Electrochimica Acta 54 (2009) 6913–6927

Contents lists available at ScienceDirect

Electrochimica Acta
journal homepage: www.elsevier.com/locate/electacta

On the modeling of water transport in polymer electrolyte membrane fuel cells

Hao Wu a , Xianguo Li a,∗ , Peter Berg b
a
Department of Mechanical Engineering, University of Waterloo, Waterloo, ON, Canada
b
Faculty of Science, University of Ontario Institute of Technology, Oshawa, ON, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Water management is a critical issue in polymer electrolyte membrane (PEM) fuel cells, and water trans-
Received 6 February 2009 port through the membrane, catalyst layer and gas diffusion layer has significant impact on the cell
Received in revised form 26 June 2009 performance and durability. In this study, the mechanism of water transport processes in PEM fuel cells
Accepted 27 June 2009
has been analyzed through 3-D unsteady non-isothermal simulations, along with a comprehensive exam-
Available online 5 July 2009
ination of various modeling approaches in literature. It is shown that the finite rates of sorption/desorption
of water in membrane affect the cell current output and the cell response time. Water dissolved in the
Keywords:
membrane should be taken as the proper mechanism of water formation in the cathode of practical PEM
Non-isothermal
Water production assumptions
fuel cells. Capillary pressure and relative permeability have significant impact on the distribution of liq-
Membrane sorption/desorption uid water saturation and transport, implying a need for their determination under specific PEM fuel cell
Capillary pressure conditions.
Relative permeability © 2009 Elsevier Ltd. All rights reserved.

1. Introduction
With the progress in computing technology over the past two
decades, numerical simulation is playing an increasingly impor-
tant role in the development of proton exchange membrane fuel
cells (PEMFCs). It helps better understand the underlying chemi-
cal/electrochemical mechanisms and transport phenomena within
the cells which are hardly accessible by experimental means.
The transport phenomena inside a PEMFC are complex and many
uncertainties related to electrochemical dynamics and transport
behavior in micro-scale porous domains are still not fully under-
stood and are undergoing active investigation. Accompanying this
process, various modeling assumptions and approaches have been
proposed in the literature, some of which requires further inves-
tigation to confirm its validity, and this is the focus of the present
study.
Most of the modeling discrepancies in literature originated from
the intricate water transport in different cell regions. Generally, pre-
vious water transport studies can be categorized into two main
groups: water transport through the polymer electrolyte mem-
brane; and two-phase water transport in the catalyst and backing
layer. Each of these two main groups can be further divided into
several subgroups.
The first main group of modeling efforts, water transport in
polymer electrolyte membranes, can be classified into three mod-
eling approaches, namely, the diffusive model [1–4], convective
∗ Corresponding author.
E-mail address: x6li@uwaterloo.ca (X. Li).
(or hydraulic) model [5–7], and chemical potential model [8–13].
A detailed review regarding the diffusive and hydraulic model is
available, e.g., [14]. Hence, only the chemical potential model is
briefly summarized below. In the diffusive/hydraulic models, the
proton concentration is assumed constant over the membrane
domain and the electric potential within the membrane is exclu-
sively determined through Ohm’s law. In the chemical potential
model, however, the concentration of mobile protons (or hydro-
nium) is assumed to vary with the water concentration. Utilizing
the dusty fluid model [8,9], generalized Stefan–Maxwell equa-
tions [10,11], or concentrated solution theory [12,13], the coupled
system (water, protons, electric potential) is then solved simulta-
neously. Generally speaking, the chemical potential model can be
considered as a superclass of diffusive/hydraulic models; the diffu-
sive/hydraulic models are only valid in certain situations (constant
proton concentration), while the chemical potential model is a more
comprehensive approach which applies to a much larger range.
Nevertheless, all above mentioned chemical potential models are
invariably confined to the membrane region along with many sim-
plifications. Further, several parameters and correlations related to
this model class remain unknown, such as the diffusion coefficient
of hydronium, the interaction properties of water and hydronium
with the solid matrix, etc. Hence, the application of the chemical
potential model in full cell modeling needs to be explored further.
The second main group of modeling efforts, two-phase water
transport in the catalyst and backing layer, can also be divided
into three modeling approaches. It includes the mixture model,
the two-fluid model, and the volume of fluid (VOF) model. Wang
and Cheng [15] were one of the first groups to apply multi-phase
mixture theory to PEMFC research. The mixture model is a kind of

0013-4686/$ – see front matter © 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.electacta.2009.06.070
6914 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927
single-fluid model. It solves a single set of conservation equations
for the phase mixture assuming phase equilibrium. Subsequently,
the volume fraction of the phases, as well as the relative velocity
among different phases, are obtained in a post-processing man-
ner. The two-fluid model, on the other hand, solves individual sets
of equations for each phase while the interaction among different
phases is explicitly taken into account through limited phase trans-
fer terms. In PEMFC modeling, the latter model is usually simplified
by combining the continuity and momentum equation with the
help of Darcy’s law and a capillary pressure function. Compared to
the mixture model, the advantage of the two-fluid models is that
only one extra equation for liquid saturation is added, while allow-
ing for the simulation of non-equilibrium phase transfer processes.
Here, however, the choice of the empirical expressions for relative
permeability and capillary pressure are crucial and many different
relationships have been proposed in literature.
He et al. [16] developed a 2-D, two-fluid model for the cath-
ode side with interdigitated channels. Considerable simplifications
with a constant interfacial drag coefficient and liquid diffusivity
are made which is deemed reasonable only for small saturations.
Later, Natarajan et al. [17] improved He et al.’s model by utilizing a
linear function for relative permeability and an exponential expres-
sion for the capillary pressure. This revised model can be used to
study conventional gas distributors where the liquid saturation is
more significant. In the model of Wang and Wang [18], power law
relations and Leverett J-functions were used for the relative per-
meability and capillary pressure, respectively. Leverett J-functions
are traditionally used in geological engineering to study water
transport behavior in materials that have homogeneous wetting
properties, like soil and rock. However, due to the lack of experi-
mental data in the early days, it has been extensively used in PEMFC
modeling as an approximation. This situation is changed with the
emergence of several recent experimental studies [19–24], that are
employed in the present study.
Recently, the VOF model has attracted increasing attention from
the PEMFC modeling community. The biggest advantage of a VOF
model is its ability to trace the trajectory of the liquid droplet move-
ment. However, due to the nature of the extremely small time-steps
and intensive computing related to VOF methods, its application so
far has been restricted to investigating the physical properties of
certain components, such as the permeability, the gas diffusivity in
the backing layer, or the analysis of liquid behavior in the porous
electrode [25] or gas flow channels [26,27]. Full geometry PEMFC
Fig. 1. Schematic of (a) computational domain; (b) membrane electrode assembly (MEA).
models that incorporate the VOF approach have not been used to
the best of the authors’ knowledge.
In summary, the “chemical potential + VOF” approaches are
promising and may finally evolve to be the main features of the
next generation of PEMFC models. At the current stage, however,
the “diffusive + two-fluid” type models still dominate. In fact, many
“diffusive + two-fluid” type full cell models have been developed in
the literature. Nonetheless, most of them were built with various
simplifications and many of them even with nonrealistic assump-
tions. For instance, experimental works [28–30] have shown that
the finite rate membrane sorption/desorption process is one of the
major transient transport process within PEMFCs, but this process
has been generally neglected in most of the previous “diffusive
+ two-fluid” models. Furthermore, the fallacious assumption that
water is produced in vapor phase during the oxygen reduction reac-
tion (ORR) has repeatedly appeared in these models. Therefore,
the present study will still concentrate on the “diffusive + two-
fluid” type models. By examining several widely used assumptions
and approaches, the goal is to improve these models, and shed a
fresh light onto future modeling endeavors, avoiding inappropriate
assumptions and approaches. In the next section, the basic formula-
tion of the “diffusive + two-fluid” model will be presented in terms
of conservation laws of mass, momentum, species and energy. Then,
comparisons among different approaches will be made in Section
4 and conclusions will be drawn accordingly in Section 5.
2. Model formulation
A three-dimensional model with a typical straight channel is
sufficient for comparison purpose, as outlined in the last section. It
has been shown in [31] that for bipolar plate made of graphite, its
effect on performance is negligible for such single-channel models
because of its high electrical and thermal conductivity. Further-
more, the effects of the cooling channel can also be minimized
by appropriate specifications of temperature boundary conditions.
Thus, both the bipolar plate and the cooling channel have been
excluded from the current study. However, the effect of bipolar plate
and cooling channel can have major impacts on performance for
bipolar plates made of other alternative materials with low elec-
trical and thermal conductivity, and for a stack due to cooling, end
plate effects and in-plane currents resulting from cell coupling. The
computational domain used in this study is shown in Fig. 1.
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927 6915

Fig. 2. Schematic of equilibrium water sorption model (a) ill-posed; (b) well-posed.

