Feedback-Optimal Controller for Cultivation Processes

Adaptive Predictive Control by Open-Loop-

Feedback-Optimal Controller for Cultivation

Processes

by

Meijie Li

A Thesis submitted in partial fulfillment

of the requirements for the degree of

Doctor of Philosophy

in Biochemical Engineering

Approved Dissertation Committee

Supervisor:

Prof. Dr. Mathias Winterhalter

(Jacobs University Bremen)

Supervisor:

Prof. Dr.-Ing. Volker C. Hass

(Hochschule Furtwangen University)

Reviewer:

Prof. Dr. h. c. Roland Benz

(Jacobs University Bremen)

Reviewer:

Dr. Florian Kuhnen

(Hochschule Bremen)

Date of Defense: 16.10.2015

Statutory Declaration

I, Meijie Li hereby declare that I have written this PhD thesis independently, unless where clearly state

otherwise. I have used only the sources, the data and the support that I have clearly mentioned. This

PhD thesis has not been submitted for conferral of degree elsewhere.

I confirm that no rights of third parties will be infringed by the publication of this thesis.

Bremen, August 20, 2015

Signature __________________

i

Acknowledgements

The work presented in this thesis, was performed during my engagement as a research associate

at the Institute of Environmental and Biotechnology at the University of Applied Sciences

Bremen (Hochschule Bremen). Having this opportunity, I would like to express my deepest

gratitude to the people who have provided me with professional advices and moral support.

First of all, I would like to thank Prof. Dr. Matthias Winterhalter, who has offered excellent

mentoring, many suggested improvements of this thesis and participated in the evaluation of

this work. Next, I would like to thank Prof. Dr. Volker Hass, who has provided me the

opportunity to work with his outstanding research group and many inspirational insights

throughout my research work. Many thanks to Prof. Dr. Roland Benz for reviewing my thesis

and being a member of my dissertation committee. A special thanks to Dr. Florian Kuhnen,

who has provided me with continuously guidance and participated in numerous fruitful

discussions in the research area.

I would also like to thank all of my colleagues who have been very helpful and provided a good

working atmosphere at the institute.

Last but not least I would like to thank my family and friends, for their unwavering support

during the many long days which went into this endeavor.

ii

Abstract

The development of a safe and resource efficient production process is probably the most

important step in the value chain starting from a biomolecular product to a final market product.

Commonly, bioprocess optimization requires numerous experiments, which can be time and

material costly. Predictions based on calculations can reduce the manufacturing experimental

costs, assist in reaching the production of the desired product to a high concentration with high

productivity and yield. Process description, optimization and control using mathematical

models are an innovative and efficient approach in the process development.

On the other hand, the development of mathematical models is also time and cost consuming.

The use of a highly adaptive general model for bio-processes can reduce the time for model

development. Because it takes only the adjustment of model parameters to achieve the

adaptation of a process model to a new organism or product. Development of a general process

model with high adaptability was one of the main goals of the project "ProTool"*. A general

process model is not bound to a certain cell line and microorganism, but adapts itself to a variety

of organisms and different scenarios. This process model will be used for data interpretation,

process monitoring, recipe optimization and the verification of control concepts as well as the

basis of a virtual representation of the process (i.e. training simulator).

Correspondingly, the development of an advanced controller to achieve an optimized process

control adapting the general model, which is named as the open-loop-feedback-optimal

(OLFO) controller, was another main goal of the project "ProTool". The development of the

OLFO controller is the core task of my dissertation. A structured model consisting of four

biomass compartments is working as the general model which is able to describe the cultivation

of different organisms. Based on this model, online parameter identification is carried out

periodically with the actual process data collected from the plant and laboratory analysis. And

the new estimated process parameters and the updated process state variables are used in

calculating an optimal control profile. The model parameters are updated at desired period

using the extended database to reduce the plant model mismatch to improve the performance

of the optimizer.

* This work is funded by the BMBF within the cluster project BioKatalyse2021, item 0315169(A-E) Catalyse 2021.

iii

This work outlines the advantages to apply the OLFO controller in fermentation processes.

First, it can be used to work with the virtual bioreactor as a training tool for students and staffs

in industry to gain insights into fermentation processes and process control strategies which

can significantly shorten the training cycle and training cost. Second, the OLFO can provide

calculated optimized result for the real fermentation process to reach its maximum productivity

and high standards of product purity with minimum development and production cost. Finally,

the OLFO controller also shows the potential to function as a software observer to detect the

key processes state variables, e.g. biomass, substrate and product concentration, which are

normally measured offline by using the most commonly selected process variables, e.g. pH

values, O2 and CO2 composition in off gas flow etc., which can be readily measured online

using standard sensors. The results obtained in both simulation and real processes show the

efficiency of the OLFO controller for online fermentation process control.

iv

Contents

1 Introduction 1

1.1 Motivation and objective 2

1.2 State of art 4

1.2.1 Development of yeast fermentation technologies 4

1.2.2 Development of control strategies for fermentation processes 5

1.2.3 Soft Sensors 10

2 Laboratory set-up 14

2.1 Organisms and media 14

2.2 The Bioreactor 14

2.3 An overview of the variables system 16

2.3.1 Direct inputs to the fermentation unit 17

2.3.2 Measured quantities of the fermentation unit 18

2.4 Technical aspects of the OLFO embedding 19

2.4.1 Establishment of the OLFO controller structure 20

2.4.2 Options of communication with the process control system 21

3 The OLFO controller 22

3.1 General process model 22

3.1.1 The biological submodel 25

3.2 Basic theory of parameter identification 29

3.3 Optimization criterion 31

3.4 Work flow of the OLFO controller 33

4 Research results and application examples of OLFO in fed-batch processes 36

4.1 Basic research on parameter identification and optimization 36

4.1.1 Different sampling time of offline measurements 37

v

4.1.2 Different weighting factors 39

4.1.3 Absence of partial measurements 40

4.1.4 Different initial values and boundaries 41

4.1.5 Influence of each parameter on the optimization 41

4.1.6 The performance of the OLFO controller 42

4.2 OLFO works with the virtual bioreactor 43

4.3 OLFO application for yeast fed-batch cultivation 48

4.4 Software observer 58

5 Summary and Outlook 63

5.1 Summary of the overall work 63

5.2 Perspective on future OLFO controller developments and

applications

64

Appendix 1: Performance of parameter identification when lack of partial

measurements

65

Appendix 2: Performance of the OLFO controller as a software observer 67

Appendix 3: Data processing toolbox with OpenOffice 72

References 78

vi

List of Figures

Figure Description Page

1.1. Scheme of the soft sensor as defined in reference (Luttmann et al., 2012). The

figure depicts only one hardware sensor, although in reality there can be several

of those.

11

1.2. Schematic depiction of the soft sensor implementation as described in reference

(Warth et al., 2010).

13

2.1. Instrument of a laboratory stirrer tank bioreactor. Adapted from reference (B.

Braun Biotech International GmbH).

15

2.2. Cultivation system in laboratory (includes a stirred tank bioreactor, a computer

with the OLFO embedding in process control system and a control unit).

16

2.3. A scheme depicting the variables of the whole system. 17

3.1. Basic structure of the OLFO controller. Three elements: a process model, a model

parameter identification and an optimization part. )(ˆ)( txCtCx is minimized

to estimate the parameters. Based on the identified model, optimal control profiles

are calculated in the optimization part and transferred to the bioreactor and

process system. Adapted from reference (B. Frahm et al., 2002).

23

3.2. : Structure of the compartment model (primary biomass (XPR): protein, DNA

and RNA; secondary biomass (XSE): enzymes for synthesis of secondary

metabolites; structured biomass (XS): membrane, cell wall and complex

polysaccharide; inactive biomass (XI): defect enzyme, DNA and RNA; G:

glucose; Eth: ethanol; O2: oxygen; CO2: carbon dioxide; HE: yeast extract; P:

product.). The figure was redrawn from reference (Witte, 1996).

24

3.3. The substrate flux is the key element of the "Lyx" process model. A total substrate

consumption rate rS limits the speed of all subsequent reactions. Distribution

functions describe the substrate flux flows into different metabolic pathways,

depending on the actual state.

25

3.4. A scheme of the transition function. 27

3.5. Work flow of the OLFO controller. When new data becomes available from the

bioreactor, an automated parameter identification takes place, and then the

optimization part calculates the optimal control profiles based on the

34

vii

identification model. Finally, the control profiles are transferred to the direct

control level.

4.1 Parameter identification result with different sampling time. S-exp. is the

experimentally measured substrate concentration, S-sim. is the related parameter

estimation result. P and X are the product and biomass concentrations

respectively. There are three groups of simulation results shown here: the first

group simulation with P, X, S every 30 minutes by thickest lines, the second

group simulation with P, X, S every 60 minutes and the third group simulation

with P, X every 60 minutes, S every 180 minutes by thinnest lines. All of the

simulation results fit the measurements well, without significant difference.

38

4.2 The parameter identification result without individual weighting factors. The

experimental data (referred to as pixels) are connected with the corresponding -

best matching - model data (solid as lines) are shown together. The upper part

shows S: substrate, P: product and X: biomass; the lower part shows the

corresponding simulation and experimental data of off gas composition. Without

appropriate weighting factors, the model cannot fit the measurements.

39

4.3 The best parameter identification result with suitable individual weighting

factors: 2CO : 0.1,

2O : 0.1, S : 0.001, P : 1.0, X : 1.0. is the weighting

factor for each individual variable which depends on its dimension and reliability.

The experimental data (referred to as pixels) are connected with the

corresponding - best matching - model data (solid as lines) are shown together.

The upper part shows S: substrate, P: product and X: biomass; the lower part

shows the corresponding simulation and experimental data of off gas

composition. By putting the appropriate weight on certain state variables renders

an obviously better description of parameter identification.

40

4.4 The graphical user interface of the virtual bioreactor. 44

4.5 The user interface of the virtual bioreactor to set the interval of sampling, the

interval to generate the measurement and the beginning of the real process time.

45

4.6 (Left) Parameter identification graphs with 2.5 hours interval, with corresponding

(right) optimization of the estimated models (upper) and the optimal feeding

profile (lower). Substrate (S), product (P) and biomass (X) concentrations are

further labeled based on their generated source virtual bioreactor (experimental)

or OLFO (simulated).

47

4.7 Precalculated optimization results based on the calibrated model from a previous

fermentation experiment. The upper part shows the optimized feeding rate fS and

49

viii

the inlet gas rate fgas; the lower part shows the corresponding prediction of

substrate S and product P and volume of the vessel VR. Only the profile of the

first 2.5 hours was used in the fermentation as the first approve of optimization

result from OLFO controller.

4.8 The first parameter identification result after 4.5 hours. The experimental data

(referred to as pixels) are connected with the corresponding - best matching -

model data (solid as lines) are presented together. The upper part shows S:

substrate, P: product and X: biomass; the lower part shows the corresponding

simulation and experimental data of off gas composition. The simulation data and

the measurements are in a good agreement.

50

4.9 Optimization after the first parameter identification. The upper part shows

optimized profile for substrate feeding rate ' fS.opt ' and inlet gas flow rate '

fgas.opt ', the lower part shows the prediction of product concentration ' P.opt ',

substrate concentration ' S.opt ' and the volume of the nutrient in the fermenter '

VR.opt '. The optimized feeding rate ' fS.opt ' was activated to control the substrate

pump rate from the 5th hour.

51

4.10 The second parameter identification after 6.5 hours. The experimental data


model data (solid as lines) are shown together. The upper part shows S: substrate,

P: product and X: biomass; the lower part shows the corresponding simulation

and experimental data of off gas composition. The simulation data and the

measurements still remained in a good agreement.

52

4.11 Optimization after the second parameter identification. The upper part shows




VR.smo '. This optimization result shows a high deviation from the previously

one. It was discarded after evaluation.

53

4.12 The third parameter identification after 8.5 hours. The experimental data (referred

to as pixels) are connected with the corresponding - best matching - model data

(solid as lines) are shown together. The upper part shows S: substrate, P: product

and X: biomass; the lower part shows the corresponding simulation and

experimental data of off gas composition. The simulation data and the

measurements remained in a good agreement.

54

4.13 Optimization after the third parameter identification. The upper part shows


55

ix



VR.opt '. The optimization also fitted the current state of the fermentation. Then

the optimized feeding rate ' fS.opt ' became activated to control the pump rate.

Particularly, the switching time was used. The process was turned into anaerobic

phase at about the 9th hour.

4.14 The fourth parameter identification after 10.5 hours. The experimental data


model data (solid as lines) are shown together. The upper part shows S: substrate,

P: product and X: biomass; the lower part shows the corresponding simulation

and experimental data of off gas composition. The simulation data and the

measurements remained in a good agreement.

56

4.15 Optimization after the fourth parameter identification. The upper part shows




VR.opt '. The resulting optimal control profile does not deviate significantly from

the previous one.

57

4.16 The top figure present offline measurements of the whole process. It can be seen

that the constraint of the maximum substrate concentration of 10 g/L is violated

between about 12th h to 20th h. Further on, one can also see that the prediction of

the substrate consumption with the manual optimization is correct. At the end of

the cultivation, the substrate concentration is reduced to below 1 g/L. The lower

figure presents the actually applied substrate feeding rate during the whole

fermentation process. After 2.5th h (the end of the batch phase), the precalculated

profile is activated. This was modified at about 5th h due to the correction of the

optimal control. After 9th h, it was modified to a constant rate as the process has

been switched to the anaerobic production phase according to the optimal control.

It must be considered that, the substrate concentration in the feed has increased

from 200 g/L to 610 g/L at the switching time. The reduction of substrate feeding

rate after 19th h is a response to the proposed zero feeding profile from the

optimization. At the meantime, a manual optimization described in the text was

performed from 21th h in order to consume all the substrate in the fermenter at the

scheduled end of the process.

58

4.17 Schematic of a software sensor. Inputs are mostly online measurements which

are facile to obtain. Outputs are the simulated values of variables which are hardly

59

x

or costly to obtain. The parameter identification part of the OLFO controller here

can function as a software observer.

4.18 The parameter estimation results around anaerobic phase. In the upper part shows

S: substrate, P: product and X: biomass; in the lower part shows the

corresponding simulation and experimental data of off gas composition. The

figures of the left sides show the original parameter estimation results. On the

figures of the right sides, the offline measurements for aerobic phase, the product

and substrate measurements for anaerobic phase are not involved in the

calculation of the parameter estimation. They are presented here as a reference,

to show how effective the OLFO controller is when it is used as a software

observer.

61

4.19 The final parameter estimation result with online measurements and biomass

concentration in anaerobic phase (12th - 25th hour). The other parts of the offline

measurements as on the figure of the right side are used as reference, not involved

in the simulation. In the upper part shows S: substrate, P: product and X: biomass;

in the lower part shows the corresponding simulation and experimental data of

off gas composition.

62

xi

List of Tables

Figure Description Page

2.1. Online measurements of laboratory bioreactor Biostat C. 15

2.2. Composition of the laboratory equipment used in bioreactor Biostat C (B:

container, V: valve, P: pump; M: Motor).

16

3.1 Description of the variables in the biological submodel. 26

The stoichiometric coefficients of the biological submodel. 27

4.1 Influence of different parameters on optimization. 42

4.2 Comparison between different fermentation processes. 43

List of Abbreviations

ANN Artificial Neural Network

CSTR Continuous Stirred Tank Reactor

FDA Food and Drug Administration

GMV Generalized Minimum Variance

HPLC High Performance Liquid Chromatography

MPC Model (based) Predictive Control

OLFO Open-Loop-Feedback-Optimal

PAT Process Analytical Technology

PCA Principal Component Analysis

PID Proportional-Integral-Derivative

PLS Partial Least Square

1. Introduction

Fermentation processes have been around for many millennia. Cooking, bread making, and wine

making are some of the fermentation processes that humans rely upon for survival and pleasure (Cinar

et al., 2003). Although fermentation operations are abundant and important in industries and academia,

high costs associated with many fermentation processes have become the bottleneck for further

development and application of the products. Developing an economically and environmentally

sound optimal cultivation method becomes the primary objective of fermentation process research

nowadays.

There are three types of fermentation operational modes: batch, fed-batch and continuous processes.

In this work, we focus on the fed-batch operation mode, since it offers a great opportunity for process

control when manipulating the feed rate profile which affects the productivity and the yield of the

desired product (Lee et al., 1999). From the concept of its implementation, the substrate concentration

can be maintained in the culture liquid at arbitrarily desired levels (in most cases at low levels). The

unfavorable effects, such as substrate inhibition, crabtree effect and catabolite repression can be

avoided. In addition, the favorable effects such as high cell density and extension of operation time

can be pursued. Furthermore, fed-batch fermentations can be the best option for some systems in

which the nutrients or any other substrates are only sparingly soluble or are too toxic to add the whole

requirement for a batch process at the start (Carrillo-Ureta, 2003).

