Computers and Chemical Engineering 98 (2017) 3149

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Actuation of spatially-varying boundary conditions for reduction of

concentration polarisation in reverse osmosis channels
Pesila Ratnayake, Jie Bao
School of Chemical Engineering, The University of New South Wales, UNSW, Sydney, NSW 2052, Australia

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

Article history: Reduction of concentration polarisation is of great value in reverse osmosis membrane systems. Concen-
Received 7 April 2016 tration polarisation leads to a reduction in ux, which corresponds to a reduction in separation efciency.
Received in revised form This paper studies an approach by which to reduce concentration polarisation in reverse osmosis chan-
30 September 2016
nels using a steady, spatially variant slip velocity prole. A method is developed to identify the most
Accepted 29 November 2016
Available online 9 December 2016
effective wall slip velocity prole for increasing the diffusive driving force away from the wall, which in
turn increases mass transfer away from the membrane and reduces concentration polarisation. The non-
linear partial differential equations (PDEs) that govern uid and mass transport behaviours are difcult
Keywords:
Concentration polarisation to solve. In this work, an approximate solution to the nonlinear system is developed using systems of lin-
Reverse osmosis earised ordinary differential equations (ODEs) to approximate the behaviour of the PDEs and determine
Innite-dimensional systems the steady-state actuation prole that most effectively increases mass transfer at the wall. This leads to
Distributed actuation a systematic method for determination of a spatially-varying actuation prole (the most effective slip
velocity prole) that decreases concentration polarisation in reverse osmosis membrane channels.
2016 Elsevier Ltd. All rights reserved.

1. Introduction

Reverse osmosis is a major component of water treatment, and is particularly important in treatment of water with high concentrations
of monovalent salts such as sea water and brackish water (Chen et al., 2004). As uids move through semipermeable membranes from
the feed channel into the permeate channel, solutes rejected by the membrane remain in the feed channel (Alexiadis et al., 2005). This
leads to a buildup of solutes close to the membrane surface, known as concentration polarisation (Alexiadis et al., 2005; Chen et al., 2004).
The concentration polarisation layer accumulates until the rate of convective movement of solutes towards the membrane balances the
convective downstream movement of solutes close to the membrane surface, the diffusive mass transport away from the membrane
surface, and any movement of solutes through the membrane (Fimbres-Weihs and Wiley, 2010). Concentration polarisation is associated
with an increased osmotic pressure on the feed side of the membrane, resulting in a decrease in permeate ux and therefore decreasing
the efciency of the pressure-driven separation (Alexiadis et al., 2005). Additionally, concentration polarisation is a cause of fouling in
membrane systems, which leads to membrane damage and reduced permeate ux (Alexiadis et al., 2005; Chen et al., 2004). Therefore,
methods to reduce concentration polarisation are of great value.
A method by which uid and solute behaviour can be altered is hydrodynamic manipulation, wherein the uid ow prole is directly
manipulated either by application of body forces (Du and Karniadakis, 2000) or by changing the ow prole at the boundary (Jovanovic,
2006). Hydrodynamic boundary manipulation is typically achieved using techniques such as wall-tangential (forcing uid to move along
the wall) and/or wall-normal actuation (injecting or removing uid at the wall). Wall-tangential actuation can be achieved by electro-
osmosis, moving walls such as rotating cylinders, or appropriately angled uid jets. Wall-normal actuation has been proposed using uid
jets at the wall to add uid to and/or remove uid from the channel.
Several approaches have been used in order to study and control uid ow (Aamo and Krstic, 2003). DAlessandro et al. (1999) developed
an algorithm by which they maximised entropy in uid ow as a measure of mixing. Armaou and Christodes (2000) developed a non-
linear controller for wave suppression, based on the Kortewegde VriesBergers and KuramotoSivashinsky equations, using distributed

Corresponding author.
E-mail address: J.Bao@unsw.edu.au (J. Bao).

http://dx.doi.org/10.1016/j.compchemeng.2016.11.045
0098-1354/ 2016 Elsevier Ltd. All rights reserved.
32 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

actuators. By further study, Christodes and Armaou (2000) showed that the maximum number of actuators required for stabilisation
of the KuramotoSivashinsky equation is 5, subject to the value of the instability parameter. Baker et al. (2000) developed a nonlinear
controller for stabilisation of Burgers equation and the 2D NavierStokes equations for incompressible ow using Galerkins method and
approximate inertial manifolds to derive systems of ODEs. Bamieh et al. (2001) achieved maximisation of mixing for both nite and innite
time horizons, based on a linear time-invariant system by using a linear quadratic (LQ) formulation to nd the H -norm of the system and
a maximising feedback controller. Aamo et al. (2003) developed ow destabilisation by reversing the sign of a Lyapunov-based stabilising
controller. Balogh et al. (2005) developed a method by which to maximise a mixing index using wall-normal actuation in 3D pipe ow
based on an optimal Lyapunov-based control law. Ouyang et al. (2013) maximised an objective function related to uid mixing using a
pseudo-feedback approach that nds the most effective electro-osmotic input waveform. A major limitation of the bulk of work done on
ow control with distributed actuation with respect to mass transport is that solute behaviour is not directly controlled, and solute mixing
is assumed to have improved based on an increase in hydrodynamic mixing or a more even distribution of non-diffusive tracer particles.
Previous work by Liang et al. (2014) studies the effect of wall-tangential slip velocities on mass transfer at the membrane wall, as
well as friction factor and permeate ux under steady-state conditions. The slip velocity proles studied included uniform slip velocities,
both in the streamwise (downstream) and the upstream directions, as well as non-uniform proles created by electro-osmosis using 2
cylindrical electrodes with equal and opposite charge located under the membrane channel. They found that streamwise uniform slip
velocities increased mass transfer and permeate ux and decreased friction factor, while upstream uniform slip velocities had the opposite
effect (Liang et al., 2014). Similarly, for non-uniform proles, increased mass transfer and permeate ux and decreased friction factor were
observed in regions where the slip velocity was in the streamwise direction, with the opposite effect observed in regions where the slip
velocity was in the upstream direction (Liang et al., 2014). Additionally, the non-uniform slip velocity prole tended to have a larger effect
than the uniform slip velocity prole. However, these proles were designed based on simplicity of analysis (uniform) and simplicity of
actuation (non-uniform). Therefore these proles are likely to be suboptimal, and may be improved by optimisation of the slip velocity
prole.
Partial differential equations (PDEs) are often converted into innite-dimensional systems of ordinary differential equations (ODEs) by
spatial discretisation or decomposition into an innite set of basis functions (Curtain and Zwart, 1995; Callier and Winkin, 1992; Moghadam
et al., 2014, 2012, 2013). This allows for systems governed by PDEs to be approximated and studied using techniques developed for systems
of ordinary differential equations. Previous work by the authors of this paper used linearised models of 2D laminar channel ow to nd the
most effective spatially and temporally oscillatory wall-tangential velocity prole for velocity oscillations in regions of the uid (Ratnayake
et al., 2015).
Other numerical studies consider reduced-order approaches for heat and mass transport, as well as reactor systems. Adomaitis (2003)
developed an approach for study of distributed parameter systems by producing trial functions for reduced-basis discretisation, studying
chemical vapour deposition as an example. Reduced order approaches allow for signicant reductions in computational load by determining
or focusing on the spatial modes that have the largest effect on the system behaviour (Adomaitis, 2003). Setiawan et al. (2015) developed
a reduced order model for mass transport behaviour in membrane channels using reverse osmosis by limiting the number of spatial
modes studied and approximating the hydrodynamic behaviour as static. This greatly reduced computation time whilst achieving similar
simulation results to nonlinear computational uid dynamics (CFD) approaches. However, these approaches are limited by their operating
conditions, and modelling far from the one predened steady state can lead to signicant error.
Yao et al. (2010) developed an approach for computation of cyclic steady states in oscillatory or periodic processes, using an integration
scheme as well as prediction and correction in order to efciently determine the behaviour of the system at each period. This approach was
able to accurately predict the steady-state behaviour of oscillatory distributed parameter systems, but requires online measurement in
order for the integration to be conducted accurately. Yao et al. (2013) developed a networked model predictive control (MPC) approach to
regulate distributed parameter systems, using the dominant eigenmode and the Galerkin method to convert the system into an ODE. This
approach was able to effectively control the diffusion-reaction process presented, using either static or adaptive communication policies,
but requires online measurement for feedback control.
Concentration polarisation is of great interest in reverse osmosis membrane systems because it contributes signicantly to the oper-
ational costs of producing water. Song et al. (2003) showed that for high concentration feeds, such as those in seawater desalination,
concentration polarisation can have a signicant impact on permeate ux. Zhu et al. (2009a) discussed the importance of considering
concentration polarisation and the osmotic pressure at the membrane surface when calculating the minimum operating (transmembrane)
pressure required in order to achieve ux through the membrane length. Zhu et al. (2009a,b, 2010) develop a number of cost optimisation
strategies for reverse osmosis membrane systems and plants, noting that concentration polarisation impacts on costs by increasing the
required cross-ow velocity. Additionally, they note that mineral scaling and fouling form constraints to the optimisation of permeate ux
(Zhu et al., 2009a).
Many nonlinear distributed parameter systems such as membrane systems, heat exchangers and chemical reactors can be improved
by using distributed boundary actuation, since distributed boundary actuation can be used to redirect ow to achieve better performance.
However, these systems often lack real-time distributed measurement. This is particularly the case for measurement of concentration.
Whilst some new technologies such as electrical resistance tomography (ERT) have been shown to provide distributed real-time measure-
ment (Shari and Young, 2011), applications of such technologies are limited to date. Therefore, this paper proposes a novel approach for
changing steady states in nonlinear distributed parameter systems using spatially-varying boundary actuation without requiring online
measurement. This is demonstrated by decreasing steady-state concentration polarisation using slip velocities applied at the surface of a
reverse osmosis membrane. By performing steady-state analysis on successive locally linearised state-space equations, a small, spatially-
variant steady-state perturbation slip velocity prole is calculated such that it provides the greatest increase in mass transfer away from
the wall per unit slip velocity input. By applying the small perturbation, a new (nearby) steady state is reached. By repeating this process
at each successive steady state, a large increase in overall mass transfer coefcient is achieved. This approach approximately determines
the perturbation slip velocity prole that has the greatest effect of increasing mass transfer at the wall. This paper also proposes a new,
generalised innite-dimensional linearised model of hydrodynamics and mass transfer in 2D channel ow. This generalised linearised form
allows for fast simulation and analysis about any given steady-state prole. A new method is proposed in this paper by which to identify
the most effective tangential wall (slip) velocity prole for reduction of concentration polarisation by increasing the diffusive driving force
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 33

