You are on page 1of 13

Available online at www.sciencedirect.

com

Computers and Chemical Engineering 32 (2008) 2243–2255

Non-equilibrium stage modeling and non-linear dynamic effects in the


synthesis of TAME by reactive distillation
Amit M. Katariya, Ravindra S. Kamath, Kannan M. Moudgalya, Sanjay M. Mahajani ∗
Department of Chemical Engineering, Indian Institute of Technology Bombay, Mumbai 400076, India
Received 29 November 2006; received in revised form 28 October 2007; accepted 4 November 2007
Available online 19 November 2007

Abstract
Tertiary-amyl methyl ether (TAME) is a potential gasoline additive that can be advantageously synthesized using the reactive distillation (RD)
technology. This work emphasizes on non-linear effects in dynamic simulations of reactive distillation column. For certain configurations, dynamic
simulation with equilibrium stage (EQ) model leads to sustained oscillations (limit cycles) which have been reported in our earlier work [Katariya,
A. M., Moudgalya, K. M., & Mahajani, S. M. (2006). Nonlinear dynamic effects in reactive distillation for synthesis of TAME. Industrial and
Engineering Chemistry Research, 45 (12), 4233–4242]. Feed condition and Damkohler number are the important parameters that influence the
existence of these effects. To confirm the authenticity of the observed non-linear behaviors, a more realistic and rigorous dynamic NEQ model for
a packed column is developed which uses a consistent hardware design. The steady state behavior of the NEQ model is examined by varying the
number of segments and the column height. The dynamic simulation and the bifurcation study with stability analysis indicate that the parameter
space, in which oscillations may be observed, is shifted in the case of NEQ model.
© 2007 Elsevier Ltd. All rights reserved.

Keywords: Reactive distillation; Dynamic simulation; Continuation analysis; Non-equilibrium model; Hopf bifurcation; Oscillations

1. Introduction dealing with departures from equilibrium in multistage towers


is through the use of stage and/or overall efficiencies or use of
Computer-aided design and simulation of multi-component height equivalent to a theoretical plate (HETP) in case of packed
multistage separation processes such as distillation, gas absorp- towers. Though, this may be a useful approach for simulating
tion and reactive distillation are important aspects of modern an existing column for which there is a good deal of data avail-
chemical engineering. Currently, such simulations are based able, it may not be possible to predict safely how the column
on the well-known equilibrium stage model. The EQ model will perform under quite different operating conditions (Baur,
assumes that the vapor and liquid leaving a stage are in equi- Higler, Taylor, & Krishna, 2000). Furthermore, it is difficult to
librium. Equilibrium stage simulations are frequently termed use this approach to simulate new processes in the design stage
rigorous, but this appellation is not entirely justified because in for which no plant data exists.
actual operation, columns rarely, if ever, operate at equilibrium. It is advantageous to use NEQ model over the EQ model due
The degree of separation is, in fact, determined as much by to some of the following reasons. It eliminates the need for effi-
mass and energy transfer between the phases being contacted ciencies and HETPs. The operating strategies for the influence
on a tray or within sections of a packed column, as it is by of chemical reactions on separations can be accounted in a bet-
thermodynamic equilibrium considerations. The usual way of ter way. The over-design or under-design can also be avoided
as the tray and packed columns are modeled with greater accu-
racy thereby reducing the capital and operating costs. Also, as
Abbreviations: RD, reactive distillation; EQ, equilibrium stage modeling; mentioned before the NEQ model is more realistic as compared
NEQ, non-equilibrium stage modeling; MeOH, methanol; 2M1B, 2-methyl 1- to the EQ model and represents a more accurate modeling of
butene; 2M2B, 2-methyl 2-butene; TAME, tertiary-amyl methyl ether; i-PENT,
iso-pentane; Da, Damkohler number; DAE, differential algebraic equation.
reactive systems.
∗ Corresponding author. Tel.: +91 22 25767246; fax: +91 22 25726895. The non-equilibrium (NEQ) model assumes that the
E-mail address: sanjaym@che.iitb.ac.in (S.M. Mahajani). vapor–liquid equilibrium is established only at the interface

0098-1354/$ – see front matter © 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2007.11.009
2244 A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255

