Module-I

Finite Element Methods

In the finite element method elements are grouped as 1D, 2D and 3D

elements. Beams and plates are grouped as structural elements. One
dimensional elements are the line segments which are used to model
bars and truss. Higher order elements like linear, quadratic and cubic
are also available. These elements are used when one of the dimension
is very large compared to other two. 2D and 3D elements will be
discussed in later chapters.

Seven basic steps in Finite Element Method

These seven steps include
  Modeling 
  Discretization 
  Stiffness Matrix 
  Assembly 
  Application of BC’s 
  Solution 
 Results 

Let’s consider a bar subjected to the forces as shown

First step is the modeling lets us model it as a stepped shaft consisting
of discrete number of elements each having a uniform cross section.
Say using three finite elements as shown. Average c/s area within each
region is evaluated and used to define elemental area with uniform
cross-section.

A1= A1’ + A2’ / 2 similarly A2 and A 3 are evaluated

Second step is the Discretization that includes both node and element
numbering, in this model every element connects two nodes, so to
distinguish between node numbering and element numbering elements
numbers are encircled as shown.
Above system can also be represented as a line segment as
shown below.

Here in 1D every node is allowed to move only in one direction, hence

each node as one degree of freedom. In the present case the model as
four nodes it means four dof. Let Q1, Q2, Q3 and Q4 be the nodal
displacements at node 1 to node 4 respectively, similarly F1, F2, F3, F4
be the nodal force vector from node 1 to node 4 as shown. When these
parameters are represented for a entire structure use capitals which is
called global numbering and for representing individual elements use
small letters that is called local numbering as shown.

This local and global numbering correspondence is established using

element connectivity element as shown
 Now let’s consider a single element in a natural coordinate system that
varies in  and , x1 be the x coordinate of node 1 and x 2 be the x
coordinate of node 2 as shown below.

Let us assume a polynomial

Now

After applying these conditions and solving for constants we have

a0=x1+x2/2 a1= x2-x1/2

Substituting these constants in above equation we get

Where N1 and N2 are called shape functions also called as interpolation
functions.

These shape functions can also be derived using nodal displacements

say q1 and q2 which are nodal displacements at node1 and node 2
respectively, now assuming the displacement function and following
the same procedure as that of nodal coordinate we get

U = Nq
 U = Nq
Where N is the shape function matrix and q is displacement matrix.
Once the displacement is known its derivative gives strain and
corresponding stress can be determined as follows.

element strain displacement matrix

From the potential approach we have the expression of  as

Third step in FEM is finding out stiffness matrix from the
above equation we have the value of K as
But

Therefore now substituting the limits as -1 to +1 because the value of 

varies between -1 & 1 we have

Integration of above equations gives K which is given as

Fourth step is assembly and the size of the assembly matrix is given
by number of nodes X degrees of freedom, for the present example that
has four nodes and one degree of freedom at each node hence size of
the assembly matrix is 4 X 4. At first determine the stiffness matrix of
each element say k1, k2 and k3 as
Similarly determine k2 and k3

The given system is modeled as three elements and four nodes we have
three stiffness matrices.

Since node 2 is connected between element 1 and element 2, the

elements of second stiffness matrix (k2) gets added to second row
second element as shown below similarly for node 3 it gets added to
third row third element
Fifth step is applying the boundary conditions for a given system. We
have the equation of equilibrium KQ=F

K = global stiffness matrix

Q = displacement matrix
F= global force vector

Let Q1, Q2, Q3, and Q4 be the nodal displacements at node 1 to node 4
respectively. And F1, F2, F3, F4 be the nodal load vector acting at node
1 to node 4 respectively.

Given system is fixed at one end and force is applied at other end.
Since node 1 is fixed displacement at node 1 will be zero, so set q1 =0.
And node 2, node 3 and node 4 are free to move hence there will be
displacement that has to be determined. But in the load vector because
of fixed node 1 there will reaction force say R1. Now replace F1 to R1
and also at node 3 force P is applied hence replace F3 to P. Rest of the
terms are zero.

Sixth step is solving the above matrix to determine the displacements

which can be solved either by

  Elimination method 
 Penalty approach method 
Details of these two methods will be seen in later sections.

Last step is the presentation of results, finding the parameters like

displacements, stresses and other required parameters.
 BASIC PROCEDURE

Rayleigh-Ritz Method

As discussed, one can solve axially loaded bars of arbitrary cross-section and material composition along
the length using the lumped mass-spring model. As shown in Figure 12 of Exercise 2.4, one can
approach the exact solution very closely by dividing the bar into more elements. One of the
disadvantages of the lumped models is that we can only compute the deflection at the locations of the
lumped masses (we call these points nodes), and we know nothing about what happens within the
element. Consequently, if we want to get the smooth shape of the deflection curve, we need to take a
very large number of elements. The Raleigh-Ritz method offers an alternative method to overcome these
problems. This method also uses the MPE principle.

