Approximate solvers for the shallow water equations

In this chapter we discuss approximate Riemann solvers for the shallow water equations:
\begin{align} h_t + (hu)_x & = 0 \label{SWA_mass} \\ (hu)_t + \left(hu^2 + \frac{1}{2}gh^2\right)_x & = 0. \label{SWA_mom} \end{align}

To examine the Python code for this chapter, and for the exact Riemann solution, see:

We investigated the exact solution of the Riemann problem for this system in Shallow_water.html. Recall that the system has two genuinely nonlinear characteristic fields, so the solution consists of two waves, each of which may be a shock or a rarefaction.

Recall that the flux Jacobian is
\begin{align} f'(q) & = \begin{pmatrix} 0 & 1 \\ -(q_2/q_1)^2 + g q_1 & 2 q_2/q_1 \end{pmatrix} = \begin{pmatrix} 0 & 1 \\ -u^2 + g h & 2 u \end{pmatrix}. \end{align}
Its eigenvalues are
\begin{align} \label{SWA:char-speeds} \lambda_1 & = u - \sqrt{gh} & \lambda_2 & = u + \sqrt{gh}, \end{align}
with corresponding eigenvectors
\begin{align} \label{SWA:fjac-evecs} r_1 & = \begin{bmatrix} 1 \\ u-\sqrt{gh} \end{bmatrix} & r_2 & = \begin{bmatrix} 1 \\ u+\sqrt{gh} \end{bmatrix}. \end{align}

Roe solver

To define a linearized Riemann solver, we consider the quasilinear form $q_t + f'(q)q_x=0$, and replace the Jacobian $f'(q)$ by some matrix $\hat{A}(q_\ell,q_r)$ depending on the left and right states. This matrix should satisfy three properties:

  1. Consistency: $\hat{A}(q_\ell,q_r) \to f'(q)$ as $q_\ell, q_r \to q$.
  2. Hyperbolicity: $\hat{A}$ must be diagonalizable with real eigenvalues, so that we can define the waves and speeds needed in the approximate solution.
  3. Conservation: $\hat{A}(q_\ell,q_r)(q_r - q_\ell) = f(q_r) - f(q_\ell)$.

The first two properties are essential for any linearized approximate Riemann solver. As mentioned in Burgers_approximate.html, the third property is the defining property of a Roe solver. It ensures that if $q_r$ and $q_\ell$ can be connected by a single shock (i.e., if they lie on a common Hugoniot locus), then the approximate solver will in fact yield the exact solution -- a single discontinuity propagating at the speed given by the Rankine-Hugoniot jump condition. This property is quite useful, since at most points (in space and time) where a shock appears, it is an isolated shock corresponding to one characteristic family. Only at isolated points in space and time do multiple shocks interact.

Here we derive the Roe solver for the shallow water equations in a simple and straightforward way by applying the conservation requirement. Roe's original paper (Roe, 1981) (which dealt with the Euler equations rather than the shallow water equations) gave a different and possibly more general derivation that is also reviewed in Chapter 15 of (LeVeque, 2002).

One natural approach is to look for a matrix $\hat{A}$ that is the flux Jacobian evaluated at some average state:
$$\hat{A}(q_\ell,q_r) = f'(\hat{q}).$$
This guarantees property 2, and will ensure property 1 as long as $\hat{q}$ is chosen as some kind of average of $q_\ell$ and $q_r$. For the shallow water system, this means that the linearized equations take the form
\begin{align} \label{SWA_quasilinear} h_t + (hu)_x & = 0 \\ (hu)_t + (g\hat{h}-\hat{u}^2)h_x + 2\hat{u}(hu)_x & = 0. \nonumber \end{align}
Observe that (\ref{SWAquasilinear}) is linear not in the physical variables $h, u$, but in the conserved variables $h, hu$. In the context of a Riemann solver, $\hat{q}$ will be some kind of average of the initial states $q\ell, q_r$.