between the bulk liquid and vapor phases, and employs a


Nomenclature transport-based approach to predict the flux of mass and energy
across the interface. Various authors have presented steady state
ai activity coefficient of ith component
non-equilibrium stage models for tray (Baur, Higler, et al., 2000;
Ac interfacial area for vapor–liquid mass transfer
Baur, Krishna, & Taylor, 2003; Baur, Taylor, & Krishna, 2000;
(m2 )
L L Higler, Taylor, & Krishna, 1999;) as well as packed (Asprion,
Ctk , Ctk total concentration of liquid and vapor phase on
2006; Jakobsson, Hasanen, & Aittamaa, 2004; Peng, Lextrait,
k th stage (mol/m3 )
Edgar, & Eldridge, 2002; Sundmacher & Hoffmann, 1996) reac-
D, B molar flow rates of distillate and bottom
tive distillation columns. For the purpose of design, optimal
EkL , EkV liquid and vapor energy holdup of k th stage (J)
operation, and the control of the reactive distillation process,
hLtk , hV
tk liquid and vapor side heat transfer coefficients for a rigorous theoretical dynamic model is required. The modeling
k th stage (W/m2 K)
and simulation with NEQ model is a computationally rigor-
Hk , hk , hfk molar enthalpy of vapor, liquid and feed for k
ous activity as it involves large number of highly non-linear
th stage (J/mol)
L , H̃ V partial molar enthalpy liquid and vapor for ith equations like pressure drop correlations, packing holdup corre-
H̃i,k i,k lations, etc. Hence, there are very few publications on dynamic
component and k th stage (J/mol)
simulation of tray (Baur, Taylor, & Krishna, 2001; Schenk,
kfm forward rate constant of m th reaction
Gani, Bogle, & Pistikopoulos, 1999) and packed (Kreul, Gorak,
(mol/equiv. s)
Dittrich, & Barton, 1998; Noeres, Dadhe, Gesthuisen, Engell,
kL , kV liquid and vapor mass transfer coefficient matri-
& Grak, 2004; Peng, Edgar, & Eldridge, 2003; Xu, Flora, &
ces (m/s)
Rempel, 2005) reactive distillation columns using NEQ model.
Kam equilibrium constant of m th reaction
They differ in the way the mass and heat transfer resistances
KIi,k interface vapor–liquid equilibrium constant of ith
are incorporated in the model. The main differences are (1)
component and k th stage
the use of driving force for the mass transfer: some use con-
MkL , MkV liquid and Vapor molar holdup of k th stage (mol)
centration gradient, whereas others use the correct gradients of
NLk , NV k vectors of liquid and vapor mass transfer fluxes chemical potentials and fugacities; (2) the diffusivity models:
of the order, C − 1 (mol/m2 s) Fick’s law or Stefan–Maxwell approach; (3) number of phases
L V
Ni,k , Ni,k liquid and vapor mass transfer fluxes for ith involved: two phase or pseudo homogeneous model and three
component and k th stage (mol/m2 s) phase heterogeneous model.
Qk heat loss from k th stage (J/s) TAME, a popular fuel additive, is commercially produced
Qr , Qc reboiler and condenser duty (J/s) by reactive distillation through the reaction of methanol with
R reflux ratio isoamylene coming from C5-stream of the refinery. It is a widely
Rm,k rate of m th reaction and k th stage (mol/s) studied model system to understand the complex behavior of
SkV , SkL molar flow of vapor and liquid side streams of k reactive distillation. A few case studies of TAME synthesis in
th stage (mol/s) RD using both EQ and NEQ models have appeared in the liter-
T L , T V , T I liquid, vapor and interface temperatures (K) ature (Baur, Taylor, et al., 2000; Baur et al., 2003; Katariya,
Vk , Lk , Fk molar flow rates of vapor, liquid and feed of k Kamath, Moudgalya, & Mahajani, 2006; Mohl et al., 1999;
th stage, respectively (mol/s) Ouni, Jakobsson, Pyhalahti, & Aittamaa, 2004; Peng et al., 2003;
W weight of the catalyst (kg or equiv.) Subawalla & Fair, 1999). Most of these except that by Peng et al.
xD , xB mole fraction of distillate and bottom (2003) are restricted to steady state analysis. Peng et al. (2003)
xfi,k , yfi,k liquid and vapor feed mole fraction of ith com- have compared the dynamic rate-based and equilibrium models
ponent and k th stage for a packed reactive distillation column for the production of
xi,k , yi,k liquid and vapor mole fractions of ith component tert-amyl methyl ether (TAME) and proposed a new approach
and k th stage to simplify the dynamic rate-based model by assuming the mass
xIi,k , yIi,k interphase liquid and vapor mole fractions of ith transfer coefficients to be time invariant. It can reduce the num-
component and k th stage ber of equations by up to two-third and still accurately predict
the dynamic behavior. A high-index problem in the models may
Greek letters arise if the pressure drop is not related to vapor and liquid flow
δ dirac-delta function rates (Kreul et al., 1998).
k volume or weight of the catalyst for k th stage Synthesis of TAME by reactive distillation is known to exhibit
γi activity coefficient of ith component non-linear dynamic effects such as multiple steady states and
γi,m stoichiometric coefficient of ith component and m relevant literature is reviewed in our earlier work (Katariya,
th reaction Moudgalya, & Mahajani, 2006). We showed for the first time
that under certain conditions, the EQ model based dynamic sim-
ulation of the reactive distillation column exhibits another type
of non-linear effect, i.e. sustained oscillations or limit cycles
(Katariya, Moudgalya, et al., 2006). In order to further examine
the authenticity of this observed non-linear dynamic effect, here
A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255 2245

design information such as column diameter, tray or packing


we present a rigorous dynamic non-equilibrium model for the type and geometry, etc., is mandatory. A packed column has
synthesis of TAME in packed RD columns. The model includes been selected for the NEQ simulations. Each continuous section
all the essential dynamic terms comprising vapor and liquid of the packed column is divided into a number of segments,
holdups. Since the TAME synthesis is carried out at high pres- each of which acts as a non-equilibrium stage. The packing
sure (4.5 bar), it is important to consider the vapor mass and selected for the reactive and non-reactive rectifying and strip-
energy holdups in the modeling equations, which are otherwise ping sections are KATAPAK-S and Sulzer-BX, respectively. The
neglected in the earlier studies due to index issues (Peng et al., hydraulic and mass transfer correlations for the selected pack-
2003) and low pressure operations (Noeres et al., 2004). Also, ing are obtained from Kolodziej, Jaroszynski, and Bylica (2004)
time variant mass and heat transfer coefficients are considered in and Rocha, Bravo, and Fair (1993, 1996), respectively, and are
our model, which are made time invariant in the earlier studies given in Appendix A.
due to computational difficulties and simulation time. We also The preliminary column design for the NEQ model is derived
present a comparative study of steady state and dynamic simu- from the steady state results of the EQ model. The column diame-
lations using both EQ and NEQ models. Detailed index analysis ter is estimated by applying the fractional approach to flooding.
of the NEQ model is carried out and the variables responsible The height of each packed section is calculated by multiply-
for the higher index in each case are identified and accord- ing the HETP with the corresponding number of theoretical
ingly, model simplifications are made without compromising stages. The hardware design for the selected conceptual column
on essential dynamic terms. The work by Reepmeyer, Repke, configuration is shown in Fig. 1.
and Wozny (2004) may be referred to understand and handle
some of the numerical issues involved in the dynamic simula-
tions. Also to systematically investigate the non-linear dynamics 2.1. Kinetics and thermodynamics
of the system, bifurcation behavior of the simplified NEQ
model with stability analysis has been carried out in some cases The following three reactions have been considered while
which, to the best of our knowledge, has not been reported till modeling the process for the synthesis of TAME, which includes
date. two synthesis reactions for TAME from the isomers of isoamy-
lene and one isomerization reaction. The side reactions such
2. Model description and hardware specification as dimerization of methanol and formation of TAA have been
neglected. The rate equations are as given below. The tempera-
A rigorous NEQ model has been developed to examine the ture dependent rate constants and the equilibrium constants for
effect of column hardware and heat and mass transfer resistances the reactions are obtained from Rihko and Krause (1995) and
on the non-linear behavior of the RD column. The purpose is to Faisal, Syed, and Datta (2000).
compare the performance and behavior with that obtained by
the EQ model in our earlier studies (Katariya, Moudgalya, et MeOH + 2M1B ⇔ TAME,
al., 2006). We refer to Powers, Vickery, Arehole, and Taylor  
(1988) for detailed model implementation and computational a2M1B 1 aTAME
R1 = kf 1 − 2
aspects. In case of NEQ models, the specification of hardware aMeOH Ka1 aMeOH

