Time Discretisation - Taylor-Galerkin Schemes
V. Selmin
Multidisciplinary Computation and Numerical Simulation
Outline
Outline
• Spatial discretisation: summary
• Basic properties of numerical schemes
• Time discretisation
• Taylor-Galerkin schemes - Basic Taylor-Galerkin schemes - Extension to non-linear problems - Extension to multi-dimensional problems - Two-steps Taylor-Galerkin schemes
• Multi-stages algorithms
Spatial Discretisation
Spatial discretisation-Summary
Structured Grids versus Unstructured Grids
Structured grids:Same number of cells around a node
Unstructured grids:The number of cells around a node is not the same
Spatial Discretisation
Finite Difference Discretisation:
Finite Volume Discretisation:
Taylor-series expansion ),()()()()(
)( 221,1,,1,1, kj
k
kjkj
j
kjkjkj
WFWFWFWFWF
Integral formulationDivergence theorem
eednWWFdWWF ),(),(
Spatial Discretisation
Spatial discretisation-Summary
Reference element Physical element
0)()()( dVfuLdVuRuWV VV
)()(1
r
iii fwdf ξξξ
ξ
Finite Element Discretisation:
Physical space Reference space
Physical element Reference element
Function approximation
Integral method Integration by parts
Weighted residuals Galerkin method
PDE discretisation method
Numerical integration Gauss method
Numerical integration
Basic properties
Basic Properties
Truncation errorDifference between the original partial differential equation (PDE) and the discretised equation (DE).
ConsistencyConsistency deals with the extent to which the discretised equations approximate the partial differential equations. A discretised representation of the PDE is said to be consistent if it can be shown that the difference between the PDE and its discretised representation vanishes as the mesh is refined:
StabilityNumerical stability is a concept applicable in a strict sense only to marching problems.A stable numerical scheme is one for which errors for any source (round-off, truncation, …) are not permitted togrow in the sequence of numerical procedures as the calculation proceeds from one marching step to the next.
Convergence of Marching ProblemsLax’s Equivalence Theorem: Given a properly posed initial value problem and a discretised approximation to it that satisfies the consistency conditions, stability is necessary and sufficient condition for convergence.
),( ht
0)(lim)(lim0,0,
TEDEPDE
thth
DEPDETE
Discretisation in Time
Discretisation in time
Model equation:
Unsteady/steady problems
If the solution u is steady, u solution of
is also solution of the following pseudo-unsteady problem:
Finite Difference:
1- FD approximation of the time derivative
Spatial discretisation →
Time discretisation →
2
2
x
u
x
f
t
u
2
2
x
u
x
f
)(2
2
x
u
x
f
t
ulimt
Steady Euler equations: elliptic for subsonic flowshyperbolic for supersonic flows
j
jjj
x
f
x
uRR
t
u
2
2
,
)(2
43
,)(2
,)(
211
211
1
tt
uuu
t
u
tt
uu
t
u
tt
uu
t
u
nj
nj
njj
nj
njj
nj
njj
),(
),,()(
tnxiuu
tnxuxuni
n
Discretisation in time
2- Taylor-series expansion
Taylor-series expansion of →
Replace the time derivatives by using the equation
That leads to the following equation
which has to be discretised in space
...!
)(1
1
1
n
ll
nllnn
t
u
l
tuu1nu
0
uaf
,)()(
,
2
22
2
2
x
ua
t
u
xa
x
ua
tt
u
x
ua
t
u
nnnn
nn
...)(!
)(1
1
1
n
ll
nll
lnn
x
ua
l
tuu
Taylor series expansion
Discretisation in time
The equality
may be rewritten in the more concise form
A family of temporal schemes may be buit by using the Padé polynomials approximation of the exponential function.It consists to approximate the function H(v) by the ratio of two polynomials of order p and q, respectively, with an
error of
...!
