Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 95 (2014) 213 224

2nd International Conference on Sustainable Civil Engineering Structures and Construction

Materials 2014 (SCESCM 2014)

Effect of elemental shape and modeling of Mixed Hybrid FEM on

numerical solution
Akito Uedaa, Yuto Ishidaa, Shunji Kanieb, *
a
Division of Engineering and Policy for Sustainable Environment, Hokkaido University, Japan
b
Faculty of engineering, Hokkaido University, Japan

Abstract

Mixed Hybrid Finite element method, MHF, is known to be an efficient and powerful analyzing technique with satisfactory
accuracy when differentiated variables should be evaluated, such as flow velocity based on velocity potential theory or stress
calculated from displacement. On the other hand, it is known that the calculation accuracy and efficiency are affected by
elemental shape and modeling. The aim of this study is to confirm the efficiency and accuracy of MHF by comparing it with
FEM, and to discuss optimum elemental shapes and modeling. MHF, unlike FEM, evaluates the velocity potential and the flux on
the element boundaries not at the element nodes; consequently, the material or mass balance of each element due to flow-in and
flow-out through the boundaries can be strictly estimated. The flow velocity within an element, however, should be interpolated
by those values on the boundaries, and a distinctive shape function, known as the Raviart-Thomas shape function, becomes
necessary. The authors established the MHF formulation for arbitrary shaped elements with 3 and 4 boundaries, and numerically
investigated its efficiency and accuracy with various shapes of elements. Through this study, it was confirmed that MHF can
evaluate accurate flow velocity with a fewer number of elements than those of FEM. In addition, the effect of shape and
modeling of elements on the numerical results were discussed to obtain better accuracy even if the total number of the elements is
kept constant.
2014The
2014 TheAuthors.
Authors. Published
Published by Elsevier
by Elsevier Ltd.
Ltd. This is an open access article under the CC BY-NC-ND license
Peer-review under responsibility of organizing committee of the 2nd International Conference on Sustainable Civil Engineering
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Structures
Peer-review and
underConstruction
responsibility Materials 2014.
of organizing committee of the 2nd International Conference on Sustainable Civil Engineering
Structures and Construction Materials 2014
Keywords:FEM; Mixed Hybrid FEM; Raviart-Thomas shape function; Transfer matrix; modeling

* Corresponding author. Tel.: +81-80-5597-5101.

E-mail address:kanie@eng.hokudai.ac.jp

1877-7058 2014 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Peer-review under responsibility of organizing committee of the 2nd International Conference on Sustainable Civil Engineering
Structures and Construction Materials 2014
doi:10.1016/j.proeng.2014.12.181
214 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224

1. Introduction

The finite element method, FEM, has been widely used for numerical simulations. In structural mechanics, for
example, displacement is usually assumed at each node and the displacement within an element is interpolated by
shape functions. The strain and stress in the element is calculated based on this interpolation. Assuming that a node
is shared by two elements on their boundary, the displacement at the node is common for those two elements, and
the displacement is consequently evaluated continuously without any difference in the displacements at the node.
However, the strains at the node are not always continuous because they are calculated based on the shape functions
with their partial differentials of displacements; in other words, the strains as a gradient of displacement are
calculated from each element. In the case of fluid flow simulations the evaluation of flow velocity or flux, which is
estimated as a partial differencial of velocity potentials, becomes more important to satisfy the mass balance within
an element, compared to the evaluation of the velocity potentials themselves. When a usual FEM is adopted for such
a simulation, the flow velocity or the flux interpolated from the velocity potentials is not always continuous on the
boundary as well as the node. As a result, the mass balance in each element due to flow-in and flow-out through the
boundaries cannot be estimated appropriately with computational accuracy [1].
Mixed Hybrid FEM, on the other hand, has an advantage to keep the continuity in the flow velocity on
boundaries between elements, since the flow amount at each boundary is defined as an objective variable in the
formulation and the mass balance in each element is strictly satisfied. Consequently, MHF can compute the whole
flow in the model accurately by tracing the path of fluid [2].
In MHF, however, it is necessary to introduce different shape functions from the FEM since the flow velocity
within an element should be interpolated from the flow amount defined at each boundary, such as the Raviart-
Thomas shape function (RT; low-order) [3] and the Brezzi-Douglas-Marini shape function (BDM; high-order). It is
notable that those shape functions are expressed as vectors with dimensions concerning the edge lengths of the
boundary in order to interpolate the fluid flux from the amount of flow through the boundary [4,5].
Through this study, the effect of shape and modeling of elements on the numerical results are discussed to obtain
greater accuracy, even if the total number of the elements is kept constant. As a conclusion, it is confirmed that MHF
can evaluate accurate flow velocity with fewer number of elements than those of FEM.