2.1. Water transport mechanisms
In the operation of PEMFC, water may exist in three different
phases depending on the region, i.e., water vapor, liquid water
and dissolved (ionomer) water. To fully appreciate the multi-phase
transport characteristics of the “diffusive + two-fluid” model, water
transport within each component of the cell is described below.
2.1.1. Water vapor transport
For better cell performance, the gaseous reactants are usually
humidified before entering the gas flow channel. The water vapor
diffuses between the catalyst layer (CL) and the gas flow channel,
i.e. through the gas diffusion layer (GDL). Compared to convective
effects, diffusion is relatively weak in the gas flow channel, but it
is more dominant in the porous backing layer (GDL and CL). Fick’s
law is usually used to describe such species diffusion process. It can
serve as a good approximation for a multi-species system, provided
that a corrected diffusivity is used [32]. Consequently, water vapor
transport is governed by the following general form of a convection-
diffusion equation
∂ they are described below.
(Cv ) + ∇ · (−Dv ∇ Cv ) + ∇ · (u
 g Cv ) = Sv ,
∂t
where Cv is the water vapor concentration,  is the gas phase poros-
 g is the gas phase velocity, and
ity, Dv is the water vapor diffusivity, u
Sv is the water vapor source term.
2.1.2. Dissolved water transport—diffusive model
The pore sizes of the polymer electrolyte are of the order of only
10 nm. Clusters of water molecules tend to be localized and less
connected in such small pores. Therefore, water is usually assumed
to be in a “dissolved” phase in the electrolyte region rather than in a
liquid state. At the anode catalyst layer, water vapor is absorbed into
the electrolyte and the water molecules tend to move with the pro-
tons towards the cathode catalyst by means of the electro-osmotic
drag. At the cathode catalyst, water is generated at the solid cata-
lyst surface and the enhanced local water concentration tends to
counteract the water movement from the anode side. If the con-
centration gradient is large enough, water will diffuse back to the
anode side.
In the diffusive model, the above process is described by
∂
n 
d
(m Cd ) + ∇ · (−Dd ∇ Cd ) + ∇ · im = Sd , (2)
∂t F
where Cd is the dissolved membrane water concentration, m is the
volume fraction of the polymer membrane (m < 1 in the catalyst
layer), Dd is the dissolved water diffusivity in the electrolyte, im is
the membrane phase current density, and Sd is the source term of
the dissolved phase.
The numerical implementation of the above equation has been
conducted in two different approaches with potential pitfalls, and
(1) Approach 1: Equilibrium membrane water sorption
The equilibrium water sorption approach can be traced back to
the very beginning of PEMFC research with the pioneering work of
Springer et al. [1], and it is still overwhelmingly adopted by today’s
models. The main assumption of this approach is that membrane
water uptake occurs instantaneously and the dissolved water is
always in equilibrium with other water phases in the catalyst layer,
i.e. vapor and liquid. The mechanism of equilibrium water sorption
is schematically illustrated in Fig. 2(a). It shows that the water con-
tent in the dissolved phase,  (or Cd ), is always at its equilibrium
value, e , determined by the vapor phase (Cv and T) and described

Table 1
Empirical and constitutive relationships used in the present study.

Description Expression
   
√ a − 0.89
0.3 + 6a [1 − tanh (a − 0.5)] + 3.9 a 1 + tanh , if s ≤ 0,
Membrane water content [2,13] = 0.23
16.8s + |(a=1) (1 − s), if s > 0
m 
Dissolved water concentration (mol m−3 ) [14] Cd = EW
−5 (T −273)2 +1.445×10−7 (T −273)3 ]
Water saturation pressure (atm) [1] P sat = 10[−2.1794+0.02953(T −273)−9.1837×10
cv T
Water activity [14] a= P sat
  0.15  −2.5 
Membrane water diffusivity (m2 s−1 ) [2] Dd = 4.1 × 10−10 25
1 + tanh 1.4
2.5
Electro-osmotic drag coefficient (H2 O/H+ ) [1] nd = 22
 
Proton conductivity (S m−1 ) [1] m = (0.005139 − 0.00326) exp 1268 1
303
− 1
T
T
Henry’s law [9] Cid = Ci
H
 Cd
i  ˛a nF   
−3 2.5
Bulter-Volmer reaction rate (A m ) [14] Ri = (1 − s) aj0,ref i
Ci,ref
exp T
i − exp − ˛c nF
T
i

Current density (A m−2 ) [14] J = Js + Jm = − eff ∇ − m

eff
∇
 s
  Pref s m
1.5
Effective gas diffusivity (m2 s−1 ) [38] Dieff = Di T
Tref P
((1 − s))

Effective thermal conductivity (W m−1 K−1 ) [14] keff = l kl + m km + s ks + g kg

6916 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927

by an algebraic relation, the sorption isotherm

f1 (Cd ) =  = e = f2 (Cv , T ). (3)

