Separation and Purification Technology 62 (2008) 582–589

Contents lists available at ScienceDirect

Separation and Purification Technology

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

A Monte Carlo-based error propagation analysis

of Simulated Moving Bed systems
Anjushri S. Kurup, Hariprasad J. Subramani 1 , Michael T. Harris ∗
School of Chemical Engineering, Purdue University, West Lafayette, IN 47907, United States

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

Article history: A Monte Carlo-based uncertainty analysis for the performance assessment of a Simulated Moving Bed
Received 9 July 2007 (SMB) system, which is widely employed in the petrochemical, sugar, and pharmaceutical industries, is
Received in revised form 23 February 2008 presented. The error propagation analysis is carried out to predict the performance of the SMB system
Accepted 26 February 2008
for two different binary separations: (a) glucose–fructose separation in a pilot-scale SMB unit and (b)
separation of p-xylene from xylene isomers in an industrial-scale SMB unit. Existing SMB models in liter-
Keywords:
ature, which utilize a linear isotherm for the glucose–fructose separation [D.C.S. Azevedo, A.E. Rodrigues,
Chromatography
Fructose–glucose separation in a SMB pilot unit: modelling, simulation, design and operation, AIChE J.
Simulated Moving Bed
Monte Carlo method
47 (2001) 2042–2051] and a nonlinear isotherm for the separation of xylene isomers [M. Minceva, A.E.
Uncertainty analysis Rodrigues, Modeling and simulation of a simulated moving bed for the separation of p-xylene, Ind. Eng.
Glucose Chem. Res. 41 (2002) 3454–3461], are employed for the analysis. While the purity and productivity of
Fructose glucose and fructose are chosen as the performance (or output) variables in the former case study, the
Xylene purity and recovery of p-xylene are chosen as those in the latter. In both cases, the isotherm parameters,
often obtained from chromatography experiments with appreciable levels of uncertainty, are chosen as
the process (or input) variables. From the sensitivity analysis, simple models are developed to predict the
variations in the performance of the SMB as a function of the isotherm parameters.
© 2008 Elsevier B.V. All rights reserved.

1. Introduction
Accurate experimental estimation of physical, chemical, and
physico-chemical properties is very essential for precise design
of chemical processes. Typical examples include the estimation
of vapor–liquid equilibria for distillation, kinetic rate constants
for reactor design, and adsorption isotherms for chromatographic
separations. However, such estimations always involve signifi-
cant uncertainties. Among the most common error sources are
random, systematic or bias, and model formulation errors [1].
While random errors are due to the experimental apparatus
and techniques, there are always irreducible residual systematic
errors in even the most credible experimental measurements.
The error due to inappropriate model formulation could, how-
ever, be eliminated by accurate models. Tremendous competition
as a result of globalization of world market has compelled
industries to appreciate the need to assess the effect of these
uncertainties on the chemical processes. Such uncertainty anal-
∗ Corresponding author.
E-mail address: mtharris@ecn.purdue.edu (M.T. Harris).
1
Presently at Chevron Energy Technology Company, Houston, TX 77002, United
States.
yses are often carried out via several formal approaches such
as differential analysis, Monte Carlo analysis, response surface
methodology, and Fourier amplitude sensitivity test [2]. Of these
approaches, Monte Carlo analysis has gained significant popular-
ity these days in performance assessment of chemical processes
[3].
Adsorption-based separation, viz., chromatography, is a com-
monly employed separation technique (both batch and continuous)
in chemical industries such as petrochemicals, sugar, and more
recently pharmaceuticals. With the revolutionary progress in
MEMS technology, miniaturized versions of chromatography finds
extensive application in forensic and health-care applications
[4–5]. For efficient modeling and design of these batch/continuous
chromatography and to have better control on the performance,
accurate estimation of several parameters such as adsorption
isotherms [4], diffusion, mass transfer coefficients, etc., and
measurement of species concentration, flow rate, temperature,
pressure, etc., are of paramount importance. Any type of error (ran-
dom or systematic) in the measurement or estimation of these
quantities could significantly affect the performance assessment
of these chromatographic techniques. For example, the errors in
measurements of adsorption isotherms (be it linear or nonlinear)
could have profound effect on the performance prediction of chro-
matographic columns. There have been quite a few research works