away from the membrane wall. The nal resulting actuation prole, which approximates the optimal steady-state actuation prole for the
nonlinear reverse osmosis membrane system, is also provided.
This paper is structured as follows: the system description and problem formulation are provided in Section 2. The innite dimensional
model is constructed and the concentration polarisation reduction algorithm is provided in Section 3. The numerical results showing the
change in solute concentration gradient, as well as the slip velocity proles and discussion are provided in Section 4. The paper is concluded
in Section 5.

2. System description and problem formulation

2.1. System description

The focus of this paper is concentration polarisation in a 2D rectangular membrane channel in the domain x [0, L], y [1, 1]. All
variables and parameters used in this paper are dimensionless unless otherwise specied. The steady state ow is laminar and the ow
prole is parabolic. Fluid enters from the left and exits through the right as shown in Fig. 1.
The membrane, located at y = 1 as shown at the bottom of Fig. 1, is treated as a dissolving wall between x = L1 and x = L. The dissolving
wall condition is commonly used in membrane simulation studies. This involves approximating the permeate solvent (water) ux to be
zero, and approximating the solute prole as satisfying the condition that the solute mass fraction at the membrane is xed. Since permeate
ow through reverse osmosis membranes is small in comparison to the bulk ow, the effect of permeation on the solute mass fraction
prole is typically small in reverse osmosis (Fimbres-Weihs and Wiley, 2010). Additionally, the condition of xed wall solute mass fraction
has been shown to generate a solute mass fraction prole that closely resembles solute mass fraction proles in system with permeation
(Fimbres-Weihs and Wiley, 2010). For systems with permeation, solute behaviour and hydrodynamics are coupled due to the effect of
wall solute mass fraction on solvent ux. The dissolving wall approximation removes the need for this coupling, making analysis more
straightforward and reducing computational load. In this study, the actuation prole that increases mass transfer away from the membrane
wall is determined for x [L1 , L2 ]. The region upstream from L1 is the inlet region, where no mass transfer occurs at the wall. The region
downstream of L2 behaves as a dissolving wall, but is included to minimise errors in the actuation prole due to discontinuities introduced
by applying a periodic boundary condition in the x-direction. These upstream and downstream regions act as a buffer and allow for more
accurate calculation of the actuation prole.
Concentration polarisation is often dened as the ratio of wall concentration to bulk concentration in permeable wall systems (Liang et al.,
2014). However, for dissolving wall systems this measure has little value since both values are xed. Instead, concentration polarisation is
based on the concentration gradient at the wall, which is a measure of the concentration adjacent to the membrane wall. This is directly
related to the mass transfer coefcient, such that an increase in mass transfer coefcient corresponds to a decrease in concentration
polarisation in dissolving wall systems.
Since mass transport and heat transport have largely similar behaviours, the approach and the ndings presented in this paper also
have implications for systems dealing with heat transfer in channel ows such as studied by Liu et al. (2010).

2.2. Problem formulation

This study aims to decrease concentration polarisation in the channel presented in Section 2.1 by applying a steady-state slip velocity
prole. Spatially variant slip velocities can be achieved via electro-osmosis (Liang et al., 2014) or by micropumps with an appropriate
orientation (Abhari et al., 2012). The approach proposed in this paper requires the system to be at steady state before commencing. Once
the nal steady-state slip velocity prole is applied as a result of this approach, the system can be run at steady state with the slip velocity
in effect for the remainder of its operating time. Therefore, this approach can be understood as a form of startup procedure, initiated once
the system has reached its initial steady state.
For systems with real-time measurement readily available, feedback control approaches such as gain scheduling can be used. However,
distributed measurement of uid ow and solute concentration throughout the uid domain are generally unavailable in distributed
parameter systems, particularly in industrial applications. This makes distributed feedback control limited or impractical. Therefore, this
study uses an open-loop method for reduction of concentration polarisation based on increase in mass transfer at the wall using steady-state
analysis.

Fig. 1. Diagram of a channel segment with arbitrary slip velocity at the bottom wall.
34 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

Since the system is nonlinear, the use of a single linearised system model may cause errors when applied over large changes in the
operating conditions. A spatially-nonuniform perturbation slip velocity actuation prole is calculated based on successive linearised state-
space equation, reecting the steady-state hydrodynamic and solute proles, such that it locally maximises the ratio of 2-norms of steady-
state mass transfer enhancement at the wall to steady-state perturbation slip velocity magnitude. This drives the system to a new steady
state or stage. At this new stage, a new spatially-nonuniform perturbation actuation prole is calculated based on the new linearised
state-space equation and so on. By repeating this process and adding to the overall slip velocity, hydrodynamic and solute concentration
proles for each applied perturbation slip velocity input, large changes can be made to the steady-state mass transfer without requiring
measurement.
This can be summarised as the following steps:

1. Linearise the governing partial differential equations about the initial steady state, establishing perturbation variables and initial steady-
state proles.
2. Convert the linearised partial differential equations into a system of linear ordinary differential equations using spatial discretisation
and spectral decomposition.
3. Establish output equation and objective function based on average mass transfer coefcient.
4. Determine perturbation slip velocity prole that maximises increase in mass transfer coefcient about the steady state.
5. Calculate the new steady state vectors and overall average mass transfer coefcient based on a small perturbation with prole determined
in Step 4, and apply new steady-state vectors to update the system of linear ODEs in Step 2 to the new steady state.
6. Repeat Steps 45 based on the new state-space equation until a sufcient increase in mass transfer coefcient is achieved.

The loop is terminated once a sufcient number of iterations has been conducted (or a goal of improving the steady-state mass transfer at
the wall by a certain amount has been achieved). By using a stage-wise steady-state approach, this analysis avoids dealing with dynamics
when changing between steady states. This is more practical since online measurement is unavailable to follow the system dynamics.
The main consideration that affects this approach is the magnitude of the perturbation slip velocity for each iteration, addressed in Step
4. For large inputs, linearised models can produce large errors in system response compared to the nonlinear models on which they are
based. By applying Steps 45 above innitely many times for inputs of innitesimal magnitude, the errors due to linearisation tend to zero
and this approach solves the nonlinear system exactly. However, performing Steps 45 innitely many times in this way is impractical. For
practicality, this study applies a nite number of small inputs to the system based on linearised equations, updating the linearised model
after applying each small input. This leads to a large increase in mass transfer away from the wall, and therefore reduction of concentration
polarisation, achieved with only small linearisation errors. Steps 15 are described in Section 3.
The result of this algorithm is a large increase in steady-state mass transfer from the wall due to a steady-state, spatially variant slip
velocity prole applied at the membrane surface as shown in Section 4.

