)

*
+,
.
!

/
0
0
1






!
"

#$

"
#

%


"

!
$

"
&

( %'





















 


2
3
4
45
4
6
7
98
:
;
<
98
:
>=
?
4@ 5
39
A5
8
B
CD
E5
;8
C9
F5
>
FG
4F
FH











I [
`L \
W ]
a
JO ^
T _
MN YO L
U X V
N
W
X
J
K
N
YWL
Z

I
J
KM L
N
O
QP
O
R
S
ST
L
J
U

Seminar for
Applied
Mathematics

Boundary Conditions for

Regularized 13-Moment-Equations
Manuel Torrilhon

Compressible Viscous Gas Flow

gas variables: density, velocity and temperature U = f; v; T g

+ div v
=0
@t

T
@t v
+ div v v + pI +
=0

@t e + 12 v2 + div e + 12 v2 + p v + v + q = 0
ideal monatomic gas, internal energy e (T ) =

3 k
2 m T;

k
T
and pressure p(; T ) = m

Empirical Consititutive Relations

@T
qi = (T )
@xi
b
c
d
ef
g
e
h
i
j
k
kl

m
n
op
q
r
s
t
u
v
vw

x
y
z
{
|
}
~


ij

@vhi
= 2(T )
@x ji

Failure of NSF Models

asymmetry of density shock wave profiles
Q

1.1

M0

(+)

1 +0
2

()
Q = I (+)
I

(x)

NSF
1.0

I ()

direct simulations

x
0.9
1.0

2.0

M0

3.0

micro-channel mass flux

y

ux
force

m = v dy

measurements

v(y)

NSF
big channel

D1

small channel

Kinetic Gas Theory

stochastic description based on velocity distribution function f :

R5

0 01

1
f f f f g b db d! dc1

@f
1
@f
+ ci
=
@t
@xi
"

[0; T ] R3 ! R

moments of the distribution function: Fijk = m

ci cj ck f dc

fluid variables F = ; Fi = vi ; Fii = 3T + vi2 ; Fij ij ; Fijj qi

moments satisfy an equation hierarchy with closure problem
scaling parameter: Knudsen number " =
b

mean-free-path
observation scale

limit " ! 0 , equilibrium flow f ! fM , Maxwell distribution

non-equilibrium modelling 0 < " < 1
Find partial differential equations that approximate
the multi-scale behavior in a continuum model

Classical Approximations
asymptotic analysis: fCE = fM + " f (1) + "2 f (2) + : : :
(0)

leads to e.g. qi = qi

(1)

+ " qi

2 (2)

+ " qi

Fourier

+ : : : for non-equilibrium variables

Burnett

super-Burnett

parabolic systems including higher order derivatives

! Burnett and super-Burnett are linearly unstable, see BOBYLEV (1982)

N

X
Hermite series of the distribution function: fG = fM 1 +
ijk ci cj ck

with relation ijk $ Fijk for a fixed number of N moments

Grad considers evolution equations for a large set of N moments, e.g. 13

large first order hyperbolic systems in divergence form, always stable

introduces subshocks into shock structure, see GRAD (1952), WEISS (1995)

STRUCHTRUP/TORRILHON (2003)

higher system

Euler eqns

importance
of the moment

13 moments

Regularization of Moment Equations

importance
of the moment

basic
equations:
Euler eqns

third expansion: Super Burnett

second expansion: Burnett
first expansion: NSF

O( " )
O( " 2 )
O( " 3 )

{z

moments of
the distribution function

importance
of the moment

{z

conservative
variables

basic
equations:
N moments
first expansion:
regularized
N moment case
O( e
")
|

moments of
the distribution function

conservative
variables

{z

conservative
variables

moments of
the distribution function

Regularized 13-Moment-Equations
basic conservation equation
@t
+ div v
=0

@t v
+ div v vT + pI +
=0

1 2
1 2
@t e + 2 v + div e + 2 v + p v + v + q = 0
Extended Consititutive Relations (Regularized 13-Moment-Equation, linearized)

stress/heatflux follow a driven wave equation system with relaxation and dissipation

