Numerical methods 1 An Introduction to Numerical Methods For Weather Prediction by Joerg Urban...
-
Upload
bailey-andrews -
Category
Documents
-
view
215 -
download
1
Transcript of Numerical methods 1 An Introduction to Numerical Methods For Weather Prediction by Joerg Urban...
Numerical methods 1
An Introduction to Numerical Methods
For Weather Predictionby
Joerg Urbanoffice 012
Based on lectures given by Mariano Hortal
Numerical methods 2
Shallow water equations in 1 dimension
)(
)(
x
hK
xx
uh
x
hu
t
hx
uK
xx
hg
x
uu
t
u
advection adjustement diffusion
u … velocity along x directionh … absolute heightg … acceleration due to gravityK …diffusion coefficient
Non linear equations
Velocity is equalin layers vertical direction (shallow)
h
Numerical methods 3
Linearization
u=U0+u’h=H +h’
Const. + perturb. in the x-comp. of velocityConst. + perturb. in the height of the free surface
Substitute and drop products of perturbations
''''
''''
2
2
0
2
2
0
hx
Kux
Hhx
Uht
ux
Khx
gux
Uut
Small perturbations
Numerical methods 4
Classification of PDE’s• Boundary value problems( ) ( )
( ) ( )
L f x
B g x
Dx
Dx
D
ΓD
f, g: known function; L, B: differential operatorφ: unknown function of x
0
( ) ( )
t
L f t
g
φ: unknown function of t
t=0 t
• Initial value problems (most important to us)
open domain
boundary
Numerical methods 5
• Initial and boundary value problems
• Eigenvalue problems
Classification of PDE’s (II)
0
( ) ( , )
( )
( ) ( , )t
L f x t
g x
B h t
( )L φ: unknown eigenfunctionλ: eigenvalue of operator L
Numerical methods 6
Existence and uniqueness of solutions
),( ytfdt
dy ; y(t0)=y0 (initial value problem)
•Does it have a solution?•Does it have only one solution?•Do we care?
If it has one and only one solution it is called a well posed problem
Numerical methods 7
Picard’s TheoremLet
y
f
f and be continuous in the rectangle
byy
atttR
0
00: then, the initial-value problem
00 )(
),(
yty
ytfdt
dy Has a unique solution y(t)on the interval attt 00
Finding the solution (not analytical)Numerical methods (finite dimensions)
Numerical methods 8
Discretization
• Finite differences
• Spectral
• Finite elements
Transform the continuous differential equationinto a system of ordinary algebraic equations where the unknowns are the numbers fj
M
jjj xfxf
1
)(~)(
Njxfxf j .....1),(~)(
Numerical methods 9
• Convergence
• Consistency
• Stability
• Lax-Richtmeyer theorem
Discretized equation ---------> continuous equation
The Lax-Richtmeyer theorem
Discretized solution ---------> continuous solution
discretization finer and finer
Discretized solution bounded
If a discretization scheme is consistent and stablethen it is convergent, and vice versa
Numerical methods 10
Finite Differences - Introduction1 j-1 j j+12 N N+1
<------------------------------- L ---------------------------------->
Taylor series expansion:
2 31
2 31
1 1' '' ( ) ''' ( ) ....
2! 3!1 1
' '' ( ) ''' ( ) ....2! 3!
j j j j j
j j j j j
x x x
x x x
1,........, 1j jx x j N
x
Numerical methods 11
Finite differences approximations1
1 1
1 1 2'' ''' ( ) ..2! 3!
' .j jj j jwhere E x xE
x
forward approximation Consistent if … are bounded
12 2
1 1 2'' ''' ( ) ..2! 3!
' .j jj j jwhere E x xE
x
backward approximationadding both
1 1 1 2''' ( ) ..3!
