Journal of Hydraulic Research Vol. 45, No. 2 (2007), pp.

147–164
© 2007 International Association of Hydraulic Engineering and Research

A Boltzmann-based finite volume algorithm for surface water flows on cells

of arbitrary shapes
Un algorithme de volume fini -basé sur les équations de Boltzmann pour les
écoulements de surface dans des cellules des formes arbitraire
J.H. LIANG, MPhil Student, Department of Civil Engineering, The Hong Kong University of Science and Technology,
Hong Kong SAR

M.S. GHIDAOUI, Professor, Department of Civil Engineering, The Hong Kong University of Science and Technology, Hong Kong
SAR, Tel.: (852) 2358 7174; e-mail: [email protected] (author for correspondence)

J.Q. DENG, Formerly a PhD Student, Department of Civil Engineering, The Hong Kong University of Science and Technology,
Hong Kong SAR

W.G. GRAY, Professor, Department of Environmental Science and Engineering, The University of North Carolina at Chapel Hill,
Rosenau Hall, CB 7431, Chapel Hill, NC 27599-7431, USA

ABSTRACT
An explicit two-dimensional conservative finite volume model for shallow water equations is formulated and tested. The algorithm for the mass and
momentum fluxes at the control surface of the finite volume is obtained from the solution of the Bhatnagar–Gross–Krook (BGK) Boltzmann equation.
Unlike classical methods, BGK schemes do not require an ad-hoc splitting of advection and diffusion. The BGK scheme is second order in both time
and space. The formulation of the BGK algorithm is performed for a cell of arbitrary irregular shape, but the test cases are conducted using a structured
grid of quadrilateral cells. Two approximate Riemann solvers, the HLLC scheme and the two-stage Hancock-HLLC scheme, where HLL stands
for Harten, Lax and van Leer and C stands for contact discontinuity, are also considered. The second-order accuracy of HLL and Hancock-HLLC
schemes is obtained by MUSCL approach, where MUSCL is the acronym for Monotone Upstream-centered Schemes for Conservation Laws. The
data reconstruction for all three schemes is carried out by the Van Leer limiter. The test cases involve strong shocks and expansion waves. The accuracy
of the schemes are measured using an absolute error norm and a waviness error norm. The HLLC scheme is highly oscillatory for Courant number
larger than 0.5, while the BGK and the Hancock-HLLC schemes are applicable for Courant numbers as high as 1.0. For a fixed value of the central
processing unit (CPU) time, the absolute error of the Hancock-HLLC is slightly smaller than that of the BGK while the waviness error of the BGK is
quite close to that of Hancock-HLLC. This is because (i) the Hancock-HLLC is a two-step method while the BGK is a single-step method (i.e., the
Hancock-HLLC requires storage of intermediate variables, but the BGK does not), and (ii) the Hancock-HLLC schemes requires larger number of
grid points than the BGK scheme for the same level of accuracy. For example, to achieve an absolute error of 0.01, the BGK requires about 600 grid
points while the Hancock-HLLC requires about 800 grid points. Both the BGK and Hancock-HLLC schemes have similar convergence properties.
Unlike exact or approximate Riemann solvers, BGK fluxes accounts for both waves and diffusion. The ability of the BGK scheme to model diffusion
is illustrated using a viscous flow problem. Excellent agreement between the analytical and computed viscous flow solution is found. Although
the BGK and Hancock-HLLC schemes perform similarly for hyperbolic problems, BGK schemes have the added advantage of being able to solve
hyperbolic–parabolic problems without the need for an ad-hoc operator splitting. This is important given that the artificial splitting of advection and
diffusion is known to cause artificial widening in shear layers and introduces artificial transient in regions with sharp gradients. Such problems arise
when the splitting operation fails to faithfully represent the correct coupling between the physics of advection and the physics of waves.

RÉSUMÉ
Un modèle conservatif bidimensionnel explicite en volumes finis pour des équations en eau peu profonde est formulé et examiné. L’algorithme
pour le flux de la masse et des quantités de mouvement sur la surface de contrôle du volume fini est obtenu à partir de la solution de Bhatnagar-
Brut-Krook (BGK) de l’équation de Boltzmann. À la différence des méthodes classiques, les schémas de BGK n’exigent pas d’éclatement ad hoc
de l’advection et de la diffusion. Le schéma de BGK est du second ordre dans le temps et l’espace. La formulation de l’algorithme de BGK
est conçu pour une cellule de forme irrégulière arbitraire, mais les essais sont conduits en utilisant une grille structurée des cellules quadri-
latères. Deux solveurs approximatifs de Riemann, le schéma de HLLC et le schéma de Hancock-HLLC à deux étages, où HLL est mis pour
Harten, Lax et van Leer et C pour interface de discontinuité, sont également utilisés. La précision du second ordre de HLL et des schémas de
Hancock-HLLC est obtenue par l’approche MUSCL, où MUSCL est l’acronyme de “Monotone Upstream-centered Schemes for Conservation
Laws”. La formulation des données pour chacun des trois schémas est effectuée par le limiteur de van Leer. Les cas test impliquent des chocs
forts et ondes d’expansion. La précision des schémas est mesurée en utilisant une norme d’erreur absolue et une norme d’erreur oscillatoire.
Le schéma de HLLC est fortement oscillant pour un nombre de Courant supérieur à 0.5, tandis que les schémas BGK et de Hancock-HLLC

Revision received July 8, 2005/Open for discussion until October 31, 2007.

147
148 Liang et al.

sont applicables pour des nombres de Courant jusqu’à 1.0. Pour une valeur donnée du temps d’unité centrale (unité centrale de traitement), l’erreur
absolue de Hancock-HLLC est légèrement plus petite que celle du BGK tandis que l’erreur de caractère oscillatoire de BGK est tout à fait proche
de celle de Hancock-HLLC. C’est parce que (i) Hancock-HLLC est une méthode à deux étages tandis que le BGK est une méthode de pas simple
(i.e., Hancock-HLLC a besoin du stockage intermédiaire des variables, mais pas BGK), et (ii) les schémas de Hancock-HLLC exigent un plus grand
nombre de points de grille que le schéma de BGK pour la même précision. Par exemple, pour obtenir une erreur absolue de 0.01, BGK exige environ
600 points de grille tandis que Hancock-HLLC en a besoin d’environ 800. Les schémas BGK et Hancock-HLLC ont des propriétés de convergence
semblables. À la différence des solveurs exacts ou approximatifs de Riemann, BGK prend en compte les ondes et la diffusion. La capacité du schéma
de BGK à modéliser la diffusion est illustrée dans un problème d’écoulement visqueux. Un excellent accord entre les solutions analytique et calculée
d’écoulement visqueux est trouvée. Bien que les schémas BGK et Hancock-HLLC marchent tout aussi bien pour des problèmes hyperboliques, les
schémas de BGK ajoutent l’avantage de pouvoir résoudre des problèmes hyperbolique-paraboliques sans besoin de dédoublement ad-hoc de l’opérateur.
C’est important car le dédoublement de l’advection et de la diffusion est connu pour causer un épaississement artificiel des couches de cisaillement et
introduire des transitions artificielles dans les régions avec de forts gradients. De tels problèmes surgissent quand l’opération de dédoublement n’arrive
pas représenter fidèlement le couplage correct entre la physique de l’advection et la physique des ondes.

Keywords: Boltzmann equation, unsteady open channel flow, dam break, bore, entropy, numerical model.

1 Introduction without the need to split the physics of advection and the physics
of diffusion is tested in this paper.
Problems in physics and engineering are often of hyperbolic– The purpose of the present paper is (i) to generalize the BGK
parabolic type. Examples in hydraulics include unsteady flows algorithm to irregular grids so as to make the scheme better suited
in surface water and closed conduits (waterhammer). The hyper- for channels with complex geometry, (ii) to compare the resulting
bolic part is a propagation phenomena. The path of propagation scheme with approximate Riemann solvers and (iii) to illustrate
of flow disturbances is determined by the wave and flow speeds. that BGK schemes incorporate the effects of waves and diffusion
The parabolic part is a diffusion phenomena linked to turbulent (i.e., no operator splitting is needed to separate the wave part
and molecular diffusion. The classical approach to numerical from the diffusion part). The generalized algorithm is used to
modeling requires that the problem be split into a hyperbolic simulate a range of two-dimensional open channel flows with
part and a parabolic part. Schemes such as exact or approx- irregular geometry. Where possible, laboratory observations or
imate Riemann solvers are used for the hyperbolic part. The simulated solutions obtained by other schemes are compared to
parabolic part is then solved by other techniques (central dif- the BGK-based solutions.
ferencing schemes). For example, hydraulic software such as
Delft3D (Delft Hydraulics, 2003) and the Princeton Ocean
Model (POM) (Mellor, 2004) use the splitting technique so as 2 Boltzmann methods
to discretize the advective terms separately from the diffusive
terms. Numerical models for fluid mechanics and hydraulics are conven-
It is diffcult to faithfully reproduce the physical balance and tionally formulated by starting from flow equations obtained by
coupling between advection and diffusion when problems are applying the conservation laws to a control volume with a macro-
solved on the basis of operator splitting. Karlsen et al. (2001) scopic length scale lmac with the condition that lmac is larger or
report that operator splitting method induces artificial widening equal to the continuum scale, lcon . Examples of such conserva-
of shear layer in regions where advection is significant. In addi- tion laws for a continuum include the Navier–Stokes equations
tion, Xu et al. (2005) note that the artificial decoupling of the and the shallow water equations. More recently, the mesoscopic
hyperbolic and parabolic terms is problematic in regions where approach, where numerical algorithms for fluid mechanics and
there is strong interactions between advective and viscous trans- hydraulics are formulated by starting from the Boltzmann equa-
port. They report that the application of schemes which are based tion, has been applied to a wide range of problems in fluid
on operator splitting trigger artificial transient in regions of strong mechanics (e.g., Reitz, 1981; Frisch et al., 1986; Xu et al., 2001;
gradients such as boundary layers, strong shock and density inter- Su et al., 1998; Chen and Doolen, 1998; Kumar et al., 1999)
faces. Therefore, it is desirable to develop schemes that do not but to a lesser extent in hydraulics (e.g., Deng, 2000; Ghidaoui
require ad-hoc splitting of the physics of advection from the et al., 2001). Numerical models obtained from the mesocopic
physics of diffusion. approach are, generally, noted for their inherent ability to sat-
Operator splitting methods are not required when Bhatnagar– isfy the entropy condition (which precludes the emergence of
Gross–Krook (BGK) schemes are used to evaluate the fluxes at physically non-realizable solutions) and for the relative ease
the cell interfaces (Xu et al. 2005). BGK-based fluxes include with which they can be implemented for multi-dimensional flows
both advective and diffusive transport. The ability of BGK and applied to problems with complex geometry and boundary
schemes to solve hyperbolic–parabolic problems without the conditions.
ad-hoc splitting of advection and diffusion was successfully The essence of the mesoscopic approach and its connection
applied to compressible-viscous flows by Xu et al. (2005) and to to the macroscopic approach is as follows. At the macroscopic
3
scalar transport by advection and diffusion (Deng et al., 2001). length scale, the smallest fluid volume is of order lmac . This
The ability of BGK schemes to solve surface water problems volume contains a large number of fluid particles (atoms and
 Boltzmann-based finite volume algorithm 149

