Инженерный вестник Дона, №4 (2019)

ivdon.ru/ru/magazine/archive/n4y2019/5940

Fast Matlab-based Finite Element Solver for the Problem of Bending of

Mindlin Plates

Alexander P. Suvorov, Nelly N. Rogacheva

Moscow State Civil Engineering University, Moscow

Abstract: In this paper we describe efficient Matlab-based implementation of the finite element
solver for the problem of bending of Mindlin plates. This solver is tested on fine meshes
consisting of a large number of linear rectangular finite elements. The performance of this solver
is so good that even for fine meshes considered the CPU time is small enough.
Keywords: simply-supported plate, Mindlin plate theory, sparse solver, Matlab, finite element
method

Introduction
Classical theory of thick plates and plates of moderate thickness is presented
in references [1–4]. Various extensions of the classical theory can be found in
references [5–10]. Recall that in the theory of thick plates (Mindlin plate theory)
the transverse normal can rotate in such a way that after rotation it does not remain
perpendicular to the mid-surface of the plate and thus the shear deformations are
not zero. In contrast, in the theory of thin plates (Kirchhoff plate theory) the
vertical element of the plate always remains perpendicular to the mid-surface of
the plate and therefore the shear deformations are zero. For better accuracy of the
solution for moderately thick and thick plates it is recommended to use Mindlin
plate theory.
In [11] we analyzed bending moments in simply-supported plate and we
showed that for this type of plate two kinds of boundary conditions are possible in
case one uses the Mindlin plate theory. These are SS-A and SS-B types of
boundary conditions and they are different in how the angles of rotations φ x , φ y are
prescribed in the planes of edges of the plate. In particular, for SS-A plate zero
rotations are prescribed but for SS-B plate zero rotations are not prescribed but the
twisting moments in the planes of edges of the plate are equal to zero.

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

The problem of bending of Mindlin plate is solved using the finite element
method and for that the plate is subdivided into linear rectangular finite elements.
The square plate is considered and the number of elements along the horizontal and
vertical axis is taken the same. The plate is subjected to the action of uniformly
distributed vertical load (pressure). To obtain the solution of the present problem,
the Matlab code by Ferreira [2] was taken as a basis but this code turned out to be
slow enough for fine meshes. The objective of this paper is to propose some
improvements in this code so that it remains efficient and fast even for fine
meshes.

Problem Description and Results for Shear Force Distribution

As in [11] we consider a simply-supported plate that is acted upon by the
uniform vertical pressure q . The plate of square form lies in the x − y plane and its
side length is equal to a . The thickness of the plate is h (Fig. 1).
We use the following notation. Let u , v and w denote the displacements of
the mid-surface of the plate along the x , y and z axes respectively. Therefore, w
is the transverse displacement or deflection. Angles of rotations of the normal to
the mid-surface around the y and x axes are denoted by φ x and φ y . We also denote

the bending moment around the axes y and x by M x and M y respectively, and the

twisting moment is denoted by M xy . In addition, Qx signifies the shear force acting

in the cross-section x = const , and Q y is the shear force acting in the cross-section

y = const .

A simply-supported plate can be defined as a plate in which the transverse

displacement w is zero and also the corresponding bending moments at the edges
of the plate are equal to zero, i.e.,
w = 0, x = 0, a, y = 0, a
M x = 0, x = 0, a (1)
M y = 0, y = 0, a

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

Fig. 1. – Geometry of square plate

As in [11] two types of simply-supported boundary conditions are
considered. For the first type of boundary conditions, denoted here by SS-A, the
zero angles of rotations are prescribed on the edges of the plate:
φ y = 0, x = 0, a,
(2)
φ x = 0, y = 0, a.

For the second type of boundary conditions, denoted by SS-B, the zero
twisting moment is prescribed on the edges of the plate:
M xy = 0, x = 0, a, y = 0, a (3)

Boundary conditions of SS-A type are much more common, while the
analysis of the plate with SS-B boundary conditions is presented rarely and in [11]
we showed some important differences in the solutions for the problem of bending
of plates with SS-A and SS-B types of boundary conditions. Note that in the
terminology of Wang et al. [1] the plates with SS-A и SS-B types of boundary
conditions are denoted by SS and SS* respectively.
The solution of this problem is obtained with the help of finite element
method and is based on the theory of plates of moderate thickness or Mindlin plate
theory. The square plate is discretized into linear rectangular elements (4-node

