1D Numerical Methods With Finite Volumes

Guillaume Riflet
MARETEC IST

1 The advection-diffusion equation

The original concept, applied to a property within a control volume V , from

which is derived the integral advection-diffusion equation, states as

{Rate of change in time} = {Ingoing − Outgoing fluxes}

(1)
+ {Created − Destroyed} .

Annotated in a correct mathematical encapsulation, equation 1 yields

∫ I
d Vt P dV
=− P (vr · n) dS
dt ∂Vt
I
− −K (∇P · n) dS
∂Vt
∫
+ (Sc − Sk) dV, (2)
Vt

where P is the transported property concentration, K is the diffusivity co-

efficient, ∂Vt is the surface of the moving control volume Vt , vr is the flow
velocity relative to the moving surface of the control volume Vt , n is the
outward normal to the control surface and Sc and Sk are the source and
sink terms, respectively. The first term on the right hand side (RHS) of
equation 2 states ingoing and outgoing fluxes due to advection, the second
term in the RHS states ingoing and outgoing fluxes due to diffusion, and the
last term on the RHS accounts for source and sink terms. Note that Fick’s
law, − K ∇P , is applied to mathematically describe diffusion [1].
By resorting to the Reynolds transport theorem [2], illustrated in figure 1,
stating that the derivative in time of a property in a given moving control
volume VA is equal to the derivative in time of the same property in another
given moving control volume VB , coincident in one time instant with VA ,
summed to the flow of the property through the control volumes given by

1
1 The advection-diffusion equation 2

Fig. 1: Illustration of control volumes VA , VB and VC near time instants

t = t0 + δt, t = t0 and t = t0 − δt, respectively, and their relative
deformation rate of change, represented by the vector fields vA|B and
vB|C plotted on the surfaces of volumes VB and VC , respectively. The
control volumes have different motions and are coincident at time
instant t = t0 .

their relative velocity vA|B , i.e.

∫ ∫ I
d d
P dV = P dV + P vA|B · n dS, (3)
dt VA dt VB ∂VB

and applying it to equation 2,

∫ I I
d V P dV ( )
+ P vVt |V · n dS = − P (vr · n) dS
dt ∂V ∂Vt
I
− −K (∇P · n) dS
∂Vt
∫
+ (Sc − Sk) dV, (4)
Vt

where V is a control volume held fixed in time relative the laboratory refer-
ence frame and vVt |V is the relative velocity of the moving control volume
Vt relative to the laboratory reference frame or to V , which is the same.
The flow velocity relative the laboratory reference frame v is then given
by
v = vr + vVt |V ,
and minding that ∂V = ∂Vt and V = Vt at one time instant, equation 4
1 The advection-diffusion equation 3

rewrites
∫ I
d P dV
V
=− P (v · n) dS
dt
I ∂V

− −K (∇P · n) dS
∂V ∫

+ (Sc − Sk) dV, (5)

V

By resorting to the divergence theorem, stating that the divergence of a

vector field inside any finite gaussian volume is equal to its flux through the
boundary of the volume, i.e.
∫ ∫
∇.E dV = E · n dS, (6)
V ∂V

being V the gaussian volume, ∂V its boundary, E the vector field, and E · n
its flux through the boundary. n is defined as the external normal unit
vector to the boundary. Hence, by applying equation(6) to equation(2), the
finite volume formulation of equation (7) is obtained:
∫ ∫
d V P dV
=− ∇ · (v P ) dV
dt
∫ V

− ∇ · (− K ∇P ) dV
V
∫
+ (Sc − Sk) dV. (7)
V

By resorting to the Leibniz integration rule,

∫ ∫
d d f (x, t)
f (x, t)dV = dV,
dt V V dt
which is true as long as the integral volume V (and consequently x) is held
fixed in time, in one hand. By joining all the integrals in V into a single
integral on the other, equation 7 yields
∫ { }
∂P
+ ∇ · (v P ) + ∇ · (− K ∇P ) − Sc + Sk dV = 0.
V ∂t
Since the above integral holds zero for all gaussian volume V , then its in-
tegrand must be zero, thus yielding the differential equation of advection-
diffusion:
∂P
+ ∇ · (v P ) + ∇ · (− K ∇P ) − Sc + Sk = 0. (8)
∂t
1 The advection-diffusion equation 4

A few comments: first, the control volume is held fixed in time. Second,
the time derivative applied to a function varying in time only in its explicit
component (such as P (x, t)), may be annotated as a partial derivative,
∂
∂t , without loss of generality. Third, the total derivative and the partial
derivative are related by the Leibniz chain rule,
 
d P (x(t), t)  ∂P (x(t0 ), t) 
dt  = ∂t 
t0 t0


d x(t)  ∂P (x(t0 ), t0 )
+
dt t0 ∂x

d y(t)  ∂P (x(t0 ), t0 )
+
dt t0 ∂y

d z(t)  ∂P (x(t0 ), t0 )
+ .
dt 
t0 ∂z

If dx
dt = 0, then the partial derivative equals the total derivative, which is
the case of P within the fixed volume V .
By noting that

