Original paper

Bioprocess Biosyst Eng 25 (2003) 349355 DOI 10.1007/s00449-003-0318-0

The application of a Pareto optimisation method in the design of an integrated bioprocess

349 Abstract A rapid method for designing integrated bioprocesses, using a combination of a windows of operation and a Pareto optimisation approach, is described in this paper. Within bioprocesses, multiple objectives are common, and achieving a satisfactory trade-off amongst the design objectives is crucial. Conventional optimisation results in the identication of the best operating policy for a given desired performance but gives little insight into how the process performance changes in the vicinity of the solution. In this paper, we explore the use of a Pareto optimisation technique to locate the optimal conditions for an integrated bioprocessing sequence and the benets of rst reducing the feasible space by the development of a series of windows of operation to provide a smaller search area for the optimisation. The nal results are then presented in performance trade-off graphs and look-up tables, which give the design engineer an easily manageable solution set to work with. In this way, the decision-making procedure for design is made faster and more transparent. Two case studies illustrate the results from this integrated design methodology, some of which are counter-intuitive compared with the general design experience. C t l W0 W w concentration of solids in the supernatant, g/L total processing time, h fed-batch fermentation growth rate, h initial feasible operating condition set feasible set performance set

1 Introduction Historically, it has been difcult to simulate complete bioprocess sequences owing to the lack of predictive unit operations models. This has led to the use of empirical methods for process development and design and a lack of computer-based design tools being used in the biotechnology-based industry. With recent pressures to achieve faster development of new products to market and a need to realise higher process yields, there has been increased interest in the use of mathematical models to describe whole process performance. There are now a few software packages available specically for such modelling tasks [5,7,8]. Keywords Simulation, Pareto optimisation, Modelling, Though process simulation packages are becoming Bioprocess, Windows of operation more widely used in the sector, take-up has been rather slow, in part owing to the amount of data required for the List of symbols models and hence the time and resources needed to deowrate of the harvesting centrifuge, L/h Qh velop accurate process simulations. We have been examP pressure of homogenisation, bar (gauge) ining ways of achieving fast parameter estimation and N number of passes in homogenisation generation of design data using so-called ultra-scale Qd owrate of centrifuge for debris removal, L/h down mimics of processing steps with which to acquire S level of debris removal achieved, % critical process parameter values, e.g. [11, 14]. AdditionR overall downstream product recovery ally, we have used graphical methods to map the feasible realised, % space for operation through so-called windows of operY overall process productivity, unit/h ation, for the visualisation of process behaviour subject to dened operating constraints [15, 16, 18]. As the sophistication of unit operation models increases, it will be possible for biochemical engineers to Received: 13 March 2002 / Accepted: 23 January 2003 contemplate the use of models for predictive design and Published online: 1 March 2003 Springer-Verlag 2003 examine the impact of process changes on overall process performance. Ultimately, such an examination will lead to Y.H. Zhou, N.J. Titchener-Hooker (&) consideration of how best to maximise the performance The Advanced Centre for Biochemical Engineering, Department of of a process, given specic process constraints. This task Biochemical Engineering, University College London, Torrington is, however, complicated by the fact that the operations Place, London, WC1E 7JE, UK may be batch or semi-batch, and the mathematical E-mail: [email protected] Tel.: +44 (0) 20 7679 3796 models are then strongly non-linear. The resulting mulThe Engineering and Physical Science Research Council (EPSRC) tiple objective optimisation problem may not necessarily be readily soluble, or no algorithm may currently exist Innovative Manufacturing Research Centre (IMRC) in Bioprocessing is located in the Department of Biochemical Engineering. for its resolution.

