Академический Документы
Профессиональный Документы
Культура Документы
Kai Schneider
M2P2-CNRS & CMI, Universit de Provence, Marseille, France
Outline
Motivation
Introduction
Adaptivity in space and time
Adaptive multiresolution method
Local time stepping / Controlled time stepping
Applications to reaction-diffusion equations
Applications to compressible Euler and Navier-Stokes
Conclusions and perspectives
Motivation
Context: Systems of nonlinear partial differential equations (PDEs) of hyperbolic or parabolic type.
Turbulent reactive or non-reactive flows exhibit a multitude of active spatial and temporal scales.
Scales are mostly not uniformly distributed in the space-time domain,
Efficient numerical discretizations could take advantage of this property -> adaptivity in space and time
Reduction of the computational complexity with respect to uniform discretizations
while controlling the accuracy of the adaptive discretization.
Introduction
l,i
U
0i<2l
l = Ul,i
U
U dV
l,i
Z
1
=
|l,i |
l,i
= (l,i)0i<2l 0 l L
+
%,
)
)
%,
)
+
)
(
(
Pll1 Pll+1 =
!
(
),
)
,
#
:U
L 7 (U
0 , D1, . . . , DL)
M
*
"!
+
(+
l (U
l1, Dl )
U
N 1
P
l,i U
l,i
Dl,i = U
Pll1
Pll+1
l1 = Pll1 U
l
U
l+1 = Pll+1 U
l
U
|Dl,i| < l
Dl,i
l = (
ul,i)0lL, il
U
B B D
B B
B
D
B
B
B
B
D
B
B
B
A
A
A
A
A
A
A
A
A
A
A
A
A
A
A
@
@
@
@
@
@
@
@
@
HH
HH
HH
HH
HH
HH
X
XX
XXX
XXX
XX
X
4
7
<;
<=
46
<
"2
/
.
t U = D(U ) U = (, ~v , e)t
2nd
"
/
.
/
(
2
2
l,i = D
l,i
(l, i) t U
l,i
U
1
=
|l,i |
U dV
l,i
l,i
D
1
:=
|l,i|
1
D dV =
|l,i|
l,i
i
(f (U ) + (U, U )) nl,i ds + S
l,i
2.
%
%
.
O(N log N ) N
E(t)
T()
1 T() M
E(t) U
n+1 = M
n
U
/
.
1
l = 2( 2 1)(lL) L
d
l,i||H1 = 1
||
l = 2 2 (lL) L
d
l,i||L2 = 1
||
l = 2d(lL) L
l,i||L1 = 1
||
l,i
l,i >
< U,
Dl,i
1st. stage
2nd. stage
update
t
1st. stage
2nd. stage
update
evolution (expensive)
interpolation (cheap)
return to the stored value (no cost)
MR/CTS/LTS scheme
Combination of MR, CTS and LTS strategies:
1. MR/CTS is applied to determine the time step t required to
attain a specified accuracy with a global time stepping;
2. the MR/LTS cycle is computed using the obtained step size
t for the evolution of the cell averages on the finest scale;
3. another MR/CTS time step is then done to adjust the next
time step, and so on.
Numerical validation
Error analysis
Stability
Convection-diffusion equation: tu + xu =
1 2
u,
P e xx
x2
L
t
,
x
2
4P e1 + x
(7)
||
uL
L
uL
L
uL
L
ex u
M R || ||
ex u
F V || + ||
FV u
M R ||
(8)
Accuracy
L
Discretization error: ||
uL
L
ex u
F V || 2
Perturbation error: ||
uL
L
FV u
M R || n =
T
(Cohen et al 2002)
t
(9)
Numerical validation
% CPU time compression
MR
100
% Memory compression
80
80
60
60
40
40
20
20
7 8 9 10 11 12 13
L
L -error
1e+01
1e+00
FV
MR
O(x2 )
MR
100
7 8 9 10 11 12 13
L
1
L -error
1e+00
1e-01
1e-01
1e-02
1e-02
1e-03
1e-03
1e-04
1e-04
1e-05
7 8 9 10 11 12 13
L
FV
MR
O(x2 )
7 8 9 10 11 12 13
L
Numerical validation
Viscous Burgers equation: t u + x
Analogously, we set
u2
2
1 2
u
Re xx
2(+1)L
=C
, C>0
L+2
Re + 2
100
MR
100
MR
(10)
1e+00
1e-01
80
80
60
60
1e-03
40
40
1e-04
20
20
1e-05
7 8 9 10 11 12 13
L
1e-02
7 8 9 10 11 12 13
L
1e-06
% Memory compression
FV
MR
O(x2 )
7 8 9 10 11 12 13
L
L1 -error
total
coherent
||
incoherent
||max
(M. Farge and K. Schneider. Flow, Turbulence and Combustion, 66, 2001.).
