Rapid testing of stabilised finite element
formulations for the Reissner-Mindlin plate
problem using the FEniCS project
J. S. Hale*, P. M. Baiz
18th June 2012
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Outline
I Main research goal
I Plate theories
I Numerical demonstration of locking using FEniCS
I Mixed formulation
I Stabilisation
I A few results
I Summary
Our Research
A meshless method for the Reissner-Mindlin plate problem that:
I is based on a sound variational principle
I is free from shear-locking
I avoids problems of previous approaches
I can be extended to the more complicated shell problem
“A locking-free meshfree method for the simulation of shear-deformable plates based on a mixed variational formulation”, Accepted in Computer Methods in Applied Mechanics and Engineering
Our Research
A meshless method for the Reissner-Mindlin plate problem that:
I is based on a sound variational principle
I is free from shear-locking
I avoids problems of previous approaches
I can be extended to the more complicated shell problem
“A locking-free meshfree method for the simulation of shear-deformable plates based on a mixed variational formulation”, Accepted in Computer Methods in Applied Mechanics and Engineering
Our Research
A meshless method for the Reissner-Mindlin plate problem that:
I is based on a sound variational principle
I is free from shear-locking
I avoids problems of previous approaches
I can be extended to the more complicated shell problem
“A locking-free meshfree method for the simulation of shear-deformable plates based on a mixed variational formulation”, Accepted in Computer Methods in Applied Mechanics and Engineering
Our Research
A meshless method for the Reissner-Mindlin plate problem that:
I is based on a sound variational principle
I is free from shear-locking
I avoids problems of previous approaches
I can be extended to the more complicated shell problem
“A locking-free meshfree method for the simulation of shear-deformable plates based on a mixed variational formulation”, Accepted in Computer Methods in Applied Mechanics and Engineering
Our Research
A meshless method for the Reissner-Mindlin plate problem that:
I is based on a sound variational principle
I is free from shear-locking
I avoids problems of previous approaches
I can be extended to the more complicated shell problem
“A locking-free meshfree method for the simulation of shear-deformable plates based on a mixed variational formulation”, Accepted in Computer Methods in Applied Mechanics and Engineering
Our Research
A meshless method for the Reissner-Mindlin plate problem that:
I is based on a sound variational principle
I is free from shear-locking
I avoids problems of previous approaches
I can be extended to the more complicated shell problem
“A locking-free meshfree method for the simulation of shear-deformable plates based on a mixed variational formulation”, Accepted in Computer Methods in Applied Mechanics and Engineering
Finite Elements
Of course, there are many successful approaches to the Reissner-Mindlin problem in the Finite Element literature.
How can FE approaches inform the design of a new meshless method?
Unifying themes with FEniCS
I Mixed variational form (robust, general)*
I Stabilisation (bubbles, parameters)*
I Reduction and projection operators to eliminate extra unknowns (‘tricks’ become rigorous)**
*Easy with FEniCS, **Doable, but could be easier
Finite Elements
Of course, there are many successful approaches to the Reissner-Mindlin problem in the Finite Element literature.
How can FE approaches inform the design of a new meshless method?
Unifying themes with FEniCS
I Mixed variational form (robust, general)*
I Stabilisation (bubbles, parameters)*
I Reduction and projection operators to eliminate extra unknowns (‘tricks’ become rigorous)**
*Easy with FEniCS, **Doable, but could be easier
Finite Elements
Of course, there are many successful approaches to the Reissner-Mindlin problem in the Finite Element literature.
How can FE approaches inform the design of a new meshless method?
Unifying themes with FEniCS
I Mixed variational form (robust, general)*
I Stabilisation (bubbles, parameters)*
I Reduction and projection operators to eliminate extra unknowns (‘tricks’ become rigorous)**
*Easy with FEniCS, **Doable, but could be easier
Finite Elements
Of course, there are many successful approaches to the Reissner-Mindlin problem in the Finite Element literature.
