A Comparison and Unification of Ellipsoidal Statistical and Shakhov BGK Models
A Comparison and Unification of Ellipsoidal Statistical and Shakhov BGK Models
A Comparison and Unification of Ellipsoidal Statistical and Shakhov BGK Models
m559
Adv. Appl. Math. Mech., Vol. 7, No. 2, pp. 245-266 April 2015
Abstract. The Ellipsoidal Statistical model (ES-model) and the Shakhov model (S-
model) were constructed to correct the Prandtl number of the original BGK model
through the modification of stress and heat flux. With the introduction of a new pa-
rameter to combine the ES-model and S-model, a generalized kinetic model can be
developed. This new model can give the correct Navier-Stokes equations in the con-
tinuum flow regime. Through the adjustment of the new parameter, it provides abun-
dant dynamic effect beyond the ES-model and S-model. Changing the free parameter,
the physical performance of the new model has been tested numerically. The unified
gas kinetic scheme (UGKS) is employed for the study of the new model. In transition
flow regime, many physical problems, i.e., the shock structure and micro-flows, have
been studied using the generalized model. With a careful choice of the free parameter,
good results can be achieved for most test cases. Due to the property of the Boltz-
mann collision integral, the new parameter in the generalized kinetic model cannot
be fully determined. It depends on the specific problem. Generally speaking, the S-
model predicts more accurate numerical solutions in most test cases presented in this
paper than the ES-model, while ES-model performs better in the cases where the flow
is mostly driven by temperature gradient, such as a channel flow with large boundary
temperature variation at high Knudsen number.
AMS subject classifications: 65M10, 78A48
Key words: Kinetic models, unified gas kinetic scheme, rarefied flow.
1 Introduction
The monatomic rarefied gas behavior can be described by the Boltzmann equation. How-
ever, the collision term of the Boltzmann equation is a multiple integral term which is
∗ Corresponding author.
Email: [email protected] (S. Z. Chen), [email protected] (K. Xu), [email protected] (Q. D. Cai)
very complicated for analysis and numerical computation. Bhatnagar et al. [1] simpli-
fied the collision term in the Boltzmann equation and proposed BGK model in which the
Boltzmann collision term is replaced by a relaxation term. This relaxation term mimics
the main relaxation process from nonequilibrium state towards to a local equilibrium one.
The local equilibrium state known as Maxwell distribution function is determined by the
local conservative flow variables, namely, the mass, momentum and energy. Due to its
simplicity, the BGK model becomes an important kinetic model for analysis and numer-
ical simulation of nonequilibrium flows. However, the Chapman-Enskog expansion of
the BGK model derives the Navier-Stokes equations with a unit Prandtl number, which
is different from the physical reality in the continuum flow regime. For a monatomic gas,
the accepted Prandtl number is about 2/3 in a wide range of flow conditions.
In order to fix the Prandtl number, many kinetic models have been proposed in the
past decades. The main approach is to modify the relaxation term. For example, the
Ellipsoidal Statistical BGK model [2] employs a Gaussian distribution as the relaxation
equilibrium state instead of the Maxwell distribution. This model is not very popular
until Andries [3] proved the entropy condition of the ES-model. In the ES-model, besides
the conservative flow variables, the local stress tensor also involves in the relaxation ter-
m. By changing the free parameter in the ES-model, it can present an arbitrary Prandtl
number. Moreover, the nonnegative property of the Gaussian distribution is a favorable
physical property.
Another very popular kinetic model is the Shakhov model [4]. Unlike the ES-model,
it adjusts the heat flux in the relaxation term. With the implementation of Hermite poly-
nomial, the low order moments of relaxation term in S-model are identical to the original
BGK one, namely, the conservative variables are maintained and the stress tensor keeps
isotropic one as that from the BGK model. Only the third order moments of relaxation
term change. In other words, the S-model modifies the BGK model by adjusting heat
flux to present the correct Prandtl number. But, it allows negative value of distribution
function and its H-theorem was only proved in near equilibrium condition [4].
In 1990, Liu [5] proposed a new kinetic model by considering the gain term and lost
term of the Boltzmann equation separately, where the Chapman-Enskog distribution is
directly used to evaluate the relaxation term. Therefore, the space derivatives are in-
volved in the collision term. Liu model changes both the heat flux and stress tensor
of the relaxation process and provides a correct Prandtl number in the continuum flow
regime. Due to its relatively complicated formulation, this model has not been widely
used.
Although all above models provide correct Prandtl number in the continuum flow
regime, their properties are very different in the transition regime [6–10]. Garzo [7] re-
ported a singular behavior of Liu model and attributed it to the negative distribution
function. Graur [8] studied the heat transfer problem and found that the ES-model pro-
vides better results than the S-model through the comparison with the results from the
Boltzmann equation. The ES-model keeps the distribution function positive, while the S-
model and Liu model always allow un-physical negative distribution function. It seems
that the nonnegative properties of the ES-model are important and promising. Moreover,
the ES-model satisfies the H theorem, while the H-theorem of the S-model is only proved
in the near local equilibrium state [9]. However, some other studies did not tell the same
story. Mieussens [11] and Kudryavtsev [10] both reported the early rising of temperature
profile in the shock structure solution derived by the ES-model.