In fed-batch fermentation operations, the substrate feeding profiles are adjusted to maximize an

appropriate performance objective with minimum experimental effort. Normally a fed-batch process

begins as a batch process but with only about 30-50% of the final volume of medium and specified

cells being inoculated (Hass & Pörtner, 2009). Until a given optimal initial biomass is reached or

substrate in the medium is consumed, the substrate is continuously fed into the bioreactor during the

fermentation period without withdrawing any fermentation broth until the reactor maximum volume

or setting goal/time is reached. In such way, the substrate concentration can be maintained at a fairly

low level.

For fed-batch systems, a practically relevant goal is to follow a predetermined trajectory for the

controlled variable which maximizes (or minimizes) a particular performance objective. It is also

known as open loop control strategy. In this context, a particular fed-batch process may aim at

maximizing cell production or target product concentration at the end of the cultivation. Proper

Chapter 1 Introduction

2

control could ensure high yield of pure product at reduced manufacturing costs. However,

determining an optimal control policy to produce the maximum yield is a challenging task, as the

dynamic governing equations are often nonlinear and include some particular physical constraints.

Nowadays, due to a lack of appropriate models and controllers, the substrate feeding profiles are

mostly adjusted based on heuristics and operational experience, which normally doesn't lead to an

optimal result. In industry automated control is essentially established by developing a reference

profile for substrate feed rate based on operational experience. The reference profile is then

implemented in the plant with suitable adjustments to account for the actual conditions of the

bioreactor. This approach is empirical in nature and operator dependent, and therefore leading to

variations in the product yield (Srinivasa & Moreshwar, 2009).

An alternative, to the current industrial approach, is to develop a mathematical model of the

fermentation processes which can facilitate calculation of the optimum substrate flow rate profile to

maximize the product yield in an actual fermentation. Normally model based control is superior to

the conventional empirical approach, since the use of a model enforces the formulation of quantitative

hypothesis on the process, which can be quantitatively checked by experiments.

However, modelling of fermentation processes is still not a totally resolved problem and

consequently, troublesome to monitor and control. Generally, this kind of processes is nonlinear. The

involved biological mechanisms are far from being well understood, and the available online sensors

are usually very expensive and/or inaccurate. Typically, such models are developed by conducting

offline identification experiments on the process. These experiments for identification often result in

inaccurate model parameter estimates. However, the performance of the control system depends on

the accuracy of the identified model. Since fermentation processes can be highly nonlinear and vary

temporally in their behavior, the model parameters and states should be updated online, to minimize

the plant model mismatch. This scheme of parameter estimation and optimization is carried out

periodically online based on the plant measurements and laboratory analysis results. This ensures that

the model used in the optimization calculations is close to the behavior of the real fermentation

process.

1.1 Motivation and objective

Due to the fact that the use of process models requires a certain level of knowledge and specialization,

numerical models have not been widely used in industry, but mostly remain in the research state.


3

Especially small and medium size enterprises still apply predominantly empirical methods of process

development. A general process model with a finite parameterization can greatly simplify the task of

model adaptation to a certain organism and creates a great value for the task of process design. Since

the adaptation of a process model to a new organism or product can be achieved by the adjustment of

parameters only. A general process model with high adaptability is one of the main goals of the project

"ProTool". A general process model is not bound to a certain cell line and microorganism, but adapts

to the variety of organisms and different situations. This process model will be used for data

interpretation, recipe optimization and the verification of control concepts as well as the basis of a

virtual representation of the process (training simulator). Accordingly, the development of an

advanced controller to achieve optimized process control for the general model, which is so called

the open-loop-feedback-optimal (OLFO) controller, is another main goal of "ProTool".

Consequently, such a model will significantly enlarge the possibility for industry especially for small

and medium size enterprises to benefit from model based process development.

The aim of my work is to develop the OLFO controller with a general model as a software tool which

facilitates the use of bio-technological models to support the maximum level of productivity in

process design. The OLFO controller takes benefit of the process model. In this control scheme, the

process model can be identified online. Based on a sufficiently accurate model calibration, the

optimization process is implemented to provide optimal control profiles (Munack, 1986; Witte, 1996).

Various theoretical and experimental publications show the high potential of this strategy (Frahm et

al., 2003; Luttmann et al., 2012; Munack, 1986; Witte, 1996). However, all of these OLFO controller

applications are designed for a single process. Here, my task was to shed light on how the parameter

identification and optimization in OLFO controller perform in the context of a general model.

The work in this thesis is organized within five chapters. The rest of the first chapter presents the

motivation of this work and state of art. Chapter 2 covers materials and methods that are used for the

work. Chapter 3 demonstrates the basic structure and elements of the OLFO controller. Chapter 4 is

divided into four sections to show the application examples of the OLFO controller for the fed-batch

processes. The first section presents the basic research results of this work. The second section

presents how the OLFO controller works with the virtual bioreactor. The third section formulates the

OLFO application for yeast fed-batch cultivation with the aim to produce maximum amount of

ethanol. The experimental procedure is also highlighted in this section. The last section details the

OLFO controller used as software sensor for detecting biomass, substrate and ethanol concentration

in aerobic phase and for detecting substrate and ethanol concentration in anaerobic phase. Chapter 5

contains the general summary and the perspectives emerging from this work.


4

1.2 State of art

1.2.1 Development of Yeast fermentation technologies

Saccharomyces cerevisiae (S. cerevisiae) or common yeast, is probably the oldest domestic organism

known to the human kind. Through the past several millennia, yeast was frequently utilized for

carrying fermentation processes in food and beverage preparation (Alba-Lois & Segal-Kischinevzky,

2010; Branduardi & Porro, 2012). Nowadays it is not exactly known at what point in human history

yeast was first time utilized to carry fermentation processes, however the oldest reported

archeological artifacts of jars containing remains of wine date to 5400 B.C. (McGovern, 2009). The

ancient cultures in Summeria and Babylonia are probably best known as the oldest to utilize yeast for

beer production (Damerow, 2012; Hornsey, 2003). Ancient Egyptians are known to be the first culture

to have used yeast for dough leavening (Redford, 2001).

Although yeast was frequently utilized, its unicellular eukaryotic nature and its role in fermentation

processes became better understood only in the course of the past 150 years. First attempts to describe

the microscopic appearance of yeast, date back to 1680 when van Leeuwenhoek observed yeast under

microscope (Dobell & Leeuwenhoek, 1932). However, these observations were not sufficient to

characterize yeast as a living matter. The developments in microscopy at the beginning of the 19th

century, allowed more detailed observation of yeast cells. This led to the discovery of Cagnaird de la

Tour in 1835 that during the fermentation process yeast cells reproduce by gemmation (budding)

(Branduardi & Porro, 2012). Further studies, published in the next few years by T. Schwann, F. Ktzing

and C. Exleben, showed that “the globular, or oval, corpuscles which float so thickly in the yeast as

to make it muddy” were living organisms (Barnett, 1998). Although these observations could

associate fermentation with the presence of yeast, the correlation between the fermentation and the

yeast metabolism was revealed by Louis Pasteur in 1857 (Pasteur, 1857). This remarkable notion,

became the foundation of the work of Eduard Buchner who in late 1880s showed that yeast ”extracts”

contain functional molecules, that can carry fermentation processes (Barnett & Lichtenthaler, 2001).

Buchner first coined the term enzyme, while his contribution in the field earned the Noble Prize in

chemistry in 1907.

Resolving the metabolic pathways in yeast presented a difficult challenge in biochemistry until the

second half of the twentieth century. Nowadays it is well-understood that glucose uptake by the yeast

cell follows the glycolysis pathway which leads to the formation of pyruvic acid. Abundance of

oxygen in the culture medium, facilitates the respiratory pathway in which pyruvic aid is further

decomposed in order to generate the energy necessary for the growth of the organism. On the other


5

hand, scarcity of oxygen or anaerobic conditions facilitate the conversion of pyruvic acid to ethanol

and carbon dioxide (Nelson et al., 2008).

When the obtained products of the alcoholic fermentation (e.g. ethanol) are present above a particular

threshold in the fermenting medium, they begin to exhibit toxicity towards the yeast cells (Gray,

1941). The threshold usually varies between 10 - 15 vol. %, however considering all varieties of yeast

strands, the range may vary from 5 - 21 vol. % of ethanol.

When designing large scale fermentation processes, the direct mapping from large scale events to

molecular scale metabolic pathways may not be practically accessible. This is because the cell

cultures vary in terms of strength, efficiency of conversion, tolerance towards toxic levels of materials

etc. The interest to develop accurate models of the fermentation process and to introduce control over

the process has been highly desirable in order to have efficient processes in terms of product yield

and time. This motivation has led to laborious research and development of process control and

monitoring fermentation technologies during the past decades. Part of this research relevant for this

thesis is summarized in the next sections of this chapter.

1.2.2 Development of control strategies for fermentation processes

Nowadays, the industry scale fermentation processes are frequently adopting mathematical model

driven techniques for process control in order to achieve cost efficient product manufacturing.

However, this practice is still relatively new practice, as at earlier stages, many industry scale

bioprocesses were typically designed and operated based on state trajectories obtained from previous

successful process campaigns. The data was aggregated on a computer and holistically analyzed for

understanding trends (Albert & Kinley, 2001). Thus, in a new process operation, the data is typically

traced using an open-loop control and it is then used as an offline reference for process monitoring

and fault detection. Sometimes, this approach may have run time-predictive capability. However, it

is often argued that the offline approach may not always lead to optimal control (see for more details

section 1.2.3) (Chu & Constantinides, 1988; Ponnuswamy et al., 1987; Soroush & Valluri, 1994).

Designing optimal control based on online calculated inputs is very challenging. Some of these

challenges include: (a) lack of accurate models that describe cell growth and product formation; (b)

the bioprocess dynamics is highly non-linear; (c) slow process response; (d) deficiency of reliable

online sensors for quantification of state variables. Some of the former challenges have been

successfully addressed with various high performance model based control algorithms. These control


6

algorithms such as: the optimal adaptive feedback control, optimal predictive control and the Open-

Loop-Feedback Optimal control are addressed in the next sections (Shimizu, 1993; Stanke &

Hitzmann, 2013).

a) Optimal adaptive feedback control

Modak and Lim presented a system systematic analysis to identify the optimal mode of operation of

two different objectives, namely maximizing the yield and productivity (Modak & Lim, 1992). In

their work, the yield is the production per unit of substrate fed to the reactor, expressed as the ratio

between the harvested amount of product and the added amount of substrate. In addition the

productivity is the production per unit of time, expressed as the ratio between the harvested amount

of product and the duration of the process operation.

Many processes are characterized by a conflict between yield and productivity, for a given amount of

substrate, the productivity is an increasing function of substrate concentration and yield is a decrease

function of substrate concentration. The optimization problem is to find the optimized amount of

substrate and optimized substrate concentration from the statistics of measurement which correspond

to the best tradeoff between yield and productivity (Jadot, Bastin, & Van Impe, 1998; Shimizu, 1993).

Modak and Lim reported an optimization study of the fed-batch process, and also proposed feedback

linearization control law to track the calculated substrate concentration (Modak & Lim, 1987).

Van Impe and coworkers introduced optimal adaptive feedback control strategy for biotechnological

processes (Bastin & van Impe, 1995; Jadot et al., 1998). The strategy combines the advantages of

both the optimal control and adaptive control approaches. The authors also adopted the feedback

linearization control law, but under the form of an adaptive regulator, which is designed by using of

Lyapunov theory. In this control structure, the biomass concentration is provided by a model based

observer. The controllers derived in this way combine a nearly optimal performance with good

robustness properties against modeling uncertainties and process disturbances.

Bošković and coworkers suggested a stable adaptive control, whose parameter are adjusted using only

one of the output errors rather than both in the previous method. Thus the convergence of the output

error to zero is guaranteed and yields the acceptable performance (Bošković, 1995; Bošković, 1996).

Alternately to the above optimal strategy on optimizing yield and productivity conflict, some

advanced mathematical algorithms are developed to calculate optimal feeding strategies for complex


7

models (Carrasco & Banga, 1997; Tholudur & Ramirez, 1996; Tremblay et al., 1992). The relatively

simple bioreactor systems which are expressed in differential equation models the optimization

problem can be solved analytically from the Hamiltonian function (Van Breusegem & Bastin, 1990).

In other studies singular arc properties were used to solve the optimal control problem (Chikkula &

Lee, 2000; Park & Ramirez, 1988; Van Impe & Bastin, 1998). However, the approaches from these

studies become too complicated when the number of state and control variables increases and the

complexity of the systems grows.

In 2007 Pan and coworkers reported the lazy learning-based online identification and adaptive

Proportional-Integral-Derivative (PID) control for Continuous Stirred Tank Reactor (CSTR)

processes (Pan et al., 2007). The developed method consists of two-layer supervised algorithm. The

lower layer consists of a conventional PID controller and a plant process, while the upper layer is

composed of identification and tuning modules. Using a lazy learning algorithm, a locally valid linear

model denoting the current state of system is automatically exacted for adjusting the PID controller

parameters based on input/output data. This scheme can adjust the PID parameters in an online

manner even if the system has nonlinear properties. In this online tuning strategy, the concepts of

generalized minimum variance (GMV) and quadratic program with constraints are also considered.

The scheme has been tested on a CSTR chemical process from an AAS platform and showed a good

control system performance.

Zeng and coworkers adopted model reference adaptive control for fermentation process (Zeng &

Dahhou, 1993). The control objective is to get the state of the system to track the state of a given

reference model despite the disturbances and system parameter uncertainties. With the adaptive state

estimator, the states and parameters are updated using Lyapunov technique. The structure of the

adaptive controller is determined by the requirement to obtain stable reference model tracking.

b) Optimal predictive control

In recent years, predictive control has been accepted as a useful advanced industrial control technique

(Frahm et al., 2003). The control task is to give a series of control signals minimizing a quadratic

deviation between a reference signal and the system output in a given prediction horizon. According

to the preceding horizon strategy, only the first control value is applied and the procedure is repeated.

In predictive control methods, all controller configurations are based on a forecast of the process

output signal, using a predefined mathematical model.


8

Several nonlinear predictive control algorithms have been presented by various researches. Most of

these approaches adopt the complex physical models of the system and solve a constrained nonlinear

programming problem subject to the system dynamic equation constraints in addition to the state and

the input constraints.

Flatness-based predictive control scheme have been adopted for fed-batch fermentation process

(Mahadevan et al., 2001). The idea of differential flatness was first introduced by Fliess and

coworkers (Fliess et al., 1995). This allowed an alternate representation of the system where trajectory

planning and nonlinear controller design is straightforward. With this approach, the optimization is

transformed into low dimensional nonlinear problem through the use of flat outputs. The optimization

approach is demonstrated in the repeated optimization of nonlinear dynamic systems under the

parameter feedback which is similar to nonlinear model predictive control. Here, the biomass and

production optimization are successfully solved. The proposed scheme is also used in conjunction

with a nonlinear Luenberger observer to generate the optimal trajectories under parametric

uncertainty. Rodrigues and Filho have presented a same approach for product optimization in fed-

batch penicillin production process with predictive controller and achieved successful result

(Rodrigues & Filho, 1999).

Dahhou and coworkers presented the adaptive predictive control for continuous stirred tank reactor

(CSTR) (Dahhou et al., 1992). A discrete adaptive controller using online estimation is developed.

The new estimation algorithm formulated consists of two estimation steps: the estimation of the

specific growth rate and the attribution of the latter variations to growth and feed effects. Good

simulation results have been obtained in regulation and tracking, disturbance rejection, showing the

efficiency of this adaptive predictive control scheme.

Roux et al. reported the four approaches adaptive predictive control with empirical models (Karra et al.,

2008). The empirical model (ARMA) scheme is used as process model. The two-tier modeling scheme is

designed in which the deterministic and stochastic components of the model are updated online by two

separate recursive pseudo linear regression schemes. The deterministic and stochastic components of the

model are then combined to form a linear time varying state-space model, which is then used to formulate

the predictive control problem at each sampling instant.

Foss and coworkers decomposed the operation process into a set of operation regimes, and simple local

state-space model structures are developed for each regime (Foss et al., 1995). These are combined into

a global model structure using an interpolation method. Unknown local model parameters axe identified

using empirical data. The control problem is solved using a model predictive controller based on this


9

model representation. The performance of the model-based controller is comparable to that of the exact

process model and linear model. It is experienced that a non-linear model with good prediction

capabilities can be constructed using elementary and qualitative process knowledge combined with a

sufficiently large amount of process data.

In 2009 Ashoori et al. reported Model Predictive Control (MPC) based on a detailed unstructured model

for penicillin production in a fed-batch fermenter (Ashoori et al., 2009). The novel approach used here

is to use the inverse of penicillin concentration as a cost function instead of a common quadratic

regulating one in an optimization block. Moreover, to avoid high computational cost, the nonlinear model

is substituted with neuro-fuzzy piecewise linear models obtained from a method called locally linear

model tree (LoLiMoT). The acceptable performance is shown in the experimental result.