3. Model development and reduction of concentration polarisation

3.1. Linearisation

Optimisation of systems governed by nonlinear partial differential equations is very challenging. This study simplies the problem by
linearising the model and focusing on local optimisation to achieve reductions in concentration polarisation. The hydrodynamic model
used in this study is the NavierStokes equation for 2D ow of incompressible Newtonian uids, written in terms of the stream function
(x, y, t) as follows:
3 3
 4 4 4
 3 3 3 3
4
+ = +2 + + + , (1)
x 2 t y2 t Re x4 x 2 y2 y4 x y2 y y x3 x2 y x x y3

where /y(x, y, t) = vx (x, y, t) is the uid velocity in the x-direction, /x(x, y, t) = vy (x, y, t) is the uid velocity in the y-direction,
and Re = 
U d/ is the uid density (kg m3 ), U is the average crossow velocity (m s1 ), d is the hydraulic
is the Reynolds number. 
diameter (twice the channel height) (m) and  is the uid dynamic viscosity (kg m1 s1 ). The NavierStokes equation is linearised about
the operating point to obtain a perturbation equation as follows:
3 3
 4 4 4
 3 3 3 3 3
  4    s  s  s  s  s 
+ = +2 + + +
x 2 t y2 t Re x4 x 2 y2 y4 x y2 y y x y2 y3 x x y3 x3 y
3 3 3
s  s  s 
+ + , (2)
y x 3 x2 y x x x2 y
where  (x, y, t) = (x, y, t) s (x, y) is the perturbation stream function and s (x, y) is the steady-state stream function. Linearisation is
a reasonable approximation for the behaviour of the uid, provided that the perturbation from the steady state is small such that terms
excluded due to linearisation (e.g., 3  /xy2  /y) would not be signicant compared to the terms in (2). Initially, the operating point
is a parabolic ow prole such that s /y(x, y) = 3(1 y2 )/2, s /x(x, y) = 0, with a uid velocity of zero at both top and bottom walls.
Integration yields that the initial steady-state stream function is s (x, y) = (3y y3 )/2.
Similarly, the mass transfer behaviour can be obtained by using the mass transport equation for incompressible uids (Setiawan et al.,
2015):
 2 2

cA 4 cA cA cA cA
= + + , (3)
t Pe x2 y2 y x x y
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 35

Fig. 2. Initial solute mass fraction prole between L1 and L2 .

where cA (x, y, t) is the solute mass fraction, Pe = U d/DAB is the Peclet number and DAB is the diffusivity of the solute in water (m2 s1 ). Eq.
(3) is linearised to obtain the following equation:
 
2 2
cA 4 cA cA s cA s cA cA,s  cA,s 
= + + + , (4)
t Pe x 2 y2 y x x y x y y x

where cA (x, y, t) is the perturbation solute mass fraction and cA,s (x, y) is the steady-state solute mass fraction such that cA (x, y, t) =
cA (x, y, t) cA,s (x, y). The initial steady-state solute mass fraction prole in the lower half of the channel is shown in Fig. 2.
The steady-state solute mass fraction prole is based on the dissolving wall boundary condition. This prole is represented mathemat-
ically as follows (Setiawan et al., 2015):
 3

1 (y+1) Pe
 ,
3 12(xL1 )
cA,s (x, y) = (h(x L1 ) h(x L)) 1 , (5)
 3

where L1 is the start position of the membrane, and (a) and (a, b) are the complete and incomplete gamma functions respectively,
dened as follows:


(a) = a1 e d, (6a)
0


(a, b) = a1 e d. (6b)
b

The Heaviside function h(x) is dened as follows:

1 x > 0,
h(x) = 0 x < 0, (7)

1/2 x = 0.

The boundary conditions of  (x, y, t) are homogeneous in y and ensure periodicity in x, and cA (x, y, t) has the following boundary
conditions to ensure a xed solute mass fraction at the bottom wall and no mass transfer at the top wall as follows (Setiawan et al., 2015):

(x, y = 1, t) = 0, (8a)

(x, y = 1, t) = 0, (8b)
 
(x = 0, y, t) = (x = L, y, t), (8c)

(x, y = 1, t) = w (x, t), (8d)
y

(x, y = 1, t) = 0, (8e)
y

(x, y = 1, t) = 0, (8f)
x

(x, y = 1, t) = 0, (8g)
x
 
(x = 0, y, t) = (x = L, y, t), (8h)
y y
 
(x = 0, y, t) = (x = L, y, t), (8i)
x x
cA (x, y = 1, t) = 0, (8j)
36 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

cA (x = 0, y, t) = cA (x = L, y, t), (8k)

cA
(x, y = 1, t) = 0. (8l)
y
As the actuation is applied as a velocity at the bottom wall, the following substitution is used:

(x, y, t) = (x, y, t) f (y)w (x, t), (9)

where f(y) = (y3 y2 y + 1)/4 to satisfy the same boundary conditions as  (x, y, t), and w (x, t) is the perturbation applied velocity at
the bottom wall. The linearised system in this study is therefore between the perturbation variables (x, y, t), w (x, t) and cA (x, y, t), and
dependent on the steady-state hydrodynamic and solute mass fraction proles s (x, y) and cA,s (x, y) respectively.

3.2. Conversion into innite-dimensional system of ODEs

For the analysis in this paper, all steady-state proles and perturbation variables need to be transformed into vectors that capture the
spatial variations in the system behaviour. The hydrodynamic system is spatially discretised in the y-direction based on the Chebyshev
collocation method. This converts (x, y, t), s (x, y, t) and f(y) into vectors as follows (Setiawan et al., 2015):
 T
a(x, t) = (x, y0 , t) (x, y1 , t) (x, yN , t) , (10a)
 T
s (x) = s (x, y0 ) s (x, y1 ) s (x, yN ) , (10b)
 T
fN = f (y0 ) f (y1 ) f (yN ) , (10c)

where yn = cos(2 n/N) for n = 0, 1, . . ., N. Differentiation in the y-direction (/y) is replaced by the matrix DN dened in (11), such that
q /yq is expressed as the qth power DqN of the differentiation matrix (Setiawan et al., 2015).

dn (1)n+l

, n=/ l,

dl yn yl



yn

 , 1 n = l N 1,
2
2 1 yn
DN (n, l) = dnl = (11)

2N 2 + 1

n = l = 0,


,


6


2N 2 + 1 , n = l = N,
6
where

2, n = 0 or N,
dn = (12)
1 otherwise.

The system is spectrally discretised in the x-direction using the FourierGalerkin method. This expresses variables that vary in x as the
sum of sine and cosine waves (or complex exponentials) of different wavenumbers as follows:



a(x, t) = aRm (t) + jaIm (t) ejm x , (13a)
m=




R I
s (x) = s,m + js,m ejm x , (13b)
m=




w (x, t) = R
wm I
(t) + jwm (t) ejm x , (13c)
m=

where m = 2 m/L and j = 1.
Similarly, the deviation solute mass fraction cA (x, y, t) and the steady-state solute mass fraction s (x, y) are spatially discretised in y
and spectrally decomposed as follows:
 T
b(x, t) = cA (x, y0 , t) cA (x, y1 , t) cA (x, yN , t) , (14a)
 T
s (x) = cA,s (x, y0 ) cA,s (x, y1 ) cA,s (x, yN ) , (14b)




b(x, t) = bRm (t) + jbIm (t) ejm x , (15a)
m=




s (x) = Rs,m + jIs,m ejm x . (15b)
m=
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 37

The discretisation in the y-direction yields the following relationships when applied to the boundary conditions:

aRm,0 (t) = 0, aIm,0 (t) = 0, (16a)

R I
aRm,1 (t) = lT1 am (t), aIm,1 (t) = lT1 am (t), (16b)
R I
aRm,N1 (t) = lT2 am (t), aIm,N1 (t) = lT2 am (t), (16c)

aRm,N (t) = 0, aIm,N (t) = 0, (16d)

R I
bRm,0 (t) = lT3 bm (t), bIm,0 (t) = lT3 bm (t), (16e)

bRm,N (t) = 0, bIm,N (t) = 0, (16f)

where
   1  
lT1 d01 d0(N1) d02 d0(N2)
= , (17a)
lT2 dN1 dN(N1) dN2 dN(N2)

1
 
lT3 = d00 d01 d0(N1) . (17b)

The states can therefore be reduced by using the following relationships:

R I
am (t) = Ra aRm (t), am (t) = Ra aIm (t), (18a)
R I
aRm (t) = Ia am (t), aIm (t) = Ia am (t), (18b)
R I
bm (t) = Rb bRm (t), bm (t) = Rb bIm (t), (18c)
R I
bRm (t) = Ib bm (t), bIm (t) = Ib bm (t), (18d)

where

01(N3)
T
l1
 

Ra = 0(N3)2 IN3 0(N3)2 , Ia = IN3 , (19a)

lT
2
01(N3)

lT3
 
Rb = 0(N1)1 IN1 0(N1)1 , Ib = IN1 . (19b)

