Chap 11 Iso Elements

F.E.M.
- Chapter 11
Isoparametric elements
J.-P. Ponthot
University of Liège - Belgium
December 23, 2021

Introduction
F.E.M - Chapter 11 -Isoparametric elements 2 / 154

Classical FE Theories
● Straight sides
● General quadrangle → difficult to write

→ limited to rectangular geometry
● Shape functions do depend on element geometry. So they are

element-dependent (even for CST)

Classical F.E. theories use straight lines or flat surfaces as boundaries

y
(a,b) Complex shapes (curved) lead to a waste of
elements
(⇒ DOF ⇒ Cost!!)
x Especially in 3 D!
(-a,-b)
→ The mesh should be dictated by the accuracy of the approximation on u or σ

but not by the geometry.

y
● 3D problems nearly impossible to
handle without isoparametric
elements when facets have to be
orthogonal to the reference axes
● One 8-node brick per voxel

produces jagged edges
x ● This is called staircase artifact
z
Duck mesh
Staircase Staircase CAD model Isoparametric

coarse mesh fine mesh elements

Without isoparametric elements
Mesh for biomechanical model based on medical imaging:
● One 8-node brick per voxel
● Produces jagged edges called the

staircase artifact

With isoparametric elements
Use of ISOPARAMETRIC ELEMENTS allows for smooth meshes !

Brain mesh
Voxel based
(staircase artifact)

Example with isoparametric elements
Isoparametric elements provide a solution
Using isoparametric elements renders possible the accurate description of curved

surfaces!


How does it work ?

● Concept proposed by Taig∗ , Irons∗∗ and Zienckiewicz∗∗∗

● Use a natural coordinate system to describe the geometry: ξ, η ∈ [−1, 1]
● Use a change of coordinates which maps the quadrilateral (brick in 3D) into
the biunit square (triunit cube) also named reference or parent element
● Makes it possible to have non rectangular elements (general quadrilateral)
∗
I.C. Taig. Structural Analysis by the Matrix Displacement Method. English Electric Aviation, Report No SO-17, 1961
∗∗
B.M. Irons. Engineering Application of Numerical Integration in Stiffness Method. AIAA J., 14:2035-2037, 1966
∗∗∗
O.C. Zienckiewicz and Y.K. Cheung. The Finite Element Method in Structural and Continuum Mechanics. Mc Graw-Hill, 1968

Elements with curved edges are also possible:
x h
7
4 h 7 3 4 3
6 Allows much better

8 x approximation of the
x 8
6 curved geometries
1 5 2
5 2
1
3 nodes per edge → parabolic sides

More special elements
This technique also allows the generation of special dedicated elements:

Geometry interpolation

● Use of a natural coordinate system to describe the geometry
h x
y
h
2
3 f 3 (0,1)
1 assumed one
to one (1,0)
x 1 (0,0) 2 x
Physical space
or domain Parent space
(Reference or Parent element)
 
∂x ∂x
 ∂ξ ∂η 
● One to one mapping ⇒ det J > 0 where J =  ∂y ∂y 
∂ξ ∂η
Assume polynomial geometry description

h x
y
x = α1 + α2 ξ + α3 η 2
h
3 f 3 (0,1)
y = β1 + β2 ξ + β3 η
1 assumed one
to one (1,0)
Remember displacements in classical F.E. Physical space
x 1 (0,0) 2 x
u = α1 + α2 x + α3 y
v = β1 + β2 x + β3 y
Stipulate that the following nodal relations are satisfied

(nodal compatibility with geometry description)
xI = α1 + α2 ξI + α3 ηI for I = 1 . . . N
yI = β1 + β2 ξI + β3 ηI i.e. for each node

● Geometry interpolation
x = α1 + α2 ξ + α3 η
x = Pa
y = β1 + β2 ξ + β3 η
a h x
y
z }| { 2
h
α1
3 f 3 (0,1)
x P  α2  1 assumed one
z }| { z }| {
  to one (1,0)
x 1 ξ η 0 0 0   α3 
 x 1 (0,0) 2 x
= Physical space
y 0 0 0 1 ξ η  β1  

 β2 
β3
● Nodal compatibility with geometry description :
x1 = α1 y1 = β1
x2 = α1 + α2 y2 = β1 + β2
x3 = α1 + α3 y3 = β1 + β3

     
x1 1 0 0 α1
or  x2  =  1 1 0   α2  xN = Qa
x3 1 0 1 α3
| {z }
Invertible (one to one mapping)
    
α1 1 0 0 x1
 α2  =  −1 1 0   x2  a = Q−1 xN Same for β
α3 −1 0 1 x3
From x = P a and a = Q−1 xN , one gets x = P Q−1 xN = N xN
where N is the shape function matrix Q−1 xN
z { z }| {}|
1 0 0 x1
x P −1 1 0  x2 
z}|{
z }| {
  
x 
1 ξ η 0 0 0 −1 0 1  x3 
=  
y 0 0 0 1 ξ η  1 0 0  y1 
 
 −1 1 0  y2 
−1 0 1 y3

From x = P a and a = Q−1 xN , one gets x = P Q−1 xN = N xN

where N is the shape function matrix
nodal values
z }| {

x1
local value Discrete interpolant  
z }| {
z }| {  y1 
x N1 0 N2 0 N3 0  x2 
=   x = N xN
y 0 N1 0 N2 0 N3  
 y2 
 x3 
y3
with : NI I=1,2,3
N1 = 1 − ξ − η
N2 = ξ 1
N3 = η I
Discrete interpolation of the geometry based on nodal values

Interpolation/shape functions: N = P Q−1
N1 = 1 − ξ − η
N2 = ξ X
NI (ξ, η) = 1
N3 = η
NI (xJ ) = δIJ
identical to triangular coordinates
Do not depend on node coordinates! → same for all elements

Higher order interpolation → curved side elements
N.B. Shape functions are identical to ”triangular” coordinates: ξ1 , ξ2 , ξ3 with

ξ1 + ξ2 + ξ3 = 1

Triangular coordinates

Alternative way to construct shape functions
h x
y
h
2
● Geometry interpolation (same for y)
3 f 3 (0,1)
1 assumed one
x = α1 + α2 ξ + α3 η to one (1,0)
x 1 (0,0) 2 x
Physical space
● Nodal compatibility with geometry description:
x1 = α1 ⇒ α1 = x1
x2 = α1 + α2 α2 = x2 − x1
x3 = α1 + α3 α3 = x3 − x1
⇒ x = x1 + (x2 − x1 )ξ + (x3 − x1 )η
3
X
= (1 − ξ − η )x1 + ξ x2 + η x3 = NI x I
| {z } |{z} |{z}
N1 N2 N3 I=1

Isoparametric elements General
description

Fundamental interpolation formula
● Geometry
N
X
xi = NI (ξ, η)xIi N = number of nodes
I=1
Use same (iso = same in Greek) interpolation for displacements
● Displacement
N
X
ui = NI (ξ, η)uIi Isoparametric element
I=1

Other fields can also be interpolated the same way:

PN
T = I=1 NI (ξ, η)T I TEMPERATURE
PN
h = I=1 NI (ξ, η)hI THICKNESS
PN
δui = I=1 NI (ξ, η)δuIi VIRTUAL DISPLACEMENT
PN
vi = I=1 NI (ξ, η)viI VELOCITY
PN
ai = I=1 NI (ξ, η)aIi ACCELERATION
...
You just have to know nodal values !
⇒ Extremely versatile

Isoparametric representation
Elements with N Nodes
● Element geometry
N
X N
X N
X N
X
1= NI (ξ, η, ζ) x= NI x I y= NI yI z= NI zI
I=1 I=1 I=1 I=1
not trivial (see later)
● Displacement interpolation
N
X
ui = NI uIi
I=1

● Matrix form of the above relations (2D case)

 
    N1
1 1 1 ... 1  
 x   x1 x2 ... xN  N2 
    .. 
 y = y1 y2 ... yN  . 
    
 ux   ux1 ux2 ... uxN  .. 
 . 
uy uy1 uy2 ... uyN NN

More rows may be added to interpolate other field quantities from nodal values
   
1 1 1 ... 1  
 x   x1 x2 ... xN  N1
     N2 
 y   y1 y2 ... yN  
     ..  h is the thickness
 ux = ux1 ux2 ... uxN  . 
     T is the temperature
 uy   uy1 uy2 ... uyN   .. 
    . 
 h   h1 h2 ... hN 
NN
T T1 T2 ... TN

Isoparametric triangles

The linear triangle
N1 = 1 − ξ − η
N2 = ξ N1 = 1 − ξ − η
N3 = η

The quadratic triangle
N1 = (1 − ξ − η)(1 − 2ξ − 2η) N4 = 4(1 − ξ − η)η

N2 = η(2η − 1) N5 = 4ηξ
N3 = ξ(2ξ − 1) N6 = 4ξ(1 − ξ − η)

The six node triangle: shape function plots
N1 = (1 − ξ − η)(1 − 2ξ − 2η) N4 = 4η(1 − ξ − η)

Isoparametric quadrangles

Quadrilateral coordinates ξ, η
Shape functions ??

