Positive Scharfetter-Gummel finite volume method for

convection-diffusion equations on polygonal meshes

El-Houssaine Quenjel

To cite this version:

El-Houssaine Quenjel. Positive Scharfetter-Gummel finite volume method for convection-diffusion
equations on polygonal meshes. 2021. �hal-03259655�

HAL Id: hal-03259655

https://hal.science/hal-03259655v1
Preprint submitted on 14 Jun 2021

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Positive Scharfetter-Gummel finite volume
method for convection-diffusion equations on
polygonal meshes
El Houssaine QUENJEL
Chair of Biotechnology, LGPM, CentraleSupélec, CEBB, 3 rue des Rouges Terres, 51110 Pomacle, France
el-houssaine.quenjel@centralesupelec.fr

June 14, 2021

Abstract
In this paper, we develop and study a fully implicit positive finite volume scheme allowing
to accurately approximate the nonlinear highly anisotropic convection-diffusion equations on
almost arbitrary girds. The key idea is to relate the unstable fluxes, including the convective
ones, to the normal monotonic diffusive flux thanks to a technique used in the Scharfetter-
Gummel discretizations. Then, we obtain a nonlinear two-point-like scheme with positive
coefficients on primal and dual meshes. We check that the underlined structure naturally
ensures the nonnegativity of the approximate solutions. We also establish energy estimate,
which enables the existence of the numerical solutions. This study is accompanied with a
series of numerical results and simulations. They highlight the satisfied discrete maximum
principle, the optimal accuracy, and the robustness with respect to the mesh and to important
anisotropic ratios.

1 Introduction
The convection-diffusion equation is fundamental in modeling and numerically solving a broad
range of physical, biological and engineering problems. Typical ones are fluid flows in com-
plex porous media, chemotaxis and semiconductor device simulation. In the current work,
the primary unknown describes the diffusion of a quantity e.g. a tracer, mass density, fluid
saturation, or temperature while its displacement is governed by a given velocity field.
We next state the mathematical model we are interested in, we survey the relation with the
existing literature works and we highlight the originality as well as the novelty of this work.

1.1 Model problem

Let Ω be a bounded connected polygonal open domain of Rd with d ≥ 1. We denote by
∂Ω = ΓD ∪ ΓN its boundary. Let tf > 0 denote time. We consider the nonlinear unsteady

1
convection-diffusion problem written as
 
ut − ∇ · κ(u)Λ∇γ(u) − γ(u)V = f in Ω × (0, tf ), (1.1)
 
κ(u)Λ∇γ(u) − γ(u)V · n = 0 on ΓN × (0, tf ), (1.2)
u = 0 on ΓD × (0, tf ), (1.3)
0
u(·, 0) = u in Ω. (1.4)

The function u(x, t) represents the primary unknown. The diffusivity function is κ. The
potential and the flux transport function γ are considered identical. The space-dependent
tensor Λ refers to the diffusion matrix. The vector V is a given velocity field. The right-hand
side f accounts for source or sink contributions. The vector n is the unit outer normal to ΓN .
The numerical analysis of the scheme derived from the model (1.1)-(1.4) requires assumptions
on the data. We set Qtf = Ω × (0, tf ).
(A0 ) The initial condition u0 is nonnegative i.e. u0 ≥ 0.
(A1 ) The diffusion function κ is nondecreasing continuous from R+ into R+ . Assume
that κ does not degenerate on a whole interval of R+ . Consider that γ is increasing
and continuous from R+ into R+ with γ(0) = 0. It is extended
Rs by 0 on (−∞, 0).
0 1
Assume that J(u ) is L -integrable on Ω, where J(s) = 0 γ(v) dv. Suppose that
√ 0
R sκγ
p is also L1loc -integrable on R+ . Then, define the Kirchhoff transform µ(s) =
0
κ(v)γ (v) dv. Assume that there exists Cγ > 0 such that
0

γ(s) ≤ Cγ + µ(s), ∀s ≥ 0. (1.5)

(A2 ) The symmetric matrix Λ is a measurable function from Ω to L∞ (Ω)d×d . There exist
λ, λ > 0 such that

λ |ζ|2 ≤ Λ(x)ζ · ζ ≤ λ |ζ|2 ∀ a.e. x ∈ Ω, ∀ζ ∈ Rd .

(A3 ) The velocity V belongs to C 1 (Ω) with ∇ · V ≥ 0.

(A4 ) The function f ≥ 0 is in L2 (Qtf ).

1.2 Literature works

Several discretizations methods are available in the literature for approximating the convection-
diffusion equation (1.1) with various focuses. For instance, in the context of porous media
flows, we find the famous (Two-Point Flux Approximation) TPFA method [16, 17] thanks to
its practical implementation, computational speed and robustness. As its names states, the
two-point scheme approaches the flux using only two degrees of freedom per every interface
shared by two cells and neglects other contributions by imposing an orthogonality condition
on the mesh. Out of the orthogonal setting, multi-point approximations of the flux were
developed to take into account general meshes and fully anisotropic tensors, e.g. see [13, 14]
and the references are therein. Also, combined finite volume/finite element schemes were
proposed in [1, 3, 18, 24]. In this work we will be concerned with the Discrete Duality Finite
Volume (DDFV) method due to special approximation of the diffusive flux [2, 6, 12, 21]. We
also conceive a positive nonlinear finite volume Scharfetter-Gummel (S. G.) strategy based on
the DDFV fluxes (see the next subsection).

2
 The standard S. G. scheme was firstly introduced in [29] for the 1D drift-diffusion system
modeling charge carrier transport in semiconductors. It has been extended to the context
of finite volume methods in [5, 10, 23]. Brezzi et al. proposed a mixed finite element S. G.
scheme in the linear case based on the non-logarithmic Landau formulation [8]. Da Veig et
al. suggested a S. G. approach to handle convection terms in finite volumes and mimetic
discretization methods for elliptic problems [11]. However, the scheme of [11] does not satisfy
the discrete maximum principle. In terms of accuracy, the S. G. methodology provides a better
alternative to the upwind scheme for the transport term, especially in the diffusive regime.
On the other hand, we recently developed some recent positive nonlinear schemes with
limitations. The work based on suitable upwind flux approximations was carried out in [26].
It is only first order accurate in space which is due to the additional artificial viscosity spanned
by upwinding. A similar idea has been generalized to DDFV framework in [28]. It was
observed that the accuracy is reduced when the solution is only continuous. Their extensions
to complex flows in porous media have been investigated in [7, 19]. Contrary to upwind based-
discretizations, central positive nonlinear schemes have been proposed in [9, 25, 27]. These
strategies make use of some singular potential functionals to reinforce the solution positivity
in the case of nonlinear diffusion or linear diffusion with a drift. They are merely problem-
dependent. Generally, it is not known how to intrinsically extend them to the context of the
porous medium equation when the convective effects are present.

1.3 Originality and contribution

The present work proposes a new finite volume scheme for (1.1)-(1.4), which is stable, accurate
and robust. Stability allows to ensures the production of physical solutions and enables the
coercivity of the scheme. Accuracy is referred to as the quadratic convergence rate when the
solution is regular. By robustness, we mean that the stability and accuracy are independent
of the mesh type and anisotropic tensors.
The originality of this work lies in the flux approximation based on the multi-point DDFV
method. Precisely, at each interface between two control volumes, the DDFV strategy partic-
ularly provides normal and tangential fluxes for the diffusion. The first one is stable whereas
the second one is unstable on generic meshes. Being inspired by the Scharfetter-Gummel
technique, widely used in the TPFA context for diffusion problems with drift, a key element
in our contribution is to connect this oscillatory term as well as the convection to the stable
monotonic normal diffusive flux in a consistent way. To our knowledge, this technique has
not been employed yet for the diffusion in the multi-point setting. This leads to a scheme
having a two-point structure with positive coefficients. Such a formulation naturally entails
the nonnegativity of the solution. Additionally, the scheme is still coercive. The coercivity
result particularly serves to prove the existence of approximate solutions.
The numerical section brings out the major strengths of our novel scheme. Indeed, we test
the accuracy and the robustness by varying the mesh and anisotropy. The precision turns out
to be of second order for sufficiently smooth analytical solutions and optimal otherwise. This
is the case of the convective porous medium equation with a low regularity solution. Beside to
the positivity feature, our new scheme provides convergence rates similar to the ones produced
by unstable quasi-linear schemes of the literature.
To sum up, the significant assets of the current scheme are provided in the following list.
• Include nonlinear diffusivity and transport functions (κ, γ).
• Handle meshes with quite generic shapes (e.g. distorted, non-conforming, etc).

