You are on page 1of 30

FL48CH16-Torrilhon ARI 17 September 2015 15:43

Review in Advance first posted online


V I E W
E on September 25, 2015. (Changes may
R

still occur before final publication

S
online and in print.)

C E
I N

A
D V A

Modeling Nonequilibrium
Gas Flow Based on Moment
Equations
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Manuel Torrilhon
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Department of Mathematics and Center for Computational Engineering Science, RWTH


Aachen University, Aachen 52064, Germany; email: mt@mathcces.rwth-aachen.de

Annu. Rev. Fluid Mech. 2016. 48:429–58 Keywords


The Annual Review of Fluid Mechanics is online at kinetic gas theory, Boltzmann equation, continuum models, rarefied gases,
fluid.annualreviews.org
microflows, channel flows, shock waves
This article’s doi:
10.1146/annurev-fluid-122414-034259 Abstract
Copyright  c 2016 by Annual Reviews. This article discusses the development of continuum models to describe
All rights reserved
processes in gases in which the particle collisions cannot maintain ther-
mal equilibrium. Such a situation typically is present in rarefied or diluted
gases, for flows in microscopic settings, or in general whenever the Knudsen
number—the ratio between the mean free path of the particles and a macro-
scopic length scale—becomes significant. The continuum models are based
on the stochastic description of the gas by Boltzmann’s equation in kinetic
gas theory. With moment approximations, extended fluid dynamic equations
can be derived, such as the regularized 13-moment equations. Moment equa-
tions are introduced in detail, and typical results are reviewed for channel
flow, cavity flow, and flow past a sphere in the low–Mach number setting
for which both evolution equations and boundary conditions are well estab-
lished. Conversely, nonlinear, high-speed processes require special closures
that are still under development. Current approaches are examined, along
with the challenge of computing shock wave profiles based on continuum
equations.

429

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

1. INTRODUCTION
The accurate modeling and simulation of nonequilibrium processes in rarefied gases or microflows
are among the main challenges in fluid mechanics today. The traditional models developed in the
eighteenth and nineteenth centuries are known to lose their validity in extreme physical situations.
In these classical models, the nonequilibrium variables, stress and heat flux, are coupled to gradients
of velocity and temperature, as given in the constitutive relations of Navier-Stokes and Fourier
(NSF). These relations are valid close to equilibrium; however, in rarefied gases or microflows,
the particle collisions are insufficient to maintain equilibrium. Away from equilibrium, the inertia
and higher-order multiscale relaxation in the fluid become dominant and essential.
It is generally accepted that kinetic theory based on a statistical description of the gas provides
a valid framework to describe processes in a rarefied regime or at small scales. Introductions to
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

kinetic theory and its core, the Boltzmann equation, can be found in many textbooks (e.g., Vincenti
& Kruger 1965, Chapman & Cowling 1970, Ferziger & Kaper 1972, Cercignani 1975). The main
variable used to describe the gas is the distribution function or probability density of the particle
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

velocities, which describes the number density of particles with a certain velocity in each space
point at a certain time. However, in many situations, this detailed statistical approach yields a
far too complex description of the gas. It turns out to be desirable to have a continuum model
based on partial differential equations for the fluid mechanical field variables. This model should
accurately approximate the multiscale phenomena present in kinetic gas theory in a stable and
compact system of field equations.
The main scaling parameter in kinetic theory is the Knudsen number, K n = λ/L, computed
from the ratio of the mean free path between collisions λ and a macroscopic length L. Complete
equilibrium is given for K n → 0, described by the dissipationless Euler equations, whereas in
rarefied gases and microflows the Knudsen numbers become significant because of either a large
mean free path or microscopic length scales. In many such processes, gas dynamics with NSF fail
above K n ≈ 0.01, and some examples are discussed below.
Figure 1 shows the range of the Knudsen number. The numbers are only for orientation and
may depend on the process and context at hand. For large Knudsen numbers, the collisions between
the particles can be neglected, and they move ballistically in a free flight. The regime between NSF
and free flight can be split into two parts. One part is the kinetic regime in which the nonequilibrium
is so strong that a detailed description by a distribution function becomes necessary. In this regime,
either the Boltzmann equation is solved directly or the direct simulation Monte Carlo (DSMC)
method is employed (see Bird 1998). The other part is the so-called transition regime in which a
fluid description is still possible, although a larger set of field variables or higher derivatives may

Euler NSF Transition Kinetic Free


equations equations regime regime flight

10 –3 10 –2 10 –1 10 0 10 1 10 2
EQUILIBR
EQUI LI BRIU
IUM N O N E QU
NON Q U IL
I L IBR
I B R IU
IUM Kn

Figure 1
Overview of the range of the Knudsen number and various model regimes. At very low Knudsen numbers, nondissipative flow can be
described by Euler equations. The classical Navier-Stokes-Fourier (NSF) theory of gas dynamics extends only slightly into
nonequilibrium. At large Knudsen numbers, particles move freely. In the middle of the range, there is a kinetic regime and a transition
regime in which continuum models such as moment equations are being developed.

430 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

be needed. In this regime, Boltzmann simulations or the DSMC method becomes increasingly
expensive. The limit for continuum models is currently somewhere between Kn = 0.5 and 1,
depending on the process. The aim of the research field is to establish continuum models and push
their limit further. The transition regime is typically split into a slip regime at lower Knudsen
numbers, K n ≈ 0.05, in which the equations of NSF are augmented with higher-order boundary
conditions (Karniadakis & Beskok 2001). However, as discussed below, with the onset of slip, bulk
effects also occur that cannot be described by advanced boundary modeling only.
This article reviews the achievements of moment equations and, in particular, regularized
moment equations applied to flows in the transition regime. Grad (1949, 1958) introduced the
moment method to kinetic gas theory by using a Hilbert expansion of the distribution function
in Hermite polynomials. The field equations based on moments use an extended set of variables
to describe the state of a gas, starting with the fluid dynamic quantities density, velocity, and
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

temperature and including higher-order moments of the distribution function. In the context of
phenomenological thermodynamics, extended theories proved to be useful for various purposes
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

(e.g., see Jou et al. 1996, Müller & Ruggeri 1998). Struchtrup & Torrilhon (2003) regularized
Grad’s moment equations for the case of 13 variables: density, velocity, temperature, stress de-
viator, and heat flux. Hybrid approaches similar in spirit to regularizing moment equations have
been studied by Grad (1958), Karlin et al. (1998), Jin & Slemrod (2001), and Müller et al. (2003).
The regularization procedure adds second-order derivatives to Grad’s evolution equations that
model the multiscale dissipation of Boltzmann’s equation. The final equations overcome various
deficiencies of the original moment equations such as subshock artifacts.
Using the Chapman-Enskog expansion, one can derive the classical relations of NSF as first-
order corrections (Chapman & Cowling 1970). Second- and third-order corrections have also
been derived (Burnett 1935, Shavaliyev 1993) and provide the Burnett and super-Burnett relations
for the stress tensor and heat flux. Similar to moment equations, Burnett-type equations aim at
describing rarefied gases and microflows as well. However, Burnett equations show instable wave
modes in general (Bobylev 1982, Rosenau 1989, Uribe et al. 2000). They also cannot properly
describe boundary value problems (see Struchtrup 2005b). Still, Burnett-type equations have
subsequently been used to compute rarefied gas flows (e.g., Agarwal et al. 2001, Lockerby &
Reese 2003, Xu 2003, Xu & Li 2004. Bao & Lin 2008, Garcia-Colin et al. 2008). In these works,
Burnett-type equations are often based on kinetic model equations such as BGK, in which case
they are stable (Struchtrup 2005c). Also, several attempts exist to stabilize the Burnett equations
(e.g., see Zhong et al. 1993, Jin & Slemrod 2001, Bobylev 2006, Söderholm 2007).
This article focuses on moment equations but does not try to explain all the details and list
all the references. Further insight and references can be obtained, for example, from a textbook
(Struchtrup 2005c), a special journal issue (Torrilhon 2009), and a review (Struchtrup & Taheri
2011), as well as the many works listed below.

2. REGULARIZED MOMENT EQUATIONS


This section focuses on the successful case of the regularized 13-moment (R13) equations. Some
further examples of moment equations are also discussed in subsequent sections.

2.1. Kinetic Gas Theory


The starting point for any system of moment equations is the kinetic gas theory of monatomic,
ideal gases in which we denote the distribution function by f (x, t, c) describing the probability

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 431

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

density to find a particle at space point x and time t with velocity c. The mass of the particles is
given by m. The distribution function is governed by the Boltzmann equation
∂f ∂f gi ∂ f
+ ci + = S( f, f ), (1)
∂t ∂ xi m ∂ci
where the collision integral is abbreviated by S( f, f ) (Boltzmann 1872, Cercignani 1975). An
external force field is given by gi . Originating from kinetic gas theory, all variables of a moment
theory are based on integral projections of f(x, t, c). The mass density,

ρ(x, t) = m f (x, t, c) d c, (2)
R3

as well as the velocity and internal energy,


 
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

m m 1
vi (x, t) = ci f (x, t, c) d c, ε(x, t) = (ci − vi )2 f (x, t, c) d c, (3)
ρ(x, t) R3 ρ(x, t) R3 2
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

form the basic equilibrium variables of the gas. The pressure is given by p = 23 ρε owing to the
assumption of monatomic gases, and we use the temperature θ = (k/m)T in energy density units
(i.e., p = ρθ). The first two nonequilibrium variables, the stress tensor and heat flux, follow by

σi j (x, t) = m C<i C j > f (x, t, C) d C (4)
R3

and

1
q i (x, t) = m
Ci C 2 f (x, t, C) d C (5)
2
and are based on the microscopic peculiar velocity Ci = ci − vi . In total, we have 13 three-
dimensional fields, and the R13 equations are evolution equations for Equations 2–5.
Indexed quantities are mostly used to describe vectors, tensors, and the summation convention
for operations between them. Bold letters represent vectors in invariant formulations. Angular
brackets denote the deviatoric (trace-free) and symmetric part of a tensor (see also Struchtrup
2005c), and parentheses are used to abbreviate the normalized sum of index-permutated (sym-
metrized) tensor expressions.

2.2. Derivation of Regularized Moment Equations


The order-of-magnitude method presented by Struchtrup (2004, 2005a) provides a framework
to derive closed moment equations. It does not depend on Grad’s closure relations and does not
directly utilize the result of asymptotic expansions. In short, the method finds the proper equations
within a certain asymptotic order of accuracy in the Knudsen number by the following three steps:
1. The determination of the order of magnitude of the moments. This is based on Chapman-
Enskog expansions of the nonequilibrium moments and follows the ideas of Müller et al.
(2003).
2. The construction of a moment set with the minimum number of moments at each order.
This is achieved by introducing new variables by the linear combination of the moments
originally chosen. The new variables are constructed such that the number of moments at a
given order is minimal.
3. The deletion of all terms in all equations that would lead only to contributions of larger
orders than that required in the conservation laws for energy and momentum.
The order-of-magnitude method gives the Euler and NSF equations at zeroth and first order
and thus agrees with the Chapman-Enskog method in the lower orders (see Struchtrup 2004).