Construction of shape functions
for Lagrange elements

Shape functions for Lagrange element
Can be done classically :
● ASSUME (same for y and also same for displacements)
x = α0 + α1 ξ + α2 η + α3 ξη (1) (bilinear form)
● STIPULATE (nodal compatibility)

xI = α0 + α1 ξI + α2 ηI + α3 ξI ηI I = 1, 2, 3, 4
invert
xN = Qa −−−→ a = Q−1 xN
         
x1 1 −1 −1 1 α0 α0 1 1 1 1 x1
 x2   1 1 −1 −1   α1  invert  α1  1  −1 1 1 −1   x2 
→ =
 x3   1 1 1 1
  −−−→ 
  α2 
=  
 α2  4  −1 −1 1 1   x3


x4 1 −1 1 −1 α3 α3 1 −1 1 −1 x4
Inserting in (1) reveals

1
NI (ξ, η) = (1 + ξξI )(1 + ηηI )
4

Construction of shape functions for Lagrange elements
ALTERNATIVELY : Simply use products of Lagrange polynomials, hence the name
Lagrange elements
● In 1D :
1 ⇒Compact Form
L11 = (1 − ξ)
2
1
1 NI = (1 + ξξI )
L12 = (1 + ξ) 2
2
Lab → a = degree & b = node number
ξI → nodal coordinate of node I & (ξi = −1 and ξ2 = +1)
● In 2D :
I
I

Construction of shape functions for Lagrange elements
● In 2D :
1
ξ1 = −1 and η1 = −1 → N1 (ξ, η) = L11 (ξ)L11 (η) = (1 − ξ)(1 − η)
4
1
ξ2 = +1 and η2 = −1 → N2 (ξ, η) = L12 (ξ)L11 (η) = (1 + ξ)(1 − η)
4
1
ξ3 = +1 and η3 = +1 → N3 (ξ, η) = L12 (ξ)L12 (η) = (1 + ξ)(1 + η)
4
1
ξ4 = −1 and η4 = +1 → N4 (ξ, η) = L11 (ξ)L12 (η) = (1 − ξ)(1 + η)
4
1
Compact form NI = (1 + ξξI )(1 + ηηI )
4
● In 3D :
1
8 Node Brick NI = (1 + ξξI )(1 + ηηI )(1 + ζζI )
8
ξI , ηI , ζI are isoparametric coordinates of node I (i.e each of them is either −1 or
+1).
4-Node bilinear quadrilateral
1
N1 = 4 (1 − ξ)(1 − η)
1
N2 = 4 (1 + ξ)(1 − η)
1 1
N3 = 4 (1 + ξ)(1 + η) N1 = (1 − ξ)(1 − η)
1 4
N4 = 4 (1 − ξ)(1 + η)

9-Node biquadratic quadrilateral
1
Corner node : N1 (ξ, η) = L21 (ξ)L21 (η)
= ξη(η − 1)(ξ − 1)
4
1
Midside node : N5 (ξ, η) = L2 (ξ)L1 (η) = η(η − 1)(1 − ξ 2 )
2 2
2
Middle node : N9 (ξ, η) = L22 (ξ)L22 (η) = (1 − η 2 )(1 − ξ 2 )

General rule :Na (ξ, η) = L2b (ξ)L2c (η)
a b c a b c
1 1 1 6 3 2 a = node number of the element (1
2 3 1 7 2 3 to 9)
3 3 3 8 1 2 b = number along ξ (1,2 or 3)
4 1 3 9 2 2 c = number along η (1,2 or 3)
5 2 1
Correspondence between 9 node element numbering (a) and ”discrete coordinates”

(b,c) in the ξ and η coordinate system.
1 N5 = 21 (1 − ξ 2 )(η − 1)η
N1 = 4 (1 − ξ)(1 − η)ηξ
1 N6 = 12 (1 + ξ)(1 − η 2 )ξ
N2 = 4 (1 + ξ)(η − 1)ηξ
...
...
N9 = (1 − ξ 2 )(1 − η 2 )

Shape function plots
N1 = 41 (ξ − 1)(η − 1)ξη N5 = 12 (1 − ξ 2 )(η − 1)η
N5 = 12 (1 − ξ 2 )(η − 1)η N9 = (1 − ξ 2 )(1 − η 2 )

Serendipity elements

8 Node ”Serendipity” element

Shape function construction for Serendipity quad
Mid-side nodes → EASY

4 7
3
8 1
6
1
1 5 2
N5=1/2(1-x2)(1-h) N8=1/2(1-x)(1-h2)
Intuitive construction of shape function → SERENDIP

Shape function construction for Serendipity quad
Corner nodes
● Start with bilinear shape function from Q4
● Use shape function from mid-side node (N5 , N8 , . . . ) to cancel out shape
function at all nodes but one
Systematic generation of ”Serendipity” shape functions

Serendipity Transition elements
(5,6,7 nodes !)

Transition elements
● Using similar techniques, one can construct transition elements
Q5 h
h
4 3 Linear 3
4
x
x Linear x
Linear
1 5 2 1
5 2
quadratic
● Allows connecting linear elements with quadratic ones in the same mesh

Transition elements
1. Start from Q4 shape function

2. Use N 5 4 7
3
8
6
1
1 5 2
N5=1/2(1-x )(1-h)
2
3. Modify N1 and N2
Step 1 Step 2
0.5 0.5
1 0.5
^ 1=1/4(1-x)(1-h)
N ^1-1/2 N5
N
Any Quad with 4 to 9 nodes can be constructed this way!

⇒ extremely versatile

Classical 2D elements
Serendipity family of
quadrilateral elements
(4, 8 and 12 nodes)
Lagrange family of
quadrilateral elements
(4, 9 and 16 nodes)
Standard
triangular
elements
(3,6,9 and 10
nodes)

Pascal triangle
Polynomial terms included in typical elements
1 1
x h x
2 2 16 nodes h
x xh h x 2
xh h2
2 3
x3 x2h xh h
x 3
x 2
h xh 2
h 3
12 nodes
4 3 3 4
x 2 2
x h x h xh h x 4
x 3
h x 2 2
h 3
h4
. xh
4 4 .. ...
...
3 2
x h x h x2h3 xh 4 4
...
. .. x h xh
m 4 2
x h x3h3 x2h4
m
...
...
x h m
hm
...
m ... x Serendipity
...
4 3 3 4
x h xh xh xhm x m
h m
... quadrilaterals xh
...
m 2 4 4 2 m
x h xh xh
...
...
xmh3 x3hm
...
xmh4 x4hm 1
Lagrange ... x
...
h
2
quadrilaterals xmhm x xh h2 10 nodes
3 2 2 3
x xh xh h
4 3 4
x x h x2h2 xh3 h
... ... ... Triangles
...
...
...
m m-1 m-2 2 2 m-2 m

x h x h x h xhm-1 h
x ...

3D Lagrange Elements

3D Lagrange Elements
Linear Quadratic
5 z 5 z 5
x(x,y,z) h=-1 16 x=-1
5
13 16
6 6 13 17 17 8
8 8 6 25
1 5 8 nodes 22 8
6 Lagrange or 18 23 7 20
2 15
Serendipity 1 18 14 20 1
9 24 12
7 4 1 26
4 2 19 4 4
x h 2
2 3
10 11 27 nodes
3 NI (x,h,z)=1/8(1+xxI)(1+hhI)(1+zzI) x 3 h Lagrange
x=(x,h,z)
z=-1 central node = 27
9 1 12
21 4
2
10 11
3
23 = 8 nodes 33 = 27 nodes
Cubic 43 : 64 nodes
Quartic 53 : 125 nodes