3
 • Incorporate highly anisotropic and heterogeneous diffusion tensors.
• Honor the physical range of the solution.
• Existence of approximate solutions.
• Second order numerical accuracy.
• Robustness with respect to the mesh and the anisotropic ratio.

1.4 Outline of the paper

In Section 2, we present and analyze the positive finite volume scheme discretizing the convection-
diffusion model (1.1)-(1.4). More specifically, we first describe the discrete setting consisting
of the spatial mesh, the discrete gradient and the discretization of the time interval. We de-
rive the proposed flux approximation by means of the Scharfetter-Gummel technique. The
final scheme is obtained by considering a fully implicit Euler approximation. Next, we study
its crucial mathematical properties. Precisely, we check that the discrete maximum principle
holds true. We demonstrate the coercivity of the scheme, and prove the existence of the nu-
merical solutions. Section 2 is devoted to the numerical results where a particular emphasis is
placed on the accuracy assessment, the validation of the proved maximum principle and the
simulation of the quarter five spot problem with barriers. Finally, in Section 4, we cloture
with a conclusion.

2 The finite volume scheme and its analysis

In this second section, we will elaborate the proposed nonlinear finite volume scheme. We are
also going to study some of its substantial mathematical properties. Before that, we need to
specify the meshed domain and some relevant notations.

2.1 Domain discretization and discrete gradient

The cornerstone point of the discrete duality finite volume methodology resides in the con-
struction of a whole gradient in 2D using only geometrical objects of the mesh. As a result,
two different directions are locally required to approach the gradient in a consistent and a
stable way. To this purpose, the normal component is approximated on the primal mesh. The
tangential component is conceived on the dual mesh.
The primal mesh P is a partitionSof Ω consisting of nonoverlapping open subset, usually
called control volumes (A), such that A∈P A = Ω. Let xA be the centeroid of A. The notation
∂A refers to the boundary of the cell A, and |A| the area of A. The family EA denotes the
set of edges of A. We assume that each inner interface σ is only shared by two adjacent cells.
So, σ ∈ EA yields the existence of a unique B ∈ P such that σ = A ∩ B. We sometimes write
σ = A|B or σAB to indicate the cell neighbors of σ which are A and B. By convention, this
notation can also cover the exterior edges. Indeed, if A is a boundary cell we still denote by
B the boundary edge, seen as a degenerate cell. The boundary cells are all gathered in the
set ∂P. The symbol |σ| defines the length of the interface σ. The unit normal to σ pointing
from A to B is labeled by nAB . Keeping the underlined orientation, the unit tangent is given
by tAB .
The dual mesh P ? is constructed and centered on the primal vertices. Let A? denote the
dual control volume. Its center coincides with the vertex xA? . In fact, A? is a subdomain

4
obtained by connecting the centeroids of the primal cells sharing the vertex xA? . The dual
edges of A? are then collected in the set EA? . The length of σ ? ∈ EA? is designated by |σ ? |.
Analogously to the primal case, the vector nA? B ? (resp. tA? B ? ) denotes the unit dual normal
(resp. tangent) to σ ? = A? |B ? .

Figure 1: Schematic draw showing the primal mesh (left with solid lines), its
associated dual (medium with dotted lines) and diamond meshes (right with dashed
lines) as well as the location of degrees of freedom.

The diamond mesh D consists of diamond cells. It is formed around the primal and dual
interfaces. Given σ = A|B with vertices xA? , xB ? , we associate a unique diamond D ∈ D.
Practically,
S the cell D is the polygon whose vertices are {xA , xA? , xB , xB ? }. It can be checked
that D∈D D = Ω. Let θD ∈]0, π/2] be the angel defined by θD = arcsin(nAB , tAB ). The
diamond area |D| is measured by |D| = sin(θD ) |σ| |σ ? | /2. The primal, dual and diamond
meshes are depicted in Figure 1. The aforementioned geometrical objects per each diamond
are illustrated in Figure 2.

Figure 2: Left : example of a whole diamond cell constructed around the interior
primal edge σAB together with its corresponding geometrical entities. Right : half
of diamond where σAB is a boundary edge.

We consider that the borders ΓD and ΓN are articulated on some vertices of the primal
mesh. We specify the sets of the different boundary cells
n o
∂P D = B ∈ ∂P/xB ∈ ΓD , ∂P N = ∂P \ ∂P D ,
n o
∂P ?,D = A? ∈ P ? /xA? ∈ ΓD , P ? = P ? \ ∂P ?,D .

5
 Let us set T = P ∪ P ? and denote by hW the diameter of the subset W ∈ D ∪ T . Define
!
1 hD hK
ξT = max max , max p , max p .
D∈D sin(θD ) D∈D |D| K∈T |K|

In particular, this indicator gives information on how flat diamonds are. To control their
shapes from degeneracy and too flatting we assume a regularity condition on the mesh. It
claims that
ξT ≤ ξ,
for some positive constant ξ.
The degrees of freedom are placed at the cell centers and at the vertices. We denote by
RT the space of discrete unknowns written as


sT = (sA )A∈T , (sA? )A? ∈P ? .

The function reconstruction on Ω is prescribed by the discrete operator IT defined from RT

into L2 (Ω) as follows
1 X
IT sT = sK 1K ,
2
K∈T
where 1K is the characteristic function of the subset K.
As mentioned in the beginning of the present subsection, the definition of the discrete
gradient formula is a key element in the flux approximation procedure when the mesh is quite
general. In the DDFV case, it is locally expressed per each diamond cell and uniquely defined
in the corresponding basis (nAB , nA? B ? ). In fact, the reconstructed gradient ∇D is a linear
operator defined from RT into L2 (Ω)2 by
 
D D 1 sB − sA sB ? − sA?
∇|D sT := ∇ sT = nAB + nA? B ? , ∀D ∈ D. (2.1)
sin(θD ) |σ ? | |σ?|
The function reconstruction and the approximate gradient are related through the discrete
Poincaré inequality. Its proof can be found in [2, Lemma 3.3]. It is recalled hereafter.
Lemma 2.1. There exists a constant Cp depending only on the diameter of Ω and the mesh
regularity ξ such that for all sT ∈ RT one has

kIT sT kL2 (Ω) ≤ Cp k∇D sT kL2 (Ω)2 .

Since |D| = sin(θD ) |σ| |σ ? | /2 the expression of the gradient is equivalent to

1  
∇D sT = σ(sB − sA )nAB + σ ? (sB ? − sA? ) nA? B ? , ∀D ∈ D.
2 |D|
The anisotropic tensor Λ(·) is attached to the diffusion and therefore to the gradient. It is also
approximated per diamonds using the integral average
Z
D 1
Λ = Λ(x) dx, ∀D ∈ D.
|D| D
In the rest of the paper, we adopt the notations

|σ|2 ΛD nAB · nAB |σ ? |2 ΛD nA? B ? · nA? B ? |σ| |σ ? | ΛD nAB · nA? B ?

