AN APPROACH TO MECHANISM RECOGNITION

FOR MODEL BASED ANALYSIS
OF BIOLOGICAL SYSTEMS
AN APPROACH TO MECHANISM RECOGNITION
FOR MODEL BASED ANALYSIS OF BIOLOGICAL SYSTEMS
vorgelegt von
Master of Science in Process Engineering
Mariano Nicolás Cruz Bournazou
aus Mexiko City, Mexiko


von der Fakultät III- Prozesswissenschaften
der Technischen Universität Berlin
zur Erlangung des akademischen Grades
Doktor der Ingenieurwissenschaften
- Dr.-Ing. -

genehmigte Dissertation




Promotionsausschuss:
Vorsitzender: Prof. Dr.-Ing. G. Tsatsaronis
Gutachter: Prof. Dr.-Ing. G. Wozny
Gutachter: Prof. Dr. P. Neubauer
Gutachter: Prof. G. Lyberatos

Tag der wissenschaftlichen Aussprache 24.01.2012


Berlin 2012
D 83































dedicada a Heberto y Helig











ACKNOWLEDGEMENTS
I want to express my gratitude to my supervisor, Professor Günter Wozny, for his
constant support and useful advice on a professional and personal level and to my co-
supervisor, Professor Peter Neubauer, for finding a perfect application for MR and for
his intellectual input to this work.

I would also like to thank Professor Kravaris and Professor Lyberatos at the University
of Patras for his collaboration and hospitality.

Special thanks go to Dr. Harvey Arellano-Garcia, Dr. Stefan Junne and Dr. Tilman Barz
for interesting discussions and support during the critical phases of this project (which
were quite numerous).

I must of course thank all my other friends and colleagues in the Chair of Process
Dynamics and Operations and in the Chair of Bioprocesses.

I would also like to thank my second family, conformed of all my friends spread around
the world, who have always motivated me to follow my goals and offered a shoulder to
console my sorrows.

Finally, I would like to thank my parents, Mariano and Effi, and my family for always
being at my side despite the distance and especially to Alexis Cruz, who one day might
realize his great contribution to each one of the achievements in my life.


M. Nicolás Cruz B.















Ich Mariano Nicolas Cruz Bournazou erkläre an Eides Statt, dass die vorliegende
Dissertation in allen Teilen von mir selbständig angefertigt wurde und die benutzten
Hilfsmittel vollständig angegeben worden sind.



Mariano Nicolas Cruz Bournazou Berlin, 1. Februar 2012



Content
i

CONTENT
Zusammenfassung ....................................................................................................................... v
Abstract ....................................................................................................................................... vii
Figure content ............................................................................................................................. ix
Table content............................................................................................................................... xi
List of Abbreviations ................................................................................................................ xii
List of symbols ......................................................................................................................... xvii
1 Introduction ......................................................................................................................... 1
1.1 The gap between research and industry ................................................................. 1
1.2 Hierarchical modeling ............................................................................................... 3
1.3 Understanding process dynamics ............................................................................ 4
1.4 The bridge between industry and research ............................................................ 5
1.5 Related work............................................................................................................... 7
1.6 Project Goal ............................................................................................................... 9
1.7 Advantages of Mechanism Recognition............................................................... 10
1.8 The good, the bad, and the useful model ............................................................ 11
2 Modeling ............................................................................................................................. 13
2.1 Definition ................................................................................................................. 13
2.2 Model complexity .................................................................................................... 14
2.3 Engineering approach to complex systems ......................................................... 15
2.4 Modeling in systems biology .................................................................................. 16
2.4.1 Systems biology ................................................................................................... 16
2.4.2 Modeling of genetic regulatory systems ........................................................... 17
2.5 Mathematical model for a batch biochemical reactor ........................................ 19
3 Model Reduction ............................................................................................................... 21
3.1 Introduction ............................................................................................................. 21
3.2 Basic approaches to Model Reduction ................................................................. 22
3.2.1 Reaction invariants.............................................................................................. 22
3.2.2 Switching functions and the reaction invariant .............................................. 24
3.2.3 Sensitivity analysis ............................................................................................... 25
Content
ii

3.2.4 Lumping................................................................................................................ 26
3.2.5 Perturbation theory ............................................................................................. 27
3.2.6 Time scale analysis .............................................................................................. 28
4 Optimal Experimental Design ......................................................................................... 31
4.1 The experiment ........................................................................................................ 33
4.1.1 The Maximum Likelihood ................................................................................. 34
4.1.2 Model identifiability ............................................................................................ 35
4.2 The Fisher Information Matrix.............................................................................. 37
4.2.1 The confidence Interval ..................................................................................... 37
4.2.2 Approximation of parameter variance-covariance matrix ............................. 39
4.2.3 Limitations of the Fisher Information Matrix ................................................ 40
4.3 Model discrimination............................................................................................... 42
4.3.1 Model discrimination in Mechanism Recognition .......................................... 44
5 Code generation, simulation and optimization ............................................................. 47
5.1 Code generation ....................................................................................................... 47
5.1.1 MOSAIC .............................................................................................................. 47
5.1.2 SBPD..................................................................................................................... 48
5.2 Simulation ................................................................................................................. 49
5.2.1 sDACL .................................................................................................................. 49
5.3 Optimization............................................................................................................. 50
6 An approach to Mechanism Recognition ...................................................................... 51
6.1 A short introduction to Mechanism Recognition ............................................... 51
6.1.1 Illustrative Example ............................................................................................ 53
6.2 Methodology for Mechanism Recognition .......................................................... 56
6.3 Program steps ........................................................................................................... 57
6.3.1 Submodels ............................................................................................................ 57
6.3.2 General structure ................................................................................................. 57
6.3.3 Submodel distinguishability ............................................................................... 58
6.3.4 Initial interval ....................................................................................................... 59
6.3.5 MR initialization .................................................................................................. 59
6.3.6 Detection of switching points ........................................................................... 60
6.3.7 Initial conditions of the interval k+1 ............................................................... 62
6.3.8 Detection of the next switching point ............................................................. 62
Content
iii

6.3.9 Flow diagram ....................................................................................................... 63
7 Mechanism Recognition applied on Sequencing Batch Reactors .............................. 65
7.1 Introduction ............................................................................................................. 65
7.1.1 Activated Sludge .................................................................................................. 65
7.1.2 Sequencing Batch Reactor ................................................................................. 66
7.1.3 Nitrate Bypass Generation ................................................................................ 67
7.1.4 Monitoring of wastewater processes ................................................................ 68
7.2 Submodel building ................................................................................................... 68
7.3 A proposed 9state model ....................................................................................... 69
7.3.1 Storage .................................................................................................................. 69
7.3.2 Reduction of the extended ASM3 model to a 9state model ......................... 70
7.3.3 Mathematical representation of the 9state model .......................................... 71
7.3.4 Stoichiometric matrix ......................................................................................... 73
7.3.5 Limitations of the reduced models ................................................................... 74
7.4 A proposed 6state model ....................................................................................... 74
7.5 A proposed 5state model ....................................................................................... 75
7.6 Results ....................................................................................................................... 75
7.6.1 Simulations Results ............................................................................................. 77
7.7 Mechanism Recognition in SBR processes .......................................................... 78
7.8 Recognition of organic matter depletion ............................................................. 79
7.8.1 Conditions for proper process description with Mechanism Recognition 79
7.8.2 Conditions for accurate switching point detection ........................................ 80
7.8.3 MR initialization .................................................................................................. 82
7.8.4 Detection of switching points ........................................................................... 82
7.9 Conclusions .............................................................................................................. 83
8 Mechanism Recognition in Escherichia coli cultivations ................................................. 85
8.1 Escherichia coli cultivations ....................................................................................... 85
8.2 Models for the description of Escherichia coli cultivations .................................. 86
8.2.1 Division of physiological states ........................................................................ 87
8.3 Modeling Escherichia coli batch fermentations with Mechanism Recognition . 90
8.3.1 General model ..................................................................................................... 90
8.3.2 Submodels for dividing metabolic states ......................................................... 92
8.4 Material and methods ............................................................................................. 96
Content
iv

8.4.1 Strain and culture conditions ............................................................................. 96
8.4.2 Online analysis ..................................................................................................... 97
8.4.3 Offline analysis .................................................................................................... 99
8.4.4 Data treatment ................................................................................................... 103
8.5 Experimental validation ........................................................................................ 104
8.5.1 Conditions for proper process description with MR ................................... 104
8.5.2 Conditions for accurate switching point detection ...................................... 105
8.5.3 Data set ............................................................................................................... 106
8.5.4 Recognition of overflow and substrate limitation regimes ......................... 109
8.5.5 Simulations vs. experimental data ................................................................... 110
8.5.6 Results ................................................................................................................. 111
8.6 Conclusions ............................................................................................................ 113
8.7 Future work ............................................................................................................ 114
9 Conclusions and outlook ................................................................................................ 115
9.1 Conclusions ............................................................................................................ 115
9.2 Outlook ................................................................................................................... 116
9.2.1 General theory for submodel generation ....................................................... 116
9.2.2 Switching point identification .......................................................................... 117
9.2.3 Global optimization .......................................................................................... 117
9.2.4 Online monitoring............................................................................................. 118
10 Appendix .......................................................................................................................... 119


An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


v

ZUSAMMENFASSUNG
Ziel dieser Arbeit ist die Entwicklung innovativer Ansätze zur Beschreibung komplexer
Prozesse mit Hilfe von reduzierten Modellen. Die resultierenden Beschränkungen für
die Vorhersage des Prozessverhaltens auf Basis von reduzierten Modellen werden
durch den Einsatz von Methoden zur Mechanismenerkennung genutzt, um Indikatoren
für relevante Änderungen im Prozessgeschehen zu erzeugen.
Empirische Kenntnisse, Analogien zu anderen Modellen aus der Literatur, Methoden
zur Bewertung des Zustand eines Systems und Ansätze zur Modellreduktion werden
kombiniert, in einem Versuch ein Set exakter Teilmodelle mit einer großen Robustheit
und Identifizierbarkeit zu generieren. Der Ansatz zur Mechanismenerkennung ist ein
Werkzeug zur effizienten Nutzung von Kenntnissen aus der Grundlagenforschung und
der Modellierung und ermöglicht ein tieferes Verständnis für den gesamten Prozess.
Biologische Prozesse stellen ein wichtiges Anwendungsgebiet für die
Mechanismenerkennung dar. Im Rahmen dieser Arbeit werden zwei Fallstudien
vorgestellt, für die sowohl die Anwendbarkeit als auch die Vorteile dieser Methode
nachgewiesen werden. Es wird gezeigt, dass die systematische Analyse des Prozesses
und seiner gemessenen sowie auf Basis von Modellen vorausberechneten Zustände, die
Beschreibung und Überwachung des Prozesses mit einer höheren Effizienz erlaubt.
Die erste Fallstudie beschreibt die Überwachung des Belebtschlammverfahrens in
Sequencing Batch Reaktoren. Dazu wird das dem aktuellen Forschungsstand
entsprechende Modell (ASM3 erweitert für die zweistufige Nitrifikation und
Denitrifikation) auf ein einfaches Teilmodell reduziert. Das resultierende Modell ist
effizient anzuwenden, liefert eine exakte Beschreibung des Prozesses in einem
wohldefinierten Bereich und erlaubt die Erkennung des Abbaus organischer Stoffe.
Die zweite Fallstudie ist die Kultivierung von Escherichia coli im Batch-Prozess. Ein
erfolgreich validiertes Modell wird analysiert und reduziert. Die Methodik der
Mechanismenerkennung ermöglicht die Erzeugung von drei Teilmodellen, die in der
Lage sind, Batch-Kultivierungen mit einfachen ODE-Systemen zu beschreiben.
Abschließend wird die Fähigkeit der Mechanismen Erkennung als
Unterstützungswerkzeug für die Zusammenarbeit zwischen Grundlagenforschung und
Industrie analysiert.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


vii

ABSTRACT
This work aims at finding new manners to accurately describe complex processes based
on simple models. Furthermore, the approach to Mechanism Recognition proposes to
exploit the description limitations of these submodels and to use them as indicators of
non-measurable variables.
Empirical knowledge, analogies to other models from literature, methods to analyze the
state of information of the system and model reduction techniques are brought together
in an effort to create an adequate set of accurate models with a significantly larger
tractability. It is worth stressing the approach to Mechanism Recognition does not
intend to substitute human reasoning or make up for lack of process knowledge. On the
contrary, this method is merely a tool to efficiently apply the knowledge obtained from
basic research to gain a better insight of the industrial process.
The approach to Mechanism Recognition finds an important field of application in
biological processes. In this work two case studies are presented to manifest the
advantages and applicability of this method. It is shown how the correct analysis of the
process, the state of information, and the models applied to describe the process results
in new methods to describe and monitor the process with higher efficiency.
The first case study presented is the monitoring of the Active Sludge Process in
Sequencing Batch Reactors. For this, the state of the art model ASM3 extended for two
step nitrification-denitrification is reduced to create a simple model which can easily
describe the process in a defined range and detect depletion of carbonate matter.
The second case study is Escherichia coli batch and fed-batch cultivations. A model
obtained from literature is analyzed and reduced. The methodology of Mechanism
Recognition allows creating a set of three submodels able to describe batch cultivations
with simple systems of Ordinary Differential Equations. Furthermore, the restrictions of
the complex model are set under scrutiny to understand its dynamics and limitations.
Finally, special attention is paid to the capability of Mechanism Recognition as a tool to
enhance collaboration between basic research and industry.

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


ix

FIGURE CONTENT
Figure 1.1: Hierarchical modeling scheme. .............................................................................. 3
Figure 2.1 : E. coli transcriptional regulatory network. [53]. ................................................ 14
Figure 2.2: Incremental approach for reaction kinetics identification [58] ....................... 16
Figure 2.3: Hypothesis-driven research in systems biology [59]. ........................................ 17
Figure 3.1. Behavior of a switching function in dependence of the limiting species. ..... 25
Figure 3.2: Three-component monomolecular reaction system, the numbers on the
arrows represent the back- and forward reaction constants. .............................................. 26
Figure 3.3: Lumping a monomolecular three-component reaction into a two-component
reaction ........................................................................................................................................ 27
Figure 3.4: Phase diagram of full order model (3.12). Comparison with reduced models
in a chemostat process .............................................................................................................. 30
Figure 4.1: Effect of sensitivities in parameter estimation accuracy. σ
P
and σ
y
represent
standard deviation of parameters and measurements respectively. .................................... 36
Figure 4.2: Confidence interval from the Lin model, obtained with Montecarlo
simulation. ................................................................................................................................... 39
Figure 4.3: Criteria for optimization [92] ............................................................................... 40
Figure 4.4: Shape of the confidence interval for different variance values from the Lin
model (appendix A). The confidence interval can be approximated by an ellipse near the
exact value. .................................................................................................................................. 41
Figure 4.5: Objective function of a nonlinear model (appendix A) with respect to
changes in a two dimensional parameter set. ........................................................................ 41
Figure 5.1: High level modeling with MOSAIC [46] ............................................................ 48
Figure 5.2: Modular structure of the toolbox. The toolbox is designed in a modular .... 49
Figure 6.1: Model fit a) without setting bounds b) with setting bounds for physical
parameters. [119] ........................................................................................................................ 54
Figure 6.2: Comparison experiment/simulation using a) just one model. B) various
models [119] ............................................................................................................................... 55
Figure 6.3 Cleaning strategy based on MR [43] ..................................................................... 55
Figure 6.4: Flow diagram of MR algorithm ........................................................................... 64
Figure 7.1: SBR cycle [136] ....................................................................................................... 66
Figure 7.2. Nitrification-denitrification process described as a two -step reaction. ......... 67
Figure 7.3. Substrate concentration S
S
and stored energy Sto against time. ...................... 76
Figure 7.4. Biomass against time. Changes in the biomass are very small (less than 10%).
...................................................................................................................................................... 76
Figure 7.5. NO
X
concentration against time. ......................................................................... 77
Figure 7.6. a) Oxygen concentration in the medium against time. ..................................... 77
Figure 7.7: Description of the 5state model in both regimes, with and without substrate.
...................................................................................................................................................... 78
Figure content


x

Figure 7.8: Minimal length for initialization of MR .............................................................. 82
Figure 7.9. Detection of the regime switching point. ........................................................... 83
Figure 8.1: Integration of the kinetic model proposed by Lin [91] .................................... 91
Figure 8.2: Complex model (Lin et al.) fitted to experimental batch cultivation data. .... 91
Figure 8.3: Comparison between the complex model (dots) vs. the overflow submodel
(lines) initializing in four different intervals. .......................................................................... 93
Figure 8.4: Comparison between the complex model (dots) vs. the substrate limiting
submodel (lines) initializing in four different intervals. ........................................................ 94
Figure 8.5: Comparison between the complex model (dots) vs. the cell starvation
submodel (lines) initializing in four different intervals. ........................................................ 95
Figure 8.6: Bioreactor KL2000 at E. coli batch cultivation [203] ........................................ 97
Figure 8.7: EloCheck
®
............................................................................................................... 99
Figure 8.8. Calibration curve for glucose determination .................................................... 100
Figure 8.9. Calibration curve of acetate ................................................................................ 101
Figure 8.10: Mechanism of the reactions involved in the assay ........................................ 102
Figure 8.11: Experimental results batch experiment G1. Part I: Dry biomass and glucose
concentrations .......................................................................................................................... 107
Figure 8.12: Experimental results batch experiment G1. Part II: Specific concentration
of acetic acid ............................................................................................................................. 108
Figure 8.13: Experimental results batch experiment G1. Part III: Outgas concentrations
.................................................................................................................................................... 108
Figure 8.14: Experimental results batch experiment G1. Part IV: Metabolite
concentration ............................................................................................................................ 109
Figure 8.15: OverFlow submodel fitted against experimental data. ................................. 110
Figure 8.16: Submodel for the description of growth under substrate limitation fitted
against experimental data. ....................................................................................................... 111
Figure 8.17: Starvation condition described by the corresponding submodel fitted
against experimental data. ....................................................................................................... 111
Figure 8.18: Experimental validation of the MR approach. .............................................. 112
Figure 8.19: Identifiability test considering white noise, standard deviation of 5% in all
measurements ........................................................................................................................... 113


An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


xi

TABLE CONTENT
Table 4.1: Criteria for confidence interval quantification [92]. ........................................... 40
Table 4.2: Types of sum of square [22] .................................................................................. 43
Table 7.1: Reaction rates of the extended ASM3 .................................................................. 70
Table 7.2: 9state model constants and its values as shown in the Matlab code ............... 73
Table 7.3: Stoichiometric matrix of the 9state model .......................................................... 73
Table 7.4. Comparison of the computation time. ................................................................. 77
Table 7.5. Singular function evaluations speed ..................................................................... 78
Table 8.1: Parameters considered for the model fit ............................................................. 95
Table 8.2. Composition of solution A .................................................................................. 102


List of Abbreviations


xii

LIST OF ABBREVIATIONS
Acs Acetyl-CoA synthase
ADHII Alcohol Dehydrogenase
AMP Adenosine monophosphate
AOB Ammonium Oxidizing Bacteria
ASM Activated Sludge Model
ASP Active Sludge Process
BOD Biological Oxygen Demand
Bpox Pyruvate oxidase
CAB Computer Aided Biology
CAPE Computer Aided Process Engineering
CFD Computational Fluid Dynamics
COD Chemical Oxygen Demand
CRB Cramer-Rao Bound
DAE Differential Algebraic Equation
DFG German Research Foundation
DNA Deoxyribonucleic acid
DOT Dissolved Oxygen Tension
EDTA Ethylenediaminetetraacetic acid
EMA European Medicines Agency
FDA Food and Drug Administration
FIM Fisher Information Matrix
GRN Gene Regulatory Network
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


xiii

HET Heteroptrophic organisms
HPLC High-Performance Liquid Chromatography
IA Incremental Approach
IMM Interactive Multiple Model
KDD Knowledge Discovery of Data
LSQ Least Squares
MBR Membrane Bioreactor
MBDoE Model Based Design of Experiments
MD Model Discrimination
MR Mechanism Recognition
mRNA Messenger Ribonucleic Acid
MTT Thiazolyl Blue
MWF Multi-Wavelength Fluorescence
MXL Maximum Likelihood
NAD
+
Nicotinamide adenine dinucleotide (NadH)
NB Nitrobacter
NBND Nitrate Bypass Nitrification-Denitrification
NDF Numerical Differentiation Formula
NS Nitrosomona
NH
+
4
Ammonia
NIRS Near-Infrared Spectroscopy
NO
-
2
Nitrite
NO
-
3
Nitrate
NOB Nitrite Oxidizing Bacteria
NSF Numerical Differentiation Formula
List of Abbreviations


xiv

OC Orthogonal Collocation
OCFE Orthogonal Collocation on Finite Elements
ODE Ordinary Differential Equations
OED Optimal Experimental Design
OF OverFlow Metabolism Model
PAT Process Analytical Technology
PCA Principal Component Analysis
PCP Process Constant Parameter
PDE Partial Differential Equation
PES Phenazine Ethosulfate
PLS Partial Least Squares
ppG Phosphoenol Pyruvate Glyoxylate
ppGpp Guanosine tetraphosphate
PSO Particle Swarm Optimization
PSSH Pseudo Steady State Hypothesis
PTS phosphotranspherase
QSSA Quasi Steady State Assumption
RWP Regime-Wise constant Parameter
SBML Systems Biology Markup Language
SBR Sequencing Batch Reactors
SF Switching Function
SL Substrate Limitation Model
SQP Sequential Quadratic Problem
ST Starvation model
TCA Tricyclic Acid Cycle
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


xv

WWTP Waste Water Treatment Plants

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


xvii

LIST OF SYMBOLS
VARIABLES
area [
2
]
acetate

A-criterion [ ]
linearly independent row vector [ ]
specific cake resistance

1

2

covariance matrix [ ]
C concentration

,

,

3

substrate consumption coefficient [ ]
carbon dioxide transfer rate

D dilution rate [ ]

Identifiability threshold [ ]
∆ pressure difference [ ]
E enzyme [ ]
Distinguishability threshold [ ]
F feed rate

,

f Function [ ]
Φ objective function [ ]
gravity acceleration

2

List of symbols


xviii

Γ initial velocity of the projectile
m
s

H hypothesis [ ]
stochastic error [% ]

systematic error [% ]
K limiting constant

k monomolecular rate matrix [ ]
friction constant

membrane thickness []
mass []
µ growth rate
−1
, [
−1
]
dynamic viscosity

2

oxygen transfer rate

P product [ ]
℘ probability distribution function [ ]
Q uptake

q specific uptake



R resistance [
−1
]
r reaction rate [ ]
residual [ ]

concentrations of soluble species (substrates
and products)/Substrate
[ ]
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


xix

correction constant [ ]
blocked area per unit filtrate volume

2

3

standard deviation [ ]
time , , , []

time span , , , []
parameter vector [ ]
input variables vector [ ]
Volume []
weighting matrix [ ]
constant input variables vector [ ]

Culture medium weight [ ]
concentrations of the particulate compounds [ ]
state variables vector [ ]
yield coefficient

Y stoichiometric coefficient [ ]
measurement values vector [ ]
z reaction invariant [ ]







List of symbols


xx

SUBSCRIPTS AND SUPERSCRIPTS
0 initial value

incoming
aer
aeration phase
anox
anoxic phase
Bio biomass
C cake

calculated value

capacity
E experimental

estimated
G general structure
Gluc glucose
H
heterotrophous
L lower
M membrane

Maximal

Measured value

nominal

Oxygen
S Substrate


An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


1

1 INTRODUCTION
1.1 THE GAP BETWEEN RESEARCH AND INDUSTRY
Globalization has changed market conditions drastically. Advances in transport and
communication bring companies together in worldwide competition. Cutting edge
technology is now essential for chemical and biochemical companies to survive. To
achieve this, substantial efforts have to be invested in research and development, not
only for direct applications, but also as long term investments to earn basic knowledge.
Industry is forced to make such investments to strive for its success in the world
markets, setting new standards in product performance. In the year 2010, BASF
invested almost 1.5 billion Euros in research and development [1].
Governments also need to make important investments on research, promoting mostly
basic research, which is not attractive to industry because it represents a long term
investment. The German Research Foundation (DFG) invested in the same year 2010
approximately 2.3 billion Euros [2], including support to universities, long term projects,
and specific research fields.
In spite of the parallel effort of both parties aiming at a common goal, collaboration
projects between academia and industry confront many complications. While industry
demands mostly fast solution to real process problems, academia is more interested in
long term projects offering novel knowledge. It can be said that industry is in search of
smart solutions while academy is looking for interesting problems. Finding novel
methods to bring industry and the research community together is essential for their
efficient development. Basic research offers a strong platform for development of
industrial applications, and industry provides not only economic support but also new
challenges and interesting applications.
Process modeling in chemistry and biotechnology offers a handful of examples of the
advantages of joint work. The development of a complex model, including estimation
and validation, may take several years. In addition, model identifiability or observability,
and application range cannot be assured beforehand. A company cannot afford to make
such long term and uncertain investments. These models have to be developed in basic
research. Still, accurate models allow optimal design and operation of plants, reducing
energy consumption, hazard, and environmental impact, while allowing better
monitoring and control [3]. Today, many of the models and software tools developed in
universities and research institutes are used in industry (Aspen®, Gproms®, Matlab®).
Introduction


2

In return, industry offers, in addition to economical support, the required facilities for
parameter estimation and model validation. The data collected daily in chemical plants
provides valuable information to researchers. Additionally, information about large scale
processes and long term performance can only be obtained from real plants.
Development of new tools that facilitate the communication and interaction between
industry and basic research lead to more efficient collaboration and better individual
performance. Instruments to benefit from the advances achieved in basic research by
allowing an adequate information transfer between both parties are crucial for an
efficient development of modern process technology. Modeling is not an exception.
New methods need to be created to bring complex models closer to industry and also to
create ways to use the information earned in industry for basic research purposes.
As maximization of process efficiency becomes essential to remaining competitive in
the market, complex models, which enable profit increase while fulfilling environmental
and safety regulations, are gaining application in industry. Process complexity and safety
restrictions have driven design and control to demand accurate and robust models.
Current black-box models and heuristic rules cannot provide the information required
in modern engineering. Regulations are changing, demanding model-based knowledge
of the process. The new regulations of the Process Analytical Technology (PAT)
initiative of the Food and Drug Administration (FDA) and the European Medicines
Agency (EMA) show the importance that modeling applied to process monitoring and
control is gaining in the pharmaceutical and generally in the biotechnological industry.
Due to the difficult measurements required, the application of model based control and
monitoring is essential.
The FDA makes the following statement in its Guidance for Industry, January 2011 [4]:
“A successful validation program depends upon information and knowledge from product and process
development. This knowledge and understanding is the basis for establishing an approach to control of
the manufacturing process that results in products with the desired quality attributes. Manufacturers
should:
- Understand the sources of variation
- Detect the presence and degree of variation
- Understand the impact of variation on the process and ultimately on product
attributes
- Control the variation in a manner commensurate with the risk it represents to the
process and product “
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


3

1.2 HIERARCHICAL MODELING
The contradiction between models in research and industry can also be seen from the
point of view of hierarchical modeling. Figure 1.1 depicts the typical layer representation
of a chemical process. These three different layers have a diverse level of significance
for industry and research, whereas industry is more interested in plant wide behavior
aiming at robust and secure process operation, basic research is more interested in the
lower layer where the study of microscalar phenomena takes place.

Figure 1.1: Hierarchical modeling scheme.
Particularly in biological systems, a gap can be seen between industry and basic research
[5]. Biological systems are extremely complex and very difficult to predict. Depending
on the level of system understanding, cells can be described with a simple Boolean
equation, from a kinetic down to a genomic level. In addition, regulations in food and
pharmaceutical industry are extremely strict. For this reason, industry is only interested
in practical, simple, and robust models. On the other hand, the main goal for building a
model in research is to gain process information [6]. This second category of models is
commonly too complex and requires advanced, expensive and time demanding
Layer 3
Process Systems
Engineering (PSE)
Design of the
complete process
E-7
V-2
V-3
P-2
P-3
E-9
P-5
P-6
E-10
P-8
P-9
P-10
V-5
P-11
P-12
P-13
E-13
V-6 P-14
P-15
P-16
Layer 2
Unit design
(design of the reactor,
destilation colum,
etc.)
Layer 1
E-14
P-13
E-13
V-6 P-14
P-16
Model 1
Model 2
Model 1
Model 2
Model 3
Expreimental
set up in
laboratory scala
Expreimental
setup in
miniplant scala
Expreimental
setup in pilot
plant scala
• Mass transport
• Phase eq.
• Unifac
• Uniquac
• Newtonian
fluids
• Bingham fluid
• Mass balance
• Energie balance
• Adiabat
• Isotherm
• State eq.
Experimental
set up in
laboratory scala
Experimental
set up in
miniplant scala
Experimental
set up in
plant scala
Experimental
set up in
laboratory scala
Experimental
set up in
miniplant scala
Experimental
set up in
plant scala
Introduction


4

measurement techniques. Finally, application of such complex models requires highly
trained personnel. Still, industry needs to take advantage of knowledge gained in basic
research in general and of application of complex models in particular.
A defined methodology to strategically simplify complex models, considering both the
requirements of a particular industrial process and the quality of the data available, is
missing in biotechnology. Although many reduction methods are applied for control
process purposes [7], a general approach for model reduction considering online and
offline measure possibilities, experimental conditions, and deep understanding of the
system is not to be found in literature.
1.3 UNDERSTANDING PROCESS DYNAMICS
Mathematical models can be described as the result of an effort to represent behavior of
nature with mathematical equations. Despite the inability of mathematics to precisely
describe physical phenomena, the approximate description achieved by models has
shown to be very useful. In the words of P. G. Box [8]:“all models are bad, but some are
useful”. Models are applied in all fields of science and have become an essential tool for
data acquisition and processing, understanding of complex systems, and prediction of
their behavior. In process engineering, models are used for process design, monitoring,
control, and optimization.
Modeling and simulation have developed rapidly over the last years [9]. Advanced
measurement techniques and fast computer processors enable the creation of very
complex models processing enormous amounts of information [10]. Nevertheless,
sophisticated models contain an important number of parameters and thus require large
amounts of very specific data in order to be identifiable. In most cases, experimental
effort for parameter estimation increases exponentially as the model grows in
complexity. Not only the measurement techniques become more complicated and
expensive but also the identifiability of the parameters is reduced with each new
parameter added to the model [11]. Online measurement limitations may also hinder the
application of complex models for model-based control. In addition, complex models
require costly hardware to make such complicated calculations as well as expensive
software to simulate and optimize the model efficiently. Furthermore, with an increasing
number of parameters to optimize, initial value consistency gains importance for
simulation convergence. Speaking of parameter estimation, a large number of
parameters increases the size of the optimization problem and number of local minima
[12]. Finally, all the complications mentioned above restrict application to highly trained
personnel.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


5

Batch processes commonly show highly nonlinear behavior and require more advanced
models for their description. Complications related to batch process simulation and
control are well known [13-15]. These dynamic and highly nonlinear processes require
accurate first principle models to be properly described. Nonetheless, in rigorous
modeling the choice of the mechanistic model to be used for the simulation is based on
the dominant physical phenomenon of the process. These phenomena, which dictate
the process dynamics, change over time. Hence, the appropriate approach is to simulate
the process with various models also changing over time. In other words, models should
change based on how and when these phenomena change. This is the principal reason
why most dynamic processes can be simulated effectively for short time periods but not
for the complete process. Nonetheless, in many cases only certain conditions of the
process are of interest. Simplifying the model to adapt it to the strictly important
conditions may reduce the complexity drastically. Unfortunately, this cannot be foreseen
and the model can only be adjusted once experimental data is available.
1.4 THE BRIDGE BETWEEN INDUSTRY AND RESEARCH
When speaking of industrial systems, there are many processes that operate without
detailed model-based knowledge of its dynamics. In the past, predictions were carried
out mainly on the basis of empirical knowledge. Experience and over-sizing combined
with improvements during operation led to fairly successful results. However, in recent
years an increasing trend to bring existing plants to meet new market demands can be
established. These demands include, for example, improved quality or compliance with
new standards for environmental restrictions. Unfortunately, simple nonlinear
regressions based on direct measurements are not suitable for these goals.
On the other hand, complex models present a number of disadvantages which hinder
their implementation in industrial processes. Low identifiability, complex measurement
techniques, large calculation costs and the need for highly trained staff are only some of
the problems to be faced in order to apply complex first principle models to industrial
processes. It is well known that mechanistic models offer a number of advantages over
“black-box” modeling, e.g. a higher process comprehension and a more accurate scale-
up capability [16-21]. Also rigorous models provide the basis needed for efficient quality
control. If correctly implemented, mechanistic models help to predict risks,
environmental impact and improve design and operation through simulation and
optimization. Still, rigorous models developed in basic research are rarely applied in
industry. Models have to be tractable, observable, robust, and simple but also accurate
and reliable for its use in industrial applications. In order to build models that have all
the aforementioned features and are also based on rigorous knowledge of the system, a
close cooperation between the research community and industry is essential.
Introduction


6

This work represents an important step towards the development of a systematic
approach to the adaptation of complex models for their application in industrial
processes. Model reduction is a promising approach to close the gap between models
developed in basic research and models required in industry.
One of the most difficult decisions to make for a modeler is the level of description
accuracy required for a model to be useful [22]. As we will see later in detail, an
agreement must be met between model accuracy and modeling, parameter fitting and
simulation efforts. Deciding how accurate a model needs to be to accept it as an
adequate description of the process is still an open question in engineering.
The difference between the outputs predicted by the model and the outputs measured
from the system is called residual [23]. Considering the exact parameters are known, the
causes for residual different from zero can be grouped in two main categories [24]:
- uncertainties (stochastic error)
Disturbances and unknowns are intrinsic errors of the system and
cannot be predicted. These show a normalized distribution with norm
equal to zero and a variance dependent on the conditions of the system,
measuring methods and further unknown factors.
- model structure (systematic, error)
When the structure of the model is incorrect, meaning it fails to consider
all important factors of the process and to represent the correct
dynamics, there exists no parameter set, which can make the model fit
the data.
Modelers usually tend to build models with too many parameters and to settle with
locally optimal parameter values. This trend is slowly changing with the development of
efficient global optimization techniques [25]. Global dynamic optimizers offer the
possibility to find the definite parameter set which best describes the observations [26].
The most significant contribution of global optimization to model structure analysis is
that one can rigorously demonstrate that the model is inconsistent with experimental
data regardless of its parameter values. Nonetheless, methods to detect the source of
systematic error are required and approaches to detect the instance of the structure
causing the error require further development.
Despite many efforts to develop automatic modeling programs [27, 28], the selection of
the structure of a model still requires individual analysis of each case and vast experience
in modeling added to deep knowledge of the system to be modeled. This is partly
overcome by adding new equations and parameters to patch errors in the structure of
the model. Nonetheless, these “patches” are usually responsible for unneeded parameter
correlation and reduction of model identifiability. To name one example, a straight line
can be described exactly by a fifth order polynomial, but such a model will never be
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


7

identifiable because there is an infinite combination of values of the polynomial, which
can describe a straight line. It is unidentifiable because an infinite combination of
parameter sets exists, which fit the system.
Creating new tools to analyze the structure of models and find correct representations
of the system is the main goal of this work. To achieve this goal, many disciplines need
to be brought together in an effort to attack model defects from different angles to
detect failures and to propose solutions. Finding communication paths between the
different disciplines to take advantage of the information gained in each case and
achieve the best possible model for each system is essential. Furthermore, as will be
shown in this manuscript, a combination of simple models may offer important
advantages.
1.5 RELATED WORK
Especially in process engineering, the use of models to obtain precise process
information based on indirect measurements has been utilized since the beginnings of
the discipline. There exists a handful of methods aiming at fast and robust description
of processes. Various fields in science require fast calculations to achieve optimal
control of systems with high dynamics. From missile tracking to burnout reactions,
many approaches have been successfully applied mostly using statistical methods and
repeated linear approximations of the system. Furthermore, the use of a combination of
more than one model in an effort to describe specific instances of a system or complete
processes has been proposed in various forms. Qualitative process theory [29, 30],
Interactive Multiple Model (IMM) [31, 32], jump Markov linear systems [33], qualitative
algebra and graph theory methods [34], semiquantative simulation [35], variable
structure theory [36], are just some examples. However, these methods relay on simple
models with no physical foundation with fast, but short term prediction being its
ultimate objective.
As limitation by computation burden losses significance due to the increasing capacity
of modern microchip architecture and cloud computing systems, the application of large
nonlinear models is gaining popularity. Approaches to reject hypothetical reaction
pathways in chemistry using first principle models in combination with global
optimization have been published [37]. Also online applications like model based fault
isolation and identification consider the application of rigorous models to detect
malfunctions in the system [38, 39]. These methods use software redundancy with
mechanistic models in an effort to detect fault behavior in complex systems.
Furthermore, fault detection techniques have rapidly evolved [23] and are being applied
in many fields of industry, e.g. PUMon (a tool for online monitoring based on neural
networks) is being developed at Bayer [40]. Nonetheless, despite the long story of
Introduction


8

similar methods to gain knowledge from limited data sets [41], its application in
complex dynamic systems is still limited.
Furthermore, a systematic methodology for the identification of non measurable
process variables, using a comparison between different first principle models
describing selected regimes in dynamic processes, is not to be found in literature [42].
Mechanism Recognition (MR) differs from all previous approaches in that the physical
properties of the system are considered. Most methods for system description with
more than one model aim strictly at computation expenses reduction, leaving system
understanding aside. On the contrary, MR is concerned with the characteristics of the
submodel and its relation to the physical system. Furthermore, MR aims at discerning
and selecting the phenomena dictating the dynamics of the system.
MR has been successfully applied for small systems [43]. Still, the first application is
limited to models with one state variable. In this example, different models were
obtained from literature each one describing a different regime of the process (section
6.1.1). Because of the simplicity of the models applied, no general structure was required
and input-output consistency was inherently fulfilled by the single input single output
condition of all models. The results obtained suggest that the method can be also
applied for systems with a higher number of state variables. Nevertheless, when
obtaining models from literature, a continuous computation of all state variables cannot
be assured (input-output consistency). Since the models are obtained from different
sources, the number and types of state variables contained by each model may differ.
Hence, it is not possible to assure calculation of all state variables in every regime.
Furthermore, complex models require a general structure to increase its identifiability
this cannot be generated for models with different characteristics.
The core of MR is model building, most precisely, submodel building. Once physical
meaning of each submodel has been experimentally validated, and its interaction with all
other submodels has been understood, the application of MR is straight forward and
some of the aforementioned techniques can be applied. Still, the practice of modeling
should not be underestimated. Novel software toolboxes for model building
modularization and reusability [44, 45] together with efficient integrators [46] facilitate
the exercise of modeling significantly. Furthermore, a number of software packages for
automatic model building [47] and automatic model reduction [45] confirm the trend to
a general, systematic, and automated modeling approach. Nevertheless, modeling is still
a field which requires intensive human intervention. The engineer must make use of his
knowhow and intuition to be able to develop efficient models which mirror reality and
are consistent with scientific evidence. This work provides significant evidence that
despite technological advances, modeling is still a challenging and exciting discipline
[48]. The challenges of modeling and experimental validation will be discussed, different
manners to create and analyze models (chapter 3) and its relation with the observations
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


9

of the system (chapter 4) will be presented in an effort to increase the efficiency of
model development.
1.6 PROJECT GOAL
The main goal of this project is to find new approaches for a target-oriented model
simplification. By these means, complex models created in basic research can be adapted
for application in industrial processes. Various methods for model reduction are to be
studied in combination with mathematical tools for experimental information
quantification (confidence intervals, optimality criteria, etc.) to fulfill specific
requirements of particular industrial problem.
Secondly, this works aims at finding new means to accurately describe complex
processes based on simple models. In order for a simple model to mirror a complex
system, three essential conditions must be fulfilled:
- deep comprehension of the dynamics of the system
- The complete system, but more important, the phenomenon
governing systems behavior must be deeply understood.
- minimal systematic error
- Equations and structure of the model must describe only the
most important dynamics, with the minimal number of
parameters possible and minimal systematic (e.g. modelization)
error.
- high model identifiability
- The data set must deliver enough information to estimate the
parameter set with high accuracy. It is essential to understand
that identifiability depends not only on the data set (state
information), but also on the structure of the model.
Now let us assume that a specific variable or process parameter cannot be measured due
to physical limitations. Let us also assume that we have created a model, which satisfies
the above mentioned conditions. This means that it is able to describe the strictly
defined regime of the system with high accuracy. This very special characteristic is
exploited by MR. If it is precisely known which regime can be described by the model, a
process running outside this regime can be easily detected.
MR provides insight into the system, allowing a deeper understanding of process
dynamics and process monitoring to operate in optimal conditions. The biggest
challenge for the application of MR is how to create a simple but accurate model
specifically adapted to the particular conditions of each regime. This is also the main
topic throughout this manuscript.
Introduction


10

The validity of the approaches proposed will be tested in two case studies of high
relevance in the field of water treatment and recombinant cultivations.
Finally, it is worth recalling that physical understanding of the system, either chemical or
biological, is the keystone to this approach. MR does not intend to substitute human
reasoning or make up for lack of process knowledge. On the contrary, MR is merely a
tool to efficiently apply this knowledge in order to gain a better insight of the system
under study.
1.7 ADVANTAGES OF MECHANISM RECOGNITION
The information contained by the complex model has to be used with intelligence to fit
the process needs while increasing the identifiability and the observability of the
submodels. Furthermore, a reduced model comprehends much more information than
the same model built using the classical top down approach (from black-box to grey-
box to first principle models). The most important advantages when creating a
submodel through an intelligent reduction of a complex model are:
- Specified adaptations for each process:
- A defined model reduction can be carried out for a specific
process. By these means the model is adapted to each particular
case. Again, because of the mathematical basis, the information
gained can be exported to systems and used for different
conditions.
- Phenomenon identification:
- The model reduction can also be conducted to determine a
selected phenomenon of the process. This allows the
identification of non measurable variables and increases the
information obtained by the experiments.
- Knowledge about the accurate experiment is gained through model
reduction:
- The creation of reduced models and their parameter estimation
delivers important information to be implemented in the
complex model. For example, the nonlinear interrelation of the
states in the complex model can be understood better if the
behavior of its reduced models is analyzed.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


11

1.8 THE GOOD, THE BAD, AND THE USEFUL MODEL
It is common to evaluate models as “good” or “bad” and these terms are also used in
this work following convention. Still, it is essential to be aware that all the approaches to
model evaluations might fail. Although it is true that some special characteristics of a
model must be analyzed before using it, experience has shown that it is very difficult to
predict the functionality of a model. Particularly in engineering, the most important
question to answer is whether or not certain model characteristics can be exploited
aiming at specific goals. In many cases, the simplest model has shown to perform much
better than complex, nonlinear ones. Reasons for this are explored in this work.
Engineering, being a practice and industry oriented discipline, is mainly interested in
usefulness of models. For an engineer the principal aspect to take into account is if a
model can bring some advantages in process efficiency or not. For a model to be useful,
it is necessary and sufficient that it be robust, reliable, and descriptive.
A model that robustly describes the simplest part of a system properly is far better than
a complex model that mirrors the complete process but has a high probability of failure.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


13

2 MODELING
2.1 DEFINITION
This works considers mathematical models exclusively and their application in the
description of physical phenomena. For sake of generality, we limit our concept of
mathematical model to the definition made by Aris [49]:
“A mathematical model is a representation, in mathematical terms, of certain aspects of a
nonmathematical system. The arts and crafts of mathematical modeling are exhibited in the construction
of models that not only are consistent in themselves and mirror the behavior of their prototype, but also
serve some exterior purpose.”
Furthermore, the study of this work is limited to mechanistic models expressed in the
form of Differential Algebraic Equation (DAE) systems applied exclusively for
description of process engineering in chemistry and biotechnology (2.1). Finally we
delimit to controlled physical, chemical, and biological systems.
, , , , , = 0
2.1
where is a vector with the derivatives of the state variables, is a vector with n
s

time-dependent variables which define the system, a vector of n
u
time-dependent
input variables, is a vector with n
w
constant input variables, is a vector with P
parameters, and represents time.
The initial conditions are also to be defined.

0
,
0
,
0
, , ,
0
= 0
2.2
where t
0
is the time at point 0.
Contrary to black box models, mechanistic models are based on physical knowledge of
the system to be described. In engineering for example, rigorous modeling includes
mass and energy balances, detailed reaction pathways, etc. Models are the core of
Computer Aided Process Engineering (CAPE) [50] and Computer Aided Biology
(CAB). The quality of every work on simulation, optimization, design, and model based
control, depends on the characteristic of the model. Models are evaluated by its
simplicity, accuracy, robustness, generality, and computation burden. It is worth
reminding, that there is no such thing as the “best” model for all applications. The
Modeling


14

“best” model can only be selected after the objective of the simulation and the state of
information (chapter 4) has been specified.
In engineering, models are not only used to describe the behavior of systems, they are
also essential to map complex systems into smaller dimension more comprehensible to
humans. Finally, they also serve to obtain indirect measurements and observe non
observable events. This last category of models is also known as software sensors [51].
Software sensors substitute measurements, which are not possible due to physical
limitations, with models which predict the behavior of the non measurable variable
based on indirect measurements.
2.2 MODEL COMPLEXITY
A common mistake is to consider the most complex model to be the most appropriate
for description of a system. In most of the cases it has shown to be quite the opposite.
Experience shows that the fewer the parameters in a model, the better [52]. Still, the
first solution that comes to mind when a model fails to describe a system is to add new
parameters. Instead, this should be considered the last resource and should be done only
after all other options have been exhausted.

Figure 2.1 : E. coli transcriptional regulatory network. [53].
Model complexity is closely related to instability, over parameterization, parameter
correlation, and low parameter identifiability. The effort required to develop and fit a
model has to be justified by its application. It is useless to apply Computational Fluid
Dynamics (CFD) to the simulation of a 1L reactor knowing that the concentration
gradients can be neglected. On the other hand, simulating a reaction in a tank with
10,000 L without considering mass transfer limitations may yield catastrophic results.
Summarizing, the key dynamics of a system need to be identified, isolated and analyzed
before any model is built. Currently, the three conditions (section 1.6) are limited mainly
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


15

due to the scarcity of measurement possibilities but also due to the insufficiency of
adequate mathematical tools. It is at this point that the MR approach can contribute to
modern model building.
A model with hundreds of parameters including exponential, hyperbolic, and
discontinuous functions might seem advanced and sophisticated, but this illusion
quickly vanishes when the model has to be validated and used for design or
optimization. Much better is a correct approximation, than an accurate misconception.
The real challenge for modeling is to develop a general and systematic approach to find
the simplest manner to describe complex systems aiming at the strictly required
accuracy. The meaning of model simplification becomes more important everyday with
the increasing complexity of processes analyzed in research Figure 2.1.
2.3 ENGINEERING APPROACH TO COMPLEX SYSTEMS
In chemical engineering, the implementation of different methods to deal with large
complex systems has a long history. Engineers have developed methods like hierarchical
modeling, model reusability, model inheritance, etc. An extensive discussion of these
methods and their application for the simulation of chemical plants is presented by
Barton [3]. In biological systems, the modularization of separated instances of the
system is not always possible. In traditional process engineering, a pump can be
modeled in a modular form and then added to the flow sheet of the plant and reused as
many times as needed [54]. Contrary to this, biological systems tend to show different
behavior under in vitro conditions compared to their in vivo state [55]. Still, some
approaches intend an analysis and modeling of biological systems with methods taken
from engineering [56, 57].
An alternative method to create optimal model structures has been published by
Bardow [58]. This method called the Incremental Approach (IA) suggests building the
model in an inductive manner. In a sense, IA could be considered a hierarchical
approach extended to an even lower layer to first principle phenomena. Although its
application finds important limitations, e.g. quality of data required and bias, the general
concept behind IA is worth our attention. In principle, IA extends the philosophy of
hierarchical modeling to the molecular level.
Inverse problem theory is the most common approach for model building and
specifically models fit to data. First, the differential model is evaluated (integrated) with
a certain parameter set, and then the data is compared against the output previously
computed. The residual between model outputs and data is calculated and a new set of
parameters is tested. These steps are followed iteratively, usually solving some least
square type of optimization problem (section 4.1.1) until the residual is considered to be
minimal. An important disadvantage of this approach is that it is not possible to directly
Modeling


16

analyze the internal structure of the model. Although, various methods exist to
indirectly investigate parameter sensitivities, correlations, and bifurcation among others,
a true insight in the structure of the model is still not possible.
Bardow [58] proposes finding the parameter value needed to fit each new data point.
Estimating new parameter values for each data assures that the differential equation
presents the correct derivative. This process would be similar to fitting one parameter
for each measured point independently. The result we obtain is a curve showing the
ideal parameter values. This curve, although very noisy in most of the cases and without
any physical meaning, is very helpful when building a new set of equations. The modeler
can visualize the behavior of the parameters and decide if they can be represented by
constants, or algebraic or differential functions.

Figure 2.2: Incremental approach for reaction kinetics identification [58]
IA proposes to build the model in a deductive way. A drawback of this approach is that
process information is required for each step in the model building process. Still, this
method can be very useful if advanced measurement techniques are available. Besides,
IA offers a very well described systematic procedure for model building, which is usually
underestimated in process engineering and biotechnology Figure 2.2.
2.4 MODELING IN SYSTEMS BIOLOGY
2.4.1 SYSTEMS BIOLOGY
We refer to the definition by Kitano [6, 59]: “Systems biology aims at understanding biological
systems at system level”. Systems biology emphasizes the fact that the only possible manner
to understand living organisms is to consider the system as a whole.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


17


Figure 2.3: Hypothesis-driven research in systems biology [59].
Identifying genes and proteins is only the first step, whereas real understanding can only
be achieved by uncovering the structure and dynamics of the system. Kitano states four
key properties:
- System structure
- System structure identification refers to understanding both, the
topological relationship of the network components as well as
the parameters for each relation.
- System dynamics
- System behavior analysis suggests the application of standardized
techniques such as sensitivity, stability, stiffness, bifurcation, etc.
- The control method
- System control is concerned with establishing methods to
control the state of biological systems.
- The design method
- System design is the effort to establish new technologies to
design biological systems aimed at specific goals, e.g. organ
cloning techniques.
The relevance of modeling in systems biology is clearly stated in Figure 2.3.
2.4.2 MODELING OF GENETIC REGULATORY SYSTEMS
System biology has triggered an impressive contest between various methods aimed at
an adequate description of the dynamics of living organisms studying its Gene
Regulatory Network (GRN), the most representative being [60]:
- Directed and undirected graphs
- Bayesian networks
Modeling


18

- Boolean networks
- Generalized logical networks
- Linear and nonlinear differential equations
- Piecewise linear differential equations
- Qualitative differential equations
- Partial differential equations
- Stochastic master equations
Each approach offers different advantages and no definitive method can be defined as
the “best” by the systems biology community. The Assessment of Network Inference
Methods attempt to analyze all pros and cons of the different GRN inference methods.
The goal is to compare the different approaches against equal data sets to obtain
quantifiable information of the difference in performance between the methods [61].
Because complete understanding of the system is essential for a proper evaluation, the
most promising results have been obtained with simulated data sets, but much work is
to be done before an adequate comparison can be achieved.
As stated before, differential equation systems settle the standard modeling method in
engineering. For this reason, the most interesting model approaches for MR are the
ones based on differential equations. In fact, linear and piecewise linear differential
approaches are perfectly suitable for model reduction.
Systems of Ordinary Differential Equations (ODE) have been widely applied for the
description of GRN. Usually the system comprises rate equations of the form

=

,
2.3
where x can be the vector of concentration of proteins, mRNAs, or other molecules, u
the vector of inputs, and f
i
is a nonlinear function. Also time delays can be added if
necessary. Typical types of equations used are, Monod type, switching, Heaviside, and
logoid functions among others. An important advantage of nonlinear ODEs is the
possibility to describe multiple steady states and oscillations in the system [62]. Besides
the requirement of testing the global convergence of the optimal solution, the bottle
neck is still the state information of the parameter set creating identifiability problems.
Nevertheless, some successful applications have been published showing the
possibilities of ODEs to describe GRN [63].
It is worth recalling that MR aims at simple model building and GRN modeling is far
from this. Still, both GRN modeling and model analysis and reduction techniques have
shown exponential development in the last years. Therefore, it can be expected, that
systematic conversion of complex GRN models in simple submodels suitable for MR
will be possible in near future. Someday, detailed descriptions of complete GRN will be
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


19

the basis for perfectly defined submodels applied in industry to make fast, robust and
accurate predictions of complex processes.
2.5 MATHEMATICAL MODEL FOR A BATCH
BIOCHEMICAL REACTOR
MR finds its most important application in dynamic systems. A process in constant
change presents different behaviors and governing phenomena also change over time. It
is at this stage where the different process conditions can be selected and the submodels
can be built. Biochemical batch reactions have been selected to validate MR and its
application for the description of industrial processes. For this reason, a short discussion
of the general form of the mathematical model is presented.
The biochemical reactions involve consumption of various chemical species (substrates)
and production (intermediate or final metabolic products) and biomass growth.
Products from a microbial group are often the reactants of other microbial groups. This
results in a sequence of individual process steps, which is part of a scheme, where some
steps may be independent of those that follow [64].
Assuming that biochemical reactions, generally described through

,

+
,

=1

2.4
take place in a batch biochemical reactor, the following differential equations can be
derived .

