with the derivatives and rearranging the terms, Eq.
pling points N ji by means of a series of coefficients to be deter-
(19) becomes:
mined. This hypothesis can be written mathematically as:
! !
8 9 8 9 X
n X
m X
n
> v1 > >
> a1 N11 þ a2 N21 þ    þ am Nm
1 >
> aj Npj Np1 Npk p1
>
> > > > i Ni
>v >
< > > >
< a1 N12 þ a2 N22 þ    þ am Nm
>
> 1 i
2= 2
= i¼1 j¼1 i¼1
¼ ð15Þ ! !
> .
. > .. X
n X
m X
n
> > > > >
>
> . > aj N pj p1
>
: > >
; >
>
. >
>
> þ 2 N i Npk p2
i Ni þ   
vn : 1 2 m ;
a1 Nn þ a2 Nn þ    þ am Nn i¼1 j¼1
! !
i¼1

X
n X
m X
n
in which aj (j = 1, . . . , m) are coefficients yet to be determined. If pj
aj N m Ni p1
Npk pm
þ i Ni
these coefficients are found, an estimate of the values at the sam- i¼1 j¼1 i¼1
pling points may be computed using Eq. (4), which now takes the X
n X
n X
n
following form: ¼ w1 Npk p1
i N i þ w2 N pk p2
i N i þ    þ wm N pk pm
i Ni ð20Þ
8 n ! 9 i¼1 i¼1 i¼1
> X X m
>
>
> aj N 1 N1i
j >
>
>
> >
> Reorganizing properly, the entire system of equations can be writ-
>
> >
>
8 9 >
> i¼1 j¼1
! >
> ten using matrices as follows:
> 1 >
w >
> >
>
>
> >
> >
> Xn Xm >
>
>
<w2 >= >
< aj N j2 N2i >
= ðNNT ÞðNNT Þa ¼ ðNNT Þw ð21Þ
. ¼ i¼1 j¼1 ð16Þ
>
> .. >
> >
> >
>
T
The product NN represents a (m  m) square matrix M. Pre-
>
> >
> > >
: ; > >
>
..
.
>
>
> multiplying Eq. (21) twice by M1, the vector of unknown coeffi-
m
w >
> ! >
>
>
> Xn X
m >
> cients is finally determined by:
>
> >
m>
>
> aj Nm Ni >
j
>
: ; 1
i¼1 j¼1 a ¼ ðNNT Þ w ð22Þ
The error in this estimate of the sampling point values is still gi- Therefore the vector of nodal values v can be calculated from Eq.
ven by Eq. (6). The unknown coefficients aj of Eq. (16) should be (15) as follows:
determined in such a way as to minimize that error. So, using
the Least Square Method, the derivatives of Eq. (6) with respect v ¼ NT a ð23Þ
to the unknown coefficients should vanish: which by means of Eq. (22) becomes:
Pm 2
@E @ j¼1 ðwj
 jÞ
w 1
¼ ¼0 ð17Þ v ¼ NT ðNNT Þ w ð24Þ
@ak @ak
or, according to the second case of generalized inverse, simply:
Computing the derivatives, the following system of equations is
obtained: v ¼ Nþ w ð25Þ
@E @v @v The only condition that is really assured by the solution given
¼ 2ðw1  w  1 Þ 1  2ðw
 2  w2 Þ 2   
@a1 @a1 @a1 by Eq. (24) is that the evaluation error at the sampling points,
@v m Eq. (6), when the computed nodal values are interpolated back to
 m  wm Þ
    2ðw ¼0
@a1 these points, yields the smallest possible value. However, when
@E @v @v the number of nodes exceeds the number of sampling points
¼ 2ðw  1  w1 Þ 1  2ðw 2  w2 Þ 2   
@a2 @a2 @a2 (n > m), the solution given by Eq. (24), despite assuring minimal er-
ð18Þ ror at the sampling points, might produce nodal results that do not
@v m
 m  wm Þ
    2ðw ¼0 satisfy the rigid body condition [1]. For instance, given constant
@a2
@E @v @v values at two Gauss points (w1 = w2 = w) of a one-dimensional
¼ 2ðw  1  w1 Þ 1  2ðw 2  w2 Þ 2    bar element with 3 nodes, the scheme of Eq. (24) may produce a
@am @am @am
set of nodal values (v1, v2, v3) all of them different from the
@v
 m  wm Þ m ¼ 0
    2ðw constant value w, but with a zero error at the Gauss points as
@am
illustrated in Fig. 2.
The trend set by the two internal points (p1 and p2) in Fig. 2 is
obviously a straight line (constant in this case) and one would ex-

