9.

The Finite Element Method

The Finite Element Method is a discretization method for
solving ODE / PDE on a domain V .

In a first step, the domain V is decomposed in N ”simple”

elements:
[N
V = Ve
e=1

By ”simple” we mean the geometrical shape: triangles,

tetrahedrons, etc.
The elements don’t have to be identical - only of the same
kind.
If we want to assembly two parts built of different elements,
transition elements can be employed.
9. The Finite Element Method

Avoid distorted elements (with a big aspect ratio).

Start with a coarse mesh and refine it step by step.
To refine the mesh locally, adaptive mesh refinement
algorithms are used.

If we consider only one element V e , we choose a nodal

approximation for u e :

u e = hNi{un } = N1 un1 + N2 un2 + · · · + Nk unk

where hNi is a row vector of shape functions (specific to that

kind of element) and {un } is a column vector of nodal values.
9. The Finite Element Method

With δu e = hNi{δun } as weighting function, we construct the

integral form corresponding to Galerkin’s method on every
element: Z
e
W = δu e (L(u e ) + fv )dV
Ve
After that, we assembly the elementary integral forms into the
global integral form W and ask it to vanish:
N
X
W = We = 0
e=1

This delivers a huge but sparse linear system in the nodal

values.
9. The Finite Element Method

If the operator L is linear, W e becomes:

Z
e
W = hNi{δun } (L(hNi{un }) + fv ) dV =
e
ZV
= hδun i{N} (hL(N)i{un } + fv ) dV =
Ve
Z  Z 
= hδun i {N}hL(N)idV {un } + hδun i {N}fv dV
Ve Ve

Now we can apply integration by parts:

9. The Finite Element Method

Z 
e
W = −hδun i {L1 (N)}hL2 (N)idV {un } +
Ve
Z  Z !
+hδun i {N}fv dV + hδun i {N}fs dS
Ve Sfe

The integrand {L1 (N)}hL2 (N)i is a matrix containing

quadratic terms in the shape functions and their derivatives.
In general, we can express
 
  N
∂N  ∂N 
{L1 (N)}hL2 (N)i = N, , . . . [D]  ∂x  = [B]T [D][B]
∂x ..
.
9. The Finite Element Method

where we denote by [B] the matrix

 
N
 ∂N 
[B] =  ∂x 
..
.

The weak integral form becomes:

Z  Z
e T
W = −hδun i [B] [D][B]dV {un } + hδun i {N}fv dV +
Ve Ve
Z
+hδun i {N}fs dS
Sfe
9. The Finite Element Method

 
R R R
W e =−hδun i ( V e [B]T [D][B]dV ){un }− V e {N}fv dV − S e {N}fs dS
f
R T
If
R we denote V eR[B] [D][B]dV = [k] and
Ve
{N}fv dV + S e {N}fs dS = {f }, W e becomes
f

W e = −hδun i ([k]{un } − {f })

where [k] is called elementary stiffness matrix (and only

depends on the shape functions and the linear operator L),
{un } is the vector of the elementary nodal values and {f } is
the elementary vector of forces.
9. The Finite Element Method
If we assembly now the global integral form as sum over all
elements of the elementary integral forms
N
X
W = We
e=1

we have to take care, because some of the nodal values refer

to nodes common to two or more elements; there, the
equations ”overlap”. At the end, we obtain

W = −hδUn i ([K ]{Un } − {F }) = 0

where [K ] is the global stiffness matrix, {Un } is the vector of

the global nodal values, and {F } is the global vector of forces.
9. The Finite Element Method
As the components of hδUn i are arbitrary, we get the linear
system
[K ]{Un } = {F }
where [K ] is often a banded (or at least a sparse) matrix.

Example: The 1-D bending of a beam

fv (x)

EI
ℓ

z
9. The Finite Element Method

Consider a beam of length ℓ, made of a linear elastic material

(Young’s modulus E , cross section inertia moment I ), subject
to a distributed load fv (x) = q0 cos πx
2ℓ
. Let w (x) denote the
bending line.

In the Euler-Bernoulli beam theory, the elastic potential energy

is Z ℓ
EI ′′
π= (w (x))2 − fv (x)w (x)dx
0 2
where the first part
R ℓ denotes the elastic energy stored in the
deformed beam; 0 fv (x)w (x)dx denotes the work done by
fv (x) along the deflection w (x).
9. The Finite Element Method
In the equilibrium state, the stationarity principle is satisfied:

δπ = W = 0

so Z ℓ
W = EIw ′′ (x)δw ′′ (x) − fv (x)δw (x)dx = 0
0
We decompose now the beam into N 1-D elements, each of
length ℓe = Nℓ and assembly
N Z
X ℓe
W = (EI w e ′′ δw e ′′ − f e δw e )d x̄ = 0
e=1 0

where x̄ is the local coordinate (0 ≤ x̄ ≤ ℓe ).