432 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

When using collision integrals based on Maxwell molecules, one finds that the second-order
equations are the original Grad’s 13-moment equations, whereas at third order, we arrive at the
regularized 13-moment equations (see Struchtrup 2004). It follows that the R13 system satisfies
some optimality when processes are to be described with third-order accuracy. For linear kinetic
equations, the method has been generalized to the kinetic description of Kauf et al. (2010), and it
is precisely shown how the closure of the distribution function is implied by the kinetic equation
itself.
The original derivation of Struchtrup & Torrilhon (2003) was based on pseudo-timescales
to separate the relaxation of the stress tensor and heat flux from higher-order moments. The
order-of-magnitude approach does not rely on such an assumption; instead, it uses the order of
magnitude of the moments to separate different sets of variables and close the system of moment
equations.
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

2.3. Evolution Equations


The evolution equations for the density, velocity, and internal energy are given by the conservation
laws of mass, momentum, and energy,
Dρ ∂vk
+ρ = 0, (6)
Dt ∂ xk

Dvi ∂p ∂σi j
ρ + + = ρg i , (7)
Dt ∂ xi ∂xj

Dε ∂vi ∂vi ∂q i
ρ +p + σi j + = 0, (8)
Dt ∂ xi ∂xj ∂ xi

including the stress tensor σi j and the heat flux qi . The R13 equations can be viewed as a generalized
constitutive theory for the stress tensor and heat flux, which are given by
∂σi j ∂σi j vk 4 ∂q i ∂vi ∂v j  ∂mi j k
+ + + 2p + 2σki + = −ν σi j (9)
∂t ∂ xk 5 ∂ xj ∂ xj ∂ xk ∂ xk

for the stress and


 
∂q i ∂q i vk ∂vi 5 1 ∂σ j k 1 ∂p 5 ∂θ
+ + qk − p δi j + σi j − σi j + p
∂t ∂ xk ∂ xk 2 ρ ∂ xk ρ ∂xj 2 ∂ xi
  (10)
6 ∂vj 1 ∂R̄ik 2
+ δ(i j q k) + mi j k + = −ν q i
5 ∂ xk 2 ∂ xk 3

for the heat flux using the abbreviation


1
Ri j = 7θ σi j + Ri j + Rδi j . (11)
3
The right-hand sides follow from the collision integral in Equation 1 for Maxwell molecules. The
only remaining parameter is the local mean collision frequency ν of the gas particles, which is
specified below.
The remaining unspecified quantities are the contributions of higher-order moments mijk , Rij ,
and R. Neglecting these expressions turns the system of Equations 6–10 into the classical 13-
moment case of Grad (1949, 1958). Struchtrup & Torrilhon (2003) derived gradient expressions

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 433

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

for mijk , Rij , and R, which regularize Grad’s equations:


p ∂(σi j / p) 20
mi j k = −2 θ + q i σ j k ,
ν ∂ xk 15 p
24 p ∂(q  j / p) 192 20
Ri j = − θ + q i q j  + σki σ j k , (12)
5 ν ∂ xj 75 p 7ρ
p ∂(q k / p) 56 5
R = −12 θ + q k q k + σi j σi j .
ν ∂ xk 5p ρ
The relations given in Equation 12 with Equations 9 and 10 complete the conservation laws in
Equations 6–8 to form the system of regularized 13-moment equations.
The evolution equations above are not in the form of a balance law. However, rewriting the
equations gives
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

∂ ∂
ρ+ ( vk ) = 0, (13)
∂t ∂ xk
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

∂ ∂
(ρvi ) + (ρvi vk + p ik ) = ρg i , (14)
∂t ∂ xk

∂ ∂
( p i j + ρvi v j ) + (ρvi v j vk + 3 p (i j vk) + Mi j k ) = −ν σi j , (15)
∂t ∂ xk
and
∂t Q i + ∂ j (2Q (i v j ) − E vi v j + Mijk vk + 12 ( p i j vk2 + 5θ pδi j + R̄i j )) = −ν(σi j v j + 23 q i ) (16)
as the R13 system equivalent to Equations 6–10. Here, the evolved quantities are the so-called
convective moments. In this formulation, it was more natural to use the pressure tensor p i j =
pδi j + σi j as a variable. Hence, the energy equation and the equation for the stress tensor are
merged into a single equation, and the energy equation may be recovered by taking the trace. The
equation for the heat flux is written as an equation for the total energy flux, Q i = Evi + p ik vk + q i ,
with the total energy E = 12 ρv 2j + 32 p. Additionally, the abbreviation Mi j k = 65 δ(i j q k) + mi j k is used.

2.4. Relation to Classical Fluid Dynamics


The mean collision frequency is given by ν and appears on the right-hand side of the R13 equations.
Introducing the mean free path at a reference state by

θ0
λ0 = , (17)
ν0

we have a fundamental microscopic length scale. The isothermal equilibrium speed of sound θ0
serves as an inherent velocity scale. The Knudsen number is then defined by
λ0
Kn = (18)
x0
with some length scale x0 . In applications close to equilibrium, the Knudsen number is a small
number, K n < 1, which justifies an asymptotic expansion in powers of Kn. The first order of the
expansion for the R13 system gives the relations of NSF of compressible gas dynamics in agreement
with the Chapman-Enskog expansion of the Boltzmann equation itself. After the dimensional
quantities are reinserted, the expressions are
(1) p ∂vi (1) 15 p ∂θ
σi j = −2 and q i = − . (19)
ν ∂ xj 4 ν ∂ xi

434 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Thus, for very small Knudsen numbers, the R13 system will essentially behave like the NSF equa-
tions. We can identify the transport coefficients and relate them to the mean collision frequency
to obtain a practical expression for ν. We assume the viscosity is given by
 s
θ
μ(θ ) = μ0 (20)
θ0
with a viscosity exponent 1/2 ≤ s ≤ 1. Hence, we may use
p θs
ν(θ ) = = ρ θ 1−s 0 (21)
μ(θ) μ0
for the collision frequency that leads to asymptotically correct transport coefficients. The def-

inition of the collision frequency leads to the expression λ0 = μ0 /(ρ  0 θ0 ) for the mean free
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

path. Other authors (e.g., Chapman & Cowling 1970) use λ0 = (4/5) (8/π )λ0 . The definition in
Equation 17 is employed to avoid additional factors in the dimensionless equations.
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Sometimes it is asked which transport coefficients are used in the R13 system. However, the
term transport coefficients (i.e., coefficients in Equation 19) is not defined for the system because
the relations in Equation 19 are entirely substituted by the evolution equations (Equations 9 and
10).
Further computations of the coefficients of the asymptotic expansion of the R13 system (Equa-
tions 6–12) show that the second- and third-order coefficients also match those of the Chapman-
Enskog expansion of the Boltzmann equation, at least in the case of Maxwell molecules (see
Torrilhon & Struchtrup 2004). Thus, any R13 result will asymptotically differ from a full
Boltzmann simulation for Maxwell molecules only with an error of O(K n4 ). By construction,
the expansion considers only bulk flow without the influence of the boundary. For boundary
layer–dominated processes, the error may be considerably larger, as discussed below.
Torrilhon & Struchtrup (2004) studied the stability of the R13 system based on linear waves.
In contrast to Burnett equations, the R13 system shows full stability for all wave numbers. The
instability of the Burnett system indicates problems with the unreflected usage of the Chapman-
Enskog expansion. However, the NSF system is stable; hence, the expansion technique must not
be completely abandoned. There is reason to believe that the first expansion steps are more robust
over the higher-order contribution. However, for some collision models, namely BGK relaxation,
the Burnett equations also show linear stability (Struchtrup 2005c). A study about the quality of
the failure of the Chapman-Enskog expansion is available in Rosenau (1989).

2.5. Further Systems of Moment Equations


The presented regularized 13-moment equations have been derived for monatomic ideal gases that
interact with the Maxwell potential. They slightly differ from the original equations in Struchtrup
& Torrilhon (2003) because of the use of the order-of-magnitude derivation (Struchtrup 2004),
which presents a more justifiable approach from today’s perspective. Additionally, the insights of
Rana et al. (2013) and Torrilhon & Struchtrup (2008a) about nonlinear boundary value problems
led to minor changes in which higher-order terms have been altered such that the overall accuracy
of the equations remains unchanged.
The order-of-magnitude approach allows one to derive macroscopic transport equations in
many different settings. Struchtrup (2005a) demonstrated how Grad’s 13-moment equations are
generalized to an arbitrary interaction potential. Recently, Struchtrup & Torrilhon (2013) derived
the linearized R13 equations for hard spheres, and Rahimi & Struchtrup (2014) presented a reg-
ularized moment theory for polyatomic gases. A first step toward regularized moment equations

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 435

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

for rarefied gas mixtures was presented by Gupta & Torrilhon (2015), who used a precise inte-
gration of the Boltzmann collision operator (Gupta & Torrilhon 2012). Similarly, Magin et al.
(2010) pioneered regularized moment equations for plasma flows, based on the plasma moment
equations of Zhdanov (2002). All these new sets of equations still need to be evaluated in detail.
Additionally, suitable numerical methods to solve moment equations are still being developed.
Finite volume approaches have been presented by Brown (1999), Torrilhon (2006), and Cai & Li
(2010), and finite differences have been used by Gu & Emerson (2007) and Ivanov et al. (2013).
The regularization technique has also been applied to other moment systems, such as the 10-
moment system by McDonald & Groth (2008). A larger system of regularized moment equations
is given by the R26 theory, which was derived and studied by Gu & Emerson (2009) and Gu
et al. (2010). The G26 results show several improvements over the R13 system at the expense of
extended equations. It is also possible to numerically generate large systems of regularized moment
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

equations, as demonstrated by Cai & Li (2010) and Cai et al. (2012). In this way, it is possible to
approximate the actual solution of the Boltzmann equation quite accurately.
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Moment equations without regularization have been used in various settings (e.g., see Reinecke
& Kremer 1990, Reitebuch & Weiss 1999, Au et al. 2001, Struchtrup 2002), especially in the
context of extended thermodynamics (Müller & Ruggeri 1998). Recent developments include,
for example, a theory of boundary conditions based on fluctuations (Barbera et al. 2004), which
is also applied to gas mixtures (Barbera & Brini 2011). A variant of the moment method has
been proposed by Eu (1980) and is used, for example, by Myong (2001, 2011). A computational
approach of Seeger & Hoffmann (2000) is the use of cumulant theory.

3. MODELING OF BOUNDARY CONDITIONS


Compared to hydrodynamics with the laws of NSF, the R13 equations have an extended set of
variables and corresponding equations. In fact, new boundary conditions need to be derived from
kinetic gas theory.
The lack of boundary conditions was seen as a major obstacle in the use of moment equations,
even though wall conditions derived from kinetic gas theory have been used by Goldberg (1954)
and Grad (1958) and also by Marquez & Kremer (2001). In recent years, this approach was used
by Gu & Emerson (2007) for regularized equations and was subsequently extended and refined
by Torrilhon & Struchtrup (2008a).