In this equation, the presence of liquid water, which gives rise

to the controversy over Schroeder’s paradox (see further below),
is neglected. The functions f1 and f2 need to be determined (at
equilibrium) experimentally. They are listed in Table 1.
The problem arises when one needs to define the boundary con-
ditions of the system at points 2 and 3 (refer to Fig. 2a). For lack
of better knowledge, it seems appropriate to set both the dissolved
water flux, which enters the GDL (point 3), and the water vapor flux,
which enters the bulk membrane (point 2), to zero, represented
by zero gradients and assuming diffusion dominates. However, at
constant temperature and given Eq. (3) which implies
df1 dCd d df2 dCv
= = , (4)
dCd dy dy dCv dy
we see immediately that this would also imply zero gradients at
points 1 and 4, and hence zero fluxes also. This leads to an ill-posed
(steady-state) model since water is produced in the catalyst layer
but it is prevented from leaving the domain.
In the model of Springer et al. [1] and some other earlier 1-D/2-D
models, the catalyst layer is treated as an interface and the dissolved
water transport in the CL is entirely neglected. Therefore, the ill-
posed issue is simply bypassed in these models. Fig. 3. Numerical implementation of the water source owing to back diffusion.
For subsequently developed 2-D/3-D models, however, the cat-
alyst layer is usually explicitly accounted for with finite thickness transport equation, Eq. (6), which accounts for the water gain or
and, thus, this issue becomes unavoidable. To tackle this problem, loss from back diffusion. In some commercial CFD software pack-
Kulikovsky [2] proposed an approach which solves the dissolved ages, such as the package FLUENT used in this study, it is hard to
water transport only in the bulk membrane region with its bound- implement internal flux boundary conditions. In such cases, the
ary values being determined by the local water activities at the mass fluxes are usually converted to source/sink terms and applied
adjacent CL, an approach similar to Springer et al.’s [1]. In the only to the first layer cells adjacent to the interface. As an example,
catalyst layer, the dissolved water concentration is converted math- the implementation of the water source term in the anode catalyst
ematically into water vapor concentration by use of an empirical layer (ACL) is illustrated in Fig. 3. In the first layer grids of the ACL,
sorption isotherm relation [2] both drag force and back diffusion are present, and the back diffu-
dCd m P sat d sion is determined by the conditions of the first layer grids at the
= , (5) membrane side via
dCv EW T da
A
where m denotes the dry membrane density, EW is the equiv- Sdiff = (−Dd ∇ Cd + ndim ) . (9)
alent molecular weight (e.g., 1.1 kg mol−1 for Nafion) of the dry V
membrane, P sat is the saturation pressure, and a is the local water The first term on the right-hand side represents the water back
activity. diffusion from the membrane region, the second term represents
Correspondingly, the governing equation of water vapor trans- the electro-osmotic drag force in the membrane, A is the interfa-
port, Eq. (1), is revised to account for the dissolved water transport cial surface area of the unit cell, and V is the volume of the unit
in CLs and takes the shape of an effective water transport, including cell. In the other layer grids of the ACL, the water source includes
both dissolved water and vapor: the electro-osmotic drag only, as shown in Fig. 3.
∂ eff On the other hand, the same mathematical conversion technique
( Cv ) + ∇ · (−Dveff ∇ Cv ) + ∇ · (u
 g Cv ) = Sv , (6) can be applied to the dissolved water in the bulk membrane region
∂t
as well. As such, the dissolved water in all regions is converted
where the effective porosity (eff eff
g ) and effective diffusivity (Dv ) are mathematically to water vapor and the dissolved water transport
determined as equation, Eq. (2), can be eliminated entirely [4,14]. In such mod-
 dC  els, the back diffusion process is self-consistent. Hence, no extra
d
eff =  + m , (7) boundary conditions or source terms are required.
dCv
 dC  It should be noted that even though the equilibrium model
d
Dveff = Dv + Dd . (8) is well-posed when following this mathematical treatment, each
dCv water phase is not modeled explicitly and its boundary condi-
With such a treatment, the equilibrium model now becomes tions are not addressed explicitly. Therefore, the effective water
well-posed and the implementation is schematically illustrated flux within the CL and across the boundaries really consists of two
in Fig. 2(b). Since the dissolved water transport in the catalyst fluxes: water in the ionomer and water vapor in the gas pores. How-
layer, which is rolled into an effective total water transport, would ever, since only a combination of the two is modeled, we have no
be decoupled from that of the membrane using this formula- explicit control over the fluxes of each phase at the membrane–CL
tion, an explicit coupling between these two regions is necessary. and CL–GDL interfaces. Physically speaking, this is a big disadvan-
This is usually done by specifying boundary conditions at the tage of this combined model. A more important drawback of the
membrane–CL interface. Two boundary conditions are required equilibrium model is that the last term (d/da) in the mathemat-
here: a Dirichlet boundary condition for the dissolved water trans- ical conversion equation, Eq. (5), is not strictly valid for two-phase
port equation, Eq. (2), and a mass flux condition for the vapor modeling since the membrane water content, , now becomes a
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927
function of both water activity, a, and liquid saturation, s (refer to
Table 1 for the expression of ). However, Eq. (5) has so far been
adopted in all the previous two-phase equilibrium models as an
approximation.
Approach 2: Non-equilibrium membrane water sorption
Non-equilibrium water sorption model circumvents all draw-
backs of the equilibrium model, and it is not only mathematically
better posed but also physically more meaningful.
It has been shown that the time scale for membrane to reach its
sorption equilibrium state in humid air is on the order of 100–1000 s
[28,29]. In the recent work of Onishi et al. [30], the membrane water
uptake takes even longer. They reveal that the water content of a
membrane (with an appropriate thermal history) which is in con-
tact with saturated vapor, is actually the same as one in contact
with liquid water as the membrane relaxes to its equilibrium state
over several weeks or months. Therefore, the so-called Schroeder’s
paradox does not appear to exist. All these studies indicate that
the equilibrium sorption assumption made in most of the previous
studies is inaccurate or invalid for PEM fuel cell catalyst layers, given
the timescales of key water transport and production mechanisms
[9].
Rather than making an equilibrium assumption for the water
content, Berg et al. [9] proposed an approach in which the flux of
water into and out of the electrolyte is assumed to be proportional
to the difference between the local ionomer water content and its
equilibrium sorption value. A similar approach was later adopted in
[33–35] but, in general, the study of non-equilibrium water sorp-
tion is still relatively new and many characteristics of this approach
remain poorly understood.
Even though the underlying physical phenomena of non-
equilibrium sorption is much more complicated than that of an
equilibrium system, the numerical implementation of the non-
equilibrium approach is much easier. The dissolved water transport
equation, Eq. (2), is solved in the entire computational domain with
two source terms imposed in the catalyst layer
 m
Svd = a (e − ), (anode)
Sd = EW (10)
R m
Sd,reac + Sld = i + d (e − ), (cathode)
2F EW
where Sd,reac is the dissolved water production from the electro-
chemical reaction which will be described in detail later; e and
 are the equilibrium and the actual membrane water content,
respectively; a and d are the rate coefficient of membrane absorp-
tion and desorption, respectively, and their numerical values at
the anode and cathode catalyst layers need to be established, as
described later. It should also be noted that no liquid water is found
at the anode side in this study. Thus we assume the water sorp-
tion in the ACL is from the vapor phase (Svd ). On the other hand,
fully humidified inlet conditions are assumed in this study and,
hence, dissolved water only enters the liquid phase during water
desorption in the CCL (Sld ).
A recent study of Satterfield and Benziger [29] shows that the
physical mechanism of membrane absorption is different from that
of desorption which is mainly limited by the interfacial mass trans-
port. Water absorption presents a two-step behavior: uptake for
the initial 35% of water absorption is described by the same inter-
facial transport rate coefficient as that of desorption, while for the
value above 35%, water absorption is controlled by the dynamics of
membrane swelling and relaxation. It is found that the absorption
process is 10 times slower than that of desorption in the second
stage. In the numerical modeling, however, it is hard to deter-
mine the 35% threshold without knowing the final hydration level
a priori. Therefore, identical rate constants for absorption and des-
orption have been used in our base case to simulate the first stage of
the sorption dynamics. The second stage of water sorption, reflected
6917
by reducing the absorption rate coefficient to (a = 0.1d ), has only
been investigated in limited cases for comparison purposes.
2.1.3. Liquid water transport—two-fluid model
Before liquid water transport is presented, it is worthwhile to
emphasize that water produced during the oxygen reduction reac-
tion (ORR) is not in the vapor phase. In fact, unless the membrane
is heated to a very high temperature so that the protons are no
longer hydrated (at this point the ionic polymer is very likely to
start degrading), the ionic group (SO3 − H+ ) in Nafion will retain at
least one water molecule per proton. Under typical fuel cell opera-
tion, the number of water molecules per proton is rarely below 3.
With at least three water molecules closely bound to a proton, these
water molecules cannot be considered to be in a gaseous state since
their corresponding mean free path is comparable to that of liquid
water rather than gas. Hence when a water molecule is generated at
the solid catalyst surface, it joins a group of water molecules already
present at the ionic group which are already in a condensed state.
Therefore, in a PEMFC water is generated at the catalyst surface in
the form of dissolved water. How these water molecules leave the
ionic phase at the ionomer/gas phase interface, whether as a gas or
liquid, will depend on the gas phase. If the gas phase is not saturated
and sufficient energy is available, water can evaporate and leave as
vapor. If the gas phase is saturated, water will leave this interface
as a liquid which is then transported away by capillary diffusion or
other forced mechanisms.
When liquid water is present in the gas pore, protons could in
principle migrate through liquid water [36] and the electrochem-
ical reaction could occur in the liquid phase. Consequently, water
may be produced in the liquid phase directly. However, the dif-
fusivity of protons in liquid water is significantly smaller than in
the ionomer. Therefore, it can be assumed that the amount of liq-
uid water production is insignificant compared to that of dissolved
water production. Later in this study, comparison will be made for
the mechanism of water production in the ionomer to that of water
production in the liquid phase, assessing their implications for the
numerical results.
The two-fluid governing equation for the liquid water transport
in the backing layer and gas flow channel is derived from the volume
averaged continuity equation, and it has the following form [35,37]
∂
(l s) + ∇ · (−Ds ∇ s) + ∇ · (l f (s)u
 g ) = Sl , (11)
∂t
where s is the liquid saturation, l is the liquid density, Ds is the
liquid water diffusivity, and f (s) is the interfacial drag coefficient.
In the porous backing layer, Ds and f (s) are derived based on the
capillary pressure theory and by applying Darcy’s law for both liquid
and gas phases, and can be expressed as [35,37]
l Krl K dPc
Ds = − , (12)
l ds
g Krl
f (s) = , (13)
l Krg
where l and g are the dynamic viscosity of the liquid and gas
phase, respectively; K is the intrinsic permeability of the porous
materials; Krl and Krg are the relative permeability of the liquid
and gas phase, respectively; and Pc is the capillary pressure. Details
regarding the capillary pressure and relative permeability will be
examined in Sections 4.3 and 4.4.
Liquid flow in the gas flow channel can be very complicated,
depending on the nature of the channel flow. In this study, an
approach similar to [35,37] is used which assumes homogeneous
mist flow in the channel, the interfacial drag coefficient f (s) is set
to be unity [35] and the liquid diffusivity is fixed at an large value.
With such specifications, the liquid saturation almost vanishes in
the channel. Thus, it has the same effect as imposing an Dirichlet
6918 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927

boundary condition (s = 0) on the channel–GDL interface which has The values of Dw and d̄ are not important here since they will cancel
been used by most of the previous studies, but with this condition out in Eq. (18). The diffusivity of water vapor and the pore size from
in place, liquid water can still enter the gas flow channel. a network model d̄ = 4d0 [41] are used here, where d0 is the fiber
The source term on the right-hand side of Eq. (11) reads diameter of the carbon material. Consequently, Shce is estimated to
be in the range of 2.04 × 10−3 to2.45 × 10−1 .
Sl = Sl,react + Svl − Mw Sld , (14)
Furthermore, to differentiate the condensation and evapora-
where Sl,react is the water source from the electrochemical reac- tion processes, a Langmuir-type correction is incorporated and the
tion provided that liquid water production is assumed, Mw is the resulting condensation/evaporation source term is related to the
molecular weight of water and Mw Sld represents the water source local liquid saturation in the form of
from membrane desorption, and Svl is the interfacial mass transfer  Shc Dw
Apore (1 − s)(w − sat ) if v ≥ sat , (condensation),
between liquid and vapor phases during evaporation and conden- Svl = d̄ (20)
She Dw
sation. From kinetic theory [39,40], assuming an ideal gas and Apore s(w − sat ) if v < sat , (evaporation),
d̄
neglecting interactions between individual molecules, the net mass
transfer of the evaporation and condensation can be estimated where Shc and She are the phase transfer rate coefficients of
using the Hertz–Knudsen–Langmuir equation condensation and evaporation, respectively. The condensation/
 evaporation process is very similar to that of the membrane water
Mw Pv Pl absorption/desorption. Therefore, as a first approximation the rate
Svl = A c − e , (15) constants Shc and She are assumed to be identical in this study.
2  Tv Tl

