Chemical Engineering and Processing 43 (2004) 14751485

Freeze drying process: real time model and optimization

Edinara Adelaide Boss a,,1 , Rubens Maciel Filho a , Eduardo Coselli Vasco de Toledo a,1
a

Laboratory of Optimization, Design and Advanced Control, Chemical Process Department, Chemical Engineering Faculty, State University of CampinasUNICAMP, C.P. 6066, Campinas 13081-970, Brazil Received 29 May 2003; received in revised form 29 January 2004; accepted 29 January 2004 Available online 18 March 2004

Abstract Freeze drying is a separation process based on the sublimation phenomenon. This process has the following advantages compared to the conventional drying process: the material structure is maintained, moisture is removed at low temperature (reduced transport rates), product stability during the storage is increased, the fast transition of the moisturized product to be dehydrated minimizes several degradation reactions. Freeze drying process has not been studied well enough. In order to put it to practice, a mathematical model based on fundamental mass and energy balance equations has been developed, based on a deterministic mathematical model proposed by Liapis and Sadikoglu [Drying Technol. 15 (34) (1997) 791], and used to calculate the amount of removed water and amount of residual water. The proposed model contains the freeze drying equations, which are solved by the orthogonal collocation and polynomial approximationJacobi method. The results show that the dynamic mathematical model represents well the process and is especially well suited for real time optimization. As a case study to illustrate the model utilization in a real time optimization procedure, the freeze drying process was optimized by the method of Successive Quadratic Programming (SQP) used for solution of non-linear equations, for skimmed milk and soluble coffee. The optimization procedure showed to be an important tool to improve the process performance since lower energy consumption and hence lower cost has been achieved to obtain the product with the same quality. 2004 Elsevier B.V. All rights reserved.
Keywords: Freeze-drying; Dynamic process; Mathematical model; Optimization

1. Introduction Freeze drying is an important drying process for products which have to conserve their avor, bioactivities, and other properties [2]. The freeze drying experiments are time consuming and are usually expensive to be carried out in all possible operating ranges. Because of that, expecting great reduction of experiments considerable increase in the development and use of mathematical models that can predict the behavior of the freeze drying process satisfactorily is observed. For most multivariable processes in which numerous potentially inuential factors are involved as is the case for the freeze drying process, it is not always easy to determine the inuence of each variable, as well as the interactions between them.

In the present work, a suitable mathematical model for the freeze drying process was developed and validated. The deterministic model is then used and the parametric sensitivity analysis of this process through factorial design in two levels is proposed as a tool to identify the effect of the process variables and the interaction among them. Finally, the skimmed milk and soluble coffee processes of freeze drying were optimized though the use of the mathematical model.

2. Process description Freeze drying is a process by which a solvent is removed from a frozen material or frozen solution by sublimation of the solvent and by desorption of the sorbed solvent, generally under reduced pressure. This process involves the following three stages: freezing, primary drying, and secondary drying periods as shown in Fig. 1. The performance of the overall freeze drying process depends signicantly on the freezing stage. The material to be

Corresponding author. E-mail address: boss@lopca.feq.unicamp.br (E.A. Boss). 1 Tel.: +55-19-37883909; fax: +55-19-37883965.

0255-2701/$ see front matter 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.cep.2004.01.005

1476

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

Fig. 1. Freeze drying periods.

processed is cooled down to a temperature below the solidication. The shape of the pores, the pore size distribution, and the pore connectivity of the porous network of the dried layer formed by the sublimation of the frozen water during the primary drying stage depend on the ice crystals formed during the freezing stage. An excellent option is to obtain large dendritic ice crystals formed and a homogeneous dispersion due to the high mass transfer rate of the water vapor in the dried layer [3]. The product could then be dried more quickly. The solvent is removed by sublimation under vacuum and heat addition in the primary drying stage. A signicant amount of the sublimation latent heat is also consumed when the water molecules sublimate and enter to the vapor phase. Because of this, the temperature of the frozen product is reduced. It is then necessary to supply heat to the product, which could be provided by conduction, convection and/or radiation. During the secondary drying stage, the solvent is removed from the chamber and a small amount of sorbed water can be removed by desorption. The bound water is removed by heating the product under vacuum. Through the parametric sensitivity analysis carried out on a literature based model a new deterministic mathematical model has been proposed. In this paper, the skim milk and soluble coffee processes data are used for the model validation as well as they are used as case studies for the process optimization.