The characteristic speeds of the linear hyperbolic system (\ref{SWA_quasilinear}) are just the characteristic speeds of the original shallow water system, evaluated at the average state $\hat{q}$:
\begin{align} \label{SWA:avg-char-speeds} s_1 & = \hat{u} - \sqrt{g\hat{h}} & s_2 & = \hat{u} + \sqrt{g\hat{h}}, \end{align}
with corresponding eigenvectors
\begin{align} \label{SWA:avg-fjac-evecs} \hat{r}_1 = r_1(\hat{q}) & = \begin{bmatrix} 1 \\ \hat{u}-\sqrt{g\hat{h}} \end{bmatrix} & \hat{r}_2 = r_2(\hat{q}) & = \begin{bmatrix} 1 \\ \hat{u}+\sqrt{g\hat{h}} \end{bmatrix}. \end{align}

The linearized solver simply decomposes the jump $q_r-q_\ell$ in terms of the eigenvectors, to define the waves:
\begin{align} \label{SWA:wave-decomp} q_r - q_\ell = \alpha_1 \hat{r}_1 + \alpha_2 \hat{r}_2 = {\mathcal W}_1 + {\mathcal W}_2 \end{align}
and uses the speeds given by (\ref{SWA:avg-char-speeds}). We refer to $\alpha_1, \alpha_2$ as the wave strengths; they can be found by solving the system above, which amounts to
\begin{align} \alpha = \hat{R}^{-1}(q_r-q_\ell), \end{align}
where $\hat{R} = (\hat{r}_1 | \hat{r}_2)$ is the matrix of right eigenvectors, and thus $\hat{R}^{-1}$ is the matrix whose rows are the left eigenvectors. This leads to the expressions
\begin{align} \alpha_1 & = \frac{(\hat{u}+ \hat{c}) \delta_1 - \delta_2}{2\hat{c}} \\ \alpha_2 & = \frac{-(\hat{u} - \hat{c}) \delta_1 + \delta_2}{2\hat{c}}, \end{align}
where $\hat{c} = \sqrt{g\hat{h}}$ and $\delta = q_r -q_\ell$.

Derivation of Roe averages

It remains only to specify the average state $(\hat{h},\hat{u})$. We do so by solving the conservation condition:
\begin{align} \label{SWA:roe_cons} f'(\hat{q}) (q_r - q_\ell) & = f(q_r) - f(q_\ell). \end{align}
For convenience we recall here the relevant quantities:
\begin{align} f'(\hat{q}) & = \begin{pmatrix} 0 & 1 \\ -\hat{u}^2 + g \hat{h} & 2 \hat{u} \end{pmatrix}, \end{align} \begin{align*} q_r - q_\ell & = \begin{pmatrix} h_r - h_\ell \\ h_r u_r - h_\ell u_\ell \end{pmatrix}, & f(q_r) - f(q_\ell) & = \begin{pmatrix} h_r u_r - h_\ell u_\ell \\ h_r u_r^2 - h_\ell u_\ell^2 + \frac{g}{2}(h_r^2-h_\ell^2)\end{pmatrix}. \end{align*} The first equation of (\ref{SWA:roe_cons}) is an identity, satisfied no matter how $\hat{u}$ and $\hat{h}$ are chosen. We therefore turn to the second equation:
\begin{align} \label{SWA:2nd} (-\hat{u}^2 + g\hat{h})(h_r-h_\ell) + 2\hat{u}(h_r u_r - h_\ell u_\ell) & = h_r u_r^2 - h_\ell u_\ell^2 + \frac{g}{2}(h_r^2-h_\ell^2). \end{align} Since (\ref{SWA:2nd}) must hold for any value of $g$, we can equate separately the terms involving $g$ and those not involving $g$. For the terms with $g$ in them, we get:
\begin{align} \label{SWA:h_roe1} \hat{h}(h_r-h_\ell) & = \frac{1}{2} (h_r^2 - h_\ell^2), \end{align}
which gives the arithmetic average
\begin{align} \label{SWA:h_roe} \hat{h} = \frac{h_r + h_\ell}{2}. \end{align} Equating terms that do not involve $g$, we get a quadratic equation for $\hat{u}$:
$$(h_r-h_\ell)\hat{u}^2 - 2(h_r u_r - h_\ell u_\ell) \hat{u} + h_r u_r^2 - h_\ell u_\ell^2 = 0.$$
The two roots of this equation are
$$\hat{u}_\pm = \frac{h_r u_r - h_\ell u_\ell \mp \sqrt{h_r h_\ell} (u_\ell-u_r)}{h_r-h_\ell}.$$
Simplifying these expressions further, we find
\begin{align} \label{SWA:u_roe} \hat{u}_\pm = \frac{\sqrt{h_r}u_r \pm \sqrt{h_\ell}u_\ell}{\sqrt{h_r} \pm \sqrt{h_\ell}}, \end{align}
where the same sign (plus or minus) must be taken in both numerator and denominator. Either of these expressions can be used for $\hat{u}$, as long as $h_r \ne h_\ell$. The value $\hat{u}_+$ is preferable for two reasons: first, it is well defined even when $h_r = h_\ell$; second, it has the appearance of a weighted average of $u$, and its value always lies between $u_\ell$ and $u_r$.