=
,

1
, ⋯,

,
1
, ⋯,

, = 1, …,

=1

=
,

1
, ⋯,

,
1
, ⋯,

, = 1, …,

=1

2.5
where:

, = 1, …, are the concentrations of the chemical species (substrates and/or
products) in the reactor,

, = 1, …, are the concentrations of the microbial masses
in the reactor,

1
, ⋯,

,
1
, ⋯,

, = 1, … , are the reaction rates,
,
and
Modeling


20

,
are the stoichiometric coefficients for substrate consumption and microbial growth,
respectively.
It should be noted that the consumption of a substrate (e.g. particulate matter) may not
be associated with biomass growth. Moreover, a single microbial group may grow on
more than one substrate and vice versa. Therefore, in the general case, the number of
the substrates involved in a bioreaction scheme will not be equal to the number of
microbial masses grown, i.e. ≠ .
Introducing vector notation for the concentrations and the rates
=

1

, =

1

, , =

1

1
, ⋯,

,
1
, ⋯,

1
, ⋯,

,
1
, ⋯,

and denoting by C and Y the x and x matrices of the stoichiometric coefficients,
model 2.5 takes a more compact form:
( , )
( , )
= ·
= ·
S C r S X
X Y r S X



2.6

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


21

3 MODEL REDUCTION
3.1 INTRODUCTION
A model is a poor mathematical representation of a physical system. Lack of accurate
knowledge of the process to be modeled, insufficient measurement techniques and
extensive computation time hinder an exact representation of the phenomena to be
described [65]. Nevertheless, models are widely used in science and their contribution to
a better understanding of engineering processes and their proper design, optimization
and control is unquestionable. From this it can be deduced that the best model to
describe a certain process is not necessarily the most accurate, but the one that describes
only the relevant aspects of the system so as to get a good description with minimal
effort [66]. Different methods have been developed to detect the key dynamics in order
to create an accurate but relatively simple model.
The process can be described as a bottom up approach in the hierarchical modeling
sense. Once a detailed model has been built, model reduction leads to model
simplification. Because of the information gained from the detailed model, the reduction
follows mathematical and physical principles. By these means species are neglected and
dynamics are simplified based on their influence on the overall system.
Model reduction is keystone in engineering, a widely applied approach for reduction of
nonlinear models is the linearization based on Taylor series, which has proven to be
very useful for processes in steady state conditions. Unfortunately, dynamic nonlinear
systems require more complex approaches. Model reduction aims at distinguishing the
important from the negligible modes in an effort to reduce the model to a more
attractable form maintaining its key dynamics [67]. Some of the most important
advantages of reducing a model are:
- increased identifiability/observability.
- increment of model robustness
- reduction of model stiffness
- reduction of computation expenses
As a result, not only the experimental effort for parameter estimation is drastically
reduced, but, most important, the measurement effort during process monitoring and
control is minimized. It may be even possible to convert non observable models into
models adequate for model-based process control. A very important difference between
steady state and dynamic processes is the selection of the “important” dynamics.
Model Reduction


22

In model reduction for continuous processes based on time scale analysis, the fast
modes are neglected. Only the slow modes, which determine the path while reaching the
equilibrium point are maintained [68]. On the contrary, when a system is far from
equilibrium (as is usually the case in dynamic processes) the fast modes are of major
importance and the very slow modes can be considered constant (quasi-steady state). In
addition, these new constants may give rise to further reaction invariants, which should
be considered because of their model reduction potential.
Literature dealing with model reduction under steady state conditions has been widely
published. However, less work has been done in model reduction of dynamic processes.
On the one side, most industrial chemical processes run at steady state (constant
operating conditions). On the other hand, the condition of stability enables many
assumptions which simplify the reduction problem significantly, linearization and
Lyapunov stability being just two examples. In biotechnology, most processes are batch
or fed-batch processes which allow no steady state assumptions. New approaches to
model reduction need to be developed to permit reduction of models for dynamic
processes. A representative example of the reduction potential of dynamic models is
shown in section 7.3.
3.2 BASIC APPROACHES TO MODEL REDUCTION
The reduction of a particular model maintaining its characteristics is a common task in
all fields of engineering. Representative examples are Lumping [69, 70], Sensitivity
Analysis [71] and Time-Scale Analysis [64, 72, 73]. In this chapter, a short introduction
to some of the methods used for model reduction is presented. Besides many efforts,
model reduction techniques still relay on process knowledge and experience to achieve
the correct reduction of the model. In most of the cases, a combination of methods is
required to achieve the simplest form possible.
3.2.1 REACTION INVARIANTS
In chemical and biochemical models, the method of reaction invariants is applied to
detect the modes which stay constant in the process. Firstly, the method finds linear
dependencies in the model. Secondly, reaction invariants help detect the invariant
manifold which in turn is a valuable tool to find the slow manifold of a system of
dynamic equations.
In the general model of equations 2.6, there are ( +) differential equations that are
affected by reaction rates. As long as + > and the differential equations are
independent of each other, there will be + − linear combinations of the
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


23

concentrations that are completely unaffected by the reaction rates and therefore
completely unaffected by the progress of the chemical reactions. In the literature, these
are referred to as reaction invariants; they capture the reaction stoichiometry relations,
which are not affected by the reaction rates.
The reaction invariants can be easily calculated from the general form 2.6 .
Assuming + > and

(
=
(
(
¸ ¸
Rank n
C
Y
,
one can find ( +−) linearly independent row vectors

, = 1, ⋯, ( +−)
of length ( +) such that

0
C
Y
v
o
(
=
(
(
¸ ¸
= 1, ⋯, ( +− )
This means that the +−x( +) matrix

1
 A
v
o
o
(
(
=
(
(
¸ ¸

has rank ( +− ) and satisfies
0
C
A
Y
(
=
(
(
¸ ¸

3.1
It can then be easily verified, as a result of (2) and (3), that the quantity

(
=
(
¸ ¸
S
z A
X

3.2
remains constant throughout the entire batch:

0 = z

3.3
and so

( ) (0)
0
( ) (0)
( (
= ¬ >
( (
¸ ¸ ¸ ¸
S t S
A A t
X t X

3.4
Model Reduction


24

3.2.2 SWITCHING FUNCTIONS AND THE REACTION
INVARIANT
Switching Functions (SF) are widely applied in biological systems. Its most common
form corresponds to the simplified Michaelis-Menten equation 3.5.
=

+

3.5
where r, r
m
, K
S
,

are the reaction rate, the maximum reaction rate, the inverse of
enzyme affinity and the substrate concentration, respectively.
SF enable activation or deactivation of different reactions depending on the
concentration of species. The objective is to create a continuous and smooth function,
which is equal to 1, when the concentration of the limiting component is high, and
equal to 0 when the limiting component has been depleted. The inhibition functions
show the opposite behavior, the equation is equal to 0, in case of high concentrations,
and to 1, when the species is not present. Referring to equation 3.5, will have the
value

when the concentration of substrate is high and 0 when the concentration of
substrate is near 0. We characterize the switching function behavior in three phases
Figure 3.1.
- active and constant (species concentration is significantly higher than the
limiting constant > 100)
- transition phase (when the ratio species/constant is between 100 and 1*10
-2
)
- inactive and constant (species concentration is significantly lower than the
limiting constant <1*10
-2
)
This particular characteristic of the SF opens interesting possibilities for model
reduction. The presence of SF in a mathematical model may engender temporal reaction
invariants during specific time intervals in a process. To be more precise, the process
presents different reaction invariants under limitation of different species.
To visualize this idea, considering the simplified Michaelis-Menten equation in 3.5, it
can be seen that if > 100 ∀ > 0, then ≈

= constant. We can therefore
create reduced versions of the original model, which behave exactly as the original
model under specific conditions. This can be very useful when applying general
mathematical models to defined conditions. When a model describes a wide range of
process conditions all species limitations, which may or may not occur during the
process, have to be considered. However, once the conditions of the process are well
defined, some limitations may be neglected and new reaction invariants can be detected.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


25


Figure 3.1. Behavior of a switching function in dependence of the limiting species.
Although no general method exists for this procedure, it is possible to achieve
important reductions with some engineering experience and mathematical background.
An example of the capabilities of this approach is presented in section 7.1.1, were the
state of the art model for Active Sludge Process (ASP) is reduced to one third of its size.
3.2.3 SENSITIVITY ANALYSIS
The basic principle of model reduction through sensitivity analysis is to eliminate the
components which are not relevant for the accuracy of the model. To correctly reduce
the order of the system it is important to eliminate only the species, which have both a
weak effect on the model outputs as well as on the important species of the model.
Local sensitivities can be easily estimated via finite differences. If the Jacobian matrixes
of both state variables and parameters can be obtained, dynamic sensitivities can be
computed tailored to the numeric integration of the DAE system.
sDACL [46] is a code for efficient integration of general DAEs with sensitivity
generation. Cheap computation is achieved through partial discretization tailored to the
integration of general DAEs based on Orthogonal Collocation on Finite Elements
(OCFE). This algorithm provides exact sensitivity information of the numeric
integration, and can be implemented in any one-step integration method. sDACL has
been customized to the staggered method for state and sensitivity integration showing
similar computational efficiency. Finally, the algorithm is able to take advantage of the
sparsitiy properties common in engineering models. With efficient simulation tools like
sDACL, an efficient analysis of the dynamic sensitivities is possible.
0 2 4 6 8
0
20
40
S
u
b
s
t
r
a
t
e

c
o
n
c
.

[
g
r
/
m
l
]
Time
0 2 4 6 8
-0.5
0
0.5
1
r
e
a
c
t
i
o
n

r
a
t
e

[
]
Phase 1
Phase 2
Phase 3
Model Reduction


26

3.2.4 LUMPING
Lumping is widely applied when dealing with systems with a large number of
components. A process were the advantages of lumping has be demonstrated is steam
cracking [74], where the millions of different components need to be grouped in a few
categories to allow its simulation. Lumping can be divided in two main categories,
proper and improper lumping. When speaking of proper lumping, each one of the
reactants to be lumped appears exclusively in one lumped entity. On the other hand,
improper lumping considers the possibility of one or more reactants contributing to
different lumped entities. Although proper lumping is mostly used in mathematics, most
reactions are only to be described precisely by improper lumping methods. A nice
illustrative example is of proper lumping given by [70]:
Let us assume we have the following three component system:

Figure 3.2: Three-component monomolecular reaction system,
the numbers on the arrows represent the back- and forward reaction constants.
We consider the system to be described by the following monomolecular reaction
scheme:

1

= −
1,1

1

1,2

2

1,3

3

2

= −
2,1

1

2,2

2

2,3

3

1

= −
3,1

1

3,2

2

3,3

3

3.6
where k is the monomolecular rate matrix
k =
13 −2 −4
−3 12 −6
−10 −10 10

3.7


A
3
A
2
A
1
10 10
4 6
2
3
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


27


and fulfils the conditions mentioned bellow:
- Nonnegative rate constant
- Mass conservation
- There exists an equilibrium composition


> 0, = 1,2, …, such
that:

= 0
For the system to be lumpable with proper lumping it is necessary and sufficient to find
a

matrix such that:
∗ =


3.8
In this example:

1 1 0
0 0 1

13 −2 −4
−3 12 −6
−10 −10 10
=
10 −10
−10 10

1 1 0
0 0 1

3.9
In conclusion, it is possible to represent the system with two components, one
representing
3
and the second representing a combination of
1
and
2
Figure 3.3.

Figure 3.3: Lumping a monomolecular three-component reaction into a
two-component reaction
Unfortunately, real processes can rarely be described by proper lumping since for most
systems there is no M matrix which fulfils eq. 3.8. On the other hand, the theory of
semi-proper and un-proper lumping can become rapidly complicated and are not
discussed in this work [75]. However, it is very important to consider the possibility of
lumping different species, especially in systems with a large amount of components or
cells with similar behavior.
3.2.5 PERTURBATION THEORY
In some problems, it may be the case that a small parameter can be detected (usually ε)
such that the solution does not differ significantly if ε = 0. Perturbation theory is a
systematic mathematical approach to find the solution for the case ε = 0. Perturbation
theory might be seen as the pure representation of model reduction. The model is
analyzed to eliminate the parameters which cause very small changes in the dynamics of
A
3
A
2
A
1
10 10
4 6
2
3
A’
2
A’
1
10
10
Model Reduction


28

the model. Again the real challenge is to find these parameters in complex nonlinear
processes. Still, the method of perturbation theory allows fairly accurate results if
process knowledge and an iterative method are combined.
A good example is modeling of a projectile thrown vertically into the air. An important
question to be answered is if air friction should be considered in the model.

2

2
+
Γ

+1 = 0

0
= 0;

0

= 1
3.10
where = [∗ /] is the friction constant, Γ = [m/s] the initial velocity of the
projectile, = [] its mass, = [/
2
] gravity acceleration, and = [] the time.
It is possible to consider that the relation = Γ/ is very small, for low initial
velocities, a body with big mass, and low air friction, and apply regular perturbation
theory to find an approximate solution to eq. 3.10. The solution obtained.
= −

2
2
+ ∗ −

2
2
+

3
6
+
2

3
6

4
24
+
3

3.11
Eq. 3.11 is the exact same solution one would obtain after calculating the Taylor series
expansion of the exact solution.
3.2.6 TIME SCALE ANALYSIS
Reaction systems are characterized by an interactive combination of fast and slow
reactions. This gives rise to stiff equation systems which are difficult to solve with
numerical methods. Substituting the fast reactions by algebraic equations, under the
assumption that these species are instantaneously at equilibrium, has proven to be a very
effective model reduction method. The Quasi Steady State Assumption (QSSA) has
been widely applied in chemical engineering and biochemistry [76]. For the QSSA
method empirical knowledge of the system to be analyzed is required. The engineer
must decide which modes are fast enough to be considered steady state. Even though
some mathematical methods have been developed to identify fast modes in equation
systems, a generalized implementation of the method has proved to be difficult.
Stamatelatou [64] published a nice application of time scale analysis for biochemical
reaction steps in CSTR. In this paper, Stamatelatou provides a systematic answer to find
and remove the fast dynamics of acidogenesis. A simple example taken from this paper
should illustrate the concept of the slow manifold.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


29



Consider the following model describes a chemostat process:

= − ∗ + ∗

= − ∗ + −
3.12
where X and S are the biomass and substrate concentrations respectively, D is the
dilution rate, Y is the biomass yield, = ∗

is the feed rate of the substrate, S
0
is
the substrate concentration in the feed, =

/

∗ /(_ + ) ∗ is the
reaction rate,

is the maximum specific growth rate constant of the biomass, K
S
is
the saturation constant, and assume that it operates under low dilution rate relative the
growth of microorganisms: /

≪ 1.
The work of the experienced engineer is to detect the fast and the slow dynamics.
Stamatelatou deliberately selects the wrong dynamics to show the limitations of the
approach. Nonetheless for sake of clarity, the correct dynamic is considered to be the
fast one. In this simple system, it is very easy to foresee that biomass growth has a
slower dynamics than substrate concentration.
Defining the reaction invariant:
= +( −
0
)
3.13
We can reformulate eq. (3.12) with the following two expressions, depending on which
is considered to be the fast and the slow dynamic of the process. Considering that the
dynamics of the biomass is fast enough to be substituted by an algebraic equation, the
equation system can take the form:

= − ∗

= ∗ (
0
−) −

∗ ∗ (
0
− +)
= −
0
+ ∗
3.14
If on the contrary, the dynamics of substrate is considered to be faster, the following
expression is more appropriate:

= − ∗

= − ∗ +


0


3.15
Model Reduction


30

=

+
0

Now let us make a phase diagram where the trajectories of S vs. X are plotted for
different initial conditions. In addition, we can see the trajectories of the two reduction
hypothesis made (setting z equal to zero).

Figure 3.4: Phase diagram of full order model (3.12). Comparison with reduced models in a
chemostat process
The method of the invariant manifold can be very effective to reduce stiff systems.
Nevertheless, finding the fast and slow dynamics rapidly becomes a difficult task in large
systems.
Finally, it is worth recalling, that the field of model reduction is showing important
advantages towards reduction of complex nonlinear systems. New techniques in
combination with advances in computer technology permit the solution of large systems
in a systematic manner [77-80]. In near future, submodel building will become an
automatic task in the MR program. By these means, a set of submodels should be
automatically available to serve recognition purposes. Still, today the proper reduction of
complex models requires much work and collaboration of different disciplines. The
efficient recognition of a process regime still depends on the proper building of a set of
submodels able to describe the process states accurately while being fast to compute,
robust, and differentiable from each other.
20 40 60 80 100 120 140 160 180 200 220
100
200
300
400
500
600
biomass X[mg/l]
s
u
b
s
t
r
a
t
e

S

[
m
g
/
l
]
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


31

4 OPTIMAL EXPERIMENTAL
DESIGN
In chapter 3, methods to analyze model structure, parameter set, and its interaction in
order to create models striving to defined objectives are discussed. Nevertheless, since
the data set against which the model is fitted is at least as important as the model, a
discussion regarding how to design experiments, in order to allow a better analysis of
models and estimation of the parameter set has great significance.
The principal quality of models is its capacity to mirror a defined system. It is therefore
not possible to determine the usefulness of a model without testing its description
accuracy. In order to do so, observations of the real system have to be made and the
similarities between model predictions, and observations of the real system, are to be
compared. Models are created under numerous assumptions and for strictly limited
systems. It is important that experiments carried out to test models consider these
limitations.
The data set is obtained from observations carried out in defined experiments.
Searching for the parameter vector of the model, which best describes the data set, is
known as parameter estimation or model fitting. In physics, this is known as the inverse
problem. Inverse problem theory studies the appropriate manner to infer the values of
the parameter set using the results obtained from observations [24]. The theory of
inverse problems has been well described for linear systems, and although it can also be
applied to nonlinear problems, it becomes rapidly complex and requires high
computation expenses. For this reason, more conventional approaches are applied for
engineering applications. Even though some assumption need to be made and the
calculations represent merely approximations of the state of information, experience has
shown that simple methods deliver fairly accurate results and are useful in many
processes [81].
To understand a system and model its behavior, a long iterative approach is required.
First the system has to be observed, to achieve this, some experiments are carried out.
Next, a hypothesis is made based on the observations. Following, a model is created on
the foundation of the hypothesis, information from literature, and models created for
the same or similar systems. The model has then to be fitted to the data set. An
antithesis is created and the model is adapted. New experiments are required based on
the information gained during the development of this process. Variables to be
Optimal Experimental Design


32

measured, frequency of sampling, experimental conditions, among others are updated
and a new set of experiments is planed and performed. The model is fitted to the new
data sets, and this iterative procedure continues until the accuracy of the predictions is
accepted or it is considered that the system cannot be described.
Optimal Experimental Design (OED) deals with the selection of the optimal
experimental setting. Methods like Principal Component Analysis (PCA) or Partial Least
Squares (PLS) search for data correlation to reduce the dimension of the data set [82].
Also more advanced methods in Knowledge Discovery of Data (KDD) like data mining
[83] have been developed for treatment of large data sets. These methods study the data
characteristics to find new relations between variables and create black-box type models
which describe it.
For modeling purposes, alternative methods have been developed which consider the
propagation of data uncertainties through the model and its reflection on the parameter
set. In literature, these methods are grouped under the field of Model Based Design of
Experiments (MBDoE).
MBDoE can be divided in two main categories: 1) parameter accuracy and 2) model
discrimination. In parameter accuracy, the optimization problem searches for the
experimental settings with the highest state information content over a parameter set,
considering the characteristics of the model being fitted to increase model identifiability
[81] (section 4.1.2). MBDoE takes advantage of the dynamics of the process to increase
the sensitivity of the model outputs with respect to the parameters. Large parameter
sensitivities reduce the confidence intervals of the parameter set, increasing the accuracy
of the parameter estimation. In other words, it increases the information obtained from
the experiment to fit the model parameters with the highest possible accuracy.
In the general case, various models are proposed as possible candidates to describe the
system and the model with the highest probability to describe correctly the system is
selected. In this case, the optimization problem has a different objective. The goal is to
maximize the difference between the outputs of the candidates. The set of experiments
is chosen in an effort to maximize the distinguishability between candidates. It is worth
to highlight that model distinguishability is closely related to model identifiability, since
accurate parameter estimations increase the probability of correct discrimination
between candidates.
MBDoE, has been widely investigated and has shown important progress in basic
research and industry. Not only theoretical work but also an important number of
applications of MBDoE have been published, the most representative examples are [84-
87].
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


33

4.1 THE EXPERIMENT
In practice, some parameters included in the model are not known beforehand and need
to be estimated to enhance model description; the model is fitted to experimental data.
The accuracy on the estimated parameters is influenced by two factors.
- The first one is the quality of the data obtained from the experiments.
Measurements are inevitably subject to uncertainties. For this reason,
data sets should not be considered observations but a “state of
information” acquired on observable variables.
- The second one is the sensitivity of the objective function with respect
to the parameters. The structure of the model defines the interrelations
and effects of the parameter values in the output vector. Outputs which
are highly sensitive to changes in the parameters allow a precise
estimation of the “exact” parameter set.
An optimal set of experiments should be designed considering both factors mentioned
above.
Experiments can be defined by the following conditions [81].
Variables that can be controlled:
-

() time-dependent control values
-

constant control values
- ()measurements in the selected time intervals
- (0)initial conditions
-

experiment duration
Parameters that cannot be controlled:
-

Systematic error
- Stochastic error
Systematic errors are errors on measurements which can be detected quantified and
reduced. They present a mean value different from zero and error bias. On the other
hand, stochastic errors can be defined as the difference between two identical
experiments.



Optimal Experimental Design


34

Experiment outputs are represented by the vector ) (t y :

=

0
,

, w
E
,

, ,

∀ ∈

4.1
Where the sub index E represents the variables of the experiment, and

represents
the time span of the experiment.
Model outputs are to be compared against the experiment outputs. For this reason,
model outputs are represented in a form similar to experiments outputs. Model outputs
(model predictions) are represented by the vector

.

= , , , , ,
4.2
Let us considering that all assumptions mentioned above are true, all sources of
systematic error have been minimized, and that all control variables have the same value
for the model and the experiment. If this is true, the experiments output vector ) (t y
equals the sum of the models output vector

and a stochastic error

=

+
4.3
The stochastic error is considered to have multivariate normal distribution with mean
zero, where

represents the variance covariance matrix of n measurements:

=

1,1
2

1,
2
⋮ ⋱ ⋮

,1
2

,
2

4.4
4.1.1 THE MAXIMUM LIKELIHOOD
Many methods for the quantification of the distance between model predictions and the
observations can be found in literature [88]. The most common equation to quantify the
residual is the Least Squares (LSQ).
= (

)
2

4.5
where

and

represent the vectors of the calculated output and of the
measured value respectively with length

.
The LSQ is a quadratic equation making it appropriate for most optimizers and easy to
compute. Nevertheless, it fails to consider important factors, as scaling multiple outputs
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


35

and considering measurement accuracy. For this reason, the Maximum Likelihood
(MXL) is preferred in process engineering:
=
1
2

, −

2

4.6
The variance-covariance matrix

serves two purposes. On the one hand the outputs
related to very noisy data has a smaller impact on the criterion. On the other hand,
different scales between the outputs are also weighted. For MR, the theory for
approximation of the confidence region for parameter estimation, which is based on the
MXL criterion, is also essential. The probability density function of the parameter set
and its correlation define the confidence region for parameter estimation. The size of
the confidence interval is a direct indicator of model identifiability.
4.1.2 MODEL IDENTIFIABILITY
As stated in chapter 3, the meticulous examination of model dynamics and relation
between states, allows a better understanding of the behavior of the model, and presents
alternative ways to obtain similar dynamics with simpler and more tractable models.
Nevertheless, the reduction of a model depends on its parameter set. For this reason
model reduction is closely related to parameter identification. In short words, the model
has to be correctly fitted before it can be reduced. Moreover, model reduction affects
the identifiability of the model. When a model is reduced some parameters are
eliminated. Depending on the correlation of the parameter set, the confidence interval
of the new parameters will be reduced. In other words, model reduction can also be
applied when complex models present low identifiability. At this point an important
question is to determine how to reduce a model correctly without knowing its exact
parameter values. Unfortunately, science has no explicit answer to this question.
In the previous section different methods to analyze and alter the structure of models
were discussed. However, it is of great importance to keep the objective of modeling in
mind. Since a model is used to mirror a physical system, the accuracy of the prediction
of the model needs to be quantified. To be more precise, the capability of the model to
describe the process behavior excluding the effects of uncertainties has to be evaluated.
Observations of the system to be described have to be made in order to test the
predictions of a model. These observations contain various sources of error. Summed to
human error, even most advanced measurement techniques present uncertainties and
produce data with deviations from the physical quantity. Therefore, it is crucial to
distinguish model prediction error (systematic error) from observation error (stochastic
error.
Optimal Experimental Design


36

For these reasons the effects of data variance on the probability distribution function of
the parameter set has to be considered. The structure of the model is responsible for the
propagation of data variance to the confidence interval of the parameters Figure 4.1. It
is the model itself which determines the level of accuracy obtained through parameter
estimation considering the quality of the data set. In other words, the quality of the
model can only be established after knowing how exact can the parameter set be
estimated. This again is directly dependent on the state of information. In conclusion,
not only the objective for the implementation of the model, but also the quality of the
experimental data have to be considered to select the optimal structure of the model.

Figure 4.1: Effect of sensitivities in parameter estimation accuracy. σP and σy represent standard
deviation of parameters and measurements respectively.
Asprey [89] defines identifiability as the quality of a model to present a monotonic
behavior. In other words each parameter set shows a unique combination of outputs.
To quantify model identifiability the objective function Φ

is maximized eq. 4.7 keeping
the difference of models output smaller than some threshold eq. 4.8.
Φ

= −

( −

)
4.7
s.t.

, − ,

, −,

=1
<

U t u e ¬ ) (

, , , , , , = 0
P i
U
i i
L
i
,..., 1 ; = s s u u u

4.8
σ
y
σ
y
σ
P
σ
P
high sensitivity
low sensitivity
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


37

P i
U
i i
L
i
,..., 1 ;
* * *
= s s u u u

where Φ

is the objective function to maximize, and

the parameter sets with


, and

an arbitrary weighting matrix.. is a vector of n
y
outputs, is a
vector with the derivatives of the state variables, is a vector with n
s
time-dependent
variables which define the system, a vector of

time-dependent input variables,
is a vector with

constant input variables, and the superscripts L and U represent
the lower and upper bound respectively.
The approach proposed by Asprey represents a high computational burden. In addition
global optimality should be guaranteed, which increases computation time even more.
Nonetheless, since identifiability studies are realized in the early stages of model
developments and discrimination, this approach might offer important reduction in later
experimental and design effort.
The standard procedure to increase mode identifiability is to detect the parameters with
small sensitivity and high correlation and take them out of the optimization variables.
The number of parameters to be estimated can then be iteratively increased as the
identifiability of the model increases. This procedure enables a more accurate parameter
estimation in early stages and thus a better design of experiments. Nevertheless, the
selection of these parameters, which are assumed to be highly correlated and with low
sensitivity, is carried out based on the Fisher Information Matrix (FIM), which is an
approximation based on first order derivative information. This very rough
approximation is only correct near the vicinity of the parameter values used to calculate
the FIM. Therefore, ideally, all parameters should be considered in the parameter
estimation.
4.2 THE FISHER INFORMATION MATRIX
4.2.1 THE CONFIDENCE INTERVAL
The propagation of data uncertainties through the model and its effect on parameter
estimation has been subject of short discussion in previous sections. When a model is to
be fitted to a data set, the accuracy of the estimation of the parameter vector depends
not only on the accuracy of the data set, which can be quantified by the variance-
covariance matrix, but also on the structure of the model. It is essential to understand
two characteristics of the model structure:
- The impact of each parameter on the dynamics of the model, therefore
its influence in the outputs.
Optimal Experimental Design


38

- The interaction between all parameters of the parameter set, hence how
they correlate to each other.
The most precise formulation of the problem is to calculate the probability density
function of the estimated parameter set.
Assuming that:
- All parameters are constant
- Input variables behold no uncertainties
- Stochastic error can be described by a normal distribution over model
predictions with the exact parameter set (which is unknown)
- There is no correlation between measurement uncertainties
- Measurement uncertainties can be described by a normal distribution
with mean equal to zero.
The probability density function of the estimated parameter set can then be estimated
with the following equation:

= 2
−2
exp−
1
2

,


,

2
C
y

=1

=1
det

()

1
2

=1


4.9
where ℘ represents the probability density function,

the estimated parameter vector,

,

and
,

the measured and the calculated variables respectively,

represents the
variance of the measurements, and

represents the variance covariance matrix of the
residuals.

-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
P1
P
2
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


39

Figure 4.2: Confidence interval from the Lin model,
obtained with Montecarlo simulation.
The most effective method to compute the real nonlinear confidence interval is to carry
out Montecarlo simulations [90]. Figure 4.2 shows the confidence interval of a two
dimensional parameter set. The model used for this example is a model for E. coli fed-
batch cultivations presented by Lin [91] (see Appendix A), the selected parameters are:

, maximal Substrate uptake capacity, and

, and Yield acetate to biomass. Even
though, for this calculation, white noise was added to the data set (simulated normal
distributed variance with zero mean), propagation of the uncertainties through the
model, results in a none normally distributed confidence interval of the parameter set.
This effect is caused by the nonlinearities of the model.
4.2.2 APPROXIMATION OF PARAMETER VARIANCE-
COVARIANCE MATRIX
As stated before, the accuracy of the parameter estimation depends on the sensitivity of
the objective function to changes in the parameter set. To be able to plan experiments
which reduce the confidence interval of the parameter, increasing the accuracy of the
parameter estimation, a cheap technique to compute the confidence interval is required.
Based on the Cramer-Rao Bound (CRB) theorem:


−1
(

)
4.10
where

and FIM are the variance-covariance matrix of the parameter set and the
inverse of the FIM respectively, and θ
es
the vector with P estimated parameters. Hence
we can calculate the lowest bound of the variance-covariance matrix in a cheap manner.
In practice, the FIM is computed to approximate the confidence interval and find the
optimal experimental set.

=
(, , )

,

−1

(, , )

,

=1

4.11
where

is the estimated parameter vector, is the state variables,

is the time point
of the measurement, is the control variables, and

is the variance-covariance matrix
of the measurements. In other words, FIM gives a linear approximation of the
confidence interval region of the parameter set to be estimated.
From equation 4.11 can be deduced, that the FIM depends on the control variables of
the experimental setup. We now have all the information we require to build an
optimization problem and find the optimal experimental setup which maximizes the
state of information obtained from experiments, an approximation to this can be
Optimal Experimental Design


40

calculated with the FIM. Finally, the FIM must be compressed into scalar form to create
an appropriate objective function for optimization. To achieve this, different criteria are
implemented in an effort to create an efficient measurement of the variance-covariance
matrix in scalar form. A graphical description of the most common criteria for the two
dimensional case can be seen in Figure 4.3.

Figure 4.3: Criteria for optimization [92]
Summarizing, the objective of design of experiments for parameter accuracy is to find
an experimental set which increases the information content of the data and hence
allows a more accurate parameter estimation. The most common criteria are shown in
Table 4.1
Table 4.1: Criteria for confidence interval quantification [92].
A- criterion
1

()
[93]

D- criterion
(
1

)

[94]

E-criterion max [95]
M- criterion max


1
2
[96]


4.2.3 LIMITATIONS OF THE FISHER INFORMATION MATRIX
It is essential to consider the limitations of the FIM at all time in order to avoid wrong
interpretation of the results or its misuse in experimental design. FIM is the result of
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


41

two important mathematic approximations. The validity of these approximations is
restricted to a small range near the region of calculation Figure 4.4 shows how the
nonlinearities of the model affect the shape of the confidence interval as it grows.

Figure 4.4: Shape of the confidence interval for different variance values from the Lin model
(appendix A). The confidence interval can be approximated by an ellipse near the exact value.
Once the variance of the data set has been quantified and set constant, the confidence
region of the intervals is determined by the objective function of the parameter
estimation program and its sensitivity to changes on its parameters. If we utilize the
MXL to estimate the residual of the models prediction, and the model is linear, the
resulting objective function will be a quadratic function. This means that we only need
to know the Hessian of MXL and the variance covariance matrix of the measurements
to calculate the exact confidence interval, again strictly for the case of linear models. On
the contrary, nonlinear models cannot be described so easily.

Figure 4.5: Objective function of a nonlinear model (appendix A) with respect to changes in a
two dimensional parameter set.
MXL
θ
1
θ
2
Optimal Experimental Design


42

Depicted in Figure 4.5. is the form of the objective function where the shape differs
from a parabolic function. Still, to avoid the expensive computation of the Hessian
tensor, which may be even impossible for large problems, the gradient information of
the state variables with regard to the state vector and to the parameter set is used to
obtain a fair approximation. This approximation, which as stated before is only accurate
near the vicinity of the exact parameter set, is the FIM.
4.3 MODEL DISCRIMINATION
In Engineering, the use of mechanistic models for the simulation of complex systems is
a widespread procedure [48]. Models are created on the basis of the physicochemical
phenomena occurring in the process. Rigorous models are, however, compared against
the so-called black box models laborious to develop, adapt, and fit.
The challenge increases when processes presenting a non-linear and dynamic behavior
are to be modeled. As a result, various models, which can differ in structure, size,
application and / or complexity, are usually to be found in literature. The questions that
arise are:
- How to quantify the capability of the different models to fit the process?
- How to select the model that best fits the needs of the process?
- How to create a minimal set of experiments to achieve question one and
two?
In statistics, the field concerned with finding the most appropriate model for a specific
process is called Model Discrimination (MD). MD deals with maximizing the probability
of a correct model selection considering data uncertainty. To achieve this, an optimal set
of experiments is planed with the aid of statistical tools. This approach is widespread in
biology and biotechnology [97-99], in physics [100] and in process engineering [101].
Solutions for the adaptation of linear and nonlinear systems and for the discrimination
of dynamic models [102-104] can also be found in literature.
The first work to address the problem of finding the most appropriate model to
describe a set of data was presented by Hunter [105]. Hunter defined the goal of MD as:
“Perform the experiment which will most strain the incorrect model to explain the data”. MD is
briefly presented in this chapter, for a better understanding of this approach we refer to
[52, 81, 106].
The general procedure for MD is rather intuitive. The first step is to select some models
which could describe the system, called candidates. Following, experimental setting is
designed in an effort to maximize the probability of choosing the correct model. To
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


43

achieve this, the experimental conditions which maximize the difference in the outputs
of each candidate are selected.
In statistics, Bayes‟ theorem calculates the precise probability function of a model being
the most appropriate to describe the data set. Nevertheless, it requires accurate
information of the distribution density function and only quantifies the probability
associated with a model prediction. The engineer needs to consider various factors
when selecting a model and requires an effective and simple method.
To state the problem we first discuss how to quantify the “goodness” of a model. This
is not an easy task. The obvious procedure is to calculate the residual . The residual
is composed of two different types of error.
= η +η
S

4.12
The stochastic error η is caused by the uncertainties in the system and measurement
methods and it should ideally be randomly distributed around the true value with mean
= 0 . The systematic error η
S
is caused by defects on the model, e.g. neglected
phenomena, not considered external disturbances, incorrect structure, etc.
In order to calculate which model describes the processes with highest certainty
possible, the variance of the data needs to be known and all models should have its
“exact” parameter set. Both errors, stochastic and systematic, are expressed by a
probability distribution function hence discrimination between models can only be
correct within a certain probability.
Furthermore, the state of information and the identifiability increase together with the
number of experiments. Therefore, the probability of selecting the “best” model
increases with each new experiment. For this reason, in order to make the most certain
decision, all candidates should be considered throughout the complete discrimination
process. However, this represents an enormous experimental effort [106]. Franceschini
[52, 81, 106] rather proposes to conduct a gradual discrimination of candidates in an
iterative process. This procedure reduces computation costs in each iteration and allows
a more specific design in the next iteration.
Table 4.2: Types of sum of square [22]
Error variance

2
=



Akaike‟s Final Prediction Error =

2

+ +1
− −1

Akaike‟s Information Criterion = ∗ ln

2
+ 2 ∗
Optimal Experimental Design


44

Shortest Data Descriptor = ln

2
+ +1 ∗ ln⁡()
Where is the Maximum Likelihood estimation eq. 4.6 value of the minimum sum
squared, the number of measurements, and the number of parameters.
Unfortunately, the residual only shows the ability of the model to describe the data set.
This is by far not the “best” model. Robustness, physical meaning, and extrapolation
capacity are a few examples of important model characteristics which should be taken
into account during MD.
Other forms of the sum of square intend to quantify further aspects of the models to
help proper discrimination. The most representative criteria are presented in Table 4.2.
The reader is referred to [107] for detailed summary.
Verheijen [22] proposes five selection criteria:
- Physicochemical criteria: physical meaning of the model
- Flexibility criteria: application in similar processes and reusability.
- Computational criteria: computation expenses and robustness
- Statistical criteria: accuracy of the prediction (level of state information).
- Engineering criteria: usefulness of the model in the real process
Still, he misses to mention the importance of identifiability as selection criterion.
Between two candidates with the same description accuracy, the model with the highest
identifiability should be selected.
4.3.1 MODEL DISCRIMINATION IN MECHANISM
RECOGNITION
For recognition purposes MD is a straight forward procedure, since calculating the
MXL is sufficient for selecting the best model. In MR, candidates have been generated
to detect different regimes. For this reason description accuracy is the only
discrimination criteria to consider. In addition, model distinguishability between the
submodels should be guaranteed beforehand (during the model building stage).
During the recognition process, the program for MR computes the minimal length of
the interval required to assure, firstly identifiability and next distinguishability. Model
distinguishability in each time regime depends on initial conditions and process state
information. An essential condition to apply MR to a process is to guarantee model
distinguishability in every interval. This has to be considered when building the
submodels.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


45

In this section the difficulties to discriminate between various models have been
discussed. From this, the complexity of model building can be deduced. A scientist
cannot be expected to build correct models, when he is not even able to select the best
among a group of models. In order to create an adequate model, the required
characteristics of the model and the objective of modeling have to be previously
defined.

To summarize, this section intends to emphasize two important aspects in modeling:
- Description ability is not the unique quality of a model worth
considering during model building. Many characteristics have to be taken
into account and there is never a simple manner to define the “best”
model.
- The state information of the system (whether laboratory, pilot plant or
real process) is essential for the selection of the optimal features of the
model. It is necessary to know the state of information of the data set
before or at list while the model is being built.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


47

5 CODE GENERATION, SIMULATION
AND OPTIMIZATION
In this project, various software packages and programming languages were used. Model
and code generation were supported by MOSAIC (section 5.1.1 ) and SBPD (section
5.1.2). Simulations were carried out with standard ODE solvers in the simplest cases
whereas the integrator sDACL (section 5.2.1) was applied for efficient integration and
sensitivity computation with piece-wise constant inputs and parameters. Finally,
different optimization algorithms (section 5.3) were required to solve the problems
addressed in this thesis.
5.1 CODE GENERATION
5.1.1 MOSAIC
MOSAIC [44, 108] is a modeling environment, which combines equation-based
modeling, use of symbolic mathematic language, and code generation, following a new
modeling approach for equation reuse and support of different nomenclature
conventions. The toolbox supports automatic code generation in a variety of
programming languages (C, Fortran, Matlab, among others) and stores the information
of the equation systems in the generalized markup languages XML and MathML.
The MOSAIC modeling environment is a tool to implement customized models aiming
at:
- the minimization of modeling errors,
- the minimization of the programming effort,
- the avoidance of errors and effort in documentation,
- the encouragement and support of cooperative work.
Furthermore, MOSAIC promotes modular modeling at a high level Figure 5.1. This
feature offers important advantages for hierarchical modeling as well as for efficient and
simple development and analysis of reduced models in general and submodels in
particular.
Code generation, simulation and optimization


48


Figure 5.1: High level modeling with MOSAIC [46]
5.1.2 SBPD
SBPD is a Systems Biology Toolbox for MATLAB developed [45]. The toolbox offers
an open and extensible environment for the analysis and simulation of biological and
biochemical systems limited to ODE systems. An important advantage is support of the
Systems Biology Markup Language (SBML) models, enabling model exchange within
the systems biology community. Models are represented in an internal model format
and can be described by entering biochemical reaction equations. The Systems Biology
Toolbox for MATLAB is open source and freely available from
http://www.sbtoolbox.org.
The toolbox is built in a modular way, as depicted in Figure 5.2. The base elements are
objects of classes SBmodel and SBdata, which are used to represent models and
experimental data.
Documentation-level modeling
Equations
Notation
Context
Model in XML/MathML
Matlab C/C++ BzzMath Aspen ACM CC UAM BzzMath
Fortran
Other languages
gPROMS
Equations
Notation
Notation Article
MOSAIC
Article
Overview
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


49


Figure 5.2: Modular structure of the toolbox. The toolbox is designed in a modular
way. Optionally, freely available third party software packages are used, e.g.
to (i) perform bifurcation analysis and (ii) realize an interface to SBML [45].
5.2 SIMULATION
This work shows that MR is a powerful tool to create simplified models and use them to
describe complex processes. Solving the set of differential algebraic equations accurately
even in the case of complex systems with highly nonlinear behavior is essential to enable
the application of MR. Furthermore, because of the calculation of the FIM and
optimization of nonlinear programs using optimality criteria, first and second order
dynamic sensitivities offer important advantages [46].
5.2.1 SDACL
Logsdon [109] showed the equivalence of the Orthogonal Collocation (OC) to an
implicit Runge–Kutta method, achieving the highest accuracy order. Also, being an
implicit numeric method, OC presents great stability properties for systems with index
one or higher. OC has shown very good performance especially when solving highly
nonlinear models. In addition, extending the method to the OCFE has shown
important reductions in computation expenses for the full discretization approach in
dynamic optimization, transforming dynamic equations into algebraic ones. Prove of its
efficiency solving typical process engineering problems is the important number of
solvers for general differential-algebraic optimization problems developed based on this
method [110]. OCFE may also be implemented on direct applying the partial
discretization approach to solve stiff implicit DAE systems [111]. Besides, solvers based
on OCFE have the ability of self-starting in high orders, an important characteristic for
efficient partial discretization approach especially when having a high number of
discontinuities or even multiple shooting.
For simulation in the MR program, the approach for large-scale dynamic models which
takes advantage of the sparsity of the system and generates first and second order
sensitivities is applied. The program sDACl developed by Barz [46] at the chair of
Code generation, simulation and optimization


50

process dynamics and operation of the TU-Berlin has been selected for the integration
of the DAE systems.
sDACL is a code for efficient integration of general DAEs with sensitivity generation.
Cheap computation is achieved through partial discretization tailored to the integration
of general DAEs based on OCFE. This algorithm provides exact sensitivity information
of the numeric integration, and can be implemented in any one-step integration method.
sDACL has been costumized to the staggered method for state and sensitivity
integration showing similar computational efficiency. Finally, the algorithm is able to
take advantage of the sparsitiy properties common in engineering models.
5.3 OPTIMIZATION
For the application of MR, optimization programs need to be solved including
parameter estimation, optimal control time and global optimization. The interface
environment Tomlab was utilized, to implement the different optimization programs
with the code written in Matlab®. Tomlab is an optimization environment containing
different state-of-the-art optimization software packages [112]. The external solvers are
distributed as compiled binary MEX DLLs on PC-systems, and compiled MEX library
files.
PARAMETER ESTIMATION
For parameter estimation purposes, the family of residual optimizers was selected in
general and the Gauss-Newton method in particular. Solving the parameter fit problem
with residual information has shown to be a very efficient method to overcome difficult
regions caused by parameter correlation [113].
NONLINEAR QUADRATIC PROBLEM
During the optimization of the optimal time lengths, the optimization problem was
formulated as solved based on Sequential Quadratic Problem (SQP) algorithm [114].
GLOBAL OPTIMIZATION
Finally, some problems require solution of high nonlinear programs where global
optimality is required. For these cases, the approach to stochastic optimization based on
Particle Swarm Optimization (PSO) was selected. PSO is a stochastic optimizer which
has shown advantages over other stochastic techniques [2,4]. The principle of PSO is
based on a population (swarm) formed by a certain number of particles (i), which search
for a global optimum in the defined region of n dimensions.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


51

6 AN APPROACH TO MECHANISM
RECOGNITION
6.1 A SHORT INTRODUCTION TO MECHANISM
RECOGNITION
The method for MR has been developed to recognize different regimes in dynamic
processes. In processes which have no steady state condition, the MD approach can be
applied to different regimes during the process. Similarly to steady state processes,
where a reaction pathway is selected among a number of candidates applying MD, in
dynamic processes, different regimes of the process with different pathways could be
detected. This is the basic principle of MR. In dynamic processes models which do not
fit the complete process data may still fit some time intervals. If the candidates set for
discrimination in each time interval are first principle models, the best candidate is also
an indicator of the predominating phenomenon in the selected regime. This approach
does not only enable a better understanding of the process and its dynamics, but also an
accurate description of complex dynamic processes with relatively simple models.
The ultimate goal of process modeling is maximization of prediction accuracy. To
achieve this, usually model complexity is increased exponentially with respect to
prediction accuracy. Drawbacks of this method are an unavoidable reduction of both,
parameter identifiability and model robustness. To overcome these problems,
experimental effort needs to be increased drastically and advanced simulation tools
(software and hardware) are required. In other words, increasing model complexity
cannot be achieved without increasing modeling and experimental effort exponentially.
Still, in order to describe the dynamics of a complex system, complex and highly
nonlinear models have to be developed.
Instead, dividing the process in different regimes with simpler dynamics and describing
them separately, allows process description with multiple but simpler models. Most
dynamic processes present two or more quasi steady state regimes, if these regimes can
be detected, isolated and modeled by different simpler models, the process is described
with remarkably less effort and higher accuracy. MR offers an adequate framework to
analyze experimental information and model dynamics. The models for MR need to be
An approach to Mechanism Recognition


52

built on the basis of process expertise, knowhow and also taking advantage of modern
mathematical tools. It is worth reminding that conclusions drawn by the recognition
program depend strictly on the expertise during submodel building.
Consider we have a model which can describe both phenomena in a batch reactor: 1)
diffusion and 2) reaction. This model would typically contain highly correlated
parameters, which would affect the accuracy of the parameter estimation. We can
reduce this model and create two submodels. A submodel 1, which can only describe
the change in concentration caused by diffusion, and a submodel 2, which is only
capable to describe concentration dynamics caused by reaction. Not only have both
submodels less parameters, but because of their limitations (only diffusion or only
reaction), a much more accurate detection of the regime where transport limits the
process and the regime where reaction limits the process can be achieved. In this
example, it would be possible to monitor the minimal speed of the stirrer needed to
avoid transport limitation. By these means, it is not only possible to describe the process
with simpler models and higher accuracy, but also information about the process is
gained.
At this point an approach to MR [42] to detect different dynamics based on simplified
models can be of great help. Complex models created to describe complete processes
are reduced to form submodels limited to the description of a specified phenomenon.
These reduced models, called submodels, have important advantages over the complex
model as are:
- Analysis of the complex model
- Because of the systematic mathematical approaches applied for
its reduction, submodels help analyze, understand and identify
the complex model.
- Simulation
- Reduced models (submodels) present a higher identifiability,
lower calculation costs and are more robust.
- Process understanding
- The physical basis of the complex model and of the submodels
allows a better understanding of the process and an easier
detection of the key dynamics of the system.
- Monitoring
- In conjunction with fault detection approaches, submodels help
detect the influence of non measurable phenomena in the
process.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


53

6.1.1 ILLUSTRATIVE EXAMPLE
To illustrate the concept of MR, a short example based on the first published
application of MR is presented. Further details of the process and recognition of
regimes can be found in [43, 115].
MR was applied to monitor fowling in Membrane Bioreactor (MBR) processes for waste
water treatment, which is a highly dynamic process [116, 117]:
- Backflush/relaxation is employed for approx.15 – 60 s every 3 – 10 min
of filtration
- Maintenance cleanings are carried out every 2 – 7 days and main
cleanings once or twice a year.
Various models for the description of membrane fouling can be found in literature.
Nevertheless, most models have been developed and validated under steady state
conditions. For the application of MR, five models obtained from literature were
selected [118] to fit experimental data obtained in a test cell and in a pilot plant:
- Cake filtration
∆ ∗ ∗

=
∗ ∗
2

+ ∗

= ∗ ∗

6.1
where ∆ represents the pressure difference, the time, the area, the filtered
volume, the dynamic viscosity, the specific cake resistance, is a dimensionless
constant, R is the resistance of the membrane (subscript M) and cake (subscript C)
respectively.
- Standard blocking

=
1

0
+



0

6.2
where

represents volume of solid particles retained per unit filtrate volume, the
membrane thickness,

the volumetric flow rate,

0
initial volumetric flow rate, and

0
the initial area.



An approach to Mechanism Recognition


54

- Intermediate blocking

=
1

0

+
∆ ∗ ∗


0

6.3
where represents blocked area per unit filtrate volume.
- Complete blocking

=
0

∆ ∗ ∗

6.4
- Cake filtration

0

−1

=

0



1

6.5
The results of model fitting in Figure 6.1 show, that none of the five models is able to
fit the data. Furthermore, if physical boundaries of parameter are considered, the
residuals obtained are extremely large. The physical explanation for the inability of all
models to fit the data is that dynamics of this particular process are governed by
different phenomena over time. The only solution to describe the process truthfully is
to create a complex model which encloses all phenomena: 1) cake formation, 2) pore
constriction, 3) pore blocking. In addition the model must foresee, the time periods
where each phenomenon is dominant.

Figure 6.1: Model fit a) without setting bounds
b) with setting bounds for physical parameters. [119]
MR offers an alternative and very practical solution to this problem. To avoid increment
on model complexity, the five models obtained from literature are used to describe
discrete intervals of the process. By these means, each model is applied only in the time
regime, where the model is able to describe the process. This allows not only a good
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


55

description of the complex process with simple models, but also enables an indirect
detection of the dominant phenomenon in each regime.

Figure 6.2: Comparison experiment/simulation using a) just one model. B) various models [119]
The program strictly detects the model which describes the present regime with the
highest accuracy. Nevertheless, if models enclose physical understanding of the process,
an indirect recognition of the dominant phenomenon is possible. In other words, MR
can use first principle models to make indirect measurements of non-observable
variables and bring new information to light. Figure 6.2 shows a regime detection with
mechanism recognition.

Figure 6.3 Cleaning strategy based on MR [43]
An approach to Mechanism Recognition


56

This ability to indirectly detect non-measurable variables can be exploited to monitor
processes and allow a better operation. For the case of the MBR process, a scheme for
action to take based on MR results was developed. This protocol facilitates a data based
process maintenance, which increases the efficiency of the process significantly. As seen
in Figure 6.3, the recognition approach can detect the fouling type and based on this a
cleaning strategy is selected.
6.2 METHODOLOGY FOR MECHANISM RECOGNITION
In the previous example models with only one variable defining the state of the system
are considered. This is a very practical condition since output input consistency between
all models is automatically guaranteed. Moreover, it was possible to measure the single
state variable at high frequency, which solves the problem of initial conditions for each
new regime. Models with different number and types of state variables can still be
applied for MR, but only in the special case where all models are observable (no general
structure is needed) and the frequency of the measurements is high enough to assure
continues initial value information for all state variables (no input-output consistency
required). The solution of this type of problems is straight forward and will not be
considered in this study. Unfortunately, in practice multiple state variables are to be
considered and online measurement information cannot be assured. For this reason an
approach to deal with these drawbacks needs to be developed. The method is required
to allow a regime recognition also in complex systems with multiple state variables and
limited process data.
The present study is constrained to following type of problems:
- Dynamic (non-stationary) process.
- Processes with two or more regimes.
- Detailed model available for the description of the general process of the
form:
- DAE system index zero or one
- no discontinuities
- known initial conditions
- Parameter boundaries have to assure global convergence with gradient
based optimizers.
- Initial regime is known.
- no model discrimination is required in the first interval.
- The sequence of the regimes is known.
- The minimal length of each time regime assures model distinguishability.
- The models present no bifurcation.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


57

6.3 PROGRAM STEPS
6.3.1 SUBMODELS
In order to assure a consistent process description throughout all regimes, generate a
proper general structure (section 6.3.2), and enable proper transfer of parameter values,
all models are derived from one single complex model, which aims at description of the
complete process. Model reduction techniques (chapter 3), empirical knowledge,
analogies to other models from literature, and the state of information obtained from
the process (chapter 4), are combined to create an adequate set of submodels. These
submodels should fulfill all the restrictions mentioned previously (input-output
consistency, parameter values transmission, general structure, etc) and be easily
distinguishable.
Unfortunately, most general model reduction approaches are restricted to systems where
Lyapunov stability can be proved. The development of a general method to reduce
nonlinear models and find adequate submodels is a very interesting challenge for
process modeling and novel promising approaches are already under study [73].
Nonetheless, a general method with reasonable computer expenses is still to be
developed. For this reason the reduction of the complex model into submodels is still
limited to specific solutions where modeling experience and process understanding are
essential. Phenomena delimitation and model adaptation for each regime is still a
difficult task. Much care has to be given to adequate description of the system
maintaining identifiability and distinguishability between models. Still, this thesis shows
the reduction potential of complex models and its advantages.
6.3.2 GENERAL STRUCTURE
It is essential to maximize submodel identifiability to increase the efficiency and
accuracy of the MR program. Since identifiability is deeply influenced by the number of
parameters to be estimated, it is convenient to distinguish the parameters in two groups:
- Process Constant Parameters (PCP)
- Regime-Wise constant Parameters (RWP)
In the reduction process to build the submodels, some parameters are present in all
submodels and remain constant throughout the whole process (PCP), e.g. conductivity,
limitation and inhibition constants. Since PCPs do not change over the process, the
state of information of the complete process can be used increasing the number of
measurements available for its estimation. This is not only important for PCPs, but also
RWPs can be estimated with higher accuracy. A very simple reformulation of FIM is
An approach to Mechanism Recognition


58

implemented to consider the difference in sate of information between PCPs and
RWPs.
The general structure is defined by the parameters and equations which remain
unchanged in all submodels. In the ideal case, all submodels consist of the general
structure and differ only in one parameter which is responsible for submodel
distinguishability. Such a scenario maximizes submodel identifiability, and is known as
piece-wise constant parameter modeling [120]. Also some very interesting studies have
been published for the case of hybrid systems [121]. Nonetheless, in order to confront
cases with multiple RWPs, and a general structure which does not include all the
equations of the submodels, MR is required.
6.3.3 SUBMODEL DISTINGUISHABILITY
Model distinguishability is strongly dependent on the modeler ability to create models
which enhance the differences between dynamics of different regimes. At this stage only
problems where two candidates are set to discrimination are addressed. In addition,
enough data information to assure discrimination with statistical meaning is to be
guaranteed. Furthermore, the order of the regimes is to be known and only the precise
conditions at which changes take place are unknown.
It is important to remind the reader that fast and accurate MD is key stone for process
monitoring. The amount of information required to discriminate between two models
depends on the candidates and on data quality. The best approach for model
discrimination and its ability to detect the next regime with the required promptness
needs to be investigated for each particular case. If the length of the interval selected for
MD is too short, false switching point detection may occur due to low identifiability. On
the other hand, if the interval is too long the sensitivity of MXL to the data points
indicating a regime change is decreased. Finally, it is important to consider the possible
estimation of initial values in every new regime.
In probability theory, distinguishability quantifies the probability of selecting the model
which best describes the observations. Once the certainty desired to consider the
discrimination between two candidates is fixed, the quality of the data set required to
distinguish between the model outputs can be calculated. The smaller the difference
between models outputs, the more precise need to be the measurements. In other
words, to assure distinguishability, the average of the square of the difference between
the outputs of the model, need to be at least larger than variance of the measured
variables. By using eq. 4.8 in compact form, it is possible to quantify the measurements
required to fulfill a specified threshold for model distinguishability

.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


59


1
,
1

2
,
2


1
,
1

2
,
2

=1

=

2

6.6
Where is a vector with

outputs of model 1 and 2 respectively, a vector of
u

time-dependent input variables, and the parameter set of model 1 and 2 respectively.
6.3.4 INITIAL INTERVAL
The initial interval is the process stage with less state of information. Because of the lack
of data at the beginning of the process, parameter values of previous runs have to be
selected as initial guesses. Luckily, in most of the systems, the first regime of the process
is known and cases where the initial conditions of a process are not well defined are
scarce in process engineering. Mostly, initial concentrations and state variables as
temperature and pressure can be defined at the beginning of the process. In addition
offline measurements can be carried out before the process is started to increase the
information at the beginning of the process. The application of MR with uncertain
initial conditions requires addition of backward propagation calculations to define the
initial states and is not considered in this work.
During the initial interval, PCPs need to be defined and the boundaries of the RWPs
have to be set. In biological systems for example, maximal reaction rates and uptake
capacities determined in the initial interval, define the boundary conditions of the
successive parameter values throughout the complete process. Hence, it must be
assured, that the initial regime offers enough information to identify the general
structure of the submodels. This might affect process efficiency since the initial regime
could be artificially kept active only to allow identifiability affecting optimal process
conditions. Still, the information gained in the initial interval is essential to assure a
correct recognition of the process. Again, the effort of this step is reduced as experience
and information are accumulated. Furthermore, many processes show great
reproducibility and the parameter values of previous runs can be applied to reduce the
required length of the initial interval.
6.3.5 MR INITIALIZATION
The recognition process may only be started after the parameter values of the general
structure have been estimated and boundaries, as well as linear and nonlinear constraints
have been defined. First, the FIM is tested for identifiability.
det ≥ 0
6.7
Secondly, once the state of information is enough to assure a trouble-free inversion of
the FIM, the computation of the some confidence interval criterion is started.
An approach to Mechanism Recognition


60

For both case studies presented in chapters 7 and 8, the A criterion (

) was selected
to quantify the state of information (section 4.2.1). In short words,

represents the
average variance of the confidence interval considering all unknown parameters. The
boundary for

to assure model distinguishability is calculated during the tests for
model distinguishability (section 4.3). The value obtained at this stage is used to select
the time point from which a regime switching point can be detected.

