# MAE4700/5700 Finite Element Analysis for Mechanical and Aerospace Design

Cornell University, Fall 2009 Nicholas Zabaras Materials Process Design and Control Laboratory Sibley School of Mechanical and Aerospace Engineering 101 Rhodes Hall Cornell University Ithaca, NY 14853-3801

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 1

Function norms • In preparation for defining error norms in finite element analysis, we next summarize some definitions. • Consider a function f(x) defined in the interval [a,b]. We define the L2 norm of this function as follows:
⎛ 2 ⎞ || f ( x) ||L 2 = ⎜ ∫ f ( x)dx ⎟ ⎝a ⎠
b 1/ 2

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 2

Finite element error
• Assume that you know the exact solution in the problem of interest (e.g. our 1D deformation problem). Then the L2 error in the finite element solution (displacement) is defined as:
|| e ||L 2 =|| u ( x) − u
h exact

⎛ ⎞ 2 h exact ( x) ||L 2 = ⎜ ∫ ( u ( x) − u ( x) ) dx ⎟ ⎝a ⎠
b

1/ 2

• Since the FEM solution is defined piecewise over each element, we can write the above error as:
⎛ x2 ⎞ 2 h exact h exact || u ( x) − u ( x) ||L 2 = ∑ ⎜ ∫ ( u ( x) − u ( x) ) dx ⎟ ⎟ e ⎜ xe ⎝ 1 ⎠
e

1/ 2

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 3

Finite element error • The normalized L2 error in the finite element solution is also used:
|| u h ( x) − u exact ( x) ||L 2 || u exact ( x) ||L 2 ⎛ ⎞ 2 h exact ⎜ ∫ ( u ( x) − u ( x) ) dx ⎟ ⎠ =⎝a 1/ 2 b ⎛ ⎞ 2 exact ⎜ ∫ ( u ( x) ) dx ⎟ ⎝a ⎠
b 1/ 2

• These definitions are of course not very useful as usually the exact solution is not known!

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 4

Energy norm • In deformation problems, one is often interested in the error in the gradient of the displacement field (e.g. in the strains and stresses). • A common form to express the error in the gradient is using the energy norm defined:
⎛1 b ⎞ 2 h exact h exact || e ||en =|| u ( x) − u ( x) ||en = ⎜ ∫ E ( ε ( x) − ε ( x) ) dx ⎟ ⎝2a ⎠
1/ 2

• Recall that energy.

2 1 1 U = ∫ σ h ( x)ε h ( x)dx = ∫ E ( ε h ( x) ) dx 2 2 a a

b

b

is the strain

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 5

Energy norm • The relative error in the energy is defined similarly as:
⎛1 ⎞ 2 h exact ( x) ) dx ⎟ ⎜ ∫ E ( ε ( x) − ε h exact || u ( x) − u ( x) ||en ⎝ 2 a ⎠ een = = 1/ 2 || u exact ( x) ||en ⎛1 b ⎞ 2 exact ( x) ) dx ⎟ ⎜ ∫ E (ε ⎝2a ⎠
b 1/ 2

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 6

Convergence of the finite element method • We are interested to know how the error in the FE solution reduces as we refine the FE grid (e.g. if the finite element size is halved). • If the finite element contains the complete polynomial of order p, then the L2 error of the displacement varies as follows:
|| e ||L 2 ≤ Ch
p +1

C: constant independent of h but depends on the element polynomial order p=1: linear elements p=2: quadratic elements, etc.
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 7

Convergence of the finite element method

|| e ||L 2 ≤ Ch

p +1

• For linear elements: If you halve the element size, the L2 error decreases by 1/4 • For quadratic elements: If you halve the element size, the L2 error decreases by 1/8!
– The cost of using higher order elements does pay of!
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 8

Convergence in the energy norm

|| e ||en ≤ Ch

p

• For linear elements: If you halve the element size, the energy error decreases by 1/2 • For quadratic elements: If you halve the element size, the energy error decreases by 1/4!
– The cost of using higher order elements does pay of!
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 9

Log-log plots of the finite element error

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 10

Convergence
• The earlier formulas are valid assuming that the actual solution has smooth derivatives up to order p+1. • Assume that the actual solution of our BVP has the property that its derivatives of order λ are square-integrable in Ω but those of order λ+1 and higher are not, λ being an integer λ > 1 • As before, we assume that we use shape functions that contain complete polynomials of degree ≤ p and a uniform mesh of element of equal size h. Then:

|| e ||en ≤ Ch

s

Thus the rate of convergence is p if p<λ or λ if λ<p Note that for λ<p, the rate of convergence is independent of p and no improvement is observed by increasing p
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 11

s = min( p, λ )

Finite element function spaces
• There are an infinite number of functions in U and U0 , i.e. these spaces are of infinite dimension. • When we represent the weight functions by shape functions, then the space of weight functions U 0h becomes of finite dimension (equal to the number of nodes excluding those on essential boundary). • Similarly, the space U h in which we seek our FE solution becomes finite dimensional.

