View Online / Journal Homepage / Table of Contents for this issue

PAPER

www.rsc.org/materials | Journal of Materials Chemistry

A computational investigation of the properties of a reverse osmosis

membrane
Zak E. Hughes* and Julian D. Gale

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

Received 21st May 2010, Accepted 3rd July 2010

DOI: 10.1039/c0jm01545h
Reverse osmosis (RO) is currently one of the most widely used methods of desalination in the world and
rapidly increasing in usage. The membranes used in the RO process play a vital role in determining the
effectiveness of the desalination process. In this work, fully atomistic molecular dynamics simulations
of one of the most widely employed membranes, namely the FT30 polyamide material, have been
carried out in order gain greater understanding of the structure of the system and its interaction with
saline solution. The system studied consisted of a solvated membrane layer and a layer of bulk solution,
thus allowing the membrane interface to be simulated. The behaviour of water and salt ions in both the
bulk solution and membrane has been investigated. It was found that the diffusivities of water and the
salt ions were reduced by an order of magnitude within the membrane. Furthermore, umbrella
sampling methods have been used in order to determine the free energy surface associated with the salt
ions passing through the membrane-solution interface. The present work demonstrates that there is
a high degree of variability in the resistance to salt diffusion into the membrane associated with the
structure of the water encountered as the ion permeates the membrane. Despite this variability in the
free energy gradient, all cases ultimately exhibit a high resistance to ionic diffusion due to charge
separation. However, migration of a sodium cation/chloride anion pair fails to substantially lower the
barrier to salt diffusion, thus confirming the robust nature of the membrane selectivity for water.

Introduction
Desalination of seawater and brackish water is an area that is
growing in importance. The increasing population of the world
and changes to the climate and environment mean that traditional water sources are being placed under increasing pressure.1,2,3,4 Desalination of brackish water and seawater is one
method of tackling this problem, albeit at the cost of increased
energy usage.
Early desalination methods were based on thermal processes,
and in some parts of the world these methods are still extensively
used.2 However, the high energy costs associated with such
processes has led to a number of different desalination techniques being developed.1,2 The most popular current technique is
that of reverse osmosis (RO), where saltwater is forced through
an appropriate membrane at high pressure, allowing the water to
pass through, while the salt ions are largely rejected. Typically,
a single pass through the current membranes can achieve
approximately two orders of magnitude in reduction of the salt
concentration.1,5 Obviously the membrane material plays
a highly significant role in determining the success of the RO
process. Understanding how the different species in the feedwater interact with the membrane will allow new, more efficient
membranes to ultimately be designed.
The first widely used commercial RO membranes were made
of cellulose-acetate (CA), but these membranes have a number of
drawbacks.6 Firstly, they tend to compact under high pressure
reducing the performance of the membrane. The second even
Nanochemistry Research Institute, Department of Chemistry, Curtin
University of Technology, P.O. Box U1987, Perth, WA, 6845, Australia

7788 | J. Mater. Chem., 2010, 20, 77887799

more serious problem is that cellulose-acetate membranes deteriorate by hydrolysis; this is a particular problem at high or low
pH. Thus to ensure a good lifetime for a CA membrane it was
necessary to monitor and control the pH of the feedwater.
Because of these disadvantages, CA membranes have largely
been replaced by a new generation of composite polyamidepolysulfone (PA-PS) membranes. In these membranes the active

Fig. 1 Schematic diagram of the polyamide-polysulfone composite

membranes currently used in reverse osmosis. Distances indicated
represent the approximate thickness of the relevant layer.

This journal is The Royal Society of Chemistry 2010

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

View Online

 7,8,9 is
polyamide layer, with a thickness of the order of 2000 A,
responsible for the actual removal of salt from the seawater while
the much thicker ( 50mm) polysulfone layer simply provides
structural support and strength, see Fig. 1. Despite the fact that
there has been a large amount of work done studying composite
membranes, there are still many aspects of their behaviour that
are not fully understood, particularly at the atomistic level.
Molecular simulation of reverse osmosis membranes can,
therefore, provide important information about these
membranes and how they interact with species in the feedwater
that cannot be determined easily by experimental methods. In
particular, simulation can provide important information about
how membranes may be altered in order to have a greater flow
rate or be made more resistant to foulants present in the feedwater. However, before such modifications can be proposed the
standard behaviour of these membranes needs to be determined.
Most previous molecular simulations of reverse osmosis have
abstracted the membrane structure, coarse-graining it in some
way.10,11,12,13 There have only been a few studies where the
membrane has been simulated in full atomistic detail.5,14 The
work of Kotelyanskii et al.5 provided valuable information about
how water diffuses within one particular RO membrane, but due
to the computational resources available at that time was limited
to simulating a relatively small system of solvated polyamide (62
repeat units). In addition, this work only simulated the bulk
membrane, such that no information about the behaviour at the
membrane interface was provided. The systems generated by
Harder et al.14 started with monomers which were polymerised
during the course of the simulation. This method did generate
configurations where an there was an interface between the bulk
solution and the membrane, but how salt ions might interact with
the membrane was not investigated.
In both these previous simulations the polyamide modelled
was FT30,15,16 for the simple reason that this material is experimentally the best characterised of the polyamide membranes.
Even so, many details of the structure of FT30 are still poorly
understood. FT30 consists of two different monomers, trimesoyl-chloride (TMC) and m-phenylenediamine (MPD), that
are cross-linked to each other; Fig. 2 shows the structure of
a single repeat unit.
The motivation for the present study is to investigate, for the
first time, the behaviour of species in bulk solution, in the

Fig. 2 Schematic representation of the structure of the polyamide FT30,

illustrating the two monomers, TMC and MPD, that compose a repeat
unit.

This journal is The Royal Society of Chemistry 2010

membrane, and at/across the membrane interface at the atomistic

level. Thus the systems simulated need to consist of both
a membrane layer and a solution layer. A further aim was to
attempt to model the polysulfone part of the membrane in some
manner, which has previously been neglected in atomistic
models. With these goals in mind, a series of simulations of the
FT30 polymer have been performed. The arrangement of the
system and the salt concentration present in both the bulk
solution or/and with the membrane itself have been varied in
order to observe differences in the behaviour of the system. These
simulations provide important information about the diffusion
rates and coordination of species in both bulk solution and
within the membrane. This fundamental knowledge aims to
provide the insights necessary to lead ultimately to improved
membrane design for greater energy efficiency.