2. Comparison of MHF with FEM

2.1. Non-degenerate form and Mixed form

The discretized formulation of a governing equation for MHF starts with its mixed form; FEM is, in contrast,
usually formulated based on the non-degenerate form. In the case of velocity potential theory for incompressible
fluid flow, for example, the non-degenerate form for FEM is given by the Laplace equation.

k 2) 0 (1)

where k: coefficient of permeability ,): velocity potential.

As Equation (1) implies, it is obvious that the exact solution for the velocity potential ) should be a 2-time
differentiable. In FEM, however, the objective function ) is usually assumed to be a 1-time differentiable as a linear
function due to the weak form formulation with partial integration after multiplication of a residual weighting
function. When the flow velocity is necessary, the velocity is computed as a constant within an element because of
the first differential of ), a linear function
On the contrary, in MHF, the governing equations are expressed in mixed form such as in Equation (2), which
relates the velocity potential and the flux based on Darcys law, and Equation (3) of mass-conservation law, shown
below.
&
k) q (2)

&
q 0 (3)
 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224 215

Table 1. Complete form and Incomplete form.

Complete Complete Incomplete

El.1 El.2 El.1 El.2 El.1 El.2

El.1 El.2 El.1 El.2 El.1 El.2

The two equations shown above are formulated independently for discretization. As a result, the objective
functions such as the velocity potential )and the flux q are also interpolated independently and estimated directly
within each element.

2.2. Complete and Incomplete forms of interpolation

The complete form of interpolation means that the differentiability of the objective function is consistently
satisfied for the evaluation of itself and the differentials. In the case of usual FEM, ) is assumed linear as a 1-time
differentiable function whereas the flux is calculated a constant as a first differential of ). Since there is no
contradiction in the relation of differentiability between the original ) and its differential q, the interpolation is
referred to as a complete form of interpolation.
In MHF, on the other hand, interpolation functions for the objective function ) and the flux q can be designated
independently without consistency in their differentiability. In our calculation, a discontinuous function is assigned
to the objective function ) whereas a linear function is adopted for the flux q. In other words, the differentiability
condition that the flux q should be a first differential of ) is not satisfied. Such an interpolation form is referred to
as a incomplete form of interpolation. Typical examples of complete and incomplete interpolations for a couple of 1-
dimensional elements are illustrated in Table 1.

2.3. Designation of location for representing values

In FEM, unknown variables of ) are calculated at each node of an element, as shown in Fig. 1 (a). The flux q at
arbitrary locations within an element, including the boundary, is computed with the interpolation functions for ). In
MHF, however, all of the unknown variables such as ) and q are defined on the boundary as shown in Fig. 1 (b). In
particular, the amount of flow Q through each boundary is evaluated instead of the flux q. Since the mass-
conservation low given by Equation (3) is satisfied by the total amount of flow-in and flow-out through the
boundaries of each element, the mass balance can be strictly preserved for each element. This is the strength of
MHF and its largest difference when compared to FEM. Similarly, unknown variables for velocity potential ) are
assigned to each boundary. However, the velocity potential within the element is assumed constant.
216 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224

a b

Fig. 1. (a) designation of location for unknown variables in FEM; (b) designation of location for unknown variables in MHF

y K
q

qy qK
qK
[

q[
x
qx

Fig. 2. Point of view for the transfer matrix

3. Shape functions for interpolation

3.1. Transfer Matrix for normalized coordinates

Before introducing shape functions for MHF, this chapter focuses on the transfer matrices of shape functions for
normalized coordinates shown in Fig. 2 since all the shape functions are defined based on normalized coordinates.
Assuming that fluxes q[ and qK are given as components of flux in the directions [ and K respectively, the
components of flux in the directions x and y, and q[ and qK, can be computed through the following equations.

