Rule of the mesh in the Finite Element Method
M. Duprez
1,V. Lleras
2andA. Lozinski
31CEntre de REcherche en MAthématiques de la DÉsision, Paris
2IMAG, Montpellier
3Laboratoire de Mathématiques de Besançon
Séminaire d’Analyse et Probabilités
29/10/19
Outline
1
Framework : Finite Element Method (FEM)
2
Geometrical quality of a mesh
3
Mesh which is not conforming to the boundaries
Outline
1
Finite Element Method (FEM)
2
Geometrical quality of a mesh
3
Mesh which is not conforming to the boundaries
Framework
Considered equation
−∆u=f, inΩ u= 0, on∂Ω i.e.the problem
Findu∈H01(Ω)s.t. :
⇔(∇u,∇v)L2(Ω)= (f, v)L2(Ω)∀v∈H01(Ω).
General framework
Findu∈V s.t. : a(u, v) =l(v)∀v∈V where
• l(·)linear and continuous:|l(v)|6C1kvk ∀v∈V
• a(·,·)bilinear continuous:|a(u, v)|6C2kukkvk ∀u, v∈V andcoercive:|a(v, v)|>Ckvk2∀v∈V
1
Framework : Finite element method
LetVh=hφk∈V :k∈ {1, ..., N}i ⊂V.
System (1) will be approximate by Finduh∈Vhs.t. :
a(uh, vh) =l(vh)∀vh∈Vh
⇔
FindUh∈RNs.t. : AhUh=Fh
where
Ah=a(φk, φj)kj
Fh=l(φk)k
Thus
uh=X Uhkφk.
Michel Duprez Rule of the mesh in the Finite Element Method 2
Framework : Lagrange continuous finite element
Piecewise linear Lagrange FE order 1 (P1) Lagrange interpolation
Order 1 (P1), dimension 2 Order 2 (P2), dimension 1
3
Outline
1
Framework : Finite Element Method (FEM)
2
Geometrical quality of a mesh
3
Mesh which is not conforming to the boundaries
Challenge
Questions
â Geometrical criterion on the mesh to ensure the convergence ? ku
h− uk −→
h→0
0 where h := max
Cellin mesh
diam ( Cell )
â Optimal order in the a priori estimates for P
1(
|u − u
h|
16 Ch|u|
2,
|u − u
h|
06 Ch
2|u|
2.
4
Framework : geometrical condition
• Minimum angle condition âdim 2 : Zlámal 68’
min{angle}> α >0
• Ciarlet condition âdim n : Ciarlet 78’
hK
ρK
< γ
hK
ρK
• Maximum angle condition
âdim 2 : Babuska-Aziz 76’, Barnhill-Gregory 76’, Jamet 76’
âdim 3 : Krizek 92’
max{angle}< β < π
• Local damage âdim 2 : Kucera 2016
âdim n : Duprez-Lleras-Lozinski 2019
Michel Duprez Rule of the mesh in the Finite Element Method 5
Idea of the proof
Findu∈V s.t. :
a(u, v) =l(v)∀v∈V (1)
Finduh∈Vhs.t. :
a(uh, vh) =l(vh)∀vh∈Vh (2) Lemma (Local interpolation)
For each cellK∈ Thandu∈H2(K)
|u− Ihu|1,K6Ch|u|K,2, whereIhis theLagrange interpolation operator.
Lemma (Céa, see Ern-Guermond’s book) Letuanduhsolution to (1) and (2). Then
|u−uh|1,Ω6C inf
vh∈Vh|u−vh|1,Ω.
Thus
|u−uh|1,Ω6C inf
vh∈Vh|u−vh|1,Ω6C|u− Ihu|1,Ω6Ch|u|Ω,2.
6
Counter-example [Babuska-Aziz 1976]
Let A
1= (−1, 0), A
2= (0, 1) and A
3= (0, ε).
Consider u(x
1, x
2) = x
21. It holds
|u − I
hu|
1,K> 1 ε .