molecules) with different velocities. The bulk behavior of the particle; c is the particle velocity vector with components cx and
flow at the macroscopic scale is represented by quantities such cy in the x and y directions, respectively; Sext is the net external
as fluid velocity, density and pressure which are manifesta- force acting on the particles with components Sx and Sy ; τ the col-
3
tions of the motions of particles averaged over the volume lmac . lision (or relaxation) time which is a measure of the average time
The Navier–Stokes equations and the shallow water equations between collisions for a particle; ρ is the density of the fluid; h is
describe how these macroscopic quantities vary in space and the water depth; v is the depth-averaged fluid velocity vector with
time. The mesoscopic approach uses a probability density func- components u and v in the x and z directions; Γ(x, t) is the vis-
tion, f , to represent the distribution of particle velocities within cous stress tensor; and I is the identity tensor. When gravity and
the volume lmac3
. The spatial and temporal variations of this distri- friction are the only external forces, Sext /m = g(S0 − Sf ), where
bution function are governed by a Boltzmann-like equation. The S0 is the channel slope vector and Sf is the friction slope. Other
connection between the mesocopic and macroscopic equations external forces such as wind stresses can be added, if desired.
comes about by the recognition that bulk quantities such as flow When all macroscopic gradients of the fleld variables (e.g.,
density and velocity are obtained as moments of the distribution pressure, temperature and velocity) are zero, the flow is said
function associated with underlying motion of particles. Conser- to be in mechanical and thermal equilibrium and the velocity
vation laws, such as the Navier–Stokes equations and the shallow distribution of the particles is given by q. When all macro-
water equations, are obtainable as moments of an appropriately scopic gradients of the field variables are small but non-zero,
formulated Boltzmann-like equation. This connection between the particles are said to be in local equilibrium and their velocity
the macroscopic conservation laws and the mesoscopic theory of local distribution is given by q. When the gradients are large,
Boltzmann has been exploited by numerical modelers to formu- such as at jumps and boundary layers, the particles are in non-
late schemes on the basis of the Boltzmann equation (e.g., Reitz, equilibrium state and their distribution is given f and not q. The
1981; Frisch et al., 1986; Xu, 2001; Su et al., 1998; Chen and relation of the equilibrium and non-equilibrium distribution is
Doolen, 1998; Kumar et al., 1999; Deng, 2000; Ghidaoui et al., given by the BGK equation. Excellent discussions on the equi-
2001). librium and non-equilibrium distributions are given in Vincenti
The mesoscopic algorithm, based on the BGK Boltzmann and Kruger (1965).
equation, has been successfully formulated and applied to surface The direct connection between (1), (2), (3) and the classical
water flows in channels with simple geometry on a regular spa- shallow water equations has been established in Ghidaoui et al.
tial grid (Ghidaoui et al., 2001). The computed solutions indicate (2001). In particular, it was shown that the zero and first statistical
that the BGK-based algorithm is accurate, stable and efficient and moments of (1) along with relations (2), (3) and Sext /m = g(S0 −
has a robust ability to simulate unsteady flow problems, including Sf ), and the fact that f and fc have compact support yields
both smooth and discontinuous flows. The BGK scheme, being
 

∂ h  hv  0
based on particle motions but not waves, requires neither char- +∇ · gh2 =
acteristics decomposition nor identification of wave type at each ∂t hv h vv + 2 I − ρ1 Γ gh(S0 − Sf )
cell interface. (4)

These are the classical 2-D shallow water equations. The fact
3 Governing equations in differential form that the differential equation of Boltzmann can be employed to
obtain the differential equations of shallow water flow suggests an
The vertically-integrated, two-dimensional BGK equation for alternative framework to the standard solution methods of finite
shallow water is (Deng, 2000; Ghidaoui et al., 2001) elements, finite differences, and characteristics. The Boltzmann-
∂f Sext ∂f q−f based framework takes advantage of the fact that the Boltzmann
+ c · ∇f + · = (1) equation is single quasi-linear partial differential equation gov-
∂t m ∂c τ
erning a scalar quantity (i.e., f ). Essentially, the formulation
where
   
of a Boltzmann-based scheme entails two steps. First, a dif-
h ∞ ∞ 1 ference form of the Boltzmann equation is formulated. Second,
   c  f dcx dcz
  hv = the zero and first statistical moments of the difference form are
−∞ −∞
h vv + 2 I − ρ Γ
gh 1
cc calculated, yielding the discrete model for the shallow water
(2) equations. A two-dimensional implementation of the Boltzmann-


1 (c − v) · (c − v) based scheme for solution of the shallow water equations on an
q(x, c, t) = exp − irregular spatial grid is formulated in the following section.
πg gh
(3)
f is the irreversible (non-equilibrium) particle distribution func- 4 Governing equation in finite volume form
tion; q is the reversible (equilibrium) particle distribution func-
tion; t is time; ∇ is the two-dimensional del operator with respect Integration of (4) in space over a control area (cell) of arbitrary
to x and y, the lateral Cartesian coordinates; m is the mass of a shape centered at point (i, j) and in time from level k to level
150 Liang et al.

k + 1 gives interval [k, k + 1]. Macroscopic approaches employ upwind-

ing techniques, such as those based on the Riemann solution,

k+1
k
h h to approximate the flux term in (5). In the last decade, Riemann-
−
hv hv based schemes for shallow water flows have received much
i,j i,j
 attention as evidenced by the surge of related papers as well as
1 k+1
 hv  · n dt dl the recent book by Toro (2001). The Riemann solution resolves
=−
ij ∂ij k h vv + gh
2
I − ρ1 Γ the wave interaction between two constant states at each cell

 interface. Therefore, the use of the Riemann solution in deter-
0
+ t (5) mining the flux term in (5) helps to ensure that the upwinding
gh(S0 − Sf )
property of the scheme is physically based. However, Riemann-
based schemes have a number of drawbacks. For example, since
where ij = area of the control volume (cell) that is centered at
the Riemann solution only exists for the 1-D surface water flow
point (i, j); ∂ij = boundary of ij ; n = (dyi − dxj)/dl is the
without external forces and where the initial condition consists
local unit outward vector normal to ∂ij ; t = (dxi+dyj)/dl is the
of two constant states, its application in formulating numerical
unit tangent vector to the surface when going around
 the boundary schemes for general surface water flow problems requires both
in the counter-clockwise direction, and dl = dx2 + dy2 . Note
operator and directional splitting. Part of the Riemann solution
that (h, hv)k+1 k
i,j and (h, hv)i,j are the average cell values at time
process also involves determining the types and speeds of the
k + 1 and k, respectively.
waves originating at each cell interface.
A typical control volume is shown in Fig. 1. In general, the
Nij Mesocopic approaches use the connection between the
number of sides of ij is Nij such that ∂ij = ∪s=1 Ls , where
Boltzamm equation and the shallow water equations to approxi-
Ls = length of side number s and ∪ = union operator. Therefore,
mate the flux term in (5). The formulation of mesoscopic models
(5) becomes:
entails the following two steps. First, an approximate solution
Nij for (1) at a cell interface is sought. Second, this approximate
1  solution for the Boltzmann equation is inserted into (2), resulting
k+1
Wi,j − k
Wi,j =− Ls Fs t + Si,j t (6)
ij s=1 in an algorithm for water depth and velocity in two-dimensional
shallow water flows. Ghidaoui et al. (2001) formulated such a
where the subscript s refers to the flux and normal at side s, the mesocopic BGK scheme on a rectangular grid system. The next
force term Si,j is some average over ij and the time interval section is devoted to the generalization of the BGK scheme to
from k to k + 1, Fs is some average of the flux over side Ls and irregular grids such as the one shown in Fig. 1.
the time interval from k to k + 1, and


h
W= ; 5 Flux estimation by the BGK Scheme
hv

k+1
 hv  Using (2), the instantaneous mass and momentum fluxes from
Fs t = gh2 · ns dt; (7)
kh vv + 2
I − ρ1 Γ cell (i, j) at side s for t in the interval [t k , t k+1 ] is as follows:
s