1383-5866/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.seppur.2008.02.021
 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589
that have been reported on the uncertainty analysis of HPLC (batch)
systems [6–9].
To the best of the authors’ knowledge, there is virtually no lit-
erature available on the uncertainty analysis of continuous-scale
chromatographic separators such as True Moving Bed (TMB), Mov-
ing Column, and Simulated Moving Bed (SMB) systems that are
so commonly used in chemical industries [10]. Of all the continu-
ous chromatographic separators, SMB and its modifications have
been proven to effect near-complete separation of petrochemi-
cals, sugars, isomers, and enantiomers [5,11–18]. Given the superior
separating power of SMB, it is but for natural to study the error prop-
agation in performance assessment as a result of uncertainties in
estimation of parameters such as flow rates, adsorption isotherms,
diffusion and mass transfer coefficients. Hence, one of the goals
of the present study is to conduct systematic parametric sensi-
tivity and uncertainty analyses for the performance assessment of
SMB systems using Monte Carlo methods. Another goal is to per-
form error propagation analysis for the performance of the SMB
system for two different binary separations: (a) glucose–fructose
separation in a pilot-scale SMB unit [14] and (b) separation of p-
xylene from xylene isomers in an industrial-scale SMB unit [18].
Performance indicators (or output variables) in these analyses are
the purity and productivity of glucose and fructose in the former
case study, and the purity and recovery of p-xylene in the lat-
ter. In both cases, process parameters (or input variables) are the
isotherm parameters, often obtained from chromatography exper-
iments with appreciable levels of uncertainty.
In what follows, a brief overview of SMB systems and the math-
ematical models to describe SMB system for the separation of
glucose–fructose mixtures and a mixture of xylene isomers respec-
tively is presented. Next, the Monte Carlo methodology for the error
propagation analysis of these two SMB systems is discussed. Then,
the results of performance assessment of both the binary separa-
tion of sugars and xylene isomers are presented. Finally, concluding
remarks and directions for future work are listed.
2. Simulated Moving Bed (SMB) systems
A schematic representation of a SMB system is illustrated in
Fig. 1 that consists of a number of packed (with adsorbent parti-
cles of diameter dp ) columns of uniform cross-section connected
in a circular array, each of length Lcol and diameter dcol . The two
incoming streams (the feed, F, and the desorbent/eluent, D) and
the two outgoing streams (the raffinate, Ra, and the extract, Ex)
Fig. 1. A schematic representation of a four zone (two columns per zone) SMB
system.
583
divide the system into four sections (zones), namely P, Q, R and S,
each of which is comprised of p, q, r and s columns respectively.
The flow rates in the sections P, Q, R and S are designated as QP ,
QQ , QR and QS respectively while those of the feed, raffinate, elu-
ent/desorbent and extract are designated as QF , QRa , QD and QEx
respectively. However, only four of the above eight flow rates are
independent, as the remaining four are determined from the mass
balance at points (nodes) A, B, C and D (as shown in Fig. 1). By suit-
ably advancing the inlet and outlet ports, column by column, in
the direction of the fluid flow at a pre-set switching time, ts , the
countercurrent movement of the solids is simulated. This switch-
ing time is the key parameter, which defines the hypothetical solid
velocity ( ∼ Lcol /ts ). However, the countercurrent separation of the
components could be achieved only by judiciously specifying the
internal flow rates in the columns besides the switching time.
Literature on SMB systems is quite exhaustive. Theories like
McCabe–Thiele method [10], triangle theory [19–20], standing
wave design [21], and separation volumes [22] have also been
developed over the years that have become quite popular in
the operability design of SMB systems. Several researchers have
also reported design and optimization methodologies [11,14,19–25]
to achieve optimal separation using SMB. Attempts have also
been successful in employing SMB as an in situ reactor–separator
[12,26–27].
2.1. Mathematical model for SMB and linear/nonlinear isotherms
A mathematical model that has been proven to mimic SMB
systems [11,14–15,18] very well is employed in the present study.
The differential transient mass balance equations along with the
boundary conditions are as follows:
2
∂Ci,k ∂qi,k k ∂ Ci,k ∂Ci,k
+ = − k (1)
∂ ∂ Pek ∂2 ∂
∂qi,k
= ˛m (q∗i,k − qi,k ) (2)
∂
where Ci,k is the concentration of the ‘I’th species in the liquid in
the ‘k’th column, q∗i,k is the concentration in the solid phase in equi-
librium with that in the liquid, qi,k is the concentration in the solid,
 = z/Lcol is the dimensionless length,  = (1 − ε)/ε is the ratio of vol-
umes of solid to liquid phases, ε is the bed porosity, ˛m = ts ˛ is the
dimensionless mass transfer coefficient and  = t/ts is the dimen-
sionless time. The other dimensionless parameters in the above
equations are the ratio of liquid to solid velocities ( k = UFz ts /Lcol )
and Peclet number (Pek = UFk Lcol /Daxk ).
In the present study, either a linear isotherm (Eq. (3)) for
glucose–fructose separation (experimentally measured by Azevedo
and Rodrigues [14] which was later utilized in an optimization
study by Subramani et al. [11]) or a non-stoichiometric Langmuir
isotherm (Eq. (4)) for xylene separation (experimentally mea-