Fig. 1. The conceptual column configuration used for EQ and NEQ simulations along with the hardware design derived from it.
2246 A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255

MeOH + 2M2B ⇔ TAME, and liquid streams are not identical. Condenser and reboiler are
  treated as equilibrium stages.
a2M2B 1 aTAME Total material balance equation for the NEQ stage are as
R2 = kf 2 − 2
aMeOH Ka2 aMeOH
dMkL  C
= Lk−1 + FkL − (Lk + SkL ) + Ac Ni,k
L
  dt
a2M1B 1 a2M2B i=1
2M1B ⇔ 2M2B, R3 = kf 3 −
aMeOH Ka3 aMeOH 
r 
C
+ γi,m Rm,k k (1)
As the system consists of mixture of polar and non-polar
m=1 i=1
components, it is highly non-ideal and the use of activity based
kinetics and thermodynamics is justifiable. The UNIQUAC dMkV  C

model has been used for describing non-ideality of the liquid = Vk+1 + FkV − (Vk + SkV ) − Ac Ni,k
V
(2)
dt
i=1
phase, with binary interaction parameters taken from HYSYS.
All the thermodynamic and kinetic parameters used in the Component material balance equation are written as
study have been also reported in our earlier work (Katariya,
dMkL xi,k
Moudgalya, et al., 2006). The process design of the column and = Lk−1 xi,k−1 + FkL xfi,k − (Lk + SkL )xi,k + Ac Ni,k
L
the input conditions have been obtained from Subawalla and Fair dt
(1999). Fig. 1 shows the column configuration along with oper- r

ating and design parameters used for the study. Here, methanol + γi,m Rm,k k (3)
is fed in excess, which is required to form a minimum boiling m=1
azeotrope with inerts (e.g. isopentane) and separate them effi- dMkV yi,k
= Vk+1 yi,k+1 + FkV yfi,k − (Vk + SkV )yi,k − Ac Ni,k
V
ciently from the top of the column. Escess methanol also helps dt
to maintain the desired temperature (330–350 K) in the reactive (4)
section of the column. Also, Subawalla and Fair (1999) have
observed in their analysis that if the methanol used is less than where Ac is the interfacial area for vapor–liquid mass trans-
V and N L are vapor and liquid mass transfer fluxes
fer and Ni,k
the amount required to form an azeotrope then the conversion i,k
of amylene and the purity of the TAME are adversely affected. respectively. Only (C − 1) component material balance equa-
tions are independent, summation constraint on vapor and liquid
phase compositions is used to get the composition of the remain-
2.2. Model equations
ing components.
Energy balance equation:
A schematic representation of the NEQ stage is shown in
Fig. 2. This NEQ stage may represent a tray or a cross-section dEkL
of a packed column. The stage equations are the traditional equa- = Lk−1 hk−1 + FkL hfk − (Lk + SkL )hk − Qk
dt
tions for mass and energy balances for individual phase, in which  
mass and heat transfer rates are also included. Bulk variables C
+ Ac htk (Tk − Tk ) +
L I L L L
Ni,k H̃i,k (5)
(compositions, flow rates, molar fluxes, energy fluxes and tem-
i=1
peratures) are different from the interface variables. Equilibrium
is assumed to be only at the interface and temperatures of vapor dEkV
= Vk+1 Hk+1 + FkV Hfk − (Vk + SkV )Hk
dt
 
C
− Ac htk (Tk − Tk ) +
V V I V V
Ni,k H̃i,k (6)
i=1
V and H̃ V are partial molar enthalpies of vapor and
where H̃i,k i,k
liquid. Vapor–liquid equilibrium at interface can be as given
below.
yIi,k = KIi,k xIi,k (7)
Mass and energy conservation equations for interface can be
written as below. It is assumed that reaction does not take place
in the liquid film.
V
Ni,k = Ni,k
L
(8)

C 
C
hLtk (TkI − TkL ) + L
Ni,k L
H̃i,k = hV
tk (Tk − Tk ) +
V I V V
Ni,k H̃i,k
i=1 i=1
(9)
Fig. 2. The typical NEQ stage representing tray or section of packed column.
A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255 2247

Summation constraints for the mole fractions in bulk vapor given in Appendix A. Binary diffusion coefficients in the corre-
and liquid as well as at the interface are written as lations are calculated using method given by Wilke and Chang
(1955) for liquid phase and correlation given by Fuller, Schettler,

C
and Diddings (1966) for gas phase. Maxwell–Stefan diffusivi-
xi,k = 1.0 (10)
ties are derived from these infinite dilution diffusivities (Dijo )
i=1
using Eq. (20).

C
yi,k = 1.0 (11) Dij = (Dijo )(1+xj −xi )/2 (Dji
o (1+xi −xj )/2
) (20)
i=1
The matrix of the thermodynamic factor, [Γ ] is calculated using

C following relation:
xIi,k = 1.0 (12)
∂ ln γi
i=1 Γi,j = δi,j + xi (21)
∂xj

C
yIi,k = 1.0 (13) Damkohler number is a key parameter which is the ratio of
i=1 characteristic residence time to characteristic reaction time. In
the present work Damkohler number is defined based on the total
Fick’s law approach described earlier (Peng et al., 2003) is
feed to the column and total amount of the catalyst used in the
used to calculate the mass transfer fluxes.
column. Boiling point of the lowest boiling component is used

C as the reference temperature.
NLk = CtjL kV (xIk − xk ) + xk L
Ni,k (14)
WT kf ,ref
i=1 Da = (22)
Ftotal

