Ab FLM Revisedv3

Adaptive multi scale scheme based on numerical density of
entropy production for conservation laws.

Mehmet Ersoy∗1 , Frédéric Golay†1 , and Lyudmyla Yushchenko‡ 1
1
Université de Toulon, IMATH, EA 2134, 83957 La Garde Cedex, France.
January 8, 2013
Abstract
We propose a 1D adaptive numerical scheme for hyperbolic conservation laws based on the numerical
density of entropy production (the amount of violation of the theoretical entropy inequality). Thus it is
used as an a posteriori error which provides information on the need to refine the mesh in the regions
where discontinuities occur and to coarsen the mesh in the regions where the solution remains smooth.
Nevertheless, due to the CFL stability condition the time step is restricted and leads to time consuming
simulations. Therefore, we propose a local time stepping algorithm. We also use high order time exten-
sions using Adams-Bashforth time integration technique as well as a second order linear reconstruction
in space. We numerically investigate the efficiency of the scheme through several test cases: the Sod’s
shock tube problem, the Lax’s shock tube problem and the Shu-Osher test problem.
Keywords: Hyperbolic system, finite volume scheme, local mesh refinement, numerical density of entropy
production, local time stepping.
1 Introduction
In this paper, we are interested in numerical integration of one dimensional non linear hyperbolic systems of
conservation laws of the form
∂w ∂f (w)
(
+ = 0, (t, x) ∈ R+ × R (1)
∂t ∂x
w(0, x) = w0 (x), x ∈ R.
where w : R+ × R → Rd stands for the vector state and f : Rd → Rd the flux function.
Solving Equation (1) with high accuracy is a challenging problem since it is well-known that solutions
can and will breakdown at a finite time, even if the initial data are smooth, and develop complex structure
(shock wave interactions). In such a situation, the uniqueness of the (weak) solution is lost and is recovered
by completing the system (1) with an entropy inequality of the form:
∂s(w) ∂ψ(w)
+ ≤0, (2)
∂t ∂x
where (s, ψ) stands for a convex entropy-entropy flux pair. This inequality allows to select the physical
relevant solution. Moreover, the entropy satisfies a conservation equation only in regions where the solution
∗
[email protected]
†
[email protected]
‡
[email protected]
is smooth and an inequality when the solution develops shocks. In simple cases, it can be proved that the
term missing in (2) to make it an equality is a Dirac mass.
Numerical approximation of Equations (1) with (2) leads to the so-called numerical density of entropy
production which is a measure of the amount of violation of the entropy equation (as a measure of the
local residual as in [3, 12, 15, 14]). As a consequence, the numerical density of entropy production provides
information on the need to locally refine the mesh (e.g. if the solution develops discontinuities) or to coarsen
the mesh (e.g. if the solution is smooth and well-approximated) as already used by Puppo [18, 19, 20] and
Golay [9]. Even if the shocks are well-captured on coarse grid using finite volume scheme such indicator is
able (as shown in Puppo [20] and the reference therein) not only to provide an efficient a posteriori error,
but also to reproduce the qualitative structure of the solution and to pilot the adaptive scheme. Let us
also mention that this does not occur efficiently on contact discontinuities where only numerical diffusion
is active. Nonetheless, we mention the work by Guermond et al. [10] as a remarkable and simple way to
solve this problem by making the numerical diffusion to be proportional to the numerical density of entropy
production (i.e. adding a large numerical dissipation in the shock regions and almost no dissipation in the
regions where the solution remains smooth).
Explicit adaptive schemes are well-known to be time consuming due to a CFL stability condition. The
cpu-time increases rapidly as the mesh is refined since the CFL stability condition imposes an upper bound
δt
on where δt is the time step and h the finest mesh size. Nevertheless, the cpu-time can be significantly
h
reduced using the local time stepping algorithm (see e.g. [16, 25, 2]). In recent years, such numerical schemes
have been widely developed for equations which arise in many fluid flows: traffic flows, multi-phase flows,
multi-layer flows, etc. The local time stepping algorithm has been employed successfully for real two and
three dimensional steady and unsteady problems (see for instance [17, 24, 13, 26, 16, 20]).
The aim of this paper is to construct a 1D multi scale adaptive numerical scheme for hyperbolic conser-
vation laws. The paper is organized as follows. In Section 2, we recall the finite volume approximation, the
time integration methods and we fix notations. Thereafter, we explain how to adapt the space grid using the
numerical density of entropy production in Section 3. Combining both time integration and mesh refinement,
we show how to implement an efficient multi scale adaptive numerical scheme. The general algorithm and
two approaches to distribute the flux between two cells of different levels are compared in Section 4. Finally
in Section 5, the efficiency of the multi scale adaptive scheme is shown through Sod’s shock tube problem,
Lax’s shock tube problem and Shu-Osher test case.
2 Numerical approximation for hyperbolic conservation laws

2.1 One-dimensional finite volume formulation
We recall here the well-known construction of a numerical approximation of the following general non linear
hyperbolic conservation laws
∂w ∂f (w)
(
+ = 0, (t, x) ∈ R+ × Ω (3)
∂t ∂x
w(0, x) = w0 (x), x ∈ R.
where Ω ⊂ R, w ∈ Rd denotes the vector state (referred as the conservative variables, e.g. Equations
(14)–(17)) and f denotes the flux governing the physical description of the flow.
For the sake of simplicity,
we consider here
the one-dimensional case. The computational domain is split
into control volumes Ck = xk−1/2 , xk+1/2 of mesh size hk with xk±1/2 = xk ± hk /2. The unknowns w(t, x)
are approximated by their mean values on the cell Ck at time t:
1
Z
wk (t) ≃ w (t, x) dx .
hk C k
Integrating (3) over each cell and applying the Green’s formula we obtain:
∂wk
hk + Fk+1/2 (t) − Fk−1/2 (t) = 0 , (4)
∂t
2
where Fk±1/2 stands for the numerical flux at the interface xk±1/2 (see for instance [8, 27, 6]).
For the convenience of the reader, noting the flux difference
δFk (t) := Fk+1/2 (t) − Fk−1/2 (t)
between two interfaces of a cell Ck , Equation (4) is also written:
∂wk (t)
hk + δFk (t) = 0 . (5)
∂t
For practical purpose, in our case, F = f (R(0± , wL , wR )) is determined from the exact solution
R(0± , wL , wR ) of the Riemann problem associated with the left wL and the right wR state:
∂w ∂f (w)


 + =0,
∂t ∂x
(6)
wL if x < 0 ,
 w0 (x) =

wR if x > 0 .
Then, the semidiscrete first order numerical scheme reads (see for instance [8, 27, 6]):
∂wk (t) ∂wk (t)
+ f R(0− , wk (t), wk+1 (t)) − f R(0+ , wk−1 (t), wk (t)) .

hk + δFk (t) := hk
∂t ∂t
Equation (3) is completed with an entropy inequality of the form:
∂s(w) ∂ψ(w)
+ ≤0, (7)
∂t ∂x
where (s, ψ) stands for a convex entropy-entropy flux pair. This inequality not only allows to select the
physical relevant solution but will be used to define a relevant mesh refinement criterion (the so-called
numerical density of entropy production, see Section 3). To this end, Equation (7) is discretized using the
same scheme as Equation (3). Thus, the semi discrete formulation for the entropy inequality reads:
∂s(wk ) ∂s(wk )
+ ψ R(0− , wk (t), wk+1 (t)) − ψ R(0+ , wk−1 (t), wk (t))

hk + δψk (t) := hk (8)
∂t ∂t
where
δψk (t) := ψk+1/2 (t) − ψk−1/2 (t)
and R is the solution of the Riemann problem (6).
To also achieve a second order approximation in space, the MUSCL reconstruction can be used that we
now recall for completeness’ sake (see for instance [8, 27]). It consists in solving a Riemann problem (6) with
a linearly extrapolated initial data