sured by Minceva and Rodrigues which was later employed in an
optimization study by Kurup et al. [15]) is used to describe the
multi-component adsorption equilibria.
q∗i,k = Ki Cik (3)
Ki Cik
q∗i,k = N (4)
1+ KC
j=1 j jk
The boundary conditions are:
in 1 ∂Ci,k
Ci, = Ci,k (0, ) − (5)
Pek ∂
∂Ci,k
(1, ) = 0 (6)
∂
584 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589
The initial conditions are:
0
Ci,k (, 0) = Ci,k () and qi,k (, 0) = q0i,k ()
The mass balance at the node is then,
in
Ci,k = Ci,k−1 (1, )
except if the column follows feed or desorbent port in which case,
in [QF Ci,F + QQ Ci,k−1 (1, )] in QS Ci,k−1 (1, )
Ci,k = and Ci,k =
QR QP
respectively.
The resulting system of partial differential equations (PDEs)
are discretized in space using the finite difference approximations
to convert them into a system of coupled ordinary differential
equations—initial value problems (ODE-IVPs) by the method of
lines [11,12,15–17,26,27]. Each column length is spatially discretized
using 100 grid points. These stiff initial value ODEs were solved
using a robust algorithm developed in FORTRAN90 along with
the subroutine, DIVPAG (based on Gear’s method), in the IMSL
library. Since a periodic switching is imposed on the system, the
SMB system operates under transient conditions. However, a cyclic
(periodic) steady state with a period equal to the global switch-
ing time is eventually reached after several switchings. All the
relevant parameters for both the separation of sugars [11,14] and
xylene isomers [15,18] are extracted from published literature. In
the glucose–fructose separation problem, the performance indica-
tors or the output variables are the purities and productivities of
glucose and fructose, where, the purities (%) of glucose and fructose
are designated as PurG and PurF respectively, and their produc-
tivites (kg/m3 h) are denoted by ProG and ProF respectively. In the
case of separation of xylene isomers, the performance indicators
(output variables) are the purity and recovery of p-xylene which
are designated as PurX and RecX respectively. In the former case,
the process parameters or the input variables are the adsorption
isotherm parameters of glucose and fructose denoted by KG and KF
respectively. In the latter case, the process parameters (input vari-
ables) are the adsorption isotherm parameters of p-xylene, ethyl
benzene, and the desorbent (p-diethyl benzene) denoted by KpX ,
KEB , and KpDEB .
2.2. Description of Monte Carlo methodology
Basically in such a Monte Carlo analysis of the performance of
SMB systems, multiple evaluations of the SMB model (as described
in the previous subsection) are carried out with probabilistically
selected input variables (e.g., adsorption isotherm parameters in
the present study), and the results of these evaluations are then
utilized to determine both the uncertainty in model predictions
(e.g., purity and productivity or recovery of the products) and the
input variables that give rise to this uncertainty.
In a general framework [2], the performance (or output) vari-
ables such as PurF, ProF, and RecX are represented as yi (i = 1, 2, . . .,
N, where N is the total number of performance variables) and the
functional dependence of yi on the input variables such as KF and
KpX denoted by xj (j = 1, 2, . . . M, where M is the total number of
input variables) is given by
y = f (x)
where y = [yi ]T , f = [fi ]T , and x = [xj ]T are vector representations
of the output variables and their functional dependence on the
input variables. The performance variables, however, are evaluated
by solving the SMB model equations (transient one-dimensional
model) described in the earlier subsection. In other words, the
mathematical model (Eqs. (1)–(9)) can be represented as:
(7) g(t, z, y, x) = h(t, z, x) (11)
where g and h are vectors. The vector g in the present study of SMB
systems is a partial differential operator and requires initial and
(8) boundary conditions as described by Eqs. (5)–(9).
As a foremost step, a range and distribution (e.g., uniform, logu-
niform, normal, etc.) are chosen for the input variables xj . Second,
a sample is generated (often randomly) from the ranges and dis-
(9) tributions specified in the first step. In the third step, the model
is evaluated for the determination of the output variables for each
set of sample values of the input variables. In essence, these model
evaluations map the analysis inputs (x) to the analysis results (y)
that is further utilized for the subsequent uncertainty and sensi-
tivity analyses [2]. In the next step, mean values of the sample of
input variables (xm ) and of the evaluated output variables (ym ) are
calculated and utilized to normalize the values of the input and
the corresponding output variables. In the step that follows, the
relationships between y and x are explored using scatter plots and
regression analysis.
3. Results and discussion
3.1. Choice of independent parameters and motivation
As has been mentioned in the earlier sections, the error propaga-
tion analyses of SMB systems have been carried out for two model
separation problems, viz., separation of glucose–fructose mixture
to obtain fructose-rich solutions [11,14] and separation of xylene
isomers to obtain a p-xylene-rich product stream [15,18]. Even
though, several parameters such as switching time, flow rates in the
various sections, etc., and measurement errors in these parameters
can also have significant effect on the performance of SMB systems,
in the present study, adsorption isotherm parameters of the vari-
ous species are chosen as the input variables or process parameters.
Adsorption isotherms describe the partition of the species in the