3.1. Maxwell’s Accommodation Model


For consistency with the characteristic theory of Boltzmann’s equation, it is required that only the
incoming half of the distribution function (with n · c > 0, n the wall normal pointing into the gas)
be prescribed. The other half is given by the process in the interior of the gas. The most common
boundary condition for the distribution is Maxwell’s (1879) accommodation model. It describes
the velocity distribution function f˜(c) in the infinitesimal neighborhood of the wall by

χ fW (c) + (1 − χ ) fgas
(∗)
(c) n · (c − vW ) > 0
f˜(c) = (22)
fgas (c) n · (c − vW ) < 0,
where n is the wall normal pointing into the gas. The accommodation coefficient χ ∈ [0, 1] gives
the fraction of the particles from the wall that have been accommodated and injected into the gas
with the distribution
 
ρW /m (c − vW )2
fW (c) = √ 3
exp − , (23)
2πθW 2θW

436 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

where θ W and vW are the known temperature and velocity of the wall. We assume that the wall is
only moving in the tangential direction such that vW has no normal component. The wall density
ρ W follows from particle conservation at the wall. The remaining fraction (1 − χ ) of the particles
are specularily reflected. Because the particles that hit the wall are described by a distribution
(∗)
function fgas , the reflected part will satisfy an analogous distribution function fgas , which follows
from fgas with accordingly transformed velocities. The case of χ = 0 (specular reflection) represents
the generalization of an adiabatic wall (no heat flux, no shear stress) to the kinetic picture.
This boundary condition is used in the majority of simulations based on the Boltzmann equation
or DSMC technique (see Bird 1998). A more refined boundary condition is given, for example,
by the Cercignani-Lampis model (see Cercignani & Lampis 1971).

3.2. Wall Conditions for R13


Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

As a system of balance laws, the regularized 13-moment equations require boundary conditions
for the normal component of the fluxes. However, a detailed investigation shows that not all
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

normal components can be prescribed (see Torrilhon & Struchtrup 2008a). Only those fluxes
with odd normal components are possible, which are vn , σt1 n , qn , mnnn , mt1 t2 n , and Rt1 n , where n
denotes a single normal tensor component and ti denotes two tangential components based on
any orthonormal tangential boundary vector with respect to the wall.
From the Maxwell accommodation model, we obtain boundary conditions for these moments:

vn = 0,
σt1 n = −β(PV t1 + 12 mt1 nn + 15 q t1 ),
q n = −β(2Pθ + 5
28
Rnn + 1
15
R + 12 θ σnn − 12 PV 2t ),

mnnn = β( 52 Pθ − 1
14
Rnn + 1
75
R − 75 θ σnn − 35 PV 2t ),
(24)
σnn
mt1 t1 n + mnnn
2
= −β( 14
1
(Rt1 t1 + Rnn
2
) + θ(σt1 t1 + 2
) − 12 P (V t1 V t1 − V t2 V t2 )),

mt1 t2 n = −β( 14
1
Rt1 t2 + θ σt1 t2 − PV t1 V t2 ),
Rt1 n = β(Pθ V t1 − 12 θ mt1 nn − 11
5
θ q t1 − PV 2t V t1 + 6PθV t1 ),

where θ = θ − θW is the temperature jump at the wall, and Vt = vt − vtW is the tangential slip
velocity. Additionally, we abbreviate
σnn Rnn R
P := ρ θ + − − (25)
2 28θ 120θ

and use the modified accommodation coefficient β = χ /(2 − χ ) 2/(π θ ). When neglecting the
contributions of the moments mi j k and Rij , the first three conditions can also be used for the
system of NSF, as higher-order jump and slip conditions (e.g., see Karniadakis & Beskok 2001).
In principle, the accommodation coefficients that occur in every expression in Equation 24
could be chosen differently, as suggested by Grad (1958). Different accommodation coefficients of
the single moment fluxes (e.g., shear or heat flux) could be used to model detailed wall properties.
However, more investigations and comparisons are required for such an approach. In the results
below, χ = 1 is used.

4. STEADY-STATE, LOW–MACH NUMBER FLOWS


Many nonequilibrium gas situations occur in micromechanical systems (Gad-el-Hak 2002) in
which the dimensions are so small that the Knudsen number becomes significant. At the same

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 437

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

time, these processes are slow and often steady such that time-independent linearized equations
are sufficient. To linearize the R13 equations, we consider a ground state given by ρ 0 , θ 0 , and p0 ,
in which the gas is at rest and in equilibrium with the vanishing stress deviator and heat flux. The
linearized equations describe deviations from this ground state, and the variables (ρ, v, θ, σ, q)
are used for these deviations.
The steady, linearized conservation laws (Equations 6–8) of mass, momentum, and energy are
then given by
∇ · v = 0, ∇ p + ∇ · σ = ρ0 g, ∇ · q = 0, (26)
and the R13 expression in Equations 9–12 gives the constitutive relations
  
4 p0 2μ0 6 1
(∇q)sym + 2 p 0 (∇v)sym = − σ + σ + (∇(∇ · σ))sym − ∇ · (∇ · σ)I , (27)
5 μ0 3ρ0 5 3
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

5 2 p0 6μ0
θ0 ∇ · σ + p 0 ∇θ = − q+ q, (28)
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

2 3μ0 5ρ0
where the parameter μ0 is the viscosity of the gas at the ground state. The notation (A)sym is used
to denote the symmetric part of a matrix A. Note that ∇(∇ · σ) is the gradient of the divergence
of the stress deviator and, as such, a matrix. In contrast to the presentation of the R13 equations
above, here the relation for the higher-order fluxes is inserted directly into the equations for the
stress and heat flux. The boundary conditions in Equation 24 can be linearized in the same way.
When using the laws of NSF (Equation 19) in the conservation laws (Equation 26), we arrive
at
∇ · v = 0, ∇p = μ0 v, θ = 0, (29)
that is, the Stokes equations and a separate Laplace equation for the temperature. The steady,
linear R13 equations can be viewed as an extension of the Stokes equations that couple the flow
and temperature problem.

4.1. Understanding Microchannel Flow


The distance of 10 mean free paths correspond to 1 µm in argon at atmospheric conditions.
Hence, on the micrometer scale, the Knudsen number for flows of argon can easily be as large as
≈ 0.1. Channels are also the main building blocks of microflows (see Karniadakis & Beskok 2001,
Gad-el-Hak 2002).
Channel flows in micro- or rarefied settings have been studied widely. They exhibit a variety
of phenomena attributed to the nonequilibrium, such as an increase of the normalized mass flux
in Poiseuille flow when reducing the channel size beyond a certain value (Knudsen 1909, Deissler
1964, Ohwada et al. 1989, Hickey & Loyalka 1990, Lockerby et al. 2004, Hadjiconstantinou
2006) or Knudsen boundary layers (see Ohwada et al. 1989, Risso & Cordero 1998, Sone 2002).
To capture the effects of Knudsen layers, some author suggest the use of specifically designed
wall functions that influence the flow depending on the distance from the wall (Reese et al. 2007,
Lockerby & Reese 2008).
Knudsen layers have been studied in the regularized 13-moment equations (Struchtrup &
Thatcher 2007) and with the inclusion of boundary conditions (Gu & Emerson 2007, Torrilhon &
Struchtrup 2008a, Young 2011). In a model study using many moments, Struchtrup (2008) showed
how the Knudsen layer is approximated in moment theories. The R13 equations have also been
applied to Poiseuille and Couette flows, as well as transpiration flows (Struchtrup & Torrilhon

438 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

2008, Torrilhon & Struchtrup 2008b), circular geometries (Taheri & Struchtrup 2010a), and
an unsteady setup (Taheri et al. 2009a). Channel flow with larger moment systems has been
investigated (Gu & Emerson 2009, Gu et al. 2010). Analytical studies (e.g., Taheri et al. 2009b)
have been especially helpful in understanding rarefied flow phenomena. Flows induced in channels
by temperature gradients or transpiration flows have been studied by Sone (2002), Sharipov &
Seleznev (1998), and Taheri & Struchtrup (2010b) on the basis of regularized moment equations.

4.1.1. Phenomena. We let the channel flow be between two infinite plates at a distance L.
The x axis is placed in the middle of the channel such that the walls reside at y = ±L / 2. The
Knudsen number is based on the channel width L. It is assumed that the plates are at rest and equal
temperature. A constant driving force F is given in the direction of the channel. As a result of this
setup, all fields depend only on the coordinate y. All variables are scaled with ρ 0 and appropriate
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.


powers of θ0 , and space is scaled by the channel width L. The only parameters are the Knudsen
number and the dimensionless force F̂ = FL/(ρ0 θ0 ). Figure 2 shows an exemplary result for this
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

channel setup based on a DSMC calculation (Zheng et al. 2002). The Knudsen number in this case
is K n ≈ 0.1. Both the velocity and temperature show strong jumps at the boundary. Additionally,
the figure shows the fields of the heat flux qx parallel to the wall, that is, along the channel, and
the total pressure p.
Interestingly, the parallel heat flux is nonzero, contradicting Fourier’s law as there is no tem-
perature difference along the channel. This has also been reported by Baranyai et al. (1992), Todd
& Evans (1995, 1997), Uribe & Garcia (1999), and Aoki et al. (2002). Furthermore, the pressure
shows a nonconstant behavior across the channel, even though there is no flow occurring across
the channel, as discussed by Tij & Santos (1994), Mansour et al. (1997), Tij et al. (1998), Uribe
& Garcia (1999), and Aoki et al. (2002). Finally, the temperature field in Figure 2b also shows
an interesting effect. Owing to dissipation inside the flow, the gas in the channel is heated. This
is a nonlinear effect and is small because of the low speed of the flow. Fourier’s law predicts a
convex curve for the temperature, whereas at larger Knudsen numbers the temperature becomes
nonconvex with a dip in the middle of the channel (see also Alaoui & Santos 1992, Tij & Santos
1994, Mansour, Babras & Garcia 1997, Tij et al. 1998, Aoki et al. 2002, Zheng et al. 2002, Xu
2003, Xu & Li 2004). The normal heat flux (not shown in the figure) still shows only a single zero
value in the middle of the channel, and heat is flowing exclusively from the middle to the walls
against the temperature gradient.
We proceed with studying channel flow on the bases of the regularized 13-moment equations.
This will shed additional light on the intriguing phenomena (see also the sidebar, Added Value of
Continuum Equations).