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9 5

trated in Fig. 4. The translated values (wpj  w ~ pj ), represented by

open stars in Fig. 4, are then extrapolated to the nodes (open cir-
cles), using Eq. (24). Finally the computed translated nodal values
should be shifted back to the original reference by adding to each
node i (i = 1, . . . , n) a quantity v
~ i which corresponds to the esti-
mated value at a given node according to the trend given by the
best fit plane, as illustrated by the full circles in Fig. 3. The final
solution is illustrated in Fig. 5.
The proposed methodology assures a minimum fitting error and
at the same time preserves the general trend set by the known
sampling point values, without adding new equations or restraints.
In fact it is a means of imposing some condition in the curvature of
Fig. 2. Example not satisfying rigid body condition when extrapolating equal
the values that fit the sampling points, and satisfying condition (iv)
values. established at the beginning of Section 2. Therefore, when a con-
stant value is imposed at all sampling points, this same value is
extrapolated to the nodes thus satisfying the condition of rigid
pect to have the same constant values extrapolated to the nodes. body movement.
However, as the interpolation functions for an element with three Mathematically the proposed method can be written as:
nodes are quadratic, a higher order function of the nodal values
may reproduce the same values of the internal points with a totally v ¼ Nþ ðw  wÞ
~ þv
~ ð26Þ
different trend. in which the values with an over tilde, w ~ and v~ , represent vectors of
One possible way to enforce that the trend set by the internal a first estimate for the values at the sampling points and at the
sampling do not differ much from the curve given by the interpo- nodes, respectively, using a best fit plane through the given sam-
lation of the computed nodal values, in the case of having m < n, is pling point values.
to add more equations to the system of restrictions given by Eq. The best fit ‘‘plane’’ is actually a straight line in one-dimensional
(17). These equations could derive, for instance, from a measure (1D) problems given by a function P(n) as:
of the curvature of the fitting curve in the sampling points. How-
ever, that requires a few second derivatives and might complicate PðnÞ ¼ b1 n þ b2 ð27Þ
the minimization procedure. Therefore a simpler procedure that For 2D elements the mean plane is given by a function P(n, g):
tries to keep the internal trend, without adding new equations or
computing second derivatives, is proposed herein as explained as
follows.
The only situation in which Eq. (24) is guaranteed to satisfy the
rigid body condition and to give the same values in the nodes for
constant values in the sampling points is when these internal val-
ues are zero. This fact is used to enhance the extrapolation
procedure, by initially fitting a plane through the values at the
internal points. This is illustrated in Fig. 3 using n = 4 nodal points
and m = 3 sampling points (pale diamonds and stars along n axis),
in which the full star symbols indicate the known values (wj) at
sampling points j (j = 1, . . . , m). This plane will set a new reference
for the values which will be extrapolated. Therefore the imposed
values at the internal points wj are diminished by the values w ~j
evaluated at these points using the best fit plane (full diamond
symbols). The process can be interpreted as a translation, as illus-

Fig. 4. First nodal estimation in a translated space.

Fig. 3. Best fit plane according to internal points. Fig. 5. Final nodal estimation back to the original space.

