1

Large Scale optimization strategies for zone conguration of simulated moving beds
Yoshiaki Kawajiri Lorenz T. Biegler
Department of Chemical Engineering, Carnegie Mellon University, 5000 Forbes Ave. Pittsburgh, PA 15213

4 5

Abstract Simulated Moving Bed (SMB) processes are widely used in sugar, petrochemical, and pharmaceutical industries. However, systematic optimization of SMB, especially nding the optimal zone conguration including the standard and modied nonstandard congurations, is still a challenging problem. This paper proposes a simultaneous, fully discretized approach with an SMB superstructure using an interior-point solver. To nd the optimal structure, two superstructures are experimented for standard and nonstandard congurations. In case studies of the linear and bi-Langmuir isotherms, optimal zone congurations have been successfully obtained without introducing discrete variables. Finally, the eect of the number of columns on the optimal throughput is investigated. Key words: simulated moving bed, zone conguration, dynamic optimization, interior-point method, superstructure

7 8 9 10 11 12 13 14 15 16

17 18

19

Introduction

20 21 22 23 24

The Simulated Moving Bed (SMB) chromatographic process, originally developed and commercialized by UOP, performs a continuous and pseudo-countercurrent operation. SMB has been gaining more attention in food, sugar, and petrochemical industries. In recent years, SMB has been widely used as an enantiomeric separation technique in the pharmaceutical industry. An SMB system consists of multiple columns connected to each other in a circle, as shown in Fig. 1. The feed and desorbent are supplied continuously, and simultaneously
Corresponding author. Tel.+1-412-268-2238; fax +1-412-268-7139 Email address: kawajiri@cmu.edu (Yoshiaki Kawajiri).

25 26

Preprint submitted to Computers & Chemical Engineering

13 October 2006

Fig. 1. Schematic diagram of 8 column SMB

27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

the ranate and extract products are withdrawn also continuously. Here, instead of actual movement of the adsorbent, the countercurrent operation is simulated by intermittently switching the four streams, desorbent, extract, feed, and ranate, in the direction of the liquid ow. Because of this rotational operation, SMB does not reach a steady state but Cyclic Steady State (CSS). The operation of an SMB system is uniquely determined by the switching interval (step time) and the four velocities of the four zones, I, II, III, and IV. Furthermore, in SMB processes with more than 4 columns, the relative positions of the four streams are not unique. Fig. 2(a), (b), and (c) shows congurations of (NI , NII , NIII , NIV ) = (2, 2, 2, 2), (1, 3, 3, 1), and (1, 1, 5, 1), respectively, where Nm is the number of columns in Zone m. This creates a large number of dierent zone congurations. In addition to the standard zone congurations, there have been a number of non-standard congurations proposed to improve the performance. Fig.3(a) shows the Three-zone conguration [1,2], where the circulation from the right end to the left is cut and all liquid is withdrawn as the ranate stream. Fig.3(b) shows the Three-zone conguration with purging, where the column at the left end is isolated and the components in the column is purged into the extract stream. Therefore, we need to deal with quite a large number of choices in designing SMB systems. Optimization approaches of periodic adsorption processes, which are characterized by CSS, can be classied into three classes: The rst is the nested approach, where the mathematical model described by Partial Dierential Algebraic Equations (PDAEs) is discretized only in the spatial domain, and integrated until it reaches the CSS. Applications of this approach include SMBs and SMB reactors [3,4]. The second approach, tailored with a single discretization, discretizes the PDAE model only in the spatial domain, but the CSS proles are obtained by enforcing equality constraints, which reduces the integration cost signicantly [5]. The nal approach is full-discretization, where the PDAE model is fully discretized both in the spatial and temporal domain, producing a large number of algebraic constraints[6,7]. In our previous study, we reported successful results of the full-discretization approach for SMB processes. [8] 2