6.8
where

represents the criterion boundary obtained from the distinguishability test.
Much care should be taken not to initiate the switching point detection before the state
of information has been tested and accepted. Once the criterion boundary has been
reached (e.g. the number of measurements is enough to make a distinction between the
model outputs), the detection of the switching point may be initiated (section 6.3.6).
It is worth reminding that an important assumption is that the general structure of the
model is globally correct and structure parameters do not need to be updated. This
assumption is essential to assure consistent evaluation of piecewise constant parameter
identifiability and model distinguishability.
6.3.6 DETECTION OF SWITCHING POINTS
The position in time, where changes on process conditions induce a change in the
regime of the system, is defined as switching point. The accurate detection of the
switching points is essential for two reasons:
- fast and accurate detection of regime change
- avoid undesired or dangerous process conditions
- accuracy of initial values for the next interval
- obtain the correct initial conditions of the non measurable
variables
Once adequate submodels have been developed and the state of information has been
analyzed, detection of the switching points is no different from fault diagnosis based on
first principle models. Although not necessarily representing a fault on the system, from
a mathematical and physical point of view, the program is required to detect an irregular
behavior of the system in comparison to some model predictions.
Methods for model-based residual generation, offering so called analytical redundancy
have been widely discussed in literature [122]. Typical methods include, on the one side
state estimation techniques (the parity space approach, observer based schemes, and
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


61

fault detection filter approach) and parameter estimation techniques on the other
(parameter identification). In this work we concentrate on the application of parameter
identification methods due to the common theoretical background studied in design of
experiments. Still MR is not limited to parameter estimation methods and the
application of other faster methods like direct evaluation of the objective function offer
a fast and fairly accurate result.
Following conditions must be considered during submodel building since they are
essential to assure detectability and distinguishability of the switching point during the
process [123]:
- knowledge of the normal behavior of the system
- distinguishability of the changing behavior
- availability of at least one observation reflecting regime change
- satisfactory state of information
The evaluation of the validity of the present regime given by the submodel is divided in
two main steps:
- residual generation
- decision of the switching point (time and type of the new interval)
To increase the probability of correct detection of the switching point, the method for
robust parameter estimation [123] is proposed. In order to decide whether the process is
running in the present regime or a change is taking place, a residual calculation is carried
out to select one of the following hypotheses:

0
:

=


1
:

6.9
where

and

represent the nominal parameter set and the estimated parameter
respectively, hypothesis zero assumes there is no regime change whereas hypothesis one
assumes the switching point has been reached.
To enable the evaluation of the hypothesis considering the uncertainties of the
observations a residual between

and

is calculated taking the variance-covariance
matrix of the expected parameter set into account:
=

−1

6.10
where is the residual, and

is the variance-covariance matrix of the parameter set.
We can also take advantage of the previous calculation of FIM (section 4.2) and
substitute

by its approximation, the resulting equation is:
An approach to Mechanism Recognition


62

=

6.11
Finally, a fixed threshold

is selected to accept or reject hypothesis zero. Although
variable threshold approaches have shown to be more efficient [124], thanks to the
implementation of FIM in the equation, the residual also reflect the actual state of
information with respect to the parameters. The variation of the threshold is substituted
by the variation of FIM, which gives a more accurate description of parameter
covariance in the sense that it also considers information from previous intervals PCPs.
6.3.7 INITIAL CONDITIONS OF THE INTERVAL K+1
It is essential to take into consideration that, in real processes, regime changes do not
happen instantaneously. Although in many cases the difference in dynamics allows to
consider changes in process conditions as instantaneous, the transition phase between
two regimes must be considered. This is of relevance for the calculation of the
successive regime (k+1), due to the fact that the initial conditions of the regime (k+1)
are determined by the endpoint outputs of the previous regime (k). To increase initial
condition accuracy and reduce error bias, offline measurements to redefine the state of
the process should be carried out whenever possible. In the cases where no
observations are available, the initial values of the next interval need to be estimated. It
is clear that this represents an addition of parameters to be estimated reducing the
identifiability of the problem drastically and tight boundaries of the initial values are
required minimize the impact of these new variables in the complexity of the
optimization problem.
To find tight boundaries the switching interval can be simulated with both models to
create an envelope by selecting the maximal and minimal values for the upper and lower
bounds respectively.
6.3.8 DETECTION OF THE NEXT SWITCHING POINT
Once the best candidate has been selected, the initial values and parameters have been
estimated, and the arc in the switching interval determined, the algorithm returns to the
search for the next switching point (return to 6.3.6). This procedure is repeated until the
end of the process is reached.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


63

6.3.9 FLOW DIAGRAM
Following, the steps of the recognition strategy are presented accompanied by a flow
diagram Figure 6.4. The flow diagram proposed has been built for the special case of
batch processes, but it can still be applied to different cases with slight modifications.
The diagram is divided in three phases:
- Submodel Building
- Initial Interval
- Regime Recognition
The first phase requires the collaborated work of process and model experts to create an
adequate set of submodels (section 6.3.1) and cannot be carried out automatically since
there is no general solution to proper submodel building. In this stage the models are
tested for identifiability considering the state of information obtained from the process
and submodel distinguishability.
Once the process has started, the model describing the first regime is tested for
identifiability to determine the minimal length required to perform reliable parameter
estimations. The parameter set of the general structure is estimated and the boundaries
of the RWPs are set. This is also the most critical phase of MR since the nominal values
of the parameter set and the general structure are defined.
Finally, with the general structure well defined and the boundaries set, the detection of
the switching point is initiated. The algorithm applies one of the switching point
detection criteria (section 6.3.6) to select the time point where the process changes from
one regime to another. The third phase is recursively repeated until the end of the
process is reached.


An approach to Mechanism Recognition


64


Figure 6.4: Flow diagram of MR algorithm

success
Complex
model
Submodel
n
Submodel
1
Data set
Data
analysis
Variance-
covariance
matrix
Jacobi
matrixes
Identifi
ability
test
Distin-
guisha-
bility
test
failed
failed
success
General
structure
Parameter
estimation
FIM
calculation
Switching
point
detection
k =k+1
Regime
type
General
structure
FIM
t = t
end
?
print
Parameter
estimation
I Submodel
Building
II Initial Interval
III Regime
Recognition
Flow diagram for
Mechanism Recognition
No
Yes
Identifi-
ability
reached
?
Interval
increment
No
Yes
General
structure
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


65

7 MECHANISM RECOGNITION
APPLIED ON SEQUENCING
BATCH REACTORS
7.1 INTRODUCTION
7.1.1 ACTIVATED SLUDGE
The most applied method for biological treatment of waste water is the Activated
Sludge Process (ASP) [125]. In ASP, carbonaceous organic matter of wastewater (readily
biodegradable as well as particulate substrate) provides an energy source for bacterial
growth of a mixed population of microorganisms in an aquatic aerobic/anoxic
environment. The microbes consume carbon and oxidized end-products that include
carbon dioxide and water. In addition, microorganisms may obtain energy by oxidizing
ammonia nitrogen to nitrate nitrogen in the process known as nitrification. Moreover,
some microorganisms are able to reduce nitrite and nitrate in the anoxic phase as well,
this process constituting the second part of the so called nitrification-denitrification
process.
The family of Activated Sludge Models (ASM) represents the state-of-the-art model
framework for ASP simulation [126]. ASM1 is the most widely used in practice [127],
ASM2 [128] is applied to simulate processes that include biological phosphorus
removal [78] and the latest version, ASM3 [129], includes the quantification of energy
storage in order to describe substrate and oxygen uptake with higher accuracy. Finally, a
newer version of ASM3, referred to in this contribution as extended ASM3, where
nitrification and denitrification are considered as two-step processes taking into account
nitrite as an intermediate, has recently been presented [130].
In order to extend the ASM3, 7 process equations were included, resulting in a
stoichiometric matrix with 15 state variables and 20 reaction equations. In addition to a
low parameter identifiability, the computation of the extended ASM3 is significantly
expensive. These drawbacks represent the main obstacle for efficient optimization and
Mechanism Recognition applied on Sequencing Batch Reactors


66

model-based control. On the other hand, the extended ASM3 describes many states,
which are not to be considered in SBR process [131, 132].
7.1.2 SEQUENCING BATCH REACTOR
Continuous Waste Water Treatment plants (WWTP) offer important advantages in
comparison to batch processes. The most relevant are; lower energy consumption and
operation costs, higher load capacity, and process control stability. Nevertheless,
continuous processes confront a major limitation when dealing with an intrinsic
characteristic of wastewater treatment processes. Wastewater naturally presents
stochastic variations in both time and space. In case of large WWTPs, these variations
can be buffered to a certain level by increasing the size of the tanks, or by introducing a
flow-stabilization tank upstream. However, small and medium size WWTPs are strongly
affected by changing concentrations of contaminants in the wastewater [133]. SBR,
because of its high flexibility and operation range, offer an adept solution for such cases
[134].
Furthermore, an SBR is able to treat different kinds of effluents such as municipal,
domestic, hypersaline, tannery, brewery, dairy wastewaters, and landfill leachates among
others [134]. One or more stirred batch reactors with an aeration system are the basis
for this kind of process. In an SBR, the retention time, the duration of the aeration and
anoxic phases, the settling time, and other conditions can be fitted to a changing quality
of load as well as effluent requirements. The SBR can also be considered to be a process
which operates with variation in time, whereas a continuous process operates with
variations in space [135]. A complete cycle of the sequencing batch process consists of
five steps; 1) Idle, 2) Fill, 3) React, 4) Settle and 5) Draw Figure 6.4.

Figure 7.1: SBR cycle [136]
idle fill
react
settle
draw
1 2
3
4
5
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


67

The SBR has gained great popularity in recent years. Advances in process measurement,
as well as automation and control, increase the process efficiency, reducing operation
costs, while fulfilling the strict environmental regulations [137].
Nowadays, SBR technology has become important in particular for small and medium-
sized WWT Plants. When properly designed and operated, an SBR also offers a process
with important advantages over continuous processes, not only because of its efficiency
and economical aspects [138-140], but also because of its small footprint [135].
Besides many efforts to monitor nutrient removal [78, 141-144] in SBR processes, direct
online monitoring of Biological Oxygen Demand (BOD), Chemical Oxygen Demans
(COD), Nitrites and Nitrates as well as ammonia is difficult inaccurate and costly. [143].
Indirect methods are required to determine these concentrations for control purposes.
7.1.3 NITRATE BYPASS GENERATION
In the ASP, nitrogen is removed from wastewater by the nitrification/denitrification
process. Most of the nitrogen contained in wastewater is converted into ammonia.
Ammonia is then converted into molecular nitrogen by a two-step biological processes,
namely nitrification followed by denitrification (Figure 7.2).

4
+

2

3

2

2

3


Figure 7.2. Nitrification-denitrification process described as a two -step reaction.
In the first stage, Nitrosomonas and other ammonia oxidizers convert ammonia and
ammonium to nitrite, whereas in a second stage, Nitrobacter and other nitrite oxidizers
finish the conversion of nitrite to nitrate.
Turk and Mavinic [145] proposed the Nitrate Bypass Nitrification-Denitrification
(NBND) process, which can be achieved by inhibiting the production of nitrate and
proposed various methods for bringing about this effect. Katsogiannis et al [146]
showed that a frequent enough change between aerobic and anoxic conditions
suppresses nitrate formation. Ammonia is converted to nitrite in the presence of
oxygen (nitritation), which is then converted into nitrogen under anoxic conditions
before the second oxidation producing nitrate (nitratation) can take place. The NBND
has the following advantages over conventional nitrification-denitrification [145]:
- 40% reduction of COD demand during denitrification
- 63% higher rate of denitrification
- 300% lower biomass yield during anaerobic growth
- no apparent nitrite toxicity effects for the microorganisms in the reactor
Mechanism Recognition applied on Sequencing Batch Reactors


68

7.1.4 MONITORING OF WASTEWATER PROCESSES
Online monitoring of WWTPs is a challenging task. Variables measured in standard
WWTPs are pH, DO, respirometry, redox capacity and titrimetry, among others [147,
148]. Despite the important advances achieved in recent years, modern measure systems
still require laborious maintenance, frequent recalibration and important know how. In
addition, environmental regulations demand better process and output quality control at
low treatment cost and energy efficient process conditions. The implementation of new
approaches to enable an accurate online monitoring of key species like organic matter is
of great relevance in modern water treatment technology. Model based control and
software sensors offer interesting possibilities for the maximization of process
efficiency.
The Activated Sludge Model No. 3 (ASM3) extended for two-step nitrification and
denitrification has proven to be a very accurate model to describe ASP [149]. The
division of the nitrification-denitrification reaction in a two-step reaction is essential
when trying to describe the bypassing nitrate generation process in the ASP [146].
Moreover, the extended ASM3 enables the description of the substrate consumption
and the oxygen uptake with a higher precision than the older versions of the ASM
family because of the addition of energy storage effects. This model extension facilitates
the calculation of the NO
2
concentration as an independent variable.
In this section, diverse approaches to model reduction (chapter 3) are applied in order
to develop submodels of the ASM3 while maintaining their characteristic dynamics in
each regime. This study is based on the analysis of the process conditions and the
available empirical knowledge. The first submodel proposed is named 5state, according
to the number of the ordinary differential equations which need to be solved.
The model reduction is based on the principle that an batch cycle should stop once all
concentrations comply with the regulations. Furthermore, these concentrations should
be reached in the shortest time possible. Consequently, lower output concentrations
than required are an indicator of a suboptimal operation and its accurate detection
increases process efficiency. Another important assumption is that the bacteria never
exhaust their stored energy. Except for the recycle process, the bacteria are always in a
medium, which is rich in substrate. Therefore, the stored energy value should be
permanently high during the process and never limit bacterial growth.
7.2 SUBMODEL BUILDING
Following the systematic approach to MR, the submodels are created considering
process conditions and the characteristics of the complex model to enable a fast
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


69

description of the different regimes of the process Figure 6.4. The recognition step
should identify two distinct regimes:
1. Regime with readily biodegradable substrate in the medium
2. Regime where organic matter has been depleted
In SBR processes, it is of interest to stop the process once the concentrations of the
medium fulfill environmental regulations. An accurate detection of depletion of organic
matter offers important improvements in process efficiency. For this reason, the
submodel is built to describe the process regime under high substrate concentrations
aiming at detection of substrate depletion.
7.3 A PROPOSED 9STATE MODEL
In the first reduction step, the model proposed is named 9state, according to the
number of state variables contained in the equation system. Thus there are nine
differential equations which describe the basic variables (concentrations), namely: 1.-
carbonaceous substrate, 2.- heterotrophic bacteria, 3.- ammonia oxidizers, 4.- nitrite
oxidizers, 5.- dissolved oxygen, 6.- ammonia, 7.- nitrite, 8.- nitrate and 9.- stored
substrate.
7.3.1 STORAGE
The implementation of energy storage represents the principal improvement of ASM3
in comparison to older versions. Neglecting this equation would impede a proper
process description and would result in an incorrect model reduction. Consequently, the
storage of substrate and its effects on substrate and oxygen concentration cannot be
ignored. For this reason, the proposed 9state model includes some adaptations to the
substrate and oxygen uptake equations. The new set of equations is presented in such a
way, that both substrate uptake and oxygen uptake increments caused by the storage are
now included in the original substrate and oxygen differential equations.
By this means, the relation between both, substrate consumption rate and oxygen
consumption rates, to energy storage are linear, which is accurate as long as the
substrate concentration is above zero. This could appear to be an inconsistent
assumption, though it is almost certain that the process continues after substrate
elimination in order to achieve further ammonia degradation. Nevertheless, previous
model versions (ASM1 and ASM2) fit the data although they lack a storage variable. In
other words, the 9state model responds to the substrate limitation similarly to ASM1,
but describes substrate and oxygen uptake as well as energy storage as precisely as the
extended ASM3.
Mechanism Recognition applied on Sequencing Batch Reactors


70

ASM3 considers that bacteria store some of the substrate they consume under high
substrate concentrations for its later use under substrate limiting conditions. This
storage is not only responsible for lower biomass growth under equal substrate
consumption, but also for the bacterial growth when no substrate is present in the
medium. In the case of an SBR process, it is considered that storage of substrate does
not limit bacterial growth. The assumption is based on the fact that in a process with
optimal aeration strategy, bacteria never exhaust their stored energy for two reasons:
- In order to minimize process time and costs, the environmental
regulations should be fulfilled as soon as substrate is consumed.
- Excepting the idle phase, bacteria are always in a medium rich in
substrate. Therefore, the value of the stored energy should be high at
any moment during the process and never limit bacterial growth.
If these assumptions are valid, the relation between the substrate used for storage and
the substrate used for growth is valid as well. In the extended ASM3, the relation
between substrate and biomass growth is described by a second order differential
equation. However, as long as the concentration of the stored energy is high, this
second-order differential equation can be accurately approximated with a first order
differential equation.
7.3.2 REDUCTION OF THE EXTENDED ASM3 MODEL TO A
9STATE MODEL
In this work, the 15 ordinary differential equations of the extended ASM3 include only
the process rate variables without their explicit equation. The process rate equations are
shown in Table 7.1and have been numbered in the same order as previously presented
in [130]:
Table 7.1: Reaction rates of the extended ASM3
Heterotrophic Organisms:
r1: Hydrolysis
r2: Aerobic Storage
r3: Anoxic Storage
r4: Anoxic Storage of carbonate substrate, NO2–N2
r5: Aerobic Growth of Heterotrophic bacteria
r6: Anoxic Growth NO3-NO2
r7: Anoxic Growth NO2-N2
r8: Aerobic Endog. Resp. of Heterotrophic bacteria
r9: Anoxic Endog. Resp. NO3-NO2
r10: Anoxic End. Resp. NO2-N2
r11: Aerobic Resp. of particulate storage
r12: Anoxic Resp. of particulate storage, NO3-NO2
r13: Anoxic Resp. of particulate storage, NO2-N2

Ammonium Oxidizing Bacteria (AOB):
r14: Aerobic Growth, Nitritation
r15: Aerobic End. Resp.
r16: Anoxic End. Resp.

Nitrite Oxidizing Bacteria (NOB):
r17: Aerobic Growth, Nitratation
r18: Aerobic End. Resp.
r19: Anoxic Endog. Resp

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


71

7.3.3 MATHEMATICAL REPRESENTATION OF THE 9STATE
MODEL
ORDINARY DIFFERENTIAL EQUATIONS
The 9 ordinary differential equations are shown in equations 7.1 - 7.9. Their
corresponding rate equations are described in 7.10 - 7.14. All the constants K.x have the
same value as published in the extended version of ASM3 [130]. The values used for the
saturation constants are shown in Table 7.2.

dS
S
dt
= −
1
Y
Haer
∗ r
aae

1
Y
Hanox
∗ r
aNO 3
+r
aNO 2
∗ 1 +St
S

7.1

=

+
3
+
2

7.2

=

7.3

=

7.4

=
7.5

4

= −−

+


1

1
+

−−

+


3
−−

+


2

7.6

2

=
1

1


1

2

+
1 −

1.14

(
3

2
) 7.7

3

=
1

3


1 −

1.14

3
7.8

= −
1


1


3
+
2
∗ (−

) 7.9


S
S
Readily biodegradable substrate
X
H
Biomass of heterotrophic bacteria
X
Ns
Biomass of Nitrosomones
X
Nb
Biomass of Nitrobacter
S
O
Oxygen concentration in the medium
Mechanism Recognition applied on Sequencing Batch Reactors


72

S
NH4
Ammonia concentration
S
NO2
Nitrite concentration
S
NO3
Nitrate concentration
X
Sto
Energy Storage
Y
Haer
Yield coefficient of S
S
to X
H
in aerobic conditions
Y
Hanox
Yield coefficient of S
S
to X
H
in anoxic conditions
Y
A1
Yield coefficient nitrite to nitrosomones in aerobic conditions
Y
A2
Yield coefficient nitrite to nitrobacter in aerobic conditions
Y
A3
Yield coefficient nitrate to nitrobacter in aerobic conditions

Maximal growth of Heterotrophous

1
Maximal growth of nitrosomones

2
Maximal growth of nitrobacter

1
Maximal growth of heterotrophic bacteria on nitrire

2
Maximal growth of heterotrophic bacteria on nitrate

Substrate to storage constant

Oxygen to storage
Limitation constant

REACTION RATES
Heterotrophic growth on aerobic conditions

=

+

+
1

4

4
+

7.10
Growth of nitrosomones

=
1

+

4

4
+

7.11
Growth of nitrobacter

=
2

2

2
+
21

+

4

4
+

7.12
Heterotrophic growth on nitrate

3
=
1

+

3

3
+
3

21

21
+

4

4
+

7.13
Heterotrophic growth on nitrite

2
=
2

+

2

2
+
2

22

22
+

4

4
+

7.14

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


73

Table 7.2: 9state model constants and its values as shown in the Matlab code

7.3.4 STOICHIOMETRIC MATRIX
The stoichiometric matrix of the 9state model is presented in Table 7.3. This matrix
represents the ( + −) matrix of equations. In other words, the analysis of the
stoichiometric matrix gives us the information about the number of reaction invariants
hidden in the 9state model.
Table 7.3: Stoichiometric matrix of the 9state model

SOStar = 7; [mgO2/l] process K_NH2 = 0.1; [mgCOD/l] ASM3
K_La = 1000; [d-1] process K_S = 10; [mgCOD/l] ASM3
i_NB = 0.086; [gN/gCOD] fitted K_S1 = 0.1; [mgCOD/l] ASM3
mou_H = 0.6021; [d-1] fitted K_S2 = 0.1; [mgCOD/l] ASM3
mou_A1 = 0.6552; [d-1] fitted K_NHH = 0.05; [mgN/l] ASM3
mou_A2 = 0.3468; [d-1] fitted K_O1 = 0.2; [mgO2/l] ASM3
Y_Haer = 0.1302; [gCOD/gCOD] fitted K_NH = 0.1; [mgN/l] ASM3
Y_A1 = 0.1327; [gCOD/gN] fitted K_O = 0.8; [mgO2/l] ASM3
Y_A2 = 0.0985; [gCOD/gN] fitted K_NO21 = 0.5; [mgO2/l] ASM3
Y_A3 = 0.0331; [gCOD/gN] fitted K_NO3 = 0.5; [mgN/l] ASM3
i_NSS = 0.01; [gN/gCOD] ASM3 K_O21 = 0.2; [mgO2/l] ASM3
Y_Hanox = 0.0632; [gCOD/gCOD] fitted K_NO2 = 0.25; [mgN/l] ASM3
mou_H1 = 0.0511; [d-1] fitted K_O22 = 0.2; [mgO2/l] ASM3
mou_H2 = 0.0362; [d-1] fitted stS = 1.7; [ ] fitted
K_NH1 = 0.01; [mgCOD/l] ASM3 stO = 0.08; [ ] fitted


4

2

3


1 +St
S

1 0 0 −
1 −

0 0
St
S

0 0 1 0 1 −
3.34

1

1

1
+

1

1
0 0

0 0 0 1 1 −
1.14

2


1

2

1

3
0

2

1 +St
S

1 0 0 0


1 −

1.14

0
St
S


3

1 +St
S

1 0 0 0

1 −

1.14


1 −

1.14

St
S

Mechanism Recognition applied on Sequencing Batch Reactors


74

7.3.5 LIMITATIONS OF THE REDUCED MODELS
It should be noted that the proposed 9state model is not valid for the whole range of
conditions as the extended ASM3. Some limitations are to put up with in order to
reduce the model and speed up the simulation in the region of interest. Moreover,
because of the new storage equation, the bacteria can store energy, but cannot use it
when there is no more substrate available. This feature converts the 9state model in a
perfect indicator of substrate depletion in the reactor.
In the extended ASM3, the ammonia concentration does not limit the energy storage.
This results in a consumption of substrate even under ammonia limitation. Once again,
because of the coupled equations, the 9tate model predicts substrate consumption only
as long as ammonia is present in the medium. Finally, the growth of heterotrophic
biomass can be mathematically described as a second order differential equation. For
this reason, if the energy stored by the bacteria is low, a time delay can be seen in the
growth curve. This time delay is not predicted by the 9state model. Taken into
consideration that the storage has a value at least larger than 100 gCOD/m
3
, both
growth curves match.
7.4 A PROPOSED 6STATE MODEL
The 9state model involves 9 differential equations, with only one of them depending on
the process input (oxygen supply). The reaction invariant theory can then be applied to
the 8 differential equations, which are not affected by the oxygen input, i.e. to 7.1 - 7.4
and 7.6 - 7.9.
These equations are of the form of eq. 2.6 with = 5, + = 8. Therefore, it is
possible in this case, to find + − = 3 linearly independent reaction invariants.
For example 7.15 - 7.17,

4
+

1 +

+

+

+

+

1
7.15

2
+2 ∗
3
+
1 −

1.14

1 +

+

1

2 ∗
2

3

2

3

7.16

1 +

+

7.17
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


75

The above quantities remain constant throughout the entire batch. Therefore, it is
possible to substitute three differential equations 7.6, 7.7 and 7.9 for three algebraic ones
7.18 - 7.20.
where C
NH4
, C
NO2
and C
XSto
are the respective constants obtained after solving
equations 7.15 - 7.17 for the initial conditions. Based on the method of reaction
invariants, we find three equations which are linearly dependent. Therefore, the 6state
model mimics the 9state model perfectly well.
7.5 A PROPOSED 5STATE MODEL
Taking a closer look at the growth of the biological matter during one SBR cycle, we can
see that the overall change in biomass does not exceed 10%. Based on this observation a
further reduction of the mode is possible for the special case of a batch process. We can
eliminate these three differential equations 7.2 - 7.4 from our 9state model and consider
a constant biomass concentration throughout each cycle of the SBR so as to obtain a
model with 6 differential equations. Again it can be shown that this system has one
reaction invariant which is the same as indicated for the case of the 9state 7.20. As a
result we finally obtain a model with only 5 state equations. This 5state submodel is
almost as accurate as the 6state model. However, it can only be applied for batch
processes where the change of biomass can be neglected.
7.6 RESULTS
The submodel was set to various conditions so as to confirm their stability and accuracy.
The most representative results obtained after the comparison of ASM3 with the 5state
model are presented in Figure 7.3-Figure 7.6.
The simulation represents a batch tank ideally mixed, where the aeration can be turned
on and off in order to induce either aerobic or anoxic conditions. The initial value of the
storage is set to 400 gCOD/m
3
. This assumption is based on the fact that the sludge is
obtained from previous cycles and bacteria have already stored energy.

4
=
4

1 +

+

+

+

+

1
7.18

2
=
2
−2 ∗
3

1 −

1.14

1 +

+

+

1

2 ∗
2

3

2

3

7.19

=

1 +

7.20
Mechanism Recognition applied on Sequencing Batch Reactors


76

The system of DAE is solved with two different integrators:
- sDACL: An in house tool developed to solve DAE systems with OCFE
able to also calculate dynamic sensitivities [46].
- ODE15s: An integrator based on the numerical differentiation formulas
(NDFs) [150].
The main purpose is to set both models to drastic changes and various limitations. The
aeration is turned on and off intermittently to produce a strongly dynamic process. As a
result, a constantly changing process is obtained, which makes it very difficult to be
described identically by two models with different characteristics. The process
conditions prove that the 5state model describes accurately the limitations of dissolved
O
2
, NO
2
-
and NO
3
-
Figure 7.3 shows the results for substrate and storage. The initial value of storage was
set to 400 gCOD/m
3
as explained in section 4.2. All reduced versions describe perfectly
the behavior of storage even doe it is calculated by an algebraic equation. This proves
that the substitution of the differential equation for energy storage by an algebraic
equation does not affect the dynamics of the model.
As seen in Figure 7.4, considering constant biomass values throughout the process does
not affect the results of the other state variables. Even though growth rate is a crucial
parameter of the process biomass concentration, changes in batch processes can be
neglected.

Figure 7.3. Substrate concentration SS and
stored energy Sto against time.

Figure 7.4. Biomass against time. Changes in
the biomass are very small (less than 10%).
0 2 4 6 8 10
200
300
400
500
600
700
800
900
1000
t [h]
S
S
,
S
t
o

[
g
C
O
D
m
-
3
]


5state
ASM3
Sto
S
S
0 2 4 6 8 10
500
510
520
530
X
H

[
g
C
O
D
m
-
3
]
a)


5state
ASM3
0 2 4 6 8 10
98
100
102
104
b)
t [h]
X
NS
X
NB
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


77


Figure 7.5. NOX concentration against time.

Figure 7.6. a) Oxygen concentration in the
medium against time.
b) Ammonia concentration against time.
The concentration of NO
2
and NO
3
are kept under 20 mgN/L. The control variable is
the aeration of the tank. Figure 7.5 shows how precise the reduced model describes the
curves of nitrite and nitrate even in such drastic conditions. Finally, the results for
oxygen and ammonia calculations can be seen in Figure 7.6.
7.6.1 SIMULATIONS RESULTS
The most representative results of the comparison of both models simulated in
Matlab® are presented in Table 7.4 and Table 7.5.

Table 7.4. Comparison of the computation time.

The reduced models are up to one order of magnitude faster. For the calculation of the
dynamic sensitivities, the Jacobian matrixes have to be solved together with the equation
system. Table 7.5 shows the difference of calculation time for the three models.



0 2 4 6 8 10
0
2
4
S
N
O
-
2

[
m
g
N
L
-
1
]
a)
0 2 4 6 8 10
0
2
4
6
8
t [h]
S
N
O
-
3

[
m
g
N
L
-
1
]
b)


5state
ASM3
0 2 4 6 8 10
0
2
4
6
8
S
O

[
m
g
O
2
L
-
1
]
a)
0 2 4 6 8 10
20
30
40
50
t [h]
S
N
H
+
4

[
m
g
N
L
-
1
]
b)


5state
ASM3
ODE15s CPU time (sec)
Number of
Aer-Anox
phases
ASM3 9State 5State ASM3/5State
1 1.769 0.157 0.156 11.3
2 2.162 0.172 0.14 15.4
3 2.602 0.172 0.172 15.1
4 2.583 0.172 0.156 16.6
5 2.608 0.156 0.172 15.2
Mechanism Recognition applied on Sequencing Batch Reactors


78


Table 7.5. Singular function evaluations speed

7.7 MECHANISM RECOGNITION IN SBR PROCESSES
We now analyze the capacity of the 5state model to act as an indicator of substrate
concentration in the medium. As depicted in Figure 7.7, the 5state model is incapable to
describe the behavior of the process after substrate has been depleted. This
characteristic of the 5state model is to be expected, since the assumption made for its
reduction clearly state the condition of substrate present during the process. The 5state
model can be applied as an indirect method to measure substrate concentration.




Figure 7.7: Description of the 5state model in both regimes, with and without substrate.
Evaluation of Jacobians ASM3 2071.114
8State 565.364
CPU time 100 evaluations (sec)
5State 321.6
Calculation of the differential eq. system ASM3 2.225
8State 0.157
CPU time 1000 evaluations (sec) 5State 0.132
0 5 10 15 20 25
-100
0
100
200
300
400
500
600
700
800
t [h]
S
S
,
S
t
o

[
g
C
O
D
m
-
3
]


5state
ASM3
S
S
Sto
0 5 10 15 20 25
500
520
540
560
580
t [h]
X
H

[
g
C
O
D
m
-
3
]


5state
ASM3
0 5 10 15 20 25
40
60
80
100
120
t [h]
X
N
S
,
X
N
B

[
g
C
O
D
m
-
3
]


5state
ASM3
X
NS
X
NB
0 5 10 15 20 25
-10
0
10
20
30
t [h]
S
N
O
-
2

[
m
g
N
L
-
1
]


5state
ASM3
0 5 10 15 20 25
-5
0
5
10
15
t [h]
S
N
O
-
3

[
m
g
N
L
-
1
]


5state
ASM3
0 5 10 15 20 25
-2
0
2
4
6
t [h]
S
O

[
m
g
O
2
L
-
1
]


5state ASM3
0 5 10 15 20 25
0
50
100
150
t [h]
S
N
H
4

[
m
g
N
L
-
1
]


5state
ASM3
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


79

Because of the qualities of the submodel, simulation inaccuracies can be overcome by
the switching point detection step. Even for the case where the dynamics of the
substrate concentration is not properly predicted, the program can distinguish if the
failure on description capacity is due to a substrate mismatch or due to substrate
depletion.
7.8 RECOGNITION OF ORGANIC MATTER DEPLETION
7.8.1 CONDITIONS FOR PROPER PROCESS DESCRIPTION WITH
MECHANISM RECOGNITION
To ensure successful recognition of the different regimes during the process, the system
is required to fulfill conditions stated in section 6.2:
DYNAMIC PROCESS
- The SBR process is characterized by changes in the state of the system
throughout the complete run.
PROCESS HAS TO HAVE MORE THAN ONE REGIME
- As previously discussed, two different regimes are considered:
- biomass growth based on consumption of organic matter.
- biomass growth under organic matter limitation.
DETAILED MODEL OF THE GENERAL PROCESS AVAILABLE OF THE FORM:
- DAE system index zero or one
- no discontinuities
- known initial conditions
The extended ASM3 fulfils all the above mentioned conditions.
PARAMETER BOUNDARIES HAVE TO ASSURE GLOBAL CONVERGENCE WITH
GRADIENT BASED OPTIMIZERS
- The parameter set was tested for global optimality with stochastic
optimization based on PSO [60]. Although global optimality cannot be
assured, the parameters of the optimizer were set to maximize the
probability of global convergence.
Mechanism Recognition applied on Sequencing Batch Reactors


80

INITIAL REGIME IS KNOWN
- Because of the high concentration of organic matter at the beginning of
the cycle, it can be guaranteed that the initial regime of the process is
biomass growth based on consumption of organic matter.
THE SEQUENCE OF THE REGIMES IS KNOWN.
- Since only two regimes are considered in this process, the sequence is
known.
THE MINIMAL LENGTH OF EACH TIME REGIME ASSURES MODEL
DISTINGUISHABILITY
- Based on the variance of the data, the form of the general structure and
the conditions of the process, it can be assured that the period of
biomass growth on consumption of organic matter is long enough to
guarantee identifiability and detection of the switching point.
7.8.2 CONDITIONS FOR ACCURATE SWITCHING POINT
DETECTION
The assumptions mentioned above assure a successful model description of the process
using rigorous submodels with low systematic error. In addition, to allow a proper
detection of the switching points in the process, following conditions are to be
considered (section 6.3.6):
KNOWLEDGE OF THE NORMAL BEHAVIOR OF THE SYSTEM
- The behavior of the bacteria in each regime is described by the
submodels. In addition, the physical foundation of the extended ASM3
offers a basis for evaluation of the dynamics of the process.
DEFINITIVENESS OF THE CHANGING BEHAVIOR
- The distinguishability test proves that, at least for the simulated
experiments, the submodels are distinguishable. From this it may be
supposed, that the process offers enough amount of information to
detect the difference in bacterial behavior under both regimes.
AVAILABILITY OF AT LEAST ONE OBSERVATION REFLECTING REGIME CHANGE
- The combination of measuring variables considered, allows an accurate
detection of the change of regime.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


81

SATISFACTORY STATE OF INFORMATION
- The variance and frequency of the simulated experimental data provides
the state of information of the system required for MR.
MEASURED VARIABLES
MR was applied considering the following measured variables:
- Ammonia concentration [mgNL
-1
]
- Dissolved Oxygen in the medium [mgO
2
L
-1
]
- Nitrite concentration [mgNL
-1
]
- Nitrate concentration [mgNL
-1
]
GENERAL STRUCTURE
Cells utilize the consumed substrate for a number of functions as are growth, energy
production or storage. The reaction pathways responsible for substrate consumption
and processing depend on many conditions. From this it can be deduced that the yields
for substrate consumption achieved by bacteria vary depending on various factors. Still,
the extended ASM3 and consequently in the submodels, consider constant yield
coefficients. Taking advantage of MR the yield coefficients are considered to be RWPs.
By these means, the program can adapt the yield coefficient in each regime.
The PCPs are:
μ

Maximal growth of Heterotrophous
μ
1
Maximal growth of nitrosomones
μ
2
Maximal growth of nitrobacter
The RWP are:
Y
Haer
Yield coefficient of S
S
to X
H
in aerobic conditions
Y
A1
Yield coefficient nitrite to nitrosomones in aerobic conditions
Y
A2
Yield coefficient nitrite to nitrobacter in aerobic conditions
Y
A3
Yield coefficient nitrate to nitrobacter in aerobic conditions
INITIAL INTERVAL
A high concentration of organic matter in the feed water to the tank is guaranteed
assuring growth on organic matter as the initial interval.

Mechanism Recognition applied on Sequencing Batch Reactors


82

7.8.3 MR INITIALIZATION
The threshold D
B
set at 10 for the A criterion of the general structure is reached after
0.30 days (7.2 hours). From this point on, MR is initiated to detect depletion of organic
matter.

Figure 7.8: Minimal length for initialization of MR
The scenario selected to show the performance of MR is depicted in Figure 7.9 and
includes three short aerobic phases as well as three large anoxic intervals. This scenario
was selected in order to test MR in a process with very low state of information. The
anoxic intervals offer measurements with low information content since 1) the
parameters set to estimation control mostly the aerobic phase and 2) the concentrations
of oxygen and Nitroxides drops to zero. Still MR shows an incredibly good performance
in terms of both accuracy and robustness.
7.8.4 DETECTION OF SWITCHING POINTS
Detection of the switch between overflow and substrate limitation regime was carried
out controlling the value of the objective function. The constant threshold was set equal
to a MXL value of 0.35. Substrate depletion is accurately detected despite the
impossibility to measure the concentration of organic matter online. As it can be seen in
Figure 7.9, MR detects the point with high precision. Furthermore, the large slope of
H0 at the depletion region assures a robust and accurate detection of the regime change.
0 0.2 0.4 0.6 0.8
0
500
1000
S
S

[
g
C
O
D
m
-
3
]
0 0.2 0.4 0.6 0.8
0
20
40
time (days)
S
N
O
3
-
,
S
N
O
2
-
,
S
N
H
4
-
,
S
O
2

[
m
g
N
L
-
1
]


0 0.2 0.4 0.6 0.8
10
0
10
1
10
2
10
3
time (days)
A
c
r
i
t
G

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


83


Figure 7.9. Detection of the regime switching point.
SWITCHING INTERVAL
The experimental results confirmed that considering an instant switch of regime allows
an accurate description of the process.
7.9 CONCLUSIONS
An assiduous analysis of the process conditions, empirical knowledge, and different
approaches to model reduction have been applied to develop a submodel of the
extended ASM3. By these means, it was possible to create a model that is significantly
simpler and delivers more information about the process. The implementation of MR
for description of ASP in SBR processes clearly states the advantages of this approach.
The resulting 5state submodel mimics the behavior of the extended ASM3 for SBR
process during growth based on organic matter with high accuracy. The results suggest
that the 5state model can be applied for the simulation of the nitrate bypass reaction in
SBR processes, and thus, for model-based control and online optimization. Moreover,
the extended ASM3 was successfully simplified, reducing its calculation costs up to one
order of magnitude.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0
0.2
0.4
time (days)
H
0

[
]
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0
500
1000
S
S

[
g
C
O
D
m
-
3
]
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0
20
40
60
S
N
O
3
-
,
S
N
O
2
-
,
S
N
H
4
-
,
S
O
2

[
m
g
N
L
-
1
]
Mechanism Recognition applied on Sequencing Batch Reactors


84

Most important, the limitation of the submodel to describe the process after depletion
of organic matter can be exploited. In other words, the reduction of the extended ASM3
creates a model which can indicate the time point where organic matter is present in the
reactor. The 5state model works as an indicator of the non measurable state variable
readily biodegradable organic matter. The program is able to detect the time point when
no more carbonate matter is present in the reactor increasing process efficiency.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


85

8 MECHANISM RECOGNITION IN
ESCHERICHIA COLI
CULTIVATIONS
8.1 ESCHERICHIA COLI CULTIVATIONS
Many reports about process optimization of Escherichia coli cultivation are describing
problems caused by acetate synthesis, which can retard growth and protein production
[151-153]. In E. coli, acetate synthesis and excretion occurs during the so called
„overflow metabolism‟ [154] when substrate concentration (namely glucose) is available
in large amounts. The threshold value depends on the strain and culture conditions.
Furthermore, acetate synthesis already occurs under some conditions at weak excess.
The carbon flow through acetyl-CoA is partly shifted towards acetate production
instead of entering into the tricarboxylic acid cycle [155]. Acetate is synthesized via
phosphotransacetylation. The product acetyl-phosphate is then converted to acetate by
the acetate kinase enzyme. Also a direct conversion from pyruvate to acetate with the
pyruvate oxidase is possible [156, 157]. Acetate itself can be reconverted to acetyl-CoA
either via acetyl-phosphate or by acetyl-CoA synthetase (Acs) via acetyl-AMP. The latter
reaction complex is characterized by a 50-fold higher affinity and therefore responsible
for acetate reconversion at low concentrations [158]. It could be demonstrated in
accelerostat studies, that the first step of overflow mechanism in E. coli is the downshift
of the ACS system leading to a lower conversion of acetate via acetyl-AMP. However,
the reconversion of acetate is regarded as being essential for chemotaxis, proteolysis and
pathogenesis [159].
Several attempts have been made to circumvent acetate accumulation. Pulsed feeding
strategies based on dissolved oxygen measurements reduced acetate formation
increasing growth and product concentrations [160, 161].
The behavior of E. coli at substrate excess (overflow metabolism) when acetic acid is
formed as carbon storage, the activation of the glyoxylatic shunt, and the response to
oxygen limitation, are in the focus of recent development. While more and more
regulatory mechanisms are considered in models, the increasing number of equations
Mechanism Recognition in Escherichia coli cultivations


86

and parameters increment model complexity. In addition, since the appropriate
detection of intermediates is crucial due to their naturally low levels in bacteria, many
compartments of the metabolism cannot be measured accurately or in a sufficient
resolution by time.
The number of parameters that can be monitored and hence integrated in models as
measured variables is steadily increasing. Nowadays, very advanced techniques are
available including online respirometry [162], continuous glucose monitoring [163], on-
line high performance liquid chromatography (on-line HPLC) [165], multi-wavelength
fluorescence (MWF) [166, 167], near-infrared spectroscopy (NIRS)[162, 168, 169], as
well as conventional but very efficient techniques like Dissolved Oxygen Tension
(DOT) [164] among others. Also methods for describing the physiological state of the
cell are developed for monitoring bacterial cultivation like flow cytometry for example
[170, 171]. However, there is no possibility to monitor these parameters on-line. Junne
[172] reports an approach aiming at monitoring the polarisability and cell length at line
and correlate these parameters with the cell viability as it was performed based on flow
cytometry studies in batch, fed-batch, and continuous cultures [173]. The great
advantage of the method is the absence of cell staining which allows for a much simpler
and automated sample preparation. Still efficient methods to obtain sufficient state of
process information with online observations are to be developed. For this reason,
parallel to advances in measuring techniques, models aiming at an accurate description
of bacterial behavior at different levels are being developed.
8.2 MODELS FOR THE DESCRIPTION OF
ESCHERICHIA COLI CULTIVATIONS
Many approaches have been developed during the last thirty years to create descriptive
and predictive models for bacterial organisms. Some of them have been created for the
scientifically well-described facultative bacterium Escherichia coli. Due to its easy
cultivation, this bacterium became the “workhorse” of experimentalists for studying
novel microbiological and analytical methods and for the industrial production of
proteins and other substances of commercial interest [174].
In recent years, dynamic models including the regulatory interaction on the
metabolomic level have been developed. For example, Chassagnole [175] created a
kinetic model of the sugar uptake system (phosphotranspherase system – PTS) and the
glycolysis following the Embden-Meyerhof-Parnas pathway in E. coli. Also, Wang [176]
described the catabolite repression based on the sucrose and glycerol transport system
in E. coli. In this study, the PTS-dependence of the sucrose transport system was
considered in the model. More and more knowledge is available concerning the
interaction of messenger nucleotides, which act at the DNA replication, translation,
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


87

transcription and enzyme activity [177]. Models included the general stringent response
in E. coli, which has a major impact on the process performance [178]. The integration
of signal molecules can serve as bridge between the regulation on the metabolome and
the transcriptome. The alarmone guanosine tetraphosphate (ppGpp) acts as a main
modulator of gene transcription [179]. It is activated, when the cell starves as in large-
scale fed-batch cultivations. Therefore, the simulation of the impact of ppGpp and
other nucleotides on the cell‟s regulation is of great interest for predicting the
physiological state of the cell and consider it for modeling [126, 180]. The consideration
of alarmones contribute to a more sensitive prediction capacity and a simulation of
dynamic cell behavior [181].
For example, a recently developed model of the total central carbon metabolism in E.
coli comprises of 50 kinetic rate equations and 46 equations that simulate gene
expression in the corresponding pathways [182]. This leads to the problem that for a
wider application range of the model, parameter estimation becomes a difficult and
sometimes even impossible task. However, parameter estimation with a limited set of
experimental data leads to unreliable results while multiple solutions exist.
Among recent strategies to overcome these problems are the application of simplified
kinetics for expressing the biochemical reaction rate equations, e.g. with linear-
logarithmic (so called lin-log) approaches [183, 184]. Also the successful application of
piecewise-linear approximations for a multilevel considering model for E. coli
cultivations has recently been described [7]. Here, the model reduction maintained the
dynamic simulation capacity of the model for the starvation response of E. coli, although
precision of prediction could not be fully maintained. Other approaches to reduce the
complexity of complex models included temporal decomposition and establishment of
pools which describe the physiologic status of the cell with very few parameters [185].
Still these reduced versions at heuristic level, need to be continuously fitted to new
conditions and do not offer nor inside of process conditions, nor of the physiological
state of the bacteria. Hence, although there is a great field of application for models for
description of E. coli cultivations in research and production, their application is very
restricted due to the afore mentioned limitations.
8.2.1 DIVISION OF PHYSIOLOGICAL STATES
In an effort to create a tractable model able to describe substrate and oxygen uptake
capacities of the cells, Lin [91] proposes a model where cellular response dynamics are
lumped into two variables. For the equation system see Appendix A. This model
considers time dependent uptake capacities over time and thus enables a better fit of
experimental glucose data. Despite the important advances achieved, the model
developed by Lin does not consider variation of cell response caused by the
physiological state of the bacteria. For this reason, in this section an alternative
modeling approach is proposed considering three main cell states, namely overflow,
Mechanism Recognition in Escherichia coli cultivations


88

growth on substrate limitation and starvation (under which the regulatory interaction of
the alarmone ppGpp is active).
Recent studies in chemostat and accelerostat E. coli cultivations have proven the
dynamic development of transcription, protein expression and carbon conversion at
different dilution rates [155, 159, 186]. A clear difference of the expression of enzymes
in the glycolysis and the TCA cycle depending on the state of the culture was
determined. This information can be used to account for altered enzyme concentration
at different cultivation conditions (leading to different physiological conditions inside
the cell). For example, acetate was shown to be synthesized already at low carbon supply
via the pyruvate oxidase (Bpox) from pyruvic acid [156]. Acetate is then converted to
acetyl-CoA via the acetyl-CoA synthase (Acs) [159, 187]. The expression of Acs is
strongly reduced (more than 30-fold) at high substrate uptake rates (at the onset of
acetic acid accumulation). Over expression of this pathway has shown that acetate
accumulation could be reduced [188]. This indicates that the pathway provides acetate
accumulation at lower substrate uptake rates, although acetate is produced. The low Km
value of Acs of 200 µM supports this hypothesis. Despite of this, the usual conversion
of accumulated acetate via the acetate kinase to acetyl-phosphate is characterized by a
value of the catalytic enzyme of 7-10 mM [158]. The expression of it does not change at
altered substrate uptake [159].
Three submodels, each one describing one of the aforementioned physiological states of
the cell, are to be developed in an effort to increase prediction accuracy while
maintaining model simplicity. By this means it is possible to reduce the number of
parameters and increase model flexibility and model identifiability [189]. A further
advantage of this approach is the achievement of a reliable mechanistic recognition of
non measurable parameters. The simplicity of the models applied, the robustness and a
better evaluation of the changing states of the process are potential contributions of MR
to the description of E. coli cultivations. Besides, MR enables a closer analysis of the
dynamics of the model and the correctness of its structure.
Similar to the concept of MR, Veloso [190] developed a software sensors framework to
monitor E. coli fermentations. He applied an asymptotic observer based on a simplified
model with five state variables, namely:
- Biomass concentration
- Substrate concentration
- Acetate concentration
- Oxygen concentrations in the offgas
- Carbon dioxide concentrations in the offgas


An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


89

combined in the form:

=

1 1 1

1

2
0
0
3

4

5

6

7

8

9

10

1

2

3
∗ − ∗

+

0

0

(8.1)
where , , , , and represent biomass, glucose, acetate, dissolved oxygen, and
dissolved carbon dioxide concentrations, respectively;
1
,
2
, and μ
3
are the specific
growth rates;

are the yield (stoichiometric) coefficients;

and

are the substrate
feed rate and the glucose concentration in the feeding solution, respectively; D is the
dilution rate and

is the culture medium weight. is the carbon dioxide transfer
rate from liquid to gas phase and is the oxygen transfer rate from gas to liquid
phase.
Finally, Veloso divides the process in four possible phases, named regimes:
- regime A:
- simultaneous oxidative and overflow growth on glucose (μ
1

2
>
0; μ
3
= 0)
- regime B:
- oxidative growth on glucose (μ
1
> 0; μ
2
= μ
3
= 0)
- regime C:
- simultaneous oxidative growth on acetate and glucose (μ
1
, μ
3
>
0; μ
2
= 0)
- regime D:
- oxidative growth on acetate (μ
3
> 0; μ
1
= μ
2
= 0)
This proposition allows building observable models with a reduced number of measured
state variables. Similar to the MR approach, Veloso divides the batch process to enable
its description with simpler models. From a physical point of view, the growth rate is
considered to be constant for each regime. Assuming a piecewise constant growth rate
simplifies the model to the extent where observability is obtained.
Although theoretically applicable for process monitoring, the proposed method has
some important drawbacks.
- The model is not continuous (discontinuity hinders gradient calculation
in the switching point).
- The model proposed for monitoring is simple and has no physical basis
nor offers information about the system.
- The results obtained cannot be translated to a first principle model to
extend state of information.
Mechanism Recognition in Escherichia coli cultivations


90

- The regimes have to be known beforehand. This supposes important
process knowledge impossible to obtain in a real process.
In order to overcome the above mentioned disadvantages, the MR approach is
proposed.
8.3 MODELING ESCHERICHIA COLI BATCH
FERMENTATIONS WITH MECHANISM
RECOGNITION
In this section, MR is applied to achieve an efficient mechanistic modeling and
simulation of E. coli batch fermentations and the simulation framework is validated
against experimental data. Fermentation processes are characterized by its dynamic
behavior described by parameters such as growth rate, substrate concentration and
cellular metabolic activity. Although models able to describe individual batch and fed-
batch fermentations exist, they become unreliable when applied to different
fermentation conditions. To overcome this drawback, diverse models are used at
various regimes enabling not only a better description of the process, but also an
improved understanding of non measurable characteristics. Three models compete in
different intervals of the process. The candidate models are:
- Overflow metabolism model (OF),
- Growth on substrate Limitation model (SL),
- Starvation model (ST).
Using an adequate model sequence, acetate formation, substrate consumption and cell
growth are described with high accuracy. Moreover, the data needed to fit the models
are reduced and a standardization of the model to be applied in different process states
is enabled.
8.3.1 GENERAL MODEL
The model for the description of E. coli K12 fed-batch fermentations was extracted
from Lin [91]. This model was originally built in an effort to describe all physiological
and regulatory effects, which cause uptake capacity variation of substrate and oxygen,
with a relatively simple kinetic model. A series of experiments with glucose pulses were
carried out to determine the dynamics of the uptake capacity of glucose during
fermentation.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


91

To describe these variations, Lin creates two fictitious enzymes in which all
physiological and regulatory effects are put together, the enzymes for glucose uptake
and for oxygen uptake to biomass [91]. This model successfully describes the variations
of the uptake capacities of substrate and oxygen of determined fermentations. The
behavior of the six principal variables: biomass, substrate, acetate, DOT substrate
uptake, and oxygen uptake are depicted in Figure 8.1

Figure 8.1: Integration of the kinetic model proposed by Lin [91]
Still, the model fails to describe the extremely different behavior of E. coli under
different process conditions.

