Journal of Chromatography A, 1305 (2013) 114–122

Contents lists available at SciVerse ScienceDirect

Journal of Chromatography A
journal homepage: www.elsevier.com/locate/chroma

Computational fluid dynamic simulation of axial and radial flow

membrane chromatography: Mechanisms of non-ideality and
validation of the zonal rate model
Pranay Ghosh a , Kaveh Vahedipour b , Min Lin c , Jens H. Vogel c,1 ,
Charles Haynes d , Eric von Lieres a,∗
a
IBG-1: Biotechnology, Forschungszentrum Jülich, Germany
b
INM-4: Medical Imaging Physics, Forschungszentrum Jülich, Germany
c
Isolation and Purification Department, Global Biologics Development, Bayer Healthcare, Berkeley, USA
d
Michael Smith Laboratories, University of British Columbia, Vancouver, Canada

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

Article history: Membrane chromatography (MC) is increasingly being used as a purification platform for large
Received 1 April 2013 biomolecules due to higher operational flow rates. The zonal rate model (ZRM) has previously been
Received in revised form 28 June 2013 applied to accurately characterize the hydrodynamic behavior in commercial MC capsules at different
Accepted 1 July 2013
configurations and scales. Explorations of capsule size, geometry and operating conditions using the
Available online 4 July 2013
model and experiment were used to identify possible causes of inhomogeneous flow and their contrib-
utions to band broadening. In the present study, the hydrodynamics within membrane chromatography
Keywords:
capsules are more rigorously investigated by computational fluid dynamics (CFD). The CFD models are
Membrane chromatography
Computational fluid dynamics
defined according to precisely measured capsule geometries in order to avoid the estimation of geometry
Flow distribution related model parameters. In addition to validating the assumptions and hypotheses regarding non-ideal
Zonal rate model flow mechanisms encoded in the ZRM, we show that CFD simulations can be used to mechanistically
Model-based scale-up understand and predict non-binding breakthrough curves without need for estimation of any param-
eters. When applied to a small-scale axial flow MC capsules, CFD simulations identify non-ideal flows
in the distribution (hold-up) volumes upstream and downstream of the membrane stack as the major
source of band broadening. For the large-scale radial flow capsule, the CFD model quantitatively predicts
breakthrough data using binding parameters independently determined using the small-scale axial flow
capsule, identifying structural irregularities within the membrane pleats as an important source of band
broadening. The modeling and parameter determination scheme described here therefore facilitates a
holistic mechanistic-based method for model based scale-up, obviating the need of performing expensive
large-scale experiments under binding conditions. As the CFD model described provides a rich mecha-
nistic analysis of membrane chromatography systems and the ability to explore operational space, but
requires detailed knowledge of internal capsule geometries and has much greater computational require-
ments, it is complementary to the previously described strengths and uses of the ZRM for process analysis
and design.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction predominantly convective and permit higher flow rates to be

realized at lower pressure drops [1–3]. Recent improvements in
Membrane chromatography (MC) is widely used as a purifi- membrane surface chemistries have led to higher binding capaci-
cation platform for virus clearance and polishing. Larger ties and, consequently, MC is also finding traction in industry as an
pore sizes in membranes (1–1.2 ␮m) make the mass transfer alternative platform to conventional packed bed chromatography
for the purification of complex biomolecules such as, e.g. glyco-
proteins [4]. In a recent study, Bayer Healthcare has demonstrated a
first commercial-scale application of MC in bind and elute mode for
∗ Corresponding author at: IBG-1: Biotechnology, Forschungszentrum Jülich,
blood coagulation factors and has reported a yield improvement of
Wilhelm-Johnen-Straße 1, 52425 Jülich, Germany. Tel.: +49 2461 61 2168;
40% while maintaining high product quality as compared to packed
fax: +49 2461 61 3870.
E-mail address: e.von.lieres@fz-juelich.de (E. von Lieres).
bed chromatography [4]. As MC is becoming increasingly accepted
1
Current address: Boehringer Ingelheim, Fremont, USA. in the biopharmaceutical industry, accurate modeling strategies

0021-9673/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.chroma.2013.07.004
 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122 115

Fig. 1. Representations of (a) the classical one-dimensional Roper and Lightfoot model, and (b) a two zone ZRM configuration for axial flow MC capsules.

