Fuelcells
Fuelcells
This research project was supported by the Fuel Cell Research Alliance Baden-Wurttemberg (FABZ), by the ZSW Ulm and by the University of Karlsruhe
1
Foreword
From May 2004 to February 2005 the Fuel Cell Research Alliance Baden-Wurttemberg
(FABZ) granted to the Computer Center of the University of Karlsruhe the financial means
to implement the PDEs (partial differential equations) for the PEMFCs and SOFCs into the
FDEM program package. The PDEs for the PEMFCs should be delivered by the ZSW and
for the SOFCs by the IWE. The special point was that such strongly nonlinear systems of
PDEs, and in the case of SOFCs also on coupled domains, should be solved numerically
whereas for the first time together with the solution should be computed a reliable error
estimate, so that the engineer could trust the numerical solution. This allows to separate
in the modeling process of FCs the model errors from the discretization errors. If there is
no error estimate and there is a discrepancy between experiment and numerical results, one
does not know if the difference comes from inaccuracies of the model or of the numerical
solution method. Therefore an error estimate is a valuable new feature and an essential
advantage. This is below confirmed by the results.
The first thing we all had to learn in this cooperation project was to speak to each other.
This means that the technical engineers learned what the numerical engineers wanted and
the numerical engineers learned what the technical engineers wanted. A special problem
are always the boundary conditions and in the case of SOFCs also the coupling conditions
between the solution for gas channel and for anode. Then the technical engineers could see
from the numerical solution the quality of the used model for their FCs and they then got
the hints how to improve the model. As they got together with the numerical solution an
error estimate they had not to doubt the quality of the delivered numbers. Finally, we had
for the PEMFC and for the SOFC a useable model. If the engineer has such a model, he
can play in the computer with the many parameters that are included in such a model,
can adapt them to new measurements and can, above all with the geometrical parameters,
optimize his FC.
Because the financial means for this project were rather limited and the project time
was very short for such a difficult problem, the project ended with the implementation of
the PDEs and some smaller variations of parameters. Ultimately, this project was intended
as a transitional task until in a larger project with industrial background the optimization of
FCs on a broad basis should start. So the purpose to show that the PDEs of FCs of different
types could be efficiently solved numerically on modern parallel supercomputers, with an
error estimate for the generated solution, was fully attained. If industrial partners want to
cooperate and use the invaluable experience of the project partners they should address the
corresponding authors of this report.
This report is written by three different groups, the ZSW, the IWE and the RZ. To avoid
lengthy coordinations, each group has written its part independently in its own preferred
style. So this report consists of three parts: Part I by ZSW, Part II by IWE, Part III by RZ.
Because of the independent writing the pages of the different parts are numbered by e.g.
I.1, I.2 etc. or II.1, II.2 etc.
Above all in Part II and III there are many coloured figures. In the printed version they
are in gray scale and much information is lost. Therefore we recommend to look at these
figures at the screen of a computer in the on-line version of the paper. This is in
http://www.rz.uni-karlsruhe.de/rz/docs/FDEM/Literatur/fuelcells.pdf
3
Table of Contents
Part I: The Polymer-Electrolyte-Membrane Fuel Cell (PEMFC)
The model and the PDEs
Model with Ficks diffusion transport mechanism
Model with MTPM as transport mechanism
Discussion of results
Conclusion
Bibliography
I.2
I.3
I.8
I.13
I.16
I.17
II.1
II.1
II.2
II.7
II.13
II.13
II.17
II.17
III.1
III.3
III.30
III.65
III.66
Part I:
The Polymer-Electrolyte-Membrane Fuel
Cell (PEMFC)
The aim at PEMFC operation and development is a nearly homogenous current density distribution
and thus power distribution all over the electrochemical active cell area. In a real fuel cell the
electrochemical reaction takes place inhomogeneous over the active cell area. The reaction runs
best, where the best reaction conditions exist, expecially the highest educt concentrations and the
best electrical conductivities for electron and proton conduction. Important mechanisms resulting
in current density inhomogenities are material transport limitations, which appear especially in the
gas diffusion layers (GDLs) of PEMFCs. Material transport limitations in the GDL result beside the
transport resistance caused by the mass transport through the GDL thickness mainly from geometry
effects as e.g. the alternation of gas channels and ribs of the flowfield structure. Therefore in order
to study the behaviour of a fuel cell it is essential to consider multidimensional effects and simulate
multidimensional models.
r/2
rib
channel
c/2
p = const.
n=0
n=0
j
t GDL
GDL
reaction layer
UR
nj = 0
R
n = f(i(x))
UC
source
chosen
chosen
chosen
I.3
Within this project two models for the gas phase transport are compared: As first approach Ficks
diffusion was used as gas transport mechanism, cause of its simple model. Afterwards the so called
Mean Transport Pore Model (MTPM) was used as gas phase transport mechanism, which considers
molecular multicomponent diffusion according to Stefan Maxwell, pore diffusion according to Knudsen, convection according to Darcy and wall slip. The comparison of both models results should help
to judge, whether the simple Ficks law is sufficient to model the gas transport in a porous PEMFC
GDL or the more detailed MTPM has to be used.
Below the models governing equations and boundary conditions for the two transport mechanisms
are listed.
no
.y
no
.x
nw
.y
nw
material balances:
(x) po
RT t
(x) pw
RT t
(x) Do
RT
(x) Do
=
RT
(x) Dw
=
RT
(x) Dw
=
RT
=
po
x
po
y
pw
x
pw
y
.x
(1)
(2)
(3)
(4)
.y
(no ) (no )
x
y
.x
.y
(nw ) (nw )
=
x
y
=
.x
.y
.x
(5)
(6)
.y
One gets 6 PDEs for the dependent variables: no , no , nw , nw , po and pw . In addition the current
density i, which is used in Eq. 8 and in the boundary conditions Eqs. 16, 17, 47 and 48, is considered
as a variable just at the boundary GDLreaction layer.
The independent variables are x, y and t. The material balances are set in their general form. The
calculations were done for the steady state case.
The dependent variables are listed in Tab. 2.
In PEMFCs the GDL is compressed under the ribs. In order to consider this effect and its impact
on the porosity a location dependent porosity is used:
h
i
h
i
2
2
exp a cos( c+r
x) + rc exp a cos( c+r
x)
h
i
h
i
(x) = 0.5 0 kf + 1 + (1 kf )
(7)
2
c
2
exp a cos( c+r x) + r + exp a cos( c+r x)
The origin of the x coordinate is in the middle of the channel, as shown in Fig. 1.
The porosity via the x-location for the models standard parameters is shown in Fig. 2.
The parameters needed for the governing equations are listed in Tab. 3.
I.4
Ficks law
unit
mol m2 s1
mol m2 s1
mol m2 s1
mol m2 s1
Pa
Pa
A m2
[]
4
source
chosen
[3]
[3]
chosen
Eq. 7
chosen
chosen
chosen
I.5
0.75
0.7
porosity []
0.65
0.6
0.55
0.5
0.45
0
0.2
0.4
0.6
0.8
1
3
x 10
xlocation [m]
i = f v i0
pD
o
pref
o
exp
nF U0 UZ
dmem
mem i
RT
(8)
pD
o is the value of the oxygen partial pressure at the reaction layer (Index D). In Eq. 8 the ohmic
resistance of the membrane is already considered and it is assumed, that the membrane conductivity
mem stays constant, according to a homogenous membrane humidification.
In addition the water vapour saturation pressure pSw in [Pa] according to the Antoine equation is
needed for the calculation of the boundary values:
log(102 pSw ) = A
Eq. 9 is a numerical equation.
B
C + (T 273.15)
(9)
= pSw
C
pU
o
ch
= (p
(10)
C
pU
w )
UC
yo,a
(11)
I.6
no
= 0
(12)
. y,U R
nw
= 0
(13)
no
= 0
(14)
. x,L/R
nw
= 0
(15)
no
. y,D
nw
1
i
4F
1
= i
2F
(16)
(17)
The negative sign in Eq. 17 marks, that the oxygen and the water vapour have to be transported in
opposite directions.
The parameter needed for the boundary conditions are listed in Tab. 4.
[]
the channel
UC
Mole fraction of O2 in
yo,a
[]
air
Exchange current deni0
[A m2
Pt ]
sity
Transfer coefficient
[]
Number of electrons
n
[]
involved in electrode
reaction (Faraday)
Reference pressure
pref
[Pa]
o
2
Surface extension facfv
[mPt m2 ]
tor at 0.15 mgPt cm2
Reaction order
[]
Open circuit voltage
U0
[V]
Cell voltage
UZ
[V]
Membrane thickness
dmem
[m]
Membrane
electric
mem
[S m1 ]
conductivity
Water vapour saturapSw
[Pa]
tion pressure
A
[]
Antoine parameters
B
[]
C
[]
I.7
conditions
value
1.013 105
source
chosen
0.8
chosen
0.21
6.7 105
[2]
0.2695
4
[2]
1.013 105
79
[2]
[2]
0.75
1.0
0.6
30.0 106
10.0
[2]
chosen
chosen
chosen
[6]
calculated
Eq. 9
8.0732991
1656.39
226.86
[6]
[6]
[6]
I.8
.x
no
transport equations:
.x
.x
.x
.x
pw no po nw
pn no po nn
+
+
DKno
(Dow p)
(Don p)
.y
.y
.y
.y
.y
no
pw no po nw
pn no po nn
+
+
DKno
(Dow p)
(Don p)
.x
.x
.x
.x
.y
.y
.y
.x
.x
.x
.y
.y
.y
.x
nn
po nn pn no
pw nn pn nw
+
+
DKnn
(Dno p)
(Dnw p)
.y
.y
nw
po nw pw no
pn nw pw nn
+
+
DKnw
(Dwo p)
(Dwn p)
.x
.x
nw
po nw pw no
pn nw pw nn
+
+
DKnw
(Dwo p)
(Dwn p)
.y
.y
nn
po nn pn no
pw nn pn nw
+
+
DKnn
(Dno p)
(Dnw p)
Bw
1 po
1
p
Bo
Bo
+
po
+
pw 1
RT x
RT p x
DKno
(Dow p)
Bo
Bo
Bn
po p
+
pn 1
(18)
(Don p)
Bo
RT p x
Bo
1 po
1
p
Bo
Bw
+
po
+
pw 1
RT y
RT p y
DKno
(Dow p)
Bo
Bo
Bn
po p
+
pn 1
(19)
(Don p)
Bo
RT p y
Bw
1 pw
1
p
Bw
Bo
+
pw
+
po 1
RT x
RT p x
DKnw
(Dwo p)
Bw
Bw
Bn
pw p
+
pn 1
(20)
(Dwn p)
Bw
RT p x
1 pw
1
p
Bw
Bw
Bo
+
pw
+
po 1
RT y
RT p y
DKnw
(Dwo p)
Bw
Bw
Bn
pw p
+
pn 1
(21)
(Dwn p)
Bw
RT p y
1 pn
1
p
Bn
Bn
Bo
+
pn
+
po 1
RT x
RT p x
DKnn
(Dno p)
Bn
Bw
Bn
pn p
+
pw 1
(22)
(Dnw p)
Bn
RT p x
Bn
1 pn
1
p
Bn
Bo
+
pn
+
po 1
RT y
RT p y
DKnn
(Dno p)
Bn
Bn
Bw
pn p
+
pw 1
(23)
(Dnw p)
Bn
RT p y
material balances:
(x) po
RT t
(x) pw
RT t
(x) pn
RT t
condition for the total pressure:
.x
.y
(no ) (no )
x
y
.x
.y
(nw ) (nw )
=
x
y
.x
.y
(nn ) (nn )
=
x
y
=
(24)
(25)
(26)
p = p o + pw + pn
(27)
.x
.y
.x
.y
.x
Finally one gets 9 PDEs and one algebraic equation for the dependent variables no , no , nw , nw , nn ,
.y
nn , po , pw , pn and p. In addition, as in the previous model, the current density i, which is used in
I.9
Eq. 8 and in the boundary conditions Eqs. 16, 17, 47 and 48, is considered as a variable just at the
boundary to the reaction layer.
The comparison of Eqs. 1 to 4 with Eqs. 18 to 23 shows, that the model with the MTPM as transport
mechanism is more elaborate as the one with Ficks law.
The dependent variables for the MTPM are listed in Tab. 5
For the transport equations the following partially variable dependend parameters are needed. All
new parameters within the MTPM are listed in Table 6.