44 45 46 47 48 49 50 51 52 53 54

Fig. 2. Examples of standard SMB zone congurations: 8 column SMB

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73

Although several optimization approaches have been applied to SMBs, the SMB zone conguration problem has seen little investigation, particularly for a large number of columns. Zhang et al. [9] reported the multi-objective optimization of SMB and VARICOL processes of up to 6 columns, while nding optimal zone congurations. They employed a genetic algorithm to explore every possible zone conguration. In addition, superstructure formulations have been considered in Karlsson [10], where a two column system was optimized, and Toumi [11], where a relaxed nonlinear programming (NLP) formulation was developed for SMB. Nevertheless, the open problem of optimal zone conguration still remains especially for a large number of columns. In this work, we develop an optimization approach for both the standard and nonstandard congurations by using a superstructure of SMB alternative systems. We apply a full discretization formulation, where a central nite dierence formula is used for the spatial discretization and Radau collocation on nite elements is used for the temporal discretization [8]. The discretized equations are incorporated within a large-scale nonlinear programming (NLP) problem, which is solved using an interior-point solver, IPOPT [12]. The discussion of the mathematical model and two superstructures for the standard and nonstandard congurations are given in Section 2. The reliability and eciency of our superstructure approach are demonstrated with several case studies in Sections 3 and 4. Finally the eect of the number of columns on the optimal throughput is discussed in Section 5. 3

Fig. 3. Examples of nonstandard congurations: (a) Three-zone conguration, (b) Three-zone conguration with purging
74

Mathematical Modeling and Problem Formulation

75

2.1 Modeling of Chromatographic Column For the modeling of each chromatographic column, we use the same mathematical model as in our previous work[8]. Here both axial dispersion and diusion into adsorbent particles, which cause band broadening, are lumped into a mass transfer coecient. The mass balance in the liquid phase is given by: Cij (x, t) q j (x, t) C j (x, t) + (1 b ) i + uj i =0 (1) b t t x Mass balance in the adsorbent phase is described by the Linear Driving Force (LDF) model: Liquid phase based: or
j qi (x, t) = Kappl i (Cij (x, t) Cij,eq (x, t)) (1 b ) t

(2)

j qi (x, t) j,eq j = Kapps i (qi (x, t) qi (x, t)) (3) t Finally, the equilibrium between the liquid and solid phase is given by isotherms, which can be linear or nonlinear:

Solid phase based:

Liquid phase based: or Solid phase based:

j qi (x, t) = f (Cij,eq (x, t))

(4)

j,eq qi (x, t) = f (Cij (x, t))

(5)

i = 1, ..., Nc , j = 1, ..., NColumn . 4

76 77 78 79 80 81 82

Here b is the void fraction, Cij (x, t) is the concentration in the liquid phase of component i j in column j, qi is the concentration in the solid phase, uj is the supercial liquid velocity j,eq in the jth column, Cij,eq (x, t) is the equilibrium concentration in the liquid phase, qi is the equilibrium concentration in the solid phase, and Kapps i and Kappl i are the solidphase based and liquid-phase based mass transfer coecient, respectively. The subscripts i correspond to chemical components, superscript j is the column index, Nc is the total number of components, and NColumn is the number of columns. At the CSS, the concentration proles at the end of a cycle must be identical to those at the beginning of the cycle. The dynamics of the concentration proles in the whole cycle need to be considered. However, the problem size can be reduced by taking advantage of the symmetry of the SMB operation. By enforcing the following constraints, the proles at the end of the next step are required to be identical to those at the beginning of the current step, thus shifting the entire proles downstream: Cij (x, 0) = Cij+1 (x, tstep ),
j qi (x, 0)

83 84 85 86 87 88

j = 1, ..., NColumn 1 j+1 qi (x, tstep ), j = 1, ..., NColumn 1 CiNColumn (x, 0) = Ci1 (x, tstep ) N 1 qi Column (x, 0) = qi (x, tstep )

(6) (7) (8) (9)

89

2.2 SMB Superstructures In this section, we discuss two SMB superstructure formulations that cover multiple congurations. In this study, we require all velocities must be non-negative: uj , uj , uj , uj , uj 0, j J F D E R (10)