Mk Mk+1
wk + , wk+1 −
2 2
where Mk is an approximation to the slope obtained with the MinMod limiter :

m min ∆wk−1/2 , ∆wk , ∆wk+1/2 if m = sgn ∆wk−1/2 = sgn (∆wk ) = sgn ∆wk+1/2
Mk =
0 otherwise
and
∆wk−1/2 + ∆wk+1/2
∆wk−1/2 := wk − wk−1 , ∆wk+1/2 := wk+1 − wk , ∆wk := .
2
Thus as already described before, we define the numerical flux as

− Mk Mk+1
Fk+1/2 := f R 0 , wk + , wk+1 − ,
2 2

+ Mk−1 Mk
Fk−1/2 := f R 0 , wk−1 + , wk − .
2 2
3
2.2 Time integration
Now we focus on the numerical time integration of Equations (5) and (8). The well-known time integration
method of Runge-Kutta and Adams-Bashforth are recalled to fix notations.
2.2.1 Runge Kutta schemes

Integrating Equation (5) and (8) during the time step ]tn , tn+1 [ of length δtn and evaluating the numerical
fluxes at time tn , we obtain the well-known Euler’s scheme:
δtn
wk (tn+1 ) = wk (tn ) − δFk (tn ) ,
hk (9)
δtn
s(wk (tn+1 )) = s(wk (tn )) − δψk (tn ) .
hk
In order to increase the accuracy, a second order method can be used as follows
δtn
wk (tn+1 ) = wk (tn ) − δFk (tn+1/2 ) ,
hk
where
δtn
wk (tn+1/2 ) = wk (tn ) − δFk (tn )
2hk
and the time discretization of the entropy is the same as the time discretization of the conservative variables.
Then, as in [20], we define the numerical density of entropy production Skn :
s(wk (tn+1 )) − s(wk (tn )) δψk (tn )

Skn := + (10)
δtn hk
for the first order scheme and
s(wk (tn+1 )) − s(wk (tn )) δψk (tn+1/2 )
Skn := +
δtn hk
for the second order scheme.
We note X
P= Skn δtn hk (11)
n,k
the total numerical entropy production.
2.2.2 Adams-Bashforth schemes

In order to avoid intermediate computation and to spare computing time, one can use multi time step method
such as the Adams-Bashforth method. The Adams-Bashforth method of order m consists in replacing the
numerical flux of Equation (5) by a polynomial interpolation of the same order [11]:
m−1
X
Fk+1/2 (t) ≈ Lj (t)Fk+1/2 (tn−j ) ,
j=0
where Lj (t) denotes the coefficients of the Lagrange polynomial. Let us note that this explicit approximation
is built with the fluxes previously computed and stored. If the polynomial coefficients are integrated (e.g.
by numerical Gauss integration):
Z tn+1
1
bj (δtn ) = Lj (t) dt ,
δtn tn
4
then the integration of equation (5) and (8) leads to:
m−1
X δtn
wk (tn+1 ) = wk (tn ) − bj (δtn ) δFk (tn−j ) .
j=0
hk
For example, the second order Adams-Bashforth method is:
δtn δt2n
wk (tn+1 ) = wk (tn ) − δFk (tn ) − (δFk (tn ) − δFk (tn−1 )) .
hk 2δtn−1 hk
The first order method is exactly as Equations (9). The Adams-Bashforth methods of order 2 and 3 are
stable for Courant-Friedrich-Levy condition (CFL) less than one [1]. Practically, for stability purpose, we
limit our applications to the second order scheme.
We also perform the same discretization above for Equation (8). As done in the case of the Runge-Kutta
scheme, we define the numerical density of entropy production Skn by (10) for the first order scheme and
s(wk (tn+1 )) − s(wk (tn )) δψk (tn ) δtn

Skn := + + (δψk (tn ) − δψk (tn−1 ))
δtn hk 2δtn−1 hk
for the second order scheme. Finally, the total numerical entropy production is given by (11).
3 Mesh refinement
In order to compute with accuracy the solution of hyperbolic systems, a strategy can be to adapt automat-
ically the mesh following a local error indicator (defined later on).
3.1 Mesh refinement criterion

Many works concentrate on a posteriori error estimate, which are constructed from mathematical arguments
[28]. But paradoxically, to our knowledge, very few works use a physical criterion as [5]: the numerical
density of entropy production. More recently, Puppo [18, 19] has proved that the numerical density of
entropy production can be used as a discontinuity indicator and a local error indicator. Indeed, the entropy
inequality being related mathematically and physically to the system (3), the numerical density of entropy
production should be (and actually is, as we will see in Section 5) an accurate and a useful mesh refinement
criterion. Let us show how to use it in an automatic mesh refinement framework.
The mesh refinement criterion is the numerical density of entropy production Skn (for instance, defined
by (10) for a first order scheme, see Section 2.2 for further details) obtained from the discretization of the
entropy inequality
∂s(w) ∂ψ(w)
+ ≤0,
∂t ∂x
where ψ is the entropy flux with
∇Tw ψ = ∇Tw s(w)f ′ (w) .
As pointed out before in [18, 19], the previous inequality is useful for numerical application both as a local
error indicator (whenever it fails to be zero for smooth flow) and discontinuity detector (whenever solution
develops shocks or contact discontinuities). Thus it will be used here as a mesh refinement criterion.
3.2 Mesh refinement process

We present here a local mesh refinement procedure driven by the numerical density of entropy production.
In order to reduce the time necessary to manage the refinement, we use “macro cells” which could be refined
by generating hierarchical grids. Each cell can be split in two sub-cells. We thus produce a dyadic cells
5
graph, whose numbering (in basis 2) allows a quick computing scan to determine the adjacent cells. For
the convenience of the reader, we make use of the following notations: let kb be the index which makes
reference to the macro cell numbered k and b a binary number which contains the hierarchical information
of a sub-cell. In particular, the level of a sub-cell Ckb is defined as the length(b) − 1. For instance, a macro
cell Ck0 of level 0 will be split into two sub-cells Ck00 and Ck01 of level 1. A mesh refinement example is
proposed in Figure 1.
Figure 1: Example of hierarchical dyadic tree.
The mesh refinement procedure can be simply expressed as follows. We first set a mesh refinement
parameter S̄. For instance, it can be the mean value over the domain Ω at time tn :
1 X n
S̄ = Skb . (12)
|Ω|
kb
We then define two coefficients 0 6 αmin 6 αmax 6 1, which determine the ratio of numerical production of
entropy leading to mesh refinement or mesh coarsening.
Thus, for each cell Ckb :
• if Sknb > S̄αmax , the mesh is refined and split into two sub-cells Ckb0 and Ckb1 ,
• if Sknb0 < S̄αmin and Sknb1 < S̄αmin , the mesh is coarsened into a cell Ckb .
On one hand, if a cell Ckb is split into two sub-cells Ckb0 and Ckb1 , averaged values at time tn are projected
on each sub-cell (for the first order scheme):
wknb0 = wknb1 = wknb
and the numerical fluxes between the two sub-cells (see Figure 2(a)) are defined as:
 n

 Fkb0 −1/2 = Fknb −1/2 ,



Fknb0 +1/2 = Fknb1 −1/2 = f (wknb ) ,



 n
Fkb1 +1/2 = Fknb +1/2 .