→ large interpolation error
Michel Duprez Rule of the mesh in the Finite Element Method 7
Inclusion of meshes [Hannukaimen-Korotov-Krizek 2012]
N2
N
TriangulationTh1
N2
N
TriangulationTh2 SinceVh1 ⊂Vh2,
|u−u2h|1,Ω6C inf
vh2∈V2 h
|u−vh2|1,Ω6C inf
vh1∈V1 h
|u−vh1|1,Ω6C|u−Ih1u|1,Ω6Ch|u|Ω,2.
TriangulationTh1 TriangulationTh2
8
Local damage on the mesh : example
Assume the degenerated cells are isolated :
Michel Duprez Rule of the mesh in the Finite Element Method 9
Local damage on the mesh : exact assumption
Assumption
LetK1deg, ..., KIdegbe the degen. cells
hKdeg i
/ρKdeg i
> c0.
Assume that hPi/ρPi 6c1. hKi,j/ρKi,j 6c1. hKnd
i /ρKnd i 6c1. Pei∩Pej=∅
#Pei6M
Kind Kid
Ki,1 Ki,2
Ki,3
Ki,4
Ki,5
x0
Pi:=Kid∪Kind Pei:=Pi∪jKi,j
Theorem (D.-Lleras-Lozinski 2019)
Letu∈V anduh∈Vhbe the solutions to continuous and discrete Systems.
Then, under Assumption 1,
|u−uh|1,Ω≤Ch|u|2,Ω and |u−uh|0,Ω≤Ch2|u|2,Ω.
10
Sketch of proof
Modified Lagrange interpolation operator
Definition
Forv∈H2(Ω), definedeIh(v)∈Vhs.t.
Ieh=IhonK6∈P˜i
Ieh(x0) =Ext(Ih|Knd i )(x0) whereIhis the standard Lagrange interpolation operator.
Kind Kid
Ki,1 Ki,2
Ki,3
Ki,4
Ki,5
x0
Proposition (D.-Lleras-Lozinski 2019)
Under Assumption 1, we have for allv∈H2(Ω)∪H01(Ω)
|v−Ieh(v)|1,Ω≤Ch|v|2,Ω. âUse the local interpolation estimate ofIhonPi
Michel Duprez Rule of the mesh in the Finite Element Method 11
Sketch of proof
Proposition (D.-Lleras-Lozinski 2019)
Under Assumption 1, we have for allv∈H2(Ω)∪H01(Ω)
|v−Ieh(v)|1,Ω≤Ch|v|2,Ω.
Lemma (Céa, see Ern-Guermond’s book) Letuanduhsolution to (1) and (2). Then
|u−uh|1,Ω6C inf
vh∈Vh|u−vh|1,Ω.
Thus
|u−uh|1,Ω6C inf
vh∈Vh|u−vh|1,Ω6C|u−Iehu|1,Ω6Ch|u|Ω,2.
12
Poor conditioning of the system matrix
Proposition (D.-Lleras-Lozinski 2019)
Suppose that the meshThcontains a degenerate cellKdeg
ρKdeg=ε, hKdeg>C1h.
Then theconditioning numberassociated to the bilinear formainVhsatisfies
κ(A) :=kAk2kA−1k2> C hε
Idea :
φk φj
ε
1 a(φk,φj) = 1/ε
Michel Duprez Rule of the mesh in the Finite Element Method 13
An alternative scheme
We approximate
Findu∈V s.t. :
a(u, v) =l(v)∀v∈V (1) by the finite element formulation
Finduh∈Vhs.t. :
ah(uh, vh) =l(vh)∀vh∈Vh (2) where the bilinear formahis defined for alluh, vh∈Vhby
ah(uh, vh) :=aΩnd h
(uh, vh)+X
i
aPi(Iehuh,Iehvh)
+X
i
1 h2P
i
((Id−Ieh)uh,(Id−Ieh)vh)Pi,
withΩndh := (∪Pi)c.