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 12

Finite element function spaces
• While the weak form is exactly equivalent to the strong form for the infinite dimensional spaces U and U 0 , it is only approximately equivalent for the finite dimensional spaces U h ⊂ U and U h ⊂ U 0 , which are used in the finite element method.
0

• Therefore, the balance (equilibrium) equation, and the natural boundary conditions that emanate from the weak form are only satisfied approximately. • In this lecture, we distinguish between the weak forms defined for the exact and finite element solutions.
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 13

Finite element error
• Let us consider the weak forms defined for the exact and finite element solutions. For the elasticity problem these equations are given as follows.
Find u ( x) ∈ U : ∫ dw du AE dx = Awt |Γ t + ∫ bwdx = 0, ∀w ∈ in U 0 dx dx Ω Ω

dwh du h Find u h ( x ) ∈ U h : ∫ AE dx = Awh t |Γ t + ∫ bwh dx = 0, ∀wh ∈ in U 0h dx dx Ω Ω

• Let e = u − u (finite element error). Subtracting the two weak forms gives (note the 1st eq. is also valid for ∀wh ∈ in U h , since U 0h ⊂ U 0):
h
0

dwh de AE dx = 0, ∀wh ∈ in U 0h ∫ dx dx Ω
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 14

Finite element error estimation • How do we estimate the error when the exact solution is not known? • Let us consider the finite element spaces
U h ⊂ U ,U 0h ⊂ U 0 .

• We will prove that minimizes the energy norm of || u( x) − u ( x) || , for any u ( x) ∈U . Here, u( x) is the exact solution. u ( x) U
* * h en

u h ( x)

|| u ( x) − u * ( x) ||en
h

|| u ( x) − u ( x) ||en = min || u ( x) − u ( x) ||en
h * u* ( x )∈U h

|| u ( x) − u h ( x) ||en

• u * ( x) • u h ( x)

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 15

Finite element error estimation
|| u ( x) − u * ( x) ||2 =|| ( u ( x) − u h ( x) ) + (u h ( x) − u * ( x)) ||2 en en
e
h wh∈U 0

• Note that since and satisfy essential boundary conditions, it follows that u ( x) − u ( x) ≡ w ∈U and thus:
u h ( x) u * ( x)
* h h h 0

dwh de || u ( x) − u * ( x) ||2 =|| e + wh ||2 =|| e ||2 + || wh ||2 +2 ∫ AE dx en en en en dx dx Ω

• However, we already have shown that the last integral vanishes:
|| u ( x) − u * ( x) ||2 =|| e ||2 + || wh ||2 ≥|| e ||2 en en en en
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 16

Finite element error estimation
|| u ( x) − u h ( x) ||en ≤|| u ( x) − u * ( x) ||en

• Thus to compute the finite element error, we need to find some function u ( x) ∈U for which the error || u( x) − u ( x) || is easier to compute. • We will use as u ( x) ∈U the finite element interpolant u( x) ∈U of u(x), i.e. a function in U that interpolates u(x) between nodes (i.e. it is equal to the exact solution at the finite element nodes, u( x ) = u( x ) ).
* * en *

h

h

j

j

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 17

Finite element interpolant

• The derivative of the interpolation function u ,x in element e is given by: u ( x j +1 ) − u ( x j )
u ,x = x j +1 − x j

• By the mean value theorem there is a c, x j ≤ c ≤ x j +1:

u , x = u, x (c)
MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 18

Interpolation error

u , x = u, x (c)

• Expand the derivative of the exact solution using a Taylor’s expansion:
u, x ( x) = u, x (c) + ( x − c)u, xx (ξ ), c ≤ ξ ≤ x.

• Subtracting the above two Eqs and assuming | u, xx (ξ ) |≤ a :
u, x ( x) − u , x ( x) = ( x − c)u, xx (ξ ) ⇒| e, x |= | u, x ( x) − u , x ( x) | =| ( x − c)u, xx (ξ ) |≤ a | x − x j |≤ ah
Interpolation error e

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 19

Finite element error for linear elements

| e, x |≤ ah

e

• The interpolation error in the energy norm is then:
e 2 e 1 1 n eh 1 e | e |en = ∫ AE (e, x ) dx = ∑ ∫ Ae E e (e, x ) 2 dx ≤ nhK (ha) 2 2Ω 2 e=1 ( e−1) h 2

where we assumed

Ae E e ≤ K

• Denoting nh=L, and recalling that the energy norm of the finite element solution error is less than or equal to the energy norm of the interpolation error: || e || ≤ 1 KLa h = Ch
2 2 en

2

MAE 4700 – FE Analysis for Mechanical & Aerospace Design N. Zabaras (9/23/2009) 20