01(N1)

This gives the following state equation, with state-space matrices dependent on the current steady-state hydrodynamic and solute mass
fraction proles s and s respectively:
X1 (t)
X1 (t) 0 0 0 I
B2
X2 (t) = A21 (s ) A22 (s ) 0 X2 (t) + w(t), (20)

X3 (t) A31 (s ) A32 (s ) A33 (s ) X3 (t) 0

where
 T
X1 (t) = w(t) = w0R (t) w1R (t) w1I (t) w2R (t) w2I (t) , (21a)
 T  T  T  T  T T
X2 (t) = R
a0 (t)
R
a1 (t)
I
a1 (t)
R
a2 (t)
I
a2 (t) , (21b)

          T
R T R T I T R T I T
X3 (t) = b0 (t) b1 (t) b1 (t) b2 (t) b2 (t) , (21c)

and
 T  T  T  T  T T
s = R
s,0 R
s,1 I
s,1 R
s,2 I
s,2 , (22a)

 T  T  T  T  T T
s = Rs,0 Rs,1 Is,1 Rs,2 Is,2 , (22b)

Therefore, X1 (t) is the state corresponding to the coefcients of the sine and cosine components of the perturbation in the slip velocity
prole, X2 (t) is the state corresponding to the coefcients of the sine and cosine components of the perturbation stream function at each
38 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

layer in the channel, and X3 (t) is the state corresponding to the coefcients of the sine and cosine components of the perturbation in the
solute mass fraction at each layer in the channel. The state-space matrices A and B in (20) are dependent on the steady-state prole s and
I I
s , and are dened in Appendix A. The terms w0I (t), a0 (t), and b0 (t) are not included in the states X1 (t), X2 (t) and X3 (t) since these terms
are static and do not contribute to the solution (can be set to zero).
The state equation given in (20) is an innite-dimensional representation of the hydrodynamic and mass transport phenomena and
their dependence on slip velocity inputs. The input to the state-space equation is written as the time derivative of the slip velocity
coefcients. This is typical of boundary control formulations with innite-dimensional systems and relates to the fact that the input has
both a static and dynamic effect on the hydrodynamic states (Curtain and Zwart, 1995). Previous approaches such as (Setiawan et al., 2015)
use simplications to remove terms from (2) and (4). This makes the calculation simpler, but makes the state-space model invalid for
steady-state proles other than that for which the model is specically derived. The novelty and value of this state-space representation
is that it can be applied to any steady-state hydrodynamic (s ) and solute (s ) prole in the uid domain.
For this analysis, the initial condition is required. For the system studied in this paper, the initial condition can be obtained based on
mathematical models and analytical solutions to the nonlinear PDEs. For more complex models or geometries, it may be necessary to solve
for the steady-state hydrodynamic and solute proles using computational uid dynamics, for example, so that this information can be
fed into this analysis.

3.3. Development of objective function to increase mass transfer

The aim of this study is to decrease steady-state concentration polarisation using a slip velocity input. Decrease in concentration
polarisation is linked to increased mass transfer coefcient (Liang et al., 2014). The mass transfer coefcient kmt (x, t) (m s1 ) is proportional
to the solute mass fraction gradient, and is typically dened as follows (Liang et al., 2014; Fimbres-Weihs and Wiley, 2010):

DAB cA
kmt (x, t) =  |wall , (23)
cA,wall cA,bulk y

where cA,wall is the real solute mass fraction at the membrane surface (kg solute kg water1 ), cA,bulk is the real bulk solute mass fraction
(kg solute kg water1 ), and cA /y|wall is the solute mass fraction gradient at the wall (kg solute kg water1 m1 ). The directions x and y are
parallel to the bulk ow and perpendicular to the wall respectively (m), and t is the time (s). Since the dissolving wall approximation sets
the wall solute mass fraction as constant and the bulk solute mass fraction does not change signicantly along the length of the channel, the
term cA,wall cA,bulk is approximately constant and uniform (Fimbres-Weihs and Wiley, 2010). Since DAB can be approximated as constant
and uniform in thin membrane channels, the mass transfer coefcient can be considered as directly proportional to the local solute mass
fraction gradient at the wall. Therefore, the mass transfer coefcient given in (23) can be replaced by the dimensionless mass transfer
coefcient kmt (x, t) dened as follows:

cA
kmt (x, t) = (x, y = 1, t). (24)
y
Therefore, the overall objective function to be maximised, Jo , can be expressed as the following:

1
 T  L2  2
1 L2 L1 0 L1
kmt,slip (x, t) kmt,ss (x) dxdt
Jo (w(x, t)) = lim
T T 1
TL , (25)
L 0 0
w2 (x, t)dxdt

where kmt,slip (x, t) is the mass transfer coefcient inuenced by the overall slip velocity w(x, t) and bulk ow conditions, and kmt,ss (x) is
the steady-state mass transfer coefcient due to bulk ow conditions alone such that kmt,slip (x, 0) = kmt,ss (x), and w(x, t) is not identically
zero. The slip velocity is applied along the whole channel to allow for mass transfer enhancements to occur near L1 and L2 , but is typically
small when signicantly far from the region of [L1 , L2 ] due to a lack of effect on the solute mass fraction. Since the bulk solute mass fraction
is lower than at wall, and due to the orientation of the membrane wall, the steady-state mass transfer coefcient is negative. Therefore,
the nal value of the overall mass transfer coefcient must be more negative than the initial value in order for there to be a decrease in
concentration polarisation, represented by the following constraint:

L2
L2
lim kmt,slip (x, T )dx < kmt,ss (x)dx, (26)
T
L1 L1

In this study, linearised systems are used to approximate the nonlinear governing equations. Therefore, an approximate objective
function J is required for the linearised system at each stage:
  L2   (x, t)
2
1 0 L1
kmt dxdt

J(w (x, t)) = lim
T T (L2 L1 )
  L2 2
. (27)
0 L1
(w (x, t)) dxdt

where kmt (x, t) is the perturbation mass transfer coefcient for the linearised system. Since the solute mass fraction in the bulk uid is

lower than at wall, the steady-state mass transfer coefcient is negative at steady state. Therefore, the steady state perturbation mass
transfer coefcient should also be negative, shown in the following constraint:

L2

lim kmt (x, T )dx < 0 (28)
T
L1
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 39

In order to satisfy the overall objective function (25), the output ys in this study is the area average of the mass transfer coefcient given
in (24):

L2
1
ys = lim kmt (x, T )dx. (29)
L2 L1 T L1

The dynamic output of the perturbation system y1 (t) is dened based on the perturbation mass transfer coefcient as follows:

L2
1 cA
y1 (t) = (x, y = 1, t)dx. (30)
L2 L1 L1 y
The outputs of the perturbation system and the overall system can be written in terms of the innite-dimensional system as shown below:
Y1 (t) = C13 X3 (t), (31a)
Ys = C13s s , (31b)
where Y1 (t) and Ys are scalar-valued approximations of y1 (t) and ys respectively based on the spatial discretisation and decomposition,
and C13 and C13s are dened in Appendix B.

3.4. Determination of perturbation slip velocity prole

In order to maximise (27) subject to (28), the most effective perturbation slip velocity prole must be determined. However, the innite-
dimensional system in (20) has the time-derivative of slip velocity as its input. This can be overcome by converting all time-dependent
variables into the Laplace domain to produce the following inputoutput relationship in the Laplace domain:

Y1 (s) = C13 (sI A33 )1 A31 + A32 (sI A22 )1 (A21 + sB2 ) w(s), (32)

where Y1 (s) and w(s) are the Laplace transforms of Y1 (t) and w(t) respectively. As this study focuses on driving the system to a new steady
state, only the nal value is of interest. By applying the nal value theorem, the zero-frequency response can be obtained:

lim Y1 (t) = limsY1 (s) = limsC 13 (sI A33 )1 A31 + A32 (sI A22 )1 (A21 + sB2 ) w(s),
t s0
 s0
(33)
= C13 A1
33
A31 + A32 A1 A
22 21
U.

where w(s) = U/s. This can be expressed in the time domain as:
w(t) = Uh(t), (34)
and h(t) is the heaviside function dened in (7). This input prole was chosen since it is bounded in time, and provides the maximum ratio
of input energy at s = 0 to energy for all s =
/ 0 for any input that is initially zero. This ensures minimal loss of energy at each transition
between steady states. The singular value decomposition of this multiple-input single-output (MISO) system at zero frequency can be
written as follows:
Y1 () = VT U. (35)
where  > 0 is the only singular value of the perturbation system and V is a real unit vector in the direction of input that most effectively
increases Y1 () such that:

VT U = C13 A1
33
A31 + A32 A1 A
22 21
U. (36)
Since the aim of this study is to make Ys more negative, the sign of U should be the opposite of the sign of V. Therefore, the input vector
U has the following form:
U = V, (37)
where is the magnitude of the input vector. In order to reduce problems caused by nonlinearities, this study uses several small changes
in the steady state to bring the system to a vastly different steady state. The input magnitude is determined such that the change in
steady state caused by the input is a predened fraction 0 < 1 of the initial steady state. In this study, this is based on the 2-norm as
follows:
 
s

s
=   2
(38)
ds

ds
2

where ds and ds are the changes in the steady state due to a unit input in the direction of V, dened as follows:
ds = CH1 X1 + CH2 X2 ,

= CH1 CH2 A1 A
22 21
V,
(39)
ds = Cc X3 ,

= Cc A1
33
A31 + A32 A1 A
22 21
V,
40 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

where CH1 , CH2 and Cc are dened in Appendix C. Alternative approaches to choosing the input magnitude at each step include xed input
magnitude, xed output magnitude, xed norm of change in steady state prole, or xed norm of change in matrix A between steady
states.