Referring back to the tapering beam problem, what we were able to do with the lumped model is
essentially solving the governing differential equation that represents the deflection of axially loaded
bars. Our method of solution was of course numerical. It is worthwhile to study the differential equation
that we just solved numerically in Chapter 2.

Thus, the objectives of this Chapter are: (i) Derive the differential equation of an axially loaded
bar using the force-balance method (ii) Derive the same equation using the MPE principle (iii) Discuss
the Rayleigh-Ritz method.

3.1 Derivation of the governing differential equation of an axially loaded bar using the
force-balance method
Let A(x), the cross-section area of the bar at x, be given. There is a body-force (gravity-like force), f(x),
per unit volume of the bar. σ(x), the axial stress and u(x), the axial deflection, are two unknown
functions. We would like to derive a differential equation that describes the axially loaded bar so that we
can solve for σ(x) and u(x).

Consider a differential element of length dx at some x. The stress and area at the left end of the
differential element are σ(x) and A(x). At (x+dx), the right end, the same quantities can be approximated
dσ(x) dA(x)
as σ(x) + dx and A(x) + dx . The free-body-diagram of the infinitesimally small
dx dx
differential element shows that the internal forces (stresses multiplied by areas of cross-section) balance

3.2
the body-force acting to the right. The body force acting on the differential element is given by f ( x) A( x
)dx . Let us now expand and simplify the internal force acting to the right.

dσ(x) dA(x)
σ(x) + dx A(x) + dx
dx dx 0

= σ(x) A(x) +σ(x) dA(x) dx + A(x) dσ(x) dx +

dσ(x) dA(x)
dx
2
(1)

dx dx dx dx

The last term in the above expression is a small second-order term and hence it can be ignored as shown
stricken by an arrow in Equation (1). The first term balances the internal force acting on the left end of the
differential element. So, the second and third terms and the body-force term should sum to zero for
equilibrium

σ(x) dA(x) dx + A(x) dσ(x) dx + f (x) A(x)dx = 0 (2a)

dx dx

You can easily check that after canceling dx although in the above equation, the two terms on the left
hand side can be collapsed as one term as shown below.

d (σ (x) A(x))
+ f (x) A(x)dx = 0 (2)
dx

This leads to the following differential equation in σ(x).

d
(σ (x) A(x))+ f (x) A(x) = 0 (3)
dx

Next, we would like to express u(x) in terms of σ(x) so that we can get the governing differential
equation in u(x). From the definition of axial strain (change in length divide by the original length), we

du (x)
get the following expression for strain, ε(x) = dx , where du(x) is the deflection of the differential

element of length dx. We also know the relationship between stress and strain: σ(x) = E ε(x) where E is
 3.3

the Young’s modulus. By substituting these relationships into Equation (3), we get the governing
differential equation:

d du(x)
E A(x) + f (x) A(x) = 0 (4)
dx dx

A(x)

f(x)

whose integrand is a function (in this case a differential

dx

f(x) A(x) dx
σ(x) A(x) { σ(x) + dσ(x) dx } {A(x) + d A(x) dx }
dx dx

Figure 1 Force balance of a differential element in an axially loaded bar

We had observed in Chapter 2 that the equilibrium equations could be written using the
force balance method as well as the MPE principle. For the continuous model of an axially
loaded bar, we just derived the equilibrium differential equation using the force-balance method.
We will obtain the same equation using the MPE principle now.

3.2 Derivation of the governing equation using the MPE principle

In this method, first we need to write down the PE of the system. Since this is a continuous model, both
SE and WP are integrals over the length of the bar. Note that

∫
SE = (strain energy density) dV =
dV
∫ 12 (stress) (strain) dV
dV
 L 1 du(x) du(x)
=∫ E A(x)dx (5)
0 2 dx dx

WP = − ∫ L

0
f (x) A(x) u(x) dx (6)

du(x)
By denoting by u′, from Equations (5) and (6), the PE can be written as the sum of SE

and WP.

PE = SE +WP = ∫ L1 A(x)Eu′2 dx −
02
∫ f (x)A(x)u(x)dx
L

0
(7)

As before, we have to minimize PE with respect to the deformation variables. Here, the deflection
variable, u(x) is a continuous function, and the PE is an integral. In fact, PE in Equation (7) is called a

functional in this case an integral

relation) of some function u(x).