∇ · (v P ) = v · ∇P + P ∇ · v,

equation 8 returns
∂P
+ v · ∇P + P ∇ · v + ∇ · (− K ∇P ) − Sc + Sk = 0.
∂t
By noting that the material derivative is defined by
D ∂
≡ + v · ∇,
Dt ∂t
equation 8 finally yields
DP
+ P ∇ · v + ∇ · (− K ∇P ) − Sc + Sk = 0.
Dt
In incompressible fluids we have ∇ · v = 0, hence in such case
DP
+ ∇ · (− K ∇P ) − Sc + Sk = 0. (9)
Dt
Thus, to sum up, equation (5) is the integral equation of advection and
diffusion (with source and sink terms) in flux form, equation (7) is the
1 The advection-diffusion equation 5

Fig. 2: Finite cartesian volume, under uniform flow v, defined by

its normal vectors n1 , n2 , n3 , n4 , n5 and n6 and respective faces
A1 , A2 , A3 , A4 , A5 and A6 .

integral equation of advection-diffusion in volume form, equation (8) is the

differential equation of advection-diffusion (where the advective field is the
velocity of the fluid particles relative to a fixed reference frame) and equation
(9) is the differential equation of advection-diffusion for an incompressible
fluid. Mind, however, that equation (2) is the flux form counterpart of
equation (5) for moving control volumes. In that case, the kinematics of the
moving control volume must also be given in order to solve the equation.
For example, consider the control volumes of a undimensional channel where
the free surface of the control volume moves in time according to the water
elevation...

1.1 Integrating in a cartesian finite volume

Let us integrate each term of equation (2) for each face of the cartesian
finite-volume of figure 2:
( )
v = (u, 0, 0) ∇P = ∂P ∂x , 0, 0 V = ∆x∆y∆z
n1 = (1, 0, 0) n2 = (−1, 0, 0) n3 = (0, 0, 1)
(10)
n4 = (0, 0, −1) n5 = (0, 1, 0) n6 = (0, −1, 0)
A1 = A2 = ∆y∆z A3 = A4 = ∆x∆y A5 = A6 = ∆x∆z

Only fluxes through faces defined by n1 and n2 are non-null:

v · n1 = u v · n2 = −u
∇P · n1 = ∂P
∂x ∇P · n2 = − ∂P
∂x .
1 The advection-diffusion equation 6

∫
P dV
Thus we get, if we define P ≡ V
V ,
∫ ( )
∂ VP dV ∂ PV
=
∂t ∂t
∂P
= V,
∂t
∫
P dS
furthermore, if we define P̃i ≡ Ai
Ai , we obtain

∫ 6 ∫
∑
P (v · n) dS = P (v · ni) dS
A i=1 Ai
∫ ∫
= u P dS − u P dS
A1 A2
= u P˜1 A1 − u P˜2 A2 ,

∫ ∂P
dS
finally, if we approximate Ai ∂x
Ai ≈ ∂P
∂x , we have

∫ 6 ∫
∑
−K (∇P · n) dS = −K (∇P · ni) dS
A i=1 Ai
∫ ∫
∂P ∂P
= −K dS + K dS
A1 ∂x A2 ∂x
∂P ∂P
= −K A1 + K A2 .
∂x (A1 ) ∂x (A2 )

Gathering up the terms above back into equation (2) yields

∂P ∂P ∂P
V = −u P˜1 A1 + u P˜2 A2 + K A1 − K A2 + Sc’ − Sk’
∂t ∂x (A1 ) ∂x (A2 )

∂P A1 ˜ A2 ˜ A1 ∂P A2 ∂P Sc’ Sk’
= −u P1 +u P2 +K −K + − . (11)
∂t V V V ∂x (A1 ) V ∂x (A2 ) V V
Now the whole of the game is to find numerical schemes that evaluate the
time and space derivatives, in equation (11), of the volume averaged P , ∂P
∂t
∂P
and ∂x properties; as well as to evaluate the surface averaged property P̃ .
1 The advection-diffusion equation 7

1.2 Discretizing in a cartesian 1D finite volume

The 1D finite volume is described in figure 2 and the 1D uniform mesh is
described in figure 4. Let it be defined that P ≡ Pi and that P˜1 ≡ Pi+1/2 ,
P˜2 ≡ Pi−1/2 .

1.2.1 Time derivatives

• Time forward or explicit method

∂P P (t + ∆t) − P (t)
(t) ≈ (12)
∂t ∆t
• Time backward or implicit method
∂P P (t) − P (t − ∆t)
(t) ≈ (13)
∂t ∆t
• Midpoint method
∂P P (t + ∆t) − P (t − ∆t)
(t) ≈ . (14)
∂t 2 ∆t
As a note, it can be shown ( [3], [4] ) that the forward and backward method
have first-order precision and that the midpoint method has second-order
precision, on an uniform grid.