Methods
Due to the complex nature and large size of the system of
interest, and the fact that many of the properties of relevance are
dynamic, the most appropriate computational method for
investigating the membrane is to use classical molecular
dynamics simulations (MD). While it would be possible to use
a coarse-grained model for the membrane, which would allow
greater length- and time-scales to be simulated, this was rejected
for two reasons. Firstly, it is at the atomistic level that information on the membrane structure and behaviour is hardest to
determine from experiment. Secondly, constructing a coarsegrained model that accurately describes FT30 is extremely
difficult based on the current information; indeed any such
coarse-graining process would typically need data determined
from atomistic simulations.
The polyamide membrane was modelled using the well known
all atom optimised potentials for liquid simulations (OPLS-AA)
force field,17,18,19 which has been used to study a range of systems
where organic molecules are solvated by water.20,21,22,23,24,25 In
this force-field, intramolecular interactions are described by
harmonic bond stretching and angle bending terms, the torsional
angles by a cosine term, while the intermolecular interactions are
represented by a Coulomb interaction and Lennard-Jones
potentials.
For the water molecules, the TIP4P-Ew model26 was used.
This is a rigid four site model consisting of a Lennard-Jones
site representing the oxygen atom and three point charges
representing the two hydrogens and a Drude particle. This
water model describes most of the properties of water relevant
to this study (i.e. diffusion coefficient, radial distribution
function, etc) well at a level of computational expense that is
affordable. In addition, Joung and Cheatham27,28 recently
developed parameters for salt ions specifically for use with
TIP4P-Ew water.
As mentioned previously, it is desirable that the porous polysulfone (PS) scaffold be represented in the simulations in some
manner. However, if a fully atomistic representation was used
the simulations would increase significantly in computational
expense. Instead it was decided that the PS support would be
abstracted and represented by a base constructed from a layer of
atoms placed on a hexagonal lattice. The atoms in this layer had
no partial charge and the parameters for the van der Waals
J. Mater. Chem., 2010, 20, 77887799 | 7789

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

View Online

interactions were set to those of an aliphatic carbon in the OPLS 3 2.76 kJ mol1). These atoms
AA force field (i.e. s 3.5A,
were tethered to their positions by a harmonic potential with
a bond constant of 419 kJ mol1. These parameters were chosen
so that the support/pore would not interact with the water
strongly, ensuring that the diffusion of the water within the
system is limited by the polyamide-water interaction.
The MD code used for the simulations was DLPOLY version
2.18,29 with the Velocity Verlet algorithm used to integrate the
equations of motion. The Coulomb interactions were evaluated
using a smooth particle mesh Ewald sum30 (with a relative
accuracy of 1.0  106) and a timestep of 0.5 fs was used. The
simulations were carried out in the either the canonical or
isothermal-isobaric ensembles using the Nose-Hoover thermostat and Hoover barostat,31,32 with relaxation parameters of 1.0
and 2.0 ps, respectively.
The unbiased diffusion of salt ions across the membrane
interface is not observed on the time scales currently accessible to
classical molecular dynamics, as would be expected for the salt
rejecting membrane material simulated. In order to gain information about the resistance to salt diffusion through FT30,
umbrella sampling methods were used to calculate the free
energy barrier for an ion to enter the membrane. These simulations were also carried out using DLPOLY, but modified to
incorporate the PLUMED33 plug-in.
Experimental structural information for FT30 indicates that
19.4% of the TMC groups are cross-linked and that the water
content of the solvated polymer is about 23% of the system
mass.5,9 Based on these estimates, a model for the solvated FT30
membrane was constructed. Initially, 24 polymers, each consisting of 23 repeat units, were solvated in water and the system
was equilibrated in an NpT ensemble at 300 K and 1 atm for 2 ns.
Once equilibrated, 54 pairs of carboxylic acid groups that were
 apart were randomly chosen, and connected by an MPD
59 A
monomer. A further equilibration run of 1 ns under the same
conditions was then carried out for the cross-linked material.
Once the solvated membrane was equilibrated, three different
arrangements of system were constructed. The first model has the
polyamide membrane in a 2-D layer arrangement, in contact
with solution at both surfaces. In the second system, the
membrane was placed in contact with the atoms representing the
polysulfone support. In the third arrangement, the support was
again present, but this time a pore is present in the support. This
 and the overall pore/
pore is cylindrical with a radius of 20 A

support being 12.1 A deep. Below the pore was another box of
water representing the pure filtrate. Fig. 3 shows these three
different system arrangements. By constructing these three
different representations of the RO system, the influence of the
support and its porosity can be examined.
In addition to the three different system arrangements, the
membrane was simulated with three distinct feedwater conditions; a solution without any salt ions (i.e. pure water), a solution
of seawater concentration (0.55 molal NaCl) and a dilute salt
solution (0.0055 molal), which approximates the feedwater in the
second pass through the RO process. A solution at each
concentration was equilibrated for 1 ns, the solvated polyamide
was placed into this simulation and molecules overlapping with
the membrane structure were deleted. A water molecule was
 of a nonconsidered overlapping if the oxygen is within 2.6 A
7790 | J. Mater. Chem., 2010, 20, 77887799

Fig. 3 The three different system arrangements simulated in the present

work; (left to right) a system without any support, a system with a nonporous support layer, and a system with a porous support. The water is
coloured green, the polyamide grey, and the coarse-grained polysulfone
support blue.

hydrogen membrane atom. The combined membrane and solution were then allowed to equilibrate for a further 1 ns before
production runs were started.
The final simulation cells possessed orthorhombic boundary
 along the x and y-axes, respectively,
conditions; 49  48.50 A
with the length in the z-direction, which represents the direction
perpendicular to the plane of the membrane-solution interface,
varying depending upon the system being simulated. In the case
of the systems with either no support or a non-porous support
 with the depth of the membrane layer and
the z-axis was 203 A
the solution layer being approximately equal. For the systems
 with
with a porous base the length in the z direction was 300 A,
the part of the system above the pore having the same initial
configuration as the system with the non-porous support.
In order to determine what influence the membrane has,
a control run for the solution alone at each of the three
concentrations was also performed. The details of all the
different simulations performed are summarised in Table 1.
This journal is The Royal Society of Chemistry 2010