90 91

92 93 94 95

where uj , uj are velocities of ranate and extract withdrawn from jth column, and uj D E R and uj are velocities of desorbent and feed supplied to jth column. It is important to F note that all velocities remain constant within each step. The treatment of time-variant velocities can be found in [13,14]. 2.2.1 Standard superstructure First, we construct a superstructure of SMB that covers all possible standard zone congurations. The superstructure for the standard SMB zone congurations (standard superstructure) is shown in Fig. 4(a). Note that by taking advantage of the symmetry of SMB, the number of streams can be signicantly reduced. We choose a structure that has only one desorbent stream (u1 ) without loss of generality. Furthermore, since we consider only D the standard SMB conguration, we can enforce that there is at least one column in each 5

96

97 98 99 100 101 102

Fig. 4. SMB superstructures

103 104 105

zone. Therefore, we are able to eliminate some of the extract, ranate, and feed streams at the beginning and end of the column sequence in the superstructure. As a result, the feed, extract, and ranate are each represented by only NColumn 3 streams. Referring to Figure 4(a), the volume and component balance equations are given by:

106

uj1 uj1 uj1 + uj + uj = uj , j = 2, . . . NColumn E R D F uNColumn uNColumn uNColumn + u1 + u1 = u1 E R F D Cij1 (L, t)(uj1 uj1 uj1 ) + CF,i uj = Cij (0, t)uj , j = 2, . . . NColumn E R F CiNColumn (L, t)(uNColumn uNColumn uNColumn ) + CF,i u1 = Ci1 (0, t)u1 E R F uj D uj F uj R uj E = 0, j = 2, ..., NColumn = 0, j = 1, 2, NColumn = 0, j = 1, 2, NColumn = 0, j = NColumn 2, NColumn 1, NColumn 6

(11) (12) (13) (14) (15) (16) (17) (18)

107 108 109

where CF,i is the concentration of component i in the feed. Further, in order to prevent draining the desorbent and feed into the product streams without going through a column, the following constraint is implemented:

uj uj uj 0 j = 1, . . . , NColumn D F

(19)

110 111 112 113

If (19) is inactive, then the liquid from a column is transported to the next. If (19) is active, then the circulation loop is cut open and all liquid is withdrawn in either or both of the product streams. Not only is this constraint necessary to prevent undesirable operating parameters, but it is crucial to ensure the robustness of optimization [13,14].

114

2.2.2 Complete superstructure For non-standard SMB congurations including those with purging and cut-o of circulation ow, the standard superstructure discussed above is no longer sucient. In this case, we need to consider all possible supply/withdraw points, leading to the complete superstructure (Fig.4(b)). Note that this superstructure embeds a number of dierent non-standard congurations such as Three-zone and Three-zone with purging. For this superstructure, we again use (11)- (14), but do not apply (15)-(18). Because this superstructure has multiple desorbent streams, the constraint that prevents draining (19) is crucial; Since desorbent consumption has no eect on purity or recovery constraints, the optimizer loses the sensitvity unless desorbent consumption is included in the objective function. Again, all of the velocities are constant in each step. For treatment of the complete superstructure with time-variant velocities, refer to [13,14].

115 116 117 118 119 120 121 122 123 124

125 126

127

2.3 Optimization problem formulation

128

In this study, the following optimization problem is considered: 7

uj ,uj ,uj ,uj ,uj ,tstep D F E R

min

(20)

subject to : (1) (14), (19) (15) (18) for standard superstructure,

NColumn tstep j uj (t)CE,P rod (t)dt E

(Extract P roduct P urity) =

j=1

NColumn N c tstep j=1 i=1 0

P urmin
j uj (t)CE,i (t)dt E

(21)

NColumn tstep

(Extract P roduct Recovery) = ul u uu

129 130 131 132 133

j=1 0 NColumn tstep j=1 0

j uj (t)CE,P rod (t)dt E

Recmin
j uj (t)CF,P rod (t)dt F

(22) (23)

where is the objective function, tstep is the valve switching interval, or step time, P urmin and Recmin are the purity and recovery requirements of the desired product which should be recovered in the extract stream, respectively. The desired product is denoted by the index P rod. uu and ul are the upper and lower bounds on the zone velocities, respectively. The variables are constrained by the model equations discussed in Section 1.