Bioprocess Biosyst Eng 25 (2003)

350

Furthermore, the mathematical formulation of such optimisation problems is normally specic to the process and subject to any changes made in the process sequence as well as to the key variables concerned. This leads to the need to reformulate the problem and to develop suitable mathematical algorithms, both of which could be time consuming. In many instances, such a problem is transformed into a scalar substitute problem with respect to the designers preference or strategy [6]. To deal with multiple objectives, weightings are often used to reect the perceived importance of each objective but, since these reect a design engineers preference, they tend to quite subjective. A possible solution to such a multiple-objective optimisation problem is to develop methods that enable the designer to see rapidly how decisions will impact upon performance and, in particular, how the optimal level of output will alter. In the biological industries, regulatory pressures effectively restrict the degree of freedom that the design engineer has to alter a process design so as to achieve optimal process output. Instead, the design engineer requires a capacity to assess the balance of performance and robustness of operation while making design decisions. Therefore, unlike in the chemical sector, where optimisation routines are applied rather widely to identify the global optimum, the bioprocess engineer seeks to understand how the feasible set available for design changes as functions of the operating variables and parameters at their disposal. Optimisation can then be carried out within a fairly well-dened range of operating parameter values so as not to move the process into a region that may violate validated units. In such a process situation, it is crucial that the design engineer knows how the process responds in the vicinity of the optimal solution so as to determine how best to achieve the desired robust solution when subject to process uncertainties and changes. In this respect, conventional search methods of optimisation such as Simplex [1] are not suitable, as they are designed to handle only linear problems and, furthermore, provide only a point solution and no indication as to process sensitivities. These single-objective methods require the determination of the impact of the different process parameters on suitable outputs such as yield or time, which are then reconciled to produce a suitable objective function for the optimisation. This can be problematic in bioprocesses, where it may hard to associate a cost penalty with each process performance criterion. In such cases, past workers have used the Pareto optimal point approach [6] to handle the multiple-objective optimisation problem. Specically, the method has been used in high-performance liquid chromatographic separations by Smilde et al. [12], and in tablet formulation by Bolhuis et al. [2] and Thibault et al. [13]. The Pareto optimal approach [6] is a constraint-oriented method in which the multiple-objective optimisation problem is transformed into a scalar optimisation problem by minimising or maximising one objective with all others remaining constrained. The set of Pareto optimal solutions is then generated by replacing each criterion with a sequence of inequality constraints. A point in the feasible criteria space is called a Pareto optimal point if

there exists no other point in that space which yields an improvement in one criterion without causing a degradation in the other criterion. The Pareto optimal point method provides a functional-efcient solution set with no preference on the criteria, and therefore gives the decision maker the exibility to select the best compromise solution out of such a set. We have applied this approach to bioprocess design through a 2-dimensional (2D) example [18]. In this paper, the approach will be examined further and extended to a high-dimensional situation typical of that generated when considering a more complete bioprocess design problem where the impact of a large number of operating variables must be considered simultaneously. We propose the use of a simple graphical method combined with the Pareto optimisation procedure to provide a fast but reliable method initially to establish the available sets of operating conditions for a given process requirement, then to identify a smaller sub-set of conditions within this space, dened by imposed operating constraints and, nally, to search this region for the best operating conditions. The rst two steps will be achieved through developing a series of windows of operation: graphical representations of the process output subject to prescribed process specications and constraints [15]. For the last step, we propose the integration of this with a Pareto optimisation routine, for the systematic optimisation of the identied feasible space for the bioprocess sequence under examination. The Pareto optimisation method then identies the optimal solution within the selected smaller operating window. In so doing, it is possible to gain a better understanding of how the process responds to changes in process variables and constraints through examining the nature of the windows of operation before carrying out the optimisation. Therefore, a fast evaluation of a process design can be achieved. The subsequent optimisation yields a series of unique solutions, each associated with a particular set of process objectives. Here, we investigate the approach further and quantify both the computational advantages of the method as well as the additional insight gained. We seek to demonstrate the two-part approach through case studies based upon pilot-scale data and veried models of a typical bioprocess sequence.