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
6 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9

Pðn; gÞ ¼ b1 n þ b2 g þ b3 ð28Þ nodal values. Therefore it is possible to obtain a unique extrapola-

tion matrix E for each case. It is important to note that the extrap-
For 3D elements the best fit represents a hyper-plane given by:
olation matrix E depends only on the element type. A summary of
Pðn; g; fÞ ¼ b1 n þ b2 g þ b3 f þ b4 ð29Þ the proposed technique is presented in Table 1.
Generalizing, for d dimensions, there are d + 1 coefficients,
b1, . . . , bd+1, which define the best fit plane. These coefficients are
3. Numerical examples
obtained from a linear regression process and can be written in
matrix notation by a general expression as follows:
3.1. Example 1
b ¼ ^nþ w ð30Þ
As a first example, consider a two-dimensional quadrilateral
^ is a matrix containing the local coordinates of the sam-
in which n element with eight nodes (Q8) and 4 integration points, so that
pling points and will be fully presented later. n > m. The coordinates of nodes and sampling points are shown
Once the best fit plane is defined, using Eq. (30), the first esti- in Table 2.
mates w~ and v~ can be readily computed substituting the local coor- For this case it was chosen a constant set of values at all sam-
dinates of sampling points and the nodes into the respective pling points as shown in the last line of Table 2. In such a case
equation for the fitting plane: one would expect the same value to be computed at the nodal
~ ¼ ^nb and points. Since n > m, Eq. (12) cannot be used. If Eq. (24) is applied,
w v~ ¼ nb ð31Þ
only the minimum error requirement is satisfied. This is shown
in which n is a matrix containing the local coordinates of nodes. in Table 3, which gives the extrapolated nodal values, as well as
Matrices n and ^ n are given below for 1D, 2D and 3D cases, the re-interpolation error at the Gauss points. It can be seen that
respectively. this error is zero, but the nodal values are not constant. If the com-
2 3 2^ 3 puted nodal values are used to interpolate a surface over the whole
n1 1 n1 1 element domain, the picture in Fig. 6 is obtained. This picture does
6n 17 6 ^n 17
6 2 7 6 2 7
n¼6
6 .. .. 7
7;
^n ¼ 6
6 ..
7
.. 7 ð32Þ
4 . .5 4 . .5 Table 2
nn 1 ^nm Local coordinates of element Q8 for Example 1.
1
Node 1 2 3 4
2 2 3
n1 g1 1 3 ^n1 g^ 1 1 n 1.00 0.00 1.00 1.00
6n 6^ 7
6 2 g2 1 7
7 6 n2 g^ 2 1 7 g 1.00 1.00 1.00 0.00
n¼6 7 ^n ¼ 6
6 .
7
ð33Þ
6 .. .. .. 7; 6 . .. .. 7
7
5 6 7 8
4 . . .5 4 . . .5 n 1.00 0.00 1.00 1.00
^nm g 1.00 1.00 1.00 0.00
nn gn 1 g^ m 1
Sampling point p1 p2 p3 p4
2 3 2 3 n 0.577 0.577 0.577 0.577
n1 g1 f1 1 ^n1 g^ 1 ^f1 1 g 0.577 0.577 0.577 0.577
6n 6^ 7
6 2 g2 f2 1 7
7 6 n2 g^ 2 ^f2 1 7 Value 0.50 0.50 0.50 0.50
n¼6 7 ^n ¼ 6
6 .
7
ð34Þ
6 .. .. .. .. 7; 6 . .. .. .. 7
7
4 . . . .5 4 . . . .5
nn gn fn 1 ^nm gm
^ ^fm 1
Table 3
Substituting the expression in (31) into Eq. (26) the nodal val- Results for Example 1 using Eq. (24).

ues are calculated by: Node 1 2 3 4

Value 0.088 0.353 0.088 0.353
T T
v ¼ N ðNN Þðw  ^nbÞ þ nb ð35Þ
Node 5 6 7 8
Substituting Eq. (30) into Eq. (35) and after some arrangement Value 0.088 0.353 0.088 0.353

the proposed extrapolation method is finally expressed as: Sampling point p1 p2 p3 p4