Figure 8.2: Complex model (Lin et al.) fitted to experimental batch cultivation data.
5 10 15
0
5
10
15
Biomass Concentration
[
g
/
l
]
5 10 15
0
5
Substrate Concentration
[
g
/
l
]
5 10 15
0
0.5
1
Acetate Concentration
[
m
M
o
l
/
l
]
5 10 15
0
50
100
150
DOT
[
%
]
5 10 15
0
0.05
ratio of substrate consumption
[

]
t [h]
5 10 15
0
0.05
0.1
ratio of oxygen consumption
[

]
t [h]
0 1 2 3 4 5 6 7 8 9 10
0
2
4
6
8
10
12
14
time [h]
A
c
e
t
a
t
e

A

[
m
M
o
l
/
l
]
,

S
u
b
s
t
r
a
t
e

S

[
g
/
l
]
,
B
i
o
m
a
s
s

X

[
g
/
l
]


A
X
S
Mechanism Recognition in Escherichia coli cultivations


92

In addition, the model contains discontinues equations which represents a major
obstacle for robust simulation and optimization. Moreover, the model fails to describe
different experiments. The results of the parameter estimation against real data are
depicted in Figure 8.2. The optimization was carried out with a stochastic optimizer
PSO (section 5.3) in order to overcome local minima.
8.3.2 SUBMODELS FOR DIVIDING METABOLIC STATES
A reduction of the model in three submodels makes it possible to overcome the
previously mentioned drawbacks improving description accuracy and robustness. Each
submodel describes one of the three main states of the cell: overflow metabolism Figure
8.3, growth under substrate limitation Figure 8.4, and starvation Figure 8.5.
MODEL FOR DESCRIPTION OF THE OVERFLOW METABOLISM
Drawbacks of acetate production are widely known and have been reported in various
contributions. Acetate production is related to growth inhibition [153, 191] and to
reduction of recombinant protein production [192, 193]. Lin [91]summarizes two rate-
limiting steps contributing to overflow of acetate in aerobic glucose-based cultures of E.
coli:
- in electron transport system [151]
- rate-limiting steps in the in the TCA cycle [194, 195]
The general model is reduced to limit its description capacity to the specific conditions
of the overflow metabolism (OF):
- Substrate concentration is always higher than the minimum amount
needed for maintenance.
- Bacteria consume substrate and oxygen to its maximum possible under
consideration of substrate and product limitation.
- The expression of both enzymes, for substrate and oxygen uptake
capacity, is maximal and constant.
Acetate consumption is considered, since bacteria also consume acetate in overflow
conditions where high acetate concentration is present in the medium.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


93


Figure 8.3: Comparison between the complex model (dots) vs. the overflow submodel (lines)
initializing in four different intervals.
In Figure 8.3 it can be seen that the submodel describes the initial stage of the process
with high precision. Four intervals were selected arbitrarily to test the description
capacity of the model in different stages of the process. The prediction accuracy is
drastically reduced as the concentration of substrate drops. This suggests that the
submodel is adequate for the identification of the bacterial growth with overflow
metabolism. For the equation system see Appendix B.
MODEL FOR DESCRIPTION OF GROWTH UNDER SUBSTRATE LIMITATION
Many fermentation strategies as well as metabolic engineering approaches have been
applied to reduce the production of acetate [191, 196-202]. It is commonly accepted,
that optimal conditions for cultivation are obtained under (moderate) substrate
limitation. By these means acetate accumulation is minimized and biomass to substrate
yield maximized, while maintaining a relatively high growth rate [155]. A topology of the
regulatory network of carbon limitation has been published by Hardiman [181] on
which it is possible to estimate to which extend substrate limitation is suitable for
application and when starvation and stringent responses are initiated.
Following assumptions were made to create the substrate limitation model:
- Substrate concentration is always higher than the minimum amount
needed for maintenance.
- There is no limiting capacity for oxygen consumption.
0 5 10 15
0
5
10
15
Biomass Concentration
[
g
/
l
]
0 5 10 15
0
5
Substrate Concentration
[
g
/
l
]
0 5 10 15
0
0.5
1
Acetate Concentration
[
m
M
o
l
/
l
]
0 5 10 15
0
50
100
150
DOT
[
%
]
0 5 10 15
0
0.05
ratio of substrate consumption
[

]
t [h]
0 5 10 15
0
0.05
0.1
ratio of oxygen consumption
[

]
t [h]
Overflow
Mechanism Recognition in Escherichia coli cultivations


94


Figure 8.4: Comparison between the complex model (dots) vs. the substrate limiting submodel
(lines) initializing in four different intervals.
Contrary to the previous submodel, the submodel for substrate limitation is only able to
describe cell behavior at very low substrate concentrations Figure 8.4. The model
predicts an extremely fast growth in overflow conditions, because it lacks uptake
capacity limitation at overflow metabolism responsible for acetate synthesis and
excretion. The detailed equation system is presented in Appendix C.
MODEL FOR DESCRIPTION OF CELL STARVATION
Finally the description of the starvation stage is also considered. The cell starvation is to
be avoided at all cost in industrial E. coli cultivations. Still large scale reactors inherently
imply special concentration gradients in the medium exposing bacteria to undesired
conditions. Extreme substrate limitation may trigger the phosphoenol pyruvate
glyoxylate (ppG) shunt. Once the stringent response is activated, reactivating bacterial
growth requires long periods under substrate excess.
To reduce the models it was assumed that:
- substrate uptake is lower than required by the cell, hence no energy
limitation is required
- enzyme rate are minimal and constant
0 5 10 15
0
5
10
15
Biomass Concentration
[
g
/
l
]
0 5 10 15
0
5
Substrate Concentration
[
g
/
l
]
0 5 10 15
0
0.5
1
Acetate Concentration
[
m
M
o
l
/
l
]
0 5 10 15
0
50
100
150
DOT
[
%
]
0 5 10 15
0
0.05
ratio of substrate consumption
[

]
t [h]
0 5 10 15
0
0.05
0.1
ratio of oxygen consumption
[

]
t [h]
Sub. Limit.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


95


Figure 8.5: Comparison between the complex model (dots) vs. the cell starvation submodel
(lines) initializing in four different intervals.
As expected, the model describing cell starvation is not able to predict the behavior at
any stage of the process Figure 8.5. This indicates that the process does not reach
starvation conditions. For the equation system see Appendix D.
The structure of the submodels is shown in Table 8.1. The identifiability of the
submodels will be determined by the parameter boundaries and the consistency of the
general structure indicates the following. This again is key to accurate regime
recognition.
Unlike the process constant parameters, the regime-wise constant parameters are
optimized independently in each regime. The parameters estimated during the
recognition process are shown in Table 8.1
Table 8.1: Parameters considered for the model fit
PCPs (general structure, section 6.3.2):

specific death rates for the uptake enzyme of oxygen [g/(gl)]

Maximal specific oxygen uptake rate [g/(gh)]

Maximal specific substrate uptake rate [g/(gh)]

specific death rates for the uptake enzyme of substrate [g/(gl)]

0 5 10 15
0
5
10
15
Biomass Concentration
[
g
/
l
]
0 5 10 15
0
5
Substrate Concentration
[
g
/
l
]
0 5 10 15
0
0.5
1
Acetate Concentration
[
m
M
o
l
/
l
]
0 5 10 15
0
50
100
150
DOT
[
%
]
0 5 10 15
0
0.05
ratio of substrate consumption
[

]
t [h]
0 5 10 15
0
0.05
0.1
ratio of oxygen consumption
[

]
t [h]
Starvation
Mechanism Recognition in Escherichia coli cultivations


96

RWPs:

Yield of the enzyme for oxidative energyproduction from the biomass
growth [-]

Yield of the enzyme for substrate uptake from the biomass growth [-]

Yield of the enzyme for respirationfrom the biomass growth [-]

Yield of the enzyme for acetate from the biomass growth [-]

Yield of the enzyme for overflow from the biomass growth [-]
8.4 MATERIAL AND METHODS
(All chemicals in this study were either purchased by Sigma Aldrich GmbH, Munich,
Germany, or Carl Roth KG, Karlsruhe, Germany, if not otherwise stated.)
8.4.1 STRAIN AND CULTURE CONDITIONS
Glucose 10 g/l
Yeast -extrakt 5 g/l
K
2
HPO
4
3 g/l
KH
2
PO
4
1.5 g/l
(NH
4
)
2
SO
4
1.25g/l
MgSO
4
0.1 g/l
Na
2
EDTA*7H
2
O 0.037 g/l
NaCl 0.01 G/l
FeSO
4
*7H
2
O 0.001 g/l
Glucose was autoclaved separately to avoid Maillard-reactions.
1000 ml Erlenmeyer flasks were filled with 250 ml of medium. The medium was
inoculated with E. coli K12 W3110 picked from a frozenstock culture. The flasks were
incubated at 37°C and 150 rpm on a longitudinal shaker.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


97

BATCH BIOREACTOR EXPERIMENTS
A reactor KLF2000 (Bioengineering AG, Wald, Switzerland) was used for all
fermentations Figure 8.6. To insure sufficient oxygen supply, the reactor was sparged
with air at a rate of 1 vvm. Both, the gas inlet and the exhaust gas were steril-filtered
with autoclavable cellulose filters.

Figure 8.6: Bioreactor KL2000 at E. coli batch cultivation [203]
The inlet gas filter was autoclaved separately while the outlet gas filter was autoclaved
together with the vessel by a small steam flow. The reactor was sterilized prior to
cultivation at 120°C and 1 bar overpressure for 20 min. The level of pH was controlled
to pH 7.0 with NaOH 30% w/w during the acetate production phase and HNO
3
10%
w/w during the acetate consumption phase. The stirrer was working at 700 rpm and
polyethyleneglycol 2000 to 3000 was added in small amounts (~ 1ml) to reduce foam
whenever needed.
8.4.2 ONLINE ANALYSIS
A sterilizable temperature detector Pt100 is connected to an electronic measuring and
control unit, which is controlling the heating and cooling devices.
Mechanism Recognition in Escherichia coli cultivations


98

An Ingold single-rod measuring cell is used to measure the pH of the medium. The
concentration of the dissolved oxygen in the medium is detected with an amperometric
(polarographic) Ingold pO
2
-electrode (Mettler-Toledo Inc., Mettlach, Germany).
O
2
concentration in the exhaust gas is determined by an OXYGOR 6 N (Maihak AG,
Hamburg, Germany) device following the magneto-pneumatic measuring principle.
CO2 concentration in the exhaust gas is measured by a non-depressive infrared
photometer with a selective optical-pneumatic receiver UNOR 4 N (Maihak AG,
Hamburg, Germany).
The oxygen uptake was calculated based on the exhaust gas analysis as follows:
4 . 22 *
) (
2 2
2
F
O O G
O
V
Y Y V
Q
e o
÷
=


8.2
The carbon dioxide production rate was calculated as:
4 . 22 *
) (
2 2
2
F
CO CO G
CO
V
Y Y V
Q
o e
÷
=


8.3
Where:
F
V
- fluid volume [l]
G
V

- volumetric gas flow m
3
/h
2 O
Q
- oxygen uptake rate [mol/lh]
2 CO
Q
carbon dioxide production rate [mol/lh]
o
2
O
Y
-mole fraction of oxygen in the incoming gas phase [ ]
e
2
O
Y
- mole fraction of oxygen in the exhaust gas [ ]
ONLINE-MEASUREMENT OF THE OPTICAL DENSITY WITH ELOCHECK®
EloCheck
®
Figure 8.7 was also developed by EloSystemsGbR for biomass online
monitoring. Because of the implementation of a smaller cuvette with a thickness of 1.6
mm instead of the usual 10 mm, EloCheck
®
is able to measure the optical density of the
sample through a bypass system [204]. EloCheck
®
needs no sample dilution which
makes it possible to reintroduce the measured sample into the reactor. By this means,
the device takes a sample and measures it every 15 seconds. Both biomass curves from
EloCheck
®
and offline analysis were compared and a high similarity on the run of the
curves was evident. The data of EloCheck
®
was not used for further analysis because of
small deviation due to antifoam agent which caused temporary occlusion in the cuvette.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


99


Figure 8.7: EloCheck
®

8.4.3 OFFLINE ANALYSIS
SAMPLE PROCEDURE
About 5 ml of crude extract were withdrawn from the reactor every 30 min. Two
Eppendorf “Flex-Tubes
®
1.5 ml, per 1,000 pcs.” were filled with 1 ml sample. The
supernatant was separated through centrifugation for 10 min at 13,000 rpm. One flex-
tube supernatant was used for the glucose determination and the second one was stored
at –20°C for later acetate determination. About 3 ml were used for biomass
determination.
BIOMASS DETECTION
Biomass was determined through optical absorption measurement with a Zeiss Spektral
Photometer (PM2A, West Germany). The biomass containing sample was diluted until
the extinction was in the range of the Lambert-Beer law (0.2 to 0.5). The determination
of the extinction was performed at λ = 600 nm. The equation for the linear regression
between biomass and extinction was:
Biomass [g/l] = 0.398*extinction [nm]-0.039
ENZYMATIC DETERMINATION OF GLUCOSE
Glucose determination was performed with the liquiUV
mono
Test (Human Geselschaft
für Biochemical und Diagnostica mgH, Wiesbaden, Germany). In this test, the glucose
was the substrate for the reaction which is catalyzed by the hexokinase and the glucose-
6-phosphat-dehydrogenase at presence of Adenosintriphosphat (ATP) and
Nicotinamide Adenine Dinucleotide (NAD
+
, NadH). During this chemical conversion
NadH, is accumulated:

The amount of NadH is proportional to the concentration of the glucose. The
extinction of NadH is measured at λ = 340 nm. A calibration curve consisting of
Mechanism Recognition in Escherichia coli cultivations


100

duplicate measurements at five different concentrations was made prior to each
fermentation. In Figure 8.8, the calibration curve for the first experiment is shown:
Concentration [g/l] = 2.776*extinction [nm] – 0.014

Figure 8.8. Calibration curve for glucose determination
Optical measurement was realized by a Beckman DU640 spectrophotometer (Beckman
Coulter Inc., Fullerton, USA).
The specific substrate uptake rate was obtained by dividing the difference off the acetate
concentration between two samples, by the biomass at this time.

=

+1

+1

+

+1

2

8.4
where:

= specific glucose uptake rate [h
-1
]

= glucose concentration [g/l]

= Biomass concentration [g/l]

= time [h]
ENZYMATIC DETECTION OF ACETIC ACID
Acetic acid determination was realized with the UV method, which is based on the
reaction from acetic acid to acetyl-CoA [205] (ROCHE Test Kits Cat. No.10148261035,
R_BIOFARM AG, Darmstadt, Germany).

The determination is realized on a light absorbance wave length of λ = 340 nm.
0
0,5
1
1,5
2
2,5
3
0 0,2 0,4 0,6 0,8 1
c
o
n
c
e
n
t
r
a
t
i
o
n

[
g
/
l
]
extintion [nm]
[g/l]
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


101

Samples were taken every 30 min, centrifuged and stored at -20°C for no more than
three weeks. Optical measurement was realized by a Beckman DU640
spectrophotometer (Beckman Coulter Inc., Fullerton, USA).
The calibration curve was plotted with the data obtained out of the supplied control
solution at five different concentrations Figure 8.9.

Figure 8.9. Calibration curve of acetate
The concentration of acetate was calculated with the following formula obtained out of
the calibration curve:
Acetate concnetration = 2.412*Absorbance – 0.0186 [g/l]
The specific acetate production rate was obtained by dividing the difference off the
acetate concentration between two samples, by the biomass at this time.

=

+1


+1

+

+1

2

8.5
where:

= specific acetate production rate [mMol/(g*h)]

= acetate concentration [mMol/l]

=Biomass concentration [g/l]

= time [h]
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0 0,05 0,1 0,15 0,2
A
b
s
Conc Acetate (g/l)
Conc vs Abs
Mechanism Recognition in Escherichia coli cultivations


102

DETERMINATION OF INTRACELLULAR NADH CONCENTRATION
The intracellular NADH concentration was measured by a cycling assay according to
Nisselbaum and Green [206] modified by Bernofsky and Swan [207] and improved by
San and Bennett [208].
Briefly, a 1 ml sample was pipetted into a microcentrifuge tube. The centrifugation was
performed for 10 min at 13,000 rpm. 300 µl of 0.2 M NaOH were added for cell lysis.
After 10 min of incubation in a 50°C bath, samples where cooled to 0° C and
neutralized by carefully adding 300 µl of 0.1 M HCl while vortexing. Cellular debris was
removed by centrifugation for 5 min at 15,000 rpm.

Figure 8.10: Mechanism of the reactions involved in the assay
The cycling assay Figure 8.10 was performed using a reagent mixture consisting of equal
volumes of 1.0 M bicine buffer (pH8.0), absolute ethanol, 40 mM EDTA (pH8.0), 4.2
mM 3-[4,5-dimethylthiazol-2-yl]-2,5-diphenyltetrazolium bromide (thiazolyl blue; MTT)
and twice the volume of 16.6 mM phenazineethosulfate (PES), previously incubated at
30°C. The following volumes were added to 1 ml cuvettes: 50 ml neutralized extract, 0.3
ml water and 0.6 ml reagent mixture. The reaction was started by adding 50 ml of
alcohol dehydrogenase (ADHII isolated from rabbit) of a concentration of 500 or 100
U/ml in 0.1 M Bicine (pH 8.0) buffer. The absorbance at λ = 570 nm was recorded for
10 min. The assay was calibrated with 0.01–0.05 mM standard solution of NAD
+
.
Table 8.2. Composition of solution A
Solution Volume 650 µl
1.0 M bicine buffer (pH8.0) 650 µl
absolute ethanol 650 µl
40mM EDTA (pH8.0) 650 µl
4.2mM MTT 650 µl
16.6 mM PES 1300 µl

1 ml cuvettes were filled with the following mixture:
50 µl extract
0.6 ml solution A
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


103

0.3 ml water
To start the reaction:
50 µl of ADH (500 or 100 U/ml in 0.1 M Bicine (pH 8.0) buffer)
The biomass yield was calculated with the following equation:

/
=

+1

+1

8.6
where:

/
= Biomass yield [-]

= glucose concentration [g/l]

=Biomass concentration [g/l]
8.4.4 DATA TREATMENT
The data of every pair of fermentations with the same conditions were considered for
calculating a fit-curve. Samples that were considered as outliers were manually excluded.
In the case of the biomass curve, oscillations at the beginning of the curve were also
attenuated per hand.
Acetate samples were taken every 30 min. The data of every pair of fermentations with
the same conditions were considered for calculating a fit-curve whenever possible. The
fit curve was calculated in Matlab with a 7
th
grade polynomial. The acetate production-
rate was obtained after dividing the difference of acetate between two samples and
dividing the result through the time between both samples.

8.7
Finally, the acetate building-rate was divided with the biomass to obtain the specific
acetate building-rate.
This procedure was also applied to all other parameters which were converted into
specific rates.
( )
( )
1 2
1 2
t t
Ac Ac
t
Ac
÷
÷
=
A
A
Mechanism Recognition in Escherichia coli cultivations


104

8.5 EXPERIMENTAL VALIDATION
First we compare submodel performance against a simulated data set. The submodels
created in the previous section showed good performance when compared to simulated
data. This indicates that it is possible to find combinations of simple models with similar
dynamics to more complex versions. MR achieves an accurate detection of the change
of regime and finds consistent parameter combinations for both models (OF and SL).
Since the complex model used for submodel building is the same model used for data
generation, it can be assured that the structure of the model is globally correct (no
systematic error). In other words it is possible to prove that the structure of the
complex models describes the process dynamics completely and no uncertainties or
unknowns are present besides white noise (stochastic error). Certainty of optimal
structure for the complex model allows a direct analysis of process phenomena and the
precise detection of changing regimes.
Nevertheless, the highest accuracy prediction of these artificially generated observations
is clearly obtained with the complex model, which also represents the optimization with
the largest number of parameters to fit. One could argue that, also in these cases, MR
offers a simplified basis for simulation and optimization. However, the effort of
submodel building and MR programming does not justify its application against
simulated data. It has been clearly stated in section 1.6 that the objective of the MR
approach is to enable industrial application of complex models. Therefore it has to be
proven that MR allows using complex models applied to real experimental data. It is
essential to test the performance of MR against data sets obtained in real experiments.
8.5.1 CONDITIONS FOR PROPER PROCESS DESCRIPTION WITH
MR
MR was tested against batch cultivation with E. coli K12 W3110 bacteria. To ensure
successful recognition of the different regimes during the process, the system is required
to fulfill following conditions.
DYNAMIC PROCESS
- The batch process is characterized by changes in the state of the system
throughout the complete run.
PROCESS HAS TO HAVE MORE THAN ONE REGIME
- As previously discussed, the cultivation of E. coli is known to present
three basic regimes; overflow metabolism, growth under substrate
limitation and starvation.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


105

DETAILED MODEL OF THE GENERAL PROCESS AVAILABLE OF THE FORM:
- DAE system index zero or one
- no discontinuities
- known initial conditions
The model applied to describe this process fulfils all the above mentioned conditions
PARAMETER BOUNDARIES HAVE TO ASSURE GLOBAL CONVERGENCE WITH
GRADIENT BASED OPTIMIZERS
- The parameter set was tested for global optimality with stochastic
optimization based on PSO [209], where the parameters of the optimizer
were set to maximize the probability of global convergence.
INITIAL REGIME IS KNOWN
- Because of the high concentration of substrate at the beginning of the
process, it is assured that the initial regime of the process is growth
under overflow conditions (after a short lag phase).
THE SEQUENCE OF THE REGIMES IS KNOWN.
- The dynamics of substrate consumption and the reaction capacity of the
bacteria assure that the regime following overflow conditions is growth
under substrate limitation.
THE MINIMAL LENGTH OF EACH TIME REGIME ASSURES MODEL
DISTINGUISHABILITY
- Based on the variance of the data, the form of the general structure and
the conditions of the process, it can be assured that the period between
overflow regime and substrate limitation is long enough to guarantee
identifiability and detection of the switching to substrate limitation
regime.
8.5.2 CONDITIONS FOR ACCURATE SWITCHING POINT
DETECTION
The above mentioned assumptions assure a successful model description of the process
using MR and a rigorous model with small systematic error. Nevertheless, to allow a
proper detection of the switching points in the process, following conditions are to be
considered:
Mechanism Recognition in Escherichia coli cultivations


106

KNOWLEDGE OF THE NORMAL BEHAVIOR OF THE SYSTEM
- The typical behavior of the cell in each regime is known. In addition, the
physical foundation of the complex model offers a basis for evaluation
of the dynamics of the process.
DEFINITIVENESS OF THE CHANGING BEHAVIOR
- The distinguishability test proves that, at least for the simulated
experiments, the submodels are distinguishable. From this it may be
supposed, that the process offers enough information to detect the
difference in bacterial behavior under both regimes.
AVAILABILITY OF AT LEAST ONE OBSERVATION REFLECTING REGIME CHANGE
- At this stage (model validation) offline measurements are fitted and
interpolated to obtain “online” measurements. Again, real time process
monitoring is not considered at this stage. Nevertheless, once the
submodels have shown to describe the different regimes with precision
and allow switching detection, identifiability and distinguishability tests
should be repeated considering online measurements exclusively.
SATISFACTORY STATE OF INFORMATION
- The experiment was carried out in duplicate to assure reproducibility and
a proper form of the variance-covariance matrix. During MR the data set
of one experiment was utilized.
However, due to the nonlinearity of the model and unknown disturbances of the
system, the results obtained by the identifiability and distinguishability test cannot
guarantee that the afore mentioned conditions are fulfilled. It is worth reminding that
the process of MR is strongly dependent of the capability of the complex model to
mirror the real process.
8.5.3 DATA SET
BATCH FERMENTATIONS; 16H, 37°C
Experiments G1 and G2 (replicate) were realized at fermentation temperature of 37°C
with a precultivation of 16h.
The results depicted in Figure 8.11-Figure 8.14 suggest that the cultivation showed a
typical E. coli batch cultivation and good reproducibility when compared against its
replicate (G2).
The measurements obtained from experiment G1 were selected for MR validation. The
acetate concentration reached a maximal concentration at 23.54 mMol/l. Uptake of
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


107

acetate started at t = 5.68 h (341min). The specific acetate production rate fitted curve
had its peak at t = 94 min with 9.9 mMol/(g*h). Acetate uptake started when the
glucose concentration reached 0.5g/l.
Glucose was completely metabolized at t = 6.7 h (400 min). The maximal glucose
uptake rate reached 4.24 [g/l/h] at t = 1.8 h (109 min). The biomass curve showed the
typical shape of growth in batch fermentation with E. coli. The lag-phase endured 2 h
(120 min). Exponential growth phase was present until t = 6.6h (397 min) when
maximum acetate concentration was measured. This point was also the onset of the
stationary phase. The biomass yield had a maximal peak at t = 3h (184 min) at a value of
0.33 g/g.
Specific intracellular NAD
+
concentration reached 0.007 mMol/g at t = 4.1h (247 min).
The respiration coefficient had its maximum at the same time. The respiration
coefficient reached a minimum at t = 1.7 h (100 min) when the specific acetic acid
production rate was maximal.


Figure 8.11: Experimental results batch experiment G1. Part I:
Dry biomass and glucose concentrations
0,0
0,5
1,0
1,5
2,0
2,5
3,0
3,5
4,0
0
2
4
6
8
10
12
0 200 400 600
D
r
y

B
i
o
m
a
s
s
[
g
/
l
]
G
l
u
c
o
s
e
[
g
/
l
]
time [min]
Glucosefit[g/l] GlucoseG1[g/l] GlucoseG2[g/l]
D.Biofit[g/l] BioG1[g/l] D.BioG2[g/l]
Mechanism Recognition in Escherichia coli cultivations


108


Figure 8.12: Experimental results batch experiment G1. Part II:
Specific concentration of acetic acid



Figure 8.13: Experimental results batch experiment G1. Part III:
Outgas concentrations

0
5
10
15
20
-5
0
5
10
15
20
0 200 400 600
A
c
e
t
i
c

a
c
i
d

[
m
M
o
l
/
l
]
d
A
c
/
d
t
/
D
.
B
i
o
[
m
M
o
l
/
g
h
]
;

d
A
c
/
d
t
[
m
M
o
l
/
l
h
]
time [min]
dAc/dt/D.Biofit[mMol/g] dAc/dt/D.BioG1[mMol/gh]
dAc/dt/D.BioG2[mMol/g] dAc/dtfit[mMol/lh]
Acfit[mMol/l] AcG1[mMol/l]
-0,5
-0,3
-0,1
0,1
0,3
0,5
0,7
0,9
1,1
1,3
1,5
0
5
10
15
20
25
30
35
40
45
50
0 200 400 600
R
Q
f
i
t

[
-
]
q
O
2
,

q
C
O
2

[
m
M
o
l
/
g
h
]
time [min]
qO2fit[mMol/gh] qO2G1[mMol/gh] qO2G2[mMol/gh]
qCO2fit[mMol/gh] qCO2G1[mMol/gh] qCO2G2[mMol/gh]
RQfit[-]
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


109


Figure 8.14: Experimental results batch experiment G1. Part IV:
Metabolite concentration
8.5.4 RECOGNITION OF OVERFLOW AND SUBSTRATE
LIMITATION REGIMES
MR was applied considering the following measured variables:
- Extracellular acetate concentration (mMol/l)
- Extracellular substrate concentration (g/l)
- Biomass concentration (g/l)
INITIAL INTERVAL
As previously discussed, the initial interval is characterized by overflow glucose uptake.
The submodel adapted for this regime (OF), has a higher identifiability than the
substrate limitation model. This is mainly because the submodel considers the substrate
and oxygen uptake enzymes to be constant.
MR INITIALIZATION
Identifiability was obtained after 16 interpolated measurement points same as
distinguishability. This permitted initiation of MR after 2.3 h.
DETECTION OF SWITCHING POINTS
Detection of the switch between overflow and substrate limitation regime was carried
out with the method for fault detection through parameter identification. The constant
threshold was set equal to a residual value of 0.3.
0
10
20
30
40
50
60
0,000
0,001
0,002
0,003
0,004
0,005
0,006
0,007
0 200 400 600
A
c
/
D
.
B
i
o
[
m
M
o
l
/
g
]
N
a
d
H
/
D
.
B
i
o
[
m
M
o
l
/
l
]
time [min]
NadH/D.Biofit[mMol/g] Ac/Biofit[mMol/g]
Ac/D.BioG1[mMol/g] Ac/D.BioG2[mMol/g]
Mechanism Recognition in Escherichia coli cultivations


110

INITIAL CONDITIONS OF THE INTERVAL K+1
Due to assumption of instant switch of the regime, the initial values of the non
measured variables had to be included in the parameter estimation problem. Still, the
program showed no problem to find an accurate solution.
SWITCHING INTERVAL
The experimental results confirmed that considering an instant switch of regime allows
an accurate description of the process. Still, the physical evidence suggests the opposite.
The effects of this “mistaken” assumption are buffered due to the estimation of the
initial values of the substrate limitation regime (k+1).
8.5.5 SIMULATIONS VS. EXPERIMENTAL DATA
The submodel based on overflow metabolism assumption can fit the data obtained
from the experiments with some accuracy, the simulation with the estimated parameters
is depicted in Figure 8.15. Nonetheless, it is clear that the submodel lacks the capacity to
describe the process after hour 7. Besides, parameters, like Yield from substrate to
biomass for example, need to be lower drastically, which is reflected in the wrong
description of substrate concentration form hour 5.5. Still, it can be clearly seen, that the
submodel can describe the first phase of the process with high accuracy.

Figure 8.15: OverFlow submodel fitted against experimental data.
Contrary to the overflow submodel, the submodel, which considers that all substrate
can be consumed efficiently through the trcyclyc acid cycle, fails to describe the initial
phase of the cultivation Figure 8.16. Since the bacteria are not set to any real limitations
for substrate uptake, the model consider an unreal growth and substrate consumption.
In addition, no acetate is produced and the cultivation stops at hour 7.5 approximately.
0 1 2 3 4 5 6 7 8 9 10
0
2
4
6
8
10
OverFlow
time [h]
A
c
e
t
a
t
e

A

[
m
M
o
l
/
l
]
,

S
u
b
s
t
r
a
t
e

S

[
g
/
l
]
,
B
i
o
m
a
s
s

X

[
g
/
l
]


A
X
S
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


111


Figure 8.16: Submodel for the description of growth under substrate limitation fitted against
experimental data.
Finally, the starvation submodel was also tested Figure 8.17. This submodel has no
possibility to describe the fermentation. The optimizer maximizes the yield coefficients
to the parameter boundaries without achieving to describe the cultivation.

Figure 8.17: Starvation condition described by the corresponding submodel fitted against
experimental data.
8.5.6 RESULTS
MR was applied considering the following measured variables:
- Extracellular acetate concentration (mMol/l)
- Extracellular substrate concentration (g/l)
- Biomass concentration (g/l)
0 1 2 3 4 5 6 7 8 9 10
0
2
4
6
8
10
SL
time [h]
A
c
e
t
a
t
e

A

[
m
M
o
l
/
l
]
,

S
u
b
s
t
r
a
t
e

S

[
g
/
l
]
,
B
i
o
m
a
s
s

X

[
g
/
l
]


A
X
S
0 1 2 3 4 5 6 7 8 9 10
0
2
4
6
8
10
Starvation
time [h]
A
c
e
t
a
t
e

A

[
m
M
o
l
/
l
]
,

S
u
b
s
t
r
a
t
e

S

[
g
/
l
]
,
B
i
o
m
a
s
s

X

[
g
/
l
]


A
X
S
Mechanism Recognition in Escherichia coli cultivations


112

The following model sequence was detected by MR Figure 8.18. MR detected the
initiation of the regime of growth under substrate limitation at t = 5.8h (348 min). This
results are consistent with the evidence shown by metabolite accumulation and acetate
uptake t = 5.7 (341 min).

Figure 8.18: Experimental validation of the MR approach.
MR enables the application of Lins model for the description of E. coli batch
cultivations. Nevertheless, due to the decrease the process constant parameters in the
general structure, and the highly relaxed boundaries required, this set of submodels
cannot guarantee adequate recognition of fed-batch processes.

0 1 2 3 4 5 6 7 8 9 10
0
2
4
6
8
10
time [h]
A
c
e
t
a
t
e

A

[
m
M
o
l
/
l
]
,

S
u
b
s
t
r
a
t
e

S

[
g
/
l
]
,
B
i
o
m
a
s
s

X

[
g
/
l
]
Batch Fermentation G1


4 6 8 10 12 14 16 18 20
0
5
10
15
20
A-Criteron
t[h]
A
-
C
r
i
t
.


0 2 4 6 8 10 12 14 16 18 20
0
5
10
15
20
t [h]
C
(
X
)

[
g
r
/
l
]
Batch fermentation


5 Par
4 Par
3 Par
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


113

Figure 8.19: Identifiability test considering white noise,
standard deviation of 5% in all measurements
To test the identifiability of the submodels with the new highly relaxed boundaries and
the reduced general structure we calculate the FIM and evaluate the A-criterion in
relation to the number of measurements. The frequency considered for data recollection
was 1 min with interpolated data set. The quality of the measurements was simulated
with white noise (standard deviation 5%). The results show that the minimal period
required to obtain a parameter estimation with less than 10% average standard deviation
in all the parameters is more than 4 hours Figure 8.19.
The identifiability is reduced as we add parameters to the regime variable structure.
Although this condition can be assured in batch cultivations, it represents a major
drawback for fed-batch applications. A new model is required to build a set of
submodels with global optimal structure and higher identifiability.
8.6 CONCLUSIONS
MR proved to be able to transform the model originally proposed by Lin into one
suitable for parameter estimation and its application in batch processes. The first
advantage showed by the application of MR is that the model sequence allows a
reduction on the number of parameters to be fitted. Second, an indicator of the current
cell state is obtained. MR is able to report whether the metabolism is dominated by
overflow (acetate production) or by conversion of metabolites in the citric acid cycle.
However, the model has limitations in its application at processes characterized by a
highly dynamic change in the cell physiology and genetic expression faster than in batch
processes and in substrate-limited fed batch processes operated at rather constant
conditions. When the model had been applied to fed-batch processes where switches of
the glucose availability per cell were performed, the submodels were not suitable to
reflect the time course of fermentation parameters (data not shown). The reason for this
is a likely change in the protein expression in the organism, hence altering the
concentration of enzymes of submetabolisms in a different way. The yield coefficients
(splits to different submetabolims) are changing. A complex model with a proper
description capacity of the dynamic system is required. An improvement of the model
considering at least a part of the regulatory behavior of the cell in a model which is
based on kinetic rate equations, would reduce the systematic error of the present model.
The model should include the glycolysis as the different substrate supply is one
characteristic of the different phases. Also, the main carbon metabolism including the
TCA as a generator of major branch point intermediates and products should be
included. Finally, the acetate synthesis should be modeled in detail including the
excretion to the environment and the assimilation. Since the intracellular acetate
accumulation is regarded as being crucial and a reason for a change in the carbon
Mechanism Recognition in Escherichia coli cultivations


114

distribution in the metabolism, any suitable model should aim to predict the dynamics
of acetate synthesis.
8.7 FUTURE WORK
So far, approaches in literature are rarely covering the complete main carbon
metabolism including the submetabolisms mentioned previously. A recent paper [210]
describes the model fusion for the glycolysis, the pentose-phosphate pathway, the
tricarboxylic acid cycle, the glyoxylate shunt. Further regulatory mechanisms were
included as the interaction of the cyclic AMP receptor protein, the catabolite repressor
(Cra) , the pyruvate dehydrogenase repressor and the repressor for the operon of acetate
kinase. The model was applied for transient phases of an E. coli batch process. Another
own approach combined a model of glycolysis [175] and the TCA cycle[211] and a
module describing acetate synthesis and excretion and reconversion. Each submodel is
regarded as having several rate limiting reaction steps depending on the substrate and
product concentrations present, but are characterized by a similar degree and change of
gene expression. This is introduced as a general description factor for each submodel.
This assumption is suitable to reduce the set of parameters and to contribute to the fact
that many regulatory patterns of the cells are still not understood and cannot be related
to single reactions. Also, the accumulation of intermediates, which in most cases cannot
be measured accurately, is not considered.
Again the reduction to create a new set of submodels is required. So far, these models
are often applied at continuous cultivations of fed batch processes characterized by a
high degree of stability concerning the environment of the cells in the bioreactor.
Transient changes cannot be predicted but only in a very narrow range close to the
conditions of parameter estimation. It is believed that the combination of the
consideration of regulatory patterns, model reduction and mechanism recognition can
contribute to the great need of process related models that are able to enhance the
degree of control.
Besides the model-based contribution, also the integration of cell physiology
measurements that are representing multiple parameters of the cells can be suitable to
establish reduced models that increase control possibilities. Many attempts are made to
establish more robust on line and at line methods which are based on optical methods (on
line microscopy) for the determination of cell size and shape, and also electrooptical
methods for measuring the polarizability of cells [172, 189]. The information that can be
gained from such methods seems to be very suitable to integrate them as activation or
repression factors to submodels. In combination with MR, the magnitude of these
factors can be assumed separately for each phase of the process while reducing the
number or rate equations needed to include all possible ways of interaction at transient
cell stages.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


115

9 CONCLUSIONS AND OUTLOOK
9.1 CONCLUSIONS
The principal achievement of this thesis is the proposal of an alternative approach to
analysis and implementation of complex rigorous models in processes with restricted
state of information. Firstly, it is shown how model simplification may lead to more
efficient process description and better understanding of the system. Secondly, the
importance of considering the state of during model building is stated. This work
proposes some insight to essential questions in biotechnology and chemical engineering:
- How accurate does a model need to be?
- How can we take advantage of the limitations of the model?
- What is the simplest manner to describe dynamic processes?
Although a general answer to these questions is still not at hand, some important
advances can be concluded from the results presented in this manuscript. Attention is
directed to make models useful and not as accurate or complex as possible. Moreover, it
is shown that, if a model is simplified properly based on a systematic analysis and
process know how, a better understanding as well as a more accurate description of the
system are enabled.
The approach to Mechanism Recognition represents a step forward in what can be
considered the right direction toward the future of efficient modeling. Two case studies
prove that the use of deep system understanding as the basis for development of
simplified models for description of the process offers important advantages in
comparison to model building through direct analysis of data correlations and heuristics.
In the first case, accurate detection of organic matter depletion with standard measuring
methods is enabled. MR is able to detect depletion of organic matter allowing a better
control of the SBR process for WWTPs increasing process efficiency and water quality.
In the second case, a model for the description of E. coli cultivation is analyzed and
compared against experimental data. The submodels created, present a significantly
higher identifiability and allow a better analysis of the complex model as well as the
cultivation. In both examples the advances of MR are clearly stated offering a systematic
methodology for an improved study of the process and its mathematical representation.
Finally, an additional advantage of process description with multiple submodels is a
better understanding of the system. As long as the physical background of the
Conclusions and outlook

116

submodels is understood, an analysis of the description capacity of the submodels is
also an indicator of the phenomenon dominating the process. This characteristic can be
exploited to detect non-measurable states of the process. As shown in both case studies
presented in this thesis, special models can be developed to detect non desired
conditions which cannot be measured directly. These simplified versions of the complex
models offer a better process understanding and an increased applicability for
engineering purposes than their complex versions.
9.2 OUTLOOK
The advantages of MR for small and large scale problems have been presented and
proved in two case studies. The method is still in a developing stage and further work is
required to maximize efficiency of MR. Nevertheless, it is the scope of this thesis to
motivate further research in this field.
The essence of this project was to test a new hypothesis, search for possible manners to
perform it, and find examples for its application. Because of the nature of this project,
its real success can be quantified by the quality and extension of the outlook. In a sense,
the real objective of this work may be understood as developing a realistic proposition
of possible manners to develop efficient methods to MR. In order to create a complete
and efficient MR program able to facilitate modeling, structure analysis, parameter
estimation, design of experiments, and ultimately process monitoring in an efficient and
systematic manner, further research is required.
9.2.1 GENERAL THEORY FOR SUBMODEL GENERATION
Model reduction theory, offers a handful of techniques to study the dynamics of models
and increase its tractability. Very sophisticated mathematical methods are available in
literature. In fact, both case studies presented in this work prove that model reduction
techniques in combination with process understanding may give rise to simpler but
accurate submodels. Still, significant effort is required to build these reduced versions of
the complex model. This represents a major drawback in MR application. Building of
submodels, as it stands, is not only time consuming but misses the software tools
required for the scientists of different fields to collaborate in the creation of proper
versions. Despite important achievements in software tools, which enhance this
collaboration, e.g. MOSAIC and SBPD toolbox, much time needs to be invested in the
translation of codes and experimental information to create a common modeling
environment.

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


117

In addition, two main aspects of model reduction require further development to reduce
the effort of submodel building:
1) Model reduction of dynamic nonlinear systems.
As previously stated, in nonlinear systems under dynamic conditions a general
approach to reduce large nonlinear systems with reasonable computer expenses
is still missing. Heuristics and scientist intuition are still the key to the reduction
of such process, resulting not only in large efforts, but also in suboptimal and
locally restricted solutions with no general validity.
2) Model reduction based on the state of information.
Present approaches for model reduction do not take the state of information of
the process into account. Furthermore, for MR the difference between the
quality of the data set obtained in laboratory experiments in comparison to
considerably reduced. This information is very important to create an optimal
set of submodels and is not consider in the state of the art methods for model
reduction.
9.2.2 SWITCHING POINT IDENTIFICATION
At this early development stage, a simple but robust method taken from the field of
model-based fault identification was applied to detect the time point were the switch of
regimes takes place. The parameter identification method showed to fulfill the
requirements at this stage of development, accurately detecting the change of regimes.
Nevertheless, this stage offers large optimization potential. A study of modern
techniques and a proper adaptation for the case of MR is required to offer an efficient
switching point identification approach.
Advanced methods of model predictive control, e.g. moving horizon, should also be
tested since they offer alternative approaches for efficient parameter estimation and
accurate regime detection in processes with longer time spans.
9.2.3 GLOBAL OPTIMIZATION
In the case of highly nonlinear models, model structure analysis and parameter
estimation require proof of global convergence. In this work a stochastic optimization
method was applied in an effort to avoid local minima and increase the probability of
global solution. Still, stochastic optimization methods cannot assure convergence to
global optimality. At its current development stage, global optimizers for nonlinear
dynamic systems require large computer expenses and can by no means be applied for
online purposes. Still, global optimizers should be implemented at the early stages of
MR. Use of global optimizers is essential in model structure analysis, design of
Conclusions and outlook

118

experiments, and model distinguishability studies. It is only after knowing the global
best solution for each case that proper conclusion can be made.
9.2.4 ONLINE MONITORING
Finally, this work proves the great potential on the application of MR to enhance
monitoring of none measurable variables in dynamic processes. In particular, detection
of quasi steady state regimes, which cannot be detected by standard measurements,
offers a great field of application. The present approach has been developed taking into
account limitations of online applications in all stages. From this, it can be assumed that
MR can be applied to industrial processes where a certain regime must be maintained to
assure optimal process conditions, but the data obtained does not offer direct
information about the state of the “optimal” regime. An example is E. coli cultivations
where optimal growth is known to occur under substrate limitation where the state of
the cell cannot be measured directly. Still, indirect measurements in combination with
the correct set of submodels enable the detection of a change in the state of the cells
optimizing feed strategies to maximize cell density and recombinant protein production.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


119

10 Appendix
A. KINETIC MODEL FOR E. COLI
MODEL AS PUBLISHED BY LIN [91]

=
A 1

= −

+
A 2

= −


A 3

=


− −

∗ ∗
A 4

= −

+


A 5

=

/
∗ −

(

+)
A 6

=

/
∗ −

(

+)
A 7
The initial values

0
=

+

A 8

0
=

+

A 9

=

/

A 10
Calculation of the subsidiary variables:

=

0

A 11

=

0

A 12
Appendix

120

=

+

A 13

=

A 14

= min[

,

]
A 15

=

/

A 16

=

A 17

=



/

A 18
To avoid a larger

than allowed by the maximal oxygen uptake capacity of the
bacteria, whenever

>

we change the values to

=

and recalculate:

=

/

A 19

=


/_


/

−1

A 20

=



A 21
Because of the recalculations, the physical limitation of the real maintenance coefficient
has to be checked once more. If

>

, then

=

=

and

=
0

=

A 22

=

+

A 23

=


/

A 24

=

A 25

=


/

A 26

=

A 27
Finally, if

>

/

then,

=

/

and:

=

1 −
/

A 28
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


121

=

A 29

=


/

A 30

=

+


/

A 31
=

/

+


/

+


/

A 32
Nomenclature
- specific growth rate [h
-1
]
- acetate concentration [g L
-1
]
- dissolved oxygen tension [%]
-

,

concentration of the enzymes substrate/oxygen consumption
[gL−1]
- feed flow rate [L h
-1
]
- constant derived from the Henry‟s law [14,000 L g
-1
]
- saturation constant [g L
-1
]
-

volumetric oxygen transfer coefficient [h
-1
]
-

specific acetate production of consumption rate [g g
-1
h
-1
]
-

specific degradation rate [g g
-1
h
-1
]
-

specific oxygen uptake rate [g g
-1
h
-1
]
-

specific substrate (glucose) uptake rate [g g
-1
h
-1
]
-

,

ratio of enzyme for oxygen or substrate consumption per
biomass
- substrate (glucose) concentration [g L
-1
]
- cultivation time [h]
- fermenter volume [L]
- biomass concentration [g L
-1
]
- yield [g g
-1
]

Subscripts
- 0 Initial concentration
- acetate
- anabolic
- consumption
- capacity
- degradation
- enzyme
- energetic
Appendix

122

- inlet concentration
- maintenance
- maximum
- real maintenance
- oxygen
- overflow
- oxidative
- production
- substrate (glucose)
- biomass
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


123

B. OVERFLOW SUBMODEL
The nomenclature is identical to the complete Lin model (appendix A)

=
B 1

= −

+
B 2

= −


B 3

=


− −

∗ ∗
B 4

= −

+


B 5

=

+

B 6

=

+

B 7
The initial values

=

/

B 8
CALCULATION OF THE SUBSIDIARY VARIABLES

=

0

B 9

=

0

B 10

=

+

B 11

=

B 12

=

B 13
Because of the overflow condition, we assume:

=

:

=

/

B 14
Appendix

124

=


/_


/

−1

B 15

=

B 16

=

B 17

=

+

B 18

=


/

B 19

=

B 20

=


/

B 21

=

B 22
=

/

+


/

+


/

B 23

An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


125

C. SUBSTRATE LIMITATION
SUBMODEL
The nomenclature is identical to the complete Lin model (appendix A)

=
C 1

= −

+
C 2

= −


C 3

=


− −

∗ ∗
C 4

= −

+


C 5

=

/
∗ −

(

+)
C 6

=

/
∗ −

(

+)
C 7
The initial values

0
=

+

C 8

0
=

+

C 9

=

/

C 10
Calculation of the subsidiary variables:

=

0

C 11

=

0

C 12
Appendix

126

=

+

C 13

=

C 14

=

C 15

=

/

C 16

=

C 17

=


/

C 18

= 0
C 19

=

+

C 20

= 0
C 21

= 0
C 22

=


/

C 23

=

C 24

= 0
C 25

=

+


/

C 26
=

/

+


/

+


/

C 27
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


127

D. STARVATION SUBMODEL
The nomenclature is identical to the complete Lin model (appendix A)

=
D 1

= −

+
D 2

= −


D 3

=


− −

∗ ∗
D 4

= −

+


D 5

:
D 6

:
D 7
The initial values

=

/

D 8
Calculation of the subsidiary variables:

=

0

D 9

=

0

D 10

=

+

D 11

=

D 12

=

D 13

=

/

D 14

=

/

D 15
Appendix

128

=


/_


/

−1

D 16

=



D 17

= 0
D 18

= 0
D 19

= 0
D 20

= 0
D 21

= 0
D 22

= 0
D 23
=

/

+


/

+


/

D 24



An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


129

BIBLIOGRAPHY
PERSONAL PUBLICATIONS
Electrooptical monitoring of cell polarizability and cell size in aerobic Escherichia coli batch cultivations
Stefan Junne, M. Nicolas Cruz-Bournazou, Alexander Angersbach and Peter Götz
Journal of Industrial Microbiology & Biotechnology
Volume 37, Number 9, 935-942, DOI: 10.1007/s10295-010-0742-5

From data to mechanistic-based models
M. N. Cruz-Bournazou, H. Arellano-Garcia, J. Schöneberger, G. Wozny
Chemical and Process Engineering
2009, 30, 291–306

A Novel Approach to Mechanism Recognition in Escherichia Coli Fed-Batch Fermentations
M.N. Cruz Bournazou, H. Arellano-Garcia, J.C. Schöneberger, S. Junne, P. Neubauer
and G. Wozny
Computer Aided Chemical Engineering
Volume 27, 2009, Pages 651-656 10th

Model reduction of the ASM3 extended for two-step nitrification
M. N. Cruz Bournazou H. Arellano-Garcia G. Wozny, G. Lyberatos, C. Kravaris
Computer Applications in Biotechnology
Volume11 Part 1, 10.3182/20100707-3-BE-2012.0047

Neuer Ansatz zur Mechanismenerkennung in E. coli Fed-Batch-Kultivierungen
M. N. Cruz Bournazou, H. Arellano-Garcia, S. Junne, G. Wozny, P. Neubauer
Chemie Ingenieur Technik
Volume 82, Issue 9, page 1503, September, 2010,DOI: 10.1002/cite.201050496

Modellreduktion des erweiterten ASM3-Modells für die zweistufige Nitrifikation
M. N. Cruz Bournazou M. Sc., H. Arellano-Garcia Dr.-Ing., G. Wozny Prof. , G.
Lyberatos Prof., C. Kravaris Prof.

