Performance of High-Resolution, Nonlevel Bed,

Shallow-Water Models
Scott F. Bradford1 and Brett F. Sanders2

Abstract: Several techniques are investigated for extending Roe’s finite volume method to second-order spatial and temporal accuracy
in order to better simulate shallow water flow over nonlevel beds. Specifically, methods reported in the literature for limiting and
Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

reconstructing dependent variables, discretizing source terms, and time stepping are examined. Steady subcritical and supercritical flow as
well as unsteady flow over nonlevel beds is simulated in order to evaluate each methods’ performance in terms of error and computational
effort. There is not a single method which minimizes error and computational effort for all test cases. On the contrary, methods that
perform best in one application may be suboptimal in another. A general purpose modeling approach that balances accuracy against
computational effort is identified and recommended.
DOI: 10.1061/共ASCE兲0733-9399共2005兲131:10共1073兲
CE Database subject headings: Shallow waters; Simulation; Water flow; Hydraulic models.

Introduction Riemann problem, which can then solved to compute the fluxes
at the cell boundaries 关often with the approximate method de-
Finite volume schemes update cell average values of the con- veloped by Roe 共1981兲. An alternative method developed by
servative variables using fluxes computed at cell boundaries, Engquist and Osher 共1981兲 was found to perform very similarly
which ensures local mass conservation. For one-dimensional 共1D兲 to Roe’s method in the test cases considered in this paper. All
frictionless shallow water flow, explicit Euler time stepping presented results are obtained with Roe’s method.兴 Solving a
yields the following model with O共⌬t兲 accuracy for computa- Riemann problem imparts the physics to the numerical model
tional cell j and makes it applicable to subcritical and supercritical flow,
which further adds to the model’s appeal. Roe’s method decom-
Un+1
j − Unj 1 n poses the flux difference across the cell boundary into two trav-
+ 共F − Fnj−1/2兲 = Snj 共1兲 eling discontinuities, which are then used to estimate the flux at
⌬t ⌬x j+1/2
the boundary as
where U j = 共d j q j兲T = 1 / ⌬x兰 j−1/2
j+1/2
共d q兲Tdx and
ˆ −⌬V̂兲
F j±1/2 = FLj±1/2 + 共R⌳

冢 冣 冢 冣
j±1/2
q 0

冕 ˆ +⌬V̂兲
= FRj±1/2 − 共R⌳
j+1/2
F= g S= 1 dz 共2兲 j±1/2
uq + d2 − gd dx
2 ⌬x dx 1 ˆ 兩⌬V̂兲
= 2 关FLj±1/2 + FRj±1/2 − 共R̂兩⌳ j±1/2兴 共3兲
j−1/2

where d = flow depth; u = depth-averaged velocity; q = du; L R

where F and F denote the fluxes evaluated to the left and right
g = acceleration due to gravity; z = bed elevation; n denotes the of the boundary, respectively, and ⌬ denotes the finite difference
time level; ⌬t = time increment; ⌬x = cell length; and j + 1 / 2 and across the cell boundary. The quantities denoted with a hat are
j − 1 / 2 denote the right and left cell boundaries, respectively.
Roe averages of the data that ensure that ⌬F = Â⌬U, where A
While the cell average approach preserves the conservative ˆ ± contain the positive/negative
nature of the homogeneous equations, it also implies that the = Jacobian of F. The matrices ⌳
flow variables are discontinuous at the cell boundaries. This eigenvalues of Â, respectively, 兩⌳ˆ 兩=⌳ˆ +−⌳ˆ −; the columns of R̂
may be exploited by treating the discontinuity as a linearized are the corresponding right eigenvectors, and V̂ represents the
characteristic variables. The specific form of these have been pre-
1
Research Scientist, Image Science and Applications Branch, Naval sented by Bradford and Katopodes 共2001兲 for the 2D shallow
Research Laboratory, Washington D.C. 20375 共corresponding author兲. water equations.
E-mail: [email protected] The discretization of the bed slope term requires special care
2
Associate Professor, Dept. of Civil and Environmental Engineering, to compute accurate solutions for problems with nonlevel beds.
Univ. of California, Irvine, CA 92697. Many researchers define bed elevation at cell boundaries 共in 1D兲
Note. Associate Editor: Michelle H. Teng. Discussion open until or cell vertices 共in 2D兲 and use a centered finite difference ap-
March 1, 2006. Separate discussions must be submitted for individual
proximation of the bed slope
papers. To extend the closing date by one month, a written request must

冕
be filed with the ASCE Managing Editor. The manuscript for this paper j+1/2
was submitted for review and possible publication on August 31, 2004; 1 dz ⌬z j z j+1/2 − z j−1/2
gd dx ⬇ gd j = gd j 共4兲
approved on December 7, 2004. This paper is part of the Journal of ⌬x j−1/2 dx ⌬x ⌬x
Engineering Mechanics, Vol. 131, No. 10, October 1, 2005. ©ASCE,
ISSN 0733-9399/2005/10-1073–1081/$25.00. which is accurate to O共⌬x2兲. However, using the cell average

JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005 / 1073

J. Eng. Mech., 2005, 131(10): 1073-1081

 flow variables to compute F yields O共⌬x兲 accuracy that when fied to account for the difference in bed elevation at the cell
combined with the bed slope discretization given by Eq. 共4兲 will boundaries. Leveque 共1998兲 describes a method to account for
initiate the flow of quiescent water over arbitrary topography. source terms in steady state problems with jumps on the interior
Bermudez and Vazquez 共1994兲 advocated upwinding the bed of cells, but states that the method is not suitable for transcritical
slope term in a manner consistent with the flux treatment, which flows.
effectively degrades the bed slope discretization to O共⌬x兲 accu- The purpose of this study is to systematically compare the
racy and balances the O共⌬x兲 accurate flux discretization. performance of finite volume schemes that adopt these variable
High-resolution schemes are obtained by extending the reconstruction, time-stepping, and source term discretization tech-
O共⌬x , ⌬t兲 model to O共⌬x2 , ⌬t2兲 accuracy. The flux discretization niques to determine whether there exists an optimal approach to
is extended to O共⌬x2兲 accuracy by linearly reconstructing flow extend Roe’s finite volume method to O共⌬x2 , ⌬t2兲 accuracy and
variables in each cell 共Van Leer 1979兲 in a manner that prevents model shallow-water flow over uneven beds. The high-resolution
the development of numerical oscillations. One approach for ex- methodologies used in this study are described in the following
tending the model to O共⌬t2兲 accuracy, known as Hancock’s section, and then each is applied to a set of test problems for
Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

method 共Van Leer 1984兲, is to compute a predictor solution at performance evaluation.

the n + 1 / 2 time level that is then linearly reconstructed prior to
computing the fluxes. High-resolution models also require careful High-Resolution Methodologies
discretization of the bed slope term in order to yield accurate
solutions. Reconstruction
Unfortunately, there does not appear to be a consensus in the
literature about constructing high-resolution schemes. Bradford O共⌬x2兲 accuracy in the flux discretization is obtained by linearly
and Katopodes 共2001兲 and Bradford and Sanders 共2002兲 advo- reconstructing cell-average flow variables prior to solving the
cated using the primitive form of the shallow water equations Riemann problem. For example, at the j + 1 / 2 face the recon-
to compute the predictor and reconstructing the primitive vari- structed depth is
ables ␨ = d + z and u. Conversely, Alcrudo and Garcia-Navarro dLj+1/2 = dn+1/2
1
+ 2 ⌬d j
j
共1993兲 and Nujić 共1995兲 proposed solving the conservative
共5兲
form of the governing equations and reconstructing the con- 1
servative variables ␨ and q. Rider 共1993兲 investigated this topic j+1 − 2 ⌬d j+1
dRj+1/2 = dn+1/2
for the homogeneous, compressible Euler equations, but his where
findings are not applicable to the nonhomogeneous shallow
⌬d j = avg共␦ j−1/2,␦ j+1/2兲 共6兲
water equations. There have also been several different methods
proposed for discretizing the bed slope term. Upwinding of where ␦ j−1/2 = dnj − dnj−1
and ␦ j+1/2 = dnj+1 − dnj .
The avg function may
the bed slope term has been proposed by Bermudez and Vasquez be selected to preserve solution monotonicity by preventing spu-
共1994兲 and extended to O共⌬x2兲 accuracy by Hubbard and Garcia- rious numerical oscillations near flow discontinuities. Such an
Navarro 共2000兲, which effectively pushes the bed slope dis- average is termed a slope limiter and accomplishes this by revert-
cretization to the cell boundaries. Gallouet et al. 共2003兲 developed ing to O共⌬x兲 accuracy at data extrema, i.e., when ␦ j−1/2␦ j+1/2 艋 0
a similar method in which the bathymetry is assumed to be and limiting the magnitude of gradients used in the reconstruc-
piecewise constant and the linearized Riemann problem is modi- tion. A popular family of such schemes is given as

⌬d j = 再 sign共␦ j−1/2兲min关min共兩␦ j−1/2兩,兩␦ j+1/2兩兲,␤ min共兩␦ j−1/2兩,兩␦ j+1/2兩兲兴 ␦ j−1/2␦ j+1/2 ⬎ 0

0 otherwise
冎 共7兲

where 1 艋 ␤ 艋 2. Selecting ␤ = 1 yields the Minmod average that is the most dissipative O共⌬x2兲 limiter. Further increasing ␤ yields a more
aggressive limiter that allows steeper gradients for an increasingly larger range of nonsmooth data, i.e., when ␦ j−1/2 ⫽ ␦ j+1/2. Such methods
actually steepen gradients beyond the O共⌬x2兲 centered finite difference gradient approximation, ⌬d j = 共d j+1 − d j−1兲 / 2, and hence add
antidiffusion to the solution. The choice ␤ = 2 results in the Superbee method, which is the most aggressive O共⌬x2兲 accurate limiter.
Another possible choice is the Double Minmod average

⌬d j = 再 sign共␦ j−1/2兲min共 2 兩␦ j−1/2 + ␦ j+1/2兩,2兩␦ j−1/2兩,2兩␦ j+1/2兩兲 ␦ j−1/2␦ j+1/2 ⬎ 0

0 otherwise
1

冎 共8兲

This limiter yields the Fromm scheme, which uses the O共⌬x2兲 centered difference gradient approximation for relatively smooth data and
limits the gradient to preserve monotonicity when ␦ j+1/2 ⬍ 3␦ j−1/2 or ␦ j+1/2 ⬎ ␦ j−1/2 / 3. Therefore this method minimizes the truncation error
except at sharp fronts, but does not introduce antidiffusion to the solution. The specific choice of a limiter is investigated in this study.