View Online

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

Table 1 Details of the different simulations performed in the present study of the FT30 desalination membrane. 9,828 water molecules were present in
all the runs except P3, P4 and P7 where 16,055 were present

Run

Support present

Pore present

Membrane present

Salt inside membrane

[NaCl]/molal

Number of NaCl

Duration of
production run/ns

C1
C2
C3
P1
P2
P3
P4
P5
P6
P7
P8
S1
S2
S2

no
no
no
no
yes
yes
yes
yes
yes
yes
no
yes
yes
yes

no
no
no
no
no
yes
no
yes
no
yes
no
no
no
no

no
no
no
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes

N/A
N/A
N/A
N/A
N/A
N/A
no
no
no
no
no
yes
yes
yes

0.0055
0.55

0.0055
0.0055
0.55
0.55
0.55
0.0055
0.0055
0.55

0
1
100
0
0
0
1
1
100
100
100
1
1
100

2
2
2
6
10
8
8
6
8
6
6
8
8
6

Results and discussion

Water and saline solutions
In order to have a set of reference data that the FT30 simulations
can be compared against, control runs of pure TIP4P-Ew and
NaCl/TIP4P-Ew water solutions at two different concentrations,
0.0055 & 0.55 molal, were carried out. The diffusion coefficients
of the three species at the three different solution concentrations
are given in Table 2. Given the large number of different force
fields and water models previously used for simulation of Na+/
Cl ions in water, it is not surprising that the properties, such as
the diffusion coefficient and coordination number, can vary
significantly between different works.34
The experimental value for the self-diffusion of water is 2.3 
109 m2s1, while the diffusion coefficient of water in a simulation
will depend not only on the model chosen but also the simulation
parameters.35 The diffusion coefficient of a system of pure
TIP4P-Ew water, under the simulation conditions used in this
paper, was determined to be 2.70  109 m2s1. This is a slight
overestimation of the experimental value, but compares relatively well with other water models. The TIP3P model, used by
Harder et al.,14 gives the self diffusion coefficient of water as 5.19
 109 m2s1,36 while for the CVFF model used in the simulations

Table 2 Diffusion coefficients of water, sodium cations and chloride

anions as determined from the control runs carried out in this study.
Literature values are included for comparison
Run
(salt concentration/
molal)
C1 (0.0)
C2 (0.0055)
C3 (0.55)
Experiment
(N dilution)38
Experiment
(0.5M)39
TIP4P-Ew (1.0)28
SPC-Fw (0.55)40

Dwater/
109 m2 s1

DNa+/
109 m2 s1

DCl/
109 m2 s1

2.70  0.03
2.71  0.03
2.53  0.03
2.35

1.51  0.40
1.22  0.02
1.33

2.07  0.60
1.67  0.03
2.03

2.27

1.28

1.85

2.05

1.12
0.93

1.52
1.27

This journal is The Royal Society of Chemistry 2010

of Kotelyanskii et al.5,37 the diffusion rate varies between 2.45

8.02  109 m2s135 depending on the type and distance of cutoff
used.
Regarding the salt ions, the experimental values of the diffusion rates at infinite dilution are 1.33 and 2.03  109 m2 s1 for
Na+ and Cl, respectively,38 and these values compare well with
the diffusion coefficients of the salt ions in the 0.0055 molal
solution. Experimental work by Woolf39 determined the diffusion rates of water, the sodium ion and the chloride anion to be
2.27, 1.28 and 1.85  109 m2 s1, respectively, in a 0.5M solution
of NaCl. Simulation data for the diffusion coefficients at 0.55
molal are more difficult to find as many simulations use salt
solutions of higher concentration. One simulation40 carried out
at this concentration, using a flexible SPC water model41 together
with the Smith-Dang42,43 salt parameters, determined diffusion
coefficients of 2.01, 0.88 and 1.27  109 m2 s1 for water, Na+
and Cl, respectively. Joung and Cheatham,28 using the same
water model and salt parameters employed in this paper, calculated DNa 1.12  109 m2 s1 and DCl 1.52  109 m2 s1 for
a solution of 1 molal concentration. These values accord well
with the trend in diffusion coefficients calculated in this study. It
is known44,40,45 that the diffusivity of species in a solution
decreases as the concentration of the solution increases, and this
qualitative behaviour is correctly reproduced. Overall, the model
reproduces the qualitative behaviour of salt solutions correctly
and is quantitatively of reasonable accuracy for ion diffusion
rates.
Aside from the diffusion coefficients, the coordination number
of the salt ions within water can differ appreciably between
studies. For sodium, values between 4 and 6 have been reported
from simulation,40,46,34 while a range between 4 and 8 has been
seen experimentally. The situation with the chloride ion is
similar, simulations placing the coordination number in the
range of 7 to 12 waters.34 Lyubartsev and Laaksonen40 reported
the coordination numbers of a 0.55 M solution (flexible SPC
water, Smith-Dang salt ions) being 5.8 and 6.9 for Na+ and Cl,
respectively. They also investigated the effect of concentration on
the coordination number, finding that the coordination number
of Cl was insensitive to changes in the concentration, but that as
the ion concentration increased the coordination number of Na+
J. Mater. Chem., 2010, 20, 77887799 | 7791

View Online

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

Table 3 Average coordination numbers of the sodium and chloride ions

determined from the control runs
Run
(salt concentration/molal)

Na+ coord. number

Cl coord. number

C2 (0.0055)
C3 (0.55)

6.05  0.02
6.00  0.05

6.80  0.02
6.82  0.03

did drop slightly (to 5.3 at 4.1 M). The coordination numbers
reported in this study (Table 3) fall inside the range of typical
values given for both ions. Furthermore, we find no significant
change in coordination number over the range of concentrations
examined, with the differences lying in the range of the statistical
uncertainty.