The binary diffusion coefficients Djk in the continuum in [m2 s1 ] are ([5]):
q
M +M
0.01013 T 1.75 1 103 Mjj Mkk
i
h 1
Djk =
j, k = o, w, n
1
p V dj3 + V dk3
(28)
(x)
Djk
j, k = o, w, n
(29)
(Djk
q
M +M
1.75
0.01013
T
1 103 Mjj Mkk
(x)
i
h 1
p) =
1
V dj3 + V dk3
j, k = o, w, n
(30)
hr2 i (x)
j + Kj
p
+
1 + Kj
8
j = o, w, n
(31)
(32)
k=1
PK
k=1 pk k Mk
PK
k=1 pk Mk
(33)
j = o, w, n
(34)
I.10
the MTPM
unit
mol m2 s1
mol m2 s1
mol m2 s1
mol m2 s1
mol m2 s1
mol m2 s1
Pa
Pa
Pa
Pa
A m2
I.11
The dimensionless Knudsen numbers Kj are achieved by dividing the mean free path lengths j by
the mean pore diameter 2 hri:
j
j = o, w, n
(35)
Kj =
2hri
The mean free path lenghts j in [m] are calculated according to the kinetic theory of gases:
s
3RT
j = o, w, n
j = j
p j Mj p
(36)
Boundary conditions
Below the boundary conditions for the MTPM are listed in short form. The current density i is
calculated with Eq. 8; the water vapour saturation pressure pSw is calculated with Eq. 9.
channel boundary (UC):
In the channel a constant pressure at air composition and given relative humidity is assumed:
pU C
= pch
(37)
C
pU
w
C
pU
o
C
UC
= (pch pU
w ) yo,a
C
pU
n
pSw
= (p
ch
(38)
C
pU
w )
(1
(39)
UC
yo,a
)
(40)
no
. y,U R
nw
. y,U R
nn
= 0
(41)
= 0
(42)
= 0
(43)
= 0
(44)
= 0
(45)
= 0
(46)
no
. x,L/R
nw
. x,L/R
nn
no
. y,D
nw
. y,D
nn
1
i
4F
1
= i
2F
= 0
(47)
(48)
(49)
The parameters for the MTPM boundary conditions are the same as listed for the Ficks law boundary
conditions.
Dynamic viscosity of
o
O2
Dynamic viscosity of
w
H2 O vapour
Dynamic viscosity of
n
N2
Molecular weight of O2
Mo
Molecular weight of
Mw
H2 O
Molecular weight of N2
Mn
Knudsen diffusion coDKnj
efficients
slip factor
Knudsen numbers
Kj
Mean free path lengths
j
Mean transport pore
hri
radius
Pore radii distribution
hr2 i
I.12
16.6
12.7
[5] Da 35
[5] Da 35
numerical eq.
[m2 s1 ]
17.9
calculated
[5] Da 35
Eq. 31
[]
calculated
Eq. 32
[Pa s]
[Pa s]
calculated
22.53 106 at 60 C
Eq. 33
[5] Db 71
[Pa s]
[5] Db 8
[Pa s]
19.396 106 at 60 C
[5] Db 37
[kg mol1 ]
[kg mol1 ]
2 16.0 103
(2 1.008 + 16.0) 103
[kg mol1 ]
[m2 s1 ]
2 14.01 103
calculated
[]
[]
[m]
[m]
[m2 ]
calculated
calculated
38.5 106
1.826 1010
Eq. 34
[4]
Eq. 35
Eq. 36
measured
measured
I.14
1.7
x 10
Fick
MTPM
1.6
1.5
1.4
1.3
1.2
1.1
1
0.9
0
0.2
0.4
0.6
xlocation at the reaction layer [m]
0.8
1
3
x 10
Fig. 3: Oxygen partial pressure at the reaction layer for standard conditions
4
x 10
Fick
MTPM
2.8
2.6
2.4
2.2
2
1.8
1.6
0
0.2
0.4
0.6
xlocation at the reaction layer [m]
0.8
1
3
x 10
Fig. 4: Water vapour partial pressure at the reaction layer for standard conditions
(saturation pressure at 60 C is approximately 2104 Pa)
I.15
3300
Fick
MTPM
3200
current density [A m ]
3100
3000
2900
2800
2700
2600
2500
2400
2300
0
0.2
0.4
0.6
xlocation at the reaction layer [m]
0.8
1
3
x 10
0.95
Fick
MTPM
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0
2000
4000
6000
2
current density [A m ]
8000
10000
Conclusion
A two dimensional PEMFC model was examined and the gas transport was modelled with two
material transport mechanisms: the simple Ficks diffusion law and the more detailed MTPM.
As one result the maximum deviations of the calculations with both models were found to be in the
range of 10 to 15 %. This has to be considered, when using the simple Ficks law instead of the
MTPM. Thus for qualitative calculations the simple Ficks law could be sufficient. For more precisely
calculations the MTPM has to be used or the Ficks law has to be corrected in an adequate way.
A further result of the calculation within this project is, that at the low operation temperature of
60 C already at a cell potential of 0.6 V the water vapour saturation pressure could be exceeded
beneath the flowfield ribs and thus liquid water could limit the material transport under the ribs and
therefore two phase effects should be considered in model refinement.
Bibliography
[1] D. Arnost, P. Schneider: Dynamic transport of multicomponent mixtures of gases in porous
solids. The Chemical Engineering Journal, 57, pp 91-99, 1995
[2] H. A. Gasteiger, W. Gu, R. Makharia, M. F. Mathias, B. Sompalli: Beginning-of-life MEA
performance Efficiency loss contributions. Handbook of Fuel Cells, Part 3, Chapter 46, pp
593-610, 2003
[3] M. Jakubith: Memofix Chemie und Chemietechnik.VCH, 1992
[4] F. Thom: diploma thesis. Forschungszentrum J
ulich GmbH, 1998
[5] VDI-W
armeatlas 8. Auflage 1997
[6] M. W
ohr: Instation
ares thermodynamisches Verhalten der Polymermembran-Brennstoffzelle.
phD thesis, Universit
at Stuttgart, 1999
17
Part II
A 2D-Model of a SOFC-Anode operated
with methane
The SOFC
II.1
dA
anode
dK
gas channel
y
x
rib
lK
Gas flow
Fig. 1. Sketch: Calculation region
The SOFC
II.2
The structural and the geometry parameters according to fig.1 are given for the base case in
tab.1
Tab. 1. Structural and geometry parameters (base case)
parameter
thickness of the
anode
height of the
anode gas channel
length of the
anode gas channel
symbol
dA
unit
m
value
5010-6
source
IWE
dK
110-3
IWE
lK
3510-3
IWE
Species balances
( pK Yi , K )
Di , gas
( pK u x , K Yi , K ) ( pK u y , K Yi , K )
y
=0
+
+
x
y
y
(1)-(4)
Equation of continuity
( K ux,K )
x
( K u y,K )
y
=0
(5)
Navier-Stokes equations
The SOFC
II.3
u x , K
ux,K
+ u y,K
u x , K
pK u x , K 2 u x , K u y , K
+ 2
+
=
y
x x
x
y
3 x
u x , K u y , K
+
+
x
y y
(6)
u y , K
ux,K
+ u y ,K
u y , K
pK u y , K 2 u x , K u y , K u x , K u y , K
+ 2
+
+
=
+
y
y y
y
y x y
x
3 x
(7)
Daltons law
i,K
=1
(8)
One obtains 7 PDEs and one algebraic equation for the 8 dependent variables in the anode gas
channel: pK, ux,K, uy,K, YCH 4 , K , YCO , K , YH 2 , K , YCO2 , K and YH 2O , K . The independent variables are x
and y.
The Diffusion coefficient Di,gas of component i in the gas mixture is calculated by the Wilke
approach [2]:
Di , gas =
1 Yi
(Y j / Dij )
(9)
j i
The binary diffusion coefficients Dij are obtained by using the theory of Chapman and
Enskong (see tab 2).
Tab. 2. Binary Diffusion coefficients Dij/ [m2/s], 950 oC [3]
CH4
CO
H2
CO2
H2O
CH4
2,34e-4
7,46e-4
1,95e-4
2,91e-4
CO
2,34e-4
8,05e-2
1,75e-3
2,80e-2
H2
7,46e-4
8,05e-4
6,91e-3
9,53e-2
CO2
1,95e-4
1,75e-4
6,91e-4
2,28e-2
H2O
2,91e-4
2,80e-4
9,53e-4
2,28e-4
For the calculation of the fluid density the ideal gas law is applied:
p
R MT
in (kg/m3)
where R M =
R
and i = CH4, CO, H2, CO2, H2O
Yi Mi
i
The viscosity of the gas mixture is obtained via the equation (see [3])
3
(10)
The SOFC
=
i
II.4
Yii
in 110-7 Pas i,j =CH4, CO, H2, CO2, H2O
Y
j ij
(11)
0.5 M 0.25
1 + i j
j M i
with ij =
0.5
Mi
8 1 +
M
j
(12)
i = 844
Mi T
i2 v,i
with v,i =
in 110-7 Pas
(13)
1.16145
0.52487
2.162
+
+
i =CH4, CO, H2, CO2, H2O
kT
kT
kT
exp 0.773 exp 2.438
i
i
i
In Tab. 3 the dependent variables of the anode gas channel are listed
Tab.3. Dependent variables (anode gas channel)
Parameter
symbol
units
ux,K
m s-1
uy,K
m s-1
YCH 4 , K
YCO , K
YH 2 , K
YCO2 , K
YH 2O , K
pK
Pa
Pressure
Boundary conditions.
(14)
The SOFC
II.5
(15)
Yi , K
y
(16-20)
ux, K = 0
Di , gas
(21)
= Dieff, gas
( p A Yi , A )
R T y
(22)-(25)
u y,K = u y, A
(26)
Yi , K = Yi , A
(27)-(32)
pK = p A
(33)
A parabolic velocity profile is assumed at the inlet of the anode gas channel
Yi , K = Yi o i =CH4, CO, H2, CO2, H2O
(34)-(38)
u y,K = 0
ux,K =
(39)
4 u x , K ,max
d
2
K
y2 + 4
u x , K ,max
dK
(40)
This is the parabolic entry profile with ux,K,max in the middle of the channel
pK=patm
(41)
The other variables at the outlet of the anode gas channel are calculated by the governing
equations
Tab.4. Parameters for the governing equations and the boundary conditions (anode gas
channel)
parameter
Diffusion coefficient of
component i in the gas
symbol
units
value
source
Di,gas
m2 s-1
eq. (9)
[3]
The SOFC
II.6
mixture
Binary Diffusion
coefficients
[3]
Dij
m2 s-1
tab 2
MCH4
10-3
kg/mol
16
[3]
MCO
10-3
kg/mol
28
[3]
MCO2
10-3
kg/mol
44
[3]
MH2
10-3
kg/mol
[3]
MH2O
10-3
kg/mol
18
[3]
uox,max
m s-1
0.5
chosen
YCHo
0.33
o
YCO
YHo2
YCOo
YHo O
0.67
CH4
3.758
CO
3.69
[3]
CO2
3.941
[3]
H2
2.827
[3]
H2O
2.641
[3]
chosen
chosen
chosen
chosen
chosen
[3]
The SOFC
II.7
steam
Lennard-Jones-Parameter
of methane
CH4/k
148.6
[3]
Lennard-Jones-Parameter
of carbon monoxide
CO/k
91.7
[3]
Lennard-Jones-Parameter
of carbon dioxide
CO2/k
195.2
[3]
Lennard-Jones-Parameter
of hydrogen
H2/k
59.7
[3]
Lennard-Jones-Parameter
of steam
H2O/k
809.1
[3]
Anode
Below the governing equations for the anode are listed.
Species balances
p A u x , A YCH , A
4
R T x
) ( p
u y , A YCH 4 , A
R T y
p A YCH 4 , A
eff
DCH
,
gas
4
R T y
+
y
p A YCH 4 , A
eff
DCH
,
gas
4
R T x
+
x
CH 4
=0
(42)
eff
eff
( p A YCO , A )
( p A YCO , A )
DCO
DCO
, gas
, gas
R T y
R T x
( p A u x , A YCO , A ) ( p A u y , A YCO , A )
+
+ rCH 4 rs = 0
+
R T x
R T y
y
x
(43)
p A u x , A YH , A
2
R
T
x
) ( p
u y , A YH 2 , A
R T y
p A YH 2 , A
DHeff2 , gas
R T y
+
y
p A YH 2 , A
DHeff2 , gas
R T x
+
x
+ 3r
CH 4
+ rs = 0
(44)
p A u x , A YCO , A
2
R
T
x
) ( p
u y , A YCO2 , A
R T y
p A YCO2 , A
eff
DCO
2 , gas
R T y
+
y
p A YCO2 , A
eff
DCO
2 , gas
R T x
+
x
(45)
+r =0
s
The SOFC
II.8
p A u x , A YH O , A
2
R T x
rs = 0
) ( p
u y , A YH 2O , A
R T y
p A YH 2O , A
DHeff2O , gas
R T y
+
y
p A YH 2O , A
DHeff2O , gas
R T x
+
x
(46)
Darcys law
p A
= ux, A
x
kp
(47)
p A
= uy, A
y
kp
(48)
Daltons law
i, A
=1
(49)
One obtains 7 PDEs and one algebraic equation for the 8 dependent variables in the anode:
ux,A, uy,A, pA, YCH 4 , A , YCO , A , YH 2 , A , YCO2 , A und YH 2O , A . The independent variables are x and y.
The effective diffusion coefficients Deffi,gas are calculated from the diffusion coefficient Di,gas
and from the Knudsen Diffusion coefficient DKi by the equation
1
D
e ff
i,gas
D
i, gas
1
D iK
(50)
The porosity of the anode is estimated to be 30%. For the tortuosity a value of 3 is
assumed.