4.1.2. R13 velocity field for Poiseuille flow. The flow field is governed by two variables: the
velocity profile along the channel vx and the shear stress σxy . The relevant component of the
momentum balance in Equation 26 reads

∂ y σxy = F, (30)

and for σxy , we have from Equation 27


2 1
∂ y q x + ∂ y vx = − σxy , (31)
5 Kn
where the second-order derivatives cancel because of geometry. The underlined terms represent
the law of Navier-Stokes. However, in the R13 equations, a gradient of the tangential heat flux qx

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 439

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Velocity Temperature
0.5
a b

y/L 0

)
slip
(no ip)
NSF no sl
NSF (
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

–0.5
0 0.2 0.4 0.6 1.00 1.02 1.04
vx /(θ0 )1/2 θ/(θ0) 3/2
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Parallel heat flux Total pressure


0.5
c d

y/L 0

–0.5
–0.06 –0.04 –0.02 0 0.02 0.99 1.00 1.01
qx /(p0 θ01/2 ) p/p0

Figure 2
Exemplary fields in acceleration-driven Poiseuille flow for moderate Knudsen numbers (K n ≈ 0.1) taken
from direct simulation Monte Carlo data (Zheng et al. 2002): (a) velocity, (b) temperature, (c) parallel heat
flux, and (d ) total pressure. The red diamonds indicate the fields, and for the velocity and temperature
(panels a and b), the result of classical Navier-Stokes and Fourier (NSF), without slip boundary conditions, is
shown as a dashed blue line for comparison. The data show various rarefaction effects, such as a nonconvex
temperature profile and nonuniform pressure, as well as nongradient phenomena such as a parallel
conductive heat flux without a temperature gradient.

influences the shear stress. Consequently, this heat flux influences the profile of velocity, which
can be computed as
 
F 1 2
vx (y) = C1 + − y 2 − q x (y) (32)
2K n 4 5

with a integration constant C1 to be fixed by boundary conditions. In this result, the classical
parabolic profile (underlined) is perturbed by the tangential heat flux. For constant dimensionless
force F, the amplitude of the parabolic profile decreases with higher Knudsen number because of
viscous friction. There is no reason to assume a vanishing tangential heat flux; instead the R13

440 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

ADDED VALUE OF CONTINUUM EQUATIONS

Direct solutions of Boltzmann’s equation or stochastic particle methods such as DSMC give very accurate results
in principle for the whole range of the Knudsen number. In these simulations, the computational expense remains
a major drawback and is considered as an important argument for the use of continuum models such as moment
equations in the transition regime. However, there is another interesting fact that is often underestimated. Nonequi-
librium flows exhibit behavior that contrasts with the intuition of fluid dynamics and is difficult to understand. The
simulations of Boltzmann and DSMC may give accurate predictions of this behavior, but it is important to realize
that their results actually do not easily provide a detailed physical understanding of the process.
The reason is that our understanding of physical phenomena is often based on the structure and qualitative
behavior of mathematical equations and their solutions. Many aspects of fluid dynamics (e.g., convection, viscosity,
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

heat conduction) are closely linked to particular aspects of the equations describing the flow. Our intuition about
these phenomena arises from analytical solutions for small models that display how the coupling of the mathematical
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

terms influences the behavior of the solution. The mere simulation result of Boltzmann or DSMC does not allow this
kind of understanding. The advantage of continuum equations is that they come with manageable partial differential
equations that allow one to identify physical effects and coupling phenomena that occur in nonequilibrium flows.
Analytical solutions for model situations provide physical insight into the intriguing flow patterns and help us
understand rarefaction effects.

expression in Equation 28 for qx reads


2 1 6
∂ y σxy = − q x + K n ∂y y q x , (33)
3 Kn 5
where the derivative of the shear stress can be substituted with the force F. After multiplication
with Kn, this equation shows that the magnitude of qx reduces with Kn, and hence, the tangential
heat flux is a nonequilibrium effect. The equation for qx allows for nontrivial solutions without
the presence of a temperature gradient. The result is a nongradient transport effect. The solution
for qx is
 
3 5/9
q x (y) = − K n F + C2 cosh y , (34)
2 Kn
where axisymmetry is assumed. The hyperbolic cosine gives exponential layer solutions of the
form exp(±y/K n). These parts of the solution decay toward the interior and are identified with
the Knudsen layer at the boundary. The boundary conditions fix the amplitude C2 of the Knudsen
layer.
The solution for qx gives a negative offset in the middle of the channel and layer behavior at the
boundaries. This is precisely the behavior of the parallel heat flux in Figure 2c. The bulk solution
is directly related to the force. The integration constant will be proportional to Kn such that qx
vanishes in near-equilibrium flows. However, for larger Knudsen numbers, the exponential layers
from both boundaries grow into each other, and qx shows a nonlinear behavior everywhere in
the channel. The complete solution of qx is superimposed to the velocity profile in Equation 32,
which leads to a term growing with larger values of Kn. Hence, from a certain value of the Knudsen
number, the amplitude of the velocity profile will increase, which gives rise to the Knudsen paradox.
The profile of qx also inherits its Knudsen layer to the velocity profile; however, because of the
parabolic behavior, these layers are hardly visible in Figure 2. As boundary conditions, the flux of

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 441

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

momentum σxy and the flux of heat flux Rxy are prescribed. These are two conditions for the two
integration constants C1,2 . The first one represents the slip condition for vx . Full expressions for
these constants are not given here.

4.1.3. R13 temperature profile for Poiseuille flow. To investigate the temperature increase,
we insert the dissipation rate r computed from the classical part of the flow field (Equation 32),
F2 2
r(y) = −σxy ∂ y vx = y , (35)
Kn
as the heat source in the energy balance in Equation 26, but the rest of the equations remain linear
for simplicity. That is, we have the energy balance

∂ y q y = r(y), (36)
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

and the R13 expression in Equation 28 for the normal heat flux reads
5 2 1
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

∂y σy y + ∂y θ = − qy , (37)
2 3 Kn
which contains Fourier’s law (underlined terms). Analogous to the flow equations above, the heat
flux couples to an atypical quantity, which is the normal stress σ y y . Any result for the normal stress
is superimposed on the temperature solution according to
1 F2 4 2
θ(y) = C3 − y − σ y y (y), (38)
45 K n2 5
where the underlined term is the classical convex temperature profile in Poiseuille flow due
to dissipation. It is quadratic in the force, which identifies the temperature increase as a non-
linear effect. There is no reason to assume σ y y = 0; instead, the R13 equation for the stress
(Equation 27) gives
4 1 6
∂y q y = − σy y + K n ∂y y σy y , (39)
5 Kn 5
which has the same structure as the equation for the tangential heat flux above. The solution for
σ yy reads
 
48 2 2 4 2 2 5/6
σyy (y) = − F K n − F y + C4 cosh y . (40)
25 5 Kn
Again, the hyperbolic cosine is identified as the Knudsen layer. The normal stress enters the
temperature profile in Equation 38 such that the classical downward quartic shape competes with
an upward parabola. The quartic is reduced at larger Knudsen numbers; thus, the temperature
profile becomes dented in the middle, and a nonconvex shape arises (see Figure 2b). It is easy to
check that the normal heat flux qy exhibits a single zero in the middle for all Knudsen numbers.
Hence, within the dent of the temperature profile, heat flows from cold to warm, in contradiction
to Fourier’s law—which of course is not valid at these Knudsen numbers.
Finally, the pressure profile follows from the momentum balance as p(y) = p 0 − σ y y (y) such
that the shape of the total pressure is an upright parabola. The Knudsen layers produce downward
kinks toward the boundaries. This precisely corresponds to the shape of p given in Figure 2d.

4.2. More Slow-Flow Examples


The R13 equations can also be solved for more complex slow flows possibly using numerical
methods. Several examples are discussed below.

442 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

4.2.1. Complete Poiseuille flow. Figure 3 presents the complete solutions of the nonlinear
R13 equations from Taheri et al. (2009b) for various Knudsen numbers, with a dimensionless
acceleration force fixed at F = 0.2355. This value is chosen such that a Knudsen number Kn = 0.072
reproduces the case of Poiseuille flow calculated in Zheng et al. (2002) by the DSMC method.
The figure displays the results for various Knudsen numbers together with the DSMC solution
for Kn = 0.072. Very good agreement is observed in most fields. Deviations at the boundary in the
parallel heat flux qx and the normal stress σ y y result from a low-order approximation of the Knudsen
layer. This can be improved in a larger moment theory (see Gu et al. 2010). The mismatch of the
maximum temperature possibly arises from the DSMC result being based on hard spheres and
R13 using Maxwell molecules.
From the velocity fields at different Knudsen numbers, we can compute the mass flow rate as
shown in Figure 4a. The plot compares the mass flow rate of a simple Navier-Stokes computation
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

with no-slip boundary conditions and with results obtained from a linear Boltzmann solution
(Ohwada et al. 1989). The R13 equations predict a Knudsen minimum at Kn = 0.5, slightly below
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

the prediction of Boltzmann. The curve from the moment model then rises too strongly and
loses accuracy at approximately Kn = 1.0. The Navier-Stokes mass flux can be improved to exhibit
a Knudsen paradox by using higher-order jump and slip conditions (see Karniadakis & Beskok
2001, Hadjiconstantinou 2006, Struchtrup & Torrilhon 2008). However, no simulation based on
classical field equations can reproduce any of the transport phenomena discussed above.

4.2.2. Flow past a sphere. The slow flow around a sphere is a classical example for exterior flow.
In classical hydrodynamics, an analytical result for the flow field was given by Stokes and became
standard material in textbooks on fluid dynamics. It is also a popular test case for rarefied or
microflows (see Barber & Emerson 2003, Bailey et al. 2005). Torrilhon (2010b) demonstrated that
an analytical result is also possible for the linearized R13 equations (Equations 26–28), following
Goldberg’s (1954) work on Grad’s 13-moment equations. Similar results have been obtained for
flow around a cylinder based on the R13 system (Westerkamp & Torrilhon 2012). The analytical
solution allows a detailed study of the rarefaction effects in a realistic flow situation. For example,
the regularized 13-moment equations add Knudsen layers to the flow and also influence the bulk
solution considerably. Within these new contributions, the coupling of the fields is counter-
intuitive, such that the velocity couples to the heat flux, and temperature couples to the stress
tensor.
Figure 5a shows the streamlines and the speed contours (the amplitude of velocity) of the R13
result for the case of Kn = 0.3, and Figure 5b shows heat flow lines together with the temperature
contours. It is an interesting aspect of the rarefied solution that it predicts temperature variations
and heat flow in the flow around a sphere even though the temperature of the sphere and that of
the flow at infinity are the same. The flow is based on linearized equations, and hence, quadratic
energy dissipation through shear is also not present. The temperature variations predicted by the
R13 system are the result of the coupling of the temperature and shear stress, which take effect
for larger values of the Knudsen number. Such a linear temperature polarization is also known
for rarefied or microflows around obstacles, and the result of the R13 equations is in agreement
with, for example, that from Aoki & Sone (1987) and Takata et al. (1993) based on the linearized
Boltzmann equation. On top and below the sphere, the heat flow forms recirculation areas, which
indicate the Knudsen layer. Outside this layer, the heat flux essentially flows from the back to the
front of the sphere. In this part of the flow, the heat flux is essentially coupled to the velocity.
Consequently, the heat flux behaves independently from the temperature gradient, and heat may
flow from cold to warm.

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 443

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

0.6
1.04

0.4
vx /(θ0 )1/2

θ/(θ0) 3/2
1.02
0.2

0 1.00
0.04
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Kn
0.1 0.072
0.072 (DSMC) 0.02
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

0.15
0.4

qy /(p0 θ01/2 )
1.0
σ/p0

0 0

–0.02
– 0.1

–0.04
0

0.02
qx /(p0 θ01/2 )

σyy/p0

0 –0.01

– 0.02

–0.02
–1/2 0 1/2 –1/2 0 1/2
y/L y/L

Figure 3
Results of nonlinear regularized 13-moment equations for slow Poiseuille flow using data from Taheri et al. (2009a) with Knudsen
numbers Kn = 0.072 (red ), 0.15 (orange), 0.4 (blue), and 1.0 ( purple). The gray circles give a direct simulation Monte Carlo (DSMC)
result for Kn = 0.072 as a reference. Figure adapted with permission from Taheri et al. (2009a). Copyright 2009, AIP Publishing LLC.