1074 / JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005

J. Eng. Mech., 2005, 131(10): 1073-1081

 Predictor ⑀c = 共û⌬d + d̂⌬u兲 j−1/2 共14兲
Before computing F, a predictor solution at the n + 1 / 2 time level
Note that c = shallow water wave celerity. The corresponding
must first be computed to obtain O共⌬t2兲 model accuracy. A com-
steady momentum equation in which the bed slope is centrally
putationally efficient approach is to use the primitive form of the
discretized with Eq. 共4兲 is
governing equations to compute the predictor as
共uq兲Lj+1/2 − 共uq兲Rj−1/2 + gdn+1/2
j 共␨Lj+1/2 − ␨Rj−1/2兲 + ⑀m = 0 共15兲
⌬t
dn+1/2
j = dnj − 共u⌬d + d⌬u兲nj 共9兲 where ⑀m represents the net perturbation of momentum flux due to
2⌬x
waves entering the cell; and for subcritical flow it is

un+1/2
j = unj −
⌬t
2⌬x
共u⌬u + g⌬␨兲nj 共10兲 ⑀m =
1
2
冋 冉 d̂
共û − ĉ兲2j+1/2 ⌬d − ⌬u
ĉ
冊
冊 册
j+1/2

It is natural to select ␨ and u for reconstruction and this approach

冉
Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

d̂
is termed the primitive predictor/reconstruction model 共Bradford + 共û + ĉ兲2j−1/2 ⌬d + ⌬u 共16兲
ĉ j−1/2
and Katopodes 2001; Bradford and Sanders 2002兲.
Another viable option is to compute the predictor in the same while for supercritical flow
manner as the corrector, i.e., solve the conservative form of the
equations ⑀m = 关共û2 + ĉ2兲⌬d + 2d̂û⌬u兴 j−1/2 共17兲
For quiescent water, u = q = 0, ␨ is constant, and Eqs. 共12兲 and 共15兲
Un+1/2
j − Unj 1 n
+ 共F − Fnj−1/2兲 = Snj 共11兲 reduce to
⌬t/2 ⌬x j+1/2
ĉ j−1/2⌬d j−1/2 − ĉ j+1/2⌬d j+1/2 = 0
and use Eq. 共3兲 to compute the fluxes. This approach, combined
with the reconstruction of ␨ and q, is termed the conservative 共18兲
predictor/reconstruction model and was proposed by Alcrudo and ĉ2j−1/2⌬d j−1/2 + ĉ2j+1/2⌬d j+1/2 = 0
Garcia-Navarro 共1993兲. Although this approach provides a con- Eqs. 共18兲 are satisfied if ⌬d j−1/2 = ⌬d j+1/2 = 0, i.e., d should be re-
servative predictor, it also adds to the computational burden of the constructed such that it is continuous across cell boundaries. Such
model since two Riemann problems must be solved at each cell a smooth reconstruction of d eliminates the perturbations of in-
face per time step as opposed to only one with the primitive coming waves and therefore preserves the quiescent state. Unfor-
predictor model. A more computationally efficient alternative tunately, reconstructing d directly from cell average values will
could be obtained by computing Fn by using the arithmetic aver- not guarantee a continuous reconstruction at the cell boundaries if
age of reconstructed U to the left and right of the cell boundaries. the bed slope varies from cell to cell. Nujić 共1995兲 recommended
This approach requires the solution of a single Riemann problem reconstructing ␨ and then computing ⌬d j = ⌬␨ j − ⌬z j. Since ␨ is
at each face per time step just like the primitive predictor model. constant and therefore ⌬␨ j = 0 everywhere, reconstructing d in this
It was found in simulations not presented here that this approach manner will guarantee that Eqs. 共18兲 are satisfied. This approach
yields nearly identical results as the Alcrudo and Garcia-Navarro has been adopted by Bradford and Katopodes 共2001兲; Bradford
共1993兲 model. and Sanders 共2002兲; and Sanders et al. 共2003兲 and has been
termed the surface-gradient 共SG兲 method by Zhou et al. 共2001兲.
Bed Slope Discretization While the SG method preserves the quiescent water state, it
does not necessarily work well for flow over a nonlevel bed. For
For the O共⌬x兲 accurate Roe’s method, an upwind bed slope dis- example, q and energy per unit mass E = u2 / 2 + g␨ are constant in
cretization as proposed by Bermudez and Vasquez 共1994兲 is nec- steady, frictionless, continuous flow over a nonlevel bed. Eq. 共12兲
essary for obtaining an accurate solution as discussed in the “In- shows that q will not converge to a constant value, i.e., qLj+1/2
troduction.” However, this is not true when Roe’s method is = qRj−1/2, unless ⑀c = 0. For a constant q, Eq. 共15兲 may be written as
extended to O共⌬x2兲 accuracy as will be shown in the following
analysis. Consider the following steady continuity equation for ELj+1/2 − ERj−1/2 + ⑀m = 0 共19兲
cell j:
which shows that E will not become constant unless q is constant
qLj+1/2 − qRj−1/2 + ⑀c = 0 共12兲 and ⑀m = 0. The simultaneous satisfaction of ⑀c = ⑀m = 0 requires
⌬d j±1/2 = ⌬u j±1/2 = 0, or in other words d and u must be recon-
where qRj−1/2and qLj+1/2
represent the linearly reconstructed dis- structed such that they are continuous at cell boundaries. In uni-
charge on the interior side of the left and right cell faces, respec- form flow, high-resolution models do in fact converge to the cor-
tively, and ⑀c represents the net perturbation of continuity flux rect solution with ⌬d j±1/2 = ⌬u j±1/2 = 0 for all j. However, a
共discharge兲 due to waves entering the cell at j + 1 / 2 and j − 1 / 2. nonuniform bed slope introduces nonuniformity in d and u and
For subcritical flow, the net perturbation is their piecewise linear reconstruction cannot guarantee ⌬d j±1/2