)(1
1
1
n
ll
nllnn
t
u
l
tuu
nn ut
tu )(exp1
Padé Polynomials
)( 1qpv)(
)(
)()( 1 qp
q
p
vvQ
vPVH u
ttv )(
Explicit schemes
Implicit schemes
Discretisation in time
Taylor Galerkin Schemes The Taylor-Galerkin schemes may be considered as a generalisation of the explicit Euler scheme (Padé polynomals with q=0):
The time derivatives are replaced by the expressions obtained by using successive differentiation of the originalequation:
The third order derivative is expressed in terms of a mixed space-time form in order to allow the use of finite element for the spatial discretisation. In this term the time derivative is replaced by a finite difference approximation that maintains the global troncature error:
The time discretised equation is written according to:
where
)(62
43
33
2
221 t
t
ut
t
ut
t
utuu
nnnnn
t
u
xa
t
u
x
ua
t
u nnnn
2
22
3
3
2
22
2
2
,
t
uu
t
u nnn
1
2
221
2)1(
x
uta
x
ua
t
uuL
nnnn
x
36
20
2
222
px
ta
p
Lx
0C
Discretisation in time
Taylor Galerkin Schemes
If the convention is adopted for the scalar product on the computational domain, the Galerkin
equation at node j corresponds to
Explicitley, we got after integration by parts of the second derivatives terms
In the case of piecewise linear shape function, ETG2 and ETG3 schemes take the form
where is the Courant number,
dxwLwL ,
k
nkk
nj
nn UxNUNUUL )(,0),,( 1
2
22222
12
12
1
221
6,
2,
2,
)(
6,)(
6,
0,2
,2
,,
0
0
0
x
taL
x
UaN
t
x
N
x
Ua
tN
x
Ua
x
UUaN
t
x
N
x
UUa
tN
t
UU
Lx
UaN
t
x
N
x
Ua
tN
x
UaN
t
UU
x
x
x
n
jj
n
j
n
x
x
nn
jj
nn
j
nn
x
x
x
n
jj
n
j
n
j
nn
n
n
n
3,2
)1(6
11
2,26
11
22
0122
22
012
TGUc
UcUUc
TGUc
UcUU
nj
nj
nj
nj
nj
nj
nj
nj
x
tac
.2),(2
111
2110 jjjjjjj UUUUUUU
2TG
3TG
Discretisation in time
Taylor Galerkin Schemes
In the right-hand side of the discretised equation we may recognize the same term as the Law-Wendroff scheme
In addition, in the left-hand side of those equations, we may regognize the classical consistent mass of the
finite element theory which corrisponds to to the operator . In the TG3 scheme, it is modified by the
additional term that appears in the time discretised equation.
LW nj
nj
nj
nj U
cUcUU 2
2
01
2
6/1 2
Remarks:Due to the coupling terms, the presence of the mass matrix represent a disadvantage from the point of view of the computational time. Nevertheless, it is possible to exploit its effect in an explicit context.The following iterative procedure may be used
where
)1(1)1(1)(1
)1(1)1(1)2(1
220
)1(1
)()()()(
)()()()(
2
1)(
mnj
nj
nj
nj
mnj
nj
nj
nj
nj
nj
nj
nj
nj
nj
nj
nj
UUIMUUUU
UUIMUUUU
UcUcUU
3)1(6
11
26
11
22
2
ETGc
ETGM
Discretisation in time
Numerical Schemes Property 1- Von Neuman analysis methodThe Von Neumann procedure consits in replacing each term of the discretised equation by the Fourier
component of order k of an harmonic decomposition of :
where is the Fourier component of order k.The amplification factor G is defined by the equality:
In general, it is a complex number which may be written on the following form
where and are respectively the module and the phase of G .The stability condition of von Neuman states that, for each Fourier mode, the amplification factor must have a module limited by a quantity enough close to unity for all value of and .The explicit expression of this criteria is
The term emphasizes that in some physical process, the modes may increase exponentially and this divergence does not be confused with an unstability of the numerical method
njU
njU
)exp( xjkivU nnj
nv
nn vGv 1
ieG
t x
0)(1 tt
)( t
Discretisation in time
Numerical Schemes Property For the previous numerical schemes, the amplification factor takes the form
where and is a real number:
The stability condition for the three schemes is
The reduction of stability for TG2 is due to the consistent mass matrix . The correction contained in the TG3 scheme allows to recover the stability condition and the unit CFL property that states that the signal propagates without distorsions when .
2sin2sin),(11),( 221 ccicLcG x
xk xL
)(32
sin)1(3
2
)(22
sin3
2
0
),(
22
2
TGLWTGc
FELWTG
LW
cLx
)(31
)(23
3
1
TGLWTG
FELWTG
LW
c
1c1c
Discretisation in time
Numerical Schemes Property
Discretisation in time
Numerical Schemes Property In the case the spatial discretisation is performed by maintaining the time continuous, the following schemes are obtained:
for the finite differences, and
for piecewise linear elements
The consistent mass matrix is responsable of the better acurracy on the phase.