have become important for rational process analysis, simulation

and design.
MC has traditionally been modeled by only considering the
axial coordinate [5–8]. In most of these classic studies, the effects
of external hold-up volumes on elution band broadening are
accounted for by coupling a plug flow model (PFR) for transport
within the membrane stack with one or two continuously stirred
tank regions (CSTR) in series to account for mixing and residence
times within the extra-column spaces (see Fig. 1a). These one-
dimensional models assume flow homogeneity over the entire
membrane cross-section, which in practice is hard to achieve due to
device design constraints. Lab scale MC capsules often have axial
flow configurations, employing a stack of flat membrane sheets.
Preparative or pilot-scale capsules usually have radial flow configu- Fig. 2. Longitudinal cross-section of the Pall Mustang XT5 capsule, showing free and
rations with various winding schemes such as spiral wound, hollow porous regions as implemented in the CFD model.
fibers, or pleats. Manifold design is long known to significantly
column loading, CFD simulations provide precise hydrodynamics
impact on the breakthrough performance of MC capsules [9,10]. In a
and thereby allow for unambiguous analysis of the ability of differ-
recent report, we showed that the assumption of flow homogeneity
ent binding models to reproduce measured breakthrough curves.
over the entire membrane cross-sections does not necessarily hold
CFD may therefore be used as a powerful tool for gaining a richer
for commercially available capsules, and that the resulting inho-
and more accurate understanding of mechanisms contributing to
mogeneous mass flows can cause unwanted tailing of breakthrough
band broadening and their relative importance in different MC
curves [11]. The zonal rate model (ZRM), which treats this problem,
devices.
uniquely de-couples the effects of hydrodynamics and binding on
the resulting chromatograms by partitioning the entrance and elu-
2. Theory
tion hold-up volumes, as well as the membrane stack, into virtual
zones that are modeled as a network of inter-connected CSTRs and 2.1. Computational fluid dynamics
PFRs [12,13]. The ZRM is thereby able to quantify non-ideal flows,
as well as binding non-idealities, and their contributions to band- CFD models are based on solution of the fundamental con-
broadening, and illustrations of these capabilities were provided servation laws of mass, momentum and (sometimes) energy.
through application to the Pall Mustang XT5 (axial flow) and Mus- Mathematically, these laws are described by a set of partial dif-
tang XT140 (radial flow) capsules. Both capsules exhibit admirable ferential equations (PDEs) that have analytical solutions only in a
performance attributes, but nevertheless exhibit non-idealities at few cases, usually described by rather simple boundary conditions.
standard operating conditions. ZRM analysis of the axial-flow XT5 Hence, numerical methods are required in most applications. In
capsule suggested that non-ideal flow in the external hold-up vol- CFD the collective set of PDEs and initial and boundary conditions
umes upstream and downstream of the membrane stack, which (the model) is solved to compute fluid flow and related phenomena,
is consistent with the high aspect ratio of the XT5 capsule, con- such as transport and adsorption of solute molecules.
tribute to band broadening. In contrast, the ZRM suggested that The system boundaries are defined by the physical geometries
structural irregularities in the membrane pleats are the primary of the system, in this case the MC capsules. For example, Fig. 2
source of flow non-idealities in the radial flow XT140 capsule. For shows the internal geometry of a Pall Mustang XT5 capsule. In this
both studies, the ZRM was able to quantitatively reproduce break- capsule, spacer meshes are placed at either sides of the membrane
through data, a result that cannot be achieved using traditional stack. The geometry shown is simplified by omitting these spacer
membrane chromatography models. Moreover, the existence of meshes and by considering the membrane stack as one homoge-
varying linear velocities in the radial-flow capsule, as suggested neous region. The model therefore considers the internal geometry
by the ZRM, was consistent with magnetic resonance tomography as two free regions and one porous region with rotational symme-
(MRT) images of the internal capsule, which provided evidence of try. The capsule hold-up volumes before and after the membrane
structural irregularities in the membrane pleats. However, direct stack constitute the free regions, and the Navier–Stokes equations
evidence of linear flow variations within either capsule was not describe fluid flow in these two regions. The peak Reynolds num-
acquired, making it unclear if the ZRM structure and predictions are ber within the capsule remains below 5, clearly indicating laminar
correct at the mechanistic level. The current study aims to address flow. The incompressible Navier–Stokes equations without exter-
this shortcoming through the creation and solution of computa- nal forces are given by Eq. (1) (v denotes the fluid velocity, p the
tional fluid dynamic (CFD) models of the axial and radial membrane pressure, and  the kinematic viscosity).
chromatography modules described above. Although not widely    
employed in MC modeling, CFD provides a scale-neutral modeling ∂v 2
 + v · ∇v = −∇ pI + ∇ (∇ v + (∇ v)T ) − (∇ · v)I
strategy that is able to independently predict hydrodynamic behav- ∂t 3
ior. Linear velocities as a function of position and breakthrough
+F (1a)
curves under non-binding conditions may therefore be accurately
predicted using only knowledge of internal capsule geometries
and basic membrane properties. Moreover, when applied to
∇ ·v=0 (1b)
116 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122