solid and liquid phases and any measurement errors in their estima-
tion can have adverse effect on the performance of SMB. Also, as the
column diameter is decreased, the adsorption isotherms of differ-
ent species, though being thermodynamic properties, change due
to the differences in packing density and inaccurate measurements
of extra-column volumes [4–5]. Hence, in the first case study of
this paper, viz., separation of sugars, the linear adsorption isotherm
parameters of glucose (KG ) and fructose (KF ) respectively are cho-
sen as the input variables, and their purity (PurG and PurF) and
productivity (ProG and ProF) are chosen as the output or perfor-
mance variables. In the latter case study, the nonlinear isotherm
parameters of p-xylene (KpX ), ethyl benzene (KEB ), and p-diethyl
benzene (KpDEB ) are chosen as the input variables, and the purity
(PurX) and recovery (RecX) of p-xylene are chosen as the output
variables.
3.2. Error propagation in pilot-scale linear system (sugars
separation)
In the case of the SMB system for the sugars separation [11,14],
the operating conditions of the pilot plant are as indicated in
Table 1. Under these conditions, while the purity (PurG) and pro-
(10) ductivity (ProG) of glucose are 86.1% and 6.68 kg/m3 h respectively,
the purity (PurF) and productivity (ProF) of fructose are 93.9%
and 5.31 kg/m3 h respectively. These model predictions from the
present study are in close agreement with experiments [14]. Since,
the measurement errors of adsorption isotherm constants are typ-
ically of the order of ±10% [4], in the Monte Carlo-based error
propagation analysis of the SMB system for the sugars separation,
 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589
Table 1
SMB operating conditions for the separation of sugars [11,14]
Parameter
Number of columns, Ncol
Column configuration, 
Length of the column, Lcol (m)
Diameter of the column, dcol (m)
Operating temperature, T (K)
Peclet number, Pe
Biot number, Bi
Porosity of packing, ε
Linear adsorption isotherm parameter for glucose, KG
Linear adsorption isotherm parameter for fructose, KF
Switching time, ts (s)
Mass transfer coefficient for glucose, ˛G
Mass transfer coefficient for fructose, ˛F
Feed concentration of glucose, CG,f (mol/l)
Feed concentration of glucose, CF,f (mol/l)
Feed flow rate, QF (m3 /s)
Eluent flow rate, QE (m3 /s)
Raffinate flow rate, QRa (m3 /s)
Flow rate in section R (between feed and raffinate), QR (m3 /s)
the process parameters or the input variables namely KG and KF
are randomly sampled from a uniform distribution with a range of
±10% of the adsorption constants (cf. Table 1). All the other pro-
cess parameters are fixed constant at the values listed in Table 1.
Monte Carlo Simulations are then run at each of the sample sets,
viz., (KG , KF ) pair to predict the performance of the SMB for the
glucose–fructose separation.
The variations with KG and KF of the productivity of glucose,
ProG, and the purity of glucose, PurG, respectively are illustrated
in Fig. 2(a) and (b) in the form of a scatter plot. It is evident
from Fig. 2(a) that the productivity is a strong function of KG ,
viz., ProG sharply decreases as KG increases, which is attributed to
Fig. 2. Effect of variations in KG and KF on the productivity and purity of glucose.
585
Value
12
3/3/3/3
0.3
0.026
323
500
300
0.4
0.27
0.53
198
6.24
4.39
0.2222
0.2222
0.56 × 10−7
1.67 × 10−7
1.21 × 10−7
5.21 × 10−7
Fig. 3. Effect of variations in KG and KF on the productivity and purity of fructose.
the relatively stronger adsorption of weakly adsorbed component
(glucose) to the adsorbent (solid) as KG increases. Also Fig. 2(a)
makes plain that ProG does not follow any observable trend to
changes in KF , which is intuitively obvious because of glucose being
the weakly adsorbed component and the operating point of the
pilot plant is tuned to effect high degree of separation of fructose
[11,14]. Similarly, Fig. 2(b) explicates that the purity of glucose, PurG,
also sharply decreases with increase in KG . Also, as KF increases,
PurG decreases due to the contamination of the glucose (raffinate)
stream with fructose (cf. Fig. 2(b)). This is because, with increase in
KF , the operating value of the eluent (or desorbent) flow rate may
not be sufficient enough to completely desorb fructose from the
adsorbent.
Fig. 3 illustrates the variation with KG and KF of the productivity
(ProF) and purity (PurF) of fructose in the form of a scatter plot. It
is clear from Fig. 3(a) that while ProF does not follow any trend to
changes in KG , it does sharply decrease with increase in KF which
again is attributed to the strong adsorption of fructose and ineffi-
cient desorption due to a fixed eluent flow rate. Fig. 3(b) further
explicates that PurF sharply decreases with increase in KG (which
is due to the contamination of fructose stream with glucose which
gets relatively strongly adsorbed due to increase in KG ) and does
not follow any trend to changes in KF .
Furthermore, a better understanding of the parametric sensi-
tivity of the SMB system for the separation of sugars is obtained
by transforming the input and output variables into normalized
deviation variables [2]. The mean values of KG and KF (0.2712 and
0.5307 respectively) from the random sampling of a uniform dis-
tribution, viz., ±10% of the experimentally estimated KG and KF , are
used to normalize the deviations of the input variables from their
mean. These normalized deviations of KG and KF about the mean are
henceforward referred to as NKG and NKF respectively. Similarly,
the mean values of the PurG, PurF, ProG, and ProF (84.9%, 91.3%,
586 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589

