Finite Element Course

8 Stiffness Matrix

Prof. Raul Durand

 Finite Introduction
Elements

Introduction
In a finite element, the stiffness matrix K relates nodal displacements U with nodal forces
F as:

KU = F

Components of matrix K are partial derivatives of nodal forces with respect to nodal dis-
placements. Thus, for an element with n nodes and two degrees of freedom per node, the
equation above can be written as:
 ∂f1x ∂f1x ∂f1x ∂f1x ∂f1x ∂f1x 
∂u1x ∂u1y ∂u2x ∂u2y ··· ∂unx ∂uny 
 u1x f1x
 ∂f1y   
∂f1y ∂f1y ∂f1y ∂f1y ∂f1y 

 ∂u1x ∂u1y ∂u2x ∂u2y ··· ∂unx ∂uny  
 u1y   f1y 
 ∂f2x ∂f2x ∂f2x ∂f2x ∂f2x ∂f2x  

···
 
∂uny   u2x 
 ∂u1x ∂u1y ∂u2x ∂u2y ∂unx
 f 
   2x 
 ∂f2y ∂f2y ∂f2y ∂f2y ∂f2y ∂f2y   u  
f 

 ∂u1x ∂u1y ∂u2x ∂u2y ··· ∂unx ∂uny  =
2y 
 2y 
 ..   .. 
 
 . .. .. .. .. .. ..   .   . 
 .. . . . . . . 
   
unx  fnx 

 ∂fnx ∂fnx ∂fnx ∂fnx ∂fnx ∂fnx 
 ∂u ∂u1y ∂u2x ∂u2y ··· ∂unx ∂uny  u

 1x ny fny
∂fny ∂fny ∂fny ∂fny ∂fny ∂fny
∂u1x ∂u1y ∂u2x ∂u2y ··· ∂unx ∂uny

Raul
Durand 1
 Finite Equilibrium equations
Elements

Equilibrium equations
The static equilibrium equations in three-dimensions are
given by:
∂σ xx ∂τ xy ∂τ xz
∂x + ∂y + ∂z + bx = 0
∂τ yx ∂σ yy ∂τ yz
∂x + ∂y + ∂z + by = 0
∂τ xy ∂τ yz ∂σ zz
∂x + ∂y + ∂z + bz = 0

These three equations can be written simply as:

div(σ)
.. + b = 0

where b is a vector with body forces. For convenience the

equation above is expressed in terms of the stress vector:
¯Tσ + b = 0
∇ Stress tensor components
¯ is a differential operator given by:
where ∇
∂ ∂ ∂ 
∂x 0 0 ∂y 0 ∂z
¯T =  ∂ ∂ ∂
∇ 0 ∂y 0 ∂x ∂z 0

∂ ∂ ∂
0 0 ∂z 0 ∂y ∂x

Raul
Durand 2
 Finite Divergence theorem
Elements

Divergence theorem
In vector calculus, the divergence theorem, also known as Gauss’s theorem, relates a vector
field inside a volume to the flow of the vector field through its surface.
Using the Gauss’s theorem, the integral of a vectorial function over a closed surface can be
transformed into an integral over the corresponding volume:
˛ ˆ
F · dS = div F dV
S V

where S = n̂ dS is a vector pointing outwards the surface and n̂ the corresponding unitary
vector.
´
For example, consider S F · dS with F = (3x + z, y 2 − sin x, xz + ex ) integrated over the
surface of a box given by: 0 ≤ x ≤ 1, 0 ≤ y ≤ 3 and 0 ≤ z ≤ 2.
Since div F = 3 + 2y + x, using the divergence theorem, the integral can be easily calculated
by the volume integral:
ˆ ˆ 1ˆ 3ˆ 2
div F dV = (3 + 2y + x) dz dy dx
V 0 0 0