Serendipity elements in 3D

Serendipity elements in 3D
Node location Interpolation functions Node location

ξI ηI ζI NI (ξ,η,ζ)
8 nodes
±1 ±1 ±1 (1 + ξξI )(1 + ηηI )(1 + ζζI )/8
20 nodes
±1 ±1 ±1 (1 + ξξI )(1 + ηηI )(1 + ζζI )(ξξI + ηηI + ζζI − 2)/8
0 ±1 ±1 (1 − ξ 2 )(1 + ηηI )(1 + ζζI )/4
±1 0 ±1 (1 + ξξI )(1 − η 2 )(1 + ζζI )/4
±1 ±1 0 (1 + ξξI )(1 + ηηI )(1 − ζ 2 )/4
32 nodes
±1 ±1 ±1 (1 + ξξI )(1 + ηηI )(1 + ζζI )(9(ξ 2 + η 2 + ζ 2 ) − 19)/64
±1/3 ±1 ±1 9(1 − ξ 2 )(1 + 9ξξI )(1 + ηηI )(1 + ζζI )/64
±1 ±1/3 ±1 9(1 − η 2 )(1 + ξξI )(1 + 9ηηI )(1 + ζζI )/64
±1 ±1 ±1/3 9(1 − ζ 2 )(1 + ξξI )(1 + ηηI )(1 + 9ζζI )/64

Typical 3D solid elements
Tetrahedron Hexahedron Penta

(brick)
Linear (8 nodes) Linear (6 nodes)

Linear (4 nodes)
Quadratic (10 nodes) Quadratic (20 nodes) Quadratic (15 nodes)

(Serendipity)

Strains

Strains
Isoparametric representation: ui = NI (ξ, η)uIi

1 ∂ui ∂uj 1 ∂NI I ∂NI I
Strains: εij = 2 + =2 u + u
∂xj ∂xi ∂xj i ∂xi j
T
1 1 2
e e e
Strain-displacement matrix: ε = B q with q = u v ue v2 u3 v3 · · · uN vN
Plane stress:  e 
∂N1e ∂N2e ∂NN
 e   0 0 ··· 0 
ε11  ∂x ∂N1 e ∂x
∂N2 e ∂x
∂NN e 
e  e   0 0 ··· 0  e
ε = ε22 = q
e  ∂y ∂y ∂y 
2ε12  ∂N1 ∂N1 ∂N2 ∂N2e
e e e ∂NN e ∂NN e 
···
∂y ∂x ∂y ∂x ∂y ∂x
Same as before but NI = NI (ξ, η)
∂NI ∂NI ∂NI ∂NI
⇒ , → OK ⇒ , →???
∂ξ ∂η ∂x ∂y

Partial derivative computation
Shape functions are written in terms of ξ and η.

But Cartesian partials (with respect to x, y) are required to get strains and stresses.

Shape Function Partial Derivative
Using chain rule:

∂NIe ∂NIe ∂ξ ∂NIe ∂η
= +
∂x ∂ξ ∂x ∂η ∂x
∂NIe ∂NIe ∂ξ ∂NIe ∂η
= +
∂y ∂ξ ∂y ∂η ∂y
∂NIe ∂NIe
Easy to compute : and (Remember NIe = 41 (1 + ξξ I )(1 + ηη I ))
∂ξ ∂η
∂ξ ∂η ∂ξ ∂η
Main problem is to get
∂x ∂x ∂y ∂y

Change of variable theory
Since x = x(ξ, η), y = y(ξ, η) and we have a one to one mapping
∂x ∂x
dx = dξ + dη (1)
∂ξ ∂η
∂y ∂y
dy = dξ + dη (2)
∂ξ ∂η
in matrix form:  
∂x ∂x
dx  ∂ξ ∂η  dξ
=  ∂y ∂y  dη
dy
∂ξ ∂η
Similarly, ξ = ξ(x, y) and η = η(x, y) leads to
 
∂ξ ∂ξ
dξ  ∂y  dx
=  ∂x
∂η ∂η  dy
dη
∂x ∂y

The Jacobian and inverse Jacobian
" #
∂x ∂x
dx ∂ξ ∂η dξ T dξ
= ∂y ∂y =J
dy ∂ξ ∂η
dη dη
" #
∂ξ ∂ξ
dξ ∂x ∂y dx dξ
= ∂η ∂η = J −T
dη ∂x ∂y
dy dη
in which " #
∂x ∂y
∂(x,y) ∂ξ ∂ξ J11 J12
J = ∂(ξ,η) = ∂x ∂y =
∂η ∂η
J21 J22
" #
∂ξ ∂η
∂(ξ,η) ∂x ∂x
J −1 = ∂(x,y) = ∂ξ ∂η
∂y ∂y
J = Jacobian Also OK in 3D

The Jacobian and inverse Jacobian
X the 2 × 2 Jacobian matrix (easy to compute even in 3D

Compute
x= NI x I )
I
   
∂x ∂y ∂NI ∂NI
∂(x, y)  ∂ξ xI yI
∂ξ   ∂ξ ∂ξ 
J= =  ∂x ∂y  =  ∂NI ∂NI 
∂(ξ, η) xI yI
∂η ∂η ∂η ∂η
Then invert numerically to get
 
∂ξ ∂η
∂(ξ, η)  ∂x ∂x 
J −1 = =  ∂ξ ∂η 
∂(x, y)
∂y ∂y
These are the quantities we need for the shape functions partials !
Implication: Shape functions derivatives are no longer polynomials !

Sequence summary
● Use those shape functions partials to build the strain-displacement matrix B at

any point (ξ, η) of interest
 e 
∂N1e ∂N2e ∂NN
   0 0 ··· 0 
ε11  ∂x e ∂x ∂x e 
e  0 ∂N1 ∂N2e ∂NN  e
ε =  ε22  = 0 ··· 0 q
 ∂y ∂y ∂y 
2ε12  ∂N1e ∂N1e ∂N2e ∂N2e ∂NNe ∂NN e 
···
∂y ∂x ∂y ∂x ∂y ∂x
| {z }
B e (ξ,η) contains non-polynomial terms
e e

⇒ ε (ξ, η) = B (ξ, η)q e
with qe T = u1 v1 u2 v2 u3 v3 · · · uN vN
Unlike the 3-node triangle, here B = B(ξ, η) varies over the element.

Stiffness Matrix
for quadrangles

Stiffness Matrix for quadrangles
Z
e 1 T 1 eT e e
U = ε dV = q K q
ε D |{z}
2 V e
2
Bq
Z Z +1 Z +1
Ke = B T DB dV = h
|{z} B T
DB det J dξ dη
V −1 −1
thickness
with
 ∂N ∂ξ ∂N1 ∂η ∂N2 ∂ξ ∂N2 ∂η 
1
+ 0 + 0 ···
B =  ∂ξ ∂x ∂η ∂x ∂ξ ∂x ∂η ∂x 
.. ..
. .
3×2N
Terms in red are non explicit (taken from numerical components of J −1 ) and
generally have a non-polynomial form.
⇒ numerical integration required

Numerical integration :
Gauss integration rule

1D Gauss integration rule
Z b Z +1
1
F (x) dx = (b − a) F (ξ) dξ
a 2 −1
1 1
x(ξ) = (b − a)ξ + (b + a)
with
2 2
1 1
→ dx = (b − a)dξ det J = (b − a)
2 2
Numerical integration ⇒ Approximation

Z +1 n
X
F (ξ)dξ ≈ wi F ( ξi )
−1 |{z} |{z}
i=1 weight sampling point location

● F (ξ) is approximated by an interpolating Lagrange polynomial ψ(ξ)
F(x): Function to be approximated

y(x): Interpolating polynomial
F1
y(x) F3 Fi n+1
Fn+1 X
F(x)
F2 ψ(ξ) = Fi (ξ)Lni (ξ)
Fn i=1
x1 x2 x3 ... xi ... xn xn+1

x=-1 Sampling points x=+1
● The approximation polynomial ψ(ξ) is integrated

● Remind the basic interpolation property of Lagrange polynomials
ψ(ξi ) = Fi ∀ i, i = 1, . . . , n + 1

Newton-Cotes
n equally spaced sampling points

→ exact integration polynomial
ORDER n − 1
Gauss(-Legendre)
n optimized sampling points

→ exact integration polynomial
ORDER 2n − 1

Gauss point location and weight
n ξi (location) wi (weight)
1 0.000000000000000 2.000000000000000
2 ± 0.577350269189626 1.000000000000000
3 ± 0.774596669241483 0.555555555555556
0.000000000000000 0.888888888888889
4 ± 0.861136311594053 0.347854845137454
± 0.339981043584856 0.652145154862546
5 ± 0.906179845938664 0.236926885056189
± 0.538469310105683 0.478628670499366
0.000000000000000 0.568888888888889
6 ± 0.932469514203152 0.171324492379170
± 0.661209386466265 0.360761573048139
± 0.238619186083197 0.467913934572691
n = number
√ of Gauss points (location = roots of Legendre polynomials)
3
N.B. = 0.57735
3

Examples
Z +1 n
X
F (ξ)dξ ≈ wi F (ξi )
−1 i=1
Z +1
● One point F (ξ)dξ ≈ 2F (0)
−1
Z +1 √ √
● Two points F (ξ)dξ ≈ F (−1/ 3) + F (1/ 3)
−1
Z p
+1
5 8 5 p
● Three points F (ξ)dξ ≈ F (− 3/5) + F (0) + F ( 3/5)
−1 9 9 9

1D Gauss integration rules
Graphical representation of the 1D Gauss integration rules
n=1
n=2
n=3
n=4
n=5
x= -1 x=1
The radius of the bullets is proportional to the weight.