FEUt
c
t
U nj
j ,6
11 0
2
FDUt
c
t
U nj
j ,0
)(32
sin)1(3
2
)(22
sin3
2
0
),(
22
2
TGLWTGc
FELWTG
LW
cLx
Discretisation in time
Numerical Schemes Property
2- Modified equation methodThe Modified Equation method consistsa- To perform a Taylor series expansion about of all the terms of the discretised equation.b- To replace all the time derivatives of order greater to one and the mixted space-time by using the equation obtained at the previous stepFollowing this procedure, we obtain the partial differential equation of infinite order genuinely solved by thenumerical scheme
The modified equation may be written according to
where the are real coefficients.
Let consider a elementary solution:
where k is real and is a complex number, the and have to satisfy the following relations:
njU
12
2
20
12
12
12p
p
p
pp
p
p
px
u
x
u
t
u
ikxteetxu ),(
21 i 1 2
012
122
12
21
)1(
,)1(
pp
pp
pp
pp
k
k
Discretisation in time
Numerical Schemes Property
In the limit case where (large wave lenghts), we can negelect all the terms except the non-zero coefficientsof the lowest order which will be denoted by r . In this case
The necessary stability condition: becomes
In addition,
0k
rrr
rrr
k
k
212
2
22
1
)1(
,)1(
01
0)1( 221
rrr k
)1(212
2
)()1(
1
r
r
rr
e
xkc
k
Discretisation in time
Numerical Schemes Property
xh
Discretisation
Time Space
FD
FD
FE
FE
Timecontinuous
Eulerscheme
LW
LW-FE
LW-TG
Discretisation in time
Numerical Schemes Property
Discretisation in time
Propagation of a cosine profile
LW
LW-FE
LW-TG
To illustrate and compare the performance of the schemes discussed so far, consider the advection problem over the interval [0,1] and defined by the following initial and boundary conditions:
00),0(
0
]/)(cos[12
1)0,(
0
00
ttu
xx
xxxxxu
Extension to non linear convection
Discretisation in time
Let consider the following hyperbolic equation
Written in the quasi-linear form
it may be interpreted as a non linear convection equation for which each point of the solution propagates with a different velocity.As in the previous case, the equation is discretised in time by using the series expansion
in which the time derivatives are replaced by using the original equation and its successive differentiation
0)(
x
uf
t
u
u
ufua
x
uua
t
u
)(
)(,0)(
)(62
43
33
2
221 t
t
ut
t
ut
t
utuu
nnnnn
)()()(
)()(
222
22
3
3
22
2
t
u
xa
x
u
t
aa
x
ua
t
a
xt
u
xa
x
u
t
a
xx
ua
xtt
u
x
ua
xt
ua
xt
u
Extension to non linear convection
Discretisation in time
By using the following identities
the expression of the third derivative in time is equivalent to the following form
Then, in the nonlinear case, the equation discretised in time may be written according to
where
x
ua
u
a
t
u
u
a
t
a
t
u
x
a
x
u
t
a
,tx
ua
t
u
x
a
t
ua
xtx
f
xt
ua
x
u
t
a
x
ua
txt
f
22
22
)(
)(
)()( 223
3
t
u
xa
t
u
x
aa
t
u
x
ua
u
a
xt
u
xa
x
u
t
aa
x
ua
t
a
xt
u
x
ua
x
t
x
f
t
uuL
nn
nnn
x2
1
)(2
)1(
TGLW
xa
x
aa
x
ua
u
a
x
t
FELW
L nn
nn
nn
x 22
)(6
0
Remarks:In the case of a scalar equation (and only) the third order time derivative may be written in the following compact form:
t
ua
xt
u 22
2
3
3
The consistent massmatrix depends of theunkown
Extension to multi dimensional problems
Discretisation in time
Let consider the following hyperbolic equation
The time derivatives may be expressed according to
By using the following identities
the expression of the third derivative in time is equivalent to the following form
Then, in the case of multidimensional problems, the equation discretised in time may be written according to
where
0,0
uat
uf
t
u
)()()(
)(
3
3
2
2
t
uaau
t
aaua
t
a
t
u
uaat
u
)(,)( uau
a
t
a
t
uau
t
a
)()()(3
3
t
uaa
t
uaa
t
uua
u
a
t
u
)(2
)1(1
nnnnnn
uaat
ft
uuL
x
TGLWaaaaua
u
at
FELW
L nnnnnnn
)()()(6
02
x
Extension to multi dimensional problems
Discretisation in time
Multidimensional discretisation property
),(),(
,),(),(
),
ykxk
hyxy
ta
x
ta
aa
yxyx
yxyx
yx
λ
ν
(a
x
y
x
y
2.0ν 3.0ν
Domain of numerical stability of the LW schemes Phase velocity error of the LW schemes
Discretisation in time