冋
= ⌬u j±1/2 = 0. This results in nonzero ⑀c and ⑀m, which are balanced

⑀c =
1
2
冉
d̂
共û − ĉ兲 j+1/2 ⌬d − ⌬u
ĉ
冊 by steady-state errors in q and E as shown by Eqs. 共12兲 and 共15兲.
The O共⌬x2兲 accurate upwind method proposed by Hubbard

冊 册
j+1/2
and Garcia-Navarro 共2000兲 is an alternative to the SG approach.

冉 d̂
+ 共û + ĉ兲 j−1/2 ⌬d + ⌬u
ĉ j−1/2
共13兲
Whereas the SG method seeks to minimize ⑀c and ⑀m by re-
constructing d in as smooth a manner as possible, the upwind
method seeks to counterbalance ⑀c and ⑀m by shifting the bed
and for supercritical flow slope discretization to cell boundaries. In the upwind method, d is

JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005 / 1075

J. Eng. Mech., 2005, 131(10): 1073-1081

 冢 冣
directly reconstructed rather than ␨ in the SG method and the bed 0
slope term is discretized at cell boundaries in a manner analogous ⌬z j±1/2
S j±1/2 = 共21兲
to F as − ĉ2j±1/2
⌬x
1
S j = 2 共S j+1/2 − 兩S兩 j+1/2 + S j−1/2 + 兩S兩 j−1/2兲 + S̃ j 共20兲
where S j±1/2 are computed as and 兩S兩 j±1/2 are computed as

兩S兩 j±1/2 =
2⌬x
冉
ĉ j±1/2⌬z j±1/2 sign共û − ĉ兲 j±1/2 − sign共û + ĉ兲 j±1/2
共û − ĉ兲 j±1/2 sign共û − ĉ兲 j±1/2 − 共û + ĉ兲 j±1/2 sign共û + ĉ兲 j±1/2
冊 共22兲
Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

The last term on the right side of Eq. 共20兲 is a correction term g⌬z
needed by the O共⌬x2兲 accurate method to preserve the quiescent u− = ū + − ⌬c 共27兲
2共ū − c̄兲
state and is defined as
where c̄ and ū = arithmetic averages of the L and R states. For

冢 冣
0 supercritical flow with u ⬎ 0, c− = cL, and u− = uL while for u ⬍ 0
c− = cR and u− = uR. To the right of the cell boundary, in either flow
S̃ j = ⌬z j 共23兲
− gd j regime, c and u are
⌬x
gc̄⌬z
Note that z must also be reconstructed in a manner that is consis- c+ = c− + 共28兲
2共ū2 − c̄2兲
tent with the reconstruction of d. For subcritical flow S becomes