@qhi
@vhi
@ ij
@
@ hij
p
4
+ 2p
= ij + 2
+5
(T
)
@t
@x ji
@x ji
@xk p @x ki
12
5

@
@q fi
(T
)
@xk p @x kg

@qi
@ ij
5 @T
+T
+ 2p
= 23 p qi +
@t
@xj
@xi

R13 is highly accurate and fully stable continuum model with smooth shock profiles
earlier attempts: GRAD (1958), KARLIN ET AL. (1998), JIN & SLEMROD (2001), MLLER ET AL. (2002)

Full Regularized 13-Moment-Equations

full nonlinear 3d-equations for stress and heat flux:

@t ij + @k ( ij vk + mijk ) + 45 @hi qji + 2p@hi vji + 2@k vhi jik = p ij

@t qi +@k qi vk +

1
5
+p@
(
=)+
R
k
ik
ik
2
2 (p ik +

ik )@k

2p
( ij =)@k kj + qk @k vi +(mijk + 65 q(i jk) )@k vi = 3
qi

^ ij = Rij + 1 R ij ):
regularization terms ( R
3
mijk = 2 @hi ( jki =) +

(NSF)
8
q

10p hi jki

Rij = 24
5 @hj (qji =) +

(NSF)
32
q
q
25p hi ji
(NSF)

R = 12 @k (qk =) + p8 qk qk
(NSF)

= 2 @hi vji

(NSF)

= 15
4 @i

abbreviations: ij

qi

(NSF)
24

7 khi jik
(NSF)

+ 6 ij ij

TORRILHON/STRUCHTRUP (2004)

Accuracy and Knudsen Order

Denition : (Knudsen order) Assume U (B oltz) and U (m odel) are the respective solutions expanded in " and the dierence satises

n+1
(model)
(model)
(Boltz)
(Boltz)

+ q
= O "

Then the Knudsen order or accuracy of the model is n 2 N.

all continuum models can be assigned a Knudsen order

equilibrium flow: n (Euler) = 0
classical theory: n (NSF) = 1
Burnett expansion: n (Burnett) = 2 (unstable!!)
Grads 13-moment-equations: n (Grad) = 2

(subshocks!!)

Theorem : (R13 Accuracy) The regularized 13-moment-equations provide a

Knudsen order
n (R13) = 3;

hence, they are of super-Burnett order. Higher terms inside the equation stabilize the system.

Boundary Value Problems

consider plane channel flow between
infinite plates

(1)

(1)

vW ; W

force F

L : channel width

vx (y)

plates can be heated and moved

(0;1)
(0;1)
independently with vW
; W

(0)

(0)

vW ; W

force represent homogeneous pressure gradient

Knudsen layer phenomenon:

write steady R13-equations as first order system

in channel geometry:

bulk solution

A (U) @y U = P (U)

velocity
or temperature profile

10 relevant R13-variables for channel flow:

n
o
^ yy ; myyy
U = vx ; xy ; qx ; mxyy ; Rxy ; qy ; yy ; R

Knudsen boundary layer

velocity part

temperature part

(0)

(0)

vW ; W

temperature jump
or velocity slip

Kinetic Boundary Conditions

Maxwell accommodation model for the distribution function at the wall

()
(c)
+
(1

)
f
(c)
n (c vW ) > 0

f
gas
W
f~ (c) =
n (c vW ) < 0
fgas (c)
R13 distribution function + integration gives boundary relations for moments
continuity for ! 0 implies: only odd (in y) moments should be prescribed
consistency implies: only fluxes of the variable set should be prescribed

xy =

2
2

1
1
P VW + mxyy + qx
2
5

11
1

2
Rxy =
P VW mxyy qx P VW3 + 6P ( W )VW
2
2
5

r
5
1
1

2
^yy + yy P VW2
qy =
2P ( W ) + R
2
28
2
2
4 boundary conditions

r
on both walls

1
7
3
2
2
2
^
myyy =
P ( W ) R
yy yy P VW
2 5
not enough...
14
5
5

with

VW = vx vW

Boundary Layer Reduction

in an infinite moment hierarchy all higher variables produce boundary layers
an even moment Meven together with an odd higher moment Fo dd forms a
boundary layer pair, with kinetic boundary condition for Fo dd
1

fother termsg + @y Fodd = Meven

"
exp( y=")
Fodd = " @y Meven

Meven

boundary layer

mxyy

xyyy

...