Two dimensional product Gauss rules
Canonical form of integral:

Z +1 Z +1 Z +1 Z +1
F (ξ, η) dξ dη = dη F (ξ, η) dξ
−1 −1 −1 −1
Gauss integration rules is applied sequentially with n1 points in the ξ direction and
n2 points in the η direction:
Z +1 Z +1 Z +1 Z +1
F (ξ, η) dξ dη = dη F (ξ, η) dξ
−1 −1 −1 −1
n1 X
X n2
≈ wi wj F (ξi , ηj )
i=1 j=1
Usually n1 = n2

Graphical representation of 2D product-type Gauss integration rules

n=1 (1 x 1 rule) n=2 (2 x 2 rule)
n=3 (3 x 3 rule) n=4 (4 x 4 rule)
The radius of the bullets is proportional to the weight.

Gauss integration of stiffness matrix in 2D
Z
K e 2N ×2N = hB T2N ×3 D 3×3 B 3×2N dV
V
N → number of nodes h = thickness
Rewrite in canonical form:

Z +1 Z +1
Ke = F (ξ, η)dηdξ  
−1 −1 ∂x ∂y
∂(x, y)  ∂ξ ∂ξ 
J= =  ∂x ∂y 
dV = dx dy = det J dξ dη ∂(ξ, η)
∂η ∂η
F (ξ, η) = hB T DB det J

Gauss integration of stiffness matrix in 2D (2)
Then apply the rule to F (a 2N × 2N matrix)
Z +1 Z +1 n1 X
X n2 N
X GP
Ke = F (ξ, η)dηdξ ≈ wi wj F (ξi , ηj ) = w

ei F (ξi , ηi )
−1 −1 i=1 j=1 i=1
● NGP = n1 × n2 = number of Gauss points

● Generally n1 = n2
● Cost is proportional to the number of Gauss points, i.e. NGP = n1 × n2

Gauss integration of stiffness matrix in 3D
Z
K e 3N ×3N = B T3N ×6 D 6×6 B 6×3N dV
V
N → number of nodes
Rewrite in canonical form:

Z +1 Z +1 Z +1  
Ke = F (ξ, η)dηdξdζ ∂x ∂y ∂z
−1 −1 −1  ∂ξ ∂ξ ∂ξ 
∂(x, y, z) 
 ∂x ∂y ∂z 

J= = 
dV = dx dy dz = det J dξ dη dζ ∂(ξ, η, ζ)  ∂η ∂η ∂η 
 ∂x ∂y ∂z 
F (ξ, η, ζ) = B T DB det J ∂ζ ∂ζ ∂ζ

Gauss integration of stiffness matrix in 3D (2)
Then apply the rule to F (a 3N × 3N matrix)
Z+1Z+1Z+1 X n3
n2 X
n1 X N
X GP
Ke = F (ξ, η)dηdξdζ ≈ wi wj wk F (ξi , ηj , ζk ) = w

ei F (ξi , ηi , ζi )
−1 −1 −1 i=1 j=1 k=1 i=1
● NGP = n1 × n2 × n3 = number of Gauss points

● Generally n1 = n2 = n3
● Cost is proportional to the number of Gauss points, i.e. NGP = n1 × n2 × n3

Required order of integration

Z +1 Z +1
K=h B T DB det J dξ dη
−1 −1
" #
∂x ∂y
∂NI (ξ,η) ∂NI (ξ,η) ∂(x,y) ∂ξ ∂ξ
Terms of B : ∂x or ∂y
J= ∂(ξ,η) = ∂x ∂y
∂η ∂η
GENERAL QUAD :
∂NI −1 ∂NI −1 ∂NI
= J11 + J12
∂x ∂ξ ∂η
∂NI −1 ∂NI −1 ∂NI
= J21 + J22
∂y ∂ξ ∂η
Not a polynomial!
RECTANGLE : 2a
∂NI 1 ∂NI
=
∂x a ∂ξ 2b
∂NI 1 ∂NI
=
∂y b ∂η

Required order of integration : Q4 element
Max order for RECT Max order for QUAD

N I = 14 (1 + ηηI )(1 + ξξI ) 2 2
∂NI ∂NI
∂x or ∂y (terms of B) 1 ??? (non polynomial)
D(cst) 0 0
det J = α + βξ + γη 0 (det J = α) 1
Max order for B T DB det J P
Rect 1 0 1 0 ⇒ =2 ???
RECTANGLE:
Exact integration requires 2 Gauss points (in each direction!)
1 Gauss point → 2n − 1 = 1
2 Gauss points → 2n − 1 = 3
QUAD: Exact integration requires. . . a lot of Gauss points

Testing the element stiffness module on two quadrilateral geometries
a
4 3 4 3
a a
1 2 1 2
2a 2a
RECTANGLE RIGHT TRAPEZOID
Main purpose is to illustrate effect of changing Gauss integration rule

Rectangular Element
Rectangular element: Stiffness matrix computed by n × n rule, n = 2, 3, 4, ...

 
42 18 −6 0 −21 −18 −15 0
 18 78 0 30 −18 −39 0 −69 
 
 −6 0 42 −18 −15 0 −21 18 
 
 0 30 −18 78 0 −69 18 −39 
K =
(e)
 −21 −18 −15

 0 42 18 −6 0 
 −18 −39 0 −69 18 78 0 30 
 
 −15 0 −21 18 −6 0 42 −18 
0 −69 18 −39 0 30 −19 78

Eigenvalues of stiffness matrix: 223.64 90 78 46.3603 42 0 0 0
proper rank(K (e) ) = order(K (e) ) − NRBM
5 = 8 − 3 OK
NRBM = number of Rigid Body Modes

Trapezoidal element

Trapezoidal element (cond’t)
Eigenvalues of stiffness matrices obtained by different Gauss quadrature rules :
Rule Eigenvalues (scaled by 10−6 of K (e) )

1x1 8.77276 3.68059 2.26900 0 0 0 0 0
2x2 8.90944 4.09767 3.18565 2.64521 1.54678 0 0 0
3x3 8.91237 4.11571 3.19925 2.66438 1.56155 0 0 0
4x4 8.91246 4.11627 3.19966 2.66496 1.56199 0 0 0
1 x 1 : Rank deficient by two
rank deficiency(K (e) ) = proper rank(K (e) ) − rank(K (e) )

= order(K (e) ) − NRBM − rank(K (e) )
N.B.
● Eigenvalues generally increasing with number of G.P.
● NRBM = Number of Rigid Body Modes.

Definitions
● FULL INTEGRATION = exact integration of K if

❍ Opposing element sides (and faces in 3D) are parallel
❍ Midside nodes at the middle of element side in 2D
i.e. J must be constant over the element.
If these conditions are not satisfied, full integration does NOT mean exact
integration, but inaccuracy does not seem to be detrimental in this case !
● REDUCED INTEGRATION = one order less than full integration.

●
Examples Full Integ Reduced Integ
Q4 - as illustrated before 2×2 1×1
Q8 - see later 3×3 2×2
Q9 - see later 3×3 2×2
As we will see, Full Integration leads to NO rank defficiency !