冉 冊
gū⌬z
1 共ĉ⌬z兲 j+1/2 − 共ĉ⌬z兲 j−1/2 u+ = u− − 共29兲
Sj = ū2 − c̄2
2⌬x 共û − ĉ兲 j+1/2共ĉ⌬z兲 j+1/2 − 共û + ĉ兲 j−1/2共ĉ⌬z兲 j−1/2 − 2gd j⌬z j
The values of c± and u± are then converted to d± and q±, which are
共24兲 used to compute the fluxes. The corrector is then computed as
and for supercritical flow it is Un+1
j − Unj 1 −
+ 共F − F+j−1/2兲n+1/2 = 0 共30兲
⌬t ⌬x j+1/2
Sj =
1
冉 0
⌬x − 共ĉ ⌬z兲 j−1/2 − gd j⌬z j
2 冊 共25兲 Note that in this form, q is not continuous at the cell boundaries
and so Gallouet et al. 共2003兲 advocated that q+ and q− be arith-
metically averaged to compute the continuity fluxes. Although
For quiescent water, upwinding S introduces errors into the con- this method also preserves quiescent flow, the implicit assump-
tinuity and momentum equations that exactly balance ⑀c and ⑀m tions at the contact discontinuity are not valid in the presence of a
since in that instance ⌬z j±1/2 = −⌬d j±1/2 at the cell boundaries. hydraulic jump and therefore it is unclear how the model will
However, for supercritical flow, upwinding S does not reduce ⑀c perform in such a case.
and it is not obvious if ⑀m is reduced. Likewise, for subcritical
flow, it is not clear whether upwinding S yields more accurate
solutions.
Model Tests and Results
A third alternative is the method developed by Gallouet et al.
共2003兲, which assumes that the bathymetry is piecewise constant The methods described in the previous section are now tested
and the linearized Riemann problem is modified to account for using a set of problems involving both 1D and 2D flow scenarios.
the difference in bed elevation at the cell boundaries. A fixed Exact solutions are used for verification purposes when possible,
共zero speed兲 contact discontinuity is introduced at the cell bound- otherwise predictions made at a much finer resolution are used.
aries in addition to the usual u ± c waves and across the contact Comparisons of model predictions with experimental data are
discontinuity it is assumed that q and E are constant. Like bed specifically avoided because this would measure the model’s abil-
slope upwinding, this method pushes the discretization of the bed ity to simulate flow physics as well as its numerical accuracy. It
slope term to the cell boundaries and because of the assumption would be impossible to separate the two errors and this paper only
that z is piecewise constant, there is no bed slope contribution
within the cell. However, this also means that the bed slope dis-
cretization is only accurate to O共⌬x兲. Gallouet et al. 共2003兲 pre- Table 1. Summary of Models Tested in This Study
ferred to reconstruct ␨ and q and to solve the Riemann problem in
Abbreviation Method
terms of u and c. For subcritical flow the values to the left of a
cell boundary are P-SG Primitive predictor, surface gradient method
C-SG Conservative predictor, surface gradient method
g⌬z ⌬u C-UP Conservative predictor, upwind bed slope method
c− = c̄ − − 共26兲 C-MR Conservative predictor, modified Roe method
4共ū − c̄兲 4

1076 / JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005

J. Eng. Mech., 2005, 131(10): 1073-1081

 Fig. 1. Sketch of channel used in one-dimensional subcritical steady
flow test 共not to scale兲. Units are in meters.

concerns numerical accuracy. Methods tested here can be catego-

rized by two factors: the computation of the predictor/choice of
Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

reconstruction variables, and the bed slope discretization. The Fig. 3. Comparison of L2共E兲 as function of ⌬x for subcritical flow
primitive predictor/reconstruction SG method 共Bradford and
Katopodes 2001; Bradford and Sanders 2002兲 is denoted by

冑兺
P-SG, while the conservative predictor/reconstruction SG method
N
共Alcrudo and Garcia-Navarro 1993兲 is denoted by C-SG. The 1
upwind 共Hubbard and Garcia-Navarro 2000兲 and modified Rie- L2共q兲 = ⑀2共q j兲 共31兲
N j=1
mann 共Gallouet et al. 2003兲 bed slope discretizations are coupled
with a conservative predictor/reconstruction and are labeled C-UP where
and C-MR, respectively. These models are summarized in
Table 1. All models employ the Minmod limiter, which is the q j − qanalytical
⑀共q j兲 = 共32兲
most widely used limiter in practice. Some experiments with the qanalytical
Double Minmod limiter are also performed. The Superbee limiter
was found to yield oscillatory 共but stable兲 solutions in several and N = total number of computational cells. Fig. 2 reveals that the
cases, especially when a conservative predictor or bed slope up- C-UP model yielded approximately a 1 order of magnitude less
winding were employed, and is therefore not considered. error than the two SG methods, but the C-MR method yielded
2–3 orders of magnitude less error than the C-UP model. The SG
methods performed similarly and had the smoothest reduction in
Steady One-Dimensional Flow error as ⌬x was reduced. Fig. 3 compares the corresponding
The four models are now used to simulate steady, 1D flow over a L2共E兲, which reveals that the C-MR model again yielded the least
step to compare their errors and computational effort. Steady flow error by a great margin. In terms of this measure, the SG methods
simulations provide a valuable test because regardless of the performance is more variable with the C-SG method performing
bathymetry, q must be constant throughout the domain. In the better than the P-SG method.
absence of hydraulic jumps, E must also be constant and therefore Fig. 4 compares ⑀共q j兲 as a function of x surrounding the step
model accuracy may be easily verified. The first test problem for all of the methods with ⌬x = 0.125 m and ⌬t = 0.01875 s. The
involves subcritical, steady flow over a 0.1 m high step in a largest errors are yielded by the SG methods and located at the
100 m long channel as illustrated in Fig. 1. q is fixed to 1 m2 / s at upstream slope transitions where q is underpredicted and at the
the upstream end and the downstream depth is fixed to 1 m. Four downstream slope transitions where q is overpredicted. The C-UP
different grid spacings/time steps are used with the initial ⌬x method yielded less error and the C-MR method yielded a negli-
= 1 m and ⌬t = 0.15 s. The grid spacing and time step are succes- gible error compared to the other methods. Both SG models over-
sively halved to obtain the remaining grids. predicted E at the upstream slope transitions and underpredicted
Fig. 2 compares the L2 errors in q as a function of ⌬x, which E on the downstream slope transitions as shown in Fig. 5. In
are defined as addition, all of the models except the C-MR model underpre-