Figure 4b shows the force on the sphere over a range of Knudsen numbers in comparison to
experimental data. Note that the classical Stokes result for the force would be a straight line on
the top of the diagram. Interestingly, the Stokes result can only be slightly improved by using
higher-order boundary conditions, an approach that works best in simple geometries. The R13
results decrease to zero for an infinite Knudsen number capturing the overall behavior correctly.
Quantitative accuracy is lost at approximately Kn = 0.5.

444 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Mass flow rate in Poiseuille flow Force of flow past a sphere


1.6 1.0
Clas
a b sica
l (firs
0.8 t-or
der B
1.4 C)
C l as
sical (hybrid BC)
0.6
J F
J0 1.2
R13
FStokes
0.4 Experimental
Cl R13
as
sic
0.8 al (
no
sl 0.2
ip) Boltzmann

0.4 0
0.05 0.1 0.2 0.5 1.0 2.0 0.01 0.02 0.05 0.1 0.2 0.5 1.0 2.0 5.0 10.0
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Kn Kn

Figure 4
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

(a) Mass flow rate for acceleration-driven Poiseuille flow. Shown is a comparison of regularized 13-moment equations with the classical
fluid dynamic result and results from the linearized Boltzmann equation ( gray diamonds). The mass flux shows a minimum (Knudsen
paradox). Panel a adapted with permission from Taheri et al. (2009a). Copyright 2009, AIP Publishing LLC. (b) Force on a sphere in
uniform slow flow. Shown is a comparison of regularized 13-moment equations with experimental data (dark yellow circles) and classical
Stokes flow with higher-order boundary conditions (BC). The original force formula of Stokes does not decay to zero. Panel b adapted
with permission from Torrilhon (2010b). Copyright 2010, AIP Publishing LLC.

Streamlines/speed contours Heat flow lines/temperature contours


3
a b

0.1 0.010 –0.010


y/R 0 0.2 0.007 –0.007
0.3 0.4 0.005 –0.005
0.5 0.003 –0.003
–1 0.6
0.002 –0.002

0.7
0.001 –0.001
–2

0.8
–3
–3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3
x/R x/R

Figure 5
Results of the regularized 13-moment equations for the slow flow past a sphere of radius R at Kn = 0.3. (a) Streamlines of the velocity
field and speed contours showing strong slip at the sphere. (b) Heat lines from the heat flux field and temperature contours showing
heat circulation in the Knudsen layers and nongradient transport as heat is flowing from cold to warm. Figure adapted with permission
from Torrilhon (2010b). Copyright 2009, AIP Publishing LLC.

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 445

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

NSF DSMC Regularized 13 moments θ/θ0


1.0 1.013

0.8

0.6
y/L
0.4

0.2

0
0.995
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
x/L x/L x/L
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Figure 6
Heat flow lines from the heat flux field and temperature contours in a driven cavity of edge length L at Kn = 0.08 for the regularized
13-moment equations compared with the Navier-Stokes-Fourier (NSF) and direct simulation Monte Carlo (DSMC) results (Rana
et al. 2013). The top wall is moving with M = 0.2. All walls have the same temperature. Figure adapted from Rana et al. (2013), with
permission from Elsevier.

4.2.3. Cavity systems. A driven cavity is a classical multidimensional test scenario in computa-
tional fluid dynamics. We first consider a square cavity in two dimensions in which the top wall
is moving to the right such that a gas inside is set into rotational motion. The dimensions of
the cavity usually imply a Knudsen number, whereas a Mach number is based on the prescribed
velocity of the top lid. Rana et al. (2013) simulated this setup based on the R13 system and
compared the solutions to DSMC results. Besides the flow field and shear stresses, it is interesting
to study the temperature field and heat flow induced by the dissipation in a driven cavity. Figure 6
shows the results of the R13 system for Kn = 0.1 and M = 0.1 compared with an NSF theory and
a DSMC simulation as a reference. The R13 results match the reference very well, especially the
heat flow. In the NSF results, the heat lines are orthogonal to the temperature contours, but in
both the DSMC and R13 results, the heat flux behaves largely independent of the temperature.
In fact, the heat flows from cold to warm.
Heat conduction and convection are crucial in micromechanical systems. Rana et al. (2015)
considered a two-dimensional heat-conducting setup in which three sides of a box are kept at a
specific temperature and the bottom plate is at a temperature twice as high. Gravitational forces
are neglected. One of the results at Kn = 0.13 is shown in Figure 7 in which the solutions of NSF,
DSMC as a reference, and the regularized 13-moment equations are compared. Higher-order
boundary conditions are used in the NSF result, which induce fluid motion owing to transpiration
or creep flow. The NSF result predicts a flow pattern of two vortices, in complete disagreement
with the reference solution. Both the DSMC and R13 results predict a more complex vortex
structure. In particular, the gas is pushed out from the middle at the top and moves up at the sides.
The R13 system also correctly predicts lower temperatures toward the lower end of the sides.

5. NONLINEAR PROCESSES
The order-of-magnitude method used to derive regularized moment equations assumes linearity
assumptions at crucial stages. The flux terms of the equations obtained after dropping second-
order derivatives and leaving out relaxation terms strongly resemble Grad’s moment equations.

446 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

NSF DSMC Regularized 13 moments θ/θ0


1.0
1.731

0.8

0.6
y/L
0.4

0.2

1.138
0
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
x/L x/L x/L
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Figure 7
Velocity streamlines and temperature contours of a heated cavity at Kn = 0.13 obtained with the regularized 13-moment equations
compared to the Navier-Stokes-Fourier (NSF) and direct simulation Monte Carlo (DSMC) systems (Rana et al. 2015). In this cavity, all
walls are at rest, but the walls are kept at different temperatures. The temperature ratio between the lower and the top and side walls is
2. Figure adapted from Rana et al. (2015).

Consequently, the R13 equations suffer from deficiencies similar to those of Grad’s 13-moment
system, especially with regard to nonlinear processes.

5.1. Discussion of Grad’s Closure


Grad’s 13-moment equation can be viewed as the underlying model of the R13 system. When
Grad (1949, 1958) considered moment equations, he solved the closure problem by assuming the
general expression
 N

f (Grad) (x, t, c) = f M (c; ρ, v, θ ) 1 + λi1 i2 ···in (x, t)H i1 i2 ···in (c ) (41)
n=0

as a model for the distribution function based on the Maxwellian distribution fM . Here, H i1 i2 ···in
are Hermite polynomials, and λi1 i2 ···in are parameters of the distribution function that are linked to
higher-order moments in a linear dependency. This necessarily yields moment equations that show
only linear dependence on nonequilibrium moments, which might be insufficient for nonlinear
processes. Another criticism is the possible negativity of the distribution function in Equation 41,
which renders the ansatz somewhat unphysical. Interestingly, in the case of linearized equations,
this does not seem to matter, but in the nonlinear case, things turn out to be more subtle.
In general, in the theory of moment equations, a closure is obtained from a model for the
distribution function, which has the form

f (x, t, c) = f (model) (c; {λi1 i2 ···in (x, t)}n=1,2,...N ) (42)

such that f depends on space and time through parameters λi1 i2 ···in . The number and form of
the parameters represent the complexity of the model. The challenge is to provide a physical
model and still manage to map the parameters to a number of moments in a one-to-one relation.
Obviously, the Grad distribution in Equation 41 is not the only possibility for a model of a
distribution function.

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 447

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

5.2. Hyperbolicity
With second-order derivatives and relaxation terms neglected, a moment system in one space
dimension can be written in the form

∂t W + A(W)∂x W = 0 (43)

with the variables combined into the vector W. This is called the collisionless case and models
the ballistic motion of particles. The eigenvalues of the matrix A(W) determine possible speeds
of propagation, and the system is called hyperbolic if all speeds are real valued and A(W) is diago-
nalizable. Hyperbolic systems represent coupled nonlinear transport and are seen as appropriate
equations to model the high-dimensional linear transport of the Boltzmann equation.
It is known that Grad’s 13-moment equations are not hyperbolic for all nonequilibrium states
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

(see Müller & Ruggeri 1998, Torrilhon 2000). Indeed, only in some vicinity of equilibrium can
real eigenvalues be assured. The region of hyperbolicity can be displayed in the projected phase
space given by the dimensionless normal stress σ/ p ∈ [−1, 2] and the dimensionless heat flux
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

q x /(ρθ 3/2 )R. Figure 8 shows this plane with the region of hyperbolicity for Grad’s 13-moment
equations. Outside the shaded region, complex-valued eigenvalues occur, and the system loses
physicality. This loss is a nonlinear phenomenon. The linearized equations consider deviations
close to equilibrium and remain hyperbolic for arbitrary values of these deviations. Furthermore,
linear waves are entirely stable in the moment equations both with and without dissipation. A
one-dimensional process can be plotted as a trajectory in the projected phase space. For example,
Figure 8 also shows a collisionless shock tube simulation. Because of strong nonequilibrium, the

Hyperbolicity region of Grad’s 13-moment equations


2.0

1.0

σ/p

Shock tube
0

–1.0
–2.0 0 2.0
qx /(pθ 1/2)

Figure 8
Hyperbolicity region of Grad’s 13-moment equations in one space dimension plotted in the projected
nonequilibrium phase space. The black dot marks the equilibrium state σ = 0 and qx = 0. The contours show
the maximal signal speed. Additionally, the solution trajectory of a collisionless shock tube simulation is
shown. Outside the hyperbolicity region, unphysical oscillations occur in the solution (see also Torrilhon
2010a).

448 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

trajectory leaves the domain of hyperbolicity, and complex eigenvalues lead to strong oscillations,
which result in a breakdown of the computation.
This behavior severely limits the application range of Grad’s equations when applied to strongly
nonlinear processes. It cannot be improved by considering more moments (Au et al. 2001, Brini
2001). Instead, a new closure procedure might be necessary.