3.5. Calculation of new steady state for each perturbation

In order to approximate the nonlinear system behaviour over a large change in behaviour, this study applies several small perturbations
and changes the system dynamics as the steady state changes. The new steady state for each small perturbation can be determined by
adding the nal values ( ds and ds ) of the system response to the perturbation to the previous values of s and s respectively.
wnew
s = ws V,

snew = s + ds ,

= s CH1 CH2 A1 A
22 21
V, (40)

new
s = s + ds ,

= s Cc A1
33
A31 + A32 A1 A
22 21
V,

where ws is the overall slip velocity prole vector, which is initially zero and added to with each applied perturbation. The new mass
transfer coefcient can be calculated as follows:
Ysnew = Ys + Y1 (),
(41)
= Ys .

By updating the state-space in (20) with the new steady-state hydrodynamic and solute proles from (40), calculating (35)(41) and
repeating this process a sufcient number of times (as stated in Section 2.2), the system can be driven to a new steady state with a much
greater mass transfer away from the membrane wall due to the application of a specically designed slip velocity prole. This corresponds
to a decrease in concentration polarisation.

4. Results and discussion

The results of this study show how the solute mass fraction gradient is affected by the slip velocity at the wall. For practical purposes, the
wavenumbers studied are nite with maximum dynamic wavenumber M (and maximum steady-state wavenumber 2M). The dimensionless
parameters used in this study are shown in Table 1.
The Reynolds number chosen for this study is in the range used for reverse osmosis membrane channels (Liang et al., 2014). The Schmidt
number Sc = /DAB is chosen to approximate the behaviour of sodium chloride in water (Liang et al., 2014). In this study, the input direction
that decreases the average absolute solute mass fraction gradients at the membrane wall is determined and applied such the 2-norm of
the steady state has a relative change of = 0.001 in each step. The new steady state is determined and used in an updated model such that
a new input can be determined and added to the previous input. This process is repeated for 49 iterations (for a total of 50 steady states).
The solute mass fraction gradient proles with (nal) and without (initial) actuation and the log scale local (spatial) average magnitudes
are given in Fig. 3.
In the range of x of interest (50 < x < 150), the initial solute mass fraction gradient at the wall is always negative, representing a diffusive
driving force moving solutes away from the wall and towards the bulk ow. This driving force is greatest close to L1 (the start of the
membrane region) where the concentration polarisation region is not developed, and decreases along the length of the membrane region.
The nal solute mass fraction gradient is created by applying a nite set of slip velocity waves up to wavenumber M. A comparison of
panels (a) and (b) shows that the nal solute mass fraction gradient is more negative at almost every location along the membrane than
the initial case.
Panel (c) of Fig. 3 takes the average solute mass fraction gradient over 4 consecutive points along the membrane wall in order to show
the general trend. Panel (c) shows the main result of this paper: a large increase in the magnitude of mass transfer coefcient at the
membrane surface. Panel (c) shows that, for the initial case, the driving force near L2 is 1/100 of the driving force close to L1 . This shows
that very little diffusive mass transfer is occurring at the wall near L2 . The nal prole is greatest near the start of the membrane and
generally decreases along the length of the channel. However, the nal gradient is more than 100 the initial gradient near the start of
the membrane and over 20,000 of the initial gradient close to L2 . The relative difference in average driving force between L1 and L2 is
much smaller in the nal case than in the initial case. It is therefore clear that this approach has created a solute mass fraction gradient
that increases the area average mass transfer away from the membrane wall by orders of magnitude (3000). Greater performance can

Table 1
Dimensionless parameters used in study.

Parameter Symbol Values

Channel length L 200

Start position of membrane and reduction L1 50
End position of reduction L2 150
Maximum dynamic wavenumber M 160
Number of subdivisions in y-direction N 120
Reynolds number Re 280
Schmidt number Sc 600
Peclet number Pe 16,800
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 41

3
x 10
6

Initial solute mass

at membrane wall
4

fraction gradient
2
0
2
4
6
50 60 70 80 90 100 110 120 130 140 150
(a)
1.5
at membrane wall
Final solute mass

1
fraction gradient

0.5
0
0.5
1
1.5
50 60 70 80 90 100 110 120 130 140 150
(b)
0
10
mass fraction gradient
Local average solute

magnitude at wall

1
10
2
10
3
10
4
10
5
10
50 60 70 80 90 100 110 120 130 140 150
Position along channel x
(c)

Fig. 3. Initial (blue) (a) and nal (green) (b) solute mass fraction gradients at the membrane wall, and local average magnitude of solute mass fraction gradients at the
membrane wall (c). (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)

be achieved by using more iterations (or choosing a greater nal value at which to stop the algorithm or larger values). The slip velocity
prole that achieves this improvement is shown in Fig. 4.
The spatial prole in Fig. 4 is the cumulative slip velocity actuation prole obtained after 49 small input steps have been calculated.
Fig. 4 shows that the nal steady-state slip velocity is in the direction of bulk ow for the majority of the region of interest. This agrees
with work conducted by Liang et al. (2014) that shows that (uniform) slip velocities in the direction of bulk ow lead to mass transfer
enhancement and ux increase. Due to the truncation of wavenumbers, the input and response proles are spatially oscillatory. Analysis
with a greater range of wavenumbers is expected to result in a smoother, less spatially oscillatory actuation prole. The actuation is greatest
close to L1 and tends to decrease and become negative (moving upstream) near L2 as the initial solute mass fraction gradient becomes
small. This suggests that the change in the output is most strongly affected by actuation earlier in the channel. This is likely due to the
solute mass fraction gradient prole in the uid. Large solute mass fraction gradients (e.g., near L1 ) are present when regions of high solute
mass fraction are close to regions of low solute mass fraction. These conditions occur at the inlet of the membrane section, since the solute
mass fraction close to the membrane at the inlet is similar to that of the bulk uid as can be seen in Fig. 2. In order to achieve the same
change in regions with small solute mass fraction gradients, more energy (i.e., a greater slip velocity) is generally required. Figs. 5 and 6
show the perturbation uid velocity proles after implementation of the actuation shown in Fig. 4.

3
x 10
2
Slip velocity w(x)

2
50 60 70 80 90 100 110 120 130 140 150
Position along channel x

Fig. 4. Slip velocity prole at the membrane wall over the region of interest.
42 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

Fig. 5. Final perturbation velocity proles in streamwise directions. (For interpretation of the references to colour in the text, the reader is referred to the web version of this
article.)

Fig. 6. Final perturbation velocity proles in wall-normal direction. (For interpretation of the references to colour in the text, the reader is referred to the web version of this
article.)