14
A priori estimates and conditioning
Theorem (A prioriestimate, D.-Lleras-Lozinski 2019)
LetΩconvex andu∈V, uh∈Vhbe the solutions to Systems (1) and (2).
Then, under Assumption 1, we have (ε= 0ifn= 3) |u−
uh|1,Ω≤Cεh1−ε|u|2,Ω, ku−uhk0,Ω≤Cεh2−ε|u|2,Ω.
Proposition (Conditioning, D.-Lleras-Lozinski 2019)
Suppose that Assumption 1 holds and the union of cellsωxattached to each nodex ofThsatisfies
c1hn≤ |ωx| ≤c2hn. Then, the conditioning number of the matrixAsatisfies
κ(A) :=kAk2kA−1k26Ch−2.
Michel Duprez Rule of the mesh in the Finite Element Method 15
Sketch of proof
Lemma (Coercivity ofah)
ah(vh, vh)>Ckvhk20,Ω ∀vh∈Vh.
Lemma (Continuity ofah)
ah(uh, vh)6(C/h2)kuhk0,Ωkvhk0,Ω ∀uh, vh∈Vh.
kAk2 = sup
v∈RN
(Av,v)
|v|22 = sup
v∈RN
ah(vh, vh)
|v|22 6Chn sup
vh∈Vh
ah(vh, vh)
kvhk20 6Chn−2 kA−1k2= sup
v∈RN
|v|22
(Av,v) = sup
v∈RN
|v|22
ah(vh, vh) 6Ch−n sup
vh∈Vh
kvhk20
ah(vh, vh) 6Ch−n Thus
κ(A) :=kAk2kA−1k26C/h2.
16
Equivalent formulation
Kind
Kideg Fi
Example of patchPi=Kind∪Kideg.
Proposition (D.-Lleras-Lozinski 2019) For alluh, vh∈Vh, it holds
ah(uh, vh) =aΩnd
h (uh, vh) +P
i
|Pi|
|Knd i |aKnd
i (uh, vh) +κnP
i
|Kdegi |3 h2P
i
|Fi|2[∇uh]Fi·[∇vh]Fi
withκn:=(n+1)(n+2)2n2 .
Michel Duprez Rule of the mesh in the Finite Element Method 17
Simulation
LetΩ := (0,1)×(0,1)andf(x, y) = 2π2sin(πx) sin(πy).
Consider the system
Findu∈H1 0(Ω)s.t. :
(∇u,∇v)L2(Ω)= (f, v)L2(Ω)∀v∈H01(Ω).
Exact solution :u(x, y) = sin(πx) sin(πy)∀(x, y)∈Ω.
Cartesian meshesThs.t. :
hKdeg =handρKdeg ∼h2.
18
Simulation
10−3 10−2 10−1
10−5 10−4 10−3 10−2 10−1
1 1
1 2
h
ku−uhk0,Ω
|u−uh|1,Ω
10−2 10−1
10−1 100 101
1 1
h
Conditioning×h12
2 degenerated cells, standard scheme
Michel Duprez Rule of the mesh in the Finite Element Method 19
Simulation
10−3 10−2 10−1
10−5 10−4 10−3 10−2 10−1 100
1 1
1 2
h
ku−uhk0,Ω
|u−uh|1,Ω
|u−Iehuh|1,Ω
10−2 10−1
10−1 100 101
h
Conditioning×h12
2 degenerated cells, alternative scheme
20
Simulation
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2 10−1 100
1 1
1 2
h
ku−uhk0,Ω
|u−uh|1,Ω
10−3 10−2 10−1
10−4 10−3 10−2 10−1 100
h
ku−uhk0,Ω
|u−uh|1,Ω
|u−Iehuh|1,Ω
5.5 % of degenerated cells standard scheme (left), alternative scheme (right)
Michel Duprez Rule of the mesh in the Finite Element Method 21
Work in progress
Application in biomechanics : generation of patient specific meshes
Collaborations A. Lozinski
âLMB, Besançon C. Lobos
âChile M. Bucki
âTexisence, Grenoble V. Lleras
âIMAG, Montpellier
[Bijar-Rohan-Perrier-Payan 2016]
22
Conclusion
• New sufficient geometrical conditions âLocal damages : optimal convergence
• New operator ofinterpolation
âConvergence of a operator of interpol.6=convergence to the solution
• Alternative formulationfor a good conditioning of the systemmatrix âQuasi optimal convergence
• Necessary and sufficientgeometrical conditions remainopen âInclusion of meshes ?