The reactions occurring on the (inner) anode surface are the reforming reaction
CH4 + H2O CO + 3H2
(51)
H2 + CO2
(52)
For the rate of the reforming reaction on a Ni/YSZ-cermet an expression is employed which
was found by [4]:
CH 4
The SOFC
II.9
mol
rCH 4 = 2.45 3
p A YCH 4
1.2
m s Pa
1.2
J
57840 [ mol
]
exp
RT
(53)
For the shift reaction, a volumetric expression is used which is calculated from the
experimental results presented in [5]:
J
112131[ mol
] p 2 Y Y 0.425 mol exp 132467 [ molJ ] p2 Y Y
rs = 0.0437 mmol
exp
3
A CO,A H2O,A
A CO2 ,A H2 ,A
Pa 2s
m3Pa 2s
R T
R T
(54)
parameter
symbol
units
ux,A
m s-1
uy,A
m s-1
YCH 4 , A
YCO , A
YH 2 , A
YCO2 , A
YH 2O , A
pA
Pa
pressure
Below the relevant boundary conditions for the anode are listed.
Boundary conditions
(55)
(56)-(60)
The SOFC
YCH 4
y
=0
(61)
( p AYCO ) jCO
=
R T y
2F
eff
DCO
, gas
p AYH 2
DHeff2 , gas
R T y
DHeff2O , gas
eff
DCO
2 , gas
II.10
)=
p AYH 2O
R T y
p AYCO2
R T y
(62)
jH 2
(63)
2F
)= j
H2
(64)
2F
)= j
(65)
CO
2F
pA=patm
(66)
The other variables at the outlet of the anode gas channel are calculated by the governing
equations
The current densities for H2 and CO, jH2 and jCO, respectively, are calculated from the cell
voltage U (eq. (67)-(68)). The equivalent circuit diagram which forms the basis of the model
is depicted in figure 2.
Fig. 2. Equivalent circuit diagram for the electrical processes in the SOFC [6]
10
The SOFC
jH2 =
jCO =
(U
Nernst,H2
II.11
(67)
(68)
RA,CO
The expression for the area-specific polarization resistance of H2, RA,H2, is of the form (see
[1])
1 .1 1 0 5 [ J / m o l ]
0 .2 5
2F
(69)
= 0 .2 1 3 1 0 9 A / m 2
ex p
YH 2 , A
R A,H 2
RT
R
T
For CO, the expression is modified so that RA,CO/RA,H2 = 2.6 which corresponds to the mean
value obtained by Matsuzaki and Yasuda [7]:
1
1
R A ,C O
1 .1 1 0 5 [ J / m o l ]
0 .2 5
2F
exp
= 0 .8 2 1 0 8 A / m 2
(Y C O , A )
RT
RT
(70)
For the cathode, the expression for the area specific polarization resistance is taken from [1]:
1 .6 1 0 5 [ J / m o l ]
1
4F
0 .2 5
10
2
exp
= 1 .4 9 1 0 A / m
( 0 .2 1 )
RK
RT
RT
(71)
The Nernst voltages are calculated according to equations (72) and (73), respectively
o
U Nernst ,H2 = UZelle
, H2 +
RT YH2 , A YO2
ln
YH2O, A
2F
o
U Nernst ,CO = U Zelle
,CO +
(72)
RT YCO , A YO2
ln
YCO2 , A
2F
(73)
(74)
Tab.4. Parameters for the governing equations and the boundary conditions (anode)
parameter
symbol
units
value
source
Dieff, gas
m2 s-1
Eq. (50)
IWE
DK CH4
m2 s-1
8.0 10-5
calculated
DK CO
m2 s-1
6.20 10-5
calculated
11
The SOFC
II.12
DK H2
m2 s-1
0.0002399
calculated
DK CO2
m2 s-1
5.12 10-5
calculated
DK H2O
m2 s-1
8.0 10-5
calculated
A s mol-1
96485.4
jCO
A m-2
Eq. (67)
jH 2
A m-2
Eq. (68)
kp
m2
8.5 10-14
FZJ
Atmospheric pressure
patm
Pa
101325
J mol-1 K-1
8.3144
RA,CO
m2
Eq (70)
RA , H 2
m2
Eq (69)
RK
m2
Eq (71)
Rohm
m2
9.78 10-6
IWE
temperature
1223
chosen
Cell voltage
0.7
chosen
Nernst voltage of CO
U Nernst ,CO
Eq (73)
Nernst voltage of H2
U Nernst , H 2
Eq (72)
o
U Zelle
,CO
0.9145
calculated
o
U Zelle
,H2
0.9342
calculated
YO2
0.21
chosen
0.3
chosen
chosen
Faraday constant
Current density produced by
elecrochem. CO-Oxidation
Current density produced by
elecrochem. H2-Oxidation
Pressure loss coefficient
Area-specific Polarization
resistance of el.CO-oxidation
Area-specific Polarization
resistance of el.H2-oxidation
Area-specific Polarization
resistance of el.O2-reduction
Area-specific Ohms resistance
for 8YSZ
12
The SOFC
II.13
Parameter Variation
For the base case an anode thickness of 50m was assumed. This anode thickness is a realistic
value for electrolyte supported cells. In anode supported cells the anode thickness has to be
higher. Therefore this value should be varied from 50m to 2mm in order to investigate
transport limitations in the porous anode in dependence of the anode thickness. For the anodes
with a thickness between 0.5 and 2mm it was assumed that the open porosity is increased.
Therefore for the ratio of porosity to tortuosity / a value of 0.21 is supposed [5]. Because of
the reduced electrolyte thickness of anode supported cells, the area-specific ohmic resistance
Rohm is set to 1.63 10-6 m2 in these cases.
a)
b)
c)
d)
13
The SOFC
II.14
d)
Fig 3. molar fractions of the components in the anode gas channel (base case)
In figure 4 the molar fractions of the components in the anode are shown for the base case.
a)
b)
c)
d)
d)
Fig 4. molar fractions of the components in the anode (base case)
14
The SOFC
II.15
The gradient of the molar fractions in x-direction is similar to the gradient in the anode gas
channel. The low variation of the molar fractions in y-direction indicates that no diffusion
limitations occur in the anode with a thickness of 50 m.
As a result from the distribution of the species in the anode, the Nernst voltage of H2 increases
along the anode length, whereas the Nernst voltage of CO decreases because the ratio of CO
to CO2 decreases in x-direction (Fig. 5). The distribution of the current densities jH2 and jCO is
according to the distribution of the Nernst voltages. The mean value of the current density j of
5615 A/m2 is high compared to measured values which are around 4000 A/m2 . This indicates
that in further work the electrochemical parameters of the model should be adjusted to results
obtained by own experiments
a)
b)
c)
d)
e)
Fig 5. Nernst voltages and current densities jH2, jCO and j (base case)
In the remainder the results of the parameter variation are presented. Exemplarily the
discussion is confined to cells with an anode thickness of 2mm. Fig. 6 shows the molar
fractions of the components in the anode for this case. In evidence the reforming reaction
proceeds more quickly in x-direction because of the higher amount of Nickel present in the
anode compared to the base case. For every component high gradients occur in y-direction
which limit the chemical and electrochemical conversions. The CO2 content in the upper left
corner seems to be too high. Whether this value originates from a numerical error or a bad
15
The SOFC
II.16
a)
b)
c)
d)
d)
Fig 6. molar fractions of the components in the anode (2mm anode)
Fig. 7 shows the Nernst voltages and current densities of the anode supported cell with 2mm
anode. Because of the thinner electrolyte a higher current density can be achieved compared
to electrolyte supported cells. The high increase of UNernst,CO compared to the base case, can
be attributed to the high rate of the reforming reaction which produces CO very fast and
therefore leads to an increase in the ratio of yCO/yCO2 in the first quarter of the cell.
16
The SOFC
II.17
a)
b)
c)
d)
e)
Fig 7. Nernst voltages and current densities jH2, jCO and j (2mm anode)
Conclusions
The internal reforming process of methane on a Ni/YSZ anode was successfully simulated
with a 2D model. The impact of the anode thickness was investigated and it could be shown
that no diffusion limitations occur in an electrolyte supported cell. In anode supported cells
with 2mm anode thickness the reforming reaction proceeds faster than in an electrolyte
supported cell but the chemical and electrochemical processes are limited by pore diffusion in
the anode. Because of the thinner electrolyte in anode supported cells, the results of the
simulation show that a higher current density can be achieved in an anode supported cell,
compared to an electrolyte supported cell.
Bibliography
17
The SOFC
II.18
[1]
[2]
[3]
[4]
[5]
[6]
[7]
Y. Matsuzaki, I. Yasuda, electrochemical oxidation of H2 and CO in a H2-H2O-COCO2-System at the interface of a Ni/YSZ cermet electrode and YSZ electrolyte, Journal
of The Electrochemical Society, 147, pp. 1630-1635, 2000
18
Part III:
The numerical solution with the FDEM
1 The FDEM program package
A detailed report on the FDEM (Finite Difference Element Method) program package is in
preparation parallel to this report [1]. There the basic method is described together with
three application examples. The third example (Chapter 3.4) is of particular interest with
respect to the present report: it is the oxygen diffusion at the cathode side of a PEM fuel cell
and has for the first time demonstrated the suitability of FDEM for the numerical simulation
of fuel cells.
In the following we give a short outline of FDEM. FDEM [2] is a program package
developed at the Computer Center of the University of Karlsruhe that solves by a finite
difference method arbitrary nonlinear systems of PDEs (partial differential equations) under
arbitrary nonlinear BCs (boundary conditions) on an unstructured FEM grid. The 2-D or
3-D PDEs must be of elliptic type (boundary value problem) or parabolic type (initialboundary value problem). Here the FEM grid serves only for the structuring of the space, i.e.
the determination of the neighbouring nodes. In 2-D we use triangles, in 3-D tetrahedrons.
For each node we generate with a sophisticated algorithm by means of neighbouring nodes
difference formulas of consistency order q, optionally q = 2 or q = 4 or q = 6. By the use
of formulas of consistency order q +2 an estimate of the discretization error is obtained. For
parabolic equations we use in time direction (for stability reasons) fully implicit difference
formulas of consistency order p < 6, with error estimate by formulas of order p + 1.
This means that one gets the exact solution if in the space direction the solution of the
PDEs is a polynomial of order q and in time direction of order p. This is used to test the
implementation of the PDEs to be solved.
The knowledge of the error permits a selfadaptation of the solution method. In time direction the order p and the time step are always automatically optimized. In space direction
the solution can be adapted to a requested accuracy by grid refinement (bisection of triangle
or tetrahedron edges). There is also possible an optimization of the consistency order q
individually for each node by the comparison of the discretization errors of the different
orders q = 2, 4, 6.
For many technical applications the solution domain is composed from subdomains in
which hold different PDEs, e.g. anode material and hydrogen channel in a SOFC. It is not
possible to apply difference formulas across the boundaries of the subdomains. Therefore
we have introduced in FDEM dividing lines (which are in 3-D in effect dividing areas,
but we call them nevertheless formally dividing lines). These dividing lines are internal
boundaries over which we cannot differentiate. The solutions on both sides of the dividing
lines are coupled by coupling conditions (CCs). Thus one gets over the whole domain
III.1
(composed of several subdomains) a global solution with global error estimate. The meshes
on both sides of a dividing line do not have to coincide, one may have non-matching grids.
By interpolation with the consistency order q that holds also for the difference formulas,
the order q is maintained also for the CCs and therefore for the whole coupled domain.
Because of the possibility of non-matching grid, in the subdomains an independent local
grid refinement for the adaptation of the accuracy can be executed. The non-matching grid
also allows a relative movement of the subdomains to each other.
Because FDEM must solve arbitrary nonlinear systems of PDEs, the linearization is
executed by the Newton-Raphson method. In order to make the method as robust as possible
we check after each iteration step if the defect has decreased. If this does not hold we try
with a self-adapted relaxation factor to reduce the defect (then there is only linear instead
of quadratic convergence). For very good convergence we use (optionally) the simplified
Newton method with constant matrix to save computation time. The Newton method is
terminated if the Newton defect is smaller than a corresponding discretization error term,
that no unnecessary digits are computed.
From the discretization of PDEs result very large and sparse linear systems of equations. These are solved by the LINSOL program package that has also been developed at
the Computer Center of the University of Karlsruhe. LINSOL [3] comprises CG (conjugate gradient) methods of quite different types for the iterative solution, between which an
automatic optimization in a polyalgorithm can be executed. LINSOL also contains a direct
solver with optionally reduced fill-in that can be used as preconditioner for the iterative
solvers. The fill-in is reduced by different bandwidth optimizers which leads above all for
3-D problems to a considerable saving of computation time and storage space.
FDEM and LINSOL have been developed from the beginning for efficient data structures on distributed memory parallel computers (in contrast to shared memory computers).
Here the distribution of the data to the processors plays the decisive role. We use a 1-D
domain decomposition that can be executed automatically and runs over dividing lines. For
grid refinement a new distribution of the data is executed after each refinement step. The exchange of the data between the processors takes place by the quasi standard MPI (message
passing interface). Thus FDEM is running efficiently on shared and distributed memory
computers. FDEM has been tested on many different types of parallel computers all over
Germany.
FDEM is a program package for the solution of PDEs that has unique properties. To us
no other program package is known that unifies in a single code comparable properties concerning the flexibility of the solution method, of the solution domain, of the error estimate
and of the parallelization.
FDEM is a black box solver, i.e. it solves an arbitrary nonlinear system of PDEs and
boundary conditions (BCs) on an arbitrary domain. The domain can be composed from
subdomains that are separated by dividing lines and whose solutions are coupled by CCs.