Fig. 5. Effect of variation in NKG on the normalized purity of glucose and fructose.
Fig. 4. Effect of variation in NKG on the normalized productivity of glucose and
fructose.

6.53 kg/m3 h, and 4.89 kg/m3 h respectively) are used to normalize

the deviations of the output variables about their mean and these
normalized deviations are designated as NPurG, NPurF, NProG, and
NProF respectively.
Fig. 4 shows effect of the normalized deviation in KG (or NKG )
on the deviations in productivity of glucose (NProG) and fructose
(NProF) respectively. As explained earlier, NProG shows a dras-
tic change with changes in NKG (cf. Fig. 4(a)). There is a decrease
(increase) in ProG by about 7% when the KG value is over- (under-)
estimated by about 10%. However, NProF is insensitive to changes
in NKG (cf. Fig. 4(b)) as discussed earlier in this paper (cf. Fig. 3(a)).
Similarly, Fig. 5 illustrates the effect of NKG on the purity of glucose
(NPurG) and fructose (NPurF) respectively. While NPurG is insensi-
tive to variations in NKG (cf. Fig. 5(a)), NPurF is significantly affected
by variations in NKG (cf. Fig. 5(b)). A decrease (increase) in PurF by
about 9% occurs when KG is over- (under-) estimated by about 10%.
The effects of normalized deviation in KF (or NKF ) on the normal-
ized deviations in the productivity of glucose (NProG) and fructose
(NProF) are illustrated in Fig. 6(a) and (b). Again, that NProG is
insensitive to changes in NKF is clearly shown in Fig. 6(a) and the
drastic changes in NProF with NKF are explicated in Fig. 6(b). ProF
decreases (increases) by about 7% when KF is over- (under-) esti-
mated by about 10%. Furthermore, that NKF has very less influence
on PurG and PurF is illustrated in Fig. 7. Clearly, there is a gradual
change in NPurG with NKF as shown in Fig. 7(a) and there is almost
no dependence of NPurF on NKF (cf. Fig. 7(b)).
Furthermore, the dependence of the output (deviation) vari-
ables on the input (deviation) variables is analyzed via regression
analysis. Using the method of least squares, simple empirical mod-
els are obtained to predict the deviation in the performance of Fig. 6. Effect of variation in NKF on the normalized productivity of glucose and
SMB system (NProGp , NProFp , NPurGp , and NPurFp ) as a function fructose.
 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589
Fig. 7. Effect of variation in NKF on the normalized purity of glucose and fructose.
of the deviations in the process parameters (NKG and NKF ). These
empirical model equations are as follows:
NProGp = 9.96 × 10−3 − 0.73NKG − 2.65NK2G
NProFp = 4.60 × 10−3 − 0.70NKF − 1.13NK2F
−3
NPurGp = 5.89 × 10 − 0.11NKG − 0.18NKF
−0.06NKG NKF − 0.52NK2G − 0.98NK2F
NPurFp = 8.74 × 10−3 − 0.77NKG − 0.06NKF −
0.41NKG NKF − 2.26NK2G − 0.12NK2F
Eqs. (12)–(15) clearly illustrate the fact that while NProG and
NProF are relatively simple functions of NKG and NKF respectively,
NPurG and NPurF have a rather very complicated functional form
with nonlinear and interaction effects due to NKG and NKF . The
comparisons of the deviations in productivity of glucose (NProG)
and fructose (NProF) that are obtained from the Monte Carlo simu-
lations with the predictions from Eq. (12) (NProGp ) and Eq. (13)
(NProFp ) respectively are illustrated in Fig. 8(a) and (b). Clearly,
there is an excellent agreement between the values obtained
from Monte Carlo uncertainty analysis and predictions (R2 = 0.999).
Also that NProG and NProF are strongly dependent on NKG (cf.
Fig. 4(a)) and NKF (cf. Fig. 6(b)) respectively is clearly reflected
in the functional form of Eqs. (12) and (13). The comparisons of
the deviations in purity of glucose (NPurG) and fructose (NPurF)
that are obtained from the Monte Carlo sensitivity study with the
predictions from Eq. (14) (NPurGp ) and Eq. (15) (NPurFp ) respec-
tively are illustrated in Fig. 9(a) and (b). Again the agreement
587
(12) Fig. 8. Predictive models for the normalized productivity of glucose and fructose.
(13)
between the Monte Carlo simulations and the model predictions
is remarkably good (R2 = 0.999). It is worth noting that the adsorp-
tion isotherms of glucose and fructose are linear, because of which
(14) the Monte Carlo error propagation analysis is not that cumber-
some. Naturally, one would expect the analysis to be quite complex
when the adsorption isotherms are nonlinear. A brief discussion
of how complex the uncertainty analysis could get when the
(15) adsorption isotherms are nonlinear is presented in the next sub-
section.
3.3. Error propagation in industrial-scale nonlinear system
(xylene isomers separation)
In the case of the SMB system for the separation of xylene
isomers [15,18], the operating conditions of the pilot plant are
as indicated in Table 2. The adsorption of the xylene isomers
is described by a multi-component Langmuir (nonlinear) model
[15,18]. Under these conditions, the output variables, viz., the purity
(PurX) and recovery (RecX) of p-xylene are 99.48% and 95.81%
respectively. Just like the first case study, in the Monte Carlo-based
error propagation analysis of the SMB system for the xylenes sep-
aration, the process parameters (input variables) namely KpX , KEB ,
and KpDEB are randomly sampled from a uniform distribution with
a range of ±10% of the adsorption constants (cf. Table 2). All the
other process parameters are fixed constant at the values listed in
Table 2.
588 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589