Chemie Ingenieur Technik
Volume 82, Issue 9, page 1540, September, 2010,DOI: 10.1002/cite.201050496

ASM3 extended for two-step nitrification-denitrification: a model reduction for Sequencing Batch
Reactors
M. N. Cruz Bournazou, H. Arellano-Garcia, G. Wozny, G. Lyberatos, C. Kravaris
Journal of Chemical Technology and Biotechnology (in press: JCTB 11-0602)

Bibliography

130

REFERENCES
1. FTD, BASF erhöht Forschungsausgaben ed. F.T. Deutschland. 2011,
http://www.ftd.de/unternehmen/industrie/:investitionen-beim-chemiekonzern-basf-
erhoeht-forschungsausgaben/60007679.html
2. DFG, DFG-Jahresbericht 2010, ed.
http://www.alphagalileo.org/ViewItem.aspx?ItemId=108368&CultureCode=de. 2010.
3. Barton, P.I., The modelling and simulation of combined discrete/continuous processes. PhD diss.,
University of London, 1992.
4. fda.gov, Guidance for Industry Process Validation: General Principles and Practice. 2011.
5. Kitano, H., Systems Biology: A Brief Overview. 2002. p. 1662-1664.
6. Kitano, H., Computational systems biology. Nature, 2002. 420(6912): p. 206-210.
7. Ropers, D., V. Baldazzi, and H. de Jong, Model Reduction Using Piecewise-Linear
Approximations Preserves Dynamic Properties of the Carbon Starvation Response in Escherichia coli.
IEEE IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2009.
8. Box, G.E.P. and W.J. Hill, Discrimination among mechanistic models. Technometrics, 1967.
9(1): p. 57-71.
9. Preisig, H.A., Constructing and maintaining proper process models. Computers & Chemical
Engineering, 2010. 34(9): p. 1543-1555.
10. Li, D., O. Ivanchenko, N. Sindhwani, E. Lueshen, and A.A. Linninger, Optimal Catheter
Placement for Chemotherapy. Computer Aided Chemical Engineering. 28: p. 223-228.
11. Brendel, M., D. Bonvin, and W. Marquardt, Incremental identification of kinetic models for
homogeneous reaction systems. Chemical Engineering Science, 2006. 61(16): p. 5404-5420.
12. Balsa-Canto, E., A.A. Alonso, and J.R. Banga, An iterative identification procedure for dynamic
modeling of biochemical networks. BMC Systems Biology, 2010. 4(1): p. 11.
13. Rippin, D.W.T., Batch process systems engineering: a retrospective and prospective review.
Computers & chemical engineering, 1993. 17: p. 1-13.
14. Kravaris, C., R.A. Wright, and J.F. Carrier, Nonlinear controllers for trajectory tracking in batch
processes. Computers & chemical engineering, 1989. 13(1-2): p. 73-82.
15. Peterson, C.L., M. Feldman, R. Korus, and D.L. Auld, Batch type transesterification process
for winter rape oil. Applied engineering in agriculture (USA), 1991. 110(3): p. 971-976.
16. Choe, Y.S. and W.L. Luyben, Rigorous dynamic models of distillation columns. Industrial &
Engineering Chemistry Research, 1987. 26(10): p. 2158-2161.
17. Gep, B.O.X. and N.R. Draper, Empirical model building and response surfaces. 1987, John
Wiley and Sons, New York, USA.
18. Atkinson, A.C. and A.N. Donev, Optimum Experimental Designs. 1992: Oxford University
Press.
19. Sjoberg, J., Q.H. Zhang, L. Ljung, A. Benveniste, B. Delyon, P.Y. Glorennec, H.
Hjalmarsson, and A. Juditsky, Nonlinear Black-Box Modeling in System-Identification-A
Unified Overview. Automatica, 1995. 31(12): p. 1691-1724.
20. Suykens, J.A.K. and J. Vandewalle, Nonlinear Modeling: Advanced Black-Box Techniques.
1998: Kluwer Academic Publishers.
21. Jones, D.R., M. Schonlau, and W.J. Welch, Efficient Global Optimization of Expensive Black-
Box Functions. Journal of Global Optimization, 1998. 13(4): p. 455-492.
22. Verheijen, P.J.T., Model Selection: an overview of practices in chemical engineering. Dynamic
Model Development, 2003: p. 85–104.
23. Frank, P.M., Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy:
A survey and some new results. Automatica, 1990. 26(3): p. 459-474.
24. Tarantola, A., Inverse Problem Theory and Methods for Model Parameter Estimation. 2005,
Society for Industrial Mathematics.
25. Esposito, W.R. and C.A. Floudas, Global optimization for the parameter estimation of
differential-algebraic systems. Ind. Eng. Chem. Res, 2000. 39(5): p. 1291-1310.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


131

26. Singer, A.B., J.W. Taylor, P.I. Barton, and W.H. Green, Global dynamic optimization for
parameter estimation in chemical kinetics. The Journal of Physical Chemistry A, 2006. 110(3):
p. 971-976.
27. Mitsos, A., G.M. Oxberry, P.I. Barton, and W.H. Green, Optimal automatic reaction and
species elimination in kinetic mechanisms. Combustion and Flame, 2008. 155(1-2): p. 118–132.
28. Herold, S., T. Heine, and R. King. An Automated Approach to Build Process Models by
Detecting Biological Phenomena in (fed-) batch Experiments. 2010, Computer Applications in
Biotechnology, 11(1).
29. Forbus, K.D., Qualitative Process Theory. Artificial Intelligence, 1984. 24(1-3): p. 85-168.
30. Forbus, K.D., Qualitative process theory: twelve years after. Artificial Intelligence in
Perspective, 1994. 59(1): p. 15-123.
31. Blom, H.A.P. and Y. Bar-Shalom, The interacting multiple model algorithm for systems with
Markovian switching coefficients. Automatic Control, IEEE Transactions on, 1988. 33(8): p.
780-783.
32. Mazor, E., A. Averbuch, Y. Bar-Shalom, and J. Dayan, Interacting multiple model methods in
target tracking: a survey. Aerospace and Electronic Systems, IEEE Transactions on, 1998.
34(1): p. 103-123.
33. Doucet, A., N.J. Gordon, and V. Krishnamurthy, Particle filters for state estimation of jump
Markov linear systems. Signal Processing, IEEE Transactions on, 2001. 49(3): p. 613-624.
34. Zhang, W., C. Wu, and C. Wang, Qualitative Algebra and Graph Theory Methods for Dynamic
Trend Analysis of Continuous System. Chinese Journal of Chemical Engineering, 2011.
19(2): p. 308-315.
35. Berleant, D. and B.J. Kuipers, Qualitative and quantitative simulation: bridging the gap.
Artificial Intelligence, 1997. 95(2): p. 215-255.
36. Utkin, V., Variable structure systems with sliding modes. Automatic Control, IEEE
Transactions on, 1977. 22(2): p. 212-222.
37. Bhattacharjee, B., D.A. Schwer, P.I. Barton, and W.H. Green, Optimally-reduced kinetic
models: reaction elimination in large-scale kinetic mechanisms. Combustion and Flame, 2003.
135(3): p. 191-208.
38. Escobet, T., D. Feroldi, S. De Lira, V. Puig, J. Quevedo, J. Riera, and M. Serra, Model-
based fault diagnosis in PEM fuel cell systems. Journal of Power Sources, 2009. 192(1): p. 216-
223.
39. De Lira Ramirez, S., V. Puig Cayuela, A.P. Husar, and J.J. Quevedo Casin. LPV model-
based fault diagnosis using relative fault sensitivity signature approach in a PEM fuel cell. 2010:
IEEE Computer Society.
40. Hotop, R., S. Ochs, and T. Ross, Überwachung von Anlagenteilen: Neues Werkzeug ermöglicht
Statusermittlung. atp edition, 2010.
41. Dochain, D. and M. Perrier, Dynamical modelling, analysis, monitoring and control design for
nonlinear bioprocesses. Biotreatment, Downstream Processing and Modelling, 1997: p. 147-
197.
42. Cruz-Bournazou, M., H. Arellano-Garcia, J. Schoneberger, and G. Wozny, From data to
mechanistic-based models. Chemical and Process Engineering Selected full texts, 2009.
30(2): p. 291-306.
43. Schöneberger, J., H. Arellano-Garcia, and G. Wozny, A Novel Approach to Automated
Mechanism Recognition based on Model Discrimination. Chemical Engineering Transactions,
2007. 12
44. Kuntsche, S., T. Barz, R. Kraus, H. Arellano-Garcia, and G. Wozny, MOSAIC a web-
based modeling environment for co de generation. Computers & Chemical Engineering, 2011. 35
(11): p.2257-2273.
45. Schmidt, H., M.F. Madsen, S. Dano, and G. Cedersund, Complexity reduction of biochemical
rate expressions. Bioinformatics, 2008. 24(6): p. 848.
46. Barz, T., S. Kuntsche, G. Wozny, and H. Arellano-Garcia, An efficient sparse approach to
sensitivity generation for large-scale dynamic optimization. Computers and Chemical
Engineering, 2010. http://dx.doi.org/10.1016/j.compchemeng.2010.10.008.
Bibliography

132

47. Violet, N., E. Fischer, T. Heine, and R. King. A Software Supported Compartmental Modeling
Approach for Paenibacillus Polymyxa. Computer Applications in Biotechnology, 2010. 11(1).
48. Silvert, W., Modelling as a discipline. International Journal of General Systems, 2001. 30(3):
p. 261-282.
49. Aris, R., Mathematical modeling: a chemical engineer's perspective. Vol. 1. 1999: Academic Pr.
50. Bayer, B. and W. Marquardt, Towards integrated information models for data and documents.
Computers & chemical engineering, 2004. 28(8): p. 1249-1266.
51. Dochain, D., State and parameter estimation in chemical and biochemical processes: a tutorial.
Journal of process control, 2003. 13(8): p. 801-818.
52. Stewart, W.E., Y. Shon, and G.E.P. Box, Discrimination and goodness of fit of multiresponse
mechanistic models. AIChE journal, 1998. 44(6): p. 1404-1412.
53. Freyre-Gonzalez, J.A. and L.G. Trevino-Quintanilla, Analyzing Regulatory Networks in
Bacteria Nature Educations, 2010. 3(9): p. 24.
54. Hady, L. and G. Wozny, Computer-aided web-based application to modular plant design.
Computer Aided Chemical Engineering, 2010. 28: p. 685-690.
55. Bailey, J.E., Mathematical modeling and analysis in biochemical engineering: past accomplishments
and future opportunities. Biotechnology progress, 1998. 14(1): p. 8-20.
56. Kremling, A., S. Fischer, K. Gadkar, F.J. Doyle, T. Sauter, E. Bullinger, F. Allgöwer,
and E.D. Gilles, A Benchmark for Methods in Reverse Engineering and Model Discrimination:
Problem Formulation and Solutions. Genome Research, 2004. 14: p. 1773-1785.
57. Csete, M.E. and J.C. Doyle, Reverse engineering of biological complexity. Science, 2002.
295(5560): p. 1664.
58. Bardow, A. and W. Marquardt, Incremental and simultaneous identification of reaction kinetics:
methods and comparison. Chemical Engineering Science, 2004. 59(13): p. 2673-2684.
59. Kitano, H., Systems biology: a brief overview. Science, 2002. 295(5560): p. 1662.
60. Spieth, C., N. Hassis, and F. Streichert. Comparing mathematical models on the problem of
network inference. ACM, 2006.
61. Hecker, M., S. Lambeck, S. Toepfer, E. Van Someren, and R. Guthke, Gene regulatory
network inference: Data integration in dynamic models--A review. Biosystems, 2009. 96(1): p. 86-
103.
62. Heinrich, R. and S. Schuster, The regulation of cellular systems. 1996: Springer Us.
63. Vilela, M., I.C. Chou, S. Vinga, A. Vasconcelos, E. Voit, and J. Almeida, Parameter
optimization in S-system models. BMC systems biology, 2008. 2(1): p. 35.
64. Stamatelatou, K., L. Syrou, C. Kravaris, and G. Lyberatos, An invariant manifold approach
for CSTR model reduction in the presence of multi-step biochemical reaction schemes. Application to
anaerobic digestion. Chemical Engineering Journal, 2009. 150(2-3): p. 462-475.
65. Box, G.E.P., Statistics as a catalyst to learning by scientific method part II-a discussion. Journal of
Quality Technology, 1999. 31(1): p. 16-29.
66. Raduly, B., A.G. Capodaglio, and D.A. Vaccari, Simplification of wastewater treatment plant
models using empirical modelling techniques. Young Researchers 2004, 2004: p. 51.
67. Okino, M.S. and M.L. Mavrovouniotis, Simplification of mathematical models of chemical
reaction systems. Chem. Rev, 1998. 98(2): p. 391-408.
68. Fraser, S.J., The steady state and equilibrium approximations: A geometrical picture. The Journal
of Chemical Physics, 1988. 88: p. 4732.
69. Kuo, J.C.W. and J. Wei, Lumping Analysis in Monomolecular Reaction Systems. Analysis of
Approximately Lumpable System. Industrial & Engineering chemistry fundamentals, 1969.
8(1): p. 124-133.
70. Kaufmann, J., H. Weiss, and Schering (Berlin West), Statistische Methoden in der
experimentellen Forschung Kolloquium, Berlin Kurzfassungen von Vorträgen der Wintersemester
1988/89, 1989/90 und 1990/91. 1991, Berlin: s.n. 254 S.
71. Liao, J.C. and J. Delgado, Advances in metabolic control analysis. Biotechnology Progress,
1993. 9(3): p. 221-233.
72. Roussel, M.R. and S.J. Fraser, Invariant manifold methods for metabolic model reduction. Chaos,
2001. 11: p. 196-206.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


133

73. Kazantzis, N., C. Kravaris, and L. Syrou, A new model reduction method for nonlinear
dynamical systems. Nonlinear Dynamics, 2010. 59(1): p. 183-194.
74. Ranzi, E., M. Dente, A. Goldaniga, G. Bozzano, and T. Faravelli, Lumping procedures in
detailed kinetic modeling of gasification, pyrolysis, partial oxidation and combustion of hydrocarbon
mixtures. Progress in energy and combustion science, 2001. 27(1): p. 99.
75. Dokoumetzidis, A. and L. Aarons, Proper lumping in systems biology models. Systems
Biology, IET, 2009. 3(1): p. 40-51.
76. Segel, L.A. and M. Slemrod, The quasi-steady-state assumption: a case study in perturbation.
SIAM review, 1989. 31(3): p. 446-477.
77. Danos, V., J. Feret, W. Fontana, R. Harmer, and J. Krivine. Abstracting the ODE semantics
of rule-based models: Exact and automatic model reduction. 2009.
78. Kim, Y.H., C.K. Yoo, and I.B. Lee, Optimization of biological nutrient removal in a SBR using
simulation-based iterative dynamic programming. Chemical Engineering Journal, 2008. 139(1):
p. 11-19.
79. Kevrekidis, I.G., C.W. Gear, J.M. Hyman, P.G. Kevrekidid, O. Runborg, and C.
Theodoropoulos, Equation-Free, Coarse-Grained Multiscale Computation: Enabling Mocroscopic
Simulators to Perform System-Level Analysis. Communications in Mathematical Sciences,
2003. 1(4): p. 715-762.
80. Nagy, T. and T. Turányi, Reduction of very large reaction mechanisms using methods based on
simulation error minimization. Combustion and Flame, 2009. 156(2): p. 417-428.
81. Franceschini, G. and S. Macchietto, Model-based design of experiments for parameter precision:
State of the art. Chemical Engineering Science, 2008, 63 (19): p 4846-4872.
82. Georgescu, R., C.R. Berger, P. Willett, M. Azam, and S. Ghoshal. Comparison of data
reduction techniques based on the performance of SVM-type classifiers: IEEE, 2010.
83. Witten, I.H. and E. Frank, Data Mining: Practical machine learning tools and techniques. 2005:
Morgan Kaufmann.
84. Korkel, S., E. Kostina, H.G. Bock, and J.P. Schloder, Numerical Methods for Optimal
Control Problems in Design of Robust Optimal Experiments for Nonlinear Dynamic Processes.
85. Küpper, A., M. Diehl, J.P. Schlöder, H.G. Bock, and S. Engell, Efficient moving horizon
state and parameter estimation for SMB processes. Journal of Process Control, 2009. 19(5): p.
785-802.
86. Barz, T., V. Löffler, H. Arellano-Garcia, and G. Wozny, Optimal determination of steric mass
action model parameters for [beta]-lactoglobulin using static batch experiments. Journal of
Chromatography A.
87. Franceschini, G. and S. Macchietto, Validation of a model for biodiesel production through
model-based experiment design. Industrial & Engineering Chemistry Research, 2007. 46(1):
p. 220-232.
88. Arlot, S. and P. Massart, Data-driven calibration of penalties for least-squares regression. The
Journal of Machine Learning Research, 2009. 10: p. 245-279.
89. Asprey, S.P. and S. Macchietto, Statistical tools for optimal dynamic model building. Computers
and Chemical Engineering, 2000. 24(2-7): p. 1261-1267.
90. Binder, K. and D.W. Heermann, Monte Carlo simulation in statistical physics: an introduction:
Springer.
91. Lin, H.Y., B. Mathiszik, B. Xu, S.O. Enfors, and P. Neubauer, Determination of the
maximum specific uptake capacities for glucose and oxygen in glucose-limited fed-batch cultivations of
Escherichia coli. Biotechnol Bioeng, 2001. 73(5): p. 347-57.
92. Korkel, D.M.S., Numerische Methoden fuer Optimale Versuchsplanungsprobleme bei nichtlinearen
DAE-Modellen, in Ruprecht-Karls-Universitaet Heidelberg. 2002.
93. Kerr, M.K. and G.A. Churchill, Experimental design for gene expression microarrays.
Biostatistics, 2001. 2(2): p. 183-201.
94. Gadkar, K.G., R. Gunawan, and F.J. Doyle Iii, Iterative approach to model identification of
biological networks. BMC Bioinformatics, 2005. 6(155).
Bibliography

134

95. Felix Oliver Lindner, P. and B. Hitzmann, Experimental design for optimal parameter
estimation of an enzyme kinetic process based on the analysis of the Fisher information matrix.
Journal of Theoretical Biology, 2006. 238(1): p. 111-123.
96. Imhof, L.A., Maximin designs for exponential growth models and heteroscedastic polynomial models.
Ann. Statist, 2001. 29(2): p. 561-576.
97. Cooney, M.J. and K.A. McDonald, Optimal dynamic experiments for bioreactor model
discrimination. Applied Microbiology and Biotechnology, 1995. 43(5): p. 826-837.
98. Swinnen, I.A.M., K. Bernaerts, K. Gysemans, and J.F. Van Impe, Quantifying microbial lag
phenomena due to a sudden rise in temperature: a systematic macroscopic study. International
Journal of Food Microbiology, 2005. 100(1-3): p. 85-96.
99. Valdramidis, V.P., A.H. Geeraerd, J.E. Gaze, A. Kondjoyan, A.R. Boyd, H.L. Shaw, and
J.F.V. Impe, Quantitative description of Listeria monocytogenes inactivation kinetics with temperature
and water activity as the influencing factors; model prediction and methodological validation on dynamic
data. Journal of Food Engineering, 2006. 76(1): p. 79-88.
100. Hung, K.K., P.K. Ko, C. Hu, and Y.C. Cheng, A physics-based MOSFET noise model for
circuit simulators. Electron Devices, IEEE Transactions on, 1990. 37(5): p. 1323-1333.
101. Patel, S. and K.K. Pant, Experimental study and mechanistic kinetic modeling for selective
production of hydrogen via catalytic steam reforming of methanol. Chemical Engineering Science,
2007. 62(18-20): p. 5425-5435.
102. Box, G.E.P. and H.L. Lucas, design of experiments in non-linear situations. Biometrika, 1959.
46(1-2): p. 77-90.
103. Box, G.E.P., J.S. Hunter, and W.G. Hunter, Statistics for experimenters. 1978: Wiley New
York.
104. Espie, D. and S. Macchietto, The optimal design of dynamic experiments. AIChE Journal,
1989. 35(2): p. 223-229.
105. Hunter, W.G. and A.M. Reiner, Designs for discriminating between two rival models.
Technometrics, 1965. 7(3): p. 307-323.
106. Buzzi-Ferraris, G. and F. Manenti, Kinetic models analysis. Chemical Engineering Science,
2009. 64(5): p. 1061-1074.
107. Linhart, H. and W. Zucchini, Model selection. 1986: John Wiley & Sons.
108. Zerry, R.u.W., G., MOSAIC: ein Modellierungsserver für die Verfahrenstechnik. 2005.
109. Logsdon, J.S. and L.T. Biegler, Accurate solution of differential-algebraic optimization problems.
Industrial & engineering chemistry research, 1989. 28(11): p. 1628-1639.
110. Zavala, V.M., C.D. Laird, and L.T. Biegler, A fast moving horizon estimation algorithm based
on nonlinear programming sensitivity. Journal of Process Control, 2008. 18(9): p. 876-884.
111. Vaca, M., A. Jimenez-Gutierrez, and R. Monroy-Loperena, Application of collocation
techniques to the design of petlyuk distillation columns. Revista Mexicana de Ingeniería Química,
2006. 5(1): p. 121-127.
112. Holmstrom, K., The TOMLAB Optimization Environment in Matlab 1. 1999.
113. Binder, T. and E. Kostina, Approximate Gauss-Newton Method for Robust Parameter
Estimation. PAMM, 2009. 10(1): p. 531-532.
114. Thanedar, P.B., J.S. Arora, C.H. Tseng, O.K. Lim, and G.J. Park, Performance of some SQP
algorithms on structural design problems. 1986.
115. Drews, A., H. Arellano-Garcia, J. Schöneberger, J. Schaller, G. Wozny, and M. Kraume,
Model-based recognition of fouling mechanisms in membrane bioreactors. Desalination, 2009.
236(1-3): p. 224-233.
116. Kraume, M., U. Bracklow, M. Vocks, and A. Drews, Nutrients removal in MBRs for
municipal wastewater treatment. 2005. p. 391-402.
117. Gimbel, R., A. Nahrstedt, and S. Panglisch, Filtration zur Partikelabtrennung bei der
Wasserreinigung. Chemie Ingenieur Technik, 2008. 80(1 2): p. 35-47.
118. Drews, A., H. Arellano-Garcia, J. Schoneberger, J. Schaller, M. Kraume, and G. Wozny,
Improving the Efficiency of Membrane Bioreactors by a Novel Model-based Control of Membrane
Filtration. Computer Aided Chemical Engineering, 2007. 24: p. 345.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


135

119. Drews, A. and M. Kraume, Process improvement by application of membrane bioreactors.
Chemical Engineering Research and Design, 2005. 83(3): p. 276-284.
120. Goshen-Meskin, D. and I.Y. Bar-Itzhack, Observability analysis of piece-wise constant systems.
I. Theory. Aerospace and Electronic Systems, IEEE Transactions on, 1992. 28(4): p.
1056-1067.
121. Bemporad, A., G. Ferrari-Trecate, and M. Morari, Observability and controllability of piecewise
affine and hybrid systems. Automatic Control, IEEE Transactions on, 2000. 45(10): p.
1864-1876.
122. Isermann, R., Supervision, fault-detection and diagnosis methods–a short introduction. Fault-
Diagnosis Applications, 2011: p. 11-45.
123. Frank, P.M. and X. Ding, Survey of robust residual generation and evaluation methods in observer-
based fault detection systems. Journal of process control, 1997. 7(6): p. 403-424.
124. Sturza, M.A. and A.K. Brown. Comparison of fixed and variable threshold RAIM algorithms.
Navigation, Vol. 35, No. 4, Winter 1988-89.
125. Gray, N.F., Activated sludge: theory and practice. 1990: Oxford University Press.
126. Gernaey, K.V., M.C.M. van Loosdrecht, M. Henze, M. Lind, and S.B. Jørgensen,
Activated sludge wastewater treatment plant modelling and simulation: state of the art.
Environmental Modelling and Software, 2004. 19(9): p. 763-783.
127. Petersen, B., K. Gernaey, M. Henze, and P.A. Vanrolleghem, Evaluation of an ASM 1
model calibration procedure on a municipal-industrial wastewater treatment plant. Journal of
Hydroinformatics, 2002. 4(1): p. 15-38.
128. Gujer, W., M. Henze, T. Mino, T. Matsuo, M. Wentzel, and G. Marais, The activated
sludge model no. 2: biological phosphorus removal. Water science and technology, 1995. 31(2):
p. 1-11.
129. Gujer, W., M. Henze, T. Mino, and M. Van Loosdrecht, Activated sludge model No. 3.
Water Science & Technology, 1999. 39(1): p. 183-193.
130. Sin, G., D. Kaelin, M.J. Kampschreur, I. Takacs, B. Wett, K.V. Gernaey, L. Rieger, H.
Siegrist, and M.C. van Loosdrecht, Modelling nitrite in wastewater treatment systems: a
discussion of different modelling concepts. Water science and technology: a journal of the
International Association on Water Pollution Research, 2008. 58(6): p. 1155.
131. Velmurugan, S., W.W. Clarkson, and J.N. Veenstra, Model-Based Design of Sequencing Batch
Reactor for Removal of Biodegradable Organics and Nitrogen. Water Environment Research,
2010. 82(5): p. 462-474.
132. Balku, S., M. Yuceer, and R. Berber, Control vector parameterization approach in optimization of
alternating aerobic-anoxic systems. Optimal Control Applications and Methods, 2009. 30(6):
p. 573-584.
133. Artan, N. and D. Orhon, Mechanism and design of sequencing batch reactors for nutrient removal.
2005: Intl Water Assn.
134. Mace, S. and J. Mata-Alvarez, Utilization of SBR technology for wastewater treatment: an
overview. Ind. Eng. Chem. Res, 2002. 41(23): p. 5539-5553.
135. Office of Water Washington, D.C., Wastewater Technology Fact Sheet Sequencing Batch
Reactors. EPA 832-F-99-073, 1999.
136. Cruz Bournazou, M.N., H. Arellano-Garcia, G. Wozny, G. Lyberatos, and C. Kravaris,
ASM3 extended for two-step nitrification-denitrification: a model reduction for Sequencing Batch
Reactors. Journal of Chemical Technology and Biotechnology 2012, DOI:
10.1002/jctb.3694.
137. Wilderer, P.A., R.L. Irvine, and M.C. Goronszy, Sequencing batch reactor technology. 2000:
Intl Water Assn.
138. Luccarini, L., G.L. Bragadin, G. Colombini, M. Mancini, P. Mello, M. Montali, and D.
Sottara, Formal verification of wastewater treatment processes using events detected from continuous
signals by means of artificial neural networks. Case study: SBR plant. Environmental Modelling
& Software. 25(5): p. 648-660.
139. Ketchum, L.H., Design and physical features of sequencing batch reactors. Water Science and
Technology, 1997. 35(1): p. 11-18.
Bibliography

136

140. Bungay, S., M. Humphries, and T. Stephenson, Operating strategies for variable flow
sequencing batch reactors. Water and Environment Journal, 2007. 21(1): p. 1-8.
141. Upadhyayula, S., Alkylation of Benzene with Isopropanol over SAPO-5: A Kinetic Study. The
Canadian Journal of Chemical Engineering, 2008. 86: p. 207-213.
142. Mak, W.C., C. Chan, J. Barford, and R. Renneberg, Biosensor for rapid phosphate monitoring
in a sequencing batch reactor (SBR) system. Biosensors and Bioelectronics, 2003. 19(3): p.
233-237.
143. Barton, P.I. and C.K. Lee, Modeling, simulation, sensitivity analysis, and optimization of hybrid
systems. ACM Transactions on Modeling and Computer Simulation (TOMACS), 2002.
12(4): p. 256-289.
144. Akin, B.S. and A. Ugurlu, Monitoring and control of biological nutrient removal in a sequencing
batch reactor. Process Biochemistry, 2005. 40(8): p. 2873-2878.
145. Turk, O. and D.S. Mavinic, Preliminary assessment of a shortcut in nitrogen removal from
wastewater. Canadian Journal of Civil Engineering/Revue Canadienne de Genie Civil,
1986. 13(6): p. 600-605.
146. Katsogiannis, A.N., M.E. Kornaros, and G.K. Lyberatos, Adaptive optimization of a
nitrifying sequencing batch reactor. Water Research, 1999. 33(17): p. 3569-3576.
147. Bourgeois, W., J.E. Burgess, and R.M. Stuetz, On-line monitoring of wastewater quality: a
review. Journal of chemical technology and biotechnology, 2001. 76(4): p. 337-348.
148. Vanrolleghem, P.A. and D.S. Lee, On-line monitoring equipment for wastewater treatment
processes: state of the art. Automation in Water Quality Monitoring, 2003. 47(2): p. 1-34.
149. Kaelin, D., R. Manser, L. Rieger, J. Eugster, K. Rottermann, and H. Siegrist, Extension of
ASM3 for two-step nitrification and denitrification and its calibration and validation with batch tests
and pilot scale data. Water Research, 2009. 43(6): p. 1680-1692.
150. Shampine, L.F., M.W. Reichelt, and J.A. Kierzenka, Solving index-I DAEs in MATLAB
and Simulink. SIAM review, 1999. 41(3): p. 538-552.
151. El-Mansi, E.M.T. and W.H. Holms, Control of carbon flux to acetate excretion during growth of
Escherichia coli in batch and continuous cultures. Journal of general microbiology, 1989.
135(11): p. 2875.
152. Diez-Gonzalez, F. and J.B. Russell, The ability of Escherichia coli O157:H7 to decrease its
intracellular pH and resist the toxicity of acetic acid. Microbiology, 1997. 143: p. 1175-1180.
153. Luli, G.W. and W.R. Strohl, Comparison of growth, acetate production, and acetate inhibition of
Escherichia coli strains in batch and fed-batch fermentations. Applied and environmental
microbiology, 1990. 56(4): p. 1004.
154. Wolfe, A., The Acetate Switch [Review]. Microbiol Mol Biol Rev, 2005. 69(1): p. 12-50.
155. Kayser, A., J. Weber, V. Hecht, and U. Rinas, Metabolic flux analysis of Escherichia coli in
glucose-limited continuous culture. I. Growth-rate-dependent metabolic efficiency at steady state.
Microbiology, 2005. 151(3): p. 693.
156. Abdel-Hamid, A.M., M.M. Attwood, and J.R. Guest, Pyruvate oxidase contributes to the
aerobic growth efficiency of Escherichia coli. Microbiology, 2001. 147(6): p. 1483.
157. Phue, J.N. and J. Shiloach, Impact of dissolved oxygen concentration on acetate accumulation and
physiology of E. coli BL21, evaluating transcription levels of key genes at different dissolved oxygen
conditions. Metab Eng, 2005. 7(5-6): p. 353-63.
158. Kumari, S., C.M. Beatty, D.F. Browning, S.J.W. Busby, E.J. Simel, G. Hovel-Miner, and
A.J. Wolfe, Regulation of acetyl coenzyme A synthetase in Escherichia coli. Journal of
bacteriology, 2000. 182(15): p. 4173.
159. Nahku, R., K. Valgepea, P.J. Lahtvee, S. Erm, K. Abner, K. Adamberg, and R. Vilu,
Specific growth rate dependent transcriptome profiling of Escherichia coli K12 MG1655 in accelerostat
cultures. Journal of biotechnology, 2010. 145(1): p. 60-65.
160. Akesson, M., E.N. Karlsson, P. Hagander, J.P. Axelsson, and A. Tocaj, On-Line Detection
of Acetate Formation in Escherichia coli Cultures Using Dissolved Oxygen Responses to Feed
Transients. Biotechnology and Bioengineering, 1999. 64(5): p. 590-598.
161. Zawada, J. and J. Swartz, Maintaining Rapid Growth in Moderate-Density Escherichia coli
Fermentations. Biotech Bioeng, 2005. 89(4): p. 407-415.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


137

162. Tosi, S., M. Rossi, E. Tamburini, G. Vaccari, A. Amaretti, and D. Matteuzzi, Assessment
of in line near infrared spectroscopy for continuous monitoring of fermentation processes.
Biotechnology progress, 2003. 19(6): p. 1816-1821.
163. Kovatchev, B.P., L.A. Gonder-Frederick, D.J. Cox, and W.L. Clarke, Evaluating the
accuracy of continuous glucose-monitoring sensors. Diabetes Care, 2004. 27(8): p. 1922.
164. Whiffin, V.S.C., M. J. Cord-Ruwisch, R., Online detection of feed demand in high cell density
cultures of Escherichia coli by measurement of changes in dissolved oxygen transients in complex media
Biotechnology and Bioengineering, 2004. 85( 4): p. 422-433
165. Turner, C.G., M. E. Thornhill, N. F., Closed-loop control of fed-batch cultures of recombinant
Escherichia coli using on-line HPLC. Biotechnology and Bioengineering 1994. 44(7): p. 819-
829
166. Chorvat Jr, D. and A. Chorvatova, Multi wavelength fluorescence lifetime spectroscopy: a new
approach to the study of endogenous fluorescence in living cells and tissues. Laser Physics Letters,
2009. 6(3): p. 175-193.
167. Marose, S., C. Lindemann, and T. Scheper, Two Dimensional Fluorescence Spectroscopy: A
New Tool for On Line Bioprocess Monitoring. Biotechnology progress, 1998. 14(1): p. 63-74.
168. Gonzalez-Vara, Y.R., Enhanced production of L-(+)-lactic acid in chemostat by Lactobacillus casei
DSM 20011 using ion-exchange resins and cross-flow filtration in a fully automated pilot plant
controlled via NIR. Biotechnology and bioengineering, 2000. 67(2): p. 147-156.
169. Scarff, M., S.A. Arnold, L.M. Harvey, and B. McNeil, Near infrared spectroscopy for bioprocess
monitoring and control: current status and future trends. Critical reviews in biotechnology, 2006.
26(1): p. 17-39.
170. Hewitt, C.J. and G. Nebe-Von-Caron, The application of multi-parameter flow cytometry to
monitor individual microbial cell physiological state. Physiological Stress Responses in
Bioprocesses, 2004: p. 197-223.
171. Shapiro, H.M., Microbial analysis at the single-cell level: tasks and techniques. Journal of
microbiological methods, 2000. 42(1): p. 3-16.
172. Junne, S., M.N. Cruz Bournazou, A. Angersbach, and P. Götz, Electrooptical monitoring of
cell polarizability and cell size in aerobic Escherichia coli batch cultivations. Journal of Industrial
Microbiology and Biotechnology, 2010: p. 1-8.
173. Want, A., O.R.T. Thomas, B. Kara, J. Liddell, and C.J. Hewitt, Studies related to antibody
fragment (Fab) production in Escherichia coli W3110 fed batch fermentation processes using
multiparameter flow cytometry. Cytometry Part A, 2009. 75(2): p. 148-154.
174. Merten, O.W., Recombinant protein production with prokaryotic and eukaryotic cells: a comparative
view on host physiology: selected articles from the meeting of the EFB Section on Microbial Physiology,
Semmering, Austria, 5th-8th October 2000. 2001: Springer Netherlands.
175. Chassagnole, C., N. Noisommit-Rizzi, J.W. Schmid, K. Mauch, and M. Reuss, Dynamic
modeling of the central carbon metabolism of Escherichia coli. Biotechnology and Bioengineering,
2002. 79(1): p. 53-73.
176. Wang, F.S., A modified collocation method for solving differential-algebraic equations. Applied
mathematics and computation, 2000. 116(3): p. 257-278.
177. Pesavento, C. and R. Hengge, Bacterial nucleotide-based second messengers. Curr Opin
Microbiol, 2009. 12(2): p. 170-6.
178. Andersson, L., S.J. Yang, P. Neubauer, and S.O. Enfors, Impact of plasmid presence and
induction on cellular responses in fed batch cultures of Escherichia coli. Journal of Biotechnology,
1996. 46(3): p. 255-263.
179. Magnusson, L.U., A. Farewell, and T. Nystrom, ppGpp: a global regulator in Escherichia coli.
Trends Microbiol, 2005. 13(5): p. 236-42.
180. Lin, H.Y. and P. Neubauer, Influence of controlled glucose oscillations on a fed-batch process of
recombinant Escherichia coli. J Biotechnol, 2000. 79(1): p. 27-37.
181. Hardiman, T., K. Lemuth, M.A. Keller, M. Reuss, and M. Siemann-Herzberg, Topology of
the global regulatory network of carbon limitation in Escherichia coli. Journal of biotechnology,
2007. 132(4): p. 359-374.
Bibliography

138

182. Usuda, Y., Y. Nishio, S. Iwatani, S.J. Van Dien, A. Imaizumi, K. Shimbo, N. Kageyama,
D. Iwahata, H. Miyano, and K. Matsui, Dynamic modeling of Escherichia coli metabolic and
regulatory systems for amino-acid production. J Biotechnol, 2010. 147(1): p. 17-30.
183. Costa, R.S., D. Machado, I. Rocha, and E.C. Ferreira, Hybrid dynamic modeling of
Escherichia coli central metabolic network combining Michaelis-Menten and approximate kinetic
equations. Biosystems, 2010. 100(2): p. 150-7.
184. Srinivasan, B., D. Bonvin, E. Visser, and S. Palanki, Dynamic optimization of batch processes::
II. Role of measurements in handling uncertainty. Computers & chemical engineering, 2003.
27(1): p. 27-44.
185. Joshi, A. and B.O. Palsson, Escherichia-Coli Growth Dynamics - a 3-Pool Biochemically Based
Description. Biotechnology and Bioengineering, 1988. 31(2): p. 102-116.
186. Lahtvee, P.J., K. Adamberg, L. Arike, R. Nahku, K. Aller, and R. Vilu, Multi-omics
approach to study the growth efficiency and amino acid metabolism in Lactococcus lactis at various
specific growth rates. Microb Cell Fact, 2011. 10(12): p. 4158-4168.
187. Kumari, S., R. Tishel, M. Eisenbach, and A.J. Wolfe, Cloning, characterization, and functional
expression of acs, the gene which encodes acetyl coenzyme A synthetase in Escherichia coli. Journal of
bacteriology, 1995. 177(10): p. 2878.
188. Lin, H., N.M. Castro, G.N. Bennett, and K.Y. San, Acetyl-CoA synthetase overexpression in
Escherichia coli demonstrates more efficient acetate assimilation and lower acetate accumulation: a
potential tool in metabolic engineering. Applied microbiology and biotechnology, 2006. 71(6):
p. 870-874.
189. Cruz Bournazou, M.N., H. Arellano-Garcia, J.C. Schöneberger, S. Junne, P. Neubauer,
and G. Wozny, A Novel Approach to Mechanism Recognition in Escherichia Coli Fed-Batch
Fermentations. Computer Aided Chemical Engineering, 2009. 27: p. 651-656.
190. Veloso, A.C.A., I. Rocha, and E.C. Ferreira, Monitoring of fed-batch E. coli fermentations with
software sensors. Bioprocess and biosystems engineering, 2009. 32(3): p. 381-388.
191. Shimizu, H., R.A. Field, S.W. Homans, and A. Donohue-Rolfe, Solution structure of the
complex between the B-subunit homopentamer of verotoxin VT-1 from Escherichia coli and the
trisaccharide moiety of globotriaosylceramide. Biochemistry, 1998. 37(31): p. 11078-11082.
192. Jensen, E.B. and S. Carlsen, Production of recombinant human growth hormone in Escherichia coli:
expression of different precursors and physiological effects of glucose, acetate, and salts. Biotechnology
and bioengineering, 1990. 36(1): p. 1-11.
193. MacDonald, H.L. and J.O. Neway, Effects of medium quality on the expression of human
interleukin-2 at high cell density in fermentor cultures of Escherichia coli K-12. Applied and
environmental microbiology, 1990. 56(3): p. 640.
194. Holms, W.H., The central metabolic pathways of Escherichia coli: relationship between flux and
control at a branch point, efficiency of conversion to biomass, and excretion of acetate. Current topics
in cellular regulation, 1986. 28: p. 69-105.
195. Majewski, R.A. and M.M. Domach, Simple constrained-optimization view of acetate overflow in
E. coli. Biotechnology and bioengineering, 1990. 35(7): p. 732-738.
196. Bauer, K.A., A. Ben-Bassat, M. Dawson, V.T. De La Puente, and J.O. Neway, Improved
expression of human interleukin-2 in high-cell-density fermentor cultures of Escherichia coli K-12 by a
phosphotransacetylase mutant. Applied and environmental microbiology, 1990. 56(5): p.
1296.
197. Chou, C.H., G.N. Bennett, and K.Y. San, Effect of modified glucose uptake using genetic
engineering techniques on high-level recombinant protein production in escherichia coli dense cultures.
Biotechnology and bioengineering, 1994. 44(8): p. 952-960.
198. Chou, C.H., G.N. Bennett, and K.Y. San, Effect of modulated glucose uptake on high-level
recombinant protein production in a dense Escherichia coli culture. Biotechnology progress, 1994.
10(6): p. 644-647.
199. Jung, H., R. Räbenhagen, S. Tebbe, K. Leifker, N. Tholema, M. Quick, and R. Schmid,
Topology of the Na+/Proline Transporter ofEscherichia coli. Journal of Biological Chemistry,
1998. 273(41): p. 26400.
An approach to Mechanism Recognition
for model based analysis of Biological Systems M. Nicolas Cruz Bournazou


139

200. Konstantinov, K., M. Kishimoto, T. Seki, and T. Yoshida, A balanced DO-stat and its
application to the control of acetic acid excretion by recombinant Escherichia coli. Biotechnology and
bioengineering, 1990. 36(7): p. 750-758.
201. San Martin, M.C., N.P.J. Stamford, N. Dammerova, N.E. Dixon, and J.M. Carazo, A
structural model for the Escherichia coli DnaB helicase based on electron microscopy data. Journal of
structural biology, 1995. 114(3): p. 167-176.
202. Lara, A.R., L. Caspeta, G. Gosset, F. Bolivar, and O.T. Ramirez, Utility of an Escherichia
coli strain engineered in the substrate uptake system for improved culture performance at high glucose
and cell concentrations: An alternative to fed―batch cultures. Biotechnology and
bioengineering, 2008. 99(4): p. 893-901.
203. Lemoine, A., Design of Experiments and parameter Estimation of a kinetic model for E. coli
cultivations. Studienarbeit, 2011.
204. Angersbach, A. and V. Bunin, Continuous On-line Electrooptical Monitoring of Physiological
State of Bacterial Cultures in Biotechnological Processes.
205. DIN_EN_ISO_11213, modifizierte Stärke-Betimmung des Acetylgehaltes-Enzymatisches
Verfahren. 1995.
206. Nisselbaum, J.S. and S. Green, A simple ultramicro method for determination of pyridine
nucleotides in tissues 1. Analytical biochemistry, 1969. 27(2): p. 212-217.
207. Bernofsky, C. and M. Swan, An improved cycling assay for nicotinamide adenine dinucleotide.
Analytical biochemistry, 1973. 53(2): p. 452-458.
208. San, K.Y., G.N. Bennett, S.J. Berrios-Rivera, R.V. Vadali, Y.T. Yang, E. Horton, F.B.
Rudolph, B. Sariyar, and K. Blackwood, Metabolic engineering through cofactor manipulation
and its effects on metabolic flux redistribution in Escherichia coli. Metabolic engineering, 2002.
4(2): p. 182-192.
209. Parsopoulos, K.E. and M.N. Vrahatis, Recent approaches to global optimization problems
through particle swarm optimization. Natural computing, 2002. 1(2): p. 235-306.
210. Usuda, Y., Y. Nishio, S. Iwatani, S.J. Van Dien, A. Imaizumi, K. Shimbo, N. Kageyama,
D. Iwahata, H. Miyano, and K. Matsui, Dynamic modeling of Escherichia coli metabolic and
regulatory systems for amino-acid production. Journal of Biotechnology, 2010. 147(1): p. 17-30.
211. Singh, V.K. and I. Ghosh, Kinetic modeling of tricarboxylic acid cycle and glyoxylate bypass in
Mycobacterium tuberculosis, and its application to assessment of drug targets. Theor Biol Med
Model, 2006. 3: p. 27.



















dedicada a Heberto y Helig

.

and my family for always being at my side despite the distance and especially to Alexis Cruz. Finally. who have always motivated me to follow my goals and offered a shoulder to console my sorrows. Stefan Junne and Dr. I would like to thank my parents. for finding a perfect application for MR and for his intellectual input to this work. Harvey Arellano-Garcia.ACKNOWLEDGEMENTS I want to express my gratitude to my supervisor. I would also like to thank Professor Kravaris and Professor Lyberatos at the University of Patras for his collaboration and hospitality. I would also like to thank my second family. Dr. Nicolás Cruz B. for his constant support and useful advice on a professional and personal level and to my cosupervisor. Professor Günter Wozny. conformed of all my friends spread around the world. M. Special thanks go to Dr. who one day might realize his great contribution to each one of the achievements in my life. Tilman Barz for interesting discussions and support during the critical phases of this project (which were quite numerous). Mariano and Effi. I must of course thank all my other friends and colleagues in the Chair of Process Dynamics and Operations and in the Chair of Bioprocesses. Professor Peter Neubauer. .

.

1. Februar 2012 . dass die vorliegende Dissertation in allen Teilen von mir selbständig angefertigt wurde und die benutzten Hilfsmittel vollständig angegeben worden sind.Ich Mariano Nicolas Cruz Bournazou erkläre an Eides Statt. Mariano Nicolas Cruz Bournazou Berlin.

.

...................................4 2..................... xi List of Abbreviations ...................................................3 The gap between research and industry ..............4.2 2...........7 1...................................................... 16 Systems biology ........................................................................................2 2..........................................2 3.............. 5 Related work.............. 9 Advantages of Mechanism Recognition............1 1.................................. 21 i ............................................................................................................................................. 13 Model complexity ................ 22 Reaction invariants.. 3 Understanding process dynamics ..............................................................................1 2................................................. 19 Introduction .......................... 4 The bridge between industry and research ....................................... 11 Definition .............................................................................................................. 7 Project Goal ...............8 2 2...........................................................................................................................................................................................6 1............ 14 Engineering approach to complex systems .............................................................................. 25 Modeling .............................................................................................................................................................. xvii 1 Introduction ............................3 2...................................................... 1 Hierarchical modeling ...................................................................................................................2 3.......1 3...3 1....................................................................................................................... the bad. 16 Modeling of genetic regulatory systems. and the useful model ........................................5 3 3.................................................. ix Table content......... 21 Basic approaches to Model Reduction .....................2. 1 1..................................................................... 13 Model Reduction ............................. 10 The good............................................................................................................................................Content CONTENT Zusammenfassung ...........................4................5 1................. vii Figure content ....2.........1 3.............................................................................................................................................................................................................. 15 Modeling in systems biology .............4 1...................................................................................................................................2......... 24 Sensitivity analysis ....................2 1............ 17 Mathematical model for a batch biochemical reactor ........................................................................................................................................................ xii List of symbols ............................1 2..... 22 Switching functions and the reaction invariant ....................................................... v Abstract ...............................

............ 44 Code generation ...................Content 3..................................1 4...2..................................................................................................................... 49 sDACL ............................................. 28 The experiment ........................... 57 Submodels ................................................ 26 Perturbation theory ..................................3......3 6 6..................................1..........................................................1 4.............................................5 6..................................................................................................................3 4..........3.............................................................................3........2...1 4...........2 4................................1 6..1 5......................................................4 6...........................1............. 47 MOSAIC .........................................................................................................................2....................................... 33 The Maximum Likelihood .......1 5...............3.............. 37 The confidence Interval .............. 35 The Fisher Information Matrix........... 40 Model discrimination.......... 53 Methodology for Mechanism Recognition ................................................................................................5 3.............2 6......... 59 Detection of switching points ......... 42 Model discrimination in Mechanism Recognition......... 47 SBPD.... 37 Approximation of parameter variance-covariance matrix.......... 62 ii Optimal Experimental Design ....1 6..................................................2 4..... 58 Initial interval ............3 6............2 5......3............3........................................................... 39 Limitations of the Fisher Information Matrix ...............................................................................1 5... 62 Detection of the next switching point ........................ 57 General structure ...........................................................6 6......................................................8 Lumping................1.............................................................................. 51 ................................4 3.................................................. 60 Initial conditions of the interval k+1 ........... 34 Model identifiability .............................3....... 49 Optimization........................ 50 A short introduction to Mechanism Recognition ................................................................................2...........................................................2 6....................... 56 Program steps ....................................................................................................................3.... 31 Code generation............................... 51 Illustrative Example ..........................1...........................1 5 5.................................................... 57 Submodel distinguishability .......... 48 Simulation .2 4............................................................2.......1.....................................3 4.............................................................. simulation and optimization ..............7 6..................3............................................................................................................ 47 An approach to Mechanism Recognition ....2 5....................................................6 4 4..................................................................................... 27 Time scale analysis ..................... 59 MR initialization .............................................1 6...................2................................................2......3 6.........................

3......................................................................1 7........................................2 7..................................3...................... 63 Introduction .............................1 8.....................................3.......................................................1............3 7.............. 79 Conditions for proper process description with Mechanism Recognition 79 Conditions for accurate switching point detection ............... 83 Escherichia coli cultivations .................................................................................................................................2.........Content 6..........8 7..................2 7.. 75 Results ..............................................................................................................1 7.............. 68 Submodel building....................................... 71 Stoichiometric matrix ................... 82 Detection of switching points ..............1........................... 74 A proposed 6state model .. 82 Conclusions ..............................................................................3.............................................5 7..... 90 Submodels for dividing metabolic states .............................4 7......................3 7...............................3..1 8........................................ 73 Limitations of the reduced models ...........8..................5 7................2 7...............................................................3.............. 90 General model .....8................9 8 8...........1 7................................. 67 Monitoring of wastewater processes ..........................................2 7................................2 8........................................ 85 ........................ 75 Simulations Results .......................................3 7..... 70 Mathematical representation of the 9state model ................................................................................................. 65 Sequencing Batch Reactor ...................... 74 A proposed 5state model ..........................................8.6.........................................4 7...................................4 7....3..........3 8....................... 65 Mechanism Recognition in Escherichia coli cultivations.......................................................1 7......1........................... 80 MR initialization .........................................3 7...9 7 7.................................. 69 Storage ..................................1 7........................1................. 85 Models for the description of Escherichia coli cultivations ....................... 66 Nitrate Bypass Generation ................................................................................ 86 Division of physiological states ....6 7. 68 A proposed 9state model ........... 65 Activated Sludge................3.....1 8.........4 7................................................................................................2 8......... 77 Mechanism Recognition in SBR processes .......................................................8...................4 Flow diagram ...................................................7 7........... 92 Material and methods ...... 78 Recognition of organic matter depletion ........................................................... 87 Modeling Escherichia coli batch fermentations with Mechanism Recognition ........... 96 iii Mechanism Recognition applied on Sequencing Batch Reactors ...................................................................................................................... 69 Reduction of the extended ASM3 model to a 9state model ..................

....6 8........................ 96 Online analysis .................................................................................................................................4...4...................................5 8....4 8.4........5.........2 9..... 97 Offline analysis ......................................5................................... 118 Conclusions and outlook ......................................1 9...................................................................................... 104 Conditions for proper process description with MR ...................................................................2..............5...................5 8..1 8.................................................4.................. 117 Global optimization ........... 105 Data set ........................................................7 9 9...... 106 Recognition of overflow and substrate limitation regimes ..... 109 Simulations vs...........................3 8....................................2....... 115 Appendix .......................2.. 119 iv ......1 8...................5........................................ 114 Conclusions ..........2 8............2 8..........5.................3 9...................... 103 Experimental validation ....................... 116 Switching point identification .........3 8......................................................... 115 Outlook ...Content 8............. experimental data ......................... 99 Data treatment ............................................. 117 Online monitoring............................................................2............................. 113 Future work ..............................5.......................... 111 Conclusions .......................................................... 110 Results .........................................6 8..............................................................................................................................................................4 10 Strain and culture conditions ....................................................................................................................................................................................2 9........ 104 Conditions for accurate switching point detection ........................................................................................................................ 116 General theory for submodel generation ..........................4 8.....................1 9.............................

Analogien zu anderen Modellen aus der Literatur. liefert eine exakte Beschreibung des Prozesses in einem wohldefinierten Bereich und erlaubt die Erkennung des Abbaus organischer Stoffe. Empirische Kenntnisse. Batch-Kultivierungen mit einfachen ODE-Systemen zu beschreiben. Dazu wird das dem aktuellen Forschungsstand entsprechende Modell (ASM3 erweitert für die zweistufige Nitrifikation und Denitrifikation) auf ein einfaches Teilmodell reduziert. für die sowohl die Anwendbarkeit als auch die Vorteile dieser Methode nachgewiesen werden. Biologische Prozesse stellen ein wichtiges Anwendungsgebiet für die Mechanismenerkennung dar. Nicolas Cruz Bournazou ZUSAMMENFASSUNG Ziel dieser Arbeit ist die Entwicklung innovativer Ansätze zur Beschreibung komplexer Prozesse mit Hilfe von reduzierten Modellen. Es wird gezeigt. v . Das resultierende Modell ist effizient anzuwenden. Abschließend wird die Fähigkeit der Mechanismen Erkennung als Unterstützungswerkzeug für die Zusammenarbeit zwischen Grundlagenforschung und Industrie analysiert. Methoden zur Bewertung des Zustand eines Systems und Ansätze zur Modellreduktion werden kombiniert. Im Rahmen dieser Arbeit werden zwei Fallstudien vorgestellt. die in der Lage sind. Die Methodik der Mechanismenerkennung ermöglicht die Erzeugung von drei Teilmodellen. um Indikatoren für relevante Änderungen im Prozessgeschehen zu erzeugen. Ein erfolgreich validiertes Modell wird analysiert und reduziert. Die erste Fallstudie beschreibt die Überwachung des Belebtschlammverfahrens in Sequencing Batch Reaktoren.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Der Ansatz zur Mechanismenerkennung ist ein Werkzeug zur effizienten Nutzung von Kenntnissen aus der Grundlagenforschung und der Modellierung und ermöglicht ein tieferes Verständnis für den gesamten Prozess. Die resultierenden Beschränkungen für die Vorhersage des Prozessverhaltens auf Basis von reduzierten Modellen werden durch den Einsatz von Methoden zur Mechanismenerkennung genutzt. die Beschreibung und Überwachung des Prozesses mit einer höheren Effizienz erlaubt. Die zweite Fallstudie ist die Kultivierung von Escherichia coli im Batch-Prozess. in einem Versuch ein Set exakter Teilmodelle mit einer großen Robustheit und Identifizierbarkeit zu generieren. dass die systematische Analyse des Prozesses und seiner gemessenen sowie auf Basis von Modellen vorausberechneten Zustände.