The Liapis and Sadikoglu [2] model contains the freeze drying equations, which are solved by orthogonal collocation and polynomial approximationJacobi method. The values of parameters required by the model solution are given in Table 1, and dusty gas parameters are available in [1]. Liapis and Sadikoglu [2] model equations for mass and energy balance are demonstrated as follows: TI = Ie t + 2 TI x2 Cpg Ie CpIe , (Nt TI ) x 0xX (1)

Hv I Ie CpIe

Csw t

TII 2 TII = II 2 , t x

XxL

(2)

Table 1 Parameters, values, and expressions Parameters C01 C1 (m) C2 Cpg (kJ/kg K) CpIe (kJ/kg K) CpII (kJ/kg K) 0 Csw (kg of water/kg of solids) Csw (kg of water/kg of solids) 0 Dw, in (kg m/s3 ) kf (kW/m2 K) kIe (kW/m K) kII (kW/m K) L (m) p0 (N/m2 ) in pin0 (N/m2 ) p0 (N/m2 ) w pw0 (N/m2 ) P0 (N/m2 ) T0 (K) TUP (K) TLP (K) f (Tx ) (N/m2 ) (m2 ) Values and expressions 7.219 1015 3.85583 104 0.921 1.6166 2.59 1.93 0.6415 exp{2.3[1.36 0.036(T T 0 )}/100 0.000143931(TI3 ((1/Mw ) + (1/Min )))0.5 (1.5358 103 )P 1.412 106 (P0 Px ) + 2.165 104 (0.48819/TII ) + 0.4685 103 0.02 4.00 4.00 1.07 1.07 5.07 233.15 313.15 313.15 133.32240[exp(2445.5646/Tx + 8.23121 log10 (Tx ) 0.01677006Tx + 2 1.20514.105 Tx 6.757169)] 2840.0 2687.0 0.785 18.4858[TI1,5 /(TI + 650)] 215.0 212.21 1030.0 5.676 1011

3. Mathematical model A mathematical model based on fundamental mass and energy balance equations has been developed, based on a model originally developed by Liapis and Sadikoglu [2]. The assumptions are the same made by Lichteld and Liapis [4] and Millman et. al. [5], which are: one-dimensional heat and mass transfer; sublimation occurs at an interface parallel to surface; the solid layer is semi-innite; a binary mixture of water vapor and inert gas ows though the dried layer; at the interface, the concentration of water vapor is in equilibrium with the ice; in the porous region, the solid matrix and the enclosed gas are in thermal equilibrium; the frozen region is considered to be homogeneous, with uniform thermal conductivity, density, and specic heat, and to contain a signicant proportion of dissolved gases.

mx (kg/m s) Ie (kg/m3 ) I (kg/m3 ) II (kg/m3 ) (kW/m2 K4 )

Hs (kJ/kg) Hv (kJ/kg)

Quoted from LIAPIS and SADIKOGLU, 1997 [2].

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

1477

1 R t 1 R t

pw TI pin TI

1 Nw I Csw , Mw x Mw t 1 Nin , Min x Cpg Ie CpIe ,

0xX (3)

= 2 TI x2

0xX (Nt TI ) x 0xL

(4)

TI = Ie t

Table 2 Range of variables for constants dependent only upon the structure of porous medium and giving relative Knudsen ow permeability (C1 , m), constant dependent only upon structure of the porous medium and giving the ratio of bulk diffusivity within the porous medium to the free gas bulk diffusivity (C2 , dimensionless), effective dried layer density (Ie , kg/m3 ), dried layer density (I , kg/m3 ), frozen layer density (II , kg/m3 ), thickness of the layer (L, m), plate temperature (Tplate , K) and total pressure (P, N/m2 ) Level 1 +1 C1 3.18 5.03 104 104 C2 0.76 1.27 Ie 161 269 I 159 265 II 773 1290 L 0.015 0.025 P 3.8 6.34 Tplate 234.9 390.6

Hv I + Ie CpIe

Csw t

(5)