6
In the case of second order method, as the gradient is available, we also use:
hkb ∂wknb

n n
 wkb0 = wkb − 4 ∂x ,



∂wknb
 w = wn + hkb

 n
.

kb1 kb
4 ∂x
On the other hand, if two sub-cells Ckb0 and Ckb1 are coarsened, we initialize the new cell Ckb as (see
Figure 2(b)):
1
wknb = wknb0 + wknb1

2
and  n n
 Fkb −1/2 = Fkb0 −1/2 ,
Fknb +1/2 = Fknb1 +1/2 .


(a) Mesh refinement. (b) Mesh coarsening.
Figure 2: Mesh refinement and mesh coarsening: construction of the fluxes.
Remark 3.1 It is hard to define an optimal mesh refinement criterion since the less the threshold is small,
the more the mesh is refined unnecessarly and the more the threshold is large, the less the mesh is refined.
Therefore, a good compromise is to use the average value of the mesh refinement criterion (12). The others
threshold parameters αmin and αmax allow to set a percentage of mesh refinement and mesh coarsening with
respect to the quantity S̄. It is not surprising that these settings will deteriorate or improve the accuracy of
the numerical solution. For instance, the more αmin and αmax are small, the more accurate are the results
to the expense of the computational time.
4 The local time stepping method

Unlike the Runge-Kutta method, the Adams-Bashforth time integration can be easily implemented in the
local time stepping framework in order to reduce the computational cost. The main approach to update
averaged quantities will be recalled and the general algorithm will be provided.
4.1 The local time stepping algorithm

In this section, we combine the local time stepping algorithm (see for instance [2]) with the mesh refinement
process presented above. For the sake of clarity, we do not use the binary subscript notation introduced in
the previous section.
Let hmin be the minimum diameter of the mesh and hmax the maximum diameter of the mesh. All cells
of the mesh are sorted in groups, called ”level”, based on their diameter. Thus, the level of refinement Lk of
the k th cell Ck is defined by:
2N −Lk hmin 6 hk < 2N +1−Lk hmin ,
7
where N stands for the maximum level

hmax
N = log2 + 1.
hmin
The coarsest cells are therefore of level 1, while the finest cells are of level N .
Remark 4.1 For any mesh, the previous inequality holds and in the particular case of a dyadic mesh, we
have:
hk = 2N −Lk hmin .
Let δtn be the minimum time step at time tn according to the CFL condition associated with the smallest
cell. The macro time step ∆tn is then defined by:
∆tn = 2N −1 δtn .
We define the level of interface of two adjacent cells Lk+1/2 by:
Lk+1/2 = max (Lk , Lk+1 ) .
Assuming that the maximum level of refinement is N at the current time tn , following [2], the local time
stepping algorithm reads as follows:
foreach i ∈ {1, 2N −1 } do
Let j be the biggest integer such that 2j divides i
foreach interface xk+1/2 such that Lk+1/2 > N − j do
1. compute the integral of Fk+1/2 (t) on the time interval 2N −Lk+1/2 δtn ,
2. distribute Fk+1/2 (tn ) to the two adjacent cells,
3. update only the cells of level greater than N − j.
end
end
Remark 4.2 We can notice that according to this algorithm, the flow is computed more times in the small
cells than in the bigger ones.
Remark 4.3 The minimum time step is defined by the CFL condition associated with the smallest cell at
time tn , i.e
hmin
δtn <
max |ν|
where ν is an eigenvalue of the convection matrix Dw f (w).
Remark 4.4 This scheme uses the Osher and Sanders [17] projection. This point is developed in Section
4.2.
Let us illustrate for a global time advancement the local time stepping algorithm for a first order scheme
(the principle for a second order is essentially the same since we have to use the value of the flux stored in
the two previous local time steps). We consider a mesh composed of 4 cells of levels 1, 2 and 3 respectively,
n
i.e. here N = 3, as displayed on Figure 3. Let us note wm the state vector at time tn . At the first stage
(see Figure 3(a)), i = 1, only the small cells are advanced with a step δtn , i.e. w3n1 := w3 (tn + δtn ) and
w4n1 := w4 (tn + δtn ) are computed. While i = 2 (see Figure 3(b)) concerns the small cells of level LN and
the cells immediately larger at level LN −1 . The cells LN travel with a time step δtn , while the cells LN −1
travel with a time step 2δtn , etc.
8
(a) At time tn + δtn (b) At time tn + 2δtn .
(c) At time tn + 3δtn . (d) At time tn+1 := tn + 4δtn .
Figure 3: Illustration of a global time advancement with a first order scheme.
4.2 Computation of the flux

In this subsection we recall the procedure introduced by Osher and Sanders [17] and, Tang and Warnecke
[26] for the computation of the flux at the cell interface of two different levels. For the sake of simplicity, we
assume here only two time increments which means that we consider cells of only two different kinds, say
h(x) = hk0 = h0 for x < xk0 +1/2 and h(x) = hk0 +1 = h0 /2 for x > xk0 +1/2 . Thus, for every k 6 k0 , all cells
Ck are of level Lk0 and for every k > k0 , all cells Ck are of level Lk0 + 1 (as displayed on Figure 4). Let
∆tn := tn+1 − tn = 2δtn be the macro time step. To avoid boundary terms, we consider the example with
an initial condition with compact support.
9
(a) Osher and Sanders projection.
(b) Tang and Warnecke projection.
Figure 4: Projection’s methods.
4.2.1 A simple projection method

The Osher and Sanders scheme is defined as follows (see Figure 4(a)):

 wkn if k 6 k0 ,
n+1/2
∀k ∈ Z, wk = n ∆tn n
 wk − δF if k > k0
2hk k
and
∆tn n n+1/2

∀k ∈ Z, wkn+1 = wkn − δFk + δFk . (13)
2hk
It was pointed out by Tang and Warnecke [26] that the present scheme loses the conservative property and
it is locally non consistent in the sense of the truncation error. Indeed, at the intermediate time step, one
has
X n+1/2 X ∆tn
wk = wkn + Fn (wn , wn )
2hk0 +1 k0 +1/2 k0 k0 +1
k∈Z k∈Z
which shows that this scheme is locally non conservative. At the next intermediate step, the conservativity
is preserved, i.e. X X n+1/2
wkn+1 = wk
k∈Z k∈Z
but the global conservativity is lost.