A decomposition of the conservative variables into coherent and incoherent components is then obtained by decomposing the conservative
variables into wavelet coefficients, applying a thresholding and reconstructing the coherent and incoherent contributions from the strong
and weak coefficients, respectively.
(5)
where C and pC respectively denote the coherent part of the density and pressure fields, while I and pI denote the corresponding
incoherent parts.
Velocity u1, u2, u3, temperature T and energy e, are decomposed
using the Favre averaging technique, i.e. density weighted.
For a quantity we obtain,
()C
(6)
= C + I , where C =
()C
Finally, retaining only the coherent contributions of the conservative
variables we obtain the filtered compressible Navier-Stokes equations
which describ e the flow evolution of the coherent flow UC . The
influence of the incoherent contributions UI is in the current approach
completely negleted.
=
t
xj
( ui ) =
t
xj
e
=
( )
t
xj
uj
ui uj + p i,j i,j
( e + p) uj ui i,j
T
xj
uj
2 uk
ui
+
i,j ,
i,j =
Re xj
xi
3 xk
where denotes the dimensionless molecular viscosity and Re the
Reynolds numb er. The dimensionless conductivity is defined by
,
=
2
( 1) M a Re P r
t
x
y
z
where U = (, u1, u2, u3, e)T denotes the vector of the conservative quantities, and F , G, H are the flux vectors in the directions
x, y, and z, respectively.
n
n
n
7 Fl,i,j,k + 8 Fl,i+1,j,k Fl,i+2,j,k
l,i,j,k = U
l,i,j,k + t
U
6x
n
n
n
7 G
+
8
G
G
l,i,j,k
l,1,j+1,k
l,i,j+2,k
+ t
6y
+ t
n
n
n
7 H
+
8
H
H
l,i,j,k
l,1,j,k+1
l,i,j,k+2
6z
n+1 =
U
l,i,j,k
U
+
U
l,i,j,k
l,i,j,k
2
+
+
+
t
2
n
n
n
7 Fl,i,j,k
+ 8 Fl,i1,j,k
Fl,i2,j,k
6x
6y
6z
n
n
n
t 7 G
l,i,j,k + 8 Gl,1,j1,k Gl,i,j2,k
n
n
n
t 7 H
l,i,j,k + 8 Hl,1,j,k1 Hl,i,j,k2
%
2
.
N = 1283
/
2
Re = 50
M a = 0.3 P r = 0.7
= [30, 30]3
t = 80
2
.
./
Numerical results
As test case for the CVS method we consider in the following a three-dimensional
weakly compressible temporally developing turbulent mixing layer, see [22].
The CVS results are compared with a DNS reference computation, using a
finite volume scheme on the finest regular grid. We study the impact of the
different thresholding rules, the choice of the threshold and the influence of
the Reynolds number to assess the precision and efficiency of CVS.
We initialize the test-case by setting two layers of a fluid stacked one upon the
other one, each of them with the same velocity norm but opposed directions.
Uo
Lz
Ly
Lx
Fig. 2. Flow configuration: domain and initial basic flow u0 of the three-dimensional
mixing layer.