How can FE approaches inform the design of a new meshless method?
Unifying themes with FEniCS
I Mixed variational form (robust, general)*
I Stabilisation (bubbles, parameters)*
I Reduction and projection operators to eliminate extra unknowns (‘tricks’ become rigorous)**
*Easy with FEniCS, **Doable, but could be easier
Finite Elements
Of course, there are many successful approaches to the Reissner-Mindlin problem in the Finite Element literature.
How can FE approaches inform the design of a new meshless method?
Unifying themes with FEniCS
I Mixed variational form (robust, general)*
I Stabilisation (bubbles, parameters)*
I Reduction and projection operators to eliminate extra unknowns (‘tricks’ become rigorous)**
*Easy with FEniCS, **Doable, but could be easier
The Reissner-Mindlin Plate Problem
The Kirchhoff Limit
Reissner-Mindlin Kirchhoff
γ = λ¯t −2 (∇z 3 − θ) = 0
The Kirchhoff Limit
Reissner-Mindlin Kirchhoff
γ = λ¯t −2 (∇z 3 − θ) = 0
Reissner-Mindlin Equations
Discrete Displacement Weak Form
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
R
Ω0L(θh) : (η) dΩ + λ¯t−2R
Ω0(∇z3− θh) · (∇y3− η) dΩ =R
Ω0gy3dΩ or:
ab(θh; η) + λ¯t−2as(θh, z3; η, y3) = f (y3)
Reissner-Mindlin Equations
Discrete Displacement Weak Form
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
R
Ω0L(θh) : (η) dΩ R
Ω0L(θh) : (η) dΩ + λ¯t−2R
Ω0(∇z3− θh) · (∇y3− η) dΩ =R
Ω0gy3dΩ or:
ab(θh; η)
ab(θh; η) + λ¯t−2as(θh, z3; η, y3) = f (y3) R
Ω0L(θh) : (η) dΩ
ab(θh; η)
Reissner-Mindlin Equations
Discrete Displacement Weak Form
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
R
Ω0L(θh) : (η) dΩ + λ¯t−2R
Ω0(∇z3− θh) · (∇y3− η) R
Ω0(∇z3− θh) · (∇y3− η) dΩ =R
Ω0gy3dΩ or:
ab(θh; η) + λ¯t−2aass(θ(θhh, z, z33; η, y; η, y33)) = f (y3) R
Ω0(∇z3− θh) · (∇y3− η)
as(θh, z3; η, y3)
Reissner-Mindlin Equations
Discrete Displacement Weak Form
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
R
Ω0L(θh) : (η) dΩ + λ¯t−2R
Ω0(∇z3− θh) · (∇y3− η) dΩ =R
Ω0gy3dΩ R
Ω0gy3dΩ or:
ab(θh; η) + λ¯t−2as(θh, z3; η, y3) = f (yf(y33)) R
Ω0gy3dΩ
f(y3)
Reissner-Mindlin Equations
Discrete Displacement Weak Form
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
R
Ω0L(θh) : (η) dΩ + λ¯t−2R
Ω0(∇z3− θh) · (∇y3− η) dΩ =R
Ω0gy3dΩ or:
ab(θh; η) + λ¯t¯t¯t−2−2−2as(θh, z3; η, y3) = f (y3)
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
...
d e g r e e = 1
V_3 = F u n c t i o n S p a c e ( mesh , " L a g r a n g e ", d e g r e e ) R = V e c t o r F u n c t i o n S p a c e ( mesh , " L a g r a n g e ",
degree , dim=2 )
U = M i x e d F u n c t i o n S p a c e ([V_3 , R]) z_3 , t h e t a = T r i a l F u n c t i o n s ( U ) y_3 , eta = T e s t F u n c t i o n s ( U ) ...