The values given by (\ref{SWA:h_roe}) and (\ref{SWA:u_roe}) (with the plus sign) are known as the Roe-averaged depth and velocity. This completes the definition of the approximate Roe solver. An implementation appears below.

In [1]:
%matplotlib inline
In [2]:
%config InlineBackend.figure_format = 'svg'

import numpy as np
import matplotlib.pyplot as plt
from exact_solvers import shallow_water as sw
from utils import riemann_tools as rt
from utils.snapshot_widgets import interact
from ipywidgets import widgets
Will create static figures with single value of parameters
In [3]:
def shallow_water_roe(q_l, q_r, grav=1.):
    """
    Approximate Roe solver for the shallow water equations.
    """
    
    h_l = q_l[0]
    hu_l = q_l[1]
    u_l = hu_l/h_l
    h_r = q_r[0]
    hu_r = q_r[1]
    u_r = hu_r/h_r
    
    delta = q_r - q_l
    
    # Roe averages
    h_hat = (h_r + h_l)/2.
    u_hat = (np.sqrt(h_r)*u_r + np.sqrt(h_l)*u_l) \
                / (np.sqrt(h_r) + np.sqrt(h_l))
    c_hat = np.sqrt(grav*h_hat)
    
    s1 = u_hat - c_hat
    s2 = u_hat + c_hat
    
    alpha1 = ( (u_hat+c_hat)*delta[0] - delta[1])/(2*c_hat)
    alpha2 = (-(u_hat-c_hat)*delta[0] + delta[1])/(2*c_hat)
    
    h_m = q_l[0] + alpha1
    hu_m = q_l[1] + alpha1 * (u_hat - c_hat)
    q_m = np.array([h_m, hu_m])
    
    states = np.column_stack([q_l,q_m,q_r])
    speeds = [s1, s2]
    wave_types = ['contact','contact']
    
    def reval(xi):
        h_out  = (xi<s1)*h_l + (s1<=xi)*(xi<=s2)*h_m + (s2<xi)*h_r
        hu_out = (xi<s1)*hu_l + (s1<=xi)*(xi<=s2)*hu_m + (s2<xi)*hu_r
        return h_out, hu_out
    
    return states, speeds, reval, wave_types

Here is an example showing the solution from the Roe solver, compared with the exact Riemann solution.

In [4]:
# Dam-break example
h_l = 4; h_r = 1
u_l = 0; u_r = 0

def make_plot_functions(h_l, u_l, h_r, u_r):
    hu_l = h_l*u_l; hu_r = h_r*u_r
    q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])
    ex_states, ex_speeds, ex_reval, ex_wave_types = \
                                    sw.exact_riemann_solution(q_l,q_r)
    states, speeds, reval, wave_types = shallow_water_roe(q_l, q_r)

    pr = rt.make_plot_function([ex_states, states],
                               [ex_speeds, speeds],
                               [ex_reval, reval],
                               [ex_wave_types, wave_types],
                               variable_names=sw.conserved_variables)
    
    pp = sw.phase_plane_plot(q_l, q_r, g=1., y_axis='u', 
                             approx_states=states)

    return pp, pr

phase_plane, comparison = make_plot_functions(h_l, u_l, h_r, u_r)
interact(comparison, t=widgets.FloatSlider(min=0,max=.9,step=.1,value=0.4));

Here and throughout this section, we plot the approximate solution in green. As usual, Hugoniot loci and shock paths are plotted in red, while integral curves and rarefaction fans are plotted in blue.