Zhang and Lennox investigates the partial least squares (PLS) modelling approach in the operation of

fed-batch fermentation processes (Zhang & Lennox, 2004). The modeling approach can be integrated

within a standard model predictive control to regulate the growth of biomass within the fermenter. It is

shown that models developed using PLS can be used to provide accurate inference of quality variables

that are difficult to measure on-line, such as biomass concentration. Additionally the proposed model

can be used to provide fault detection and isolation capabilities. This model predictive controller is shown

to provide its own monitoring capabilities that can be used to identify faults within the process and also

within the controller itself. Finally it is demonstrated that the performance of the controller can be

maintained in the presence of fault conditions within the process.

A new robust Model based Predictive Control (MPC) uses a finite horizon is used in fermentation process

(Eaton & Rawlings, 1992). Solving the optimization problem allows the optimal set of controllers to be

calculated efficiently by minimizing the resulting upper bound of the worst-case infinite-horizon control

cost. This approach has the advantage of guarantee the stability of algorithm.

c) The Open-Loop-Feedback Optimal (OLFO) control

Open-loop optimal feedback control uses current measurement data, the feedback policy which uses

all past measurement data and control signal of last stage. The adaptive process control with the

OLFO-method has been successfully applied to control warm water floor heating (Munack, 1986)

and in controlling the flying trajectories of unmanned aerial vehicles considering infrequent

battlefield information updates (Shen et al., 2010). In 1985 Luttmann performed early simulation

studies based on this method (Luttmann et al., 1985). In addition, the OLFO control was implemented


10

in cultivation of Cyathus striatus (Witte, 1996). A model based, adaptive process control strategy for

animal cell cultures was developed on the basis of the OLFO by Frahm. The yield of production of

monoclonal antibodies has been improved by 70% only by change of process control strategy with

the OLFO controller (Frahm et al., 2003). A model based, adaptive process control strategy for animal

cell cultures was developed on the basis of the OLFO by Frahm (Frahm et al., 2002)

Detailed description on OLFO and its implementation in control of yeast fermentation processes is

provided in Chapter 3.

1.2.3 Soft Sensors

In 2011, the Food and Drug Administration (FDA) released recommendations for development and

implementation of analytical tools that can improve the manufacturing efficiency and the process

quality in the industry (U.S. Department of Health and Human Services Food and Drug

Administration). Implementation of these recommendations, requires that the industry lowers process

expenses and losses with an aim to make many pharmaceuticals cost-efficient (Glassey et al., 2011).

As biotechnology is involved in pharmaceutical manufacturing, these recommendations have

triggered development of novel approaches in the process monitoring and quality control.

Software sensors or also referred to as soft sensors (Figure 1.1) represent synergistic combination of

precise and reliable analyzers with estimation algorithms, i.e. software (Kadlec et al., 2009; Luttmann

et al., 2012; Mandenius & Gustavsson, 2015). In the context of the Process Analytical Technology

(PAT) recommendations, soft sensors can be applied to estimate variables that are challenging to be

measured online (Chopda et al., 2016; Glassey et al., 2011). For fermentation processes one such

problem is the monitoring of biomass in real time (Wechselberger et al., 2013). In this respect, online

quantification of biomass can be inaccurate as protein expression during the induction phase impacts

cells’ morphology and physiology, while frequent offline quantification is time-consuming (leading

to delayed data accumulation) and it increases the risk for contamination.

Soft-sensors have actual applications in the industrial lypophilisation (freeze-drying) processes and

in wastewater treatment, while they are rarely used for monitoring industry scale fermentations

(Goldrick et al., 2015; Warth et al., 2010). The perspective to apply soft-sensors in fermentation

processes dates back since to the mid 1980’s (Luttmann et al., 2012), when the respiratory quotient

(i.e. the ration of the rates of carbon dioxide production and oxygen consumption) in S. cerevisiae

was monitored and used to forecast the biomass evolution and the substrate concentration (Graindorge


11

et al., 1994; Bellgardt et al., 1986; Stephanopoulos & San, 1984).

The accuracy and the efficiency of the soft sensors is directly related to the compatibility of the two

parts (hardware and software) and on the nature of the fermentation processes. In this context, soft

sensors relay on developed analytical tools for monitoring of a particular variable and on software for

modeling and calculation of the evolution of the bioprocess through time.

Figure 1.1: Scheme of the soft sensor as defined in reference (Luttmann et al., 2012). The figure depicts only one

hardware sensor, although in reality there can be several of those.

The predictions carried by the software unit of the soft sensors can be data-driven or (mathematical)

model-driven (Luttmann et al., 2012). Data driven soft sensors are often based on chemometric

techniques which allow extraction of information from large dataset obtained from previous

experiments and statistical process control. Data driven soft sensors currently are very attractive to

industry as they have been traditionally used there. In the pharmaceutical industry, partial least square

(PLS) and principal component analysis (PCA) (Luttmann et al., 2012), artificial neural networks

(ANN) (Bolf & Jerbic, 2006) neuro-fuzzy systems and support vector machines (Kadlec et al., 2009),

are commonly used as chemometric techniques. Although these techniques allow fast predictions

once significantly large database is created for a particular fermentation process, they are limited in

bringing understanding of the observed correlations.

The mathematical models can address the steady-state or the dynamic system (Luttmann et al., 2012).

Steady-state models address mass and component balances, mass or heat transfer and elementary

balances. On the other hand, the dynamic models address the kinetics of the state variables. This

kinetics and thus yields can be affected by the count and the physiology of the involved organisms.

Thus dynamic models which capture this complexity are referred to as structured models and they

have changing yield coefficients. However, such mathematical models are computationally

expensive. On the other hand, dynamical unstructured models approximate the organism complexity


12

and use constant yield coefficients. This approach leads to lower computational costs then the

structured models and therefore makes the unstructured dynamic models more attractive for sensor

applications (Pörtner et al., 1994).

Considering the strengths and weaknesses of both data and model driven software components, it has

been recommended to develop approaches that allow combination of macroscopic correlations and

mathematical models (Teixeira et al., 2007). Such attempts in principle lead towards development of

hybrid semi-parametric models, or so called ”grey-box” models, which exhibit flexible frameworks

of heterogeneous databases of different layers of information about the cell and the process (Kadlec

et al., 2009).

Regardless of method employed in carrying the predictions, the software algorithms also require to

adopt filtering of the input experimental data prior processing, as the experimental noise and the

potential outliers can lead to poorly predicted process trajectories. Processing such uncertainties of

the incoming online measurements is typically done by using the Kalman filter for linear systems and

extended Kalman filter for non-linear systems (Bellgardt et al., 1986).

The monitoring of one or several experimental variables can be performed by various types of probes,

sensors, analyzers, spectroscopic instruments or even chromatographs (See Figure 1.2). However,

one typical approach to describe the used hardware is on the basis if the sampling and analysis occurs

in the fermentor (in situ) or outside the fermentor (ex situ). In this context in-situ hardware that are

placed in the fermentation medium after proper sterilization. This category includes many probes that

record temperature, pressure, pH, volumetric or mass flow rates and weights. Paramagnetic oxygen

analyzers (O2) and infrared adsorption photometers (CO2), electronic noses (e.g. for ethanol, sulfides

etc.) are typically used for quantification of the components of the off-gas mixture (Van Impe &

Bastin, 1998), Raman spectrometers have been effectively used as a tool to follow the glucose

concentrations (Berry et al., 2016). Biological analysis of the cell cultures can be also provided by in

situ microscopy (Havlik et al., 2013), or by applications of near infrared spectroscopy (Gustavsson et

al., 2015). Ex situ hardware (sometimes referred to as at-line) requires utilization of sterile barrier

between the analytical system and the fermentation medium. This category includes sophisticated gas

chromatographs (e.g. HPLC), mass spectrometers and flame ionization detectors. In these techniques

the withdrawal of a sample may exploit methods for bioprocess stream analysis either in closed or in

open bypass flow procedures (Kaiser et al., 2007; Peuker et al., 2004). Contamination in at-line

systems is commonly solved by utilization of filtration modules (Warth et al., 2010), or alternatively

catheter probes (Olsson et al., 1998).


13

Figure 1.2: Schematic depiction of the soft sensor implementation as described in reference (Warth et al., 2010).

From the perspective of soft sensor design, in-situ sensors seem more attractive than ex situ sensors

as they do not contribute to the contamination risk and allow faster data measurements. Currently

many already available ex situ instruments allow better precision and much larger sets that can be

measured. However, considering the advances in fluorescence microscopy, other techniques for in

situ monitoring may become commonly applied for bioprocess monitoring in near future (Ohadi et

al., 2015; Ödman et al., 2009).

2. Laboratory set up

2.1 Organisms and media

The choice of the organism for our experiments is the bacteria Escherichia coli (E.coli) and the

common yeast S. cerevisiae. E.coli was chosen on the ground that it is commonly studied organism

in industry, genetics and pathology. S. cerevisiae was chosen on the ground that it is commonly

studied eukaryotic model organisms in molecular and cell biology, much like E.coli as the model

bacterium. Since most of the studies are based the fed-batch fermentation of S. cerevisiae, so only the

procedure of S. cerevisiae fed-batch fermentation is presented here. S. cerevisiae was used to establish

an application of the OLFO controller to test the adaptive, model-based fed-batch process control.

500.0 ml of cultivation medium consisting of 10.0 g/L dextrose, 5.0 g/L pep-tone, 3.0 g/L malt extract

and 3.0 g/L yeast extract was prepared. Small culture of S. cerevisiae (ca. 0.10 g/L) is added and the

medium is inserted into a humidified incubator (30 °C) with a rotator shaker (150-180 rpm) for 18

hours. At the end of this period, the preculture medium reaches 2.58 g/L substrate, 0.11 g/L ethanol

and 0.70 g/L biomass concentration, as determined as offline measurements (vide infra). This is a

typical procedure for a preculture preparation in our laboratory.

2.2 The Bioreactor

The cultivation is performed in a 15.0 L bioreactor, three six-blade stirrers (Sartorius stedim biotech,

Germany). The reactor with its complementary parts is depicted on Figure 2.1 while description of

the used abbreviations is provided in Tables 2.1 and 2.2. The reactor is equipped with online sensors

and controllers, namely: motor rotation speed, aeration rate, temperature, pH, oxygen concentration

and foam level controllers of the fermentation media. Prior use, the pH and dissolved oxygen pO2

sensors were calibrated. To avoid contamination, the bioreactor was sterilized with hot steam (120

°C) for about 20 minutes. The previously preculture medium was transferred to the bioreactor which

already contains 6.5 L fermentation medium with a glucose concentration of about 3.0 g/L. Once the

initial glucose depleted, further glucose medium was pumped into the reactor.

The laboratory bioreactor is supplemented by a computer with the OLFO embedding, a control unit

and many transmission lines (see Figure 2.2). The whole fermentation process lasted until a prefixing

Chapter 2 Laboratory set up

15

process time 25 hours. At that time, the desired ethanol concentration was also achieved by using the

optimized feed rate profiles generated by the OLFO controller.

Figure 2.1: Instrument of a laboratory stirrer tank bioreactor. Adapted from reference (B. Braun Biotech International

GmbH).

Table 2.1: Online measurements of laboratory bioreactor Biostat C.

No. PLT Measurement type

101 QIC pH-measurement

102 FI Gas flow measurement

103 QIC Rotation speed

104 TIC Temperature measurement

105 QIC Dissolved oxygen measurement

106 LA+ Substrate flow rate measurement

108 LA+ Conductivity measurement

109 LA+ Conductivity measurement


16

Table 2.2: Composition of the laboratory equipment used in bioreactor Biostat C (B: container, V: valve, P: pump; M:

Motor).

No. Equipment Apparatus

107 V Harvest valve

117 V Drain valve

118 V Valve for cooling and heating jacket

113 P Inlet air dosage

114 P Acid dosage, peristaltic pump

115 P Base dosage, peristaltic pump

116 P Substrate, peristaltic pump

111 Tachogenerator

112 M Agitator motor

Figure 2.2: Cultivation system in laboratory (includes a stirred tank bioreactor, a computer with the OLFO embedding in

process control system and a control unit).

2.3 An overview of the variables system

During the actual experiment, several offline and online parameters are monitored, analyzed and

controlled in parallel. In this context, online parameters such as temperature, pH, pO2 etc. are

predominantly automatically controlled, while others such as foams levels require manual assistance.


17

Offline parameters are obtained in particular timer intervals and they are analyzed manually. In the

following two subsections, details on the reactor maintenance and sample analysis are provided.

Variables of the whole system are shown in Figure 2.3. It shows all the measurements, input, output

flow, controller variables and actuating variables in the whole system.

Figure 2.3: A scheme depicting the variables of the whole system.

2.3.1 Direct inputs to the fermentation unit

The fermentation unit consist of vessels, equipped with several valves to control the flow patterns of

process gas, with a motor driven stirrer as well as with pumps for the control of substrate, base, acid

and antifoam input.

All those agitators are under direct control of a control unit. The control unit may be controlled by

hand via input of setpoints for certain vital state quantities or for the operation level of some of the

agitators. Alternatively, the control unit can be operated by a computer using a serial interface

protocol.


18

2.3.2 Measured quantities of the fermentation unit

By default, the fermentation unit offers the online measurement of several state variables to get some

insight into the process state. These online quantities may be supplemented by some offline quantities.

The vessel and devices give a feedback of the operational settings (e.g. stirrer frequency or air flow

rate). Then some of the state quantities inside the fermentation fluid are observed (pH, redox potential,

temperature, dissolved oxygen content). In addition, the oxygen and the carbon dioxide content of the

cultivation exhaust gas is observed online.

1) Control and monitoring of online variables

The online variables are updated every minutes (adjustable due to different needs).

pH value: The pH controller automatically adjusts the flow rate of sodium hydroxide solution

to maintain the fermenter pH at a desired value. If the pH becomes lower than a certain

threshold, the controller switches on the pump which adds sodium hydroxide to the fermenter.

When enough sodium hydroxide is added and the pH returned to the set value, the pump is

switched off. For the experiments in this thesis, the pH was always maintained at 7.0.

Temperature: The temperature of the medium was maintained at 37 °C to ensure the optimal

growth of the cells by a thermal equilibration external jacket filled with water. As the

temperature of the medium is not constant (the fermentation process dissipates energy) the

temperature of the external jacket has to be regulated.

Dissolved oxygen pO2: The oxygen from the inlet air is absorbed at the gas-liquid interface.

The dissolved concentration is automatically controlled by a self-developed PID controller

implemented in the process control. The dissolved oxygen concentration is maintained at 60%

air saturation and it can be controlled by three variables such as: stirring frequency, aeration

rate and oxygen concentration coming from the inlet air flow. Typical aeration rates for our

system are 2.5 L/min and 5.0 L/min inlets of room air. Stirring rates are set in the range of

100 - 800 rpm. Vigorous stirring over the later boundary is not recommended since it can

cause damage to the studied species. In cases of high culture density or specific organism

characteristics (e.g. high shear sensitivity in animal cells), the desired oxygen concentrations

may be difficult to reach. Under such circumstances, the oxygen concentrations are controlled

by increasing oxygen's partial pressure in the inlet air flow.

Off gas content: The volume concentrations of CO2 and O2 in the off gas are monitored by

extractive gas analyzer S700 (SICK MAIHAK GmbH). This at-line tool extracts certain


19

portion of the off -gas which is then supplied continuously to the gas analyzer. Prior supplying

the extracted portion to the gas analyzer, the gas is pretreated with liquid separators to remove

condensable components such as water. The detection and quantification of CO2 content is

carried by the installed FINOR analyzed module which is based on non-dispersive infrared

absorption measurements of the sampled gas. On the other hand, quantification of the O2

content is carried by the OXOR - analyzed module which is based on the paramagnetic

properties of the O2 molecules (or OXOR-E/electrochemical cell).

Foam formation: It is common that during the fermenting process, foam formation occurs.

To avoid foam formation, 2.0 ml of antifoam suppression agent is added manually to the

fermenting medium prior the fermentation. During the fermentation process, foam formation

is manually controlled by addition of antifoam agent.

2) Sampling and analysis of the offline variables

Offline measurements are performed periodically by sampling aliquots of ca. 5.0 - 10.0 mL from the

reactor medium every 30 minutes. The sampling interval can be further adjusted to the evolving

scenarios. Prior analysis, the samples must be treated with Carrez reagent, in order to stop metabolic

reactions by inhibiting the activity of the active enzymes. Three parameters are recorded using simple

light spectroscopy methods:

Cell concentration: this parameter is determined through measuring the optical density (OD)

of the dry biomass. The OD of each sample was measured at 600nm wavelength with a

spectrophotometer (Biomate 3, USA).

Glucose concentration: by using d-glucose enzymatic UV-method (R-Biopharm AG,

Germany).

Ethanol concentration: by using oxidation enzymatic UV method (R-BiopharmAG,

Germany).