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

elements) and the number of elements along the x and y axes is the same (for
example, 40 by 40 elements). As was already mentioned, the finite element
program written in Matlab and presented in the book by Ferreira [2] was taken as a
basis for obtaining the solution. To be able to use this program for fine meshes, say
200 by 200 elements, some important modifications must be introduced into the
program and these improvements will be described below.
Before describing these improvements of the code let us illustrate the
importance of solving the present problem on fine meshes. As an example consider
a square simply-supported plate with a side of length a = 1 . The magnitude of the
transverse pressure is q = 1 . The Young’s modulus is E = 10920 and Poisson’s ratio
is ν = 0.3 . The thickness of the plate is h = 0.1 , and therefore, this plate is a plate of
moderate thickness with h / a = 1 / 10 . This data was used in the book by Ferreira [2].
In Figs. 2–3 we show how the shear force values for this plate change with the
increase in the number of elements used in discretization of the domain of the
plate.
The shear force Q y at the center of the side of the plate y = 0 evaluated for

finite elements meshes with various number of elements along the x or y axis is
shown in Fig. 2. The shear force Q y is shown for the plates with boundary
conditions of SS-A and SS-B types. We see that the convergence of the shear force
Q y for the SS-B plate is slower than for the SS-A plate. Indeed, a reasonably good

accuracy in the value of the shear force can be achieved with 100 elements for the
SS-A plate but for the SS-B plate we need at least 200–250 elements along each
side of the plate. Having the mesh with 100 elements we can get only a rough
estimate of the shear force Q y for the SS-B plate. Fig. 2 also shows that by using

the meshes with small number elements, say 20 by 20 elements, it is not possible
to achieve a good accuracy in distribution of shear force even for the plate with
boundary conditions of SS-A type.

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

Fig. 2. – Dependence of the shear force Q y at the center of the side y = 0 of

the square plate, subjected to uniform transverse pressure of magnitude 1, on the

number of elements

Fig. 3 shows the shear force Q y at the corner of the plate y = 0 , x = 0 . Due to

the symmetry Q y = Qx at the corner. For the SS-A plate the shear force is equal to

zero, but for the SS-B plate this force is not zero. We see here that situation with
the convergence of the shear force results is slightly better compared to the shear
force at the center of the side. Indeed, with 150–200 elements along each side of
the plate we can obtain a reasonably good approximation to the value of the shear
force at the corner.

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

Fig. 3. – Dependence of the shear force Q y = Qx at the corner y = 0 , x = 0 of

the square plate, subjected to uniform transverse pressure of magnitude 1, on the

number of elements

All these results suggest that having the mesh with large number of elements
is strongly desired in the analysis of the simply-supported plate, especially in case
of the plate with SS-B type of boundary conditions.

Implementation of Efficient FE solver in Matlab

For solving the problem described above the Matlab code presented in the
book by Ferreira [2] is used. In this code the entire stiffness matrix with all zero
terms is stored in computer memory and then the Gaussian elimination procedure
is used to solve the resulting system of linear equations. If one uses the mesh with
N elx = N ely elements along each side of the square plate and each element is a

rectangle with 4 nodes, then the total number of nodes in the mesh is N n = (N elx + 1)2 .
Taking into account the fact that each node has 3 degrees of freedom (one for

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

transverse displacement and two for rotations), the total number degrees of
freedom for this problem is N dof = 3(N elx + 1)2 . Thus, the global stiffness matrix has

N t = 9( N elx + 1) terms (numbers) which must be stored in computer memory. If, for
4

example, N elx = 100 , then N t ≈ 900000000 or 900 mln. terms and if N elx = 50 , then
N t ≈ 56250000 or 56 mln. terms. It appears that the memory resources of the

average home-based workstation with 8 GB of RAM may be insufficient to store

all this data. In fact, we have established that the maximum number of elements in
1D array that the aforementioned computer may store is about 200 mln. elements
given that 8 bytes is required for storing each element of the array.
Table 1 gives the amount of CPU time that the computer spends on solution
of the present problem for different sizes of the mesh. Original code of Ferreira [2]
was used for solving the problem. It was not possible to solve the present problem
on the mesh of 80x80 elements because the memory limits were exceeded.

Table 1
CPU time required for solving the problem with the help of original code
Mesh size Time [sec]
20x20 elements 0.19
30x30 elements 0.94
50x50 elements 13.7
70x70 elements 95.8
80x80 elements out of memory

To be able to run problems with large number of elements, we should take

advantage of Matlab capabilities of working with sparse matrices. The most
straightforward implementation is to declare the global stiffness matrix as sparse
K=sparse(GDof,GDof);
where the variable Gdof is the total number of degrees of freedom in the mesh. By
doing so, only nonzero terms of the global stiffness matrix are stored in computer