wx wx
q x 1 w[ wK q[ 1 q

wy qK
>Tk @q[ (4)
q y wx wy wx wy wy
 det >Tk @ K
w[ wK wK w[ w[ wK

where [Tk] is referred to as a transfer matrix. 1/det[Tk] in Equation (4) stands for the rate of change in area.

3.2. Linear Shape Functions for FEM

Here we introduce linear shape functions commonly used in FEM for comparison. Fig. 3 indicates typical shapes
of a quadrilateral element and a triangle element. The objective variables are assigned at each node. The shape
functions for a quadrilateral element are expressed in Equation (5), the shape functions for a triangle element in
Equation (6). The flux within an element is calculated by the differentials of the objective variables.
 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224 217

a K b K

I 1 I I
1

-1 1

I [

I -1 I 0 I 1

Fig. 3. Normal elemental shape for FEM and number of objective functions: (a) quadrilateral; (b) triangle

1
4 1  [ 1  K 
1
1  [ 1  K 
I ^I1 I 2 I3

I 4 ` 4
1

^Ii `T >N quad @T (5)
1  [ 1  K 
4
1
 
4 1  [ 1  K 

1  [  K

I ^I1 I2 I3 ` [ ^Ii `T >N tri @T (6)
K

where [Nquad] is shape function of quadrilateral element and [Ntri] is the shape function of triangle element.

3.3. Raviart-Thomas Shape Function for MHF

As described in the previous chapter, unknown variables concerning the flux in MHF are assigned to the
boundaries of each element shown in Fig. 4. In this figure, the variables named Q are representing the quantity of
fluid flow through the boundary, and we introduce Raviart-Thomas shape functions to interpolate the flux within an
element based on the quantity of fluid flow on the boundaries. Raviart-Thomas shape functions for a quadrilateral
element are shown in Equation (7).

Q
1 1 1
q[ 0 [  1 0 [  1 Q2
1
4
1
4
>RT @^Q `
quad i (7)
qK K  1 0 K  1 0 Q3
4 4 Q
4

Where [RTquad] is the Raviart-Thomas shape function for quadrilateral element. In a similar way, the Raviart-
Thomas shape function for triangles is shown as Equation (8).
218 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224

a K b K

Q3 1
1

Q4 [
Q2
Q2
-1 1 Q3

-1 Q 0 Q1 1
1

Fig. 4. Normal elemental shape for MHF and number of objective functions: (a) quadrilateral; (b) triangle

Q
q[ [ [ [  1 1
Q2 >RTtri @^Qi ` (8)
qK K  1 K K Q
3

It is of interest that the shape functions have a dimension concerning the length of edge in order to convert the
flow quantity to the flux. This is one of the important characteristics of this shape function. In addition, those shape
functions estimate flux vectors within an element, whereas usual shape functions as Equation (5) for FEM
approximate a scalar value of ).

4. Global matrix formulation and mass balance evaluation

4.1. Global matrix formulation

The governing equations in mixed form are given by Equation (2) and Equation (3). First, we multiply a
weighting function given by the Raviart-Thomas shape functions to Equation (2) and derive a discretized
formulation. Equation (9) is obtained by applying Greens theorem with partial integration. As a result, Equation (2)
can be rewritten with three terms shown on the left hand side in Equation (9).
&
Sk
kI GWndSk  k)GWd: k  qGWd: k
:k :k
0 (9)

The first term of Equation (9) includes peripheral integration of the velocity potential and the weighting function
along the boundary. By introducing Raviart-Thomas shape functions into the term, it can be expressed as a
transcribed summation as shown in Equation (10).