In fact, the physical performance of these models has not yet been evaluated exten-
sively in the transition regime. The properties, such as the H-theorem, nonnegative dis-
tribution and conservation etc., cannot cover a complete physical picture of dynamics of
the particle collision term and the evolution of the distribution function. The original
motivation for the development of the kinetic models is to fix the Prandtl number which
is well defined in the continuum flow regime. In transition regime, it is expected that
significant differences in their performance would appear in different physical problem-
s. Furthermore, the practical applications care more about the macroscopic quantities,
such as the moments of a distribution function. The H-theorem and the nonnegative dis-
tribution function only present a portion of physical conditions, but is not sufficient to
guarantee a correct dynamic evolution of macroscopic quantities. So, it is necessary to
inspect the practical performance of different kinetic models through the numerical simu-
lations in the transition regime. In order to cover a whole range of dynamic performance
of kinetic models, we are going to introduce a generalized kinetic model which combines
the ES-model and S-model. With the combination of these two models, besides obtaining
the correct Prandtl number, we have one more free parameter to be adjusted. With the
variation of this parameter, a continuum dynamic performance from the ES-model to the
S-model and beyond, can be identified.
In the past years, a unified gas kinetic scheme (UGKS) [12–16] has been well devel-
oped. The BGK model and the S-model have been employed in the UGKS. In this paper
we will use the UGKS framework to construct numerical scheme for the generalized ki-
netic model. The numerical scheme will be used to examine physical performance of
different kinetic models with the variation of the parameter, where both the ES-model
and the S-model become limiting cases. A continuous dynamic transition between these
two models can be obtained. Through investigations, the performances of different ki-
netic models under different flow conditions in the transition regime are presented in
details.
This paper is organized as follows. Section 2 introduces criteria for the construction
of kinetic model. Section 3 proposes a generalized kinetic model. Section 4 gives the
simulation results of the new model in the shock structure and microflow computations.
The parameter dependent dynamic effect will be discussed in different test cases. Section
5 presents the analysis and insight of the new model. The last section is the conclusion.
f t + u ·∇ f = J ( f ), (2.1)
where J ( f ) denotes the collision term. Here f represents the velocity distribution func-
tion which is a function of the physical space location x and the molecule velocity u.
The macroscopic quantities, such as, the mass ρ, momentum ρU (ρUi ), energy ρE, stress
tensor P ( Pij ) and heat flux q (qi ), can be derived from the distribution function f ,
ρ ∫
W = ρU = ψ f du, (2.2a)
ρE
∫
Pij = (ui − Ui )(u j − Uj ) f du, (2.2b)
∫
1
qi = (u − Ui )(u − U)2 f du, (2.2c)
2 i
where ψ is defined as follows,
1 2 )T (
ψ = 1,u, u , (2.3)
2
and du is the volume element in the velocity space. Since the mass, momentum and en-
ergy are conserved during particle collisions, the collision term satisfies the conservation
constraint, ∫
J ( f )ψdu = 0, (2.4)
where g+ is the post collision term and τ denotes the relaxation time. The purpose of con-
structing kinetic model is to get physical insight of rarefied gas system and capture the
flow behavior with minimum mathematical complexity. In the modeling process, certain
information about the gas system, such as the individual particle collision trajectory, has
to be ignored. But, the recovered macroscopic equations should be maintained as pre-
cisely as possible in comparison with the corresponding ones of the Boltzmann equation.
The moment equations are accurate in all flow regimes if the last two equation of Eq. (2.5)
is accurate. In fact, from the view of macroscopic quantities, the different gas molecules
or kinetic models are reflected in two parts, one is the source terms S p and Sq , and the
other is convection term in the last two equations of Eq. (2.5) which induce the closure
problem. In this study, we focus on the first one, the source terms, which represent the
relaxation processes of stress and heat flux. Different relaxation processes of the kinetic
models are compared with that of the Boltzmann solution.
g+ = M[ f ]+S[ f ], (3.3a)
[ ( c2 )/ ]
S[ f ] = M[ f ] (1 − Cshak )c · q − 5 (5pRT ) , (3.3b)
RT
ρ ( 1 2)
M[ f ] = √ exp − c , (3.3c)
det(2πRTI) 2RT
where M[ f ] denotes the Maxwell distribution function. and Cshak is a parameter which
is related to the Prandtl number in this model.
Replacing the Maxwell distribution by the Gaussian distribution in the S-model, the
post collision term of the generalized kinetic model is
g+ = G[ f ]+S[ f ]. (3.4)
The two coefficients, Ces and Cshak , are two independent parameters at this moment. In
order to obtain the right transport coefficients, we follow the proof of Andries [3]. In
continuum regime, the distribution function is expanded as,
then
1
p= p . (3.7)
(1 − Ces ) bgk
And heat flux is,
q = (1 − Cshak )q + qbgk , (3.8)
then
1
q= qbgk . (3.9)
Cshak
As a result, the Prandtl number for the generalized kinetic model is
Cshak C
Pr = Prbgk = shak . (3.10)
1 − Ces 1 − Ces
And the viscosity from the distribution function (Eq. (3.5)) is
τp
µ= . (3.11)
1 − Ces
Considering a spatially homogenous monatomic gas, the following three moment equa-
tions can be obtained from the generalized model,
4 Numerical results
In this section, the generalized kinetic model is solved by the unified gas kinetic scheme
(UGKS) [12–16]. At first, we will briefly introduce the UGKS.
where x′ = x − u(t − t′ ).