0
Sij = 
gh(S0 − Sf ) hv
Ḟs =  2
 · ns
h vv + gh2 I − ρ1 Γ
A numerical scheme is required to estimate the first term in s
∞ ∞

the right-hand side of (6) which represents the net mass and 1
= cs · ns fs dcns dcts
momentum fiuxes at the surface of cell (i, j) during the time −∞ −∞ c s
∞ ∞

cns
= fs dcns dcts (8)
y −∞ −∞ cns cs
c
3
d
where fs = distribution function at side number s; ns = local
n2 unit outward vector normal to side number s; cns = cs · ns is
(i,j) 2
4 the particle speed normal to side s; and ts = local tangent to side
b number s. It is noted that the fluxes in (8) are valid for an arbitrary
1
a cell shape, regardless whether the overall grid is structured or
unstructured. However, in the examples section, only structured
grids with quadrilateral cells are considered.
x The BGK scheme entails finding an approximation for fs so
Figure 1 An arbitrary quadrilateral finite volume and its interfaces (n2 that an algorithm for the fluxes in (8) can be developed. To this
is unit outward vector normal to side 2). end, the solution of (1) at side number s for t in the interval
 Boltzmann-based finite volume algorithm 151

[t k , t k+1 ] is (Ghidaoui et al., 2001): [t k , t k+1 ] is obtained by inserting (12) into (8) as follows:
t−t n 
fs (cs , t) = f(cs , Xs , t) = f(cs , Xs , t n )e− τ  hv 
Ḟs = 2 · ns
1 t t−β h vv + gh2 I − ρ1 Γ
+ q(c, x(β), β)e− τ dβ (9) s
τ tk ∞ ∞

cns
where β is dummy variable of integration and Xs = xs − =
−∞ −∞ cns cs
cs (t − t k ) position of a particle at time t k which would arrive  
∂
to side s at time t; xs = local coordinate of any point belonging × α3 + α4 c · ∇ + α5 qs dcns dcts
to side s. ∂t
∞ ∞

The distribution functions f(cs , Xs , t k ) and q(c, xs (β), β) cns
+ (α1 + α2 c · ∇)qsi dcns dcts
which appear in the right-hand side of (9) are unknown and −∞ 0 cns cs
need to be approximated. Using a second-order Taylor expan- ∞ 0

cns
sion near side number s (i.e., with respect to xs ) and since + (α1 + α2 c · ∇)qso dcns dcts (13)
−∞ −∞ cns cs
Xs − xs = −cs (t − t k ), one can develop an expression for
f(cs , Xs , t k ) as follows: where cts = cs · ts = particle speed tangent to side s.
k The right-hand side of (13) involves moments of a Gaussian
f(cs , Xs , t )
 distribution. The recursive formulas given in Appendix provide
qsi − cns (ns · ∇qsi )(t − t k ) if cns > 0 moments of Gaussian distribution of any order. These moments
= (10)
qso − cns (ns · ∇qso )(t − t k ) if cns ≤ 0 are used to evaluate the integrals in (13) and the result is:
where the subscript so indicates that the function is evaluated at
Ḟs = α1 {Asi + Aso } + α2 {ns · ∇(Bsi + Bso )
a location outside cell (i, j) while the subscript si indicates that  
the function is evaluated at a position within the cell [i.e., as one ∂
+ ts · ∇(Dsi + Dso )} + α3 + α5 Es
approaches side s from the inside of cell (i, j)]. ∂t
Note that (10) accounts for both smooth and discontinuous + α4 {ns · ∇Gs + ts · ∇Is } (14)
variations at the cell interface. In particular, if the flow exhibits no √
jumps at side s of the cell boundary, qso = qsi = qs which results where, with Vn = vn / gh,
in f(cs , Xs , t k ) being continuous at the cell interface. On the other ∞ ∞

hand, if the flow exhibits a jump at side s the cell boundary, qso  = cns
Asi = qsi dcns dcts
qsi  = qs which results in f(cs , Xs , t k ) also being discontinuous at −∞ 0 cns cs
 
the boundary. Allowing f(cs , Xs , t k ) to vary discontinuously from √ V erfc(−V ) +
2
√1 e−Vn
hsi ghsi   n n
√  
π
cell to cell provides the BGK scheme with the ability to handle = −V 2
large flow gradients such as hydraulic jumps, bores, interfaces
2 Vn v + n 2 erfc(−Vn ) + v e√πn
gh
si
between different fluids, and boundary layers. (15)
A second-order Taylor expansion of q(c, x(β), β) around xs ∞ 0

cns
and t k gives Aso = qso dcns dcts
−∞ −∞ c ns cs
∂qs  
q(c, x(β), β) = qs − c · ∇qs (β − t k ) + (β − t k ) (11) √ 2
Vn erfc(Vn ) − √1π e−Vn
hso ghso  
∂t = √  −V 2

Insertion of (10) and (11) into the right side of (9) and evalua- 2 Vn v + n 2gh erfc(Vn ) − v e√ n π so
tion of the integral over time gives the following second order (16)
approximation in both space and time for fs (cs , t) 
∞ ∞ 2
cns
Bsi = qsi dcns dcts
fs (cs , t) −∞ 0
2
cns cs

 α1 (t)qsi + α2 (t)c · ∇qsi + α3 (t)qs   2 1 

 −Vn2

 ∂q gh2si  Vn + 2 erfc(−Vn ) + Vn√e π
 + α4 (t)c · ∇qs + α5 (t) s if cns > 0 =  2  √ 2

2 (Vn v+ ghn)e−Vn
= ∂t Vn v + 2 vn n erfc(−Vn ) +
v √

 α1 (t)qso + α2 (t)c · ∇qso + α3 (t)qs π si



 + α4 (t)c · ∇qs + α5 (t) ∂qs (17)
if cns ≤ 0 
∂t ∞ 0 2
cns
(12) Bso = 2
qso dcns dcts
−∞ −∞ cns cs
t−t k
where α1 (t) = e− τ ; α2 (t) = −(t − t k )α1 (t); α3 (t) = 1 − α1 ;    2 
e−Vn
gh2so Vn2 + 1
erfc(Vn ) − Vn√
α4 (t) = τ(−1 + α1 (t)) + (t − t k )α1 (t); and α5 (t) = (t − t k ) + =  2 π 
2  √ 2
(Vn v+ ghn)e−Vn
τ(−1 + α1 (t)). Vn2 v + v2 vn n erfc(Vn ) − √
π so
The BGK-based algorithm for the net instantaneous mass and
(18)
momentum fluxes from cell (i, j) at side s for t in the interval
152 Liang et al.

∞ ∞

 √
cns cts hsi ghsi where
Dsi = qsi dcns dcts =
−∞ 0 cns cts cs 2  k
    h
∂ 
vt Vn erfc(−Vn ) + e√πn
−V 2
hu 
  ∂x
 √   hv i,j
 vt gh V 2 + 1 erfc(−Vn ) + 3Vn√e−Vn 
2
× n 2 π  (19)      
     sgn x hni,j +sgn ∇x hni,j
 −V 2   
v2t + gh 2
Vn erfc(−Vn ) + e√πn  
2
 
 min |δx hki,j |, |x hni,j |, |∇x hki,j | 

√
si
 
∞ 0      
cns cts hso ghso  sgn x (hu)ki,j +sgn ∇x (hu)ki,j 
Dso = qso dcns dcts = 1  
−∞ −∞ cns cts cs 2 =  2 
    x 

  
−V 2
 min |δx (hu)ki,j |, |x (hu)ki,j |, |∇x (hu)ki,j | 
vt Vn erfc(Vn ) − e√πn      
 
 √    k k
sgn x (hv)i,j +sgn ∇x (hv)i,j 
 vt gh V 2 + 1 erfc(Vn ) − 3Vn√e−Vn   
2
×  (20)  2 
 
n 2 π
   
  −V 2  min |δx (hv)i,j |, |x (hv)i,j |, |∇x (hv)i,j |
k k k
v2t + gh
2
Vn erfc(Vn ) − e√πn
so (25)
∞ ∞
  k
cns h
Es = qs dcns dcts ∂ 
−∞ −∞ cns cs hu 

 ∂z
hv i,j
vn
= hs (21)      
vn v + ghn
2 s sgn z hki,j +sgn ∇z hki,j
 2 

    
∞ ∞ 2
cns  min |δz hki,j |, |z hki,j |, |∇z hki,j | 
Gs = qs dcns dcts  
2      
−∞ −∞ cns cs  k
sgn z (hu)i,j +sgn ∇z (hu)i,j k 
 1 


v2n + gh = 2
z 
   
= hs min |δz (hu)i,j |, |z (hu)i,j |, |∇z (hu)i,j | 
2
(22) k k k
v2n v + 3gh
v n  
2 n s  
 sgn(z (hv)ki,j )+sgn(∇z (hv)ki,j ) 
 

  2 
∞ ∞
cns cts  
Is = qs dcns dcts min |δz (hv)ki,j |, |z (hv)ki,j |, |∇z (hv)ki,j |
−∞ −∞ cns cts cs