I1

'y
^GWi `T k 1 >RT @T >Tk @T i I 2
4
1
Sk
kI GWndSk
2 i det >Tk @  'xi I3
(10)

I
4

where 'xi and 'yi are the lengths of the elemental boundary.
Next, we consider the second term of Equation (9) by calculating the area integration of the velocity potential and
 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224 219

first differential of the weighting function. Since we assumed the velocity potential ) is constant within an element
except its boundaries and the weighting function given by Raviart-Thomas function is linear, the second term is
expressed very simply as Equation (11).

:k
k)GWd: k ^GW `T k) k (11)

Where ) k is a velocity potential representing the value of the element.

Finally, the third term of Equation (9) is rewritten by introducing Raviart-Thomas shape functions with the
transfer matrix. This term can be written as Equation (12).

Q1


&
qGWd: k ^GW ` K [
T 1
>RT @ >Tk @ >Tk @>RT @d[dK Q2
T T
(12)
:k det >Tk @ Q3
Q
4

Substituting all of the terms in Equation (9) by Equation (10) to (12), we obtain the following equation.

I1 Q1

'y
1 4
k
1
>RT @T >Tk @T i I 2  k) k  K [ 1 >RT @T >Tk @T >Tk @>RT @d[dK Q2 0 (13)
2 i det >Tk @  'xi I3 det >Tk @ Q3
I Q
4 4

By eliminating the terms relating the variation of the weighting function, the discretized form of Equation (9) is
finally written through Equation (14).

tm11I1 tm21 tm311 tm312 tm313 tm314 Q1

tm12 I2 tm2 2
tm3 tm322 tm323 tm324 Q2
) k 
21
 0 (14)
tm13 I3 3
tm 2 tm3 31 tm332 tm333 tm334 Q3

tm14 I4 tm2 4
tm3
41 tm342 tm343 tm344 Q4

The equation above relates the velocity potentials on the boundaries and within an element to the quantities of
fluid flow through the boundaries. Since we introduced the mixed form for the governing equation, mass balance of
each element is evaluated separately by Equation (3). By multiplying a weighting function to Equation (3) with
application of Greens theorem, Equation (15) is derived.
& &
Sk
qGWndSk  qGWd: k
:k
0 (15)

&
where q is a vector of flow rate interpolated linearly.
The first term of Equation (15) gives the mass balance due to flows through the boundaries, and the second term
of Equation (15) stands for the change in density of fluids in an element. Since we assume the fluid is
incompressible, the second term must be always zero. Thus, we introduced a constant to the weighting function in
Equation (15). In other words, we adopted the idea of Finite Volume Method for the mass balance. As a result,
Equation (15) is simplified as Equation (16).
220 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224

&
Sk
qGWndSk 0 (16)

After application of peripheral integration along the boundaries, we can finally obtain Equation (17) as the
material balance equation.

Q1  Q2  Q3  Q4 0 (17)

Combining Equation (14) and Equation (17) together, the matrix expression of the mixed form of the governing
equation is given by Equation (18).

tm311 tm312 tm313 tm314 tm21

tm1 
tm11 tm11 tm11 tm11
1
tm321 tm322 tm323 tm324 tm2 2 Q1 I1

tm12 tm12 tm12 tm12 tm12 Q2 I2
tm3
tm332 tm333 tm334 tm2 3 Q3 I3 (18)
31

tm13 tm13 tm13 tm13 tm13 Q4 I
tm341 4
tm342 tm343 tm344 tm2 4 ) 0
 k
tm14 tm14 tm14 tm14 tm14
1 1 1 1 0

If the superposition principle is adopted, the global matrix can be constructed.