' .2
j jj jwhere E xE
x
centered differences
" "', ,j j
Consistent if … are bounded"' ,j
Numerical methods 12
Finite differences approximations (2)
Also
1 1 2 2 44 1' ( )
3 2 3 4j j j j
j O xx x
1 1 22
2'' ( )
( )j j j
j O xx
fourth order approximation to the first derivative
second order approximation of the second derivative
Using the Taylor expansion again we can easily get the second derivative:
Numerical methods 13
The linear advection equation
00
xU
t00
xU
t+ initial and boundary conditions
We start with a guess: )()(),( tTxXtx
Substituting we get: 00 1
UCdt
dT
Tdx
dX
X
U
TUdt
dT
Xdx
dX
0
Eigenvalue problems for
tU
x
eTT
eXX0
0
0
dt
dand
dx
d
With periodic B.C. λ can only have certain (imaginary) values where k is the wave number ik
The general solution is a linear combination of several wave numbers
Numerical methods 14
The linear advection equation (2)
The analytic solution is then:
)(),( 0)(
0)(
0000 tUxfeeTXtx tUxtUx
Propagating with speed U0
For a single wave of wave number k the frequency is ω=kU0
No dispersion
Energy: L
dxtE0
2
2
1)(
022 0
20
0
20
L
L Udx
x
U
t
EFor periodic B.C.:
Numerical methods 15
Space discretization
xxjj
j
211
dt
d
xU
tjjjj
211
0 ( ) ( )jikj xt t e
0)sin(
0
xk
xkikU
dt
d
whose solution is ikcte 0 with 0
sin( )( )
k xc k U
k x
U0
c
kΔx
The phase speed c depends on k dispersion
kΔx= π ---> λ=2Δx ==> c=0
centered second-order approximation
Try:
results in
π
Numerical methods 16
Group velocity
dk
dcg
)cos()(
)(
0*
00
xkUdk
kcdc
Udk
kUdc
g
g Continuous equation
Discretized equation
=-U0 for kΔx=π
Approximating the space operator introduces dispersion
Numerical methods 17
Time discretization
11 1
0 2
n n n nj j j jj
Ut t x
n
Tryxikj
etninj e 0
Substituting we get 1 sin( )0U ti t i k xx
e
Courant-Friedrich-Levy number
ω=a+ib
If b>0, φjn increases exponentially with time (unstable)
If b<0, φjn decreases exponentially with time (damped)
If b=0, φjn maintains its amplitude with time (neutral)
Also another dispersion is introduced, as we have approximated the operator ∂/ ∂t
In addition to our 2nd order centered approx. for the space derivative we use a 1st order forward approx. for the time derivative:
1 110 2
n nj jn n
j j tUx
Numerical methods 18
Three time level scheme (leapfrog)
)( 1111
j
nj
nnj
nj
)( 1111
j
nj
nnj
nj
This scheme is centered (second order accurate) in both space and time
Try a solution of the form
xikjnk
nj e 0
exponential
If |λk| > 1 solution unstableif |λk| = 1 solution neutralif |λk| < 1 solution damped
Substituting 2 2 1 0 where sin( )k kip p k x
21 pipk 11
11
2
2
pip
pip
k
kΔx--->0Δt --->0
physical mode
computational mode
Numerical methods 19
Stability analysisEnergy method
We have defined ;2
1)(
0
2L
dxtE
L
LUdx
x
U
t
E
00
202
0 0]22
For periodic boundary conditions
We have discretized t and
hence the discretized analog of E(t) is En
N
j
nj
n xE1
2)(2
1φn
N+1≡ φn1
If En=const, than the scheme is stable t
Numerical methods 20
Example of the energy method
xU
t
jnn
jn
jn
j
10
1upwind if U0>0downwind if U0<0 x
j-1 j
1 2 2 11 0( ) ( ) ( )( )n n n n n n
j j j j j j
tU
x
21
21
2221 ))(1(})(){()()( j
nnj
nj
nj
nj
nj
j
0
1
21
1 )()1(
N
jj
nj
nnn xEE
En+1=En ifα=0 => U0=0 no motionα=1 Δt= Δx/U0
En+1 > En unstable
En+1 < En if α > 0 => U0 > 0 upwindα < 1 U0 Δt/ Δx < 1 CFL cond. damped
Numerical methods 21
Von Neumann methodConsider a single wave jikxn
kknj ectx ),(
if |λk| < 1 the scheme is damping for this wave number kif |λk| = 1 k the scheme is neutralif |λk| > 1 for some value of k, the scheme is unstable
alternatively jikxtniknj eectx ),(
if Im(ω) > 0 scheme unstableif Im(ω) = 0 scheme neutralif Im(ω) < 0 scheme damping
Vf= ω/k vg=∂ω/∂k
Numerical methods 22
Stability of some schemes• Forward in time, centered in space (FTCS) scheme
• Upwind or downwind
xU
t
nj
nj
nj
nj
211
0
1
using Von Neumann, we find
1 sin( ) 1k ki k x k scheme unstable
xU
t
nj
nj
nj
nj
10
1 upwind if U0 > 0downwind if U0 < 0
Using Von Neumann, we find2
0
1 (1 ) hence 1 2 ( 1) (1 cos( ))ik xk ke k x
α(α-1) > 0 unstableα < 0 downwindα > 1 CFL limit