 (26)
vn vt
= hs (23) x (·)ki,j = (·)ki+1,j − (·)ki,j (27)
vt vn v + (vt n + vn t) gh
2 s
∇x (·)ki,j = (·)ki,j − (·)ki−1,j (28)
Relations (14)–(23) provide the BGK-based algorithm for mass
1# $
and momentum flux at side s of cell (i, j). The quantitative eval- δx (·)ki,j = x (·)ki,j + ∇x (·)ki,j (29)
uation of theses fluxes requires the values of the water height, 2
normal velocity and tangential velocity at s, si and so to be z (·)ki,j = (·)ki+1,j − (·)ki,j (30)
known [i.e., (h, vn , vt )s , (h, vn , vt )si , and (h, vn , vt )so ] and their ∇z (·)ki,j = (·)ki,j − (·)ki−1,j (31)
respective gradients to be estimated.
Suppose that the solution at time level k has just been and
obtained. Therefore, the flow velocity and height at the cen- 1# $
δz (·)ki,j = z (·)ki,j + ∇z (·)ki,j . (32)
ter of each cell (i, j) at time level k are known. The desired 2
values of (h, vn , vt )s , (h, vn , vt )si , and (h, vn , vt )so and their The values of (h, u, v)si and their gradients at the boundary
gradients can be obtained by a spatial interpolation scheme between, say, cell (i, j) and cell (i+1, j) are obtained by applying
between the known nodal cell values. Ghidaoui et al. (2001) (24) to cell (i, j) and evaluating the resulting expression at x =
applied a second-order interpolation scheme, which uses nonlin- xsi . The projection (vn , vt )si and the components of the gradients
ear limiter, and found that this interpolation method accurately of h, vn , and vt along the normal and tangential directions at side
resolves steep surface water gradients and successfully sup- s are obtained using the rotation matrix which transforms from
presses spurious oscillations. Use of this non-linear second-order (x, z) to (n, t) coordinates according to:
interpolation scheme provides the spatial variation of flow veloc-



fn ns · i ns · k fx
ities and water depth within any cell (i, j) at time level k is as = (33)
follows: ft si n s · k ns · i fz si
 k  k  k where i and k are the unit vectors in the x and z directions,
h h (x − xi,j ) · ∇h
respectively. Calculation of functions (vn , vt ) and the normal and
 hu  =  hu  +  (x − xi,j ) · ∇(hu)  (24)
tangential derivatives at x = xso are obtained using the same
hv hv i,j (x − xi,j ) · ∇(hv) i,j
rotation matrix and Eq. (24).
 Boltzmann-based finite volume algorithm 153

Since (h, vn , vt ) can experience a jump from cell to cell (i.e., where
at x = xs ), the values of (h, vn , vt )s and their gradients can-
−t
not, in general, be obtained by evaluating (24) at x = xs . γ1 = τ − τ e τ ;
Instead, (h, vn , vt )s and their gradients can be obtained from a −t
γ2 = −τ 2 + (τ 2 + τt) e τ ;
weighted average (h, vn , vt )si and (h, vn , vt )so and their gradients.
−t
In particular, γ3 = −τ + τe τ ; (38)

k ∞ ∞
 −t −t
h 1 γ4 = −τt + 2τ 2 (1 − e τ ) − tτe τ ;
= qsi dcns dcts −t
hv s −∞ 0 cs γ5 = τt − τ (1 − e );
2 τ (39)
∞ 0
 ∞ ∞

1 1
+ qso dcns dcts asi = qsi dcns dcts
−∞ −∞ cs −∞ 0 cs
 k 
hksi erfc(−Vn ) hsi erfc(−Vn )
= √ −V 2 = √ −V 2 (40)
2 verfc(−Vn ) + n gh√eπ n 2 verfc(−Vn ) + n gh√eπ n
si si
 k ∞ 0

erfc(Vn ) 1
hkso aso = qso dcns dcts
+ √ 2
gh e−Vn (34) −∞ −∞ c s
2 verfc(Vn ) − n √
π 
so
hso erfc(Vn )
Similarly, the gradients at water depth and velocities at x = xs = √ −V 2 (41)
2 verfc(Vn ) − n gh√eπ n
are given as the weighted average of the gradients at water depth so
and velocities at x = xsi and at x = xso : bsi = Asi (42)

k %  &k
h h erfc(−Vn ) bso = Aso (43)
∇ =∇ √ −V 2 ∞ ∞

hv 2 verfc(−V ) + n gh√e n cts
s n π si dsi = qsi dcns dcts
%  &k −∞ 0 cts cs
h erfc(Vn ) 
+∇ √ −V 2 (35) hsi vt erfc(−Vn )
2 verfc(Vn ) − n gh√eπ n ×   √
e−Vn2 (44)
so 2 vt v + t gh 2
erfc(−Vn ) + n vt gh √
π si
Now that the values of (h, vn , vt ) and their spatial gradients at ∞ 0

xsi , xs and xso have been determined, Asi + Aso ; ∇ · (Bsi + Bso ); cts
dso = qso dcns dcts
Dsi + Dso ; and Gs ns + Is ts can all be evaluated. Therefore, the −∞ −∞ c ts cs

only remaining unknown in the right-hand side of (14) is the hso vt erfc(Vn )
∂Es /∂t. Referring to (21), the term ∂Es /∂t becomes known once ×   √
e−Vn
2 (45)
2 vt v + t gh erfc(Vn ) − n vt gh
√
is the time derivatives of (vn , vt , h) at x = xs are determined. 2 π so
∞ ∞


This is accomplished by enforcing the condition that mass and 1 1
momentum are collision invariant for all x and t. In particular, es = qs dcns dcts = hs (46)
−∞ −∞ c s v

 k+1 ∞ ∞   s
0 ∂ gs = Es (47)
= −α1 + α4 c · ∇ + α5 ∞


0 k −∞ −∞ ∂t ∞
cts vt

 hs = qs dcns dcts = hs
1 −∞ −∞ cts cs vt v + t gh
× qs dcns dcts dt 2 s
c (48)
k+1 ∞ ∞
+ (−α1 + α2 c · ∇)
k −∞ 0 Expression (38) along with (39)–(48) completes the estimate of

 the time derivative of flow variables at cell interfaces (i.e., pro-
1
× qsi dcns dcts dt vides ∂es /∂t). This in turn provides ∂Es /∂t. Therefore, all the
c
k+1 ∞ 0 terms in the right-hand side of (14) are known.
The time integration of the instantaneous flux, Ḟs , given by
+ (−α1 + α2 c · ∇)
k −∞ −∞ (14) from time level k to time level k + 1 gives:


1
× qso dcns dcts (36)
c Fs t = γ1 {Asi + Aso } + γ2 {ns · ∇(Bsi + Bso )
Using moments in Appendix, Eq. (36) becomes as follows: ∂
+ ts · ∇(Dsi + Dso )} + t − γ1 + γ5 Es
∂t
∂es 1
= {γ1 {asi + aso } + γ2 {ns · ∇(bsi + bso ) + γ4 {ns · ∇Gs + ts · ∇Is } (49)
∂t γ5
+ ts · ∇(dsi + dso )} − γ1 es
Insertion of Eq. (49) into the right side of (6) completes the deriva-
+ γ4 {ns · ∇gs + ts · ∇hs }} (37) tion of the BGK-based algorithm for shallow water equations
154 Liang et al.