The fact that the scheme is not consistent in the sense of truncation error can be observed in the simplest
case of the one-dimensional transport equation (as done in [26]):
∂t w + c∂x w = 0
10
with a positive constant c. In that case, using the Godunov solver, the scheme (13) writes for k = k0 :
wkn+1
0 +1
− wkn0 +1 c n
n+1/2

− wk0 +1 − wkn0 + wk0 +1 − wkn0 =0.
∆tn 2hk0
Assuming w smooth enough, writing the Taylor expansion of w at point xk0 +1 at time tn+1 and tn+1/2 , we
deduce the following equation
c ∆t
∂t w + c∂x w = − ∂t w + O(h, ∆t).
4 h
∆t
Thus, if converge to a constant d > 0 then the previous equation does not converge to the initial transport
h
equation.
4.2.2 A natural projection method

Contrary to the Osher and Sanders [17] approach, Tang and Warnecke [26] propose the following scheme:
n+1/2 ∆tn n
∀k ∈ Z, wk = wkn − δF
2hk k
and
δFkn if k 6 k0

n+1/2 ∆tn
∀k ∈ Z, wkn+1 = wk − n+1/2 .
2hk δFk if k > k0
One can easily check that the global conservativity is also lost and
X X ∆tn n+1/2
wkn+1 = wkn + F (wn , wn ) .
2hk0 +1 k0 +1/2 k0 k0 +1
k∈Z k∈Z
Although the scheme is not conservative, they claim that it is consistent in the sense of the truncation error.
Even if the Osher and Sanders projection leads to non conservative and non consistent scheme at the
space-time grid between two levels of refinement, it has the advantage to be less time consuming than the
second one. Let us also note that the apparent loss of consistency on the local truncation error seems not
to affect the actual error of the scheme (as we will see in Section 5.1.1 and see also for instance [4, 22]
for interesting issues). On the contrary, Puppo and Semplice [20] do not loose the consistency and the
conservativity because their procedure does not produce mass and the fluxes are well computed at the cell
interface between two levels of refinement. Nevertheless, their approach requires more evaluations of the flux
than the Osher and Sanders projection since they use updated neighbourhood state (non projected as done
in the Osher and Sanders approach) .
5 Numerical experiments
We now present some results using the adaptive multi scale scheme constructed in Section 3 and Section 4.
The discussion is limited on studying the robustness, accuracy and the gain of cpu-time of several schemes.
To this end, we compare results to those from the exact solution or a reference solution. These include the
Sod’s shock tube problem, the Lax’s shock tube problem, and the Shu-Osher test case.
For all test cases, numerical solutions are computed using the one-dimensional gas dynamics equations
for ideal gas:
∂ρ ∂ρu
+ =0 (14)
∂t ∂x
2
∂ρu ∂ ρu + p
+ =0 (15)
∂t ∂x
∂ρE ∂ (ρE + p) u
+ =0 (16)
∂t ∂x
p = (γ − 1)ρε (17)
11
where ρ, u, p, γ, E are respectively the density, the velocity, the pressure, the ratio of the specific heats (set
u2
to 1.4) and the total energy E = ε + (where ε is the internal specific energy). Using the conservative
2
T
variables w = (ρ, ρu, ρE) , we classically define the entropy by

p
s(w) = −ρ ln
ργ
and the entropy flux by

ψ(w) = u s(w)
for Equations (14)–(17).
Notations and settings

• In what follows, we perform several numerical tests using first and second order schemes. Thus, we
will refer to AB1 as the first order scheme, AB2 as the second order Adams-Basforth scheme, RK2 as
the second order Runge-Kutta scheme. AB2 and RK2 use a MUSCL reconstruction. Moreover, all
computations are made with a dynamic grid except if the acronym ends with the capital letter “U”
which refers to a uniform fixed grid. All computations are made with an uniform time step except if
the acronym ends with the capital letter “M” which refers to the local time stepping algorithm.
• We will also compare the adaptive numerical solution to the one computed on a uniform fixed grid.
To have a coherent support of comparison, the solution on the fixed grid will be computed with NLmax
cells. NLmax stands for the average number of cells used during a simulation of an adaptive scheme
with a maximum level Lmax .
• All of the presented results display the density which is positive. Thus, for the convenience of the
reader, in all figures the numerical density of entropy production is plotted with the reversed sign to
overlap with the density.
• For all numerical tests, the initial mesh is adapted to the initial data during few iterations (except
for the numerical solution computed on uniform fixed grid) and we have used the following threshold
parameters:
Mesh refinement parameter αmax : 0.01 ,
Mesh coarsening parameter αmin : 0.001 ,
1 X n
Mesh refinement parameter S̄ : Skb .
|Ω|
kb
• ρex will make reference to the exact solution of the Sod’s problem or the Lax’s problem.
5.1 The Sod’s shock tube problem

We first consider the classical Sod’s shock tube problem [23] which models the one-dimensional flow resulting
from the rupture of a membrane separating air with different density and pressure inside a shock tube. The
advantage of this test case is that an exact solution exists which enables a detailed comparison between the
approximate results and the exact solution.
The test consists of a one-dimensional Riemann problem with the following Riemann data

(1, 0, 1), x 6 0,
x ∈ [−1, 1], (ρ, u, p)(0, x) = (18)
(0.125, 0, 0.1), x > 0 .
It consists of two non linear waves (one right shock and one left rarefaction) connected through a contact
discontinuity. This basic test provides information on how well a scheme captures and resolves shocks and
12
contact discontinuities and how well the correct density of the rarefaction wave is reproduced and therefore
it is used as a first test.
The computational domain here is [−1, 1] with prescribed free boundary conditions. For each numerical
computation, the following parameters have been used:
CFL : 0.25,
Simulation time (s) : 0.4,
Initial number of cells : 200,
Maximum level of mesh refinement : Lmax .
5.1.1 The Osher-Sanders and the Tang-Warnecke projections

In Section 4.2, we have recalled the computation of the numerical flux between two levels of refinement. We
compare here the Tang-Warnecke (TW) [26] and the Osher-Sanders (OS) [17] projections. In particular, we
numerically show that even if the TW projection leads to a consistent scheme, for Lmax large enough (here 6),
numerical oscillations are observed (see Figure 5(b)-Figure 5(c)). The same behavior is also obtained for
the OS projection as displayed on Figure 5(a)-Figure 5(c). We can reduce these effects using a smoothing
grid technique. It means that we prevent two adjecent cells from having a level difference greater than two.
This method minimizes the loss of the consistency and conservativity error. We have displayed on Figure 5
the density of the solution at time t = 0.4 using the first order scheme AB1M (using AB2M leads to the
same conclusion). To compare these methods, we have also computed the cpu-time.
We have without smoothing effect:
cpu-time for TW projection = 330.27 (s),

cpu-time for OS projection = 85.02 (s),
and with smoothing effect:
cpu-time for TW projection = 288.21 (s),

cpu-time for OS projection = 80.04 (s).
As displayed on Figure 5 and cpu performance above, even if the results using the TW projection are more
accurate than the OS one: the TW projection is time consuming because it is necessary to update the fluxes
on large cells at small intermediate times if they border a smaller cell. As a consequence, we will use in
what follows the OS projection. Let us also note that such a problem appear when Lmax is large. In order
to avoid such oscillations, we will limit to Lmax 6 5 which is large enough for practical applications.
5.1.2 Numerical Convergence

We study the numerical convergence of AB1, AB1M, AB2, AB2M and RK2 schemes. We compare the
cpu-time and the accuracy of the numerical solutions for level 1 6 Lmax 6 5 (i.e. on fixed and dynamic
grid). The analytical solution of the Riemann problem (14)–(17) with Riemann data (18) is computed using
5 000 uniform cells.
The numerical density of entropy production as a discontinuity and error indicator.