1.2.2 Space derivatives

• Diffusion typically uses the midpoint method for the spatial derivative
∂P
∂x (see equation (14)).
{ ∂P Pi+1 −Pi
∂x (i+1/2) = ∆x
∂P Pi −Pi−1 ,
∂x (i−1/2) = ∆x

A1 ∂P A2 ∂P
−K +K
V ∂x (i+1/2) V ∂x (i−1/2)
( )
K Pi+1 − Pi Pi − Pi−1
=− −
∆x ∆x ∆x
K
=− (Pi+1 − 2 Pi + Pi−1 ) , (15)
∆x2
• Advection typically uses the upwind method (backward method if u <
0, and forward method if u ≥ 0),
2 Some notes on numerical methods 8

{
Pi , if u≥0
Pi+1/2 =
{ Pi+1 , if u<0
, (16)
Pi−1 , if u≥0
Pi−1/2 =
Pi , if u<0

A1 A2
u P −u P
V i+1/2 V i−1/2
= 0 Pi+1 + u
∆x Pi + − ∆x u
Pi−1 , if u ≥ 0
= u
Pi+1 + − ∆xu
Pi + 0 Pi−1 , if u < 0
∆x
u u u
− u + u
= − ∆x ∆x
Pi+1 + Pi − ∆x ∆x
Pi−1 , (17)
2 ∆x 2
• or the central i method (i.e. midpoint method),
Pi+1 +Pi
Pi+1/2 = 2
Pi +Pi−1 , (18)
Pi−1/2 = 2
A1 A2 u u
u Pi+1/2 − u Pi−1/2 = Pi+1 + 0 Pi − Pi−1 . (19)
V V 2 ∆x 2 ∆x

2 Some notes on numerical methods

This section aims at commenting on stability (robustness), positivity, mass

conservation, boundary conditions and initial conditions for the several types
of schemes reviewed in this work. We’ll only look at the fluxes through the
boundaries of the control volumes and consider null sources and sinks in
order to properly discuss mass conservation. But, in the first hand a quick
revision on finite-difference methods are suggested so as clearly derive the
approximation order of the forward, backward and mid-point discretization
of the derivative.

2.1 Finite-difference solvers

In order to solve numerically equations like equation (8) it is necessary to
discretize the derivative operators. After discretizing the derivative opera-
tors, it is also useful to have an estimate of the error of the solution. When
one knows its analytical solution, then it suffices to compute the RMS (root
mean square error) in order to know exactly what is the error. However,
often, numerical solvers are used because there are no known analytical solu-
tion to the PDE (partial differential equation), thus there is no alternative
2 Some notes on numerical methods 9

but to estimate the error. A traditional approach consists in using con-

stant discrete time steps dt and constant discrete spatial steps dx combined
with the Taylor serie expansion definition of analytical functions [4]. The
Taylor serie expansion of an analytical real function forward in its variable
coordinate is
∞
∑ (dt)n
f (t + dt) = f (n) (t) (20)
n!
n=0

and backward in its variable coordinate is

∞
∑ (−dt)n
f (t − dt) = f (n) (t) . (21)
n!
n=0

2.1.1 Explicit method

The Taylor series in equation (20) expanded up to the second order (with
third order error),

′ ′′ (dt)2 ( )
f (t + dt) = f (t) + f (t) dt + f (t) + o (dt)3 . (22)
2
A first derivative approximation may be obtained from equation 22,

′ f (t + dt) − f (t)
f (t) = + o (dt) . (23)
dt
The first derivative approximation in equation (23) yields first order error.
This numerical method is often referred to as the explicit method or the
forward in time method. Note that

o((dt)3 )
= o((dt)2 ),
dt
o(dt) = o(−dt) = −o(dt),
are general properties of truncature error operator.

2.1.2 Implicit method

Likewise, the Taylor series in equation (21) expanded up to the second order
(with third order error) yields,

′ (−dt)2
′′
( )
f (t − dt) = f (t) + f (t) (−dt) + f (t) + o (−dt)3 . (24)
2
2 Some notes on numerical methods 10

A first derivative approximation may be obtained from equation 24,

′ f (t) − f (t − dt)
f (t) = + o (dt) . (25)
dt
This numerical method is referred to as the implicit method or the backward
in time method and is also first order error.

2.1.3 Mid-point method

By subtracting equation (24) from equation (22), the mid-point method is
obtained:
′
( )
f (t + dt) − f (t − dt) = f (t) 2dt + o (dt)3 , (26)
′ f (t + dt) − f (t − dt) ( )
f (t) = + o (dt)2 . (27)
2dt
Equation (27) is referred to as the mid-point method and has a second order
error.

2.1.4 Second derivative discretization

By summing equation (24) from equation (22), a discretization of the second
derivative is obtained:
′′
( )
f (t + dt) + f (t − dt) = 2 f (t) + f (t) (dt)2 + o (dt)4 . (28)
Note that the third order terms in the forward and backward Taylor series
expansion cancel out. In fact, every odd order term are cancelled when
summing the forward and backward Taylor series expansion, whereas its the
even terms that are cancelled when subtracting the backward Taylor series
from the forward Taylor series. Thus, equation (28) yields
′′ f (t + dt) − 2 f (t) + f (t − dt) ( )
2
f (t) = + o (dt) . (29)
(dt)2

2.2 Explicit schemes

Unidimensional meshes yield, by discretizing equation (2) forward in time,
Pi (t + ∆t) = Ai Pi−1 (t) + (1 − Bi )Pi (t) + Ci Pi+1 (t). (30)
The adimensional coefficients Ai and Ci are associated with the ingoing
fluxes from neighbouring cells, while coefficient Bi is associated with the
outgoing fluxes to neighbouring cells. Physically, they represent the per-
centage of mass coming from, and going to, neighbouring cells and, as such,
they should be bounded between 0% and 100%.
2 Some notes on numerical methods 11

2.2.1 positivity
Indeed, if the coefficients affected to P were negative, that would violate the
principle of positive mass, or positivity, and the numerical solution would
return negative oscillations of the property concentration. Thus, to ensure
that positivity is not violated, it is required that:


 Ai ≥ 0,
∀i Bi ≥ 0, (31)


Ci ≥ 0.