Next we will show that if PE is minimized with respect to all kinematically admissible
displacement u(x), then that u(x) satisfies the differential equation (4). To show this, consider the
~
kinematically admissible displacement u (x) = u(x) +αδu(x) where the variation from the exact
~
solution u(x) is given by the function δu(x) times the parameter α . Since u (x) must satisfy the same
~
kinematical boundary conditions as u(x), it follows that δu(x = 0) = 0 . With u (x) substituted in the
place of u(x) in the PE expression in Equation (7), for a given δu(x) , we can regard the potential energy

to be a function of the parameter α , i.e., PE(α) . Then, minimizing PE(α) with respect to α and
setting α = 0 gives the desired governing differential equation:

PE(α) = ∫ 12 EA(x)(u′+αδu′)
L 2 dx − ∫
L
f (x)A(x)(u +α u)dx
0 0

d (PE) =
dα
∫ L

0
EA(x)(u′+αδu′) δu′dx − ∫ L

0
f (x)A(x)(δ u )dx = 0

By substituting α = 0 , we get

d (PE)
dα α=0
= ∫
0
L EA(x)(u′) δu′dx − ∫ f (x)A(x)(δ u )dx = 0
L

3.5

Integrating the expression in the last equation by parts and using the boundary conditions on δu(x) , we

arrive at (note: we substitute u′ = du(x) to get back to our original notation)

 dx

L du(x)
∫
d

EA(x)( ) + f (x) A(x) δudx = 0 (8)

0 dx dx
Since this last integral must vanish for all kinematically admissible δu when the potential energy of the
deformed beam is minimized, it follows that the integrand itself must vanish, i.e.:

d du(x)
EA(x)( ) + f (x) A(x) = 0 (9)
dx dx

which is the same as Equation (4).

We have demonstrated above that the MPE principle can be applied to continuous elastic systems
as well. In fact, in doing so, we have utilized a fundamental mathematical approach in the calculus of
variations. We could also have derived Equation (9) by applying what is known as Euler-Lagrange equation of
calculus of variations. The Euler-Lagrange equation helps us minimize a functional (the PE expression in
Equation (7) in our case) with respect to a function (in our case u(x)). It is given by

d ∂(PE) ∂(PE)
− =0 (10)
dx ∂u′ ∂u

You should verify that Equation (10) also leads to Equation (9).

Once again, the MPE principle gave us the solution with less work and more systematically as
compared to the force-balance method. It is systematic in the following sense. If you were to derive the
governing equilibrium differential equation for a beam, all you need is its PE, as opposed to the force-
balance method where you need to know much more about the internal forces. Much of the theoretical
basis for the finite element method is rooted in the method we used above. In particular, Equation (10) is
a fundamental equation in calculus of variations – an important mathematical tool in FEM formulations.
Refer to any book on calculus of variations for more details. References to two books are given in the
bibliography at the end.
3.3 Rayleigh-Ritz method
In Chapter 2, we solved a problem numerically the differential equation of which we derived in this
chapter. We noted that the lumped-model method gives us deflections at only some discrete points
(nodes), and we know nothing in between the nodes. Rayleigh-Ritz method is an alternative numerical
method to solve the same equation in a simple way to know what happens in between as well.

There is one more thing to bear in mind. The lumped-model method gave us a nice set of linear
equations, which we can easily solve. Also, we reduced a continuous system to a discretized system so
that we can easily implement it on the computer. We don’t want to lose these advantages in the Rayleigh-
Ritz method. Thus, the Rayleigh-Ritz method is another way to discretize the continuous model.

Let us refer to Equation (7). We need to minimize PE to find u(x). If u(x) were to be a scalar
variable, we could have minimized PE very easily as we did several times in Chapter 2. So, we have to
employ a trick to get u(x) to become scalar variables somehow. We can do that as follows.

Note from Figure 12 of Chapter 2 that as we increased the number of elements, the deflection
curve converged to a continuous shape. And that shape looks like a parabola. So, the unknown function
u(x) can be assumed to be a quadratic equation of the form shown below.

u(x) = a + a x + a x2 (10)
0 1 2

But, what we don’t know are three scalars viz. a0, a1, and a2. That is perfectly agreeable to us, because
we can substitute for u(x) from Equation (10) into the expression for PE given in Equation (7). Then, we
get PE in terms of scalar quantities as we wanted. Now invoke the MPE principle.

Extremize PE(a0 , a1 , a2 ) with respect to a0 , a1 , & a2 (11)

The conditions for solving the above are:

∂( PE) = 0 i = 0, 1, 2 (12)
∂ai

Equations (12) result in three linear equations in a0, a1, and a2, which can easily be solved. In fact, you

would note at once that a0 = 0 as u(x=0) = 0. That is our assumed function for u(x) should satisfy the

boundary condition. Or in other words, it should be a kinematically admissible deformation. If you didn’t
appreciate kinematic admissibility in Chapter 2, here is the second chance!
Exercise 3.1
For the same tapered bar problem considered in Chapter 1, use the Rayleigh-Ritz method. That is, write
Equations (7), and (12) to solve for a0, a1, and a2.
· Work it out by hand so that you can understand more.
· Try it out with Maple also so that you can solve more interesting and larger problems.
· Check the Rayleigh-Ritz solution with the lumped-model solution with a large number of
elements.

