Séminaire MLMS
φ-FEM :
A fictitious domain method for finite element methods on domains defined by level-sets
M. Duprez
1,V. Lleras2andA. Lozinski3,1Inria, Strasbourg, France
2IMAG, Montpellier, France
3Laboratoire de Mathématiques de Besançon, France
Outline
• The "standard" Finite Element Method
• Geometrical constraint on the mesh
• Previous Fictitious Domain Methods
• φ-FEM
• Conclusions and perspectives
Weak formulation
Strong formulationof the Poisson equation
Findu∈H2(Ω)s.t. :
−∆u=f, inΩ u= 0, on∂Ω
H2(Ω)≈differentiable two times
Multiplying by a "test" functionv&by integration by part, one has Weak formulation
( Findu∈H01(Ω)s.t. :
⇔RΩ∇u· ∇v=RΩf v ∀v∈H01(Ω).
H01(Ω)≈differentiable one times, equal to zero on the boundary
Michel Duprez φ-FEM : fictitious domain method with levelset functions 1
Standard FEM formulation
Weak formulation
( Findu∈H01(Ω)s.t. :
⇔RΩ∇u· ∇v=RΩf v ∀v∈H01(Ω).
Standard FEM formulation ( Finduh ∈Vhs.t. :
⇔RΩ∇uh· ∇vh =RΩf vh ∀v∈Vh. whereVh is a subspace ofH01(Ω)offinite dimensional.
φ-FEM : fictitious domain method with levelset functions 2
Standard FEM formulation
Finite element space
LetVh =hψk∈H01(Ω) :k∈ {1, ..., N}i ⊂H01(Ω). Equivalence with a matrix system
( Finduh =PUhkψk ∈Vhs.t. : R
Ω∇uh· ∇ψk=RΩf ψk ∀k ⇔
( FindUh ∈RN s.t. : AhUh =Fh
where
Ah= (RΩ∇ψk· ∇ψj)kj Fh = (RΩf ψk)k
Uh= (Uhk)k Final solution
uh =XUhkψk.
Michel Duprez φ-FEM : fictitious domain method with levelset functions 3
Lagrange continuous finite element
FE space :Vh={cont. piecewise pol. functions on aregular mesh}
Piecewise linear Lagrange FE order 1 (P1)
Order 1 (P1), dimension 2 Order 2 (P2), dimension 1
φ-FEM : fictitious domain method with levelset functions 4
Lagrange continuous finite element : shape function
Why use polynomials ?
because the computation of the coefficientRΩ∇ψk· ∇ψl of the FE matrix is explicit andexact
Important :ifψkis not polynomial, this computation is not exact anymore
Why a mesh ?
because we cannot good approximate a function by a polynomial on the whole domain
Michel Duprez φ-FEM : fictitious domain method with levelset functions 5
Regular mesh : Geometrical condition
• Simplex (triangle, tetrahedron)
The standard FEM works under theCiarlet condition[Ciarlet 78’]
hK
ρK
< γ
hK
ρK
• In general (hexahedron,...)
The standard FEM works if the Jacobian matrix to an reference element is positive.
Important remark :If the mesh contains degenerated cells :
• It is not guaranty that standard FEM converges
• Theconditioning numberof the FE matrix is bad
φ-FEM : fictitious domain method with levelset functions 6
What is the conditioning number ?
Conditioning numberof a matrixA:
Cond(A) :=kAk2kA−1k2
For symmetric definite positive matrices :
Cond(A) := bigest eigen value smallest eigen value>1
Example ifA=
1 0 0 0 1 0 0 0 ε
!
then Cond(A) = 1/ε.
Ifεgoes to zero,Agoes to a non-invertible matrix Conclusion
the conditioning number measures theinvertibility of the matrix.
Michel Duprez φ-FEM : fictitious domain method with levelset functions 7
Poor conditioning of the system matrix
Proposition
• Non-degenerated mesh :Suppose that the mesh satisfies the geometrical conditions, then
Cond(A)6C/h2 wherehis the size of the cells.
• Only one degenerated cell :
Cond(A)>C/hε whereεis the size the degenerated cell.