The fluid flow within the porous membrane is modeled using These same two models will be used in this work. The Langmuir
Brinkman’s equations, considering the membrane stack as a sin- binding kinetics model is given by.
gle porous domain with porosity ε and permeability K. Brinkman’s
equations for this region are given by Eq. (2). ∂q
= ka c(qm − q) − kd q (4)
    ∂t
∂v∗ 2
 + v∗ · ∇ v∗ = −∇ pI + ∇ (∇ v∗ + (∇ v∗ )T ) − (∇ · v∗ )I
∂t 3 In Eq. (4), ka and kd are the adsorption and desorption rate
constants, and qm is the maximum binding capacity. The Lang-
 ∗
− v +F (2a) muir model assumes single-component interaction with one type
K of binding site for solute molecules that do not interact with each
other.
∇ · v∗ = 0 (2b) The spreading model (Eq. (5)) assumes two different bound
states due to a re-orientation or re-conformation of the bound
The above equations include momentum loss due to shear molecules. It is given by
stress. Here, v* is the interstitial velocity given by v/␧. No slip condi-
tions, i.e. v = 0, are applied at the inner capsule boundaries. A given ∂ q ∂ q1 ∂ q2
= + (5a)
linear flow rate is specified at the inlet, and vanishing viscous stress ∂t ∂t ∂t
is used as the boundary condition at the outlet of the capsule. The
velocity field and the pressure profile are continuous, resulting in
a stress discontinuity at the interface between the void and porous ∂q1
= (ka,1 c − k12 q1 )(qm − q1 − ˇq2 ) − kd,1 q1 + k21 q2 (5b)
regions. Membrane movements that might be caused by this stress ∂t
are not considered this study.
The transport of solute molecules is described by a classic con-
∂q2
tinuity (mass balance) equation, given by Eq. (3). = (ka,2 c + k12 q1 )(qm − q1 − ˇq2 ) − (k21 − kd,2 )q2 (5c)
∂t
dc 1 − ε dq
= −v∇ c + D∇ 2 c − (3) Here, q1 and q2 are the concentrations in bound states 1 and
dt ε dt
2, respectively, ˇ is the ratio of the sorbent surface area occupied
In Eq. (3), c and q denote the solute concentrations in the mobile
in state 2 relative to state 1, ka,1 , kd,1 , ka,2 and kd,2 are binding
and stationary phases, respectively, v the flow velocity, and D the
constants, that are defined in analogy to the Langmuir model, and
dispersion coefficient. The velocity vector v in Eq. (3) is taken from
k12 and k21 describe the rates of state changes.
the solution of Eq. (1) or Eq. (2), depending on the region (free or
porous). The stationary phase concentration, q, is accounted for per
unit volume of solid membrane. Possible binding models used to 3. Materials and methods
compute q are described in the next section. We have previously
shown that axial dispersion within the membrane stack makes a Bovine serum albumin (A 7638, Sigma–Aldrich Corp, St. Louis,
negligible contribution to the total system dispersion [12,14]. D is MO, USA) was used in breakthrough experiments at a concentration
therefore set to the molecular diffusivity of the protein molecule in of 1 g/l and a flow rate of 12 CV/min for both the axial and radial flow
water at column operating temperature. MC capsules. The protein was dissolved in 25 mM Tris buffer at pH
The CFD model is solved in two steps. In the first step, the 8.0 (Sigma–Aldrich, USA) for the loading step. Loading was followed
velocity field and the pressure profile, which are both stationary, by a washing step with 25 mM Tris buffer at pH 8.0. Then, 1 M NaCl
are computed by solving the coupled Navier–Stokes and Brinkman in 25 mM Tris buffer pH 8.0 was used to elute the bound BSA from
equations. In the second step, time-dependent concentration pro- the membranes. The units were cleaned with 1 N NaOH after each
files are computed by solving the mass balance equation, based on run, as specified by the manufacturer. In a revised protocol, the
the previously computed velocity field. cleaning step was performed with 1 M NaCl instead of 1 N NaOH.
For CFD modeling, the geometry of the XT140 capsule was Mustang Q XT5 anion-exchange membrane chromatography
re-constructed from MRT data using the commercial image capsules (axial flow) and Mustang Q XT140 anion-exchange mem-
processing software CorelDRAW graphics suite. The CFD model is brane chromatography capsules (radial flow) were purchased from
implemented in the commercial software COMSOL multiphysics. Pall Inc. (East Hills, NY, USA) (see Fig. 12). Both capsules contain
COMSOL spatially discretizes the PDE with finite elements and uses modified hydrophilic polyethersulfone (PES) membranes whose
a BDF method for time integration. The CFD geometries of both cap- surfaces are coated with a cross-linked polymer that contains
sules are finely meshed with unstructured triangular elements at pendant Q groups. The effective bed height of the membrane stack
a minimum and maximum element size of 0.0001 m and 0.001 m, in either capsule is 2.20 mm. The pore size and porosity ε of the
respectively. Zero solute concentration within the entire system membrane are 0.8 ␮m and 0.70 ± 0.05, respectively (manufacturer
was provided as initial condition, and vanishing viscous stress was data). Internal dimensions were manually measured by opening the
used as outlet boundary condition. The direct PARDISO solver is capsules. The total hold-up volume of the XT5 capsule was calcu-
chosen for solving the resulting large linear-equation systems. lated to be 6 ml from the weight difference between a dry capsule
and a capsule filled with water. Due to the symmetry of the capsule,
2.2. Binding models the hold-up volumes before and behind the membrane are assumed
to be 3 ml each. The XT140 capsule has hold-up volumes of 105 ml
Protein adsorption on functionalized membrane is a complex before and 45 ml after the membrane (manufacturer data). In addi-
process and several models, accounting for different physical tion, an experimental 9.4 T magnetic resonance tomography (MRT)
mechanisms, have been published [2]. A review of the general per- device was used to measure internal geometries.
formance of different kinetic models (Langmuir, bi-Langmuir, steric The XT5 capsule was attached to an ÄKTAexplorer system that
mass action, spreading) can be found in a previous publication was controlled by the Unicorn software (GE Healthcare, Uppsala,
[14]. In another publication [11], we compared a ZRM utilizing the Sweden). The XT140 capsule was attached to an ÄKTAprocess sys-
kinetic Langmuir model and with one utilizing the spreading model tem that was controlled by the Unicorn software (GE Healthcare,
[15,16] to evaluate their ability to reproduce breakthrough data. Uppsala, Sweden).
 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122 117