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

memory and the problem with memory shortage will be solved. However, after
running the modified code, we observe that the computation speed is low for
problems on fine meshes. Surprisingly, most of the CPU time is now spent not on
the solution of the system of equations (with the help of Matlab operator
“backslash”) but on assembling the global stiffness matrix. Assembling the global
stiffness matrix in the code by Ferreira [2] is divided into two stages: first the
bending part of the stiffness matrix for all elements is assembled, and after that the
shear part of the stiffness matrix for all elements is assembled:
loop over all elements
K_b = create bending part of stiffness matrix
K(elementDof,elementDof)=...
K(elementDof,elementDof)+K_b;
end loop
loop over all elements
K_s = create shear part of stiffness matrix
K(elementDof,elementDof)=...
K(elementDof,elementDof)+K_s;
end loop

Here elementDof is an array with row and column numbers that

correspond to location within the global stiffness matrix into which the local
stiffness matrix of the element must be placed. In other words, elementDof
stores all degrees of freedom of the element using global numbering.
It should be noted that the bending part of the stiffness matrix K_b has 64
nonzero terms while the shear part of the stiffness matrix K_s has 104 nonzero
terms out of total 124 terms (since there are 12 degrees of freedom for each
element). Therefore, K_s is less sparse and its assembly takes noticeably more
time that the assembly of K_b. It is also important to note that assembly of the
stiffness matrices for the first, say, 1000 elements takes considerably less time than
the assembly of the matrices for the last 1000 elements in the loop. This is true
because the structure of the global stiffness matrix gets more and more complex as

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

we sequentially go through all elements in the loop and add more and more
nonzero terms to the global stiffness matrix.
In view of the reasons mentioned above, we can somewhat improve the code
performance (in terms of speed) by executing first the loop where the shear part of
the stiffness matrix is assembled, and then placing the loop in which the bending
part of the stiffness matrix is assembled, i.e. by reversing the original order:
loop over all elements
K_s = create shear part of stiffness matrix
K(elementDof,elementDof)=...
K(elementDof,elementDof)+K_s;
end loop
loop over all elements
K_b = create bending part of stiffness matrix
K(elementDof,elementDof)=...
K(elementDof,elementDof)+K_b;
end loop

This new code will be faster since here we first assemble the non-sparse
local stiffness matrix K_s into a relatively empty global stiffness matrix and after
that the sparse matrix K_b is placed into the relatively populated global stiffness
matrix. Thus, we avoid the situation in which the non-sparse matrix is placed into
the populated global stiffness matrix, which we had in the first code.
It is also worth mentioning that combining two loops over all elements into
one loop has led to a slight slow-down of the code:
loop over all elements
K_s = create shear part of stiffness matrix
K(elementDof,elementDof)=...
K(elementDof,elementDof)+K_s;
K_b = create bending part of stiffness matrix
K(elementDof,elementDof)=...
K(elementDof,elementDof)+K_b;
end loop

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

Resulting CPU times (after this improvement) are shown in Table 2 for
different meshes (again we remark that almost all CPU time is spent on the
assembly of the global stiffness matrix).
Table 2
CPU time required for solving the problem
with the help of the code after the first improvement
Mesh size Time [sec]
50x50 elements 7.9
70x70 elements 29
100x100 elements 117
150x150 elements 707

This is of course an improvement of the code performance compared to the

original version but it is still not satisfying as we need a further speed-up.
А considerable improvement of the code can be achieved if we create the
sparse matrix by using triplets i, j, v:
K = sparse(i,j,v);
where i is an array that contains the row numbers of all nonzero elements of the
matrix K, j is an array that contains the column numbers of the nonzero elements
of the matrix in the same order, and v is an array that stores the nonzero elements
of the matrix in the same order.
At the beginning, the arrays similar to i and j can be formed locally, for
each element, using the array elementDof, which stores all degrees of freedom
of the element. Below we describe how it can be done in Matlab.
Using the function meshgrid we can obtain two-dimensional arrays in
which degrees of freedom for the element are repeated row-wise or column-wise
[X,Y]=meshgrid(elementDof,elementDof);
Note that here in the array X the degrees of freedom are changing along the
row, while in the array Y the degrees of freedom are varying along the column.

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

After that we transform the two-dimensional arrays X and Y into one-

dimensional arrays X1 and Y1 by using the function reshape
X1=reshape(X,1,12^2);
Y1=reshape(Y,1,12^2);
and here we use the fact that there are 12 degrees of freedom for each element. The
function reshape forms 1-D array column-wise (first, it copies elements from
the first column of a two-dimensional array, then from the second column, etc.).
Thus, the array X1 will contain dof1 (degree of freedom number 1) repeated 12
times, then dof2 repeated 12 times, etc. On the hand, the array Y1 will contain the
sequence dof1, dof2, …, dof12 repeated 12 times.
Next, for each finite element we transform the element stiffness matrix into
1D array V1, also by using the function reshape
V1=reshape(K_s+K_b,1,12^2);
where K_s and K_b are the shear and bending parts of the element stiffness
matrix.
Obviously, we need to create similar arrays X1, Y1, V1 for each element.
Then, for each element we save its indices X1, Y1 inside the global arrays II and
JJ in proper places and save the components of the stiffness matrix of each
element inside the global array VV:
II((e-1)*12*12+1:e*12*12)=X1;
JJ((e-1)*12*12+1:e*12*12)=Y1;
VV((e-1)*12*12+1:e*12*12)=V1;
where e is the element number. Obviously, the global arrays II, JJ will have
repeated terms since most degrees of freedoms in the mesh are shared by several
elements. But that is not a problem for assembling the global stiffness matrix since
the Matlab function sparse will accumulate values of VV (add them up) if the