The maximum perturbation uid velocity (2 103 ) required is much smaller than the maximum bulk ow velocity (1.5), but is
signicant close to the wall where the initial uid velocity is close to zero. Close to L1 , the slip velocity is negative (blue region in Fig. 5) and
over a small space in x becomes positive (red region in Fig. 5). This causes uid with low solute mass fraction to be dragged from the bulk
towards the membrane (blue wall-normal velocity in Fig. 6). This is much easier to accomplish close to L1 since the solute mass fraction
gradient is greater, and uid with low solute mass fraction is closer to the membrane at the start of the membrane compared to further
along the channel where the solute mass fraction prole is more developed. Some of this low solute uid is sent upstream and ejected
towards the bulk uid as shown in the red region in Fig. 6 which is not directly valuable but is necessary to conserve continuity. The value
is that much of the remainder of the uid dragged from the bulk goes toward an increase in streamwise velocity close to the membrane
surface. The streamwise perturbation velocity is small but positive between L1 and L2 , causing uid with low solute mass fraction to
move downstream, replacing uid of high solute mass fraction and thereby increasing mass transfer at the membrane and decreasing
concentration polarisation. The value of the prole determined in this study is that the prole shown in Fig. 4 is a result of a sequence of
local optimisations and is therefore more effective than the proles used in the study by Liang et al. (2014) For sufciently small and
sufciently large M and N, the input prole obtained by this approach should approximate the most effective steady state input prole
for increasing diffusive mass transfer based on the nonlinear 2D NavierStokes and mass transport equations. The nature of the velocity
proles shown in Figs. 5 and 6 illustrates how the actuation prole obtained by this approach (as seen in Fig. 4) creates conditions that are
favourable to increasing mass transfer away from the wall and therefore reduces concentration polarisation.
The nonlinearity in this system can be seen in the singular value  at each iteration step as shown in Fig. 7. At each step, a new linearised
steady state is evaluated. In this study, the singular value  at the nal step is 83% of the singular value at the rst step. The step is
sufciently small since the relative difference in  between steps is roughly 0.38%. Using a single linear approximation over this range
would signicantly overestimate the effectiveness of a slip velocity input since the system becomes less sensitive to inputs as the system
moves further from the initial steady state in this direction. This approach allows for a more accurate determination of the effect of the
input on the system behaviour at steady state. The signicant change in system response between the initial condition and the nal stages
seen in Fig. 7 provides justication of the value of a stage-wise approach over a single linearised model.
This study was conducted by calculating the steady state and evaluating the input direction and magnitude. Implementation of the
slip velocity proles calculated in this study could be conducted by applying each input and allowing the system to reach steady state
before applying the next step. Approximating the system as a rst-order system with a time constant  equal to the largest time constant
of roughly 4 min (based on a reverse osmosis channel with hydraulic diameter d = 0.002 m, and Reynolds number 280), each stage takes
approximately 20 min to reach 5 when the system is approximately at steady state. For roughly 50 stages, this takes less than 17 h to
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 43

250

200

Singular value
150

100

50

0
0 5 10 15 20 25 30 35 40 45 50
Iteration

Fig. 7. Singular value at each iteration.

reach the nal steady state. Since reverse osmosis membrane processes can operate for 23 months between washing (Jin et al., 2013), this
is a reasonable time frame for the improvement that can be obtained. An advantage of this study is that, since no online measurement is
required, the slip velocity proles can be calculated off-line and applied to experimental or industrial systems without requiring extensive
real-time computation and calculation of steady-states.
As mentioned earlier in the paper, the approach of small perturbations to achieve a large change in performance is designed to minimise
the errors due to linearisation of nonlinear equations. Additionally, when updating the state-space to the new steady state, errors are
introduced into the state-space equation. This is partially managed by choosing smaller magnitudes for the rst few perturbations, and
increasing the magnitude of the later perturbations once a signicant improvement has been made. However, these errors are unavoidable
without solving the nonlinear equations directly, and the approach proposed in this paper is expected to provide a better performance
than applying a large input based on the initial steady state.
In practice, however, the desired number of iterations may be signicantly less than used in this study (approximately 5 iterations in
total). While using less iterations with larger values of provide larger errors due to linearisation, it is likely that the performance using a
few large inputs will only be slightly poorer than using many small inputs. This is because the shape of the nal actuation prole provides
hydrodynamic effects that increase mass transfer at the wall by bringing low concentration uid towards the wall and increasing the uid
ow near the wall as shown in Figs. 5 and 6, and the prole shape for the rst iteration in this study, has similar features. Therefore, the
number and size of the steps should not greatly affect whether an improvement is made, but will have some effect on the extent of the
improvement and the accuracy of the estimation of the improvement. However, due to the potential for errors at each iteration or stage,
it may be valuable to introduce a robustness parameter when determining the direction of the next perturbation slip velocity, such that a
suboptimal slip velocity prole is chosen, but the potential for errors in the new state-space equation is reduced. This could require more
input energy to achieve equivalent performance, but could be more reliable for estimation of the system behaviour or subsequent steady
states.
This approach only addresses using step inputs since these inputs apply more energy at the zero-frequency than any other input, and
the only input component of value for a linear system when changing steady state is the zero-frequency component. However, nonlinear
systems may change their zero-frequency response due to input signals with sustained oscillatory components. Therefore, it is not yet
known whether additional, temporally-oscillatory actuation proles can further enhance mass transfer in this system given the same
time-average input energy, particularly in the absence of online measurement.

5. Conclusions

In this paper, a method has been developed for changing steady states in nonlinear distributed parameter systems using spatially-
varying boundary actuation in the absence of distributed online measurement, demonstrated by decreasing concentration polarisation
in reverse osmosis membrane systems using a steady-state slip velocity input prole. This is achieved by successive linearisations of the
NavierStokes and mass transport PDEs, and conversion into an innite-dimensional system that can be used for arbitrary steady-state
hydrodynamic and mass transport proles. By using small input magnitudes and determining the new steady-state, this approach can
account for much of the nonlinearities associated with the governing PDEs.
The actuation prole determined by this approach draws uid with low solute mass fraction from the bulk at the start of the membrane
section of the channel and uses this uid to increase the downstream uid velocity close to the membrane surface. This in turn increases
average mass transfer away from the wall using approximately the least input energy possible using non-oscillatory inputs. This increase
the average mass transfer at the wall is expected to signicantly increase ux and decrease concentration polarisation in membrane
systems. Nonlinearities in the system and the calculated stage-wise control action can be observed in the relative change in the singular
value of the system response at each stage, with later stages showing a 17% decrease in singular value compared to the initial condition.
At each subsequent stage, there is a clear decrease in sensitivity of the system to further increases in mass transfer by manipulating wall-
tangential velocity. Analysis based on a single linearised state-space model would have overestimated the effectiveness of the actuation,
and hence performed more poorly than expected. The stage-wise approach allows for small changes in the system to be studied (changes
in singular value of approximately 0.38% between each stage) and reducing linearisation-based errors. This approach therefore allows for
the steady-state control of large nonlinear systems or nonlinear PDEs without specialised nonlinear techniques.
This work can be further extended to permeable wall systems to better understand how this approach can affect concentration pola-
risation. Furthermore, wavelets may be used as the basis functions to eliminate the need for periodicity in the geometry or better handle
the discontinuity in the wall solute concentration. Since the innite-dimensional hydrodynamic and mass transport state-space equations
44 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

developed in this paper can accommodate for any steady-state prole, this approach can be applied in a wide range of scenarios such
as 3D channels, other geometries, wall-normal actuation, decreasing inhomogeneity of solute concentrations or increasing yields in ow
reactors, etc.

Acknowledgements

This work is partially supported by the Australian Research Council Discovery Projects DP110101643. The rst author also acknowledges
the support of an Australian Postgraduate Award scholarship.

Appendix A. State space model

The state-space matrices are dened as follows:

0 0 0
A = A21 (s ) A22 (s ) 0 , (A.1a)
A31 (s ) A32 (s ) A33 (s )

I
B = B2 , (A.1b)
0

where each submatrix has several components for example:

A21,(0,0) (s ) A21,(0,1) (s ) A21,(0,k) (s )
A

21,(1,0) (s ) A21,(1,1) (s ) A21,(1,k) (s )

. .. .. ..
A21 (s ) = .
. ..
. . . . . (A.2a)
A
21,(m,0) (s ) A21,(m,1) (s ) A21,(m,k) (s )

.. .. .. .. ..
. . . . .