Convection of a product cosine hill To illustrate and compare the performance of the schemes discussed so far, the advection of a product cosine hill in a pure rotation velocity field is considered. The initial condition is
The unknown has to be prescribed on the inflow boundary and leave free on the outflow boundary
0
000
0
]/)(cos[1]/)(cos[14
1)0,(
xx
xxx
yyxxu
0,0 outinanan
Discretisation in time
Convection of a product cosine hill
LW-FE(Md) LW LW-FE LW-TG
LW-FE(Md) LW LW-FE: unstable LW-TG
Convection of a hill in a rotational field, ∆t=2π/200, after one complete rotation
Convection of a hill in a rotational field, ∆t=2π/120, after one complete rotation
Discretisation in time
Convection of a product cosine hill
Convection of a hill in a rotational field, ∆t=2π/200, after 3/4 and 1 complete rotation
LW-TG
Two step Taylor Galerkin Schemes
Discretisation in time
A third-order T-G scheme which can be effectively employed in non-linear multidimensional advection problems is obtained by considering the following two step procedure:
where the value of the parameter is left unspecified for the time being. In fact while the other coefficients in the two equations assume the value necessary for a third-order accuracy of the two discretisations combined, the parameter enters only the coefficient of fourth-order term in the overall series and therefore its value affects the amplification factor only within the fourth-order accuracy. The fully discrete counterpart of the equations above in the case of linear advection problem in one dimension is
The value of the parameter may be fixed in such a way to reproduce the excellent phase accuracy of the LW-TG scheme:
2
221
2
22
~)(
2
1
)(3
1~
t
ut
t
utuu
t
ut
t
utuu
nnnn
nnnn
nj
nj
nj
nj
nj
nj
nj
nj
Uc
UcUU
UcUc
UU
~26
11
3~
6
11
22
012
220
2
9
1
Two step Taylor Galerkin Schemes
Discretisation in time
The modified equation is
The third order accuracy is preserved by the two-step procedure. Moreover, the new scheme has the same phase response. The amplification factor takes the form
5
542
4
4
4222
)451(180
)3
1(24 x
ucc
xa
x
uctxa
x
ua
t
u
2sin2sin
2sin)1(
3
211),(),( 22
122 cciccRcG realcR ),(
Property analysis
1c2
3cTTG=LW-TG2
Two step Taylor Galerkin Schemes
Discretisation in time
Propagation of a cosine profile
Two step Taylor Galerkin Schemes
Discretisation in time
Multidimensional discretisation property LW-TG
LW-TG2
Amplification factor of the LW schemesDomain of numerical stability of the LW schemes
),(),(
,),(),(
),
ykxk
hyxy
ta
x
ta
aa
yxyx
yxyx
yx
λ
ν
(a
Multiple stages algorithms
Discretisation in time
Another important family of time integration schemes that account for only two time levels are those namedmulti-stage algorithms (generalisation of the Runge-Kutta method).Let write the spatial discretised scheme as follows:
where is the residual of the partial differential equations
The general form of an explicit multi-stage algorithm with l levels may be written according to
where and for consistency.
)(uRdt
du
)(uR
)()1(
)1()0()(
)1(2
)0()2(
)0(1
)0()1(
)0(
ln
ll
l
n
uu
Rtuu
Rtuu
Rtuu
uu
)( )()( kk uRR 1l
4
1,3
1,2
1,1
3
1,2
1,1
2
1,1
1234
323
12
Second order:
Third order:
Fourth order:
Maximum stability condition:
1lc
Hybrid Multi stages algorithms
Discretisation in time
In the case of the solution of steady problems, the objective is to have an efficient integration method which allows to get the steady state as fast as possible while the time accuracy is unimportant.That leads to a class of hybrid schemes of the multi-stage tipe where the residual of a generic l+1 stage is evaluated according to
where and are the convective and dissipative terms, respectively. In order to satisfy the consistency property, the following relations have to be satisfied
l
k
klk
klk
l uDuFR0
)1( )()(
)( kuF )( kuD
100
l
klk
l
klk
Example:
)5()1(
)1()4(5
)0()5(
)1()3(4
)0()4(
)1()2(3
)0()3(
)1()1(2
)0()2(
)0()0(1
)0()1(
)0(
)(
)(
)(
)(
)(
uu
DFtuu
DFtuu
DFtuu
DFtuu
DFtuu
uu
n
n
1
2/1
8/3
6/1
4/1
5
4
3
2
1
1
2/1
324.0
228.0
1375.0
5
4
3
2
1