where A is the liquid/vapor specific interfacial area which depends 2.2. Governing equations and constitutive relations
on the saturation, c and e are the condensation and evapora-
tion rate coefficient, respectively; Pv and Pl are the vapor and The complete set of governing equations of the current
liquid pressure, respectively, and Tv and Tl are the vapor and liq- three-dimensional multi-phase model consists of the following
uid temperature, respectively. A comprehensive investigation of conservation laws.
the condensation and evaporation process is rather complicated ∂
Continuity : (g ) + ∇ · (g u
 g ) = Sm (21)
and needs to be performed in the surrounding region of the ∂t
liquid–vapor interface at the molecular level. For PEMFC modeling,
it is impractical to incorporate such processes and a revised form 1 ∂ 1
Momentum : g ) +
(g u ∇ · (g u g u g )
of the above equation is usually used [16,17]  ∂t 2
  = −∇ Pg + ∇ · ( ∇ u
 g ) + Su (22)
Pv P sat
Svl = A ce − =A ce (w − sat ), (16)
T T
∂ eff
where ce (or ce ) is the analogous condensation/evaporation rate Species : ( Ci ) + ∇ · (−Dieff ∇ Ci ) + ∇ · (u
 g Ci ) = Si (23)
∂t
coefficient and it reads
Dissolved water: Eq. (2)
T Liquid water: Eq. (11)
= . (17)
ce m
2 Mw ⎛ ⎞
∂ 
Here m is an uptake coefficient that accounts for the combined Energy : ⎝ (cp )j T ⎠ + ∇ · (−keff ∇ T )
effects of heat and mass transport limitations in the vicinity of the ∂t
j=g,l,s
liquid–vapor interface. From the analysis of [41], this coefficient is
⎛ ⎞
about 0.006. The specific liquid/vapor interfacial area is calculated

as A = s Apore , where Apore is the pore surface area per unit volume +∇ · ⎝  )j T ⎠ = ST
(cp u (24)
which varies from 13 to 30 m2 cm−3 for different GDL materials [22]
j=g,l
and a value of 20 m2 cm−3 is used in the current study; s is an
accommodation coefficient similar to m . The study of [42] shows
that s rarely exceeds 20% for spherical particles with small water Electrons : ∇ · (−seff ∇ Vs ) = S (25)
saturation. In this study, the mass transfer uptake coefficient m is
varied in the range of 0.001–0.006 and the interfacial area accom- Protons : ∇ eff
· (−m ∇ Vm ) = −S (26)
modation coefficient s is varied in the range of 1–20% to roughly The above equations are closely coupled through the right-hand
estimate a range for the condensation/evaporation rate coefficient. side source terms, which either stem from the electrochemical
Similar to the membrane water absorption/desorption pro- reactions or from the interfacial mass transfer among different
cesses, the water condensation/evaporation dynamics are limited phases. The expressions of these source terms have been sum-
by the mass transport in the vicinity of the liquid–vapor interface. marized in Tables 2 and 3. In addition, constitutive and empirical
Then Eq. (16) can be further revised to formulas are necessary to close the set of equations and they
Shce Dw are listed in Table 1. Those parameters specific to the current
Svl = Apore (w − sat ), (18) two-phase model are provided in Table 4. Other physical and
d̄
dynamic parameters can be found in [14] and have been omitted
where d̄ is the characteristic length for water diffusion, Dw is here.
the mass diffusivity of water vapor, and Shce is a dimensionless It is worthwhile mentioning that the latent heat generated or
number accounting for mass transport capability during conden- consumed during the membrane absorption and desorption pro-
sation/evaporation. This is analogous to the Sherwood number for cess has been considered in this study. Water vapor absorption
mass transfer, and it is calculated as into a membrane is a process similar to a vapor condensation pro-
cess and involves a release of heat (exothermic). Similarly, water
T d̄ molecules leaving a membrane as vapor is similar to a liquid evap-
Shce = s m . (19)
2 Mw Dw oration process and requires heat (endothermic). Ostrovskii and
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927 6919

Table 2
Source terms in the conservation equation of water vapor (Eq. (1)), dissolved water (Eq. (2)), and liquid water (Eq. (11)).

Sv Sl Sd

Channel 0 0 0
S
Electrode − Mvwl Svl 0
S
ACL − Mvwl − Svd Svl Svd
⎧
⎪ S
vp : − vl +
Ri ⎧ ⎧
⎪
⎨ Mw 2F ⎨ vp : Svl − Mw Sld ⎨ vp : Sld
S vl Mw Ri lp : Sld
CCL lp : − lp : Svl − Mw Sld +
⎪
⎪
Mw ⎩ 2F ⎩ dp : Sld + Ri
⎩ dp : − Svl dp : Svl − Mw Sld 2F
Mw
Membrane 0 0 0

vp: Vapor production; lp: liquid production; dp: dissolved production.

Table 3
Source terms in other conservation equations, excluding those for water.

Sm Su Si S ST

Channel 0 0 0 0 0
 J2
Electrode Sv − K
g
g
u 0 0 + Svl hfg
 eff
 R J2
ACL SH2 + Sv − g
g
u SH2 = − 2Fi Ri Ri + + Svl hfg + Mw Svd hm,fg
K
 (T S)eff 
CCL SO2 + Sv −

K
g
g
u SO2 =
R
− 4Fi Ri  4F  +  Ri + J2
+ Svl hfg + Mw Sld (hm,fg − hfg )
 eff
J2
Membrane 0 0 0 0
 eff

Gostev [43] show that the differential heat of sorption decreases 2.3. Boundary and initial conditions
with the hydration level from 68 kJ mol−1 at a water content 
0 to about 45 kJ mol−1 at  5 but not below the latent heat The final set of governing equations, Eqs. (21)–(26), is solved in
of water condensation, 40.7 kJ mol−1 , in any case. In this study, a single computational domain as shown in Fig. 1. Thus, boundary
a constant value (1.5 × 40.7 60 kJ mol−1 ) that is deemed to be conditions are only needed to be specified on the outer surfaces of
larger than the real case is used and serves as an upper bound. the domain.
It is found that the inclusion of this overestimated heat source At the flow inlet boundaries (z = 0), the liquid saturation is set
does not affect the temperature field significantly (Tmax < 0.2 K). at s = 0 and mass fluxes of gas species are given by their respec-
In fact, the source of reaction heat (refer to Table 3) from the tive stoichiometric ratios, a and c , defined at a reference current
oxygen reduction reaction (ORR) is on the order of 108 W m−3 density, Iref = 1 (A cm−2 ), as
which is much more significant than the heat source from the
membrane sorption, 106 – 107 W m−3 . Furthermore, the swelling
a a Iref A c c Iref A
of the membrane due to water sorption is an endothermic pro- ṁa = and ṁc = , (27)
2FCH2 4FCO2
cess and the endothermic “swelling energy” must be smaller
than the exothermic solvation energy for water sorption to occur
[29]. Therefore, this energy term would be negligibly small as where A is the active reaction area, a and c are the anode and
well. cathode inlet gas density, respectively. They relate to the gas species

Table 4
Physical parameters used in the present study.

Parameter Value

Channel width/height/length (mm) 1.0/1.0/50.0

Thickness of CL/GDL (mm) 0.01/0.2
Thickness of membrane, Nafion112 (mm) 0.05
Porosity of CL/GDL,  0.3/0.6
Nafion content in CL, m 30 wt%
Fiber diameter of carbon paper (AvCarb GDL), d0 (␮m) 7.5
Intrinsic permeability of carbon paper, K (m2 ) [44] 8.7e − 12
Liquid water density, l (kg m−3 ) 970.0
Henry’s constant of oxygen, Ho (Pa m3 mol−1 ) [9] 2.0e − 4
Liquid water dynamic viscosity, l (Pa s) [45] 3.517e − 4
Surface tension,  (N m−1 ) [16] 0.0625
Contact angle, c (◦ ) 110.0
Thermal conductivity of GDL, k (W m−1 K−1 ) [14] kxx = kzz = 10, kyy = 1.3
Thermal conductivity of CL, k (W m−1 K−1 ) [14] 0.8725
Thermal conductivity of membrane, k (W m−1 K−1 ) [14] 0.445
Condensation/evaporation latent heat, hfg (J kg−1 ) [45] ±2.308e6

Liquid product Vapor product

Entropy change at reference conditions, Sref (J mol−1 K−1 ) [46] −163.25 −44.42
Entropy change at working conditions, S (J mol−1 K−1 ) [46] −149.142 −43.207
Gibbs free energy change at reference conditions, Gref (J mol−1 ) [46] 237,177.5 228,588.84
6920 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927

via anode catalyst layer (ACL) is simply the difference between the
 solid phase and membrane phase potentials, a = Vs − Vm , while
a,c = Ci Mi , (28)
the over-potential at the cathode catalyst layer (CCL) is calculated
i as c = Vs − Vm − Vrev , where Vrev is the theoretical reversible cell
where Mi is the molecular weight of the ith species. The inlet con- potential. It is calculated from the modified form of the Nernst equa-
centrations of hydrogen, CH2 , and oxygen, CO2 , can be obtained tion by assuming that the overall cell reaction is at thermodynamic
through the following relations equilibrium [47]:
  1/2 
(Pa − RHa P sat ) 0.21(Pc − RHc P sat ) Gref Sref T PH2 PO2
CH2 = and CO2 = , (29) Vrev = + (T − Tref ) + ln .
T0 T0 2F 2F 2F Pref Pref
where T0 is the inlet gas temperature, Pa and Pc are the inlet gas (31)
pressure at anode and cathode, respectively, RHa and RHc are the
inlet relative humidities of anode and cathode, respectively. The Here Gref is the Gibbs free energy change and Sref the entropy
coefficient 0.21 represents the molar fraction ratio between O2 and change for the overall reaction at reference temperature, Tref , and
N2 in air (0.21/0.79). pressure Pref . PH2 and PO2 are the partial pressure of hydrogen and
Solid phase potentials are prescribed on the upper surface of oxygen, respectively.
the anode electrode (Vs,a ) and on the bottom surface of the cathode In the second approach (Method 2), a zero solid potential is set
electrode (Vs,c ) in terms of Dirichlet boundary conditions. A constant at the bottom surface of the cathode, i.e. Vs,c = 0. While at the top
temperature is imposed on the channel inlet, channel walls, upper surface of the anode, the total cell potential loss total is imposed.
surface of the anode, and bottom surface of the cathode. It relates to the cell output voltage and reversible cell potential via
For the remaining boundary conditions not specifically men- total = Vrev − Vcell . In comparison, the second method of boundary
tioned here, a specific form of a Neumann condition applies (no-flux condition specification is rarely used in literature.
or symmetry condition): Both methods have been implemented in this study and it
is found that there are no observable differences in the results
∂ between these two approaches, except for a constant off-set for the
= 0, (30)
∂n distribution of phase (solid and electrolyte) potential (Vs , Vm ). Fig. 5
where  can be any variables of interest. shows the potential distribution on a line that crosses the mem-
The initial conditions required by the transient modeling are
simply the flow solutions of the previous steady-state run.