Notice that the rarefaction wave is approximated by a discontinuity.

The next example illustrates the exactness property for isolated shocks. Here we have chosen $q_\ell$ and $q_r$ such that they are related by a single 2-shock (up to rounding errors).

In [5]:
h_l = 2.20698770767; h_r = 1.
u_l = 2.27057814896/h_l; hu_r = 0.

phase_plane, comparison = make_plot_functions(h_l, u_l, h_r, u_r)
interact(comparison, t=widgets.FloatSlider(min=0,max=.9,step=.1,value=0.4));

Because the exact and approximate solutions are the same, the exact solution is not visible in these plots. The only exception is the 1-wave in the first plot; the speed of the 1-wave appears different for the Roe solution and the exact solution, but this is unimportant since the 1-wave has zero strength.

Entropy fix

As discussed previously, the use of an entropy-violating discontinuity in place of a rarefaction can be problematic in the case of a transonic rarefaction. For the Roe solver, we can again apply an entropy fix, splitting the single discontinuity into two waves, to avoid this problem. Several different entropy fixes have been proposed; here we focus on one that is quite general and effective in practice.

Suppose that $\lambda_1(q_\ell) < 0$ while $\lambda_1(q_m)>0$. In this case, it appears that the 1-wave should be a transonic rarefaction. We take the 1-wave ${\mathcal W}_1$ of the Roe solver, with speed $s_1$, and split it into two waves, ${\mathcal W}_{1\ell} = \alpha_{1\ell} r_1$ and ${\mathcal W}_{1r} = \alpha_{1r} r_1$, with speeds $\lambda_1(q_\ell), \lambda_1(q_m)$. Two requirements determine the strength of these waves. First, the total jump across them should be equal to the jump across the original wave:

$${\mathcal W}_{1\ell} + {\mathcal W}_{1r} = {\mathcal W}_1.$$

Second, we must maintain conservation:

$$\lambda_1(q_\ell) {\mathcal W}_{1\ell} + \lambda_1(q_m) {\mathcal W}_{1r} = s_1 {\mathcal W}_1.$$

This leads to

\begin{align*} {\mathcal W}_{1\ell} & = \frac{\lambda_1(q_m)-s_1}{\lambda_1(q_m)-\lambda_1(q_\ell)}{\mathcal W}_1 \\ {\mathcal W}_{1r} & = \frac{s_1 - \lambda_1(q_\ell)}{\lambda_1(q_m)-\lambda_1(q_\ell)}{\mathcal W}_1. \end{align*}

We can check in a similar way for a transonic 2-rarefaction, and modify the 2-wave accordingly. Note that only one of the two characteristic fields can be transonic in the Riemann solution for the shallow water equations.

This entropy fix is implemented in the solver below.

In [6]:
def shallow_water_roe_with_efix(q_l, q_r, grav=1., efix=True):
    """
    Approximate Roe solver for the shallow water equations,
    with entropy fix.
    """
    
    h_l = q_l[0]
    hu_l = q_l[1]
    u_l = hu_l/h_l
    h_r = q_r[0]
    hu_r = q_r[1]
    u_r = hu_r/h_r
    
    delta = q_r - q_l
    
    # Roe averages
    h_hat = (h_r + h_l)/2.
    u_hat = (np.sqrt(h_r)*u_r + np.sqrt(h_l)*u_l) /  \
            (np.sqrt(h_r) + np.sqrt(h_l))
    c_hat = np.sqrt(grav*h_hat)
    
    s1 = u_hat - c_hat
    s2 = u_hat + c_hat
    
    alpha1 = ( (u_hat+c_hat)*delta[0] - delta[1])/(2*c_hat)
    #alpha2 = (-(u_hat-c_hat)*delta[0] + delta[1])/(2*c_hat)
    