<
3
8
;
5
B
7
5
6*
;
2-
@8C
E
C
5
<< A
63
8
B
8
34C
7C
5
<2C7
3
-C
3Z
y=0
0.25
A
0.5
Slices of vorticity at y = 0
Adaptive grid
DNS
CVS Epsilon=0.3
CVS Epsilon=0.25
CVS Epsilon=0.2
CVS Epsilon=0.1
140000
120000
DNS
CVS Epsilon=0.08
CVS Epsilon=0.07
140000
100000
100000
80000
80000
80000
100000
120000
120000
60000
60000
60000
40000
40000
40000
20000
20000
20000
10
20
30
40 50
60 70 80
10 20 30 40 50 60 70 80
6@
>g
500
80
6*
500
70
500
60
6@
2000
1000
50
C
0
2-
2000
1000
40
2500
1000
30
65
2500
1500
20
DNS
CVS Epsilon=0.06
CVS Epsilon=0.05
CVS Epsilon=0.04
CVS Epsilon=0.03
3000
1500
10
H1
3500
DNS
CVS Epsilon=0.08
CVS Epsilon=0.07
1500
5
B
2*
C
0
L2
6@
28
<
2 A
65
2B
25
3,
>8
28C
3
5
4<C
-@
L1
3000
2000
10 20 30 40 50 60 70 80
t
3500
DNS
CVS Epsilon=0.3
CVS Epsilon=0.25
CVS Epsilon=0.2
CVS Epsilon=0.1
2500
3500
3000
DNS
CVS Epsilon=0.06
CVS Epsilon=0.05
CVS Epsilon=0.04
CVS Epsilon=0.03
140000
10
20
30
40
50
60
70
80
10
20
6@
Z>
40
50
60
70
80
>g
6*
H1
3
6@
C
0
2-
65
L2
5
B
2*
6@
C
0
28 -
<
2 A
68
25
>8
28C
3
5
4<C
@-
L1
30
1e+02
1e+03
DNS
CVS Epsilon = 0.08
1e+02
1e+01
1e+00
1e+00
1e+00
1e-01
1e-01
1e-01
1e+01
1e-02
1e-02
1e-02
1e-03
1e-03
1e-03
1e-04
1e-04
1e-04
1e-05
1e-05
DNS
CVS Epsilon = 0.03
1e+02
1e+01
1e+03
DNS
CVS Epsilon = 0.2
1e-05
10
10
10
>g
6*
6@
H1
C
0
2-
65
5
B
2*
6@
L2
C
0
28 -
<
=
Z
A
L1
6@
2
68
B
65
W2
t = 80
VX
>
Y
>
i
%
i
>`
O_
H1
L2
L1
!
M a = 0.3 P r = 0.7
= [60, 60]3
"
%
4
N = 2563
Re = 200
t = 80
!
1283
0 !
01230 4
* ,
56
35265 4
27
Fig. 17. Time evolution of a weakly compressible mixing layer at resolution N = 2563
in the quasi-2D regime. CVS computation with = 0.03 and norm #3. First row:
Two-dimensional cut of vorticity at y = 0, 10 isolines of vorticity between 0.1 and
1. Second row: Corresponding isosurfaces of vorticity |||| = 0.5 (black) and ||||
= 0.25 (gray). Third Row: Corresponding adaptive mesh of the CVS computation.
The corresponding time instants are t = 19 (left), t = 37 (center) and t = 78 (right).
1e+03
10000
870000
CVS
1e+02
CVS
1e+01
CVS
860000
9000
850000
8000
1e-01
1e+00
1e-02
1e-03
7000
840000
1e-04
1e-05
830000
10
100
k
6000
0 10 20 30 40 50 60 70 80
t
0 10 20 30 40 50 60 70 80
t
Fig. 18. Energy spectra in the streamwise direction at t = 80 (left). Time evolution
of the kinetic energy (center) and enstrophy (right) for the CVS computations at
Re = 200, N = 2563 .
25
Conclusions (CVS I)
Adaptive multiresolution method to solve the three-dimensional compressible NavierStokes equations in a Cartesian geometry.
Extension of the Coherent Vortex Simulation approach to compressible flows.
Time evolution of the coherent flow contributions computed efficiently using the adaptive multiresolution method.
Generic test case: weakly compressible turbulent mixing layers.
Different thresholding rules, i.e. L1, L2 and H 1 norms.
H 1 based threshold yields the b est results in terms of accuracy and
efficiency.
Drawbacks:
Explicit time discretization, imp oses a time step limitation due to
stability reasons, i.e. the smallest spatial scale dictates the actual
size of the time step ( local time stepping strategies).
Using lo cal time stepping the time step on larger scales can b e increased without violating the stability criterion of the explicit time
integration (further speed up).
Generalisation to complex geometries: volume p enalization approach
(cf. Angot et al. 1999, Schneider & Farge 2005).
http://www.cmi.univ-mrs.fr/~kschneid
http://wavelets.ens.fr
Q F
+
= 0, with Q = ~v
t ~r
e
and
where t is time,
p
~r is 2D position vector with |~r |= (x 2 + y 2 ),
= (~r , t) density,
~v = ~v (~r , t) velocity with components (v1 , v2 ),
e = e(~r , t) energy per unit of mass and
p = p(~r , t) pressure.
~v
F = u 2 + p
(e + p)~v
(r , 0) =
e(r , 0) =
1
0.125
if r r0
if r > r0 ,
2.5
0.25
if r r0
if r > r0 ,
Density
Grid
Ref.: Domingues et al., ANM, 2009, in press
adaptive schemes, Table 5 shows that the MR/CTS/LTS-RK2(3) method is 1.5 times faster than the MR-RK3 method,
and 8.33 times faster than the FV-RK3 computation, which nicely illustrates the additional speed-up of adaptive and
local time stepping.