as follows: acronym for Monotone Upstream-centered Schemes for Conser-

Sij % vation Laws. Details of the Hancock-MUSCL and the HLLC can
1 
Wijk+1 = Wijk + Skij t − Ls γ1 (Asi + Aso ) be found in Toro (2001). The data reconstruction for the BGK in
 ij s=1 the two MUSCL schemes is carried out by the Van Leer limiter.
+ γ2 [ns · ∇(Bsi + Bso ) + ts · ∇(Dsi + Dso )] The accuracy and efficiency of the three schemes are investigated
  & using the analytical solutions of the dam-break problem as well
∂
+ t − γ1 + γ5 Es +γ4 [ns · ∇Gs +ts · ∇Is ] as the oblique jump test.
∂t
The accuracy and efficiency of the schemes, for one-
(50) dimensional problems, are measured using the following error
norms:
A summary of how the algorithm advances the solution from ' Nx
time step k to time step k + 1 in any cell (i, j) is as follows: |ei |
ABSERROR = 'Nxi=1 exact (51)
Step 0: Due to the previous time step calculations, the values of i=1 |φi |
'Nx
water depth h and flow velocities v at every node (i, j) at some |ej + 1 − ej |
WAVINESS = i=1 'Nx exact (52)
level k are known. Therefore, Wijk and Skij are known. This data i=1 |φi |
is stored.
where ej = φinumerical − φiexact = difference between the numer-
Step 1: Given the data in step 0, Eqs (25) and (26) are used to ical and analytical solution at node i; φ = dependent variable
calculate the slopes of water depth and flow velocities within cell such as the water depth or velocity; ABSERROR = absolute
(i, j) and its neighbors. error; WAVINESS = waviness error; and Nx = number of nodes
along x. The absolute error is measure of the difference between
Step 2: Given the slopes obtained in step 1, Eq. (24) is used to
the analytical and numerical solution for either the water depth or
calculate the water depth and flow velocities and their respective
the flow velocity. The waviness error is a measure of the differ-
gradients just to the inside of the control surface of cell (i, j) (i.e.,
ence between the gradient of the analytical and the gradient of the
at x = xsi ) and just to the outside of the control surface of cell
numerical solution and quantifies the magnitude of the spurious
(i, j) (i.e., at x = xso ).
oscillation by a numerical scheme (Leonard, 1991).
Step 3: The data in step 2 is projected along the normal and Similarly, the accuracy and efficiency of the schemes, for two-
tangential directions to side s using Eq. (33). dimensional problems, are measured using the following error
Step 4: Given the data in step 3, Eqs (34) and (35) are used to norm:
'Ny 'Nx
calculate (h, vn , vt ) and their gradients at x = xs . j=1 i=1 |eij |
ABSERROR = 'Ny 'N (53)
i=1 |φij |
x exact
Step 5: Given the data in step 3 and step 4, Eqs (15)–(23) and j=1
(40)–(48) are used to calculate the vectors asi ; aso ; bsi ; bso ; dsi ; where Ny = number of nodes along y.
dso ; es ; hso ; Asi ; Aso ; Bsi ; Bso ; Dsi ; Dso ; Es ; Gso ; and Hso .
Step 6: Given the data in step 5, Eq. (38) provides the time 6.1.1 One-dimensional dam-break problem
derivative of (h, vn , vt ) at xs . As a result, ∂Es /∂t is evaluated. The three schemes are first tested using the 1D dam-break test.
Step 7: Given the data in step 5 and step 6, Eq. (51) gives the The dam is located midway of a horizontal and frictionless chan-
net mass and momentum flux from cell (i, j) at control surface s. nel 2000 m long. The initial conditions consist of a 10-m water
These fluxes are stored. depth upstream of the dam and a 0.1-m water depth downstream
of the dam. The ratio of the two water depths are such that the
Step 8: Repeat steps 1–7 for all sides of the cell (i, j) (i.e., for water depth and velocity at the dam location are constant with
s = 1, . . . , Nij ), where Nij = number of sides of cell (i, j). time and the local Froude number is equal to 1. Many numerical
Step 9: With the fluxes stored in step 7 and the data given in step schemes often develop a sonic point at the dam location (Zoppou
k+1
0, Eq. (50) is used to calculate Wi,j and Roberts, 2003). This test problem allows one to investigate
how the different schemes handle extreme shock and expansion
Step 10: Repeat steps 1–9 for all cells (i.e., for all (i, j)). waves and whether or not these schemes are free of spurious
oscillations and sonic points.
The computed and analytical water depth 50 s after the dam-
6 Computational results
break event are shown in Figs 2–4. The tests are performed for
different values of the Courant number, Cr . It is clear that the
6.1 Comparison with analytical solutions
HLLC with theVan Leer limiters is highly oscillatory for Cr larger
For comparison, two approximate Riemann solvers are also than 0.5. As a result, the HLLC scheme was discarded from fur-
implemented. These are the HLLC scheme and the two stage ther considerations. The Hancock-HLLC and the BGK provide
Hancock-HLLC scheme, where HLL stands for Harten, Lax and good resolution for the shock front for all values of Cr ≤ 1. Toro
van Leer and C stands for contact discontinuity. Second-order (1999) notes that first version of Godunov schemes are only stable
accuracy is achieved by MUSCL approach, where MUSCL is the for Cr ≤ 0.5, but the second version of Godunov schemes, to
 Boltzmann-based finite volume algorithm 155

12 12
Analytical
Analytical
BGK
BGK
MUSCL HLLC 10
10 MUSCL HLLC
MUSCL-Hancock HLLC
MUSCL Hancock HLLC
8
8 h(m)
6
h(m)

6 4

2
4

0
2 0 500 1000 x(m ) 1500 2000

Figure 3a Water depth 50 seconds from the moment the dambreak event
0 occurred (Cr = 0.7, N = 200).
0 500 1000 x(m) 1500 2000

Figure 2a Water depth 50 seconds from the moment the dambreak event
occurred (Cr = 0.9, N = 200). 3
Analytical
BGK
2.5 MUSCL HLLC
MUSCL Hancock HLLC
2
h(m)
3 1.5
Analytical
BGK
MUSCL HLLC 1
2.5 MUSCL-Hancock HLLC

0.5

2 0
1200 1300 1400 1500 x(m ) 1600 1700
h(m)
Figure 3b Water depth 50 seconds from the moment the dambreak event
1.5
occurred, from 1200 m to 1700 m (Cr = 0.7, N = 200).

1
11
Analytical
BGK
0.5 10.5 MUSCL HLLC
MUSCL Hancock HLLC
10
h(m )
0
1200 1300 1400 1500 x(m ) 1600 1700 9.5

Figure 2b Water depth 50 seconds from the moment the dambreak event 9
occurred, from 1200 m to 1700 m (Cr = 0.9, N = 200).
8.5

8
300 400 500 600 x(m ) 700 800

11
Analytical
Figure 3c Water depth 50 seconds from the moment the dambreak event
BGK occurred, from 300 m to 800 m (Cr = 0.7, N = 200).
10.5 MUSCL HLLC
MUSCL HancockHLLC

10
h(m )
which Hancock-HLLC belongs, are stable as long as Cr ≤ 1.0. In
9.5
addition, Zhou et al. (2001) applied Godunov schemes to surface
water problems for Cr = 0.3, 0.65 and 1.0. They report that no
9 stablity problem was encounered. The stability of BGK schemes
for linear advection–diffusion problem has been investigated by
8.5 Torrilhon and Xu (2006). The stability characteristics depend on
Cr , the grid-based Reynolds number, Rg = Ux/ν, and the ratio
8 of the flow speed to the wavespeed U/C (i.e., Mach number for
300 400 500 600 x(m ) 700 800 gas flows and Froude number for gravity flows). For the inviscid
Figure 2c Water depth 50 seconds from the moment the dambreak event case, the BGK is always stable for Cr ≤ 1.0 for small to moderate
occurred, from 300 m to 800 m (Cr = 0.9, N = 200). values of U/C and Cr ≤ 1.3 for large values of U/C.
156 Liang et al.

12
Analytical
dependent variables, are required to remove the spurious oscilla-
BGK tions. Such gradients are in the naturally included in the fluxes of
10 MUSCL HLLC the BGK scheme, as explains why the BGK results contain less
MUSCL Hancock HLLC
spurious oscillations than the Hancock-HLLC.
8 The results in Figs 2–4 show that for the same discretization
h(m) the BGK is more accurate than the Hancock-HLLC scheme. A
6 more conclusive comparison requires measurement of the central
processing unit (CPU) time needed by each scheme to achieve a
4
given level of accuracy. Figure 5(a, b) depicts the absolute and
waviness error as function of CPU time for both schemes. With
2
reference to the absolute error, the Hancock-HLLC scheme is
more efficient than the BGK scheme despite the fact that, for a
0
0 500 1000 x(m )1500 2000
given level of absolute error, the Hancock-HLLC scheme requires
a finer grid than the BGK. The efficiency of the BGK scheme is
Figure 4a Water depth 50 seconds from the moment the dambreak event
affected by the fact that its fluxes require the evaluation of the
occurred (Cr = 0.4, N = 200).
error function. In terms of the waviness error, the efficiency of
the BGK is close to that of the Hancock-HLLC scheme. Figure 6
3
Analytical shows the absolute error versus the number of reaches for both the
BGK
2.5 MUSCL HLLC
BGK and the Hancock-HLLC schemes. It is clear that the rate of
MUSCL-Hancock HLLC convergence of both schemes are comparable. Although memory
2 storage is generally not a major consideration in 1D problems, it is
h(m) noted that the Hancock-HLLC requires twice as much as memory
1.5 storage as the BGK. This is because (i) the Hancock-HLLC is a
two-step method while the BGK is a single step method (i.e.,
1
the Hancock-HLLC requires storage of intermediate variables,
0.5
but the BGK does not), and (ii) the Hancock-HLLC schemes

0
1200 1300 1400 1500 x(m ) 1600 1700 0.05
BGK

Figure 4b Water depth 50 seconds from the moment the dambreak event Hancock MUSCL HLLC
0.04
occurred, from 1200 m to 1700 m (Cr = 0.4, N = 200).

0.03
11
Analytical Absolute error
BGK
MUSCL HLLC 0.02
10.5
MUSCL HancockHLLC

0.01
10
h(m ) CPU tim e
9.5 0
0 2 4 6 8 10 12 14 16

9 Figure 5a Absolute Error Versus CPU time for the one-dimensional

Dam-Break Test.
8.5

0.01
8 BGK
300 400 500 600 x(m ) 700 800
Hacock HLLC
0.008
Figure 4c Water depth 50 seconds from the moment the dambreak event
occurred, from 300 m to 800 m (Cr = 0.4, N = 200).
0.006

Waviness
Note that the results of Hancock-HLLC contain slight spu- 0.004
rious oscillations at the point where the expansion wave meets
the constant state, while the BGK is essentially oscillation free. 0.002
Theoretically, MUSCL-type schemes are known to be oscillation-
CPU tim e
free for linear problems. However, it is widely recognized that a 0
mere non-linear limiter in the reconstruction stage is insufficient 0 2 4 6 8 10 12 14 16

to suppress all oscillations for non-linear equations. Fluxes which Figure 5b Waviness Error Versus CPU time for the one-dimensional
include the information of the whole cell, i.e., gradients of the Dam-Break Test.
 Boltzmann-based finite volume algorithm 157

1
6.1.3 Oblique hydraulic jump problem
100 1000
Number of grids This problem is used to test the capability of the BGK and
the Hancock-HLLC in solving shock waves in two-dimensional
domain. When a supercritical flow goes through a converging
0.1
channel, the flow is “chocked” and a hydraulic jump forms. The
ABSERROR angle of convergence of the channel is 8.95◦ . A structured grid
with quadrilateral cells is used in all computations (see Fig. 8).
0.01
The inflow velocity and depth are 8.57 m/s and 1 m. Therefore,
the in flow consists of a supercritical flow with a Froude number
2.74. According to Zhao et al. (1996), exact solutions of shock
BGK
height, shock wave angle, velocity and Froude number after the
shock are 1.5 m, 30.00◦ , 7.953 m/s and 2.074, respectively. The
Hancock HLLC
0.001

Figure 6 Absolute Error Versus number of grids for the One-di- tests are carried out using the following mesh sizes: 40 × 60,
mensional Dam-Break Test (Cr = 0.9, in Logarithmic Plot). 80 × 60, 160 × 60 and 320 × 60. There are no significant differ-
ence between the results by both models and the exact solutions.
Water depth contour and profile by BGK scheme are shown in
requires larger number of grid points than the BGK scheme for Figs 9 and 10. Close inspection of the longitudinal surface water
the same level of accuracy as seen in Fig. 6.