via regression analysis. Two empirical models are obtained to pre-

dict the deviation in the performance of SMB system (NRecXp and
NPurXp ) as a function of the deviations in the process parameters
(NKpX , NKEB , and NKpDEB ) as shown:

NRecXp = − 1.4 × 10−3 − 0.058NKpX

+0.125NKEB − 0.121NKpDEB (16)

NPurXp = 1.94 × 10−3 + 0.036NKpX − 0.077NKEB

+0.019NKpDEB − 0.522NK2pX − 0.60NKpX NKEB

+0.675NKEB NKpDEB (17)

From Eqs. (16) and (17), it is clear that while the recovery of
p-xylene (NRecX) has a linear relationship with the adsorption con-
stants, the purity of p-xylene (NPurX) is a very complex function of
the adsorption constants, which may be attributed to the nonlin-
ear adsorption equilibria of the xylene isomers. The comparisons of
the deviations in recovery (NRecX) and purity (NPurX) of p-xylene

Fig. 9. Predictive models for the normalized purity of glucose and fructose.

Table 2
SMB operating conditions for the separation of xylene isomers [15,18]

SMB unit geometry Operating conditions Model parameters

Ncol = 24 T = 453 K, liquid phase Pe = 2000

Lcol = 1.135 m ts = 1.16 min
dcol = 4.117 m QF = 1.45 m3 /min
 = 6/9/6/3 QEx = 1.65 m3 /min
QRa = 2.69 m3 /min
QD = 2.89 m3 /min
QS = 5.39 m3 /min
˛ = 2.32
dp = 9.2 × 10−4 m
ε = 0.39
KpX = 1.0658 m3 /kg
KmX = 0.2299 m3 /kg
KoX = 0.1884 m3 /kg
KEB = 0.3037 m3 /kg
KpDEB = 1.2935 m3 /kg