Applying this solution at cell interface, the mass flux, momentum flux and energy
flux can be obtained as follows,
∫
n· u f du
Fmass ∫ Ωu
Fmacro = Fmomentum = n · uu f du . (4.2)
Ωu
Fenergy ∫
1
n· u u2 f du
Ωu 2
The Ωu denotes the entire velocity space and n is the normal direction of the cell interface.
The flux of velocity distribution function at particle velocity uk takes the following
form:
∫
Fuk = n · u f du, (4.3)
Ωuk
where Vxi is the volume of Ωxi in the physical space, ∆Sm is the area of interface and m is
the index of surfaces of Ωxi . The collision term must be considered for the update of the
distribution function. Here, we use two steps to update the distribution function
∫ t n +1 +(n)
1 gu k − f unk
f u∗k = f unk − ∑ ∆Sm Fu + ∆t , (4.5a)
Vxi tn m
k
τn
∫ t n +1
∆t ( guk − f unk+1 guk − f unk )
+(∗) +(n)
1
f unk+1 = f unk −
Vxi tn
∑ ∆S m F u k
+
2 τ n +1
+
τn
, (4.5b)
m
∆t +(∗) τ
n +1 ∆t +(n) 2τ n − ∆t n
f unk+1 = g u + g u + f
2τ n+1 + ∆t k τ n 2τ n+1 + ∆t k 2τ n+1 + ∆t uk
∫ n +1
2τ n+1 1 t
+ ∆t Vxi tn ∑
− n +1 ∆Sm Fuk . (4.5c)
2τ m
A medium state f ∗ is estimated first. Then solve the second equation and get f unk+1 at
the next time level. The above procedure is identical for an arbitrary g+ . For the details
of the numerical reconstruction, please refer to the articles about gas kinetic scheme [12,
13, 19].
The UGKS is a direct physical modeling of flow motion in the scale of the discretized
space and the integral solution (Eq. (4.1)) for the kinetic model covers the flow evolu-
tion from kinetic to the hydrodynamics scales. The specific flux used at the cell inter-
face depends on the ratio of time step to the local particle collision time (see the ter-
m e−t/τ ). Such unique merits enable the UGKS to recover correct fluxes in the whole
Knudsen number ranges including free molecule regime, transition regime and con-
tinuum regime. In the following numerical simulations, the velocity space is uniform-
ly discretized. And Newton-Cotes quadrature rules are adopted for the integration of
velocity space.√ Without specification, √ the velocity points cover the velocity range of
[−(|Ure f |+ 4.5 RTre f ), |Ure f |+ 4.5 RTre f ].
where L is the width of the channel and subscript 0 denotes the initial value of variables.
The gas is confined between two vertical plates which locate at x = −0.5 and x = 0.5 re-
spectively. The temperature of the plates is Tw = 1. The initial flow states are shown as
follows, T0 = 1, ρ0 = 1, p0 = 1. The gravity is represented by G and is upward in the vertical
direction. Here the hard sphere molecule is adopted, namely, the viscosity-temperature
1.2 1.4
DSMC
1.1 Ces=0.5
Ces=0.25
Ces=0 (S-model)
1.35 Ces=-0.25
1 Ces=-0.5 (ES-model)
DSMC
V
0.9
T
Ces=0.5
Ces=0.25 1.3
Ces=0 (S-model)
0.8 Ces=-0.25
Ces=-0.5 (ES-model)
0.7
1.25
0.6
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
X X
1.55
2
1.5
Ces = 0.5
V
1.4
T
Ces = 0.25
DSMC Ces = 0 (S-model)
1 Ces = -0.25
Ces = 0.5 1.35 Ces = -0.5 (ES-model)
Ces = 0.25
Ces = 0 (S-model)
Ces = -0.25
Ces = -0.5 (ES-model) 1.3
0.5
1.25
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
X X
3
1.9
2.5
1.8
2
1.7
DSMC
V
1.5
T
1.6
DSMC Ces = 0.5
Ces = 0.25
1 Ces = 0 (S-model)
Ces = 0.5 1.5
Ces = -0.25
Ces = 0.25
Ces = -0.5 (ES-model)
Ces = 0 (S-model)
0.5 Ces = -0.25 1.4
Ces = -0.5 (ES-model)
0 1.3
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
X X
Figure 1: The velocity and temperature profiles for force driven Poiseuille flow derived from the generalized
kinetic model under different Knudsen numbers. The Knudsen numbers are 1, 0.1 and 0.05 respectively from
top to bottom. And the Gravity is G = 1.
coefficient is ω = 0.5, i.e., µ ∼ T ω . The gas-wall interaction is fully diffusive kinetic bound-
ary condition. Due to the unreasonable large value of G = 1, this test becomes a very
difficult simulation and the distribution function is totally distorted by the external force,
especially at high Knudsen number. Therefore, the discrete velocity point must extend
to very far place and be modified manually in the force direction. The upper bound of
√
vertical velocity extends to 8 RT0 . Here Ces varies from −0.5 to 0.5 in the generalized
kinetic model. The two special cases, the ES-model and the S-model, are included in this
set of simulations.