Idea :
ψk ψk+1
h ε h
1
FE matrix :A= (aij)ij
ak(k+1)=R
Ω∇ψk· ∇ψk+1=−1/ε
φ-FEM : fictitious domain method with levelset functions 8
Complex geometries
What can we do oncomplex geometries?
In particular how to usehexahedral meshes âQuasi-compressible elastic models : locking effect
What is the looking effect ?
• Theoretical continuous quasi-compressible models
"converge" to theoretical continuous incompressible models
• False for numerical models with tetrahedrons ! â not enough degree of freedom
Michel Duprez φ-FEM : fictitious domain method with levelset functions 9
Complex geometries
Alternative :Fictitious domain methods
⇒
Advantages
• No need to mesh
• Regular cells
Difficulties
• Adapt theweak formulation
• Conditioningof the FE matrix
φ-FEM : fictitious domain method with levelset functions 10
Outline
• The "standard" Finite Element Method
• Geometrical constraint on the mesh
• Previous Fictitious domain Methods
• φ-FEM
• Conclusions and perspectives
Previous works, first works
References
Saul’ev 63’(Dirichlet), Astrakhantsev 78’(Neumann), Glowinski 92’(first proof)
Formal idea
Total energy = (energy onΩ) +ε×(energy onΩc) (ε <<1)
⇒
Advantage
Non-conform mesh (complex and time varying geometry) Difficulty
Discontinuity,large FE matrixandbad cond. number
φ-FEM : fictitious domain method with levelset functions 11
Previous works, XFEM : cut shape functions
References
Moes-Bechet-Tourbier 2006, Haslinger-Renard 2009 Formal idea
Cut shape function :ψk−→ψk1Ω
Boundary condition onΓ: Lagrange multiplier
Conditioning of the matrix : stabilization on the boundary
⇒
Advantage Small FE matrix
Good conditioning number Difficulty
Non-classical shape functionsanddiscontinuityin the integrals
Michel Duprez φ-FEM : fictitious domain method with levelset functions 12
Previous works, Cut FEM : partial integration on the cells
References
Burman et al 2010-2014 Formal idea
Partial integration on the cell near the boundary
Lagrange multiplier or penalization for the boundary conditions Conditioning of the matrix : stabilization on the boundary
⇒
Advantage
Standard shape functions
Difficulty
• Integral on the real boundary, cut integral.
• Pkformulation
φ-FEM : fictitious domain method with levelset functions 13
Previous work : Shifted Bound. Method : Taylor expansion
Formal idea (Dirichlet)
Taylor expansionof the bound. cond.
Reference
Main-Scovazzi 2017 Real boundary condition onΓ:u =uD
Discrete bound. cond. onΓe:u+∇u·d=uD
Advantage No cut integral
Difficulty
• Construction ofd
• Pk,Neumann conditions
Michel Duprez φ-FEM : fictitious domain method with levelset functions 14
Outline
• The "standard" Finite Element Method
• Geometrical constraint on the mesh
• Previous Fictitious domain Methods
• φ-FEM
• Conclusions and perspectives
What is the idea of φ-FEM ?
Hypothesis :
Assume thatΩandΓare given by alevel-setfunctionφ: Ω :={φ <0}andΓ :={φ= 0}.
Idea ofφ-FEM :
Include the Level-set function in the formulation to take into account the boundary conditions
Michel Duprez φ-FEM : fictitious domain method with levelset functions 15
Formal approach
ϕFEM for the Poisson Dirichlet Problem
∃?ut.q.
−∆u=f u= 0au bord
−→
∃?vt.q.−∆(ϕv) =f oùΩ ={ϕ <0}, u=ϕv
φ-FEM : fictitious domain method with levelset functions 16
φ-FEM : weak formulation
Hypothesis :ΩandΓare given by a level-set functionφ: Ω :={φ <0}andΓ :={φ= 0}.
φ-FEM formulation : Findvhs.t.
Z
Ωh
∇(φvh)·∇(φwh)−
Z
∂Ωh
∂
∂n(φvh)φwh+ Stab.
Term = Z
Ωh
f φwh ∀wh∈Wh, whereWh ={cont. piecewise pol. functions on a the mesh}.