C

k = Ctj k (yk − yIk ) + yk


NV V L V
Ni,k (15) 2.3. Steady state analysis
i=1

where NLk and NV Steady state simulations with the help of developed NEQ
k are the vectors of mass transfer fluxes of
the order (C − 1) for each stage. Only (C − 1) mass fluxes are model are carried out for the design and operating parameters
independent, summation equation of the interface mole fractions given in Fig. 1. This step is mandatory for getting the initial
are used to find the mass flux of the last component. kL and kV steady state required for carrying out the dynamic simulations.
are the mass transfer matrices of order (C − 1) × (C − 1) for Same design and operating parameters as in EQ stage simu-
each stage. lations (Katariya, Moudgalya, et al., 2006) have been used to
We used a method suggested by Krishna and Standart (1976) compare the behavior (P = 4.5 bar, Qreb = 20.5 MW, R =
which involves relating [k∗ ] to the binary pairs of mass transfer 1.5). Pure methanol is fed at the bottom of the reactive zone
coefficients through solution of Maxwell–Stefan equations for whereas the pre-reacted feed was supplied at the midpoint of
film model. Matrices are calculated using following relations non-reactive stripping section.
with assumption that the matrices accounting the influence of The NEQ model equations are implemented in large scale
mass transfer rates on the mass transfer coefficients are identity. equation oriented simulator DIVA (Kroner, Holl, Marquardt, &
Gilles, 1990). DIVA uses the equation-oriented approach for
−1
[kV ] = [BV ] (16) solving all the differential and algebraic equations simultane-
ously. This comes with an inbuilt package for continuation and
L −1
[kV ] = [B ] [Γ L ] (17) stability analysis for the DAEs systems.
Initially the column height in NEQ model was divided in the
The elements of the matrix [B] have been calculated using same number of slices as the number of equilibrium stages, i.e.
the following equations. 33 (4 slices in non-reactive rectifying, 19 slices in reactive sec-
 C tion and 10 slices in non-reactive stripping section). Following
zi zk
Bii = + (18) attempts have been made to arrive at the initial steady state of
κi,C κi,k the NEQ simulations which is required for starting the dynamic
k=1
simulation.
k = i
  1. Steady state simulation of NEQ model: the results from
1 1
Bij = zi − (19) equilibrium stage model were used as initial guesses to the
κi,j κi,C
non-linear algebraic equation solver. The guess values for
where zi is the mole fraction of vapor or liquid phase and κi,j is bulk and interface variables were assumed to be same. Con-
the mass transfer coefficient of the binary pair in an appropriate vergence failed in this case.
phase. The packing selected for the non-reactive and the reactive 2. Steady state simulation of NEQ model with infinite mass and
sections are Suzler-BX and KATAPAK-S, respectively. Corre- heat transfer coefficients: the model when solved with infi-
lations for calculating the binary mass transfer coefficients are nite mass and heat transfer coefficients, is equivalent to the
2248 A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255

EQ model. This model with the initial guesses same as in an stripping section increases, the NEQ profile in the stripping sec-
attempt one above is used for the simulation and a continua- tion moves in the direction towards the EQ profile, crosses it
tion approach was used to reach the finite values of mass and and continues to move away from it. Finally a stage is reached
heat transfer coefficients. In this case also the convergence when a further increase in the number of segments does not sig-
could not be obtained due to the non-linearity and interaction nificantly influence the composition profiles. Peng et al. (2002)
of the pressure drop and holdup equations. have correlated the effect of NEQ segments with the extent of
3. Dynamic simulation of rigorous NEQ model: the integration back-mixing. At very large number of segments, back-mixing in
of the rigorous dynamic model for relatively large time to liquid and vapor phases is virtually absent and there is no effect
arrive at the steady state has been carried out. This attempt of further change in the number of segments. Thus Fig. 3 shows
was also failed due to large number of stiff DAEs. that the number of slices, i.e. the extent of back-mixing in the
4. Dynamic simulation of constant holdup NEQ model: the packed columns strongly influences the composition profile. In
dynamic model was simplified with certain assumptions real columns, back-mixing and other non-ideal conditions can-
like constant molar holdup of liquid and negligible vapor not be eliminated and hence an appropriate number of segments
and energy holdup on each segment of the column and the should be used. However, this number cannot be determined a
required steady state was obtained. This steady state was priori. For steady state simulations it was observed that for the
then used in further simulations. This approach was found number of segments greater than 189 the composition profiles
to work well in most of the cases. The number of segments do not change significantly.
in each section of the column was increased such that there As discussed before, the objective of the present work is to
was no further change in the column profiles, conversion of confirm whether the oscillations observed in the EQ model pre-
isoamylene and purity of the TAME in the bottom. dictions still persist in the case of NEQ simulations. In other
words, we examine whether the consideration of mass and heat
2.4. Influence of number of segments transfer limitations would influence the presence of non-linear
dynamic behavior. The oscillations being a non-linear dynamic
The effect of the number of segments in the packed sections of effect, may originate from the nonlinearity in the vapor–liquid
the RD column on the steady state results using the NEQ model equilibrium relation, reaction kinetics or the functional depen-
is shown in Fig. 3. When the number of segments in a particular dence of the physical properties on the compositions and/or
section is chosen to be same as the number of corresponding temperature (Kienle & Marquardt, 2003). Hence, the probabil-
theoretical stages in the EQ model, a significant difference in ity of realizing oscillations with NEQ model will be more if we
the composition profile predicted by the two models is seen but work in the same region of composition and temperature space
only in the stripping section. As the number of segments in the for which the oscillations were observed in the case of the EQ

Fig. 3. Comparison of the steady state composition and temperature profiles along the height of the column for the EQ model and NEQ model with various number
of total segments.
A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255 2249

model predictions. As mentioned before, in the case of NEQ


predictions, the composition profiles, conversion of isoamylene
and TAME purity in the bottom are significantly different from
that obtained by the EQ model and are very sensitive to the
change in column height and number of segments. Hence, we
present here two different column designs as mentioned below.
Further we perform dynamic simulations and the bifurcation
analysis in some cases, to explore the possibility of the presence
of oscillations.

