See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258258340

Solution of 2D convection-diffusion transient problems by a fractional-step FE

method

Conference Paper · October 2011

CITATION READS

1 433

2 authors, including:

Vincenzo Angelo Pennati

57 PUBLICATIONS   163 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Vincenzo Angelo Pennati on 01 June 2014.

The user has requested enhancement of the downloaded file.

 MASCOT11-IMACS/ISGG Workshop IAC-CNR, Rome, Italy

Solution of 2D convection-diffusion transient

problems by a fractional-step FE method
Antoine C. Kengni Jotsa and Vincenzo A. Pennati
Università degli Studi dell’Insubria,
Dipartimento di Scienza e Alta Tecnologia,
Via Valleggio 11, 22100 Como, Italy
[email protected]
[email protected]

Abstract
In this paper, we present a time advancing scheme based on a fractional
step method for the solution of 2D linear convection diffusion problems
in which the linear convective term is approximated by means of charac-
teristics, and the spatial approximation is treated by the FE with second
degree polynomials. The accuracy of the method is verified by solving
some numerical experiments. We also present the numerical results that
follow for the application of the flux conservation method [5] to classi-
cal time dependent parabolic problems. As the computational kernel is
of elliptic kind, a Schwarz overlapping multi-domains technique has been
adopted as preconditioner.

Keywords: Convection-diffusion, FE, Fractional-step, Flux conservation, Schwarz

preconditioner.

1. Introduction
Convection-diffusion transient equations are generally employed to describe the
transport processes involving fluid motion. By inspecting these type of equa-
tions, one find that they contain two distinct differential operators derived from
the described physical processes: the convection and the diffusion operators.
The convection operator consists of variations (first-order spatial derivatives) of
the transported variables which arise from fluid flow. On the other hand, the
diffusion operator (represented by second order spatial derivatives) is due to the
transport at the molecular level. These two terms can be treated separately
and then associated to obtain the discretized expression. The convective term
has been afforded in the framework of finite difference (FD), finite volume (FV)

1
and finite element (FE) methods, i.e., by second order centered or by first order
upwind FD schemes, by high order FV scheme such as Quick [4], and by highly
consistent FE scheme as SUPG [1]. It appears from the large amount of papers
written and schemes developed that the approximation of the convective term
is an important and difficult task. The proposed method consists of a fractional
step scheme for advancing in time, where the convective term is approximated by
characteristics and the diffusive term by traditional FE approach using second
degree polynomials. Being the most expensive kernel of elliptic kind, the final
algebraic system is solved by iterative method (Bi-CGSTAB) preconditioned by
a scalable additive preconditioner. The paper is organized as follows: in section
1, we make a brief overview of the characteristic methods, then in section 2 we
present an advancing in time fractional step method followed by numerical tests
and the conclusions.

2. Characteristic Galerkin method

We assume that Ω is a bounded domain of R2 with Lipschitz boundary ∂Ω
and we consider the time dependent initial boundary value problem: for each
t ∈ [0, T ] find u(X, t) ∈ L2 ((0, T ); H10 (Ω)) ∩ C0 ([0, T ]; L2 (Ω)) such that

∂u
+ Lu = f in (0, T ) × Ω
∂t
u = 0 on (0, T ) × ∂Ω (1)
u|t=0 = u0 in Ω
P2 ∂
where L is the second-order elliptic operator Lw := −Γ∆w + i=1 ci · ∂x i
(w),
∆ is the Laplace operator, Γ a physical parameter and c is a velocity field.
We consider the case in which Γ ≪ kckL∞ (Ω) and we also assume there exists
two positive constants µ0 and µ1 such that 0 < µ0 ≤ µ(x) = 21 div c(x) ≤ µ1 for
almost every x ∈ Ω.
Traced by Pironneau [7] and Douglas-Russel [2], the characteristic method
consists of considering the unsteady advection diffusion equation (1) from a
Lagrangian instead of Eulerian point of view. By defining the characteris-
tic lines associated to a vector field c = c(t, x), and by giving x ∈ Ω and
s ∈ [0, T ], there exists a vector function X = X(t; s, x) such that dX dt (t; s, x) =
c(t, X(t; s, x)), t ∈ (0, T ) and X(s; s, x) = x.
It has been proved in [3] that the existence and uniqueness of the charac-
teristic lines for each choice of s and x hold under mild assumption on c, for
instance c continuous in [0, T ] × ∂Ω, Lipstchiz continuous in Ω, uniformly with
respect to t ∈ [0, T ]. Geometrically, X(t; s, x) provides the position at time t of a
particle which has been driven by the field c and that occupied the position x at
time s. In particular, the uniqueness results gives X(t; s, X(s; τ, x)) = X(t; τ, x)
for each t, s, τ ∈ [0, T ] and x ∈ Ω. Let X n (x) be a suitable approximation of
X(tn ; tn+1 , x), n = 0, 1, · · · , N − 1, and un ◦ X n an approximation of u at the
point X n , then applying the backward scheme to the time derivative, we have