5.3. Nonlinear Moment Closures


Statistics provides a variety of distribution models that could be used as a model in Equation 42.
A particular interesting case is the Pearson-IV distribution, which allows for anisotropic variances
and skewness in multiple dimensions; hence, it can model stress and heat flux. Torrilhon (2010a)
demonstrated how a Pearson-IV distribution can be used to formulate moment equations for
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

rarefied gases. Whereas the equations showed global hyperbolicity for processes in one space
dimension, complex eigenvalues still occur in higher-dimensional situations.
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

The maximum entropy moment closure is an approach that has gained popularity in recent
years. Maximizing entropy in nonequilibrium has a long history (see Kogan 1969, Dreyer 1987),
and Levermore (1996) and Müller & Ruggeri (1998) emphasized the promising mathematical and
physical properties of such a closure. The result of this approach is a model of the distribution
function in the form
N 
f (ME) (x, t, c) = exp λi1 i2 ···in (x, t)c i1 c i2 · · · c in , (44)
n=0

where λi1 i2 ···in (x, t) are again the parameters of the model to be mapped to moments. This form
of the distribution function is very flexible as it can be turned into a Maxwellian and into a sum
of Dirac deltas. Levermore (1996) and Müller & Ruggeri (1998) demonstrated that the resulting
moment equations are always globally hyperbolic and possess an entropy law. However, except
for the trivial case n = 2 (see Levermore & Morokoff 1998), the mapping between the parameters
λ and the moments turned out to be computationally quite challenging and even ill-posed for
some moment values (see Junk 1998, 2002; Dreyer et al. 2001). Still, McDonald & Groth (2009)
and McDonald (2010) showed that the maximum entropy closure can be used in a rarefied gas
simulation. The mapping from the parameters to the moment is so computationally demanding
that precomputed values and interpolation techniques are used to arrive at an implementation
useful for computational fluid dynamics (see McDonald & Torrilhon 2013).
Öttinger (2005) used a generalized Hamiltonian structure to impose conditions on thermody-
namics models. Öttinger (2010b) argued that a 13-moment theory for gases should replace the
heat flux variable qi in Equation 5 by

1
s i (x, t) = m Ci C j Ck p −1
j k (x, t) f (x, t, C) d C (45)
R3 2

using the pressure tensor pij to achieve thermodynamic admissible equations that, for instance,
possess an entropy law. The resulting equations do not easily satisfy basic requirements from
kinetic theory (Öttinger 2010a, Struchtrup & Torrilhon 2010, Torrilhon 2012) and still need
to demonstrate their usefulness in simulations. Although thermodynamically admissible, these
equations turn out to have a restricted region of hyperbolicity similar to Grad’s equations (see
Torrilhon 2012). Note that the linearized R13 theory already admits an entropy law, as shown by
Struchtrup & Torrilhon (2007).

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 449

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Cai et al. (2014) presented a hyperbolic correction to Grad’s moment equations that uses a
special projection technique. In essence, this approach is considering a distribution function of
the form
N
f (x, t, c) = λα (x, t)ψα (c; λ(x, t)) (46)
α=0
with parameters λα and basis functions ψα , for example, as shifted and scaled Hermite functions.
It is then required that the differential of f can be represented with the same set of basis functions,
N
!
df(x, t, c) = T αβ (λ(x, t))ψα (c; λ(x, t))d λβ , (47)
α,β=0

with some invertible coefficient matrix T αβ . For the classical Grad distribution, this results in
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

a cutoff projection that yields globally hyperbolic equations. A similar approach is presented by
Köllermeier et al. (2014) based on quadrature projection.
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

A method that is somewhat related to the maximum entropy approach is the quadrature method
of moments originally designed for aerosol dynamics (McGraw 1997) and later generalized by Fox
(2008). In this approach, the distribution is represented by a sum of Dirac deltas
N
f (x, t, c) = ωi (x, t)δ(c − c(i) (x, t)), (48)
i=0

where the variable weights ωi and positions c(i ) are related to given moments in a one-to-one
nonlinear mapping. This is equivalent to finding a quadrature rule for an integral with a weight
function that exactly reproduces the given moments. Finding a mapping in multidimensions in-
volves the solution of a highly nonlinear system of equations, which can be done efficiently only
using additional constraints (Yuan & Fox 2011). It turns out that the corresponding moment equa-
tions are only nonstrictly hyperbolic (i.e., signal speeds coincide). Moreover, the superposition
of Dirac deltas cannot represent a Maxwellian properly. Using Gaussians instead of Diracs over-
comes this problem and also leads to strict hyperbolicity (see Yuan & Fox 2011). The resulting
equations need to be tested on nonlinear rarefied gas processes.

5.4. Shock Wave Profiles


Because of the absence of boundary effects, and because of its essential one-dimensionality, the
normal shock wave is one of the simplest flows with large deviations from thermodynamic equi-
librium. It is also one of the key elements in hypersonic flows of compressible gas dynamics. At
the same time, the strong nonlinearity of the flow makes the prediction of a shock wave profile a
challenging benchmark problem for continuum models. In a shock wave, the Knudsen number is
essentially unity, and the strength of the shock is controlled by the Mach number M 0 > 1 based
on the shock speed and the upstream sound speed. Typical shock characteristics for comparison
are the shock thickness and shock asymmetry (Pham-Van-Diep et al. 1991, Holian et al. 1993), as
well as the precise behavior of the thermodynamic fields.
Fundamental measurements of complete shock structures have been accurately conducted and
furnish data for theoretical results for comparison (e.g., Alsmeyer 1976). The experimental data
have been satisfactorily reproduced by choosing a statistical particle approach such as in DSMC
calculations (see Bird 1998). However, so far no rational continuum description of rarefied flow
has succeeded in predicting the experimental evidence accurately over a wide range of the shock
wave Mach numbers. Gas dynamics with the laws of NSF succeed in predicting at least the shock

450 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

R13; M0 = 2.0 R13; M0 = 4.0


1.0
a b
DSMC
Heat flux
Density

0.5

0
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

–10 –5 0 5 10 –10 –5 0 5 10
x/λ 0 x/λ 0
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Figure 9
Shock wave profiles for the normalized density and heat flux from regularized 13-moment equations compared with direct simulation
Monte Carlo (DSMC) results (small gray squares) at two different Mach numbers: M 0 = 2.0 and M 0 = 4.0. The simulations are based
on Maxwell molecules, and the space coordinate is scaled by the mean free path. Figure adapted from Torrilhon & Struchtrup
(2004).

thickness qualitatively (see Gilberg & Paolucci 1953), but the quantitative agreement is restricted
to Mach numbers up to M 0 ≈ 1.3. Even worse, NSF theory cannot even reproduce the shock
asymmetry qualitatively (Pham-Van-Diep et al. 1991, Au 2001).
Burnett-type equations have been used for shock waves (Zhong et al. 1993, Fiscko & Chapman
1989, Uribe et al. 1998), but they either required ad hoc modifications of the equations or resulted
in oscillatory results. Grad (1952) applied his 13-moment case to shock waves. Unfortunately, the
results match the shock thickness only up to M 0 ≈ 1.1 accurately, and beyond M 0 = 1.65, they
suffer from an unphysical subshock (see Grad 1952). Subshocks are artifacts from the hyperbolic
nature of the 13-moment equations (Müller & Ruggeri 1998, Torrilhon 2000). It turned out that
any hyperbolic moment theory will yield continuous shock structures only up to the Mach number
corresponding to the highest characteristic velocity (Ruggeri 1993). As the maximal characteristic
speeds grow with the number of moments, large moment systems are needed to describe shock
wave profiles (Weiss 1995, Au 2001).
Regularized moment equations add dissipative terms to the equations that eliminate the sub-
shocks and produce smooth shock profiles (see Torrilhon & Struchtrup 2004, Ivanov et al. 2012).
Figure 9 displays the shock structures for M 0 = 2 and M 0 = 4 calculated with the R13 system
together with a DSMC result, both for Maxwell molecules. The figure also shows the profiles of
the density and heat flux. Because the heat flux is a higher moment, it is typically more difficult
to match than the density or temperature. In the plot, the density is normalized to give values
between zero and unity. Similarly, the heat flux is normalized such that the DSMC result gives
a maximal heat flux of +0.9. For M 0 = 2, the density profile and the shape of the heat flux are
captured very well by the R13 result. In the plots for M 0 = 4, the R13 result starts to deviate from
the DSMC solution in the upstream part. For higher Mach numbers, the results of the regularized
13-moment equations deviate even more.
If high accuracy for large–Mach number shocks is required, either large moment systems
are needed or nonlinear closures must be used. The maximum entropy approach of McDonald

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 451

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Density; M0 = 4.0 Density; M0 = 8.0


0.5 0.5
a b
NSF
Kinetic
Moment
ρ– ρ0 equations ρ– ρ0
0 0
ρ1– ρ0 ρ1– ρ0

–0.5 –0.5
– 40 –30 –20 –10 0 10 20 30 40 –40 –30 –20 –10 0 10 20 30 40
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

x/λ 0 x/λ 0

Heat flux; M0 = 4.0 Heat flux; M0 = 8.0


Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

0.5 0.5
c d

qx qx
0 0
pθ 1/2 pθ 1/2

–0.5 –0.5
– 40 –30 –20 –10 0 10 20 30 40 –80 –60 –40 –20 0 20 40
x/λ 0 x/λ 0

Figure 10
Shock wave profiles computed with moment equations based on the maximum entropy closure from McDonald (2010) and McDonald
& Torrilhon (2013). The plots show the density and heat flux (red solid lines) for Mach numbers M 0 = 4.0 and M 0 = 8.0. A kinetic
result (dashed dark yellow lines) is used as a reference, and the Navier-Stokes-Fourier (NSF) solution is shown for comparison (dotted blue
lines). Figure adapted from McDonald & Torrilhon (2013), with permission from Elsevier.

(2010) and McDonald & Groth (2009) is able to reproduce shock profiles very accurately, at
least in a model gas in which the velocity is one dimensional, c R, only. Figure 10 shows the
maximum entropy moment result at M 0 = 4 and M 0 = 8 for the density and heat flux and the
corresponding kinetic reference solution. Note that the moment theory uses an extremely low
number of variables, only five. Additionally, the figure shows the dramatic failure of the classical
laws of NSF. It also possible to compute shock wave profiles in a three-dimensional gas based
on maximum entropy (see McDonald & Torrilhon 2013) with promising results, but the closure
interpolation still needs more improvement. Interestingly, the maximum entropy shocks do not
show any subshock even though they do not use dissipative gradient fluxes. It turns out that
the ill-posedness of the closure leads to singular fluxes that generate infinite signal speeds close
to equilibrium (Schärer & Torrilhon 2015), which makes the subshock carry only a very small
amplitude.

452 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

FUTURE ISSUES
1. There is a need for a comprehensive simulation tool for moment equations. This
tool should be based on adequate numerical methods and be able to compute three-
dimensional unsteady processes in complex geometries. It may also need to solve larger
moment systems to increase the accuracy of the simulation. Eventually, this should open
the possibility of using a different number of moments in different regimes of an adaptive
computation.
2. The development of moment equations should include more realistic physical properties
of a gas. Here, polyatomic gases and mixtures are just the first steps. Further aspects could
be chemical reactions, real gas effects, multiphase flow, and turbulence. Especially with
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

regard to plasma flows, interactions with electromagnetic fields present a real challenge.
3. The usability of nonlinear closures must be improved for moment equations to be applied
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

to hypersonic flow computations, for example. All approaches suggested above come
with severe drawbacks, either in their practical and numerical use or in terms of physical
accuracy. Obviously, this challenge must be coupled to the first two above.
4. Finally, moment equations are also used in very different fields of physics, such as radiative
transfer, neutron transport, and granular flow. Even in uncertainty quantification of
stochastic influences, some sort of moment equations needs to be solved. It is to be
expected that more intriguing contributions are yet to come.

DISCLOSURE STATEMENT
The author is not aware of any biases that might be perceived as affecting the objectivity of this
review.

ACKNOWLEDGMENTS
The author is grateful to Dr. Henning Struchtrup from the University of Victoria for numerous
discussions and fruitful collaborations over the years, without which this article would not exist.