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

indices inside the arrays II and JJ are repeated. Thus, we can conveniently use
the command
K=sparse(II,JJ,VV);
which will create the desired sparse global stiffness matrix K.
In the code that we used we did not actually form the array V1 but placed the
components of the stiffness matrix of the element directly inside the array VV
loop over all elements
K_s = create shear part of stiffness matrix
VV((e-1)*12*12+1:e*12*12)=...
VV((e-1)*12*12+1:e*12*12)+reshape(K_s,1,12^2);
end loop
loop over all elements
K_b = create bending part of stiffness matrix
VV((e-1)*12*12+1:e*12*12)=...
VV((e-1)*12*12+1:e*12*12)+reshape(K_b,1,12^2);
end loop

In Table 3 we present the resulting CPU times required for the solution of
the problem by using the efficient assembly of the sparse stiffness matrix described
above. Note that these times do not include the time spent on mesh generation,
which also grows as the mesh gets large.

Table 3
CPU time required for solving the problem
with the help of the code after the second improvement
Mesh size Time [sec]
100x100 elements 3.3
200x200 elements 14.3
300x300 elements 33.7
400x400 elements 63.3

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

We can see now that the time spent on the assembly of the global stiffness
matrix and solution of the system of equations is now considerably smaller and
even for very fine meshes the solution can be obtained within a reasonable amount
of time.

Conclusions
In this article we described an efficient implementation of the finite element
solver for the problem of bending of simply-supported plate. This implementation
is realized in Matlab software and it is based on the code written by Ferreira [2].
It was shown that for better accuracy of the solution of the present problem,
and in particular, shear force distribution, it is strongly advised to use fine meshes,
at least when simple linear 4-node finite elements are used in the mesh. But the
finite element code proposed by Ferreira [2] could hardly be used for the fine
meshes due to low speed of execution and limitations on computer memory. That
was a motivation for searching the ways of code improvement.
Major improvement of the code performance was achieved by using the
Matlab function sparse and its capability to create the sparse matrix from one-
dimensional arrays with accumulation of values over the repeated indices. The
latter capability is very convenient for assembling the global stiffness matrix
during finite element computations.

References

1. Wang C.M., Reddy J.N., Lee K.H. Shear deformable plates and shells:
relationships with classical solutions. Elsevier Science Ltd., 2000. 296 p.
2. Ferreira A.J.M. Matlab Codes for Finite Element Analysis: Solids and
Structures, Solid Mechanics and Its Applications, 157. Springer Science+Business
Media B.V., 2009. 235 p.

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019

 Инженерный вестник Дона, №4 (2019)
ivdon.ru/ru/magazine/archive/n4y2019/5940

3. Khennane A. Introduction to finite element analysis using MATLAB and

Abaqus. Taylor & Francis Group, LLC, Boca Raton, 2013. 487 p.
4. Mindlin R.D. Journal of Applied Mechanics, 18, 1951. pp. 31–38.
5. Ma H.H., Gao X.-L., Reddy J.N. Acta Mechanica, 220, 2011. pp. 217–235.
6. Rajapakse R.N.K.D., Selvadurai A.P.S. International Journal for Numerical
Methods in Engineering, 23, 1986. pp. 1229–1244.
7. Soldatos K.P. A refined laminated plate and shell theory with applications.
Journal of Sound and Vibration, 144(1), 1991. pp. 109–129.
8. Kabir H.R.H. A double Fourier series approach to the solution of a
moderately thick simply supported plate with antisymmetric angle-ply laminations.
Computers & Structures, 43(4), 1992. pp. 769–774.
9. Kabir H.R.H. Analysis of a simply supported plate with symmetric angle-ply
laminations. Computers and Structures, 51, 1994. pp. 299–307.
10. Chaudhuri R.A. On boundary-discontinuous double Fourier series
solution to a system of completely coupled PDEs. International Journal of
Engineering Science, 27(9), 1989. pp. 1005–1022.
11. Suvorov A.P. Inženernyj vestnik Dona (Rus), 2019, №5. URL:
ivdon.ru/ru/magazine/archive/N5y2019/5892.

© Электронный научный журнал «Инженерный вестник Дона», 2007–2019