Error 0.00 0.00 0.00 0.00
v ¼ ½Nþ ðI  ^n^nþ Þ þ n^nþ w ð36Þ
where I is a m  m identity matrix.
Note that the term NNT results in a square matrix (m  m). Thus,
if m > n then matrix NNT will be bigger than the original N (m  n)
and becomes singular, since the additional m  n lines will be lin-
ear combinations of the original ones, and the determinant will be 0.6
zero. In this case, Eq. (12) should be used.
Each case described above ends up with a matrix–vector prod- 0.4
uct in order to perform the extrapolation process and obtain the
0.2

Table 1 0
Summary of the proposed extrapolation technique. -0.088
-1
Case Formula Equation 0
m>n E = N+ Eq. (13) 1
1 -1 0
m=n E = N1 Eq. (14)
m<n E ¼ Nþ ðI  ^
n^
nþ Þ þ n^
nþ Eq. (36)
Fig. 6. Surface field obtained from extrapolated values in Example 1 using Eq. (24).

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9 7

not keep the trend set by the Gauss point values, despite reproduc- Table 6
ing them accurately. In order to get the proper solution, one could Results for Example 2.

apply the proposed Eq. (36). The correct result obtained in this case Node 1 2 3 4
is shown in Table (4) and the interpolated surface from the com- Value – Eq. (24) 1.732 0.866 0.000 0.866
puted nodal values is a plane as expected and depicted in Fig. 7. Value – Eq. (36) 0.577 1.155 0 1.155
Node 5 6 7 8
Value – Eq. (24) 1.732 0.866 0.000 0.866
3.2. Example 2 Value – Eq. (36) 0.577 1.155 0 1.155
Sampling point p1 p2 p3 p4
This example uses the same Q8 element as in the previous sec- Error – Eq. (24) 0.00 0.00 0.00 0.00
Error – Eq. (36) 0.00 0.00 0.00 0.00
tion. However a different set of values (not constant) is imposed to
the sampling points as shown in Table 5. These values were chosen
deliberately as fitting in a plane, i.e., they are co-planar in a 3D
space.
As n > m, again the solutions were obtained with Eqs. (24) and
(36). The results for both equations are shown in Table 6. It is seen
1
that both solutions satisfy the condition of error minimization.
However, when Eq. (24) is used the curvature trend is totally
wrong, especially close to the nodal points as shown in Fig. 8. In or-
der to get the correct extension of the plane from the Gauss points 0
to the nodes, Eq. (36) should be used as illustrated by the results in
Fig. 9.
-1
1
-1 0.5
3.3. Example 3 -0.5 0
0 0.5 -0.5
1 -1
In this example a cantilever beam, as depicted in Fig. 10, was Fig. 8. Surface generated by interpolation of extrapolated values in Example 2 using
analyzed using rectangular isoparametric plane stress Q8 ele- Eq. (24).
ments. The beam was subject to a uniformly distributed load along
its length at the top. The analysis was performed using a 2  2
Gauss reduced integration. This test is similar to the one presented
in [12,20] were the shear stress along the beam axis is evaluated

Table 4 1
Results for Example 1 using Eq. (36).

Node 1 2 3 4 0
Value 0.500 0.500 0.500 0.500
Node 5 6 7 8
Value 0.500 0.500 0.500 0.500
-1

Sampling point p1 p2 p3 p4 1
Error 0.000 0.000 0.000 0.000 -1 0.5
-0.5 0 0
0.5 -0.5
1 -1

Fig. 9. Surface generated by interpolation of extrapolated values in Example 2 using

Eq. (36).
2

Fig. 10. Cantilever beam discretization used in Example 4.

-1

0
1 for comparison between locally smoothed and unsmoothed nodal
1 -1 0
values.
Fig. 11 shows the distribution of shear stress. The unsmoothed
Fig. 7. Surface field obtained from extrapolated values in Example 1 using Eq. (36).
element stress was calculated at different points inside the ele-
ments using the nodal displacements, the displacement–strain ma-
trix (B) and the constitutive matrix (D). On the other hand, the
Table 5
Data for Example 2.
smoothed values were obtained using the proposed local extrapo-
lation. A good approximation between the smoothed values and
Sampling point p1 p2 p3 p4
the analytical ones given by conventional beam theory were found.
Error 1.00 0.00 1.00 0.00
Similar results were obtained using three-dimensional hexahedral