134

2.4 Solution strategy

135 136 137 138 139 140 141 142 143 144 145 146

In our previous study [8], two optimization approaches for SMBs are compared. The rst approach is single discretization, where the PDAE model is discretized only in the spatial domain using the centered nite dierence method, and the resulting dierential equations are integrated. The dierential equations need to be integrated with the sensitivity equations in each iteration. The second approach is full discretization, where the PDAE model is discretized both in the spatial and temporal domains using centered nite dierence and Radau collocation on nite elements, respectively. This approach does not require integration. Our study has found that full-discretization approach signicantly reduces the computational eort. It is important to note that although the NLP problem of the full-discretization approach is large, the linear system to be solved in each iteration, i.e., the linearized Karush-Kuhn-Tucker (KKT) conditions, has a sparse structure and can be solved very quickly. Again in this study, we employ the full-discretization method with a Newton-based NLP solver. For this solver, we choose IPOPT, an interior-point solver with a lter line search method which exploits the sparse structure of linearized KKT conditions. For details of IPOPT, refer to [12].

147 148 149 150

151

Case Study 1: Standard SMB Zone Conguration

These optimization cases will be considered for two dierent separation systems. The rst system is the separation of fructose and glucose [15] with the liquid-based isotherm:
j qi (z, ) = Ki Cij,eq (z, )
152 153 154 155

(24)

The linear driving force equation based on the liquid phase (2) is used and the model parameters are shown in Table 1. As a more challenging system, we also consider separation of 1,1-bi-2 naphthol enantiomers, where the adsorption equilibrium is given by a solid-based bi-Langmuir isotherm [16]: H1,i Cij (z, ) H2,i Cij (z, ) + j j j j 1 + K11 C1 (z, ) + K12 C2 (z, ) 1 + K21 C1 (z, ) + K22 C2 (z, )

j,eq qi =

(25)

156 157 158 159 160 161 162 163

This isotherm takes into account the saturation of adsorption as well as interactions between two components. For this system, the mass transfer resistance equation is based on the solid phase (3). The model parameters are listed in Table 2. Throughout this study, the number of nite elements in the temporal domain NF ET and collocation points NCOL are set to 5 and 3 , respectively, and the total number of nite elements NColumn NF EX is xed. Discussion of the accuracy of this discretization method and choice of numbers of nite elements can be found in [8]. This optimization problem is implemented within the AMPL modeling environment [17]. As our rst case study, we consider maximizing the sum of all feed streams, NColumn uj . i F The results of the nonlinear isotherm with dierent feed compositions are tabulated in Table 3, and corresponding optimal congurations and concentration proles are shown in Fig. 5. As can be seen in the table, the optimal zone conguration is dependent on the feed composition. If the feed composition of the desired extract component (Component 1) is large, NI must be increased to retain this component in Zone IV and not to leave it behind to drain into the ranate stream (Fig. 5(a)). If the impurity component (Component 2) is contained in the feed at a relatively higher concentration, then NIII must be increased to avoid contaminating the extract product (Fig. 5(b) and (c)). Interestingly, the NLP solution selects only one stream of each kind (uj , uj , and uj ) with F E R a dominant nonzero ow and the remaining streams have negligible values, typically less than 0.01% of the dominant ow. This is observed as long as appropriate tolerance and scaling are chosen. Therefore we are able to obtain the optimal zone conguration with an NLP formulation and without the addition of logic constraints or integer variables. This follows because streams of dierent types should be placed as far away as possible from each other to prevent contamination; the optimizer successfully nds such congurations. We note that this is also observed for optimal feed location problem of distillation columns, where the NLP formulation of the superstructure without logic constraints or binary variables is sucient to nd the optimal feed location [18]. 9