LITERATURE CITED
Agarwal RK, Yun KY, Balakrishnan R. 2001. Beyond Navier-Stokes: Burnett equations for flows in the
continuum-transition regime. Phys. Fluids 13:3061–85. Erratum. 2002. Phys. Fluids 14:1818
Alaoui M, Santos A. 1992. Poiseuille flow driven by an external force. Phys. Fluids A 4:1273–82
Alsmeyer H. 1976. Density profiles in argon and nitrogen shock waves measured by the absorption of an
electron beam. J. Fluid Mech. 74:497–513
Aoki K, Sone Y. 1987. Temperature field induced around a sphere in a uniform flow of a rarefied gas. Phys.
Fluids 30:2286–88
Aoki K, Takata S, Nakanishi T. 2002. Poiseuille type flow of a rarefied gas between two parallel plates driven
by a uniform external force. Phys. Rev. E 65:026315
Au JD. 2001. Lösung nichtlinearer Probleme in der Erweiterten Thermodynamik. PhD Thesis, Tech. Univ. Berlin
Au JD, Torrilhon M, Weiss W. 2001. The shock tube study in extended thermodynamics. Phys. Fluids 13:2423–
32
Bailey CL, Barber RW, Emerson DR, Lockerby DA, Reese JM. 2005. A critical review of the drag force on a
sphere in the transition flow regime. AIP Conf. Proc. 762:743–48

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 453

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Bao F, Lin J. 2008. Burnett simulation of gas flow and heat transfer in micro Poiseuille flow. Int. J. Heat Mass
Transf. 51:7593–600
Baranyai A, Evans DJ, Daivis PJ. 1992. Isothermal shear-induced heat flow. Phys. Rev A 46:7593–600
Barber RW, Emerson DR. 2003. Numerical simulation of low Reynolds number slip flow past a confined
microsphere. AIP Conf. Proc. 663:808–15
Barbera E, Brini F. 2011. Heat transfer in gas mixtures: advantages of an extended thermodynamics approach.
Phys. Lett. A 375:827–31
Barbera E, Müller I, Reitebuch D, Zhao NR. 2004. Determination of boundary conditions in extended
thermodynamics via fluctuation theory. Contin. Mech. Thermodyn. 16:411–25
Bird GA. 1998. Molecular Gas Dynamics and the Direct Simulation of Gas Flows. New York: Oxford Univ. Press.
2nd ed.
Bobylev AV. 1982. The Chapman-Enskog and Grad methods for solving the Boltzmann equation. Sov. Phys.
Dokl. 27:29–31
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Bobylev AV. 2006. Instabilities in the Chapman-Enskog expansion and hyperbolic Burnett equations. J. Stat.
Phys. 124:371–99
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Boltzmann L. 1872. Weitere Studien über das Wärmegleichgewicht unter Gasmolekülen. Wien. Ber. 66:275–
370
Brini F. 2001. Hyperbolicity region in extended thermodynamics with 14 moments. Contin. Mech. Thermodyn.
13:1–8
Brown S. 1999. Approximate Riemann solvers for moment models of dilute gases. PhD Thesis, Univ. Mich., Ann
Arbor
Burnett D. 1935. The distribution of velocities in a slightly non-uniform gas. Proc. Lond. Math. Soc. 39:385–430
Cai Z, Fan YW, Li R. 2014. Globally hyperbolic regularization of Grad’s moment system. Commun. Pure
Appl. Math. 64:464–518
Cai Z, Li R. 2010. Numerical regularized moment method of arbitrary order for Boltzmann-BGK equation.
SIAM J. Sci. Comput. 32:2875–907
Cai Z, Li R, Qiao Z. 2012. NRxx simulation of microflows with Shakhov model. SIAM J. Sci. Comput.
34:A339–69
Cercignani C. 1975. Theory and Applications of the Boltzmann Equation. Edinburgh: Scott. Acad.
Cercignani C, Lampis M. 1971. Kinetic models for gas-surface interactions. Transp. Theory Stat. Phys. 1:101–14
Chapman S, Cowling TG. 1970. The Mathematical Theory of Non-Uniform Gases. Cambridge, UK: Cambridge
Univ. Press
Deissler RG. 1964. An analysis of second order slip flow and temperature jump boundary conditions for
rarefied gases. Int. J. Heat Mass Transf. 7:681–94
Dreyer W. 1987. Maximization of the entropy in non-equilibrium. J. Phys. A 20:6505–17
Dreyer W, Junk M, Kunik M. 2001. On the approximation of the Fokker-Planck equation by moment systems.
Nonlinearity 14:881–906
Gad-el-Hak M, ed. 2002. The MEMS Handbook. Boca Raton, FL: CRC
Eu BC. 1980. A modified moment method and irreversible thermodynamics. J. Chem. Phys. 73:2958–69
Ferziger JH, Kaper HG. 1972. Mathematical Theory of Transport Processes in Gases. Amsterdam: North-Holland
Fiscko KA, Chapman DR. 1989. Comparison of Burnett, super-Burnett and Monte Carlo solutions for hyper-
sonic shock structure. Proc. 16th Int. Symp. Rarefied Gas Dynamics, ed. EP Muntz, D Weaver, D Campbell,
pp. 374–95. Reston, VA: AIAA
Fox RO. 2008. A quadrature-based third-order moment method for dilute gas-particle flows. J. Comput. Phys.
227:6313–50
Garcia-Colin LS, Velasco RM, Uribe FJ. 2008. Beyond the Navier-Stokes equations: Burnett hydrodynamics.
Phys. Rep. 465:149–89
Gilberg D, Paolucci D. 1953. The structure of shock waves in the continuum theory of fluids. J. Ration. Mech.
Anal. 2:617–42
Goldberg R. 1954. The slow flow of a rarefied gas past a spherical obstacle. PhD Thesis, New York Univ.
Grad H. 1949. On the kinetic theory of rarefied gases. Commun. Pure Appl. Math. 2:331–407
Grad H. 1952. The profile of a steady plane shock wave. Commun. Pure Appl. Math. 5:257–300

454 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Grad H. 1958. Principles of the kinetic theory of gases. In Thermodynamics of Gases, ed. S Flügge, pp. 205–94.
Berlin: Springer-Verlag
Gu XJ, Emerson D. 2007. A computational strategy for the regularized 13 moment equations with enhanced
wall-boundary conditions. J. Comput. Phys. 225:263–83
Gu XJ, Emerson D. 2009. A high-order moment approach for capturing non-equilibrium phenomena in the
transition regime. J. Fluid Mech. 636:177–216
Gu XJ, Emerson D, Tang GH. 2010. Analysis of the slip coefficient and defect velocity in the Knudsen layer
of a rarefied gas using the linearized moment equations. Phys. Rev. E 81:016313
Gupta VK, Torrilhon M. 2012. Automated Boltzmann collision integrals for moment equations. AIP Conf.
Proc. 1501:67–74
Gupta VK, Torrilhon M. 2015. Higher order moment equations for mixtures of rarefied gases. Proc. R. Soc. A
471:20140754
Hadjiconstantinou NG. 2006. Comment on Cercignani’s second-order slip coefficient. Phys. Fluids 15:2352–54
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Hickey KA, Loyalka SK. 1990. Plane Poiseuille flow: rigid sphere gas. J. Vac. Sci. Technol. A 8:957–60
Holian BL, Patterson CW, Mareschal M, Salomons E. 1993. Modeling shock waves in an ideal gas: going
beyond the Navier-Stokes level. Phys. Rev. E 47:R24–27
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Ivanov IE, Kryukov IA, Timokhin MY. 2013. Application of moment equations to the mathematical simulation
of gas microflows. Comput. Math. Math. Phys. 53:1543–50
Ivanov IE, Kryukov IA, Timokhin MY, Bondar YA, Kokhanchik AA, Ivanov MS. 2012. Stud of the shock
wave structure by regularized Grad’s set of equations. AIP Conf. Proc. 1501:215–22
Jin S, Slemrod M. 2001. Regularization of the Burnett equations via relaxation. J. Stat. Phys. 103:1009–33
Jou D, Casas-Vazquez J, Lebon G. 1996. Extended Irreversible Thermodynamics. Berlin: Springer-Verlag.
2nd ed.
Junk M. 1998. Domain of definition of Levermore’s five-moment system. J. Stat. Phys. 93:1143–67
Junk M. 2002. Maximum entropy moment systems and Galilean invariance. Contin. Mech. Thermodyn. 14:563–
76
Karlin IV, Gorban AN, Dukek G, Nonnenmacher TF. 1998. Dynamic correction to moment approximations.
Phys. Rev. E 57:1668–72
Karniadakis GE, Beskok A. 2001. Micro Flows: Fundamentals and Simulation. New York: Springer
Kauf P, Torrilhon M, Junk M. 2010. Scale-induced closure for approximations of kinetic equations. J. Stat.
Phys. 41:848–88
Knudsen M. 1909. Die Gesetze der Molekularströmung und der inneren Reibungsströmung der Gase durch
Röhren. Ann. Phys. 333:75–130
Kogan MN. 1969. Rarefied Gas Dynamics. New York: Plenum
Köllermeier J, Schärer RP, Torrilhon M. 2014. A framework for hyperbolic approximation of kinetic equations
using quadrature-based projection methods. Kinet. Rel. Models 7:531–49
Levermore CD. 1996. Moment closure hierarchies for kinetic theories. J. Stat. Phys. 83:1021–65
Levermore CD, Morokoff WJ. 1998. The Gaussian moment closure for gas dynamics. SIAM J. Appl. Math.
59:72–96
Lockerby DA, Reese JM. 2003. High-resolution Burnett simulations of micro Couette flow and heat transfer.
J. Comput. Phys. 188:333–47
Lockerby DA, Reese JM. 2008. On the modeling of isothermal gas flows at the microscale. J. Fluid Mech.
604:235–61
Lockerby DA, Reese JM, Emerson DR, Barber RW. 2004. Velocity boundary condition at solid walls in
rarefied gas calculations. Phys. Rev. E 70:017303
Magin TE, Martins G, Torrilhon M. 2010. Regularized Grad equations for multicomponent plasmas. AIP
Conf. Proc. 1333:99–104
Mansour MM, Babras F, Garcia AL. 1997. On the validity of hydrodynamics in plane Poiseuille flows. Physica
A 240:255–67
Marquez W Jr, Kremer GM. 2001. Couette flow from a thirteen field theory with jump and slip boundary
conditions. Contin. Mech. Thermodyn. 13:207–17
Maxwell JC. 1879. On stresses in rarefied gases arising from inequalities of temperature. Philos. Trans. R. Soc.
170:231–56