2.2.2 stability and robustness

On the other hand, if the absolute value of the coefficients were above unity,
that would transport more mass than there is (credit is risky) and would
turn the scheme unstable. When instabilities occur, they are immediately
spotted, due to the disproportionate growth of the property concentration,
and the computer usually quickly returns an overflow error. Thus, to ensure
the stability of the method, it is required that:


 ∥Ai ∥ ≤ 1,
∀i ∥Bi ∥ ≤ 1, (32)


∥Ci ∥ ≤ 1.

Some numerical schemes can only ensure this condition for a limited range
of parameters. A numerical scheme robustness is measured by the range of
values that the parameters may take that ensure the stability condition in
equation 32.

2.2.3 mass conservation

Furthermore, all outward fluxes in one control volume should become in-
ward fluxes, in the same exact proportion, in the neighbouring volumes.
If the outward/inward symmetric fluxes aren’t exactly the same, then the
numerical scheme gains or looses mass over time, and is considered ”not con-
servative”. Thus, to ensure that the scheme is conservative, it is required
that:
∀i Ai+1 − Bi + Ci−1 = 0, (33)
as long as the B coefficients don’t contain any sources nor sink terms. Equa-
tion 33 states the following:
2 Some notes on numerical methods 12

What Pi+1 gains from Pi (Ai+1 ) and what Pi−1 gains from Pi
(Ci−1 ) is equal to what Pi looses (Bi ).

2.2.4 numerical diffusion

Once the continuous equation of advection-diffusion are discretized with
finite-difference solvers, errors are introduced. A particular expression of
these errors is that the finite-difference solver will always show a certain
degree of spurious diffusion along the advective direction, which is called
numerical diffusion. Numerical diffusion is maximum when the diffusive
term is zeroed and only advection occurs. The mecanism goes that any
particular tracer that doesn’t undergo diffusive processes, but is mainly ad-
vected by the tracer, should follow a steady path and the particule of tracer
should never grow in size nor dilute. However, whenever a finite-difference
solver for the advection equation is applied, the particulate tracer present
in a control volume is always split, at each time increment, between the
remaining part in the control volume and the outgoing part to the neigh-
bouring control volumes. This mechanism, which is illustrated in figure 3,
creates numerical diffusion and is the main reason modelers often try to
implement higher-order schemes that exhibit less numerical diffusion.

2.3 Implicit schemes

We can adapt the time forward equation into a time backward equation by
simply considering that all outgoing and ingoing fluxes occur at t + ∆t, thus
yielding
−Ai Pi−1 (t + ∆t) + (1 + Bi )Pi (t + ∆t) − Ci Pi+1 (t + ∆t) = Pi (t). (34)
This system of N equations, where N is the number of cells in the domain,
is harder to solve because we have N + 2 unknowns. Boundary conditions
give the other two missing equations. The time forward scheme was simpler
because it was a one equation to one unknown problem. The time backward
scheme describes a tridiagonal matrix system of the type
M x = T,
where M is the tridiagonal matrix, x is the unknown and T is the inde-
pendent term. Such systems can be solved resorting to the Gauss-Seidel
elimination method or, more efficiently, to the Thomas algorithm. The gain
is that the implicit method, applied to equation 2, is unconditionally stable,
although it can still violate positivity should the conditions of equation 31
not be met.
2 Some notes on numerical methods 13

Fig. 3: Advected boat by a uniform current along time (from top to bot-
tom). The dotted boat represents a realistic advection. The all-
black boat represents advection by an upwind time-forward scheme
in a unidimensional grid. The all-black boat falls in pieces, after each
time iteration, as a consequence of numerical diffusion. Thus, finite-
difference solvers aren’t adequate to model pure advection. However,
they work adequately to model advection and diffusion.
2 Some notes on numerical methods 14

2.3.1 Thomas algorithm

The Thomas algorithm is used to solve tridiagonal systems described by
ai xi−1 + bi xi + ci xi+1 = di ∀i ∈ {1, 2 ... n} ,
where x0 and xn+1 is the reference solution for the boundary conditions. In
matricial form, we can establish that
    
1 0 x0 d0
a1 b1 c1   x1   d1 
    
 a2 b2 c2   x2   d2 
    
 a3 b3 c3  ·  =  · ;
    
 · · ·  ·   · 
    
 an bn cn   xn   dn 
0 1 xn+1 dn+1
which suitably simplifies to
      ′
b1 c1 0 x1 d1 − a1 d0 d1
a2 b2 c2   x2   d   d′2 
    2   
 a3 b3 ·  ·  =  ·  =  · ,
      
 · · cn−1   ·   ·  ·
0 an bn xn dn − cn dn+1 d′n
after noting that d0 = x0 and dn+1 = xn+1 .
Thus, any non-trivial boundary conditions must go into the independent
term in d′1 and d′n . The algorithm follows two sweeps, one forward and one
backwards. First, the forward sweep:


 c1 ,
 i = 1,
Qi = b 1

 ci
 , i = 2, 3, . . . , n − 1 ;
bi − Qi−1 ai
 ′

 d
 1, i = 1,
Ri = b1
′
d − Ri−1 ai


 i , i = 2, 3, . . . , n;
bi − Qi−1 ai
which yields the following system:
    
1 Q1 0 x1 R1
 1 Q   x2   R 2 
 2    
 · ·  ·  =  · .
    
 1 Qn−1   ·   · 
0 1 xn Rn
2 Some notes on numerical methods 15

Then, the backward sweep is applied,

{
Ri , i=n
xi =
Ri − Qi xi+1 , i = n − 1, n − 2, . . . , 1 .