Find (z3h, θh) ∈ (V3h× Rh) such that for all (y3, η) ∈ (V3h× Rh):
...
d e g r e e = 1
V_3 = F u n c t i o n S p a c e ( mesh , " L a g r a n g e ", d e g r e e ) R = V e c t o r F u n c t i o n S p a c e ( mesh , " L a g r a n g e ",
degree , dim=2 )
U = M i x e d F u n c t i o n S p a c e ([V_3 , R]) z_3 , t h e t a = T r i a l F u n c t i o n s ( U ) y_3 , eta = T e s t F u n c t i o n s ( U ) ...
ij(u) :=1/2(∇u + (∇u)T) L(ij) := D((1 − ν) + ν tr I ) ab(θh, η) :=
Z
Ω0
L(θh) : (η) dΩ
...
e = l a m b d a t h e t a: 0 . 5*( g r a d ( t h e t a ) + g r a d ( t h e t a ) . T )
L = l a m b d a e: D*(( 1 - nu )*e + nu*tr ( e )*I d e n t i t y ( 2 ) )
a_b = i n n e r ( L ( e ( t h e t a ) ) , e ( eta ) )*dx ...
ij(u) :=1/2(∇u + (∇u)T) L(ij) := D((1 − ν) + ν tr I ) ab(θh, η) :=
Z
Ω0
L(θh) : (η) dΩ
...
e = l a m b d a t h e t a: 0 . 5*( g r a d ( t h e t a ) + g r a d ( t h e t a ) . T )
L = l a m b d a e: D*(( 1 - nu )*e + nu*tr ( e )*I d e n t i t y ( 2 ) )
a_b = i n n e r ( L ( e ( t h e t a ) ) , e ( eta ) )*dx ...
as(θh, z3; η, y3) = Z
Ω0
(∇z3− θh) · (∇y3− η) dΩ
...
a_s = i n n e r ( g r a d ( z_3 ) - theta , g r a d ( y_3 ) - eta )*dx
...
as(θh, z3; η, y3) = Z
Ω0
(∇z3− θh) · (∇y3− η) dΩ
...
a_s = i n n e r ( g r a d ( z_3 ) - theta , g r a d ( y_3 ) - eta )*dx
...
ab(θh; η) + λ¯t−2as(θh, z3h; η, y3) = f (y3)
...
a = a_b + l m b d a*t* * -2*a_s f = f o r c e*y_3*dx
u_h = s o l v e( a = = f , bcs= [bc1 , bc2]) z_3h , t h e t a _ h = u_h . s p l i t ()
Done!
ab(θh; η) + λ¯t−2as(θh, z3h; η, y3) = f (y3)
...
a = a_b + l m b d a*t* * -2*a_s f = f o r c e*y_3*dx
u_h = s o l v e( a = = f , bcs= [bc1 , bc2]) z_3h , t h e t a _ h = u_h . s p l i t ()
Done!
102 103 dofs
10−1 100
eL2
t = 0.1 t = 0.01 t = 0.001
Figure: Locking; Fix discretisation, decrease ¯t
Locking
Locking
Inability of the basis functions to represent the limiting Kirchhoff mode.
Vb= {(y3, η) ∈ (V3× R) | ∇y3− η= 0}
Vb∩ Vh= {0}
Locking
Locking
Inability of the basis functions to represent the limiting Kirchhoff mode.
Vb= {(y3, η) ∈ (V3× R) | ∇y3− η= 0}
Vb∩ Vh= {0}
Locking
Locking
Inability of the basis functions to represent the limiting Kirchhoff mode.
Vb= {(y3, η) ∈ (V3× R) | ∇y3− η= 0}
Vb∩ Vh= {0}{{0}0}
for d e g r e e in r a n g e( 0 , 6 ):
V_3 = F u n c t i o n S p a c e ( mesh , " L a g r a n g e ", d e g r e e )
R = V e c t o r F u n c t i o n S p a c e ( mesh , " L a g r a n g e ", degree , dim=2 )
...