As emphasized by Puppo [18, 19], the numerical density of entropy production can be used as a discontinu-
ity and a local error indicator (everywhere where the solution remains smooth). Thus, the mesh refinement
parameter (12) combined with the numerical density of entropy production (see also Remark 3.1) can provide
a useful tool to refine/adapt the mesh around discontinuities and sharply varying regions. In Figures 6-7,
we plot the density ρ for levels Lmax = 2 and 5, that we compare to the numerical solution computed on a
fixed grid with NLmax cells which are respectively 254 and 681.
The adaptive solution for Lmax = 2 requires NLmax = 254 cells while for the level Lmax = 5 uses
NLmax = 681 cells.
13
As we observe on Figure 6-Figure 7, the numerical density of entropy production for levels Lmax = 2 and
5 captures the contact discontinuity (see Figures 6(d)-7(d)) and the shock discontinuity (see Figures 6(e)-
7(e)). Moreover, the numerical density of entropy production reproduces the behavior of the error made on
ρ (as displayed on Figures 6(b)-7(b)) and the mesh is automatically refined in those areas. Far from such
regions, the mesh is coarsened since the exact solution and the adaptive one coincide with each other (precise
order and error will be provided in the sequel). The accuracy is of course improved when we increase the
value of Lmax as we can compare on Figure 6-Figure 7.
The total production of entropy P (11) corresponds to the discrete lt1 lx1 norm of the numerical density of
entropy production. Through several numerical tests, we observe that P decreases as we refine. For instance,
for Lmax = 2, one has P = 0.0098 while for Lmax = 5, one has P = 0.0073.
Numerical order.
We now perform several numerical tests to compute the rate of convergence of the first and second order
adaptive scheme. We use the discrete lt1 lx1 norm of the error on the density (see Figure 8) and the discrete
lx1 on the error on the density at time t = 0.4. As the number of cells is not fixed, we consider the average
number of cells NLmax for levels Lmax = 1, 2, 3, 4, 5 used as abscissa on Figures 8(a)–8(d).
As expected (see for instance [7, 20]), the convergence of second order methods are of first order on fixed
grid due to the presence of discontinuities in the solution (as displayed on Figures 8(b)-8(d)). Nevertheless,
the adaptive scheme improves the rate of convergence for the first as well as for the second order scheme
(see Figures 8(a)–8(d)). Moreover, using local time stepping or uniform time step, the numerical order are
similar (see Figures 8(a)–8(d) and Table 1).
The rate of convergence is considerably increased by the adaptive scheme and, in our experiences, it can
be improved by changing the threshold parameters αmin and αmax at the expense of the cpu-time.
Let us also emphasize that with the present threshold parameters, the order of the AB1M scheme is
approximately 2.
Accuracy and cpu-time.

Until now, the relevance of the adaptive scheme have been obtained relatively to the number of cells. We
now focus on the efficiency of the adaptive scheme with respect to the cpu-time1 . For the previous numerical
tests, we plot the error on the density with respect to the cpu-time (see Figures 8(e)-8(f)). As expected, the
use of the local time stepping algorithm reduces considerably the cpu-time for a given error on the density.
For instance, using Lmax = 2, the AB1 and AB1M scheme uses the same number of average cells (254) and
computes with the same accuracy (lt1 lx1 error on the density is 0.0041) but AB1M computes in 0.65 s against
1.17 s for AB1. It means that AB1M computes approximatively 2 times faster than AB1 which proves the
efficiency of the local time stepping algorithm. The same statement is also observed with the second order
scheme (see Table 1).
kρ − ρex kl1t l1x vs NLmax kρ − ρex kl1x vs NLmax kρ − ρex kl1t l1x vs cpu-time
AB1U 0.70 0.72 0.41
AB1 1.86 1.93 0.64
AB1M 2.15 2.16 0.77
AB2U 0.77 0.77 0.42
AB2 1.74 1.84 0.57
AB2M 2.21 2.18 0.76
RK2U 0.76 0.76 0.42
RK2 1.68 1.80 0.55
Table 1: Numerical order on the density error using lt1 lx1 norm and lx1 at time t = 0.4 with respect to the
averaged number of cells, and the lt1 lx1 norm on the density error with respect to the cpu-time (in log10 scale).
1
Intel(R) Core(TM) i5-2500 CPU @ 3.30GHz
14
5.2 The Lax’s shock tube problem
The Lax’s shock tube is similar to the Sod’s shock tube problem, except that the amplitude of the shock is
stronger and the initial condition has a discontinuity in the velocity:

(0.445, 0.698, 3.528), x 6 0,
x ∈ [−1, 1], (ρ, u, p)(0, x) = (19)
(0.5, 0, 0.571), x > 0 .
This second test is used in order to confirm similar conclusion obtained for the Sod’s shock tube problem.
As in the previous case, the exact solution is known and computed from the Riemann problem (14)–(17)
with Riemann data (19) using 5 000 uniform cells.
The computational domain here is [−1, 1] with prescribed free boundary conditions. For each numerical
computation, the following parameters have been used:
CFL : 0.5,
Maximum level of mesh refinement : Lmax .
As done before, we first represent the numerical solution compared to the exact one for levels Lmax = 2
and 5. On Figures 9-10, we represent the density of the solution at time t = 0.13, plot its numerical density
of entropy production, the error made on the density and the level of refinement. We also perform as done
previously, the numerical order of first and second order schemes on Figures 11(a)–11(d) and compare the
accuracy versus cpu-time on Figures 11(e)–11(f) and Table 2.
Even if the Lax’s shock tube problem introduce a huge shock wave with respect to the Sod’s shock tube
problem the same conclusion holds, namely, the numerical density of entropy production reproduces the
behavior of the error done as we can observe on Figures 9-10 and thus the mesh is automatically refined
where it is necessary. Moreover, we observe again an improvement of the numerical approximation when the
adaptive scheme is used and exceptional convergence of first order adaptive scheme (see Figure 11(a) and
Table 2). Finally, the computational time is drastically reduced when the local time stepping algorithm is
used (see Figures 11(e)–11(f) and Table 2).
kρ − ρex kl1t l1x vs NLmax kρ − ρex kl1x vs NLmax kρ − ρex kl1t l1x vs cpu-time
AB1U 0.66 0.69 0.27
AB1 3.20 3.31 0.66
AB1M 3.55 3.56 0.82
AB2U 0.60 0.62 0.32
AB2 3.24 3.40 0.70
AB2M 3.31 3.11 0.78
RK2U 0.73 0.70 0.38
RK2 3.25 3.43 0.65
Table 2: Numerical order on the density error using lt1 lx1 norm and lx1 at time t = 0.13 with respect to the
averaged number of cells, and the lt1 lx1 norm on the density error with respect to the cpu-time (in log10 scale).
5.3 The Shu-Osher test case

The next test case concerns the Shu and Osher’s problem [21] in which a Mach 3 shock moves to the right
and collides with an entropy disturbance moving to the left. We focus on the solution at time t = 0.18 s
which develops 3 shocks in a highly oscillating regime. Thus, it is hard to compute an accurate numerical
solution with a fixed uniform grid which requires a large number of cells to precisely capture oscillating
regions. In this section we compare adaptive schemes of order 1 and 2. Especially, we numerically verify
that for a given accuracy, the local time stepping algorithm is less time consuming than the standard one.
15
The Shu and Osher’s initial conditions are