2
the following implicit discretization scheme for the problem (1): set u0 := u0 ,
n+1 n
◦X n
then for n = 0, 1, · · · , N − 1, solve u −u ∆t − Γ∆un+1 = f (tn+1 ) in Ω and
n+1
u = 0 on ∂Ω.
The backward Euler scheme can be chosen also for discretizing dX dt (t; tn+1 , x) =
c(t, X(t; tn+1 , x)) and this give the following approximation of X(tn ; tn+1 , x):
X n (x) = x − c(tn+1 , x)∆t.
Remark 2.1 It is proven in [8] that when a finite element subspace of the
Sobolev space H10 (Ω) is taken as the space of continuous and piecewise linear
h2
polynomials vanishing on ∂Ω we have that ku(tn )−unh k0 = O(h+∆t+ ∆t ),where
unh is the approximate solution of (1), h is the characteristic parameter of the
elements in the triangulation Ih of Ω.

3. A fractional step for convection diffusion equa-

tions
3.1. Weak and algebraic formulation
In order to fix the ideas let us consider the problem (1) rewritten in term of a
temperature T that traduces the propagation of the temperature:
∂T
+ c · ∇T − Γ ∆T = f in Ω (2)
∂t
T|t=0 = T0 (3)
T = 0 on ∂Ω . (4)

where c = (u, v) is a known velocity field, ∇ the gradient operator and in

this case Γ is a known thermal diffusion parameter. Applying the implicit
discretization and the characteristics in the convective term of equation (2), it
comes for n = 0, · · · , N − 1,

Tn+1 − Tn (X̃)
− Γ ∆Tn+1 = f n+1 (5)
∆t

where X̃ is the foot of characteristic.

Multiplying (5) by a suitable test function ϕ and integrating by parts, the weak
formulation of (5) with (4) is the following:
Z Z Z
Tn+1 − Tn (X̃)
ϕ dΩ − Γ ∆Tn+1 ϕ dΩ = f n+1 ϕ dΩ ∀ ϕ ∈ H10 (Ω) (6)
Ω ∆t Ω Ω

i.e. Z Z Z
Tn+1 ϕ dΩ − Tn (X̃)ϕ dΩ + ∆t Γ ∇Tn+1 · ∇ϕ dΩ =
Ω Ω
Z Ω
(7)
∆t f n+1
ϕ dΩ ∀ ϕ ∈ H10 (Ω) .
Ω

3
 Taking as basis functions traditional 2D polynomials of second degree with six
nodes on the boundary, it comes
X X Z X Z
n+1 n+1
[ Ti ϕi ϕj dK − ∆tΓ Ti ∇ϕi · ∇ϕj d K ] =
K∈Ih i K i K
X XZ X Z
n n+1
[ Ti (X̃)ϕi ϕj d K + ∆t fi ϕi ϕj dK ] i, j = 1, · · · , Nh
K∈Ih i K i K
(8)
i.e. in algebraic form:
Z
MTn+1 + Γ ∆tATn+1 = ∆tMf n+1 + Tn (X̃)ϕ dΩ
Ω

thus Z
( M + Γ ∆tA )Tn+1 = ∆tMf n+1 + Tn (X̃) ϕ dΩ (9)
Ω
R R
where M := Ω ϕi ϕj dΩ is the mass matrix and A := Ω ∇ϕi · ∇ϕj dΩ is the
stiffness matrix. To apply the formulation (9), we have to respect the following
iterative procedure:

Iterative procedure
step 1) The nodal values of T are given at time instant t = tn
step 2) Given the velocity c at time instant t = tn+1
R
step 3) Compute Ω Tn (X̃)ϕ dΩ (indication of its computation in the next section)
step 4) Compute the nodal values of T at time instant t = tn+1
step 5) Go to step 1 and repeat the above procedure until the desired solution is
obtained.