102 103 104 dofs
10−6 10−5 10−4 10−3 10−2 10−1 100
eH1(θ)
p = 1 p = 2 p = 3 p = 4 p = 5
Figure: Fix ¯t, increase polynomial order p
100 101 102 1/h
10−1 100
eL2
t = 0.01
Figure: Fix ¯t, refine mesh by decreasing h
Error estimate
ku − uhk
ku − uhk ≤ C(Ω0, κ, E , ν)hp
¯t |u|
Conclusion
We can never fully eliminate locking with these approaches. It would be better to remove the dependence on ¯t entirely.
ku − uhk
Error estimate
ku − uhk ≤ CC(Ω(Ω00, κ, E , ν), κ, E , ν)hp
¯t |u|
Conclusion
We can never fully eliminate locking with these approaches. It would be better to remove the dependence on ¯t entirely.
C(Ω0, κ, E , ν)
Error estimate
ku − uhk ≤ C(Ω0, κ, E , ν)hp
¯t¯t |u|
Conclusion
We can never fully eliminate locking with these approaches. It would be better to remove the dependence on ¯t entirely.
¯t
Error estimate
ku − uhk ≤ C(Ω0, κ, E , ν)hhp
¯t |u|
Conclusion
We can never fully eliminate locking with these approaches. It would be better to remove the dependence on ¯t entirely.
h
Error estimate
ku − uhk ≤ C(Ω0, κ, E , ν)hpp
¯t |u|
Conclusion
We can never fully eliminate locking with these approaches. It would be better to remove the dependence on ¯t entirely.
p
Error estimate
ku − uhk ≤ C(Ω0, κ, E , ν)hp
¯t |u|
Conclusion
We can never fully eliminate locking with these approaches. It would be better to remove the dependence on ¯t entirely.
Mixed Form
Treat the shear stress as an independent variational quantity:
γh= λ¯t−2(∇z3h− θh) ∈ Sh
Discrete Mixed Weak Form
Find (z3h, θh, γh) ∈ (V3h, Rh, Sh) such that for all (y3h, η, ψ) ∈ (V3h, Rh, Sh):
ab(θh; η) + (γh; ∇y3− η)L2 = f (y3) (∇z3h− θh; ψ)L2−¯t2
λ(γh; ψ)L2 = 0
Mixed Form
Treat the shear stress as an independent variational quantity:
γh= λ¯t−2(∇z3h− θh) ∈ Sh
Discrete Mixed Weak Form
Find (z3h, θh, γh) ∈ (V3h, Rh, Sh) such that for all (y3h, η, ψ) ∈ (V3h, Rh, Sh):
ab(θh; η) + (γh; ∇y3− η)L2 = f (y3) (∇z3h− θh; ψ)L2−¯t¯t22
λ(γh; ψ)L2 = 0
¯t2
J. S. Hale
Stability
Theorem (Brezzi 1974)
The classical saddle point problem (¯t = 0) is stable, if and only if, the following conditions hold:
1. (Z-Ellipticity of a)(Z-Ellipticity of a) There exists a constantα ≥ 0 such that:
a(v, v ) ≥ αkv k2X ∀v ∈ Z where Z is the kernel of the bilinear form b:
Z := {v ∈ X | b(v , q) = 0 ∀q ∈ M}
2. (inf-sup condition on b) The bilinear form b satisfies an inf-sup condition:
inf
q∈M
sup
v∈X
b(v, q)
kv kXkqkM =β > 0 ...
(Z-Ellipticity of a)
Stability
Theorem (Brezzi 1974)
The classical saddle point problem (¯t = 0) is stable, if and only if, the following conditions hold:
1. (Z-Ellipticity of a) There exists a constantα ≥ 0 such that:
a(v, v ) ≥ αkv k2X ∀v ∈ Z where Z is the kernel of the bilinear form b:
Z := {v ∈ X | b(v , q) = 0 ∀q ∈ M}
2. (inf-sup condition on b)(inf-sup condition on b) The bilinear form b satisfies an inf-sup condition:
inf
q∈M
sup
v∈X
b(v, q)
kv kXkqkM =β > 0 ...