    h_m = q_l[0] + alpha1
    hu_m = q_l[1] + alpha1 * (u_hat - c_hat)
    u_m = hu_m/h_m
    q_m = np.array([h_m, hu_m])
        
    transonic = None
    if efix:
        # Check if either wave appears transonic
        lambda_1_l = u_l - np.sqrt(grav*h_l)
        lambda_1_m = u_m - np.sqrt(grav*h_m)
        lambda_2_m = u_m + np.sqrt(grav*h_m)
        lambda_2_r = u_r + np.sqrt(grav*h_r)
        
        if lambda_1_l < 0 < lambda_1_m: # 1-wave appears transonic
            transonic = 1
        elif lambda_2_m < 0 < lambda_2_r: # 2-wave appears transonic
            transonic = 2
    
    if transonic == 1:
        beta = (lambda_1_m - s1)/(lambda_1_m - lambda_1_l)
        alpha1_l = beta*alpha1
        #alpha_1_r = (1-beta)*alpha1
        h_m_new = q_l[0] + alpha1_l
        hu_m_new = q_l[1] + alpha1_l * (u_hat-c_hat)
        q_m_new = np.array([h_m_new, hu_m_new])
        states = np.column_stack([q_l,q_m_new,q_m,q_r])
        speeds = [lambda_1_l, lambda_1_m, s2]
        wave_types = ['contact']*3
        
        def reval(xi):
            h_out  = (xi<speeds[0])*states[0,0] + \
                     (speeds[0]<=xi)*(xi<speeds[1])*states[0,1] + \
                     (speeds[1]<=xi)*(xi<speeds[2])*states[0,2] + \
                     (speeds[2]<=xi)*states[0,3]
            hu_out = (xi<speeds[0])*states[1,0] + \
                     (speeds[0]<=xi)*(xi<speeds[1])*states[1,1] + \
                     (speeds[1]<=xi)*(xi<speeds[2])*states[1,2] + \
                     (speeds[2]<=xi)*states[1,3]
            return h_out, hu_out
        
    elif transonic == 2:
        beta = (lambda_2_r - s2)/(lambda_2_r - lambda_2_m)
        #alpha2_l = beta*alpha2
        alpha2_r = (1-beta)*alpha2
        h_m_new = q_r[0] - alpha2_r
        hu_m_new = q_r[1] - alpha2_r*(u_hat+c_hat)
        q_m_new = np.array([h_m_new, hu_m_new])
        states = np.column_stack([q_l,q_m,q_m_new,q_r])
        speeds = [s1,lambda_2_, lambda_2_r]
        wave_types = ['contact']*3

        def reval(xi):
            h_out  = (xi<speeds[0])*states[0,0] + \
                     (speeds[0]<=xi)*(xi<speeds[1])*states[0,1] + \
                     (speeds[1]<=xi)*(xi<speeds[2])*states[0,2] + \
                     (speeds[2]<=xi)*states[0,3]
            hu_out = (xi<speeds[0])*states[1,0] + \
                     (speeds[0]<=xi)*(xi<speeds[1])*states[1,1] + \
                     (speeds[1]<=xi)*(xi<speeds[2])*states[1,2] + \
                     (speeds[2]<=xi)*states[1,3]
            return h_out, hu_out
        
    else:    
        states = np.column_stack([q_l,q_m,q_r])
        speeds = [s1, s2]
        wave_types = ['contact','contact']
    
        def reval(xi):
            h  = (xi<s1)*h_l + (s1<=xi)*(xi<=s2)*h_m + (s2<xi)*h_r
            hu = (xi<s1)*hu_l + (s1<=xi)*(xi<=s2)*hu_m + (s2<xi)*hu_r
            return h, hu
    
    return states, speeds, reval, wave_types

An implementation of this solver for use in Clawpack can be found here.

Roe solver examples

Here are some examples of problems with transonic rarefactions, showing the behavior of the solver. The first example involves two rarefaction waves; the left-going one is transonic.