Furthermore, the input and output variables are transformed

into normalized deviation variables. The mean values of KpX ,
KEB , and KpDEB (1.0705, 0.4125, and 1.2918 respectively) from the
random sampling of a uniform distribution, viz., ±10% of the exper-
imentally estimated KpX , KEB , and KpDEB , are used to normalize the
deviations of the input variables from their mean. These normalized
deviations of KpX , KEB , and KpDEB about the mean are henceforward
referred to as NKpX , NKEB , and NKpDEB respectively. Similarly, the
mean values of the PurX and RecX (99.15% and 96.43% respectively)
are used to normalize the deviations of the output variables about
their mean and these normalized deviations are designated as
NPurX and NRecX respectively. Then, the dependence of the output
(deviation) variables on the input (deviation) variables is analyzed Fig. 10. Predictive models for the normalized recovery and purity of p-xylene.
 A.S. Kurup et al. / Separation and Purification Technology 62 (2008) 582–589
that are obtained from the Monte Carlo simulations with the pre-
dictions from Eq. (16) (NRecXp ) and Eq. (17) (NPurXp ) respectively
are illustrated in Fig. 10(a) and (b). Clearly, there is a reasonably
good agreement between the values obtained from Monte Carlo
uncertainty analysis and predictions. However, it should be noted
that while the fit is excellent for linear isotherms (cf. Figs. 8 and 9)
with R2 = 0.999 for both normalized deviations in productivities and
purities of glucose and fructose, in the case of nonlinear isotherms
the best fits for the normalized deviations in recovery and purity
of p-xylene that could be obtained are with R2 = 0.88 and 0.73
respectively (cf. Fig. 10). This clearly indicates how drastically the
performance parameters could get affected when there are uncer-
tainties in the measurement of nonlinear adsorption isotherms.
Nevertheless, the Monte Carlo uncertainty analysis serves as a
reasonably simple tool to study error propagation even in such
complex separation systems.
4. Conclusion and future work
Sensitivity and uncertainty analyses of chemical processes are
very essential given the competitiveness of the current global mar-
ket [1–3]. In this paper, a Monte Carlo-based uncertainty analysis
for the performance assessment of a Simulated Moving Bed system,
which is widely employed in petrochemicals, sugar, and phar-
maceutical industries, has been presented. The error propagation
analysis has been carried out to assess the performance variability
of the SMB system due to errors in the estimation of adsorption
isotherm parameters for two case studies: (a) glucose–fructose
separation in a pilot-scale SMB unit [11,14] and (b) separation of p-
xylene from xylene isomers in an industrial-scale SMB unit [15,18].
Simple empirical models have also been developed via regression
analysis of the parametric sensitivity study to predict the varia-
tions in the performance of the SMB as a function of the isotherm
parameters.
One could also envision performing error propagation analysis
by other techniques such as the so-called “Differential Analysis” [2]
wherein Taylor series approximation of models can be developed
to understand the effect of perturbations in the input variables on
the output variables. For instance, the novel analytical solutions for
linear adsorption isotherms in the presence of mass transfer resis-
tance [28] can be employed in a differential analysis to perform
uncertainty analysis of the glucose–fructose separation [11,14].
However, it should be noted that evaluating the partial derivatives
in the Taylor series approximation of the models is a demanding
task [2]. Also, for nonlinear systems like xylene separation [15,18],
obtaining the analytical solutions in the first place is very dif-
ficult let alone performing a differential uncertainty analysis. In
such cases, Monte Carlo-based uncertainty analysis can readily be
employed.
In summary, the Monte Carlo-based error propagation analysis
has been demonstrated, in this paper, to be a robust and success-
ful technique to carry out sensitivity and uncertainty analyses of
complex chemical processes such as the two industrial-scale SMB
systems reported in this paper for the separation of sugars and
xylene isomers respectively. The Monte Carlo-based methodology
is so robust and relatively straight-forward that its implemen-
tation for the error propagation analysis of multi-functional or
hybrid chemical processes (e.g., reactive SMB, reactive distilla-
tion, etc.) could readily be envisioned as future research direc-
tions.
589
References
[1] V.R. Vasquez, W.B. Whiting, Accounting for both random errors and system-
atic erros in uncertainty propagation analysis of computer models involving