Michel Duprez Rule of the mesh in the Finite Element Method 23
Outline
1
Framework : Finite Element Method (FEM)
2
Geometrical quality of a mesh
3
Mesh which is not conforming to the boundaries
Framework : previous results
Fictitious domain methods : non-matching meshes
Previous results
• First work âSaul’ev 63’
• XFEM
âMoes-Bechet-Tourbier 2006 âHaslinger-Renard 2009
• CutFEM
âBurman-Hansbo 2010-2014 âLozinski 2019
Methods
• Extended finite element space
• Standard penalizations
• Nitsche penalizations
• Ghost penalty
• Lagrange multipliers
Michel Duprez Rule of the mesh in the Finite Element Method 24
Framework : previous results
Lagrange multipliers approximated byP0-FE on the cut trianglesThγ:
Finduh∈Vh, λh∈Wh={µh∈L2(Ωγh) :µh|T∈P0(T)∀T ∈ ThΓ}:
Z
Ω
∇uh· ∇vh+ Z
Γ
λhvh= Z
Ω
f vh ∀vh∈Vh
Z
Γ
µhuh=σh X
E∈Tγ h
[λh][µh] ∀µh∈Wh
âBurman-Hansbo 2010
Difficulty :
The actual formulation contain aintegral on the real boundary.
25
φ-FEM : formal approach
Initial problem
Consider a domainΩand
( Findus.t. :
−∆u=f in Ω u= 0 on ∂Ω
φ-FEM (formal) approach
Assume thatΩandΓare given by alevel-setfunctionφ: Ω :={φ <0}andΓ :={φ= 0}.
Consider the problem
Findvs.t. :
−∆(φv) =f in Ω Thenu:=φvis solution to the initial problem.
Michel Duprez Rule of the mesh in the Finite Element Method 26
φ-FEM : weak formulation
Assume thatΩandΓare given by a level-set functionφ: Ω :={φ <0}andΓ :={φ= 0}.
Suppose thatφis nearΓas the signed distance toΓ.
Consider the finite element approximation Z
Ωh
∇(φvh)· ∇(φwh)− Z
∂Ωh
∂
∂n(φvh)φwh+ Stab.
Term = Z
Ωh
f φwh ∀wh∈Wh,
Thenuh:=vhφas approximation of the initial problem.
27
φ-FEM : a priori error estimates
Consider the finite element approximation :ah(vh, wh) =lh(wh) ∀wh∈Wh, where
ah(vh, wh) = Z
Ωh
∇(φhvh)· ∇(φhwh)− Z
∂Ωh
∂
∂n(φhvh)φhvh+G1h(˜uh, vh) lh(wh) =
Z
Ωh
f φhwh+G2h(wh),
withφh=Ih(φh),G1handG2hstands for the theghost penaltygiven by
G1h(vh, wh)=σh X
E∈FΓ
Z
E
h ∂
∂n(φhvh) i h ∂
∂n(φhwh) i
+σh2P
T∈TΓ h
R
T∆(φhvh)∆(φhwh) G2h(wh)=−σh2 X
T∈TΓ h
Z
T
f∆(φhwh)
and
ThΓ={T ∈ Th:T∩Γh6=∅} (Γh={φh= 0});
FΓ={E(an internal edge ofTh)such that∃T ∈ Th:T∩Γh6=∅andE∈∂T}.