Which PDEs under which BCs and CCs are to be solved is determined by the entering of
the PDEs, BCs and CCs as Fortran code into prescibed program frames. The domain with
its boundaries and dividing lines is read from a file as 2-D or 3-D grid according to given
rules.
Because we assume nonlinear PDEs and solve them by the Newton-Raphson method,
III.2
also the Jacobi matrices for the PDEs, BCs and CCs must be entered as Fortran code into
program frames. The Jacobi matrices reflect the dependence of the PDEs, BCs and CCs
from their arguments. For strongly nested nonlinear systems of PDEs the formulation and
implementation of the Jacobi matrices is the main part of the task.
For the simulation of a fuel cell one will start at first with a simplified 2-D model to get
insight in the numerical behaviour of the PDEs. Then the 2-D model is refined as necessary.
Here one will at first solve the stationary equations, then eventually the non-stationary ones.
With the thus gained experiences one will try to implement a simplified, later a refined 3-D
model. The transition from 2-D to 3-D is with respect to computation time and above all
to the necessary storage space a very large step. The transition from the stationary to the
non-stationary problem results only in a corresponding factor in computation time, 100 time
steps need 100 times the computation time of a single stationary step.
The above mentioned procedure would be a research program for several years. Because
of the relatively short available time for the project (it was intended as an intermediate
financing to gain experience and prepare a larger cooperation project) only stationary 2-D
fuel cells were simulated. However, the experiences gained for the formulation of the many
parameters in the system of PDEs for the different types of fuel cells are very promissing
and prepare the research direction for further more sophiscated models of fuel cells.
III.3
Figure 1: Upper part: domain of solution in the environment. Lower part: Details of the
domain of solution.
III.4
parameter
thickness of GDL
width of channel
width of rib
Symbol
English German
tGDL
dGDL
c
k
r
s
name
in code
dGDLc
kc
sc
We got from the ZSW at first a report with the PDEs and BCs in German and for
these notations we wrote the code. Later we got the English version. Unfortunately, ZSW
has changed the symbols for some parameters so that they fitted to their English names.
Therefore we give the symbols for the parameters of the English version, of the German
version and additionally the name of the parameter in the Fortran code which usually fits to
the German name.
Table 2: Dependent variables for node i, mfd means molar flux density.
no.
1
2
3
4
5
6
7
8
9
10
11
parameter
mfd of oxygen in x-dir.
mfd of oxygen in y-dir.
mfd of water vapour in x-dir.
mfd of water vapour in y-dir.
mfd of nitrogen in x-dir.
mfd of nitrogen in y-dir.
oxygen partial pressure
water vapour partial pressure
nitrogen partial pressure
total pressure
current density
symbol
n xo
n yo
n xw
n yw
n xn
n yn
po
pw
pn
p
i
name in code
u(i,1)
u(i,2)
u(i,3)
u(i,4)
u(i,5)
u(i,6)
u(i,7)
u(i,8)
u(i,9)
u(i,10)
u(i,11)
Table 1 gives the structural and geometry parameters of Table I1. We do not repeat here
the units and values. Table 2 which corresponds to Table I5 gives the dependent variables
for ZSW4. In the black-box solver FDEM the variables for node i and solution component j
are denoted by u(i,j). In the Fortran code many loops run over the node number i. If the
innermost loop runs over the first index of a multidimensional array, in Fortran the elements
are accessed contiguously which is the most efficient access. Therefore the node number i
is in the first position in u(i,j).
We can see from Table 2 that we have 11 dependent variables. One of them, namely the
current density i (do not confuse with node index i), plays a very special role. It occurs only
in the boundary condition at the reaction layer. If it appeared explicitly we could express
it by the other variables. But it appears in a bad nonlinear way in the Tafel equation I(8).
Therefore we proceed as follows: we take the current density i as a variable in the whole
III.5
no.
1
2
3
4
5
6
7
8
9
10
11
variable
n xo
n yo
n xw
n yw
n xn
n yn
po
pw
pn
p
i
equation
I(18) in form (2)
I(19) in similar form
I(20) in similar form
I(21) in similar form
I(22) in similar form
I(23) in similar form
I(24)
I(25)
I(26)
I(27)
(1)
domain, but it has a physical meaning only at the reaction layer. In the other nodes the
current density i plays the role of a dummy variable and there we put
i=0.
(1)
This is the PDE and BC except for the nodes at the reaction layer (boundary in
Fig. 1).
As we can see at Table 2 we have 11 variables and therefore we need 11 equations.
In Table 3 we show which equation is used in which position in the system of PDEs. As
FDEM requests the PDEs in the form Pi u = 0 for PDE no. i, we take all terms to the l.h.s.
of the equations. Therefore I(18) is used in the form
pn n xo po n xn
pw n xo po n xw
n xo
+
+
+
DKno
Dow p
Don p
po p
Bo
Bw
Bo pw
1 po
+
1
+
RT x
RT p x
DKno Dow p
Bo
Bn
po p
Bo p n
1
=0.
+
Don p
Bo
RT p x
(2)
Similarly we write in equs. I(19)I(27) all terms to the l.h.s.. Now these equations
can be coded in Fortran. Before we can write down code we must give to the coefficients
Fortran names. The coefficients that occur in the PDEs are compiled in Tables I3 and I6.
There are coefficients that are constants, their names will end by c for constant. There
are coeffients that depend on the node i and there are coefficients that depend on solution
component index j or j, k. In Tables I3 and I6 all information about the coefficients is given.
Here we give in Table 4 only the Fortran names for the coefficients. In the PDEs appear the
following coefficients DKnj with (j) as o, w, n, thus DKno(i), DKnw(i), DKnn(i) for
each node i, Beff (j) similarly as Beff o(i), Beff w(i), Beff n(i). As the coefficients Dj,k
III.6
Table 4: Fortran names for the coefficients that occur for the PDEs.
symbol
R
F
T
o
(x)
a
kf
Djk
V do
V dw
V dn
Bj
j
Fortran
name
Rc
Fc
Tc
epsoc
epsxc(i)
aec
kfc
tau
D(j,k)
Vdo
Vdw
Vdn
Beff(j)
nue(j)
symbol
o
w
n
Mo
Mw
Mn
DKnj
Kj
j
<r>
< r2 >
RT
Fortran
name
eta
etaoc
etawc
etanc
Moc
Mwc
Mnc
Dkn(j)
omc
Kn(j)
lam(j)
rmc
rmqc
RTc
(3)
Table 5: Fortran names for the coefficients that occur for the BCs.
symbol
pch
uc
yo,a
i0
n
ref
o
fv
pSw
Fortran
name
pKac
phic
yokc
ioc
alphac
nc
porefc
fvc
pwsc
symbol
U0
Uz
dmem
mem
A
B
C
Fortran
name
gammac
U0c
UZc
dmemc
kappac
Ac
Bc
Cc
into Fortran code. These p(i, j) for equation j in node i are pre-computed and in a later
step entered in a program frame that places them into the right position in the linear system
of equations for the computation of the Newton correction. The meaning of p(i, 1) is the
Newton residual or Newton defect of the first equation in the block of 11 equations for
node i. It is the r.h.s. in the linear system of the computation of the Newton correction.
The next problem we want to discuss are the boundary conditions for the 4 types of
boundaries to of Fig. 1. At the boundaries for some variables BCs are given. However,
we need at each boundary node 11 equations for the 11 variables. So we take the missing
equations quite naturally from the set of the interior equations, where no other conditions,
e.g. symmetry at boundary , are given.
The set of parameters that occur in the BCs is given in Table I4. We give in Table 5 the
Fortran names for these parameters.
The BCs for the MTPM transport equations are given in equs. I(37)I(49). We have to
supplement these equations by interior or other PDEs (symmetry) so that we have always
11 equations for the 11 variables. In contrast to Part I we do not use (upper) indices like
U C, U R for channel or rib because the index i of the node tells where the node is.
BCs at boundary (open to channel): Here we need pSw with the Fortran name pwsc
pre-computed from I(9) which is resolved for pSw and gives a constant value. For a node i
on the boundary we have the BCs I(37)I(40). As we need 11 BCs for the 11 variables,
we supplement the conditions by the PDEs I(18)I(23) and equ. eqrefequ1. In Table 6 we
list the variables and the equations at boundary that includes the left corner. In order to
show how such a BC is translated to Fortran, we show the code for the equation for pn in
Table6 for a node i on the boundary :
p(i, 9) = u(i, 9) (pKac u(i, 8)) (1D0 yokc)
III.8
(4)
no.
1
2
3
4
5
6
7
8
9
10
11
variable
n xo
n yo
n xw
n yw
n xn
n yn
po
pw
pn
p
i
equation
I(18) in form (2)
I(19) in similar form
I(20) in similar form
I(21) in similar form
I(22) in similar form
I(23) in similar form
UC = 0
po (pch pw )yo,a
S
pw pw = 0
ch
UC) = 0
pn (p pw )(1 yo,a
ch
pp =0
(1)
The right corner of boundary is avoided as the separation between boundary and
is put between two nodes. If at the separation of the two boundaries a node would be
placed, there would be no unique BC for this node. However, this procedure means that the
geometrical location how far the BC extends to the right and BC to the left changes
with the mesh size in the x-direction.
BCs at boundary (under rib): Here the impermeability conditions I(41)I(43) are
given. Therefore we need another 8 equations for the 11 variables. In Table7 we list the
variables and the equations at boundary that includes the right corner.
BCs at boundary (symmetry lines): Because of the symmetry no x-transport of
species takes place, this results in the BCs I(44)I(46). For the other variables, except
current density i for which i = 0 holds also here, we put the x-derivative to zero. Table 8
shows the variables and the equations for boundary , excluding the corner nodes.
BCs at boundary (reaction layer): The BCs for boundary are given in I(47)I(49).
However, in these equations appears the current density i for which holds the Tafel equation I(8). This extremely nonlinear equation cannot be resolved explicitly for the current
density i. Therefore we take it in the form
nF (U0 Uz dmem
i)
po
mem
=0
(5)
exp
i fv i0
RT
pref
o
and use this as the equation for the current density i. The other two BCs I(47), I(48) are
used in the form
1
i = 0,
4F
1
i = 0.
n yw +
2F
n yo
III.9
(6)
(7)
no.
1
2
3
4
5
6
7
8
9
10
11
variable
n xo
n yo
n xw
n yw
n xn
n yn
po
pw
pn
p
i
equation
I(18) in form (2)
n yo = 0
I(20) correspond. form
n yw = 0
I(22) correspond. form
n yn = 0
I(24) correspond. form
I(25) correspond. form
I(26) correspond. form
I(27) correspond. form
(1)
no.
1
2
3
4
5
6
7
8
9
10
11
variable
n xo
n yo
n xw
n yw
n xn
n yn
po
pw
pn
p
i
equation
n xo = 0
n yo /x = 0
n xw = 0
n yw /x = 0
n xn = 0
n yn /x = 0
po /x = 0
pw /x = 0
pn /x = 0
p/x = 0
(1)
III.10
no.
1
2
3
4
5
6
7
8
9
10
11
variable
n xo
n yo
n xw
n yw
n xn
n yn
po
pw
pn
p
i
equation
I(18) in form (2)
(6)
I(20) correspond. form
(7)
I(22) correspond. form
I(49)
I(24) correspond. form
I(25) correspond. form
I(26) correspond. form
I(27) correspond. form
(5)
For the other variables we again take the corresponding equations of Table 3. Table 9
shows the variables and equations at boundary .
Again we show how equation (5) is coded in Fortran for a node i on the boundary :
at first auxiliary values are computed before the i-loop:
hU 2 = alphac nc F c
hU 3 = fvc ioc
Then in the i-loop for a node i, we have
p(i, 11) = u(i, 11) hU 3 (u(i, 7)/porefc)
gammac exp(hU 2 (U 0c U Zc
(8)
III.11
(9)
u1
u2
..
.
u=
P1 u
P2 u
..
.
Pu =
um
(10)
Pm u
=
u
P1
u1
...
..
.
P1
um
..
.
Pm
u1
...
Pm
um
=
,...,
uxy
P1
u1,xy
...
..
.
P1
um,xy
..
.
Pm
u1,xy
...
Pm
um,xy
(11)
We will not give here the formulas for all Jacobians of the system. We rather want to
illustrate the procedure by a few examples. The first PDE P1 u in (10) is in our case the first
equation in Table 3, i.e. I(18), that we use in the formulation of equ. (2). The first variable in
1
Table 3 is n xo . Thus in (11) P
xo . But before we
u1 is the derivative of equ. (2) with respect to n
formulate the derivatives we must check if there are not hidden variables in coefficients
that appear in the equation. Fortunately, the coefficients do not depend on nxo , so we get
Pn
1
pw
P1
+
.
=
+
x
n o
DKno Dow p Don p
(12)
P1
1
However, if we e.g. want to compute P
po = u7 , i.e. the dependency of the first PDE
from the variable po , the permeabilities Bo and Bw depend in a complicated way from po :
In I(31) Bj , j = o, w, n, depends on j , I(32) that depends on pk , k = o, w, n, and Bj
depends also on , I(33) that depends on pk , k = o, w, n and finally Bj depends on Kj ,
i
, i, j = o, w, n and
I(35) that depends by j on pj , I(36). Therefore we precompute the p
j
similarly
( 1 )
pj
and
Ki
pj .