-1/4 < α(α-1) < 0 => 0 ≤ α ≤ 1 stable damped scheme
Numerical methods 23
Stability of some schemes (cont)
• Leapfrog xU
t
nj
nj
nj
nj
2211
0
11
Using von Neumann we find |α|≤1 as stability condition
0
tU
x
As a reminder α is the Courant-Friedrich-Levy number:
Numerical methods 24
• Lax Wendroff scheme
Stability of some schemes (cont)
From a Taylor expansion in t we get:
2
22)(
!2
1),(),(
tt
tttxttx
2
220
20 )(
!2
1),(),(
xUt
xtUtxttx
)2(2
)(2 11
2
111 n
jnj
nj
nj
nj
nj
nj
Applying Von Neumann we can find that |α| ≤1 -----> stable
Discretization:
Substitution the advection equation we:
Numerical methods 25
Stability of some schemes (cont)• Implicit centered scheme
1 1 1 1 1 11 1 1 10
2
n n n n n nj j j j j jU
t x x
using von Neumann
1)sin(1
)sin(12
xki
xki
)tan(0 t
t
U
c
We replace the space derivation by the average value of the centred space derivation at time level n-1 and n+1
Always neutral, however an Expensive implicit equation need to solved
Dispersion worse than leapfrog
Numerical methods 26
“Intuitive” look at stabilityIf the information for the future time step “comes from” inside the interval used for the computation of the space derivative, the scheme is stable.Otherwise it is unstable
x, x’ … point where the information comes from (xj-U0Δt) Interval used for the computation of ∂φ/∂xj-1 j j+1
x
U0ΔtDownwind scheme (unstable)
j-1 j j+1x’ x
x if α < 1x’ if α > 1
Upwind scheme (conditionally stable)
CFL number ==> fractionof Δx traveled in Δt seconds
Leapfrog (conditionally stable)
Implicit (unconditionally stable)
j-1 j j+1
x’ x
Numerical methods 27
Dispersion and group velocity
ωΔt
π/2 π
U0
vg vf
Leapfrog
K-N
Implicit
Vf= ω/k vg=∂ω/∂k
Numerical methods 28
Effect of dispersionIn
itial
Lea
pfro
gim
plic
it
Numerical methods 29
Two-dimensional advection equation
000
yV
xU
t
Using von Neumann, assuming a solution of the form)(0 lykxinn e
we obtain
y
ylV
x
xkUt
)sin()sin( 00
using )sin,cos(),( 00 RRVUV
we obtain, for |λ| ≤ 1 the condition
2R
st
where Δs= Δx= Δy
This is more restrictive than in one dimension by a factor 2
Numerical methods 30
0
x
uu
t
u
x
u(x,0)
x
u(x,t)
Change in shape evenfor the continuous form
One Fourier component ukeikx no longer moving with constantspeed but interacting with other componentsFourier decomposition valid at each individual timebut it changes amplitude with time
k
ikxtik eetuu )(
No analytical solution!
Non linear advection equation Continuous form
Numerical methods 31
Energy conservation
Define again: L
dxuE0
2
2
1
0)(2
10
222
Lufdxx
uu
t
E
x
uu
t
u
t
uu
x
uu
t
u
periodic B.C.==
Discretization in space
First attempt:1 1 1 1 2 2
j j j j jj j
u u u u uuu u u
x x t x
2 2 21 1
1 1( )
2 2
'0j j j j j
j
u x u u u ut
E
t
1 1
~ averaging
1 1
3 2j
j j j
u
j jju u uu u u
t x
Second attempt:
2 2 2 21 1 1 1
1( )
'0
6 j j j j j j j jj
u u u uE
u u u ut
terms joined by arrows cancel from consecutive j’s
Numerical methods 32
Aliassing
Consider the productdx
xdxA
)()(
k
kk
k kxxkxx )sin()()sin()( in the interval 0≤x ≤2π
2
1]})sin[(])sin[({
)cos()sin(
21212
221
2
1 2
1
1 2
21
xkkxkkk
xkkxkA
kk k
k
k kkk
Minimum wavelength 2 2m M
m
L Nx k
<---------------L------------->
1 2 n N+1Maximum wave number representable with the discretized grid
Aliasing occurs when the non-linear interactions in the advection term produce a wave which is too short to be represented on the grid.
Numerical methods 33
Aliassing (cont.)Trigonometrical manipulations lead to: sin(kxj)=-sin[(2kM-k)xj]
wave number k wave number 2kM-k
x
x
x
Therefore, it is not possible to distinguish wave numbers k and (2kM-k)on the grid.
Numerical methods 34
Non-linear instabilityIf k1+k2 is misrepresented as k1 there is positive feedback, which causes instability
k1 = 2kM - (k1+k2) ----------> 2k1=2kM-k2
2kM 2k1 kM
2Δx ≤ λ1 ≤ 4ΔxThese wavelengths keep storing energy and total energy is not conserved
We can remove energy from the smallest wavelengths by - Fourier filtering - Smoothing - Diffusion - Use some other discretization (e.g. semi-Lagrangian)