τAB = , τA? B ? = , ηD = .
2 |D| 2 |D| 2 |D|

6
Define the 2 × 2 symmetric matrix RD and the vector δ D sT by
   
τ ηD sA − sB
RD = AB , δ D sT = . (2.2)
ηD τA? B ? . s A? − s B ?

It can be easily checked that

1
|D| ΛD ∇D sT · ∇D sT = RD δ D sT · δ D sT . (2.3)
2
To establish the coercivity of the scheme we need to make sure that the left hand side
of (2.3) is only controlled by the usual R2 -norm of δ D sT independently of the mesh. This is
addressed in the following result.

Lemma 2.2. There exist two positive constants m0 and m1 depending only on λ, λ and the
mesh regularity ξ such that
2 2
m0 δ D sT ≤ RD δ D sT · δ D sT ≤ m1 δ D sT , ∀sT ∈ RT , ∀D ∈ D.

Proof. Let sT ∈ RT and D ∈ D. Thanks to the ellipticity of Λ, given in (A2 ), the equality
(2.3) entails
2 2
2λ |D| ∇D sT ≤ RD δ D sT · δ D sT ≤ 2λ |D| ∇D sT .
Following the result [27, Lemma 2.1.], there exist C0 > 0 and C1 > 0 depending only on the
mesh regularity such that
2 2 2
C0 δ D s T ≤ |D| ∇D sT ≤ C1 δ D s T .

As a consequence, the required inequality is obtained by setting

m0 = 2λC0 , m1 = 2λC1 ,

which concludes the proof.

Let Ntf ∈ N \ {0}. We denote J0, Ntf K = [0, Ntf ] ∩ N. Without loss of generality, the time
interval (0, tf ) is partitioned into Ntf equidistant sub-intervals whose size is denoted by 4t.
Their end-points are given by an increasing sequence (tn )n∈J0,Ntf K where t0 = 0 and tNtf = tf .
Any time-dependent variable denoted by sn+1 T means that it is computed at the (n + 1)th level.
The temporal discretization is of chief important regarding stability. In this paper, we will
take into account a fully implicit scheme to ensure the unconditional stability in time. In other
words, no CFL condition is required for the numerical analysis of the scheme.

2.2 Generalized Scharfetter-Gummel finite volume scheme

In this part, we elaborate the proposed nonlinear finite volume scheme. We forget the time for
the moment. Our key contribution particularly targets the space discretization of the model.
First, we derive a consistent and coercive approximation of the fluxes based on the DDFV
method. We next show how to extend the Scharfetter-Gummel strategy to handle both : the
oscillatory diffusive fluxes and the convective term. Then, we provide an optimal positive
structure to the whole approximated flux. We only focus on the scheme construction in the
context of the primal mesh. Same ideas are naturally carried out in the case of the dual mesh.

7
 Let A ∈ P be a control volume. Integrating the divergence term of (1.1) on A and applying
Green’s formula yield
Z  
XA = − ∇ · κ(u)Λ∇γ(u) − γ(u)V dx
A
X Z  
=− κ(u)Λ∇γ(u) − γ(u)V · nAB dσ(x).
σ=A|B∈EA σ

Approaching the continuous gradient and the tensor Λ by their discrete counterparts permits
us to approximate XA as follows
X Z  
XA ≈ − κ(u)ΛD ∇D γ(uT ) − γ(u)V · nAB dσ(x). (2.4)
σ=A|B∈EA σ

Developing the right hand side (RHS) of (2.4) one obtains at each interface σ
Z  
− κ(u)ΛD ∇D γ(uT ) − γ(u)V · nAB dσ(x)
σ Z


D


= κ (uT ) τAB γ(uA ) − γ(uB ) + ηD γ(uA? ) − γ(uB ? ) + γ(u)V · nAB dσ(x),
σAB

where the diffusivity function is approximated thanks to a centered scheme on the diamond
D, that is
1 
κD (uT ) = κ(uA ) + κ(uB ) + κ(uA? ) + κ(uB ? ) . (2.5)
4
A well-known fashion to discretize the convective term in the RHS of (2.4) consists in the
standard upwind scheme. In contrast to the centered scheme, it takes into account the sign
of the projected velocity on nAB based on physical principles. This approach only offers an
accuracy of first order on the solution due to the additional numerical viscosity. To improve
the accuracy, at least where the diffusion is important, one can make use of the Scharfetter-
Gummel scheme. The later scheme has been intensively studied in the case of TPFA methods.
Its generalization to multi-point finite volumes schemes has been addressed in [11] for the linear
unsteady convection-diffusion problem by separating the discretization of the connective term
from the diffusion part. In the current work, we develop a novel idea enabling to connect the
unsigned tangential diffusive flux together with the convective flux to the normal monotonic
diffusive flux. To translate that into formulas, let us first set
Z
1

ϕ(s) = s coth(s/2)/2 (s 6= 0), γAB = γ (uA ) + γ (uB ) , VAB = V · nAB dσ(x).
2 σAB

Note that ϕ is extended by continuity at 0 i.e. ϕ(0) = 1. The proposed flux, which is
subsequently denoted by FAB (uT ), centers the convection and weightens the monotonic normal
flux i.e.
Z  
− κ(u)ΛD ∇D γ(uT ) − γ(u)V · nAB dσ(x)
σ

 
≈ FAB (uT ) := κD (uT )τAB ϕ gAB (uT ) γ(uA ) − γ(uB )


D ηD γ(uA? ) − γ(uB ? )
+ κ (uT )γAB
+ γAB VAB , (2.6)
γAB + ε

8
where gAB (uT ) refers to terms with uncontrolled signs that are gathered under the following
form 

ηD γ(uA? ) − γ(uB ? ) VAB
+ if κD (uT ) 6= 0


τAB κD (uT )


 τAB γAB + ε
gAB (uT ) = . (2.7)



0 if κD (uT ) = 0


The role of the parameter ε is to ensure the continuity of the function gAB with respect to
uT . To avoid it influence from a numerical perspective, it can be fixed to ε = 10−14 . The case
corresponding to κD (uT ) = 0 has no impact since its associated flux is vanished.
As a consequence, the final proposed approximation of XA is derived thanks to a general-
ization of the Scharfetter-Gummel scheme. It is written in the weighted two-point formulation
X  
XA ≈ κD (uT )τAB αAB γ(uA ) − αBA γ(uB ) . (2.8)
σ=A|B∈EA

For readability, the remained coefficients αAB , αBA are nonlinearly expressed as



αAB = α − gAB (uT ) , αBA = α gAB (uT ) ,

where α is the Lipschitz-continuous function of Bernoulli defined by

r r r 
α(r) = r = coth − 1 , ∀r ∈ R \ {0}, α(0) = 1.
e −1 2 2
It further satisfies
α(−r) − α(r) = r, ∀r ∈ R \ {0}. (2.9)
Remark 2.1.
(i) Observe that α(r) ∼ max(−r, 0). This means that the Scharfetter-Gummel scheme re-
∞
duces to the uwpwind scheme in regions where convection is highly dominated.
(ii) Even if the compact approximation (2.8) seems simple, the formula (2.6) is quite practical
for the derivation of the energy estimate, since it segregates the unstable fluxes from the
monotonic flux.
Proceeding similarly on the dual cells we find
Z  
− ∇ · κ(u)Λ∇γ(u) − γ(u)V dx
A?
X  
≈ FA? B ? (uT ) := κD (uT )τA? B ? αA? B ? γ(uA? ) − αB ? A? γ(uB ? ) .
σ=A? |B ? ∈EA?

One can check that FAB (uT ) and FA? B ? (uT ) are conservative using (2.9).