Michel Duprez Rule of the mesh in the Finite Element Method 28
φ-FEM : a priori error estimate
Continuous problem a(u, w) =l(w)∀w∈V
Finite element formulation
ah(vh, wh) =lh(wh)∀wh∈Vh(k)⊂V
Theorem (D.-Lozinski 2019)
Suppose that the meshThis uniform (with some weak assumptions), and
f∈Hk(Ωh∪Ω). Letu∈Hk+2(Ω)be the continuous solution andwh∈Vh(k)be the discret solution. Denotinguh:=φhwh, it holds
|u−uh|1,Ω∩Ωh≤Chkkfkk,Ω∪Ωh
withC=C(φ). Moreover, supposingΩ⊂Ωh
ku−uhk0,Ω≤Chk+1/2kfkk,Ωh.
29
φ-FEM : key lemmas
Lemma (Hardy inequality, D.-Lozinski 2019)
We assume that the domainΩis given by the level-setφregular enough. Then, for anyu∈Hk+1(O)vanishing onΓ,
u φ k,O
≤Ckukk+1,O.
Lemma (Local interpolation, Ern-Guermond’s book) Fork∈N∗, each cellK∈ Thandu∈H2(K)
|u− Ihu|1,K6Chk|u|K,k+1, whereIhis the Lagrange interpolation operator.
Michel Duprez Rule of the mesh in the Finite Element Method 30
φ-FEM : conditioning
Theorem (Conditioning)
Assume thatThis uniform (and satisfies the assumptions). Then the condition numberκ(A) :=kAk2kA−1k2of the matrixAassociated to the bilinear formah
onVh(k)satisfies
κ(A)≤Ch−2.
Here,k · k2stands for the matrix norm associated to the vector 2-norm| · |2.
31
φ-FEM : Simulations
LetΩbe the circle of radius√
2/4centered at the point(0.5,0.5)and the surrounding domainO= (0,1)2. The level-set functionφgiving this domainΩis taken as
φ(x, y) = (x−1/2)2−(y−1/2)2−1/8.
We useφ-FEM to solve numerically Poisson-Dirichlet problem with the exact solution given by
u(x, y) =φ(x, y)×exp(x)×sin(2πy).
φ <0 Ω
0 1
1
Michel Duprez Rule of the mesh in the Finite Element Method 32
φ-FEM : Simulations
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2 10−1 100
1 1
1 2
h
ku−uhk0,Ωh/kuk0,Ωh
|u−uh|1,Ωh/|u|1,Ωh
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2 10−1 100 101
1 1
1 2
h
ku−uhk0,Ωh/kuk0,Ωh
|u−uh|1,Ωh/|u|1,Ωh
FIGURE–Relative errors ofφ-FEM fork= 1. Left :φ-FEM with ghost penaltyσ= 20; Right :φ-FEM without ghost penalty (σ= 0).
33
φ-FEM : Simulations
10−2 10−1
101 102 103 104 105
1 2
h
Conditioning
10−2 10−1
102 103 104 105 106 107
1 4
h
Conditioning
FIGURE–Condition numbers forφ-FEMk= 1. Left :φ-FEM with ghost penaltyσ= 20; Right :φ-FEM without ghost penalty (σ= 0).
Michel Duprez Rule of the mesh in the Finite Element Method 34
φ-FEM : Simulations
10−510−410−310−210−1 100 101 102 103 10−5
10−4 10−3 10−2 10−1 100 101
σ
h= 1.41×10−2 h= 7.07×10−3 h= 3.54×10−3 h= 1.77×10−3
10−510−410−310−210−1 100 101 102 103 10−2
10−1 100 101
σ
h= 1.41×10−2 h= 7.07×10−3 h= 3.54×10−3 h= 1.77×10−3
FIGURE–Influence of the ghost penalty parameterσon the relative errors forφ-FEMk= 1. Left :ku−uhk0,Ωh/kuk0,Ωh; Right :
|u−uh|1,Ωh/|u|1,Ωh.