3.2. Computation of the step 3 of the iterative procedure

R P R P
We have Ω Tn (X̃)ϕ dΩ = K K ( i Tni (X̃)ϕi )ϕj dK.
R
To compute Tn (X̃) and Ω Tn (X̃)ϕ dΩ, we proceed as follows:

• Compute the foot of characteristics X̃ by means of one of these expressions

X̃ = X − ∆t cn+1 (X)
X̃ = X − ∆t[ 2cn+1 (X) − cn (X) ]
∆t n+1 (10)
X̃ = X − ∆t cn+1 (X − c (X))
2
∆t n+1 ∆t n+1
X̃ = X − ∆t[ 2 cn+1 (X − c (X)) − cn (X − c (X)) ] ,
2 2

4
 Figure 1: Backward characteristics, foot X̃ of the trajectory.

where X is the physical coordinate of the node of the element in the

triangulation Ih at time instant t = tn+1 , see Figure 1 (the coordinate of
the centroid of the element has to be considered too)
• Find the element in the triangulation Ih that contains the foot X̃ of char-
acteristic and obtain the value Tn (X̃) by interpolation of the nodal values
of T at instant tn
• Use a Gauss quadrature formula i.e.
Z X X
7 X 6
( Tni (X̃)ϕi )ϕj ∼
= ( Tni (X̃l )ϕi (ξl , ηl )) ϕj (ξl , ηl ) J(ξl , ηl ) wl , j = 1, · · · , 6
K i l=1 i=1

where wl are the weights of the Gauss points, J(ξl , ηl ) is the modulus of the Jaco-
bian of the transformation involving the element K and the reference element,
all evaluated at the seven integrations points (vertex, mid-edge and centroid
points).
Other indications about the implementations of the method of characteristics
can be referred to [9].
Remark 3.1 The computation of the foot of characteristics should be consistent
with the boundary condition considered. For instance, the foot of characteristics
can never lie outside the physical domain Ω.
Remark 3.2 As the more expensive computational kernel from the formulation
(9) is of elliptic kind, a Schwarz additive overlapping multi-domains technique
as preconditioner can be adopted [5, 6].
Remark 3.3 A generalization of the presented approach to the non linear convective-
diifusive equations afford to solve efficiently incompressible fluid dynamics prob-
lems such as the Navier-Stokes and Shallow Water ones without using the arti-
ficial compressibility approach [6].

5
Remark 3.4 It has been presented in [5] a new FE scheme by means of which,
using 2D second degree polynomials for basis and test functions, is possible to
guarantee the continuity of the mathematical fluxes in elliptic problems. Thus
we decided to carry out some numerical experiences in order to heuristically
verify the efficiency of the conservative FE [5] also in parabolic problems. In
section 4, the Tables 9 - 11 show the results of the numerical tests.

4. Numerical tests
In order to check the accuracy of the formulation (9), several problems with
known analytical solutions have been considered. The computations were per-
formed in Ω = [0, 1]2 . In Tables 1 - 7, we reported the L∞ and L2 error norms
generated by the numerical solutions. We have considered in all cases Dirichlet
boundary conditions. We denote by X̃ 1 , X̃ 2 , X̃ 3 , X̃ 4 , the foot of characteristics
obtained using the expression (10)1 , (10)2 , (10)3 and (10)4 respectively. Let de-
note by Nh the total number of nodes with respect to each domain considered,
by ∆t the time increment, by N ts the number of time steps, by c = (u, v) the
velocity and by T the temperature. We note Ah = M+ Γ ∆tA the final matrix of
the formulation (9). In Table 8, we reported the condition number of the matrix
Ah when it is not preconditioned K(Ah ) and when it is Schwarz preconditioned
K(Pas Ah ). Contemporary, we collected the iterations numbers obtained using
the Bi-CGSTAB iterative method when the algebraic system (9) is not precon-
ditioned 6= Iters and when it is Schwarz preconditioned 6= Iters − S. We define
the following expressions for the analytical velocity and temperature:

cI = (uI , v I ) such that uI = − cos(πx) sin(πy) exp(−2t)

v I = sin(πx) cos(πy) exp(−2t)

cII = (uII , v II ) such that uII = − cos(πx) sin(πy)

v II = sin(πx) cos(πy)