164 165 166 167 168 169 170 171 172

173 174 175 176 177 178 179 180 181 182

Table 1 Parameters of Fructose / glucose separation parameter

b

value 0.389 0.518 0.743

parameter L NColumn [m] Kapps Kapps CF,2

1 2

value 16.0 6.84 103 6.84 103 50.0 8.0 80.0 3

K1 K2 CF,1 [%]

[1/s] [1/s]

50.0 0.0 90.0 5 160

[%]

ul [m/h] P urmin [%] NF ET NF EX NColumn

uu [m/h] Recmin [%] NCOL

Table 2 Parameters of 1,1-bi-2 naphthol enantiomeric separation parameter

b

value 0.40

parameter L NColumn [m] Kappl H1,2 H2,2 K12 K22 CF,2 [g/l]
2

value 0.84 0.10 2.69 0.10 0.0466 3.0 3 6.424 80.0 3

Kappl H1,1 H2,1 K11 K21 CF,1

[1/s]

0.10 3.73 0.30 0.0336 1.0

[1/s]

[g/l] [g/l]

[g/l] [g/l]

[g/l]

3 0.0 97.0 5 80

ul [m/h] P urmin [%] NF ET NF EX NColumn

uu [m/h] Recmin [%] NCOL

183 184 185 186 187 188 189 190

We also investigate how the objective function inuences the optimal solution. (Table 4). As the minimization of the desorbent velocity is introduced in the objective function, NIV increases. This is because reducing the desorbent velocity requires a corresponding increase of the recycle stream, which could lead to elimination of the faster component at the downstream end of Zone IV that would contaminate the extract. To compensate, the optimizer also increases NIV to prevent the elimination of the faster component. Again, only one stream of each kind is dominant and the rest are negligible. All computations are carried out on a Xeon 3.4 GHz Linux machine. 10

Table 3 Feed velocity maximization with 12 columns for 1,1-bi-2 naphthol enantiomeric separation Feed composition (Component 1:Component 2) uj [m/h] (Only dominant ow is shown) F (NI , NII , NIII , NIV ) Extract purity [%] Extract recovery [%] Number of iterations CPU time [min]
uD=3.600m/h
4 C (x,t) 2 0 0
i

80:20 % u6 = 1.608 F (3,5,3,1) 97.0 80.0 61 4.51

50:50 % u6 = 1.068 F (2,5,4,1) 97.0 80.0 69 5.85

30:70% u6 = 0.871 F (2,5,4,1) 97.0 80.0 47 3.02

uF=1.608m/h 4.021 m/h 5.629 m/h 2.824 m/h 11 uR=2.805m/h

0.8

6.424 m/h

0.1

0.2

3 uE=2.403m/h

0.3

0.4
x [m]

0.5

0.6

0.7

(a) Component 1: Component 2 = 80:20

1 uD=3.962m/h

4 Ci(x,t) 2 0 0

6.424 m/h

4.127 m/h

uF=1.068m/h 5.195 m/h

2.462 m/h
0.7

0.1

2 uE=2.297m/h

0.2

0.3

0.4
x [m]

0.5

0.6

11 uR =2.734m/h

0.8

(b) Component 1: Component 2 = 50:50

1 uD=4.160m/h

4 C (x,t) 2 0 0

6.424 m/h

4.198 m/h

uF=0.872m/h

5.069 m/h

2.264 m/h
0.8

0.1

2 uE=2.226m/h

0.2

0.3

0.4
x [m]

0.5

0.6

0.7

11 uR=2.806m/h

(c) Component 1: Component 2 = 30:70

Fig. 5. Optimal conguration for 1,1-bi-2 naphthol enantiomeric separation: (a) Component 1:2=80:20, (b) Component 1:2=50:50, (c) Component 1:2 = 30:70
191

Case Study 2: Non-standard congurations

192 193 194 195 196

In this section, case studies for nonstandard congurations are carried out with the complete superstructure. The same model parameters and isotherms in the previous section are used here. The result of NColumn = 8 with the linear isotherm is shown in Table 5. In the optimal conguration of feed maximization ( = NColumn uj ), the circulation j F is cut open, the leftmost column is isolated and the components and desorbent in the 11