2 Computational methods The approach developed for the simulation of the integrated bioprocess relies upon the use of veried unit operation models to describe a sequence of a typical bioprocess operations. We have reported elsewhere on the use of each model of the individual operations for the design of a bioprocess sequence (e.g. [16, 17]) where we have used the models to simulate the output performance of a process train for the production of an intracellular product from a yeast organism, a common source of industrially relevant products. A typical process for the production and isolation of a soluble intracellular enzyme from a S. cerevisiae strain was used in this study. The micro-organism is cultured in a fed-batch fermenter. At the end of the fermentation stage,

Y.H. Zhou, N.J. Titchener-Hooker: The application of a Pareto optimisation method

cells are harvested using a high-speed disc-stack centrifuge removing fermentation liquors that might complicate further processing steps. The harvested cell paste is resuspended in a buffer to an appropriate concentration prior to disruption of the cells by means of high-pressure homogenisation to release the intracellular protein. During disruption, cell wall debris will be formed, and this must be removed from the soluble protein mixture by highspeed centrifugation. Subsequent processing would include high-resolution methods to achieve a puried end product but, for simplicity, only the rst part of the process is considered. The process simulation was implemented on a Pentium II PC using LabView (version 4.0, National Instruments, Austin, USA), a graphical programming environment, allowing the construction of a user-friendly interface. The simulation was constructed using the veried models described in the literature data [3, 4, 9, 10, 14] for both the individual unit operations and the integrated processes. The process performance was measured in terms of a total volumetric enzyme productivity, Y (unit activity/unit time) and the levels of debris removal achieved relative to the total amount of solids present in the feedstream, S (%). Other criteria were the concentration of solids in the supernatant leaving each centrifuge step, C, and the total downstream process operating time, t. All these criteria are functions of the operating variables and of the fermentation conditions selected. In practice, one might wish to achieve a high level of productivity but ensure that the level of solids carry-over in each centrifuge step was minimised so as to avoid contamination of later processing stages and a loss in product quality. The total process time is important, since many biological materials are strongly labile. Processing to a stable state in a short time interval can be a critical driver in reaching an operating decision.

from 0.06/h (effectively a doubling of cell mass every 11.5 h) to a growth rate of 0.26/h (doubling of mass every 2.6 h). At this point, the trade-offs in process design can be assessed quantitatively using the windows of operation approach. The next stage in the procedure is to identify the most promising growth rate to use; effectively that which generates the largest window of operation for the given constraints and output specication. The most suitable growth rate of the three simulated was l=0.19/h, since this gave the largest window of operation. In our previous paper [18], we used this space as the basis for determining the Pareto optimal solutions. However, in the general case, the design problem may have a higher dimensionality, and the computational time required for locating the Pareto optimal solutions could be an issue. For this reason, in this paper we carried out a further examination of the feasible region at a growth rate of 0.19/h, seeking a smaller sub-window of operation in which a better level of performance would be achieved by tightening the process specications. This resulted in the generation of a series of smaller windows within the earlier feasible region and formed the basis for the Pareto optimisation procedure. For example, if two levels of specications are chosen as Y1>5106 unit/h, Y2>18106 unit/h and S1>90%, S2>96%, four sub-windows, WI, WII, WIII, WIV, are generated, as shown in Fig. 1. It is clear that WI represents the lowest levels of productivity and debris removal, while WIV shows the highest levels of productivity and debris removal achievable. Hence