An approach to Mechanism Recognition for model based analysis of Biological Systems

M. Nicolas Cruz Bournazou

ABSTRACT
This work aims at finding new manners to accurately describe complex processes based on simple models. Furthermore, the approach to Mechanism Recognition proposes to exploit the description limitations of these submodels and to use them as indicators of non-measurable variables. Empirical knowledge, analogies to other models from literature, methods to analyze the state of information of the system and model reduction techniques are brought together in an effort to create an adequate set of accurate models with a significantly larger tractability. It is worth stressing the approach to Mechanism Recognition does not intend to substitute human reasoning or make up for lack of process knowledge. On the contrary, this method is merely a tool to efficiently apply the knowledge obtained from basic research to gain a better insight of the industrial process. The approach to Mechanism Recognition finds an important field of application in biological processes. In this work two case studies are presented to manifest the advantages and applicability of this method. It is shown how the correct analysis of the process, the state of information, and the models applied to describe the process results in new methods to describe and monitor the process with higher efficiency. The first case study presented is the monitoring of the Active Sludge Process in Sequencing Batch Reactors. For this, the state of the art model ASM3 extended for two step nitrification-denitrification is reduced to create a simple model which can easily describe the process in a defined range and detect depletion of carbonate matter. The second case study is Escherichia coli batch and fed-batch cultivations. A model obtained from literature is analyzed and reduced. The methodology of Mechanism Recognition allows creating a set of three submodels able to describe batch cultivations with simple systems of Ordinary Differential Equations. Furthermore, the restrictions of the complex model are set under scrutiny to understand its dynamics and limitations. Finally, special attention is paid to the capability of Mechanism Recognition as a tool to enhance collaboration between basic research and industry.

vii

.............................................. ............................................................................................2: Three-component monomolecular reaction system.................. The toolbox is designed in a modular ......... 76 Figure 7...................................................4: Shape of the confidence interval for different variance values from the Lin model (appendix A).............. 25 Figure 3...........3................................... .............................4: Phase diagram of full order model (3...... 77 Figure 7.......................... with and without substrate................. 14 Figure 2........................................................................... Behavior of a switching function in dependence of the limiting species.......... ........................................5.. ...... 41 Figure 5.......... 39 Figure 4.. 78 ix ....................................... Nicolas Cruz Bournazou FIGURE CONTENT Figure 1.....1: High level modeling with MOSAIC [46] ...............4: Flow diagram of MR algorithm .......1: Hierarchical modeling scheme............ obtained with Montecarlo simulation.............2...... .....6..1......................... 66 Figure 7.and forward reaction constants.. Nitrification-denitrification process described as a two -step reaction....................3: Lumping a monomolecular three-component reaction into a two-component reaction ..... NOX concentration against time................................................2: Incremental approach for reaction kinetics identification [58] ......................... [119]........5: Objective function of a nonlinear model (appendix A) with respect to changes in a two dimensional parameter set.. Changes in the biomass are very small (less than 10%)............................ 41 Figure 4.................................................................... a) Oxygen concentration in the medium against time...................................................... [53]....................................... 36 Figure 4................................... 77 Figure 7........................................1: Model fit a) without setting bounds b) with setting bounds for physical parameters. the numbers on the arrows represent the back...... B) various models [119] ............................ Biomass against time...3: Criteria for optimization [92] ........................ .......... 67 Figure 7..................3: Hypothesis-driven research in systems biology [59]... 54 Figure 6........ Comparison with reduced models in a chemostat process ................................................................... 26 Figure 3.... 27 Figure 3. 3 Figure 2.................................. 16 Figure 2....................3 Cleaning strategy based on MR [43]............................. 64 Figure 7.................................................... coli transcriptional regulatory network.................. .12).........................................................1 : E........ 55 Figure 6............. 55 Figure 6........................................2: Modular structure of the toolbox..............................2: Comparison experiment/simulation using a) just one model..................................1: SBR cycle [136]......... ......................................1: Effect of sensitivities in parameter estimation accuracy..............................................................................An approach to Mechanism Recognition for model based analysis of Biological Systems M.......... 17 Figure 3.......................................................... 30 Figure 4......................... .. 76 Figure 7................................2: Confidence interval from the Lin model..... 48 Figure 5..................................................4...................................... Substrate concentration SS and stored energy Sto against time............ 49 Figure 6...................................... σP and σy represent standard deviation of parameters and measurements respectively............... The confidence interval can be approximated by an ellipse near the exact value..................... .....................7: Description of the 5state model in both regimes......................... 40 Figure 4.......

................ Part I: Dry biomass and glucose concentrations ............. 101 Figure 8.... 83 Figure 8......... ................ ............. 109 Figure 8............................18: Experimental validation of the MR approach........................................... 91 Figure 8........... Part IV: Metabolite concentration ............9................ .. 82 Figure 7.....................................................................................17: Starvation condition described by the corresponding submodel fitted against experimental data...................... 91 Figure 8....................................................13: Experimental results batch experiment G1............................. Part II: Specific concentration of acetic acid ............6: Bioreactor KL2000 at E..........................16: Submodel for the description of growth under substrate limitation fitted against experimental data...................... the overflow submodel (lines) initializing in four different intervals....... .......14: Experimental results batch experiment G1...................................... 110 Figure 8.................. 112 Figure 8........... 102 Figure 8........12: Experimental results batch experiment G1.............................................................3: Comparison between the complex model (dots) vs. coli batch cultivation [203] ....................4: Comparison between the complex model (dots) vs............. 108 Figure 8.. .10: Mechanism of the reactions involved in the assay .. ...........................5: Comparison between the complex model (dots) vs................................................................... Part III: Outgas concentrations .........................................................................................................................19: Identifiability test considering white noise..............................9............................2: Complex model (Lin et al...... ..................................... Calibration curve for glucose determination ......... 95 Figure 8........................1: Integration of the kinetic model proposed by Lin [91] .. 107 Figure 8..8: Minimal length for initialization of MR ........................................................Figure content Figure 7.... 93 Figure 8.... 108 Figure 8............................. the substrate limiting submodel (lines) initializing in four different intervals..................................... 97 Figure 8.......................................................................................................... 111 Figure 8....... the cell starvation submodel (lines) initializing in four different intervals....... standard deviation of 5% in all measurements .... 99 Figure 8................................................ 111 Figure 8....................................................11: Experimental results batch experiment G1..................................................................8.............. 100 Figure 8.15: OverFlow submodel fitted against experimental data..........) fitted to experimental batch cultivation data........7: EloCheck................. .......... 94 Figure 8............ Calibration curve of acetate ....... 113 x ................... Detection of the regime switching point..................................................................................................... .................................................

.. Composition of solution A .An approach to Mechanism Recognition for model based analysis of Biological Systems M.. 40 Table 4................................................................ 43 Table 7............. Comparison of the computation time..... ....................... .................... 73 Table 7...................1: Parameters considered for the model fit .....................4.....2: 9state model constants and its values as shown in the Matlab code .. 78 Table 8.............................................................................................1: Criteria for confidence interval quantification [92].................3: Stoichiometric matrix of the 9state model .................................................. 102 xi ............. 70 Table 7.................5....................... 73 Table 7...................1: Reaction rates of the extended ASM3 ... 95 Table 8... 77 Table 7.................................. Singular function evaluations speed .................................................2...................................... Nicolas Cruz Bournazou TABLE CONTENT Table 4.....2: Types of sum of square [22] ................

List of Abbreviations LIST OF ABBREVIATIONS Acs ADHII AMP AOB ASM ASP BOD Bpox CAB CAPE CFD COD CRB DAE DFG DNA DOT EDTA EMA FDA FIM GRN Acetyl-CoA synthase Alcohol Dehydrogenase Adenosine monophosphate Ammonium Oxidizing Bacteria Activated Sludge Model Active Sludge Process Biological Oxygen Demand Pyruvate oxidase Computer Aided Biology Computer Aided Process Engineering Computational Fluid Dynamics Chemical Oxygen Demand Cramer-Rao Bound Differential Algebraic Equation German Research Foundation Deoxyribonucleic acid Dissolved Oxygen Tension Ethylenediaminetetraacetic acid European Medicines Agency Food and Drug Administration Fisher Information Matrix Gene Regulatory Network xii .

Nicolas Cruz Bournazou HET HPLC IA IMM KDD LSQ MBR MBDoE MD MR mRNA MTT MWF MXL NAD+ NB NBND NDF NS NH+4 NIRS NO-2 NO-3 NOB NSF Heteroptrophic organisms High-Performance Liquid Chromatography Incremental Approach Interactive Multiple Model Knowledge Discovery of Data Least Squares Membrane Bioreactor Model Based Design of Experiments Model Discrimination Mechanism Recognition Messenger Ribonucleic Acid Thiazolyl Blue Multi-Wavelength Fluorescence Maximum Likelihood Nicotinamide adenine dinucleotide (NadH) Nitrobacter Nitrate Bypass Nitrification-Denitrification Numerical Differentiation Formula Nitrosomona Ammonia Near-Infrared Spectroscopy Nitrite Nitrate Nitrite Oxidizing Bacteria Numerical Differentiation Formula xiii .An approach to Mechanism Recognition for model based analysis of Biological Systems M.

List of Abbreviations OC OCFE ODE OED OF PAT PCA PCP PDE PES PLS ppG ppGpp PSO PSSH PTS QSSA RWP SBML SBR SF SL SQP ST TCA Orthogonal Collocation Orthogonal Collocation on Finite Elements Ordinary Differential Equations Optimal Experimental Design OverFlow Metabolism Model Process Analytical Technology Principal Component Analysis Process Constant Parameter Partial Differential Equation Phenazine Ethosulfate Partial Least Squares Phosphoenol Pyruvate Glyoxylate Guanosine tetraphosphate Particle Swarm Optimization Pseudo Steady State Hypothesis phosphotranspherase Quasi Steady State Assumption Regime-Wise constant Parameter Systems Biology Markup Language Sequencing Batch Reactors Switching Function Substrate Limitation Model Sequential Quadratic Problem Starvation model Tricyclic Acid Cycle xiv .

An approach to Mechanism Recognition for model based analysis of Biological Systems M. Nicolas Cruz Bournazou WWTP Waste Water Treatment Plants xv .

.

. 3 𝐿 𝑔 𝑚 [] 𝑀𝑜𝑙 𝑠 [] [] [𝑏𝑎𝑟 ] [] [] 𝐿 𝑚𝐿 .An approach to Mechanism Recognition for model based analysis of Biological Systems M. 𝑠 𝑕 [] [] 𝑚 𝑠 2 xvii . Nicolas Cruz Bournazou LIST OF SYMBOLS VARIABLES 𝐴 𝐴𝑐 𝐴𝑐𝑟𝑖𝑡 𝑎 𝛼 𝐶 C 𝑐 𝐶𝑇𝑅 D 𝐷𝐵 ∆𝑝 E 𝜀 F f Φ 𝑔 area acetate A-criterion linearly independent row vector specific cake resistance covariance matrix concentration substrate consumption coefficient carbon dioxide transfer rate dilution rate Identifiability threshold pressure difference enzyme Distinguishability threshold feed rate Function objective function gravity acceleration [𝑚2 ] 𝑚𝑀𝑜𝑙 𝐿 [] [] 1 𝑚2 [] 𝑚𝑀𝑜𝑙 𝑀𝑜𝑙𝐶 𝑔 .

List of symbols Γ H 𝜂 𝜂𝑠 K k 𝜅 𝐿 𝑚 µ 𝜈 𝑂𝑇𝑅 P ℘ Q q R r 𝑟𝑒𝑠 𝑆 initial velocity of the projectile hypothesis stochastic error systematic error limiting constant monomolecular rate matrix friction constant membrane thickness mass growth rate dynamic viscosity oxygen transfer rate product probability distribution function uptake specific uptake resistance reaction rate residual concentrations of soluble species (substrates and products)/Substrate m s [] [% ] [% ] 𝑚𝑔 𝐿 [] 𝑚 ∗ 𝑘𝑔 𝑠 [𝑚𝑚] [𝑘𝑔] 𝑕−1 . [𝑑 −1 ] 𝑁 ∗ 𝑠 𝑚2 𝑀𝑜𝑙 𝑠 [] [] 𝑔 𝐿 𝑔 𝑔 ∗ 𝐿 [𝑚−1 ] [] [] [] xviii .

[𝑑] [] [] [𝐿] [] [] [] [] [] 𝑔 𝑔 [] [] [] xix . [𝑑] 𝑠 . 𝑕 . 𝑚𝑖𝑛 . 𝑕 . Nicolas Cruz Bournazou 𝑆𝑡 𝑠 𝜎 𝑡 𝑡𝑠𝑝 𝜃 𝑢 𝑉 𝑊 𝑤 𝑊 𝑚 𝑋 𝑥 𝑌 Y 𝑦 z correction constant blocked area per unit filtrate volume standard deviation time time span parameter vector input variables vector Volume weighting matrix constant input variables vector Culture medium weight concentrations of the particulate compounds state variables vector yield coefficient stoichiometric coefficient measurement values vector reaction invariant [] 𝑚2 𝑚3 [] 𝑠 .An approach to Mechanism Recognition for model based analysis of Biological Systems M. 𝑚𝑖𝑛 .

List of symbols SUBSCRIPTS AND SUPERSCRIPTS 0 𝛼 aer anox Bio C 𝑐𝑎𝑙𝑐 𝑐𝑎𝑝 E 𝑒𝑠 G Gluc H L M 𝑚𝑎𝑥 𝑚𝑒𝑠 𝑛𝑜𝑚 𝑂 S initial value incoming aeration phase anoxic phase biomass cake calculated value capacity experimental estimated general structure glucose heterotrophous lower membrane Maximal Measured value nominal Oxygen Substrate xx .

Industry is forced to make such investments to strive for its success in the world markets. The development of a complex model. Today. It can be said that industry is in search of smart solutions while academy is looking for interesting problems. accurate models allow optimal design and operation of plants. collaboration projects between academia and industry confront many complications. Gproms®. many of the models and software tools developed in universities and research institutes are used in industry (Aspen®. Basic research offers a strong platform for development of industrial applications. In addition. Cutting edge technology is now essential for chemical and biochemical companies to survive. not only for direct applications. including estimation and validation. Still. may take several years. These models have to be developed in basic research. 1 . and specific research fields. including support to universities. Matlab®). BASF invested almost 1. long term projects. The German Research Foundation (DFG) invested in the same year 2010 approximately 2. In the year 2010.5 billion Euros in research and development [1]. promoting mostly basic research. Governments also need to make important investments on research. setting new standards in product performance. Process modeling in chemistry and biotechnology offers a handful of examples of the advantages of joint work.1 THE GAP BETWEEN RESEARCH AND INDUSTRY Globalization has changed market conditions drastically.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Advances in transport and communication bring companies together in worldwide competition. and environmental impact. While industry demands mostly fast solution to real process problems. To achieve this. reducing energy consumption. hazard. while allowing better monitoring and control [3]. academia is more interested in long term projects offering novel knowledge. but also as long term investments to earn basic knowledge. and industry provides not only economic support but also new challenges and interesting applications. and application range cannot be assured beforehand. In spite of the parallel effort of both parties aiming at a common goal. model identifiability or observability. A company cannot afford to make such long term and uncertain investments. Nicolas Cruz Bournazou 1 INTRODUCTION 1. substantial efforts have to be invested in research and development. which is not attractive to industry because it represents a long term investment.3 billion Euros [2]. Finding novel methods to bring industry and the research community together is essential for their efficient development.

This knowledge and understanding is the basis for establishing an approach to control of the manufacturing process that results in products with the desired quality attributes. the required facilities for parameter estimation and model validation. Current black-box models and heuristic rules cannot provide the information required in modern engineering. January 2011 [4]: “A successful validation program depends upon information and knowledge from product and process development. Due to the difficult measurements required. Instruments to benefit from the advances achieved in basic research by allowing an adequate information transfer between both parties are crucial for an efficient development of modern process technology. Process complexity and safety restrictions have driven design and control to demand accurate and robust models. The new regulations of the Process Analytical Technology (PAT) initiative of the Food and Drug Administration (FDA) and the European Medicines Agency (EMA) show the importance that modeling applied to process monitoring and control is gaining in the pharmaceutical and generally in the biotechnological industry. information about large scale processes and long term performance can only be obtained from real plants. New methods need to be created to bring complex models closer to industry and also to create ways to use the information earned in industry for basic research purposes. As maximization of process efficiency becomes essential to remaining competitive in the market. demanding model-based knowledge of the process. in addition to economical support. Manufacturers should: Understand the sources of variation Detect the presence and degree of variation Understand the impact of variation on the process and ultimately on product attributes Control the variation in a manner commensurate with the risk it represents to the process and product “ 2 . are gaining application in industry. Development of new tools that facilitate the communication and interaction between industry and basic research lead to more efficient collaboration and better individual performance. the application of model based control and monitoring is essential. Modeling is not an exception. The FDA makes the following statement in its Guidance for Industry. The data collected daily in chemical plants provides valuable information to researchers. Additionally.Introduction In return. complex models. industry offers. which enable profit increase while fulfilling environmental and safety regulations. Regulations are changing.

Figure 1. Particularly in biological systems. Unifac Uniquac Newtonian fluids • Bingham fluid • Mass balance • • • • Energie balance Adiabat Isotherm State eq.2 HIERARCHICAL MODELING The contradiction between models in research and industry can also be seen from the point of view of hierarchical modeling. In addition.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Depending on the level of system understanding. Figure 1. cells can be described with a simple Boolean equation. etc. These three different layers have a diverse level of significance for industry and research. expensive and time demanding 3 .1 depicts the typical layer representation of a chemical process. This second category of models is commonly too complex and requires advanced. Biological systems are extremely complex and very difficult to predict. destilation colum. whereas industry is more interested in plant wide behavior aiming at robust and secure process operation. from a kinetic down to a genomic level. simple. and robust models. P-3 E-9 P-11 Layer 3 Process Systems Engineering (PSE) Design of the complete process P-2 V-2 V-5 P-10 P-16 Experimental Expreimental set up in setup in pilot plant scala P-9 P-5 V-3 P-8 P-12 plant scala E-7 E-13 P-6 E-10 P-15 P-13 V-6 P-14 Layer 2 Unit design (design of the reactor. industry is only interested in practical. basic research is more interested in the lower layer where the study of microscalar phenomena takes place. For this reason. regulations in food and pharmaceutical industry are extremely strict. a gap can be seen between industry and basic research [5].) Expreimental set up in setup scala miniplant in miniplant scala Experimental P-16 E-14 E-13 P-13 V-6 P-14 Model 1 Layer 1 Experimental Expreimental set up in set up scala laboratory in Model 1 Model 2 laboratory scala Model 3 Model 2 • • • • • Mass transport Phase eq. the main goal for building a model in research is to gain process information [6]. Nicolas Cruz Bournazou 1.1: Hierarchical modeling scheme. On the other hand.

Finally. but some are useful”. a general approach for model reduction considering online and offline measure possibilities. experimental conditions. experimental effort for parameter estimation increases exponentially as the model grows in complexity.3 UNDERSTANDING PROCESS DYNAMICS Mathematical models can be described as the result of an effort to represent behavior of nature with mathematical equations. Online measurement limitations may also hinder the application of complex models for model-based control. Finally. and prediction of their behavior. A defined methodology to strategically simplify complex models. considering both the requirements of a particular industrial process and the quality of the data available. G. monitoring. In process engineering. control. Still. understanding of complex systems. Not only the measurement techniques become more complicated and expensive but also the identifiability of the parameters is reduced with each new parameter added to the model [11]. a large number of parameters increases the size of the optimization problem and number of local minima [12]. and deep understanding of the system is not to be found in literature. 1. models are used for process design. industry needs to take advantage of knowledge gained in basic research in general and of application of complex models in particular. all the complications mentioned above restrict application to highly trained personnel. In addition. with an increasing number of parameters to optimize.Introduction measurement techniques. Although many reduction methods are applied for control process purposes [7]. Speaking of parameter estimation. Advanced measurement techniques and fast computer processors enable the creation of very complex models processing enormous amounts of information [10]. In most cases. Box [8]:“all models are bad. the approximate description achieved by models has shown to be very useful. initial value consistency gains importance for simulation convergence. complex models require costly hardware to make such complicated calculations as well as expensive software to simulate and optimize the model efficiently. Furthermore. Nevertheless. 4 . and optimization. application of such complex models requires highly trained personnel. sophisticated models contain an important number of parameters and thus require large amounts of very specific data in order to be identifiable. Modeling and simulation have developed rapidly over the last years [9]. In the words of P. Despite the inability of mathematics to precisely describe physical phenomena. is missing in biotechnology. Models are applied in all fields of science and have become an essential tool for data acquisition and processing.

Nicolas Cruz Bournazou Batch processes commonly show highly nonlinear behavior and require more advanced models for their description. complex models present a number of disadvantages which hinder their implementation in industrial processes. 5 .An approach to Mechanism Recognition for model based analysis of Biological Systems M. Nonetheless. These dynamic and highly nonlinear processes require accurate first principle models to be properly described. this cannot be foreseen and the model can only be adjusted once experimental data is available. It is well known that mechanistic models offer a number of advantages over “black-box” modeling. Nonetheless. Hence. models should change based on how and when these phenomena change. which dictate the process dynamics. These phenomena. Experience and over-sizing combined with improvements during operation led to fairly successful results. observable. rigorous models developed in basic research are rarely applied in industry. Still. Unfortunately. there are many processes that operate without detailed model-based knowledge of its dynamics. and simple but also accurate and reliable for its use in industrial applications. in rigorous modeling the choice of the mechanistic model to be used for the simulation is based on the dominant physical phenomenon of the process. environmental impact and improve design and operation through simulation and optimization. In order to build models that have all the aforementioned features and are also based on rigorous knowledge of the system. If correctly implemented. mechanistic models help to predict risks. change over time. In the past. In other words. These demands include. a higher process comprehension and a more accurate scaleup capability [16-21]. predictions were carried out mainly on the basis of empirical knowledge. This is the principal reason why most dynamic processes can be simulated effectively for short time periods but not for the complete process. large calculation costs and the need for highly trained staff are only some of the problems to be faced in order to apply complex first principle models to industrial processes.g. complex measurement techniques. robust. Simplifying the model to adapt it to the strictly important conditions may reduce the complexity drastically. in many cases only certain conditions of the process are of interest. Also rigorous models provide the basis needed for efficient quality control. On the other hand. Low identifiability. for example. However.4 THE BRIDGE BETWEEN INDUSTRY AND RESEARCH When speaking of industrial systems. a close cooperation between the research community and industry is essential. Unfortunately. Complications related to batch process simulation and control are well known [13-15]. the appropriate approach is to simulate the process with various models also changing over time. 1. Models have to be tractable. simple nonlinear regressions based on direct measurements are not suitable for these goals. e. in recent years an increasing trend to bring existing plants to meet new market demands can be established. improved quality or compliance with new standards for environmental restrictions.

measuring methods and further unknown factors. there exists no parameter set. The difference between the outputs predicted by the model and the outputs measured from the system is called residual [23]. This is partly overcome by adding new equations and parameters to patch errors in the structure of the model.Introduction This work represents an important step towards the development of a systematic approach to the adaptation of complex models for their application in industrial processes. This trend is slowly changing with the development of efficient global optimization techniques [25]. model structure (systematic. Nonetheless. - Modelers usually tend to build models with too many parameters and to settle with locally optimal parameter values. One of the most difficult decisions to make for a modeler is the level of description accuracy required for a model to be useful [22]. methods to detect the source of systematic error are required and approaches to detect the instance of the structure causing the error require further development. the selection of the structure of a model still requires individual analysis of each case and vast experience in modeling added to deep knowledge of the system to be modeled. meaning it fails to consider all important factors of the process and to represent the correct dynamics. To name one example. Model reduction is a promising approach to close the gap between models developed in basic research and models required in industry. Considering the exact parameters are known. Global dynamic optimizers offer the possibility to find the definite parameter set which best describes the observations [26]. which can make the model fit the data. parameter fitting and simulation efforts. the causes for residual different from zero can be grouped in two main categories [24]: uncertainties (stochastic error) Disturbances and unknowns are intrinsic errors of the system and cannot be predicted. error) When the structure of the model is incorrect. Deciding how accurate a model needs to be to accept it as an adequate description of the process is still an open question in engineering. but such a model will never be 6 . 28]. Nonetheless. These show a normalized distribution with norm equal to zero and a variance dependent on the conditions of the system. these “patches” are usually responsible for unneeded parameter correlation and reduction of model identifiability. The most significant contribution of global optimization to model structure analysis is that one can rigorously demonstrate that the model is inconsistent with experimental data regardless of its parameter values. an agreement must be met between model accuracy and modeling. a straight line can be described exactly by a fifth order polynomial. Despite many efforts to develop automatic modeling programs [27. As we will see later in detail.

Finding communication paths between the different disciplines to take advantage of the information gained in each case and achieve the best possible model for each system is essential. fault detection techniques have rapidly evolved [23] and are being applied in many fields of industry. Furthermore. as will be shown in this manuscript. Interactive Multiple Model (IMM) [31. Furthermore. 39]. 30]. the application of large nonlinear models is gaining popularity. There exists a handful of methods aiming at fast and robust description of processes. the use of a combination of more than one model in an effort to describe specific instances of a system or complete processes has been proposed in various forms. these methods relay on simple models with no physical foundation with fast. but short term prediction being its ultimate objective. Creating new tools to analyze the structure of models and find correct representations of the system is the main goal of this work. 32]. the use of models to obtain precise process information based on indirect measurements has been utilized since the beginnings of the discipline. These methods use software redundancy with mechanistic models in an effort to detect fault behavior in complex systems. PUMon (a tool for online monitoring based on neural networks) is being developed at Bayer [40]. semiquantative simulation [35]. Various fields in science require fast calculations to achieve optimal control of systems with high dynamics. many disciplines need to be brought together in an effort to attack model defects from different angles to detect failures and to propose solutions. It is unidentifiable because an infinite combination of parameter sets exists. From missile tracking to burnout reactions. Also online applications like model based fault isolation and identification consider the application of rigorous models to detect malfunctions in the system [38.g. e. Nicolas Cruz Bournazou identifiable because there is an infinite combination of values of the polynomial. As limitation by computation burden losses significance due to the increasing capacity of modern microchip architecture and cloud computing systems. Approaches to reject hypothetical reaction pathways in chemistry using first principle models in combination with global optimization have been published [37]. which fit the system. are just some examples. a combination of simple models may offer important advantages. Qualitative process theory [29. which can describe a straight line. variable structure theory [36]. many approaches have been successfully applied mostly using statistical methods and repeated linear approximations of the system.5 RELATED WORK Especially in process engineering. Nonetheless. jump Markov linear systems [33]. despite the long story of 7 . qualitative algebra and graph theory methods [34]. Furthermore. 1. To achieve this goal. However.An approach to Mechanism Recognition for model based analysis of Biological Systems M.

The core of MR is model building. and its interaction with all other submodels has been understood. On the contrary. the first application is limited to models with one state variable. Nevertheless. MR is concerned with the characteristics of the submodel and its relation to the physical system.1). complex models require a general structure to increase its identifiability this cannot be generated for models with different characteristics.1. systematic. different models were obtained from literature each one describing a different regime of the process (section 6. Mechanism Recognition (MR) differs from all previous approaches in that the physical properties of the system are considered. leaving system understanding aside. is not to be found in literature [42]. different manners to create and analyze models (chapter 3) and its relation with the observations 8 . it is not possible to assure calculation of all state variables in every regime. This work provides significant evidence that despite technological advances. modeling is still a field which requires intensive human intervention. and automated modeling approach. The engineer must make use of his knowhow and intuition to be able to develop efficient models which mirror reality and are consistent with scientific evidence. Hence. Most methods for system description with more than one model aim strictly at computation expenses reduction. Once physical meaning of each submodel has been experimentally validated. when obtaining models from literature. 45] together with efficient integrators [46] facilitate the exercise of modeling significantly. a number of software packages for automatic model building [47] and automatic model reduction [45] confirm the trend to a general. Still. Because of the simplicity of the models applied. Nevertheless. Furthermore. its application in complex dynamic systems is still limited. the number and types of state variables contained by each model may differ. Still. modeling is still a challenging and exciting discipline [48]. The challenges of modeling and experimental validation will be discussed. The results obtained suggest that the method can be also applied for systems with a higher number of state variables. a systematic methodology for the identification of non measurable process variables. Furthermore.Introduction similar methods to gain knowledge from limited data sets [41]. Furthermore. MR aims at discerning and selecting the phenomena dictating the dynamics of the system. the application of MR is straight forward and some of the aforementioned techniques can be applied. the practice of modeling should not be underestimated. using a comparison between different first principle models describing selected regimes in dynamic processes. a continuous computation of all state variables cannot be assured (input-output consistency). Novel software toolboxes for model building modularization and reusability [44. MR has been successfully applied for small systems [43]. no general structure was required and input-output consistency was inherently fulfilled by the single input single output condition of all models. submodel building. In this example. Furthermore. most precisely. Since the models are obtained from different sources.

Equations and structure of the model must describe only the most important dynamics. This very special characteristic is exploited by MR. modelization) error. complex models created in basic research can be adapted for application in industrial processes.) to fulfill specific requirements of particular industrial problem. three essential conditions must be fulfilled: deep comprehension of the dynamics of the system . By these means.The data set must deliver enough information to estimate the parameter set with high accuracy.6 PROJECT GOAL The main goal of this project is to find new approaches for a target-oriented model simplification. It is essential to understand that identifiability depends not only on the data set (state information). 1.g. etc. high model identifiability . This is also the main topic throughout this manuscript. Various methods for model reduction are to be studied in combination with mathematical tools for experimental information quantification (confidence intervals. The biggest challenge for the application of MR is how to create a simple but accurate model specifically adapted to the particular conditions of each regime. 9 . but also on the structure of the model. This means that it is able to describe the strictly defined regime of the system with high accuracy. which satisfies the above mentioned conditions. In order for a simple model to mirror a complex system. If it is precisely known which regime can be described by the model.An approach to Mechanism Recognition for model based analysis of Biological Systems M. this works aims at finding new means to accurately describe complex processes based on simple models. Nicolas Cruz Bournazou of the system (chapter 4) will be presented in an effort to increase the efficiency of model development. the phenomenon governing systems behavior must be deeply understood. optimality criteria. allowing a deeper understanding of process dynamics and process monitoring to operate in optimal conditions. MR provides insight into the system. with the minimal number of parameters possible and minimal systematic (e. Secondly. a process running outside this regime can be easily detected. minimal systematic error . but more important. Let us also assume that we have created a model. - - Now let us assume that a specific variable or process parameter cannot be measured due to physical limitations.The complete system.

either chemical or biological.The creation of reduced models and their parameter estimation delivers important information to be implemented in the complex model.Introduction The validity of the approaches proposed will be tested in two case studies of high relevance in the field of water treatment and recombinant cultivations.7 ADVANTAGES OF MECHANISM RECOGNITION The information contained by the complex model has to be used with intelligence to fit the process needs while increasing the identifiability and the observability of the submodels. a reduced model comprehends much more information than the same model built using the classical top down approach (from black-box to greybox to first principle models). This allows the identification of non measurable variables and increases the information obtained by the experiments. - - 10 . the information gained can be exported to systems and used for different conditions. Finally. is the keystone to this approach. By these means the model is adapted to each particular case. it is worth recalling that physical understanding of the system. MR is merely a tool to efficiently apply this knowledge in order to gain a better insight of the system under study. The most important advantages when creating a submodel through an intelligent reduction of a complex model are: Specified adaptations for each process: . On the contrary. Again.A defined model reduction can be carried out for a specific process. because of the mathematical basis. Furthermore. Knowledge about the accurate experiment is gained through model reduction: . the nonlinear interrelation of the states in the complex model can be understood better if the behavior of its reduced models is analyzed. MR does not intend to substitute human reasoning or make up for lack of process knowledge.The model reduction can also be conducted to determine a selected phenomenon of the process. 1. Phenomenon identification: . For example.

THE BAD. Particularly in engineering. reliable. is mainly interested in usefulness of models. Reasons for this are explored in this work. For an engineer the principal aspect to take into account is if a model can bring some advantages in process efficiency or not. the simplest model has shown to perform much better than complex. 11 . nonlinear ones. AND THE USEFUL MODEL It is common to evaluate models as “good” or “bad” and these terms are also used in this work following convention. Although it is true that some special characteristics of a model must be analyzed before using it. Nicolas Cruz Bournazou 1. In many cases. it is necessary and sufficient that it be robust.8 THE GOOD. Engineering. being a practice and industry oriented discipline. A model that robustly describes the simplest part of a system properly is far better than a complex model that mirrors the complete process but has a high probability of failure. the most important question to answer is whether or not certain model characteristics can be exploited aiming at specific goals. For a model to be useful. and descriptive. it is essential to be aware that all the approaches to model evaluations might fail. experience has shown that it is very difficult to predict the functionality of a model.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Still.

.

depends on the characteristic of the model. the study of this work is limited to mechanistic models expressed in the form of Differential Algebraic Equation (DAE) systems applied exclusively for description of process engineering in chemistry and biotechnology (2. 𝑤 is a vector with nw constant input variables. etc.1 DEFINITION This works considers mathematical models exclusively and their application in the description of physical phenomena. 𝑥 𝑡 . Models are evaluated by its simplicity. It is worth reminding. For sake of generality. rigorous modeling includes mass and energy balances. in mathematical terms. Finally we delimit to controlled physical. accuracy. generality. mechanistic models are based on physical knowledge of the system to be described. and biological systems. 𝜃 is a vector with P parameters. The quality of every work on simulation. 𝑡 = 0 2.2 13 . 𝑡0 = 0 where t0 is the time at point 0. design. 𝑢 𝑡0 . 𝜃. The 2.1). robustness. 𝑥 𝑡 is a vector with ns time-dependent variables which define the system.1 where 𝑥 𝑡 is a vector with the derivatives of the state variables. 𝑤. The arts and crafts of mathematical modeling are exhibited in the construction of models that not only are consistent in themselves and mirror the behavior of their prototype. of certain aspects of a nonmathematical system. 𝑓 𝑥 𝑡0 . chemical. 𝑢 𝑡 . we limit our concept of mathematical model to the definition made by Aris [49]: “A mathematical model is a representation. 𝑢 𝑡 a vector of nu time-dependent input variables. 𝜃. Models are the core of Computer Aided Process Engineering (CAPE) [50] and Computer Aided Biology (CAB). that there is no such thing as the “best” model for all applications. The initial conditions are also to be defined. Nicolas Cruz Bournazou 2 MODELING 2. 𝑓 𝑥 𝑡 . Contrary to black box models. and model based control. In engineering for example.An approach to Mechanism Recognition for model based analysis of Biological Systems M. and 𝑡 represents time. detailed reaction pathways. 𝑤. 𝑥 𝑡0 . but also serve some exterior purpose. optimization.” Furthermore. and computation burden.

Modeling “best” model can only be selected after the objective of the simulation and the state of information (chapter 4) has been specified. the three conditions (section 1. they are also essential to map complex systems into smaller dimension more comprehensible to humans. and low parameter identifiability. isolated and analyzed before any model is built.1 : E. Still. On the other hand. coli transcriptional regulatory network. the better [52]. In most of the cases it has shown to be quite the opposite. This last category of models is also known as software sensors [51]. which are not possible due to physical limitations. this should be considered the last resource and should be done only after all other options have been exhausted. In engineering. It is useless to apply Computational Fluid Dynamics (CFD) to the simulation of a 1L reactor knowing that the concentration gradients can be neglected. Currently.000 L without considering mass transfer limitations may yield catastrophic results. they also serve to obtain indirect measurements and observe non observable events. the first solution that comes to mind when a model fails to describe a system is to add new parameters. [53]. over parameterization. models are not only used to describe the behavior of systems. Summarizing. the key dynamics of a system need to be identified. Software sensors substitute measurements. simulating a reaction in a tank with 10. Finally. parameter correlation.6) are limited mainly 14 . 2.2 MODEL COMPLEXITY A common mistake is to consider the most complex model to be the most appropriate for description of a system. Figure 2. Model complexity is closely related to instability. Instead. with models which predict the behavior of the non measurable variable based on indirect measurements. The effort required to develop and fit a model has to be justified by its application. Experience shows that the fewer the parameters in a model.

hyperbolic. usually solving some least square type of optimization problem (section 4. and discontinuous functions might seem advanced and sophisticated.g. In traditional process engineering. These steps are followed iteratively. the modularization of separated instances of the system is not always possible. but this illusion quickly vanishes when the model has to be validated and used for design or optimization. The meaning of model simplification becomes more important everyday with the increasing complexity of processes analyzed in research Figure 2.3 ENGINEERING APPROACH TO COMPLEX SYSTEMS In chemical engineering.1. An alternative method to create optimal model structures has been published by Bardow [58]. Much better is a correct approximation. quality of data required and bias. model reusability. The residual between model outputs and data is calculated and a new set of parameters is tested. In biological systems. a pump can be modeled in a modular form and then added to the flow sheet of the plant and reused as many times as needed [54]. etc. IA extends the philosophy of hierarchical modeling to the molecular level. This method called the Incremental Approach (IA) suggests building the model in an inductive manner. Inverse problem theory is the most common approach for model building and specifically models fit to data. the implementation of different methods to deal with large complex systems has a long history. Engineers have developed methods like hierarchical modeling. In a sense. the differential model is evaluated (integrated) with a certain parameter set. A model with hundreds of parameters including exponential. than an accurate misconception.1) until the residual is considered to be minimal. and then the data is compared against the output previously computed. 57].1. Contrary to this. Although its application finds important limitations. Still. e.An approach to Mechanism Recognition for model based analysis of Biological Systems M. It is at this point that the MR approach can contribute to modern model building. First. biological systems tend to show different behavior under in vitro conditions compared to their in vivo state [55]. An extensive discussion of these methods and their application for the simulation of chemical plants is presented by Barton [3]. The real challenge for modeling is to develop a general and systematic approach to find the simplest manner to describe complex systems aiming at the strictly required accuracy. some approaches intend an analysis and modeling of biological systems with methods taken from engineering [56. model inheritance. Nicolas Cruz Bournazou due to the scarcity of measurement possibilities but also due to the insufficiency of adequate mathematical tools. An important disadvantage of this approach is that it is not possible to directly 15 . IA could be considered a hierarchical approach extended to an even lower layer to first principle phenomena. 2. In principle. the general concept behind IA is worth our attention.

Besides.4 MODELING IN SYSTEMS BIOLOGY 2. correlations. This curve.1 SYSTEMS BIOLOGY We refer to the definition by Kitano [6. is very helpful when building a new set of equations. 59]: “Systems biology aims at understanding biological systems at system level”. various methods exist to indirectly investigate parameter sensitivities. although very noisy in most of the cases and without any physical meaning. The result we obtain is a curve showing the ideal parameter values. 2. Figure 2. This process would be similar to fitting one parameter for each measured point independently. The modeler can visualize the behavior of the parameters and decide if they can be represented by constants. this method can be very useful if advanced measurement techniques are available. Estimating new parameter values for each data assures that the differential equation presents the correct derivative. IA offers a very well described systematic procedure for model building.2: Incremental approach for reaction kinetics identification [58] IA proposes to build the model in a deductive way. and bifurcation among others. which is usually underestimated in process engineering and biotechnology Figure 2. Although. Still. or algebraic or differential functions. A drawback of this approach is that process information is required for each step in the model building process. Bardow [58] proposes finding the parameter value needed to fit each new data point.Modeling analyze the internal structure of the model.2. a true insight in the structure of the model is still not possible. 16 . Systems biology emphasizes the fact that the only possible manner to understand living organisms is to consider the system as a whole.4.

bifurcation.3. The design method . Kitano states four key properties: System structure . whereas real understanding can only be achieved by uncovering the structure and dynamics of the system.4. etc. Identifying genes and proteins is only the first step.2 MODELING OF GENETIC REGULATORY SYSTEMS System biology has triggered an impressive contest between various methods aimed at an adequate description of the dynamics of living organisms studying its Gene Regulatory Network (GRN). Nicolas Cruz Bournazou Figure 2.System design is the effort to establish new technologies to design biological systems aimed at specific goals. the topological relationship of the network components as well as the parameters for each relation. The control method .System behavior analysis suggests the application of standardized techniques such as sensitivity. - - - The relevance of modeling in systems biology is clearly stated in Figure 2. stiffness. System dynamics .g. the most representative being [60]: Directed and undirected graphs Bayesian networks 17 .3: Hypothesis-driven research in systems biology [59]. stability. 2. organ cloning techniques.System structure identification refers to understanding both. e.An approach to Mechanism Recognition for model based analysis of Biological Systems M.System control is concerned with establishing methods to control the state of biological systems.

Because complete understanding of the system is essential for a proper evaluation.3 where x can be the vector of concentration of proteins. detailed descriptions of complete GRN will be 18 . mRNAs. differential equation systems settle the standard modeling method in engineering. Monod type. 𝑢 𝑑𝑡 2. Also time delays can be added if necessary. Usually the system comprises rate equations of the form 𝑑𝑥𝑖 = 𝑓𝑖 𝑥. linear and piecewise linear differential approaches are perfectly suitable for model reduction. and fi is a nonlinear function. the bottle neck is still the state information of the parameter set creating identifiability problems. Systems of Ordinary Differential Equations (ODE) have been widely applied for the description of GRN. or other molecules. it can be expected. For this reason. It is worth recalling that MR aims at simple model building and GRN modeling is far from this. but much work is to be done before an adequate comparison can be achieved. An important advantage of nonlinear ODEs is the possibility to describe multiple steady states and oscillations in the system [62]. that systematic conversion of complex GRN models in simple submodels suitable for MR will be possible in near future. Besides the requirement of testing the global convergence of the optimal solution. Nevertheless. and logoid functions among others.Modeling - Boolean networks Generalized logical networks Linear and nonlinear differential equations Piecewise linear differential equations Qualitative differential equations Partial differential equations Stochastic master equations Each approach offers different advantages and no definitive method can be defined as the “best” by the systems biology community. Someday. Still. The Assessment of Network Inference Methods attempt to analyze all pros and cons of the different GRN inference methods. switching. both GRN modeling and model analysis and reduction techniques have shown exponential development in the last years. u the vector of inputs. As stated before. some successful applications have been published showing the possibilities of ODEs to describe GRN [63]. Therefore. Typical types of equations used are. The goal is to compare the different approaches against equal data sets to obtain quantifiable information of the difference in performance between the methods [61]. Heaviside. the most interesting model approaches for MR are the ones based on differential equations. the most promising results have been obtained with simulated data sets. In fact.

𝑘 ∗ 𝑆𝑙 2. 𝑗 = 1. ⋯ . Assuming that biochemical reactions. ⋯ . 𝑋1 . 𝑙 are the concentrations of the chemical species (substrates and/or products) in the reactor.𝑘 ∗ 𝑟𝑘 𝑆1 . robust and accurate predictions of complex processes. 𝑗 = 1. Nicolas Cruz Bournazou the basis for perfectly defined submodels applied in industry to make fast. where some steps may be independent of those that follow [64].𝑘 ∗ 𝑟𝑘 𝑆1 . This results in a sequence of individual process steps. It is at this stage where the different process conditions can be selected and the submodels can be built. 𝑚 are the concentrations of the microbial masses in the reactor. … . generally described through 𝑆𝑖 𝑟 𝑘 𝑙 𝑌 .An approach to Mechanism Recognition for model based analysis of Biological Systems M. Biochemical batch reactions have been selected to validate MR and its application for the description of industrial processes. ⋯ . which is part of a scheme. A process in constant change presents different behaviors and governing phenomena also change over time. 𝑋𝑚 . … . 𝑙 2. 𝑛 𝑆𝑖 = 𝑘=1 𝑛 𝐶𝑖.𝑘 and 19 . 𝑖 = 1. 𝑆𝑙 . For this reason. 𝑋1 . … . 𝑖 = 1. 𝑋𝑗 . 𝑟𝑘 𝑆1 .5 MATHEMATICAL MODEL FOR A BATCH BIOCHEMICAL REACTOR MR finds its most important application in dynamic systems. … . ⋯ .5 𝑋𝑖 = 𝑘=1 𝑌 . … . 𝑋𝑚 . ⋯ .4 take place in a batch biochemical reactor. 𝑋𝑚 . 𝑚 𝑗 where: 𝑆𝑖 . a short discussion of the general form of the mathematical model is presented. 𝑘 = 1.𝑘 ∗ 𝑋𝑗 + 𝑗 𝑙=1 𝑙≠𝑖 𝐶𝑙. the following differential equations can be derived . ⋯ . 𝑆𝑙 . 𝑋1 . 𝐶𝑖. The biochemical reactions involve consumption of various chemical species (substrates) and production (intermediate or final metabolic products) and biomass growth. 2. 𝑆𝑙 . 𝑛 are the reaction rates. Products from a microbial group are often the reactants of other microbial groups.

𝑋1 . 𝑟 𝑆. ⋯ .𝑘 are the stoichiometric coefficients for substrate consumption and microbial growth. 𝑋𝑚 ⋮ 𝑆 = ⋮ .6 20 . ⋯ . model 2. Introducing vector notation for the concentrations and the rates 𝑆1 𝑋1 𝑟1 𝑆1 . Moreover. i. ⋯ . particulate matter) may not be associated with biomass growth.e. X ) 2.g. in the general case. 𝑙 ≠ 𝑚. X )  X  Y  r (S . 𝑗 respectively. 𝑋 = ⋮ .5 takes a more compact form:  S  C  r (S . the number of the substrates involved in a bioreaction scheme will not be equal to the number of microbial masses grown. a single microbial group may grow on more than one substrate and vice versa. 𝑆𝑙 .Modeling 𝑌 . 𝑋 = 𝑋𝑚 𝑆𝑖 𝑟𝑛 𝑆1 . ⋯ . 𝑋1 . Therefore. It should be noted that the consumption of a substrate (e. 𝑆𝑙 . 𝑋𝑚 and denoting by C and Y the 𝑙x𝑛 and 𝑚x𝑛 matrices of the stoichiometric coefficients.

A very important difference between steady state and dynamic processes is the selection of the “important” dynamics. Nevertheless. model reduction leads to model simplification. Because of the information gained from the detailed model. models are widely used in science and their contribution to a better understanding of engineering processes and their proper design. The process can be described as a bottom up approach in the hierarchical modeling sense. Nicolas Cruz Bournazou 3 MODEL REDUCTION 3. a widely applied approach for reduction of nonlinear models is the linearization based on Taylor series. optimization and control is unquestionable. which has proven to be very useful for processes in steady state conditions. insufficient measurement techniques and extensive computation time hinder an exact representation of the phenomena to be described [65]. Once a detailed model has been built. Model reduction is keystone in engineering. 21 .1 INTRODUCTION A model is a poor mathematical representation of a physical system. increment of model robustness reduction of model stiffness reduction of computation expenses As a result. Unfortunately. From this it can be deduced that the best model to describe a certain process is not necessarily the most accurate. not only the experimental effort for parameter estimation is drastically reduced. but. By these means species are neglected and dynamics are simplified based on their influence on the overall system. Lack of accurate knowledge of the process to be modeled. It may be even possible to convert non observable models into models adequate for model-based process control. but the one that describes only the relevant aspects of the system so as to get a good description with minimal effort [66]. the reduction follows mathematical and physical principles. Different methods have been developed to detect the key dynamics in order to create an accurate but relatively simple model.An approach to Mechanism Recognition for model based analysis of Biological Systems M. the measurement effort during process monitoring and control is minimized. Some of the most important advantages of reducing a model are: increased identifiability/observability. dynamic nonlinear systems require more complex approaches. most important. Model reduction aims at distinguishing the important from the negligible modes in an effort to reduce the model to a more attractable form maintaining its key dynamics [67].

However. Sensitivity Analysis [71] and Time-Scale Analysis [64. the fast modes are neglected.2.Model Reduction In model reduction for continuous processes based on time scale analysis.6. In addition.2 BASIC APPROACHES TO MODEL REDUCTION The reduction of a particular model maintaining its characteristics is a common task in all fields of engineering. Representative examples are Lumping [69. Firstly.1 REACTION INVARIANTS In chemical and biochemical models. when a system is far from equilibrium (as is usually the case in dynamic processes) the fast modes are of major importance and the very slow modes can be considered constant (quasi-steady state). most industrial chemical processes run at steady state (constant operating conditions). which determine the path while reaching the equilibrium point are maintained [68]. 3. 3. On the other hand. A representative example of the reduction potential of dynamic models is shown in section 7. 72. Secondly. Besides many efforts. the method of reaction invariants is applied to detect the modes which stay constant in the process. a combination of methods is required to achieve the simplest form possible. linearization and Lyapunov stability being just two examples. the method finds linear dependencies in the model.3. In this chapter. there will be 𝑙 + 𝑚 − 𝑛 linear combinations of the 22 . 73]. a short introduction to some of the methods used for model reduction is presented. 70]. In the general model of equations 2. On the contrary. reaction invariants help detect the invariant manifold which in turn is a valuable tool to find the slow manifold of a system of dynamic equations. the condition of stability enables many assumptions which simplify the reduction problem significantly. less work has been done in model reduction of dynamic processes. New approaches to model reduction need to be developed to permit reduction of models for dynamic processes. these new constants may give rise to further reaction invariants. which should be considered because of their model reduction potential. Only the slow modes. On the one side. most processes are batch or fed-batch processes which allow no steady state assumptions. As long as 𝑙 + 𝑚 > 𝑛 and the differential equations are independent of each other. In most of the cases. model reduction techniques still relay on process knowledge and experience to achieve the correct reduction of the model. In biotechnology. there are (𝑙 + 𝑚) differential equations that are affected by 𝑛 reaction rates. Literature dealing with model reduction under steady state conditions has been widely published.

(𝑙 + 𝑚 − 𝑛) This means that the 𝑙 + 𝑚 − 𝑛 x(𝑙 + 𝑚) matrix 1  A         has rank (𝑙 + 𝑚 − 𝑛) and satisfies C  A  0 Y    3. In the literature.6 . these are referred to as reaction invariants.2 3. The reaction invariants can be easily calculated from the general form 2.An approach to Mechanism Recognition for model based analysis of Biological Systems M. one can find (𝑙 + 𝑚 − 𝑛) linearly independent row vectors 𝑎𝑣 . as a result of (2) and (3). (𝑙 + 𝑚 − 𝑛) of length (𝑙 + 𝑚) such that     0 Y    C  𝑣 = 1. they capture the reaction stoichiometry relations. 𝑣 = 1. ⋯ . ⋯ . that the quantity  S z  A  X  remains constant throughout the entire batch:  z 0 3.1 It can then be easily verified.3 and so  S (t )   S (0)  A   A  X (0)  t  0  X (t )    23 3. Nicolas Cruz Bournazou concentrations that are completely unaffected by the reaction rates and therefore completely unaffected by the progress of the chemical reactions.4 . which are not affected by the reaction rates. Assuming 𝑙 + 𝑚 > 𝑛 and Rank    n C  Y    .

SF enable activation or deactivation of different reactions depending on the concentration of species. which may or may not occur during the process. then 𝑟 ≈ 𝑟𝑚 = constant. when the concentration of the limiting component is high. The inhibition functions show the opposite behavior. and to 1.5.5. When a model describes a wide range of process conditions all species limitations. We can therefore create reduced versions of the original model. which behave exactly as the original model under specific conditions. the process presents different reaction invariants under limitation of different species.2. rm . We characterize the switching function behavior in three phases Figure 3. which is equal to 1.2 SWITCHING FUNCTIONS AND THE REACTION INVARIANT Switching Functions (SF) are widely applied in biological systems. Referring to equation 3. However.1. have to be considered. K S . To visualize this idea. To be more precise. The presence of SF in a mathematical model may engender temporal reaction invariants during specific time intervals in a process. respectively. it can be seen that if 𝑆 > 100 ∀ 𝑡 > 0. the equation is equal to 0. the maximum reaction rate. 𝑟 = 𝑟𝑚 𝐶𝑆 𝐾𝑆 + 𝐶𝑆 3. considering the simplified Michaelis-Menten equation in 3.5 where r. in case of high concentrations. 𝜗 will have the value 𝜗𝑚 when the concentration of substrate is high and 0 when the concentration of substrate is near 0. when the species is not present. The objective is to create a continuous and smooth function. and equal to 0 when the limiting component has been depleted.5. This can be very useful when applying general mathematical models to defined conditions. some limitations may be neglected and new reaction invariants can be detected. active and constant (species concentration is significantly higher than the limiting constant > 100) transition phase (when the ratio species/constant is between 100 and 1*10-2) inactive and constant (species concentration is significantly lower than the limiting constant <1*10-2) This particular characteristic of the SF opens interesting possibilities for model reduction. Its most common form corresponds to the simplified Michaelis-Menten equation 3.Model Reduction 3. the inverse of enzyme affinity and the substrate concentration. 24 . 𝐶𝑆 are the reaction rate. once the conditions of the process are well defined.

1. [gr/ml] 1 20 Phase 1 Phase 3 0. it is possible to achieve important reductions with some engineering experience and mathematical background.3 SENSITIVITY ANALYSIS The basic principle of model reduction through sensitivity analysis is to eliminate the components which are not relevant for the accuracy of the model. To correctly reduce the order of the system it is important to eliminate only the species.2. which have both a weak effect on the model outputs as well as on the important species of the model.1. This algorithm provides exact sensitivity information of the numeric integration. Cheap computation is achieved through partial discretization tailored to the integration of general DAEs based on Orthogonal Collocation on Finite Elements (OCFE). the algorithm is able to take advantage of the sparsitiy properties common in engineering models.An approach to Mechanism Recognition for model based analysis of Biological Systems M.1. sDACL [46] is a code for efficient integration of general DAEs with sensitivity generation. Local sensitivities can be easily estimated via finite differences.5 0 0 0 2 4 6 -0. and can be implemented in any one-step integration method. Finally. dynamic sensitivities can be computed tailored to the numeric integration of the DAE system. Nicolas Cruz Bournazou Phase 2 40 Substrate conc. Although no general method exists for this procedure. If the Jacobian matrixes of both state variables and parameters can be obtained. an efficient analysis of the dynamic sensitivities is possible.5 8 Time Figure 3. 25 reaction rate [] . sDACL has been customized to the staggered method for state and sensitivity integration showing similar computational efficiency. Behavior of a switching function in dependence of the limiting species. An example of the capabilities of this approach is presented in section 7. 3. were the state of the art model for Active Sludge Process (ASP) is reduced to one third of its size. With efficient simulation tools like sDACL.

the numbers on the arrows represent the back.3 ∗ 𝐴3 𝑑𝑡 𝑑𝐴1 = −𝑘3.7 26 . each one of the reactants to be lumped appears exclusively in one lumped entity. When speaking of proper lumping.6 3.3 ∗ 𝐴3 𝑑𝑡 𝑑𝐴2 = −𝑘2.2.1 ∗ 𝐴1 − 𝑘3. We consider the system to be described by the following monomolecular reaction scheme: 𝑑𝐴1 = −𝑘1.2 ∗ 𝐴2 − 𝑘2.3 ∗ 𝐴3 𝑑𝑡 where k is the monomolecular rate matrix k= 13 −3 −10 −2 −4 12 −6 −10 10 3. Although proper lumping is mostly used in mathematics.2 ∗ 𝐴2 − 𝑘3. A process were the advantages of lumping has be demonstrated is steam cracking [74].1 ∗ 𝐴1 − 𝑘2. improper lumping considers the possibility of one or more reactants contributing to different lumped entities. Lumping can be divided in two main categories. proper and improper lumping. On the other hand. most reactions are only to be described precisely by improper lumping methods.2: Three-component monomolecular reaction system.4 LUMPING Lumping is widely applied when dealing with systems with a large number of components.1 ∗ 𝐴1 − 𝑘1. where the millions of different components need to be grouped in a few categories to allow its simulation.and forward reaction constants.Model Reduction 3. A nice illustrative example is of proper lumping given by [70]: Let us assume we have the following three component system: A3 10 4 2 6 10 A1 3 A2 Figure 3.2 ∗ 𝐴2 − 𝑘1.