Table 1
Model parameters employed in the CFD simulations.

Parameter Value

Density,  [kg/m3 ] 1 × 103

Viscosity,  [Pa·s] 1 × 10−3
Permeability, k [m2 ] 1 × 10−13
Porosity, ε 0.7
Dispersion coefficient, D [m2 /s] 7.13 × 10−9

4. Results and discussion

4.1. Model assumptions and parameters

Strict mechanistic CFD simulations require one to model the

entire physical dimensions of the membrane chromatography cap-
sules under study. A full 3D simulation, including all physical Fig. 3. Comparison of simulation using the Roper and Lightfoot model, CFD sim-
complexities, with the available hardware and software (COMSOL ulated and measured breakthrough curve for bovine serum albumin loaded under
on a shared memory machine with 16 compute cores and 128 GB non-binding conditions on the XT5 capsule.
memory) would therefore be a challenging endeavor due to the
expense and long duration. Hence, judicious modeling assumptions
were made to reduce the computational effort with minimal impact the aims of both identifying the hydrodynamic and binding mech-
on the accuracy of the model in reflecting the real device: anisms leading to band broadening, and determining if the ZRM
properly captures those non-idealities and their underlying causes.
1. For the XT5 capsule, rotational symmetry was assumed, allowing For these studies, replacement of 1 N NaOH with 1 M NaCl in the
the CFD model to be transformed into cylindrical co-ordinates, cleaning protocol and reducing exposure to the storage buffer was
which reduces the model to 2D. This simplification effectively found to be crucial for achieving reproducible well-behaved exper-
assumes that the small capsule spacers do not contribute signif- imental breakthrough curves.
icantly to band broadening.
2. For the XT140 capsule, we previously showed that a single virtual 4.2.1. Non-binding conditions
zone along the axial coordinate was sufficient to quantitatively First, a forward simulation was performed with the CFD model of
predict breakthrough performance with the ZRM. Hence, the CFD the XT5 capsule, assuming no adsorption in the porous domain. This
model is set-up for a 2D cross-section, perpendicular to the axial ab initio simulation was solely based on the capsule geometry and
coordinate, which again reduces the model to 2D. parameters in Table 1. A dead time of 3 s was added to the solution
3. For either capsule, the membranes were considered static, and in order to account for a combined plug-flow dead time in the tub-
possible movements under dynamic conditions were not con- ing and in the Äkta system. The CFD simulated breakthrough curve
sidered. closely matches experiment (see Fig. 3), which illustrates the accu-
racy of the CFD model, including the aforementioned assumptions
We will show below that these assumptions are reasonable for that reduced the order to a 2D geometry, we observed in predicting
their respective capsule, so that quantitative predictions of experi- breakthrough for non-binding conditions.
mental behavior are achieved. The physical parameters of the CFD Our CFD simulations of the velocity field in the membrane stack
model are provided in Table 1. of the XT5 capsule (Fig. 4b) show a near constant linear velocity
throughout the membrane stack, which implies ideal flow in this
4.2. Analysis of an axial-flow MC system region. Non-ideal flow leading to a distribution of residence times
is predicted in the hold-up volumes before and behind the mem-
CFD model results were compared to experimental break- brane stack, and the CFD simulations reveal that this non-ideality
through curves and to ZRM results reported previously [11] with is caused by the geometry and high aspect ratio of the capsule.