• In the first case, we vary the column height, especially that of


the stripping section, such that the composition and temper-
ature profiles and isoamylene conversion/TAME purity are
close to the EQ predictions. The number of segments used
here is such that the back mixing is absent (i.e. 189 segments)
and there is no further change in the composition profile with
increase in number of segments.
• In the second case, the number of segments, which repre- Fig. 4. Effect of height of the stripping section on the isoamylene conversion.
sent the extent of back-mixing in vapor and liquid phases, is
varied such that composition and temperature profiles of EQ The optimum stripping height in this case was found to be
and NEQ model match reasonably well. This is the case with about 1.62 m. A height of 1.98 m was selected for the new
partial back-mixing. design since it not only gives a conversion of isoamylene close
to the optimum but also the conversion and end composition
3. Case 1: NEQ model without back-mixing are very similar to that given by the earlier design with EQ
model.
3.1. Influence of height of the stripping section The steady state composition and temperature profiles of the
NEQ model using this new design are plotted along with that of
The steady state result using the NEQ model, with total num- the EQ model in Fig. 5. Even though the top and bottom com-
ber of slices 189, showed a much lower isoamylene conversion positions and temperature profiles are similar, certain sections
(67.6%) compared to 84.6% obtained in the EQ model. Since of the stripping zone show different compositions.
the primary objective was to compare the non-linear dynamic
effects of EQ and NEQ models, getting similar conversions 3.2. Dynamic simulation
and end compositions is essential. So, an attempt was made
to change the hardware design (diameter and heights of sec- For the NEQ model, the liquid and vapor flow rates in the
tions) estimated from the EQ model to a new design such that packed sections are not responsible for the higher index as alge-
EQ and NEQ models give similar results, and the composition braic equations for these variables in terms of pressure drop and
and temperature profiles roughly lie in the same domain. From holdup correlations are incorporated in the model. However,
the previous analysis, it is clear that the stripping zone plays the liquid and vapor flows associated with the condenser and
a crucial role in the column behavior. The effect of height of the reboiler can pose high-index problems as those are mod-
the stripping section on isoamylene conversion was investigated eled using equilibrium assumptions. One will arrive at index
using continuation analysis and is shown in Fig. 4. Surprisingly, two DAEs when the equations for the holdup as a function
conversion of isoamylene increases with decrease in height of of vapor and/or liquid flows, e.g. controller equations are not
the column. This is clearly a counter-intuitive effect since we explicitly considered in the model. If these equations are not
expect that a larger packed height should result in a better available (open loop column) then some of the differential equa-
separation and as per the principles of RD, a better separa- tions need to be converted to algebraic equations by neglecting
tion of the product TAME from the reactants should result in the dynamics to eliminate the index problem. It can be proved
enhanced amylene conversion. However, an optimum in conver- with the help of a detailed index analysis that at least the fol-
sion was observed beyond which conversion of amylene again lowing differential equations need to be converted to algebraic
decreases as height is decreased. This is because an increase equations:
in the number of stages in the stripping section results in bet-
ter separation of not only TAME but C5 olefins also, which 1. Energy balance for the condenser: this is because condenser
are the reactants. C5-olefins under otherwise similar conditions load does not appear in any other algebraic equation.
find the way out from the bottom thereby causing a reduc- 2. Energy balance for the reboiler: this is to account for reboiler
tion in their concentrations in the reactive zone. Hence, the duty, bottom flow rate or vapor flow from the reboiler depend-
overall isoamylene conversion decreases with an increase in ing upon the bottom specification.
the height of the stripping section. This particular effect has 3. Total material balance for the reboiler and condenser: this
also been confirmed through EQ stage simulations as well. is to account for either vapor or liquid flow.
2250 A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255

Fig. 5. Comparison of composition and temperature profile for the EQ and the NEQ model with the new design.

Apart from this rigorous model, we define constant holdups the NEQ models take almost equal computation times since the
dynamic NEQ model as the one in which differential equations total number of equations (differential and algebraic) is the same.
for all the total material and energy balances are converted to However, the rigorous NEQ model was much more difficult to
algebraic equations, as was done in the case of the EQ model converge for larger step changes because of the stiffness issues.
(Katariya, Moudgalya, et al., 2006). Starting from a steady state The convergence properties of the constant holdups NEQ model
with the same operating conditions, the dynamics of EQ model, were very similar to that of the EQ model with almost no conver-
rigorous NEQ model and the constant holdups NEQ model for a gence problem up to ±5% step changes in operating parameters.
2% step increase in the pre-reacted feed flow rate have been stud- The computation time for the NEQ models was observed to be
ied. As seen from Fig. 6, both the NEQ models show a slightly almost 15 times higher than that of the EQ model.
different dynamics but reach the same steady state while the Fig. 7 shows the response of the average values of liquid and
EQ model reaches a different steady state as expected. Both vapor side heat and mass transfer coefficients to a step increase
in the pre-reacted feed flow rate using rigorous NEQ model. It
has been seen that for very small changes in the feed there are
significant changes in the heat and mass transfer coefficients.
This justifies the fact that time variant heat and mass transfer
coefficients have to be considered while simulating the NEQ
model for reactive distillation.
To confirm the authenticity of the sustained oscillations
observed in the dynamic EQ model, similar analysis was
repeated with constant holdup NEQ model using the new hard-
ware design (i.e. 189 total number of segments, stripping section
height =2.43 m, reactive section height =8.41 m and rectifying
section height =2.98 m). As seen in Fig. 8, unlike the EQ model,
oscillations were not observed and the system always reaches the
corresponding steady state. Thus, the oscillatory behavior that
existed in the EQ model disappears in the NEQ model for the
desired isoamylene conversion and TAME purity in the bottom.
However, it must be noted that the parameter space wherein the
non-linear dynamic effects are observed in EQ model, is likely
Fig. 6. Dynamic response of EQ, rigorous and ‘constant holdups’ NEQ model to shift in the case of NEQ model simulations. Such a possibility
for a 2% step change in feed flow. can be ascertained only by studying the bifurcation behavior with
A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255 2251

Fig. 7. Dynamic response of (a) mass transfer coefficient and (b) heat transfer coefficients, in the rigorous NEQ model to a 2% step change in feed flow.

models are close to each other, with approximately same con-


version of isoamylene and the purity of TAME in the bottom.
Fig. 9 shows the steady state composition and temperature pro-
files when the column is divided in 47 segments (6 segments in
non-reactive rectifying section in 2.43 m height, 23 segments
in reactive section in 8.41 m height, and 18 segments in non-
reactive stripping section in 3.4 m height), with almost same
conversion of isoamylene and TAME purity in the bottoms. A
very good match with the base case EQ profiles is observed.
The response of the rigorous NEQ model with partial back-
mixing, constant holdups NEQ model with partial back-mixing
and constant holdup EQ model, to a change in 2-methyl-1-butene
concentration can be seen in Fig. 10. Significant differences
are observed in the responses. As mentioned before, sustained
oscillations are realized only in the case of EQ model. Whereas
oscillations disappear in both the NEQ models under similar
operating conditions.