351

WI:={(P, Qd)|Y1 Y Y2; S1 S S2} WII:={(P, Qd)|Y1 Y Y2; S>S2} WIII:={(P, Qd)|Y>Y2; S1 S S2} WIV:={(P, Qd)|Y>Y2; S>S2}

3 Results and discussion 3.1 Window of operation approach for process design A set of windows for the process studied can be identied in terms of the acceptable centrifuge owrate range for the removal of cell debris and the pressure of disruption used to release the product held within the cells. The design problem was specied such that two output criteria must be satised: an enzyme productivity of Y>5106 units of enzyme activity per hour of processing time and a level of debris removal of S>90%. Two equipment constraints were applied: the maximum pressure of homogenisation is 600 bar (this being the mechanical limit of the particular machine), and the minimum centrifuge owrate is 20 L/h, below which heating of the process uid is difcult to control. As shown before [18], individual windows corresponding to the region in which this level of performance can be realised can then be generated at the specic growth conditions of fermentation employed at the front end of the process. The results [18] showed how the ability to achieve the design specication changes as the rate of growth of the organisms within the fermenter increases

Figure 1 then gives an indication of the characteristics of each sub-window of operation and enables decisions on which space to focus on for subsequent optimisation to be determined quickly. The sub-window of operation WIV

Fig. 1. Simulated windows of operation for homogenisation and

centrifugation at a single growth rate of 0.19/h sub-windows shown as functions of productivity, Y, and cell debris mov levels, S. WI, 5106<Y<18106 unit/h, 90%<S<96%; WII, 5106<Y<18106 unit/h, S>96%; WIII, Y>18106 unit/h, 90%<S<96%; WIV, Y>18106 unit/h, S>96%

Bioprocess Biosyst Eng 25 (2003)

corresponding to high levels of productivity and of cell debris removal was therefore chosen for the subsequent optimisation.

352

3.2 Pareto optimal solutions for process design Using the sub-window of operation (WIV) identied above, we can now investigate the use of the Pareto method for determining the location of the optimal operating points within the feasible regions. This was done through a series of linked studies built from the denition of the total feasible set through to the determination of the max Y 4 S100 Pareto optimal points for particular operating strategies. k Y ;S2W The process implications of the results will be discussed where k is a lower bound and 100% is the high bound for each case. for S. Increasing k from 96% to 100% gradually, the set of 3.2.1 Pareto optimal points was achieved by solving problem Case I The optimisation problem considered here was to maxi- (4). The results are displayed in Figs. 2 and 3. Figure 2 mise both the productivity and the level of debris removal illustrates the Pareto optimal points in terms of the opby manipulating the set-points of selected operating vari- erating conditions. Not unexpectedly, the required operating conditions are located along the boundary of the ables while keeping all other unit operations xed. Two window of operation. The level of process performance criteria, those of productivity, Y, and the level of debris removal achieved, S, were considered subject to two op- corresponding to these points is shown in Fig. 3. This erating variables; homogeniser pressure, P, and centrifuge demonstrates clearly the compromise between the levels of productivity and debris removal achieved. If the process owrate, Qd. operation moves from point A (P=480 bar, Qd=22 L/h) Referring to Fig. 1, we denote by W0 the operating conditions available to start with for design, i.e. the initial toward point B (P=600 bar, Qd=22 L/h) in Fig. 2, i.e. from low owrate and the medium pressure to the highest feasible operating condition set satisfying the minimum pressure achievable for this homogeniser, then the level of specications of Y and S, that is Y1 and S1. Referring to Fig. 1, the operating conditions available to start with for debris removal (as shown in Fig. 3) is very high, above 98%, while the productivity increases by 10%. These design, i.e. the feasible operating conditions set, were: changes are shown by the move from A* to B* in Fig. 3. In X0  fP; Qd j175  P  600; 20  Qd  200g; X0 2 R2 Fig. 2, moving from B to C requires no increase in pres1 sure and results in a small rise in productivity but an appreciable decrease in the level of debris removal. Figure 3 shows that, for an increase of 6% in the level of Setting the minimum process performance as the pro- productivity achieved by moving from B* to C*, the degree ductivity Y=18106 unit/h and the level of debris removal of debris remaining in the supernatant increases fourfold. of S=96%, the window of operation approach tightened the feasible region down to the sub-window IV:
X  fP; Qd j400  P  600; 20  Qd  80g; X 2 R2 2 The performance set is then a projection of the feasible set, W, dened by w : fY ; SjY f1 P; Qd ; S f2 P; Qd ; P; Qd 2 Xg 3 where f1 and f2 are functions associated with the whole process characteristics with strong non-linearity and are not known explicitly. However, w can be achieved easily as the output of the process simulations. Computationally, the generation of the output set w and the solution set using W was much faster than that using W0. This would be particularly true when the dimension of the feasible set becomes larger, since the higher the dimension, the more simulations are needed, and the greater the data set for the corresponding Pareto optimal search