Raul
Durand 3
 Finite Stifness matrix derivation
Elements

Stifness matrix derivation

Virtual work approach
One way to obtain the stiffness matrix is by using the principle of Virtual Work.
First, consider the equilibrium equation:

div σ
.. + b = 0

Assumign a virtual displacement field δu, the virtual work equation for a solid with volume
V is:
ˆ ˆ
δu · div σ
.. dV + δu · b dV = 0
V V

Also, using the product rule we have that:

div (δu · σ)
.. = δu · div σ
.. + ∇δu : σ
..

so the first integral can be expressed as:

ˆ ˆ ˆ
δu · div σ
.. dV = div (δu · σ)
.. dV − ∇δu : σ
.. dV
V V V

Raul
Durand 4
 Finite Stifness matrix derivation
Elements

and the original expression becomes:

ˆ ˆ ˆ
∇δu : σ
.. dV − div (δu · σ)
.. dV − δu · b dV = 0
V V V

Using the divergence theorem, the second integral in the last equation can be expressed as
a surface integral, thus:
ˆ ˛ ˆ
∇δu : σ
.. dV − δu · σ
.. · n̂ dS − δu · b dV = 0
V S V

The product σ .. · n̂ = t is the tensor decomposition in the direction of n̂, where t represents a
surface traction force. In addition, tensor σ.. can be expressed by a vector σ, so the equation
above becomes:
ˆ ˛ ˆ
∇δu · σ dV − δu · t dS − δu · b dV = 0
V S V

The term ∇δu represents the virtual strain vector δε, thus:
ˆ ˛ ˆ
δε · σ dV − δu · t dS − δu · b dV = 0
V S V

This equation states the equality between internal and external work.
Raul
Durand 5
 Finite Stifness matrix derivation
Elements

Introduction of finite elements

Considereing the volume of a finite element, strains and displacements in the last equation
can be related with nodal displacements. For example, using interpolation, a vector with
virtual displacements of a point can be expressed in terms of virtual nodal displacements

δu = N δU

where N = [IN1 IN2 IN3 . . . INn ] is a matrix in terms of shape functions.

Also, virtual strains can be related to virtual nodal displacements by:

δε = B δU

Substituting δu and δε into the virtual work equation we have:

ˆ ˛ ˆ
(B δU) σ dV − (N δU) t dS − (N δU)T b dV = 0
T T
V S V

Since the vector with virtual node displacements δU contains values from a constant field,
the equation above becomes:
ˆ ˆ ˆ
T T
B σ dV − N t dS − NT b dV = 0
V S V

Raul
Durand 6
 Finite Stifness matrix derivation
Elements

Introduction of the consitutive relation

Using the constitutive relation, the stress vector can be expressed as σ = DBU, leading to
the finite element equations for static analysis:
ˆ ˆ ˆ
BT DB dV U = NT t dS + NT b dV
V S V

This equation relates forces with displacements and can be written simply as KU = F where
K represents the stiffness matrix of a finite element:
ˆ
K= BT DB dV
V

Also, the right hand side of the finite element equilibrium equation represents the external
force vector:
ˆ ˆ
T
F= N t dS + NT b dV
S V

and is constituted by equivalent nodal forces corresponding to surface tractions and body
forces.
In nonlinear analyses, the stifness matrix relates increments of force and displacements as:
K∆U = ∆F

Raul
Durand 7
 Finite Minimum potential energy approach
Elements

Minimum potential energy approach

The total potential energy for a deformed solid with volume V is given by the strain energy
U plus the work potential W :
ˆ ˆ ˆ
Π = U + W = 12 ε · σ dV − b · u dV − u · t dS
V V S

where vector u represents the displacement field inside the volume.

Introducing finite element relations as u = NU, ε = BU and σ = Dε and minimizing the
total potential energy we get the same expression as the one obtained by the use of the
Virtual Work principle.

Raul
Durand 8
 Finite Stiffness matrix in 1D and 2D
Elements

Stiffness matrix in 1D and 2D

Since matrix B depends on natural coordinates and the element domain is simple in local
space, the stiffness matrix is integrated in the local space. Thus, the volume differential is
written as dV = J dξdηdζ, so:
˚
K= BT DB J dξdηdζ

can be expressed for lower dimensions.