3. Numerical implementation

3.1. Implementation of boundary conditions for solid potential

As mentioned in Section 2.3, boundary conditions (B.C.) for

the solid phase potential are specified at the top surface of the
anode and bottom surface of the cathode. In practice, there are two
approaches to accomplish this and they are schematically shown in
Fig. 4.
In the first approach (Method 1), the solid phase potential at
the upper surface of the anode is set at Vs,a = 0 for convenience,
then the solid phase potential at the cathode bottom surface equals
the cell output voltage, i.e. Vs,c = Vcell . The over-potential in the

Fig. 4. Schematic of two different methods in the specifications of boundary condi- Fig. 5. Phase potential distribution across the MEA (at x = 3.75 × 10−4 m and z =
tions for solid potential. 0.025 m) for the boundary conditions specified by: (a) Method 1; (b) Method 2.
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927
Table 5
Comparison of computational time by implementing the two different methods in
the specification of boundary conditions for the solid potential
2-D, Comsol Base case Parallel
(s/case) (min/100 iter) (min/100 iter)
Method 1 808 5.12 131.7
Method 2 733 4.48 114.3
brane electrode assembly (MEA). Generally speaking, the potential
distribution from Method 1 is more physically meaningful since
it demonstrates the real potential distributions within the cell. In
contrast, the potential distribution from Method 2 is more intuitive
since it reveals in a straightforward manner the potential loss from
each component of the cell.
Since both approaches are applicable, the issue of computational
expense is of concern. Comparisons of the computational time by
using two different B.C. specifications are performed for three cases
and the results are summarized in Table 5. To ensure that the final
conclusion is geometry and solver independent, the comparison is
firstly made for a simplified 2-D model (for more details, please see
[31]) in Case 1, using a finite element based commercial software,
COMSOL Multiphysics. Case 2 is simply our base case, a 3-D domain
with a single straight channel and is solved in Fluent. In Case 3, par-
allel processing is conducted on a Beowulf cluster system with 8
CPUs, using Fluent. A full-size cell (25 cm2 ) with 25 parallel chan-
nels is studied and the resulting mesh possesses about 2.2 million
nodes. The results from all three cases indicate that the implemen-
tation of Method 1, the most popular B.C. specification in literature,
is less efficient. For instance, assuming a typical parallel case that
requires about 1000 iterations, using the Method 1 will waste about
3 × 8 CPU hours of computational time compared to Method 2.
The reason for the accelerated computation of Method 2 may lie
in the initialization of the flow field. As shown in Fig. 5, the poten-
tial varies in a relatively narrow range when using Method 2, thus
the initial guess is comparatively more close to the final solution
than that of Method 1. Moreover, it is found that Method 1 is more
dependent on the initialization of the phase potential. It easily cause
divergence if the initial guess is too far off from the final solution. In
contrast, Method 2 is more tolerant towards the initialization pro-
cess. Actually, no special care is required for the initialization of the
solid phase potential. This may be treated as another advantage of
Method 2.
3.2. Numerical procedure
The governing equations (Eqs. (21)–(26)) are discretized by
a second-order upwind approach and solved with a finite vol-
ume method based on the commercial software package FLUENT
(6.3.26), including its user coding abilities. For steady-state model-
ing, the SIMPLE algorithm is selected for the coupling between the
pressure and velocity field. An algebraic multi-grid (AMG) method
with a Gauss–Seidel type smoother is used to accelerate the conver-
gence. In case of transient modeling, the PISO algorithm is chosen
which is based on predictor–corrector splitting for unsteady prob-
lems. Adaptive time stepping is employed with a minimum 10−4 s
and a maximum 1 s time step size. Strict convergence criteria with
a residue of 10−8 have been used for all variables. In addition, one
extra user defined criterion is implemented to ensure that the aver-
age current density differences between any of the two layers is less
than 1 ‰.
3.3. Grid independent solution
Stringent numerical tests are performed to ensure that the solu-
tions are independent of grid size. Determination of the number of
6921
Fig. 6. Effect of N y on the error of average current density.
grid points in each direction (N x , N y , and N z ) has been conducted
in three steps. During each step, the number of grid points is varied
only in one direction while the values at another two directions are
fixed. To determine N y , for example, the number of grid points in the
x and z directions are fixed at N x = 20 and N z = 50. Then the value
of N y is varied from 5 up to 25 with a double sided non-uniform grid
scheme with an increasing factor of 1.4. In addition, an uniform grid
scheme (N y = 40) is also investigated for comparison purposes.
It is found that the model quickly diverges when N y below or
equals 5 since the discretization error would be too large and the
solution is very likely to oscillate and diverge. On the other hand,
it also fails to achieve a converged solution when the non-uniform
grid number beyond 25. This is probably due to the increasingly
high grid aspect ratio which tends to impair the stability of the
system. Therefore, N y = 25 has been referred to as the “exact solu-
tion” and the relative errors induced in other cases are defined
accordingly as
 
 N 25 − N y 
y 
Err =   × 100% (32)
N 25 
where  represents the variables to be compared.
In Fig. 6, the effect of N y on the error of average current den-
sity is demonstrated. It is seen that as long as the non-uniform
grid scheme is used, the precision of the solution increases with
the grid number. The error is approximately 1.88% for N y = 15,
which is even more accurate than the uniform grid scheme (with
N y = 40) which has an error of 2.57%. Therefore, the non-uniform
grid scheme with 15 grid points is chosen for N y . The values of N x
and N z for a grid independent solution are determined in a simi-
lar manner. Generally speaking, the model is much less sensitive
to the grid numbers in the x and z directions and relatively coarse
uniform grid schemes with N x = 20 and N z = 50 are found suffi-
cient in these directions. The corresponding solution errors in the
x and z directions are Errx = 0.49% and Errz = 0.009%, respectively.
Consequently, the total error for the final mesh is roughly estimated
as
Err ≤ |Errx |2 + |Erry |2 + |Errz |2 = 1.94% (33)
It should be noted that the effects of grid number on the com-
putational efforts (such as the number of iterations required for
convergency and the time required for each iteration, etc.) are also
considered when selecting the current grid scheme. More details
regarding the grid independent study can be found in [48].
6922 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927

4. Results and discussion

In the present model, there are seven input parameters that can
be varied to study different working conditions. They are the stoi-
chiometric ratios a , c , inlet gas relative humidities RHa , RHc , inlet
gas pressures Pa , Pc , and cell output voltage, Vcell . Two parame-
ters are used to investigate the non-equilibrium liquid/vapor phase
transfer characteristics, namely Shc and She , and two parameters to
study the non-equilibrium water sorption/desorption process, i.e.
a and d .
In our base case, the output voltage is set at 0.65 V; the inlet
gaseous reactants are assumed to be fully humidified (RHa = RHc =
1.0) at a pressure of 2 atm and a temperature of 80 ◦ C; the stoichio-
metric ratio at anode and cathode sides are 1.2 and 2.0, respectively;
the dimensionless condensation/evaporation rate coefficients (Shc ,
She ) are set at 2.0 × 10−3 ; and the sorption/desorption rate coeffi-
cients (a , d ) are chosen to be 1.0 s−1 [35]. Further, water production
is assumed to be in the dissolved phase unless otherwise specified.