The components of each submatrix are as follows:

A21,(0,0) (s ) = A2,(0,0) (s )fN , (A.3a)

A22,(0,0) (s ) = A2,(0,0) (s )Ia , (A.3b)

 
A21,(0,k) (s ) = AR2,(0,k) (s )fN AI2,(0,k) (s )fN , (A.3c)
 
A22,(0,k) (s ) = AR2,(0,k) (s )Ia AI2,(0,k) (s )Ia , (A.3d)
 
AR2,(m,0) (s )fN
A21,(m,0) (s ) = , (A.3e)
AI2,(m,0) (s )fN
 
AR2,(m,0) (s )Ia
A22,(m,0) (s ) = , (A.3f)
AI2,(m,0) (s )Ia
 
ARR
2,(m,k)
(s )fN ARI
2,(m,k)
(s )fN
A21,(m,k) (s ) = , (A.3g)
AIR
2,(m,k)
(s )fN AII2,(m,k) (s )fN
 
ARR
2,(m,k)
(s )Ia ARI
2,(m,k)
(s )Ia
A22,(m,k) (s ) = , (A.3h)
AIR
2,(m,k)
(s )Ia AII2,(m,k) (s )Ia

4 1  2 2
A2,(0,0) (s ) = M0 Ra DN 20 I , (A.4a)
Re
    
AR2,(0,k) (s ) = 2M01 Ra diag D3N 2k DN s,k
I I
k diag k s,k D3N 2k DN diag 3k I k D2N s,k
I
DN
I

+ diag DN s,k 3k I k D2N , (A.4b)

    
AI2,(0,k) (s ) = 2M01 Ra diag D3N 2k DN s,k
R R
k + diag k s,k D3N 2k DN + diag 3k I k D2N s,k
R
DN
R

diag DN s,k 3k I k D2N , (A.4c)
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 45

1
   
AR2,(m,0) (s ) = 2Mm I
Ra diag m s,m D3N 20 DN diag 3m I m D2N s,m
I
DN , (A.4d)
1
   
AI2,(m,0) (s ) = 2Mm Ra diag R
m s,m D3N 20 DN + diag 3m I m D2N R
s,m DN , (A.4e)

1
   
ARR
2,(m,k)
(s ) = Mm Ra diag D3N 2mk DN s,mk
I I
k diag mk s,mk D3N 2k DN
  
diag 3mk I mk D2N s,mk
I I
DN diag DN s,mk 3k I k D2N
   
+ diag D3N 2m+k DN s,m+k
I I
k diag m+k s,m+k D3N 2k DN diag 3m+k I m+k D2N s,m+k
I
DN
  4  2 2 
I
+ diag DN s,m+k 3k I k D2N + mk D 2m I , (A.4f)
Re N

1
   
ARI
2,(m,k)
(s ) = Mm Ra diag D3N 2mk DN s,mk
R R
k diag mk s,mk D3N 2k DN
   
diag 3mk I mk D2N s,mk
R R
DN diag DN s,mk 3k I k D2N diag D3N 2m+k DN s,m+k
R
k
 R
   
+ diag m+k s,m+k D3N 2k DN + diag 3m+k I m+k D2N s,m+k
R R
DN diag DN s,m+k 3k I k D2N ,

(A.4g)

1
    
AIR
2,(m,k)
(s ) = Mm Ra diag D3N 2mk DN s,mk
R R
k + diag mk s,mk D3N 2k DN + diag 3mk I mk D2N s,mk
R
DN

R
   
+ diag DN s,mk 3k I k D2N diag D3N 2m+k DN s,m+k
R R
k + diag m+k s,m+k D3N 2k DN
  
+ diag 3m+k I m+k D2N s,m+k
R R
DN diag DN s,m+k 3k I k D2N , (A.4h)

1
   
AII2,(m,k) (s ) = Mm Ra diag D3N 2mk DN s,mk
I I
k diag mk s,mk D3N 2k DN
   
diag 3mk I mk D2N s,mk
I I
DN diag DN s,mk 3k I k D2N diag D3N 2m+k DN s,m+k
I
k
 I
 
+ diag m+k s,m+k D3N 2k DN + diag 3m+k I m+k D2N s,m+k
I
DN
  4  2 2 
I
diag DN s,m+k 3k I k D2N + mk D 2m I . (A.4i)
Re N

A31,(0,0) (s ) = A3,(0,0) (s )fN , (A.5a)

A32,(0,0) (s ) = A3,(0,0) (s )Ia , (A.5b)

 
A31,(0,k) (s ) = AR3,(0,k) (s )fN AI3,(0,k) (s )fN , (A.5c)
 
A32,(0,k) (s ) = AR3,(0,k) (s )Ia AI3,(0,k) (s )Ia , (A.5d)
 
AR3,(m,0) (s )fN
A31,(m,0) (s ) = , (A.5e)
AI3,(m,0) (s )fN
 
AR3,(m,0) (s )Ia
A32,(m,0) (s ) = , (A.5f)
AI3,(m,0) (s )Ia
 
ARR
3,(m,k)
(s )fN ARI
3,(m,k)
(s )fN
A31,(m,k) (s ) = , (A.5g)
AIR
3,(m,k)
(s )fN AII3,(m,k) (s )fN
 
ARR
3,(m,k)
(s )Ia ARI
3,(m,k)
(s )Ia
A32,(m,k) (s ) = , (A.5h)
AIR
3,(m,k)
(s )Ia AII3,(m,k) (s )Ia

A3,(0,0) (s ) = 0, (A.6a)
  
AR3,(0,k) (s ) = 2Rb diag DN Is,k k + diag k Is,k DN , (A.6b)
  
AI3,(0,k) (s ) = 2Rb diag DN Rs,k k diag k Rs,k DN , (A.6c)

AR3,(m,0) (s ) = Rb diag m Is,m DN , (A.6d)

AI3,(m,0) (s ) = Rb diag m Rs,m DN , (A.6e)
46 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

    
ARR
3,(m,k)
(s ) = Rb diag DN Is,mk k + diag mk Is,mk DN + diag DN Is,m+k k + diag m+k Is,m+k DN , (A.6f)

    
ARI
3,(m,k)
(s ) = Rb diag DN Rs,mk k + diag mk Rs,mk DN diag DN Rs,m+k k diag m+k Rs,m+k DN , (A.6g)

    
AIR
3,(m,k)
(s ) = Rb diag DN Rs,mk k diag mk Rs,mk DN diag DN Rs,m+k k diag m+k Rs,m+k DN , (A.6h)

    
AII3,(m,k) (s ) = Rb diag DN Is,mk k + diag mk Is,mk DN diag DN Is,m+k k diag m+k Is,m+k DN . (A.6i)

A33,(0,0) (s ) = A4,(0,0) (s )Ib , (A.7a)

 
A33,(0,k) (s ) = AR4,(0,k) (s )Ib AI4,(0,k) (s )Ib , (A.7b)

 
AR4,(m,0) (s )Ib
A33,(m,0) (s ) = , (A.7c)
AI4,(m,0) (s )Ib

 
ARR
4,(m,k)
(s )Ib ARI
4,(m,k)
(s )Ib
A33,(m,k) (s ) = , (A.7d)
AIR
4,(m,k)
(s )Ib AII4,(m,k) (s )Ib

4 
A4,(0,0) (s ) = R D2N 20 , (A.8a)
Pe b
  
AR4,(0,k) (s ) = 2Rb diag DN s,k
I I
k diag k s,k DN , (A.8b)

  
AI4,(0,k) (s ) = 2Rb diag DN s,k
R R
k + diag k s,k DN , (A.8c)


AR4,(m,0) (s ) = Rb diag m s,m
I
DN , (A.8d)


AI4,(m,0) (s ) = Rb diag m s,m
R
DN , (A.8e)

   
ARR
4,(m,k)
I
(s ) = Rb diag mk s,mk I
DN + diag DN s,mk I
k diag m+k s,m+k DN
 4 2 2 
I
diag DN s,m+k k + mk D 2m I , (A.8f)
Pe N

    
ARI
4,(m,k)
R
(s ) = Rb diag mk s,mk R
DN + diag DN s,mk R
k + diag m+k s,m+k R
DN + diag DN s,m+k k , (A.8g)

    
AIR
4,(m,k)
R
(s ) = Rb diag mk s,mk R
DN diag DN s,mk R
k + diag m+k s,m+k R
DN + diag DN s,m+k k , (A.8h)

   
AII4,(m,k) (s ) = Rb diag mk s,mk
I I
DN + diag DN s,mk I
k + diag m+k s,m+k DN
 4 2 2 
I
+ diag DN s,m+k k + mk D 2m I , (A.8i)
Pe N

where

1 m = k,
mk = (A.9)
0 otherwise.

Submatrix B2 is a block diagonal matrix, dened as follows:

B2 = diag (Ra fN , Ra fN , Ra fN , . . .) (A.10)

 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 47

Appendix B. Formulation of output

The output in this study is the solute mass fraction gradient at the membrane wall between x = L1 and x = L2 . By converting this system
using the Fourier-Galerkin method, this output can be written in terms of the states given in (21).

L2
1 cA
y1 (t) = (x, y = 1, t)dx,
L2 L1 L1 y
 
  R
 2 (sin (m L2 ) sin (m L1 )) R
Y1 (t) = 0 0 1 DN Ib b0 (t) + 0 0 DN Ib bm (t)
m (L2 L1 )
 m=1
2 (cos (m L2 ) cos (m L1 ))
 I
+ 0 0 DN Ib bm (t),
m (L2 L1 )
  T T
0 0 1 DN Ib

  T
2 (sin (1 L2 ) sin (1 L1 ))
0 0 DN Ib
1 (L2 L1 ) (B.1)

  
2 (cos (1 L2 ) cos (1 L1 )) T
0 0 DN Ib
1 (L2 L1 )
Y1 (t) =
  T X3 (t),
2 (sin (2 L2 ) sin (2 L1 ))
0 0 DN Ib
2 (L2 L1 )

   T
0 2 (cos (2 L2 ) cos (2 L1 ))
0
2 (L2 L1 )
DN Ib


..
.
Y1 (t) = C13 X3 (t).