6.1.2 Transition from subcritical to supercritical flow

over a channel hump
To test the ability of the schemes to simulate flow over varied
bottom geometry, the test rig in Navarro et al. (1992) and Meselhe
et al. (1997) is simulated. The elevation of the channel bed, y in
meters, is given as follows (Navarro et al., 1992; Meselhe et al.,
1997):

5 sin2 (π x−125 ) if 125 m ≤ x ≤ 875 m
y(x) = 750 (54)
0 elsewhere

The channel is rectangular with total length of 1000 m, which is

discretized into 41 computational cells of 25 m each. The water
depth at the upstream and downstream boundaries are 10 m and
7 m, respectively.
The analytical and computed water surfaces along the chan-
nel are plotted in Fig. 7. Good agreement between the computed
results and exact solutions is evident. In particular, both schemes
Figure 8 Mesh for the Oblique Hydraulic Jump problem.
accurately capture the transition from subscritical to supercritical
at the hump and the transition from supercritical to subcriti-
cal at the downstream of the hump. This high accuracy further 30
demonstrates the ability of both schemes in modeling mixed flow
regimes.
25

12
Exact Solution 20
BGK
10
Hancock-HLLC
15
8
h(m )
6 10

4
5

0 5 10 15 20 25 30 35 40
0 200 400 x(m ) 600 800 1000
Figure 9 Water depth contours for the Oblique Hydraulic Jump test
Figure 7 Water surface profile for flow over a hump. produced by the BGK Scheme.
158 Liang et al.

Path 3
14 (boundary)
13
12
11
10
9
8
Path 2
7

y(m)
6
5
4 inflow
3
2
1
Path 1
0
0 10 20 30(centerline)
x(m)

Figure 13 Curvilinear expanding channel (Fr = 2.0 at entrance;

Figure 10 Surface water profile for the Oblique Hydraulic Jump test S0x = S0y = 0.0; n = 0.012; cells = 60 × 21(x × y)).
produced by the BGK Scheme.

6.2 Comparison with experimental and numerical data

1.6 6.2.1 Supercritical flow in a curvilinear expanding channel

Exact Solution
This test case consists of a steady supercritical flow in a curvi-
BGK
linear expanding channel, which was studied experimentally by
Hancock-HLLC
1.4 Rouse et al. (1951). The channel is symmetrical about the center-
line. Figure 13 depicts the plan view of half of the flume where the
h(m)
curvilinear expanding side-wall is given by the following formula
1.2  3/2
y 1 x 1
= + (55)
b1 2 2b1 2
1
where b1 = 10.0 m = channel width at the inlet (i.e., at x = 0).
At the inlet, the water depth h1 is 1.0 m, the transverse velocity
component is 0.0 m/s, and the Froude number is 2.0. The bottom
0.8
slopes of the channel in x and z directions are zero. The Man-
0 5 10 15 20 25 x(m ) 30 35 40
ning coefficient is n = 0.012. The domain is discretized into 60
Figure 11 Longitudinal depth profile for the Oblique Hydraulic Jump
reaches along the x direction and 21 reaches along the z direction.
test.
Figure 14 depicts the computed and measured water depth
along Paths 1, 2 and 3. The agreement between the computed
and measured water depths is in general satisfactory.
1.40E-02
Hancock-MUSCL HLLC
BGK
1.20E-02 6.2.2 Transition from subcritical to supercritical flow
in a converging channel
1.00E-02
A steady flow in a converging channel, where a transition from
8.00E-03 subcritical to supercritical regime occurs, is simulated. The
ABSERROR experimental data for this test case were carried out by Coles
6.00E-03
and Shintaku (1943) and the information related to the test
4.00E-03 rig and flow conditions are described in Molls and Chaudhry
(1995). Their laboratory flume (Fig. 15) consisted of two
2.00E-03
straight, rectangular sections joined by a 1.49-m long, straight-
CPU time
0.00E+00 walled contraction. The upstream section was 0.629 m wide,
0 50 100 150 200 250 and the downstream section was 0.314 m wide. The computa-
Figure 12 Absolute Error Versus CPU time for both schemes for the tional domain is divided into 40 × 20 grids. The flow rate was
Oblique Hydraulic Jump Test. 0.0451 m3 /s. The Chezy coefficient was 84 m1/2 /s. The chan-
nel is symmetrical about the centerline. The flow is subcritical
at the inlet and throughout most of the converging channel and
supercritical near the outlet. In the computation, the upstream
profile (Fig. 11) shows that that the Hancock-HLLC exibits slight depth (h = 0.1762 m) and downstream depth (h = 0.115 m)
oscillations for 26 m < x < 31 m. Figure 12 shows that both are specified. Figure 16 shows that the computed and measured
schemes have comparable efficiency in the sense of the absolute water profiles along Path 1 (i.e., channel centerline) are in good
error. agreement.
 Boltzmann-based finite volume algorithm 159

1.2 Experiment (Path 1)

Experiment (Path 2)

Experiment (Path 3)
1 BGK (Path 1)

BGK (Path 2)

BGK (Path 3)
0.8 MUSCL Hancock HLLC(Path 1)

MUSCL Hancock HLLC(Path 2)

h(m)
MUSCL Hancock HLLC(Path 3)
0.6

0.4

0.2

0
0 5 10 15 x(m) 20 25 30

Figure 14 Comparison between computed and measured water depths (Fr = 2.0 at entrance; S0x = S0y = 0.0; n = 0.012; cells = 60 × 21 (x × y)).

0.3

0.2
Path 2
0.1
(boundary)
inflow
0
y(m)

Path 1
-0.1 (centerline)

-0.2

0 0.5 1 1.5 2
x(m)

Figure 15 Geometry of converging channel (flow rate = 0.0451 m3 /s;

S0x = S0y = 0.0; C = 84.0; cells = 40 × 20 (x × y)).

0.2
Experiment
BGK Figure 17 Water surface plot by the BGK scheme (Fr > 1.0 for all
MUSCL-Hancock HLLC x, i.e., supercritical flow) (flow rate = 0.0451 m3 /s; S0x = S0y = 0.0;
0.18
C = 84.0; cells = 40 × 20 (x × y); upstream depth h = 0.0314 m;
h(m) upstream velocity u = 2.25 m/s).
0.16

0.14
experiment for a highly supercritical flow. The experimental flow
rate was 0.0451 m3 /s with a Froude number of 4. The upstream
0.12
boundary conditions were specified with h = 0.0314 m, u =
2.25 m/s and v = 0 m/s.
0.1
0 0.5 1 x(m) 1.5 2
Figure 17 shows the simulated water surface by both mod-
els, where the formation of oblique waves is evident. The
Figure 16 Computed and measured profiles along the centerline of
the channel (flow rate = 0.0451 m3 /s; S0x = S0y = 0.0; C = 84.0; high density of water depth contours at the location of oblique
cells = 40 × 20 (x × y); upstream depth h = 0.1762 m; downstream waves (see Fig. 18) illustrates the ability of the scheme to
depth h = 0.115 m). resolve shocks and steep gradients. Figures 19 and 20 show
that both schemes are in excellent agreement. In addition,
6.2.3 Supercritical flow in a converging channel while the wave height is well represented by both schemes,
In the same converging channel as described above and shown the wave location is not. Similar discrepancies in wave loca-
in Fig. 15, Coles and Shintaku (1943) carried out another tion have been found by other researchers (e.g., Bhallamudi
160 Liang et al.

0.3 7 Application of the BGK scheme to viscous flows

0.2 The examples of the last section test the ability of the BGK and
the Hancok-HLLC schemes in resolving shallow water waves.
As was stated in the introduction, macroscopic schemes, such

5
0.
0.1

6
04 0.0
05

0.0
0.