Table 4 Optimization of fructose and glucose separation with standard SMB superstructure for dierent objective functions with 8 columns Objective function uj [m/h] (Only dominant ow is shown) F u1 [m/h] D (NI , NII , NIII , NIV ) Extract purity [%] Extract recovery [%] Number of iterations CPU time [min]
uD=3.942m/h 1 8.000 m/h
0.5 0 0

NColumn i

uj F

NColumn i

uj + 0.5u1 D F

u1 D u3 = 0.500 (xed) F 0.477 (2,2,2,2) 90.0 80.0 113 4.00

u3 = 1.1584 F 3.942 (1,3,3,1) 90.0 80.0 75 2.76

u3 = 0.6233 F 0.6878 (2,2,2,2) 90.0 80.0 88 2.81

C (x,t)/C

F,i

5.029 m/h

uF=1.159m/h

6.188 m/h

4.058 m/h

1 uE=2.971m/h

4
x [m]

(a) =
1 5

NColumn i=1

7 uR=2.130m/h

uj F

C (x,t)/C

uD=0.688m/h 1 3.374 m/h

0.5 0 0

F,i

2.727 m/h

uF=0.623m/h 3.351 m/h

2.686 m/h

2 uE=0.646m/h

4
x [m]

6 uR=0.665m/h
NColumn i=1

(b)

NColumn i=1

uj + 0.5 F

uj D

Ci(x,t)/CF,i

uD=0.477m/h 1 2.620 m/h

0.5 0 0

2.148 m/h

uF=0.500m/h 2.648 m/h

2.143 m/h

Glucose Fructose

2 uE=0.473m/h

4
x [m]

6 uR=0.505m/h

(c) =

NColumn i=1

uj D

Fig. 6. Optimal congurations and concentration proles obtained with standard superstructure for fructose and glucose separation with 8 columns
197 198 199 200 201 202

(rst) extract stream are purged (see Fig. 7). As a result, this conguration consumes more desorbent, but the throughput is 20% higher than that of the standard SMB zone conguration. Furthermore, the (second) extract is withdrawn at the right end of the SMB. Note that although desorbent and extract streams appear at multiple locations for purging, the feed and ranate have only one dominant stream; the remaining streams are negligible. We observe throughout our case study that desorbent is supplied to multiple 12

Table 5 Optimization of fructose and glucose separation with complete superstructure for dierent objective functions with 8 columns Objective function uj [m/h] (Only dominant ow is shown) F u1 [m/h] D Extract purity [%] Extract recovery [%] Number of iterations CPU time [min]
uD=8.000m/h uD=6.610m/h 6.610 m/h 1 8.000 m/h
0.5 0 0

NColumn i

uj F

NColumn i

uj + 0.5u1 D F

u1 D u3 = 0.500 (xed) F 0.477 90.0 80.0 87 2.95

u3 = 1.390 F 14.61 90.0 80.0 79 3.02

u3 = 0.6233 F 0.688 90.0 80.0 74 6.23

uF=1.390m/h

Ci(x,t)/CF,i

8.000 m/h

4.521 m/h

1 uE=8.000m/h

4
x [m]

Glucose Fructose

7 uR=3.479m/h

8 uE=4.521m/h

Fig. 7. Optimal conguration and concentration proles obtained by complete superstructure approach for throughput maximization ( = NColumn uj ) of glucose and fructose separation j=1 F with NColumn =8. Vertical dotted line shows the port where the circulation is cut open.
203

columns only when the columns are isolated for purging. However, when desorbent minimization is included in the objective function, the optimal conguration becomes the standard SMB. The optimal solutions of = NColumn uj + j F NColumn j NColumn j uD are the same as those of the standard SMB zone uD and = j 0.5 j conguration problem (Table 4). This is consistent with our previous studies that deal with time-variant velocities[13,14]. Again in this case study, the non-dominant ow rates are negligible.