Solvated FT30
The experimental density of FT30 is measured to be 1.38 g cm35
under standard conditions. Density profiles of the three different
arrangements of systems are shown in Fig. 4, 5 and 6. The density
of the water, the FT30 and support and the total density system
have been calculated. Defining the membrane bulk as being the
 of the membrane region, then the average system
central 80.0 A

Fig. 5 Density profiles of FT30 membrane taken from run P2, (a) the
full system, (b) the membrane-support interface and (c) the membranebulk solution interface.

Fig. 4 Density profiles of FT30 membrane taken from run P1, (a) the
full system, (b) and (c) close-ups of the profiles at the membrane-bulk
solution interfaces.

7792 | J. Mater. Chem., 2010, 20, 77887799

density was calculated as 1.34  0.01 g cm3, 97% of the experimental value, while the average water density in this region is
0.28  0.01 g cm3. The membrane-bulk solution interface is z
 wide, with no sharp transition point. Thus it is difficult
1020 A
to specify the membrane density precisely.
In contrast, the membrane-support interface is quite clearly
defined for the non-porous case, P2, with a peak in the density of
both the polyamide and the water. Here the polyamide is
responsible for the strongly ordered peak immediately adjacent
to the support, while the water ordering occurs in the second
peak when moving away from the interface. This is consistent
with the hydrophobic nature of the description of the coarsegrained polysulfone system. The perturbation of the membrane
by the support is found to be reasonable localised at the interface,
with little influence on the interface with the solution. The density
profile of the pore in run P3 shows an interesting feature; while
the total density around the pore is symmetrical, the water
density is not, there being a trough on the side of the pore with
the membrane. The reason for this is that some of the polyamide
chains have entered the mouth of the pore, thus disrupting the
sharp interface seen for the non-porous support.
The surface of the membrane at the interface with the solution
is found to be rough, commensurate with the thickness observed
in the density profiles, with pores, valleys and other features
This journal is The Royal Society of Chemistry 2010

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

View Online

Fig. 6 Density profiles of FT30 membrane taken from run P3, (a) the
full system, (b) the membrane-porous support interface and (c) the
membrane-bulk solution interface.

present. This is illustrated in Fig. 7 where snapshots of the

surface of the membrane in run P2 have been taken at 2 ns
intervals. These snapshots also show that that surface of the
polyamide membrane has some freedom of movement and
therefore dynamically evolves, with some pores closing and new
ones opening and with chains shifting position. This nonuniformity might be considered to be a problem for achievement
of overall salt-rejection, and indeed leads to variability in the
properties of the membrane, as will be demonstrated later.
The overall diffusion coefficient of water within the membrane
is 0.21  0.02  109 m2 s1, while Table 4 shows the breakdown
of diffusion coefficients for the different systems. These results
match well with previous simulations,5 where for a cross-linked
polymer at a density of 1.38 g cm3 a diffusion coefficient of 0.20
 109 m2 s1 was calculated. While the study of Harder et al.
gave a higher diffusion coefficient of 0.50  109 m2 s1 for the
water within the membrane, the ratio of the diffusion rate of
water within the membrane to that of the bulk was found to be
0.1, for the TIP3P water model used, similar to the present study.
This approximate order of magnitude reduction in the diffusion
rate of the water within the membrane compared to the diffusion
rate of water in bulk solution is consistent across both the results
reported in this paper and previous literature, despite the fact
that the polyamide membrane was constructed using three
This journal is The Royal Society of Chemistry 2010

Fig. 7 Snapshots of the surface the membrane forms with the bulk
solution taken from run P2 looking down the z-axis at (a) 0 ns, (b) 2 ns,
(c) 4 ns, (d) 6 ns, (e) 8 ns and (f) 10 ns. Water molecules have been
removed for clarity and the colouring depends on the z-coordinates of the
atoms, red bottom, grey middle, blue top, where the z-coordinate
becomes increasingly positive when moving from the membrane into
solution. The snapshots illustrate both that the membrane-water interface shows atomic roughness and some degree of fluidity.

different methods. It is therefore reasonable to conclude that any

effect of the preparation method has on the behaviour of the
system is relatively minor. The effect of a support, either porous
or non-porous, on the diffusion rate of the water within the
membrane is found to be negligible. The water molecules in the
membrane form into pockets which are connected to each other
by chains of water molecules hydrogen bonded to each other in
single file. A typical arrangement is shown in Fig. 8. In addition,
even when the system has no pressure gradient applied, the
diffusion of water molecules both into and out of the membrane
is observed. By comparing configurations 1 ns apart, it is possible
to estimate that the average rate of exchange of water molecules
between the bulk and the membrane is approximately 149
molecules ns1.
The addition of salt to the bulk solution does not change the
above properties particularly; the structure of the membrane
remains the same and the diffusion coefficient of water within the
J. Mater. Chem., 2010, 20, 77887799 | 7793

View Online
Table 4 Diffusion coefficients of water, Na+ and Cl within the membrane and solution, taken from the different simulation runs

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

Diffusion coefficient/109m2s1
Simulation

In solution

Run

Water

P1
P2
P3
P4
P5
P6
P7
P8
S1
S2
S3

2.71 
2.73 
2.64 
2.76 
2.66 
2.46 
2.55 
2.49 
2.74 
2.69 
2.54 

0.07
0.07
0.08
0.07
0.09
0.06
0.07
0.05
0.06
0.07
0.12

In membrane
Na

Cl

Water

Na

Cl

1.41  0.48
1.05  0.23
1.20  0.03
1.28  0.06
1.21  0.09

1.32  0.13

2.10 
1.93 
1.61 
1.75 
1.58 

1.76 

0.18  0.02
0.22  0.02
0.23  0.04
0.23  0.02
0.24  0.05
0.20  0.01
0.23  0.03
0.17  0.01
0.23  0.02
0.22  0.02
0.19  0.01

0.10  0.04
0.10  0.03
0.14  0.03

0.07  0.02
0.07  0.02
0.04  0.01

Fig. 8 Snapshot taken from a simulation illustrating the water structure

inside the membrane, consisting of clusters connected by linear chains of
water molecules. Here oxygen and hydrogen are shown in red and white,
respectively, while the polyamide is omitted for clarity.

Fig. 9 Density profile of FT30 membrane in contact with seawater (0.55

molal NaCl, run P8).