Comparison for the numerical solutions of the 2D Euler equations for t = 0.5 with L = 10 and = 2 103 .
Method
Error
E
(%)
0.60
0.67
1.09
0.66
0.59
0.66
0.63
CPU
Time
(103 sec)
45
10
9
8
65
12
9
Memory
(%)
100
18
16
18
100
18
18
(%)
100
23
19
18
100
18
14
In Fig. 16, we plot the evolution of the time step, together with the CFL number. We observe that all the methods
with a fixed time step and the MR/CTS-RK2(3) method guarantee the condition CFL 0.5. However, the MR/CTSRK2(3) scheme forces the time step to decrease in the region around t = 0.1, where max is larger. After t = 0.2, for
smaller values of max , the time step increases again. This behavior illustrates the ability of CTS methods to adapt the
step size according to needs of the numerical solution.
For the MR/CTS/LTS-RK2(3) method, the CFL locally reaches 0.65, since the time step on the finest grid can
only be modified at the end of a time cycle (see Fig. 16, left). Nevertheless, this fact does not affect much the quality of
the solution, since the error on the kinetic energy is roughly the same as the one obtained using the MR-RK3 method.
1.0e-03
0.7
-4
t = 6.10
MR/CTS RK2(3)
MR/CTS/LTS RK2(3)
9.0e-04
-4
t = 6.10
MR/CTS RK2(3)
MR/CTS/LTS RK2(3)
CFL=0.5
0.6
8.0e-04
CFL
0.5
7.0e-04
0.4
6.0e-04
0.3
5.0e-04
4.0e-04
0.2
0.1
0.2
0.3
t
0.4
0.1
0.2
0.3
0.4
Figure 16: 2D Euler equations: time evolution of the time step t (left) and the CFL number (right) for the different adaptive methods.
5. Conclusion
In the present paper, different time stepping strategies for space adaptive MR methods for hyperbolic conservation
laws were investigated and applied to the compressible Euler equations in one and two space dimensions. We have
compared explicit time stepping using a fixed time step with a level-dependent time-stepping MR/LTS method, where
21
1.00
1.00
-0.75
-0.50
Domain position
2
3
2.00
1.00
-0.75
0.50
1.00
1.00
0.75
0.50
y
4
3.00
1.00
0.75
-0.50
4
x
Density
Grid
AMR simulation
Uniform
x = 1/1024
r1,2,3 = 2, 2, 2
x = 1/1024
Reference solution
x = 1/4096
Level
L=8
L=9
L=10
Level
L=8
L=9
L=10
Le1 ()
[102 ]
4.13
2.79
1.84
Overhead
per it. cell
0.58
0.52
0.63
MR
Grid Compression
(%)
24.98
13.23
6.58
Overhead
per it. (%)
14.6
6.8
4.2
Le1 ()
[102 ]
4.00
2.66
1.57
Overhead
per it. cell
0.13
0.03
0.12
AMR
Grid Compression
(%)
68.2
44.4
26.2
Overhead
per it. (%)
8.7
1.3
3.1
12.5
Q(~r , t = 0) =
~0 otherwise.
2.5
The computational domain is = [0, 1] [0, 1] [0, 1].
The computations are performed until t = 0.84.
The physical parameters are Ma = 1 and = 1.4.
MR computations
Numerical Parameters: L = 7, = 0.001, RK2 scheme, MUSCL
AUSM+up flux, CFL = 0.8.
Evolution of density at
t = 0.042, 0.084, 0.126, 0.210, 0.252, 0.294, 0.336, 0.378, 0.420
(from left to right and top to bottom).
AMR computations
Numerical Parameters: 2 levels with refinement factor 2 are used,
finest level: 120 120 120 grid (1.73 M cells), coarse grid of
30 30 30 cells, minmod-limiter, CFL = 0.8, until physical time
t = 0.84, 58 time steps.
Reaction-diffusion equations
2D thermo-diffusive flames
3D flame balls
Governing equations
Non-dimensional thermodiffusive equations
~ 2 T = s
tT + ~v T
1 2
~
tY + ~v Y
Y =
Le
Ze2
(1)
(2)
Ze(T 1)
(reaction rate)
2 Le
1 + (T 1)
h
4
4 i
1
1
s(T ) = T + 1 1
(heat loss due to radiation)
(T, Y ) =
Y exp
Tb Tu
Ea
Y = Y1 , T =
, Le =
(Lewis), =
, Ze =
(Zeldovich)
Tb Tu
D
Tb
RTb
~v given by the incompressible NS equations. When the fluid is at rest, ~v = ~0.