Reduced integration and
Hourglass modes

Reduced integration
Reduced integration might appear as interesting

● It saves computation time (stiffness evaluation ÷ NGP )
● It renders some element “hybrid” (those elements are generally less stiff than
“exact” kinematically admissible elements – but you loose the monotonic
convergence property)
But it can lead to some troubles

Numerical integration: rank of K (1)
N
X GP N
X GP N
X GP
T
T

K= w
ei F (ξi , ηi ) = w
e det J B DB = αi B DB
| i {z }i | {z }i i
i=1 i=1 scalar αi matrix product i=1
with B = B(ξi , ηi , ζi )
● What is the rank of K ?
❍ It depends on the rank of B T DB.
❍ It depends on the number of Gauss points NGP .
● Properties of the rank (dimensions are assumed compatible !)
❍ rank(A + B) ≤ rankA + rankB
❍ rank(AB) ≤ min(rankA; rankB)
● From the last property, the maximum rank of B T DB is the dimension of D,
i.e. 3 in 2D and 6 in 3D.
● If you use 1 GP ⇒ K = αB T DB T⇒ rankK = rankD
T = 3 in 2D and 6 in 3D.
● If you use more GP ⇒ K = α1 B DB 1 + α2 B DB 2 + . . . according to
the first property, you can possibly increase the rank of K.
R 1
R 1
R T B T DBq dV = 1 qT Kq should
Remember U = V W dV = 2 V εT Dε dV = 2 V q 2
be correctly evaluated !
Gauss point
● Through the connectors q, we have NDOF (2N in 2D,
3N in 3D - N is the number of nodes) independent
modes of displacements (dimension of K and q), in-
cluding NRBM Rigid Body Modes (3 in 2D; 6 in 3D).
● Through ε = Bq, there exists connections between the NDOF (dimensions of q) and
strain components ND (3 in 2D; 6 in 3D - dimensions of D and ε).
● So, each Gauss point introduces maximum 3 contributions in 2D (εxx , εxy , εyy ) and
6 in 3D to the strain energy storage, that is M = NGP × ND contributions.
● If there are not enough storage mechanisms (M = NGP × ND ) to feed the
NDOF − NRBM independent strain modes of K, then K will be rank deficient, and
there will be spurious mechanisms.
● So one should check NDOF − NRBM ≤ NGP × ND →gives the minimum NGP .

In other words: NGP NGP
X X
T
K= w
ei F (ξi , ηi ) = ei B DB i det Ji
w
i=1 i=1
YIELD: M = NGP × ND internal equilibrium relations
with NGP : number of integration points
ND : order of stress/strain matrix D (σ = Dε) (3 in 2D; 6 in 3D)
RULE: Always verify that NDOF −NRBM ≤ M = NGP × ND

with NDOF : number of D.O.F. (element)
NRBM : number of Rigid Body Modes (3 in 2D; 6 in 3D)
If not verified ⇒rank deficiency of K! ⇒ SINGULARITIES ( = Hourglass modes)
Example: Q4: NDOF = 8, ND = 3 and NRBM = 3.

Number of singularities NHG = NDOF − NRBM − NGP × ND
• 1 GP: NHG = 8 − 3 − 3 = 2 !
• 2 GP: NHG = 8 − 3 − 6 = −1 ⇒ OK !

Reduced integration Q4/Q8
If NDOF − NRBM > NGP × ND ⇒ rank deficiency of K
N.B. for a T3 with 1 GP (NDOF = 6), we have 6 − 3 ≤ 3 ⇒ OK !

Reduced integration Q9 (and Q8)
If NDOF − NRBM > NGP × ND ⇒ rank deficiency of K

Reduced integration: element level (1)
Mathematically: U = 21 qT Kq = 12 qT Fint = 21 P
If singularities are present in K, ∃ qHG such that:
Fint = KqHG = 0
If qHG is non trivial (HG stands for HourGlass - see later):

● OK for Rigid Body Modes
● NOT OK for other modes (Spurious modes)
Work associated with spurious deformation modes :
T T
2U = P = qHG Fint = qHG KqHG = 0
→ qHG produces no strain (εHG = BqHG = 0) and does not work in the discrete
model (like RBM) → this is a manifestation of the rank deficiency of K

⇒ Zero energy associated to that mode ⇒

● Spurious modes
● Hourglass modes
● Zero energy modes
● Kinematic modes
● Mechanisms instabilities
● Chickenwiring
● Keystoning
Reported in literature more than 50 years ago in Finite Difference (Maenchen &
Sack, 1964)

What does singularity mean ?
Bilinear quad with 1 Gauss point

1
u = (1 ± ξ)(1 ± η)uI = NI uI
X 4
1
v = (1 ± ξ)(1 ± η)v I = NI v I
4
u = α0 + α1 ξ + α2 η + α3 ξη
OR :
v = β0 + β1 ξ + β2 η + β3 ξη

εxx = α1 + α3 η 
=⇒ εyy = β2 + β3 ξ what is contained in the quad = linear strains
 (Continuum Mechanics)
γxy = (α2 + β1 ) + α3 ξ + β3 η
5 independent modes : α1 , α3 , β2 , β3 , (α2 + β1 ) ⇒ rank(K (e) ) ≤ 5

+3 RBM → OK since K 8×8
∂u ∂u ∂ξ 2 ∂u
N.B. Formally : εxx = = = (2a = length of the element), but
∂x ∂ξ ∂x 2a ∂ξ
this changes nothing here.

Reduced integration : local level
Let’s use 1 GP at the centroid (ξ = 0 & η = 0)

→ strains as seen from the GP (constants !):
ε̃xx = α1 with 1 GP, I am completely

missing the effects
ε̃yy = β2
of α3 and β3 (whatever their value !),
γ̃xy = α2 + β1 i.e. the linear part of the strain
Conclusion • If I have a deformation that activates α3 or β3 , this deformation will

remain undetected by the numerical integration scheme ⇒ zero strain energy will
be associated to that mode (since the numerical scheme does not ’see’ it).
• If I have a deformation that activates α3 or β3 and does NOT
activate α1 , α2 , β1 , β2 (i.e. they are all equal to 0), zero strain energy will be
associated to the element deformation which is wrong !
• My discrete model does not grab all the features of the mathematical
(continuous) model !!!

What is the spurious mode?
● NUMERICALLY: Find eigenvectors conjugated to zero eigenvalues of K

∂2u
● PHYSICALLY: From displacement field: α3 = ∂ξ∂η =
   
4NI u1
2  z }| {   u2 
1 ∂ I 1 −1  
4 ∂ξ∂η (1 ± ξ)(1 ± η) ui  = 4 (u1 − u2 + u3 − u4 ) = 4 [−1 1 − 1 1]  u 
3
u4
α3 6= 0 → u1 = −u2 = u3 = −u4 Mode: h = [−1 1 − 1 1]
HOURGLASS
4 3
1 2
What happens in the whole structure strongly depends on boundary conditions!

HOURGLASS
4 3
Hourglass mode h = [−1 1 − 1 1]

1 2
From shape function calculation, we know that :

u = α0 + α1 ξ + α2 η + α3 ξη
v = β0 + β1 ξ + β2 η + β3 ξη Hourglass
and that : mode
         z }| {
α0 1 1 1 1 u1 0 1 1 1 1 −1
 α1  1  −1 1 1 −1   u2   0  1  −1 1 1 −1  1 
 =   ⇒    
 α2  4 −1 −1 1 1   u3
   0  = 4  −1 −1 1 1   −1 
α3 1 −1 1 −1 u4 −1 1 −1 1 −1 1
Thus, for the hourglass mode:
α0 = α1 = α2 = 0
α3 6= 0

Thus we have : α0 = α1 = α2 = 0 and α3 6= 0.

Also, since there is no vertical displacement, β0 = β1 = β2 = β3 = 0.
So the strain detected by the element is :
ε̃xx = α1 = 0 NO STRAIN AT ALL !

ε̃yy = β2 = 0 ⇒ NO STRAIN ENERGY !
γ̃xy = α2 + β1 = 0 OBVIOUSLY WRONG !
While the correct strains are :

Z +1 Z +1
average strain 1
εxx = α3 η −−−−−−−−→ ε̄xx = α3 ηdξdη = 0
4 −1 −1
εyy = 0
Z +1 Z +1
average shear 1
γxy = α3 ξ −−−−−−−−→ γ̄xy = α3 ξdξdη = 0
4 −1 −1
With 1 Gauss Point, you only capture average values over the element.
This is not enough for quadrilateral elements !
Hourglass modes for quadrilaterals
Q4 with 1 GP.
Q4
h = [1 -1 1 -1]
Q8 and Q9 with 2x2 GP

Q9(Q8) (non communicable*) .
Q9 Q9 with 2x2 GP.
⋆ Assemblage will have the proper rank even if the element has not!