We have I(31)
Bo = DKno
o + Ko < r 2 >
+
1 + Ko
8
(x)
1
.
(13)
With the pre-computed values we can compute with the chain rule for differentiation
e.g.
1
< r 2 > (x) ( )
p
.
(1 + Ko )2
8
po
(14)
In the same way all Bi /pj and also Bi /p are pre-computed and stored for all
nodes i. Note that the expressions for (Djk p) I(30) do not depend on p.
Now we can formulate P1 /u7 in the terminology of (11) which is our equ. (2) differentiated to po
(1 + Ko )(
Bo
= DKno
po
o
po
Ko
po )
( o + Ko )
III.12
Ko
po
n xn
1 p
n xw
P1
+
=
po
Dow p Don p RT p x
Bo
w
Bo B
pw
p
po
Bo
po Bw po
+
+ Bo
[]
po RT p x
Dow p
Bo2
Bo
n
Bo B
1 p
po p
pn
po Bn po
+ Bo []
.
2
Don p
Bo
RT p x
RT p x
(15)
Here [*] denotes the brackets in (2) which saves writing down twice this large expression. Quite naturally, in the coding one would precompute [*] as an auxiliary variable, so
this form is closer to the coding than the written-out form. If you nevertheless imagine
written out [*] in (15) you get the impression, in which a complicated way the first PDE of
the system depends on the 7th variable. In a similar way the first 6 PDEs depend on the 7th
to 10th variable. It would not be possible to handle these Jacobians without a clear scheme
for intermediate derivatives as it has been presented above. These expressions then must be
translated into Fortran code as it has been shown for equ. (2).
In the same way we must deliver the Jacobians for the BCs. In many cases the interior
PDEs are also used for the BCs, so the Jacobians are the same in these cases.
In Table 9 the equations are shown for boundary , the reaction layer. Here the Tafel
equation (5) is the BC for the last variable, the current density i. Therefore P11 /u11 is
the derivative of (5) with respect to i. That we do not get too complicated expressions and
because intermediate results can be used also in other Jacobians, we precompute auxiliary
values
hu2 = nF ,
hu3 = o i0 ,
i)/(RT )] .
hu4 = exp[hu2 (U0 Uz dmem
mem
With these values we get for the boundary from equ. (5)
po
dmem
p11
= 1 + hu3
.
(16)
hu4 hu2
ref
i
mem RT
po
All these Jacobians are entered in a prescribed program frame of FDEM that then executes with this information the solution process as described in detail in [1]. So by the PDEs
and BCs and their Jacobians from the black-box solver a solver for the special problem is
created.
Here we present the results that we obtained for the standard data given in Part I for the
(more complicated) Mean Transport Pore Model (MTPM) as an example for the information
that FDEM delivers as result. We computed on the HP XC6000 parallel supercomputer of
the University of Karlsruhe. The processors are Intel Itanium2 processors with 1.5 GHz. We
computed with 32 processors in parallel. The grid was 200 201 nodes in x-,y-direction.
We computed with consistency order q = 4. As we have 11 unknowns per node, see
Table 2, we have totally 442, 200 unknows. The computation time was 4123sec on the
master processor 1.
III.13
In Figs. 221 the results are plotted as coloured contour plots together with their errors.
These plots are black-and-white in the printed version. So it might be recommended to look
at the pictures at the computer in the on-line version of the paper.
Fig. 1 shows the domain of solution. At the lower boundary we have in the left half
the opening to the oxygen channel and in the right half the domain is closed by the rib.
Therefore we have in the middle of the lower boundary an abrupt change in the BCs that
causes a singularity in the solution because there is mathematically no smooth transition
from the one type of BC to the other type. This singularity is clearly visible in the figures.
For the fuel cell research the behaviour of the partial pressure po of oxygen, pw of
water vapour, pn of nitrogen and the current density i at the reaction layer are of greatest
interest. Therefore we show in Figs. 2226 these quantities together with their errors along
the reaction layer. We can see that the singularity of the lower boundary has smoothed
out at the upper boundary of the solution domain. At the top of Fig. 26 the value imean is
presented which is the mean value of the grid points at the reaction layer and is a measure
of the output of the fuel cell. As we have the singularity in the problem, the local error at
the singularity will not decrease with finer grid as fast as we expect from the consistency
order q = 4.
The discussion of the results from the point of view of the fuel cell researcher is presented in Part I.
III.14
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
n-dot-o-x
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
5.4E-02
4.8E-02
4.2E-02
3.6E-02
3.0E-02
2.4E-02
1.8E-02
1.2E-02
6.0E-03
error-n-dot-o-x
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.15
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
n-dot-o-y
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
5.4E-02
4.8E-02
4.2E-02
3.6E-02
3.0E-02
2.4E-02
1.8E-02
1.2E-02
6.0E-03
error-n-dot-o-y
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.16
0.00
-0.02
-0.04
-0.06
-0.08
-0.10
-0.12
-0.14
-0.16
-0.18
n-dot-w-x
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
5.4E-02
4.8E-02
4.2E-02
3.6E-02
3.0E-02
2.4E-02
1.8E-02
1.2E-02
6.0E-03
error-n-dot-w-x
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.17
-0.02
-0.04
-0.06
-0.08
-0.10
-0.12
-0.14
-0.16
-0.18
n-dot-w-y
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
5.0E-02
4.4E-02
3.9E-02
3.4E-02
2.8E-02
2.2E-02
1.7E-02
1.1E-02
6.0E-03
error-n-dot-w-y
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.18
3.0E-05
2.5E-05
2.0E-05
1.5E-05
5.0E-06
0.0E+00
-5.0E-06
n-dot-n-x
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
1.0E+00
9.0E-01
8.0E-01
7.0E-01
6.0E-01
5.0E-01
4.0E-01
3.0E-01
2.0E-01
1.0E-01
error-n-dot-n-x
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.19
1.2E-05
1.0E-05
8.0E-06
6.0E-06
4.0E-06
2.0E-06
0.0E+00
n-dot-n-y
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
6.0E-01
5.0E-01
4.0E-01
3.0E-01
2.0E-01
1.0E-01
error-n-dot-n-y
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.20
1.7E+04
1.6E+04
1.5E+04
1.4E+04
1.3E+04
1.2E+04
p-o
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
2.1E-02
1.8E-02
1.5E-02
1.2E-02
9.0E-03
6.0E-03
3.0E-03
error-p-o
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.21
2.3E+04
2.2E+04
2.1E+04
2.0E+04
1.9E+04
1.8E+04
1.7E+04
p-w
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
2.4E-02
2.0E-02
1.6E-02
1.2E-02
8.0E-03
4.0E-03
error-p-w
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.22
67200
66800
66400
66000
65600
p-n
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
2.1E-03
1.8E-03
1.5E-03
1.2E-03
9.0E-04
6.0E-04
3.0E-04
error-p-n
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.23
101300.5
101300.4
101300.3
101300.2
101300.1
p
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
4.0E-07
3.5E-07
3.0E-07
2.5E-07
2.0E-07
1.5E-07
1.0E-07
5.0E-08
error-p
0.0002
0.00015
0.0001
5E-05
0.00025
0.0005
0.00075
0.001
III.24
p-o
17000
16000
15000
14000
13000
12000
11000
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
error-p-o
3,0E-02
2,5E-02
2,0E-02
1,5E-02
1,0E-02
5,0E-03
0,0E+00
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
Figure 22: Oxygen partial pressure p o along the reaction layer and its global relative error.
III.25
p-w
25000
24000
23000
22000
21000
20000
19000
18000
17000
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
error-p-w
3,0E-02
2,5E-02
2,0E-02
1,5E-02
1,0E-02
5,0E-03
0,0E+00
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
Figure 23: Water vapour partial pressure p w along the reaction layer and its global relative
error.
III.26
p-n
68000
67500
67000
66500
66000
65500
65000
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
error-p-n
3,0E-03
2,5E-03
2,0E-03
1,5E-03
1,0E-03
5,0E-04
0,0E+00
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
Figure 24: Nitrogen partial pressure p n along the reaction layer and its global relative error.
III.27
p
102000
101900
101800
101700
101600
101500
101400
101300
101200
101100
101000
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
error-p
5,0E-07
4,0E-07
3,0E-07
2,0E-07
1,0E-07
0,0E+00
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
Figure 25: Total pressure p along the reaction layer and its global relative error.
III.28
imean = 2981,92
3300
3200
3100
3000
2900
2800
2700
2600
2500
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
error-i
2,0E-02
1,6E-02
1,2E-02
8,0E-03
4,0E-03
0,0E+00
0,0000 0,0001 0,0002 0,0003 0,0004 0,0005 0,0006 0,0007 0,0008 0,0009 0,0010
x
Figure 26: Current density i along the reaction layer and its global relative error. On top of the
upper frame is the value i mean = 2981.92.
III.29
Table 10 which corresponds to Table II1 gives the names of the geometrical parameters
in the Fortran code. The last letter c denotes that these values are constants. We do not
repeat here the units and values, they are given in Part II. Table 11 shows the 8 dependent
variables with their Fortran names. It combines Tables II3 and II5. In Part II the variables
have different symbols for gas channel, index K (for Kanal, German word for channel)
and index A for anode. However, these are in reality the same variables, only in different
domains. That they are computed by different PDEs in the different domains does not
matter. As the node number i tells where the variable is used, a different index K or A is not
necessary. In the black-box solver FDEM the variables for node i and solution component j
are denoted by u(i,j). In the Fortran code many loops run over the node number i. If the
III.30
parameter
thickness of anode
height of gas channel
length of gas channel
symbol
dA
dK
lK
name in code
dac
dkc
lkc
Table 11: Dependent variables with their Fortran names for node i, mf means mole fraction.
no.
1
2
3
4
5
6
7
8
variable
flow velocity in x-direction
flow velocity in y-direction
mf of methane
mf of carbon monoxide
mf of hydrogen
mf of carbon dioxide
mf of steam
pressure
symbol
ux
uy
YCH4
YCO
YH2
YCO2
YH2 O
p
name in code
u(i,1)
u(i,2)
u(i,3)
u(i,4)
u(i,5)
u(i,6)
u(i,7)
u(i,8)
innermost loop runs over the first index of a multidimensional array, in Fortran the elements
are accessed contiguously which is the most efficient access. Therefore the node number i
is in the first position in u(i,j).
no.
1
2
3
4
5
6
7
8
variable
ux
uy
YCH4
YCO
YH2
YCO2
YH2 O
p
our notation
ux
uy
Y3
Y4
Y5
Y6
Y7
p
equation
II(5) in form (17)
II(7) in form (18)
II(1) in form (19)
II(2) in similar form
II(3) in similar form
II(4) in similar form
II(8) in form (20)
II(6) in similar form to (18)
III.31
Table 12 shows the variables and the equations that we use in the channel. That we do
not have to write the chemical indices like CH4 in YCH4 we have replaced them by numbers
that correspond to the numbering in Table 12, thus instead of YCH4 we write Y3 , instead of
YCO we write Y4 etc. These simplified indices are also used for component indices in the
coefficients of the PDEs.
For the numerical solution we must differentiate out the product terms. Therefore the
first equation II(5) (continuity equation) in Table 12 becomes
uy
ux
ux +
+
uy +
=0.
x
x
y
y
(17)
The second equation in Table 12, the y-momentum equation II(7) becomes, as we must
take all terms of the PDEs to the l.h.s.
uy
uy
p 4 uy
2 ux
+
uy
+
ux
x
y
y
y 3 y
3 x
2
2 uy
1 2 ux
ux uy
4 uy
+
+
=0.
+
3 y 2
3 xy
x2
x y
x
(18)
The third equation in Table 12, the species balance II(1) for CH4 is with the simplified
chemical indices of Table 12
uy
ux
Y3
p
Y3
p
ux Y3
pY3
pux
uy Y3
pY3
puy +
x
x
x
y
y
y
D3,gas p
Y3
Y3 +
p +
y
y
y
2
p Y3 2 Y3
p
+
Y3 + 2
p =0.
D3,gas
y 2
y y
y 2
(19)
The equations no. 4 to 6 in Table 12 are obtained by replacing in Y3 and D3,gas the
index 3 by 4, 5, 6. The 7th equation II(8), Daltons law, becomes
Y3 + Y4 + Y5 + Y6 + Y7 1 = 0 .
(20)
The 8th equation II(6), the y-momentum equation, is written similarly to the x-momentum
equation (18).
In (19) appears the diffusion coefficient D3,gas which is given in II(9) which is in our
notation
1 Yi
,
j=i Yj /Dij
Di,gas =
III.32
i, j = 3, 4, 5, 6, 7 .
(21)
However, there appears in (19) also D3,gas /y. To get a corresponding expression
from (21), we write out the expression for D3,gas
D3,gas =
Y4
D34
1 Y3
6
+ DY36
+
Y5
D35
Y7
D37
1 Y3
,
sum 3
(22)
where we have abbreviated the denominator by sum i where sum means that the term
Yj /Dij with j = i, in our case Y3 /D33 , does not appear in sum 3. With the quotient rule
for differentiation we get from (22)
D3,gas
Y3
2
= (sum 3)
+
sum 3
y
x
1 Y5
1 Y6
1 Y7
1 Y4
+
+
+
.