experimental measurements with Monte Carlo methods, Risk Anal. 25 (2006)
1669–1681.
[2] J.C. Helton, Uncertainty and sensitivity analysis techniques for use in perfor-
mance assessment for radioactive waste disposal, Reliab. Eng. Syst. Saf. 42
(1993) 327–367.
[3] O.A. Guell, J.A. Holcombe, Analytical applications of Monte Carlo techniques,
Anal. Chem. 62 (1990) 529A–542A.
[4] P. Jandera, S. Buncekova, K. Mihlbachler, G. Guiochon, V. Backovska, J. Planeta,
Fitting adsorption isotherms to the distribution data determined using packed
micro-columns for high-performance liquid chromatography, J. Chromatogr. A
925 (2001) 19–29.
[5] H.J. Subramani, A.S. Kurup, Micro-Simulated Moving Bed (␮SMB) systems: a
numerical study, Chem. Eng. J. 120 (2006) 169–179.
[6] D.B. Hibbert, J. Jiang, M.I. Mulholland, Propagation of uncertainty in high-
performance liquid chromatography with UV–vis detection, Anal. Chim. Acta
443 (2001) 205–214.
[7] E. Hund, D.L. Massart, J. Smeyers-Verbeke, Comparison of different approaches
to estimate the uncertainty of a liquid chromatographic assay, Anal. Chim. Acta
480 (2003) 39–52.
[8] N.M. Faber, Quantifying the effect of measurement errors on the uncertainty
in bilinear model predictions: a small simulation study, Anal. Chim. Acta 439
(2001) 193–201.
[9] B. Spangenberg, K.F. Klein, J. Mannhardt, Proposals for error-reduction in planar
chromatography, J. Planar Chromatogr. Mod. TLC 15 (2002) 204–209.
[10] D.M. Ruthven, C.B. Ching, Countercurrent and simulated countercurrent
adsorption separation processes, Chem. Eng. Sci. 44 (1989) 1011–1038.
[11] H.J. Subramani, K. Hidajat, A.K. Ray, Optimization of Simulated Moving Bed
and Varicol processes for glucose–fructose separation, Chem. Eng. Res. Des. 81
(2003) 549–567.
[12] A.S. Kurup, H.J. Subramani, K. Hidajat, A.K. Ray, Optimal design and operation
of SMB bioreactor for sucrose inversion, Chem. Eng. J. 108 (2005) 19–33.
[13] K. Hashimoto, S. Adachi, H. Noujima, H. Maruyama, Models for separation of
glucose–fructose mixture using a simulated moving bed adsorber, J. Chem. Eng.
Jpn. 16 (1983) 400–406.
[14] D.C.S. Azevedo, A.E. Rodrigues, Fructose–glucose separation in a SMB pilot unit:
modelling, simulation, design and operation, AIChE J. 47 (2001) 2042–2051.
[15] A.S. Kurup, K. Hidajat, A.K. Ray, Optimal operation of an industrial-scale Parex
process for the recovery of p-xylene from a mixture of C8 aromatics, Ind. Eng.
Chem. Res. 44 (2005) 5703–5714.
[16] A.S. Kurup, K. Hidajat, A.K. Ray, Comparative study of modified Simulated Mov-
ing Bed systems at optimal conditions for the separation of ternary mixtures
of xylene isomers, Ind. Eng. Chem. Res. 45 (2006) 6251–6265.
[17] F. Wongso, K. Hidajat, A.K. Ray, Optimal operating mode for enantioseparation
of SB-553261 racemate based on SMB technology, Biotechnol. Bioeng. 87 (2004)
704–722.
[18] M. Minceva, A.E. Rodrigues, Modeling and simulation of a simulated mov-
ing bed for the separation of p-xylene, Ind. Eng. Chem. Res. 41 (2002) 3454–
3461.
[19] M. Mazzotti, G. Storti, M. Morbidelli, Robust design of countercurrent adsorp-
tion separation. 3. Nonstoichiometric systems, AIChE J. 42 (1996) 2784–
2796.
[20] M. Mazzotti, G. Storti, M. Morbidelli, Optimal peration of simulated moving bed
units for nonlinear chromatographic separations, J. Chromatogr. A 769 (1997)
3–24.
[21] Z. Ma, N.H.L. Wang, Standing wave analysis of SMB chromatography: linear
systems, AIChE J. 43 (1997) 2488–2508.
[22] D.C.S. Azevedo, A.E. Rodrigues, Design of a simulated moving bed in the pres-
ence of mass transfer resistances, AIChE J. 45 (1999) 956–966.
[23] J. Strube, A. Jupke, A. Epping, H. Schmidt-Traub, M. Schulte, R. Devant, Design,
optimization and operation of SMB chromatography in the production of enan-
tiometrically pure pharmaceuticals, Chirality 11 (1999) 440–450.
[24] C. Heurer, E. Kusters, T. Plattner, A. Seidel-Morgenstern, Design of the simu-
lated moving bed process based on adsorption isotherm measurements using
a perturbation method, J. Chromatogr. A 827 (1998) 175–191.
[25] G. Dunnebier, K.U. Klatt, Optimal operation of simulated moving bed chromato-
graphic processes, Comput. Chem. Eng. 23 (1999) S195–S198.
[26] H.J. Subramani, K. Hidajat, A.K. Ray, Optimization of reactive SMB and Varicol
systems, Comput. Chem. Eng. 27 (2003) 1883–1901.
[27] W. Yu, K. Hidajat, A.K. Ray, Modeling, simulation and experimental study of a
simulated moving bed reactor for the synthesis of methyl acetate ester, Ind.
Eng. Chem. Res. 42 (2003) 6743–6754.
[28] V.M.T. Silva, M. Minceva, A.E. Rodrigues, Novel analytical solution for a simu-
lated moving bed in the presence of mass-transfer resistance, Ind. Eng. Chem.
Res. 43 (2004) 4494–4502.