becomes. Furthermore, reducing the size of the search space to W enabled simulations with small intervals in terms of the values of the operating variables to be performed so as to achieve a more precise location of the optimal solutions results. In this case study, the intervals of the operating variables used in the simulation were fDP 50 bar; DQd 20 L=hg for generating X0 ; and fDP 10 bar; DQd 2 L=hg for X respectively. We chose to select the productivity as the design cost function and the level of debris removal as the constraint. This may be represented as:

Fig. 2. The Pareto optimal point-set for the pressure of homogenisation and the ow rate of centrifugation within WIV requested to achieve a minimum performance of Y>18106 unit/ h, S>96%. Points A, B and C have been identied on the subsequent plot (Fig. 3) to show the matching performance levels that result from operating at these conditions

Y.H. Zhou, N.J. Titchener-Hooker: The application of a Pareto optimisation method

Fig. 3. The Pareto optimal solutions of productivity and debris

removal corresponding to Fig. 2. The process performances at points A*, B* and C* were achieved using the operating conditions at points A, B and C respectively in Fig. 2

Point B* is clearly critical in this design (P=600 bar, Qd=22 L/h,) since the rate of increase in productivity falls away from B* towards point C*, as does the ability of the centrifuge to remove efciently the contaminating cell debris. Based on the simulation results, a series of alternative design strategies can therefore be adopted with decisions made directly by consulting Figs. 2 and 3.

3.2.2 Case II To demonstrate the generality of this approach, a more complicated problem is considered next. It shows the common case in the bioprocess design where more variables are involved. The key variables were the growth rate of the fermenter, l, the owrate of the harvesting centrifuge, Qh, the pressure of the homogeniser, P, the number of passes of recycle used in homogenisation, N, and the centrifuge owrate employed for debris removal, Qd. The working volume of the fermenter was assumed to be 100 L. The process performance was again measured in terms of process productivity, Y, and the level of debris removal achieved, S. The minimum process performance was set at a productivity of Y=18106 unit/h, and the level of debris removal at S=96%. This was done in order to yield a large initial feasible set for the search. Two process constraints were considered in turn. First, the solids concentration in the supernatant stream leaving the centrifuge, C, must be less than 1 g/L in order to minimise the load on any subsequent polishing step needed before further purication, such as chromatography, can be carried out satisfactorily. Subsequently, we examined the impact of limiting the total downstream process operating time, t, to be less than 5 h so as to accommodate the downstream processing within a normal working shift. (An alternative scenario would be to consider the time-dependent effects such as product denaturation and to determine a maximum process time based upon an acceptable level of loss in product activity.) The simulated results for achieving Y, S, C and t were obtained by running the process simulation in the chosen range of the operating conditions. Initially, Pareto optimal points satisfying the concentration requirement were calculated for each growth rate, irrespective of total process time. As expected, the choice