(1 Y3 )
D34 y
D35 y
D36 y
D37 y
(23)
The other Di,gas /y are computed similarly. For the anode we will need also Di,gas /x
that we get by replacing /y by /x.
In (18) appears the coeffient (viscosity) that is given by II(11)
i
Y
i i ,
j Yj ij
i, j = 3, 4, 5, 6, 7 ,
(24)
where ij are values that are computed from II(12)II(14), but do not depend on the
variables Yj . However, there appears also /x and /y in (18). To get these expressions we write out (23):
Y3 3
+
Y3 33 + Y4 34 + Y5 35 + Y6 36 + Y7 37
sumps3
Y4 4
+
Y3 43 + Y4 44 + Y5 45 + Y6 46 + Y7 47
sumps4
..
.
Y7 7
.
Y3 73 + Y4 74 + Y5 75 + Y6 76 + Y7 77
sumps7
III.33
(25)
Y3
=(sumps3)2 sumps33
x
x
Y3
Y4
Y7
+ 34
+ + 37
Y3 3 33
+
x
x
x
2
sumps3
x
Y4
sumps4
Y4 4
+
sumps44
x
x
(sumps4)
..
.
(sumps7)
Y7
sumps7
Y7 7
sumps77
x
x
(26)
.
p
,
RM T
with
R
,
i Yi Mi
RM =
i = 3, 4, 5, 6, 7
(27)
R
Y3 M3 + Y4 M4 + Y5 M5 + Y6 M6 + Y7 M7
(28)
sumr
and we have
=
=
x
RT
p sumr
,
RT
p
Y3
Y4
Y7
sumr + p M3
+ M4
+ + M7
.
x
x
x
x
(29)
(30)
Table 13: Symbols for the parameters and their Fortran names for the PDEs and BCs of the
channel.
Fortran
Fortran
symbol name
symbol name
symbol
0
Di,gas
digas
YCH4
y03c
H2
0
Dij
dijc
YCO
y04c
H2 O
MCH4
m3c
YH0 2
y05c
CH4 /k
0
MCO
m4c
YCO
y06c
CO /k
2
0
MCO2
m5c
YH2 O
y07c
CO2 /k
MH2
m6c
CH4
sig3c
H2 /k
MH2 O
m7c
CO
sig4c
H2 O /k
0
u0x,max
u0xc
CO
sig5c
2
Here i or j must be replaced by i, j=3, 4, 5, 6, 7.
Fortran
name
sig6c
sig7c
epk3c
epk4c
epk5c
epk6c
epk7c
Fortran
Fortran
symbol name
symbol name
symbol
rho
ij
psiij
/x
/x
rhox
i
mui
/y
/y
rhoy
,i
omvi
Di,gas /x
RM
rm
mue
Di,gas /y
Here i or j must be replaced by i, j=3, 4, 5, 6, 7.
Fortran
name
muex
muey
digasx
digasy
But we must give for the coding also names for pre-computed coefficients that are
given in Table 14. Partly they are mere parameters, partly they contain the variables Yi .
These pre-computed values are stored in corresponding arrays for the nodes i.
Now we give the Fortran coding for the first 3 equations in Table 12. These are the
Newton residuals P1 u to P3 u in (10). But at first we must say how derivatives of variables
are denoted in the code: for node i the x- and y-derivatives of u(i,1), i.e. of the first solution
component (here ux ) are denoted and stored as ux(i,1) and uy(i,1), second derivatives are
denoted by uxx(i,1), uyy(i,1), uxy(i,1). Similarly the derivatives of other components u(i,j)
are denoted by ux(i,j), uy(i,j), uxx(i,j), uyy(i,j), uxy(i,j).
The Fortran code for P1 u, equation (17), looks like this for node i:
p(i, 1) = rhox(i) u(i, 1) + rho(i) ux(i, 1) +
rhoy(i) u(i, 2) + rho(i) uy(i, 2)
The Fortran code for P2 u, equation (18), looks like this for node i:
III.35
(31)
no.
1
2
3
4
5
6
7
8
variable
ux
uy
YCH4
YCO
YH2
YCO2
YH2 O
p
our notation
ux
uy
Y3
Y4
Y5
Y6
Y7
p
equation
II(46) in form (34)
II(48) in form (35)
II(42) in form like (34)
II(43) in form like (34)
II(44) in form like (34)
II(45) in form like (34)
II(49) all terms to l.h.s.
II(47) in form like (35)
(32)
(33)
Table 16: Symbols for the parameters and their Fortran names for the PDEs and BCs for the
anode.
Fortran
Fortran
symbol name
symbol name
symbol
ef f
Di,gas
digas
kp
kpc
UN ernst,CO
K
DCH4
dk3c
patm
patc
UN ernst,H2
K
0
DCO
dk4c
R
rc
UZelle,CO
K
0
DH
dk5c
RA,CO
raco
UZelle,H
2
2
K
DCO2
dk6c
RA,H2
rah2
YO2
K
DH
dk7c
RK
rk
2O
F
fc
Rohm
rohmc
jCO
jco
T
tc
rCH4
jH2
jh2
U
uc
rs
Here i must be replaced by 3, 4, 5, 6, 7.
Fortran
name
uneco
uneh2
uzcoc
uzh2c
yo2c
epsc
tauc
rch4
rs
of equation II(46) which is the first equation in Table 15, we multiply the equation by
RT which brings this term as a factor to the last term in II(46). We must differentiate
ef f
out all terms. In II(46) there are effective diffusion coefficients Di,gas
that are computed
from II(50). Here, i is the component index that takes values 3 to 7 in our notation. In
the equations and in the node we drop the upper index eff because in the anode domain
the Di,gas are computed from II(50) so that we do not need a special identification in the
formulas or code. So II(46) becomes
uy
ux
Y3
p
p
ux Y3
pY3
pux
uy Y3
pY3
x
x
x
y
y
2
2 Y3
p Y3
p
Y3
puy + D3,gas
+p 2 +
Y3 + 2
y
y 2
y y
y
2
p
2 Y3
Y3
p Y3
p
Y3 +
p + D3,gas
+
p
Y
+
2
+
3
y
y
x2
x x
x2
D3,gas p
RT
Y3
Y3 +
p
r3 = 0 .
x
x
x
dA
D3,gas
y
(34)
The equation II(48) which is the equation in the second position in Table 15 becomes
p
+ uy = 0 .
y kp
(35)
Now we have prepared the basic types of the equations for the anode, the other equations
have quite similar form, see Table 15. The functions rCH4 = r3 and rs that appear in the
equations II(43)(46) are computed from II(53), (54).
III.37
In Table 16 we have listed the parameters of Table II4 with their Fortran names. As
ef f
by Di,gas in the anode domain. However, they are
mentioned above, we named the Di,gas
computed there from II(50) in which appear the Di,gas computed from II(9) or from (22)
for the example of D3,gas . In order to distinguish the two Di,gas we denote the Di,gas from
Di,gas =
Di,gas
K
Di
K
Di,gas + Di
(36)
with i = 3, 4, 5, 6, 7 in our notation. From (36) we get the partial derivative (with
intermediate calculation)
DiK
Di
Di,gas
=
,
K
x
( Di + Di )2 x
D
(37)
with xi as shown for y3 in (22). The formula for Di,gas /y is obtained by replacing /x by /y.
The formal translation of these equations is executed as we have shown it above for the
translation of (19) to (33) so that there is no new information.
Now we have finished the discussion of the programming of the PDEs in the interior
of the domain for the channel, Table 12, and for the anode, Table 15. The next problem is
the discussion of the BCs. As we have 8 variables, we need at each boundary a system of
8 equations. Often at a boundary only a few conditions are precribed, e.g. the pressure at
the exit of the channel. Then we fill up the set of the equations by appropriate equations.
The domain with its boundaries is shown in Fig. 27. We discuss the boundaries without
their corners and discuss the conditions at the corners later.
We start with the BCs for boundary , the left entry of the anode, which is assumed
to be permeable. For ux we take the Darcy law II(47) in the form of the 8th equation of
Table 15. The value uy = 0 is prescribed which fulfils with the condition for p: p/y = 0
Darcys law II(48). The condition p/y = 0 means constant pressure along . We will
later take the pressure from the channel at the lower corner of . The Yi are prescribed by
their entry values Yi0 . In Table 17 the boundary equations for boundary are compiled.
For boundary , the boundary between anode and electrolyte, the conditions II(61)
(65) are given. For ux we prescribe Darcys law like on boundary . There is uy = 0
because of closed wall. For p we prescribe Darcys law II(48), which is no.2 in Table 15.
The other conditions are given. For Y4 we have II(62) which is in our notation and product
differentiation
1
D4
RT
Y4
p
Y4
+p
y
y
III.38
j4
1
=0.
2F
(38)
The other conditions are in similar form. The equations for boundary are shown in
Table 17. For these BCs the values jH2 = j5 II(67) and jCO = j4 II(68) must be precomputed for the nodes of boundary , except for the first node (left corner) where for the
entry values Yi a singular value would result. However, for the computation of j4 and j5
we need intermediate values RA,H2 , RA,CO , RK , UN ernst,H2 and UN ernst,CO which are
given in II(69)(73). These values are also pre-computed for the nodes of boundary . For
o
= 0 we get a singular value for II(73). Therefore we do not
the given entry value YCO
2
compute the j4 , j5 for the entry values of Yi .
For boundary , exit of anode, which is assumed to be open, there is prescribed only
p = patm . So we use for ux Darcys law II(47) (no.8 in Table 15) and for the equations in
position 27 those of the interior, Table 15. The equations are compiled in Table 17.
Now follow the BCs for the channel. At boundary , the right exit of the channel,
there only p = patm is prescribed, II(41). The other equations are like in the interior of the
channel, i.e. the same as positions 17 in Table 12. The boundary equations for boundary
are compiled in Table 18.
The BCs for boundary , the rib wall, are given in II(15)(21), these are the no-slip
conditions for the velocity components and zero gradient normal to the wall for the Yi . For
the pressure we take the x-momentum equation II(6) like in the interior (no.8 in Table 12).
The boundary equations for boundary are compiled in Table 18.
The boundary conditions for boundary , the channel entry, are given in II(34)(41).
Here values are prescribed for all variables except the pressure which is computed from the
x-momentum equation II(6) (no.8 in Table 12). For ux a parabolic profile is prescribed with
the maximal velocity ux,K,max in the middle of the channel and no-slip condition ux = 0 at
the rib wall and at the anode, equ. II(40). However, we must have the same ux in the channel
and in the anode at the interface DL, see Fig. 27. We compute at the anode interface ux from
Darcys law II(47) and we take this as slip for the channel flow at the anode interface. We
call this the modified parabola II(40). Else we would have a jump in ux at the interface.
The boundary equations for boundary are compiled in Table 18.
The next problem are the coupling conditions (CCs) at the dividing line DL, see Fig. 27.
The DL is geometrically one line, but logically two lines: side 1 belongs to the channel,
side 2 to the anode. The CCs are given in II(22)(33). However, we must clearly distinguish
which conditions are applied at side 1 and side 2. The variables are computed at one side
and their value is transferred to the other side. In the code the logical nodes on side 1 and
2 have different node numbers i, so the variables can be easily distinguished. In the writing
down of the CCs we distinguish the variables and also the variable-dependent coefficients
by the index K for channel and A for anode. We have compiled the CCs in Table 19 which
we now will discuss.
The value of ux is computed from Darcys law in x-direction II(47), the equation is that
of position 8 in Table 15, i.e. it is computed with the variables in the anode, index A. So this
is the equation for ux on side 2. This value is taken to the channel, side 1, by the condition
ux,K = ux,A , this is the slip velocity.
III.39
no. var.
boundary
boundary
boundary
1
ux no.8 in Tab. 15
no.8 in Tab. 15
no.8 in Tab. 15
2
uy
uy = 0
uy = 0
no.2 in Tab. 15
3
Y3
Y3 Y30 = 0
Y3 /y = 0
no.3 in Tab. 15
0
4
Y4
Y4 Y4 = 0
II(62) in form (38) no.4 in Tab. 15
5
Y5
Y5 Y50 = 0 II(63), similar form no.5 in Tab. 15
6
Y6
Y6 Y60 = 0 II(64), similar form no.6 in Tab. 15
7
Y7
Y7 Y70 = 0 II(65), similar form no.7 in Tab. 15
8
p
p/y = 0
no.2 in Tab. 15
p patm = 0
no.8, Tab. 15 means the equation in row 8 of Table 15.
Table 18: Boundary equations for the channel for the boundaries .
Similar conditions hold for uy . The Yi are computed in the anode from the equality
of the diffusion transport for channel and anode, II(22)(26). From II(22) we get in our
notation, differentiating out the products and taking all terms to the l.h.s., the equation for
Y3 :
D3,gas,A
Y3,A
pA
Y3,A + pA
x
x
D3,gas,K
Y3,K
pK
Y3,K + pA
y
y
=0.
(39)
Here D3,gas,A is computed from (36) with the Yi,A and D3,gas,K is computed from
type (22) with the Yi,K . Similar relations hold for the Y4 to Y7 . These values are taken by
Yi,K = Yi,A to the channel side 1.