As shown in Fig. 1, when the Knudsen number is small, say, Kn = 0.05, the differ-
ence between results from different kinetic models and the DSMC is small. But, as the
Knudsen number becomes large, the temperature profiles separate from each other. For
all Knudsen numbers, the results from the S-model are closer to the DSMC results than
that from the ES-model. The profiles with different Ces cover the results of ES-model and
S-model. And when the Ces is larger than 0, the temperature profile moves from the S-
model result to the DSMC result. It is clearly shown that the generalized kinetic model
can predict more accurate results in comparison with the S-model and ES-model if Ces is
specified properly.
In this case, we also consider the monatomic gas between two plates. The left and right
plates move with the same speed (300), but with opposite directions. The left one moves
downward and the right one moves upward. The temperatures of gas and plates both
are 273. The reference viscosity is defined as,
√
30 ρλ 2πRT
µre f = . (4.7)
(7 − 2ω )(5 − 2ω ) 4
The viscosity-temperature coefficient ω is 0.81. The solutions are compared with the
DSMC results [21]. The Prandtl number is 2/3. In current simulation, the spatial coordi-
nate is ranged from −50 to 50 and is discretized with 100 cells. The Knudsen number is
varied by the changing of gas density.
Fig. 2 shows the velocity and temperature profiles from the different kinetic model-
s. When Knudsen number is 0.01, the results from different kinetic models are almost
identical with the DSMC results. While Knudsen number becomes large, the velocity
and temperature profiles separate gradually. The cases in which Ces = 0 or 0.25 present
most satisfactory results compared with the DSMC results, especially for Kn = 0.1. When
Knudsen number is 1, the temperature profiles move away from the DSMC results. Based
on these numerical data, we show that the generalized kinetic model with different pa-
rameters present the same continuum limit when Knudsen number is small. In the medi-
um Knudsen number regime, the S-model perform the best. For low speed Couette flow,
we consider the case presented in [22]. Hard sphere model is adopted. The initial value
is 1 for the density and temperature. The wall speed is only ±0.2. The Knudsen number
is defined by Eq. (4.6). Since the velocity profiles are very close to each other, we on-
ly present temperature profiles. As shown in Fig. 3, the generalized kinetic model with
Ces = 0.25 gives the best results.
300 340
200
320
100
V
T
DSMC 300
-100 DSMC
Ces=0.5
Ces=0.25 Ces=0.5
Ces=0.0 Ces=0.25
-200 Ces=-0.25 Ces=0.0
Ces=-0.5 Ces=-0.25
280 Ces=-0.5
-300
-40 -20 0 20 40 -40 -20 0 20 40
X X
(a) Kn = 0.01
300 360
200
100
340
V
0
T
DSMC
-100 DSMC
Ces=0.5
320 Ces=0.5
Ces=0.25
Ces=0.0 Ces=0.25
-200 Ces=-0.25 Ces=0.0
Ces=-0.5 Ces=-0.25
Ces=-0.5
-300
-40 -20 0 20 40 -40 -20 0 20 40
X X
(b) Kn = 0.1
150 420
415
100 410
405
50
400
395
V
0
T
390
DSMC 385
-50 DSMC
Ces=0.5
380 Ces=0.5
Ces=0.25
Ces=0.0 Ces=0.25
-100 375 Ces=0.0
Ces=-0.25
Ces=-0.5 Ces=-0.25
370 Ces=-0.5
-150 365
-40 -20 0 20 40 -40 -20 0 20 40
X X
(c) Kn = 1
Figure 2: The velocity and temperature profiles for Couette flow problem.
7
∆T×1000
∆T×1000
5
6
DSMC(Meng) DSMC(Meng)
4
Ces=0.5 Ces=0.5
Ces=0.25 Ces=0.25
5
Ces=0.0 Ces=0.0
Ces=-0.25 Ces=-0.25
3 Ces=-0.5 Ces=-0.5
4
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
X X
10
∆T×1000
8 DSMC(Meng)
Ces=0.5
Ces=0.25
Ces=0.0
7 Ces=-0.25
Ces=-0.5
(c) Kn = 0.5
Figure 3: The temperature profiles (∆T = T − T0 ) for low speed Couette flow problem. The DSMC results are
extracted from the reference [22].