… . it is possible to represent the system with two components. one representing 𝐴3 and the second representing a combination of 𝐴1 and 𝐴2 Figure 3.8. On the other hand. real processes can rarely be described by proper lumping since for most systems there is no M matrix which fulfils eq.2. the theory of semi-proper and un-proper lumping can become rapidly complicated and are not discussed in this work [75].3: Lumping a monomolecular three-component reaction into a two-component reaction Unfortunately.3. 𝑖 = 1.8 3. Perturbation theory is a systematic mathematical approach to find the solution for the case ε = 0.5 PERTURBATION THEORY In some problems. it is very important to consider the possibility of lumping different species. 3. 𝑛 such 𝑖 ∗ that: 𝐾𝐴 = 0 For the system to be lumpable with proper lumping it is necessary and sufficient to find a 𝐾 ′ matrix such that: 𝑀 ∗ 𝑘 = 𝑘 ′ ∗ 𝑀 In this example: 1 0 1 0 0 1 13 −3 −10 −2 −4 10 12 −6 = −10 −10 10 −10 1 1 0 10 0 0 1 3. especially in systems with a large amount of components or cells with similar behavior. Nicolas Cruz Bournazou and fulfils the conditions mentioned bellow: Nonnegative rate constant Mass conservation There exists an equilibrium composition 𝐴∗ > 0. it may be the case that a small parameter can be detected (usually ε) such that the solution does not differ significantly if ε = 0. Perturbation theory might be seen as the pure representation of model reduction. A3 10 4 2 3 6 10 A’1 10 10 A’2 A1 A2 Figure 3. However.9 In conclusion. The model is analyzed to eliminate the parameters which cause very small changes in the dynamics of 27 .2. 3.An approach to Mechanism Recognition for model based analysis of Biological Systems M.

3. 3.11 is the exact same solution one would obtain after calculating the Taylor series expansion of the exact solution. Again the real challenge is to find these parameters in complex nonlinear processes. 𝑔 = [𝑚/𝑠 2 ] gravity acceleration. for low initial velocities. a generalized implementation of the method has proved to be difficult. a body with big mass. The engineer must decide which modes are fast enough to be considered steady state. 𝑥 𝑡 = 𝑡 − 𝑡 2 𝑡 2 𝑡 3 𝑡 3 𝑡 4 + 𝜀 ∗ − + + 𝜀 2 ∗ − + 𝑂𝜀 3 2 2 6 6 24 3. Stamatelatou provides a systematic answer to find and remove the fast dynamics of acidogenesis. 3. 𝑑2 𝑥 𝜅Γ 𝑑𝑥 + ∗ +1 =0 𝑑𝑡 2 𝑚𝑔 𝑑𝑡 𝑑𝑥0 𝑥0 = 0. Γ = [m/s] the initial velocity of the projectile. The Quasi Steady State Assumption (QSSA) has been widely applied in chemical engineering and biochemistry [76]. under the assumption that these species are instantaneously at equilibrium. A good example is modeling of a projectile thrown vertically into the air. and low air friction. =1 𝑑𝑡 3. 28 . For the QSSA method empirical knowledge of the system to be analyzed is required. It is possible to consider that the relation 𝜀 = 𝑘Γ/𝑚𝑔 is very small. Stamatelatou [64] published a nice application of time scale analysis for biochemical reaction steps in CSTR. An important question to be answered is if air friction should be considered in the model. A simple example taken from this paper should illustrate the concept of the slow manifold. and 𝑡 = [𝑠] the time. Even though some mathematical methods have been developed to identify fast modes in equation systems. and apply regular perturbation theory to find an approximate solution to eq.11 Eq. The solution obtained. has proven to be a very effective model reduction method. This gives rise to stiff equation systems which are difficult to solve with numerical methods.2. Substituting the fast reactions by algebraic equations. In this paper.10. 𝑚 = [𝑘𝑔] its mass.Model Reduction the model. Still.10 where 𝜅 = [𝑚 ∗ 𝑘𝑔/𝑠] is the friction constant.6 TIME SCALE ANALYSIS Reaction systems are characterized by an interactive combination of fast and slow reactions. the method of perturbation theory allows fairly accurate results if process knowledge and an iterative method are combined.

Considering that the dynamics of the biomass is fast enough to be substituted by an algebraic equation. (3. Nonetheless for sake of clarity.12 where X and S are the biomass and substrate concentrations respectively. depending on which is considered to be the fast and the slow dynamic of the process. Defining the reaction invariant: 𝑧 = 𝑋 + 𝑌(𝑆 − 𝑆0 ) 3. KS is the saturation constant. 𝐹 = 𝐷 ∗ 𝑆𝑂 is the feed rate of the substrate.14 If on the contrary. the following expression is more appropriate: 𝑑𝑧 = −𝐷 ∗ 𝑧 𝑑𝑡 𝑑𝑋 𝑧 − 𝑋 = −𝐷 ∗ 𝑋 + ∗ 𝑟 𝑑𝑡 𝑆 − 𝑆0 29 3.13 We can reformulate eq. The work of the experienced engineer is to detect the fast and the slow dynamics.15 . Stamatelatou deliberately selects the wrong dynamics to show the limitations of the approach. it is very easy to foresee that biomass growth has a slower dynamics than substrate concentration. Nicolas Cruz Bournazou Consider the following model describes a chemostat process: 𝑑𝑋 = −𝐷 ∗ 𝑋 + 𝑌 ∗ 𝑟 𝑑𝑡 𝑑𝑆 = −𝐷 ∗ 𝑆 + 𝐹 − 𝑟 𝑑𝑡 3. the dynamics of substrate is considered to be faster.12) with the following two expressions. Y is the biomass yield. D is the dilution rate. S0 is the substrate concentration in the feed. and assume that it operates under low dilution rate relative the growth of microorganisms: 𝐷/𝜇𝑚𝑎𝑥 ≪ 1. the equation system can take the form: 𝑑𝑧 = −𝐷 ∗ 𝑧 𝑑𝑡 𝑑𝑆 𝑌 = 𝐷 ∗ (𝑆0 − 𝑆) − ∗ 𝑟 ∗ (𝑆0 − 𝑆 + 𝑧) 𝑑𝑡 𝑋 𝑋 = 𝑧 − 𝑆0 + 𝑆 ∗ 𝑌 3.An approach to Mechanism Recognition for model based analysis of Biological Systems M. In this simple system. the correct dynamic is considered to be the fast one. 𝑟 = 𝜇𝑚𝑎𝑥 /𝑌𝑚𝑎𝑥 ∗ 𝑆/(𝐾_𝑆 + 𝑆) ∗ 𝑋 is the reaction rate. 𝜇𝑚𝑎𝑥 is the maximum specific growth rate constant of the biomass.

and differentiable from each other. The efficient recognition of a process regime still depends on the proper building of a set of submodels able to describe the process states accurately while being fast to compute. Still. we can see the trajectories of the two reduction hypothesis made (setting z equal to zero).12). X are plotted for different initial conditions. New techniques in combination with advances in computer technology permit the solution of large systems in a systematic manner [77-80].Model Reduction 𝑆 = 𝑧 − 𝑋 + 𝑆0 𝑌 Now let us make a phase diagram where the trajectories of S vs. it is worth recalling. finding the fast and slow dynamics rapidly becomes a difficult task in large systems. submodel building will become an automatic task in the MR program. Comparison with reduced models in a chemostat process The method of the invariant manifold can be very effective to reduce stiff systems. Finally. In addition. In near future. Nevertheless. that the field of model reduction is showing important advantages towards reduction of complex nonlinear systems. a set of submodels should be automatically available to serve recognition purposes. 600 500 substrate S [mg/l] 400 300 200 100 20 40 60 80 100 120 140 biomass X [mg/l] 160 180 200 220 Figure 3. today the proper reduction of complex models requires much work and collaboration of different disciplines. robust. By these means. 30 .4: Phase diagram of full order model (3.

and its interaction in order to create models striving to defined objectives are discussed. this is known as the inverse problem. Even though some assumption need to be made and the calculations represent merely approximations of the state of information. In physics. since the data set against which the model is fitted is at least as important as the model. An antithesis is created and the model is adapted. To understand a system and model its behavior. In order to do so. is known as parameter estimation or model fitting. It is therefore not possible to determine the usefulness of a model without testing its description accuracy. Nevertheless. a hypothesis is made based on the observations. Following. which best describes the data set. and observations of the real system. parameter set. are to be compared. The model has then to be fitted to the data set. more conventional approaches are applied for engineering applications. Models are created under numerous assumptions and for strictly limited systems. The theory of inverse problems has been well described for linear systems. in order to allow a better analysis of models and estimation of the parameter set has great significance. Inverse problem theory studies the appropriate manner to infer the values of the parameter set using the results obtained from observations [24]. information from literature. New experiments are required based on the information gained during the development of this process. methods to analyze model structure. some experiments are carried out. to achieve this. Nicolas Cruz Bournazou 4 OPTIMAL EXPERIMENTAL DESIGN In chapter 3. First the system has to be observed. it becomes rapidly complex and requires high computation expenses. Next. a long iterative approach is required. Searching for the parameter vector of the model. Variables to be 31 .An approach to Mechanism Recognition for model based analysis of Biological Systems M. The data set is obtained from observations carried out in defined experiments. a model is created on the foundation of the hypothesis. and although it can also be applied to nonlinear problems. It is important that experiments carried out to test models consider these limitations. experience has shown that simple methods deliver fairly accurate results and are useful in many processes [81]. The principal quality of models is its capacity to mirror a defined system. and models created for the same or similar systems. For this reason. a discussion regarding how to design experiments. observations of the real system have to be made and the similarities between model predictions.

it increases the information obtained from the experiment to fit the model parameters with the highest possible accuracy. various models are proposed as possible candidates to describe the system and the model with the highest probability to describe correctly the system is selected. Not only theoretical work but also an important number of applications of MBDoE have been published. among others are updated and a new set of experiments is planed and performed. For modeling purposes. MBDoE takes advantage of the dynamics of the process to increase the sensitivity of the model outputs with respect to the parameters. 32 . considering the characteristics of the model being fitted to increase model identifiability [81] (section 4. Optimal Experimental Design (OED) deals with the selection of the optimal experimental setting. the optimization problem searches for the experimental settings with the highest state information content over a parameter set. The set of experiments is chosen in an effort to maximize the distinguishability between candidates. The goal is to maximize the difference between the outputs of the candidates.2). the most representative examples are [8487]. Also more advanced methods in Knowledge Discovery of Data (KDD) like data mining [83] have been developed for treatment of large data sets.1. MBDoE can be divided in two main categories: 1) parameter accuracy and 2) model discrimination. In this case. the optimization problem has a different objective. and this iterative procedure continues until the accuracy of the predictions is accepted or it is considered that the system cannot be described. experimental conditions. The model is fitted to the new data sets. It is worth to highlight that model distinguishability is closely related to model identifiability. In the general case. In parameter accuracy. These methods study the data characteristics to find new relations between variables and create black-box type models which describe it. In other words. since accurate parameter estimations increase the probability of correct discrimination between candidates.Optimal Experimental Design measured. MBDoE. these methods are grouped under the field of Model Based Design of Experiments (MBDoE). alternative methods have been developed which consider the propagation of data uncertainties through the model and its reflection on the parameter set. In literature. Large parameter sensitivities reduce the confidence intervals of the parameter set. has been widely investigated and has shown important progress in basic research and industry. Methods like Principal Component Analysis (PCA) or Partial Least Squares (PLS) search for data correlation to reduce the dimension of the data set [82]. increasing the accuracy of the parameter estimation. frequency of sampling.

The accuracy on the estimated parameters is influenced by two factors. - An optimal set of experiments should be designed considering both factors mentioned above. On the other hand.1 THE EXPERIMENT In practice. For this reason. The structure of the model defines the interrelations and effects of the parameter values in the output vector. stochastic errors can be defined as the difference between two identical experiments. Nicolas Cruz Bournazou 4. Variables that can be controlled: 𝑢𝐸 (𝑡) time-dependent control values 𝑤𝐸 constant control values 𝑦(𝑡)measurements in the selected time intervals 𝑦(0)initial conditions 𝑡𝑠𝑝 experiment duration Parameters that cannot be controlled: 𝜂𝑠 Systematic error 𝜂 Stochastic error Systematic errors are errors on measurements which can be detected quantified and reduced. The first one is the quality of the data obtained from the experiments. Measurements are inevitably subject to uncertainties. some parameters included in the model are not known beforehand and need to be estimated to enhance model description. Outputs which are highly sensitive to changes in the parameters allow a precise estimation of the “exact” parameter set. Experiments can be defined by the following conditions [81].An approach to Mechanism Recognition for model based analysis of Biological Systems M. 33 . They present a mean value different from zero and error bias. data sets should not be considered observations but a “state of information” acquired on observable variables. The second one is the sensitivity of the objective function with respect to the parameters. the model is fitted to experimental data.

where 𝐶𝑦 represents the variance covariance matrix of n measurements: 2 𝜎𝑦1. 𝑡 4. the experiments output vector y (t ) equals the sum of the models output vector 𝑦 𝑐𝑎𝑙𝑐 𝑡 and a stochastic error 𝜂 𝑦 𝑚𝑒𝑠 𝑡 = 𝑦 𝑐𝑎𝑙𝑐 𝑡 + 𝜂 4. and 𝑡𝑠𝑝 represents the time span of the experiment.3 The stochastic error 𝜂 is considered to have multivariate normal distribution with mean zero.𝑦1 𝐶𝑦 = ⋮ 2 𝜎𝑦𝑛 .4 4.1 THE MAXIMUM LIKELIHOOD Many methods for the quantification of the distance between model predictions and the observations can be found in literature [88]. 𝑡𝐸 ∀𝑡 ∈ 𝑡𝑠𝑝 4. The LSQ is a quadratic equation making it appropriate for most optimizers and easy to compute. it fails to consider important factors. all sources of systematic error have been minimized.𝑦1 2 ⋯ 𝜎𝑦1. and that all control variables have the same value for the model and the experiment.5 where 𝑦 𝑐𝑎𝑙𝑐 and 𝑦 𝑚𝑒𝑠 represent the vectors of the calculated output and of the measured value respectively with length 𝑛𝑦 . 𝜂𝑠 . model outputs are represented in a form similar to experiments outputs.𝑦𝑛 4. 𝜃. 𝜂. For this reason. Nevertheless. If this is true. The most common equation to quantify the residual is the Least Squares (LSQ).Optimal Experimental Design Experiment outputs are represented by the vector y (t ) : 0 𝑦 𝑚𝑒𝑠 𝑡 = 𝑓 𝑦𝐸 . Model outputs are to be compared against the experiment outputs. Model outputs (model predictions) are represented by the vector 𝑦 𝑐𝑎𝑙𝑐 𝑡 .2 Let us considering that all assumptions mentioned above are true. 𝑦 𝑐𝑎𝑙𝑐 𝑡 = 𝑓 𝑥 𝑡 . 𝑢𝐸 𝑡 .𝑦𝑛 ⋱ ⋮ 2 ⋯ 𝜎𝑦𝑛 . 𝑥 𝑡 .1. 𝑢 𝑡 .1 Where the sub index E represents the variables of the experiment. wE . as scaling multiple outputs 34 . 𝑤. 𝐿𝑆𝑄 = (𝑦 𝑐𝑎𝑙𝑐 − 𝑦 𝑚𝑒𝑠 𝑡 )2 4.

Moreover. The size of the confidence interval is a direct indicator of model identifiability. Depending on the correlation of the parameter set. Summed to human error. However. the Maximum Likelihood (MXL) is preferred in process engineering: 𝑐𝑎𝑙𝑐 𝑡. 𝑥 − 𝑦 𝑚𝑒𝑠 𝑡 1 𝑦 𝑀𝑋𝐿 = 2 𝐶𝑦 2 4.6 The variance-covariance matrix 𝐶𝑦 serves two purposes. Since a model is used to mirror a physical system. model reduction can also be applied when complex models present low identifiability.2 MODEL IDENTIFIABILITY As stated in chapter 3. science has no explicit answer to this question. The probability density function of the parameter set and its correlation define the confidence region for parameter estimation. different scales between the outputs are also weighted. Nicolas Cruz Bournazou and considering measurement accuracy. These observations contain various sources of error. In short words. Unfortunately. the reduction of a model depends on its parameter set. To be more precise. even most advanced measurement techniques present uncertainties and produce data with deviations from the physical quantity. the meticulous examination of model dynamics and relation between states. For this reason model reduction is closely related to parameter identification. the theory for approximation of the confidence region for parameter estimation. In the previous section different methods to analyze and alter the structure of models were discussed. On the one hand the outputs related to very noisy data has a smaller impact on the criterion. is also essential. 4. For this reason. and presents alternative ways to obtain similar dynamics with simpler and more tractable models. it is crucial to distinguish model prediction error (systematic error) from observation error (stochastic error. the accuracy of the prediction of the model needs to be quantified. On the other hand. which is based on the MXL criterion. At this point an important question is to determine how to reduce a model correctly without knowing its exact parameter values. Nevertheless. it is of great importance to keep the objective of modeling in mind. For MR. 35 . Observations of the system to be described have to be made in order to test the predictions of a model. allows a better understanding of the behavior of the model. the model has to be correctly fitted before it can be reduced.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Therefore. model reduction affects the identifiability of the model. the confidence interval of the new parameters will be reduced. When a model is reduced some parameters are eliminated.1. In other words. the capability of the model to describe the process behavior excluding the effects of uncertainties has to be evaluated.

𝑢 𝑡 . To quantify model identifiability the objective function Φ𝐼 is maximized eq. In conclusion. 𝑡 = 0  iL   i   iU . 𝜃.7 𝑦 𝑢 𝑡 . 𝜃 ∗ 𝑦 < 𝜀𝑦 4. 𝑥 𝑡 .1: Effect of sensitivities in parameter estimation accuracy. The structure of the model is responsible for the propagation of data variance to the confidence interval of the parameters Figure 4. i  1. Φ𝐼 = 𝜃 − 𝜃 ∗ 𝑇 𝑊𝜃 (𝜃 − 𝜃 ∗ ) s. 4. Asprey [89] defines identifiability as the quality of a model to present a monotonic behavior. but also the quality of the experimental data have to be considered to select the optimal structure of the model. In other words each parameter set shows a unique combination of outputs. 𝜃 ∗ 𝑖=1 𝑇 𝑊 𝑦 𝑢 𝑡 .8 u(t )  U 𝑓 𝑦 𝑡 . 𝜃 − 𝑦 𝑢 𝑡 ... 4. not only the objective for the implementation of the model. This again is directly dependent on the state of information. 𝑤.t. It is the model itself which determines the level of accuracy obtained through parameter estimation considering the quality of the data set. the quality of the model can only be established after knowing how exact can the parameter set be estimated.1. P 36 . 𝜃 − 𝑦 𝑢 𝑡 .Optimal Experimental Design For these reasons the effects of data variance on the probability distribution function of the parameter set has to be considered. 𝑥 𝑡 . In other words. σP and σy represent standard deviation of parameters and measurements respectively.8.7 keeping the difference of models output smaller than some threshold eq.. low sensitivity σP σy high sensitivity σP σy Figure 4. 𝑛 𝑦 4..

In addition global optimality should be guaranteed. which increases computation time even more. and 𝑊𝜃 an arbitrary weighting matrix.. 𝜃 and 𝜃 ∗ the parameter sets with 𝜃 ≠ 𝜃 ∗ . Nevertheless. which can be quantified by the variancecovariance matrix. The standard procedure to increase mode identifiability is to detect the parameters with small sensitivity and high correlation and take them out of the optimization variables. this approach might offer important reduction in later experimental and design effort. but also on the structure of the model. The number of parameters to be estimated can then be iteratively increased as the identifiability of the model increases..2 THE FISHER INFORMATION MATRIX 4. The approach proposed by Asprey represents a high computational burden. therefore its influence in the outputs. When a model is to be fitted to a data set.1 THE CONFIDENCE INTERVAL The propagation of data uncertainties through the model and its effect on parameter estimation has been subject of short discussion in previous sections. since identifiability studies are realized in the early stages of model developments and discrimination. It is essential to understand two characteristics of the model structure: The impact of each parameter on the dynamics of the model. This procedure enables a more accurate parameter estimation in early stages and thus a better design of experiments. 𝑢 𝑡 a vector of 𝑛𝑢 time-dependent input variables.2. 𝑥 𝑡 is a vector with ns time-dependent variables which define the system. all parameters should be considered in the parameter estimation. This very rough approximation is only correct near the vicinity of the parameter values used to calculate the FIM.An approach to Mechanism Recognition for model based analysis of Biological Systems M. the accuracy of the estimation of the parameter vector depends not only on the accuracy of the data set... Nicolas Cruz Bournazou  i*L   i*   i*U . 4. the selection of these parameters. is carried out based on the Fisher Information Matrix (FIM). which is an approximation based on first order derivative information. Therefore. i  1.. 37 . Nonetheless. and the superscripts L and U represent the lower and upper bound respectively. 𝑤 is a vector with 𝑛𝑤 constant input variables. ideally. 𝑥 𝑡 is a vector with the derivatives of the state variables. P where Φ𝐼 is the objective function to maximize. 𝑦 𝑡 is a vector of ny outputs. which are assumed to be highly correlated and with low sensitivity.

𝜃𝑒𝑠 the estimated parameter vector. and 𝐶𝑟𝑒𝑠 represents the variance covariance matrix of the residuals.𝑘 the measured and the calculated variables respectively.2 0. Assuming that: All parameters are constant Input variables behold no uncertainties Stochastic error can be described by a normal distribution over model predictions with the exact parameter set (which is unknown) There is no correlation between measurement uncertainties Measurement uncertainties can be described by a normal distribution with mean equal to zero.2 0 P1 0.9 where ℘ represents the probability density function. The most precise formulation of the problem is to calculate the probability density function of the estimated parameter set.5 P2 -1 -1.𝑘 and 𝑦𝑖.6 -0. 𝐶𝑦 represents the variance of the measurements.8 1 38 .4 0.𝑘 Cy det 𝐶𝑟𝑒𝑠 𝑡(𝑘) 𝑘=1 1 − 2 4. 𝑚𝑒𝑠 𝑐𝑎𝑙 𝑦𝑖.4 -0. 1 0.6 0.5 -0.5 -2 -2. hence how they correlate to each other.𝑘 𝜃𝑒𝑠 − 𝑦𝑖.Optimal Experimental Design - The interaction between all parameters of the parameter set. The probability density function of the estimated parameter set can then be estimated with the following equation: ℘ 𝜃𝑒𝑠 = 2𝜋 −2𝑛𝑁 𝑛 𝑁 2 𝑁 1 exp − 2 𝑖=1 𝑘=1 𝑐𝑎𝑙 𝑚𝑒𝑠 𝑦𝑖.5 0 -0.8 -0.

and 𝑌𝑋𝐴 .11 where 𝜃𝑒𝑠 is the estimated parameter vector. the accuracy of the parameter estimation depends on the sensitivity of the objective function to changes in the parameter set.𝑡 𝑘 4. coli fedbatch cultivations presented by Lin [91] (see Appendix A). white noise was added to the data set (simulated normal distributed variance with zero mean). 𝑁 𝐹𝐼𝑀 𝜃𝑒𝑠 = 𝑘=1 𝑑𝑦(𝑥.2 APPROXIMATION OF PARAMETER VARIANCECOVARIANCE MATRIX As stated before. for this calculation.2.10 where 𝐶𝑃 and FIM are the variance-covariance matrix of the parameter set and the inverse of the FIM respectively. that the FIM depends on the control variables of the experimental setup. FIM gives a linear approximation of the confidence interval region of the parameter set to be estimated.2: Confidence interval from the Lin model. increasing the accuracy of the parameter estimation. and 𝐶𝑦 is the variance-covariance matrix of the measurements. an approximation to this can be 39 . The most effective method to compute the real nonlinear confidence interval is to carry out Montecarlo simulations [90]. 𝑢. Nicolas Cruz Bournazou Figure 4. We now have all the information we require to build an optimization problem and find the optimal experimental setup which maximizes the state of information obtained from experiments. obtained with Montecarlo simulation. 𝑡𝑘 is the time point of the measurement. the FIM is computed to approximate the confidence interval and find the optimal experimental set. Hence we can calculate the lowest bound of the variance-covariance matrix in a cheap manner. In practice. the selected parameters are: 𝑞𝑆𝑚𝑎𝑥 . 𝑡) 𝑑𝜃 𝜃𝑒𝑠 . 𝑥 is the state variables. 𝑢. This effect is caused by the nonlinearities of the model. propagation of the uncertainties through the model. results in a none normally distributed confidence interval of the parameter set. Based on the Cramer-Rao Bound (CRB) theorem: 𝐶𝑝 ≥ 𝐹𝐼𝑀−1 (𝜃𝑒𝑠 ) 4. and Yield acetate to biomass. The model used for this example is a model for E. and θes the vector with P estimated parameters. Figure 4.An approach to Mechanism Recognition for model based analysis of Biological Systems M. From equation 4.11 can be deduced. 𝑡) 𝑑𝜃 𝑇 𝜃𝑒𝑠 . a cheap technique to compute the confidence interval is required. Even though. In other words.𝑡 𝑘 −1 𝐶𝑦 𝑡𝑘 𝑑𝑦(𝑥. 4. 𝑢 is the control variables. maximal Substrate uptake capacity. To be able to plan experiments which reduce the confidence interval of the parameter.2 shows the confidence interval of a two dimensional parameter set.

The most common criteria are shown in Table 4.1: Criteria for confidence interval quantification [92]. Finally.criterion E-criterion M. the objective of design of experiments for parameter accuracy is to find an experimental set which increases the information content of the data and hence allows a more accurate parameter estimation.3.3 LIMITATIONS OF THE FISHER INFORMATION MATRIX It is essential to consider the limitations of the FIM at all time in order to avoid wrong interpretation of the results or its misuse in experimental design. FIM is the result of 40 . A.2.1 Table 4. the FIM must be compressed into scalar form to create an appropriate objective function for optimization.Optimal Experimental Design calculated with the FIM. different criteria are implemented in an effort to create an efficient measurement of the variance-covariance matrix in scalar form.criterion 1 𝑡𝑟𝑎𝑐𝑒(𝐶) 𝑛 𝑑𝑒𝑡(𝐶 𝑛 ) max 𝑒𝑖𝑔𝑒𝑛𝑣𝑎𝑙𝑢𝑒 𝐶 max 𝐶𝑖𝑖 2 1 − 1 [93] [94] [95] [96] 4. Figure 4. To achieve this.3: Criteria for optimization [92] Summarizing.criterion D. A graphical description of the most common criteria for the two dimensional case can be seen in Figure 4.

An approach to Mechanism Recognition for model based analysis of Biological Systems

M. Nicolas Cruz Bournazou

two important mathematic approximations. The validity of these approximations is restricted to a small range near the region of calculation Figure 4.4 shows how the nonlinearities of the model affect the shape of the confidence interval as it grows.

Figure 4.4: Shape of the confidence interval for different variance values from the Lin model (appendix A). The confidence interval can be approximated by an ellipse near the exact value.

Once the variance of the data set has been quantified and set constant, the confidence region of the intervals is determined by the objective function of the parameter estimation program and its sensitivity to changes on its parameters. If we utilize the MXL to estimate the residual of the models prediction, and the model is linear, the resulting objective function will be a quadratic function. This means that we only need to know the Hessian of MXL and the variance covariance matrix of the measurements to calculate the exact confidence interval, again strictly for the case of linear models. On the contrary, nonlinear models cannot be described so easily.

MXL

θ1

θ2

Figure 4.5: Objective function of a nonlinear model (appendix A) with respect to changes in a two dimensional parameter set.

41

Optimal Experimental Design

Depicted in Figure 4.5. is the form of the objective function where the shape differs from a parabolic function. Still, to avoid the expensive computation of the Hessian tensor, which may be even impossible for large problems, the gradient information of the state variables with regard to the state vector and to the parameter set is used to obtain a fair approximation. This approximation, which as stated before is only accurate near the vicinity of the exact parameter set, is the FIM.

4.3 MODEL DISCRIMINATION
In Engineering, the use of mechanistic models for the simulation of complex systems is a widespread procedure [48]. Models are created on the basis of the physicochemical phenomena occurring in the process. Rigorous models are, however, compared against the so-called black box models laborious to develop, adapt, and fit. The challenge increases when processes presenting a non-linear and dynamic behavior are to be modeled. As a result, various models, which can differ in structure, size, application and / or complexity, are usually to be found in literature. The questions that arise are: How to quantify the capability of the different models to fit the process? How to select the model that best fits the needs of the process? How to create a minimal set of experiments to achieve question one and two?

In statistics, the field concerned with finding the most appropriate model for a specific process is called Model Discrimination (MD). MD deals with maximizing the probability of a correct model selection considering data uncertainty. To achieve this, an optimal set of experiments is planed with the aid of statistical tools. This approach is widespread in biology and biotechnology [97-99], in physics [100] and in process engineering [101]. Solutions for the adaptation of linear and nonlinear systems and for the discrimination of dynamic models [102-104] can also be found in literature. The first work to address the problem of finding the most appropriate model to describe a set of data was presented by Hunter [105]. Hunter defined the goal of MD as: “Perform the experiment which will most strain the incorrect model to explain the data”. MD is briefly presented in this chapter, for a better understanding of this approach we refer to [52, 81, 106]. The general procedure for MD is rather intuitive. The first step is to select some models which could describe the system, called candidates. Following, experimental setting is designed in an effort to maximize the probability of choosing the correct model. To

42

An approach to Mechanism Recognition for model based analysis of Biological Systems

M. Nicolas Cruz Bournazou

achieve this, the experimental conditions which maximize the difference in the outputs of each candidate are selected. In statistics, Bayes‟ theorem calculates the precise probability function of a model being the most appropriate to describe the data set. Nevertheless, it requires accurate information of the distribution density function and only quantifies the probability associated with a model prediction. The engineer needs to consider various factors when selecting a model and requires an effective and simple method. To state the problem we first discuss how to quantify the “goodness” of a model. This is not an easy task. The obvious procedure is to calculate the residual 𝑟𝑒𝑠. The residual is composed of two different types of error. 𝑟𝑒𝑠 = η + ηS
4.12

The stochastic error η is caused by the uncertainties in the system and measurement methods and it should ideally be randomly distributed around the true value with mean 𝜇 = 0 . The systematic error ηS is caused by defects on the model, e.g. neglected phenomena, not considered external disturbances, incorrect structure, etc. In order to calculate which model describes the processes with highest certainty possible, the variance of the data needs to be known and all models should have its “exact” parameter set. Both errors, stochastic and systematic, are expressed by a probability distribution function hence discrimination between models can only be correct within a certain probability. Furthermore, the state of information and the identifiability increase together with the number of experiments. Therefore, the probability of selecting the “best” model increases with each new experiment. For this reason, in order to make the most certain decision, all candidates should be considered throughout the complete discrimination process. However, this represents an enormous experimental effort [106]. Franceschini [52, 81, 106] rather proposes to conduct a gradual discrimination of candidates in an iterative process. This procedure reduces computation costs in each iteration and allows a more specific design in the next iteration.
Table 4.2: Types of sum of square [22]

Error variance Akaike‟s Final Prediction Error Akaike‟s Information Criterion

2 𝜎𝑟𝑒𝑠 = 𝑀𝑋𝐿

𝑛 − 𝑝 𝑛 + 𝑝 + 1 𝑛 − 𝑝 − 1

2 𝐹𝑃𝐸 = 𝜎𝑟𝑒𝑠 ∗

2 𝐴𝐼𝐶 = 𝑛 ∗ ln 𝜎𝑟𝑒𝑠 + 2 ∗ 𝑝

43

This is by far not the “best” model.Optimal Experimental Design Shortest Data Descriptor 2 𝑆𝐷𝐷 = ln 𝜎𝑟𝑒𝑠 + 𝑝 + 1 ∗ ln⁡ (𝑛) Where 𝑀𝑋𝐿 is the Maximum Likelihood estimation eq. The most representative criteria are presented in Table 4. In addition. Unfortunately. since calculating the MXL is sufficient for selecting the best model. the residual only shows the ability of the model to describe the data set. 𝑛 the number of measurements. 4. During the recognition process. Between two candidates with the same description accuracy.6 value of the minimum sum squared. the model with the highest identifiability should be selected. and extrapolation capacity are a few examples of important model characteristics which should be taken into account during MD. This has to be considered when building the submodels. Verheijen [22] proposes five selection criteria: Physicochemical criteria: physical meaning of the model Flexibility criteria: application in similar processes and reusability. An essential condition to apply MR to a process is to guarantee model distinguishability in every interval. Other forms of the sum of square intend to quantify further aspects of the models to help proper discrimination. Computational criteria: computation expenses and robustness Statistical criteria: accuracy of the prediction (level of state information). In MR. model distinguishability between the submodels should be guaranteed beforehand (during the model building stage). the program for MR computes the minimal length of the interval required to assure. he misses to mention the importance of identifiability as selection criterion. For this reason description accuracy is the only discrimination criteria to consider. and 𝑝 the number of parameters.3. Model distinguishability in each time regime depends on initial conditions and process state information.1 MODEL DISCRIMINATION IN MECHANISM RECOGNITION For recognition purposes MD is a straight forward procedure. 4.2. The reader is referred to [107] for detailed summary. candidates have been generated to detect different regimes. firstly identifiability and next distinguishability. Engineering criteria: usefulness of the model in the real process Still. 44 . physical meaning. Robustness.

the complexity of model building can be deduced. Nicolas Cruz Bournazou In this section the difficulties to discriminate between various models have been discussed. It is necessary to know the state of information of the data set before or at list while the model is being built. this section intends to emphasize two important aspects in modeling: Description ability is not the unique quality of a model worth considering during model building. The state information of the system (whether laboratory. when he is not even able to select the best among a group of models.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Many characteristics have to be taken into account and there is never a simple manner to define the “best” model. To summarize. In order to create an adequate model. A scientist cannot be expected to build correct models. pilot plant or real process) is essential for the selection of the optimal features of the model. the required characteristics of the model and the objective of modeling have to be previously defined. From this. - 45 .

.

1 CODE GENERATION 5. 47 . which combines equation-based modeling.1.1. Simulations were carried out with standard ODE solvers in the simplest cases whereas the integrator sDACL (section 5.1 ) and SBPD (section 5. Model and code generation were supported by MOSAIC (section 5. The toolbox supports automatic code generation in a variety of programming languages (C. Nicolas Cruz Bournazou 5 CODE GENERATION. This feature offers important advantages for hierarchical modeling as well as for efficient and simple development and analysis of reduced models in general and submodels in particular.2). the minimization of the programming effort. among others) and stores the information of the equation systems in the generalized markup languages XML and MathML. various software packages and programming languages were used. 5. the avoidance of errors and effort in documentation.2. the encouragement and support of cooperative work. Furthermore. use of symbolic mathematic language.1. MOSAIC promotes modular modeling at a high level Figure 5. following a new modeling approach for equation reuse and support of different nomenclature conventions. 108] is a modeling environment.1. Matlab. Fortran. SIMULATION AND OPTIMIZATION In this project. Finally. and code generation. The MOSAIC modeling environment is a tool to implement customized models aiming at: the minimization of modeling errors. different optimization algorithms (section 5.An approach to Mechanism Recognition for model based analysis of Biological Systems M.1 MOSAIC MOSAIC [44.3) were required to solve the problems addressed in this thesis.1) was applied for efficient integration and sensitivity computation with piece-wise constant inputs and parameters.

The toolbox is built in a modular way. 48 .org.2 SBPD SBPD is a Systems Biology Toolbox for MATLAB developed [45].sbtoolbox.1. as depicted in Figure 5. which are used to represent models and experimental data. The toolbox offers an open and extensible environment for the analysis and simulation of biological and biochemical systems limited to ODE systems. enabling model exchange within the systems biology community. simulation and optimization Documentation-level modeling Article Overview Article Equations Notation Context MOSAIC Notation Equations Fortran Model in XML/MathML Other languages Notation gPROMS C/C++ BzzMath Matlab Aspen ACM CC UAM BzzMath Figure 5. The base elements are objects of classes SBmodel and SBdata.1: High level modeling with MOSAIC [46] 5. The Systems Biology Toolbox for MATLAB is open source and freely available from http://www. Models are represented in an internal model format and can be described by entering biochemical reaction equations.Code generation. An important advantage is support of the Systems Biology Markup Language (SBML) models.2.

Nicolas Cruz Bournazou Figure 5. 5. being an implicit numeric method. For simulation in the MR program. an important characteristic for efficient partial discretization approach especially when having a high number of discontinuities or even multiple shooting. The program sDACl developed by Barz [46] at the chair of 49 . Also.2 SIMULATION This work shows that MR is a powerful tool to create simplified models and use them to describe complex processes. extending the method to the OCFE has shown important reductions in computation expenses for the full discretization approach in dynamic optimization. because of the calculation of the FIM and optimization of nonlinear programs using optimality criteria. Solving the set of differential algebraic equations accurately even in the case of complex systems with highly nonlinear behavior is essential to enable the application of MR. the approach for large-scale dynamic models which takes advantage of the sparsity of the system and generates first and second order sensitivities is applied. first and second order dynamic sensitivities offer important advantages [46]. OC presents great stability properties for systems with index one or higher. OCFE may also be implemented on direct applying the partial discretization approach to solve stiff implicit DAE systems [111]. Furthermore. OC has shown very good performance especially when solving highly nonlinear models. achieving the highest accuracy order.An approach to Mechanism Recognition for model based analysis of Biological Systems M. solvers based on OCFE have the ability of self-starting in high orders.2. Prove of its efficiency solving typical process engineering problems is the important number of solvers for general differential-algebraic optimization problems developed based on this method [110]. 5. e.g. In addition. Optionally. freely available third party software packages are used.1 SDACL Logsdon [109] showed the equivalence of the Orthogonal Collocation (OC) to an implicit Runge–Kutta method. to (i) perform bifurcation analysis and (ii) realize an interface to SBML [45].2: Modular structure of the toolbox. Besides. transforming dynamic equations into algebraic ones. The toolbox is designed in a modular way.

Finally. PSO is a stochastic optimizer which has shown advantages over other stochastic techniques [2. and compiled MEX library files. The interface environment Tomlab was utilized. optimal control time and global optimization. sDACL is a code for efficient integration of general DAEs with sensitivity generation. Tomlab is an optimization environment containing different state-of-the-art optimization software packages [112]. The external solvers are distributed as compiled binary MEX DLLs on PC-systems. the family of residual optimizers was selected in general and the Gauss-Newton method in particular. 50 . Solving the parameter fit problem with residual information has shown to be a very efficient method to overcome difficult regions caused by parameter correlation [113].4]. The principle of PSO is based on a population (swarm) formed by a certain number of particles (i). the algorithm is able to take advantage of the sparsitiy properties common in engineering models. which search for a global optimum in the defined region of n dimensions. GLOBAL OPTIMIZATION Finally. For these cases. the approach to stochastic optimization based on Particle Swarm Optimization (PSO) was selected. optimization programs need to be solved including parameter estimation. simulation and optimization process dynamics and operation of the TU-Berlin has been selected for the integration of the DAE systems. 5. NONLINEAR QUADRATIC PROBLEM During the optimization of the optimal time lengths. PARAMETER ESTIMATION For parameter estimation purposes.3 OPTIMIZATION For the application of MR.Code generation. some problems require solution of high nonlinear programs where global optimality is required. to implement the different optimization programs with the code written in Matlab®. sDACL has been costumized to the staggered method for state and sensitivity integration showing similar computational efficiency. and can be implemented in any one-step integration method. the optimization problem was formulated as solved based on Sequential Quadratic Problem (SQP) algorithm [114]. Cheap computation is achieved through partial discretization tailored to the integration of general DAEs based on OCFE. This algorithm provides exact sensitivity information of the numeric integration.

if these regimes can be detected. Similarly to steady state processes. The ultimate goal of process modeling is maximization of prediction accuracy. Still. the best candidate is also an indicator of the predominating phenomenon in the selected regime. This is the basic principle of MR. In dynamic processes models which do not fit the complete process data may still fit some time intervals. If the candidates set for discrimination in each time interval are first principle models. Drawbacks of this method are an unavoidable reduction of both. Most dynamic processes present two or more quasi steady state regimes. To overcome these problems. allows process description with multiple but simpler models. In other words. in order to describe the dynamics of a complex system. experimental effort needs to be increased drastically and advanced simulation tools (software and hardware) are required. different regimes of the process with different pathways could be detected. complex and highly nonlinear models have to be developed. parameter identifiability and model robustness. To achieve this. increasing model complexity cannot be achieved without increasing modeling and experimental effort exponentially. Nicolas Cruz Bournazou 6 AN APPROACH TO MECHANISM RECOGNITION 6. in dynamic processes. dividing the process in different regimes with simpler dynamics and describing them separately. usually model complexity is increased exponentially with respect to prediction accuracy. where a reaction pathway is selected among a number of candidates applying MD. the MD approach can be applied to different regimes during the process. isolated and modeled by different simpler models. Instead. but also an accurate description of complex dynamic processes with relatively simple models. In processes which have no steady state condition. MR offers an adequate framework to analyze experimental information and model dynamics. the process is described with remarkably less effort and higher accuracy. The models for MR need to be 51 . This approach does not only enable a better understanding of the process and its dynamics.1 A SHORT INTRODUCTION TO MECHANISM RECOGNITION The method for MR has been developed to recognize different regimes in dynamic processes.An approach to Mechanism Recognition for model based analysis of Biological Systems M.

which is only capable to describe concentration dynamics caused by reaction. - - - 52 . which can only describe the change in concentration caused by diffusion. A submodel 1.Because of the systematic mathematical approaches applied for its reduction. called submodels. which would affect the accuracy of the parameter estimation. and a submodel 2. Consider we have a model which can describe both phenomena in a batch reactor: 1) diffusion and 2) reaction.Reduced models (submodels) present a higher identifiability. lower calculation costs and are more robust. but also information about the process is gained.In conjunction with fault detection approaches.An approach to Mechanism Recognition built on the basis of process expertise. knowhow and also taking advantage of modern mathematical tools. In this example. submodels help detect the influence of non measurable phenomena in the process. a much more accurate detection of the regime where transport limits the process and the regime where reaction limits the process can be achieved. but because of their limitations (only diffusion or only reaction). At this point an approach to MR [42] to detect different dynamics based on simplified models can be of great help.The physical basis of the complex model and of the submodels allows a better understanding of the process and an easier detection of the key dynamics of the system. submodels help analyze. Not only have both submodels less parameters. understand and identify the complex model. By these means. Simulation . We can reduce this model and create two submodels. it would be possible to monitor the minimal speed of the stirrer needed to avoid transport limitation. it is not only possible to describe the process with simpler models and higher accuracy. have important advantages over the complex model as are: Analysis of the complex model . These reduced models. It is worth reminding that conclusions drawn by the recognition program depend strictly on the expertise during submodel building. Process understanding . This model would typically contain highly correlated parameters. Monitoring . Complex models created to describe complete processes are reduced to form submodels limited to the description of a specified phenomenon.

R is the resistance of the membrane (subscript M) and cake (subscript C) respectively. 𝛼 the specific cake resistance.1. 115].1 where ∆𝑝 represents the pressure difference. MR was applied to monitor fowling in Membrane Bioreactor (MBR) processes for waste water treatment. five models obtained from literature were selected [118] to fit experimental data obtained in a test cell and in a pilot plant: Cake filtration ∆𝑝 ∗ 𝑡 ∗ 𝐴 𝜈 ∗ 𝛼 ∗ 𝑘 𝑉 𝑡 = ∗ + 𝜈 ∗ 𝑅𝑀 𝑉 𝑡 2 𝐴 𝑉 𝑡 𝑅𝐶 = 𝛼 ∗ 𝑘 ∗ 𝐴 6. 𝑉 the filtered volume. 𝜈 the dynamic viscosity.2 where 𝑋𝑃 represents volume of solid particles retained per unit filtrate volume. Nicolas Cruz Bournazou 6. a short example based on the first published application of MR is presented. Nevertheless. 𝐿 the membrane thickness. 𝑡 the time. 𝐴 the area.15 – 60 s every 3 – 10 min of filtration Maintenance cleanings are carried out every 2 – 7 days and main cleanings once or twice a year.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Standard blocking 𝑡 𝑉 𝑡 = 1 𝑉0 + 𝑋𝑃 ∗ 𝑡 𝐿 ∗ 𝐴0 6. For the application of MR. Various models for the description of membrane fouling can be found in literature. Further details of the process and recognition of regimes can be found in [43.1 ILLUSTRATIVE EXAMPLE To illustrate the concept of MR. which is a highly dynamic process [116. 𝑉0 initial volumetric flow rate. 𝑉 𝑡 the volumetric flow rate. 53 . 117]: Backflush/relaxation is employed for approx. 𝑘 is a dimensionless constant. most models have been developed and validated under steady state conditions. and 𝐴0 the initial area.

Figure 6. each model is applied only in the time regime.1: Model fit a) without setting bounds b) with setting bounds for physical parameters.An approach to Mechanism Recognition - Intermediate blocking 𝑡 𝑉 𝑡 = 1 𝑉0 + ∆𝑝 ∗ 𝑠 ∗ 𝑡 𝜈 ∗ 𝑅𝑀 ∗ 𝑉0 6. This allows not only a good 54 . [119] MR offers an alternative and very practical solution to this problem. Complete blocking 𝑡 𝑉 𝑡 Cake filtration 𝑉0 −1 𝛼 ∗ 𝑋0 𝛼 ∗ 𝑉 1 𝑉 𝑡 𝑟 = − ∗ 𝐴 ∗ 𝑅𝑀 𝐴 ∗ 𝑅𝑀 𝑉 𝑡 𝑉 𝑡 = 𝑉0 − ∆𝑝 ∗ 𝑠 ∗ 𝑉 𝑡 𝜈 ∗ 𝑅𝑀 6. the time periods where each phenomenon is dominant. In addition the model must foresee. Furthermore.1 show. To avoid increment on model complexity.3 where 𝑠 represents blocked area per unit filtrate volume. By these means. 3) pore blocking. The physical explanation for the inability of all models to fit the data is that dynamics of this particular process are governed by different phenomena over time. 2) pore constriction.5 The results of model fitting in Figure 6. The only solution to describe the process truthfully is to create a complex model which encloses all phenomena: 1) cake formation. the residuals obtained are extremely large. that none of the five models is able to fit the data. the five models obtained from literature are used to describe discrete intervals of the process. where the model is able to describe the process. if physical boundaries of parameter are considered.4 6.

3 Cleaning strategy based on MR [43] 55 . but also enables an indirect detection of the dominant phenomenon in each regime. Figure 6.2: Comparison experiment/simulation using a) just one model. Nicolas Cruz Bournazou description of the complex process with simple models. if models enclose physical understanding of the process.2 shows a regime detection with mechanism recognition. In other words. Figure 6. B) various models [119] The program strictly detects the model which describes the present regime with the highest accuracy. MR can use first principle models to make indirect measurements of non-observable variables and bring new information to light. Nevertheless. Figure 6.An approach to Mechanism Recognition for model based analysis of Biological Systems M. an indirect recognition of the dominant phenomenon is possible.

Detailed model available for the description of the general process of the form: . This protocol facilitates a data based process maintenance. The minimal length of each time regime assures model distinguishability. the recognition approach can detect the fouling type and based on this a cleaning strategy is selected. Moreover.2 METHODOLOGY FOR MECHANISM RECOGNITION In the previous example models with only one variable defining the state of the system are considered. The method is required to allow a regime recognition also in complex systems with multiple state variables and limited process data. Unfortunately. but only in the special case where all models are observable (no general structure is needed) and the frequency of the measurements is high enough to assure continues initial value information for all state variables (no input-output consistency required). - 56 . The solution of this type of problems is straight forward and will not be considered in this study. This is a very practical condition since output input consistency between all models is automatically guaranteed.no discontinuities . which solves the problem of initial conditions for each new regime. The present study is constrained to following type of problems: Dynamic (non-stationary) process.DAE system index zero or one . . in practice multiple state variables are to be considered and online measurement information cannot be assured. Processes with two or more regimes. a scheme for action to take based on MR results was developed. 6. The models present no bifurcation.no model discrimination is required in the first interval. For this reason an approach to deal with these drawbacks needs to be developed. The sequence of the regimes is known. Models with different number and types of state variables can still be applied for MR.3. it was possible to measure the single state variable at high frequency.An approach to Mechanism Recognition This ability to indirectly detect non-measurable variables can be exploited to monitor processes and allow a better operation. As seen in Figure 6. For the case of the MBR process. which increases the efficiency of the process significantly.known initial conditions Parameter boundaries have to assure global convergence with gradient based optimizers. Initial regime is known.

e. general structure. all models are derived from one single complex model. 6. the state of information of the complete process can be used increasing the number of measurements available for its estimation.3 PROGRAM STEPS 6. Much care has to be given to adequate description of the system maintaining identifiability and distinguishability between models.An approach to Mechanism Recognition for model based analysis of Biological Systems M. etc) and be easily distinguishable. and enable proper transfer of parameter values. it is convenient to distinguish the parameters in two groups: Process Constant Parameters (PCP) Regime-Wise constant Parameters (RWP) In the reduction process to build the submodels. most general model reduction approaches are restricted to systems where Lyapunov stability can be proved. Unfortunately. This is not only important for PCPs.3. and the state of information obtained from the process (chapter 4). Nicolas Cruz Bournazou 6. The development of a general method to reduce nonlinear models and find adequate submodels is a very interesting challenge for process modeling and novel promising approaches are already under study [73]. Model reduction techniques (chapter 3). Phenomena delimitation and model adaptation for each regime is still a difficult task. These submodels should fulfill all the restrictions mentioned previously (input-output consistency. analogies to other models from literature.3. this thesis shows the reduction potential of complex models and its advantages. conductivity. For this reason the reduction of the complex model into submodels is still limited to specific solutions where modeling experience and process understanding are essential. Since identifiability is deeply influenced by the number of parameters to be estimated.3. some parameters are present in all submodels and remain constant throughout the whole process (PCP). empirical knowledge. A very simple reformulation of FIM is 57 . limitation and inhibition constants. generate a proper general structure (section 6. which aims at description of the complete process.1 SUBMODELS In order to assure a consistent process description throughout all regimes. Since PCPs do not change over the process.2).g. a general method with reasonable computer expenses is still to be developed.2 GENERAL STRUCTURE It is essential to maximize submodel identifiability to increase the efficiency and accuracy of the MR program. parameter values transmission. Nonetheless. Still. are combined to create an adequate set of submodels. but also RWPs can be estimated with higher accuracy.

4. the quality of the data set required to distinguish between the model outputs can be calculated. 58 . the order of the regimes is to be known and only the precise conditions at which changes take place are unknown. all submodels consist of the general structure and differ only in one parameter which is responsible for submodel distinguishability. Finally. Furthermore. By using eq. MR is required. and a general structure which does not include all the equations of the submodels. it is important to consider the possible estimation of initial values in every new regime.8 in compact form. 6. It is important to remind the reader that fast and accurate MD is key stone for process monitoring. The amount of information required to discriminate between two models depends on the candidates and on data quality. need to be at least larger than variance of the measured variables. enough data information to assure discrimination with statistical meaning is to be guaranteed. Nonetheless. In the ideal case. The general structure is defined by the parameters and equations which remain unchanged in all submodels. The smaller the difference between models outputs. it is possible to quantify the measurements required to fulfill a specified threshold for model distinguishability 𝐷 𝐵 . the average of the square of the difference between the outputs of the model. Also some very interesting studies have been published for the case of hybrid systems [121]. false switching point detection may occur due to low identifiability.3 SUBMODEL DISTINGUISHABILITY Model distinguishability is strongly dependent on the modeler ability to create models which enhance the differences between dynamics of different regimes. and is known as piece-wise constant parameter modeling [120]. If the length of the interval selected for MD is too short. in order to confront cases with multiple RWPs. if the interval is too long the sensitivity of MXL to the data points indicating a regime change is decreased. In probability theory. the more precise need to be the measurements. to assure distinguishability. Such a scenario maximizes submodel identifiability.3. distinguishability quantifies the probability of selecting the model which best describes the observations.An approach to Mechanism Recognition implemented to consider the difference in sate of information between PCPs and RWPs. Once the certainty desired to consider the discrimination between two candidates is fixed. At this stage only problems where two candidates are set to discrimination are addressed. On the other hand. In other words. The best approach for model discrimination and its ability to detect the next regime with the required promptness needs to be investigated for each particular case. In addition.