qx

Rxy

xyy

...

qy

^ yy
R

yyy

...

yy

myyy

yyyy

...

boundary
layer pair

...

...

...

 

 








 
 

vx
bulk solution

xy





 



bulk values imply boundary conditions





any truncated theory produces cut pairs, single variables are reduced to their
bulk solution

Nullspace Conditions
consider the R13 system in first order form with variable vector U 2 RN
A (U) @y U = P (U)

reduction of boundary layer pair from a larger system produces a singular

matrix A(U) , thus eigenvalues i = 0 and left eigenvectors
fxi gi=1;::: with xi A(U) = 0

on the differential equations, this produces intrinsic relations of the variables

for i = 1; :::
xi P(U) = 0

intrinsic relations correspond to bulk solutions and supplement boundary conditions

= 2 for R13 with linear constitutive equations:
x1 P(U) = 0

x2 P (U) = 0

16
mxyy = F
15
^ yy = 6 Rxy 6
R
5
p + yy

Order-Preserving Transformation
fully non-linear R13 system fails to exhibit zero eigenvalues
coherence implies: transformation necessary to find bulk solutions
order preserving transformation:

(R13)
(R13)
(Boltz)
(Boltz) !
4
~
~

q
+ q
= O Kn

2
constitutive relations may be altered within an error O Kn

(NSF)
2

@
v
=

+
O
Kn
e.g., by replacing:
y x
xy
xy

^ yy =
algebraisation: R

@y qy

66
5p

qy @y

36
7

xy @y vx + : : :

qy

final system
exhibits nullspaces as before

36
5

x1 P (U) = 0

x2 P (U) = 0

xy

mxyy =

16
32
qy F
45p
15

^ yy = 136 qy2 72 2
R
25p
35

Complete Set of Boundary Conditions

TORRILHON/STRUCHTRUP (2007)

kinetic boundary conditions on both sides of the channel

1
1
xy =
1 P VW + mxyy + qx
2
5

r
1
11
2
3
P

q
Rxy =

P
V
+ 6P ( W )VW

xyy
x
W
W
2

2
5

r
5
1
1
2
2
^

2P ( W ) + R
qy =
yy + yy P VW
3

28
2
2

r
2
1
7
3
2
2
^
P ( W ) R
myyy =

yy yy P VW
4

5
14
5
5
r

specific accommodation coefficients f 1 ; 2 ; 3 ; 4 g are fitted to be f0:9; 0:5; 0:9; 0:5g

bulk solution values supplement conditions on both sides
mxyy =

32
16
qy F
45p
15

^ yy = 136 qy2 72 2
R
25p
35

larger values of Kn seem to require larger values of accommodation coefficients

TORRILHON/STRUCHTRUP (2007)

Poiseuille Channel Flow

channel flow with walls at rest
and same temperature

vx

1.04

0.2

0.0
L=2

the case Kn = 0:068 is compared

to DSMC results
R13 produces Knudsen minmum
in total mass flux

Kn = 0:068

Kn = 0:15

0.4

given force F = 0:23 corresponds to

homogeneous pressure gradient

Kn = 0:068

Kn = 1:0

1.02

Kn = 1:0

Kn = 0:4

Kn = 0:15

Kn = 0:4

L=2

1.00
L=2

L=2

L=2

L=2

qy

0.1

Kn

0.02
0.0

0.0

-0.02
-0.1

L=2

L=2

-0.04
L=2

1.6

qx

^
J

0.0

yy

0.0
1.4

Kn

Kn

-0.04

1.2

-0.01

R13
-0.08

0.8

NSF
0.4

0.05

0.1

0.2

L=2
0.5

1.0

d
Kn

2.0

L=2

-0.02
L=2

R13 Couette Channel Flow I

channel flow with walls at
opposite velocities and
same temperature
two different cases of
velocities vW = 1:26 and
vW = 0:63 at different Kn