4.2. Difference in mass balance evaluation between FEM and MHF

In MHF, evaluation of mass balance in each element is done based on the quantities of fluid flow through the
boundaries. Accordingly, the mass balance is strictly managed in MHF. In FEM, however, flux on each boundary
should be estimated by applying shape functions to the velocity potentials calculated at each node. Equation (19) is
showing the relation between velocity potentials and fluxes computed at four nodes of the element.

wN 4 I1
1
wx wy wN 1 wN 2 wN 3

q x w[ w[ w[
w[ w[ w[ I 2
k wx wy

wN 3 wN 4 I 3
(19)
q y wN 1 wN 2

wK wK wK wK wK wK I 4

By integrating the flux vector in Equation (19) along the boundaries, the quantitative volume of fluid flow
through the boundaries can be obtained. Although the flux calculated by Equation (19) is continuous within an
element, it is not continuous on the boundary because the flux is interpolated from both elements connected by the
boundary.

5. Effect of elemental shape and modeling on the numerical accuracy

5.1. Effect of elemental shape

In order to evaluate the effect of the elemental shape on the accuracy in numerical simulation, we assume a
fundamental model for incompressible fluid flow. The shape of the calculation domain is a trapezoid shown in Fig. 5,
and the velocity potentials at the left and the right boundaries are set at 1 and 0 respectively as a Dirichlet condition.
 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224 221

The length of the left boundary is equal to 1 whereas the length of the right boundary is given as h. Thus, the fluid
flows in the domain from left to right. We varied h, the length of right boundary, to confirm the accuracy in quantity
of fluid flow. The fluid flux within the domain differs corresponding to h. When h is equal to 1, the shape of the
calculation domain is assumed to be rectangular and the flow velocity becomes constant at any location in the
domain.
The authors adopted a quadrilateral element and triangular elements for the calculation domain. Fig. 6 (a) to (d)
show the various calculation models for comparison. Model D consists of one quadrilateral element and three
models, coined Model A to Model C, consist of triangular elements.
Fig. 7 illustrates the relation between the shape of domain and the computation error. Model D with a
quadrilateral element estimates the quantity of fluid flow with high accuracy regardless of h. When triangular
elements are adopted such as Model A to Model C and the shape of domain is a rectangular with h = 1, estimated
results coincide with the exact solution. However, the numerical errors increase with the growth of h. Generally
speaking, an increase in the number of discretization of the calculation domain improves the numerical accuracy.
As Fig. 7 shows, Model B and Model C with 4 triangular elements give better results than Model A with 2 triangular
elements. Interesting is that Model B makes much better estimations than Model C for large values of h, although
the total number of elements are.
The authors consider that those numerical errors in triangular element models are caused by the characteristics in
interpolation for flux within an element. Even if the quantities of fluid flow on the boundaries along a triangular
element are evaluated as variables, the interpolated flux becomes constant at any location in the triangular element.
Therefore, the whole flow in the calculation domain is approximated as a combination of constant flows computed
in each triangular element. Discretization of the calculation domain for Model A unexpectedly produced better
approximations than Model C. However, it should be noted that the distribution of triangular elements within the
domain may deteriorate the accuracy in simulation results even if the total number of discretized elements is kept
constant.
I =1

I =0
1

Fig. 5. Calculation domain with boundary conditions

a b c d
h/3
h/3

h
h

h/3

Fig. 6. (a) Model A; (b) Model B; (c) Model C; (d) Model D

222 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224

Fig. 7. Relation between shape of domain and computation error

b
a =1

=0

Fig. 8. (a) Upper half of the domain; (b) Calculation model

5.2. Effect of elemental shape

We considered the effect of discretized modeling on the simulation results in a practical application model of
MHF. A rectangular domain is assumed with a rectangular-shaped tunnel located at the center of the domain. The
outside boundaries of the domain are pressured and velocity potentials of I  are given to the perimeter as a
Dirichlet condition. In contrast, no pressure is acting on the inside boundaries of the tunnel and velocity potentials of
I  can be assumed. Then, water permeates from the outside boundaries into the tunnel. Fig. 8 (a) illustrates the
upper half of the domain, and Fig. 8 (b) shows a calculation domain.
We discretized the calculation domain equally both in the water flow direction and in the transverse direction.
Fig. 9 demonstrates a discretized modeling with 64 elements. Applying quadrilateral elements, we evaluated the
total amount of water leaked into the tunnel. The estimated results are compared with those by the FEM using the
same discretized modeling.
Fig. 10 shows the relation between the numerical error in % and the total number of elements. It can be seen that
MHF computed better estimations with few errors and the calculation result converges stably at considerably fewer
elements than FEM. This best illustrates the advantage of MHF.
 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224 223