TI = xy(1 − x)(1 − y) exp(−2t)

TII = xy(1 − x)(1 − y)

TIII = − cos(πx) sin(πy) exp(−2t)

6
 Table 1: First test with Nh = 889, ∆t = 0.001, N ts = 10, c = cI and T = TI .

X̃ L∞ − norm L2 −norm
X̃ 1 3.281E − 4 9.416E − 4
X̃ 2 3.084E − 4 9.168E − 4
X̃ 3 3.346E − 4 9.792E − 4
X̃ 4 3.142E − 4 9.460E − 4

Table 2: Second test with Nh = 109, ∆t = 0.001, N ts = 10, c = cI and T = TI .

X̃ L∞ − norm L2 −norm
X̃ 1 2.077E − 4 4.418E − 4
X̃ 2 2.072E − 4 4.414E − 4
X̃ 3 2.078E − 4 4.420E − 4
X̃ 4 2.072E − 4 4.4152E − 4

Table 3: Third test with Nh = 241, ∆t = 0.1, N ts = 10, c = cI and T = TI .

X̃ L∞ − norm L2 −norm
X̃ 1 4.523E − 4 1.140E − 3
X̃ 2 4.430E − 4 1.133E − 3
X̃ 3 4.543E − 4 1.148E − 3
X̃ 4 4.488E − 4 1.139E − 3

Table 4: Fourth test with Nh = 241, ∆t = 0.1, N ts = 10, c = cI and T = TII .

X̃ L∞ − norm L2 −norm
X̃ 1 6.486E − 3 1.746E − 2
X̃ 2 6.455E − 3 1.744E − 2
X̃ 3 6.499E − 3 1.749E − 2
X̃ 4 6.164E − 3 1.466E − 2

Table 5: Fifth test with Nh = 733, ∆t = 0.0725, N ts = 10, c = cI and T = TI .

X̃ L∞ − norm L2 −norm
X̃ 1 2.760E − 4 8.901E − 4
X̃ 2 2.626E − 4 8.691E − 4
X̃ 3 2.792E − 4 9.134E − 4
X̃ 4 2.653E − 4 8.876E − 4

7
Table 6: Sixth test with Nh = 457, ∆t = 0.01, N ts = 1, N ts = 25 and N ts = 50,
c = cII and T = TI .

X̃ L∞ − norm L2 −norm
N ts = 1 X̃ 1 4.873E − 4 1.395E − 3
N ts = 25 X̃ 1 9.872E − 3 3.279E − 2
N ts = 50 X̃ 1 9.948E − 3 2.884E − 2

Table 7: Seventh test with Nh = 457, ∆t = 0.001, N ts = 1 and N ts = 500,

c = cII and T = TI .

X̃ L∞ − norm L2 −norm
N ts = 1 X̃ 1 2.571E − 4 6.010E − 4
N ts = 500 X̃ 1 8.310E − 3 2.400E − 2

Table 8: Eighth test with Nh = 1569, ∆t = 0.005, N ts = 200, c = cI and T = TI

X̃ L∞ −norm K(Ah ) K(Pas Ah ) 6= Iters 6= Iters − S

X̃ 1 1.397E − 5 8.796E + 1 2.213E + 1 67 8

Remark 4.1 In Table 8, we see that the Schwarz additive preconditioner works
well in the sense that the condition number K(Pas Ah ) and the iteration number
6= Iters−S diminish substantially with respect to the not preconditioned counter
parts.

In order to verify the accuracy of the conservative FE [5] scheme to parabolic

problems, we solved three problems like (5) with the following known solutions
u:
uI = (x + y) exp(−t)
uII = xy(1 − x)(1 − y) exp(−2t)
uIII = − cos(π x) sin(π y) exp(−2 t) .
In all the computations, we assigned Dirichlet boundary conditions and chose
the parameters: 72 elements with Nh = 169 and ∆t = 0.005. In Tables 9 -
11 are reported the L∞ error norm values calculated by traditional (FE) and
conservative finite elements (FEC) for N ts = 2, 20, 100 and 200. M ax u is
the maximum value of the known solutions u after N ts time steps.

