quadrilateral et hexahedral pseudo-conform finite …fdubois/organisation/cnam-20-decembre... ·...
TRANSCRIPT
Quadrilateral et HexahedralPseudo-conform Finite Elements
E. DUBACH, R. LUCE, J.M. THOMASLaboratoire de Mathématiques Appliquées, UMR 5142, Pau (France)GDR MoMas
Méthodes Numériques pour les Fluides. Paris 20 décembre 2006
What is the problem?
Loss of convergence for the RT, BDM, BDFM
when we use quadrilateral or hexahedral elements.
( )
hhhhhh
hhhhh
hhhh
KhhhKhhh
LdxpdxA
Mqdxqdxq
MLp
KPqLqMKRTdivHL
∈∀=∇−
∈∀=+
×∈
∈Ω∈=∈Ω∈=
∫∫
∫∫
ΩΩ
−
ΩΩ
vvvu
u
u
vv
0
0fdiv
of solution, find
)();(,)();,(
:modelStandart
1
0
2
0
uh does not converge in H(div) when the mesh is based on
quadrilateral or hexahedral elements.
−5.5−5−4.5−4−3.5−3−2.5−2−1.5−1−6
−5
−4
−3
−2
−1
0
1
2
log(pas)
log
(err
eu
r)
||u−uh||0,Ω
==> 0.99621 ||p−ph||0,Ω
==> 0.98933 ||div(u−uh)||
0,Ω==> 0.14476
2D
Loss of convergence on div(u)
3D
Loss of convergence on u and div(u)
Where does the problem come from?
element reference
tion triangula theofelement
:ˆ
:
K
K
F is bilinear
21det(DF) PQJ F ∩∈=
F is trilinear
42det(DF) PQJ F ∩∈=
)x(ˆ(x)),x(u)xDF()x(
u(x) ppJ F
== 1
∫∫∫∫ ==∂∂ KKKK
dxpdivpdxdivsdppdsˆˆ
ˆuu,ˆˆn.un.u
Piola transform:
u(x))xM()x(u =
u(x))x()x(u divJdiv F=Non linear relations
ΩΩ≤−
,,uπu
10uChh
ΩΩ≤−
,,uu)π(u10
divCdiv h
ΩΩ≤−
,,uuu10
Chπ
ΩΩ≤−
,1,0)( uuπu divCdiv h
2D 3D
Solutions
- Increase the space of discretisation to control the non-linear part of
the transformation F. (Arnold-Boffi-Falk, 2004 (2D))
2D Parallelogram (RT0) Quadrilateral (ABF0)
Degrees of freedom: 4 Degrees of freedom: 4+2=6
3D Parallelepiped (RT0) Hexahedron
Degrees of freedom: 6 Degrees of freedom: 36!!
By using the same way, we obtain
- « Remesh » the quadrilaterals or hexahedrals
Yu. Kuznetsov and S. Repin (J. Numer. Math 2005)
- Build pseudo-conform finite elements without adding
degrees of freedom
Solutions
We have chosen this last way
Obviously it is a possibility
Let us examine the Q1 conform approximation on quadrilaterals. Why?
Cas 2D
No loss of convergence with the Q1 finite element but the background is
different of the P1 finite element on triangles
Work on is equivalent
to work on T
T
11ˆ PPˆ ∈⇔∈TT
pp
polynomialnot is
QQˆ11ˆ
K
KK
p
pp ∉⇒∈
Consequence: All the integrals must be calculated on K^
xdJvDFpDFdxvp
pDFp
FhK
hhK
h
hh
ˆˆ.ˆ.
ˆ
1
ˆ
1
1
∇∇=∇∇
∇=∇−−
−
∫∫
Goal:
Construct a new finite element satisfying:
• ph is polynomial on K
• Degrees of freedom are the same
• On parallelogram, recover the classical “Q1 finite element”
Price to pay:
• The degree of the polynomials must be increased.
• The approximation is pseudo-conform
Geometry and notations
ralquadrilateconvex a :R2⊂K
[ ] 4,...,1 ,, :edges ,:vertices 1 == + iiiii aaa γ
)(4
1: ofcenter 43210 aaaaa +++=K
)(4
1
)(4
1
by given of basis the),(
43212
43211
2
21
aaaa
aaaa
++−−=
−++−=
e
e
Ree K
KK
KK
ˆ squareunit theof ndeformatioa is
into ed transformis , By. ofpart affine :
#
### FFF
KK#
F#
1ralquadrilateconvex K 21 <+=⇔ ddK
d
),,(Kelement finitenew a of onConstructi ## ,1
#
KKP Σ
41),( #
,1 # ≤≤→=Σ iww iKaWe want
? of Choice #KP
1# QPK
=a
We obtain a non conforming approximation that does not satisfy the patch-test.