The pressure pk is computed from the x-momentum equation II(6) in the channel which
is the equation of position 8 in Table 12. This value is transferred by pA = pK to the anode
side. This discussion of the CCs at the DL is an illustrative example of the use and of the
possibilities of a dividing line. It couples by CCs the two domains channel and anode in
III.40
Table 19: Coupling conditions (CCs) at dividing line DL. Index K is variable in channel, index
A in anode.
no. var.
side 1, channel
side 2, anode
1
ux ux,K ux,A = 0
no.8 in Tab. 15
2
uy uy,K uy,A = 0
no.2 in Tab. 15
3
Y3 Y3,K Y3,A = 0
II(22) in form (39)
4
Y4 Y4,K Y4,A = 0
II(23) in similar form
5
Y5 Y5,K Y5,A = 0
II(24) in similar form
6
Y6 Y6,K Y6,A = 0
II(25) in similar form
7
Y7 Y7,K Y7,A = 0
II(26) in similar form
8
p
no.8 in Tab. 12
PA PK = 0
no.8 in Tab. 12 means the equation in position 8 in Table 12.
which hold quite different PDEs. We get a global solution over the two domains with global
error estimate.
The next problem are the corners. They belong to two boundaries. At first sight it
seems to be of no importance which conditions we take. However, the present problem is
very critical and we had experienced that we got large errors and even divergence of the
Newton iteration if we did not take the correct conditions at the corners. We report here
those conditions that we take presently which does yet not mean that these are the optimal
conditions. One has to look at Fig. 27 for the corners.
Upper left corner, intersection of boundaries and : we take the equations of boundary , Table 17.
Upper right corner, intersection of boundary and : For position 1 to 7 we take the
equations of boundary , Table 17, for position 8 we take ppatm = 0. At first we had used
the conditions of boundary for 1 to 7 and p patm = 0 for 8. With this choice the errors
were 1 to 2 orders of magnitude larger. This shows the extreme sensibility of the problem
to such seemingly tiny changes and it shows that the corner nodes play an important role
for the solution of the whole problem.
Lower right corner, intersection of boundary and : we take the equations of boundary , Table 18, for positions 1 to 7 and for position 8 we take p patm = 0.
Lower left corner, intersection of boundary and : we take the equations of boundary .
The next problem are the left and right end nodes of the dividing line. Here we have for
the one geometrical node again two logical nodes at the end and we have to differ between
side 1 (channel) and side 2 (anode). We depict our choice of all CCs in Table 20, it is
self-explaining.
Now we have prepared all the information for the coding of the r.h.s. of the linear system
for the computation of the Newton correction. You have seen that there was still much to
do to formulate clearly the problem. The BCs and CCs have been given shortly in Part II,
but these are only the given conditions that must be filled up by other equations to get
III.41
Table 20: CCs at left and right end of DL. Index K is variable in channel, index A in anode.
left end
right end
side 1
side 2
side 1
side 2
no. var.
channel
anode
channel
anode
1
ux ux,K ux,A = 0
2
uy
uy = 0
uy = 0
position
position
3
Y3
Y3,K Y30 = 0 Y3,K Y30 = 0
1 to 7
1 to 7
4
Y4
Y4,K Y40 = 0 Y4,K Y40 = 0
like
like
0
0
5
Y5
Y5,K Y5 = 0 Y5,K Y5 = 0
side 1
side 2
6
Y6
Y6,K Y60 = 0 Y6,K Y60 = 0
Table 19
Table 19
7
Y7
Y7,K Y70 = 0 Y7,K Y70 = 0
8
p
no.8, Tab. 12
PA PK = 0 p patm = 0 p patm = 0
no.8, Tab. 12 means the equation in position 8 in Table 12.
the necessary 8 equations. And in Part II nothing is said about the corners. So the user
of FDEM is forced to formulate exactly all the details of his problem to get the equations
that finally are entered in FDEM. The formal translation to Fortran code is easy as has been
shown exemplarily in equs. (31)(33) for the channel flow.
The matrix Qd of the linear system is composed by FDEM from the difference formulas
and the Jacobian matrices, that are also needed for the computation of the error estimate,
see the basic paper [2] or the detailed report [1]. The meaning and the form of the Jacobian
matrices has been explained above in the equations (9)(11) and the corresponding context.
The system of PDEs for the SOFC is extremely nonlinear, above all by the nonlinear coefficients, e.g. the Di,gas or . It will be impossible to give here the formulas for the Jacobian
matrices, they would fill many pages with formulas that nobody would read. We rather want
to show the procedure how to get the Jacobian matrices in our case.
Let uns consider equ. (17), this is the first equation in Table 12 and therefore P1 u in
the notation of (10). for the Jacobians we must look for the dependencies of (17) from the
variables. Equ. (17) depends explicitly from ux = u1 , the first variable, from its x-derivative
ux /x = u1,x , from uy = u2 , from its y-derivative uy /y = u2,y in the terminology
of (9)(11). If would be a constant, these would be the (explicit) dependencies of the
operator P1 u (17). However, depends by (29) from sumr and sumr depends by (28)
from the variables Y3 = u3 to Y7 = u7 , and depends on p = u8 . These are implicit
dependcies. So we have purely formal
III.42
(p,
, sumr(Yi ),
(
)),
x
x
x
x
p
sumr Yi
(p, , sumr(Yi ),
(
))) .
y
y
r
y
(40)
P1
=
=
u1
ux
x
P1
P1
=
=
u2
uy
y
P1
ux
P1
=
=
u3
Y3
x
P1
to
similarly
Y4
ux
P1
P1
=
=
u8
p
x
( y
)
( x
)
uy
+
+ ux
+ uy
,
y Y3
Y3
Y3
P1
Y7
( y
)
( x
)
uy
+
+ ux
+ uy
.
y p
p
p
(41)
This ends the dependencies of P1 from the variables ux to p. Now come the dependencies from the x-derivatives of the variables:
P1
ux
x
P1
Y3
x
P1
p
x
=
= ux
=
(there is no dep. on
( x
)
Y3
x
( x
)
ux p
x
similarly for
III.43
uy
),
x
Y4
x
to
Y7
x
(42)
P1
uy
y
P1
Y3
y
P1
p
y
= ,
= uy
=
( y
)
Y3
y
( y
)
uy p
y
similarly for
Y4
to
Y7
(43)
All intermediate quantities are pre-computed and stored, e.g. sumr, sumr/x, sumr/y
and their derivatives with respect to the variables. From (28)(30) we get e.g.
p sumr
=
,
Y3
RT Y3
sumr
= M3 ,
Y3
x
(
1 p sumr
x )
=
,
Y3
RT x Y3
x
)
(
1 ( sumr
x )
p Yx ,
=
Y3
3
RT
x
x
( sumr
x )
Y3
x
(44)
= M3 .
This should be sufficient to show you how the Jacobians (11) are computed and why
we do not list them all here. We have developed a formalism where we reduce the complicated expression by a chain of intermediate values for which we apply the chain rule of
differentiation, so that even the most complex expressions get back to their simple origins.
Nevertheless, the formulation of the Jacobians is a hard mental training.
It is now also clear that the Jacobians are the main part of the implementationand
that they are rather error-prone. Therefore we have developed the Jacobi tester, where we
check the elements of the Jacobi matrices by difference quotients, see [1]. And indeed, there
were a lot of errors found, but finally the Jacobi tester did no longer show errors. Then the
numerical solution started.
However, the situation was not quite as simple as it seems to be. For the PEMFC
we got rather final models that worked nearly immediately. The situation is much more
complicated for the SOFC, especially because of the methane reforming process and by the
coupling of the two domains of anode and gas channel. Only by solving the PDEs we could
check their validity. At the same time also the used values for the many coefficients that
came from different sources were not safe values and had to be adapted. So we had to
improve the model of the PDEs and the values of the coefficients until we finally came up
with the model that has been presented here. Because changing the model means changing
the Jacobi matrices this is a hard work.
III.44
Table 21: Maxima of the solution components and of the relative global error estimates for the
computation with the 2 mm anode.
no.
1
2
3
4
5
6
7
8
var.
ux
uy
YCH4
YCO
YH2
YCO2
YH2 O
p
channel
solution
error
0.6733E+00 0.36E-02
0.1749E-01 0.39E-01
0.3300E+00 0.19E-02
0.2022E+00 0.15E-01
0.5754E+00 0.64E-02
0.8065E-01 0.82E-02
0.6700E+00 0.89E-02
0.1013E+06 0.11E-05
anode
solution
error
0.2287E-01 0.76E-01
0.1749E-01 0.91E-02
0.3300E+00 0.96E-02
0.2022E+00 0.42E-01
0.5752E+00 0.18E-01
0.1269E+00 0.36E-01
0.6700E+00 0.31E-01
0.1082E+06 0.16E-02
The numerical experiments showed that the SOFC PDEs are very critical because of
their extreme non-linearity which can be seen by the complicated Jacobian matrices. We
computed solutions for 3 different thicknesses dA of the anode, 0.05, 1 and 2 mm. We could
get the solution for 2 mm only by computing the solution for 0.05 mm, using this as starting
solution for 1 mm and this solution as starting solution for 2 mm. When we used the entry
values of x = 0 for the variables as starting values, the Newton iteration diverged. We used
the consistency order q = 4. With the order q = 2 we got errors 10 to 100 times larger.
When we used the order q = 6, the Newton iteration diverged, even with the solution of
order 4 as initial solution. We used a grid of 80 nodes in the x-direction. When we used for
accuracy tests a grid of 160 nodes we could get the solution of order 4 only by computing
a solution for order 2 and using this as starting solution for order 4. These experiences
illustrate how critical the SOFC PDEs are.
The results that we present below were computed on the HP XC6000 parallel supercomputer of the University of Karlsruhe. The processors are Intel Itanium2 processors with
1.5 GHz. We computed with 8 processors in parallel. The grid was 80 nodes in x-direction
and 41 nodes in y-direction in the channel and 41 nodes in the anode. We computed with
consistency order 4. We will present below the results for the anode with dA = 2 mm,
/ = 0.21, Rohm = 1.63 106 , see the section Parameter Variation in Part II. We
used as starting values the solution for dA = 1 mm and needed 4 Newton iterations. The
execution time on the master processor 1 was 510 sec, of which are needed 508 sec by the
linear solver LINSOL. As we have 80 82 nodes and 8 unknowns per node, the number of
unknows was 52480.
In Table 21 the maxima of the solution components and the maxima of the estimated
relative global errors are compiled for channel and anode. Before we will discuss these
numbers we look at the result plots in Figs. 2859. These figures are in grayscale in the
printed version of the report, but they are in colour in the on-line version. Therefore you
should look at these figures at the computer screen.
We discuss at first the results for the channel, Figs. 2843. Our discussion is purely
III.45
about the numerical behaviour. The physical behaviour is discussed in Part II. The plot for
ux , Fig. 28, shows that in the middle of the channel ux grows from the entry value of 0.5
m/s to its maximal value 0.67, see Table 21, and then goes down again. The error plot,
Fig. 29, shows that the maximal relative error of 0.36E-2 or 0.36% occurs in a small region
at the exit. Except this region the error is much smaller. The plot of uy , Fig. 30, shows that
the maximal value of 0.017 m/s, Table 21, occurs close to the entry in the upper part which
results from the influence of the anode, see Fig. 46. The error plot, Fig. 31, shows a small
error below 1%, the max. error of 1.7% of Table 21 is not visible and obviously occurs
in a very small region at the exit of the channel. The plots for the chemical components,
Figs. 3241, speak for themselves. It is interesting to see that YCH4 and YH2 O drop quite fast
from their entry values to lower values by the chemical reactions of the reforming process.
The errors of the Y s are quite small, their maximal values occur either in the entry or in the
exit region of the channel. The pressure is shown in Fig. 42, you can see in the scale to the
right of the plot that the change from entry to exit is only in the 5th digit, this is in the range
of the discretization error whose max. value is 0.11E-5 from Table 21. Therefore the error
plot, Fig. 43, looks a little weird, but this comes only from the smallness of the error.
Now we want to discuss shortly the results for the anode, Figs. 4459. The velocity
components ux and uy are small. The maximal error of ux of 7.6%, the largest error that
occurs in Table 21, is practically not visible in Fig. 45 because it occurs obviously just at the
entry. This indicates that there is a singularity or an incompatibility in the BCs at the entry
of the anode. The second largest error in Table 21 occurs in the anode for YCO . Fig. 51
shows that this error occurs in a small region at the exit of the anode, indicating a singularity
or incompatibility in the BCs at the exit of the anode. Also for the other Y s the max. errors
occur in very small regions at the entry or exit of the anode. The variation of the pressure in
the anode, Fig. 58, is much larger than in the channel and is surely caused by the diffusion
process of the chemical species.
From the numerical point of view these results are very convincing by the small estimates of the errors. The system of the SOFC PDEs is strongly non-linear and numerically
very critical, and it is solved on a coupled domain with different PDEs in channel and anode.
How should we trust our solution without the error estimate?
III.46
Channel
u-x
0.0010
0.65
0.60
0.55
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-u-x
0.0010
3.0E-03
2.5E-03
2.0E-03
1.5E-03
1.0E-03
5.0E-04
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 29: Contour plot of the global relative error of u x in the channel.