The TI is the dried layer temperature (K), Hv the enthalpy of vaporization of sorbed water (J/kg), I the dried layer density (kg/m3 ), Ie the effective dried layer density (kg/m3 ), CpIe the effective dried layer heat capacity (J/(kg K)), Cpg the gas heat capacity (J/(kg K)), Nt the total mass transfer ux (kg/(m2 s)), TII the frozen layer temperature (K), II the frozen layer thermal diffusivity (m2 /s), R the universal gas constant, pW is the partial pressure of water (N/m2 ), pin the partial pressure of inert gas (N/m2 ), MW the molecular weight of water, Min the molecular weight of inert gas, NW the water vapor mass transfer ux (kg/(m2 s)), Nin the inert gas mass transfer ux (kg/(m2 s)), and is the voidage fraction. Eqs. (1) and (5) are the same in form, differing in the application intervals and the data used which depends on the drying period that they refer to. The model allows for extensive simulation, so that it is possible to identify the design as well as operational variables, which more signicantly impact the systems behavior. 3.1. Statistical results Due to the relatively large number of variables to be analyzed, it was necessary to submit the process to an initial screening design since a full factorial design with two levels would require a lot of computer simulations. The methodology of PlackettBurman [6] is an adequate tool for this initial screening, since it makes it possible to determine the inuence of various factors with only a small number of trials, instead of using more extensive factorial design, which would furnish more complete information, but would also involve complexity. The variables analyzed are: model constant dependent upon the structure of the porous medium and relative Knudsen ow permeability (C1 , m), constant dependent upon the structure of the porous medium and giving the ratio of bulk diffusivity within the porous medium to the free gas bulk diffusivity (C2 , dimensionless), effective dried layer density (Ie , kg/m3 ), dried layer density (I , kg/m3 ), frozen layer density (II , kg/m3 ), thickness of the layer (L, m), plate temperature (Tplate , K), and total pressure (p, N/m2 ). Each independent variable is tested at two levels, a high (+) and a low () level, as shown in Table 2.

As it can be seen in Table 3, the effects of the C2 , Ie , and L are signicant for a 95% condence level, and the same occurred in Table 4 without the C2 effect that is signicant for 81% condence level. Bearing in mind the results of Tables 3 and 4, the variables chosen to the complete factorial design are C2 , Ie , and L. The variable P was also included in the factorial since it is an important variable of the process. The factorial design is a punctual analysis. Therefore, a planning was accomplished for each time as it is shown in Table 3, which account for the estimated effects for each time regarding the amount of removed water. The same occurs in Table 4 for the amount of residual water. The negative values, as it is showed in Table 3, are based on the decrease of the analyzed response. Once the relevant variables were selected by screening, simulations have been planned to obtain a quadratic model, consisting of 24 trials plus a star conguration with a central point, with the amount of removed water and the amount of residual water as responses. Table 5 presents the range of the variables. Figs. 25 show the effects of the complete factorial design on the amount of removed water and amount of residual water. These gures show the response surface and the contour diagrams. The response surface is a three-dimensional graphic that show the inuence of two variables in relation to the response and the contour diagrams show the same as the response surface but in a plane graphic. The results from the complete factorial design show that Ie should have its values as small as possible because when its value is increased, in this range, it reduces the amount of removed water and increases the amount of residual water, that is to say, high density restricts the sublimation. The thicker the layer of the material on the tray, the thinner the available supercial area for contact what leads to a decrease in the drying rate. Due to this, it is appropriate to use this variable as small as the process allows. The pressure and the C2 value should be maintained around the central point in order not to harm neither the amount of water to be removed nor the amount of residual water. 3.2. Developed mathematical model Through the analysis of the variables behavior and their interactions is proposed a deterministic mathematical model

1478

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

Table 3 Hourly effects of the amount of removed water Mean/interaction C1 C2 a Ie a I II La P Tplate

a

1h 0.1367 4E05 0.0049 0.0106 4E06 9E06 0.02 4E05 5E06

2h 0.1907 7E05 0.0089 0.0198 7E06 2E05 0.0374 7E05 9E06

3h 0.2432 0.0001 0.012 0.0277 9E06 2E05 0.0528 1E04 1E05

4h 0.2945 0.0001 0.0146 0.0347 1E05 3E05 0.0664 0.0001 1E05

5h 0.3448 0.0002 0.0168 0.041 1E05 3E05 0.0785 0.0001 2E05

6h 0.3448 0.0002 0.0168 0.041 1E05 3E05 0.0785 0.0001 2E05

7h 0.4428 0.0002 0.0201 0.0515 1E05 4E05 0.0993 0.0002 2E05

8h 0.4907 0.0002 0.0213 0.056 1E05 4E05 0.1083 0.0002 2E05

9h 0.5381 0.0002 0.0224 0.06 1E05 4E05 0.1164 0.0002 2E05

Signicant for a 95% condence level.