Fig. 4. Comparison of ⑀共q兲 as function of x for subcritical flow with

Fig. 2. Comparison of L2共q兲 as function of ⌬x for subcritical flow ⌬x = 0.125 m

JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005 / 1077

J. Eng. Mech., 2005, 131(10): 1073-1081

Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 7. Comparison of L2共E兲 as function of x for supercritical flow

Fig. 5. Comparison of ⑀共E兲 as function of x for subcritical flow with
⌬x = 0.125 m

dicted E everywhere downstream of the first upstream bed slope that did not decrease as the grid was refined. Fig. 9 compares the
change with the largest errors given by the P-SG model. corresponding ⑀共E j兲 and unlike the subcritical case, E is under-
For a given ⌬x, all methods converged at nearly the same rate predicted at the upslope and overpredicted at the downslope by
to a steady state in terms of q. In addition, the use of the Double the SG methods. In addition, the SG models underpredicted E
Minmod limiter instead of the Minmod limiter yielded no signifi- everywhere downstream of the first upstream bed slope change.
cant difference with the SG or upwind methods in terms of either However, the C-MR method drastically underpredicted E every-
error or rate of convergence. where downstream of the first upstream bed slope transition.
For supercritical flow, the upstream depth and velocity are Fig. 10 compares the convergence rate of the methods and
fixed to 0.1 m and 2.5 m / s, respectively. The step height is arbi- shows that the C-UP method converged the fastest, while the
trarily reduced to 0.05 m, but is otherwise unchanged. The same C-MR model had difficulty converging. The use of the Double
four grid spacings are used in this case, but the largest ⌬t Minmod limiter had no significant impact on the convergence rate
= 0.25 s and was successively halved for the finer grids. However, of the P-SG model, but it hindered the convergence of the C-SG
the C-MR method required an even smaller time step to remain and C-UP models.
stable and in this case ⌬t = 0.075 s for the coarsest grid and was A third test consists of transcritical flow over a step with a
successively halved for finer grids. Fig. 6 compares L2共q兲 for the stationary hydraulic jump. In this case, the step height is arbi-
models as a function of ⌬x and once again, the C-UP model trarily changed to 0.065 m, but is otherwise unchanged. At the
yielded less error than the SG models. The error of the SG models upstream end of the channel, q = 0.1 m2 / s is specified, while at
varied widely with the C-SG model converging to the analytical the downstream end h = 0.17 m is specified. The same ⌬x and ⌬t
solution much more rapidly than the P-SG method as the grid was used in the supercritical flow test are used in this test as well. For
refined. The C-MR yielded the worst error by a great margin. Fig. the sake of brevity, only errors in q are considered and in this case
7 compares L2共E兲 and in terms of this norm, the C-UP model the C-MR model would not converge and is therefore not consid-
again performed much better than the SG models while the C-MR ered. Fig. 11 compares L2共q兲 for the remaining methods and for
model performed very poorly and did not improve as the grid was the coarsest grid the C-UP scheme yielded the least error. How-
refined. ever, as the grid was refined, the error at the hydraulic jump
Fig. 8 compares ⑀共q j兲 for all of the methods and unlike the dominated over the error at the changes in bed slope and the SG
subcritical case q is overpredicted at the upslope and underpre- methods converged as rapidly to the analytical solution as the
dicted at the downslope by the C-MR and SG models. The C-MR
model yielded a large local error at the upstream slope transition

Fig. 8. Comparison of ⑀共q兲 as function of x for supercritical flow

Fig. 6. Comparison of L2共q兲 as function of x for supercritical flow with ⌬x = 0.125 m

1078 / JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005

J. Eng. Mech., 2005, 131(10): 1073-1081

Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 9. Comparison of ⑀共E兲 as function of x for supercritical flow Fig. 11. Comparison of L2共q兲 as function of ⌬x for transcritical flow
with ⌬x = 0.125 m

C-UP method. Fig. 12 illustrates ⑀共q j兲 for ⌬x = 0.125 m and re- ods performed similarly while the C-UP method performed
veals that the largest errors are at the jump location and are asso- slightly worse. Fig. 15 compares L2共u兲 for the same three meth-
ciated with the C-UP method. ods and follows the same trends as L2共␨兲. The use of the Double
Fig. 13 compares the convergence to steady state of the three Minmod limiter yielded slightly less error than the Minmod lim-
methods. The SG methods converged at roughly the same rate iter for all three methods with the improvement mainly at the
regardless of the limiter choice. The C-UP method converged sharp discontinuities.
more slowly and its convergence was slightly hindered by the use
of the Double Minmod limiter.
Steady Two-Dimensional Flow