9. The Finite Element Method
element e
node i node i + 1

wi wi +1

w (x)
wi′
wi′+1

Consider an element e. As the element has two nodes, we can

describe the bending line w (x) by considering the 4 nodal
values wi = w (xi ), wi +1 = w (xi +1), wi′ = w ′ (xi ), and
wi′+1 = w ′ (xi +1). In the local coordinate ξ = ℓx̄e , this reads:
w e (0) = wi , w e (1) = wi +1 , w e ′ (0) = wi′ , and w e ′ (1) = wi′+1 .
9. The Finite Element Method
For the element e, we define the vector of the nodal values to
be:  

 wi 
 ′  
e wi
{un } =

 wi +1 
 ′  
wi +1
and the corresponding shape functions to be the Hermite
polynomials on [0, ℓe ]:
    
1 0 −3 2   1 
 1 − 3ξ 2 + 2ξ 3 
0 ℓe −2ℓe ℓe   ξ  ℓe ξ − 2ℓe ξ 2 + ℓe ξ 3 
{N(ξ)} = 0 0
 =
3 −2 ξ 2 
   3ξ 2 − 2ξ 3 

   
0 0 −ℓe ℓe ξ3 −ℓe ξ 2 + ℓe ξ 3
9. The Finite Element Method
If we define
w e (ξ) = hNi{une }
we obtain the approximation of the bending line w (x) based
on the four nodal values, multiplied by the Hermite
polynomials. Passing from x̄ to ξ, the integral form becomes:
XN  Z 1 Z 1 
e EI T e e e
W =− hδun i e )3
[B] [B]d ξ{un } − ℓ {N}fv d ξ = 0
e=1
(ℓ 0 0

where
 
−3 6  
d2 −2ℓe 3ℓe  1
[B]T = 2 hN(ξ)i = 2 
 3

dξ −6 ξ
−ℓe 3ℓe
9. The Finite Element Method
From here, we obtain the elementary stiffness matrix:
 e e

Z 1 6 3ℓ −6 3ℓ
EI 2EI 3ℓe 2(ℓe )2 −3ℓe (ℓe )2 
[k] = e 3 [B]T [B]d ξ = e 3  
(ℓ ) 0 (ℓ ) −6 −3ℓe 6 −3ℓe 
3ℓe (ℓe )2 −3ℓe 2(ℓe )2

We observe that the elementary stiffness matrix is symmetric.

To carry out the computation, consider now N = 2. We have

thus two elements, each of length ℓe = 2ℓ , and three nodes (at
x1 = 0, x2 = 2ℓ , and x3 = ℓ, respectively). The corresponding
nodal values are (w1 , w1′ ) in x1 , (w2 , w2′ ) in x2 and (w3 , w3′ ) in
x3 .
9. The Finite Element Method
For both elements, the elementary stiffness matrix is the same:
 
24 6ℓ −24 6ℓ
4EI  6ℓ 2ℓ2 −6ℓ ℓ2 
[k1 ] = [k2 ] = 3 

ℓ −24 −6ℓ 24 −6ℓ
6ℓ ℓ2 −6ℓ 2ℓ2

The elementary vectors of the nodal values for the two

elements are:
   

 w   w2 
 ′  ′

1
 
w w2
{un1 } = 1
{un2 } =

 w2   w3 
 ′   ′
 
w2 w3
9. The Finite Element Method

We observe the overlapping on two positions. The global

vector of the nodal values is:
 

w1 

 ′

w 
 1 
w2
{Un } =

w2′ 


 


w 
 ′
3

w3

When we construct the global stiffness matrix [K ], we have to

overlap the elementary stiffness matrices [k1 ] (in pink) and [k2 ]
(in yellow) on two common rows and columns.
9. The Finite Element Method

 
24 6ℓ −24 6ℓ 0 0
 6ℓ 2ℓ −6ℓ ℓ2 0 0

4EI 
 −24 −6ℓ 24 + 24 −6ℓ + 6ℓ −24 6ℓ
[K ] = 3 
ℓ  6ℓ ℓ2 −6ℓ + 6ℓ 2ℓ2 + 2ℓ2 −6ℓ ℓ2