8
 Table 9: Nineth test with u = uI and L∞ − norm after N ts time steps.

N ts = 2 N ts = 20 N ts = 100 N ts = 200
L∞ −norm by FE 4.691E − 5 7.278E − 5 5.092E − 5 3.084E − 5
L∞ −norm by FEC 8.894E − 5 1.748E − 4 1.250E − 4 7.588E − 5
M ax u 1.980E + 0 1.800E + 0 1.210E + 0 7.357E − 1

Table 10: Tenth test with u = uII and L∞ − norm after N ts time steps.

N ts = 2 N ts = 20 N ts = 100 N ts = 200
∞
L −norm by FE 5.169E − 5 5.637E − 5 2.661E − 5 9.793E − 6
L∞ −norm by FEC 1.438E − 3 4.604E − 3 2.223E − 3 8.215E − 4
M ax u 6.120E − 2 5.117E − 2 2.200E − 2 8.450E − 3

Table 11: Eleventh test with u = uIII and L∞ − norm after N ts time steps.

N ts = 2 N ts = 20 N ts = 100 N ts = 200
L∞ −norm by FE 2.633E − 3 2.420E − 3 1.080E − 3 4.007E − 4
L∞ −norm by FEC 1.907E − 2 2.770E − 2 1.251E − 2 4.600E − 3
M ax u 9.800E − 1 8.100E − 1 3.600E − 1 1.353E − 1

Remark 4.2 Checking Tables 9 - 11, we see that the L∞ error norm generated
by applying the FEC to parabolic problems seem not optimal and this is not
surprising, infact the same observation was made for elliptic problems [5].

Remark 4.3 We would remind that the conservative approach [5] does not re-
spect the equilibrium between source term and mathematical (numerical) fluxes.
In order to obtain a method genuinely conservative, for instance like those de-
veloped in the finite volume element (FVE) framework, a generalization of the
techniques presented in [10] could be afforded.

5. Conclusions
The results obtained by the presented method and shown in the Tables 1 - 7
demonstrate the accuracy and the efficiency of it, in particular the reduction
of the most computationally expensive kernel to elliptic one and the use of a
preconditioned iterative solver (by an additive scalable Schwarz preconditioner)
make the approach followed very interesting also for the solution of very impor-
tant problems, such as the 2D incompressible Navier-Stokes and the 1D Shallow
Water ones without using the artificial compressibility approach [6].

9
 References
[1] A.N. Brooks & T.J.R. Hughes, Streamline-upwing/Petrov-Galerkin formu-
lations for convections dominated flows with particular emphasis on the in-
compressible Navier-Stokes equations, Comput. Methods Appl. Mech. Eng.,
32, (1982), 199-259.
[2] J. Douglas, Jr. & T.F. Russell, Numerical methods for convection-dominated
diffusion problems based on combining the method of characteristic with
finite element and finite difference procedures, SIAM J. Numer.Anal., 19,
(1982), 871-885.
[3] P. Hartman, Ordinary differential equations, John Wiley and Sons, Balti-
more, 1973.
[4] B.P. Leonard, A stable and accurate convective modeling procedure based
on quadratic upstream interpolation, Comput. Methods Appl. Mech. Eng.,
19, (1978), 59-98.
[5] A.C. Kengni Jotsa, A new method for elliptic problems in CFD resembling
the discontinuous FE approach, Proceedings MASCOT 2010 Special Session,
IMACS/ISGG, IMACS Series in Computational and Applied Mathematics,
ULPGC, Las Palmas de Gran Canaria, Spain, (2010), 219-228.
[6] A.C. Kengni Jotsa, Solution of the 2D Navier-Stokes equations by a new
FE fractional step method, PhD thesis, Università degli studi dell’ Insubria,
March 2012.
[7] O. Pironneau, On the transport-diffusion algorithm and its applications to
the Navier-Stokes equations, Numer. Math., 38, (1982), 309-332.
[8] O. Pironneau, Méthodes des Élements Finis pour les fluides, Masson, Paris,
1988.
[9] A. Priestley, Evaluating the spatial integral in the Lagrange-Galerkin
method in finite elements in fluids, New trends and Applications, K. Mor-
gan, E. Oñate, J. Périaux and J. Peraire eds., Pineridge Press, Swansea,
(1993), 146-155.
[10] A. Quarteroni & R. Ruiz.Baier, Analysis of a finite volume element method
for the Stokes problem, Numer. Math., Springer-Verlag, 2011.

10

View publication stats