III.47
Channel
u-y
0.0010
0.016
0.014
0.012
0.010
0.008
0.006
0.004
0.002
0.000
-0.002
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-u-y
0.0010
3.5E-02
3.0E-02
2.5E-02
2.0E-02
1.5E-02
1.0E-02
5.0E-03
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 31: Contour plot of the global relative error of u y in the channel.
III.48
Channel
Y-CH4
0.0010
0.32
0.30
0.28
0.26
0.24
0.22
0.20
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-Y-CH4
0.0010
1.6E-03
1.4E-03
1.2E-03
1.0E-03
8.0E-04
6.0E-04
4.0E-04
2.0E-04
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 33: Contour plot of the global relative error of Y CH4 in the channel.
III.49
Channel
Y-CO
0.0010
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-Y-CO
0.0010
1.4E-02
1.2E-02
1.0E-02
8.0E-03
6.0E-03
4.0E-03
2.0E-03
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 35: Contour plot of the global relative error of Y CO in the channel.
III.50
Channel
Y-H2
0.0010
0.55
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-Y-H2
0.0010
6.0E-03
5.0E-03
4.0E-03
3.0E-03
2.0E-03
1.0E-03
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 37: Contour plot of the global relative error of Y H2 in the channel.
III.51
Channel
Y-CO2
0.0010
0.075
0.070
0.065
0.060
0.055
0.050
0.045
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-Y-CO2
0.0010
7.0E-03
6.0E-03
5.0E-03
4.0E-03
3.0E-03
2.0E-03
1.0E-03
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 39: Contour plot of the global relative error of Y CO2 in the channel.
III.52
Channel
Y-H2O
0.0010
0.65
0.60
0.55
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-Y-H2O
0.0010
8.0E-03
6.0E-03
4.0E-03
2.0E-03
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 41: Contour plot of the global relative error of Y H2 O in the channel.
III.53
Channel
p
0.0010
101356
101354
101352
101350
101348
101346
101344
101342
101340
101338
101336
101334
101332
101330
101328
101326
0.0009
0.0008
0.0007
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Channel
error-p
0.0010
1.0E-06
8.0E-07
6.0E-07
4.0E-07
2.0E-07
0.0008
0.0006
0.0004
0.0002
0.0000
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 43: Contour plot of the global relative error of p in the channel.
III.54
Anode
u-x
0.0032
0.0030
2.2E-02
2.0E-02
1.8E-02
1.6E-02
1.4E-02
1.2E-02
1.0E-02
8.0E-03
6.0E-03
4.0E-03
2.0E-03
0.0E+00
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-u-x
0.0030
7.0E-02
6.0E-02
5.0E-02
4.0E-02
3.0E-02
2.0E-02
1.0E-02
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 45: Contour plot of the global relative error of u x in the anode.
III.55
Anode
u-y
0.0032
0.0030
1.6E-02
1.4E-02
1.2E-02
1.0E-02
8.0E-03
6.0E-03
4.0E-03
2.0E-03
0.0E+00
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-u-y
0.0030
8.0E-03
7.0E-03
6.0E-03
5.0E-03
4.0E-03
3.0E-03
2.0E-03
1.0E-03
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 47: Contour plot of the global relative error of u y in the anode.
III.56
Anode
Y-CH4
0.0032
0.0030
0.32
0.30
0.28
0.26
0.24
0.22
0.20
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-Y-CH4
0.0030
9.0E-03
8.0E-03
7.0E-03
6.0E-03
5.0E-03
4.0E-03
3.0E-03
2.0E-03
1.0E-03
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 49: Contour plot of the global relative error of Y CH4 in the anode.
III.57
Anode
Y-CO
0.0032
0.0030
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-Y-CO
0.0030
4.0E-02
3.5E-02
3.0E-02
2.5E-02
2.0E-02
1.5E-02
1.0E-02
5.0E-03
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 51: Contour plot of the global relative error of Y CO in the anode.
III.58
Anode
Y-H2
0.0032
0.0030
0.55
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-Y-H2
0.0030
1.6E-02
1.4E-02
1.2E-02
1.0E-02
8.0E-03
6.0E-03
4.0E-03
2.0E-03
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 53: Contour plot of the global relative error of Y H2 in the anode.
III.59
Anode
Y-CO2
0.0032
0.0030
0.12
0.11
0.10
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-Y-CO2
0.0030
3.0E-02
2.5E-02
2.0E-02
1.5E-02
1.0E-02
5.0E-03
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 55: Contour plot of the global relative error of Y CO2 in the anode.
III.60
Anode
Y-H2O
0.0032
0.0030
0.65
0.60
0.55
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-Y-H2O
0.0030
3.0E-02
2.5E-02
2.0E-02
1.5E-02
1.0E-02
5.0E-03
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 57: Contour plot of the global relative error of Y H2 O in the anode.
III.61
Anode
p
0.0032
0.0030
107000
106000
105000
104000
103000
102000
101000
100000
99000
98000
97000
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Anode
error-p
0.0030
1.4E-03
1.2E-03
1.0E-03
8.0E-04
6.0E-04
4.0E-04
2.0E-04
0.0028
0.0026
0.0024
0.0022
0.0020
0.0018
0.0016
0.0014
0.0012
0.0010
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Figure 59: Contour plot of the global relative error of p in the anode.
III.62
jmean = 6860,67
10000
8000
6000
4000
2000
0
0,000
0,005
0,010
0,015
0,020
0,025
0,030
0,035
0,020
0,025
0,030
0,035
0,020
0,025
0,030
0,035
jCO
4000
3000
2000
1000
0
0,000
0,005
0,010
0,015
jH2
6000
5000
4000
3000
2000
1000
0
0,000
0,005
0,010
0,015
III.63
Channel
u-x
u-y
Y-CH4
Y-CO
Y-H2
Y-CO2
Y-H2O
p
Anode
u-x
u-y
Y-CH4
Y-CO
Y-H2
Y-CO2
Y-H2O
p
80x41
max.
mean
3,6E-03
9,4E-04
2,6E-01
4,3E-03
8,1E-04
5,0E-04
5,1E-03
1,2E-03
1,7E-03
6,3E-04
3,8E-03
8,0E-04
2,8E-03
6,6E-04
1,3E-06
8,0E-07
160x81
max.
mean
9,1E-02
4,6E-03
1,1E+00
7,1E-03
8,2E-04
4,2E-04
3,3E-02
8,5E-03
1,3E-02
5,4E-03
2,0E-02
5,8E-03
1,8E-02
6,0E-03
7,9E-06
5,2E-06
80x41
max.
mean
1,9E-01
2,9E-03
2,0E-02
3,3E-03
7,8E-04
4,8E-04
6,1E-03
1,2E-03
1,7E-03
6,1E-04
7,2E-03
7,7E-04
4,0E-03
6,5E-04
4,0E-05
1,5E-05
160x81
max.
mean
1,7E-01
1,4E-03
1,9E-02
3,6E-03
8,2E-04
4,2E-04
3,3E-02
8,4E-03
1,3E-02
5,3E-03
2,2E-02
5,9E-03
1,8E-02
5,8E-03
9,6E-05
2,8E-05
Figure 61: Comparison of max. and mean error for grid 80 41 and 160 81 for channel and
anode.
For the output of a SOFC the current densities at the interface of the anode to the electrolyte are an essential result. Fig. 60 shows the current densities j [A/m2 ], equ. II(74),
jCO , equ. II(68) and jH2 , equ. II(67). We have also computed the arithmetic mean of j of
the node values at boundary which is jmean = 6860.67 A/m2 . For these quantities we
cannot give error estimates as we did it for the PEMFC in Fig. 26, there the current density
is named i. For the PEMFC i was a variable that had a meaning only at the reaction layer. If
we needed error estimates for jCO and jH2 (j is trivial by equ. II(74)), we had to introduce
them as variables that have a physical meaning only at the interface to the electrolyte, just
as we did it for i in the PEMFC. However, this would increase the number of variables to
10. But as the entries in the formulas for the js have errors in the 1% range, we can assume
that the js are accurate also in the 1% range.
As mentioned above the system of SOFC PDEs is numerically very critical and there
is the suspicion of a hidden singularity or incompatibility of the BCs. If this would be the
case the errors would grow if we use a finer grid because the singularity is better detected,
or at least would not go down as expected from the consistency order q = 4. Halving the
step size would theoretically reduce the error by a factor (1/2)4 = 1/16 or roughly one
order of magnitude. Therefore we computed the solution with a grid of 160 81 for the
channel and for the anode and compared the errors to those of 80 41. As we have seen
in the (coloured) figures the error is usually small in the main part of the domain and gets
larger values only in very small regions at the entry or exit. Therefore we investigated also
the mean error, i.e. the arithmetic mean value of the errors in the domain. The max. and
mean errors for grid 80 41 and 160 81 for channel and anode are depicted in Fig. 61.
What can we learn from these 64 numbers? Where we have small max. errors, the mean
error is not much smaller, so we have a flat error distribution. However, where we have
III.64
large max. errors (compared to their neighbouring values), the mean error is by one to two
orders of magnitude smaller, see the following examples in Fig. 61:
1.1E0 7.1E-3 ,
channel, uy : 2.6E-1 4.3E-3,
1.7E-1 1.4E-3 .
anode,
ux : 1.9E-1 2.9E-3,
This shows clearly that there are error peaks that indicate a singularity or incompatibility, presumably in the BCs. This is what the comparison of max. and mean error tells
us.
Now we compare the errors of the two grids. Where we have large errors, e.g. for uy in
the channel, the max. and mean errors for the fine 160-grid are both larger than those of the
coarser 80-grid. For uy we have for max. and mean error:
2.6E-1 1.1E0,
4.3E-3 7.1E-3 .
All max. errors go up from the 80-grid to the 160-grid except ux , uy in the anode. Also
most of the mean errors go up from the 80-grid to the 160-grid. This shows that there is not
the theoretically expected behaviour of an error reduction by a factor of 1/16. This confirms
the assumption of a singularity/incompatibility in the system of PDEs and BCs and CCs.
This needs further investigation of the problem. However, because the financial means were
very short there was not the time to do this. Nevertheless we can trust our results for the
80-grid. This ends the discussion of numerical solution of the SOFC PDEs.
4 Concluding remarks
In chapter 1 we gave a short survey of the FDEM program package and its possibilities. The
essential advantage of this exceptional black-box PDE solver is the error estimate. This is a
unique feature. To our knowledge nobody else can solve such general nonlinear systems of
PDEs with a reliable error estimate. FDEM is an unprecedented generalization of the FDM.
As large technical problems make the use of large parallel supercomputers mandatory: Also
in this respect the FDEM program package is exemplary because it is efficiently parallelized
with MPI for shared and distributed memory parallel computers.
In Chapter 2 we report on a fruitful cooperation with the ZSW Ulm to solve the PDEs
of PEMFCs. We solved several models with increasing complexity. Only if one solves the
PDEs of a certain model numerically, the properties of the model come to light. Our partner,
in this case ZSW, is responsible for the PDEs, we are responsible for the numerical solution.
The partner will immediately ask us: How good is the solution, can I trust it? He takes it for
granted, that we can solve his PDEs. But he expects also that we can tell him the quality of
the solution. As the PEMFCs were rather critical, ZSW asked us for solutions on finer grids
until they were satisfied with the errors. In chapter 2 we described how we implemented
their most elaborate model in FDEM and we discussed the properties of the PDEs from the
point of view of numerics.
In Chapter 3 we report about the numerical solution of the SOFC PDEs that were partly
established and improved in a vital cooperation with the IWE of our university. The problem
of SOFCs is rather complicated by the methane reforming process and hence there are many
parameters in the system that must carefully be adapted to get useful solutions. So much of
III.65
the project time was used to tune the model. We report here on the implementation of the
final model. Again the knowledge of the error of the numerical solution was an invaluable
help that opened the eyes to us and to IWE. As in the PDEs are many coefficients that
themselves depend in a multistage way on the variables, the generation and implementation
of the Jacobian matrices that describe the dependency of the PDEs from the variables was a
really hard task. The PDEs of the SOFC are numerically very critical as has been shown in
Chapter 3.
For us the cooperation with the ZSW and IWE was extremely useful to demonstrate
with practical industry-near problems the usefulness of FDEM and we want to thank the
colleagues of these institutes for their engaged cooperation.
As it will be our task to demonstrate the usefulness of FDEM by possibly many examples, we are looking for further cooperation partners to solve in common research projects
their problems. The problems may be any type of numerical simulation. If you are interested you may contact us.
References
[1] W. Schonauer, T. Adolph, FDEM: The evolution and application of the Finite Difference Element Method (FDEM) program package for the solution of partial differential equations. This report will be available as hard copy and in electronic form at the
libraries of the Universities of Karlsruhe and Hannover (there TIB, Technische Informationsbibliothek) and online at
http://www.rz.uni-karlsruhe.de/rz/docs/FDEM/Literatur/fdem.pdf
[2] W. Schonauer, T. Adolph, FDEM: How we make the FDM more flexible than the
FEM, J. of Computational and Applied Mathematics 158 (2003), pp. 157167.
[3] LINSOL, see
http://www.rz.uni-karlsruhe.de/rd/linsol.php
Remark: The on-line version of this report (with coloured figures) will be available at
http://www.rz.uni-karlsruhe.de/rz/docs/FDEM/Literatur/fuelcells.pdf
III.66