For 1D problems, dV = AdL = AJdξ, so the stiffness matrix becomes:
ˆ
K = A BT E B J dξ
L

For 2D problems, dV = hdA = hJdξdη, so the stiffness matrix becomes:

¨
K= BT DB hJ dξdη

For plane strain problems it is usually considered h = 1.

For axisymmetric problems the volume integral has to be taken over the whole ring of
material, so dV = 2πrdrdz. In this case the stiffness matrix becomes:
¨
K = 2π BT DB r drdz
Raul
Durand 9
 Finite Stiffness matrix in 1D and 2D
Elements

Expressing the volume differential in local coordinates, the integral turns into:
¨
K = 2π BT DB rJ dξdη

Since the term r is remaining inside the integral, a simple integration procedure evaluates
the term r at the element centroid, thus:
¨
K = 2πr̄ BT DB J dξdη

Raul
Durand 10
 Finite Appropriate order of quadrature
Elements

Appropriate order of quadrature

The quadrature rule sufficient to provide the exact in-
tegrals of all terms of the element stiffness matrix when
the element is undistorted is called full integration.
An element is considered undistorted when its geometry
and nodal positions are proportional to the geometry in
local coordinates.
The full integration rule is not enough to integrate dis-
torted elements. For them, Jacobian matrix J is not
constant throughout the element.
Full integration rule may result in over-estimated stiff-
ness matrices.
Appropriate order of integration for
The use of quadrature rules lower than full integra- some quadrilateral elements
tion may result in rank deficiency of the stiffness ma-
trix and may generate spurious deformation modes in a
FEM analysis.

Raul
Durand 11
 Finite Global system
Elements

Global system

Using the superposition principle, a global stiffness ma-

trix K can be assembled from interconnected finite ele-
ments as:

KU = F

where U and F are now global displacements and forces

vectors, respectively.
Vector U contains all displacement degrees of freedom
from the system. It includes support conditions and
other prescribed displacements.
Global stiffness matrix is usually sparse. Proper node
numbering can reduce the bandwidth in stiffness matri-
ces.
Frequently, sparse solvers are required for large finite
element analyses.

Raul
Durand 12
 Finite Global system
Elements

The global vector of forces F is assembled from natural

boundary conditions and body forces. However, force
values at positions corresponding to prescribed displace-
ments are unknown before solving the system. These
values represent reactions forces.
The sequence of global node numbers corresponding to
a particular element is called as connectivity. Global
node numbers should be listed in the same order as
local numbering.
Element connectivity can be stored in a vector. For
example, for element 2 in the figure, the connectivity
vector is:
h i
c(2) = 6 7 2

A mapping vector may be used to store the positions of

element dofs in the global vectors U and F:
h i
(2)
mdof = 3 4 11 12 13 14

Raul
Durand 13
 Finite Stiffness matrix properties
Elements

Stiffness matrix properties

For element stiffness matrices:
Due to rigid body motion, stiffness matrix is singular; hence, it cannot be inverted.
There is no unique solution for U to satisfy F = KU.
There is at least one non-trivial vector U for which KU = 0.
For global stiffness matrix:
The matrix size is equal to the number of total degrees of freedom and it is usually sparse.
The global stiffness matrix is singular. However, the reduced stiffness matrix obtained after
applying displacement boundary conditions should not be singular. This reduced matrix
may be singular in case of insufficient essential boundary conditions.
Symmetric stiffness matrices show that forces are directly proportional to displacements and
vice-versa.
Diagonal terms of the stiffness matrix are usually positive, i.e. forces applied in one direction
do produce displacements in the same direction. In nonlinear analyses, diagonal terms can
be zero or negative if the structure is experiencing softening.