2.4 Technical aspects of the OLFO embedding

Two main aspects govern the technical boundaries of the OLFO algorithm. First, the practical

relevance of the OLFO controller must be proven. This can be done by application of the algorithm

for the control of a real fermentation experiment. In the research lab, a biostat C laboratory

fermentation unit, being controlled by the process control system WinErs* is used. Consequently, the

OLFO controller has to be designed to operate along that system.

*WinErs is manufactured by the engineering consultant Ingenieurbüro Dr.-Ing Schoop GmbH, www.schoop.de.


20

A second aspect of software design of the OLFO components is its potential use as the part in a

commercial software suite. Independent of how this software suite will be designed, it will for sure be

implemented within or along some process control system, performing the primary control of the

cultivation process.

Consequently, the algorithm can only use those measured quantities which are delivered by the process

control system. Further, it can only manipulate on the standard inputs of the fermentation unit.

2.4.1 Establishment of the OLFO controller structure

The OLFO controller is realized using a program package which is based on the programming

language C++ along with the industrial process control system WinErs, which is set up to control a

biostat C laboratory fermenter.

The process control system embeds a process model, which can be switched to serve as the virtual

counterpart of the fermenter. At an experiment, process data recorded online and offline will be

collected by the control system. During the experiment, the process data will be exported in a suitable

format for the exchange with the process model. This allows for a rapid data interpretation in the light

of the numerical model.

In order to perform an estimation of model parameters based on a cultivation process, the complete

information about the cultivation (state variables and the complete history of control settings) must be

passed to the estimation algorithm. This at least affords some defined data format and if later on, the

parameter estimation has to run online, some sort of communication between the parameter estimation

algorithm and the process control system should be set up.

The commercial process control system WinErs offers two interfaces for communication with devices

or external software. The common property of these interfaces is their binding to the operation cycle

of the system. That means, that these interfaces are designed to obtain a set of inputs each cycle,

returning an output each cycle. These interfaces are not designed to transmit a set of data vectors from

time to time. Since model calibration plus control optimization usually takes longer than one process

cycle, it would be very helpful, to achieve a decoupling from the process cycle. Actually, the

engineering company Dr-Ing. Schoop GmbH implemented a driver for the export of experimented

data and the import of control quantities. The present implementation of the OLFO is based on that

driver.


21

2.4.2 Options of communication with the process control system

Since the OLFO later on will compute the optimal control pattern, a way must be found to transfer this

information to the process control system. However in the context of a commercial software tool, it is

helpful if the parameter estimation and thus the optimization would run based on the online

measurements only.

The DLL* interface allows to send 32 floating point and 32 binary quantities to user defined

functionality and return 32 floating point and 32 binary quantities each cycle.

The driver offers the same functionality in principle, with the difference, that the number of inputs and

outputs is significantly restricted. The trade-off is the fact, that the floating point variables are

transmitted in a lower digital resolution. This affords to restrict their range in order not to lose

numerical precision, which is especially important, when parameter estimations have to be performed

based on that data. A little exaggerated demonstration may underline the problem. In a hypothetical

scenario, when substrate concentrations in the range of 0.0 – 50.0 g/L are expected, the reuiered

precession is 1.0 g/l. If the associate process variable is defined to have a range of 0.0 g/L to 1.0 ∙ 105

g/L with an 8 bit digital resolution, then the minimal difference between two numbers is given by

LgLgy /16.392/255

100.1 5

which is above the desired precision. If the range definitions on the device side and on the control side

are different, a false translation of the transmitted values will be performed.

In principle, WinErs offers the functionality to implement the OLFO-loop. Online measurements and

control settings can be transmitted each operation cycle and the optimized control pattern can also be

transmitted each cycle.

* DLL is an abbreviation for the term Dynamic Link Library. It is an implementation of run time loadable computer

libraries specific for the MS-Windows operating system.

3. The OLFO controller

The OLFO controller is an adaptive, model-based, long-term-predictive controller (Frahm & Pörtner,

2002; Frahm et al., 2002a; Frahm et al., 2002b; Hass et al., 2002). It helps to find a quick pathway to an

optimal process recipe. The basic structure of the OLFO controller is shown in Figure 3.1. Major

elements of the OLFO controller are: a process model, a model parameter identification process and

an optimization process. Within the OLFO controller, the process model is calibrated online based on

online process data and offline laboratory data at runtime of the experiment to reduce the mismatch

between the process model and the actual fermentation process. The updated calibrated model is used

for the process optimization to generate the optimum substrate feed rate. According to the evaluation

result, the optimum substrate feed rate will be implemented in the process control system or discarded.

The cycle of parameter identification, optimization and the implementation of the optimum result are

repeated periodically in real time. The interval of the cycle should be set at the beginning of the

process. It can also be adjusted during the process according to the practical situation.

3.1 General process model

A general process model serves as core element of the tool set. The result of the OLFO controller relies

heavily on the process model. Since it is very difficult to develop models which take into account the

numerous factors influencing the parameters which characterize the microorganism growth. Moreover,

since the process involves living organisms, the process dynamics is strongly nonlinear and time

varying. In this respect, choosing an adequate process model and model structure applicable to the

OLFO controller is essential (Li et al., 2012).

In general, the bioprocesses can be modelled as structured or unstructured models. Unstructured

models consider only the physiology of the cells due to changes in their environment, such as the

concentrations of the main substrates and metabolites. These models neither recognize nor represent

the composition or what we call the quality of the biophase. The advantage of using simple

unstructured models is that, these models have only a few model parameters and are easily to be

controlled. However, the unstructured models exhibit several weaknesses. They do not show any lag

Chapter 3 The OLFO controller

23

phase and do not provide with any insight into the variables which influence growth. Further, they do

not make attempt to utilize or recognize knowledge about cellular metabolism and regulation.

Figure 3.1: Basic structure of the OLFO controller. Three elements: a process model, a model parameter identification and

an optimization part. )(ˆ)( txCtCx is minimized to estimate the parameters. Based on the identified model, optimal

control profiles are calculated in the optimization part and transferred to the bioreactor and process system. Adapted from

reference (B. Frahm et al., 2002).


24

Structured models are typically complex and have a higher cost in their creation when compared to

unstructured models. A variety of models of different structure and complexity has been proposed in

the literature (Posten & Munack, 1990; Pörtner et al., 1994). However complex, structured models offer a

higher variety in the cultivation characterization. The cell population composition changes

significantly and these composition changes influence kinetics are possibly described in structured

model, which make it a suitable model type for the OLFO controller as a general tool.

Taking into account the tasks to be accomplished, the model aims to describe the cultivation process

as simple as possible, but also be as detailed as necessary, which can easily be reached by a

compartment structure, which is the simplest structured model (Witte, 1996). This four-compartments

model has been proved to have a high potential as a general model. It can describe different organisms

with the same model, just by adaptation of model parameters (Schwarz, 2009; Witte, 1996). The

compartment model is divided into four compartments of different functions. It is illustrated in Figure

3.2.

The four-compartments represent the following parts: portion of the biomass; the expiry of the primary

metabolism; secondary metabolism, the part of the product formation catalyzed. In addition, a

compartment of inactive biomass forms the structural part of the biomass.

Figure 3.2: Structure of the compartment model (primary biomass (XPR): protein, DNA and RNA; secondary biomass

(XSE): enzymes for synthesis of secondary metabolites; structured biomass (XS): membrane, cell wall and complex


25

polysaccharide; inactive biomass (XI): defect enzyme, DNA and RNA; G: glucose; Eth: ethanol; O2: oxygen; CO2: carbon

dioxide; HE: yeast extract; P: product.). The figure was redrawn from reference (Witte, 1996).

A flexible process model which describes the cultivation process in a laboratory setting was developed

based on the compartment model in our group. This model should be a structured model consisting of

four biomass compartments. At first, the developed model is a single compartment (unstructured)

model representing cell growth, substrate uptake, product formation, oxygen consumption and carbon

dioxide formation. The model is able to describe the cultivation of different organisms. Its capability

to describe fed-batch cultivations of S. cerevisiae and E.coli has been confirmed (Schwarz, 2009;

Schwarz et al., 2010). The substrate concentration has inhibitory effect over the cell grow rate and

production formation rate. However this inhibitory effect is not included in the current general model

yet. Its ability to describe other organism cultivations is still in development (Schwarz, 2009).

Consequently the new model already fulfils some very important claims concerning its handling, but

still needs to be expanded to be able to describe the cultivations better.

3.1.1 The biological submodel

The biological submodel describes a general bio technological cultivation process. In the cultivation

process, the model provides three metabolic pathways of substrate consumption: biomass growth,

product formation and energy generation (See Figure 3.3). The substrate consumption through the cell

membrane is described by higher order kinetics. The flow of the substrate into these different metabolic

pathways is described by distribution functions. With this structure, the metabolism of the cell can be

formally written to adapt different conditions. A version of this model, which is named "Lyx" was used

in this work (Schwarz et al., 2010). The development of the procedure for model-based process

optimization and optimal control are done based on this process model. In addition, a structured model

with increasing complexity is constantly under development.


26

Figure 3.3: The substrate flux is the key element of the "Lyx" process model. A total substrate consumption rate rS limits

the speed of all subsequent reactions. Distribution functions describe the substrate flux flows into different metabolic

pathways, depending on the actual state.

The consumption of substrate in different metabolic pathways is the main idea to construct the process

model "Lyx". The key functions of the biological submodel are defined by turning model variables

which depend on the physiological state. The submodel is governed by several components as listed

below.

The core variables, parameters and stoichiometric coefficients that are used in this

biological submodel are listed in Table 3.1 and Table 3.2. YanXPX , YaeSXP , YaeXlo, YaeXhi,

KS and rSmax are parameters to be estimated in the OLFO controller.

The substrate consumption rate is SS ry .

The total substrate uptake is described by a monod type expression:

𝑟𝑆 = 𝑟𝑆𝑚𝑎𝑥∙

𝑆

𝐾𝑆 + 𝑆+ 𝑆𝑤𝐴𝑛

While rS is reduced in case of anaerobic growth.

A modulation function controls the switch between aerobic and anaerobic uptake of substrate S:

)( 5021

1anslan KPOK

rsanrSan

e

KKSwAn

Table 3.1: Description of the variables in the biological submodel.


27

Variables Descriptions

rSmax Maximum glucose consumption rate per biomass (under aerobic conditions).

rS The substrate uptake rate.

KS The Monod constant of glucose consumption (when substrate concentration converges

to KS, then the glucose consumption rate is reduced to 1/2 of the maximum value.

KrSan

Coefficient, the maximum glucose consumption rate decreases under anaerobic

condition (which means when pO2 = 0, the maximum glucose consumption rate is:

rSmax ∙ KrSan).

rXPX Glucose consumption for biomass growth.

rXPP Glucose consumption for product formation.

rSE Glucose consumption for energy generation.

YSXP describes the proportion of glucose consumption which does not go in the part for

power generation.

YanSXP boundary of YSXP when pO2 converges to 0.

YaeSXP boundary of YSXP when pO2 converges to 1.

Kslan determines the gradient of this transition function.

Table 3.2: The stoichiometric coefficients of the biological submodel.

Stoichiometric coefficients Values Core parameters Values

yOX 0 Kslan 1

yCOX 0 KSlcr 20

yOP 0 K50an 5

yCOP 2 KrSan 1

yOE 6 K50cr 0.45

yCOE 6

yS 180

yXX 120

With properties

rSanpO

KSwAn 02

lim

and

0.1lim2

SwAnpO

which under anaerobic conditions reduces the maximum substrate uptake rate to

rSanS Kr max

Half of the transition function:


28

2

1)(

2

1maxmax SrSanS rKr

The transition function represents the glucose consumption rate under different dissolved oxygen

levels as in Figure 3.4.

Figure 3.4: A scheme of the transition function.

The fraction, the yield factors under aerobic and anaerobic conditions:

aeSXPSXPPO

anSXPSXPPO

KPOK

anSXPaeSXPanSXPSXP

YYYY

e

YYYY

anslan

22

502

lim,lim

1

0

)(

of the total substrate uptake goes into growth and product formation while the reaction (1-YSXP ) goes

into the energy formation:

SSXPSE rYr )1(

The part of the substrate uptake rate that goes not into energy generation

SXPS Yr

splits into a fraction YXPX that serves biomass growth and a faction

)1( XPXY

that serves product formation.

aeXPXXPXPO

anXPXXPXPO

K

anXPXaeXPXanXPXXPX

YYYY

e

YYYY

anKPOslan

22

)502(

lim,lim

1

0


29

Consequently the rate of product formation is

)1( XPXSXPSXPP YYrr

while the rate of biomass growth is

SXPSXPXXPX YrYr

The yield factor changing with crabtree effect:

aeXhiaeXPXS

aeXloaeXPXS

KSK

aeXloaeXhiaeXloaeXPX

YYYY

e

YYYY

crSlCr

lim,lim

1

0

)( 50

In aerobic phase, the consumption of O2 and generation of CO2 are related with the rates for

substrate consumption, biomass growth and product generation.

The expression for O2 consumption is given as

The expression for CO2 is given bellow:

SECOEXPPCOPXPXCOX ryryry

The substrate used for energy formation:

1lim

1

00

1lim)1(

0

)(

2

0

2

502

2

anSXPanSXPSXPPO

KPOK

anSXPaeSXPanSXPSXP

SE

SXPPO

SSXPSE

YYY

e

YYYY

POwhenr

YrYr

anslan

Total substrate uptake and biomass/product formation rates

SEXPPXPXS rrrr

Boundary for SXPX rr is as follows:

0S S

02 PO anSXPanXPX YY anSXPanXPX YY

SEOEXPPOPXPXOX ryryry


30

2PO aeSXPaeXlo YY aeSXPaeXhi YY

The yield is given as:

SS

XPXXX

rY

rY

producedegluofamount

producedbiomassofamount

cos

It presents the practical circumstances of the cultivation process.

3.2 Basic theory of parameter identification

Factors such as variations in the quality of raw materials, characteristics of the initial charge media,

disturbances in process conditions and the time varying behavior of the fermentation process lead to

mismatch between the model and the actual plant, leading to deterioration in the performance of the

fermenter optimization system. The success of a feedback control scheme depends heavily on the

reliability of the parameters of the model. Consequently the model parameters need to be updated

online to reduce the model mismatch. Also the optimal trajectories should be recalculated based on

the updated model and state information to maximize the product yield.

The model parameters could be adapted to the experimental data so that the course of biomass and

product formation, glucose utilization, waste gas composition and dissolved oxygen concentration of

the experiments show very good agreement. The measured values of the state variables are compared

with the model predicting values of the same state variables. The model parameters should be well

estimated by a certain identification method to minimize the cost function J, which includes weighted

least squares:

2

1 1

exp

,, )()(

n

i

m

j

ji

sim

jiij yywpJ

where p is parameter vector, i is index of measured variable, j is index point of time, m is number of

measurements, n is number of measured variable, w is weighting factor, exp

, jiy is measurement data of

the data set i at points of time j, and sim

jiy , is model output of data set i at points of time j.

In principle, all products of the factors i and j are summed with individual weighting. However, it can

be assumed that the measurement errors are uncorrelated, both among the different measurements and

also among the various time points. Then the weighting matrix has only diagonal elements.


31

More rational form with weighting factors is:

2

1 1

exp

,,

2

1 1

exp

,,, )(

n

i

m

j i

ji

sim

jin

i

m

j

ji

sim

jiji

yyyywSSR

where i is the weighting factor used in the OLFO controller. It is used to define the different roles

that each experimental variable plays in the calculation of the residuals. Their values are determined

based on the magnitude of relative experimental errors. That is the form used in this work.

Based on the great dynamic of the model parameters, a stable parameter identification should be

provided, which fulfilled the following requirements:

1. The parameter identification process must be fast, so that the parameter identification can be

implemented online.

2. The parameter identification process must be robust and stable, for the purpose to adapt the

general process model. Estimation results must be reasonable, and therefore one has to enable

the possibility that the user can define the boundary conditions of parameters. In addition, the

identified parameters should be precise enough to perform the optimization based on the last

calibrated model.

The differential equations of the model can be solved by DASSL algorithm (Brenan et al., 1995) or 4th

order Runge Kutta algorithm (Press et al., 2007) with variable step size. Parameter identification part

can be done by the robust Nelder-Mead algorithm (Nelder & Mead, 1965) or Gradient algorithm

(Snyman, 2005), and the optimization part can be done with the same algorithm. Different algorithms

are set as control pattern within the OLFO controller, which can be chosen freely by the operator

according to the practical circumstances. The model parameters are estimated with the following

measurements:

1. The values of the concentration of biomass, product and substrate in the broth are measured

through lab analysis every hour or few hours. The sampling interval can be adjusted according

to practical situation;

2. The composition of vent gas and dissolved oxygen concentration from the control system are

measured every few minutes. The measuring interval is adjustable according to practical

situation.