Concerning the temporal discretization we consider a fully implicit Euler scheme in order
to reinforce the unconditional stability of the scheme. This allows to avoid small time steps in
highly heterogeneous permeability regions.
The initial datum and the source term are approximated thanks to the mean integral i.e.
Z Z tn+1 Z
1 n+1 1
u0K = u(x) dx, fK = f (x, t) dx dt, K = A, A? ; n ≥ 0.
|K| K 4t |K| tn K

9
 To sum up, the discrete convection-diffusion process is governed by the nonlinear finite
volume scheme consisting of the algebraic equations et each time level n:

|A| n+1 X
uA − unA + FAB (un+1 n+1


T ) = |A| fA , ∀A ∈ P, (2.10)
4t
σAB ∈EA
|A| n+1 X
uA? − unA? + FA? B ? (un+1 ? n+1
∀A? ∈ P ? ,


T ) = |A | fA? , (2.11)
4t ?
σA ? B ? ∈EA?

un+1
B = 0, ∀B ∈ ∂P D , un+1
A? = 0, ∀A? ∈ ∂P ?,D , (2.12)
FAB (un+1
T ) = 0, ∀B ∈ ∂P N . (2.13)

The first line represents the balance equation on each primal cell. The second one accounts for
the balance equation on the dual control volumes. The latter ones are necessary to equalize
the number of unknowns with the system size and to ensure the scheme stability as we are
going to see below. The equation (2.12) is the discrete counterpart of the imposed Dirichlet
boundary condition. The last raw is nothing more than the homogeneous Neumann boundary
condition, which allows to determine un+1
B whenever B is in ∂P N .

Remark 2.2.
• When κ is constant and γ is linear, note that the flux FAB (un+1
T ) always remains non-
linear contrary to the traditional DDFV flux which is linear.
• Assume that the mesh is orthogonal in the sense of Eymard et al. [16] and the tensor Λ is
a scalar function. Such a situation may happen on Cartesian meshes or on triangulations
which the angle condition. In particular, this yields ηD = 0 for all D ∈ D. As a result,
the system (2.10)-(2.13) reduces to the nonlinear conventional Scharfetter-Gummel finite
volume scheme [5, 22, 29] on the primal and on the dual meshes separately.
• We emphasize the fact that the boundary conditions are strongly simplified in order to
highlight the key points of our contribution. The results are still valid if the Dirichlet
constraint is prescribed by a quite general function uDir . We refer to [2] for the derivation
and the analysis of the discrete Dirichlet condition using center or averaged values of uDir .

2.3 Stability analysis and existence result

In this subsection, we prove that the new nonlinear finite volume scheme is stable in the
following sense. First, the numerical solution is nonnegative as in the continuous setting.
Also, the discrete gradient of the Kirchhoff function is uniformly bounded with regard to the
energy norm. Both properties are independent of the chosen mesh. The coercivity property
allows to prove the existence result.
To start off, let us look at the central statement about the nonnegativity of the solution.

Proposition 2.1. Any solution (un+1T )n∈J0,Ntf −1K to the nonlinear system (2.10)-(2.13) re-
spects its physical range, that is

un+1
K ≥ 0, ∀A ∈ T , ∀n ∈ J0, Ntf − 1K.

Proof. It is suffices to establish

n+1
uK ≥ 0, ∀K ∈ P, (2.14)

10
by induction on n. The property on the dual mesh can be drawn in the same manner. Choose
A ∈ P so that un+1A = minK∈P un+1K . There is no thing to prove if uT
n+1
is constant. Avoiding
n+1
the trivial case, the claim (2.14) is equivalent to showing uA ≥ 0. Let us prove the latter in-
equality by contradiction. Assume that un+1 A < 0. Multiplying the line of (2.10) corresponding
to A by un+1
A yields

|A| n+1 X
uA − unA un+1 FAB (un+1 n+1
= |A| fAn+1 un+1


A + T )uA A ≤ 0. (2.15)
4t
σAB ∈EA

Let us study the left hand side of (2.15). First, it obvious that

|A| n+1
uA − unA < 0.


4t

The assumption un+1

A = minK∈P un+1K < 0 gives γ(un+1
A ) = 0 since γ is extended by 0 on
(−∞, 0). The function γ is being increasing and the coefficients αn+1
BA are positive. Therefore,
we obtain
X X
FAB (un+1 κD (un+1 n+1 n+1 n+1 n+1


T )= T )τAB αAB γ(uA ) − αBA γ(uB )
σAB ∈EA σAB ∈EA
X
D
κ (un+1 n+1 n+1 n+1


= T )τAB αBA γ(uA ) − γ(uB )
σAB ∈EA
X
+ γ(un+1 κD (un+1 n+1 n+1


A ) T )τAB αAB − αBA
σAB ∈EA
X
κD (un+1 n+1 n+1 n+1


= T )τAB αBA γ(uA ) − γ(uB ) ≤ 0.
σAB ∈EA

Accordingly, the left hand side of (2.15) is negative whereas the source term is nonpositive. The
relationship is not true except if un+1
A = 0. This is a contradiction with the hypothesis un+1
A <
0. To conclude, the nonlinear scheme prohibits the arsing of undershootss by construction.

Remark 2.3. If γ is linear, i.e. γ(u) = u, the whole numerical scheme (2.10)-(2.13) can be
recast under the symmetric matrix form R(UTn+1 )UTn+1 = q(UTn ), at each time level n. Hence,
the main entries of R are

|A| X
R(UTn+1 )AA = + κD (un+1 n+1
T )τAB αAB > 0,
4t
σAB ∈EA
D
R(UTn+1 )AB = −κ (un+1 n+1
T )τAB αBA ≤ 0.

The coefficients R(UTn+1 )A? A? and R(UTn+1 )A? B ? are defined in the same way. Since homo-
geneous boundary conditions are prescribed, the right hand side q(UTn ) includes the old level
solution and the source term. Furthermore, R(UTn+1 ) is strictly diagonally dominant and its
graph is connected. As a result, R(UTn+1 ) has the M -Matrix structure i.e. it is invertible and
all the entries of its inverse are nonnegative [4]. Consequently, any solution to the proposed
finite volume scheme is nonnegative. This gives an alternative proof to Proposition 2.1.

In the following result, we prove the scheme coercivity.

11
Proposition 2.2. The energy norm of the approximate gradient of the Kirchhoff functional
is controlled by a constant C depending only on the geometrical regularity of the mesh ξ, λ, λ
and the source function f . In other words, there holds
Ntf −1
X
4tk∇D µ(un+1 2
T )kL2 (Ω)2 ≤ C.
n=0

Proof. Multiply the balance equation (2.10) by 4tun+1

A and sum over the primal cells. Simi-
larly, multiply (2.11) by 4tun+1
A? and sum on the dual control volumes of P ? . Add the resulting
equations on both meshes, perform a discrete integration-by-parts and use the identity (2.6)
to obtain

Z0 + Z1 + Z2 + Z3 + Z4 + Z0? + Z1? + Z2? + Z3? + Z4? = 0,

where Z0 , · · · , Z4 write
Ntf −1
X X
|A| un+1 n+1
− unA γ(uA


Z0 = A ),
n=0 A∈P
Ntf −1
X X
 
κD (un+1 n+1 n+1 n+1
γ(un+1 n+1


Z1 = 4t T )τAB ϕ g (u
AB T ) γ(uA ) − γ(uB ) A ) − γ(uB ) ,
n=0 D∈D
Ntf −1
γ(un+1 n+1


n+1 ηD A? ) − γ(uB ? )
X X
D
(un+1 γ(un+1 n+1


Z2 = 4t κ T )γAB n+1

A ) − γ(uB ) ,
n=0 D∈D
γAB + ε
Ntf −1 Ntf −1
X X X X X
n+1
Z3 = 4t γAB VAB , Z4 = 4t |A| fAn+1 γ(un+1
A ).
n=0 A∈P σAB ∈EA n=0 A∈P