In [7]:
h_l = 1; h_r = 1
u_l = 0.5; u_r = 2

def compare_efix(h_l,u_l,h_r,u_r):
    hu_l = h_l*u_l; hu_r = h_r*u_r
    q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])
    ex_states, ex_speeds, ex_reval, ex_wave_types = \
                        sw.exact_riemann_solution(q_l,q_r)
    states, speeds, reval, wave_types = \
                        shallow_water_roe_with_efix(q_l, q_r)
    no_efix_states, no_efix_speeds, no_efix_reval, no_efix_wave_types = \
                        shallow_water_roe(q_l, q_r)

    plot_function = \
        rt.make_plot_function([ex_states, no_efix_states, states],
                              [ex_speeds, no_efix_speeds, speeds],
                              [ex_reval, no_efix_reval, reval],
                              [ex_wave_types, no_efix_wave_types, wave_types],
                              ['Exact','Roe w/o fix','Roe with fix'],
                              layout='vertical',
                              variable_names=sw.conserved_variables)
    interact(plot_function, t=widgets.FloatSlider(min=0,max=.9,step=.1,value=0.4));
    return q_l, q_r, states, no_efix_states

q_l, q_r, states, no_efix_states = compare_efix(h_l,u_l,h_r,u_r)
In [8]:
fig, ax = plt.subplots(1,2,figsize=(10,4))
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=no_efix_states,ax=ax[0])
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=states,ax=ax[1],color='orange')
ax[0].set_title('Without entropy fix')
ax[1].set_title('With entropy fix');

The next example is a dam-break problem; the rarefaction wave is transonic because initially the water is all moving to the right.

In [9]:
h_l = 4; h_r = 1
u_l = 1; u_r = 1

q_l, q_r, states, no_efix_states = compare_efix(h_l,u_l,h_r,u_r)
In [10]:
fig, ax = plt.subplots(1,2,figsize=(10,4))
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=no_efix_states,ax=ax[0])
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=states,ax=ax[1],color='orange')
ax[0].set_title('Without entropy fix')
ax[1].set_title('With entropy fix');

Positivity

In the shallow water system, it is both physically and mathematically essential that the depth remain non-negative. Unfortunately, as the next example shows, the Roe solver does not always guarantee this.

In [11]:
h_l = 1; h_r = 1
u_l = -1.5; u_r = 1.5

hu_l = h_l*u_l; hu_r = h_r*u_r
q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])
ex_states, ex_speeds, ex_reval, ex_wave_types = \
                sw.exact_riemann_solution(q_l,q_r)
states, speeds, reval, wave_types = \
                shallow_water_roe_with_efix(q_l, q_r)

plot_function = \
    rt.make_plot_function([ex_states, states],
                          [ex_speeds, speeds],
                          [ex_reval, reval],
                          [ex_wave_types, wave_types],
                          variable_names=sw.conserved_variables)
interact(plot_function, t=widgets.FloatSlider(min=0,max=.9,step=.1,value=0.4));
In [12]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=states, hmin = -1)
plt.plot([0,0], [-2,2], 'r--')
plt.text(-0.8,1.5,'Unphysical states',color='red')
Out[12]:
Text(-0.8,1.5,'Unphysical states')

In fact, this is a feature of all linearized Riemann solvers: for some Riemann problems, they will produce negative intermediate states. The HLL solver described next does guarantee a positive solution for all initial data.

HLL solver

Derivation and implementation

Recall that an HLL approximate solver is based on a fictitious solution consisting of two waves (discontinuities) with a single intermediate state $q_m$ between them. The waves are thus
\begin{align} {\mathcal W}_1 & = q_m - q_\ell \\ {\mathcal W}_2 & = q_r - q_m. \end{align}
The middle state is determined by the requirement of conservation:
\begin{align} s_1 {\mathcal W}_1 + s_2 {\mathcal W}_2 & = f(q_r) - f(q_\ell), \end{align}
which can be solved for $q_m$ to find
\begin{align} \label{SWA:hll_middle_state} q_m = \frac{f(q_r) - f(q_\ell) - s_2 q_r + s_1 q_\ell}{s_1 - s_2}. \end{align}