Fig. 9. Example of equally divided model (8 8)

Fig. 10. Difference from convergence under equally divided

Fig. 11. (a) Basic model; (b) Parallel to streamline model; (c) Perpendicular to streamline model

5.3. Effect of meshing technique

In the previous section, the advantages of MHF were confirmed through a comparison in the numerical accuracy
and in the convergence of the solution vis--vis those by FEM. In this section, the effect of meshing technique on
the numerical accuracy is discussed. Specifically, it is examined whether the numbers of discretization in the flow
direction and in the transverse direction may influence on the accuracy. Fig. 11 illustrates typical meshing models.
The basic model shown in Fig. 11 (a) is discretized equally in both directions. In the parallel to streamline model in
Fig. 11 (b), the calculation domain is discretized into small portions along the streamline whereas the domain is
divided into only 2 elements in the transverse direction. On the contrary, the perpendicular to streamline model in
Fig. 11 (c) is discretized into many portions along the transverse direction. Though the total number of discretization
is frequently discussed, we believed that the effect of meshing technique on the numerical accuracy must be verified.
224 Akito Ueda et al. / Procedia Engineering 95 (2014) 213 224

Fig. 12. Different from divergence by dividing way in MHF

The relations between the numerical error in % and the total number of elements are compared in Fig. 12. The
error itself is quite small as less than 4 % at most for any model. If the calculation domain is discretized equally in
both directions, stable convergence can be expected with the increase in the number of elements. However, the
parallel to streamline model always gives a better estimation than the perpendicular to streamline model regardless
of the number of elements. It is an important finding because fine meshing along the streamline raises the accuracy
of evaluation if the streamline can be forecasted prior to calculation.

6. Conclusion

In this paper, the authors provided a formulation procedure for Mixed Hybrid FEM as well as the characteristics
of MHF. As a result, discretized matrix equations are successfully derived for both quadrilateral and triangular
elements. After comparisons in numerical simulations, it was verified that a quadrilateral element has an advantage
over a triangular element for good estimation with small errors. In addition, the distribution of elements in a
calculation domain is an important factor when triangular elements are adopted.
By comparing the simulation results by MHF with those by FEM, it was confirmed that MHF has a great
advantage in the evaluation of quantity of flow with a lower number of elements. For improvement in the
calculation efficiency concerning the load computation, discretization technique may be one of the factors if the
direction of the streamline can be forecasted in advance of the simulation.

References

[1] Durlofsky, L. J. (), Accuracy f mixed and control volume finite element approximations to Darcy velocity and related quantiti es, Water
Resour. Res., 1994, 30,965-973, doi:10.1029/94WR00061
[2] Cordes, C., and W. Kinzelbach, Continuous groundwater velocity field and path lines in linear, bilinear, and trilinear finite elements, Water
Resour. Res., 1992, 28 (11), 2903-2911,doi:10.1029/92WR01686
[3] Raviart, P. A., and J. M. Thomas (A mixed finite element method for 2nd order elliptic problemsin MathmaticaAspects of the Finite
Element Method, Lect. Notes in Math, Springer, New York, 1989, vol. 606, pp. 292-315,
[4] Matringe, S. F., R. Juanes, and H. A. Tchelepi, Robust streamline tracing for the simulation of porous media flow on general triangular and
quadrilateral grids, J. Comput. Phys., 2006, 219(2), 992-1012, doi:10. 1016/j.jcp.206.07.004,
[5] Anis Younes, Philippe Ackerer, and Frederick Delay, Mixed Finite Elements for Solving 2D Diffusion-Type Equations, Rev. Geophys.,
2010, 48, Rg1004, doi:10. 1029/2008RG000277