2.4 Boundary conditions

When solving the discretized equations (30) or (34) in for a finite domain
with i ∈ {1, 2, · · · , N }, necessity urges to impose a boundary condition for
P0 and PN +1 , as well as to establish the advective fluxes in coefficients A1
and CN . Imposed boundary conditions are also known as Dirichlet boundary
conditions and can be described as
P (t) = f (t),
at the boundaries, where f is an imposed solution for P . A particular case
is the null-value condition,
P (0) = 0 = P (N + 1).
Gradient dependent boundary conditions are also known as Neumann bound-
ary conditions and are best described by
∇P (t) = g(t),
at the boundaries, where g is an imposed solution for ∇P . A particular case
is the null-gradient condition,
P (0) = P (1)
P (N + 1) = P (N ).
In a closed system (with walls at the boundary), null-gradient is adequate to
simulate correctly diffusion, and nullyfying advection in coefficients A1 and
CN is adequate to simulate correctly advection. In an open system however,
the correct value of P must be known at the boundaries in order to correctly
simulate advection-diffusion. In a uniform steady flow, admitting that the
initial value of P is null outside the domain where the flow is incoming,
the null-value condition at the open boundaries allows to correctly simulate
advection, but diffusion will always be overestimated at the boundaries.
A more interesting way to look at the open-boundaries is to build a bal-
ance of the fluxes that cross the finite volume’s faces. Each flux gathers a
physical meaning, such as inwards advection or outwards diffusion. Depend-
ing on the modeler’s boundary conditions, the modeler must choose to keep
or to nullify outward fluxes, or to keep or nullify or impose inward fluxes.
3 The problem 16

2.5 Initial conditions

In order to completely define a problem, besides equation (2), we also need
to define its boundary conditions and its initial conditions. In the time
backward or time forward scheme, one initial condition suffices.

f (t + ∆t) − f (t)
f ′ (t) ≈
∆t
⇔ f (t + ∆t) ≈ f (t) + ∆t f ′ (t).

f (t + ∆t) − f (t)
f ′ (t + ∆t) ≈
∆t
′
⇔ f (t + ∆t) ∆t − f (t + ∆t) ≈ f (t) .

For both previous methods, in order to compute f (m ∆t) for any m ,

f ((m − 1) ∆t) is required. But with the midpoint method, two initial con-
ditions are required as,

f (t + 2 ∆t) − f (t)
f ′ (t + ∆t) ≈
2 ∆t
⇔ f (t + 2 ∆t) ≈ 2 ∆t f ′ (t + ∆t) + f (t) ,

thus, in order to compute f (m ∆t) for any m, f ((m − 2) ∆t) and f ((m − 1) ∆t)
are both required. Another issue with the midpoint method is that consec-
utive time steps may come decoupled if the two initial conditions are un-
similar. To re-enable coupling, use of filters such as Robert-Asselin filters
may be required [5], at the expense of some precision in the solution,

f ∗ (t) = f (t) + α (f (t + ∆t) − 2 f (t) + f (t − ∆t)) ,

where α typically ranges the 1%.

3 The problem

We want to model the transport of a sourceless and sinkless tracer in a

square-shaped duct and subjected to a uniform one-dimensional flow. The
advection-diffusion model of the tracer is mathematically described by equa-
tion (2). Also, we want to model the system using a finite volume approach.
The tracer is injected at initial time at some position inside the channel.
The boundary condition in this problem are two-fold: i) the duct is open at
both ends and inward and outward fluxes of the tracer are allowed; ii) the
4 The numerical models 17

Fig. 4: Unidimensional uniform mesh.

duct contains a grid at each end that filters the tracer while letting pass the
fluid. Boundary condition i is called open boundary condition and boundary
condition ii is equivalent to a fully closed boundary condition. The partic-
ularity of the open boundary condition is that we must impose a tracer
concentration of PbL (t) on the left end of the duct, and a tracer concentra-
tion of PbR (t) on the right end, be it Dirichelet or Neumann conditions. On
the other hand, the main interest of the closed boundary condition is to test
if the numerical methods conserve the tracer’s total mass inside the domain
as expected.

3.1 The mesh to the problem

Given that the fluid flow is uniform and along a closed square-shaped pipeline,
the best way to describe the geometry of the system is by cutting N uni-
form slices of volume area V , where the leftmost slice is indexed 1 and the
rightmost slice is indexed N . This sliced square-shaped pipeline defines our
mesh of uniform cells of length ∆x, width ∆y and height ∆z such that
∆x ∆y ∆z = V . A one-dimensional projection of the mesh is illustrated in
figure 4.