Fig. 4. (a) Definition of central region (1) and outer region (2) in the CFD geometry of the XT5 capsule. Each region occupies 50 percent of the frontal surface area of the
membrane stack. (b) CFD simulation of the magnitude of the velocity field inside the XT5 capsule.
118 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122

Fig. 5. Averaged inlet side boundary concentration at the central and outer regions
of the membrane (as defined in Fig. 4a) in the CFD simulation of the XT5 capsule
under non-binding conditions.
Fig. 6. Best fit of various forms of the CFD model to breakthrough curves for bovine
serum albumin loaded onto the XT5 capsule. CFD results are shown when either the
Langmuir or spreading model for protein binding is employed.
Fig. 4a shows a virtual segmentation of the XT5 capsule into
two regions, each of which occupies 50% of the frontal surface area
of the membrane stack, and Fig. 5 shows the CFD simulated aver- CSTR volumes are similar to those volumes. They reflect the sizes
age boundary concentrations at the inlet side of the membrane in of corresponding CSTRs needed to match both the mass flows and
these two regions. One can clearly see that the inner and outer average solute residence times through each zone of the device. In
boundary concentration profiles differ not only in the time when the CFD simulation, 57% of the total mass passes though the cen-
the concentration front initially reaches the membrane, but also in tral membrane zone, while the remaining 43% passes through the
the initial slope and in the subsequent tailing. The flux through the outer region. The ZRM predicts this result essentially exactly (57%
outer region is delayed and more dispersive, due to an increased and 43%, respectively).
residence time in a larger fraction of the hold-up volume on the These results therefore show that ZRM analysis of the XT5
inlet side of the membrane, and the same mechanism is active in capsule, though not as rigorous as CFD, does not just provide a
the hold-up volume on the outlet side. The CFD simulations there- good correlative fit, but reveals real and important properties of
fore provide a detailed snapshot of flow patterns and sources of the physical system without requiring knowledge of the actual
non-uniform solute residence times at resolutions that cannot be system geometry. In particular, the ZRM provides reliable infor-
achieved by standard MC modeling approaches. mation on the percentage of the entering mass flow that cannot be
Flow non-idealities in the extra-membrane volumes are specif- treated as an ideal plug passing through the stack (and thus whose
ically addressed by the ZRM, and in accordance with CFD results breakthrough would not obey classic MC models). It likewise pro-
a symmetric two-zone configuration of the ZRM (Fig. 1b) was vides a reliable estimate of the elution band broadening caused by
required to accurately reproduce measured breakthrough data for the increased residence times of mass elements passing through
the XT5 capsule. The validity of the ZRM at the mechanistic level can the outer zone(s). The more rigorous CFD simulations obviously
be further evaluated by comparing the CFD computed and ZRM pre- provide this and more information as well, but with considerably
dicted hold-up volumes and mass flows through each of the outer more computational effort and time.
and central regions. In the CFD simulations, the hold-up volumes
upstream of the central membrane region (1) and outer membrane 4.2.2. Binding conditions
region (2) sum to the total upstream distributor volume (they are The CFD model was next combined with either the Langmuir
1.90 ml and 1.10 ml, respectively), as shown in Fig. 4a. The cor- or spreading model to simulate breakthrough under load (binding)
responding upstream inner and outer CSTR volumes used in the conditions. The parameters of either binding model were estimated
ZRM are best-fit values (1.69 ml and 1.24 ml, respectively). While by fitting CFD simulations to measured breakthrough curves for
not required or expected to be identical to the partitioning of the BSA (Tables 2 and 3). Fig. 6 shows the best fit of each CFD model
XT5 capsule in the CFD analysis shown in Figs. 4a and 5, the ZRM to a breakthrough curve for BSA loaded at a flow rate of 12 CV/min.

Fig. 7. (a) cross-sectional MRT scan of the XT140 capsule. The membrane pleats are clearly visible in gray, due to their water content and irregularities shown in different
colors, (b) 2D CFD geometry reconstructed from the MRT image.
 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122 119

Table 2
Parameters of the Langmuir model, as determined by fitting the corresponding
Roper and Lightfoot model (RLM), CFD and ZRM simulations to a binding break-
through curve of the XT5 capsule. The ZRM related parameter values are taken from
a previous publication [11].