(inf-sup condition on b)
Displacement Formulation Locking as ¯t → 0
Mixed Formulation Not necessarily stable
Solution
Combine the displacement and mixed formulation to retain the advantageous properties of both
Displacement Formulation Locking as ¯t → 0
Mixed Formulation Not necessarily stable
Solution
Combine the displacement and mixed formulation to retain the advantageous properties of both
as = αadisplacement+ (¯t−2− α)amixed
Displacement
Mixed
Stabilised Mixed Weak Form
Discrete Mixed Weak Form
ab(θh; η) + (γh; ∇y3− η)L2 = f (y3) (∇z3h− θh; ψ)L2−¯t2
λ(γh; ψ)L2 = 0
Stabilised Mixed Weak Form (Brezzi and Arnold 1993, Boffi and Lovadina 1997)
ab(θ; η) + λαas(θ, z3; η, y3) + (γ, ∇y3− η)L2 = f (y3) (∇z3− θ, ψ)L2−λ(1¯t2
−α¯t2)(γ; ψ)L2 = 0
Stabilised Mixed Weak Form
Discrete Mixed Weak Form
ab(θh; η) + (γh; ∇y3− η)L2 = f (y3) (∇z3h− θh; ψ)L2−¯t2
λ(γh; ψ)L2 = 0
Stabilised Mixed Weak Form (Brezzi and Arnold 1993, Boffi and Lovadina 1997)
ab(θ; η) + λαaλαass(θ, z(θ, z33; η, y; η, y33)) + (γ, ∇y3− η)L2 = f (y3) (∇z3− θ, ψ)L2−λ(1¯t2
−α¯t2)
¯t2
λ(1−α¯t2)(γ; ψ)L2 = 0
¯t2
λαas(θ, z3; η, y3)
¯t2 λ(1−α¯t2)
J. S. Hale
Rapid Testing with FEniCS - FEniCS@Imperial Day 22
Stability revisited
Theorem (Brezzi 1974)
The classical saddle point problem (¯t = 0) is stable, if and only if, the following conditions hold:
1. (Z-Ellipticity of a)(Z-Ellipticity of a) There exists a constantα ≥ 0 such that:
a(v, v ) ≥ αkv k2X ∀v ∈ Z where Z is the kernel of the bilinear form b:
Z := {v ∈ X | b(v , q) = 0 ∀q ∈ M}
2. (inf-sup condition on b) The bilinear form b satisfies an inf-sup condition:
inf
q∈M
sup
v∈X
b(v, q)
kv kXkqkM =β > 0 ...
(Z-Ellipticity of a)
Stability revisited
Theorem (Brezzi 1974)
The classical saddle point problem (¯t = 0) is stable, if and only if, the following conditions hold:
1. (Z-Ellipticity of a) There exists a constantα ≥ 0 such that:
a(v, v ) ≥ αkv k2X ∀v ∈ Z where Z is the kernel of the bilinear form b:
Z := {v ∈ X | b(v , q) = 0 ∀q ∈ M}
2. (inf-sup condition on b)(inf-sup condition on b) The bilinear form b satisfies an inf-sup condition:
inf
q∈M
sup
v∈X
b(v, q)
kv kXkqkM =β > 0 ...
(inf-sup condition on b)
...
S = V e c t o r F u n c t i o n S p a c e ( mesh , " DG ", 0 , dim=2 ) ...
g a m m a = T r i a l F u n c t i o n ( S ) psi = T e s t F u n c t i o n ( S ) ...
a = a_b + a l p h a*l m b d a*a_s + i n n e r ( gamma , g r a d ( y_3 ) - eta ) + i n n e r ( g r a d ( z_3 ) - theta , psi ) - t* *2/( l m b d a*( 1 . 0 - a l p h a*t* *2 ) )*i n n e r ( gamma , psi ) ...
