The main focus of this thesis has been the development of robust and predictive models describing the downstream processing of an intracellular protein product using membrane filtration. Verification experiments have been conducted to test the accuracy of model predictions and the models developed have been encoded within SPEEDUP (AspenTech Ltd., Massachusetts, U.S.A.) to demonstrate the application of process simulation within a bioprocess environment. The main conclusions of the studies will now be drawn together.
9.2 Modelling techniques
The initial phase of this study focused on the available approaches to model crossflow membrane filtration for use in a typical bioprocess flowsheet. Most modelling of membrane separation bioprocesses is achieved using a concentration polarisation model or a resistance model. A concentration polarisation model is only applicable in the pressure independent region and a resistance model is only applicable in the pressure dependent region. All experiments conducted in this study reside in the pressure independent region and for this reason a polarisation model was used for predictive purposes. Models based on a polarisation approach were developed by characterising the suspension properties and determining model parameters by conducting single microfiltration experiments for a given membrane pore size. The benefits of using such an approach include a significant reduction in process development time allowing for the rapid piloting of bioprocesses. However, the reliability of the models developed is questionable, as was demonstrated in Chapter 4. This stems from the specific nature of bioprocess feed-streams and their interaction with the membrane. A lack of a true understanding of the prevailing fouling mechanisms means that empirical parameter determination should not be based on single microfiltration experiments. The minimum number of experiments required to develop a robust model will depend on the number of input variables and also the range of variable settings. Such features point the model developer towards a statistical approach to modelling.
Statistical approaches have also been used to model membrane filtration bioprocesses but are less common than polarisation models. Such models use a 'black box' approach by developing an empirical linear or non-linear relationship between n input variables and m output variables. A statistical model based on linear regression was used to describe the permeate flux properties of ceramic membranes during the microfiltration of yeast whole cells. The main benefits of using such an approach include the reliability and accuracy of the model predictions and the ease of implementation. However, statistical models tend to be system specific and require a high degree of experimentation to train the simulator. They also tend to be scale dependent.
In conclusion, it is probably most appropriate to model crossflow membrane filtration systems using a statistical approach. The reasons for using such an approach include
• a high degree of specific interactions between membranes and typical bioprocess feed-streams make the development of theoretical models difficult
• the ease of implementation of statistical models not requiring the estimation of intermolecular forces, particle size distributions, suspension rheological properties and diffusion coefficients which are difficult to measure
• accurate models as a result of continuous updating by including information from all experiments
• the robust nature of models within the confines of variable settings
Statistical models give no indication of possible fouling or separation mechanisms occurring in filtration modules. The behaviour of biological feed-streams in high shearing devices will depend largely on their viscoelastic properties. Newtonian fluids are easier to model in such environments. Non-Newtonian suspensions are less predictable. The benefits of using fluid mechanics to develop rigorous flow models include an understanding of the flow patterns in membrane filtration devices which may explain possible fouling mechanisms. Such models may also explain the interaction between the particle size and the membrane pore size and the importance of hydrodynamic forces in the boundary layer with regard to re-suspended flowing cakes or static non-flowing cakes, the cake thickness and porosity. The development of rigorous flow models will make for models which are scale independent and membrane module geometry independent. Thermodynamic forces such as intermolecular forces and entropie pressure will only
be significant if they match or exceed hydrodynamic forces. Such models may elucidate the effects of certain components such as colloidal proteins, lipids, anti-foam agents and other soluble components on the filterability of biological suspensions.
9.3 Physical property measurements and model development
The possibility of using certain physical property characterisations including particle size measurements and viscosity measurements alongside single microfiltration experiments to determine permeate flux rates and soluble product transmission was examined in Chapters 5, 7 and 8. Using the particle-to-pore size ratio (k) and information from Chapter 4, it was concluded that an empirically determined concentration polarisation model based on single microfiltration experiments should be applicable to yeast homogenate for the range of X studied. Verification studies confirmed the applicability of the polarisation model to the yeast homogenate system. Similar results were obtained for polyethyleneimine (PEI) treated yeast homogenate. Yeast whole cells and yeast homogenate displayed an apparent linear response to k with respect to permeate flux rates and soluble product transmission. However, PEI treated yeast homogenate displayed curvature and the best membrane performance occurred at a value of ^ = ~ 10. This is in complete contradiction to studies conducted in Chapters 4 and 5, and also studies conducted by Kawakatsu et a i , (1993) where the worst membrane performance occurred at a value of A, = ~ 10. Thus, it is erroneous to conclude that compressible systems will display inferior membrane performance at values of A = ~ 10 (Kawakatsu et al., 1993). This is a direct consequence of the specific nature of the filtration of biological systems.
Viscosity measurements provide a means of determining the changes in the steady-state permeate flux rate as a function of the cell concentration. However, viscosity measurements give no indication of membrane filtration performance when the host system is changed. This was demonstrated in Chapter 8 where equating the system viscosities but changing the biological system produced vastly different permeate flux rate and transmission properties.
In conclusion, the use of physical property characterisations to predict membrane performance is limited by the highly specific nature of biological feed-streams and their interaction with the membrane.
9.4 Simulation studies
Simulation studies provide a means of studying the effect of a range of design and operating variables on the performance of unit operations. The benefits of using such techniques are enormous and have been highlighted in Chapter 6. Simulations studies were conducted on a 3- stage filtration process for the recovery of alcohol dehydrogenase (ADH) from yeast homogenate. The studies assessed the impact of the recirculation rate, the membrane module length, the starting cell concentration and diafiltration volumes on the product yield and product purity. The models used were those developed in Chapter 5.
To be of real benefit to the bioprocess engineer, process simulation studies should be conducted in conjunction with pilot-scale verification trials. The ability to use simulations to direct pilot-plant research activity is crucial. It is also important that the models developed are scaleable and any sources of inaccuracies quickly found and rooted out. This powerful combination of process simulation and pilot-scale verification will allow for the rapid development of near optimal process routes. Pilot-scale verification of the models developed in this thesis have not been conducted because of a lack of appropriate process-scale equipment.
9.5 Overall conclusions
The development of robust and predictive models for the downstream processing of complex biological feedstocks using membrane filtration has been demonstrated. Such models are empirical in nature because of specific interactions between feed-streams and the membrane which are poorly understood. The unit operation models developed require single experiments to determine empirical parameters. However, the reliability of such an approach is questionable. Statistical models, on the other hand, provide more accurate and reliable predictions but require several experiments to develop and are probable not scaleable. Simulations provide a useful means of understanding fully the implications of design and operating parameters on unit operation performance. The benefits of such techniques in the context of whole processes has been demonstrated by several other workers in the research group (Clarkson, 1994, Zhou et al.,
membrane processes allow continuous separations in sterile environments
membrane operating regimes are less harsh than alternative unit operations such as centrifugation allowing for complete bio-containment
the product fraction is devoid of any particulate material if membrane pore sizes are selected appropriately simplifying subsequent downstream operations
feedstock pre-treatment may allow for competitive membrane processes i.e. permeate flux rates in excess of ~ 100 L m'^ h'^ (Warren et aL, 1991)
simulations of bioprocesses provide a suitable medium for teaching undergraduate students, postgraduate students and industrial delegates.