Raul
Durand 14
 Finite Solving procedure
Elements

Solving procedure
The system of equations KU = F can be reorganized as:
" # ! " #
K11 K12 U1 F1
=
K21 K22 U2 F2

where U1 contains the unknown displacements and F2 the

unknown reactions. U2 contains prescribed displacements
and F2 contains prescribed forces.
This equation can be decomposed in the following two equa-
tions:

K11 U1 + K12 U2 = F1
K21 U1 + K22 U2 = F2
Example of coefficients distribution
The first equation leads to the following system which is in a finite element global stiffness
used to find the unknown displacements U1 : matrix

K11 U1 = F1 − K12 U2

Once U1 is found, reaction forces F2 are calculated using

the second equation.

Raul
Durand 15
 Finite Exercises
Elements

Exercises

1. Using computer software, compute the stiffness ma-

trix for the triangular element. Use E = 20 GPa and
ν = 0.25.

2. The figure shows a finite element mesh for a plane

stress analysis. Each element has dimensions 1 m× 1 m 2

and thikness equal to 0.2 m. Only selft weight is consid-

ered in the analysis where γ = 25 kN/m3 . With the aid
of computer software calculate nodal displacements and 1
reaction forces. Consider full integration. The material
properties are E = 20 GPa and ν = 0.25.

3. In the last exercise, calculate the strain and stress vectors at the integration points of
element 1.

Raul
Durand 16
 Finite Locking
Elements

Locking
In some circumstances finite elements show stiffer be-
havior which is referred as locking.
Locking occurs in lower order elements because their
kinematics is not enough to represent the correct so-
lution. In other words, it happens when an element
cannot interpolate a field correctly from nodal values
and shape functions.
The most common types of locking are shear and pres-
Shear locking in a fully integrated 8 node
sure locking. Shear locking occurs when elements are
hexahedron element
subject to bending, and pressure locking occurs when
the material is incompressible.
Figure at top right shows an 8 node hexahedron element
subject to pure bending. The element is not able to as-
sume the curved shape to correctly represent the shape No locking in a fully integrated 20 node
change. In this case incorrect shear strain is introduced. hexahedron element

Figure at bottom right shows a 20 node element that

correctly assumes the shape change. No artificial shear
strain is introduced and there is no shear locking asso-
ciated with this type of element.
Raul
Durand 17
 Finite Reduced integration
Elements

Reduced integration
A lower order than the required quadrature rule to integrate an element is called reduced
integration and involves a loss in accuracy.
Displacement based finite elements are likely to over-estimate the stiffness matrix. The use
of reduced integration tends to soften an element, thus countering the stiffer effect.
On the other hand, reduced integration is convenient to reduce the expense to generate a
stiffness matrix since it uses fewer integration points.
The use of reduced integration is frequent in non-linear analyses in order to reduce stiffness
and to better approximate to real behavior.
The use of reduced integration could generate mesh instability and produce the so-called
spurious modes (hourglass modes).

Raul
Durand 18
 Finite Hourglass modes
Elements

Hourglass modes
A spurious deformation mode may also be called an instability, a
spurious singular mode, a mechanism, a kinematic mode, a zero-
energy mode and an hourglass mode. Spurious deformations arise
in cases such as the use of a low-order quadrature rule.
The term zero-energy mode refers to nodal displacements U that
do not correspond to rigid body motion but nevertheless produces
zero strain energy:
ˆ
Hourglass mode for an element
Ue = 21 UT KU = 12 εT Dε dV = 0 initially rectangular

If Ue = 0 for a mode U different from a rigid body mode, then

U is known as a spurious mode and the corresponding element
is called unstable.
Unstable elements may be combined to form a structure that
is stable but susceptible to certain load patterns, thus resulting
displacements may be excessive at some nodes.
Deformed mesh showing the
An element that displays a mechanism is said to be rank deficient.
hourglassing effect
That is, the rank of K is less than the number of dofs minus the
number of rigid modes.

Raul
Durand 19
 Finite Hourglass modes
Elements

Independent displacement modes of a bilinear element. 1 to 3 are rigid modes. 4 to 6 are constant strain
modes. 7 and 8 are spurious modes.

Possible spurious modes in 8 and 9 node elements integrated using a 2x2 quadrature rule.