of growth rate was dominated by l=0.19/h for which there were 30 Pareto optimal points. We found a small set of points was obtained for l=0.12/h but that the performance was much reduced compared with l=0.19/h. There were no solutions available for other growth rates. Data for the higher growth rate only were used in the subsequent analysis. Constraining the problem by addition of a process operating time requirement of less than 5 h reduced the set of Pareto optimal points to 27, as shown by the open symbols in Fig. 4. The disqualied solutions satised the solid concentration constraint but did this by operating at low owrates for cell debris removal. This leads to a long overall processing time and is the principal contributing factor for the solutions violating the time constraint. The operating conditions corresponding to the identied Pareto optimal points are tabulated in Table 1. The highest productivity achievable is then 21.2106 unit/h. The trade-offs between the objectives are obvious, with the level of debris removal falling off as the productivity demands are increased. The Pareto optimal solution points obtained show the unique combinations of productivity and debris removal that may be achieved. At low levels of productivity, the degree of cell rupture necessary yields a cell mass containing a high proportion of undisrupted cells and little extent of micronisation. As a result, the level of solids removal approaches the limits of centrifugal separation of about 0.2% on a mass basis for a dilute yeast suspension. As the productivity demand increases, so the level of solids carryover in the centrifuge supernatant rises, reecting the smaller size of cell fragments generated at the more extreme conditions of cell disruption. Figure 4 quanties the exact relationship for a particular set of design objectives so that the desired process performance can be identied quickly and the corresponding operating conditions obtained from the look-up table (Table 1). The centrifuge owrates for cell harvesting, Qh, are all in the range of 200400 L/h. For cell dis-

353

Fig. 4. Pareto optimal solutions for Y>18106 unit/h and S>96%.

Five operating variables (growth rate in fermentation, the owrate of centrifuge in cell harvest and debris removal, pressure and number of passes of disruption) are considered simultaneously. The solid data points are for those solutions that meet the further constraint of a total process time limit of 5 h. Open symbols are for those solutions which violate the time limit constraint (growth rate 0.19/h)

Bioprocess Biosyst Eng 25 (2003) Table 1. The results of Pareto

optimal solutions and corresponding operating conditions for Case 1 where, by identication of a subwindow of operation for the optimisation problem, the growth rate has been xed to 0.19/h

Performance Productivity Y (106 unit/h) 18.2 18.2 18.3 18.3 18.4 18.6 18.6 19.0 19.1 19.1 19.2 19.4 19.4 19.5 19.5 19.9 19.9 20.0 20.1 20.2 20.2 20.3 20.4 21.0 21.0 21.2 21.2 Debris removal S (%) 99.8 99.8 99.7 99.7 99.7 99.7 99.7 99.6 99.6 99.6 99.5 99.5 99.5 99.3 99.3 99.3 99.3 99.1 99.1 99.1 99.0 99.0 99.0 98.8 98.8 98.4 98.4

Constraints Solid conc. C (g/L) 0.11 0.11 0.17 0.17 0.19 0.19 0.19 0.24 0.24 0.24 0.29 0.29 0.29 0.39 0.39 0.4 0.4 0.53 0.53 0.53 0.56 0.56 0.56 0.68 0.68 0.95 0.95 Processing time t (h) 4.9 4.8 3.7 3.6 4.6 4.5 4.4 4.1 3.9 3.8 4.9 4.7 4.6 3.5 3.4 4.9 4.8 4.6 4.5 4.4 3.9 3.7 3.6 4.9 4.8 3.5 3.4

Operating variables Flowrate Qh (L/h) 300 400 300 400 200 300 400 200 300 400 200 300 400 300 400 300 400 200 300 400 200 300 400 300 400 200 400 Pressure P (bar) 350 350 350 350 400 400 400 350 350 350 400 400 400 400 400 400 400 450 450 450 400 400 400 450 450 450 450 No. of passes N 8 8 7 7 6 6 6 8 8 8 7 7 7 6 6 8 8 6 6 6 7 7 7 8 8 6 6 Flowrate Qd (L/h) 40 40 60 60 40 40 40 60 60 60 40 40 40 60 60 40 40 40 40 40 60 60 60 40 40 60 60