5
0 .0 6 7 5 as those based on the Riemann solution, can only solve hyper-
35

75
06
0.
bolic (wave) equations. Application of Riemann schemes to

0.0775

0.06
y(m)

0.
03

0.06
25
0

5
hyperbolic–parabolic equations requires that the problem is split

0.
06
into a hyperbolic part to which the Riemann solution can be

75 0.06
-0.1
applied and a parabolic part for which another scheme needs to be

5
devised. This section illustrates that the BGK schemes does not
-0.2
require any splitting and that the fluxes given by Eq. (50) incor-
porate the effects of diffusion via the collision time. In fact, it is
0 0.5 1 1.5 2 well known that the viscous terms in the Navier–Stokes equations
x(m) are recovered from the collision term in the BGK equation by set-
Figure 18 Water depth contour by the BGK scheme (Fr > 1.0 for all ting τ = µ/p, where µ = dynamic viscosity and p = pressure
x, i.e., supercritical flow) (flow rate = 0.0451 m3 /s; S0x = S0y = 0.0; (e.g., Vincenti and Kruger, 1965). Similarly, the viscous terms
C = 84.0; cells = 40 × 20 (x × y); upstream depth h = 0.0314 m; in the shallow water equations are recovered from the collision
upstream velocity u = 2.25 m/s). term in the BGK equation by setting τ = µ/ρgh = ν/gh (e.g.,
Ghidaoui et al., 2001; Zhou, 2004). The results of this shows
0.1 that the BGK scheme provides the viscous flow solution by set-
ting the collision time in the right-hand side of Eq. (50) equal
0.08 to ν/gh.
h(m)
Let us consider a steady-uniform-laminar flow in a channel
0.06 of width W and slope S0 for which an analytical solution can be
derived. The governing equations for a steady-uniform-laminar
0.04 flow in a channel with no bottom friction are obtained from (4)
and the result is as follows:

0.02 Experiment
 hu 
∇· 2
h u2 + gh2 I − ρ1 Γ
BGK
MUSCL-Hancock HLLC
0


0 0.6 1.2 x(m) 1.8 2.4
0 d2 u
Figure 19 Computed and measured profiles along Path 2 (i.e., bound- = ⇒ −ν 2 = gS0 (56)
ghS0 dy
ary) (Fr > 1.0 for all x, i.e., supercritical flow; flow rate = 0.0451 m3 /s;
S0x = S0y = 0.0; C = 84.0; cells = 40 × 20 (x × y); upstream depth
Integrating the above ordinary differential equation and imposing
h = 0.0314 m; upstream velocity u = 2.25 m/s). the no slip condition [i.e., u(y = ±W/2) = 0] and the flow
symmetry condition (i.e., du/dy = 0 at y = 0) gives:
 
gSo W 2
0.1
u(y) = −y 2
(57)
ν 4
0.08 The analytical solution of the velocity profile in viscous prob-
lems with non-zero bottom friction can also be derived. The
0.06
h(m) shear between the fluid and the channel bottom is given by
τw = u2 ρCf /2 = (νPo /2R)u, where Cf = friction coeffi-
0.04
cient; R = hydraulic radius and Po = Poiseuille coefficient. The
0.02 Experiment
Poiseuile coefficient depends on the geometry of the channel and
BGK has a value of 16 for a circular channel and 24 for wide rectangu-
0
MUSCL-Hancock HLLC
lar channel (White, 1991). Inserting τ = u2 ρCf /2 = (νPo /2R)u
0 0.6 1.2 x(m) 1.8 2.4
into (4) and solving gives
Figure 20 Computed and measured profiles along Path 1 (i.e., center-  √
2gS0 R eW po /2R/2
line) (Fr > 1.0 for all x, i.e., supercritical flow) u(y) = − √
Po ν 1 + eW po /2R

 √  
× e Po /2Ry + e− Po /2Ry − 1 (58)
and Chaudhry, 1992; Molls and Chaudhry, 1995). The precise
identification of the location of the oblique wave in a channel The tests are carried out for the cases of a rectangular channel
requires that velocity distribution and wavespeed be accurately 4 m wide. The water depth is 0.5 m. The channel slope So is 10-
determined. 7 for the test without bottom friction and 3 × 10−5 for the test
 Boltzmann-based finite volume algorithm 161

equation. In addition, Chen et al. (2003) noted “. . . it is simpler

to model complex physics first at the Boltzammn (kinetic) level,
and then use coarse graining procedures [i.e., moments] only
after modeling is done.” Such an approach has been successfully
used in gas dynamics, in multi-phase and multi-component flows
as well as porous media flows. The BGK approach has not yet
been fully explored for shallow water problems. In this paper, the
relation between the BGK model and the shallow water equations
is used to derive numerical schemes by the discretization of the
BGK equation and not the shallow water equations. The algo-
rithm for shallow water equation is obtained by taking moments
of the discrete BGK equation. The resulting scheme incorpo-
rates the physics of waves and diffusion because both of these
Figure 21 The computed and analytical dimensionless velocity profile physical phenomena are naturally included in the BGK equation.
solution for the viscous flow—viscous effects at the vertical walls are The BGK scheme is second order in both time and space. The
included and the bottom friction is zero.
formulation of the BGK algorithm is performed for a cell of arbi-
trary irregular shape, but the test cases are conducted using a
1
structured grid of quadrilateral cells. Two approximate Riemann
u
solvers are also considered. These are: the HLLC scheme and the
Analytical
0.5 two-stage Hancock-HLLC scheme. The second-order accuracy
BGK
is achieved by MUSCL approach. The data reconstruction for all
three schemes is carried out by the Van Leer limiter. The test cases
0 involve strong shocks and expansion waves. The accuracy of the
0 0.2 0.4 0.6 0.8 x 1
schemes are measured using an absolute error norm, a waviness
-0.5
error norm, and a maximum error norm. The main results are as
follows.
1. The HLLC scheme is highly oscillatory for Courant num-
-1
ber larger than 0.5, while the BGK and the Hancock-HLLC
Figure 22 The computed and analytical dimensionless velocity profile
schemes are applicable for Courant numbers as high as 1.0.
solution for the viscous flow—viscous effects at the vertical walls are
2. For one-dimensional hyperbolic problems and for a fixed
included and bottom friction is non-zero.
value of the CPU time, the absolute error of the Hancock-
HLLC is slightly smaller than that of the BGK while the
with bottom friction. The fluid is water with ν = 10−6 m2 /s and waviness error of the Hancock-HLLC is slightly smaller than
Po = 24. The results in Figs 21 and 22 show good agreement that of BGK. Overall, the Hancock-HLLC model is about
between the BGK scheme and the analytical solution. 10–15% faster to execute than the BGK model.
The test cases in this paper illustrate that the BGK scheme 3. For two-dimensional hyperbolic problems both schemes are
resolves both waves and diffusion. That is, BGK schemes pro- of comparable efficiency.
vide the complete solution for hyperbolic and parabolic equations 4. The BGK requires less storage space than the Hancock-HLLC.
without any need for operator splitting. It is noted that turbulent 5. Both the BGK and Hancock-HLLC schemes have similar
diffusion can easily be solved by BGK schemes. This is accom- convergence properties.
plished by setting the collision time in the right-hand side of (52) 6. An attractive feature of BGK schemes is their ability to model
equal to (ν + νe )/gh, where νe = eddy viscosity. This approach diffusion as well as waves without any need for operator
has been successfully used by (e.g., Chan et al., 2004; Ghidaoui splitting. That is, BGK schemes solve hyperbolic–parabolic
et al., 2006; Zhou, 2004). In addition, a recent paper by Chen problems. The ability of the BGK algorithm to model dif-
et al. (2004) shows that the classical turbulence models such as the fusion is illustrated using viscous flow problems. Excellent
eddy viscosity zero closure model and the k–ε model are recov- agreement between the analytical and computed viscous flow
erable from the collision term in the BGK equation. In addition, solution is found.
they also showed that the time lag in the response of turbulent
field to flow unsteadiness naturally emerges from the Boltzmann Although the BGK and Hancock-HLLC schemes perform
framework. similarly for hyperbolic problems, BGK schemes have the added
advantage of being able to solve hyperbolic–parabolic problems
without the need for an ad-hoc operator splitting. This is impor-
8 Conclusions tant given that the artificial splitting of advection and diffusion is
known to cause artificial widening of shear layers and introduce
It is well known that the Navier–Stokes and the shallow water artificial transient in regions with sharp gradients. Such problems
equations can be obtained by taking moments of the BGK arise when the splitting operation fails to faithfully represent the
162 Liang et al.

Table 1 Comparison of computed with Exact Solution. The exact The moments of q for 0 ≤ cns < ∞ are as follows:
values are taken from Zhao et al. 1996.
1 ∞ ∞ 1
µn |>0 =
0
q dcns dcts = erfc(−Vns )
Scheme Velocity Water Shock Froude h −∞ 0 2
Angle (◦ )
(m/s) Height (m) Number 1 ∞ ∞
µ1n |>0 = cns q dcns dcts
Exact 7.952 1.500 30.00 2.074 h −∞ 0
BGK 7.949 1.498 30.10 2.074 √  2

gh e−Vns
MUSCL 7.949 1.498 30.10 2.074 = Vns erfc(−Vns ) + √
Hancock HLLC 2 π
∞ ∞
1
µ2n |>0 = c2 q dcns dcts
h −∞ 0 ns
correct coupling between the physics of advection and the physics   2

of waves. gh 1 Vns e−Vns
= Vns +
2
erfc(−Vns ) + √
A note of caution. The present work neither intends nor shows 2 2 π
that the BGK approach is superior to other approaches for all sur- √
1 ∞ ∞ 3 gh gh
face water problems. Rather, this work explores the feasibility of µ3n |>0 = cns q dcns dcts =
h −∞ 0 2
using the BGK approach to model surface water problems and 
 2
shows that this approach provides an alternative framework for 3Vns 2
(Vns + 1) e−Vns
× Vns + 3
erfc(−Vns ) + √
modeling the physics of waves and diffusion in open channels in 2 π
a consistent manner and without any need for operator splitting. ∞ ∞
1
This is an important feature of BGK schemes and “paves the way” n |>0 =
µm cm q dcns dcts
h −∞ 0 ns
for the development of surface water models which avoid the
√ 
problems associated with ad-hoc operator splitting methods. In  gh
= gh Vns µn |>0 +
m−1
(m − 1)µm−2n |>0
addition, we note that researchers in other fields have successfully 2
used the Boltzmann framework to formulate numerical algo- The moments of q for −∞ < cns ≤ 0 are as follows:
rithms and to understand and formulate theories for flows involv-
ing complex physics such as multiphase flows and turbulence. 1 ∞ 0 1
µ0n |<0 = q dcns dcts = erfc(Vns )
The application of the BGK approach to surface water flows h −∞ −∞ 2
has thus far been fruitful and further research is needed to fully ∞ 0
1
explore the merits of this approach in the water resources field. µ1n |<0 = cns q dcns dcts
h −∞ −∞
√  2

gh e−Vns
Acknowledgment = Vns erfc(Vns ) − √
2 π
∞ 0
We acknowledge the financial support by the Research Grants 1
µ2n |<0 = c2 q dcns dcts
Council (RGC) of Hong Kong: project number DAG01/02.EG24 h −∞ −∞ ns
and project number HKUST6227/04E.   2