7794 | J. Mater. Chem., 2010, 20, 77887799

0.66
0.48
0.03
0.07
0.05
0.14

membrane is unaltered. As might be expected for a membrane

with good salt rejection properties, the salt ions do not show any
significant attraction to the membrane surface. There is no
congregation of salt ions at the membrane interface and no salt
ions enter the membrane during the timescale of the present
simulations. Fig. 9 shows the density profile for the system
without a support; the slight overlap of the polymer density with
the salt density is due to the surface roughness of the polymer
rather than any salt ions actually entering the membrane. Again
the diffusion coefficients of the salt ions in these simulations
match well with those of the control runs. The only anomalous
result is that for the sodium ion in run P5 where the diffusion
coefficient is lower than it should be. This is due to the fact that
the ion shows some interaction with the membrane surface
(though without ever actually entering the membrane) and is
compounded by the fact that at this high dilution the statistics
are rather poor. As per the situation for pure water, the presence
of the support in the simulation does not influence the results for
the saline solutions.
A number of runs with salt ions within the membrane layer
were also carried out to examine their properties, should they be
able to overcome the barrier to cross the membrane-solution
interface. In the first run, S1, a single Na+ and single Cl were
placed in the polyamide, some distance apart. For the second
run, S2, the initial coordinates of the Na+ were those of the initial
coordinates of the Cl ion in S1, while the starting position of the
Cl ion in S2 was the same as the starting position of Na+ in S1
(i.e. the initial positions of the ions in run S2 were reversed
relative to run S1). This attempted to ensure that any effect on
the ion due the initial configuration would be able to be observed.
For the third run, S3, the concentration of salt ions within the
membrane was increased so that seven ions of each species were
present in the membrane. Table 4 shows the diffusion coefficients
of the ions and water molecules within the membrane.
It appears that the presence of salt ions within the membrane
does not change the rate at which water molecules within the
membrane diffuse at these salt concentrations. Likewise, the
small changes that are seen in the diffusion coefficients of the ions
within the membrane with increasing salt concentration are
probably nothing more than statistical variation. It does seem
that Cl diffuses somewhat more slowly within the polyamide
than Na+; a result also found by Kotelyanskii et al.,5 and the
This journal is The Royal Society of Chemistry 2010

View Online
Table 5 Coordination numbers of ions to different groups when within the FT30 membrane
Coordination number
Species

Run

Water

Na

S1
S2
S3
S1
S2
S3

4.02 
5.19 
4.58 
3.97 
6.15 
4.82 

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

Cl

0.60
0.24
0.29
1.15
0.20
0.38

Carbonyl

Hydroxyl

Aromatic ring

Amide

2.09  0.62
0.74  0.22
1.28  0.21

0.01  0.03
0.21  0.14
0.24  0.15
1.62  0.52
0.03  0.03
0.84  0.23

0.65  0.28
0.22  0.16
0.71  0.30

0.30  0.26
0.03  0.03
0.26  0.15

opposite of the situation in aqueous solutions. Both ions experience at least an order of magnitude reduction in diffusivity
within the membrane relative to solution.
Table 5 gives the coordination environment of the salt ions
when they are within the solvated FT30. Both ions have their
solvation shell disrupted with some of the water molecules being
replaced by groups from the membrane. In the case of the sodium
ion, at least one water molecule in the solvation shell is replaced
by a carbonyl oxygen, while there is also some coordination
between the ion and hydroxyl oxygens, but to a lesser degree. The
situation with the chloride ion is less clear, but at least one, and
more usually two, water molecules are removed from the solvation shell, with the ion principally interacting with the hydrogen
of the carboxylic acid and the aromatic rings, as well some minor
coordination to the amide hydrogens. Kotelyanskii et al.5
proposed that the sodium ion was coordinated to five water
molecules and a carbonyl oxygen while the chloride ion was
coordinated to four waters, an amide hydrogen, a carbonyl
hydrogen and a phenyl ring. While these results are, on average,
similar to those presented in this paper, Kotelyanskii and coworkers proposed that these compositions of the solvation shell
were largely fixed, not changing either with time or environment.
The results of this study contradict this, with both much more
variety in the composition of the solvation shells being observed
and with the nature of the coordination environment changing
over the course of the simulation. The findings of the present
study are supported by the results of solid state 23Na NMR
experiments47 where both a negative chemical shift (suggesting
that some of nearest atomic neighbours are polymer atoms) and
significant heterogeneity in the molecular-scale Na+ environments were observed. The cause of the difference in the results
of this paper and Ref. 5 may simply be due to sampling, as the
simulations in the current study are not only longer, but also
sample a larger number of salt ions than the previous work.

 within the membrane layer. Five runs were

a point z 12 A
carried out for each ion, using different initial x- and y-coordinates, in order to ensure that different parts of the membrane
surface were sampled. As no difference is seen in the behaviour of
the non-porous support systems and the porous support systems,
the former arrangement was used for the umbrella runs as it is
computationally cheaper. The solution used in the umbrella
sampling runs had a salt concentration of 0.55 molal, with a few
salt ions already dispersed within the membrane layer, this setup
was used as it most closely resembles that of the real membranes,
where there is some diffusion of salt ions into the membrane.47
The ions within the membrane layer were well distributed, and
were well separated from the ions brought across the membranesolution interface. The duration of these simulations was 550 ps
for each umbrella, while the other simulation details are as
described in the method section.
The position of the membrane interface is difficult to pinpoint
exactly since the membrane surface is not sharply defined and the
ion is free to move in the x and y dimensions. However, the centre
point of each umbrella can be used as an approximation of the
ion position relative to the membrane-solution interface.
As the membrane surface has a relatively large surface area,
five umbrella runs were carried out for each ion in order to
sample different regions of the interface. Fig. 10 shows the free
energy profile for the five different runs of an Na+ ion entering
the membrane. The results cover quite a wide range of energies,
but there are some features that seem to be common to most
runs. As the ion approaches the membrane there is a slight

Free energy surface for ion migration across the membranesolution interface
On the time scale of the present simulations no salt ions were
observed entering or leaving the membrane layer. Thus, in order
to determine whether this was a result of a thermodynamic
barrier or a kinetic barrier, and to gain an idea of the magnitude
of any barrier, umbrella sampling methods were used to
construct the free energy surface for ions migrating into the
membrane from solution.
 were set up in the
A set of 40 umbrellas at a spacing of 0.5 A

z-axis, spanning from z7 A above the membrane surface to
This journal is The Royal Society of Chemistry 2010

Fig. 10 Free energy profile for a Na+ion to enter the FT30 membrane
determined using umbrella sampling methods. An increasing negative
position indicates the ion is moving into the membrane.