Unsteady One-Dimensional Flow Two-dimensional steady flow simulations with the P-SG, C-SG,
and C-UP methods are now performed in a curving channel
An analysis using an unsteady flow problem is difficult to per- shown in Fig. 16. The discretization is also shown in Fig. 16
form because analytical solutions are unavailable for problems along with the bathymetry for the subcritical flow case in which
with arbitrary bathymetry. Therefore, the converged numerical so- the flow rate Q is specified as 1 m3 / s at the upstream boundary
lution is used as a measure of each model’s accuracy. A dambreak and d = 1 m is specified at the downstream boundary. Fig. 17 com-
problem in the same channel as before is used as the test. In this pares ⑀共Q j兲 as a function of the channel center line distance and
case the step is 0.05 m high and the dam is initially located 40 m reveals that the C-SG method yielded the largest errors at the
downstream from the left channel end. Both channel ends are upstream and downstream faces of the step. The P-SG and C-UP
treated as solid walls and the initial depth upstream of the dam is methods have errors that are smaller and comparable in magni-
1 m and downstream it is 0.1 m. The solution is computed at t tude, but of opposite sign.
= 15 initially with ⌬x = 1 m and ⌬t = 0.15 s. Both ⌬x and ⌬t are For supercritical flow, u = 3 m / s and d = 0.1 m are specified at
halved for three additional resolutions. the upstream boundary and the step height is reduced to 0.05 m.
Fig. 14 compares L2共␨兲 for t = 15 s, which is computed using Fig. 18 compares ⑀共Q j兲 for the three methods and in this case the
the converged numerical solution in place of an analytical solu- C-UP method yielded significantly less error than the SG methods
tion, for the P-SG, C-SG, and C-UP methods. Note that the SG while the C-SG method once again yielded the most error.
and C-UP methods converged to identical solutions within ma-
chine precision. This figure shows that the P-SG and C-SG meth-

Fig. 12. Comparison of ⑀共q兲 as function of x for transcritical flow

Fig. 10. Convergence history for supercritical flow with ⌬x = 0.125 m

JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005 / 1079

J. Eng. Mech., 2005, 131(10): 1073-1081

Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 13. Convergence history for transcritical flow Fig. 15. Comparison of L2共u兲 as function of ⌬x for unsteady
dambreak flow

Summary and Conclusions between continuous and discontinuous flow make it difficult, if
not impossible, to design a high-resolution model that is optimal
High-resolution finite volume methods typically reconstruct cell in both cases.
average data at cell boundaries in order to compute boundary The most appropriate high-resolution model for general appli-
fluxes by solving a Riemann problem. This approach yields a cations can be identified based on the results of this study. Firstly,
model with excellent mass conservation and phase accuracy that despite the stellar performance of the C-MR method for steady,
simulates wave propagation very well. For uniform flow, high- subcritical flow, its poor performance for supercritical and discon-
resolution schemes converge rapidly to analytical solutions in tinuous flow makes it unsuitable for general applications. Second,
both the subcritical and supercritical flow regimes. However, the C-SG method was equally or less accurate than the P-SG
when the bottom slope varies from cell to cell, the flow becomes method in all cases except steady, 1D, supercritical flow. How-
nonuniform and the reconstruction of data at cell boundaries be- ever, the SG methods’ errors were within 1 order of magnitude of
comes discontinuous. It was shown that such discontinuities in- each other for all test cases. The use of a conservative predictor
troduce errors into the cell average values of q and E even though by the C-SG and C-UP methods required approximately 30%
the model computes correct flux values at the cell boundaries. more CPU time than the P-SG method for 1D simulations and
Therefore, for continuous nonuniform flow, it is desirable to re- approximately 60% more for the 2D problem. The additional
construct data at the cell boundaries in the smoothest possible computational expense and the marginal differences in accuracy
manner to minimize discontinuities and corresponding error. This make it difficult to recommend the C-SG model over the P-SG
is contrary to accurately simulating discontinuous flow such as a model. However, the additional computational expense of the
steady hydraulic jump over a flat bed. In this case, the reconstruc- C-SG method may be eliminated if the predictor fluxes are com-
tion should capture the sharp discontinuities in d, ␨, and u, which puted without solving a Riemann problem as previously dis-
unfortunately introduces a fixed error in the continuous variable cussed. In this case, either the P-SG or C-SG methods are recom-
q. This is illustrated in Fig. 11 in which L2共q兲 initially increased mended. Third, the C-UP model performed the best in terms of
as the grid was refined and the hydraulic jump better resolved.
L2共q兲 ultimately decreased with further grid refinement, but the
error at the jump remained fairly constant. The conflicting needs

Fig. 14. Comparison of L2共␨兲 as function of ⌬x for unsteady Fig. 16. Channel geometry and discretization for two-dimensional
dambreak flow steady flow simulations

1080 / JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005

J. Eng. Mech., 2005, 131(10): 1073-1081

 well. Finally, the Double Minmod limiter hindered convergence
to a steady state when coupled with the use of a conservative
predictor. This limiter exacerbated reflections of waves leaving
the domain, which caused oscillatory 共but stable兲 solutions for
several test cases. Therefore, it is recommended that the Double
Minmod limiter not be coupled with conservative predictor
schemes such as the C-SG and C-UP models. For the P-SG
model, the Double Minmod limiter should be used if sharply re-
solving discontinuities is the most important priority.