0.3
-0.9
Ces=0.5
0.2 Ces=0.25
Ces=0
Ces=-0.25
0.1 -0.92
Ces=-0.5
∆T
LVDSMC
U
-0.94 Ces=0.5
-0.1
Ces=0.25
Ces=0
LVDSMC Ces=-0.25
-0.2
Ces=-0.5
-0.96
-0.3
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
X X
(a) Kn = 0.1, θ = ϕ
1.5 -0.4
1 Ces=0.5 -0.45
Ces=0.25
Ces=0 -0.5
0.5 Ces=-0.25
Ces=-0.5
-0.55
∆T
LVDSMC
U
0
-0.6
Ces=0.5
-0.5 Ces=0.25
-0.65 Ces=0
LVDSMC Ces=-0.25
-1 Ces=-0.5
-0.7
-1.5 -0.75
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
X X
(b) Kn = 0.2, θ = 4ϕ
-0.45
1
-0.5
Ces=0.5
Ces=0.25
Ces=0
0.5 -0.55
Ces=-0.25
Ces=-0.5
∆T
LVDSMC
U
0 -0.6
Ces=0.5
-0.65 Ces=0.25
-0.5 LVDSMC
Ces=0
Ces=-0.25
Ces=-0.5
-0.7
-1
-0.75
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
X X
(c) Kn = 0.5, θ = 4ϕ
Figure 4: The velocity
√ and temperature profiles for unsteady boundary heating problem at θt = 3π/2. The ϕ is
defined as ϕ = π 2/16.
between two horizontal boundaries. The lower boundary at y = 0 is fully diffusive with a
temperature given by Tw = T0 (1 − 0.5cos(2πx )). An identical boundary is located at y = 1.
0.5 0.5
-0.05
-0.1
-0.1
0.4 -0.02 0.4
05
0 2
-0 .
0 -0.
0.3 0.3
02
2
0.0
-0 .
-0.05
Y
Y
-0.1
-0.1
-0 . -0
2 .2
0.2 0.2
0
2
0
-0.0
05
2
-0.05
0.0
05
0.
2 0. 0.1
1
0.
0.0
0.1 0.1
-0.02
-0 0.2
.3
5
-0.2
-0.05
-0.1
-0
-0.5 0.2
-0.2
.35
0.35 -0 .
-0.1
5 0.35
0
0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
X X
S-model ES-model
Figure 5: The temperature deviation contour of a spatially boundary temperature variation problem. The dash
line is DSMC data extracted from the reference [24].
The Knudsen number based on the separation between the two boundaries is 1. Working
gas is argon with reference viscosity defined by Eq. (4.7). Owing to the symmetries in the
x and y directions, the simulation domain is chosen as [0,1/2]×[0,1/2]. Fig. 5 shows the
temperature deviation contours derived from ES-model and S-model. The temperature
deviation is defined as ∆T = ( T − T0 )/0.5. The background data is extracted from the ref-
erence [24]. Unlike the previous two test cases, ES-model predicts more accurate results
than the S-model does in comparison with DSMC results.
1 1
DSMC DSMC
0.8 0.8
Normalized temperature
S-model S-model
Normalized density
ES-model ES-model
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-20 -15 -10 -5 0 5 10 15 -20 -15 -10 -5 0 5 10 15
X X
Figure 6: The shock structure for ES-model and S-model at Ma = 8 and ω = 0.81.
1 1
Ces = 0.5
Ces = 0.25
Ces = 0.0 (S-model) DSMC
0.8 0.8
Ces = -0.25
Ces = -0.5 (ES-model) Normalized temperature
Normalized density
DSMC
0.6 0.6
0.4 0.4
Ces = 0.5
Ces = 0.25
Ces = 0.0 (S-model)
0.2 0.2 Ces = -0.25
Ces = -0.5 (ES-model)
0 0
-20 -15 -10 -5 0 5 10 15 -20 -15 -10 -5 0 5 10 15
X X
Figure 7: The shock structure for the generalized kinetic model with different Ces at Ma = 8 and ω = 0.81.
Fig. 7 shows the tendency how the shock structure changes while the Ces varies from
−0.5 to 0.5. Note that Pr = 2/3 is fixed in all these results. The generalized kinetic model
presents a set of shock structures with the same Prandtl number. When the Ces = −0.5,
the generalized kinetic model presents the ES-model. As the value of Ces becomes larger,
the temperature profile becomes steeper. Meanwhile, the density profile grows wider.
When Ces is larger than 0, the temperature profile still becomes steepening. But when Ces
exceeds 0.15, the density profile turns out to be twisted near the upstream. Although the
annoyed twisting density profile makes this range of the free parameter unacceptable,
the strong dependence of the Ces is confirmed.
5 Analysis
The individual ES-model and S-model were constructed to set the Prandtl number as a
free parameter. Physically, the Prandtl number of a monatomic gas has a fixed value,
especially in the continuum flow regime. Therefore, the Prandtl number should not be
taken as a free parameter for monatomic ideal gas. With a fixed Prandtl number, theoret-
ically there is not any freedom in the ES-model and S-model. In this study we propose a
generalized kinetic model. Besides a fixed Prandtl number for monatomic gas, the new
model provides one more free parameter. This free parameter can present a continuum
spectrum of kinetic models with a fixed Prandtl number. And it provides ways to mimic
more complicated physical relaxation process. With certain choices of this free parameter,
say Ces , the S-model and ES-model become a subset of the new model.