4.1. Bifurcation analysis of NEQ model with partial


Fig. 8. Dynamic response for change in amylene feed composition for Da = 3.0
using the NEQ model with new design.
back-mixing

Comparison of the bifurcation diagrams for EQ and NEQ


respect to all possible parameters and their combinations, using models is shown in Fig. 11. Both the curves almost overlap
continuation method coupled with stability analysis, which is quantitatively but they have different stability behaviors. The
computationally an intensive task. EQ model shows unstable solution branch with the presence of
The presence of non-ideality in terms of the partial back- Hopf bifurcation whereas NEQ model under similar condition
mixing may also influence the column performance. Hence the shows the stable solution. This comparison shows that modeling
model with partial back-mixing has been considered in the next assumptions have a significant impact on the observed oscilla-
section for the realistic comparison of non-linear dynamics. tions in case of EQ model. The difference in the stability of the
two curves in Fig. 11 does not imply that the oscillations have
4. Case 2: NEQ model with partial back-mixing disappeared in NEQ model. Fig. 12 shows the bifurcation dia-
gram with reboiler duty as a parameter. The presence of the Hopf
As mentioned earlier, it is difficult to perform the bifurcation bifurcation point is realized in this case. This probably implies
and stability analysis of the model with 189 NEQ segments. Also that the parameter space wherein the oscillations were observed
to have a realistic comparison of non-linear behavior observed in the case of EQ model has been shifted while dealing with the
in the EQ stage simulations (Katariya, Moudgalya, et al., 2006), NEQ model. Hopf bifurcation is observed at higher reboiler duty.
NEQ model with partial back-mixing is considered. For studying Fig. 13 shows the bifurcation diagram with respect to Damkohler
the detailed bifurcation behavior with stability analysis, simu- number at the corresponding higher reboiler duty. Upto certain
lations were carried out with reduced number of segments such value of the Da (= 4.39), stable steady state is observed, which
that the steady state column profiles with both EQ and NEQ then converts to unstable steady state with possible oscillations.
2252 A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255

Fig. 9. Steady state composition and temperature profiles: comparison of EQ and NEQ model with partial back-mixing. Da = 3.0Q = 20.5 MW, R = 1.5, P =
4.5 bar. Amylene conversion (EQ) = 0.8457; (NEQ) = 0.8461. TAME purity in bottom (EQ) = 0.8782; (NEQ) = 0.8776.

This behavior is qualitatively similar to that of the EQ model as column stages, condenser and reboiler. Methanol–isopentane
reported earlier (Katariya, Moudgalya, et al., 2006). mixture, which is realized as a distillate, has been observed to
From the foregoing discussion, it is clear that the oscillations exhibit phase splitting under certain conditions. Also the work by
observed in the EQ model are not because of ignoring the trans- Zayer, Kulkarni, Kienle, Kumar, and Pushpavanam (2007) iden-
port processes. To understand the cause behind this effect, it may tifies the role of energy balance formulation in the dynamics of
be useful to study separately each of the modeling entities, such CSTR and reactive flash. This work may be extended to the mul-

Fig. 10. Comparison of the step (10% increase) response in 2M1B concentration
in feed for EQ, rigorous NEQ and constant holdup NEQ models: plot of TAME Fig. 11. Comparison of the bifurcation diagrams of EQ and NEQ models: plot of
purity in bottom vs. time (operating and design parameters: Da = 3.0, Q = bottom temperature vs. Damkohler number as continuation parameter (operating
20.5 MW, R = 1.5, P = 4.5 bar.) and design parameters: Q = 20.5 MW, R = 1.5, P = 4.5 bar).
A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255 2253

model, it is found that the number of segments (extent of


back-mixing) in the stripping section strongly influences the
performance of the NEQ results. A counter-intuitive behavior in
the form of isoamylene conversion increasing with decrease in
the height of the stripping section is observed. Also the dynamic
response of time variant mass and heat transfer coefficients show
significant variation when small disturbances are introduced.
This implies that consideration of time variant mass and heat
transfer coefficient is important. Synthesis of TAME in RD may
be associated with non-linear dynamic effects like limit cycles,
which are confirmed by dynamic simulation using an EQ model
in our earlier studies. However, NEQ model simulations in the
same parameter space and operating region do not reveal such
phenomena. The oscillatory behavior that existed in the EQ
model has been shifted to a new parametric space in the case
of NEQ model that considers partial back-mixing.
To summarize, in order to explain the oscillations observed
Fig. 12. Bifurcation diagrams of NEQ models: plot of bottom temperature vs. re- in a simple EQ model, we have studied in detail a complex and
boiler duty as continuation parameter (operating and design parameters: Da = computationally rigorous NEQ model that incorporates concen-
3.0, R = 1.5, P = 4.5 bar).
tration dependent heat and mass transfer coefficients. Steady
state and dynamic simulations, along with bifurcations studies,
have confirmed the existence of oscillations in the NEQ model
as well. We believe that the effect of possible liquid phase split-
ting and dynamic energy balance may provide an explanation to
this phenomenon.

Acknowledgement

Authors would like to acknowledge Max Planck Institute for


Dynamics of Complex Technical Systems, Magdeburg, Sand-
torstr.1, 39106, Germany for providing the DIVA simulation
environment which has been used for bifurcation analysis in the
present work.