4 The numerical models

Considering u, ∆x and ∆t constants throughout the problem, let us define

Cr ≡ u∆x
∆t
, Dif ≡ K ∆t
∆x2
, CrL ≡ ∥Cr∥+Cr
2 , CrR ≡ ∥Cr∥−Cr
2 . As a note, Cr is
known in the literature as the Courant number and P ec ≡ ∥Cr∥ Dif is known as
the Péclet number. They are both adimensional numbers and are relevant
to characterize the positivity and stability of the numerical schemes. In
particular, it can be seen from inequation 31 that a Péclet number below or
equal to 2 is necessary to avoid negative oscillations in the central difference
4 The numerical models 18

scheme. {
Cr if u ≥ 0
CrL = ,
0 if u < 0
{
0 if u ≥ 0
CrR = .
−Cr if u < 0
Let us define the concentration of the tracer imposed at the boundaries by:
{
fL if Dirichelet
PbL = ,
P1 if Neumann
{
fR if Dirichelet
PbR = .
PN if Neumann
The stability and positivity criteria introduced below are the simultaneous
combination of inequations 31, 32 and 33 applied to this case.

4.1 Upwind explicit

Pi (t + ∆t) = (CrR + Dif ) Pi+1 (t)

+ (1 − CrR − CrL − 2 Dif ) Pi (t)
+ (CrL + Dif ) Pi−1 (t),

stability criteria:
0 ≤ ∥Cr∥ + 2 Dif ≤ 1.
closed boundary condition:

P1 (t + ∆t) = (CrR + Dif ) P2 (t)

+ (1 − CrR − CrL − 2 Dif ) P1 (t)
+ (CrL + Dif) PbL (t),

PN (t + ∆t) = (CrR + Dif) PbR (t)

+ (1 − CrR − CrL − 2 Dif ) PN (t)
+ (CrL + Dif ) PN −1 (t),

open boundary condition:

P1 (t + ∆t) = (CrR + Dif ) P2 (t)

+ (1 − CrR − CrL − 2 Dif ) P1 (t)
+ (CrL + Dif ) PbL (t),
4 The numerical models 19

PN (t + ∆t) = (CrR + Dif ) PbR (t)

+ (1 − CrR − CrL − 2 Dif ) PN (t)
+ (CrL + Dif ) PN −1 (t).

4.2 Central differences explicit

Pi (t + ∆t) = (−Cr/2 + Dif ) Pi+1 (t)

+ (1 − Cr/2 + Cr/2 − 2 Dif ) Pi (t)
+ (Cr/2 + Dif ) Pi−1 (t),

stability and positivity criteria:

0 ≤ ∥Cr∥ ≤ 2 Dif ≤ 1.

closed boundary condition:

P1 (t + ∆t) = (−Cr/2 + Dif ) P2 (t)

+ (1 − Cr/2 + Cr/2 − 2 Dif ) P1 (t)
+ (Cr/2 + Dif) PbL (t),

PN (t + ∆t) = (−Cr/2 + Dif) PbR (t)

+ (1 − Cr/2 + Cr/2 − 2 Dif ) PN (t)
+ (Cr/2 + Dif ) PN −1 (t),

open boundary condition:

P1 (t + ∆t) = (−Cr/2 + Dif ) P2 (t)

+ (1 − Cr/2 + Cr/2 − 2 Dif ) P1 (t)
+ (Cr/2 + Dif ) PbL (t),

PN (t + ∆t) = (−Cr/2 + Dif ) PbR (t)

+ (1 − Cr/2 + Cr/2 − 2 Dif ) PN (t)
+ (Cr/2 + Dif ) PN −1 (t),
4 The numerical models 20

4.3 Upwind implicit

(−CrR − Dif ) Pi+1 (t + ∆t)

+ (1 + CrR + CrL + 2 Dif ) Pi (t + ∆t)
+ (−CrL − Dif ) Pi−1 (t + ∆t)
= Pi (t),

unconditionally stable:

0 ≤ ∥Cr∥ + 2 Dif ≤ +∞.

closed boundary condition:

(−CrR − Dif ) P2 (t + ∆t)

+ (1 + CrR + CrL + 2 Dif ) P1 (t + ∆t)
= P1 (t) − (−CrL − Dif) PbL (t),

+ (1 + CrR + CrL + 2 Dif ) PN (t + ∆t)

+ (−CrL − Dif ) PN −1 (t + ∆t)
= PN (t) − (−CrR − Dif) PbR (t),

open boundary condition:

(−CrR − Dif ) P2 (t + ∆t)

+ (1 + CrR + CrL + 2 Dif ) P1 (t + ∆t)
= P1 (t) − (−CrL − Dif ) PbL (t),

(1 + CrR + CrL + 2 Dif ) PN (t + ∆t)

+ (−CrL − Dif ) PN −1 (t + ∆t)
= PN (t) − (−CrR − Dif ) PbR (t),

4.4 Central differences implicit

(Cr/2 − Dif ) Pi+1 (t + ∆t)

+ (1 + Cr/2 − Cr/2 + 2 Dif ) Pi (t + ∆t)
+ (−Cr/2 − Dif ) Pi−1 (t + ∆t)
= Pi (t),
4 The numerical models 21

positivity criteria:
0 ≤ ∥Cr∥ ≤ 2 Dif ≤ +∞.
closed boundary condition:
(Cr/2 − Dif ) P2 (t + ∆t)
+ (1 + Cr/2 − Cr/2 + 2 Dif ) P1 (t + ∆t)
= P1 (t) − (−Cr/2 − Dif) PbL (t),

+ (1 + Cr/2 − Cr/2 + 2 Dif ) PN (t + ∆t)

+ (−Cr/2 − Dif ) PN −1 (t + ∆t)
= PN (t) − (Cr/2 − Dif) PbR (t),
open boundary condition:
(Cr/2 − Dif ) P2 (t + ∆t)
+ (1 + Cr/2 − Cr/2 + 2 Dif ) P1 (t + ∆t)
= P1 (t) − (−Cr/2 − Dif ) PbL (t),

(1 + Cr/2 − Cr/2 + 2 Dif ) PN (t + ∆t)

+ (−Cr/2 − Dif ) PN −1 (t + ∆t)
= PN (t) − (Cr/2 − Dif ) PbR (t).