As mentioned in Section 3, Ces and Cshak are related to the relaxation of moments of the
distribution function and they should behave similar to the counterpart from Boltzmann
equation. To shed light on this topic, we examine the Boltzmann collision term for VHS
molecule. The Boltzmann equation is written as following,
∂(n f ) ∂(n f )
+u· = J ( n f ), (5.1)
∂t ∂x
where Ω is the solid angle for scattering molecule, cr is the relative velocity between
two colliding molecules and σ is the collision cross section. For a spatially homogenous
monatomic gas problem, the moments equation of the Boltzmann equation gives the
relaxation process of the moments. For the quantity Q, the relaxation process can be
written as
m∂⟨n f ,Q⟩
= m∆[ Q], (5.3)
∂t
∫
where ⟨n f ,Q⟩ = n f Qdu and m is the mass of molecule. For example, if Q = u2 and for
Maxwell molecule, i.e., µ ∼ T, the corresponding relaxation equation is
∂P11 ∂p11
= = m∆[u2 ]. (5.4)
∂t ∂t
The collision integral can be obtained explicitly for Maxwell molecule [21] and reads as
following,
∂p11 p
= − p11 . (5.5)
∂t µ
For the other molecules, there is no explicit solution. However, some qualitative results
can be deduced from a given distribution function. Here, we only consider two kinds of
distribution functions for VHS molecule. The diameter of VHS molecule is given by
where υ = ω − 1/2. The first distribution function is the one employed in Grad’s thirteen
moments method [25] and it reads
( P − pI q ( c2 ))
f = M[ f ] 1 + c · ·c+ ·c −1 . (5.7)
2pRT pRT 5RT
Since the quantities, such as the stress and the heat flux, are invariants for Galilean trans-
formation, we consider the situation when U = 0. When (P − pI)/(2pRT ) and q/( pRT )
are much less than 1, namely, in continuum regime, by substituting Eq. (5.7) into collision
integral (Eq. (5.2)), the above distribution function gives
8
√ 4−υ ( RT ) 2 −υ Γ(4 − υ)p,
1
∆[uu] = −(n/m)σre f c2υ
r,re f (5.8a)
15 π
[1 ] 8 2
√ 4−υ ( RT ) 2 −υ Γ(4 − υ) q,
1
∆ uu2 = −(n/m)σre f c2υ
r,re f (5.8b)
2 15 π 3
where Γ denotes the Gamma function. The viscosity of the VHS molecule and the mean
collision rate (1/τ) per molecule in an equilibrium gas of VHS molecules are given by the
reference [21]. Here we reformulate them as following,
√
15m π4υ ( RT ) 2 +υ
1
µ= , (5.9a)
8Γ(4 − υ)σre f c2υ r,re f
1 −υ
√
2 − υ Γ (2 − υ ) /
1
r,re f σre f 4 ( RT )
= 4nc2υ π. (5.9b)
τ
Using the above results, the relaxation process of moments of the Boltzamnn equation
can be written as,
∂n f 1 ∂pij p ∂qi 2p
= − (n f −(τ J (n f )+ n f )), = − pij , =− qi . (5.10)
∂t τ ∂t µ ∂t 3µ
Actually, for VHS molecule in a local equilibrium state, the Ces can be derived as [21],
1 30 p (7 − 2ω )(5 − 2ω )
= , Ces = 1 − . (5.11)
τ (7 − 2ω )(5 − 2ω ) µ 30
Here Ces is confined in a domain of [0.2,0.5] for VHS molecules and can be taken as a
constant. For the force driven Poiseuille flow and unsteady boundary heating problem,
when Ces is in such range, the generalized kinetic model presents better results. But when
the Knudsen number changes, the optimized Ces also changes. We attribute this kind of
observation to the nature of Boltzmann equation.
Hereafter we consider another kind of distribution function and show that Ces is high-
ly dependent on the specific form of distribution function. Assume the distribution func-
tion is composed of two delta function, say,
where u0 = (u0 ,0,0) and α ∈ [0,1], u denotes molecule velocity in x direction. Then the
pressure, stress tensor and heat flux can be expressed by α and u0 ,
1 1
p = mnα(1 − α)u20 , P11 = 3p, q1 = mnα(1 − α)(1 − 2α)u30 . (5.13)
3 2
The collision rate is
1 1−2υ
r,re f α (1 − α ) u0
= 2nπd2re f c2υ , 0 ≤ υ < 1/2, (5.14a)
τ
1
= nπd2re f cr,re f , υ = 1/2. (5.14b)
τ
Note that the 1/τ is not continuous when υ = 1/2, because the collision cross section
between two molecules with identical velocity is infinite when υ = 1/2. It is inappropriate
to count this kind of collision. So we will discuss the case of 0 ≤ υ < 1/2. Substituting
Eq. (5.12) into Eq. (5.2), the collision terms give,
Here, the relaxation process is totally different from the near equilibrium state as shown
before. Ces in this case can be formally written as
Obviously, it is not a constant. Furthermore, it is less than 0 and can even go to minus
infinity. The situation described here always happens in near free molecular flow, espe-
cially for the case where spatial variation is large. As an example in the one dimensional
case, the solution of free molecular flow is,
∂f ∂ f0 ∂ f0
f ( x,u,t) = f 0 ( x − ut,u,t), = −t + .
∂u ∂x ∂u
If the initial condition is nonuniform, discontinuous distribution function will emerge.
For boundary temperature variation problem, the value of Ces is preferred to recover
the ES-model, namely, Ces = −0.5. The Fig. 8 shows the distribution function at (0,0.5).