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 455

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

McDonald JG. 2010. Extended fluid-dynamic modelling for numerical solution of micro-scale flows. PhD Thesis,
Univ. Toronto
McDonald JG, Groth CPT. 2008. Extended fluid dynamic model for micron-scale flows based on Gaussian moment
closure. Presented at AIAA Aerosp. Sci. Meet., 46th, Reno, NV, AIAA Pap. 2008-691
McDonald JG, Groth CPT. 2009. Towards realizable hyperbolic moment closures for viscous heat-conducting
gas flows based on a maximum-entropy distribution. Contin. Mech. Thermodyn. 21:467–93
McDonald JG, Torrilhon M. 2013. Affordable robust moment closures for CFD based on the maximum-
entropy hierarchy. J. Comput. Phys. 251:500–23
McGraw R. 1997. Description of aerosol dynamics by the quadrature method of moments. Aerosol Sci. Technol.
27:255–65
Müller I, Reitebuch D, Weiss W. 2003. Extended thermodynamics: consistent in order of magnitude. Contin.
Mech. Thermodyn. 15:411–25
Müller I, Ruggeri T. 1998. Rational Extended Thermodynamics. New York: Springer. 2nd ed.
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Myong RS. 2001. A computational method for Eu’s generalized hydrodynamic equations of rarefied and
microscale gasdynamics. J. Comput. Phys. 168:47–72
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Myong RS. 2011. A full analytical solution for the force-driven compressible Poiseuille gas flow based on a
nonlinear coupled constitutive relation. Phys. Fluids 23:012002
Ohwada T, Sone Y, Aoki K. 1989. Numerical analysis of the Poiseuille and thermal transpiration flows
between two parallel plates on the basis of the Boltzmann equation for hard-sphere molecules. Phys.
Fluids A 1:2042–49
Öttinger HC. 2005. Beyond Equilibrium Thermodynamics. New York: Wiley
Öttinger HC. 2010a. Öttinger replies. Phys. Rev. Lett. 105:128902
Öttinger HC. 2010b. Thermodynamically admissible 13 moment equations from the Boltzmann equation.
Phys. Rev. Lett. 104:120601
Pham-Van-Diep GC, Erwin DA, Muntz EP. 1991. Testing continuum descriptions of low-Mach-number
shock structures. J. Fluid Mech. 232:403–13
Rahimi B, Struchtrup H. 2014. Capturing non-equilibrium phenomena in rarefied polyatomic gases: a high-
order macroscopic model. Phys. Fluids 26:052001
Rana AS, Mohammadzadeh A, Struchtrup H. 2015. A numerical study of the heat transfer through a rarefied
gas confined in a microcavity. Contin. Mech. Thermodyn. 27:433–46
Rana AS, Torrilhon M, Struchtrup H. 2013. A robust numerical method for the R13 equations of rarefied gas
dynamics: application to lid driven cavity. J. Comput. Phys. 236:169–86
Reese JM, Zheng Y, Lockerby DA. 2007. Computing the near-wall region in gas micro- and nanofluidics:
critical Knudsen layer phenomena. J. Comput. Theor. Nanosci. 4:807–13
Reinecke S, Kremer GM. 1990. Methods of moments of Grad. Phys. Rev. A 42:815–20
Reitebuch D, Weiss W. 1999. Application of high moment theory to the plane Couette flow. Contin. Mech.
Thermodyn. 11:217–25
Risso D, Cordero P. 1998. Generalized hydrodynamics for a Poiseuille flow: theory and simulations. Phys.
Rev. E 58:546–53
Rosenau P. 1989. Extending hydrodynamics via the regularization of the Chapman-Enskog solution. Phys.
Rev. A 40:7193–96
Ruggeri T. 1993. Breakdown of shock wave structure solutions. Phys. Rev. E 47:4135–40
Schärer RP, Torrilhon M. 2015. On singular closures for the 5-moment system in kinetic gas theory. Commun.
Comput. Phys. 17:371–400
Seeger S, Hoffmann H. 2000. The cumulant method in computational kinetic theory. Contin. Mech. Therm.
12:403–21
Sharipov F, Seleznev V. 1998. Data on internal rarefied gas flows. J. Phys. Chem. Ref. Data 27:657–706
Shavaliyev MS. 1993. Super-Burnett corrections to the stress tensor and the heat flux in a gas of Maxwellian
molecules. J. Appl. Math. Mech. 57:573–76
Söderholm LHS. 2007. Hybrid Burnett equations: a new method of stabilizing. Trans. Theory Stat. Phys.
36:495–512
Sone Y. 2002. Kinetic Theory and Fluid Dynamics. Basel: Birkhauser

456 Torrilhon

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Struchtrup H. 2002. Heat transfer in the transition regime: solution of boundary value problems for Grad’s
moment equations via kinetic schemes. Phys. Rev. E 65:041204
Struchtrup H. 2004. Stable transport equations for rarefied gases at high orders in the Knudsen number. Phys.
Fluids 16:3921–34
Struchtrup H. 2005a. Derivation of 13 moment equations for rarefied gas flow to second order accuracy for
arbitrary interaction potentials. Multiscale Model. Simul. 3:221–43
Struchtrup H. 2005b. Failures of the Burnett and super-Burnett equations in steady state processes. Contin.
Mech. Thermodyn. 17:43–50
Struchtrup H. 2005c. Macroscopic Transport Equations for Rarefied Gas Flows. New York: Springer
Struchtrup H. 2008. Linear kinetic heat transfer: moment equations, boundary conditions, and Knudsen
layers. Physica A 387:1750–66
Struchtrup H, Taheri P. 2011. Macroscopic transport models for rarefied gas flows: a brief review. IMA J.
Appl. Math. 76:672–97
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

Struchtrup H, Thatcher T. 2007. Bulk equations and Knudsen layers for the regularized 13 moment equations.
Contin. Mech. Thermodyn. 19:177–89
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Struchtrup H, Torrilhon M. 2003. Regularization of Grad’s 13 moment equations: derivation and linear
analysis. Phys. Fluids 15:2668–80
Struchtrup H, Torrilhon M. 2007. H-theorem, regularization, and boundary conditions for linearized 13
moment equations. Phys. Rev. Lett. 99:014502
Struchtrup H, Torrilhon M. 2008. High order effects in rarefied channel flows. Phys. Rev. E 78:046301.
Erratum. 2008. Phys. Rev. E 78:069903E
Struchtrup H, Torrilhon M. 2010. Comment on “Thermodynamically admissible 13 moment equations from
the Boltzmann equation.” Phys. Rev. Lett. 105:128901
Struchtrup H, Torrilhon M. 2013. Regularized 13 moment equations for hard sphere molecules: linear bulk
equations. Phys. Fluids 25:052001
Taheri P, Rana AS, Torrilhon M, Struchtrup H. 2009a. Macroscopic description of steady and unsteady
rarefaction effects in boundary value problems of gas dynamics. Contin. Mech. Thermodyn. 21:423–43
Taheri P, Struchtrup H. 2010a. An extended macroscopic transport model for rarefied gas flows in long
capillaries with circular cross section. Phys. Fluids 22:112004
Taheri P, Struchtrup H. 2010b. Rarefaction effects in thermally-driven microflows. Physica A 389:3069–80
Taheri P, Torrilhon M, Struchtrup H. 2009b. Couette and Poiseuille microflows: analytical solutions for
regularized 13-moment equations. Phys. Fluids 21:017102
Takata S, Sone Y, Aoki K. 1993. Numerical analysis of a uniform flow of a rarefied gas past a sphere on the
basis of the Boltzmann equation for hard-sphere molecules. Phys. Fluids A 5:716–37
Tij M, Sabbane M, Santos A. 1998. Nonlinear Poiseuille flow in a gas. Phys. Fluids 10:1021–27
Tij M, Santos A. 1994. Perturbation analysis of a stationary nonequilibrium flow generated by external force.
J. Stat. Phys. 76:1399–414
Todd BD, Evans DJ. 1995. The heat-flux vector for highly inhomogeneous nonequilibrium fluids in very
narrow pores. J. Chem. Phys. 103:9804–9
Todd BD, Evans DJ. 1997. Temperature profiles for Poiseuille flow. Phys. Rev. E 55:2800–7
Torrilhon M. 2000. Characteristic waves and dissipation in the 13-moment-case. Contin. Mech. Thermodyn.
12:289–301
Torrilhon M. 2006. Two-dimensional bulk microflow simulations based on regularized 13-moment equations.
SIAM Multiscale Model. Simul. 5:695–728
Torrilhon M. 2009. Editorial: special issue on moment methods in kinetic gas theory. Contin. Mech. Thermodyn.
21:341–43
Torrilhon M. 2010a. Hyperbolic moment equations in kinetic gas theory based on multi-variate Pearson-IV-
distributions. Commun. Comput. Phys. 7:639–73
Torrilhon M. 2010b. Slow rarefied flow past a sphere: analytical solutions based on moment equations. Phys.
Fluids 22:072001
Torrilhon M. 2012. H-theorem for nonlinear regularized 13-moment equations in kinetic gas theory. Kinetic
Relat. Models 5:185–201

www.annualreviews.org • Nonequilibrium Flow Based on Moment Equations 457

Changes may still occur before final publication online and in print
FL48CH16-Torrilhon ARI 17 September 2015 15:43

Torrilhon M, Struchtrup H. 2004. Regularized 13-moment equations: shock structure calculations and com-
parison to Burnett models. J. Fluid Mech. 513:171–98
Torrilhon M, Struchtrup H. 2008a. Boundary conditions for regularized 13-moment-equations for micro-
channel-flows. J. Comput. Phys. 227:1982–2011
Torrilhon M, Struchtrup H. 2008b. Modeling micro-mass and -heat transfer for gases based on extended
continuum models. ASME J. Heat Transfer 131:033103
Uribe FJ, Garcia AL. 1999. Burnett description for plane Poiseuille flow. Phys. Rev. E 60:4063–78
Uribe FJ, Velasco RM, Garcia-Colin LS. 1998. Burnett description of strong shock waves. Phys. Rev. Lett.
81:2044–47
Uribe FJ, Velasco RM, Garcia-Colin LS. 2000. Bobylev’s instability. Phys. Rev. E 62:5835–38
Vincenti WG, Kruger CH. 1965. Introduction to Physical Gas Dynamics. New York: Wiley
Weiss W. 1995. Continuous shock structures in extended thermodynamics. Phys. Rev. E 52:R5760–63
Westerkamp A, Torrilhon M. 2012. Slow rarefied gas flow past a cylinder: analytical solution in comparison
Access provided by Ecole Polytechnique Federal Lausanne on 10/05/15. For personal use only.

to the sphere. AIP Conf. Proc. 1501:207–14


Xu K. 2003. Super-Burnett solutions for Poiseuille flow. Phys. Fluids 15:2077–80
Annu. Rev. Fluid Mech. 2016.48. Downloaded from www.annualreviews.org

Xu K, Li ZH. 2004. Microchannel flow in the slip regime: gas-kinetic BGK-Burnett solutions. J. Fluid Mech.
513:87–110
Young JB. 2011. Calculation of Knudsen layers and jump conditions using the linearised G13 and R13 moment
methods. Int. J. Heat Mass Transfer 54:2902–12
Yuan C, Fox RO. 2011. Conditional quadrature method of moments for kinetic equations. J. Comput. Phys.
230:8216–46
Zhdanov VM. 2002. Transport Processes in Multicomponent Plasma. London: Taylor & Francis
Zheng Y, Garcia AL, Alder JB. 2002. Comparison of kinetic theory and hydrodynamics for Poiseuille flow.
J. Stat. Phys. 109:495–505
Zhong X, MacCormack RW, Chapman DR. 1993. Stabilization of the Burnett equations and applications to
hypersonic flows. AIAA J. 31:1036–43

458 Torrilhon

Changes may still occur before final publication online and in print

You might also like