Parameter CFD ZRM RLM

ka [l/(g·s)] 6.434 × 10−2 6.4 × 10−2 6.13 × 10−2

kd [1/s] 6 × 10−3 6 × 10−3 7.2 × 10−4
qm [g/l] 278.19 284.04 284.83

Table 3
Parameters of the spreading model, as determined by fitting the corresponding
Roper and Lightfoot model (RLM), CFD and ZRM simulations to a binding break-
through curve of the XT5 capsule. The ZRM related parameter values are taken from
a previous publication [11].
Fig. 8. Comparison of CFD simulation to measured breakthrough curve of XT140
Parameter CFD ZRM RLM capsule under non-binding conditions.
ka1 [l/(g·s)] 8.073 × 10−2 8.08 × 10−2 8.37 × 10−2
kd1 [1/s] 1.06 × 10−5 1.06 × 10−5 0.82 × 10−5
structural irregularities are marked in the figure. The pleats indi-
k12 [l/(g·s)] 11.08 × 10−4 7.37 × 10−4 10.67 × 10−4
k21 [1/s] 10.01 × 10−3 9.41 × 10−3 13.45 × 10−3 cated by red arrows have larger cross-sectional areas as compared
qm [g/l] 286.56 289.003 289 to average pleats. The sealing region, which is the start and end
ˇ 1.15617 1.144 1.166 point of the winding process (yellow bars), is thicker than the rest
of the pleats. Finally, some pleats are broadened at the outer loops
(green bars). A CFD geometry that accounts for all of these struc-
CFD simulations using the Langmuir model fail to reproduce the tural irregularities was reconstructed from the MRT image using
observed tailing of the breakthrough curve near sorbent saturation, COREL draw, as shown in Fig. 7b. The porosity in the membrane
indicating non-ideal binding mechanisms contribute to broaden- region was increased within the outer loops in order to main-
ing of BSA breakthrough, even though the CFD model accurately tain a constant average membrane volume (however, no impact of
describes the hydrodynamics in the XT5 capsule. The spreading these porosity variations on the simulated chromatograms could
model accurately captures these non-idealities, and when com- be observed, data not shown). A two-dimensional CFD model was
bined with the accurate hydrodynamics provided by CFD results thereby achieved with the hold-up volumes upstream and down-
in an excellent representation of the true breakthrough curve. stream of the membrane set at their experimental values of 62 ml
It is important to note here that using the classical Roper and and 36 ml, respectively. Due to additional channels, the XT140
Lightfoot model along with the spreading isotherm also regresses capsule has total hold-up volumes of 105 ml and 45 ml. The resid-
to the binding breakthrough curve well (data not shown). However, ual hold-up volume upstream of the membrane is accounted for
the parameters of the binding models are different when estimated by a CSTR before the CFD model with average residence time
using the Roper and Lightfoot model, because the description of  = (105–65)/28 = 1.53 s. The protein concentration CCFD,in at the
non-ideal flow is then lumped into the binding model. Hence, these model capsule inlet is then given by
parameters cannot be used for predicting the performance of differ-
ent capsules with the same membrane but changed hydrodynamics CCFD,in = C(1 − e−t/ ) (6)
(model-based scale-up), as will be further discussed in Section
Similarly the output of the CFD simulation is connected with a
4.3.2. The binding parameters obtained through CFD-based param-
second CSTR in order to model the remaining downstream hold-up
eter estimation are summarized in Tables 2 and 3. The parameters
volume with average residence time  = (45–36)/28 = 0.32 s:
previously obtained through ZRM analysis and using the classical
Roper and Lightfoot model are also tabulated for comparison. The dc CCFD,out − c
= (7)
binding parameter sets are quite similar in both cases, even though dt 
different flow models and independent optimization algorithms First, a forward simulation was performed with this CFD model
have been used. of the XT140 capsule, assuming no binding in the porous domain.
The same set of initial parameters was used for the CFD and A dead time of 7.5 s was added to the solution in order to account
ZRM models. However, the CFD based optimization of the spread- for a plug flow in the tubing of the Äkta system. Fig. 8 shows the
ing model took 2 weeks on 4 compute cores, whereas the ZRM predicted breakthrough curve, which closely matches the experi-
based optimization took only 10 min on one compute core. Hence, mental data. The stationary velocity profile in the XT140 capsule
though the CFD method is more information rich, the reduced com- will be discussed in next section, as the solute residence times
putational complexity of the ZRM can be a big advantage when in the membrane are very small under non-binding conditions.
different binding models need to be evaluated for suitability. Hence, structural irregularities hardly impact on the simulated
breakthrough curves. This was also observed in the ZRM analysis
4.3. Analysis of an radial-flow MC system of the same capsule [11], where a one-zone configuration was able
to accurately correlate with the experimental non-binding curve.
A similar analysis was conducted on the larger radial-flow MC However, below we show that the widely accepted practice of using
device to identify differences in hydrodynamics and sources of flow non-binding breakthrough data, alone, is not sufficient to quantify
non-ideality, and to assess the fundamental value of the ZRM when flow inhomogeneities and then predict breakthrough in large-scale
applied to this configuration [11]. MC capsules.