 0 0 −24 −6ℓ 24 −6ℓ 
0 0 6ℓ ℓ2 −6ℓ 2ℓ2
9. The Finite Element Method

We have now to compute the generalized external loads on

each element:
Z 1 Z 1
e e e π(x̄k + ℓe ξ)
{fk } = ℓ {N}fv n d ξ = ℓ q0 cos {N(ξ)}d ξ
0 0 2ℓ

with x̄1 = 0, x̄2 = 2ℓ . For the first element, we get:

 √ √ 

 192(16 − √8 2 − π 2) √ 

e  e 2 
q0 ℓ 16ℓ (96 − 48 2 − π − 4π 2)
e
{f1 } = 4 √ √ 3 √
π  2(−1536 + 768√ 2 +√ 2π + 96π√ 2) 
 e 
8ℓ (192 − 96 2 + 2π 2 − 16π 2)
9. The Finite Element Method
Or, with ℓe = 2ℓ :
 √ √ 

 96(16 −
√ 8 2 − π 2)√ 

q0 ℓ  2 
e 4ℓ(96 − 48√ 2 −√ π − 4π 2)√
{f1 } = 4
π  −1536 + 768√ 2 +√ 2π 3 + 96π√ 2 
 
2ℓ(192 − 96 2 + 2π 2 − 16π 2)

For the second element, we get:

 √ √ 3 √ 

 2(768 √2 − 2π
√ 2 − 96π √2 − 192π) 

e  e 
q0 ℓ 8ℓ (96 2 − 2π − 16π 2 − 16π)
e
{f2 } = 4 √ √
π   4(96π + π 3 + 48π√ 2 − 384√ 2)  
 e 
64ℓ (−4π + 12 2 − π 2)
9. The Finite Element Method
Or, with ℓe = 2ℓ :
 √ √ 3 √ 

 768 √2 − 2π
√ 2 − 96π 2
√ − 192π 

q ℓ  
e 0 2ℓ(96 2 − 2π −√16π 2 − 16π)
√
{f2 } = 4
π   2(96π + π 3 + 48π√ 2 − 384
√ 2)  
 
16ℓ(−4π + 12 2 − π 2)
We assembly the global force vector:
 √ √ 

 96(16 −√8 2 − π 2)√  

 


 4ℓ(96 − 48 2 − π 2
− 4π 2) 


 √ 

q0 ℓ −1536 + 1536√ 2 − 192π
{F } = 4
π   2ℓ(192 − 32π 2 − 16π) 

2(96π + π 3 + 48π √2 − 384√2)

 

 
 16ℓ(−4π + 12√2 − π √2) 
 
9. The Finite Element Method
The global integral form

W = −hδUn i ([K ]{Un } − {F }) = 0

leads to the 6 × 6 linear system:

[K ]{Un } − {F } = 0

If we consider now the boundary conditions, some of the nodal

values are given, leading to zero variations:

w1 = 0 =⇒ δw1 = 0

w1′ = 0 =⇒ δw1′ = 0
w3 = 0 =⇒ δw3 = 0
9. The Finite Element Method

The global nodal vector becomes:

 

 0 
0
 

 
  
w2
{Un } =

 w2′ 


 


 0 
 ′ 
w3

The zero variations also cancel the first, second, and fifth line
in our 6 × 6 linear system. We obtain a reduced 3 × 3 linear
system.
9. The Finite Element Method

    √ 
48 0 6ℓ w2  4  −1536 + 1536√ 2 − 192π 
 0 4ℓ2 ℓ2  w2′ = q0 ℓ 2ℓ(192 − 32π√ 2 − 16π)
2 2  ′ 4π 4 EI  √ 
6ℓ ℓ 2ℓ w3 16ℓ(−4π + 12 2 − π 2)

or, after simplification,

    √ 
48 0 6ℓ w2  4  12(−8 + 8 √ 2 − π) 
 0 4ℓ2 ℓ2  w2′ = 4q0 ℓ 2ℓ(12 − 2π√ 2 − π)
 ′ π 4 EI  √ 
6ℓ ℓ2 2ℓ2 w3 ℓ(−4π + 12 2 − π 2)
9. The Finite Element Method

Solving the system, we obtain the nodal values:

√ q0 ℓ4 q0 ℓ4
w2 = (8 2 − 11) 4 ≃ 0, 3137 4
π EI π EI
√ q0 ℓ 3
q0 ℓ3
w2′ = 2(9 − 2π 2) 4 ≃ 0, 2285 4
π EI π EI
3
q0 ℓ q0 ℓ3
w3′ = 8(3 − π) 4 ≃ −1, 1327 4
π EI π EI
From here, we can construct the approximate polynomial
solutions w 1 (ξ) and w 2 (ξ) on the two elements.