The dual terms Z0 , · · · , Z4 are written in the same fashion. They are omitted for legibility.
The accumulation term is commonly estimated with a telescopic series leading to
X 

|A| J(uN 0
≥ − J(u0 ) L1 (Ω) ,


Z0 ≥ A − J(u A
A∈P

where the convexity of J and Jensen’s inequality are applied. Same inequality holds for Z0 .
Next, the function ϕ satisfies ϕ(s) ≥ 1 for all s. As a consequence
Ntf −1
X X
 
κD (un+1 n+1 n+1 n+1
γ(un+1 n+1


Z1 = 4t T )τAB ϕ gAB (uT ) γ(uA ) − γ(uB ) A ) − γ(uB )
n=0 D∈D
Ntf −1
X X  
κD (un+1 n+1 n+1
γ(un+1 n+1


≥ 4t T )τAB γ(uA ) − γ(uB ) A ) − γ(uB ) .
n=0 D∈D

Z1? is analogously estimated. Define

n 
 o
D± = D ∈ D, sgn ηD γ(un+1 n+1 n+1 n+1


A ? ) − γ(uB ? ) γ(uA ) − γ(uB ) = ±1 ,

12
sgn is the sign function i.e. sgn (x) = x/ |x| for all x 6= 0. Using this notation, one finds that
X 

κD (un+1 n+1 n+1 n+1 n+1


Z2 ≥ T ) ηD γ(uA? ) − γ(uB ? ) γ(uA ) − γ(uB ) .
D∈D−

Similar inequality is valid for Z2? . Let R

e D = diag RD be a diagonal matrix defined by
extracting the diagonal of RD introduced in (2.2). It is now possible to gather the previous
terms to compute
Ntf −1

X X
Z := Z1 + Z2 + Z1? + Z2? ≥ 4t  κD (un+1 )R
e D δ D γ(un+1 ) · δ D γ(un+1 )
T T T
n=0 D∈D+

X
+ κD (un+1 D n+1 D n+1 
T )RD δ γ(uT ) · δ γ(uT ) .
D∈D−

Using the Cauchy-Schwarz inequality on the term

1 
√  
? √ 
γ(un+1
A ) − γ(un+1
B ) |σ| ΛDn
AB · γ(un+1
A ? ) − γ(un+1
B ? ) |σ | Λ Dn ? ? ,
A B
2 |D|
we deduce that
e D δ D γ(un+1 ) · δ D γ(un+1 ) ≥ 1 RD δ D γ(un+1 ) · δ D γ(un+1 ).
R T T T T
2
By virtue of this inequality, applying a couple of times Lemma 2.2, bearing in mind the fact
that κ is nondecreasing and using the dentition of the Kirchhoff functional µ we infer
f
Nt −1
m0 X X 2
Z≥ 4t κD (un+1 D n+1
T ) δ γ(uT )
2
n=0 D∈D
Ntf −1
m0 X X 2
≥ 4t max κ(s) δ D γ(un+1
T )
8 s∈{un+1 n+1 n+1 n+1
A ,uB ,uA? ,uB ? }
n=0 D∈D
f
Nt −1
m0 X X 2
≥ 4t δ D µ(un+1
T )
8
n=0 D∈D
Ntf −1
m0 X X
≥ 4t RD δ D µ(un+1 D n+1
T ) · δ µ(uT )
8m1
n=0 D∈D
Ntf −1
X m0
= m0 4tk∇D µ(un+1 2
T )kL2 (Ω)2 , m0 = .
4m1
n=0

Rearranging and introducing the positivity assumption (A2 ) on the divergence of the velocity
field we deduce
f
Nt −1 Z
1 X X
n+1 2


Z3 = 4t γ(uA ) V · n dσ(x)
2 ∂A
n=0 A∈P
Ntf −1 Z
1 X X
n+1 2


= 4t γ(uA ) ∇ · V dx ≥ 0.
2 A
n=0 A∈P

13
It can be checked that Z3? ≥ 0. Finally, thanks to the growth condition on γ mentioned in
(1.5), the Cauchy-Schwarz, Young and discrete Poincaré inequalities there holds

f
Nt −1
1 m0 X
2
|Z4 | ≤ Cγ kf kL1 (Qtf ) + 0 kf kL2 (Qt ) + 4tk∇D µ(un+1 2
T )kL2 (Ω)2
m f 4
n=0

The estimation of |Z4? | is similarly obtained. Thereby, combining all the contributions, one
ends up with
Ntf −1
X 4 4 4
4tk∇D µ(un+1 2
T )kL2 (Ω)2 ≤ C := 0
Cγ kf kL1 (Qtf ) + 0 2
kf k2L2 (Qt ) + 0 J(u0 ) L1 (Ω)
,
m (m ) f m
n=0

proving the coercivity as required.

We next wonder if the scheme has a solution. The answer is positive. It is stated and
proved in the following result.

Proposition 2.3. There exits at least one nonnegative solution (un+1

T )n∈J0,Ntf −1K to the non-
linear finite volume scheme (2.10)-(2.13).

Proof. An inductive argument on n is used for the existence proof. The initial solution trivially
exists. Assume that unT is known and let us prove the existence of un+1
T . Set m = #(P ∪ P ).
?
m m
The space R is equipped with the usual norm denoted |·|Rm . Let T : R → R be the m

continuous functional defined by

X
T(vT )|A = |A| vA − unA + 4t FAB (vT ) − 4t |A| fAn+1 , ∀A ∈ P,

σAB ∈EA
X
?
unA? FA? B ? (vT ) − 4t |A? | fAn+1 ∀A? ∈ P ? ,


T(vT )|A? = |A | vA? − + 4t ? ,

σA? B ? ∈EA?
N
FAB (vT ) = 0, ∀B ∈ ∂P .

If σAB is an edge belonging to the Dirichlet boundary, we maintain the condition vB = vA? =
vB ? = 0. Then, un+1T is solution to the finite volume scheme (2.10)-(2.13) if un+1
T solves
n+1
T(uT ) = 0. Define the homeomorphism G : Rm → Rm such that G(vT ) = γ −1 (vT ). Set
γT = γ(vT ) and Υ = T ◦ G. It follows that Υ(γT ) = T(vT ). In particular, a solution to
Υ(γT ) = 0 is automatically a solution to T(vT ) = 0 and vice-versa.
Reproducing the guidelines of the previous proof, making use of the discrete Poincaré
inequality and retain the inequality (1.5) of Assumption (A1 ) we deduce
X X
|A| vA − unA γ(vA ) + 4t



Υ(γT ) · γT = FAB (vT ) γ(vA ) − γ(vB )
A∈P σAB ∈EA
X X
|A| fAn+1 γ(vA ) |A? | vA? − unA? γ(vA? )


− 4t +
A∈P A? ∈P ?
X X
|A? | fAn+1


+ 4t FA? B ? (vT ) γ(vA? ) − γ(vB ? ) − 4t ? γ(vA? )

σA? B ? ∈E ? A? ∈P ?
0
≥ C4t k∇ D
µ(vT )k2L2 (Ω)2 −C ≥ C4t kIT γT k2L2 (Ω) − C 0,

14
for some positive constants C4t , C4t , C and C 0 . The equivalence of norms on the finite dimen-
sional space Rm ensures the existence of C4t,T > 0 depending on the mesh parameters such
that
Υ(γT ) · γT ≥ C4t,T |γT |2Rm − C 0 .
For all γT such that |γT |2Rm ≤ 1 + C4t,T /C, one infers that Υ(γT ) · γT > 0. As a consequence
of Brouwer’s fixed point criterion [15, Lemma of page 493], the vector field T admits at least
one zero. Hence, there exists a solution un+1
T to the proposed finite volume scheme for all n.
Thanks to Proposition 2.1, this numerical solution is necessarily nonnegative and the proof is
concluded.