32

3.3 Optimization criterion

After parameter identification, an estimated model is ready for the optimization. The models which

can be used in the optimization process must have sufficiently good prediction quality. A virtual

validation of the quality of parameter identification is done manually. After successful parameter

identification, an estimated model is ready for the optimization. Based on the prediction of the adapted

model and a suitable optimization criterion, a performance functional Jopt can be formed to get an

optimum reference value. Nelder-Mead algorithm or Gradient algorithm is used to minimize the cost

function to get an optimal substrate feed rate for the remaining part of the fermentation.

A suitable optimization criterion can be a maximum time-space yield, maximum income of the whole

process, a good medium utilization, to keep a certain growth rate or to maintain a constant substrate

feed rate etc. An optimum of this performance functional Jopt can be calculated by Pattern Search

Algorithms, e.g. Nelder-Mead algorithm, genetic algorithm etc. This optimum is saved as 'Control

Pattern', i.e. the optimal reference value of substrate feed profile, temperature, pH value, dissolved O2

concentration or combination of those optimal set points.

While conducting this work, the optimization of substrate feed rate is chosen as the 'Control Pattern'.

During the feed optimization part, input for the predictive algorithm is not artificially preset data for

optimization, but experimental data, the model, the set of previously identified parameters, a set of

feed flow starting values, and the optimization time interval. Output is an optimized set of feeding

profiles that best fulfills the optimization criterion provided by the model prediction.

Normally the fed-batch cultures are operated either at a constant feed rate, at a linearly increasing feed

rate or at an exponentially increasing feed rate. This process is operated in an aerobic growth phase

and an anaerobic production phase. In the growth phase one observes that the biomass grows

exponentially. During the production phase, the biomass growth slows down considerably.

Accordingly, the substrate feeding profile can be divided into exponential feeding stage and constant

feeding stage. The substrate feeding profile is modified by four optimization variables: the switching

time from aerobic phase to anaerobic phase tS; the initial pumping rate f0; the exponential factor for

the increase of the pumping rate α and the constant pumping rate for the anaerobic part f1. The substrate

feeding profile is defined as follows:

FS

SII

tttf

tttttff

;

));(exp(

1

0


33

The optimal control is based on the concept that one can obtain an optimum ethanol yield when the

ideal switching point from the aerobic biomass growth phase to the anaerobic ethanol formation phase

is found. The feeding rate functions are defined by four optimization variables which are generated by

the optimization process. The switching time tS is defined as:

)(1 IFIS tttt

where tF and tI define the final and initial process time.

The initial value of the feeding rate is limited by the maximum feeding rate.

max20 ff

The exponential factor of the feeding is between zero and a maximum value.

max3

The constant feeding rate in the anaerobic phase is then limited by the maximum feeding rate.

max41 ff

The four optimization variables 321 ,, and 4 defined the substrate feed rate functions. They all

vary between 0 and 1.

The following boundary conditions are considered as constraints of the optimization: maximum

working volume of the fermenter; the substrate supply is finite and the maximum pumping capacity

maxf is limited.

This optimization pattern is used in most of our experiments related with the OLFO controller. The

number of optimization variables is low, so the optimization only occupies limited amount of the

computing power of the system.

3.4 Workflow of the OLFO controller

The work flow of the OLFO controller is illustrated by Figure 3.5. The typical steps to implement

the OLFO controller are as follows:

1. The fed-batch fermentation process is started by charging the media into the fermentation

reactor, starting the agitator and initiating the air ow through the broth.


34

2. All the plant operating parameters like air flow rate, pH value, temperature, agitation rate etc.

are measured. Their initial values are stored in the OLFO controller and available for the

calculations.

3. Periodically, the broth samples are collected and analyzed in the lab for the concentration of

substrate, product and biomass. The analysis results are transformed into suitable form and

stored in the OLFO controller.

4. With the initial conditions (broth volume, biomass concentration, product concentration,

substrate concentration), the optimal substrate feeding rate profiles are calculated.

5. While the fed-batch fermentation is in progress, the following steps are implemented:

(a) Once the predetermined schedule of the initial stage of fermentation is done, an online

parameter identification process is carried out based on the actual process data collected

from the plant and laboratory analysis. The parameters are estimated by minimizing the

error between the measured and simulation values for concentration of biomass, product,

substrate, dissolved oxygen in the broth and composition (O2 and CO2) of off gas flow. A

nonlinear optimization technique, e.g. Nelder-Mead method or Gradient algorithm is used

to minimize the plant model mismatch.

(b) The new identified parameters and the updated process state variables are used in

calculating the optimal substrate feed rate.

(c) The optimization part predicts the future course of process states using the model and its

new adapted parameters. A control profile is precalculated based on an optimization

criterion (e.g. controlling substrate concentration at a certain level, maximizing time-space-

yield etc.). With the approval from the evaluation, the updated control profiles are

transferred to the process control system and fermenter.

(d) The aforementioned steps a, b and c are executed at every optimization calculation period

using the updated database. This periodic re-estimation of the model parameters and update

of the state variables while the fed-batch is in progress as it helps in reducing the plant-

model mismatch leading to improved performance of the optimizer.


35

Figure 3.5: Work flow of the OLFO controller. When new data becomes available from the bioreactor, an automated

parameter identification takes place, and then the optimization part calculates the optimal control profiles based on the

identification model. Finally, the control profiles are transferred to the direct control level.

Since the OLFO controller is developed as a general process optimization tool-box, there are more

demands than the normal version. Above all, a general process model with a high adaptability is

required, which affords the ability to describe the characters of multiple cultivation processes.

Accordingly, the parameter identification can adapt the parameters change due to the metabolism

change or the change of the system. A version of the OLFO controller that works stable in conjunction

with a general model and fit sophisticated requirements toward adaptability and flexibility is the

purpose of my work.

4. Research results and application examples of

OLFO in fed batch processes

In this chapter, a compartment model is used to present the evolution of fermentation processes. The

Gradient algorithm is formulated to calculate the process model mismatch at each interval to update

the model parameters. The Nelder-Mead method is used to optimize the objective function dynamically

throughout the course of the fed-batch fermentation process.

At the first stage, the result of the OLFO controller was not used in the real fermentation. The

experimental data from the fermentation was used for the evaluation of the performance and

characteristics of the parameter identification part and optimization part. During this research, some

illuminated features of the OLFO controller are discovered. Those results are summarized in the first

section of this chapter. Those discovers are the basis of the design of the subsequent experiments.

At the advanced stage, as a software tool which facilitates the use of biotechnological models, the

OLFO controller is mainly used in the following ways. First, the OLFO can combine with the virtual

bioreactor to be used as a training tool set in academia and industry. Second, the OLFO provides

optimized feed rate profiles for the real fermentation processes. Third, the OLFO can be used also as

a software observer for the process variables which are difficult or costly to obtain.

4.1 Basic research on parameter identification and optimization

With suitable weighting factors for measurements, the parameter identification can achieve a satisfying

result and is relatively insensitive to different initial values and boundaries. The scheme of parameter

identification is carried out online periodically based on the plant measurements and laboratory

analysis results. This ensures a model of higher accuracy used in calculating the optimum feed rate.

Particular constraints that may largely influence the parameter estimation result, such as weighting

factors on state variables, initial values and boundaries of parameters etc. must be taken into account.

To ensure a successful parameter identification, a series of simulation experiments have been

accomplished to evaluate the significance of the constraints and their proper use. The most relevant

results are presented in this thesis.

Chapter 4 Research results and application examples of OLFO in fed batch processes

37

A series of simulations have been carried out to demonstrate the influence of the following constraints

on the parameter identification process:

1. The density of offline measurements (different sampling time of offline measurements);

2. Different weighting factors;

3. Different initial values of parameters;

4. Different boundaries of the parameters;

5. Absence of partial measurements;

6. With more online measurements, less offline measurements.

4.1.1 Different sampling time of offline measurements

The rate to take the samples has some effects on the fermentation process. To take a sample, it increases

the probability of contamination and also changes the volume of the broth in the reactor as the volume

of the sample is difficult to be taken into account precisely. It is advisable to take samples as less as

possible. Therefore it is necessary to learn about the minimum amount of offline measurements which

is requisite for the accuracy of parameter estimation.

Normally, the concentration of product, biomass and substrate can only be measured offline, whereby

the results are available for control purposes only after some delay. The delay to get the measurements

of biomass concentration is about a few minutes. The delay for the product measurements varies.

However it costs at least half an hour to get the measurements of substrate concentration.

Following this, it becomes evident to determine if it is feasible to reach the accuracy of parameter

estimation by taking the samples in a longer interval. With a long sampling interval, it can reduce not

only the probability of contamination but also the manpower and costs on the analysis of the samples

which leads to the cost reduction of the whole process.

In this investigation, three different conditions are set up:

1. Measurements of product, biomass and substrate concentration are sampled every 30 minutes;

2. Measurements of product, biomass and substrate concentration are sampled every 60 minutes;


38

3. Measurements of product and biomass concentration are offered every 60 minutes, the substrate

concentration every 180 minutes.

In order to reduce the risks of contamination, sampling costs and time, three different sampling

strategies were used and their effect on the parameter identification performance was evaluated (see

Figure 4.1). From the results of this investigation, all of the simulation results fit the measurements

well, without significant difference. As a reference, an accurate parameter identification was obtained

for sampling intervals of 30 minutes for P, X and S. Reducing the sampling intervals for P and X to

60 minutes and for S to 180 minutes, did not affect significantly the parameter identification results

which appeared to be in the difference range of -1.3% to 2.0% from the reference simulation. The good

agreement with the reference simulation showed that longer sampling intervals could be applied for

the further studies.

Figure 4.1: Parameter identification result with different sampling time. S-exp. is the experimentally measured substrate

concentration, S-sim. is the related parameter estimation result. P and X are the product and biomass concentrations

respectively. There are three groups of simulation results shown here: the first group simulation with P, X, S every 30

minutes by thickest lines, the second group simulation with P, X, S every 60 minutes and the third group simulation with

P, X every 60 minutes, S every 180 minutes by thinnest lines. All of the simulation results fit the measurements well,

without significant difference.

From the subsequent investigation described in Chapter 4.4 Software observer, the OLFO controller

can simulate the fermentation well even without the offline measurements on aerobic phase. For


39

anaerobic phase, the success of parameter estimation still requires one of the offline measurements.

To overcome this problem, the amount of gas produced could be used as an online measurement. This

investigation also demonstrates that the density of the offline measurements makes slight difference

on the parameter estimation result.

4.1.2 Different weighting factors

Proper weighting factors of process variables are essential to ensure a successful parameter

identification. The objective function without weighting factors might lead to neglect the magnitude

distance between different variables, which means the measurements of high magnitude are

overweighed.

From the results of this investigation (Figures 4.2 and 4.3), we observe that, the weighting factors of

parameters have great influence on the result of parameter identification. By putting the appropriate

weight on certain state variables renders an obviously better description of parameter estimation.


40

Figure 4.2: The parameter identification result without individual weighting factors. The experimental data (referred to as

pixels) are connected with the corresponding - best matching - model data (solid as lines) are shown together. The upper

part shows S: substrate, P: product and X: biomass; the lower part shows the corresponding simulation and experimental

data of off gas composition. Without appropriate weighting factors, the model cannot fit the measurements.

Figure 4.3: The best parameter identification result with suitable individual weighting factors: 2CO : 0.1,

2O : 0.1, S

: 0.001, P : 1.0, X : 1.0. is the weighting factor for each individual variable which depends on its dimension and

reliability. The experimental data (referred to as pixels) are connected with the corresponding - best matching - model data

(solid as lines) are shown together. The upper part shows S: substrate, P: product and X: biomass; the lower part shows the

corresponding simulation and experimental data of off gas composition. By putting the appropriate weight on certain state

variables renders an obviously better description of parameter identification.

Moreover, weighting factors can also be used to heighten the influence of some important variables or

more reliable measurements or weaken the influence of some unreliable measurements, in order to

make the estimated model close to the real fermentation process as much as possible.


41

4.1.3 Absence of partial measurements

The purpose of this evaluation is to figure out how the parameter identification evolves in the absence

of partial online measurements or offline measurements at some time range as a result of technical

malfunction, such as a temporary defect in the measuring device.

From the simulation results, we observe that the absence of particular online measurement has

significant influence on the parameter identification, especially for the parameter YaeXhi & rSmax. In

this example, the deviation of the parameter rSmax is about 14% and of YaeXhi is about 91% (See

Appendix 1). Therefore one should pay closer attention to the accuracy of the online measurements

especially the percentage of CO2 and O2 in the off gas. The OLFO controller is less sensitive to offline

measurements as manifested by the low mean deviation of about 2% with the same example in this

research. There one can note, online measurements play more important roll compared with offline

measurements in the parameter identification process. The accuracy of online measurements should

be more regarded.

4.1.4 Different initial values and boundaries

From the results of the research, with appropriate weighting factors of state variables, the different

initial values have nearly no influence on the parameter identification result except the value of YaeXlo

in the first time range. On the other hand, the optimization result is relatively stable to the change of

YaeXlo (see the optimization result analysis with the change of parameters). We can notice that, the

identified parameters are relatively insensitive to the different initial values of state variables in the

'step by step' case which is the way how the OLFO controller works. For the 'one step' parameter

identification, the result is extremely sensitive to the initial values. The simulation can fit the

measurement well only with very good initial values.

We can conclude from the simulation results, the boundaries of the parameters have very little

influence on the result of parameter identification, which can be ignored for the optimization.

Nevertheless, a parameter identification with a reasonable variables boundaries is still advisable, since

the variables make sense only when they are within operating boundaries which are set according to

the actual physical conditions.


42

4.1.5 Influence of each parameter on the optimization

Some studies have been done to evaluate the influence of different factors on the parameter

identification and further on the optimization results. In each simulation, only one parameter is

changed to evaluate the influence of this parameter on the optimization. A part of the results of this

research is presented in Table 4.1.

Table 4.1: Influence of different parameters on optimization.

Parameters Pi

Opti. fs

i=1

(rSmax)

i=2

(YaeSxp)

i=3

(YaeXhi)

i=4

(YaeXlo)

fsopt(1.1Pi) -21.04% -14.70% -14.11% 0.28%

fsopt(0.9Pi) 20.29% 15.41% 15.34% -0.23%

fsopt(0.95Pi) -10.36% -7.44% -7.08% 0.12%

fsopt(1.05Pi) 10.08% 7.66% 7.59% -0.12%

fsopt(0.98Pi) -4.15% -2.98% -2.78% /

fsopt(1.02Pi) 3.93% 3.02% 3.09% /

fsopt(0.995Pi) -1.02% -0.76% -0.59% /

fsopt(1.005Pi) 1.02% 0.76% 0.87% /

fsopt(xPi) is the calculated differences in [%] of the feeding profiles optimized with a reference

parameter Pi and a modified reference parameter xPi where x is a coefficient that varies between 0.9

and 1.1. The interval of re-optimization is set 3 hours as an example in this work. In table 4.1, the

optimized feed profile in 3 hours is used for the evaluation. The optimized feed profile with the

expected parameters is used as reference. The change of the parameters rSmax, YaeSxp, YaeXhi have

similar influence on the optimization results. The optimization result is stable upon influence of the

parameter YaeXlo.

4.1.6 The performance of the OLFO controller

In order to demonstrate the performance of the OLFO controller in comparison with other fermentation

processes, some experiments results are presented in this thesis. A batch, a standard fed-batch and a

fed-batch with the OLFO controller were carried out under approximately the same experimental

precondition in order to compare their outcome. The standard fed-batch experiment is implemented to

keep a constant substrate concentration at about 1 g/L. The performance of the OLFO controller is

compared to this experiment in order to demonstrate the achieved process improvement when using

the exponentially fed batch strategy by the OLFO controller. In the comparison, the optimization


43

criterion JF (tf ) = 15 S(tf ) X(tf ) P (tf ) was used in the OLFO controller. The comparison of the outcome

of different experiments is presented in Table 4.2.

Table 4.2: Comparison between different fermentation processes.

Biomass conc.

[g/L]

Substrate conc.

[g/L]

Ethanol conc.

[g/L] JF(tf )

Batch 5.9 0 4.8 -10.7

Standard fed-batch 8.6 0.1 4.8 -11.9

Fed-batch with OLFO 7.8 0.1 6 -12.3

Note: Experiments were carried out in a 20.0 L stirred bioreactor (virtual bioreactor), 60% pO2 at the

beginning, after 6th hour, ca. 10% pO2, 32 °C. The biomass concentration in the inoculum (200.0 ML)

was 25.0 g/L. Substrate concentration in medium (5.0 L) was 2.5 g/L for fed-batch fermentation, 20.0

g/L for batch fermentation. The fed-batch fermentations last 10 hours, the batch fermentation takes 6

hours.