35
φ-FEM : Simulations
10−3 10−2 10−1
10−10 10−8 10−6 10−4 10−2
1 2
1 3
h
ku−uhk0,Ωh/kuk0,Ωh
|u−uh|1,Ωh/|u|1,Ωh
10−2 10−1
10−10 10−8 10−6 10−4 10−2
1 3
1 4
h
ku−uhk0,Ωh/kuk0,Ωh
|u−uh|1,Ωh/|u|1,Ωh
FIGURE–Relative errors ofφ-FEM. Left :k= 2; Right :k= 3.
Michel Duprez Rule of the mesh in the Finite Element Method 36
φ-FEM : Simulations
10−510−410−310−210−1 100 101 102 103 10−8
10−6 10−4 10−2 100 102
σ
h= 1.41×10−2 h= 7.07×10−3 h= 3.54×10−3 h= 1.77×10−3
10−510−410−310−210−1 100 101 102 103 10−6
10−4 10−2 100 102 104
σ
h= 1.41×10−2 h= 7.07×10−3 h= 3.54×10−3 h= 1.77×10−3
FIGURE–Influence of the ghost penalty parameterσon the relative errors forφ-FEM andk= 2. Left :ku−uhk0,Ω/kuk0,Ω; Right :
|u−uh|1,Ω/|u|1,Ω.
37
φ-FEM : Simulations
We now choose domainΩgiven by the level-set
φ(x, y) =−(y−πx−π)×(y+x/pi−π)×(y−πx+π)×(y+x/pi+π).
It is thus the rectangle with corners 2π2
π2+1,ππ32−π+1
,(0, π),
−π2π2+12 ,−ππ32−π+1
, (0,−π). We useφ-FEM to solve numerically Poisson-Dirichlet problem inΩwith the right-hand side given by
f(x, y) = 1.
Ω
0 4
4
Michel Duprez Rule of the mesh in the Finite Element Method 38
φ-FEM : Simulations
10−2 10−1
10−6 10−5 10−4 10−3 10−2 10−1 100
1 1
1 2
h
kuref−uhk0,Ω/kurefk0,Ω
|uref−uh|1,Ω/|uref|1,Ω
10−2 10−1
10−8 10−7 10−6 10−5 10−4 10−3 10−2
1 2
1 3
h
kuref−uhk0,Ω/kurefk0,Ω
|uref−uh|1,Ω/|uref|1,Ω
FIGURE–Relative errors ofφ-FEM. Left :k= 2; Right :k= 3. The reference solutionuref is computed by a standardF EM on a sufficiently fine fitted mesh onΩ.
39
Conclusion
Results
Optimal convergence of φ-FEM in the H
1semi-norm Quasi-optimal convergence of φ-FEM in the L
2norm
â
Optimal convergence numericallyDiscrete problem well conditioned
Perspectives
Neumann or Robin boundary conditions â
First results with a mixed formulationDynamic equation : heat equation
Michel Duprez Rule of the mesh in the Finite Element Method 40
Thanks for your attention !
φ-FEM : mesh assumptions
Assumption
The approximate boundaryΓhcan be covered by element patches{Πi}i=1,...,NΠ
having the following properties :
Each patchΠiis a connected set composed of a mesh elementTi∈ Th\ ThΓ and some mesh elements cut byΓh. More precisely,Πi=Ti∪ΠΓi with ΠΓi ⊂ ThΓcontaining at mostMmesh elements ;
ThΓ=∪Ni=1ΠΠΓi ;
ΠiandΠjare disjoint ifi6=j.
Michel Duprez Rule of the mesh in the Finite Element Method 41
φ-FEM : mesh assumptions
Assumption
The boundaryΓcan be covered by open setsOi,i= 1, . . . , Iand one can introduce on everyOilocal coordinatesξ1, . . . , ξdwithξd=φsuch that all the partial derivatives∂αξ/∂xαand∂αx/∂ξαup to orderk+ 1are bounded by someC0>0.
Morover,|φ| ≥monO \ ∪i=1,...,IOiwith somem >0.
42