J. Mater. Chem., 2010, 20, 77887799 | 7795

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

View Online

the membrane surface, with some pathways proving far less

energetically unfavourable than others. In all of the runs the
largest jump in the free energy profile occurs when the ion is some
way into the membrane layer, where the membrane starts to
become more dense, and after this peak there is often a local
minimum. Fig. 11(a) shows the coordination number of the ion
to water and carbonyl oxygens as a function of umbrella position
for one run; here the disruption of the water solvation shell as the
membrane is approached can be seen clearly. In addition, while
the overall coordination number of the sodium is relatively
constant, water molecules and carbonyl oxygens exchange with
each other in the solvation shell of the ion depending on the local
environment. Fig. 11(b) shows the number of water molecules
coordinated to the ion for three of the different runs (numbers 1,
3 and 4). For run 1, which has a low energy barrier, there are
fewer sharp changes in the coordination number compared with
the other two runs. This indicates that the height of the energy
barrier for the diffusion of the system is linked to the amount of
exchange the solvation shell of the ion undergoes as it diffuses
into the membrane. Comparing Fig. 10 with Fig. 11(b) seems to
confirm this, since points at which there are sharp changes in

Fig. 11 Coordination number of Na+ ion as a function of the umbrella

position. (a) shows the coordination number of the sodium ion to water
molecules, carbonyl oxygens and the overall coordination number for the
Na+ umbrella sampling run 3. (b) shows the coordination of the sodium
ion to water molecules in runs 1, 3 and 4.

increase in the free energy, due to the disruption of the solvation

shell around the ion. The lack of any sharp barrier at the point
where ions enter the membrane indicates the non-uniformity of

Fig. 12 Free energy profile for a Cl ion to enter the FT30 membrane
determined using umbrella sampling methods. An increasing negative
position indicates the ion is moving into the membrane.

7796 | J. Mater. Chem., 2010, 20, 77887799

Fig. 13 Coordination number of Cl ion as a function of the umbrella

position. (a) Shows the coordination number of the chloride ion to water
molecules, hydroxyl groups, aromatic rings and the overall coordination
number for the Cl umbrella sampling run 1. (b) Shows the coordination
of the chloride ion to water molecules in runs 1, 3 and 4.

This journal is The Royal Society of Chemistry 2010

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

View Online

Fig. 14 Snapshots taken from a sequence of umbrella sampling runs

with a Cl ion crossing the membrane-solution interface. The system is
viewed looking down the y-axis with the interface normal running
vertically. The position of the chloride ion with respect to the membrane
 (b) 4.4 A,
 (c) 0.81 A,
 (d) 5.6 A,
 (e) 9.2 A
 and (f)
interface is (a) 8.4 A,
 with negative position indicating that the ion is within the
12.4 A,
membrane. The water molecules, with the exception of those in the first
solvation shell of the chloride ion, have been removed for clarity, the
membrane has been coloured blue, the chloride ion green, the water
oxygens and the water hydrogens red and white, respectively. The
snapshots illustrate the type of pathway taken by both ions when
diffusing across the membrane-solution interface, with water molecules
and polymer groups exchanging with each other in the solvation shell of
the ion.

composition of the solvation shell of the ion profile coincide with

increases in the free energy.
The free energy profile of the Cl ion, Fig. 12, shows both
similarities and differences to that of the sodium ion. The free
 into the membrane
energy required to penetrate a depth of 8 A
varies over a relatively large range, z2595 kJ mol1. Unlike the
case of Na+, however, there does not seem to be any sharp
increase in the free energy followed by a local minima. Instead
the profiles tend to show a smoother, more gradual increase in
free energy, up to a plateau value in two cases. As in the case of
the sodium ion, the exchange of water molecules and polymer
groups (in this case aromatic rings and the hydroxyl hydrogen) is
seen to occur, while the overall coordination number remains
relatively constant, as shown in Fig. 13(a). This behaviour can
also been seen Fig. 14, where snapshots have been taken from an
umbrella sampling run illustrating the type of diffusion pathway
the ions take and how different groups coordinate to the chloride
ion as it passes through the membrane-solution interface.
This journal is The Royal Society of Chemistry 2010

Additionally, the same relationship between the magnitude of the

free energy barrier and change in water coordination number,
Fig. 13(b), seen for the sodium ion also applies for the chloride
ion; run 3 is the lowest energy pathway and has the smoothest
water coordination profile.
It is difficult to determine if the free energy barrier to the
diffusion of ions into the membrane is greater for one ion than
the other from the results obtained as they vary over such a wide
range. One major factor in determining the magnitude of the free
energy barrier for either ion is the amount of exchange the
solvation shell of the ion undergoes as the ion diffuses into
the membrane. Runs where there are sudden, large changes in the
solvation shell are the runs that have the higher energy barriers.
In turn, this can be correlated with the variable structure of the
membrane-solution interface. In some regions there are larger
channels of water molecules that permeate the membrane and
these obviously represent the preferred pathways for ion diffusion.
A factor that contributes to the energy barrier is the fact that
as a single ion is driven to diffuse into the membrane the counter
ion is left behind in solution. For the above calculations each ion
is assumed to enter the membrane independently, but the energy
barrier for the ions to enter the membrane as an ion pair might be
lower. Attempts were made to test this theory by drawing an ion
pair through the membrane. However, the lifetime of an ion pair
is too short, with the ion pairs dissociating before they can enter
the membrane. In addition, when an ion is pulled into the
membrane no diffusion of any counter ion into the membrane is
observed.
Although it is difficult to determine the free energy for an ion
pair to enter the membrane, it is possible, however, to determine
the free energy profile of a Cl ion entering the membrane after
a sodium ion has already diffused into the membrane. This was
done by taking, from three of the Na+ umbrella sampling runs (1,
3 and 4), configurations where the sodium ion was in a local
energy minimum, and then performing further umbrella
sampling runs on a chloride ion, to compute the free energy

Fig. 15 Free energy barrier for a Cl ion to enter the FT30 membrane
after a Na+ ion has already entered the membrane, determined using
umbrella sampling methods. The vertical lines show the average positions
of the sodium ions throughout the three different umbrella sampling
runs.