Appendix A. Calculation of mass transfer coefficients

A.1. Non-reactive SULZER BX packing (Rocha et al., 1993,


Fig. 13. Bifurcation diagrams of NEQ models: plot of bottom temperature vs. 1996)
Damkohler number as continuation parameter (operating and design parameters:
Q = 20.82 MW, R = 1.5, P = 4.5 bar). Void fraction:  = 0.9; packing area (m2 /m3 ): a = 492; chan-
nel flow angle: θ = 60; channel side (mm): 8.9.
tistage columns and more specifically to TAME synthesis. It has
been noticed that the assumption of pseudo-steady state energy • Gas phase calculations:
balance, especially for the condenser and reboiler, strongly in-
   
fluences the dynamic behavior, often resulting in oscillations. kg S (Uge + ULe )ρg S 0.8 μg 0.33
A detailed investigation on these aspects is expected to give a = 0.054 (23)
Dg μg D g ρg
better insight into the nonlinear dynamics of TAME synthesis
in reactive distillation. A preliminary study of these topics is Ugs
Uge = (24)
available in Katariya (2007). (1 − hL ) sin θ
ULs
5. Conclusion ULe = (25)
hL sin θ
A rigorous dynamic NEQ stage model has been formulated where Uge and ULe is the effective gas and liquid velocity in
and solved for the synthesis of TAME by reactive distillation. m/s, respectively, Ugs and ULs the superficial gas and liquid
The results of steady state and dynamic simulations using both velocity in m/s, respectively, kg the mass transfer coefficient
EQ model and NEQ models with and without partial back- in m/s (for binary pair), S the characteristic length, i.e. side
mixing are compared. From the steady state analysis of NEQ dimension of the corrugation crass-section (m), hL the frac-
2254 A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255

tional liquid holdup, μg and μL the gas and liquid viscosity in crimp height (mm): 11.5; crimp wavelength (mm): 21.8; thick-
Pa s and Dg and DL are the gas and liquid diffusion coefficient ness of single sandwich (mm) =21.8.
in m2 /s (for binary pair).
• Liquid phase calculations: • Mass transfer coefficient: for ReL = 13 − 320
 
DL CE ULe 0.5 ShL = 3.846 × 10−3 Re0.667
L ScL0.5 ,
kL (m/s) = 2 (26)
πS for 630 < Reg < 2181 (39)
where CE is the factor slightly less than unity to account for
those part of the packed bed that do not encourage the rapid ShL = 9.457 × 10−5 Re0.667
L Re0.482
g ScL0.5 ,
surface renewal (CE = 0.9) for 2181 < Reg < 5900 (40)
• Hydraulic calculations:
   1/3 Shg = 0.0476 Re0.736 Re0.229 Scg0.33 ,
Ft 2/3 3μL ULs g L
hL = 4 (27)
S ρL sin θgeff for ReL = 15.2 − 360 and Reg = 610 − 5920 (41)

Ft = 29.12(WeL FRL )0.15


μg μL kg de
Scg = , ScL = , Shg = ,
S 0.359 ρg D g ρg D L Dg
× 0.2 (28)
ReL 0.6 (1 − 0.93 cos γ)(sin θ)0.3 kL ν 2 gog de goL de
ShL = , Reg = , ReL = ,


DL μg μL
ρL − ρg ( P/ z)
geff = g 1− (29) gog deK
ρl ( P/ z)flood RegK = , deK = 4K/a,
μg
2 ρ S
ULs L  1/3
WeL = (30) 1 mu2L
σ K= and ν2 =
2
1 + (4/ap D) ρL2 g
ULs
FRL = (31) where de = 4/ap is the hydraulic packing diameter (m) for
Sg
mass transfer.
ULs SρL • Hydraulic calculations:
ReL = (32)
μL
 
5 ReL = 7.3 − 530 and RegK = 620 − 5900
P Pd 1
= (33)
z z 1 − K 2 hL
  hd = 0.0273 Re0.331
L (42)
Pd
= AUgs2
+ BUgs (34)   0.5
z 4Fs 2σ(1 − cos γ)
hstat = (43)
0.177ρg S gρL (1 − ρg /ρL ) sin θ
A= (35)
S2 (sin θ)2 • Pressure drop calculations:
Limited to 70% of the flooding point, i.e. for given liquid
88.774μg load, gas velocity does not exceed 70% of its value corre-
B= (36)
S 2  sin θ sponding to flooding.
K2 = 0.614 + 71.35S (37) P Pdry
⎛ = , ReL < 94 (44)
 0.111 ⎞ H H
ae U 2
β= = ⎝1 − 1.203 Ls ⎠ (38) P Pdry
ap Sg = 0.716 exp(0.00357 ReL ), 94 ≤ ReL ≤ 264
H H
where ae is the effective interfacial area (m2 /m3 ), ap the area (45)
of packing (m2 /m3 ), σ the surface tension (N/m), cos γ =
P Pdry
0.9 for σ < 0.055, cos γ = 5.211.1016.835σ for σ > 0.055 and = 0.443 exp(0.00539 ReL ),
P/ z is the pressure drop per unit section height (Pa/m). H H
264 < ReL < 630 (46)
A.2. Reactive KATAPAK packing (Kolodziej et al., 2004)

Void fraction (m3 /m3 ):  = 0.622; packing area or specific Pdry ψaρg w2og
= (47)
surface area (m2 /m3 ): a = 128.2; corrugation angle: θ = 45; H 83 K
A.M. Katariya et al. / Computers and Chemical Engineering 32 (2008) 2243–2255 2255