Hourglass modes for quads/bricks
Full (or standard) and Reduced quadrature rules.
Element Q4 Q8 Q9 Brick 8 Brick 20 Brick 27

Full integration 2x2 3x3 3x3 2x2x2 3x3x3 3x3x3
Reduced integration 1x1 2x2 2x2 1x1x1 2x2x2 2x2x2
Number of Hourglass modes 2 1 3 12 6 27
Number of Hourglass modes = NHG = NDOF − NRBM − NGP × ND

( (
3 in 2D 3 in 2D
ND = NRBM =
6 in 3D 6 in 3D
Example: Brick 27 ⇒ NDOF = 27 × 3 = 81

• 8 GP (= 2 × 2 × 2) : NHG = 81 − 8 × 6 − 6 = 27 Hourglass modes
• 27 GP (= 3 × 3 × 3): NHG = 81 − 27 × 6 − 6 = −87 No Hourglass mode

Hourglass modes for quads
Element Hourglass modes

Q4 h1x = [1 − 1 1 − 1]
(2 modes) h1y = [0 0 0 0]
h2x = [0 0 0 0]
h2y = [1 − 1 1 − 1]
Q8 h1x = [1 0 − 1 − 2 2 1 0 − 1]
(1 mode) h1y = [−1 2 − 1 0 0 1 − 2 1]
Q9 h1x = [1 0 − 1 − 2 0 2 1 0 − 1]
(3 modes) h1y = [−1 2 − 1 0 0 0 1 − 2 1]
h2x = h3y [1 − 1 1 − 1 0 − 1 1 − 1 1]
h2y = h3x [0 0 0 0 0 0 0 0 0]

At the structural level
No energy associated to the hourglass modes → spread like a communicable

disease at the structural level
From Belytschko, Liu and Moran, 1996
Mesh in pure hourglass mode of deformation (as opposed to actual computation

where they are influenced by the boundary conditions)

Hourglass modes
Rank sufficiency for some 2D isoparametric elements

(no Hourglass modes)
Element nodes NDOF NDOF − 3 Min NGP recommended rule

3-node triangle 3 6 3 1 centroid
6-node triangle 6 12 9 3 3-midpoint rule
10-node triangle 10 20 17 6 7-point rule
4-node quadrilateral 4 8 5 2 2x2
n= number of nodes (NDOF = 2n)
NB: Gauss rules for triangles, see later.

Triangular elements

Triangular elements
Peculiarities of implementation of isoparametric

triangular elements
● Special Gauss quadrature rules needed

● Computation of Jacobian and shape function x-y derivatives complicated by
having 3 natural coordinates (ξ1 , ξ2 , ξ3 with ξ1 + ξ2 + ξ3 = 1)

Gauss rules for straight sided triangles
Centroid rule: 1 point, degree 1

3
Area A
Z V
e
1 1 1 1
F (ξ1 , ξ2 , ξ3 ) dV ≈ F ( , , ) 1
A Ve 3 3 3 2
Three-interior-point rule : 3 points, degree 2

Z
1 1 2 1 1 1 1 2 1 1 1 1 2
F (ξ1 , ξ2 , ξ3 ) dV ≈ F ( , , ) + F ( , , ) + F ( , , )
A Ve 3 3 6 6 3 6 3 6 3 6 6 3
Midpoint rule : 3 points, degree 2

Z
1 1 1 1 1 1 1 1 1 1
F (ξ1 , ξ2 , ξ3 ) dV ≈ F ( , , 0) + F (0, , ) + F ( , 0, )
A Ve 3 2 2 3 2 2 3 2 2

Gauss rules for straight sided triangles
The 5 simplest Gauss rules drawn over

straight sided triangles

Triangles of Arbitrary geometry
Curved sided triangles

The element of area, dV, of an isoparametric tri-
angular element with N nodes can be expressed dV
as
dV = J dξ1 dξ2 dξ3
where J is the ”Jacobian determinant”, J = det J

 
1 1 1
 XN
∂NI
N
X ∂NI
N
X ∂NI 
 
1  xI xI xI 
J = det  ∂ξ1 ∂ξ2 ∂ξ3 
2  I=1 I=1 I=1 
 XN XN N
X ∂NI 
 ∂NI ∂NI 
yI yI yI
∂ξ1 ∂ξ2 ∂ξ3
I=1 I=1 I=1

Triangles of Arbitrary geometry (cont’d)
Centroid rule
Z
1 1 1 1 1 1
F (ξ1 , ξ2 , ξ3 ) dV ≈ J( , , )F ( , , )
Ve 3 3 3 3 3 3
Midpoint rule
Z
F (ξ1 , ξ2 , ξ3 ) dV ≈
Ve
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
J( , , 0)F ( , , 0) + J(0, , )F (0, , ) + J( , 0, )F ( , 0, )
3 2 2 2 2 3 2 2 2 2 3 2 2 2 2
etc.

Triangles of Arbitrary geometry (cont’d)
The 5 simplest Gauss rules drawn over arbitrary

(curved side) triangles
number annotated near sample point is the weight

Computation of shape function derivatives
6
Illustrated for 6-node 5
quadratic triangle 1 4
2
 
    N1
1 1 1 1 1 1 1  
Recall the isopara-     N2 
 x   x1 x2 x3 x4 x5 x6  
metric element  =  N3 
 y   y1 y2 y3 y4 y5 y6  
definition     N4 
ux ux1 ux2 ux3 ux4 ux5 ux6  
N5
uy uy1 uy2 uy3 uy4 uy5 uy6
N6
N1 = ξ1 (2ξ1 − 1) N2 = ξ2 (2ξ2 − 1)
N3 = ξ3 (2ξ3 − 1) N4 = 4ξ1 ξ2
N5 = 4ξ2 ξ3 N6 = 4ξ3 ξ1

Partial derivative computation (cont’d)
 
ξ1 (2ξ1 − 1)
 ξ2 (2ξ2 − 1) 
 
 ξ3 (2ξ3 − 1) 
w(ξ1 , ξ2 , ξ3 ) = w1 w2 w3 w4 w5 w6 



 4ξ1 ξ2 
 4ξ2 ξ3 
4ξ3 ξ1
where w = w(ξ1 , ξ2 , ξ3 ) can be any quantity interpolated quadratically over the
triangle. Then
XN XN
∂w ∂NI ∂NI ∂ξ1 ∂NI ∂ξ2 ∂NI ∂ξ3
= wI = + + wI
∂x ∂x ∂ξ1 ∂x ∂ξ2 ∂x ∂ξ3 ∂x
I=1 I=1
XN XN
∂w ∂NI ∂NI ∂ξ1 ∂NI ∂ξ2 ∂NI ∂ξ3
= wI = + + wI
∂y ∂y ∂ξ1 ∂y ∂ξ2 ∂y ∂ξ3 ∂y
I=1 I=1
∂ξi ∂ξi
Once again, quantities like and with i = 1, 2, 3 cannot be explicitely
∂x ∂y
obtained.

 
∂ξ1 ∂ξ1
" # ∂x ∂y 
∂w ∂w
N
X ∂NI
N
X ∂NI
N
X ∂NI  ∂ξ2 ∂ξ2 
 
= wI wI wI  
∂x ∂y ∂ξ1 ∂ξ2 ∂ξ3  ∂x ∂y 
I=1 I=1 I=1  ∂ξ3 ∂ξ3 
∂x ∂y
Make w = 1, x, y and stack results rowwise:
 
N
X XN XN
∂NI ∂NI ∂NI
  ∂ξ1 ∂ξ1
 
∂1 ∂1

 
 I=1 ∂ξ1 I=1
∂ξ2
I=1
∂ξ3  ∂x ∂y   ∂x ∂y 
 N N N    
 X ∂NI X ∂NI X ∂NI  ∂ξ2 ∂ξ2   ∂x ∂x 
 x x xI  = 
 I I  ∂x ∂y   ∂x ∂y 
 I=1 ∂ξ1 I=1
∂ξ2
I=1
∂ξ3    
 N N N
 ∂ξ3 ∂ξ3 ∂y ∂y
 X ∂N X ∂N X ∂NI 
 I
yI
I
yI yI  ∂x ∂y ∂x ∂y
∂ξ1 ∂ξ2 ∂ξ3 | {z }
I=1 I=1 I=1 unknowns

This system can be reduced to :

  
1 1 1 ∂ξ1 ∂ξ1
 XN
∂NI
N
X ∂NI
N
X ∂NI  ∂x  
∂y 
   0 0
 xI xI xI   ∂ξ2 ∂ξ2
 
 ∂ξ1 ∂ξ2 ∂ξ3   = 1 0 
 I=1 I=1 I=1  ∂x 
∂y
 XN XN N
X ∂NI  0 1
 ∂NI ∂NI  ∂ξ3 
∂ξ3
yI yI yI
∂ξ1 ∂ξ2 ∂ξ3 ∂x ∂y
I=1 I=1 I=1 | {z }
unknowns are
grouped here

Replacing the shape functions of the 6-node triangle

 
1 1 1
 x1 (4ξ1 − 1) + 4x4 ξ2 + 4x6 ξ3 x2 (4ξ2 − 1) + 4x5 ξ3 + 4x4 ξ1 x3 (4ξ3 − 1) + 4x6 ξ1 + 4x5 ξ2 
y1 (4ξ1 − 1) + 4y4 ξ2 + 4y6 ξ3 y2 (4ξ2 − 1) + 4y5 ξ3 + 4y4 ξ1 y3 (4ξ3 − 1) + 4y6 ξ1 + 4y5 ξ2
 