Based on the above analysis, two peak structure corresponds to a negative value of Ces .
Therefore, we qualitatively conclude that the ES-model is more appropriate for this prob-
lem.
As show above, Ces is not a constant due to the nature of the Boltzmann equation.
Even with distribution function Eq. (5.7), Ces is nonconstant if high order terms are taken
into consideration. With such understanding, we construct a variable Ces in the shock
structure calculation in order to get a good agreement with DSMC. As show in Fig. 9, a
2 0
u
0 -2
v
-2
-4 -4
Figure 8: The distribution function at location (0,0.5) for spatially varying boundary temperature heating
problem.
good shock structure can be obtained and the corresponding Ces is plotted for this calcu-
lation. The stress is normalized by ρ1 RT1 and the heat flux is normalized by ρ1 ( RT1 )3/2 .
The temperature, stress and heat flux profile are improved a lot, while the density pro-
file changes only a little bit. The early raising of temperature profile in the upstream is
suppressed efficiently. In the past years, some similar kinetic models [5, 26] have been
proposed. But people always try to find a universal optimized parameter which might
not exist under current framework of kinetic models. Variables or functional instead of
a fixed parameter might be more suitable for kinetic models. Based on these numerical
results and analysis, we believe that the new free parameter in the generalized kinetic
model has significant physical insight which deserves its further study.
Owing to the existence of negative Ces , the ES-model might be incapable for such
simulation, since the ES-model always constrains Ces in the interval [−0.5,1) to keep a
positive eigenvalue of T. Here, an alternative of Gaussian distribution can be adopted in
the kinetic model,
G[ f ] ≈ M[ f ](1 + c · T′ · c), (5.16)
where
1
Tij′ = Tij , i ̸= j, (5.17)
2( RT )2
and
1
Tij′ = ( Tij − trace(T)/3), i = j. (5.18)
2( RT )2
The lower bound of Ces for Gaussian distribution can be removed. Surprisingly, although
the above expansion cannot guarantee the positivity of the distribution function, the nu-
merical results from the above expansion are very close to that where a full Gaussian
distribution function is used. The increase of the computational cost of the combined
the the ES-model and S-model is only about 15% compared with the S-model. If the ex-
pansion of ES-model is used, the coefficients of different models can be merged before
calculation. So the computational cost is almost identical to that for S-model.
1
16
Normalized density & temperature
DSMC(ρ) 14
0.8
DSMC(T)
S-model(ρ) DSMC
12
Normalized stress
S-model(T)
Generalized(ρ) Generalized
0.6 Generalized(T) 10 S-model
Ces ES-model
8
0.4
6
4
0.2
0 0
-20 -15 -10 -5 0 5 10 15 -20 -15 -10 -5 0 5 10 15
X X
0
Normalized heat flux
-50
-100 DSMC
Generalized
S-model
ES-model
-150
Figure 9: The shock structure from the generalized kinetic model with a variable Ces at Ma = 8 and ω = 0.81.
11
10
8
∆T×1000
5
Kn=0.1 DSMC Ces = 0.2
4 Kn=0.1 ESLBM
Kn=0.1
Kn=0.2 DSMC
Kn=0.2 ESLBM
Kn=0.2
3 Kn=0.5
Kn=0.5 DSMC
Kn=0.5 ESLBM
2
-0.4 -0.2 0 0.2 0.4
X
Figure 10: Comparison of the temperature profiles for low speed Couette flow problem. The DSMC and ESLBM
results are extracted from the reference by [22].
6 Conclusions
In this paper, we have developed a generalized kinetic model through the combination
of the ES-model and the S-model. With a fixed Prandtl number, this new model pro-
vides an additional free parameter, which can be used to cope with the complicated non-
equilibrium flow physics and recover the physical solution more accurately. By adjusting
the free parameter, different relaxation times between different moments of a gas dis-
tribution function can be simulated. With the variation of this free parameter, the new
model covers the BGK model, ES-model and Shakhov model. Meanwhile, it provides a
continuum spectrum of kinetic models with different dynamics.
The numerical study and analysis indicates that, due to the relaxation of Boltzmann
collision term (Eqs. (5.8) and (5.15)), an important property for a kinetic model to capture
physically valid solutions is the ratios between the relaxation rates of different moments
of a gas distribution function. In most cases, the S-model presents reasonable numerical
results. In the near continuum regime, parameters deduced from linear problem is rec-
ommended, such as Eq. (5.11). In transition regime, the optimized parameter is larger
than 0 when isothermal boundary condition is applied, or less than 0 when large tem-
perature gradients exist in the boundary. As shown in Fig. 10, the generalized model can
present accurate results, especially for low speed and weakly non-equilibrium problems.
We believe that the kinetic models presented in this paper, along with the unified
scheme, will be very useful for the microflow studies in vacuum industry.
Acknowledgments
We would like to thank Yonghao Zhang and Lei Wu for their helpful comments and
suggestions. This work was supported by Hong Kong Research Grant Council (621011,
620813) and SRFI11SC05 at HKUST and the National Natural Science Funds for Distin-
guished Young Scholar group under Grant No. 11221061.