4.2 OLFO works with the virtual bioreactor

A reliable process model is served as the plant and embedded in a self-constructed software namely

virtual bioreactor. This process model is developed by Prof. V. Hass and has been verified in his work

group during long years of experiments (Hass, 2005). However due to its complexity, it does not fit as

the general model in the OLFO controller. The general process model "Lyx" as mentioned in section

3.1 is different from the process model which is embedded in the virtual bioreactor. Training with

virtual bioreactor leads to improved operating capability with laboratory cultivations for the groups

who have obtained such training in comparison to the groups who have not (Gerlach et al., 2014). The

combination of virtual bioreactor and the OLFO controller can be used as a training tool for academia

and industry with an aim to offer the user insights into fermentation process control.

The implementation of a model calibration requires a data set with characteristic measured variables

of the process. A model calibration can be implemented successfully only when all the control

variables that have influences on the process are available. Those data must be synchronized with time.

The subsequent optimization generates optimized control profiles. These profiles can be applied to the

process.


44

Synchronization of the OLFO sub processes

process data

formatting model calibration

process optimization

is done by a monitor process programmed by Perl. The sub processes are implemented in the C++

language or in the FORTRAN language. The offline measurements are transformed into a form

suitable for WinErs by a simple tool-box based on OpenOffice which I developed as part of this work

(See Appendix 3); and the online measurement data are automatically generated in a suitable form from

the OLFO controller; The interval of sampling, the interval to generate the measurement and the

beginning of the real process time (the time point of inoculate) should be set ahead. The unit

transformation is realized automatically. For example, the unit of feeding profile in virtual bioreactor

is ml/min, but in model is m3/s. After the optimization process, a new generated optimal feeding profile

can be used for the direct control level after evaluation. A button is set to active the feeding profile

acceptance (See Figure 4.5).

Figure 4.4: The graphical user interface of the virtual bioreactor


45

Events-sequence of the virtual bioreactor with

the OLFO controller is as follows:

1. New data are acquired;

2. Update of the four parameters by

minimizing the sum of squares;

3. Calculation of a new optimal feeding

profile (from now until Tf );

4. Verification and release of the new profile;

5. Replacement of the old feeding profile

(from the current point in time).

An example of the optimal control of a virtual yeast cultivation is described to exhibit the basic

procedure of the OLFO controller:

1. The coupling with WinErs is realized as follows:

(a) The measurements from the cultivation can be exported via a special driver;

(b) The OLFO controller works as an external analysis software to calibrate the model;

(c) The OLFO controller is used to create optimal control profile;

(d) After evaluation and approval, the driver imports the updated optimal control profile for

the direct control level to be applied in the process.

2. The control of the OLFO controller includes:

(a) To trigger the OLFO controller, the reference time for the fermentation process (after

inoculation) and the interval for the automatic data export should be set;

(b) The activation of an optimized control profile.

3. Calibration of the following four model parameters:

(a) rSmax: maximum substrate uptake rate.

(b) YaeSxp: maximal yield under aerobic condition.

Figure 4.5: The user interface of the virtual

bioreactor to set the interval of sampling, the

interval to generate the measurement and the

beginning of the real process time.


46

(c) YaeXlo: maximal yield of biomass when substrate concentration is low.

(d) YaeXhi: maximal yield of biomass when substrate concentration is high.

4. Exponential substrate supply is parameterized.

5. )()()(15)( ffftF tPtXtSJf

is formulated as the cost function aiming at consuming all the

substrate in the fermenter and producing biomass and product as much as possible at the end of

the process. The numbers before the variables stand for the weight that you want to put on the

variables. Bigger number stands for more weight. This cost function is randomly set and only

used as a demonstration.

The complete experiment runs for 10 "virtual" hours (speed-up is possible) and depicts a fermentation

unit in the aerobic phase. During this demonstration, parameter identification is performed every 2.5

hours and implemented with the new generated input data by the virtual bioreactor. Based on the

updated estimated model of the virtual fermentation process, OLFO controller generates optimal

feeding profiles for the remaining time of the process. This profile is then used as a feedback for the

virtual fermentation process (Figure 4.6). The parameter identification as part of the whole procedure

is demonstrated in a stepwise manner (Figure 4.6.A). Based on the parameter identification results, the

simulation of the prediction of the estimated model (Figure 4.6.B, upper diagrams) and the newly

generated optimal feed rate profiles (Figure 4.6.B, lower diagrams) are plotted for each interval. The

current general model does not contain any inhibition function that accounts for cases of substrate

inhibition. However the OLFO controller, implements a penalty factor that sets in control once a high

level of substrate concentration is present. In this example, the penalty factor is manually set to 1000.

This means that once the substrate concentration is over the 4 g/L threshold value, the penalty factor

will be activated which will lead to a subsequent reduction on the substrate feeding rates. As shown in

Figure 4.6, at a time point near the 6th hour, the substrate concentration which is reaching the threshold

value, suddenly starts to deplete. This is due to the penalty function which is activated at this point.

On the other hand, the biomass grew faster and also consumed the substrate faster, which also led to

the rapid depletion of the substrate concentration.

The current tandem system combines several advantages. The tandem system offers an alternative to

learn and gain experience through "trial and error" without using fermentation material. Moreover, the

simulation can be accelerated in order to shorten the training process which makes it a very efficient

toolbox considering that the real fermentation processes are very time consuming.


47

Figure 4.6: (Left) Parameter identification graphs with 2.5 hours interval, with corresponding (right) optimization of the

estimated models (upper) and the optimal feeding profile (lower). Substrate (S), product (P) and biomass (X) concentrations

are further labeled based on their generated source virtual bioreactor (experimental) or OLFO (simulated).


48

4.3 OLFO application for yeast fed-batch cultivation

A cultivation of yeast cells with the aim to produce maximum amount of ethanol was supported by the

OLFO controller. The parameter estimation are based on the model "Lyx" as mentioned in section 3.1.

The optimization calculations are based on each updated calibrated model. For a better comparability,

the process optimization is always carried out with identical boundary conditions. These constraints

are related to the duration of cultivation, the initial concentrations and operational limits of the

instrumentation (e.g. maximum flow rate of the pump). The only variable that can potentially cause

fluctuations on the model parameters is the updated measurements.

The evaluation of process quality is determined by the choice of the corresponding cost parameters

so as to maximize the benefit at the end of the process. It depends on the market price of the product

(in this case ethanol) eP, mass specific substrate costs eS and volume specific costs of product isolation

eI. Substrate mass is the initial concentration in the vessel and the additional feed solution being

pumped in during fermentation. Product is contained in the final volume of the vessel, where the

maximum volume has been reached. Consequently, this function evaluates as

)()( ,

00

2 mvapPRPOHRI

t

t

in

SSRSPRP HCVTCVedtfCCVeCVeBF

I

∆T describes the temperature increase from the fermentation temperature to the boiling point. Using

vaporization factor of molar enthalpy mvapH , , the energy lost by the distillation of ethanol

evaporation can be estimated.

In this experiment, eS and eI were set 0, which means cost function was set to maximize the ethanol

concentration at the end of the process.

As mentioned before, as the process model is formulated without substrate inhibition, a penalty factor

is set to control the situation with high substrate concentration. In this example, the penalty factor is

set 1000. The maximum substrate concentration of 10.0 g/L is used as a constraint in the optimization.

A violation of this boundary condition produces a high punitive contribution to the quality criterion

to reduce the substrate feeding rate.

There are all in total nine parameters set free: initial concentration for substrate, biomass and product,

KS for substrate uptake, rSmax for substrate uptake, and four yield coefficients: YaeSxp, YaeXlo,

YaeXhi, YanXpx (active for anaerobic phase). The weighting factors for different measurements were

set according to the measurements density. After the batch fermentation part, the parameter

identification and optimization is implemented every 2 hours.


49

The measured concentration of substrate, biomass and product at the beginning of batch are 2.6g/L,

0.7g/L and 0.0g/L respectively. The other settings of the process were basically identical to those in

chapter 2. Laboratory set-up. After about 2.5 hours, the substrate was about to deplete which means

the end of the batch phase. At that moment, the OLFO controller was activated.

1. Based on the model calibration of a previous fermentation experiment, an optimal control was

calculated for the process. The precalculated optimization result (Figure 4.7) was used between

the 2.5th-5th hours.

Figure 4.7: Precalculated optimization results based on the calibrated model from a previous fermentation experiment.

The upper part shows the optimized feeding rate fS and the inlet gas rate fgas; the lower part shows the corresponding

prediction of substrate S and product P and volume of the vessel VR. Only the profile of the first 2.5 hours was used in the

fermentation as the first approve of optimization result from OLFO controller.

2. Considering the time delay for parameter identification, optimization and evaluation for both

part, the implementation of the first parameter identification based on the current measurements

occurred after 4.5 hours. The simulation data and the measurements are in a good agreement

(Figure 4.8 ). From the optimization result (Figure 4.9), an exponential substrate feeding profile


50

is proposed for the aerobic phase. A switching time to the anaerobic production phase at about

9th hour of the process time is proposed. The flow rate of the substrate feeding in the anaerobic

production phase is lower, but with a higher inlet concentration. The mass supply of substrate

in the anaerobic phase is larger. The optimized feeding rate 'fS.opt' from the folder 'opti1' was

activated to control the substrate pump rate from the 5th hour.

Figure 4.8: The first parameter identification result after 4.5 hours. The experimental data (referred to as pixels) are

connected with the corresponding - best matching - model data (solid as lines) are presented together. The upper part shows

S: substrate, P: product and X: biomass; the lower part shows the corresponding simulation and experimental data of off

gas composition. The simulation data and the measurements are in a good agreement.


51

Figure 4.9: Optimization after the first parameter identification. The upper part shows optimized profile for substrate

feeding rate ' fS.opt ' and inlet gas flow rate ' fgas.opt ', the lower part shows the prediction of product concentration '

P.opt ', substrate concentration ' S.opt ' and the volume of the nutrient in the fermenter ' VR.opt '. The optimized feeding

rate ' fS.opt ' was activated to control the substrate pump rate from the 5th hour.

3. The second parameter identification was implemented after 6.5 hours. The simulation data and

the measurements still remained in a good agreement (Figure 4.10). However the system

reported a convergence error of the algorithm by minimizing the sum of squares. This does not

necessarily mean that the updated model parameters must be discarded. But one should be

cautious about using the follow up optimization result based on this calibrated model. This

updated optimal control profile (Figure 4.11) showed a high deviation from the previously one.

The new optimized feeding profile 'fS.opt' from the folder 'opti2' was discarded. The previously

calculated control profile was continued.


52

Figure 4.10: The second parameter identification after 6.5 hours. The experimental data (referred to as pixels) are

connected with the corresponding - best matching - model data (solid as lines) are shown together. The upper part

shows S: substrate, P: product and X: biomass; the lower part shows the corresponding simulation and experimental

data of off gas composition. The simulation data and the measurements still remained in a good agreement.


53

Figure 4.11: Optimization after the second parameter identification. The upper part shows optimized profile for substrate

feeding rate ' fS.opt ' and inlet gas flow rate ' fgas.opt ', the lower part shows the prediction of product concentration ' P.opt

', substrate concentration ' S.opt ' and the volume of the nutrient in the fermenter ' VR.smo '. This optimization result shows

a high deviation from the previously one. It was discarded after evaluation.

4. The third parameter identification was implemented after 8.5 hours. The simulation data and

the measurements still remained in a good agreement (Figure 4.12). The optimization also fitted

the current state of the fermentation (Figure 4.13). Then the optimized feeding rate 'fS.opt' from

the folder 'opti3' became activated to control the pump rate at least from the 9th to 11th hour.

Particularly, the switching time was used. The process was turned into anaerobic phase at about

the 9th hour.


54

Figure 4.12: The third parameter identification after 8.5 hours. The experimental data (referred to as pixels) are connected

with the corresponding - best matching - model data (solid as lines) are shown together. The upper part shows S: substrate,

P: product and X: biomass; the lower part shows the corresponding simulation and experimental data of off gas

composition. The simulation data and the measurements remained in a good agreement.


55

Figure 4.13: Optimization after the third parameter identification. The upper part shows optimized profile for substrate


', substrate concentration ' S.opt ' and the volume of the nutrient in the fermenter ' VR.opt '. The optimization also fitted the

current state of the fermentation. Then the optimized feeding rate ' fS.opt ' became activated to control the pump rate.

Particularly, the switching time was used. The process was turned into anaerobic phase at about the 9th hour.

5. After two more hours, once again one parameter identification and optimization were

implemented. The simulation data and the measurements still remained in a good agreement. A

convergence warning about the model calibration occurred as well. Besides, there is one

parameter at its low boundary. The process was already in the anaerobic phase and therefore no

more optimized parameters for aerobic growth phase can be proposed. Nevertheless, the

resulting optimal control profile does not deviate significantly from the previously one. The

change in the pump power is negligible compared to the last valid optimal control profile and

therefore does not require correction of the current control profile.

6. After a process time of 18th hour, a high substrate concentration was showed from the sample

analysis. Therefore, an additional model calibration and optimization was performed. A zero

substrate supply was response to this optimization, as a maximum glucose concentration 10.0g/L


56

was set as a constraint of the optimization. At this point, the substrate concentration was already

significantly higher than the constraint 10.0 g/L. According to the model prediction, this

optimization proposal will cause the complete consumption of the substrate long before the

scheduled end of the process. So this proposal was not adopted. To gain time for a reassessment

of the situation, the substrate feeding rate was cut in half. Soon after that it was replaced by a

manually controlled simulation. The manual control based on the calculation that the residual

substrate concentration would be close to zero at the planned end of the process. A feeding rate

of 15% of the maximum pump rate was generated from this calculation. The analysis of the

process data at the final stage of the cultivation confirms the accuracy of this prediction.

Figure 4.14: The fourth parameter identification after 10.5 hours. The experimental data (referred to as pixels) are

connected with the corresponding - best matching - model data (solid as lines) are shown together. The upper part shows

S: substrate, P: product and X: biomass; the lower part shows the corresponding simulation and experimental data of off

gas composition. The simulation data and the measurements remained in a good agreement.


57

Figure 4.15: Optimization after the fourth parameter identification. The upper part shows optimized profile for substrate


', substrate concentration ' S.opt ' and the volume of the nutrient in the fermenter ' VR.opt '. The resulting optimal control

profile does not deviate significantly from the previous one.

7. At the 24th hour, we reached an ethanol concentration of 87.8 g/L, which has reached our set aim

for this experiment. This yeast fed-batch cultivation ended at about the 25th hour with an ethanol

concentration of 85.5 g/L. The development of the whole process and the actually applied

substrate feeding rate during the whole fermentation process are summarized at Figure 4.16.

We can see from Figure 4.16 that the prediction of the substrate consumption is correct. All the

substrate in the fermenter is consumed at the scheduled end of the process. The optimal feeding profile

generated by OLFO led the process to reach the set aim for this experiment. As the process is less

operator-dependent in comparison to conventional control techniques, the result can be easily

reproduced using similar conditions.


58

Figure 4.16: The top figure present offline measurements of the whole process. It can be seen that the constraint of the

maximum substrate concentration of 10 g/L is violated between about 12th h to 20th h. Further on, one can also see that the

prediction of the substrate consumption with the manual optimization is correct. At the end of the cultivation, the substrate

concentration is reduced to below 1 g/L. The lower figure presents the actually applied substrate feeding rate during the

whole fermentation process. After 2.5th h (the end of the batch phase), the precalculated profile is activated. This was

modified at about 5th h due to the correction of the optimal control. After 9th h, it was modified to a constant rate as the

process has been switched to the anaerobic production phase according to the optimal control. It must be considered that,

the substrate concentration in the feed has increased from 200 g/L to 610 g/L at the switching time. The reduction of

substrate feeding rate after 19th h is a response to the proposed zero feeding profile from the optimization. At the meantime,

a manual optimization described in the text was performed from 21th h in order to consume all the substrate in the fermenter

at the scheduled end of the process.

4.4 Software observer

Biomass, substrate and product concentrations are conventionally measured offline. Their acquisitions

are normally through the analysis of samples which are usually carried out offline at certain time-

interval and normally manually. Such operations, besides being costly and laborious often threat

fermentation culture with contamination. For this purpose, technological developments aiming to solve

this problem are needed.


59

Software observers are suitable for online monitoring of those offline measurements. They are an

alternative to the conventional sample-based analysis. It is used for automatic online calculation of

estimated values for process variables. A software observer is not a real analysis unit but is

implemented as a software. It uses online measurements which are available on any bioreactor, for

example outlet gas flow, to estimate those variables which are difficult to get online. These estimated

values can then be used for monitoring or control, as if they came from a real measuring device.

Application of the software observer holds some advantages: it reduces sampling and laboratory cost

and possibility for contamination; it has no time delay in acquiring the offline parameters: substrate,

biomass and product concentration; monitoring the culture course (online monitoring), which also lead

to early detection of possible defects and deviations. In general, it provides better chance to prevent

loses in the process and it facilitates the cell growth rate control, automatic control of time point of

innoculation or harvest. In general, it increases the possibility to apply automatic strategies for process

control.