Thenuh:=vhφas approximation of the initial problem.
Michel Duprez φ-FEM : fictitious domain method with levelset functions 17
φ-FEM : weak formulation
Stabilization terms(to ensure a good conditioning number)
σh X
E∈FΓ
Z
E
∂
∂n(φhvh) ∂
∂n(φhwh)
+σh2PT∈TΓ
h
R
T(∆(φhvh) +f)∆(φhwh) where[·]is the jump on the interfaceE
( ThΓ={T ∈ Th :T∩Γh 6=∅} (Γh={φh= 0});
FΓ ={E(an internal edge ofTh)such that∃T ∈ Th : T∩Γh 6=∅andE∈∂T}.
φ-FEM : fictitious domain method with levelset functions 18
φ-FEM : a priori error estimate
Theorem (D.-Lozinski 2020, optimal error)
Letube the continuous solution andvhtheφ-FEM solution Deformation error:|u−uh|1,Ω≤Chkkfkk,Ωh
Displacement error:ku−uhk0,Ω≤Chk+1/2kfkk,Ωh withC=C(φ)andkthe polynomial order.
Proposition (Optimal conditioning)
The finite element matrix ofφ-FEM satisfies Cond(A)≤Ch−2.
Michel Duprez φ-FEM : fictitious domain method with levelset functions 19
φ-FEM : Simulation of the Poisson-Dirichlet problem
DomainΩ: circle of radius√
2/4centered at(0.5,0.5) Surrounding domain:O = (0,1)2
Level-set function:φ(x, y) = (x−1/2)2−(y−1/2)2−1/8 Exact solution:u(x, y) = exp(x)×sin(2πy)
Artificial external force :f :=−∆u Boundary :uD :=u(1 +φ)
Remark(non-homogeneous case) Ifu:=uD onΓ,
we replaceφhvh in the scheme byφhvh+uD onΩΓh.
φ < 0 Ω
0 1
1
φ-FEM : fictitious domain method with levelset functions 20
φ-FEM : Simulations in P
110−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2
1 2
h
standard. FEM on the ext. domain standard-FEM
Φ-FEM
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2 10−1
1 1
h
standard. FEM on the ext. domain standard-FEM
Φ-FEM
L2relative error≈displacement error H1relative error≈deformation error
standard FEM on the extended standard FEM φ-FEM
domain with a perturbation ( Dirichlet :(1 +φ)u)
Michel Duprez φ-FEM : fictitious domain method with levelset functions 21
φ-FEM : Poisson-Dirichlet problem
Explanation of the numerical results
• Why better than standard FEM ?
âStandard FEM : polygonal approximation of the boundary â φ-FEM : better approx. of the bound. with a levelset
• Numerical cost ofφ-FEM :
â Good point : small size of the FEM matrix â Bad point : quadrature more expensive
→integral of polynomial product
standard FEM φ-FEM
φ-FEM : fictitious domain method with levelset functions 22
φ-FEM : Stabilization and conditioning
φ-FEM inP1with stabilization φ-FEM inP1without stabilization
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2
1 1
1 2
h H1 rel. error (deformation) L2 rel. error (displacement)
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2
1 1
1 2
h
H1 rel. error (deformation) L2 rel. error (displacement)
10−2 10−1
102 103 104 105
1 2
h
Conditioning
10−2 10−1
102 104 106 108 1010 1012
1 10
h
Conditioning
Michel Duprez φ-FEM : fictitious domain method with levelset functions 23
φ-FEM : high polynomial order
Remarque :φ-FEM works high polynomial orders
P2 P3
10−2 10−1
10−10 10−8 10−6 10−4 10−2
1 2
1 3
h
L2 rel. error (displacement) H1 rel. error (deformation)