∂ξ1 ∂ξ1
 ∂x ∂y   
  0 0
 ∂ξ2 ∂ξ2  
 = 1 0 
 ∂x ∂y 
 ∂ξ3 ∂ξ3  0 1
∂x ∂y
or J P = R where J and R are known

Solve J P = R for the partials

 
∂ξ1 ∂ξ1
 ∂x ∂y 
 ∂ξ2 ∂ξ2 
 
P = 
 ∂x ∂y 
 ∂ξ3 ∂ξ3 
∂x ∂y
Plug these in the chain rule to get the x-y partial derivatives of the shape
functions, and use these to form the strain-displacement matrix B

Stress recovery in isoparametric
elements

Stress recovery for isoparametric elements
● In isoparametric elements, stresses are computed at Gauss points, thus

providing a discrete stress field
● A continuous field can be obtained by extrapolating stresses from Gauss points
to corner nodes (see after)
● Stresses obtained this way often display large interelements jumps
● Jumps can be eliminated by interelement averaging at nodes
● This continuous field can be used for:
❍ error indication
❍ nice drawing of (continuous) stress field

How to extrapolate stresses in 1D ?
w2N
• Known: w1GP and w2GP
w2GP
w1GP
• Unknown: w1N and w2N
w1N 1 1
N1 = (1 − ξ) N2 = (1 + ξ)
2 2
w is any stress component or any quan-
Node 1 GP 1 GP 2 Node 2 tity evaluated at the GP (von Mises stress,
ξ
−1
p
− 3=3
p
3=3 +1 plastic strain, etc)
• For any point w(ξ) = 21 (1 −√ξ)w1N + 12 (1 + ξ)w
√
N
2
• For GP1 w1GP = 2 (1 + √3 )w1N + 2 (1 − √33 )w2N
1 3 1
• For GP2 w2GP = 12 (1 − 33 )w1N +

√2
1
(1 + 3
3
)w N
!2
GP 3
√
3

w1 1 1+ 3 1− 3 w1N
• Matrix form = √ √
w2GP 2 1 − 33 1 + 33 w2N
| {z } | {z }
known
N √ √ unknown
GP
invert w1 1 1+ 3 1− 3 w1
−→ = √ √
w2N 2 1− 3 1+ 3 w2GP

How to extrapolate in 1D: extrapolation
w2N
w2GP
• Unknown: w1N and w2N
w1GP
w1N • Known: w1GP and w2GP
• Define an ”embedded element”
between the Gauss points with its own
Node 1 GP 1 GP 2 Node 2
ξ
coordinate system ξ ′ (see yellow in figure).
p p One has:
−1 − 3=3 3=3 +1
w(ξ ′ ) = 12 (1 − ξ ′ )w1GP + 21 (1 + ξ ′ )w2GP
ξ0
p p
− 3 +1 + 3
−1
′
√ √
• Extrapolate to the nodes ξnodes = ± 3 (N.B. 3 ≈ 1.73 > 1)
√ 1 √ 1 √
⇒ w1N = w(− 3) = (1 + 3)w1 + (1 − 3)w2GP
GP
2 2
√ 1 √ 1 √
w2N = w(+ 3) = (1 − 3)w1 + (1 + 3)w2GP
GP
2 2
SAME RESULTS AS BEFORE !

Problems linked to extrapolation
● Extrapolation leads to
stress jumps at the in-
Gauss point Value computed at GP terface ! Can be an error
σ Node Value extrapolated from GP indicator. A unique value can
be obtained by averaging:
>0
1
Stress jump ! σ
einterface = (σelement 1
>0
2
+ σelement 2 )
0
< 0!
Element 1 Element 2 ● Two positive values can be ex-
trapolated and deliver a neg-
Interface ative value (troublesome for
Von Mises stress !)
q
1 2 + σ2 + σ2 ≥ 0
σV M = 6 (σ11 − σ22 )2 + (σ22 − σ33 )2 + (σ33 − σ11 )2 + σ12 23 31

Extrapolation to the corner nodes in 2D (1)
● In isoparametric elements stresses are computed at Gauss points

(because it is where B matrices are computed as well & σ = Dε = DBq)
h
3 3
4 4 h’ 3’
4’
(e) (e’)
x x’
1’
1 1 2’
2 2
● Stress field is only known at discrete points

● How to get a continuous stress field ? (i.e. a stress field that is known at the
nodes and that can be interpolated through σ(ξ, η) = NI σ I where σ I are the
nodal values)

● Define an embedded element (in yellow) with its own coordinate system ξ ′ , η ′
● Inside this element, a continuous stess field can be generated using shape
functions
● Shape functions of ”Gauss Element”
3
4 h’ 3’
N1
(e′ )
= 1
− ξ ′ )(1 − η ′ ) 4’
4 (1
(e′ ) 1
N2 = 4 (1 + ξ ′ )(1 − η ′ ) (e’)
(e′ )
N3 = 1
+ ξ ′ )(1 + η ′ ) x’
4 (1
(e′ )
N4 = 1
− ξ ′ )(1 + η ′ )
1’
4 (1 1 2’
2
● Use these shape functions to EXTRAPOLATE the values from the Gauss point
to the nodes.
w(ξ ′ , η ′ ) = NI (ξ ′ , η ′ )wIGP

Natural coordinates of bilinear quadrilateral nodes
Corner nodes ξ η ξ’
√ η’
√ Gauss nodes ξ η ξ’ η’
−1 −1
1 -1 -1 − 3 − 3 1’ √
3
√
3
-1 -1
√ √ +1 −1
2 +1 -1 + 3 − 3 2’ √
3
√
3
+1 -1
√ √ +1 +1
3 +1 +1 + 3 + 3 3’ √
3
√
3
+1 +1
√ √ −1 +1
4 -1 +1 − 3 + 3 4’ √
3
√
3
-1 +1
As seen from element

As seen from embedded (yellow) element

● Consider any quantity wGP (stress component, equivalent von Mises stress. . . )
known at the Gauss point. We can write:
w(ξ ′ , η ′ ) = NI (ξ ′ , η ′ )wIGP
● To extrapolate, replace the ξ’ and η’ corner coordinates of the actual element

 N   1
√ 1 1
√ 1   GP 
w1 1 + 2 3 − 2√ 1 − 2 3 − 2√ w1
 w2N   − 1
1 + 1
3 − 1
1 − 1
3   w2GP 
 =  2√ 2 2√ 2   
 w3   1 −
N 1 1 1
− 2√   w3 
1 GP
2 3 − 2√ 1 + 2 3
w4N − 12 1 − 21 3 − 12 1 + 21 3 w4GP
| {z } | {z }
Extrapolated Discrete
quantity quantity
at corner nodes at Gauss points
→ continuous field

● After extrapolation at element level, stress jumps from one element to its
neighbors are present. σelem 1 + σelem 2 + . . .
● Nodal averaging i.e. σnode = will
number of elements connected to the node
lead to a unique nodal value.
Stress jumps after extrapolation (contin- Unique nodal value after averaging (con-
uous representation per element). tinuous representation over the whole
Blue dots are nodal averages. structure).

Other ”Gauss Element” configurations
● Similar procedures can be defined for other element types or in 3D
7 3 7 3 3
4 7’ 3’ 4 7’ 3’ 5
4’ 4’ 3’
9 9’ 6’ 6 9 6 2’ 2
8 8’ 8 8’ 6’ 6
2’ 1’
1’ 5’ 1’ 5’ 2’
1 4
1 5 2 1 5 2
Q9 Q8 T6
(3x3 GP) (3x3 GP) (3 GP)

Convergence requirements

Convergence
Convergence means that is the mesh is repeatedly refined, the sequence of discrete
(FEM) solutions converges towards exact displacements, strains and stresses.
Exact means perfect agreement with the mathematical model on which the
element is based or there is no discretization error
(= verification of the model).
Whether or not the mathematical model is a good representation of physical reality

is another problem (= validation of the model).
Finite Elements converge in a perverse way: they converge as we refine

the mesh, and then they diverge because of roundoff.
(Bruce Irons)

Convergence requirements
Convergence = Consistency + Stability
(analog of Lax-Wendroff theorem in finite differences)
Further breakdown of convergence requirements
● Consistency
Completeness: at the element level
Compatibility: C 1 on the element interiors
C 0 on the element boundaries
● Stability
Rank sufficiency: at the element level
Positive Jacobian: at the element level

Completeness of isoparametric
elements

Completeness
COMPLETENESS requires that the element interpolation functions are capable ot

representing an arbitrary linear polynomial (polynomial of order m, if K e involves
mth derivative) when the nodal DOF’s are assigned values in accordance with it.