4.3.1. Non-binding conditions 4.3.2. Binding conditions

Fig. 7a shows an MRT image at an internal axial position of In Section 4.2.2, we showed that CFD simulations employing
the XT140 capsule. Although most pleats are similarly structured, the spreading model accurately reproduce BSA breakthrough from
certain differences in folded pleat can be observed; some specific the axial-flow XT5 capsule. As a more powerful illustration of the
120 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122

Fig. 9. Comparison of measured breakthrough data of the XT140 capsule under binding conditions and of model-based predictions, computed with parameters of the
spreading model that have been estimated from breakthrough data of the XT5 capsule. The hydrodynamics in both capsules was modeled using (a) the CFD model, and (b)
the classical Roper and Lightfoot model.

value of CFD modeling, truly predictive simulation of breakthrough

from the larger scale and altered geometry (radial flow) XT140 MC
capsule was performed by transferring the binding parameters esti-
mated from data for the XT5 capsule to a CFD model of the XT140
capsule; an attempt to predict the breakthrough curve under bind-
ing conditions was then made. As shown in Fig. 9a, the result of
this simulation, for which no model parameters measured directly
on the XT140 device (basic device geometry values excepted) were
used, matches the experimental data very well. Due to the careful
decomposition of scale-dependent flow non-idealities and scale-
independent binding non-idealities, both the initial increase and
the tailing of the breakthrough curve are accurately predicted. For
comparison, the same model-based scale-up was attempted using
the Roper and Lightfoot model. As shown in Fig. 9b, this approach
fails to quantitatively predict the experimental data, which fur-
ther underlies the importance of carefully separating the impact
of non-idealities that are related to binding and hydrodynamics.
In addition to its predictive power, the CFD model provides
a wealth of mechanistic information related to causes of band-
broadening as well. For example, Fig. 10 reports magnitudes of Fig. 10. CFD simulation of the magnitude of the velocity field inside the XT140
the velocity field within the cross-section of the XT140 capsule. capsule.
The fluid enters radially at the outer boundary of the capsule,
passes through the curved membrane pleats, and exits through
channels at the inner capsule boundary. Variations in fluid veloc-
ity are predicted in the void regions upstream and downstream the membrane volume can by roughly split into two sub-regions
of the membrane, with the largest gradients occurring upstream (Fig. 11a). The first sub-region, which covers approximately 15% of
of the folded regions of the membrane. This suggests the added the total membrane area and mainly comprises the sealing (inner)
membrane surface area and throughput provided by the pleat and the outer membrane bends, is characterized by fluid veloci-
design may come at the expense of non-ideal fluid hydrodynamics. ties less than 85% of the average. In the larger sub-region, mainly
Moreover, in contrast to the XT5 capsule, the fluid velocity is not covering the linearly configured pleat centers, velocities are more
constant within the porous membrane. Instead, based on velocities, uniform and close to (within 15% of) the average.

Fig. 11. Distribution of the volumetric flow relative to the total volumetric flow, f, over the corresponding linear velocity relative to the average linear velocity, v: (a) velocity
distribution obtained from the CFD analysis of the XT140 capsule, (b) velocity distribution obtained from the ZRM analysis of the XT140 capsule, taken from previous
publication [11].
 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122 121