Exercise 3.2
Consider the overhanging simply supported beam shown below in Figure 2. In order to use the Rayleigh-
2π x
Ritz method, we would like to approximate the deflected profile, v(x) as a cos where L is the
L
length of the beam. Use the minimum potential energy principle to compute the unknown constant, a .
· Draw the assumed deflected profile. Is it a kinematically admissible function?
· Write down the expression for the strain energy of the beam.
· What is the work potential due to each force (use yx=0 , yx=40 , and yx=80)?
· Compute the expression for the total potential energy in terms of a .
· Compute the value of a .

3.8

If a single assumed function is not adequate to represent the deformation, one can use more than
one function for different parts of the structure. Each of these functions will have unknown coefficients
which can be determined by minimizing PE. If more than one function is used, one needs to ensure
continuity of the functions at points where they connect with each other. The following exercise uses this
technique.

Exercise 3.3
Repeat the tapered bar problem if the area of cross-section varies as follows. Area at the top is the same as
before (i.e., A0). The cross-section area remains constant up to the middle of the bar (x=0.5), and then
increases parabolically to become three times A0 at the bottom.

A1 (x) = A0 for 0 ≤ x ≤ 0.5

2
A (x) = A (3 −8 +8x ) for 0.5 ≤ x ≤ 1
2 0

Use two different polynomials for the ranges (0 ≤ x ≤ 0.5) and (0.5 ≤ x ≤ 1) to approximate u(x) with two
piece-wise continuous polynomials. Note that you should ensure continuity at x = 0.5 so that u(x) and its
derivative are continuous.

Exercise 3.4
Comfy Beds, Inc. is considering a new design for the box-spring system. It consists of top and bottom
grids of thin strips of metal connected by linear helical springs. A portion of this new box-spring system
is shown in the figure. Use Rayleigh-Ritz method to determine the maximum deflections of the top and
bottom beams. (see Figure 3).

y1 = a1 x1 (x1 −l1 )
Use as the basis functions where y1 and y2 are the deformations of the top and
y2 = −a2 x22
bottom beams respectively. x1 and x2 are zero at the left end of each beam.
(a) Do the above basis functions satisfy the kinematic admissibility conditions? Explain how.

L EI 2 y2
(b) The strain energy for a beam is given by ∫ 2 dx . Write the total strain energy stored in the
0 2dx
two beams and the spring in terms of a1 and a2.

+ What is the work potential due to the applied force, F of 5 lb? (again in terms of a1 and a2).

+ Use the principle of the minimum potential energy to find the equilibrium values of a1 and a2.
Both beams have rectangular cross-section of thickness 0.1 in and a width of 1 in. The Young's modulus
is 30E6 psi, and the spring constant, k is 10 lb/in. The applied force F is 5 lb. l1 and l2 are respectively 40
in and 30 in.

l1/2 l1/2

Force = F

C
A B

D
E

l2

Figure 3 The schematic of the springs used by Comfy Beds, Inc.

The Rayleigh-Ritz method is a powerful method to use if we know a priori, the nature of the
function for the deformation. However, we may not be able to guess such a function or several piece-wise
functions for any given problem. The FEM enables us to come up with such functions systematically.
Those functions are called shape functions. They serve the following purpose.

• Approximate the continuous deformation using piece-wise functions defined over elements.
• Shape functions depend on some scalar quantities and those scalar quantities are nothing but the
value of the deformation at the nodes.
• Interpolation, i.e., knowing what happens within the element is readily available through shape
functions.
 The following Table summarizes the basic concepts we laid out in Chapters 2 and 3. In
the next chapter, we will study the shape functions and apply this concept to the axially loaded
bars once again. This is the real beginning of our FEM discussion.

Table 1 Comparison of three approaches to deformation analysis

Lumped-model Rayleigh-Ritz FEM
Discretization Divide into segments Discretization concept In principle, it is the
(“element”). The is different. You do same as the lumped
value of the convert a continuous model, i.e., the
deformation at the problem into a discretization is
discrete points discrete problem. But, physical.
(“nodes”) are the the discrete (scalar)
unknown scalar unknowns are
quantities to be coefficients of the
determined using the assumed polynomials
MPE principle. (basis functions).
Interpolation Not possible. You need to know the The procedure is
nature of the function systematic.
so that you can
approximate the Shape functions are
deformation curve used for interpolation
with one or more trial locally for small
(guess) functions elements.
globally.

The procedure is not

systematic.