Governing equations
Planar flames
Flame propagation at the velocity vf
When the fresh mixture is advected at v = vf steady planar flame
T
Y
0
A
AB: fresh mixture, BC: preheat zone, CD: reaction zone d = O(Ze1), DE: burnt mixture
Governing equations
Thermodiffusive instability
y
y
<0
x
Fresh gas
Burnt gas
x
Fresh gas
Burnt gas
heat
mass
2D Flame front
Temperature
stable
Le = 1.0
Unstable
Le = 0.3
Reaction rate
Adaptive grid
Application to TD flames
The flame ball configuration
Simplest experiment to study the interaction of chemistry and transport of
gases (experimental: Ronney 1984, theory: Buckmaster-Joulin-Ronney 1990-91)
Enables to study the flammability limit of lean gaseous mixtures
COLD PREMIXED GAS
radiation
heat conduction
HOT BURNT GAS
reactant diffusion
Application to TD flames
Interaction flame front-adiabatic wall: the 1D case
Lean mixture H2-air, Ze = 10, = 0.64, = [0, 30]
Radiation neglected
Adiabatic walls Neuman boundary conditions
Objective: study the inflence of Le
Profiles of T and for Le = 0.3 (animation 1)
Profiles of T and for Le = 1 (animation 2)
Profiles of T and for Le = 1.4 (animation 3)
Application to TD flames
Interaction flame front-adiabatic wall: the 1D case
2
Le = 0.3
Le = 0.7
Le = 0.8
Le = 0.9
1.5
Le = 0.95
Le = 1
Le = 1.4
6
5
4
3
2
0.5
10 12
t
Flame velocity vf for Le < 0.95
10 12
t
Flame velocity vf for Le 0.95
Application to TD flames
Interaction flame ball-adiabatic wall
Radiation neglected, Ze = 10, = 0.64, = [50, 50]d
Adiabatic walls Neuman boundary conditions
At t = 0, the radius of the flame ball is r0 = 2.
2D: Evolution of T and mesh for Le = 0.3 (animations 1-2)
2D: Evolution of T for Le = 1 (animation 3)
2D: Evolution of T for Le = 1.4 (animation 4)
3D: Evolution of T and mesh for Le = 1 (animations 5-6)
Analogy with capillarity for a fluid droplet
Application to TD flames
Interaction flame ball-adiabatic wall
140
Le = 0.3
Le = 1
Le = 1.4
120
100
80
60
40
20
0
0 10 20 30 40 50 60 70
t
R=
d in 2D
8000
7000
6000
5000
4000
3000
2000
1000
0
Le = 0.3
Le = 1
Le = 1.4
10
t
R=
d in 3D
Application to TD flames
Interaction flame ball-adiabatic wall: Performances
d
Le
Nmax
% CPU % Mem
2
2
2
0.3
1
1.4
2562
2562
2562
25.50%
21.50%
21.00%
14.10%
11.75%
11.10%
1283
12.98%
4.38%
Application to TD flames
Interaction flame ball-vortex
Application to TD flames
Interaction flame ball-vortex
40
35
30
d:
30
25
20
20
15
15
10
10
6= 0
=0
35
25
R=
40
MR
F V 25622
F V 512
3D flame ball, Le = 1
Temperature
Adaptive grid
Temperature
Concentration
grid xy
grid yz
Iso-surfaces and isolines on the cut-plane for temperature (top) and concentration (bottom) with
L=8 scales, Le=0.3, Ze=10,k=0.1.
Method
% CPU time
% Memory
Integral reaction
rate
MR
2.7
1.05
669.09
MR/LTS
2.3
1.05
669.11
Conclusions
Finite volume discretization with explicit time integration (both of second-order) to
solve evolutionary PDEs in Cartesian geometry.
Efficient space-adaptive multiresolution method (MR) with local time stepping (LTS).
CPU speed-up and memory reduction, while controling the accuracy.
Further speed-up due to an improved time advancement using larger time steps on
large scales without violating the stability condition of the explicit scheme.
However, synchronization of the tree data structure necessary.
Time-step control (CTS) for space adaptive schemes (embedded Runge-Kutta schemes)
and combination with LTS.
Applications to reaction-diffusion equations, compressible Euler and Navier-Stokes equations.
Next: develop level dependent time step control which allows to adapt the time step within a
cycle of the level dependent time stepping MR/LTS.
http://www.cmi.univ-mrs.fr/~kschneid
http://wavelets.ens.fr