Table 4 Hourly effects of the amount of residual water Mean/interaction C1 C2 a Ie b I II Lb P Tplate

a b

13 h 0.1500 0.0049 0.0106 0.0158 0.0050 0.0050 0.0803 0.0051 0.0050

14 h 0.1277 0.0047 0.0092 0.0152 0.0047 0.0047 0.0728 0.0048 0.0047

15 h 0.1085 0.0045 0.0080 0.0145 0.0045 0.0045 0.0659 0.0046 0.0045

16 h 0.0919 0.0042 0.0069 0.0137 0.0043 0.0043 0.0597 0.0043 0.0043

17 h 0.0776 0.0040 0.0061 0.0129 0.0040 0.0040 0.0539 0.0040 0.0040

18 h 0.0654 0.0037 0.0054 0.0121 0.0037 0.0037 0.0487 0.0037 0.0037

19 h 0.0549 0.0034 0.0048 0.0113 0.0034 0.0034 0.0438 0.0035 0.0034

20 h 0.0459 0.0031 0.0044 0.0105 0.0031 0.0032 0.0394 0.0032 0.0031

Signicant for 81% condence level. Signicant for a 95% condence level.

Fig. 2. Response surface and contour diagrams of the amount of removed water as a function of: C2 (C2) and L.

Fig. 3. Response surface and contour diagrams of the amount of residual water as a function of: C2 (C2) and L.

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

1479

Fig. 4. Response surface and contour diagrams of the amount of removed water as a function of: C2 (C2) and Ie (ROIE), C2 (C2) and P, L and Ie (ROIE), L and P.

by the insertion of a new term in the Liapis and Sadikoglu [2] model. This term is C1 P (6) 9.5C2 Ie L C1 and P are proportional to the increase of the drying rate and C2 , Ie , and L are inversely proportional to the increase of the drying rate. C2 , Ie , and L have their effects, on the average, 9.5 times bigger than C1 and P. The constant unit is kg/(N m2 ).

Canceling the quadratic term from Eqs. (1) and (5) and inserting the term from Eq. (6), results in Eqs. (7) and (11). Eq. (8) presents the energy balance for the freezing period and Eqs. (9) and (10) the material balances for the drying period. H v I TI = t Ie CpIe Csw t Cpg Ie CpIe (Nt TI ) x C1 P , 9.5C2 Ie L (7)

0xX

1480

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

Fig. 5. Response surface and contour diagrams of the amount of residual water as a function of: C2 (C2) and Ie (ROIE), C2 (C2) and P, L and Ie (ROIE), L and P.

TII 2 TII = II 2 , X x L t x 1 pw 1 Nw I Csw = , R t TI Mw x Mw t 1 R t pin TI 1 Nin = , Min x 0xX

(8) 0xX (9) (10)

Hv I TI = t Ie CpIe

Csw t

Cpg Ie CpIe

(Nt TI ) x

C1 P , 9.5C2 Ie L (11)

0xL

Eqs. (7) and (11) are the same in form, differing in the application intervals and the data used which depends on the drying period that they refer to.

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485 Table 5 Range of variability for constant dependent upon structure of porous medium and giving the ratio of bulk diffusivity within the porous medium to the free gas bulk diffusivity (C2 , dimensionless), effective dried layer density (Ie , kg/m3 ), thickness of the layer (L, m) and total pressure (P, N/m2 ) Level 1 +1 0 2a +2a
a

1481

The initial condition for Eq. (25) is X = 0, t=0 t = tX=L , 0 x L TI x TI x

x=0

(26)

The initial and boundary condition of Eq. (11) is TI = (x), qI = kIe qII = kIe (27) (28) (29)

C2 0.76 1.27 1.01 0.5 1.52

Ie 161 269 215 107.5 322.5

L 0.015 0.025 0.02 0.01 0.03

P 3.8 6.34 5.07 2.535 7.605

, ,

x = 0, t > tX=L x = L, t > tX=L

x=L

2 = (24 )1/4 (star point), where 4 is the number of variables.

for the radiation heat transfer

4 qI = F(TUP (TI (t, 0))4 ),

The initial and boundary conditions of Eqs. (7)(10) are TI = TII = Tx = T 0 , pw = p0 , w

0 Csw = Csw ,

t > tX=L

(30)

t = 0, 0 x L t = 0, 0 x X

(12) (13) (14) (15)