J. Mater. Chem., 2010, 20, 77887799 | 7797

View Online

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

profile to bring this ion to the same depth into the membrane as
the sodium ion minimum. The results of these three umbrella
simulations are shown in Fig. 15. The free energy profiles for
these three runs fall within the range of energy values generated
for a chloride ion entering the membrane without any corresponding sodium ion and are consistent with those results.
However, what is surprising is that the presence of the sodium
ion within the membrane does not appear to assist in the diffusion of Cl ions into the membrane, indicating that charge
separation is only one in factor in determining the free energy
profile; other factors, such as the desolvation of the ion, will also
play their part.

Conclusions
Molecular dynamics simulations of the FT30 polyamide reverse
osmosis membrane have been carried out in order to determine
the detailed properties of this system, at the atomic level, relevant
to desalination. Diffusion coefficients of the species in both the
bulk solution and within the membrane layer itself have been
calculated, showing that water diffuses approximately an order
of magnitude slower within the membrane layer than in bulk
solution. The presence of a support, either porous or nonporous, had a negligible effect in these simulations where there is
no pressure difference across the system. The diffusion of Na+
and Cl ions in the polyamide is slower than water, with the
reduction relative to solution being similar. However, the diffusion of chloride ions is slowed to a greater extent than for
sodium, leading to a reversal in the relative diffusion rates in
comparison to saline solution. The observation that the rates of
diffusion for all species are decreased by approximately the same
factor has implications for the design of more efficient
membranes based on the polyamide system. To reduce the energy
consumption, the ideal situation would be to increase the flow
rate of water for a given pressure, while maintaining the saltrejection at the same level. If the diffusivities of all species are
similarly influenced by the membrane structure, then the selectivity may remain constant even at higher flow rates, thus making
it feasible to reduce the energetic cost.
How the ions interact with the polyamide has been investigated by determination of the coordination numbers of the ions
with various groups within the membrane. Even when the ions
are within the membrane, water molecules account for over 50%
of the species coordinated to the ions. From the polymer, Na+
coordinates most strongly with the carbonyl oxygen, and shows
a lesser degree of interaction with the hydroxyl oxygen. The
chloride ion loses a greater percentage of water molecules from
its solvation shell, and shows some coordination with the
hydroxyl hydrogen, the aromatic rings and the amide hydrogens.
The greater coordination of polymer groups to the Cl ion is
likely to be the reason for the greater relative reduction of the
diffusion coefficient of the chloride ion compared to the sodium
ion within the membrane.
In order to investigate the behaviour at the membrane interface a number of runs were carried out using umbrella sampling
methods to determine the free energy surface associated with an
ion diffusing into the polyamide. The results show that the rate of
increase in free energy varies quite considerably, making it
difficult to determine if one ion would find it easier to enter the
7798 | J. Mater. Chem., 2010, 20, 77887799

membrane than the other. One result seen in every run was that
even before the ion entered the membrane the disruption of the
solvation shell around the ion resulted in an energetic penalty. In
addition, the profiles did not show a sharp peak in the free energy
at the membrane interface but a more gradual increase as the ion
diffused into the membrane. The size of the free energy increase
partly depends upon the degree of disruption of the solvation
shell that an ion undergoes, with the higher energy pathways
showing both more frequent exchanges of the species coordinated to the ions and a greater degree of change in the composition of the solvation shell at any one time.
Although the free energy profiles for both sodium and chloride
diffusion into the membrane show considerable variability, corresponding to the different range of environments possible within
FT-30, there are ultimately no facile pathways through the
 thick membrane used in the present simuapproximately 100 A
lations. Given that the experimental membrane used is typically
twenty times this thickness, this will be even more true of the real
material. Attempts to force an ion to migrate through the
membrane show an increasing free energy profile with no signs of
 This behaviour can be
a plateau, at least to a depth of 10 A.
explained in terms of the increasing Coulomb energy of separation of the ion and the counter-balancing charge remaining
above the interface, with a variable effective dielectric constant
depending on the water structure in the membrane and its ability
to reorient to screen the interaction. It might be expected that this
Coulomb interaction would reach the point where it is strong
enough to allow the counter ion to also enter the membrane
itself. However, this is not observed in any of the present simulations. Even when using umbrella sampling to examine the free
energy to bring a chloride anion toward a net sodium ion that
already resides within the membrane, there remains a significant
resistance to creating a separated ion pair. This finding is
important since it suggests the absence of a low energy pathway
for sodium chloride through the less hydrophilic membrane by
ion pairing.
Molecular dynamics studies can provide detailed information
about RO membranes that can often be difficult to determine by
experimental methods. The simulations carried out in this work
describe the behaviour of water molecules, sodium cations and
chloride anions and their interaction with the FT30 membrane
under standard conditions (i.e. atmospheric pressure). This
provides a reference point for future simulations investigating the
influence of changes to the membrane structure, as well as the
behaviour of the membrane under applied pressure, as would be
the case during reverse osmosis. Furthermore, future investigations can now consider the interaction of foulants with the
membrane-solution boundary region having characterised the
properties of the clean interface. Through combining the insights
obtained from simulation with the information obtained by
experimental methods, it is hoped that this will allow new, more
efficient membranes to be developed in the future.

Acknowledgements
This research was supported by the Commonwealth Scientific
and Research Organisation (CSIRO) under the Water For A
Healthy Country Flagship. Additionally the authors would like
to thank the cluster members, especially Stephen Grey, for their
This journal is The Royal Society of Chemistry 2010

View Online

comments and discussion. JDG would like to thank the

Australian Research Council (ARC) for a Professorial fellowship. Both iVEC and the National Computational Infrastructure
(NCI) are acknowledged for the provision of computational
resources.