ψ = 6.275 Re−0.293
gK , 550 < RegK < 1550, Kroner, A., Holl, P., Marquardt, W., & Gilles, E. D. (1990). DIVA: An open
architecture for dynamic simulation. Computer and Chemical Engineering,
ψ = 6.561 Re−0.171
gK , 1550 ≤ RegK < 6000 (48) 14(11), 1289–1295.
Mohl, K. D., Kienle, A., Gilles, E. D., Rapmound, P., Sundmacher, K., &
where wog is the superficial velocity (m/s) of gas and Fs = Hoffmann, U. (1999). Steady state multiplicities in reactive distillation
columns for the production of fuel ethers MTBE and TAME: Theoretical
0.018. analysis and experimental verification. Chemical Engineering Science, 54,
1029–1043.
References Noeres, C., Dadhe, K., Gesthuisen, R., Engell, S., & Grak, A. (2004). Model-
based design, control and optimisation of catalytic distillation processes.
Asprion, N. (2006). Nonequilibrium rate-based simulation of reactive systems: Chemical Engineering and Processing, 43, 421–434.
Simulation model, heat transfer, and influence of film discretization. Indus- Ouni, T., Jakobsson, K., Pyhalahti, A., & Aittamaa, J. (2004). Enhancing
trial and Engineering Chemistry Research, 45, 2054–2069. productivity of side reactor configuration through optimizing the reaction
Baur, R., Higler, A. P., Taylor, R., & Krishna, R. (2000). Comparison of equi- conditions. Chemical Engineering Research and Design, 82, 167–174.
librium stage and nonequilibrium stage models for reactive distillation. Peng, J., Edgar, T. F., & Eldridge, R. B. (2003). Dynamic rate based and equilib-
Chemical Engineering Journal, 76, 33–47. rium model for a packed reactive distillation column. Chemical Engineering
Baur, R., Krishna, R., & Taylor, R. (2003). Bifurcation analysis for TAME syn- Science, 58, 2671–2680.
thesis in reactive distillation column: Comparison of pseudo-homogeneous Peng, J., Lextrait, S., Edgar, T. F., & Eldridge, R. B. (2002). A comparison
and heterogeneous reaction kinetics models. Chemical Engineering and of steady state equilibrium rate based models for packed reactive distil-
Processing, 42, 211–221. lation columns. Industrial and Engineering Chemistry Research, 41(11),
Baur, R., Taylor, R., & Krishna, R. (2000). Bifurcation analysis for TAME syn- 2671–2680.
thesis in a reactive distillation column: Comparison of pseudo-homogeneous Powers, M. F., Vickery, D. J., Arehole, A., & Taylor, R. (1988). A nonequilib-
and heterogeneous reaction kinetics models. Chemical Engineering and rium stage model of multicomponent separation processes. V. Computational
Processing, 42, 211–221. method for solving model equations. Computer and Chemical Engineering,
Baur, R., Taylor, R., & Krishna, R. (2001). Dynamic behavior of reactive dis- 12(12), 1229–1241.
tillation columns described by a nonequilibrium stage model. Chemical Reepmeyer, F., Repke, J., & Wozny, G. (2004). Time optimal start-up strate-
Engineering Science, 56, 2085–2102. gies for reactive distillation columns. Chemical Engineering Science, 59,
Faisal, H., Syed, C. E., & Datta, R. (2000). TAME: Thermodynamic analysis of 4339–4347.
reaction equilibria in the liquid phase. Journal of Chemical and Engineering Rihko, L. K., & Krause, A. O. I. (1995). Kinetics of heterogeneously catalysed
Data, 45, 319–323. tert-amyl methyl ether reactions in liquid phase. Industrial and Engineering
Fuller, E. N., Schettler, P. D., & Diddings, J. C. (1966). A new method for predic- Chemistry Research, 34, 1172.
tion of binary gas phase diffusion coefficients. Industrial and Engineering Rocha, J. A., Bravo, J. L., & Fair, J. R. (1993). Distillation columns contain-
Chemistry, 58, 19–27. ing structured packings: A comprehensive model for their performance. 1.
Higler, A. P., Taylor, R., & Krishna, R. (1999). Nonequilibrium modeling of Hydrulic model. Industrial and Engineering Chemistry Research, 32(4),
reactive distillation: Multiple steady states in MTBE synthesis. Chemical 641–651.
Engineering Science, 54, 1389–1395. Rocha, J. A., Bravo, J. L., & Fair, J. R. (1996). Distillation columns contain-
Jakobsson, K., Hasanen, A., & Aittamaa, J. (2004). Modeling of countercurrent ing structured packings: A comprehensive model for their performance.
hydrogenation process. Chemical Engineering Research and Design, 82, 2. Mass-transfer model. Industrial and Engineering Chemistry Research,
203–207. 35(5), 1660–1667.
Katariya, A. M. (2007). Non-linear dynamics of reactive distillation for synthesis Schenk, M., Gani, R., Bogle, D., & Pistikopoulos, E. N. (1999). A hybrid
of TAME. PhD Thesis. IIT Bombay, Powai. modeling approach for separation systems involving distillation. Chemical
Katariya, A. M., Kamath, R. S., Moudgalya, K. M., & Mahajani, S. M. (2006). Engineering Research and Design, 77, 519–534.
Study of non-linear dynamics in reactive distillation for TAME synthesis Subawalla, H., & Fair, J. R. (1999). Design guidelines for solid catalysed reactive
using equilibrium and non-equilibrium models, ESCAPE-16 and PSE’ 2006. distillation systems. Industrial and Engineering Chemistry Research, 38,
Katariya, A. M., Moudgalya, K. M., & Mahajani, S. M. (2006). Nonlinear 3696–3709.
dynamic effects in reactive distillation for synthesis of TAME. Industrial Sundmacher, K., & Hoffmann, U. (1996). Development of new catalytic distil-
and Engineering Chemistry Research, 45(12), 4233–4242. lation process for fuel ethers via a detailed nonequilibrium model. Computer
Kienle, A., & Marquardt, W. (2003). Nonlinear Dynamics and Control of Reac- and Chemical Engineering, 51(10), 2359–2368.
tive Distillation Proceses. In K. Sundmacher & A. Kienle (Eds.), Reactive Wilke, C.R., & Chang, P. (1955). Corelation of diffusion coefficients in dilute
distillation: Status and future directions (p. 241). Weinheim, Germany: solutions. AIChE Journal, 1, 264–270.
Wiley–VCH. Xu, Y., Flora, T. N., & Rempel, G. L. (2005). Comparison of a pseudo-
Kolodziej, A., Jaroszynski, M., & Bylica, I. (2004). Mass transfer and hydraulics homogeneous nonequilibrium dynamic model and a three-phase nonequi-
for KATAPAK-S. Chemical Engineering and Processing, 43(3), 457–464. librium dynamic model for catalytic distillation. Industrial and Engineering
Kreul, L. U., Gorak, A., Dittrich, C., & Barton, P. I. (1998). Dynamic catalytic Chemistry Research, 44, 6171–6180.
distillation: advanced simulation and experimental validation. Computer and Zayer, K. P., Kulkarni, A. A., Kienle, A., Kumar, M. V., & Pushpavanam, S.
Chemical Engineering, 22, S371–S378. (2007). Non-linear behavior of the reactor separator networks: Influence
Krishna, R., & Standart, G. L. (1976). A multiconponent film model incarporat- of the energy balance formulation. Industrial and Engineering Chemistry
ing an exact matrix solution of Maxwell–Stefan equations. AIChE Journal, Research, 46, 1197–1207.
22, 383.

You might also like