Raul
Durand 20
 Finite Stress recovery
Elements

Stress recovery
Stress recovery consists on calculating stresses along an element. This analysis step is usually
called as postprocessing because it happens after calculation of displacements.
Stresses are calculated from computed displacements, and are thus derived quantities:

σ = DBU

In elements based on displacement fields, the accuracy of derived quantities is generally lower
than primary quantities as displacements. However, stresses are superconvergent at Gauss
points because they have the same degree of accuracy than displacements.

√ are ξ = η = 0 in bilinear and trilinear

For rectangular elements, superconvergent points
(solid) elements, and where ξ, η and ζ are ±1/ 3 in eight or nine node plane elements and
20 or 27 solid elements. For distorted elements, Gauss points are not the optimal locations
but provide good values.
Frequently, stresses are extrapolated to element nodes. Then they can be interpolated over
the element using shape functions.

Raul
Durand 21
 Finite Stress recovery
Elements

Shear strain for a beam loaded in bending using bilinear elements

Shear strain for a beam loaded by transverse tip force using 8 node elements

Raul
Durand 22
 Finite Stress recovery
Elements

Assuming that stress values at m sampling (superconvergent) points can be expressed in

terms of nodal values and shape functions:
n
X
σ̄ j (ξ, η, ζ) = Ni (ξ, η, ζ) σ̂ i j = 1...m
i=1
in matrix form:
N1p1 N2p1
   
Nnp1


 σ̄ 1 
 ... 
 σ̂ 1 

N1p2 N2p2 Nnp2
   
 σ̄ 2
 
 
 ... 

σ̂ 2 

.. = .. .. .. ..  ..
. .
 





  . . . . 



N1pm N2pm pm

 σ̄   
m

... Nn
 σ̂ n 

or in short σ̄ = Nσ̂
If the number of sampling points, m, is equal to the number of nodes, n, then nodal values
can be approximated by:
σ̂ = N−1 σ̄

If m > n then:
σ̂ = N+ σ̄
where N+ is the Moore and Penrose pseudoinverse and in this case it is calculated by:
 −1
N+ = NT N NT

Raul
Durand 23
 Finite Equivalent nodal forces
Elements

Equivalent nodal forces

Equivalent nodal forces, based on the element stress

field, are used to compare with external forces (natural
boundary conditions) in order to check the equilibrium
of forces.
Thus, stresses usually calculated at integration points,
are used to estimate equivalent nodal forces. Using Vir-
tual Work principle in a finite element we can write:
ˆ
δU · F = δε · σ dV
V

Later, writting the strain vector as a function of virtual

nodal displacements, the equation above becomes:
ˆ
T
δU F = δUT BT σ dV
V

Then, the equivalent nodal forces can be calculated by:

ˆ Xm
F= BT σ dV ≈ BT σ J wi
V i=1

Raul
Durand 24
 Finite Exercises
Elements

Exercises
1. The nodal coordinates of a three-node bar element are (x1 , y1 ) = (−1, −1), (x2 , y2 ) =
(1, 1) and (x3 , y3 ) = (0, 1). At the end of a finite element analysis, the axial stresses
calculated at the three integration points were: σ1 = 5 kPa, σ2 = 10 kPa and σ3 =
15 kPa. Compute the equivalent nodal loads.
2. Recover the nodal stresses for the last exercise.
3. The nodal coordinates of a four-node quadrilateral element are (x1 , y1 ) = (0, 0), (x2 , y2 ) =
(2, 0), (x3 , y3 ) = (2, 2) and (x4 , y4 ) = 0, 2. The displacements vector for this element is
shown below. Assuming full integration, find the strain and the stress vector at each
integration points. Later, find the internal forces vector F(e) . Use E = 20GPa and
ν = 0.1.
h iT
U(e) = 0.0 0.0 0.001 0.0 0.002 −0.001 −0.001 −0.001

4. In the last exercise, recover the stress values at nodal points.

Raul
Durand 25