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
8 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9

60 tained by extrapolation. For reference, results are compared with

Unsmoothed shear stress the analytical solution provided by [21]. Since the analytical solu-
50 Beam theory tion is for a semi-infinite mass, a second reference curve is used
Proposed smoothing
40 which is obtained from a previous analysis with a very fine domain
discretization, i.e. (100  100).
30 Figs. 13–16 show stress results at integration points and nodal
points along the normalized vertical coordinate z compared with
Shear stress

20
analytical and reference values. Figs. 13 and 14 show the results
10 using triangular elements. Those triangular elements were ob-
tained dividing quadrilateral elements from the mesh in Fig. 12.
0 It can be seen that for a coarse mesh of constant stress T3 elements
the results at integration points (open circles) are not accurate.
-10
However the nodal values obtained by extrapolation and averaging
-20 (full squares) are not in much disagree with the reference curve. In
the other hand, for linear stress distribution T6 elements, the val-
-30 ues at integrations points are more accurate and the nodal values
0 5 10 15 20 25 30 35 40
obtained by extrapolation are in good agreement with the refer-
Distance
ence curve.
Fig. 11. Variation of shear stress s along the neutral axis of a cantilever beam. Figs. 15 and 16 show the results for quadrilateral elements. For
Q4 elements, the stresses along their local y axis remain constant,
so values at integration points are not accurate. However the nodal
elements with 20 nodes and a 2  2  2 quadrature. The obtained values obtained by extrapolation and averaging show more accu-
results are quite as good as the presented by [12], despite the sim- rate results. For the case of Q8 elements, the integration points val-
plicity in the application of the proposed method as well as its
generality.

3.4. Example 4

A surface strip load over a semi-infinite mass is analyzed in this

case. The purpose of this example is to determine the vertical
stress distribution along a vertical line below the load center and
compare the results with reference values. A coarse mesh illus-
trated in Fig. 12 represents a half of the semi-infinite mass. In
the analysis, triangular elements T3 and T6 with 1 and 3 points
quadrature respectively and quadrilateral elements Q4 and Q8
both with 2  2 quadrature are used. All element types except
Q4 fall in the case where the number of sampling points are bigger
than the number of nodal points.
After the analysis, integration points which are closer to the left
side of the domain were used to approximate stresses along the
vertical line below the load center. In a similar fashion, nodes along
the left side of the domain were used to extract stress values ob-

Fig. 13. Stress rz along depth using T3 elements and 1 point quadrature.

Fig. 12. Finite element mesh for the analysis of a infinite strip load. Fig. 14. Stress rz along depth using T6 elements and 3 points quadrature.

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002
 R. Durand, M.M. Farias / Advances in Engineering Software 67 (2014) 1–9 9

So, during a finite element analysis all extrapolation matrices for

all element types in the mesh may be computed only once and
stored to be used later for the whole mesh, saving computational
time.
From case studies, it can be concluded that finite elements that
develop at least a linear stress field can show accurate nodal values
obtained by extrapolation. Although the examples shown here
used Gaussian integration points as sampling points, the developed
formulation is applicable to any disposition of sampling points and
nodes. In this regard, it is desirable that sampling points contain
accurate values in order to get good extrapolated results at nodal
points. It should be noted also that integration points are not al-
ways the best points for sampling quantities as stresses inside
the element [9].

Acknowledgment

The authors acknowledge the financial support of the Brazilian

Fig. 15. Stress rz along depth using Q4 elements and 2  2 quadrature.
National Research Council (CNPq).