1.26

vx

Kn = 0:1

1.3

0.63

vW = 0:63

1.2

vW = 1:26

0.0

vW = 0:63
1.1

-0.63

vW = 1:26
-1.26
L=2

L=2

1.0
L=2

L=2

qy

0.0

here: Kn = 0:1

Kn = 0:1

Kn = 0:1

vW = 0:63

0.2
0.1

-0.1

vW = 0:63

0.0

vW = 1:26
-0.2

-0.1
-0.2

-0.3
L=2

L=2

0.2

vW = 1:26

Kn = 0:1

L=2

L=2

0.0

qx

yy

0.1

vW = 0:63

-0.02

0.0

vW = 0:63
vW = 1:26

-0.04
-0.1

vW = 1:26

Kn = 0:1
-0.2
L=2

Kn = 0:1

L=2

-0.06
L=2

L=2

R13 Couette Channel Flow II

channel flow with walls at
opposite velocities and
same temperature
two different cases of
velocities vW = 1:26 and
vW = 0:63 at different Kn

1.26

vx

Kn = 0:25

0.63
0.0

vW = 1:26

vW = 0:63
1.2

vW = 0:63

-0.63

vW = 1:26
-1.26
L=2

L=2

1.0
L=2

L=2

qy

0.0

here: Kn = 0:25

Kn = 0:25

1.4

Kn = 0:25

0.4

vW = 0:63

0.2
-0.2

vW = 0:63

0.0

vW = 1:26

-0.2

-0.4
-0.4

L=2

L=2

vW = 1:26

Kn = 0:25

L=2

L=2

0.0

0.4

yy

qx

0.2

vW = 0:63
0.0

-0.1

vW = 0:63

vW = 1:26

-0.2
-0.4

L=2

vW = 1:26

Kn = 0:25
0

Kn = 0:25

L=2

-0.2
L=2

L=2

R13 Couette Channel Flow III

channel flow with walls at
opposite velocities and
same temperature
two different cases of
velocities vW = 1:26 and
vW = 0:63 at different Kn

1.26

1.6

Kn = 0:5

vx

Kn = 0:5

0.63
1.4

vW = 0:63
0.0

vW = 1:26

vW = 0:63

1.2
-0.63

vW = 1:26
-1.26
L=2

L=2

1.0
L=2

L=2

0.0

here: Kn = 0:5

Kn = 0:5

0.4

qy

vW = 0:63

-0.2

0.2

vW = 0:63

0.0
-0.4

vW = 1:26
-0.2

-0.6

-0.4

L=2

L=2

qx

vW = 1:26

Kn = 0:5

L=2

L=2

0.0

yy

0.4
0.2
0.0

vW = 0:63

vW = 1:26

vW = 0:63
-0.2

-0.2
-0.4

L=2

vW = 1:26

Kn = 0:5
0

Kn = 0:5

L=2

-0.4
L=2

L=2

Achievements with R13

accuracy of super-Burnett order in full non-linear, multidimensional setting
(MT & HS 2004)

(R13)
(R13)
(Boltz)
(Boltz)

+ q
= O Kn4

linearly stable for all wave numbers and frequencies (HS & MT 2003)
follows from an order-of-magnitude argument without expansion (HS 2004)
good agreement with dynamic form factors of light scattering (MT 2006)
smooth shock wave profiles with improved quantitative agreement (MT & HS 2004)
allows efficient numerical multi-dimensional simulations (MT 2006)
comes with entropy law in the linear case (HS & MT 2007)
channel micro-flow simulations possible (Gu & Emerson 2007)