Completeness
Completeness is a plausible requirement as the following argument indicates : as

the finite element mesh is further and further refined, the exact solution over one
element is essentially linear (think of Taylor’s series) and its derivatives approach
constant values over each element domains.
To ensure that these constant values are representable, the shape function must
contain all constants (corresponds to rigid body modes) and linear monomials
(corresponds to εij = cst or rotations).
For the engineer :
completeness = ability to represent RBM and ε = cst

Completeness of isoparametric elements
Question: Are isoparametric elements able to represent arbitrary linear

polynomials? (convergence requirement) or equivalently, are they able to represent
rigid body modes and εij = cst modes?
By construction: P
x = P NI (ξ, η) xI (1)
y = P NI (ξ, η) y I (2)
ui = NI (ξ, η) uIi (3)

Assume a linear displacement everywhere (leads to constant strains ε)
ui (x, y) = C0 + C1 x + C2 y
⇒ at the nodes uIi (xI , y I ) = C0 + C1 xI + C2 y I
Is it true in the other direction?
Does uIi (xI , y I ) = C0 + C1 xI + C2 y I

implies ui (x, y) = C0 + C1 x + C2 y
everywhere inside the element?
This condition is called completeness
(This is not obvious)

If uIi is assigned uIi = C0 + C1 xI + C2 y I , one gets

X X
I
ui (x, y) = NI (ξ, η) ui = NI (C0 + C1 xI + C2 y I )
X X X
I
= C0 NI + C 1 NI x +C2 NI y I
| {z } | {z }
x see (1) y see (2)
??
= C0 + C1 x + C2 y
X
⇒ Completeness is achieved iff NI (ξ, η) = 1
Easily checked on a case by case basis.

Stability of isoparametric
elements

Stability
Rank sufficiency
The discrete model must possess the same solution uniqueness attributes than the
mathematical model
For displacement finite elements:

The rigid body modes (RBMs) must be preserved ⇒ no zero-energy modes other
than RBMs
Can be tested by looking at the rank of the stiffness matrix.

(Practically : use enough Gauss Points !)

Stability
Positive Jacobian Determinant
The determinant of the Jacobian matrix that relates Cartesian and natural
coordinates must be everywhere positive within the element.
NI (ξ, η) are trivially smooth with respect to ξ, η but they also have to be smooth
W.R.T. x, y.
This is guaranteed by det J > 0 everywhere inside the element.

⇒ one to one change of variable.

Positive Jacobian requirement
QUAD 4
>180° All interior

=180° angles less
than 180°
NI fails to be NI is smooth
smooth
One to one mapping <=> a<180°
QUAD 8/9
● All interior angles < 180◦
● Midside node should not move ”too far ” of the center of the side (middle third
recommendation).

Positive Jacobian: Q4 element
Displacing a corner node of 4-node quad

Positive Jacobian: Q9 element
Displacing a midside node of 9-node quad

Positive Jacobian: 6 node triangle
Displacing a midside node of 6-node equilateral triangle

Convergence requirement
In practice: How much Gauss Points?
● ”Exact” integration requires a lot of GP’s → never used.

Consider a 3D brick N nodes (8,20,27,32,...)
B = [6, 3N ] [. . . ] = dimension of matrix
Z +1 Z +1 Z +1 n
X
3
K= BT DB det J dξ dη dζ ≃ wi (BT DB det J )i

−1 −1 −1 i=1
9N 2
B T DB → [3N, 6] × [6, 6] × [6, 3N ] = [3N, 3N ] ≈ 2 terms (symmetry !)
+ J + det J + J −1 at each GP!
9N 2
ex: for n3 GP [e.g.: 3x3x3=27], n3 loops over 2
terms
Considerable computation required!

Example : for the 27 nodes element

2
N = 27 ⇒ 9N2 ≈ 3280 components in the stiffness matrix K
Then for (n = number of GP in each space direction)
9N 2
n n3 3
n 2
1 1 3 280 Hourglass modes
2 8 26 244 are present
3 27 88 560
4 64 209 920
5 125 410 000 −→ terms to compute
+ J , det J , J −1 at each Gauss point, for each element in the structure !

● Less GP is better (and cheaper!)

❍ ”Ignores” some straining modes
❍ Element is less stiff (”exact” KA elements are too stiff)
❍ Hybrid element model (not exactly Kinematically Admissible)
❍ 6= GP ⇒ = 6 results
❍ Problem dependent
exception: some non-linear problems Experts only

● minimum number of Gauss Points:

❍ Avoid rank deficiency and Hourglass modes (except for experts)
❍ Convergence: εij = cst should be OK
Z Z
K = B T DB det J dξ dη dζ ≈ B T DB h det J dξ dη dζ
V | {z }
VOLUME
⇒ Volume should be computed exactly
⇒ Use full integration

Chap 11 Iso Elements

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chap 11 Iso Elements

Uploaded by

Copyright:

Available Formats

F.E.M.

December 23, 2021

F.E.M - Chapter 11 -Isoparametric elements 2 / 154

● General quadrangle → difficult to write

● Shape functions do depend on element geometry. So they are

F.E.M - Chapter 11 -Isoparametric elements 3 / 154

Classical F.E. theories use straight lines or flat surfaces as boundaries

→ The mesh should be dictated by the accuracy of the approximation on u or σ

F.E.M - Chapter 11 -Isoparametric elements 4 / 154

● One 8-node brick per voxel

Staircase Staircase CAD model Isoparametric

F.E.M - Chapter 11 -Isoparametric elements 6 / 154

Mesh for biomechanical model based on medical imaging:

● One 8-node brick per voxel

● Produces jagged edges called the

F.E.M - Chapter 11 -Isoparametric elements 7 / 154

Use of ISOPARAMETRIC ELEMENTS allows for smooth meshes !

F.E.M - Chapter 11 -Isoparametric elements 8 / 154

F.E.M - Chapter 11 -Isoparametric elements 9 / 154

Isoparametric elements provide a solution

Using isoparametric elements renders possible the accurate description of curved

F.E.M - Chapter 11 -Isoparametric elements 11 / 154

F.E.M - Chapter 11 -Isoparametric elements 12 / 154

F.E.M - Chapter 11 -Isoparametric elements 13 / 154

● Concept proposed by Taig∗ , Irons∗∗ and Zienckiewicz∗∗∗

F.E.M - Chapter 11 -Isoparametric elements 14 / 154

Elements with curved edges are also possible:

6 Allows much better

3 nodes per edge → parabolic sides

F.E.M - Chapter 11 -Isoparametric elements 15 / 154

This technique also allows the generation of special dedicated elements:

F.E.M - Chapter 11 -Isoparametric elements 16 / 154

F.E.M - Chapter 11 -Isoparametric elements 17 / 154

Assume polynomial geometry description

Stipulate that the following nodal relations are satisfied

F.E.M - Chapter 11 -Isoparametric elements 19 / 154

F.E.M - Chapter 11 -Isoparametric elements 20 / 154

F.E.M - Chapter 11 -Isoparametric elements 21 / 154

From x = P a and a = Q−1 xN , one gets x = P Q−1 xN = N xN

F.E.M - Chapter 11 -Isoparametric elements 22 / 154

Interpolation/shape functions: N = P Q−1

Do not depend on node coordinates! → same for all elements

N.B. Shape functions are identical to ”triangular” coordinates: ξ1 , ξ2 , ξ3 with

F.E.M - Chapter 11 -Isoparametric elements 23 / 154

F.E.M - Chapter 11 -Isoparametric elements 24 / 154

● Nodal compatibility with geometry description:

F.E.M - Chapter 11 -Isoparametric elements 25 / 154

F.E.M - Chapter 11 -Isoparametric elements 26 / 154

Use same (iso = same in Greek) interpolation for displacements

F.E.M - Chapter 11 -Isoparametric elements 27 / 154

Other fields can also be interpolated the same way:

You just have to know nodal values !

F.E.M - Chapter 11 -Isoparametric elements 28 / 154

Elements with N Nodes

not trivial (see later)

F.E.M - Chapter 11 -Isoparametric elements 29 / 154

● Matrix form of the above relations (2D case)

F.E.M - Chapter 11 -Isoparametric elements 30 / 154

F.E.M - Chapter 11 -Isoparametric elements 31 / 154

F.E.M - Chapter 11 -Isoparametric elements 32 / 154

F.E.M - Chapter 11 -Isoparametric elements 33 / 154

N1 = (1 − ξ − η)(1 − 2ξ − 2η) N4 = 4(1 − ξ − η)η