3 Numerical results
This numerical section puts into practice the extended positive Scharfetter-Gummel discretiza-
tion allowing the resolution of the convection-diffusion problem (1.1)-(1.4). The main aim is
to highlight the capability of the developed finite volume scheme to respect the physical range
of the solution while the optimal accuracy is maintained.
In all the test, the computational domain is the unit square Ω = [0, 1] × [0, 1]. It is
discretized using five sequences of various meshes taken from the famous FVCA5 benchmark
for diffusion problems [20]. Theses partitions are made of the triangular, randomly disturbed,
locally refined, and Kershaw meshes. We label them with Tri, LocRef, Rand and Kersh
respectively. Notice that LocRef is an example of nonconforming mesh and Kersh mesh
shows the distortion along the x-direction.

Figure 3: From left to right: triangular, random, locally refined, and Kershaw
meshes.

The scheme (2.10)-(2.13) gives rise to a nonlinear algebraic system solved thanks to New-
ton’s method. Its convergence is achieved if the difference between the successive iterates in
the `2 -norm is smaller than 10−10 .
To evaluate the discrete errors we compute
N 1/2
tf
X
Esol = max kuex (., tn ) − IT unT kL2 (Ω) , Egrad =  4tk∇uex − ∇D uT k2 2
L (Ω)2
 .
n∈J0,Ntf K
n=0

3.1 Example 1 : linear Fokker-Planck equation

In this example, we are interested in the heat equation with an additional linear convection.
This model is usually called linear Fokker-Planck equation (FPE). Our goal is to study the nu-
merical convergence of the proposed ”Positive Scharfetter-Gummel Finite Volume” (PSGFV)

15
approach with regard to different meshes depicted in Figure 3. Also, the scheme precision is
assessed in the presence or the absence of anisotropy effects in the y-direction. Concerning te
used data-set, we first consider a diagonal tensor Λ as
 
ax 0
Λ= .
0 ay

We take κ(u) = 1 and γ(u) = u. Setting the right-hand side f = 0 and the mono-directional
velocity V = ay~ey = (0, ay ), it is possible to manufacture a two-dimensional solution to the
model (1.1)-(1.4) as follows
1
cos(πx) exp(−π 2 ax t) + 1 exp(y),


uex (x, y, t) = (x, y) ∈ Ω, t ∈ [0, tf ].
2
The problem is subject to no-flux Neumann boundary condition on the whole boundary. The
initial state stems from this analytical solution. For all the tests of the current example, the
final time is set to tf = 0.15 and the time step is proportional to the squared mesh size, i.e.
4t = 0.1h2T . This 4t allows to evaluate the spatial accuracy on the refined meshes.
The numerical results are graphically shown in Figures 4-7 in the log-scale using an increas-
ing anisotropic ratio varying from 0.1 to 100. Regardless anisotropy, a quadratic convergence
is obtained in the L2 norm and rates greater than 1 are noticed for the gradients.

Figure 4: Accuracy results for FPE with ay = 0.1

Figure 5: Accuracy results for FPE with ay = 1

16
 Figure 6: Accuracy results for FPE with ay = 10

Figure 7: Accuracy results for FPE with ay = 100

In Table 1 we report the minimum of the computed solution after the first time iteration
together with the required Newton’s iterations for the all considered situations. It is observed
that our approach reinforces the positivity of the solution independently of the mesh and
anisotropy. The computational cost, leaded by the Newton’s solver increases on the coarse
meshes as ay becomes important.

17
 Triangular meshes
ay = 0.1 ay = 1 ay = 10 ay = 100
2.8E-2 74 2.7E-2 73 2.8E-2 97 3.6E-2 155
7.4E-3 199 7.2E-3 282 7.4E-3 307 9.9E-3 460
1.8E-3 772 1.8E-3 772 1.8E-3 919 2.5E-3 1247
4.6E-4 3075 4.5E-4 3074 4.6E-4 3090 6.3E-4 3126
1.1E-4 12290 1.1E-4 12290 1.1E-4 12291 1.5E-4 12297
Random meshes
ay = 0.1 ay = 1 ay = 10 ay = 100
1.4E-2 134 1.4E-2 151 1.4E-2 193 3.6E-2 155
3.9E-3 363 3.9E-3 367 3.9E-3 535 9.9E-3 460
1.1E-3 1351 1.1E-3 1351 1.1E-3 1407 2.5E-3 1247
3.9E-9 7928 3.8E-9 7928 1.8E-4 7930 6.3E-4 3126
1.4E-9 31120 1.3E-9 31120 1.3E-9 31120 1.5E-4 12297
Locally refined meshes
ay = 0.1 ay = 1 ay = 10 ay = 100
2.8E-2 65 2.8E-2 65 2.8E-2 68 2.6E-2 86
7.5E-3 201 7.5E-3 201 7.5E-3 203 7.4E-3 209
1.9E-3 772 1.9E-3 772 1.9E-3 772 1.9E-3 777
4.8E-4 3075 4.8E-4 3075 4.7E-4 3075 4.7E-4 3077
1.8E-9 12289 1.8E-9 12289 1.1E-4 12289 1.2E-4 12293
Kershaw meshes
ay = 0.1 ay = 1 ay = 10 ay = 100
3.4E-2 76 3.4E-2 68 3.4E-2 99 3.4E-2 165
8.9E-3 171 8.9E-3 228 8.9E-3 267 8.8E-3 413
4.0E-3 371 3.9E-3 371 3.9E-3 480 3.9E-3 643
2.2E-3 657 2.2E-3 657 2.2E-3 657 2.2E-3 882
1.4E-3 1025 1.4E-3 1025 1.4E-3 1025 1.4E-3 1060

Table 1: FPE : lower bound of the numerical solution and total number of Newton’s
iterations with respect to anisotropy on the four used meshes.

3.2 Example 2 : heterogeneous rotating anisotropy

In this test we compare the numerical solution produced by the linear DDFV scheme and the
one obtained by our nonlinear finite volume method when applying a heterogeneous rotating
anisotropic tensor. Precisely, we are interested in assessing the positivity preservation for
both schemes. The continuous model consists of the linear heat equation together with fully
homogeneous Dirichlet boundary conditions. The convection is neglected. The diffusivity
function κ is constant and the potential γ is linear i.e. κ(u) = 1 and γ(u) = u. The anisotropy
matrix is given by  2
δx + y 2 (1 − δ)xy

1
Λ= 2 ,
x + y 2 (1 − δ)xy x2 + δy 2

18
where δ accounts for the anisotropy ratio that is fixed to 10−3 . The initial condition is chosen
as u(x, y, 0) = sin(πx) sin(πy) and there is no contribution of the source term, i.e. f = 0.
Recall that the DDFV scheme is characterized by the linear fluxes
D D n+1 D D n+1
FAB (un+1
T ) = − |σ| Λ ∇ uT · nAB , FA? B ? (un+1 ?
T ) = − |σ | Λ ∇ uT · nA? B ? .

The taken time step is 4t = 2.3 · 10−4 . The final time is tf = 0.1. We use the third
randomly disturbed mesh for the simulation.
The outputs of the test are shown on Figure 8 for three instances t = 0.01, 0.05, 0.1. The first
row corresponds to the results of the standard DDFV scheme while the second row indicates
the results of the proposed PSGFV scheme. As shown in the upper figures, a violation of the
discrete maximum principle is recorded. Indeed, the red dots illustrate the positions where the
DDFV solution presents negative and nonphysical values. These undershoots are accumulating
in time. As in Remark 2.3, the DDFV scheme can be written in a linear matrix formulation
HUTn+1 = q(UTn ), where H does not have the M-matrix structure when Λ is anisotropic or
the mesh is non-orthogonal. On the other hand, no oscillations are noticed in case of the new
scheme. In terms of accuracy, one can clearly see that both solutions are quite similar.