The initial and boundary conditions of Eq. (11) are pw = (x), pin = (x), Csw = (x), pw = pw0 , t = tX=L , 0 x L t = tX=L , 0 x L t = tX=L , 0 x L pin = pin0 = P0 Pw0 , (31) (32) (33) x = 0, t = tX=L (34) pw x pin x = 0, = 0, x = L, t < tX=L x = L, t > tX=L (35) (36)

pin = p0 , in TI x

t = 0, 0 x L
x=0

qI = kIe and

x = 0, t > 0

4 qI = F(TUP (TI (t, 0))4 ),

t>0

(16)

for the radiation heat transfer to the upper dried surface kII TII TI kIe x x=X x x=X + V(II CpII TII |x=X I CpI TI |x=X ) + Nt Cpg Tx = H s Nt , TI = Tx = TII , qII = kII TII x x = X, 0 < t tX=L (17) (18) (19) (20) (21) x = X, t > 0
x=L

x=L

x=L

x = L, t > 0 t>0 t>0

4 qII = F(TLP (TII (t, 0))4 ),

qII = kf (TLP (TII (t, L))),

where is the StefanBoltzmann constant (kW/(m2 K4 )), TUP the upper plate temperature (K), TLP the lower plate temperature (K), and kf is the lm thermal conductivity. The pressure can be calculated as follows: pw = pw0 , pw = f(Tx ), pin x
x=X

pin = pin0 = P0 Pw0 , x = X, 0 < t tX=L x = X, 0 < t tX=L

x = 0, t > 0 (22) (23) (24)

= 0,

The material balance at the interface which denes the interface velocity, V, is given by V = dX Nw |x=X = dt II I (25)

The functions (x), (x), (x), and (x) provide the proles of TI , pw , pin , and Csw at the end of the primary drying stage or at the beginning of the secondary drying stage obtained by the solution of the model equation for the primary drying stage. The solution procedure involves the discretization of the axial variables by orthogonal collocation and integration in respect to time, through a RungeKutta based algorithm. The solution to the proposed model (CURVE 2) describes results depicted in Figs. 6 and 7. Curve 1 refers to the model of Liapis and Sadikoglu [2] predictions. Tables 6 and 7 show the standard deviation of the model by Liapis and Sadikoglu [2] as well as of the proposed mathematical model. It can be noticed that the proposed mathematical model better describes the real data compared to the model by Liapis and Sadikoglu [2]. Figs. 8 and 9 show the amount of removed water and the amount of residual water predicted by the proposed model compared to the soluble coffee industrial data. The plant is a continuous one and a company, which name is not being disclosed, supplied the operational values as well as the process parameters. Figs. 8 and 9 show that the proposed mathematical model represents well the freeze drying process of soluble coffee of a real plant in continuous operation.

1482

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485 Table 7 Standard deviation of the model by Liapis and Sadikoglu [2] and the proposed model in relation to the real data for the secondary period Time (h) Real data Curve 2 Curve 1 Standard DP2 Standard DP1 14 15 16 17 18 19 20 21 0.1085 0.0966 0.0860 0.0766 0.06820 0.0608 0.0541 0.0482 0.1032 0.0945 0.0860 0.0780 0.0702 0.0628 0.0557 0.0490 0.0488 0.0416 0.0354 0.0302 0.0257 0.0219 0.0187 0.0159 4.9212 2.1942 0.0465 1.8011 2.8873 3.3745 2.9575 1.7231 55.0364 56.9654 58.8099 60.5847 62.2747 63.8848 65.4529 66.9296

Fig. 6. Amount of removed water vs. the time during the primary period of the freeze drying process of skimmed milk. CURVE1: model by Liapis and Sadikoglu [2]. REALDATA: real data; CURVE2: proposed model.

Curve 1, model by Liapis and Sadikoglu [2] DP1 : standard deviation of the existent model; Curve 2, proposed model DP2 : standard deviation of the developed model; the medium standard deviation of the proposed model is of 2.49%.; the medium standard deviation of the model by Liapis and Sadikoglu [2] is of 261.2%.

Fig. 8. Amount of removed water vs. time during the primary drying stage of soluble coffee freeze drying.