4.1. Equilibrium vs. non-equilibrium water sorption/desorption

processes

Both equilibrium and non-equilibrium water sorption model-

ing has been conducted, using the approaches described in Section
2.1.2. For equilibrium sorption modeling, the flux of back diffusion
is converted to source terms in both ACL and CCL as illustrated
in Fig. 3. Since the equilibrium model does not explicitly solve
the dissolved water transport in the catalyst layer, all previous
equilibrium models have assumed that the water production is
either in the liquid or in the vapor phase. Therefore, to make the
comparison more meaningful the water production of the non-
equilibrium model is also assumed to be in the liquid phase in this
subsection.
Fig. 7 demonstrates the variation of membrane water content on
a line that crosses the CCL-membrane-ACL. The difference in water
content between the actual value () and the equilibrium value
(e ) in the CCL and ACL represents the driving force for membrane
water desorption and absorption, respectively. For the equilibrium
model, these values coincide. However, absorption and desorp-
tion still occurs, only it is instantaneous balanced. For vanishing
sorption rate coefficients, however, absorption and desorption no
longer takes place and the water dynamics decouple. In Fig. 7(a), the
absorption and desorption rate are assumed to be equal. It shows
that the difference between e and  decreases with increasing
water absorption/desorption rate. Finally, as the system reaches
equilibrium at around  = 100 s−1 the water content distribution
of the non-equilibrium model almost coincides with that of the
equilibrium model. In Fig. 7(b), two cases (lines 2 and 4) with
their absorption rate reduced to 1/10 of their desorption rate are
investigated to simulate the second stage of the two-step dynam-
ics of membrane sorption, as mentioned early. It is seen that for
both cases (lines 2 and 4) the water content difference between
the actual and the equilibrium value decreases in the CCL but
increases in the ACL compared to the cases corresponding to equal
rate for the sorption and desorption processes (lines 1 and 3). This
means that the water desorption in the CCL is accelerated but the
water absorption in the ACL is slowed down in the second sorp-
tion stage, which agrees well with the experimental observations
in [29].
As shown in Fig. 7, the water content in the catalyst layer differs
between the equilibrium and non-equilibrium approaches at the
final equilibrium state. This is probably due to the mathematical
conversion approximation of the equilibrium model, as discussed
in Section 2.1.2. To better demonstrate the water content evolution
with increasing sorption rate and to show the differences between
the two approaches, 2-D contour plots along the x–z cross section of
Fig. 7. Distribution of the actual and the equilibrium membrane water content on
a line across the anode catalyst layer (ACL)-membrane-cathode catalyst layer (CCL)
(at x = 3.75 × 10−4 m and z = 0.025 m) for the membrane water sorption and des-
orption rate coefficient of a and d (s−1 ): (a) a = d ; (b) a = 0.1d .
the ACL are shown in Fig. 8. It is seen that the actual water content
() increases along the flow direction at low sorption rates because
the sorption resistance is relatively high at this stage and water
tends more to back-diffuse from the cathode side. At high sorp-
tion rates or in the equilibrium model, however, the water content
presents a parabolic-type distribution. It decreases first, owing to
the electro-osmotic drag effect, then increases towards the outlet
due to back diffusion. Comparing Fig. 8(c) and (d), it clearly shows
that once the sorption equilibrium is approached, the water con-
tent distribution in both models are quantitatively and qualitatively
similar. Therefore, Figs. 7 and 8 can be regarded as good numerical
validations for both equilibrium and non-equilibrium models.
Fig. 9 shows the transient variation of the cell current output
during a step change in relative humidity from RH = 1 to RH = 0.5.
The response time for the cell to reach 99% of its steady-state value
is about 20.9, 23.8, 25.4, and 42.2 s for the equilibrium model and
for the non-equilibrium model with sorption rates of  = 100, 10,
and 1 s−1 , respectively. Not surprisingly, the equilibrium model
has the fastest response because one of the slowest transient pro-
cesses, the membrane water sorption, is simply neglected when
the equilibrium assumption is adopted. Therefore, Fig. 9 underlines
another large advantage of the non-equilibrium water sorption
model, namely, the transient process of water sorption is accounted
for explicitly.
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927
Fig. 8. Distribution of the actual and the equilibrium membrane water content along the middle x–z cross section of the anode catalyst layer (ACL): (a) a,d = 0.01 s−1 ; (b)
a,d = 1.0 s−1 ; (c) a,d = 100 s−1 ; (d) equilibrium model.
4.2. Comparison of water production mechanisms
As presented earlier, water is produced in the dissolved phase
during the electrochemical reaction and, in principle, a small
amount of water can be produced in the liquid phase. Neverthe-
less, water production can not be in the vapor phase for PEM fuel
cells during normal operation. Regardless, we are going to compare
three different water production mechanisms, namely, vapor, liq-
uid and dissolved water production, and examine what potential
error the vapor and liquid production assumptions may induce.
Basically, there are four main factors that have been considered
in the current model to differentiate between different water pro-
duction mechanisms. The first and most important factor comes
from the water source terms as shown in Table 2. Secondly, since
the entropy changes for vapor and liquid water production are dif-
Fig. 9. Cell dynamic response corresponding to different water sorption rates and
with respect to a step change in relative humidity from RHa,c = 1.0 to RHa,c = 0.5.
6923
ferent (refer to Table 4), the heat generation from the reversible
electrochemical reaction, T S/4F (the first energy source term in
the cathode catalyst layer as shown in Table 3), should be revised
accordingly. The third difference among these assumptions has
been implicitly accounted for by the latent heat generation or
absorption during phase change, as shown in the energy source
terms in Table 3. The last important factor comes from the determi-
nation of the reversible cell potential, Eq. (31). As shown in Table 4,
the value of Gref and Sref are different for different water prod-
ucts [46]. Gref and Sref for dissolved water production are not
available in the open literature, the values for liquid production are
thus used temporarily in this study.
The vapor and liquid water production mechanisms are related
and differentiated by condensation/evaporation processes, which
are determined by the phase transfer rate coefficient, Shce . The
dissolved and liquid water production models are related and dif-
ferentiated by water absorption/desorption processes, and thus the
absorption/desorption rate coefficient, , is of significance. There-
fore, the comparison among three water production assumptions is
best conducted for two separate phase transfer mechanisms, which
are characterized by Shce and , respectively.
Fig. 10 demonstrates the variation of average cell current den-
sity versus the phase transfer rate coefficient, Shce , for the liquid
and vapor production assumptions. It shows that the liquid pro-
duction model is almost independent of the phase transfer rate
coefficient throughout the whole range. This is sensible since the
system is already fully saturated and evaporation is not likely to
occur even at high phase transfer rates. The slight fluctuation of the
current density with increasing rate coefficient might be caused
by local under-saturation due to thermal effects. The vapor pro-
duction model, on the other hand, is significantly affected by the
phase transfer rate coefficient. It shows that the current density
increases as the rate coefficient Shce is reduced because more and
more vapor product tends to remain in its original gas phase, thus
alleviating the flooding level. Finally, as Shce reaches zero no con-
densation occurs and the model returns to a single-phase model.
6924 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927
Fig. 10. Variation of average current density with phase transfer rates, Shce , corre-
sponding to vapor and liquid water production assumptions.
In contrast, if the rate coefficients are increased towards infinity,
water vapor will condense to liquid water. When the system reaches
equilibrium, it is found that the current density of the vapor pro-
duction model reaches its minimum value which is almost identical
to the liquid production model. This is expected since both models
should converge towards identical solutions in the limit of infinite
evaporation/condensation rate coefficients. The range of our esti-
mated phase transfer rate is also marked in Fig. 10. It shows that
the estimated range is very close to the equilibrium state but more
likely corresponding to a quasi-equilibrium region. Therefore, the
assumption of phase equilibrium used in the mixture model [15]
should work reasonably well. From this point of view, although the
liquid and vapor production assumptions are not physically correct,
they can serve as a tool to identify the phase equilibrium state and
help us understand the phase transfer processes.
Fig. 11 demonstrates the variation of the average cell current
density versus the water absorption/desorption rate coefficient, ,
and the dissolved and liquid production assumptions. It is seen that
the general trends are similar for both assumptions, the cell cur-
rent density increases with decreasing sorption rate because the
membrane tends to be better hydrated when the sorption rate is
low and, thus, the absorption/desorption resistance is high. For the
Fig. 11. Variation of average current density with absorption/desorption rates, ,
corresponding to liquid and dissolved water production assumptions.
same reason, the dissolved water production model exhibits bet-
ter cell performance than the liquid production model throughout
the whole range because more water is present in the membrane
when water is produced in the dissolved phase. On the other hand,
as the absorption/desorption rate increases towards equilibrium,
the difference between the liquid and dissolved water production
models gradually diminishes and the current density approaches
the value at the sorption equilibrium state. At  = 100 s−1 which
is deemed to be very close to the sorption equilibrium, the current
density of the liquid and dissolved production models are 8063.2
and 8164.14 A m−2 , respectively, and it seems that the difference
between the two models does not vary further with increasing sorp-
tion rate. This final distinction between the liquid and dissolved
production models may arise from the estimate of the changes in
standard entropy and Gibbs free energy, as well as the reaction
heat T S/4F for the dissolved production model, where the values
of liquid production have been used as an approximation. More-
over, the current density predicted from the equilibrium model is
8104.85 A m−2 and is also identified in the figure. Again, the differ-
ence between the equilibrium model and non-equilibrium liquid
production model is probably due to the mathematical conver-
sion approaximation of the equilibrium model. In addition, Fig. 11
indicates that using the traditional equilibrium model slightly
overestimates the current density at the equilibrium state but sig-
nificantly underestimates the actual current density if the real
scenario takes place at non-equilibrium.
4.3. Effect of the empirical expressions for capillary pressure
As discussed in Section 1, owing to the lack of experimental data
the determination of the capillary pressure in the porous backing
layer of PEMFCs has been traditionally performed in terms of the
Leverett function [49]


Pc =  cos(c ) (1.417s − 2.120s2 + 1.263s3 ), (34)
K
where  is the surface tension between the liquid water and gas
phase, and c is the assumed uniform contact angle of the porous
materials.
The Leverett approach was derived based on experimental data
of homogeneous soil or a sand bed with uniform wettability. In
the backing layer of the PEMFC, however, the pore size ranges
from nanometer to micrometer, and its wettability is strongly
affected by both the hydrophilic carbon substrate and the level of
the hydrophobic PTFE coating. Therefore, the Leverett approach is
incapable of precisely predicting the capillary pressure in PEMFCs.
Many experiments [19–24] have been carried out recently, trying
to assess the real situation in PEMFCs. However, due to the dif-
ferences in their measurement approaches, facilities, experimental
conditions, and the materials being investigated, their results do
not agree with each other very well. To the best of our knowledge,
there exist no correlations to-date which can quantify the capil-
lary pressure for all materials used in PEMFCs under all possible
conditions. In this study, therefore, we are not trying to propose a
general capillary pressure modeling approach. Instead, two of the
newly developed capillary pressure correlations that are specific to
PEMFCs are examined and the resulting saturation characteristics
are compared with the standard Leverett function approach.
The first new correlation is proposed by Kumbur et al. [22–24],
based on their drainage capillary pressure-saturation measure-
ments of the SGL 24 serials composite carbon paper and E-TEK Elat
1200W carbon cloth materials, and it has the following form
 293 6 

Pc =  20.4C (s), (35)
T K
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927 6925

(s) = (wt%)[0.0469 − 0.00152(wt%) − 0.0406s2 + 0.143s3 ]

+ 0.0561 ln(s).