Figure 8: Solution of the heat equation computed by the linear DDFV scheme
(top row) and the nonlinear PSGFV scheme (bottom row) for three times t =
0.01, 0.05, 0.1.

3.3 Example 3 : porous medium equation with drift

The intention of this experiment is to demonstrate how well the developed scheme can ap-
proximate the low regular solution of the degenerate porous medium equation with drift. This
test-case is being inspired from [9].
For the moment, the tensor Λ is diagonal as in the test-case of Subsection 3.1. The
nonlinearities are taken as
κ(u) = 2 |u| , γ(u) = u.

19
Then µ(u) = |u| u. We put the velocity vector V = (ax /2, 0). Under this setting, a one-
dimensional continuous solution to the equation (1.1) is expressed by

uex (x, y, t) = max(3ax t − x, 0), (x, y) ∈ Ω, t ∈ [0, tf ].

We consider the final time tf = 0.2. For accuracy measurement, the time step is computed
by 4t = 0.1h2T . To close the model, the boundary and initial conditions agree with the exact
solution. Note that this solution is not smooth enough in space. Accordingly, the errors Esol
are expected to decrease as the mesh is refined with a rate which is strictly less than 2 (around
3/2). This fact is obviously concertized on Figures 9-12. Improved convergence orders are
obtained for most cases, especially when using the Kershaw mesh.
In Table 2 we display the lower bound on the numerical solution as well as the behavior
of the Newton solver in terms of its total iterations by mesh and anisotropic tensor. First,
the algorithm does not come with an extra computational cost for low values of ay . More
iterations are required as strong anisotropy takes place. The solution remains nonnegative
in all the cases, which is in excellent accordance with our expectation. This is not the case
of the accurate VAG (Vertex Approximate Gradient) scheme presented in [9, Test 3] where
undershoots were observed for a similar problem.

Figure 9: Accuracy results for PME with ay = 0.1

Figure 10: Accuracy results for PME with ay = 1

Next, we compare the robustness of the well known DDFV scheme and the proposed
PSGFV methodology. To this purpose, we keep the porous medium equation with the same

20
 Figure 11: Accuracy results for PME with ay = 10

Figure 12: Accuracy results for PME with ay = 100

Triangular meshes Random meshes

ay = 0.1 ay = 1 ay = 10 ay = 100 ay = 0.1 ay = 1 ay = 10 ay = 100
0 126 0 126 0 125 0 148 0 66 0 66 0 66 0 74
0 383 0 383 0 509 0 606 0 197 0 97 0 261 0 262
0 1535 0 1535 0 1539 0 1977 0 710 0 710 0 764 0 930
0 6144 0 6144 0 6144 0 6324 0 2993 0 2693 0 2693 0 2843
0 24577 0 24577 0 24577 0 24590 0 15856 0 15856 0 15856 0 15856
Locally refined meshes Kershaw meshes
ay = 0.1 ay = 1 ay = 10 ay = 100 ay = 0.1 ay = 1 ay = 10 ay = 100
0 96 0 100 0 126 0 122 0 112 0 112 0 117 0 137
0 383 0 383 0 403 0 491 0 435 0 435 0 435 0 439
0 1534 0 1534 0 1534 0 1642 0 827 0 838 0 978 0 978
0 6144 0 6144 0 6144 0 6144 0 1347 0 1354 0 1524 0 1742
0 24576 0 24576 0 24576 0 24576 0 2045 0 2045 0 2170 0 2524

Table 2: PME : lower bound of the numerical solution and total number of Newton’s
iterations with respect to anisotropy on and the level the four used meshes.

nonlinearities. The convection is neglected and the anisotropic tensor is now set to
   
cos(θ) − sin(θ) 1 0 cos(θ) sin(θ)
Λ= ,
sin(θ) cos(θ) 0 20 − sin(θ) cos(θ)

21
where θ = π/8. The initial state of the solution is considered to be u0 = 0, and the imposed
boundary Dirichlet conditions are chosen such that

u(x, y, t) = max(2t − x, 0), (x, y) ∈ ∂Ω, t ∈ [0, tf ].

We underline that the DDFV scheme for the porous medium equation is characterized by
the nonlinear fluxes
D D D D
FAB (un+1 n+1
T ) = − |σ| Λ ∇ µ(uT ) · nAB , FA? B ? (un+1 ? n+1
T ) = − |σ | Λ ∇ µ(uT ) · nA B .
? ?

We utilize the first element of Kershaw meshes. The simulation stops as it reaches tf =
0.198 by the time step 4t = 0.073. The results are exhibited on Figure 13. Since the DDFV
scheme is not positive, undershoots on the numerical solution are expected to occur as the
left part of the same figure indicates. On the other hand, the PSGFV scheme is free of such
instabilities. This is once again confirmed by the right part of Figure 13.

Figure 13: DDFV solution (left) and PSGFV (righ) for the anisotropic PME at
tf = 0.198. The red dots account for the locations of occurred undershoots.

3.4 Example 4 : quarter five spot problem

We now look at the quarter five spot problem widely known in the context of multi-phase
flows in porous media. In this study, it serves to accurately capture, quantify and track the
behavior of water diffusion with a heterogeneous porous medium such as wood, underground,
etc, regardless the used mesh. Then, the solution may account for the water content or
saturation.
The computational domain is Ω = (0, 1) × (0, 1). It is composed of three parts defined by


Ω1 = Ω \ Ω1 ∪ Ω2 , Ω2 = (0.75, 1) × (0.625, 0.75), Ω2 = (0.25, 0.75) × (0.25, 0.375).

The boundary ∂Ω is split into the
outflow boundary ΓDl , the injection one ΓDr and the
Neumann one ΓN = ∂Ω \ ΓDl ∪ ΓDl where
[
ΓDl = {x = 0} × {y ≥ 0.9} {x ≤ 0.1} × {y = 1},
[
ΓDr = {x ≥ 0.9} × {y = 0} {x = 1} × {y ≤ 0.1}

We refer to Figure 14 for an illustration of these subsets. In this test-case, we neglect convection

22
 Figure 14: Schematic illustration of the porous medium with barriers Ω2 and Ω3
with low permeabilities. The injection zone is ΓDr . The production one is located
at ΓDl . The rest of the boundary is impermeable.

effects and source term. Only diffusion is taken into account with a nonlinear diffusivity
function
u2
κ(u) = 2 ,
u + 8(1 − u)2
and γ(u) = u. Highly heterogeneous permeabilities are considered in these subdomains as
follows
 −3 
10 0
Λ(x) = I2 , x ∈ Ω1 , Λ(x) = , x ∈ Ω2 , Λ(x) = 10−4 I2 , x ∈ Ω3 ,
0 10−1
where I2 is the identity matrix. The region of small permeability acts as a barrier.
Initially, there is no water in the medium. This amount to setting u0 = 0. We put Ω in
contact with water at ΓDr by imposing uDr = 1 in the course of time simulation. Letting an
outflow condition on ΓDl , the fluid can exit the medium. ΓDl is refereed to as an exaction zone
playing the role of well. The remained part ΓN is impervious.
The domain Ω is discretized utilizing a triangular mesh made of 3584 elements and a locally
refined mesh consisting of 2560 cells. The used time step is 4t = 0.001. In this experiment, we
consider that the permeability is constant by cells. It is then discontinuous across the primal
edges. To deal with that, we implemented the coefficient τAB , τA? B ? and ηD that are borrowed
from the m-DDFV method. Such a strategy allows to maintain the quadratic convergence
of our scheme, otherwise the optimal accuracy is only of first order. One can consult [6, 6.
Examples] for more details on this method.
In Figure 15, snapshots of the saturation profile are exhibited for different times. The first
row corresponds to the result on the triangular mesh whereas the second one refers to the
results on the locally refined mesh. As expected, in the both cases, the fluid spreads from the
injection zone towards the exit by avoiding the areas of lower permeabilities. It is clearly seen
that Ω3 takes more time to be occupied by water due its small permeability.
We here emphasize two key observations. On the one hand, no numerical instabilities such
as undershoots are noticed during the simulation and the saturation honors its physical ranges,
which is already justified by the scheme construction in Proposition 2.1. On the other hand,
the Newton solver requires around 10 iterations to converge at the beginning while it needs 3
or 2 after that.