4. Optimization of the freeze drying process

Fig. 7. Amount of residual water vs. the time during the secondary period of the freeze drying process of skimmed milk. CURVE1: model by Liapis and Sadikoglu [2]; REALDATA: real data; CURVE2: proposed model. Table 6 Standard deviation of the model by Liapis and Sadikoglu [2] and the proposed model in relation to the real data for the primary period Time (h) Real data Curve 2 Curve 1 Standard DP2 Standard DP1 3 4 5 6 7 8 9 10 11 0.1841 0.2634 0.3248 0.3750 0.4175 0.4543 0.4867 0.5158 0.5420 0.2180 0.2724 0.3212 0.3655 0.4061 0.4436 0.4785 0.5110 0.5415 0.2698 0.3263 0.3702 0.4060 0.4364 0.4626 0.4858 0.5065 0.5252 18.4136 3.4265 1.1196 2.5425 2.7281 2.3491 1.7023 0.9252 0.0942 46.5345 23.9149 13.9708 8.2658 4.5125 1.8292 0.2013 1.8012 3.1014

The optimization is an important tool to take decisions in order to nd out better operating condition in relation to design parameters as well as process restrictions.

Curve 1, Model by Liapis and Sadikoglu [2] DP1 : standard deviation of the model by Liapis and Sadikoglu [2]; Curve 2, proposed model DP2 : standard deviation of the proposed model; the medium standard deviation of the proposed model is of 3.7%; the medium standard deviation of the model by Liapis and Sadikoglu [2] is of 11.6%.

Fig. 9. Amount of residual water vs. time during the secondary drying stage of soluble coffee freeze drying.

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

1483

Fig. 10. Amount of removed water vs. time during the primary drying stage of freeze drying of skimmed milk. Comparison of the non-optimized data (WOTM) and optimized data (OTM).

Fig. 11. Amount of residual water vs. time during the secondary drying stage of freeze drying of skimmed milk. Comparison of the non-optimized data (WOTM) and optimized data (OTM).

In this work, the optimization of the freeze drying process for the skimmed milk as well as for the process of the soluble coffee production will be studied, using the mathematical model proposed in the previous section. The method of Successive Quadratic Programming (SQP) will be used for resolution of the non-linear equations that are included in the proposed mathematical model. 4.1. Successive Quadratic Programming There are several subroutines for implementation, which can solve the problems of SQP non-linear programming. For the freeze drying optimization process, the chosen option was a routine based on the software NLPQL developed by Schittkowski [7], in 1985. This routine solves general problems of non-linear programming and it is well suited to the proposed mathematical model. 4.2. The freeze drying optimization process of skimmed milk In the freeze drying process, the biggest problem is the time of drying, since the longer the time, the higher the energy costs. Because of this, the goal function is to minimize the drying time. Based on the factorial design, three variables were manipulated: thickness of the layer on the tray of the freeze dryer (L), the temperature of the plates under the tray (Tplate ), and the pressure in the freeze drying chamber (P). The values before optimization are: 313.14 K, 5.07 N/m2 , and 0.02 m. As the freeze drying of the two drying periods cannot be optimized simultaneously, the optimization was focused on the rst period which refers to the time when the largest removal of water usually occurs. After having optimized the rst period the results were analyzed for the second period. In Figs. 10 and 11, the real (pilot plant) results compared to the optimized process of the skimmed milk were presented. The optimized values are: 314.12 K, 7.48 N/m2 , and 0.018 m.

Observing Figs. 10 and 11, it can be seen that a signicant improvement in the freeze drying process of the skimmed milk was obtained. Time consumption was reduced by half compared to previous results for the same amount of removed water. This also allows the freeze drying process to be sensible to changes of values of some manipulated variables. The amount of residual water in the optimized process is much smaller than in the non-optimized, conrming larger efciency after optimization. 4.3. The freeze drying optimization process of soluble coffee The optimization of the freeze drying process for the soluble coffee was accomplished in the same way as the optimization of the freeze drying process for the skimmed milk. The values before optimization for Tplate , P, and L cannot be disclosed for industrial condence. Figs. 12 and 13 present the real (industrial) results compared to the optimized ones for the freeze drying process for the soluble coffee. A closer look at Figs. 12 and 13 allows noticing that a signicant improvement in the freeze drying process of

Fig. 12. Amount of removed water vs. time during the primary drying stage of freeze drying of soluble coffee. Comparison of the real data (REALDATA) and optimized data (OTMDATA).