354

ruption the pressure is situated at a medium level, while the number of passes tends to higher values. This is the opposite of normal experience, where homogenisers usually operate at a high pressure and medium number of passes. This simulation result reects the competing processes of product release and debris micronisation. From a design perspective, it is desirable that the former be maximised but that the latter be minimised. The integrated model predicts that, at a moderate pressure, the extent of micronisation is lower for a given degree of product release than would be achieved under operation at higher pressures which result in intrinsically ner debris irrespective of pass number. Since the ease of cell debris removal is related to the square of particle size, the more rapid reduction in particle diameter associated with the higher pressure operation inevitably leads to a fast falling off in separation efciency and is reected in a lower S value and an increase in process time. The centrifuge owrate for debris removal, Qd, is maintained above the lowest permissible value of 20 L/h, reecting the major inuence that this variable has on the overall process time. A minimum downstream processing time of 3.4 h which satised all the specications was obtained. While all the optional solutions follow a common trend, it is only by referral to the associated performance and operating variable table (Table 1) that one can see fully the impact of imposing such process constraints. For example, in Fig. 4, maintaining a constant ~99% level of debris removal can be achieved with only a small change in productivity. However, as can be seen from the highlighted section in Table 1, this can only be realised by making

signicant changes in the set points of key operating variables, for example perhaps necessitating an increase in the centrifuge owrate for cell harvesting by 100% from 200 to 400 L/h. By contrast, the overall process time may only be reduced by 6% in such a process change. Such insight is vital for maintaining efcient levels of operation, and demonstrates the levels of sensitivity of bioprocesses to changes in variables and also the non-linear nature of the processes. Use of conventional optimisation approaches would not make such interactions obvious and provides an additional reason why the Pareto method is particularly suited to bioprocess designs where the impact of process trade-offs must be quantied and visualised. (Identifying the sub-window of operation for l=0.19/h would have reduced the size of the search space and the dimension of the problem by one variable and thereby increased the speed of obtaining the optimal solutions, as will be discussed later.)

4 Computational effort The computation involved in the Pareto optimisation algorithm is mainly dependent on the matrix element selection. With a high-dimension problem and/or large-scale matrix, the computation time will increase linearly. In this paper, we investigated the effect of reducing the computational effort by rst generating a suitably small window of operation and then calculating the Pareto optimal solution set within this. Two case studies were carried out to examine whether it is worthwhile computationally to use the combination of the windows of operation followed by

Y.H. Zhou, N.J. Titchener-Hooker: The application of a Pareto optimisation method