(3.857143, 2.629369, 10.3333), x 6 0.1
(ρ, u, p)(0, x) =
(1 + 0.2 sin(50x), 0, 1), x > 0.1
and the computational domain is [0, 1] with prescribed free boundary conditions.
As a reference solution, we compute the solution on a uniform fixed grid (20 000 cells) with the RK2
scheme. The density of the reference solution at time t = 0.18 s and its numerical density of entropy
production are displayed on Figure 12(a). This solution being computed on a very fine fixed grid, as
predicted by the theory, the density of entropy production is almost concentrated at the shocks. Even if
small productions are present between 0.5 6 x 6 0.75, one can consider such a solution as an “exact” one.
For each numerical computation, we have used the following parameters:
CFL : 0.219,
Maximum level of mesh refinement : Lmax = 4.
On Figure 12(b), we plot the density of the reference solution, the one by AB1, AB2 and RK2 schemes
and their numerical density of entropy production. Starting from 500 cells, the adaptive schemes lead to
very close solutions for each scheme and the numerical density of entropy production vanishes everywhere
where the solution is smooth and every solution fit to the reference solution. However, focusing closely on
the oscillating area between 0.5 6 x 6 0.7, one can observe that, as expected, RK2 is better than AB2 which
is better than AB1. Those results are also confirmed by the computation of the numerical order displayed
on Figure 12(d): it was not true in the two previous test cases (see Figures 8-11) as emphasized in Section
5.1 (for which the numerical order for AB1, for instance, was greater than second order methods).
Table 3 summarizes the computation of the total entropy production, the discrete lx1 norm of the error
on the density, the cpu-time, the average number of cells and the maximum number of cells at final time
t = 0.18 s for a level Lmax = 4.
As said before, the hierarchy in terms of accuracy is very well respected as we also see on Table 3 for
the error on the density or the total production of entropy. We see again a case where the total production
of entropy is an accuracy indicator.
It is well-known that the AB2 scheme is less stable and less accurate than the RK2 scheme. Nonetheless,
in the framework of the local time stepping, for almost the same accuracy the AB2M scheme computes 3
times faster than the RK2 which is a significant gain in time.
P kρ − ρref kl1x cpu-time NLmax maximum number of cells

−2
AB1 0.288 4.74 10 181 1574 2308
AB1M 0.288 4.80 10−2 120 1572 2314
AB2 0.287 2.75 10−2 170 1391 2023
AB2M 0.286 2.74 10−2 108 1357 1994
RK2 0.285 2.08 10−2 299 1375 2005
Table 3: Comparison of numerical schemes of order 1 and 2
6 Concluding remarks and perspectives

In this paper, first and second order methods in space and time are coupled with an adaptive algorithm
employing local time stepping, obtaining an adaptive numerical scheme in which the grid is locally refined
or coarsened according to the entropy indicator. Several numerical tests have been performed and show an
impressive improvement with respect to uniform grids even if a large number of cells is used.
16
All numerical tests also show that the numerical density of entropy production combined with the pro-
posed mesh refinement parameter is a relevant local error indicator (everywhere where the solution remains
smooth) and discontinuity detector (large shocks and oscillating solutions are very well-captured). Moreover,
we have shown that the implementation of the local time stepping algorithm can significantly reduce the
computational time keeping the same order of accuracy.
We have recalled the Tang-Warnecke projection and the Osher-Sanders projection. In particular, even if
the Tang-Warnecke projection is consistent, the solution develops numerical instabilities for a large maximum
level of refinement. This statement has been numerically reproduced and we have proposed a numerical
smoothing grid technique which prevents two adjacent cells from having a level difference greater than
two. As a consequence the amplitude of such oscillation is minimized. Nevertheless, since both methods
lead to oscillating solutions, the Osher-Sanders projection offers a good compromise between accuracy and
computing cost reduction. Finally, we plan to improve the efficiency of the adaptive scheme to capture
accurately the contact discontinuities and to extend this work for 2D/3D numerical applications.
Acknowledgements
The authors wish to thank the referees for their careful reading of the previous version of the manuscript
and useful remarks.
This work is supported by the "Agence Nationale de la Recherche" through the COSINUS program (ANR
CARPEiNTER project n◦ ANR-08-COSI-002).
References
[1] Allahviranloo T., Ahmady N., Ahmady E., Numerical solution of fuzzy differential equations by
predictor–corrector method. Inform. Sci., 177(7):1633–1647, 2007.
[2] Altmann C., Belat T., Gutnic M., Helluy P., Mathis H., Sonnendrücker E., Angulo W., Hérard J.M.,
A local time-stepping discontinuous Galerkin algorithm for the MHD system. In CEMRACS 2008—
Modelling and numerical simulation of complex fluids, volume 28 of ESAIM Proc., pages 33–54. EDP
Sci., Les Ulis, 2009.
[3] Berger M.J., Oliger J., Adaptive mesh refinement for hyperbolic partial differential equations. J. Comp.
Phys., 53(3):484–512, 1984.
[4] Cockburn B., Gremaud P.A., A priori error estimates for numerical methods for scalar conservation laws.
Part II: Flux-splitting monotone schemes on irregular Cartesian grids. Math. Comp., 66(218):547–572,
1997.
[5] Croisille J.P., Contribution à l’étude théorique et à l’approximation par éléments finis du système
hyperbolique de la dynamique des gaz multidimensionnelle et multi-espèce. PhD thesis, Université de
Paris VI, 1990.
[6] Eymard R., Gallouët T., Herbin R., Finite volume methods. In Handbook of numerical analysis, Vol.
VII, Handb. Numer. Anal., VII, pages 713–1020. North-Holland, Amsterdam, 2000.
[7] Gallouët T., Hérard J.M., Seguin N., Some recent finite volume schemes to compute Euler equations
using real gas EOS. Internat. J. Numer. Methods Fluids, 39(12):1073–1138, 2002.
[8] Godlewski E., Raviart P.A., Numerical approximation of hyperbolic systems of conservation laws, volume
118 of Applied Mathematical Sciences. Springer-Verlag, New York, 1996.
[9] Golay F., Numerical entropy production and error indicator for compressible flows. C.R. Mécanique,
337:233–237, 2009.
17
[10] Guermond J.L., Pasquetti R., Popov B., Entropy viscosity method for nonlinear conservation laws. J.
Comp. Phys., 230(11):4248–4267, 2011.
[11] Hairer E., P. Nørsett S., Wanner G., Solving ordinary differential equations. I, volume 8 of Springer
Series in Computational Mathematics. Springer-Verlag, Berlin, second edition, 1993. Nonstiff problems.
[12] Houston P., Mackenzie J.A., Süli E., Warnecke G., A posteriori error analysis for numerical approxi-
mations of Friedrichs systems. Numer. Math., 82(3):433–470, 1999.
[13] Jenny P., Lee S.H., Tchelepi H.A., Adaptive multiscale finite-volume method for multiphase flow and
transport in porous media. Multiscale Model. Simul., 3(1):50–64, 2005.
[14] Karni S., Kurganov A., Local error analysis for approximate solutions of hyperbolic conservation laws.
Adv. Comput. Math., 22(1):79–99, 2005.
[15] Karni S., Kurganov A., Petrova G., A smoothness indicator for adaptive algorithms for hyperbolic
systems. J. Comp. Phys., 178(2):323–341, 2002.
[16] Müller S., Stiriba Y., Fully adaptive multiscale schemes for conservation laws employing locally varying
time stepping. SIAM J. Sci. Comput., 30(3):493–531, 2007.
[17] Osher S., Sanders R., Numerical approximations to nonlinear conservation laws with locally varying
time and space grids. Math. Comp., 41(164):321–336, 1983.
[18] Puppo G., Numerical entropy production on shocks and smooth transitions. In Proceedings of the Fifth
International Conference on Spectral and High Order Methods (ICOSAHOM-01) (Uppsala), volume
17(1-4), pages 263–271, 2002.
[19] Puppo G., Numerical entropy production for central schemes. SIAM J. Sci. Comput., 25(4):1382–1415
(electronic), 2003/04.
[20] Puppo G. , Semplice M., Numerical entropy and adaptivity for finite volume schemes. Commun.
Comput. Phys., 10(5):1132–1160, 2011.
[21] Shu C. W., Osher S., Efficient implementation of essentially nonoscillatory shock-capturing schemes. J.
Comput. Phys., 77(2):439–471, 1988.
[22] Simeoni C., Remarks on the consistency of upwind source at interface schemes on nonuniform grids.
SIAM J. Sci. Comput., 48(1):333–338, 2011.
[23] Sod G.A., A survey of several finite difference methods for systems of nonlinear hyperbolic conservation
laws. J. Comput. Phys., 27(1):1–31, 1978.
[24] Sonar T., Hannemann V., Hempel D. Dynamic adaptivity and residual control in unsteady compressible
flow computation. Math. Comput. Modelling, 20(10-11):201–213, 1994.
[25] Tan Z., Zhang Z., Huang Y., Tang T., Moving mesh methods with locally varying time steps. J. Comp.
Phys., 200(1):347–367, 2004.
[26] Tang H., Warnecke G., A class of high resolution difference schemes for nonlinear Hamilton-Jacobi
equations with varying time and space grids. SIAM J. Sci. Comput., 26(4):1415–1431, 2005.
[27] Toro E.F., Riemann solvers and numerical methods for fluid dynamics. Springer-Verlag, Berlin, second
edition, 1999. A practical introduction.
[28] Zhang X.D., Trépanier J.Y., Camarero R., A posteriori error estimation for finite-volume solutions of
hyperbolic conservation laws. Comp. Meth. App. Eng., 185:1–19, 2000.
18
1 1
0.8 0.8
0.6 0.6
Density
Density
0.4 0.4
0.2 0.2
0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
(a) Numerical solution using the OS projection. (b) Numerical solution using the TW projection.
0.429
TW projection
OS projection
0.428
0.427
Density
0.426
0.425
0.424
0.423
0.422
-0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
x
(c) Without smoothing effect.
0.429
TW projection
OS projection
0.428
0.427
Density
0.426
0.425
0.424
0.423
0.422
-0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
x
(d) With smoothing effect.
Figure 5: Sod’s shock tube problem: numerical oscillations with/without smoothing effect at time t = 0.4 s
with Lmax = 6. The Tang-Warnecke (TW) and the Osher-Sanders (OS) projections are used.
19
1.2 0.6 5 0.1
ρ on adaptive mesh with Lmax = 2 level
n
ρ on uniform fixed mesh N = 254 4.5 Sk
Numerical density of entropy production