MINI2 MINI1 TRIA0220 TRIA1B20
10−2 10−1 100 101 102 103 104 105 α
10−3 10−2 10−1 100 101 102
eH1(z3h)
Convergence in z3h for varying α with t = 10−3, h = 1/8
TRIA1B20 TRIA0220 MINI2 MINI1
Figure: Various elements, Fix h, vary α.
10−2 10−1 100 101 102 103 104 105 α
10−3 10−2 10−1 100 101 102
e
O(h−1K) O(h−2K)
eL2(θ) eH1(θ) eL2(z3) eH1(z3)
Figure: TRIA0220 element. Fix h, vary α.
h = C e l l S i z e ( m e s h )
a l p h a = h* *(-2 . 0 )*c o n s t a n t
100 101 102 1/h
10−5 10−4 10−3 10−2 10−1 100
eH1(z3h)
TRIA0220. Convergence of z3h in H1 norm Varying α recipes and t = 10−3
α = 1.0 α = 5(1h−1−ν)
α = 5(1−ν)h−2 α = 100 1/h 1/h2 1/h3
Figure: Trying different α recipes; convergence can be improved (Lovadina)
Conclusions
I The FEniCS project allows for rapid testing of different Finite Element strategies.
I ∼400 lines of code; Displacement, Mixed, Projections, Errors, Command Line Interface, Output Results etc.
I The stabilisation parameter α should be chosen based on some local discretisation dependent length measure.
I Convergence rates can even be improved by a ‘good’ choice of α
I Based on these results, we have designed a novel meshfree method based on a stabilised weak form with a Local Patch Projection technique to eliminate the shear-stress unknowns a priori
Conclusions
I The FEniCS project allows for rapid testing of different Finite Element strategies.
I ∼400 lines of code; Displacement, Mixed, Projections, Errors, Command Line Interface, Output Results etc.
I The stabilisation parameter α should be chosen based on some local discretisation dependent length measure.
I Convergence rates can even be improved by a ‘good’ choice of α
I Based on these results, we have designed a novel meshfree method based on a stabilised weak form with a Local Patch Projection technique to eliminate the shear-stress unknowns a priori
Conclusions
I The FEniCS project allows for rapid testing of different Finite Element strategies.
I ∼400 lines of code; Displacement, Mixed, Projections, Errors, Command Line Interface, Output Results etc.
I The stabilisation parameter α should be chosen based on some local discretisation dependent length measure.
I Convergence rates can even be improved by a ‘good’ choice of α
I Based on these results, we have designed a novel meshfree method based on a stabilised weak form with a Local Patch Projection technique to eliminate the shear-stress unknowns a priori
Conclusions
I The FEniCS project allows for rapid testing of different Finite Element strategies.
I ∼400 lines of code; Displacement, Mixed, Projections, Errors, Command Line Interface, Output Results etc.
I The stabilisation parameter α should be chosen based on some local discretisation dependent length measure.
I Convergence rates can even be improved by a ‘good’ choice of α
I Based on these results, we have designed a novel meshfree method based on a stabilised weak form with a Local Patch Projection technique to eliminate the shear-stress unknowns a priori
Conclusions
I The FEniCS project allows for rapid testing of different Finite Element strategies.
I ∼400 lines of code; Displacement, Mixed, Projections, Errors, Command Line Interface, Output Results etc.
I The stabilisation parameter α should be chosen based on some local discretisation dependent length measure.
I Convergence rates can even be improved by a ‘good’ choice of α
I Based on these results, we have designed a novel meshfree method based on a stabilised weak form with a Local Patch Projection technique to eliminate the shear-stress unknowns a priori