the Pareto search, especially for a high-dimensional case. References 1. Beale EML (1988) Introduction to optimisation, Mackley L (ed). The computation effort was measured by the CPU time John Wiley, Chichester, pp 3948 used. 2. Bolhuis GK, Duineveld CAA, de Boer JH, Coengracht PMJ (1995) The rst example used the data from Case I, where the Simultaneous optimisation of multiple criteria in tablet formuladimension was 2 and the size of the simulation data set tion: Part II. Pharm Tech Europe 7(8):4251 3. Bulmer M, Clarkson AI, Titchener-Hooker NJ, Dunnill P (1996) was small. When the Pareto optimal solution searched Computer-based simulation of the recovery of intracellular enfrom the entire window of operation, W0, the CPU time zymes and its pilot-scale verication. Bioprocess Eng 15:331337 needed was 3.08 s. However, when using the smaller 4. Clarkson AI, Bulmer M, Titchener-Hooker NJ (1996) Pilot-scale window of operation, W, to search for the Pareto optimal verication of a computer-based simulation for the centrifugal solution, the CPU time for achieving the reduced window recovery of biological particles. Bioprocess Eng 14:8190 5. Cooney CL, Petrides D, Barrera M, Evans L (1988) The impact of and the subsequent Pareto optimal solution search was chemistry on biotechnology. Computer-aided design of a bio0.22 s. So by reducing the search to a sub-region, the chemical process, Chap. 5. ACS Symposium Series 362. American computational effort was reduced by 93%. Chemical Society, Washington DC The second example (Case II) comprised a 5-dimen6. Eschenauer H, Koski J, Osyczha A (1990) Multicriteria design optimisation: procedures and applications. Springer, Berlin Heisional design with a large associated simulation data set. delberg New York Without division of the window of operation, the total 7. Evans LB, Field RP (1989) Bioprocess simulation: a new tool for CPU time was 1498 s. The CPU time for the problem obprocess development. Chem Eng Aust 14:2023 tained by rst generating a sub-window of operation 8. Petrides D, Cooney CL, Evans LB, Field RP, Snoswell M (1989) Bioprocess simulation: an integrated approach to process develcorresponding to the data at a growth rate of 0.19/h, and opment. Comput Chem Eng 13:553561 thereby reducing the size of the computational problem, 9. Siddiqi SF, Titchener-Hooker NJ, Ayazi Shamlou P (1996) Simuwas 4.54 s, a reduction of 99.7% of the CPU demands of lation of particle size distribution changes occurring during highthe original problem. These data reect the value of repressure disruption of bakers yeast. Biotechnol Bioeng 50:145 ducing the size of the windows of operation prior to im150 10. Siddiqi SF, Bulmer M, Ayazi Shamlou P, Titchener-Hooker NJ plementing the Pareto optimisation search.
(1995) The effects of fermentation conditions on yeast cell debris particle size distribution during high pressure homogenisation. 5 Bioprocess Eng 14:18 Conclusions 11. Siddiqi SF, Titchener-Hooker NJ, Ayazi Shamlou P (1997) High pressure disruption of yeast cells: the use of scale down operations We used veried models to simulate the performance of a for the prediction of protein release and cell debris size distritypical bioprocess sequence. The output of these simulabution. Biotechnol Bioeng 55:642649 tions was displayed in the form of 2D feasible spaces, 12. Smilde AK, Knevelman A, Coenegracht PMJ (1986) Introduction windows of operation, which provide the starting point for of multi-criteria decision-making in optimization procedures for high-performance liquid chromatographic separations. J Chrolocating the set of optimal conditions of operation. We matogr A 369:110 used a multiple objective optimisation method, that of J, Taylor D, Fonteix C (2001) Multicriteria optimization Pareto optimal points, to search for the best combinations 13. Thibault for the production of gluconic acid. In : Proceedings of the 8th in operating conditions required to satisfy a series of siInternational Conference on Computer Applications in Biotechmultaneous process objectives. This search can be made nology, Quebec City, 2427 June. Elsevier, Amsterdam more efcient by reducing the size of feasible space by rst 14. Varga EG, Titchener-Hooker NJ, Dunnill P (1998) Use of scaledown methods to rapidly apply natural yeast homogenisation identifying a series of sub-windows, each with different models to a recombinant strain. Bioprocess Eng 19:373380 process characteristics. A sub-window can then be selected 15. Woodley JM, Titchener-Hooker NJ (1996) The use of windows of by the user as the basis for the search for the optimal operation as a bioprocess design tool. Bioprocess Eng 14:263268 solution set. This two-phase approach was demonstrated 16. Zhou YH, Holwill ILJ, Titchener-Hooker NJ (1997) A study of the computer simulation for the design of integrated downstream to provide the design engineer with a fast means of pinbioprocess. Bioprocess Eng 16:367374 pointing optimal levels of process performance and the 17. Zhou YH, Titchener-Hooker NJ (1999) Visualising integrated corresponding operating set-points with signicant savbioprocess designs through windows of operation. Biotechnol Bioeng 65:550557 ings in computational effort. The simple approach can be 18. Zhou YH, Titchener-Hooker NJ (1999) Simulation and optimisaimplemented easily and provides a rapid and reliable tion of integrated bioprocesses: a case study. J Chem Technol. method for bioprocess optimisation. Biotechnol 74:289292

355