Downloaded by George Washington University on 02 March 2012

Published on 05 August 2010 on http://pubs.rsc.org | doi:10.1039/C0JM01545H

References
1 C. Fritzmann, J. Loewenberg, T. Wintgens and T. Melin,
Desalination, 2007, 216, 176.
2 A. D. Khawaji, I. K. Kutubkhanah and J.-M. Wie, Desalination,
2008, 221, 4769.
3 C. Charcosset, Desalination, 2009, 245, 214231.
4 I. El Saliby, Y. Okour, H. K. Shon, J. Kandasamy and I. S. Kim,
Desalination, 2009, 247, 114.
5 M. J. Kotelyanskii, N. J. Wagner and M. E. Paulaitis, J. Membr. Sci.,
1998, 139, 116.
6 M. G. Khedr, Desalination, 2003, 153, 295304.
7 B. Mi, O. Coronell, B. J. Marinas, F. Watanabe, D. G. Cahill and
I. Petrov, J. Membr. Sci., 2006, 282, 7181.
8 P. S. Singh, S. V. Joshi, J. J. Trivedi, C. V. Devmurari, A. P. Rao and
P. K. Ghosh, J. Membr. Sci., 2006, 278, 1925.
9 B. Mi, D. G. Cahill and B. J. Marinas, J. Membr. Sci., 2007, 291, 7785.
10 S. Murad and J. Lin, Chem. Eng. J., 1999, 74, 99108.
11 S. Murad and L. C. Nitsche, Chem. Phys. Lett., 2004, 397, 211215.
12 R. Qiao, J. G. Georgiadis and N. R. Aluru, Nano Lett., 2006, 6, 995
999.
13 A. V. Raghunathan and N. R. Aluru, Phys. Rev. Lett., 2006, 97,
024501.
14 E. Harder, D. E. Walters, Y. D. Bodnar, R. S. Faibish and B. Roux, J.
Phys. Chem. B, 2009, 113, 1017710182.
15 J. E. Cadotte, R. J. Petersen, R. E. Larson and E. E. Erickson,
Desalination, 1980, 32, 2531.
16 R. J. Petersen, J. Membr. Sci., 1993, 83, 81150.
17 W. L. Jorgensen and J. Tirado-Rives, J. Am. Chem. Soc., 1988, 110,
16571666.
18 W. L. Jorgensen, D. S. Maxwell and J. Tirado-Rives, J. Am. Chem.
Soc., 1996, 118, 1122511236.
19 R. C. Rizzo and W. L. Jorgensen, J. Am. Chem. Soc., 1999, 121, 4827
4836.
20 W. Damm and W. E. V. Gunsteren, J. Comput. Chem., 2000, 21, 774
787.

This journal is The Royal Society of Chemistry 2010

21 P. M. Anderson and M. R. Wilson, J. Chem. Phys., 2004, 121, 8503

8510.
22 C. D. Snow, L. L. Qiu, D. G. Du, F. Gai, S. J. Hagen and V. S. Pande,
Proc. Natl. Acad. Sci. U. S. A., 2004, 101, 40774082.
23 B. Adisa and D. Bruce, J. Phys. Chem. B, 2005, 109, 75487556.
24 D. P. Tieleman, J. L. MacCallum, W. L. Ash, C. Kandt, Z. T. Xu and
L. Monticelli, J. Phys.: Condens. Matter, 2006, 18, S1221S1234.
25 C. D. Daub, K. Leung and A. Luzar, J. Phys. Chem. B, 2009, 113,
76877700.
26 H. W. Horn, W. C. Swope, J. W. Pitera, J. D. Madura, T. J. Dick,
G. L. Hura and T. Head-Gordon, J. Chem. Phys., 2004, 120, 9665
9678.
27 I. S. Joung and T. E. Cheatham, J. Phys. Chem. B, 2008, 112, 9020
9041.
28 I. S. Joung and T. E. Cheatham, J. Phys. Chem. B, 2009, 113, 13279
13290.
29 W. Smith and T. R. Forester, J. Mol. Graphics, 1996, 14, 136141.
30 U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee and
L. G. Pedersen, J. Chem. Phys., 1995, 103, 85778593.
31 S. Nose, J. Chem. Phys., 1984, 81, 511519.
32 W. G. Hoover, Phys. Rev. A: At., Mol., Opt. Phys., 1985, 31, 1695
1697.
33 M. Bonomi, D. Branduardi, G. Bussi, C. Camilloni, D. Provasi,
P. Raiteri, D. Donadio, F. Marinelli, F. Pietrucci, R. A. Broglia
and M. Parrinello, Comput. Phys. Commun., 2009, 180, 19611972.
34 M. Patra and M. Karttunen, J. Comput. Chem., 2004, 25, 678689.
35 K. F. Lau, H. E. Alper, T. S. Thacher and T. R. Stouch, J. Phys.
Chem., 1994, 98, 87858792.
36 M. W. Mahoney and W. L. Jorgensen, J. Chem. Phys., 2001, 114,
363366.
37 M. J. Kotelyanskii, N. J. Wagner and M. E. Paulaitis, Comput. Theor.
Polym. Sci., 1999, 9, 301306.
38 Y. Marcus. Ion Properties, Marcel Dekker, New York, 1997.
39 L. A. Woolf, J. Phys. Chem., 1978, 82, 959962.
40 A. P. Lyubartsev and A. Laaksonen, J. Phys. Chem., 1996, 100,
1641016418.
41 K. Toukan and A. Rahman, Phys. Rev. B: Condens. Matter, 1985, 31,
26432648.
42 L. X. Dang and D. E. Smith, J. Chem. Phys., 1993, 99, 69506956.
43 D. E. Smith and L. X. Dang, J. Chem. Phys., 1994, 100, 37573766.
44 K. J. Muller and H. G. Hertz, J. Phys. Chem., 1996, 100, 12561265.
45 H. Uchida and M. Matsuoka, Fluid Phase Equil., 2004.
46 S. Koneshan and J. C. Rasaiah, J. Chem. Phys., 2000, 113, 81258137.
47 X. Xu and R. J. Kirkpatrick, J. Membr. Sci., 2006, 280, 226233.

J. Mater. Chem., 2010, 20, 77887799 | 7799