The OLFO controller shows the potential to function as a software observer in a way to observe the

states of offline variables. A general scheme of the software observer implemented within the OLFO

controller is provided in Figure 4.17. Here one can notice that, the online measurements are used for

the calculation with the OLFO controller to get the simulated values for biomass, substrate and product

concentrations. The calculations are based on the period updated calibrated process model.

The simulation of the following example was based on the measurements from the yeast fed batch

fermentation December 2011 in our group. The basic settings of this experiment is the same as the

other experiments described in this thesis. When the OLFO controller is used as a software observer,

it works in the same way as it is used for the process optimization. Only the initial values but no latter

measured values of the observed variables are provided in the individual data set which is used for the

parameter estimation.

Figure 4.17: Schematic of a software sensor. Inputs are mostly online measurements which are facile to obtain. Outputs

are the simulated values of variables which are hardly or costly to obtain. The parameter identification part of the OLFO

controller here can function as a software observer.


60

In the aerobic phase, only the initial values of substrate, biomass and yeast concentration and online

measurements were used in the simulations. The last update before anaerobic phase was at the 10th

hour (Figure 4.18 (a)). For the whole aerobic phase, the OLFO controller simulated the behavior of

substrate, biomass and yeast concentration precisely only using the online measurements for the

parameter estimation. The process switched to anaerobic phase after about 10.5th hours. From that time

on, the outlet gas composition measurements are not meaningful any more since the inlet gas flow is

cut out. That is the reason that the OLFO controller failed to simulate the behavior of the process really

precisely (Figure 4.18 (b)). So since the 12th hour, the biomass concentration was used in the simulation

considering it is the least costly and laborious offline measurement to obtain compare with substrate

and product (in our case ethanol) concentration (Figure 4.18 (c)). The other offline measurements that

showed in these figures were not used in the simulation, but only as a reference to show the effect of

the simulation. The good fit (considering the measurements error and noise) of the simulated values

and experimental reference coming from the analysis of the offline measurements indicates that OLFO

can be used as a reliable software observer.

The simulation of the OLFO controller was implemented every 2 hours. Only the results of a few

critical moments are shown here. The development of the whole process is shown in Appendix 2.

Similar performance of the simulation was obtained with the measurements from other experiments

from our group. This implies that the OLFO controller can be used as a software observer for the

offline measurements.

This application of OLFO as a software observer for offline measurements does not aim to replace

traditional offline measuring techniques, but to offer additional monitoring capabilities for faster

response and longer sampling interval. Moreover, it can be of great value where the traditional offline

analyzing methods are inaccurate or impossible to operate under certain circumstances.

The follow up optimization result after each parameter estimation is not presented since here the OLFO

controller only functions as a software observer. Further research of those optimization results can be

considered. So that the OLFO controller could work as an observer and also used for the process

optimization at the same time.


61

Figure 4.18: The parameter estimation results around anaerobic phase. In the upper part shows S: substrate, P: product

and X: biomass; in the lower part shows the corresponding simulation and experimental data of off gas composition. The

figures of the left sides show the original parameter estimation results. On the figures of the right sides, the offline

measurements for aerobic phase, the product and substrate measurements for anaerobic phase are not involved in the

calculation of the parameter estimation. They are presented here as a reference, to show how effective the OLFO controller

is when it is used as a software observer.


62

Figure 4.19: The final parameter estimation result with online measurements and biomass concentration in anaerobic phase

(12th - 25th hour). The other parts of the offline measurements as on the figure of the right side are used as reference, not

involved in the simulation. In the upper part shows S: substrate, P: product and X: biomass; in the lower part shows the

corresponding simulation and experimental data of off gas composition.

5. Summary and Outlook

5.1 Summary of the overall work

Nowadays a large variety of products are manufactured in batch and fed-batch fermentation processes.

In comparison to the batch fermentation processes, fed-batch commonly exhibits higher efficiency.

However, the process control in fed-batch fermentation processes is complex as it depends on

understanding the specific cell culture properties. Development and validation of mathematical models

for fed-batch processes are cumbersome and time consuming, since the cell cultures exhibit complex

dynamics with unpredictable changes in the process parameters. To address such challenges and to

insure optimized process control, a new OLFO controller - a general model with an advanced controller

is developed and presented throughout this thesis.

In this work, fundamentals of the OLFO controller and the establishment of the OLFO controller

structure are firstly presented. Subsequently, coupling of the OLFO with the industrial process control

system "WinErs", was set to control on a biostatic laboratory fermenter. Both Nelder-Mead and

Gradient algorithm can be used for the parameter identification. Consistently with a related cost

function, the algorithm is also applied in the optimization part to calculate an optimal exponential feed

rate profile. In this context, several simulations with the OLFO controller have been performed to

evaluate the influences of various factors on the parameter identification and the subsequent

optimization. The results of those investigations are used as instructions for the later applications of

the OLFO controller (Li et al., 2012).

Several applications of the OLFO controller have been presented in the last part. First, it has been

developed and presented that the OLFO controller can generate an optimal feed rate profile for the

fed-batch fermentation process in the virtual fermenter. This, in principle can be used to run some

virtual experiments of process optimization with different setups. The application allows savings in

terms of experimental costs, which are appreciated for employing new cell lines. Further on, the

application can serve as training simulator for academia and industry. Second, application of the OLFO

controller in real fermentation laboratory scale processes has been validated. OLFO control over the

fermentation processes appears to be stable (Figure 4.16), even in cases when some deviation from the

normal behavior of cultivation occurs. In comparison with the conventional way of fermentation

process control, the OLFO controller is less operator dependent, and therefore leads to lower variations

Chapter 5 Summary and Outlook

64

influencing the product yield. The advantages of the OLFO controller are remarkable when one

encounters a less experienced operator or a new cell line. At last, the OLFO controller was presented

to exhibit function as a reliable software observer to estimate the offline measurements which are

usually obtained and analyzed manually. The results have been verified with our own process data and

satisfying results have been obtained (Figure 4.19).

5.2 Perspective on future OLFO controller developments and

applications

Developing the OLFO controller into an actual industrial multifunctional toolbox for fermentation

processes is probably the ultimate goal of this technology nowadays. In this context, one can identify

two areas which require more development e orts. The first is the general model, which currently lacks

substrate inhibition function, leading to an optimization result with a penalty function which is not the

real optimal result but a suboptimal outcome. The development of the OLFO controller is greatly

dependent on the development of the general model and therefore future efforts into improving the

general model is essential.

The current graphical user interface (GUI) for the OLFO controller is another area which needs

improvement. In general, GUI should be able to provide more user friendly embedding, which can

easily gain the OLFO controller wider research popularity and application. The current GUI version

does not allow facile access to some important settings, such as checking parameter identification and

optimization results in both figure and digital form in the context of real time analysis.

Appendix 1: Performance of parameter identification when lack of partial measurements

65

Appendix 1: Performance of parameter identification when

lack of partial measurements

The fed batch fermentation last 12 hours. There are updates every two hours. The difference compared to the

reference which are bigger than 2% are set in red.

The column "with all measurements" is set as reference. These are parameter identification results

with all measurements.

The column "Lack of partial online measurements" shows the parameter identification results when

there is lack of off gas concentration data from 2th to 6th hour.

The column "Lack of partial offline measurement" shows the parameter identification result when

there is lack of offline measurement from 2th to 6th hour.

Basic settings of the simulations:

Parameters name Initial value Lower limit Upper limit the weight of O2out 0,100

rSmax 5,00E-04 1,00E-05 1,00E-03 the weight of CO2out 0,100

YaeSxp 9,00E-01 1,00E-02 1,00E+00 the weight of S 0,001

YaeXlo 8,50E-01 4,00E-01 1,00E+00 the weight of P 1,000

YaeXhi 3,00E-01 5,00E-03 1,00E+00 the weight of X 1,000

The results of parameter identification under different scenarios:

Simulation

rounds

with all

measure

ments

error Lack of

partial

online

measure-

ments

error difference Lack of

partial

offline

measure-

ments

error difference

1.parameter: rSmax

1. 4,66E-04 4,03E-06 4,66E-04 4,03E-06 0,00% 4,66E-04 4,03E-06 0,00%

2. 4,62E-04 4,01E-06 4,70E-04 5,02E-06 1,80% 4,50E-04 3,96E-06 -2,64%

3. 4,37E-04 3,67E-06 4,71E-04 1,95E-07 7,66% 4,26E-04 3,63E-06 -2,45%

4. 4,26E-04 3,55E-06 3,86E-04 3,14E-06 -9,45% 4,15E-04 2,38E-06 -2,52%

5. 4,34E-04 3,48E-06 3,78E-04 5,02E-06 -12,96% 4,25E-04 3,56E-06 -2,04%

6. 4,46E-04 3,62E-06 3,82E-04 5,32E-09 -14,39% 4,38E-04 3,68E-06 -1,72%

2.parameter: YaeSxp

1. 9,22E-01 3,01E-03 9,22E-01 3,01E-03 0,00% 9,22E-01 3,01E-03 0,00%

2. 8,67E-01 2,02E-03 9,16E-01 3,17E-03 5,66% 8,62E-01 2,30E-03 -0,61%

Appendix 1: Performance of parameter identification when lack of partial measurements

66

3. 8,26E-01 1,62E-03 9,16E-01 2,82E-03 10,97% 8,22E-01 1,62E-03 -0,40%

4. 8,06E-01 1,04E-03 8,05E-01 1,31E-03 -0,10% 8,05E-01 1,03E-03 -0,18%

5. 8,23E-01 5,91E-04 8,26E-01 6,32E-04 0,27% 8,23E-01 5,92E-04 -0,06%

6. 8,40E-01 3,78E-04 8,42E-01 7,12E-05 0,21% 8,40E-01 3,78E-04 -0,02%

3.parameter: YaeXlo

1. 8,50E-01 4,92E-01 8,50E-01 4,92E-01 0,00% 8,50E-01 4,92E-01 0,00%

2. 4,00E-01 8,26E-04 4,00E-01 3,38E-01 0,01% 5,29E-01 7,51E-02 32,28%

3. 5,76E-01 1,17E-02 1,00E-00 5,18E-02 73,60% 6,03E-01 1,18E-02 4,72%

4. 7,72E-01 5,84E-03 8,03E-01 7,04E-03 4,10% 7,81E-01 5,85E-03 1,17%

5. 7,82E-01 2,96E-03 7,75E-01 3,92E-03 -0,82% 7,85E-01 2,95E-03 0,46%

6. 7,34E-01 1,82E-03 7,26E-01 2,46E-04 -1,01% 7,35E-01 1,81E-03 0,22%

4.Parameter: YaeXhi

1. 5,98E-01 1,41E-02 5,98E-01 1,41E-02 0,00% 5,98E-01 1,41E-02 0,00%

2. 3,70E-01 8,40E-03 4,62E-01 1,29E-02 24,71% 3,79E-01 9,07E-03 2,51%

3. 4,30E-01 8,99E-03 4,61E-01 4,99E-04 7,20% 4,42E-01 9,34E-03 2,79%

4. 4,56E-01 9,28E-03 7,51E-01 1,38E-02 64,51% 4,69E-01 6,57E-03 2,71%

5. 4,27E-01 8,55E-03 7,80E-01 2,12E-02 82,54% 4,32E-01 9,18E-03 1,16%

6. 3,94E-01 8,19E-03 7,56E-01 1,31E-04 91,79% 3,95E-01 8,75E-03 0,17%

Appendix 2: Performance of the OLFO controller as a

software observer

In the following section, the results of parameter estimation with the measurements from the yeast cultivation

of december 2011 are shown here. Before anaerobic phase, parameter estimation was implemented only with

online measurements. The offline measurements on the figures of the right side are used as reference, not

involved in the simulation. After the switching time, the anaerobic phase was settled. The biomass concentration

measurements were also involved in the simulation, as the use of the vent gas composition measurements and

dissolved oxygen concentration from that moment was not meaningful anymore.

The following figures show the parameter estimation results every two hours real time along the cultivation

process. The figures on the left side are the original results. On the right side, the offline measurements which

are not shown on the left side are plotted here as an object of reference in order to show the effect of the OLFO

controller when it is used as a software observer. For aerobic phase, the OLFO controller simulates the behavior

of biomass, substrate and product concentration real time only with online measurements, most important of all

the measurements of off gas composition. The simulation of all those offline measurements and the

measurements themselves show really good agreement. For anaerobic phase, the biomass measurements are

used as well for the parameter estimation since the measurements of off gas composition do not make sense at

this stage. The simulations also fit with the measurements all along.

Appendix 2: Performance of the OLFO controller as a software observer

68


69


70


71

Figure A1: Parameter estimation results with only online measurements for aerobic phase; with online measurements and

biomass measurements for anaerobic phase. The upper part of each figure shows S: substrate, P: product and X: biomass;

the lower part shows the corresponding simulation and measured values of off gas composition. The figures of the left

sides show the original parameter estimation results of the whole process. The parameter estimation updates every two

hours until the end of the process. On the figures of the right sides, the offline measurements for aerobic phase, the product

and substrate measurements for anaerobic phase are not involved in the calculation of the parameter estimation. They are

presented here as a reference, to show how effective the OLFO controller is when it is used as a software observer.

Appendix 3: Data processing toolbox with OpenOffice

72


In order to deal with the measurements to create individual documents for every required variables

automatically, a data processing toolbox based on a free open source office suite, namely OpenOffice, was

developed during my PhD. The instruction of how to set up this toolbox and the source code of this toolbox is

explained in this section.

1. Initial settings of OpenOffice toolbox

a) In the Tools browser, get a hierarchical view of the contents by clicking

Macros Organize Macros OpenOffice org Basic, then you get the following window as

Figure A2:

Figure A2: After opening the Macros, one can select the library that contains Module 1

b) Select the library that contains the macro Module 1 new.


73

c) Insert BASIC source, select the last version → open → save. Then you have a window as Figure A3:

Import Dialog → open dialog 1 → open dialog2 → save. Then you have a window as Figure A4:

Figure A3: Importing BASIC source files

Figure A4: Importing dialog files

Until here, the toolbox is ready to be used.


74

2. Toolbox User's Guide

a) In the Tools browser, get a hierarchical view of the contents by clicking Macros → Run Macros.

b) Select the library that contains the macro Module 1 (Figure A5).

Figure A5: Selecting the library that contains macro Module 1.

Select the macro Offline and then klick Run (Figure A6). Then you get the following window: The

corresponding data can be filled in as needed:


75

Figure A6: Data box for offline measurements

Row Toplimit: The top limit of row

Row Downlimit: The down limit of row

Timecolumn: the column of relative data time

Datacolumn: the column of relative data

Filename: the file name of relative data

Conversion: the conversion coefficient from g/L to mol/L, e.g. The molecular weight of glucose is

180, then you can just fill 180 in the relative blank space. When you don't want any conversion,

just fill in 1.

Export folder: The path's name to save all these files.

c) Again, Tools → Macros → Run Macros. Select the library that contains the macro Module 1. Select

the macro Online and then klick Run (Figure A7). The Row Toplimit, Row Downlimit, Timecolunm,

Datacolumn, Filename and Export folder have the same meaning as the above explanation.

Process time beginning: The beginning of the real process time. Unit: s (second) e.g. the beginning

of the inoculation is 00:30:00, then you can fill in 1800.


76

Interval: The interval to reduce the online data. E.g., to extract a data point for every three data

points, you should just fill 3 in the blank space.

Figure A7: Data box for online measurements.

Additional Information:

a) The beginning of the row number and column number are 0.

b) The unit of time in offline data sheet is automatically transferred from hour to second; the unit of time

in online data sheet is transferred from hh:mm:ss to second.

c) Sheet1 is set for offline measurement, sheet2 is set for online measurement.

d) Sheet3 and Sheet4 are reserved to save the inputs of the dialogs.

References

77

References

Alba-Lois, L., & Segal-Kischinevzky, C. (2010). Yeast fermentation and the making of beer and wine. Retrieved from

http://www.nature.com/scitable/topicpage/yeast-fermentation-and-the-making-of-beer-14372813.

Albert, S., & Kinley, R. D. (2001). Multivariate statistical monitoring of batch processes: An industrial case study of

fermentation supervision. Trends in Biotechnology, 19(2), 53-62.

Ashoori, A., Moshiri, B., Khaki-Sedigh, A., & Bakhtiari, M. R. (2009). Optimal control of a nonlinear fed-batch

fermentation process using model predictive approach. Journal of Process Control, 19(7), 1162-1173.

B. Braun Biotech International GmbH.Biostat C betriebshandbuch. BADC - Rev. 2.4 - 0600 SW 4.1:148-150.

Barnett, J. A. (1998). A history of research on yeasts 1: Work by chemists and biologists 1789–1850. Yeast, 14(16),

1439-1451.

Barnett, J. A., & Lichtenthaler, F. W. (2001). A history of research on yeasts 3: Emil Fischer, Eduard Buchner and their

contemporaries, 1880–1900. Yeast, 18(4), 363-388.