1484

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

Appendix A. Nomenclature AW C01 amount of water constant dependent only upon structure of porous medium and giving relative DArcy ow permeability (m2 ) constant dependent only upon structure of porous medium and giving relative Knudsen ow permeability (m) constant dependent only upon structure of porous medium and giving the ratio of bulk diffusivity within the porous medium to the free gas bulkdiffusivity, dimensionless. heat capacity (J/(kg K)) weight fraction of bound water in dried layer (kg of water/kg of solid) free gas mutual diffusivity in a binary mixture of 0 water vapor and inert gas Dw,in = Dw,in P 3) (kg m/s water vapor pressure-temperature functional form presented in Table 1 thermal conductivity (kW/(m K)) thickness (m) molecular weight (kg/kmol) total mass ux Nt = Nw + Nin (kg/(m2 s)) partial pressure (N/m2 ) total pressure (N/m2 ) heat transfer rate temperature (K) interface velocity position (m) position of frozen interface (m)

C1

C2
Fig. 13. Amount of residual water vs. time during the secondary drying stage of freeze drying of soluble coffee. Comparison of the real data (REALDATA) and optimized data (OTMDATA).

soluble coffee was obtained. Time was 1.5 h shorter to get the same amount of removed water. The amount of residual water in the optimized process is again smaller than in the non-optimized process; this shows greater efciency of the process after optimization. The optimized values are: 348.27 K, 58.93 N/m2 , and 0.0175 m. This result leads to a conclusion that the optimization of the freeze drying process is a tool of extreme importance, which should be used to decrease the processing time, that is to say, to decrease the cost of the process. It is also observed that the process is very sensitive and small alterations in suitable variables may allow the process to be operated with larger efciency and productivity.

Cp Csw Dw,in

f(Tx ) K L M Nt p P q T V x X

5. Conclusion The results show that the proposed mathematical model represents well the process. The methodology of PlackettBurman was shown to be very useful for the determination of signicant variables for further process optimization. In such way, it is possible to consider a large number of variables and to avoid the loss of information, which might be essential for the model development as well as in the optimization of the process. The use of this technique reduced the number of variables to be analyzed, for the considered case study from eight to four. The optimization procedure in two steps (the sequential optimization of each freeze drying period) using SQP algorithm appeared to be very suitable and robust for both case studies considered. The results showed that the optimization is an important tool to maximize the amount of removed water and minimize the time of the freeze drying process with signicant impact on the energy consumption and hence in the process costs.

Greek letters pw at the end of primary drying period voidage fraction Hs enthalpy of sublimation of frozen water (kJ/kg) Hv enthalpy of vaporization of sorbed water (kJ/kg) TI at the end of primary drying period mx mixture viscosity (kg/(m s)) Csw at the end of primary drying period pin at the end of primary drying period density (kg/m3 ) Stefan-Boltzmann constant (kW/(m2 K4 )) Superscripts 0 initial value at time zero equilibrium value Subscripts e effective value f lm g gas in inert L value at x = L

Acknowledgements The authors acknowledge the nancial support from FAPESP. Process: 99/11333-1.

E.A. Boss et al. / Chemical Engineering and Processing 43 (2004) 14751485

1485

LP UP w X 0 I II

lower plate upper plate water vapor interfacial value surface value dried region frozen region

[3]

[4] [5]

References
[6] [1] P.J. Gloor, O.K. Crosser, A.I. Liapis, Dusty-gas parameters of activated carbon desorbent particles, Chem. Eng. Commun. 59 (1987) 95105. [2] A.I. Liapis, H. Sadikoglu, Mathematical modeling of the primary and secondary drying stages of bulk solution freeze-drying in trays: [7]

parameter estimation and model discrimination by comparison of the theoretical results with experimental data, Drying Technol. 15 (34) (1997) 791810. A.I. Liapis, R. Bruttini, M.J. Pikal, Research and development needs and opportunities in freeze drying, Drying Technol. 14 (6) (1996) 12651300. R.J. Lichteld, A.I. Liapis, An absorptionsublimation model for a freeze dryer, Chem. Eng. Sci. 34 (9) (1979) 10851090. M.J. Millman, A.I. Liapis, J.M. Marchello, An analysis of the lyophilization process using a sorptionsublimation model and various operational policies, AIChE J. 31 (10) (1985) 1594 1604. R.L. Plackett, J.P. Burman, The design of optimum multifactorial experiments, Biometrika 33 (1946) 305325. K. Schittkowski, NLPQL: A FORTRAN-subroutine for solving constrained non-linear programming problems, Ann. Operat. Res. 5 (1985) 485500.