204 205 206 207 208 209

210

Eect of total number of columns on optimal throughput

211 212 213 214 215 216 217 218

In this section we consider optimal throughput problems ( = NColumn uj ) by increasi=1 F ing the number of columns while xing the total length of the columns (NColumn L). Fig.8 shows the optimal throughput for fructose and glucose separation. Note that the optimal throughput for the conguration obtained with the complete superstructure is always higher than that of the standard superstructure. The optimal conguration for NColumn = 16 is shown in Fig. 9. The two columns from the left are isolated and the components in these columns are purged into the extract stream. Again, only one stream of each kind is dominant except the desorbent and extract streams for purging. 13

2.0 Standard Complete 1.5

NColumn

i=1
1.0 0.5 0 4 6 8 12 16

uj F

NColumn

Fig. 8. Throughput versus NColumn for fructose and glucose separation

219 220 221 222 223 224 225 226 227 228

Also note that the optimal throughput becomes insensitive to NColumn at higher values of NColumn . This is due to the singular nature of the optimization problem (optimal control problem), where the control variables appear linearly in the dierential equations and objective functions. This type of problem is known to have a at optimal surface in the reduced space, and the objective value is insensitive to continuous renement of the optimal control prole. For the SMB problem, increasing NColumn leads to renement of the control variables in the spatial domain, since the number of column velocities uj increases. Similarly, in our PowerFeed previous study, we observed that the optimal throughput plateaus with renement of the time-variant control variable uj (t) in the temporal domain [8]. It should be also noted that because the optimization problem is non-convex, our approach does not guarantee nding the global optimum. In our case studies, we nd local optimal solutions with very close optimal values when NColumn are high, which again is attributed to the singular nature of the problem. Therefore, multiple starting points and IPOPT options are applied to ensure that undesirable local solutions are avoided. A similar trend is observed for the separation of 1,1-bi-2 naphthol enantiomers, except that the optimal throughput of the complete superstructure plateaus at relatively lower values of NColumn (Fig. 10). It is observed only in the complete superstructure with NColumn 8 that the non-dominant ows become non-negligible; some exceeded 3% of the dominant ones. This can be also explained by the singular nature of the problem, since the ill-conditioning of the optimal control problem intensies as NColumn increases. Nevertheless, this is not a practical issue; since high NColumn does not improve throughput but increases SMB capital cost, this solution should not be considered.

229 230 231 232 233

234 235 236 237 238 239 240 241

242

Conclusions and Future Work

243 244 245

A fully-discretized NLP formulation with an SMB superstructure and interior-point solver has been found to be ecient and reliable for the optimal zone conguration problem for the standard SMB conguration. Moreover, discrete variables were not required through14

C (x,t)/C

uD uD uD =8.00m/h =8.00m/h =6.50m/h 6.50 m/h 1 8.00 8.00 m/h m/h 0.5
i F,i

uF=1.50m/h

8.000 m/h

3.59 m/h
7 8

2 1 uE=8.00m/h uE=8.00m/h

0 0

4
x [m]

15 uR=4.41m/h

16 uE=3.59m/h

Glucose Fructose

Fig. 9. Optimal conguration and concentration proles obtained by complete superstructure approach for throughput maximization of fructose and glucose separation with 16 columns. Vertical dotted lines show the ports where the circulation is cut open.
2.0 Standard Complete

uj F
NColumn i=1

1.5

1.0

=
0.5 0 4 6 8 12 16

NColumn
Fig. 10. Throughput versus NColumn for 1,1-bi-2 naphthol enantiomeric separation
246 247 248 249 250 251 252

out our case study for practical numbers of columns. Furthermore, by introducing the complete superstructure, more ecient congurations can be found where the throughput is further increased by consuming a larger amount of desorbent. Starting points and IPOPT options have been carefully chosen to nd the optimal control proles, as the optimization problems are non-convex and have a singular nature. Finally, it was found that the number of columns has a limited eect on the optimal throughput due to the singular nature of the optimization problem. Our future work will further extend our optimization approach to nd optimal operating schemes for multi-component separations, where more than two components are fractionated into multiple streams. As our superstructure formulations embed a number of both existing and novel operating schemes, we aim to develop further ecient designs and operating schemes.