23
 Figure 15: Saturation profile on the triangular mesh (top row) and on the locally
refined mesh (bottom row) at t = 0.5, t = 1.5, t = 3 and t = 8.

4 Conclusion
In this work we presented an efficient and a robust positive nonlinear finite volume scheme for
solving convection-diffusion equations on polygonal meshes with heterogeneous and anisotropic
diffusivity tensors. The key idea consists in connecting nonlinearly the unstable tangential flux
and the convection term thanks to a technique widely used in Scharfetter-Gummel approxima-
tions. As outcomes, the scheme is naturally free of numerical instabilities such as undershoots
and the optimal accuracy is formally achieved in the diffusive regime independently of the
mesh and anisotropy. These findings have been jointed with intensive numerical validations
and illustrations focusing on the Fokker-Planck equation with a drift, the convective porous
medium equation and the highly heterogeneous quarter five spot problem.

Acknowledgments: This study was carried out in the Centre Européen de Biotechnologie
et de Bioéconomie (CEBB), supported by the Région Grand Est, Département de la Marne,
Greater Reims and the European Union. In particular, the author would like to thank the
Département de la Marne, Greater Reims, Région Grand Est and the European Union along
with the European Regional Development Fund (ERDF Champagne Ardenne 2014-2020) for
their financial support of the Chair of Biotechnology of CentraleSupélec.

References
[1] M. Afif and B. Amaziane. Convergence of finite volume schemes for a degenerate convection–
diffusion equation arising in flow in porous media. Computer Methods in Applied Mechanics and
Engineering, 191(46):5265–5286, 2002.

24
 [2] B. Andreianov, F. Boyer, and F. Hubert. Discrete duality finite volume schemes for Leray–Lions–
type elliptic problems on general 2D meshes. Numerical Methods for Partial Differential Equations,
23(1):145–195, 2007.

[3] P. Angot, V. Dolejšı́, M. Feistauer, and J. Felcman. Analysis of a combined barycentric finite
volume-nonconforming finite element method for nonlinear convection-diffusion problems. Appli-
cations of Mathematics, 43(4):263–310, 1998.

[4] A. Berman and R. J. Plemmons. Nonnegative matrices in the mathematical sciences. SIAM,
Philadelphia, PA, USA, 1994.

[5] M. Bessemoulin-Chatard. A finite volume scheme for convection–diffusion equations with nonlinear
diffusion derived from the Scharfetter–Gummel scheme. Numerische Mathematik, 121(4):637–670,
2012.

[6] F. Boyer and F. Hubert. Finite volume method for 2D linear and nonlinear elliptic problems with
discontinuities. SIAM Journal on Numerical Analysis, 46(6):3032–3070, 2008.

[7] K. Brenner, R. Masson, and E. H. Quenjel. Vertex Approximate Gradient Discretization preserving
positivity for two-phase Darcy flows in heterogeneous porous media. Journal of Computational
Physics, 409:109357, 2020.

[8] C. Buet and S. Dellacherie. On the Chang and Cooper scheme applied to a linear Fokker-Planck
equation. Communications in Mathematical Sciences, 8(4):1079–1090, 2010.

[9] C. Cancès and C. Guichard. Numerical analysis of a robust free energy diminishing finite volume
scheme for parabolic equations with gradient structure. Foundations of Computational Mathemat-
ics, 17(6):1525–1584, 2017.

[10] C. Chainais-Hillairet and J. Droniou. Finite-volume schemes for noncoercive elliptic problems with
neumann boundary conditions. IMA Journal of Numerical Analysis, 31(1):61–85, 2011.

[11] B. Da Veiga, J. Droniou, and G. Manzini. A unified approach for handling convection terms in
finite volumes and mimetic discretization methods for elliptic problems. IMA Journal of Numerical
Analysis, 31(4):1357–1401, 2011.

[12] K. Domelevo and P. Omnès. A finite volume method for the Laplace equation on almost arbitrary
two-dimensional grids. ESAIM: Mathematical Modelling and Numerical Analysis, 39(6):1203–1249,
2005.

[13] J. Droniou. Finite volume schemes for diffusion equations: introduction to and review of modern
methods. Mathematical Models and Methods in Applied Sciences, 24(08):1575–1619, 2014.

[14] J. Droniou, R. Eymard, T. Gallouët, C. Guichard, and R. Herbin. The gradient discretisation
method, volume 82. Springer, 2018.

[15] L. C. Evans. Partial differential equations, volume 19. American Mathematical Society, 2010.

[16] R. Eymard, T. Gallouët, and R. Herbin. Finite volume methods. In Handbook of Numerical
Analysis, volume 7, pages 713–1018. Elsevier, 2000.

[17] R. Eymard, T. Gallouı̈t, R. Herbin, and A. Michel. Convergence of a finite volume scheme for
nonlinear degenerate parabolic equations. Numerische Mathematik, 92(1):41–82, 2002.

[18] R. Eymard, D. Hilhorst, and M. Vohralı́k. A combined finite volume–finite element scheme for the
discretization of strongly nonlinear convection–diffusion–reaction problems on nonmatching grids.
Numerical Methods for Partial Differential Equations, 26(3):612–646, 2010.

25
[19] M. Ghilani, E. H. Quenjel, and M. Saad. Positivity-preserving finite volume scheme for compress-
ible two-phase flows in anisotropic porous media: The densities are depending on the physical
pressures. Journal of Computational Physics, 407:109233, 2020.

[20] R. Herbin and F. Hubert. Benchmark on discretization schemes for anisotropic diffusion problems
on general grids. In R. Eymard and J.-M. Herard, editors, Finite Volumes for Complex Applications
V, pages 659–692. Wiley, 2008.

[21] F. Hermeline. A finite volume method for the approximation of diffusion operators on distorted
meshes. Journal of computational Physics, 160(2):481–499, 2000.

[22] A. Jüngel. Numerical approximation of a drift-diffusion model for semiconductors with nonlin-
ear diffusion. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte
Mathematik und Mechanik, 75(10):783–799, 1995.

[23] A. Jüngel and P. Pietra. A discretization scheme for a quasi-hydrodynamic semiconductor model.
Mathematical Models and Methods in Applied Sciences, 7(07):935–955, 1997.

[24] I. S. Pop and W.-A. Yong. A numerical approach to degenerate parabolic equations. Numerische
Mathematik, 92(2):357–381, 2002.

[25] E. H. Quenjel. Analysis of accurate and stable finite volume scheme for anisotropic diffusion
equations with drift. Preprint, 2019.

[26] E. H. Quenjel. Enhanced positive vertex-centered finite volume scheme for anisotropic convection-
diffusion equations. ESAIM: Mathematical Modelling and Numerical Analysis, 54(2):591–618,
2020.

[27] E. H. Quenjel. Nonlinear finite volume discretization for transient diffusion problems on general
meshes. Applied Numerical Mathematics, 161:148–168, 2021.

[28] E. H. Quenjel, M. Saad, M. Ghilani, and M. Bessemoulin-Chatard. Convergence of a positive

nonlinear DDFV scheme for degenerate parabolic equations. Calcolo, 57(19), 2020.

[29] D. L. Scharfetter and H. K. Gummel. Large-signal analysis of a silicon read diode oscillator. IEEE
Transactions on electron devices, 16(1):64–77, 1969.

26