In the above correlation, the effects of temperature, T, com-

pression pressure, C, and mixed wettability, wt, representing the
amount of PTFE in the GDL, have been incorporated into a modified
form of the standard Leverett function. Since no micro-porous layer
(MPL) is considered in the current study, the (s) function has been
extended from its original range, 0 < s < 0.5, to 0 < s < 1 to elimi-
nate the effect of the MPL, as suggested by the authors. In our model,
the temperature is coupled through the energy equation; the com-
pression pressure is assumed to be 1 MPa; for the mixed wettability,
we found that it is not really accounted for by the above expres-
sion. This is mainly because the derivative of the capillary pressure
(dPc /ds) is the actual form that is being used in the two-fluid model.
In Kumbur’s expression, it results in

dPc

293
6
= 20.4C
ds T
  
 0.0561
× (wt%)(−0.0812s + 0.429s2 ) + . (36)
K s

It is easily seen that for small saturations as in PEMFCs, Eq. (36) is

essentially dominated by the second term (1/s) while the effect of
PTFE loading that is described by the first term is greatly subdued.
The second correlation used in this study was derived from a fit
by Ye and Nguyen [35] based on the experimental data of Nguyen
et al. [20]. Both drainage and imbibition processes have been inves-
tigated in this experiment. The capillary pressure within a Toray
TGP-H-060 GDL with 10% PTFE loading has been measured, using
a volume displacement technique, while the capillary pressure
within a thickened CL (55 ␮m) that is composed of the same mate-
rial is obtained through neutron imaging techniques. Accordingly,
two separate correlations for GDL and CL are developed [35]

Pc,gdl = 2.09(e22.2(0.321−s) − e44.9(0.321−s) ) + 35.6, (37)

Fig. 12. Saturation distribution in cathode catalyst layer (CCL) and cathode gas dif-
Pc,cl = 2431(e0.0088(0.567−s) − e92.36(0.567−s) ) − 2395. (38) fusion layer (GDL) with respect to the empirical function of capillary pressure used:
(a) standard Leverett function; (b) Kumbur’s expression; (c)-1 Ye’s expression, CCL;
It should be noted that, the experimental method used in [20] is (c)-2 Ye’s expression, GDL.

naturally limited by break-through and break-off liquid pressures of

the GDL materials. Consequently, the capillary pressure they mea-
sured lies in a relatively small range (one to two orders smaller than
that of Gostick et al.’s [19], Kumbur et al.’s [22–24], and Fairweather
et al.’s [21]). Physically, smaller capillary pressure means higher
water retention ability, thus a higher level of saturation would be
expected by this expression.
The saturation distribution resulting from Kumbur’s and Ye’s
capillary correlations are compared with the standard Leverett
function and the results are shown in Fig. 12. The standard Lev-
erett function and Kumbur’s expression are applied in both GDL and
CL. Consequently, the figures exhibit a smooth saturation decrease
from the catalyst layer to the gas flow channel (Fig. 12a and b). In
contrast, GDL and CL are differentiated by two separate capillary
expressions in Ye’s approach and two plots are presented for the CL
(Fig. 12(c)-1) and GDL (Fig. 12(c)-2) separately. It is found that Ye’s
approach results in a very high level of liquid flooding in the cata-
lyst layer but the saturation decreases steeply towards the GDL. For
the main part of the GDL domain, the saturation from Ye’s approach
is even smaller than Kumbur’s and the standard Leverett function
approach. Comparing Fig. 12(b) and (c) to (a), it is seen that the
saturation distribution predicted by the Leverett function is more
uniform. This can be observed more easily in the vicinity of the
gas flow channel. This uniformity may be looked at as a counter-
evidence of the applicability of the Leverett function in PEMFCs,
since it implicitly treats the backing layer of PEMFCs as homogenous
materials (soil or sand) with uniform wetting properties.
4.4. Effect of the empirical expressions for relative permeability
Similar to the capillary pressure, various forms of relative per-
meability correlations have been employed in previous PEMFCs
modeling efforts due to the lack of experimental support. Predictive
models for the relative permeability were developed from con-
ceptual models of flow in capillary tubes combined with models
of pore-size distribution. The common predictive models are the
power law function [50], Burdine and Mualem functions [51], Van
Genuchten function [52], Brooks-Corey function [53], etc. Among
them, the power law function is the most widely used in fuel cell
modeling studies. The general form of the power law function is
eff n eff n
Krl = (sw ) , Krg = (snw ) , (39)
where the subscript w denotes the wetting phase, and nw denotes
eff , is defined as:
the non-wetting phase. The effective saturation, sw
eff sw − Sw,irr
sw = , (40)
1 − Snw,irr − Sw,irr
where Sw,irr and Snw,irr are the irreducible saturation for the wetting
and non-wetting phase, respectively. The exponent n in Eq. (39)
6926 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927

5. Conclusions

In this study, a 3-D model with a single straight flow channel

has been developed which incorporates transport phenomena in
PEMFCs, such as multi-species transport, multi-water-phase trans-
port, heat transfer, and electrochemical reactions as well as product
water in the dissolved ionomer phase; and several approaches tra-
ditionally used in modeling studies have been examined. It is found
that the most widely used methods for the specification of the
boundary condition for the solid phase potential is numerically
not the most efficient and most stable. Instead, a better boundary
condition specification is with the cathode solid potential being
set to zero and the anode solid potential set to the total over-
potential of the cell. The equilibrium model for the membrane
sorption/desorption processes tends to underestimate the cell cur-
rent output even under steady conditions, and the cell response
time if transient phenomena are of concern, for which, non-
equilibrium model with finite sorption/desorption rates should be
used. The effect of the state of the product water during the half cell
reaction (ORR) on the cell performance has also been investigated;
the assumption of vapor or liquid water production mechanisms
may cause non-negligible discrepancies in the cell performance,
and water dissolved in membrane should be considered as the
proper mechanism of water formation in practical PEMFCs. Cap-
illary pressure and relative permeability have significant effect on
water transport, and relative permeability has even more impact
on the liquid water transport than the capillary pressure.

Acknowledgement

Financial support of Natural Sciences and Engineering Research

Council of Canada (NSERC) through a Strategic Projects Grant and
Discovery Grants is gratefully acknowledged.

Appendix A. Nomenclature

Fig. 13. Saturation distribution in cathode catalyst layer (CCL) and cathode gas dif-
fusion layer (GDL), with a power of 4.5 for the relative permeability in GDL with the a water activity
capillary pressure given by: (a) standard Leverett function; (b) Kumbur’s expression; A area (m2 )
(c)-1 Ye’s expression, CCL; (c)-2 Ye’s expression, GDL. C molar concentration (mol m−3 )
cp specific heat (J kg−1 K−1 )
is usually determined through a curve fit of experimental data. In d0 fiber diameter of carbon material (␮m)
PEMFC modeling, a factor of 3, the so-called Wyllie’s cubic power d̄ characteristic length of pore size (␮m)
law [54], is commonly adopted in literature and it is used as the D diffusivity (m2 s−1 )
base case in our current study. The saturation distribution with this EW equivalent molecular weight of dry membrane
cubic power law has been used for the results shown in Fig. 12. (1.1 kg mol−1 )
On the other hand, based on the analysis of [19,55], several F Faraday’s constant 96487 (C mol−1 )
recent studies [35,55,56] have used an exponent of n = 4.5 for the G Gibbs free energy (J mol−1 )
GDL while keeping the exponent at 3.0 in the CL. As a comparison to H Henry’s constant (Pa m3 mol−1 )
the cubic power law, this approach is implemented in this study to hfg latent heat of condensation/evaporation (J kg−1 )
investigate the effect of relative permeability on the distribution of hm,fg latent heat of water absorption/desorption (J kg−1 )
the liquid saturation. Three different capillary pressure correlations j0,ref reference exchange current density (A m−3 )
are investigated again for comparison purposes and the results are J mass flux (kg m−2 s−1 ); current density (A m−2 )
presented in Fig. 13. A significant increase of the liquid saturation k thermal conductivity (W m−1 K−1 )
is observed for all three capillary pressure correlations used when K permeability (m2 )
compared to Fig. 12. This is because the saturation is always smaller Kr relative permeability
than 1, hence an increase in the exponent in Eq. (39) tends to reduce M molecular weight (kg mol−1 )
the relative permeability exponentially, which in turn results in a n number of electrons transferred in the half cell reaction;
much reduced water removal ability for the porous backing layer. unit normal vector
Fig. 13 clearly demonstrates the importance of the determination nd electro-osmotic drag coefficient (H2 O/H+ )
of relative permeability. It indicates that the relative permeability P pressure (atm, Pa)
has an even more significant impact on the modeling results than  universal gas constant 8.314 (J mol−1 K−1 )
the capillary pressure. Therefore, apart from the measurements of RH relative humidity
capillary pressure the relative permeability for PEMFCs should also Ri reaction rate (A m−3 )
be measured, which is a relatively rare focus of attention to-date. s liquid saturation
 H. Wu et al. / Electrochimica Acta 54 (2009) 6913–6927 6927