References

[1] Zienkiewicz OC, Taylor RL, Zhu JZ. The finite element method: its basis and
fundamentals. 6th ed. Butterworth-Heinemann; 2005.
[2] Hinton E, Owen DRJ. An introduction to finite element computations. Pineridge
Press; 1979.
[3] Zienkiewicz OC, Zhu JZ. A simple error estimator and adaptive procedure for
practical engineering analysis. Int J Numer Methods Eng 1987;24(2):337–57.
[4] Yamagami T, Ueta Y. Search for critical slip lines in finite element stress field
by dynamic programming. In: Proceedings of the international conference on
numerical methods in geomechanics. Innsbruck; 1988. p. 1347–52.
[5] Aguero MDS. Surface sliding slope determination using dynamic programing.
MSc Dissertation in Geotecnics. PhD thesis, University of Brasilia; 2004.
[6] Farias MM, Aguero MDS. Slope stability analysis of three-dimensional
problems using dynamic programming and enhanced limit equilibrium
methods. In: Proceedings of the 2nd international workshop in new frontiers
in computational geotechnics. Fortaleza: Gifu University/University of Brasilia;
2005. p. 127–34.
[7] Gadala M. Recent trends in ALE formulation and its applications in solid
mechanics. Comput Method Appl Mech 2004;193(39–41):4247–75.
[8] Tan H, Nairn JA. Hierarchical, adaptive, material point method for dynamic
energy release rate calculations. Comput Method Appl Mech 2002;191(19–
20):2123–37.
[9] Barlow J. Optimal stress locations in finite element models. Int J Numer
Fig. 16. Stress rz along depth using Q8 elements and 2  2 points quadrature. Methods Eng 1976;10(2):243–51.
[10] Zlámal M. Superconvergence and reduced integration in the finite element
method. Math Comput 1978;32(143):663–85.
[11] Levine N. Superconvergent recovery of the gradient from piecewise linear
ues and nodal extrapolated values show the same trend and are in
finite-element approximations. IMA J Numer Anal 1985;5(4):407–27.
good agreement with the reference curve. [12] Hinton E, Campbell JS. Local and global smoothing of discontinuous finite
element functions using a least squares method. Int J Numer Methods Eng
1974;8(3):461–80.
4. Concluding remarks [13] Křížek M, Neittaanmäki P. On superconvergence techniques. Acta Appl Math
1987;9(3):175–98.
The present work investigates different alternatives to extrapo- [14] Oden JT, Brauchli HJ. On the calculation of consistent stress distributions in
finite element approximations. Int J Numer Methods Eng 1971;3(3):317–25.
late known data from internal sampling points, such as Gauss inte- [15] Oden JT, Reddy JN. Note on an approximate method for computing consistent
gration points, to the nodes of a finite element. Two formulations conjugate stresses in elastic finite elements. Int J Numer Methods Eng
are presented, both satisfying the criterion of least square error 1973;6(1):55–61.
[16] Zienkiewic OC, Zhu JZ. The superconvergent patch recovery and a posteriori
when the nodal values are re-interpolated to the sampling points. error estimates. Part 1: The recovery technique. Int J Numer Methods Eng
These are represented by Eqs. (12) and (36). If the number of sam- 1992;33(7):1331–64.
pling points equals the number of nodes, then the solution is un- [17] Zienkiewicz OC, Zhu JZ. The superconvergent patch recovery and a posteriori
error estimates. Part 2: Error estimates and adaptivity. Int J Numer Methods
ique and both methods give the same result, which can be
Eng 1992;33(7):1365–82.
directly obtained by a simple inversion procedure. The summary [18] Lancaster P, Tismenetsky M. The theory of matrices: with applications. 2nd
of the proposed technique presented in Table 1 shows the formulas ed. Orlando: Academic Press; 1985.
for all cases which can be quickly and efficiently implemented in [19] Barnett S. Matrices: methods and applications. Oxford: Clarendon Press; 1990.
[20] Oñate E. Structural analysis with the finite element method linear
finite element programs. statics. Dordrecht: Springer; 2009.
The formulation for the proposed extrapolation matrix depends [21] Poulos HG, Davis EH. Elastic solutions for soil and rock mechanics. Wiley;
only on the element type (shape and sampling points locations). 1973.

Please cite this article in press as: Durand R, Farias MM. A local extrapolation method for finite elements. Adv Eng Softw (2013), http://dx.doi.org/10.1016/
j.advengsoft.2013.07.002