In the error estimation, we don’t control the following term:
[ ] σγ
dIn
hh
ii
uuu −
∂∂
∑∫
[ ] 0 that such find :Solutioni
# =∫ σγ
dP hKu
Proposition:
For any convex quadrilateral K there exist a
polynomial ωK# such that for the choice:
),( ## 1 KKPVectP ω=
We have
)4...,1()()(2
)2
element. finitea is ),,(K)1
#
1
#
,1
#
#
#
##
=+=∈∀°
Σ°
+∫ iwwwdPw
P
ii
i
K
KK
i
aa
γ
γσ
uniquenot is :Remark #Kω
#1 havemust WeKPP ⊂
Moreover, it is of interest to obtain:
squareunit theonelement finite reference Lagrange the
)ˆ,,K(),,(Kelement finite the0, limite at the
parameter distorsion theony continousl depend
1,1
###
#
Σ=Σ=•
•
QPKK
K
d
dω
( )
∫ ==
=−=
∩+
#
#
#
#
4,...,10
4,...,11)(
:satisfying in choose tois choicesimplest The
K
1
K
23K
i
id
i
QP
i
i
γ
σω
ω
ω
a
ynumericallor toaccordingy explicitel calculated be can # dK
ω
Remark: We can computed the finite element basis in without
calculating ωK#
23 QP ∩
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mesh in “chevron” Mesh in “alvéole”
))(1)()(1)(1)(1( with
))221(
)221(
)2322322272221(
)1)(1)(1(3(1
),(
2
21
2
21
2
2
2
1
2#
2
#
11
4
2
2
2
2
1
4
1
2
2
2
1
#
2
2#
12
4
2
2
2
2
1
4
1
2
2
2
1
#
2
#
1
2
2
4
1
4
2
2
1
4
2
2
2
2
1
4
1
2
2
2
1
2#
2
2#
1
2
2
2
121
#
2
#
1#
ddddddden
xxddddddd
xxddddddd
xxdddddddddd
xxddddden
xxK
−−+−−−=
+−−−+−+
+−++−−+
−−+++−−+
−−−−−=ω
),,(Kelement finitenew a of onConstructi *
,1
*### KK
P Σ
≤≤→=Σ ∫#
# 41,. want We *
,1
i
idK
γ
σnww
),,( ##
#
00
*
KKPPvectP ψx×=
Let us return to the initial problem
Proposition:
For any convex quadrilateral K there exist a polynomial
vector ψψψψK# such that for the choice:
We have
)4...,1(.1
.
,)2
element. finitea is ),,(K)1
#
#
##
1
*
*
,1
*#
=
=
∈∀∈∀°
Σ°
∫∫ ∫ idsdds
PsP
P
ii ii
K
KK
γγ γ
σσγ
σ nwnw
w
squareunit theonelement finite reference RTthe
)ˆ,,K(),,(Kelement finite the0, limite At the *
11,00,1
*
,1
*### Σ×=Σ= PPP
KKd
( ) ∫∫ ==−==
•→
=•
#
#
#
##
#
##
4,...,10.,1.,0)(
:satisfying in choose tois choiceAnother
))( diagram Rham(De
)( choose tois choicesimplest The
KKK
1K
1
KK
ii
idsddiv
BDM
divHH
i
rot
γγ
σσ
ω
nψnψψ
ψ
rotψ
Remark: We can computed the finite element basis in BDM1without calculated ψψψψK#
)2
,2
,,,0,
0,
1
0,
0
1(
2
2
2
1
−
−
−
=x
xy
xy
x
y
x
y
x
x
yvectBDM
−5.5−5−4.5−4−3.5−3−2.5−2−1.5−1−6
−5
−4
−3
−2
−1
0
1
2
log(pas)
log(
erre
ur)
||u−uh||0,Ω
==> 0.99661 ||p−ph||0,Ω
==> 0.99364 ||div(u−uh)||
0,Ω==> 0.99431
The convergence orders are good
Conclusion
• If the quadrilaterals K are parallelograms we have
)ˆ,,ˆ(),,( 1,1
### Σ=Σ QKPK
KK
and the approximations are conform.
• If the quadrilaterals K are not parallelograms the approximations are not
conform.
The use of allows us to have the continuity of ph at the
nodes and only the continuity of the mean values through each edge of
the mesh.
The use of allows us to have the continuity of the
mean value of uh.n and the continuity of the first momentum of uh.n
through each edge.
• The implementation is simple, but the basis functions depend on of the
shape of K.
)ˆ,,ˆ(),,( *
11,00,1
*
,1
*### Σ×=Σ PPKPK
KK
),,( ## ,1
#
KKPK Σ
)ˆ,,ˆ( *
11,00,1 Σ× PPK
3D case
We apply the same approach with the 3D case.
• Plane faces and non-plane faces? Is it a real problem?
Problems with the hexahedrons
• Parametrisation of the faces
• Description of a hexahedron (deformation of the unit cube)
3D case
K#
K# depends on 6 parameters if the
faces are plane and 9 parameters if
the faces are not plane.