−Vns
gh 1 V ns e
= 2
Vns + erfc(Vns ) − √
2 2 π
Appendix: Moments of the Gaussian distribution √
1 ∞ 0 3 gh gh
µ3n |<0 = cns q dcns dcts =
In the following expressions Vns is the dimensionless normal h −∞ −∞ 2
√  
velocity along an element side such that Vns = vns / gh. 3Vns (Vns2
+ 1) e−Vns
2

The moments of q for −∞ < cns < ∞ are as follows: × Vns + 3

erfc(Vns ) − √
2 π

1 ∞ ∞ ∞ 0
µn =
0
q dcns dcts = 1 1
n |<0 =
µm cm q dcns dcts
h −∞ −∞
h −∞ −∞ ns
1 ∞ ∞ 
√ 
µ1n = cns q dcns dcts = ghVns  gh
h −∞ −∞ = gh Vns µm−1 |<0 + (m − 1)µ m−2
|<0

 n
2 n
1 ∞ ∞ 2 1
µn =
2
c q dcns dcts = gh Vns +2
h −∞ −∞ ns 2
∞ ∞

1  3Vns Notation
µ3n = 3
cns q dcns dcts = gh gh Vns 3
+
h −∞ −∞ 2
∞ c = Particle velocity vector
1 f = Non-equilibrium particle distribution function
µmn = cm q dcns dcts
h −∞ ns g = Gravity

√ 
 gh h = Water depth
= gh Vns µm−1 + (m − 1)µ m−2
n
2 n I = Identity tensor
 Boltzmann-based finite volume algorithm 163

m = Mass of fluid particle 10. Frisch U., Hasslacher B. and Pomeau Y. (1986). “Lattice
n = Unit outward vector normal to the surface of control Gas for the Navier–Stokes Equations”. Phys. Rev. Lett. 14,
volume 1505–1508.
p = Pressure 11. Garcia-Navarro, P., Alcrudo, F. and Saviron, J.M.
q = Equilibrium particle distribution function (1992). “1-D Open-Channel Flow Simulation Using TVD-
S = Equilibrium particle distribution function McCormack Scheme”. J. Hydraul. Engrg. 118(10),
S0 = Channel slope vector 1359–1372.
Sf = Friction slope 12. Ghidaoui, M.S., Deng, J.Q., Gray, W.G. and Xu, K.
t = Time (2001). “A Boltzmann Based Model for Open Channel
t = Unit vector tangent to the surface of control volume Flows”. Int. J. Numer. Methods Fluids 35(4), 449–494.
v = Depth-averaged velocity 13. Ghidaoui, M.S., Kolyshkin, A.A., Liang, J.H., Chan,
x = Longitudinal coordinate F.C., Li, Q. and Xu, K. (2006). “Linear and Nonlinear
∇ = Two-dimensional del operator with respect to x and y Analysis of Shallow Wakes”. J. Fluid Mech. 548, 309–340.
Γ(x, t) = Viscous stress tensor 14. Karlsen, K.H., Lie, K.-A., Natvig, J.R., Nordhaug,
µ = Dynamic viscosity H.F. and Dahle, H.K. (2001). “Operator Splitting Methods
ν = Kinetic viscosity for Systems of Convection-Diffusion Equations: Nonlinear
i,j = Area of control volume Error Mechanisms and Correction Strategies”. J. Comput.
∂i,j = Boundary of control volume Phys. 173, 636–663.
ρ = Density of fluid 15. Kumar, R., Nivarathi, S.S., Davis, H.T., Kroll, D.M.
τ = Collision time and Maier, R.S. (1999). “Application of the Lattice-
τw = Wallshear Boltzmann Method to Study Flow and Dispersion in
Channels with and without Expansion and Contraction
Geometry”. Int. J. Numer. Methods Fluids 31, 801–819.
References 16. Leonard, B.P. (1991). “The ULTIMATE Conservative
Difference Scheme Applied to Unsteady One-dimensional
1. Bhallamudi, M.S. and Chaudhry, M.H. (1992). “Two Advection”. Comput. Methods Appl. Mech. Engng. 88,
dimensional modelling of supercritical and subcritical flow 17–74.
in channel transitions”. J. Hydrau. Engrg. 30(1), 77–91. 17. Mellor, G.L. (2004). “User Guide for a Three-
2. Chan, F.C., Ghidaoui, M.S. and Li, Q. (2004). “Numer- dimensional, Primitive Equation, Numerical Ocean Model”.
ical Modeling of Island Wakes in Shallow Waters”. In: http://www.aos.princeton.edu/.
International Conference on Computational Methods in 18. Meselhe, E.A., Sotiropoulos, F. and Holly, F.M.
Water Resources. Chapel Hill, North Carolina, USA, (1997). “Numerical Simulation of Transcritical Flow in
pp. 1731–1742. Open Channels”. J. Hydraul. Engng. ASCE l23(9),774–783.
3. Chen, S. and Doolen, G.D. (1998). “Lattice Boltzmann 19. Molls, T. and Chaudhry, M.H. (1995). “Depth-averaged
Method for Fluid Flows”. Ann. Rev. Fluid Mech. 30, Open-channel Flow Model”. J. Hydraul. Engng. ASCE
329–364. 121(6), 453–465.
4. Chen, H., Kandasamy, S., Orszag, S., Shock, R., Succi, 20. Reitz, R.D. (1981). “One Dimensional Compressible Gas
S. and Yakhot, V. (2003). “Extended Boltzmann Kinetic Dynamics Calculations Using the Boltzmann Equations”.
Equation for Turbulent Flows”. Science 301, 633–636. J. Comput. Phys. 42, 108–123.
5. Chen, H., Orszag, A., Staroselsky, I. and Succi, S. 21. Rouse, H., Bhoota, B.V. and Hsu, E. (1951). “Design of
(2004). “Expanded Analogy between Boltzmann Kinetic Channel Expansions”. Trans. ASCE 116, 347–363.
Theory of Fluids and Turbulence”. J. Fluid Mech. 519, 22. Su, M., Xu, K. and Ghidaoui, M.S. (1998). “Low Speed
301–314. Flow Simulation by the Gas-Kinetic Scheme”. J. Comput.
6. Coles, D. and Shintaku, T. (1943). “Experimental Rela- Phys. 150, 17–39.
tion between Sudden Wall Angle Changes and Standing 23. Toro, E.F. (1999). Riemann Solvers and Numerical Meth-
Waves in Supercritical Flow”. B.S. Thesis, Lehigh Univer- ods for Fluid Dynamics. Springer, Berlin.
sity, Bethlehem, Pa. 24. Toro, E.F. (2001). Shock Capturing Methods for Free-
7. Delft Hydraulics (2003). User Manual Delft3D-FLOW. Surface Flows. John Wiley & Sons, Chichester.
http://www.wldelft.nl/soft/d3d/intro/index.html. 25. Torrilhon, M. and Xu, K. (2006). “Stability and Con-
8. Deng, J.Q., Ghidaoui, M.S., Gray, W.G. and Xu, K. sistency on Kinetic Upwinding for Advection-Diffusion
(2001). “A Boltzmann Based Mesoscopic Model for Con- Equations”. Submitted to SIAM J. Numer. Anal.
taminant Transport in Flow Systems”. Adv. Water Resour. 26. Vincenti, G.H. and Kruger, C.H., Jr. (1965). Introduc-
24, 531–550. tion to Physical Gas Dynamics. Krieger Publishing Co.,
9. Deng, J. (2000). “BGK Boltzmann Models for Free-surface Malabar, FL USA.
Water Flows and Contaminant Transport”. PhD Thesis, 27. White, F.M. (1991). Viscous Fluid Flow, 2nd edn.,
UMI. A Bell & Howell Company. McGraw-Hill Inc., New York.
164 Liang et al.

28. Xu, K. (2001). “A Gas-Kinetic BGK Scheme for the Shock Wave Modeling”. J. Hydraul. Engng. ASCE 122(12),
Navier–Stokes Equations and Its Connection with Artificial 692–702.
Dissipation and Godunov Method. J. Comput. Phys. 171, 32. Zhou, J.G., Causon, D.M., Mingham, C.G. and Ingram,
289–335. D.M. (2001). “The Surface Gradient Method for the Treat-
29. Xu, K., Mao, M. and Tang, L. (2005). “A Multidimensional ment of Source Terms in the Shallow-water Equations”. J.
Gas-Kinetic BGK Scheme for Hypersonic Viscous Flow”. Comput. Phys. 168, 1–25.
J. Comput. Phys. 203, 405–421. 33. Zhou, J.G. (2004). Lattice Boltzmann Methods for Shallow
30. Zhang, S.Q., Ghidaoui, M.S., Gray, W.G. and Li, N. Water Flows. Springer-Varleg, Hong Kong.
(2003). “A Kinetic Flux Vector Splitting Scheme for Shallow 34. Zoppou, C. and Roberts, S. (2003). “Explicit Schemes for
Water Flows”. Adv. Water Resour. 26(6), 635–647. Dam-Break Simulations”. J. Hydraul. Engng. ASCE 119(1),
31. Zhao, D.H., Shen, H.W., Lai, J.S. and Tabios, G.Q. (1996). 11–34.
“Approximate Riemann Solvers in FVM for 2D Hydraulic