the XT140 capsule were hypothesized to cause non-ideal flow due

to different linear velocities though the membrane stack.
In the current contribution, mechanistic CFD models are
employed to independently validate and confirm the assumptions
and hypotheses, which were developed in the previous ZRM anal-
yses. CFD modeling is based on the exact capsule geometries. Both
models, for the XT5 and XT140 capsules, could be reduced to 2
spatial dimensions, in order reduce computation requirements.
The stationary velocity profile, computed for the XT5 geometry,
shows that the central membrane region receives a different inflow
than the peripheral membrane region, due to the great width of
the capsule. Hence, the assumption of two parallel zones in the
corresponding ZRM configuration is not only reasonable, but also
required. In full agreement with ZRM results, the spreading model
for protein binding was found to describe the experimental bind-
ing breakthrough curve much better than the traditional Langmuir
model. The stationary velocity profile, computed for the XT140
geometry, shows that a small fraction of the membrane area, mainly
the sealing and bend regions, have significantly reduced veloc-
ity magnitudes. This velocity reduction explains the strong tailing
observed in the experimental breakthrough curves of this capsule.
Fig. 12. The XT5 and XT140 capsules as bought from the vendor. A 2 Euro coin is The distribution of flow magnitudes throughout the membrane
included for size comparison. can be described and parameterized by a combination of linear
and Gaussian functions. The ZRM configuration with varying linear
velocities, introduced in a previous publication [11], is therefore
In contrast, we recently reported that direct application of the
proven to be most adequate to capture the impact of hydrodyna-
ZRM to model-based scale-up failed for the Pall Mustang XT140
mics in the XT140 capsule.
capsule [11] because, unlike with the CFD model, linear velocity
The CFD simulations in this contribution do not only provide a
variations within the membrane pleats could not be determined
quantitative validation of the ZRM, but can also accurately predict
from non-binding data. Instead, the velocity distribution used in
non-binding breakthrough curves of the XT5 and XT140 capsules.
the ZRM was calibrated to breakthrough data collected under bind-
Moreover, a CFD model of the XT140 capsule was able to pre-
ing conditions. That distribution is shown in Fig. 11b and can be
cisely predict the measured binding breakthrough curve, based
approximated and parameterized by an initial linear increase with
on binding parameters obtained from an analysis of the XT5 cap-
given slope and a Gaussian peak with given center and width. Based
sule, thereby providing a method for model-based scale-up. The
on that distribution, we were able to define a multi-zone configura-
fundamental knowledge obtained about hydrodynamics can also
tion of the ZRM capable of modeling breakthrough curve shape with
potentially lead to improvement in capsule design. On the other
good accuracy [11]. Notably, the velocity distribution estimated by
hand, CFD simulations are computationally expensive and time
best fit of the ZRM to breakthrough data under binding conditions
consuming, which hinders their routine application in dynamic
is nearly identical to that predicted from first principles by our CFD
process analysis and design of experiments. The ZRM, conversely,
analysis of the XT140 capsule (see Fig. 11a). Thus, the ZRM, once cal-
provides a much faster modeling approach and, as substantiated
ibrated, can be employed to properly understand hydrodynamics
by the present contribution, not only accurately describes the
and predict the breakthrough performance at different operating
impact of hydrodynamics, but also reflects physical properties of
conditions and for different molecules in much shorter compute
the studied capsules. Therefore, both the ZRM and CFD modeling
times, as compared to the CFD approach. Knowledge of the internal
approaches have great potentials to be used as tools in process anal-
capsule geometry is not required. The two modeling approaches are
ysis and capsule design. Future work will address ZRM and CFD
therefore complementary, with the CFD model serving to provide
analyses of different capsules from other vendors, and of different
a comprehensive analysis of system hydrodynamics and sources
molecules and operating conditions.
of band broadening, and thus a rational basis for improving cap-
sule design, and the ZRM providing a reliable and rapid method for
simulating process operation and screening for optimal operating Acknowledgements
conditions.
The presented work was supported by the Cluster for Industrial
Biotechnology (CLIB) with a doctoral scholarship for Pranay Ghosh,
5. Conclusions and outlook
as well as by NSERC, the Natural Sciences and Engineering Research
Council of Canada. The authors wish to thank Birgit Stute, Andreas
Ideal hydrodynamics behavior is hard to achieve in membrane
Püttmann and Sebastian Schnittert for their valuable inputs and
chromatography capsules at all scales for various reasons, and tra-
help throughout this work.
ditional modeling approaches cannot capture non-ideal flow. The
zonal rate model (ZRM) has been originally developed as a semi-
empirical approach for independently quantifying the impacts of References
non-ideal flow and non-ideal binding in membrane chromatogra-
[1] F.T. Sarfert, M.R. Etzel, J. Chromatogr. A 764 (2000) 1.
phy capsules. In the ZRM, the physical capsule geometry is virtually [2] S.Y. Suen, M.R. Etzel, J. Chromatogr. A 686 (1994) 2.
partitioned into zones that are approximately homogeneous. In a [3] E. Klein, J. Membr. Sci. 179 (2000) 1.
previous publication [11], the ZRM was applied for analyzing Pall [4] J.H. Vogel, H. Nguyen, R. Giovannini, J. Ignowski, S. Garger, A. Salgotra, Bio-
technol. Bioeng. 109 (2012) 12.
Mustang XT5 and Mustang XT140 capsules. The high length-to- [5] D.K. Roper, E.N. Lightfoot, J. Chromatogr. A 702 (1995) 1.
width ratio of the XT5 capsule was concluded to cause non-ideal [6] C. Boi, J. Chromatogr. B 848 (2007) 1.
flow due to flow paths of different lengths, while irregular pleats in [7] S. Dimartino, C. Boi, G.C. Sarti, J. Chromatogr. A 1218 (2011) 13.
122 P. Ghosh et al. / J. Chromatogr. A 1305 (2013) 114–122

[8] J. Wang, F. Dismer, J. Hubbuch, M. Ulbricht, J. Membr. Sci. 320 (2008) 1. [12] P. Francis, E. von Lieres, C.A. Haynes, J. Chromatogr. A 1218 (2011) 31.
[9] R. Ghosh, T. Wong, J. Membr. Sci. 281 (2006) 1. [13] E. von Lieres, J. Wang, M. Ulbricht, Chem. Eng. Technol. 33 (2010) 6.
[10] A. Tejeda, J. Ortega, I. Magaña, R. Guzmán, J. Chromatogr. A 830 (1999) 293. [14] P. Francis, E. von Lieres, C.A. Haynes, Biotechnol. Bioeng. 109 (2012) 3.
[11] P. Ghosh, K. Vahedipour, M. Lin, J.H. Vogel, C.A. Haynes, E. von Lieres, Biotechnol. [15] H. Yang, M.R. Etzel, Ind. Eng. Chem. Res. 42 (2003) 4.
Bioeng. 110 (2013) 4. [16] A.J. Clark, A. Kotlicki, C.A. Haynes, A.L. Whitehead, Langmuir 23 (2007) 10.