Acknowledgment
Downloaded from ascelibrary.org by University of Sheffield on 03/27/24. Copyright ASCE. For personal use only; all rights reserved.

B.F.S. was supported on this project by the National Science

Fig. 17. Comparison of ⑀共Q兲 as function of channel centerline Foundation 共Grant No. CMS-9984579兲.
distance for steady, two-dimensional, subcritical flow

accuracy for subcritical and supercritical steady, 1D continuous References

flow. However, for discontinuous and unsteady flow in 1D, this
method yielded more error than either of the SG methods. In 2D, Alcrudo, F., and Garcia-Navarro, P. 共1993兲. “A high-resolution Godunov-
the C-UP method yielded errors similar to the P-SG method for type scheme in finite volumes for the 2D shallow-water equations.”
Int. J. Numer. Methods Fluids, 16, 489–505.
subcritical flow, but was slightly more accurate than both SG
Bermudez, A., and Vazquez, M. E. 共1994兲. “Upwind methods for hyper-
methods for supercritical flow. The loss of its accuracy advantage bolic conservation laws with source terms.” Comput. Fluids, 23,
in unsteady and 2D subcritical flow, as well as its computational 1049–1071.
premium, reduces the appeal of the C-UP method. Therefore, in Bradford, S. F., and Katopodes, N. D. 共2001兲. “Finite volume model for
terms of accuracy for a wide variety of problems and computa- nonlevel basin irrigation.” Irrig. Drain., 127共4兲, 216–223.
tional effort, the P-SG method is the most attractive of the high- Bradford, S. F., and Sanders, B. F. 共2002兲. “Finite-volume model for
resolution models. shallow-water flooding of arbitrary topography.” J. Hydraul. Eng.,
The specific choice of limiter 共Minmod or Double Minmod兲 128共3兲, 289–298.
made little difference in the steady flow simulations and both Engquist, B., and Osher, S. 共1981兲. “One-sided difference approximations
require roughly the same computational effort for a given grid. for nonlinear conservation laws.” Math. Comput., 36, 321–352.
Gallouet, T., Herard, J., and Seguin, N. 共2003兲. “Some approximate Go-
The Double Minmod limiter was slightly more accurate for the
dunov schemes to compute shallow-water equations with topogra-
unsteady flow test, but for the grid resolutions considered here phy.” Comput. Fluids, 32, 479–513.
there was less than a 1 order of magnitude reduction in the L2共d兲 Hubbard, M. E., and Garcia-Navarro, P. 共2000兲. “Flux difference splitting
and L2共u兲 errors. The Double Minmod limiter yields sharper reso- and the balancing of source terms and flux gradients.” J. Comput.
lution of discontinuities and allows for the use of a coarser grid Phys., 165, 89–125.
for a given error in d and u than the Minmod limiter, which leads Leveque, R. J. 共1998兲. “Balancing source terms and flux gradients in
to greater computational efficiency. However, as noted above, high-resolution Godunov methods: The quasi-steady wave-
sharper resolution of discontinuities results in greater error in q as propagation algorithm.” J. Comput. Phys., 146, 346–365.
Nujić, M. 共1995兲. “Efficient implementation of non-oscillatory schemes
for the computation of free-surface flows.” J. Hydraul. Res., 33, 101–
111.
Rider, W. J. 共1993兲. “Methods for extending high-resolution schemes to
non-linear systems of hyperbolic conservation laws.” Int. J. Numer.
Methods Fluids, 17, 861–885.
Roe, P. L. 共1981兲. “Approximate Riemann solvers, parameter vectors, and
difference schemes.” J. Comput. Phys., 43, 357–372.
Sanders, B. F., Jaffe, D. A., and Chu, A. K. 共2003兲. “Discretization of
integral equations describing flow in nonprismatic channels with un-
even beds.” J. Hydraul. Eng., 129共3兲, 235–244.
Van Leer, B. 共1979兲. “Towards the ultimate conservative difference
scheme. V. A second order sequel to Godunov’s method.” J. Comput.
Phys., 32, 101–136.
Van Leer, B. 共1984兲. “On the relation between the upwind-differencing
schemes of Godunov, Engquist–Osher, and Roe.” SIAM (Soc. Ind.
Appl. Math.) J. Sci. Stat. Comput., 5, 1–20.
Zhou, J. G., Causon, D. M., Mingham, C. G., and Ingram, D. M. 共2001兲.
Fig. 18. Comparison of ⑀共Q兲 as function of channel centerline
“The surface-gradient method for the treatment of source terms in the
distance for steady, two-dimensional, supercritical flow shallow-water equations.” J. Comput. Phys., 168, 1–25.

JOURNAL OF ENGINEERING MECHANICS © ASCE / OCTOBER 2005 / 1081

J. Eng. Mech., 2005, 131(10): 1073-1081