10−2 10−1
10−10 10−8 10−6 10−4 10−2
1 3
1 4
h
L2 rel. error (displacement) H1 rel. error (deformation)
φ-FEM : fictitious domain method with levelset functions 24
Neumann boundary condition
Consider thePoisson Neumannproblem : ( −∆u+u=f, inΩ
∇u·n=g, on∂Ω
If φ= 0 and ∇φ=nonΓ,
the last system is (formally) equivalent to ( y=−∇u, inThΓ
y· ∇φ=pφ+g, inThΓ ≈> ∇u·n=g onΓ
Michel Duprez φ-FEM : fictitious domain method with levelset functions 25
Neumann boundary condition : weak formualtion
Find(uh, yh, ph)∈Wh(k)such that for all(vh, zh, qh)∈Wh(k) Z
Ωh
∇uh· ∇vh+ Z
Ωh
uhvh+ Z
∂Ωh
yh·nvh
+γu
Z
ΩΓh
(yh+∇uh)·(zh+∇vh) +γp
h2 Z
ΩΓh
(y· ∇φh+1
hphφh)(z· ∇φh+ 1 hqhφh) +σh
Z
Γih
[∂nu] [∂nvh]+γdiv Z
ΩΓh
(divyh+uh)(divzh+vh)
= Z
Ωh
f vh+γp h2
Z
ΩΓh
g(z· ∇φh+ 1
hqhφh)+γdiv Z
ΩΓh
f(divzh+vh),
whereϕhis the Lagrange interpolation ofφof orderland Wh ={(uh, yh, ph)∈C0(Ωh)×C0(ΩΓh)d×L2(ΩΓh) :
(uh, yh, ph)K ∈Pk×(Pk)d×Pk−1}
φ-FEM : fictitious domain method with levelset functions 26
Neumann : optimal convergence
Theorem (D.-Lleras-Lozinski 2021, optimal error)
Letube the continuous solution anduh theφ-FEM solution Deformation error:|u−uh|1,Ω≤Chk(kfkk,Ωh+kgkk+1,ΩΓ
h) Displacement error:ku−uhk0,Ω ≤Chk+1/2(kfkk,Ωh+kgkk+1,ΩΓ
h) withC=C(φ)andkthe polynomial order.
Proposition (Optimal conditioning)
The finite element matrix ofφ-FEM satisfies Cond(A)≤Ch−2.
Michel Duprez φ-FEM : fictitious domain method with levelset functions 27
φ-FEM Neumann Poisson
• Level setfunction : distance to the boundary
•Exact solution:u(x, y) := sin(x) exp(y)
• Sourceterm :f:=−∆u+u
• ExtrapolatedNeumannboundary condition :
˜
g=∇u·∇φ|∇φ| +uφ
10−3 10−2 10−1
10−6 10−5 10−4 10−3 10−2 10−1
1 1
1 2
h
ku−uhk0,Ωh/kuk0,Ωh
|u−uh|1,Ωh/|u|1,Ωh
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
FIGURE–Left :P1; Right :P2
φ-FEM : fictitious domain method with levelset functions 28
φ-FEM Neumann Poisson - Sensibility to cut cells
0 0.2 0.4 0.6 0.8
10−4 10−3 10−2 10−1
θ0
h 0.217 0.109 0.0543 0.0271 0.0136
0 0.2 0.4 0.6 0.8
10−2 10−1
θ0
h 0.217 0.109 0.0544 0.0272 0.0136
FIGURE–Sensitivity to the rotation inφ-FEM,k= 1andl= 3. Left :L2 relative error ; Right :H1relative error.
Michel Duprez φ-FEM : fictitious domain method with levelset functions 29
φ-FEM : Linear elasticity equation - Dirichlet
Model
( −div(σ(u)) =f, inΩ, u=uD, onΓ φ-FEM model
−div(σ(φv+uD)) =f, inΩ,
10−2.5 10−2 10−1.5 10−8
10−7 10−6 10−5 10−4 10−3 10−2
1 1
1 2
h
L2 rel. errorΦ-FEM H1 rel. errorΦ-FEM
[Duprez-Roussel]
φ-FEM : fictitious domain method with levelset functions 30
Conclusion
Results
For the Poisson Dirichlet/Neumann problem : Optimal convergence
Discrete problem well conditioned
Simple implementation : standard shape functions Formulation available for any order of approximation φ-FEM works for the elasticity equations
Work in progress Dynamic system Fluid structure
Implementation of elastic models in Sofa
Michel Duprez φ-FEM : fictitious domain method with levelset functions 31