Bellgardt, K. H., Kuhlmann, W., Meyer, H. D., Schügerl, K., & Thoma. M. (1986). Application of an extended Kalman

filter for state estimation of a yeast fermentation. IEE Proceedings D - Control Theory and Applications, 133(5), 226-

234.

Bastin, G., & van Impe, J. F. (1995). Optimal adaptive control of fed-batch fermentation processes. Control Eng Pract,

3(7), 939-954.

Berry, B. N., Dobrowsky, T. M., Timson, R. C., Kshirsagar, R., Ryll, T., & Wiltberger, K. (2016). Quick generation of

raman spectroscopy based in-process glucose control to influence biopharmaceutical protein product quality during

mammalian cell culture. Biotechnology Progress, 32(1), 224-234.

Bolf, N., & Jerbic, I. (2006). Application of artificial neural networks for process identification and control. Kemija u

Industriji, 55(11), 457-468.

Bošković, J. D. (1995). Stable adaptive control of a class of nonlinearly parametrized bioreact or processes. Proceedings

of the American Control Conference, 1995, 3, 1795-1799.

Bošković, J. D. (1996). Stable adaptive control of a class of continuous-flow bioreactors. AIChE Journal, 42(1), 176-186.

Branduardi, P., & Porro, D. (2012). Yeasts in biotechnology. Yeast (pp. 347-370) Wiley-VCH Verlag GmbH & Co.

KGaA.

Brenan, K., Campbell, S., & Petzold, L. (1995). Numerical solution of initial-value problems in differential-algebraic

equations. Society for Industrial and Applied Mathematics.

Carrasco, E. F., & Banga, J. R. (1997). Dynamic optimization of batch reactors using adaptive stochastic algorithms.

Industrial & Engineering Chemistry Research, 36(6), 2252-2261.

Carrillo-Ureta, G. E. (2003). Optimal control of fermentation processes. PhD Thesis, City University London.

Chikkula, Y., & Lee, J. H. (2000). Robust adaptive predictive control of nonlinear processes using input-output models.

Ind. Eng. Chem. Res., 39, 2010-2023.

Chopda, V. R., Gomes, J., & Rathore, A. S. (2016). Bridging the gap between PAT concepts and implementation: An

integrated software platform for fermentation. Biotechnology Journal, 11(1), 164-171.

Chu, W. B. Z., & Constantinides, A. (1988). Modeling, optimization, and computer control of Cephalosporin C

fermentation process. Biotechnol. Bioeng., 32(3), 277-288.

http://www.nature.com/scitable/topicpage/yeast-fermentation-and-the-making-of-beer-14372813

References

78

Cinar, A., Parulekar, S., Ündey, C., & Birol, G. (2003). Batch fermentation: Modelling, monitoring, and control. CRC

Press, New York.

Dahhou, B., Roux, G., & Chamilothoris, G. (1992). Modelling and adaptive predictive control of a continuous

fermentation process. Applied Mathematical Modelling, 16(10), 545-552.

Damerow, P. (2012). Sumerian beer: The origins of brewing technology in ancient Mesopotamia. Cuneiform Digital

Library Journal, 2, 1-20.

Dobell, C., & Leeuwenhoek, A. v. (1932). Antony van Leeuwenhoek and his little animals. Dover Publications Inc.

Eaton, J. W., & Rawlings, J. B. (1992). Model-predictive control of chemical processes. Chemical Engineering Science,

47(4), 705-720.

Fliess, M., Lévine, J., Martin, P., & Rouchon, P. (1995). Flatness and defect of nonlinear systems: Introductory theory

and examples. Int. J. Control, 61(6), 1327-1361.

Foss, B. A., Johansen, T. A., & Sørensen, A. V. (1995). Nonlinear predictive control using local models — applied to a

batch fermentation process. Control Engineering Practice, 3(3), 389-396.

Frahm, B., & Pörtner, R. (2002). Messung und Simulation des Kohlendioxidausstoßes von Zellkulturprozessen. Chemie

Ingenieur Technik, 74(5) 702.

Frahm, B., Blank, H., Cornand, P., Oelssner, W., Guth, U., Lane, P., Pörtner, R. (2002a). Determination of dissolved CO2

concentration and CO2 production rate of mammalian cell suspension culture based on off-gas measurement. Journal of

Biotechnology, 99(2), 133-148.

Frahm, B., Lane, P., Märkl, H., & Pörtner, R. (2003). Improvement of a mammalian cell culture process by adaptive,

model-based dialysis fed-batch cultivation and suppression of apoptosis. Bioprocess and Biosystems Engineering, 26(1),

1-10.

Frahm, B., Lane, P., Atzert, H., Munack, A., Hoffmann, M., Hass, V. C., & Pörtner, R. (2002b). Adaptive, model-based

control by the open-loop-feedback-optimal (OLFO) controller for the effective fed-batch cultivation of hybridoma cells.

Biotechnology Progress, 18(5), 1095-1103.

Gerlach, I., Brüning, S., Gustavsson, R., Mandenius, C., & Hass, V. C. (2014). Operator training in recombinant protein

production using a structured simulator model. Journal of Biotechnology, 177, 53-59.

Glassey, J., Gernaey, K. V., Clemens, C., Schulz, T. W., Oliveira, R., Striedner, G., & Mandenius, C. (2011). Process

analytical technology (PAT) for biopharmaceuticals. Biotechnology Journal, 6(4), 369-377.

Goldrick, S., Ştefan, A., Lovett, D., Montague, G., & Lennox, B. (2015). The development of an industrial-scale fed-

batch fermentation simulation. Journal of Biotechnology, 193, 70-82.

Graindorge, P., Charbonnier, S., Magnin, J. P., Mauvy, C., & Cheruy, A. (1994). A software sensor of biological activity

based on a redox probe for the control of thiobacillus ferrooxidans cultures. Journal of Biotechnology, 35(1), 87-96.

Gray, W. D. (1941). Studies on the alcohol tolerance of yeasts. Journal of Bacteriology, 42(5), 561-574.

Gustavsson, R., Lukasser, C., & Mandenius, C. (2015). Control of specific carbon dioxide production in a fed-batch

culture producing recombinant protein using a soft sensor. Journal of Biotechnology, 200, 44-51.

Hass, V. C. (2005). Verbesserung der bioverfahrenstechnischen Ausbildung durch einen „Virtuellen bioreaktor”. Chemie

Ingenieur Technik, 77(1-2), 161-167.

Hass, V. C., Lane, P., Hoffmann, M., Frahm, B., Schwabe, J. -., Pörtner, R., & Munack, A. (2002). Model-based control

of hybridoma cell cultures. Computer applications in biotechnology 2001 : Modelling, monitoring and control of

References

79

biotechnological processes ; a Proceedings volume from the 8th IFAC international conference, Québec City, Canada,

24 - 27 june 2001 (pp. 323-329) Published for the International Federation of Automatic Control by Pergamon.

Hass, V. C., & Pörtner, R. (2009). Praxis der Bioprozesstechnik mit virtuellem Praktikum. Springer Spektrum

Akademischer Verlag.

Havlik, I., Lindner, P., Scheper, T., & Reardon, K. F. (2013). On-line monitoring of large cultivations of microalgae and

cyanobacteria. Trends in Biotechnology, 31(7), 406-414.

Hornsey, I. S. (2003). A history of beer and brewing. The Royal Society of Chemistry.

Jadot, F., Bastin, G., & Van Impe, J. F. (1998). Optimal adaptive control of a bioprocess with yield–productivity conflict.

Journal of Biotechnology, 65(1), 61-68.

Kadlec, P., Gabrys, B., & Strandt, S. (2009). Data-driven soft sensors in the process industry. Computers & Chemical

Engineering, 33(4), 795-814.

Kaiser, C., Carvell, J. P., & Luttmann, R. (2007). A sensitive, compact, in situ biomass measurement system controlling

and monitoring microbial fermentations using radio-frequency impedance. BioProcess International, 5(Suppl. 1), 52-55.

Karra, S., Shaw, R., Patwardhan, S. C., & Noronha, S. (2008). Adaptive model predictive control of multivariable time-

varying systems. Industrial & Engineering Chemistry Research, 47(8), 2708-2720.

Lee, J., Lee, S. Y., Park, S., & Middelberg, A. P. J. (1999). Control of fed-batch fermentations. Biotechnology Advances,

17(1), 29-48.

Li, M., Kuhnen, F., Pörtner, R., & Hass, V. C. (2012). Modellgestützte Optimalsteuerung von biotechnischen

Kultivierungen – Möglichkeiten und Grenzen. Chemie Ingenieur Technik, 84(8) 1339-1340.

Luttmann, R., Bracewell, D. G., Cornelissen, G., Gernaey, K. V., Glassey, J., Hass, V. C., Keiser, C., Preusse, C.,

Steiedner G., Mandenius, C. F. (2012). Soft sensors in bioprocessing: A status report and recommendations.

Biotechnology Journal, 7(8), 1040-1048.

Luttmann, R., Munack, A., & Thoma, M. (1985). Mathematical modelling, parameter identification and adaptive control

of single cell protein processes in tower loop bioreactors. Agricultural feedstock and waste treatment and engineering

(pp. 95-205). Springer Berlin Heidelberg.

Mahadevan, R., Agrawal, S. K., & III, F. J. D. (2001). Differential flatness based nonlinear predictive control of fed-

batch bioreactors. Control Engineering Practice, 9(8), 889-899.

Mandenius, C., & Gustavsson, R. (2015). Mini-review: Soft sensors as means for PAT in the manufacture of bio-

therapeutics. Journal of Chemical Technology & Biotechnology, 90(2), 215-227.

McGovern, P. E. (2009). Uncorking the past: The quest for wine, beer, and other alcoholic beverages. University of

California Press.

Modak, J. M., & Lim, H. C. (1987). Feedback optimization of fed-batch fermentation. Biotechnology and

Bioengineering, 30(4), 528-540.

Modak, J. M., & Lim, H. C. (1992). Optimal mode of operation of bioreactor for fermentation processes. Chemical

Engineering Science, 47(15), 3869-3884.

Munack, A. (1986). Application of receding horizon adaptive control to an underfloor heating system. IFAC Conference

on Simulation of Control Systems. 263-268.

Nelder, J. A., & Mead, R. (1965). A simplex method for function minimization. The Computer Journal, 7(4), 308-313.

Nelson, D. L., Lehninger, A. L., & Cox, M. M. (2008). Lehninger principles of biochemistry. New York: W.H. Freeman.

References

80

Ödman, P., Johansen, C. L., Olsson, L., Gernaey, K. V., & Lantz, A. E. (2009). On-line estimation of biomass, glucose

and ethanol in Saccharomyces cerevisiae cultivations using in-situ multi-wavelength fluorescence and software sensors.

Journal of Biotechnology, 144(2), 102-112.

Ohadi, K., Legge, R. L., & Budman, H. M. (2015). Development of a soft-sensor based on multi-wavelength

fluorescence spectroscopy and a dynamic metabolic model for monitoring mammalian cell cultures. Biotechnology and


Olsson, L., Schulze, U., & Nielsen, J. (1998). On-line bioprocess monitoring – an academic discipline or an industrial

tool? TrAC Trends in Analytical Chemistry, 17(2), 88-95.

Pan, T., Li, S., & Cai, W. (2007). Lazy learning-based online identification and adaptive PID control: A case study for

CSTR process. Industrial & Engineering Chemistry Research, 46(2), 472-480.

Park, S., & Ramirez, F.,W. (1988). Optimal production of secreted protein in fed-batch reactors. AIChE Journal, 34(9),

1550-1558.

Pasteur, L. (1857). Mémoire sur la fermentation alcoolique. Comptes Rendus De l'Académie Des Sciences, 45, 1032-

1036.

Peuker, T., Riedel, M., Kaiser, C., Ellert, A., Lenz, K., Elsholz, O., & Luttmann, R. (2004). At-line determination of

glucose, ammonia, and acetate in high cell density cultivations of Escherichia coli. Engineering in Life Sciences, 4(2),

138-143.

Ponnuswamy, S. R., Shah, S. L., & Kiparissides, C. A. (1987). Computer optimal control of batch polymerization

reactors. Industrial & Engineering Chemistry Research, 26(11), 2229-2236.

Pörtner, R., Bohmann, A., Lüdemann, I., & Märkl, H. (1994). Animal cell culture estimation of specific glucose uptake

rates in cultures of hybridoma cells. Journal of Biotechnology, 34(3), 237-246.

Posten, C., & Munack, A. (1990). Improved modeling of plant cell suspension cultures by optimum experiment design.

Preprints of the 11th IFAC world congress 1990 (pp. 268-273)

Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. (2007). Numerical recipes 3rd edition: The art of

scientific computing. New York, NY, USA: Cambridge University Press.

Redford, D. B. (2001). The Oxford encyclopedia of ancient Egypt Oxford University Press, Oxford Univ.

Rodrigues, J. A. D., Zaiat, M., & Filho, R. M. (1999). State estimation and parameter identification in a fed-batch

penicillin production process. Brazilian Journal of Chemical Engineering, 16(1), 41-52.

Schwarz, S. (2009). Modellgestützte prozessoptimierung von escherichia coli kultivierungen. Master thesis, University of

Applied Sciences Bremen.

Schwarz, S., Kuhnen, F., & Hass, V. C. (2010). Development of a general process model for optimisation of

biotechnological cultivations - a case study with Escherichia coli and Saccharomyces cerevisiae. CHISA 2010 & ECCE 7

(http://www.chisa.cz/).

Shen, D., Chen, G., Cruz, J., Jose B., Pham, K., Blasch, E., & Lynch, R. (2010). Cooperative controls with intermittent

communication. Proc. SPIE. 7691, Space Missions and Technologies, 76910D-76910D-11.

Shimizu, K. (1993). An overview on the control system design of bioreactors. Measurement and control (pp. 65-84).

Springer Berlin Heidelberg.

Snyman, J. (2005). An introduction to basic optimization theory and classical and new gradient-based algorithms.

Springer US.

References

81

Soroush, M., & Valluri, S. (1994). An approach to optimization and control of batch processes. American Control

Conference,1, 490-494.

Srinivasa, B. B., & Moreshwar, J. M. (2009). Method for on-line optimization of a fed-batch fermentation unit to

maximize the product yield. Patent US20090117647 A1.

Stanke, M., & Hitzmann, B. (2013). Automatic control of bioprocesses. Advances in Biochemical

Engineering/Biotechnology, 132, 35-63.

Stephanopoulos, G., & San, K. (1984). Studies on online bioreactor identification. I. Theory. Biotechnology and


Teixeira, A. P., Carinhas, N., Dias, J. M. L., Cruz, P., Alves, P. M., Carrondo, M. J. T., & Oliveira, R. (2007). Hybrid

semi-parametric mathematical systems: Bridging the gap between systems biology and process engineering. Journal of

Biotechnology, 132(4), 418-425.

Tholudur, A., & Ramirez, W. F. (1996). Optimization of fed-batch bioreactors using neural network parameter function

models. Biotechnology Progress, 12(3), 302-309.

Tremblay, M. d., Perrier, M., Chavarie, C., & Archambault, J. (1992). Fed-batch culture of hybridoma cells: Comparison

of optimal control approach and closed loop strategies. Bioprocess Engineering, 7(5), 229-234.

U.S. Department of Health and Human Services Food and Drug Administration.Guidance for industry: Process

validation: General principles and practices. Retrieved from

http://www.fda.gov/downloads/Drugs/.../Guidances/UCM070336.pdf

Van Breusegem, V., & Bastin, G. (1990). Optimal control of biomass growth in a mixed culture. Biotechnology and


Van Impe, J. F., & Bastin, G. (1998). Optimal adaptive control of fed-batch fermentation processes. In Van Impe, Jan F.

M., P. A. Vanrolleghem & D. M. Iserentant (Eds.), Advanced instrumentation, data interpretation, and control of

biotechnological processes (pp. 401-435). Dordrecht, The Netherlands: Springer Netherlands.

Warth, B., Rajkai, G., & Mandenius, C. (2010). Evaluation of software sensors for on-line estimation of culture

conditions in an Escherichia coli cultivation expressing a recombinant protein. Journal of Biotechnology, 147(1), 37-45.

Wechselberger, P., Sagmeister, P., & Herwig, C. (2013). Real-time estimation of biomass and specific growth rate in

physiologically variable recombinant fed-batch processes. Bioprocess and Biosystems Engineering, 36(9), 1205-1218.

Witte, V. C. (1996). Mathematische modellierung und adaptive prozesssteuerung der kultivierung von cyathus striatus.

PhD thesis, Technische Universität Hamburg-Harburg.

Zeng, F. Y., & Dahhou, B. (1993). Adaptive control of a nonlinear fermentation process via MRAC technique. Appl.

Math. Modelling, 17, 58-69.

Zhang, H., & Lennox, B. (2004). Integrated condition monitoring and control of fed-batch fermentation processes.

Journal of Process Control, 14(1), 41-50.

http://www.fda.gov/downloads/Drugs/.../Guidances/UCM070336.pdf

Feedback-Optimal Controller for Cultivation Processes

Documents