253 254 255 256 257

258

Acknowledgment We thank Prof. Ignacio E. Grossmann for discussion of his distillation superstructure study and Dr. Carl D. Laird for his help with the IPOPT solver. Funding from the National 15

259 260

261

Science Foundation under Grant CTS-0314647 is gratefully acknowledged.

16

262

References
[1] D. M. Ruthven, C. B. Ching, Counter-current and simulated counter-current adsorption separation processes, Chemical Engineering Science 44 (1989) 10111038. [2] C. B. Ching, K. H. Chu, K. Hidajat, M. S. Uddin, Comparative study of ow schemes for a simulated countercurrent adsorption separation process, AIChE Jounal 38 (1992) 1744 1750. [3] G. Dnnebier, K.-U. Klatt, Modelling and simulation of nonlinear chromatographic u separation processes: a comparison of dierent modelling approaches, Chemical Engineering Science 55 (2000) 373380. [4] A. Toumi, F. Hanisch, S. Engell, Optimal operation of continuous chromatographic processes: Mathematical optimization of the VARICOL process, Industrial and Engineering Chemistry Research 41 (2002) 43284337. [5] L. Jiang, L. T. Biegler, V. G. Fox, Simulation and optimization of pressure swing adsorption systems for air separation, AIChE Journal 49 (5) (2003) 11401157. [6] S. Nilchan, The optimisation of periodic adsorption processes, Ph.D. thesis, University of London, London, U.K. (1997). [7] D. Ko, R. Siriwardane, L. T. Biegler, Optimization of a pressure-swing adsorption process using zeolite 13X for CO2 sequestration, Industrial and Engineering Chemistry Research 42 (2003) 339348. [8] Y. Kawajiri, L. T. Biegler, Optimization strategies for simulated moving bed and PowerFeed processes, AIChE Journal 52 (4) (2006) 13431350. [9] Z. Zhang, K. Hidajat, A. K. Ray, M. Morbidelli, Multiobjective optimization of SMB and Varicol processes for chiral separation, AIChE Journal 48 (12) (2002) 28002816. [10] S. Karlsson, Optimization of a sequential-simulated moving-bed separation process with mathematical programming methods, Ph.D. thesis, bo Akademi University, bo, Finland A A (2001). [11] A. Toumi, Optimaler betrieb und regelung von simulated-moving-bed-prozessen, Ph.D. thesis, Universitt Dortmund, Shaker Verlag, Aachen, Germany (2005). a [12] A. Wchter, L. T. Biegler, On the implementation of an interior point lter line search a algorithm for large-scale nonlinear programming, Mathematical Programming 106 (1) (2005) 2557. [13] Y. Kawajiri, L. T. Biegler, A nonlinear programming superstructure for optimal dynamic operations of simulated moving bed processes, Industrial & Engineering Chemistry Research, to appear. [14] Y. Kawajiri, L. T. Biegler, Large scale nonlinear programming for asymmetic design and operation of simulated moving beds, Journal of Chromatography A, to appear.

263 264

265 266 267

268 269 270

271 272 273

274 275

276 277

278 279 280

281 282

283 284

285 286 287

288 289

290 291 292

293 294 295

296 297

17

298 299 300

[15] K. Hashimoto, S. Adachi, H. Noujima, H. Maruyama, Models for the separation of glucose/fructose mixture using a simulated moving bed adsorber, Journal of Chemical Engineering of Japan 16 (5) (1983) 400406. [16] M. Minceva, L. S. Pais, A. E. Rodrigues, Cyclic steady state of simulated moving bed processes for enantiomers separation, Chemical Engineering and Processing 42 (2003) 93 104. [17] R. Fourer, D. M. Gay, B. W. Kernighan, AMPL: A Modeling Language for Mathematical Programming, Duxbury Press, Belmont, CA, 1992. [18] J. Viswanathan, I. E. Grossmann, Optimal feed locations and number of trays for distillation columns with multiple feeds, Industrial and Engineering Chemistry Research 32 (1993) 2942 2949.

301 302 303

304 305

306 307 308

18