S source terms; entropy (J mol−1 K−1 ), surface area (m2 ) [5] D.M. Bernardi, M.W. Verbrugge, Am. Inst. Chem. Eng. J. 37 (1991) 1151.
Shce dimensionless condensation/evaporation rate [6] T. Berning, D. Lu, N. Djilali, J. Power Sources 106 (2002) 284.
[7] M. Eikerling, Y.I. Kharkats, A.A. Kornyshev, Y.M. Volkovich, J. Electrochem. Soc.
T temperature (K) 145 (1998) 2684.

u velocity (m s−1 ) [8] T. Thampan, S. Malhotra, H. Tang, R. Datta, J. Electrochem. Soc. 147 (2000) 3242.
V potential (V); volume (m3 ) [9] P. Berg, K. Promislow, J.S. Pierre, J. Stumper, B. Wetton, J. Electrochem. Soc. 151
(2004) 341.
[10] J.J. Baschuk, X. Li, J. Power Sources 142 (2004) 134.
Greek letters [11] J. Fimrite, H. Struchtrup, N. Djilali, J. Electrochem. Soc. 152 (2005) 1804.
[12] G.J.M. Janssen, J. Electrochem. Soc. 148 (2001) 1313.
˛ transfer coefficient
[13] A.Z. Weber, J. Newman, J. Electrochem. Soc. 150 (2003) 1008.
ε porosity [14] H. Wu, P. Berg, X. Li, J. Power Sources 165 (2007) 232.
 overpotential (V) [15] C.Y. Wang, P. Cheng, Int. J. Heat Mass Transfer 39 (1996) 3607.
 membrane water absorption/desorption rate (s−1 ) [16] W. He, J.S. Yi, T.V. Nguyen, AIChE J. 46 (2000) 2053.
[17] D. Natarajan, T.V. Nguyen, J. Electrochem. Soc. 148 (2001) 1324.
m mass transfer accommodation coefficient [18] Z.H. Wang, C.Y. Wang, Abstract 90, in: Electrochem. Soc. Meeting, Toronto, ON,
s liquid/vapor surface area accommodation coefficient 2000.
 membrane water content [19] J.T. Gostick, M.W. Fowler, M.A. Ioannidis, M.D. Pritzker, Y.M. Volfkovich, A.
Sakars, J. Power Sources 156 (2006) 375.
dynamic viscosity (kg m−1 s−1 ) [20] T.V. Nguyen, G. Lin, H. Ohn, X. Wang, ECS Trans. 3 (2006) 415.
 density (kg m−3 ) [21] J.D. Fairweather, P. Cheung, J. St-Pierre, D.T. Schwartz, Electrochem. Commun.
 electrical conductivity (S m−1 ); surface tension (N m−1 ) 9 (2007) 2340.
[22] E.C. Kumbur, K.V. Sharp, M.M. Mench, J. Electrochem. Soc. 154 (2007) 1295.
c contact angle (◦ ) [23] E.C. Kumbur, K.V. Sharp, M.M. Mench, J. Electrochem. Soc. 154 (2007) 1305.
 variables [24] E.C. Kumbur, K.V. Sharp, M.M. Mench, J. Electrochem. Soc. 154 (2007) 1315.
 stoichiometric ratio [25] A. Bazylak, D. Sinton, N. Djilali, J. Power Source 176 (2008) 240.
[26] K. Jiao, B. Zhou, P. Quan, J. Power Sources 157 (2006) 226.
[27] A. Theodorakakos, T. Ous, M. Gavaises, J.M. Nouri, N. Nikolopoulos, H. Yanagi-
Subscripts and superscripts hara, J. Colloid Interface Sci. 300 (2006) 673.
a anode; absorption [28] F. Opekar, D. Svozil, J. Electroanal. Chem. 385 (1995) 269.
[29] M.B. Satterfield, J.B. Benziger, J. Phys. Chem. B 112 (2008) 3693.
act activation [30] L.M. Onishi, J.M. Prausnitz, J. Newman, J. Phys. Chem. B 111 (2007) 10166.
c cathode; capillary pressure [31] H. Wu, X. Li, P. Berg, in: I. Dincer, X. Li (Eds.), Proceedings of the 2nd Int. Green
d dissolved water phase; desorption Energy Conference, ON, 2006 (ISBN: 0-9781236-0-3).
[32] J.M. Stockie, K. Promislow, B.R. Wetton, Int. J. Numer. Methods Fluids 41 (2003)
e equilibrium state 577.
eff effective value [33] A.A. Shah, G.S. Kim, P.C. Sui, D. Harvey, J. Power Sources 163 (2007) 793.
g gas phase [34] A. Vorobev, O. Zikanov, T. Shamim, J. Power Sources 166 (2007) 92.
[35] Q. Ye, T.V. Nguyen, J. Electrochem. Soc. 154 (2007) 1242.
i the ith species
[36] Q.P. Wang, M. Eikerling, D.T. Song, Z.S. Liu, J. Electroanal. Chem. 573 (2004) 61.
j the jth phase [37] P.C. Sui, S. Kumar, N. Djilali, J. Power Sources 180 (2008) 410.
l liquid water [38] R.B. Bird, W.E. Stewart, E.N. Lightfoot, Transport Phenomena, 2nd ed, John Wiley
& Sons Inc., New York, 2002.
ld phase transfer between liquid and dissolved water
[39] H.K. Cammenga, E. Kaldis, in: Current Topics in Materials Science, North-
m membrane; mass source term Holland, 1980 (Chapter 4).
ref reference state, reference value [40] J.G. Collier, J.R. Thome, Convective Boiling and Condensation, 3rd ed., Clarendon
s solid phase; saturation Press, Oxford, 1994.
[41] J.H. Nam, M. Kaviany, Int. J. Heat Mass Transfer 46 (2003) 4595.
sat saturation pressure (atm) [42] H. Kim, P.S.C. Rao, M.D. Annable, Water Resour. Res. 33 (1997) 2705.
T source term of temperature [43] V.E. Ostrovskii, B.V. Gostev, J. Thermal Anal. Calorim. 46 (1996) 397.
u source term of momentum [44] K. Ito, K. Ashikaga, H. Masuda, T. Oshima, Y. Kakimoto, K. Sasaki, J. Power Sources
175 (2008) 732.
v water vapor [45] F.P. Incropera, D.P. Dewitt, Fundamentals of Heat and Mass Transfer, 5th ed.,
vd phase transfer between vapor and dissolved water John Wiley & Sons, New York, 2002.
vl phase transfer between vapor and liquid water [46] X. Li, Principles of Fuel Cells, Talor & Francis, 2006.
[47] C. Marr, X. Li, ARI 50 (1998) 190.
w water [48] H. Wu, Transient transport phenomena in PEM fuel cells, Ph.D. thesis, University
source term of charge of Waterloo, 2009.
0 inlet conditions [49] M.C. Leverett, Trans. AIME 142 (1941) 151.
[50] M. Kaviany, Principles of Heat Transfer in Porous Media, 2nd ed., Wiley, New
York, 1995.
References [51] J. Chen, J. Hopmans, M. Grismer, Adv. Water Resour. 22 (1999) 479.
[52] J.B. Kool, J.C. Parker, M.T. Van Genuchten, J. Hydrol. 91 (1987) 255.
[1] T.E. Springer, T.A. Zawodzinski, S. Gottesfeld, J. Electrochem. Soc. 138 (1991) [53] P. Ustohal, F. Stauffer, T. Dracos, J. Contam. Hydrol. 33 (1998) 5.
2334. [54] T.A. Corey, Mechanics of Immiscible Fluids in Porous Media, 3rd., Water
[2] A.A. Kulikovsky, J. Electrochem. Soc. 150 (2003) 1432. Resources Publications, Colorado, 1994.
[3] H. Meng, J. Power Sources 162 (2006) 426. [55] Y. Wang, S. Basu, C.Y. Wang, J. Power Sources 179 (2008) 603.
[4] S. Um, C.Y. Wang, J. Power Sources 156 (2006) 211. [56] Y. Wang, C.Y. Wang, J. Electrochem. Soc. 154 (2007) 636.