ρex 0.5 |ρ-ρex|
1 0.08
Skn 4
0.4
Mesh refinement level

0.8 3.5 0.06
0.3
Density
3
0.6
0.2
0.04
2.5
0.4 0.1
2
0.02
0 1.5
0.2
-0.1 1 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
(a) Density and numerical density of entropy production. (b) Mesh refinement level, numerical density of entropy produc-
tion and local error.
1
ρ on adaptive mesh with Lmax = 2 0.45 ρ on adaptive mesh with Lmax = 2
ρ on uniform fixed mesh N = 254 ρ on uniform fixed mesh N = 254
0.9 ρex ρex
0.4
0.8
0.7
0.35
0.6
0.3
0.5
0.4
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0.25 0.3 0.35 0.4 0.45 0.5
x x
(c) Zoom on the rarefaction. (d) Zoom on the contact discontinuity.
0.3
ρ on adaptive mesh with Lmax = 2
0.28 ρ on uniform fixed mesh N = 254
ρex
0.26
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.6 0.65 0.7 0.75 0.8
x
(e) Zoom on the shock.
Figure 6: Sod’s shock tube problem: solution at time t = 0.4 s using the AB1M scheme on a dynamic grid
with Lmax = 2 and the AB1 scheme on a uniform fixed grid of 254 cells.
20
1.2 1.4 5 0.1
n
ρ on uniform fixed mesh N = 681 4.5 Sk

1.2
1 ρex |ρ-ρex|
0.08
Skn 1 4

0.8 0.8 3.5 0.06
Density
0.6 3
0.6
0.04
0.4 2.5
0.4 0.2 2
0.02
0 1.5
0.2
-0.2 1 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
(a) Density and numerical density of entropy production. (b) Mesh refinement level, numerical density of entropy produc-
tion and local error.
1
ρ on adaptive mesh with Lmax = 5 0.45 ρ on adaptive mesh with Lmax = 5
ρ on uniform fixed mesh N = 681 ρ on uniform fixed mesh N = 681
0.9 ρex ρex
0.4
0.8
Density
Density
0.7
0.35
0.6
0.3
0.5
0.4
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0.25 0.3 0.35 0.4 0.45 0.5
x x
(c) Zoom on the rarefaction. (d) Zoom on the contact discontinuity.
0.3
ρex
0.26
0.24
Density
0.22
0.2
0.18
0.16
0.14
0.12
0.6 0.65 0.7 0.75 0.8
x
(e) Zoom on the shock.
Figure 7: Sod’s shock tube problem: solution at time t = 0.4 s using the AB1M scheme on a dynamic grid
with Lmax = 5 and the AB1 scheme on a uniform fixed grid of 681 cells.
21
-2 -2.2
AB1 uniform mesh: order = 0.69705 AB2 uniform mesh: order = 0.76835
AB1 adaptive mesh: order = 1.861 RK2 uniform mesh: order = 0.75971
AB1M adaptive mesh: order = 2.1539 AB2 adaptive mesh: order = 1.7421
-2.2 -2.4 AB2M adaptive mesh: order = 2.2141
RK2 adaptive mesh: order = 1.6808
-2.4 -2.6
log10(error)
log10(error)
-2.6 -2.8
-2.8 -3
-3 -3.2
-3.2 -3.4
-3.4 -3.6
2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3
log10(average number of cells) log10(average number of cells)
(a) kρex − ρkl1 l1 with respect to the averaged number of cells (b) kρex − ρkl1 l1 with respect to the averaged number of cells
t x t x
for the schemes of order 1. for the schemes of order 2.
-1.6 -1.8
AB2M adaptive mesh: order = 2.1825
-1.8 -2
-2 -2.2
log10(error)
log10(error)
-2.2 -2.4
-2.4 -2.6
-2.6 -2.8
-2.8 -3
2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3
(c) kρex − ρkl1 with respect to the averaged number of cells (d) kρex − ρkl1 with respect to the averaged number of cells
x x
for the schemes of order 1 at time t = 0.4 s. for the schemes of order 2 at time t = 0.4 s.
-2 -2.2
-2.4 -2.6
log10(error)
log10(error)
-2.6 -2.8
-2.8 -3
-3 -3.2
-3.2 -3.4
-3.4 -3.6
-0.5 0 0.5 1 1.5 -0.5 0 0.5 1 1.5
log10(cpu-time) log10(cpu-time)
(e) cpu-time–kρex − ρkl1 l1 with respect to the cpu-time for (f) cpu-time–kρex − ρkl1 l1 with respect to the cpu-time for
t x t x
the schemes of order 1. the schemes of order 2.
Figure 8: Sod’s shock tube problem: numerical order.
22
9 5 0.4
1.4 ρ on uniform fixed mesh N = 227 8 Skn

4.5 0.35
ρex |ρ-ρex|
7
1.2 Skn 4 0.3

6
5 3.5 0.25
1
Density
4 3 0.2
0.8 3
2.5 0.15
2
0.6 2 0.1
1
1.5 0.05
0.4 0
-1 1 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
(a) Density and numerical density of entropy production. (b) Mesh refinement level and local error.
0.46
ρ on uniform fixed mesh N = 227
0.44 ρex
0.42
0.4
0.38
0.36
0.34
-0.5 -0.45 -0.4 -0.35 -0.3 -0.25 -0.2 -0.15 -0.1
x
(c) Zoom on the rarefaction.