References
[1] P. L. B HATNAGAR , E. P. G ROSS AND M. K ROOK , A model for collision processes in gases, Phys.
Rev., 94 (1954), pp. 511–525.
[2] L. H. H OLWAY, New statistical models for kinetic theory: Methods of construction, Phys. Fluids,
9(9) (1966), pp. 1658–1673.
[3] P. A NDRIES AND B. P ERTHAME , The ES-BGK model equation with correct Prandtl number, AIP
Conf. Proc., 585 (2001), pp. 30–36.
[4] E. M. S HAKHOV, Generalization of the Krook kinetic relaxation equation, Fluid Dynamics, 3
(1968), pp. 95–96.
[5] G. L IU , A method for constructing a model form for the Boltzmann equation, Phys. Fluids A, 2
(1990), 277.
[6] P. A NDRIES , J. B OURGAT, P. LE TALLEC AND B. P ERTHAME , Numerical comparison between
the Boltzmann and ES-BGK model for rarefied gases, Comput. Methods Appl. Mech. Eng., 191
(2002), pp. 3369–3390.
[7] V. G ARZ Ó , M. DE H ARO AND M. L ÓPEZ, Kinetic model for heat and momentum transport, Phys.
Fluids, 6 (1994), 3787.
[8] I. A. G RAUR AND A. P. P OLIKARPOV, Comparison of different kinetic models for the heat transfer
problem, Heat Mass Transfer, 46 (2009), pp. 237–244.
[9] Y. Z HENG AND H. S TRUCHTRUP, Ellipsoidal statistical Bhatnagar-Gross-Krook model with
velocity-dependent collision frequency, Phys. Fluids, 17 (2005), 127103.
[10] A. N. K UDRYAVTSEV, A. A. S HERSHNEV AND M. S. I VANOV, Comparison of different kinetic
and continuum models applied to the shock-wave structure problem, AIP Conf. Proc., 1084 (2008),
pp. 507–512.
[11] L. M IEUSSENS , Numerical comparison of Bhatnagar-Gross-Krook models with proper prandtl num-
ber, Phys. Fluids, 16 (2004), 2797.
[12] K. X U AND J UAN -C HEN H UANG , A unified gas-kinetic scheme for continuum and rarefied flows,
J. Comput. Phys., 229 (2010), pp. 7747–7764.
[13] K UN X U AND J UAN -C HEN H UANG , An improved unified gas-kinetic scheme and the study of
shock structures, IMA J. Appl. Math., 76 (2011), pp. 698–711.
[14] J. C. H UANG , K. X U AND P. B. Y U , A unified gas-kinetic scheme for continuum and rarefied flows
ii: Multi-dimensional cases, Commun. Comput. Phys., 12 (2012), pp. 662–690.
[15] J. C. H UANG , K. X U AND P. B. Y U , A unified gas-kinetic scheme for continuum and rarefied
flows iii: Microflow simulations, Commun. Comput. Phys., 14 (2011), pp. 1147–1173.
[16] S. Z. C HEN , K. X U , C. B. L EE AND Q. D. C AI , A unified gas kinetic scheme with moving mesh
and velocity space adaptation, J. Comput. Phys., 231 (2012), pp. 6643–6664.
[17] H. G RAD , On the kinetic theory of rarefied gases, Commun. Pure Appl. Math., 2 (1949), pp.
331–407.
[18] H. S TRUCHTRUP AND M. T ORRILHON , Regularization of grad’s 13 moment equations: Deriva-
tion and linear analysis, Phys. Fluids, 15 (2003), 2668.
[19] K. X U , A gas-kinetic BGK scheme for the Navier-Stokes equations and its connection with artificial
dissipation and godunov method, J. Comput. Phys., 171 (2001), pp. 289–335.
[20] J. M ENG , Y. Z HANG , AND J. M. R EESE , Assessment of the ellipsoidal-statistical Bhatnagar-
Gross-Krook model for force-driven poiseuille flows, J. Comput. Phys., 251 (2013), pp. 383–395.
[21] G. A. B IRD , Molecular Gas Dynamics and the Direct Simulation of Gas Flows, Clarendon
Press, Oxford, 1994.
[22] J. M ENG , Y. Z HANG , N. G. H ADJICONSTANTINOU , G. A. R ADTKE AND X. S HAN , Lattice
ellipsoidal statistical BGK model for thermal non-equilibrium flows, J. Fluid Mech., 718 (2013), pp.
347–370.
[23] T. M. M. H OMOLLE AND N. G. H ADJICONSTANTINOU , Low-variance deviational simulation
Monte Carlo, Phys. Fluids, 19 (2007), 041701.
[24] G. A. R ADTKE , N. G. H ADJICONSTANTINOU AND W. WAGNER , Low-noise Monte Carlo sim-
ulation of the variable hard sphere gas, Phys. Fluids, 23 (2011), 030606.
[25] S. H ARRIS , An Introduction to the Theory of the Boltzmann Equation, Dover Publications,
Inc, 2004.
[26] V. I. Z HUK , V. A. RYKOV AND E. M. S HAKHOV, Kinetic models and the shock structure problem,
Fluid Dynamics, 8 (1973), pp. 620–625.