The key remaining ingredient is the choice of the wave speeds $s_1, s_2$. These should approximate the wave speeds appearing in the true Riemann solution. In general, it is desirable for the purpose of stability that the speeds used be approximations of the fastest (left- and right-going) waves in the Riemann solution. Indeed, in the original paper of Harten, Lax, and van Leer (Harten, Lax, and van Leer) it was proposed to use lower and upper bounds on the true wave speeds. However, overestimating the wave speeds leads to increased diffusion, and various refinements to the choice of wave speeds were subsequently proposed. Here we choose the wave speeds based on a suggestion of Einfeldt (originally in the context of the Euler equations); the resulting solver is often referred to as an HLLE solver.

For the shallow water system, the 1-wave speed is always less than or equal to the 2-wave speed. If the 1-wave is a rarefaction, then the minimum wave velocity is given simply by $\lambda_1(q_\ell)$, the characteristic speed at the left edge of the 1-rarefaction fan. If the left-going wave speed is a shock, then the minimum wave velocity is the shock speed. We cannot determine the shock speed exactly without solving the Riemann problem -- just what we hope to avoid. Instead, we can use an approximate shock speed based on the Roe average:

$$\hat{s}_1 = \hat{u} - \hat{c}.$$

Since we don't know whether the 1-wave should be a shock or a rarefaction, we use a lower bound on these two speeds, by taking their minimum:

$$s_1 = \min\left(\lambda_1(q_\ell), \hat{u}-\hat{c}\right).$$

Similarly, the second wave speed is given by

$$s_2 = \max\left(\lambda_2(q_r), \hat{u}+\hat{c}\right).$$

This completes the definition of the solver. Notice that if the states $q_\ell, q_r$ are connected by a single 2-shock, then the value of $s_2$ will be the exact shock speed (since the Roe speed is exact in this case, and the shock speed will be faster than the characteristic speed ahead of it). Furthermore, by substituting the Rankine-Hugoniot conditions into (\ref{SWA:hll_middle_state}), one sees that the the middle state in this case will be $q_m = q_\ell$. Thus the HLLE solver gives the exact solution for this case. Similarly, it yields the exact solution when the initial states are connected by just a 1-shock. It shares this nice property with the Roe solver (as long as the wave speeds are chosen as just described).

In [13]:
def shallow_water_hll(q_l, q_r, grav=1.):
    """
    HLLE approximate solver for the shallow water equations.
    """
    h_l = q_l[0]
    hu_l = q_l[1]
    u_l = hu_l/h_l
    h_r = q_r[0]
    hu_r = q_r[1]
    u_r = hu_r/h_r
        
    # Roe averages
    h_hat = (h_r + h_l)/2.
    u_hat = (np.sqrt(h_r)*u_r + np.sqrt(h_l)*u_l) / \
            (np.sqrt(h_r) + np.sqrt(h_l))
    c_hat = np.sqrt(grav*h_hat)

    lambda_1_l = u_l - np.sqrt(grav*h_l)
    lambda_2_r = u_r + np.sqrt(grav*h_r)
    
    s1 = min(lambda_1_l, u_hat - c_hat)
    s2 = max(lambda_2_r, u_hat + c_hat)
    
    h_m = (hu_r - hu_l - s2*h_r + s1*h_l)/(s1-s2)
    hu_m = (hu_r*u_r - hu_l*u_l + 0.5*grav*(h_r**2 - h_l**2) \
            - s2*hu_r + s1*hu_l)/(s1-s2)
    q_m = np.array([h_m, hu_m])
    
    states = np.column_stack([q_l,q_m,q_r])
    speeds = [s1, s2]
    wave_types = ['contact','contact']
    
    def reval(xi):
        h_out  = (xi<s1)*h_l + (s1<=xi)*(xi<=s2)*h_m + (s2<xi)*h_r
        hu_out = (xi<s1)*hu_l + (s1<=xi)*(xi<=s2)*hu_m + (s2<xi)*hu_r
        return h_out, hu_out
    
    return states, speeds, reval, wave_types

An implementation of this solver for use in Clawpack can be found here.