K)
∫ ∫=# ˆ
### ˆˆ)ˆ())ˆ(()(
i i
dvdv i
γ γ
σσ nxMxxx
polynomialnot is ˆ)ˆ( : face plane non
ˆ)ˆ( :face plane 1
i
i P
nxM
nxM ∈
faces) plane of (case ?built How to #KP
),,(element finitenew a of onConstructi ## ,1
#
KKPK Σ
81),( want We #
,1 # ≤≤→=Σ iww iKa
KKV
ivdwiwwΣ
VP
Vdim
V
K
i,K
K
K
K
i
ˆ when on unisolvant is
)6...1,;8...1),(
14)(
thatsuch space polynomiala beLet
#
#
2
1
#
##
#
#
#
=
=→=→=•
⊂•
=•
∫γ σa
1,...,4)j,( pointsat jacobian the
ofmatrix cofacteur theof values theononly depend
, face theof node theof numbers serial are )4,...,1),(( where
6,...,1)()(
#
)(
,
#
4
1
#
)(,
##
1
#
#
#
=
=
==∈∀ ∫ ∑=
jl
j
ii
j
jlj
i
i
i
ii
jjl
ivvdKPv
a
a
γ
γγ
ωγ
ωσ
Building of integration formulas
==∈= ∫ ∑=
6,...,1)(;#
###
4
1
#
)(,
#** ivvdsVvP
i
iij
jljjKK
γγωσ a
Possible choices of polynomial space #KV
23
222222 ,,,,,,,,,,,,,1# QPxzzyyxzyxxyzyzxzxyzyxVK
∩⊂=•
)1)(1)(1()1)(1)(1(
)1)(1)(1()1)(1)(1(
)1)(1)(1()1)(1)(1()(
2222
2222
2222#
1#
yxzyxz
xzyxzy
zyxzyxKQVK
+−−⊕−−−⊕+−−⊕−−−⊕
+−−⊕−−−⊕=•
Proposition:
For any quadrilateral K# not too much deformed
compared to unit cube we have
6,...,1)()2
element. finitea is ),,(K)1
#
##
##
4
1
#
)(,
#
,1
#
==∈∀°
Σ°
∫ ∑=
ivvdPv
P
i
iij
jljK
KK
γγωσ a
Remark: The main difference, when the faces are not plane, concerns the integration formulas:
case plane-non thein calculatedy numericall bemust
case plane thein known explicitly are
,
,
#
#
j
j
i
i
γ
γ
ω
ω
00.2
0.40.6
0.81
0
0.5
10
0.2
0.4
0.6
0.8
1
Non plane faces
Interior faces are not plane
Destructured mesh
What’s about the approximation in Hdiv
),,(Kelement finitenew a of onConstructi *
,1
*### KK
P Σ
≤≤→=Σ ∫#
# 61,. want We *
,1
i
idiK
γ
σnww
?built How to *#K
P
KKV
jidswidΣ
VP
Vdim
V
K
j,K
K
K
K
ii
ˆ when on unisolvant is
)2,1,6...1,.;6...1,.
18)(
thatsuch space polynomiala beLet
#
##*
2
*
0,0,0
*
*
#
###
#
#
#
=
==→=→=•
⊂•
=•
∫∫ γγσσ nunuu
. on jacobian the
ofmatrix cofacteur theof valuesononly depend
6,...,1;2,1..)(
#
,
#
,
##
0,0,0
#
#
##
i
j
ijij
i
i
ii
ijddsKP
γ
ω
σωσ
γ
γγγ ===∈∀ ∫ ∫ nvnvv
Building of integration formulas
===∈= ∫ ∫ 6,...,1;2,1..;#
####
#
,
#** ijddsVP
i
iiijijKK
γγγ σωσ nvnvv
Possible choice of polynomial space 1
*# BDMV
K=
)0,
0
,
0
,
0
2
,0
2
,
0
2,,
0
,0,0
0
,0
0
,
0
0
,
0
0
,
0
0,
0
0,
1
0
0
,
0
1
0
,
0
0
1
( 2
2
2
1
−
−
−
−
−
−
−
−
=yz
xy
yz
xz
xz
xyy
xy
z
xz
xy
x
z
y
x
y
x
z
x
yx
zx
zy
vectBDM
Proposition:
For any quadrilateral K# not too much deformed
compared to unit cube we have
6,...,1;2,1..)2
element. finitea is ),,(K)1
#
###
##
#
,
#*
*
,1
*#
===∈∀°
Σ°
∫ ∫ ijddsP
P
i
iiijijK
KK
γγγ σωσ nvnvv
Problem with the integration formula when the faces are not plane.
Conclusions
• 2D case:
• The numerical results are agree with the theoretical results.• Possibility to generalise the results to FEM of higher degrees.
• 3D case:
• Necessity to work on the geometry of the hexahedrons.
• Clarify the problem of plane faces and non-plane faces.
• Use the De Rham diagram to connect the H1 case to the H(div) case
))()( :diagram Rham(De 1 divHcurlHHrot→→∇