This gives the change in solute mass fraction gradient at each step. The total solute mass fraction gradient is calculated similarly as follows:

L2
1 cA
ys (t) = (x, y = 1, t)dx,
L2 L1 L1 y

 
  2 (sin (m L2 ) sin (m L1 ))
Ys = 0 0 1 DN Rs,0 + 0 0 DN Rs,m
m (L2 L1 )
 m=1
2 (cos (m L2 ) cos (m L1 ))

+ 0 0 DN Is,m ,
m (L2 L1 )
  T T
0 0 1 DN

 T

2 (sin (1 L2 ) sin (1 L1 ))
0 0 DN
1 (L2 L1 ) (B.2)

  
2 (cos (1 L2 ) cos (1 L1 )) T
0 0 DN
1 (L2 L1 )
Ys =
  T s ,
2 (sin (2 L2 ) sin (2 L1 ))
0 0 DN
2 (L2 L1 )

   T
0 2 (cos (2 L2 ) cos (2 L1 ))
0
2 (L2 L1 )
DN


..
.
Ys = C13s s .
48 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149

Appendix C. Calculation of new steady states

In order to update the steady state, the matrices that relate the change in steady-state conditions to the states are required. These
matrices, as shown in (40) are given in detail below:
f 0 0

N

..
0 fN 0 .

CH1 = , (C.1a)
..
0 0 fN .

.. .. .. ..
. . . .
I 0 0

a

..
0 Ia 0 .

CH2 = , (C.1b)
..
0 0 Ia .

.. .. .. ..
. . . .
I 0 0

b

..
0 Ib 0 .

Cc = . (C.1c)
..
0 0 Ib .

.. .. .. ..
. . . .

References

Aamo, O.M., Krstic, M., 2003. Flow Control by Feedback: Stabilization and Mixing. Springer.
Aamo, O.M., Krstic, M., Bewley, T.R., 2003. Control of mixing by boundary feedback in 2D channel ow. Automatica 39 (9), 15971606.
Abhari, F., Jaafar, H., Yunus, N.A.M., 2012. A comprehensive study of micropumps technologies. Int. J. Electrochem. Sci. 7 (10), 97659780.
Adomaitis, R.A., 2003. A reduced-basis discretization method for chemical vapor deposition reactor simulation. Math. Comput. Model. 38, 159175.
Alexiadis, A., Bao, J., Fletcher, D.F., Wiley, D.E., Clements, D.J., 2005. Analysis of the dynamic response of a reverse osmosis membrane to time-dependent transmembrane
pressure variation. Ind. Eng. Chem. Res. 44 (20), 78237834.
Armaou, A., Christodes, P.D., 2000. Wave suppression by nonlinear nite-dimensional control. Chem. Eng. Sci. 55 (14), 26272640.
Baker, J., Armaou, A., Christodes, P.D., 2000. Nonlinear control of incompressible uid ow: application to Burgers equation and 2D channel ow. J. Math. Anal. Appl. 252
(1), 230255.
Balogh, A., Aamo, O.M., Krstic, M., 2005. Optimal mixing enhancement in 3-D pipe ow. IEEE Trans. Control Syst. Technol. 13 (1), 2741.
Bamieh, B., Mezic, I., Fardad, M., 2001. A framework for destabilization of dynamical systems and mixing enhancement. In: Proceedings of the 40th IEEE Conference on
Decisions and Control, vol. 5, pp. 49804983.
Callier, F.M., Winkin, J., 1992. LQ-optimal control of innite-dimensional systems by spectral factorization. Automatica 28 (4), 757770.
Chen, J.C., Li, Q., Elimelech, M., 2004. In situ monitoring techniques for concentration polarization and fouling phenomena in membrane ltration. Adv. Colloid Interface Sci.
107 (23), 83108.
Christodes, P.D., Armaou, A., 2000. Global stabilization of the KuramotoSivashinsky equation via distributed output feedback control. Syst. Control Lett. 39 (4), 283294.
Curtain, R.F., Zwart, H., 1995. An Introduction to Innite-Dimensional Linear Systems Theory. Springer-Verlag, Inc., New York, NY, USA.
DAlessandro, D., Dahleh, M., Mezic, I., 1999. Control of mixing in uid ow: a maximum entropy approach. IEEE Trans. Autom. Control 44 (10), 18521863.
Du, Y., Karniadakis, G.E., 2000. Suppressing wall turbulence by means of a transverse traveling wave. Science 288, 12301234.
Fimbres-Weihs, G.A., Wiley, D.E., 2010. Review of 3D CFD modeling of ow and mass transfer in narrow spacer-lled channels in membrane modules. Chem. Eng. Process.:
Process Intensif. 49 (7), 759781.
Jin, X., Li, E., Lu, S., Qiu, Z., Sui, Q., 2013. Coking wastewater treatment for industrial reuse purpose: combining biological processes with ultraltration, nanoltration and
reverse osmosis. J. Environ. Sci. 25 (8), 15651574.
Jovanovic, M.R., 2006. Turbulence suppression in channel ows by small amplitude transverse wall oscillations. In: Proceedings of the 2006 American Control Conference,
pp. 11611166.
Liang, Y.Y., Chapman, M.B., Fimbres-Weihs, G.A., Wiley, D.E., 2014. CFD modelling of electro-osmotic permeate ux enhancement on the feed side of a membrane module. J.
Membr. Sci. 470, 378388.
Liu, G., Xu, J., Yang, Y., Zhang, W., 2010. Active control of ow and heat transfer in silicon microchannels. J. Micromech. Microeng. 20 (045006), 116.
Moghadam, A.A., Aksikas, I., Dubljevic, S., Forbes, J.F., 2012. Innite-dimensional LQ optimal control of a dimethyl ether (DME) catalytic distillation column. J. Process
Control 22 (9), 16551669.
Moghadam, A.A., Aksikas, I., Dubljevic, S., Forbes, J.F., 2013. Boundary optimal (LQ) control of coupled hyperbolic PDEs and ODEs. Automatica 49 (2), 526533.
Moghadam, A.A., Aksikas, I., Dubljevic, S., Forbes, J.F., 2014. LQ (optimal) control of hyperbolic PDAEs. Int. J. Control 87 (10), 21562166.
Ouyang, H., Bao, J., Fimbres-Weihs, G.A., Wiley, D.E., 2013. Control study on mixing enhancement in boundary layers of membrane systems. J. Process Control 23 (8),
11971204.
Ratnayake, P., Setiawan, R., Bao, J., 2015. Analysis of ow control by boundary-layer manipulation using 2D frequency response. Asia-Pac. J. Chem. Eng. 10 (4), 512525.
Setiawan, R., Ratnayake, P., Bao, J., Fimbres-Weihs, G.A., Wiley, D.E., 2015. Reduced-order model for the analysis of mass transfer enhancement in membrane channel using
electro-osmosis. Chem. Eng. Sci. 122, 8696.
Shari, M., Young, B.R., 2011. 3-Dimensional spatial monitoring of tanks for the milk processing industry using electrical resistance tomography. J. Food Eng. 105 (2),
312319.
Song, L., Hu, J.Y., Ong, S.L., Ng, W.J., Elimelech, M., Wilf, M., 2003. Emergence of thermodynamic restriction and its implications for full-scale reverse osmosis processes.
Desalination 155 (3), 213228.
Yao, H.-M., Tad, M.O., Tian, Y.-C., 2010. Accelerated computation of cyclic steady state for simulated-moving-bed processes. Chem. Eng. Sci. 65 (5), 16941704.
Yao, Z., Hu, Y., El-Farra, N.H., 2013. Networked model predictive control of spatially distributed processes. In: Proceedings of the 2013 American Control Conference, pp.
20622067.
 P. Ratnayake, J. Bao / Computers and Chemical Engineering 98 (2017) 3149 49

Zhu, A., Christodes, P.D., Cohen, Y., 2009a. Effect of thermodynamic restriction on energy cost optimization of RO membrane water desalination. Ind. Eng. Chem. Res. 48
(13), 60106021.
Zhu, A., Christodes, P.D., Cohen, Y., 2009b. Energy consumption optimization of reverse osmosis membrane water desalination subject to feed salinity uctuation. Ind. Eng.
Chem. Res. 48 (21), 95819589.
Zhu, A., Rahardianto, A., Christodes, P.D., Cohen, Y., 2010. Reverse osmosis desalination with high permeability membranes cost optimization and research needs.
Desalin. Water Treat. 15 (13), 256266.