ρex
1.2
0.8
0.6
0.4
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

x
(d) Zoom on the contact discontinuity.
Figure 9: Lax’s shock tube problem: solution at time t = 0.13 s using AB1M scheme with Lmax = 2.
23
12 5 0.4
1.4 ρ on uniform fixed mesh N = 398 Skn

10 4.5 0.35
ρex |ρ-ρex|
1.2 Skn 4 0.3
8

3.5 0.25
1 6
Density
3 0.2
0.8 4
2.5 0.15
2
0.6 2 0.1
0 1.5 0.05
0.4
-2 1 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
(a) Density and numerical density of entropy production. (b) Mesh refinement level and local error.
0.46
ρ on uniform fixed mesh N = 398
0.44 ρex
0.42
Density
0.4
0.38
0.36
0.34
-0.5 -0.45 -0.4 -0.35 -0.3 -0.25 -0.2 -0.15 -0.1
x
(c) Zoom on the leg of the rarefaction.

ρex
1.2
1
Density
0.8
0.6
0.4
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

x
(d) Zoom on the contact discontinuity.
Figure 10: Lax’s shock tube problem: solution at time t = 0.13 s using AB1M scheme with Lmax = 5.
24
-2.2 -2.4
-2.6 -2.8
log10(error)
log10(error)
-2.8 -3
-3 -3.2
-3.2 -3.4
-3.4 -3.6
-3.6 -3.8
2.3 2.35 2.4 2.45 2.5 2.55 2.6 2.65 2.7 2.3 2.35 2.4 2.45 2.5 2.55 2.6 2.65 2.7
(a) kρex − ρkl1 l1 with respect to the averaged number of cells (b) kρex − ρkl1 l1 with respect to the averaged number of cells
t x t x
for the schemes of order 1. for the schemes of order 2.
-1.2 -1.4
-1.6 -1.8
log10(error)
log10(error)
-1.8 -2
-2 -2.2
-2.2 -2.4
-2.4 -2.6
-2.6 -2.8
2.3 2.35 2.4 2.45 2.5 2.55 2.6 2.65 2.7 2.3 2.35 2.4 2.45 2.5 2.55 2.6 2.65 2.7
(c) kρex − ρkl1 with respect to the averaged number of cells (d) kρex − ρkl1 with respect to the averaged number of cells
x x
for the schemes of order 1 at time t = 0.13 s. for the schemes of order 2 at time t = 0.13 s.
-2.2 -2.4
-2.6 -2.8
log10(error)
log10(error)
-2.8 -3
-3 -3.2
-3.2 -3.4
-3.4 -3.6
-3.6 -3.8
-1.5 -1 -0.5 0 0.5 -1 -0.5 0 0.5 1
log10(cpu-time) log10(cpu-time)
(e) cpu-time–kρex − ρkl1 l1 with respect to the cpu-time for (f) cpu-time–kρex − ρkl1 l1 with respect to the cpu-time for
t x t x
the schemes of order 1. the schemes of order 2.
Figure 11: Lax’s shock tube problem: numerical order.
25
5 10 7 12
Reference solution AB1
Skn Skn for AB1
4.5

6 AB2
Skn for AB2 10
4 8 RK2
Skn for RK2
5 reference solution
3.5 Skn for reference solution 8
6
4
Density
Density
3
6
2.5 3
4
2 4
2
1.5 2
1 2
1
0.5 0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
(a) Reference solution. (b) Density and numerical density of entropy production.
4.6 AB1
AB2
RK2
4.4 Reference solution
4.2
4
Density
3.8
3.6
3.4
3.2
0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7

x
(c) Zoom on oscillating region.
-0.6
AB1 adaptive mesh: order = 0.78839
AB2 adaptive mesh: order = 1.1846
-0.8
-1
log10(error)
-1.2
-1.4
-1.6
-1.8
2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4
log10(average number of cells)
(d) Numerical order: kρex − ρkl1 with respect to the averaged number of cells
x
for the schemes of order 1 and 2 at time t = 0.18 s.
Figure 12: Shu Osher test case.
26

Ab FLM Revisedv3

Uploaded by

Copyright:

Available Formats

Ab FLM Revisedv3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ab FLM Revisedv3

Uploaded by

Copyright:

Available Formats

Adaptive multi scale scheme based on numerical density of

entropy production for conservation laws.

2 Numerical approximation for hyperbolic conservation laws

2.2.1 Runge Kutta schemes

s(wk (tn+1 )) − s(wk (tn )) δψk (tn )

the total numerical entropy production.

2.2.2 Adams-Bashforth schemes

For example, the second order Adams-Bashforth method is:

s(wk (tn+1 )) − s(wk (tn )) δψk (tn ) δtn

3.1 Mesh refinement criterion

3.2 Mesh refinement process

Figure 1: Example of hierarchical dyadic tree.

wknb0 = wknb1 = wknb

Fknb +1/2 = Fknb1 +1/2 .

(a) Mesh refinement. (b) Mesh coarsening.

Figure 2: Mesh refinement and mesh coarsening: construction of the fluxes.

4 The local time stepping method

4.1 The local time stepping algorithm

We define the level of interface of two adjacent cells Lk+1/2 by:

Lk+1/2 = max (Lk , Lk+1 ) .

(c) At time tn + 3δtn . (d) At time tn+1 := tn + 4δtn .

Figure 3: Illustration of a global time advancement with a first order scheme.

4.2 Computation of the flux

(b) Tang and Warnecke projection.

Figure 4: Projection’s methods.

4.2.1 A simple projection method

but the global conservativity is lost.

4.2.2 A natural projection method

and the entropy flux by

Notations and settings

5.1 The Sod’s shock tube problem

5.1.1 The Osher-Sanders and the Tang-Warnecke projections

cpu-time for TW projection = 330.27 (s),

and with smoothing effect:

cpu-time for TW projection = 288.21 (s),

5.1.2 Numerical Convergence

The numerical density of entropy production as a discontinuity and error indicator.

Accuracy and cpu-time.

5.3 The Shu-Osher test case

P kρ − ρref kl1x cpu-time NLmax maximum number of cells

Table 3: Comparison of numerical schemes of order 1 and 2

6 Concluding remarks and perspectives

Numerical density of entropy production

Numerical density of entropy production

Mesh refinement level

Numerical density of entropy production

Numerical density of entropy production

Mesh refinement level

log10(average number of cells) log10(average number of cells)

log10(average number of cells) log10(average number of cells)

-0.5 0 0.5 1 1.5 -0.5 0 0.5 1 1.5

Figure 8: Sod’s shock tube problem: numerical order.

Numerical density of entropy production

Numerical density of entropy production

Mesh refinement level

ρ on adaptive mesh with Lmax = 2

0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Numerical density of entropy production

Numerical density of entropy production