Examples using the HLL solver

In [14]:
def compare_solutions(q_l, q_r, solvers):
    outputs = []
    states = {}

    for solver in solvers:
        if solver.lower() == 'exact':
            outputs.append(sw.exact_riemann_solution(q_l,q_r))
        if solver.lower() == 'hll':
            outputs.append(shallow_water_hll(q_l, q_r))
            states['hll'] = outputs[-1][0]
        if solver.lower() == 'roe':
            outputs.append(shallow_water_roe_with_efix(q_l, q_r))
            states['roe'] = outputs[-1][0]
        if solver.lower() == 'roe no efix':
            outputs.append(shallow_water_roe(q_l, q_r))
            states['roe no efix'] = outputs[-1][0]
            
    plot_function = \
        rt.make_plot_function([val[0] for val in outputs],
                              [val[1] for val in outputs],
                              [val[2] for val in outputs],
                              [val[3] for val in outputs],
                              solvers, layout='horizontal',
                              variable_names=sw.conserved_variables)
    
    interact(plot_function,
             t=widgets.FloatSlider(min=0,max=0.9,step=0.1,value=0.4));
    
    return states

Our first example is a standard dam-break problem.

In [15]:
# Dam-break example
h_l = 4; h_r = 1
u_l = 0; u_r = 0

hu_l = h_l*u_l; hu_r = h_r*u_r
q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])

states = compare_solutions(q_l, q_r, ['Exact','HLL'])
In [16]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',approx_states=states['hll'])

In the next example, the left and right states are constructed so as to be connected exactly by a single right-going shock. As we have discussed, the solver captures this solution exactly.

In [17]:
# Isolated 2-shock example
h_l = 2.20698770767; h_r = 1.
u_l = 2.27057814896/h_l; hu_r = 0.

hu_l = h_l*u_l; hu_r = h_r*u_r
q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])

states = compare_solutions(q_l, q_r, ['Exact','HLL'])
In [18]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=states['hll'])

The final example includes a transonic rarefaction.

In [19]:
# Transonic rarefaction example a
h_l = 1; h_r = 1
u_l = 0.5; u_r = 2

hu_l = h_l*u_l; hu_r = h_r*u_r
q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])

states = compare_solutions(q_l, q_r, ['Exact','HLL'])
In [20]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',approx_states=states['hll'])

Notice that, by design, the HLLE wave always lies exactly on the leading edge of a rarefaction wave, when compared to the true solution.

The momentum solution above has the interesting feature that the solution in the 1-rarefaction is not monotone but appears as a "bump". This can be explained by plotting the solution in the h-hu phase plane, where one sees that the integral curve has a local maximum between $q_\ell$ and $q_m$:

In [21]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='hu',approx_states=states['hll'])

Entropy and positivity

It can be shown that the HLLE solver always preserves positivity and does not require an entropy fix. Both of these properties result in part from the fact that the HLLE solver is more diffusive than the Roe solver.

In [22]:
# Transonic rarefaction example b
h_l = 4; h_r = 1
u_l = 4; u_r = 1

hu_l = h_l*u_l; hu_r = h_r*u_r
q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])

states = compare_solutions(q_l, q_r, ['Exact','HLL'])
In [23]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=states['hll'])

The final example below illustrates that the HLLE solver maintains positivity when the same initial data is used as in the example with the Roe solver above where a negative intermediate depth was obtained.

In [24]:
h_l = 1; h_r = 1
u_l = -1.5; u_r = 1.5

hu_l = h_l*u_l; hu_r = h_r*u_r
q_l = np.array([h_l, hu_l]); q_r = np.array([h_r, hu_r])

states = compare_solutions(q_l, q_r, ['Exact','HLL'])
In [25]:
sw.phase_plane_plot(q_l,q_r,g=1.,y_axis='u',
                    approx_states=states['hll'])

The above discussion concerns the nature of a single approximate Riemann solution. In a numerical method, the approximate solver is used repeatedly, so it is interesting to compare the performance of the HLLE and Roe Riemann solvers when they are used within a full discretization to solve the shallow water equations. We will do so in the FV_compare.