4.5 Hybrid schemes

{
1 upwind
α= ,
0 central differences
{
0 explicit
β= ,
1 implicit
{
β = 0.5 i.e. hybrid
Crank-Nicholson : ,
α = 0 i.e. central differences

β ((1 − α) Cr/2 − αCrR − Dif ) Pi+1 (t + ∆t)

+ (1 + β (α (CrR + CrL) + 2 Dif )) Pi (t + ∆t)
+ β (− (1 − α) Cr/2 − αCrL − Dif ) Pi−1 (t + ∆t)
=
(1 − β) (− (1 − α) Cr/2 + αCrR + Dif ) Pi+1 (t)
( ( ))
+ 1 − (1 − β) α (CrR + CrL) + 2 Dif Pi (t)
+ (1 − β) ((1 − α) Cr/2 + αCrL + Dif ) Pi−1 (t),
4 The numerical models 22

stability and positivity criteria:


 ≥0
∀α,β∈[0 1] , − (1 − α) ∥Cr∥ + 2 Dif ≤ β2 ,

≤ − ∥Cr∥ + 1
1−β

closed boundary condition:

β ((1 − α) Cr/2 − αCrR − Dif ) P2 (t + ∆t)

+ (1 + β ((1 − α) (Cr/2 − Cr/2) + α (CrR + CrL) + 2 Dif )) P1 (t + ∆t)
=
(1 − β) (− (1 − α) Cr/2 + αCrR + Dif ) P2 (t)
( ( ))
+ 1 − (1 − β) (1 − α) (−Cr/2 + Cr/2) + α (CrR + CrL) + 2 Dif P1 (t)
+ (1 − β) ((1 − α) Cr/2 + αCrL + Dif) PbL (t)
− β (− (1 − α) Cr/2 − αCrL − Dif) PbL (t),

(1 + β ((1 − α) (Cr/2 − Cr/2) + α (CrR + CrL) + 2 Dif )) PN (t + ∆t)

+ β (− (1 − α) Cr/2 − αCrL − Dif ) PN −1 (t + ∆t)
=
− β ((1 − α) Cr/2 − αCrR − Dif) PbR (t)
+ (1 − β) (− (1 − α) Cr/2 + αCrR + Dif) PbR (t)
( ( ))
+ 1 − (1 − β) (1 − α) (−Cr/2 + Cr/2) + α (CrR + CrL) + 2 Dif PN (t)
+ (1 − β) ((1 − α) Cr/2 + αCrL + Dif ) PN −1 (t),

open boundary condition:

β ((1 − α) Cr/2 − αCrR − Dif ) P2 (t + ∆t)

+ (1 + β (α (CrR + CrL) + 2 Dif )) P1 (t + ∆t)
=
(1 − β) (− (1 − α) Cr/2 + αCrR + Dif ) P2 (t)
( ( ))
+ 1 − (1 − β) α (CrR + CrL) + 2 Dif P1 (t)
+ (1 − β) ((1 − α) Cr/2 + αCrL + Dif ) PbL (t)
− β (− (1 − α) Cr/2 − αCrL − Dif ) PbL (t),
5 Adding sources and sinks to build an ecological model 23

(1 + β (α (CrR + CrL) + 2 Dif )) PN (t + ∆t)

+ β (− (1 − α) Cr/2 − αCrL − Dif ) PN +1 (t + ∆t)
=
− β ((1 − α) Cr/2 − αCrR − Dif ) PbR (t)
(1 − β) (− (1 − α) Cr/2 + αCrR + Dif ) PbR (t)
( ( ))
+ 1 − (1 − β) α (CrR + CrL) + 2 Dif PN (t)
+ (1 − β) ((1 − α) Cr/2 + αCrL + Dif ) PN +1 (t).

4.6 Characterization of the numerical methods

The implicit schemes are the most robust as they are unconditionally stable,
but they also produce the most numerical diffusion. The explicit schemes are
the less robust. The central differences method may, additionally, violate
positivity (either with the explicit, either with the implicit schemes) but
has a higher precision (second-order) and produces less numerical diffusion.
The upwind method never violates positivity, but shows a lot of numerical
diffusion and has less precision than the central differences. The hybrid
method is more robust than the explicit one, is more resilient to positivity
violation than central differences, and produces less numerical diffusion than
the implicit methods.

5 Adding sources and sinks to build an ecological model

The previous section limited to describe numerical methods suitable for ad-
vective and diffusive processes only. However, generic tracers may also have
source/sink and/or growth/decay terms. Besides involving more terms in
the equation, the stability conditions of the numerical scheme also change.
This section proposes to describe mathematically a mass conservative eco-
logical system composed by three variables of state describing a predador-
prey-nutrients system, that are transported in a 1D-channel. The stability
conditions of the numerical implementation are also considered.

5.1 the mathematical model of a prey-predator-nutrients system

The mathematical model will be described in two parts. The first part will
contain only the ecological modeling terms, acting in a single control-volume
(without any advective or diffusive process). The second part will add the
advective and diffusive terms to the equations.
5 Adding sources and sinks to build an ecological model 24

The equations describing the ecological model composed by three vari-

ables, predator, prey and nutrients (let them be zooplankton, phytoplankton
and nutrients) evolving in a single control-volume devoid of any advective
or diffusive processes are



dPzoo
= eh kh Pphy Pzoo − kmz Pzoo ,


dt



dPphy
= kg Pnut Pphy − kh Pzoo Pphy , (35)


dt




 dPnut
dt = −kg Pphy Pnut + (1 − eh ) kh Pphy Pzoo + kmz Pzoo ,

where Pzoo , Pphy and Pnut are, respectively, the concentrations of zooplank-
ton, phytoplankton and nutrients. kh is the assimilation rate of phytoplank-
ton by zooplankton, kg is the grazing rate of phytoplankton and kmz is the
mortality rate of zookplankton. eh is the zooplankton’s efficiency rate of
assimilation of phytoplankton (phytoplankton grazes with 100 percent of
efficiency). All the remains of dead zooplankton and dead phytoplankton
are wholly decomposed into nutrients by bacteria. Growth terms come with
a plus sign, and depletion terms come with a minus sign. The system of
equations 35 is globally conservative, as the sum of the initial masses of the
three properties is preserved over time.
By substituting the ecological model in equations 35 in the source and
sink terms of equation 8, the full ecological model system of differential
equations writes

 ∂Pzoo
 ∂t = −∇ · (v Pzoo ) + ∇ · (Kzoo ∇Pzoo ) + λzoo Pzoo ,
∂Pphy
= −∇ · (v Pphy ) + ∇ · (Kphy ∇Pphy ) + λphy Pphy , (36)
 ∂t
 ∂Pnut
∂t = −∇ · (v Pnut ) + ∇ · (Knut ∇Pnut ) − λnut Pnut + Scnut ,

where 

 λzoo ≡ eh kh Pphy − kmz ,


λ ≡ kg Pnut − kh Pzoo ,
phy
(37)

 λ ≡ kg Pphy ,


nut
Sc ≡ ((1 − eh ) kh Pphy + kmz ) Pzoo .
nut

In particular, λzoo and λphy are growth or decay coefficients of the zooplank-
ton and phytoplankton, respectively. Their sign can change over time. λnut
is always positive and is a decay coefficient in the nutrients equation. Fi-
nally, Scnut is a source term in the nutrients equation. Growth and source
5 Adding sources and sinks to build an ecological model 25

terms tend to stabilize the numerical solution of equations 36, using the ex-
plicit method, whereas decay terms may reduce the stability interval when
using the explicit method.

5.2 The discretization of ecological properties

Discretizing equations 36 follows the same approach that led to the dis-
cretized equations 30, except that this time growth/decay and source/sink
terms appear in the equations i.e.

Pi (t + ∆t) = Ai Pi−1 (t) + (1 − Bi + ∆t λi )Pi (t) + Ci Pi+1 (t) (38)

+∆t (Sci − Ski ) .

To meet the positivity (eq. 31) and stability (eq. 32) criteria, the following
constraints must be followed for equations 38,


 0 ≤ Ai ≤ 1,
∀i 0 ≤ Bi − ∆t λi ≤ 1, (39)


0 ≤ Ci ≤ 1.

However, even if the constraints in equations 39 are met, the numerical

solution may still undershoot (i.e. yield a negative concentration) if the sink
term is predominant over the source term, i.e.

If ∃i, Sci < Ski then the model may undershoot. (40)

In the ecological model presented in equations 36 and 37 there are no sink

terms. Thus, the model will never undershoot on account of criteria given
in equation 40. However the stability criteria provided in equations 39 still
require to be met in order to obtain a stable solution with the explicit
method. The implicit method still remains unconditionally stable (though
non-positive schemes, like the central-differences method, may yield too
much spurious negative oscillations).
The implicit method is straightforward from equation 38, by analogy
with equations 30 and 34:

−Ai Pi−1 (t + ∆t) + (1 + Bi − ∆t λi )Pi (t + ∆t) − Ci Pi+1 (t + ∆t) (41)

= Pi (t) + ∆t (Sci − Ski ) .
5 Adding sources and sinks to build an ecological model 26

References

[1] Chapra, S. C. Surface water-quality modeling. Mcgraw-hill Series In

Water Resources And Environmental Engineering (1997).

[2] Collado, F. J. Reynolds transport theorem for a two-phase flow.

Applied Physics Letters 90, 2 (2007), 024101+.

[3] Ferziger, J. H., and Perić, M. Computational methods for fluid

dynamics. Springer, Berlin, 1999.

[4] Fletcher, C. A. J., and Srinivas, K.

Computational Techniques for Fluid Dynamics 1. Springer, 1991.

[5] Kantha, L. H., and Clayson, C. A.

Numerical models of oceans and oceanic processes. Academic Press,
2000.