This might affect process efficiency since the initial regime could be artificially kept active only to allow identifiability affecting optimal process conditions. the computation of the some confidence interval criterion is started. PCPs need to be defined and the boundaries of the RWPs have to be set. Because of the lack of data at the beginning of the process. that the initial regime offers enough information to identify the general structure of the submodels. 6. Nicolas Cruz Bournazou 𝑛 𝑦 𝑖=1 𝑦1 𝑢 𝑡 . Again. the effort of this step is reduced as experience and information are accumulated. it must be assured. Luckily. During the initial interval.3. Still. det 𝐹𝐼𝑀 ≥ 0 6. 𝜃2 𝑛𝑦 2 = 𝐷 𝐵 ≥ 𝜎𝑦 6. in most of the systems. 𝜃1 − 𝑦2 𝑢 𝑡 . 6. and 𝜃 the parameter set of model 1 and 2 respectively. Hence.6 Where 𝑦 is a vector with 𝑛𝑦 outputs of model 1 and 2 respectively. the FIM is tested for identifiability. In biological systems for example. 𝜃1 − 𝑦2 𝑢 𝑡 .4 INITIAL INTERVAL The initial interval is the process stage with less state of information. as well as linear and nonlinear constraints have been defined. the first regime of the process is known and cases where the initial conditions of a process are not well defined are scarce in process engineering. Mostly.5 MR INITIALIZATION The recognition process may only be started after the parameter values of the general structure have been estimated and boundaries. many processes show great reproducibility and the parameter values of previous runs can be applied to reduce the required length of the initial interval. initial concentrations and state variables as temperature and pressure can be defined at the beginning of the process.3. once the state of information is enough to assure a trouble-free inversion of the FIM. define the boundary conditions of the successive parameter values throughout the complete process. maximal reaction rates and uptake capacities determined in the initial interval.7 Secondly. Furthermore. 𝜃2 𝑇 ∗ 𝑦1 𝑢 𝑡 . 𝑢 𝑡 a vector of 𝑛𝑢 u time-dependent input variables. parameter values of previous runs have to be selected as initial guesses. the information gained in the initial interval is essential to assure a correct recognition of the process. First.An approach to Mechanism Recognition for model based analysis of Biological Systems M. The application of MR with uncertain initial conditions requires addition of backward propagation calculations to define the initial states and is not considered in this work. In addition offline measurements can be carried out before the process is started to increase the information at the beginning of the process. 59 .

6. The value obtained at this stage is used to select the time point from which a regime switching point can be detected.2. It is worth reminding that an important assumption is that the general structure of the model is globally correct and structure parameters do not need to be updated. the program is required to detect an irregular behavior of the system in comparison to some model predictions. on the one side state estimation techniques (the parity space approach. Although not necessarily representing a fault on the system. 𝐴𝑐𝑟𝑖𝑡 represents the average variance of the confidence interval considering all unknown parameters. where changes on process conditions induce a change in the regime of the system.An approach to Mechanism Recognition For both case studies presented in chapters 7 and 8. is defined as switching point. the A criterion (𝐴𝑐𝑟𝑖𝑡 ) was selected to quantify the state of information (section 4. The accurate detection of the switching points is essential for two reasons: fast and accurate detection of regime change .obtain the correct initial conditions of the non measurable variables Once adequate submodels have been developed and the state of information has been analyzed.avoid undesired or dangerous process conditions accuracy of initial values for the next interval . Much care should be taken not to initiate the switching point detection before the state of information has been tested and accepted.6).8 where 𝐷𝐵 represents the criterion boundary obtained from the distinguishability test. from a mathematical and physical point of view.g. 𝐴𝑐𝑟𝑖𝑡 ≤ 𝐷𝐵 6. the detection of the switching point may be initiated (section 6. detection of the switching points is no different from fault diagnosis based on first principle models.3. The boundary for 𝐴𝑐𝑟𝑖𝑡 to assure model distinguishability is calculated during the tests for model distinguishability (section 4.3).1). Once the criterion boundary has been reached (e. Typical methods include. In short words. and 60 . the number of measurements is enough to make a distinction between the model outputs). Methods for model-based residual generation. observer based schemes.6 DETECTION OF SWITCHING POINTS The position in time. This assumption is essential to assure consistent evaluation of piecewise constant parameter identifiability and model distinguishability.3. offering so called analytical redundancy have been widely discussed in literature [122].

In this work we concentrate on the application of parameter identification methods due to the common theoretical background studied in design of experiments. Following conditions must be considered during submodel building since they are essential to assure detectability and distinguishability of the switching point during the process [123]: knowledge of the normal behavior of the system distinguishability of the changing behavior availability of at least one observation reflecting regime change satisfactory state of information The evaluation of the validity of the present regime given by the submodel is divided in two main steps: residual generation decision of the switching point (time and type of the new interval) To increase the probability of correct detection of the switching point.9 where 𝜃𝑛 and 𝜃𝑒 represent the nominal parameter set and the estimated parameter respectively. the resulting equation is: 61 . the method for robust parameter estimation [123] is proposed. a residual calculation is carried out to select one of the following hypotheses: 𝐻0 : 𝜃𝑛 = 𝜃𝑒 𝐻1 : 𝜃𝑛 ≠ 𝜃𝑒  6. In order to decide whether the process is running in the present regime or a change is taking place. We can also take advantage of the previous calculation of FIM (section 4.10 where 𝑟𝑒𝑠 is the residual. hypothesis zero assumes there is no regime change whereas hypothesis one assumes the switching point has been reached.2) and substitute 𝐶𝑝 by its approximation. Nicolas Cruz Bournazou fault detection filter approach) and parameter estimation techniques on the other (parameter identification).An approach to Mechanism Recognition for model based analysis of Biological Systems M. Still MR is not limited to parameter estimation methods and the application of other faster methods like direct evaluation of the objective function offer a fast and fairly accurate result. To enable the evaluation of the hypothesis considering the uncertainties of the observations a residual between 𝜃𝑛 and 𝜃𝑒 is calculated taking the variance-covariance matrix of the expected parameter set into account: 𝑟𝑒𝑠 = 𝜃𝑛 − 𝜃𝑒 𝑇 𝐶𝑝 −1 𝜃𝑛 − 𝜃𝑒 6. and 𝐶𝑝 is the variance-covariance matrix of the parameter set.

offline measurements to redefine the state of the process should be carried out whenever possible. To find tight boundaries the switching interval can be simulated with both models to create an envelope by selecting the maximal and minimal values for the upper and lower bounds respectively.6).3.7 INITIAL CONDITIONS OF THE INTERVAL K+1 It is essential to take into consideration that. 6. The variation of the threshold is substituted by the variation of FIM. This is of relevance for the calculation of the successive regime (k+1). due to the fact that the initial conditions of the regime (k+1) are determined by the endpoint outputs of the previous regime (k). the initial values and parameters have been estimated. regime changes do not happen instantaneously. Although variable threshold approaches have shown to be more efficient [124]. a fixed threshold 𝐷𝐵 is selected to accept or reject hypothesis zero. in real processes. the initial values of the next interval need to be estimated. the algorithm returns to the search for the next switching point (return to 6. It is clear that this represents an addition of parameters to be estimated reducing the identifiability of the problem drastically and tight boundaries of the initial values are required minimize the impact of these new variables in the complexity of the optimization problem.3. thanks to the implementation of FIM in the equation.11 Finally.3. which gives a more accurate description of parameter covariance in the sense that it also considers information from previous intervals PCPs.8 DETECTION OF THE NEXT SWITCHING POINT Once the best candidate has been selected. Although in many cases the difference in dynamics allows to consider changes in process conditions as instantaneous. and the arc in the switching interval determined. This procedure is repeated until the end of the process is reached. In the cases where no observations are available.An approach to Mechanism Recognition 𝑟𝑒𝑠 = 𝜃𝑛 − 𝜃𝑒 𝑇 𝐹𝐼𝑀 𝜃𝑛 − 𝜃𝑒 6. 6. 62 . To increase initial condition accuracy and reduce error bias. the transition phase between two regimes must be considered. the residual also reflect the actual state of information with respect to the parameters.

1) and cannot be carried out automatically since there is no general solution to proper submodel building. the model describing the first regime is tested for identifiability to determine the minimal length required to perform reliable parameter estimations. Once the process has started. 63 . The flow diagram proposed has been built for the special case of batch processes.3. but it can still be applied to different cases with slight modifications. with the general structure well defined and the boundaries set. The diagram is divided in three phases: Submodel Building Initial Interval Regime Recognition The first phase requires the collaborated work of process and model experts to create an adequate set of submodels (section 6.4.3. the detection of the switching point is initiated.9 FLOW DIAGRAM Following.6) to select the time point where the process changes from one regime to another.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Nicolas Cruz Bournazou 6. In this stage the models are tested for identifiability considering the state of information obtained from the process and submodel distinguishability.3. This is also the most critical phase of MR since the nominal values of the parameter set and the general structure are defined. The parameter set of the general structure is estimated and the boundaries of the RWPs are set. The algorithm applies one of the switching point detection criteria (section 6. The third phase is recursively repeated until the end of the process is reached. the steps of the recognition strategy are presented accompanied by a flow diagram Figure 6. Finally.

4: Flow diagram of MR algorithm 64 .An approach to Mechanism Recognition Flow diagram for Mechanism Recognition II Initial Interval I Submodel Building FIM calculation No Interval increment Complex model Data set Identifiability reached ? Yes General structure Submodel 1 Submodel n Data analysis Parameter estimation Switching point detection Jacobi matrixes Variancecovariance matrix III Regime Recognition k =k+1 General structure failed Identifi ability test success Regime type failed Distinguishability test Parameter estimation General structure FIM No t = tend? success Yes print Figure 6.

carbonaceous organic matter of wastewater (readily biodegradable as well as particulate substrate) provides an energy source for bacterial growth of a mixed population of microorganisms in an aquatic aerobic/anoxic environment. referred to in this contribution as extended ASM3. includes the quantification of energy storage in order to describe substrate and oxygen uptake with higher accuracy. microorganisms may obtain energy by oxidizing ammonia nitrogen to nitrate nitrogen in the process known as nitrification. where nitrification and denitrification are considered as two-step processes taking into account nitrite as an intermediate. The microbes consume carbon and oxidized end-products that include carbon dioxide and water. Finally. Moreover. the computation of the extended ASM3 is significantly expensive. ASM2 [128] is applied to simulate processes that include biological phosphorus removal [78] and the latest version.An approach to Mechanism Recognition for model based analysis of Biological Systems M.1 ACTIVATED SLUDGE The most applied method for biological treatment of waste water is the Activated Sludge Process (ASP) [125].1. In addition. 7 process equations were included. These drawbacks represent the main obstacle for efficient optimization and 65 . resulting in a stoichiometric matrix with 15 state variables and 20 reaction equations. ASM1 is the most widely used in practice [127]. ASM3 [129]. this process constituting the second part of the so called nitrification-denitrification process. In addition to a low parameter identifiability. Nicolas Cruz Bournazou 7 MECHANISM RECOGNITION APPLIED ON SEQUENCING BATCH REACTORS 7. In ASP. has recently been presented [130]. a newer version of ASM3. The family of Activated Sludge Models (ASM) represents the state-of-the-art model framework for ASP simulation [126]. In order to extend the ASM3. some microorganisms are able to reduce nitrite and nitrate in the anoxic phase as well.1 INTRODUCTION 7.

these variations can be buffered to a certain level by increasing the size of the tanks.1: SBR cycle [136] 66 . and landfill leachates among others [134]. and process control stability. One or more stirred batch reactors with an aeration system are the basis for this kind of process. the settling time. lower energy consumption and operation costs. higher load capacity. In case of large WWTPs. dairy wastewaters. The most relevant are.2 SEQUENCING BATCH REACTOR Continuous Waste Water Treatment plants (WWTP) offer important advantages in comparison to batch processes. tannery. 2) Fill. because of its high flexibility and operation range. whereas a continuous process operates with variations in space [135]. 1) Idle. A complete cycle of the sequencing batch process consists of five steps. SBR. Furthermore. In an SBR. which are not to be considered in SBR process [131. 132]. the retention time. On the other hand. offer an adept solution for such cases [134]. hypersaline. small and medium size WWTPs are strongly affected by changing concentrations of contaminants in the wastewater [133]. the extended ASM3 describes many states. or by introducing a flow-stabilization tank upstream. an SBR is able to treat different kinds of effluents such as municipal. brewery. Wastewater naturally presents stochastic variations in both time and space. The SBR can also be considered to be a process which operates with variation in time. and other conditions can be fitted to a changing quality of load as well as effluent requirements. domestic. continuous processes confront a major limitation when dealing with an intrinsic characteristic of wastewater treatment processes.4. 1 2 5 idle fill 3 settle draw 4 react Figure 7. Nevertheless. 4) Settle and 5) Draw Figure 6. the duration of the aeration and anoxic phases. 3) React.1. 7. However.Mechanism Recognition applied on Sequencing Batch Reactors model-based control.

Katsogiannis et al [146] showed that a frequent enough change between aerobic and anoxic conditions suppresses nitrate formation. Turk and Mavinic [145] proposed the Nitrate Bypass Nitrification-Denitrification (NBND) process. which is then converted into nitrogen under anoxic conditions before the second oxidation producing nitrate (nitratation) can take place.1. an SBR also offers a process with important advantages over continuous processes. Nitrification-denitrification process described as a two -step reaction. namely nitrification followed by denitrification (Figure 7.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Besides many efforts to monitor nutrient removal [78. reducing operation costs. Chemical Oxygen Demans (COD). increase the process efficiency. Advances in process measurement. + 𝑁𝐻4 − 𝑁𝑂2 − 𝑁𝑂2 − 𝑁𝑂3 − 𝑁𝑂3 𝑁2 Figure 7.2). 7. In the first stage. Most of the nitrogen contained in wastewater is converted into ammonia. Nicolas Cruz Bournazou The SBR has gained great popularity in recent years. Nowadays. When properly designed and operated. direct online monitoring of Biological Oxygen Demand (BOD). Ammonia is converted to nitrite in the presence of oxygen (nitritation). as well as automation and control. whereas in a second stage. Ammonia is then converted into molecular nitrogen by a two-step biological processes. Nitrosomonas and other ammonia oxidizers convert ammonia and ammonium to nitrite. which can be achieved by inhibiting the production of nitrate and proposed various methods for bringing about this effect. SBR technology has become important in particular for small and mediumsized WWT Plants. The NBND has the following advantages over conventional nitrification-denitrification [145]: 40% reduction of COD demand during denitrification 63% higher rate of denitrification 300% lower biomass yield during anaerobic growth no apparent nitrite toxicity effects for the microorganisms in the reactor 67 . Nitrites and Nitrates as well as ammonia is difficult inaccurate and costly. while fulfilling the strict environmental regulations [137]. not only because of its efficiency and economical aspects [138-140]. Nitrobacter and other nitrite oxidizers finish the conversion of nitrite to nitrate.2. Indirect methods are required to determine these concentrations for control purposes. 141-144] in SBR processes.3 NITRATE BYPASS GENERATION In the ASP. [143]. nitrogen is removed from wastewater by the nitrification/denitrification process. but also because of its small footprint [135].

In this section. The model reduction is based on the principle that an batch cycle should stop once all concentrations comply with the regulations. Furthermore. these concentrations should be reached in the shortest time possible. Except for the recycle process. Another important assumption is that the bacteria never exhaust their stored energy. The Activated Sludge Model No. the bacteria are always in a medium. In addition. The division of the nitrification-denitrification reaction in a two-step reaction is essential when trying to describe the bypassing nitrate generation process in the ASP [146]. Consequently. according to the number of the ordinary differential equations which need to be solved. Moreover. 3 (ASM3) extended for two-step nitrification and denitrification has proven to be a very accurate model to describe ASP [149]. This study is based on the analysis of the process conditions and the available empirical knowledge. which is rich in substrate. Therefore. among others [147.1. the submodels are created considering process conditions and the characteristics of the complex model to enable a fast 68 .2 SUBMODEL BUILDING Following the systematic approach to MR. The first submodel proposed is named 5state. The implementation of new approaches to enable an accurate online monitoring of key species like organic matter is of great relevance in modern water treatment technology. the extended ASM3 enables the description of the substrate consumption and the oxygen uptake with a higher precision than the older versions of the ASM family because of the addition of energy storage effects. Variables measured in standard WWTPs are pH. environmental regulations demand better process and output quality control at low treatment cost and energy efficient process conditions. redox capacity and titrimetry. DO. frequent recalibration and important know how. 7. modern measure systems still require laborious maintenance. Despite the important advances achieved in recent years. diverse approaches to model reduction (chapter 3) are applied in order to develop submodels of the ASM3 while maintaining their characteristic dynamics in each regime. Model based control and software sensors offer interesting possibilities for the maximization of process efficiency. the stored energy value should be permanently high during the process and never limit bacterial growth. respirometry.4 MONITORING OF WASTEWATER PROCESSES Online monitoring of WWTPs is a challenging task. lower output concentrations than required are an indicator of a suboptimal operation and its accurate detection increases process efficiency.Mechanism Recognition applied on Sequencing Batch Reactors 7. 148]. This model extension facilitates the calculation of the NO2 concentration as an independent variable.

the storage of substrate and its effects on substrate and oxygen concentration cannot be ignored. 7.nitrite oxidizers. 5.3. 4. namely: 1.. This could appear to be an inconsistent assumption. In other words. For this reason.4. 69 . substrate consumption rate and oxygen consumption rates...3 A PROPOSED 9STATE MODEL In the first reduction step.heterotrophic bacteria... Nicolas Cruz Bournazou description of the different regimes of the process Figure 6. Neglecting this equation would impede a proper process description and would result in an incorrect model reduction. the model proposed is named 9state.. 7. By this means. 3. it is of interest to stop the process once the concentrations of the medium fulfill environmental regulations. to energy storage are linear. Consequently.. the relation between both.nitrite. 6. that both substrate uptake and oxygen uptake increments caused by the storage are now included in the original substrate and oxygen differential equations. but describes substrate and oxygen uptake as well as energy storage as precisely as the extended ASM3. Regime with readily biodegradable substrate in the medium 2. which is accurate as long as the substrate concentration is above zero.stored substrate.nitrate and 9. The recognition step should identify two distinct regimes: 1.ammonia oxidizers. the 9state model responds to the substrate limitation similarly to ASM1.carbonaceous substrate. the proposed 9state model includes some adaptations to the substrate and oxygen uptake equations.. For this reason. 2. Thus there are nine differential equations which describe the basic variables (concentrations). though it is almost certain that the process continues after substrate elimination in order to achieve further ammonia degradation. The new set of equations is presented in such a way. Nevertheless.dissolved oxygen. 7.1 STORAGE The implementation of energy storage represents the principal improvement of ASM3 in comparison to older versions.ammonia. Regime where organic matter has been depleted In SBR processes. 8. the submodel is built to describe the process regime under high substrate concentrations aiming at detection of substrate depletion.An approach to Mechanism Recognition for model based analysis of Biological Systems M. according to the number of state variables contained in the equation system. An accurate detection of depletion of organic matter offers important improvements in process efficiency. previous model versions (ASM1 and ASM2) fit the data although they lack a storage variable.

If these assumptions are valid. Nitratation r18: Aerobic End. r19: Anoxic Endog.1: Reaction rates of the extended ASM3 Heterotrophic Organisms: r1: Hydrolysis r2: Aerobic Storage r3: Anoxic Storage r4: Anoxic Storage of carbonate substrate. Therefore. The process rate equations are shown in Table 7.1and have been numbered in the same order as previously presented in [130]: Table 7. the environmental regulations should be fulfilled as soon as substrate is consumed. Resp. of Heterotrophic bacteria r9: Anoxic Endog. of particulate storage r12: Anoxic Resp. Resp. NO3-NO2 r13: Anoxic Resp.2 REDUCTION OF THE EXTENDED ASM3 MODEL TO A 9STATE MODEL In this work. this second-order differential equation can be accurately approximated with a first order differential equation. as long as the concentration of the stored energy is high. The assumption is based on the fact that in a process with optimal aeration strategy.Mechanism Recognition applied on Sequencing Batch Reactors ASM3 considers that bacteria store some of the substrate they consume under high substrate concentrations for its later use under substrate limiting conditions. the relation between substrate and biomass growth is described by a second order differential equation. bacteria never exhaust their stored energy for two reasons: In order to minimize process time and costs. it is considered that storage of substrate does not limit bacterial growth. the 15 ordinary differential equations of the extended ASM3 include only the process rate variables without their explicit equation. NO2-N2 Ammonium Oxidizing Bacteria (AOB): r14: Aerobic Growth. the value of the stored energy should be high at any moment during the process and never limit bacterial growth. of particulate storage. In the case of an SBR process. However. of particulate storage. the relation between the substrate used for storage and the substrate used for growth is valid as well. Resp. r16: Anoxic End. bacteria are always in a medium rich in substrate.3. This storage is not only responsible for lower biomass growth under equal substrate consumption. In the extended ASM3. Resp 70 . NO3-NO2 r10: Anoxic End. Resp. 7. but also for the bacterial growth when no substrate is present in the medium. Resp. Nitritation r15: Aerobic End. NO2-N2 r11: Aerobic Resp. Nitrite Oxidizing Bacteria (NOB): r17: Aerobic Growth. NO2–N2 r5: Aerobic Growth of Heterotrophic bacteria r6: Anoxic Growth NO3-NO2 r7: Anoxic Growth NO2-N2 r8: Aerobic Endog. Resp. Excepting the idle phase.

1 7.7.An approach to Mechanism Recognition for model based analysis of Biological Systems M.5 𝑑𝑆𝑁𝐻4 𝑖𝑁𝑆𝑆 1 =− − + 𝑖𝑁𝐵 𝑟𝑎𝑎𝑒 − + 𝑖𝑁𝐵 𝑟𝑎𝑎𝑁𝑠 − 𝑖𝑁𝐵 𝑟𝑎𝑎𝑁𝑏 𝑑𝑡 𝑌𝐻𝑎𝑒𝑟 𝑌 𝐴1 𝑖𝑁𝑆𝑆 𝑖𝑁𝑆𝑆 − − + 𝑖𝑁𝐵 𝑟𝑎𝑁𝑂 3 − − + 𝑖𝑁𝐵 𝑟𝑎𝑁𝑂 2 𝑌𝐻𝑎𝑛𝑜𝑥 𝑌𝐻𝑎𝑛𝑜𝑥 𝑑𝑆𝑁𝑂2 1 1 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 = 𝑟𝑎𝑎𝑁𝑠 − 𝑟𝑎𝑎𝑁𝑏 + (𝑟 − 𝑟𝑎𝑁𝑂 2 ) 𝑑𝑡 𝑌 1 𝑌 2 1. All the constants K.14.8 7.2.14𝑌𝐻𝑎𝑛𝑜𝑥 𝑎𝑁𝑂 3 𝐴 𝑑𝑋𝑆𝑡𝑜 1 1 = − ∗ 𝑟𝑎𝑎𝑒 − ∗ 𝑟𝑎𝑁𝑂3 + 𝑟𝑎𝑁𝑂 2 𝑑𝑡 𝑌𝐻𝑎𝑒𝑟 𝑌𝐻𝑎𝑛𝑜𝑥 ∗ (−𝑆𝑡𝑆 ) 7.x have the same value as published in the extended version of ASM3 [130].3 7.3.2 7. dSS 1 1 = − ∗ raae − ∗ raNO 3 + raNO 2 dt YHaer YHanox 𝑑𝑋𝐻 = 𝑟𝑎𝑎𝑒 + 𝑟𝑎𝑁𝑂 3 + 𝑟𝑎𝑁𝑂 2 𝑑𝑡 𝑑𝑋𝑁𝑠 = 𝑟𝑎𝑎𝑁𝑠 𝑑𝑡 𝑑𝑋𝑁𝑏 = 𝑟𝑎𝑎𝑁𝑏 𝑑𝑡 𝑑𝑆𝑂 = 𝑑𝑡 ∗ 1 + St S 7.9.7.4 7.9 SS XH XNs XNb SO Readily biodegradable substrate Biomass of heterotrophic bacteria Biomass of Nitrosomones Biomass of Nitrobacter Oxygen concentration in the medium 71 .1 .7 7. Nicolas Cruz Bournazou 7. Their corresponding rate equations are described in 7.10 .14𝑌𝐻𝑎𝑛𝑜𝑥 𝑎𝑁𝑂 3 𝐴 𝐴 𝑑𝑆𝑁𝑂3 1 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 = 𝑟𝑎𝑎𝑁𝑏 − 𝑟 𝑑𝑡 𝑌 3 1.3 MATHEMATICAL REPRESENTATION OF THE 9STATE MODEL ORDINARY DIFFERENTIAL EQUATIONS The 9 ordinary differential equations are shown in equations 7.6 7. The values used for the saturation constants are shown in Table 7.

14 72 .13 Heterotrophic growth on nitrite 𝑟𝑎𝑁𝑂 2 = 𝜇𝐻2 ∗ 7.11 Heterotrophic growth on nitrate 𝑟𝑎𝑁𝑂 3 = 𝜇𝐻1 ∗ 𝑆𝑆 𝑆𝑁𝑂3 𝐾𝑂21 𝑆𝑁𝐻4 ∗ ∗ ∗ ∗ 𝑋𝐻 𝑆𝑆 + 𝐾𝑆 𝑆𝑁𝑂3 + 𝐾𝑁𝑂3 𝐾𝑂21 + 𝑆𝑂 𝑆𝑁𝐻4 + 𝐾𝑁𝐻 𝑆𝑆 𝑆𝑁𝑂2 𝐾𝑂22 𝑆𝑁𝐻4 ∗ ∗ ∗ ∗ 𝑋𝐻 𝑆𝑆 + 𝐾𝑆 𝑆𝑁𝑂2 + 𝐾𝑁𝑂2 𝐾𝑂22 + 𝑆𝑂 𝑆𝑁𝐻4 + 𝐾𝑁𝐻 7.10 7.Mechanism Recognition applied on Sequencing Batch Reactors SNH 4 SNO 2 SNO 3 XSto YHaer YHanox YA1 YA2 YA3 𝜇𝐻 𝜇𝐴1 𝜇𝐴2 𝜇𝐻1 𝜇𝐻2 𝑆𝑡𝑆 𝑆𝑡𝑂 𝐾 Ammonia concentration Nitrite concentration Nitrate concentration Energy Storage Yield coefficient of SS to XH in aerobic conditions Yield coefficient of SS to XH in anoxic conditions Yield coefficient nitrite to nitrosomones in aerobic conditions Yield coefficient nitrite to nitrobacter in aerobic conditions Yield coefficient nitrate to nitrobacter in aerobic conditions Maximal growth of Heterotrophous Maximal growth of nitrosomones Maximal growth of nitrobacter Maximal growth of heterotrophic bacteria on nitrire Maximal growth of heterotrophic bacteria on nitrate Substrate to storage constant Oxygen to storage Limitation constant REACTION RATES Heterotrophic growth on aerobic conditions 𝑟𝑎𝑎𝑒 = 𝜇𝐻 ∗ Growth of nitrosomones 𝑟𝑎𝑎𝑁𝑠 = 𝜇𝐴1 ∗ Growth of nitrobacter 𝑟𝑎𝑎𝑁𝑏 = 𝜇𝐴2 ∗ 𝑆𝑁𝑂2 𝑆𝑂 𝑆𝑁𝐻4 ∗ ∗ ∗ 𝑋𝑁𝑏 𝑆𝑁𝑂2 + 𝐾𝑁𝑂21 𝑆𝑂 + 𝐾𝑂 𝑆𝑁𝐻4 + 𝐾𝑁𝐻 7.12 𝑆𝑆 𝑆𝑂 𝑆𝑁𝐻4 ∗ ∗ ∗ 𝑋𝐻 𝑆𝑆 + 𝐾𝑆 𝑆𝑂 + 𝐾𝑂1 𝑆𝑁𝐻4 + 𝐾𝑁𝐻 𝑆𝑂 𝑆𝑁𝐻4 ∗ ∗ 𝑋𝑁𝑠 𝑆𝑂 + 𝐾𝑂 𝑆𝑁𝐻4 + 𝐾𝑁𝐻 7.

3468.2. 1.08. 0. 0. 0. 0. 0.0985.5.3.0511.4 STOICHIOMETRIC MATRIX The stoichiometric matrix of the 9state model is presented in Table 7. 0.1.3. [mgCOD/l] [mgCOD/l] [mgCOD/l] [mgCOD/l] [mgN/l] [mgO2/l] [mgN/l] [mgO2/l] [mgO2/l] [mgN/l] [mgO2/l] [mgN/l] [mgO2/l] [] [] ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 ASM3 fitted fitted 7. the analysis of the stoichiometric matrix gives us the information about the number of reaction invariants hidden in the 9state model. 0.1327. 0.3: Stoichiometric matrix of the 9state model 𝑆𝑆 𝑟𝑎𝑎𝑒 − 1 + St S 𝑌𝐻𝑎𝑒𝑟 0 𝑋𝐻 1 𝑋𝑁𝑆 0 𝑋𝑁𝐵 0 − 𝑆𝑂 1 − 𝑌𝐻𝑎𝑒𝑟 𝑌𝐻𝑎𝑒𝑟 3. 0. This matrix represents the (𝑙 + 𝑚 − 𝑛) matrix of equations. 0. 10.14𝑌𝐻𝑎𝑛𝑜𝑥 − 𝑟𝑎𝑁𝑂 3 − 1 0 0 0 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 1. 0.1.1302. 0.1.2.6021. [mgO2/l] [d-1] [gN/gCOD] [d-1] [d-1] [d-1] [gCOD/gCOD] [gCOD/gN] [gCOD/gN] [gCOD/gN] [gN/gCOD] [gCOD/gCOD] [d-1] [d-1] [mgCOD/l] process process fitted fitted fitted fitted fitted fitted fitted fitted ASM3 fitted fitted fitted ASM3 K_NH2 K_S K_S1 K_S2 K_NHH K_O1 K_NH K_O K_NO21 K_NO3 K_O21 K_NO2 K_O22 stS stO = = = = = = = = = = = = = = = 0. 0.0331. 0. 0. 0. 0.0632. 0. Nicolas Cruz Bournazou Table 7. 0. 0.2.01.1.14𝑌𝐻𝑎𝑛𝑜𝑥 73 .2: 9state model constants and its values as shown in the Matlab code SOStar K_La i_NB mou_H mou_A1 mou_A2 Y_Haer Y_A1 Y_A2 Y_A3 i_NSS Y_Hanox mou_H1 mou_H2 K_NH1 = = = = = = = = = = = = = = = 7. 0.14𝑌𝐻𝑎𝑛𝑜𝑥 𝑟𝑎𝑎𝑁𝑏 0 1 + St S 𝑌𝐻𝑎𝑛𝑜𝑥 1 + St S 𝑌𝐻𝑎𝑛𝑜𝑥 0 0 1 1− 0 St S 𝑌𝐻𝑎𝑛𝑜𝑥 St S 𝑌𝐻𝑎𝑛𝑜𝑥 𝑟𝑎𝑁𝑂 2 − 1 0 0 0 − 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 1.086.34 𝑌 1 𝐴 1.6552.01. Table 7. 0.8.7. 0.14 𝑌 𝐴2 𝑆𝑁𝐻4 𝑖𝑁𝑆𝑆 − 𝑖𝑁𝐵 𝑌𝐻𝑎𝑒𝑟 − 1 + 𝑖𝑁𝐵 𝑌𝐴1 −𝑖𝑁𝐵 𝑖𝑁𝑆𝑆 − 𝑖𝑁𝐵 𝑌𝐻𝑎𝑛𝑜𝑥 𝑖𝑁𝑆𝑆 − 𝑖𝑁𝐵 𝑌𝐻𝑎𝑛𝑜𝑥 𝑆𝑁𝑂2 0 1 𝑌 1 𝐴 − 1 𝑌 𝐴2 𝑆𝑁𝑂3 0 𝑋𝑆𝑡𝑜 St S 𝑌𝐻𝑎𝑒𝑟 0 𝑟𝑎𝑎𝑁𝑠 0 1 0 1− 0 1 𝑌 𝐴3 0 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 1.25.5. 0.An approach to Mechanism Recognition for model based analysis of Biological Systems M. In other words.0362. 1000. 0.05.

These equations are of the form of eq.3. the bacteria can store energy. both growth curves match. Therefore. For example 7.14 𝑌𝐻𝑎𝑛𝑜𝑥 − 𝑌𝐻𝑎𝑒𝑟 1 + 𝑆𝑡𝑆 𝑌𝐻𝑎𝑒𝑟 𝑌 1 𝐴 2 ∗ 𝑌 2 − 𝑌 3 𝐴 𝐴 − 𝑌 2 ∗ 𝑌 3 ∗ 𝑋𝑁𝐵 𝐴 𝐴 𝑆𝑆 𝑆𝑡𝑂 + 1 + 𝑆𝑡𝑆 𝑆𝑡𝑆 7. but cannot use it when there is no more substrate available. if the energy stored by the bacteria is low. Finally.17.16 7. the growth of heterotrophic biomass can be mathematically described as a second order differential equation. to find 𝑙 + 𝑚 − 𝑛 = 3 linearly independent reaction invariants. 𝑆𝑁𝐻4 + 𝑖𝑁𝑆𝑆 ∗ 𝑆𝑁𝑂2 + 2 ∗ 𝑆𝑁𝑂3 + 𝑆𝑆 𝑋𝑁𝑆 + 𝑖𝑁𝐵 ∗ 𝑋𝐻 + 𝑋𝑁𝑆 + 𝑋𝑁𝐵 + 1 + 𝑆𝑡𝑆 𝑌 𝐴1 7.7.7.5 LIMITATIONS OF THE REDUCED MODELS It should be noted that the proposed 9state model is not valid for the whole range of conditions as the extended ASM3. it is possible in this case. the ammonia concentration does not limit the energy storage.1 . with only one of them depending on the process input (oxygen supply).7. Moreover. which are not affected by the oxygen input. 7. The reaction invariant theory can then be applied to the 8 differential equations. This results in a consumption of substrate even under ammonia limitation. because of the new storage equation.15 . 𝑙 + 𝑚 = 8.4 and 7. This time delay is not predicted by the 9state model. In the extended ASM3. Some limitations are to put up with in order to reduce the model and speed up the simulation in the region of interest. because of the coupled equations. the 9tate model predicts substrate consumption only as long as ammonia is present in the medium.6 . 2.Mechanism Recognition applied on Sequencing Batch Reactors 7.e. i.9. a time delay can be seen in the growth curve.6 with 𝑛 = 5. For this reason. Taken into consideration that the storage has a value at least larger than 100 gCOD/m3.17 74 .4 A PROPOSED 6STATE MODEL The 9state model involves 9 differential equations.15 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 𝑌𝐻𝑎𝑒𝑟 𝑆𝑆 𝑋𝐻 𝑋𝑁𝑆 ∗ ∗ + − 1. This feature converts the 9state model in a perfect indicator of substrate depletion in the reactor. Once again. to 7.

it can only be applied for batch processes where the change of biomass can be neglected. 7. The simulation represents a batch tank ideally mixed. This assumption is based on the fact that the sludge is obtained from previous cycles and bacteria have already stored energy. Therefore.6.4 from our 9state model and consider a constant biomass concentration throughout each cycle of the SBR so as to obtain a model with 6 differential equations.20 where CNH 4 .17 for the initial conditions. As a result we finally obtain a model with only 5 state equations.20.3-Figure 7.2 . 𝑆𝑁𝐻4 = 𝐶𝑁𝐻4 − 𝑖𝑁𝑆𝑆 ∗ 𝑆𝑁𝑂2 = 𝐶𝑁𝑂2 − 2 ∗ 𝑆𝑁𝑂3 − 𝑆𝑆 𝑋𝑁𝑆 + 𝑖𝑁𝐵 ∗ 𝑋𝐻 + 𝑋𝑁𝑆 + 𝑋𝑁𝐵 + 1 + 𝑆𝑡𝑆 𝑌 1 𝐴 7.6.9 for three algebraic ones 7. However. CNO 2 and CXSto are the respective constants obtained after solving equations 7. The initial value of the storage is set to 400 gCOD/m3. This 5state submodel is almost as accurate as the 6state model.7 and 7. Based on this observation a further reduction of the mode is possible for the special case of a batch process. We can eliminate these three differential equations 7. where the aeration can be turned on and off in order to induce either aerobic or anoxic conditions. The most representative results obtained after the comparison of ASM3 with the 5state model are presented in Figure 7. 7. the 6state model mimics the 9state model perfectly well.15 .7. we find three equations which are linearly dependent. we can see that the overall change in biomass does not exceed 10%. Nicolas Cruz Bournazou The above quantities remain constant throughout the entire batch. 7. it is possible to substitute three differential equations 7. Again it can be shown that this system has one reaction invariant which is the same as indicated for the case of the 9state 7.7.18 1 − 𝑌𝐻𝑎𝑛𝑜𝑥 𝑌𝐻𝑎𝑒𝑟 𝑆𝑆 𝑋𝐻 ∗ ∗ + 1. Based on the method of reaction invariants.18 . 75 .5 A PROPOSED 5STATE MODEL Taking a closer look at the growth of the biological matter during one SBR cycle.6 RESULTS The submodel was set to various conditions so as to confirm their stability and accuracy.20.An approach to Mechanism Recognition for model based analysis of Biological Systems M.19 7.14 𝑌𝐻𝑎𝑛𝑜𝑥 − 𝑌𝐻𝑎𝑒𝑟 1 + 𝑆𝑡𝑆 𝑌𝐻𝑎𝑒𝑟 𝑋𝑁𝑆 2 ∗ 𝑌 2 − 𝑌 3 𝐴 𝐴 + − 𝑌 1 𝑌 2 ∗ 𝑌 3 ∗ 𝑋𝑁𝐵 𝐴 𝐴 𝐴 𝑋𝑆𝑡𝑜 = 𝐶𝑋𝑆𝑡𝑜 − 𝑆𝑆 ∗ 𝑆𝑡𝑆 1 + 𝑆𝑡𝑆 7.7. Therefore.

2. changes in batch processes can be neglected. Even though growth rate is a crucial parameter of the process biomass concentration. This proves that the substitution of the differential equation for energy storage by an algebraic equation does not affect the dynamics of the model. Substrate concentration SS and stored energy Sto against time. NO2.3. 76 . Biomass against time.4. a constantly changing process is obtained. ODE15s: An integrator based on the numerical differentiation formulas (NDFs) [150]. As seen in Figure 7. The main purpose is to set both models to drastic changes and various limitations. The process conditions prove that the 5state model describes accurately the limitations of dissolved O2.Mechanism Recognition applied on Sequencing Batch Reactors The system of DAE is solved with two different integrators: sDACL: An in house tool developed to solve DAE systems with OCFE able to also calculate dynamic sensitivities [46].3 shows the results for substrate and storage. considering constant biomass values throughout the process does not affect the results of the other state variables. Changes in the biomass are very small (less than 10%). Figure 7.and NO3Figure 7.4. All reduced versions describe perfectly the behavior of storage even doe it is calculated by an algebraic equation.Sto [gCODm -3] 800 700 5state 600 500 400 300 200 0 ASM3 Figure 7. a) SS 900 1000 530 520 510 500 0 104 Sto 102 100 2 4 t [h] 6 8 10 98 0 2 4 t [h] 6 8 10 2 4 b) XNS XNB 6 8 10 XH [gCODm-3] 5state ASM3 SS. The aeration is turned on and off intermittently to produce a strongly dynamic process. The initial value of storage was set to 400 gCOD/m3 as explained in section 4. which makes it very difficult to be described identically by two models with different characteristics. As a result.

the Jacobian matrixes have to be solved together with the equation system.1 16.5. For the calculation of the dynamic sensitivities. ODE15s CPU time (sec) Number of Aer-Anox phases 1 2 3 4 5 ASM3 1. 7. The concentration of NO2 and NO3 are kept under 20 mgN/L. Table 7.769 2.172 0.An approach to Mechanism Recognition for model based analysis of Biological Systems M.4 and Table 7. NOX concentration against time.172 0. Comparison of the computation time. 77 .2 The reduced models are up to one order of magnitude faster.1 SIMULATIONS RESULTS The most representative results of the comparison of both models simulated in Matlab® are presented in Table 7. Table 7.602 2. The control variable is the aeration of the tank. Nicolas Cruz Bournazou a) a) SNO-2 [mgNL-1] SO [mgO2L-1] 2 4 b) 6 8 5state ASM3 + [mgNL-1] 4 2 0 0 8 6 4 2 0 0 2 4 b) 6 8 5state ASM3 10 10 SNO-3 [mgNL-1] 8 6 4 2 0 0 2 4 6 50 40 30 20 8 10 S 0 NH 4 2 4 6 8 10 t [h] t [h] Figure 7.5 shows the difference of calculation time for the three models.172 ASM3/5State 11.5 shows how precise the reduced model describes the curves of nitrite and nitrate even in such drastic conditions. Finally.4 15.6 15. the results for oxygen and ammonia calculations can be seen in Figure 7.157 0. a) Oxygen concentration in the medium against time.162 2.6.6.172 0.156 0.608 9State 0.172 0.4. Figure 7.156 0.6.3 15.5. b) Ammonia concentration against time.583 2.14 0. Figure 7.156 5State 0.

the 5state model is incapable to describe the behavior of the process after substrate has been depleted.225 0. with and without substrate.132 7. Singular function evaluations speed Evaluation of Jacobians CPU time 100 evaluations (sec) Calculation of the differential eq.Sto [gCODm-3] 5state ASM3 0 5 10 15 20 25 500 400 300 200 100 0 -100 XNS. The 5state model can be applied as an indirect method to measure substrate concentration.7: Description of the 5state model in both regimes. system CPU time 1000 evaluations (sec) ASM3 8State 5State ASM3 8State 5State 2071. 78 .7 MECHANISM RECOGNITION IN SBR PROCESSES We now analyze the capacity of the 5state model to act as an indicator of substrate concentration in the medium.Mechanism Recognition applied on Sequencing Batch Reactors Table 7. since the assumption made for its reduction clearly state the condition of substrate present during the process.157 0.7.6 2.XNB [gCODm-3] 5state ASM3 SS t [h] 120 100 80 60 40 XNS 5state ASM3 5 10 15 20 25 XNB 0 0 5 10 15 20 25 t [h] t [h] SNO-2 [mgNL-1] 30 20 10 0 -10 6 5state ASM3 SO [mgO2L-1] 5state ASM3 4 2 0 -2 0 5 10 15 20 25 150 0 5 10 15 20 25 t [h] t [h] SNO-3 [mgNL-1] 10 5 0 -5 5state ASM3 SNH4 [mgNL-1] 15 100 5state ASM3 50 0 0 5 10 15 20 25 0 5 10 15 20 25 t [h] t [h] Figure 7.114 565. As depicted in Figure 7.364 321. XH [gCODm-3] 800 700 580 560 540 520 500 Sto 600 SS.5. This characteristic of the 5state model is to be expected.

the parameters of the optimizer were set to maximize the probability of global convergence. Nicolas Cruz Bournazou Because of the qualities of the submodel. PROCESS HAS TO HAVE MORE THAN ONE REGIME As previously discussed.8. the system is required to fulfill conditions stated in section 6. simulation inaccuracies can be overcome by the switching point detection step.8 RECOGNITION OF ORGANIC MATTER DEPLETION 7. the program can distinguish if the failure on description capacity is due to a substrate mismatch or due to substrate depletion. Even for the case where the dynamics of the substrate concentration is not properly predicted. 79 . . PARAMETER BOUNDARIES HAVE TO ASSURE GLOBAL CONVERGENCE WITH GRADIENT BASED OPTIMIZERS The parameter set was tested for global optimality with stochastic optimization based on PSO [60]. 7.biomass growth based on consumption of organic matter. DETAILED MODEL OF THE GENERAL PROCESS AVAILABLE OF THE FORM: DAE system index zero or one no discontinuities known initial conditions The extended ASM3 fulfils all the above mentioned conditions. two different regimes are considered: .An approach to Mechanism Recognition for model based analysis of Biological Systems M.biomass growth under organic matter limitation. Although global optimality cannot be assured.2: DYNAMIC PROCESS The SBR process is characterized by changes in the state of the system throughout the complete run.1 CONDITIONS FOR PROPER PROCESS DESCRIPTION WITH MECHANISM RECOGNITION To ensure successful recognition of the different regimes during the process.

to allow a proper detection of the switching points in the process. the physical foundation of the extended ASM3 offers a basis for evaluation of the dynamics of the process. DEFINITIVENESS OF THE CHANGING BEHAVIOR - The distinguishability test proves that. the submodels are distinguishable. the sequence is known. it can be assured that the period of biomass growth on consumption of organic matter is long enough to guarantee identifiability and detection of the switching point.2 CONDITIONS FOR ACCURATE SWITCHING POINT DETECTION The assumptions mentioned above assure a successful model description of the process using rigorous submodels with low systematic error. allows an accurate detection of the change of regime. 80 . In addition. From this it may be supposed. it can be guaranteed that the initial regime of the process is biomass growth based on consumption of organic matter.3. In addition. the form of the general structure and the conditions of the process. THE SEQUENCE OF THE REGIMES IS KNOWN. following conditions are to be considered (section 6. that the process offers enough amount of information to detect the difference in bacterial behavior under both regimes. LENGTH OF EACH TIME REGIME ASSURES MODEL MINIMAL DISTINGUISHABILITY - Based on the variance of the data.Mechanism Recognition applied on Sequencing Batch Reactors INITIAL REGIME IS KNOWN Because of the high concentration of organic matter at the beginning of the cycle. 7.6): KNOWLEDGE OF THE NORMAL BEHAVIOR OF THE SYSTEM - The behavior of the bacteria in each regime is described by the submodels. at least for the simulated experiments. AVAILABILITY OF AT LEAST ONE OBSERVATION REFLECTING REGIME CHANGE - The combination of measuring variables considered.8. THE Since only two regimes are considered in this process.

consider constant yield coefficients. By these means. 81 .An approach to Mechanism Recognition for model based analysis of Biological Systems SATISFACTORY STATE OF INFORMATION M. the extended ASM3 and consequently in the submodels. The reaction pathways responsible for substrate consumption and processing depend on many conditions. energy production or storage. The PCPs are: μ𝐻 μ𝐴1 μ𝐴2 The RWP are: YHaer YA1 YA2 YA3 Yield coefficient of SS to XH in aerobic conditions Yield coefficient nitrite to nitrosomones in aerobic conditions Yield coefficient nitrite to nitrobacter in aerobic conditions Yield coefficient nitrate to nitrobacter in aerobic conditions Maximal growth of Heterotrophous Maximal growth of nitrosomones Maximal growth of nitrobacter INITIAL INTERVAL A high concentration of organic matter in the feed water to the tank is guaranteed assuring growth on organic matter as the initial interval. From this it can be deduced that the yields for substrate consumption achieved by bacteria vary depending on various factors. Nicolas Cruz Bournazou - The variance and frequency of the simulated experimental data provides the state of information of the system required for MR. the program can adapt the yield coefficient in each regime. Taking advantage of MR the yield coefficients are considered to be RWPs. Still. MEASURED VARIABLES MR was applied considering the following measured variables: Ammonia concentration [mgNL-1] Dissolved Oxygen in the medium [mgO2L-1] Nitrite concentration [mgNL-1] Nitrate concentration [mgNL-1] GENERAL STRUCTURE Cells utilize the consumed substrate for a number of functions as are growth.

8 Figure 7.8.8: Minimal length for initialization of MR The scenario selected to show the performance of MR is depicted in Figure 7.2 0.35.6 0. the large slope of H0 at the depletion region assures a robust and accurate detection of the regime change.4 time (days) 0. MR is initiated to detect depletion of organic matter. 82 .6 0.9.SO2 [mgNL-1] AG crit 40 20 0 10 1 0 0. As it can be seen in Figure 7. 7. Furthermore.SNH4-.9 and includes three short aerobic phases as well as three large anoxic intervals.4 0.4 time (days) 0. The anoxic intervals offer measurements with low information content since 1) the parameters set to estimation control mostly the aerobic phase and 2) the concentrations of oxygen and Nitroxides drops to zero. From this point on. MR detects the point with high precision.Mechanism Recognition applied on Sequencing Batch Reactors 7. Still MR shows an incredibly good performance in terms of both accuracy and robustness.6 0.2 0. 1000 10 3 SS [gCODm-3] 500 10 0 0.8. Substrate depletion is accurately detected despite the impossibility to measure the concentration of organic matter online.8 10 0 0 0.2 0.8 2 0 SNO3-. This scenario was selected in order to test MR in a process with very low state of information.SNO2-. The constant threshold was set equal to a MXL value of 0.4 DETECTION OF SWITCHING POINTS Detection of the switch between overflow and substrate limitation regime was carried out controlling the value of the objective function.2 hours).30 days (7.3 MR INITIALIZATION The threshold DB set at 10 for the A criterion of the general structure is reached after 0.

Nicolas Cruz Bournazou 1000 SS [gCODm-3] 500 SNO3-.6 0. it was possible to create a model that is significantly simpler and delivers more information about the process.1 0.9.6 0.5 0.7 0.3 0.4 time (days) 0.1 0.SO2 [mgNL -1] 0 60 40 20 0 0 0. and different approaches to model reduction have been applied to develop a submodel of the extended ASM3.4 0. 7. The implementation of MR for description of ASP in SBR processes clearly states the advantages of this approach.2 0. and thus.9 CONCLUSIONS An assiduous analysis of the process conditions. 83 .SNO2-.4 H0 [] 0.An approach to Mechanism Recognition for model based analysis of Biological Systems M.1 0.3 0. the extended ASM3 was successfully simplified.4 0.7 0.7 0. Moreover.2 0 0 0.3 0. for model-based control and online optimization.2 0.5 0.SNH4-. Detection of the regime switching point.8 0 0.5 0. reducing its calculation costs up to one order of magnitude. By these means.8 Figure 7.2 0. The results suggest that the 5state model can be applied for the simulation of the nitrate bypass reaction in SBR processes. empirical knowledge.8 0. SWITCHING INTERVAL The experimental results confirmed that considering an instant switch of regime allows an accurate description of the process.6 0. The resulting 5state submodel mimics the behavior of the extended ASM3 for SBR process during growth based on organic matter with high accuracy.

the reduction of the extended ASM3 creates a model which can indicate the time point where organic matter is present in the reactor.Mechanism Recognition applied on Sequencing Batch Reactors Most important. 84 . the limitation of the submodel to describe the process after depletion of organic matter can be exploited. In other words. The program is able to detect the time point when no more carbonate matter is present in the reactor increasing process efficiency. The 5state model works as an indicator of the non measurable state variable readily biodegradable organic matter.

It could be demonstrated in accelerostat studies. the activation of the glyoxylatic shunt. The carbon flow through acetyl-CoA is partly shifted towards acetate production instead of entering into the tricarboxylic acid cycle [155]. The product acetyl-phosphate is then converted to acetate by the acetate kinase enzyme. Also a direct conversion from pyruvate to acetate with the pyruvate oxidase is possible [156. Acetate is synthesized via phosphotransacetylation. proteolysis and pathogenesis [159]. coli at substrate excess (overflow metabolism) when acetic acid is formed as carbon storage. coli is the downshift of the ACS system leading to a lower conversion of acetate via acetyl-AMP. However. acetate synthesis already occurs under some conditions at weak excess. and the response to oxygen limitation. 157]. Pulsed feeding strategies based on dissolved oxygen measurements reduced acetate formation increasing growth and product concentrations [160. Several attempts have been made to circumvent acetate accumulation. the increasing number of equations 85 . The latter reaction complex is characterized by a 50-fold higher affinity and therefore responsible for acetate reconversion at low concentrations [158]. acetate synthesis and excretion occurs during the so called „overflow metabolism‟ [154] when substrate concentration (namely glucose) is available in large amounts. that the first step of overflow mechanism in E. are in the focus of recent development.1 ESCHERICHIA COLI CULTIVATIONS Many reports about process optimization of Escherichia coli cultivation are describing problems caused by acetate synthesis. Acetate itself can be reconverted to acetyl-CoA either via acetyl-phosphate or by acetyl-CoA synthetase (Acs) via acetyl-AMP. In E. The behavior of E. Nicolas Cruz Bournazou 8 MECHANISM RECOGNITION IN ESCHERICHIA COLI CULTIVATIONS 8.An approach to Mechanism Recognition for model based analysis of Biological Systems M. The threshold value depends on the strain and culture conditions. While more and more regulatory mechanisms are considered in models. Furthermore. which can retard growth and protein production [151-153]. the reconversion of acetate is regarded as being essential for chemotaxis. 161]. coli.

parallel to advances in measuring techniques. The great advantage of the method is the absence of cell staining which allows for a much simpler and automated sample preparation. 171]. 167]. Also. Junne [172] reports an approach aiming at monitoring the polarisability and cell length at line and correlate these parameters with the cell viability as it was performed based on flow cytometry studies in batch. In this study. coli. 8. dynamic models including the regulatory interaction on the metabolomic level have been developed.Mechanism Recognition in Escherichia coli cultivations and parameters increment model complexity. Wang [176] described the catabolite repression based on the sucrose and glycerol transport system in E. More and more knowledge is available concerning the interaction of messenger nucleotides. Nowadays. this bacterium became the “workhorse” of experimentalists for studying novel microbiological and analytical methods and for the industrial production of proteins and other substances of commercial interest [174]. For this reason. Chassagnole [175] created a kinetic model of the sugar uptake system (phosphotranspherase system – PTS) and the glycolysis following the Embden-Meyerhof-Parnas pathway in E. which act at the DNA replication. Some of them have been created for the scientifically well-described facultative bacterium Escherichia coli. models aiming at an accurate description of bacterial behavior at different levels are being developed. In addition. and continuous cultures [173].2 MODELS FOR THE DESCRIPTION OF ESCHERICHIA COLI CULTIVATIONS Many approaches have been developed during the last thirty years to create descriptive and predictive models for bacterial organisms. there is no possibility to monitor these parameters on-line. The number of parameters that can be monitored and hence integrated in models as measured variables is steadily increasing. In recent years. 86 . continuous glucose monitoring [163]. 168. the PTS-dependence of the sucrose transport system was considered in the model. Due to its easy cultivation. very advanced techniques are available including online respirometry [162]. For example. multi-wavelength fluorescence (MWF) [166. online high performance liquid chromatography (on-line HPLC) [165]. fed-batch. However. coli. many compartments of the metabolism cannot be measured accurately or in a sufficient resolution by time. translation. near-infrared spectroscopy (NIRS)[162. 169]. Also methods for describing the physiological state of the cell are developed for monitoring bacterial cultivation like flow cytometry for example [170. as well as conventional but very efficient techniques like Dissolved Oxygen Tension (DOT) [164] among others. since the appropriate detection of intermediates is crucial due to their naturally low levels in bacteria. Still efficient methods to obtain sufficient state of process information with online observations are to be developed.

a recently developed model of the total central carbon metabolism in E. parameter estimation becomes a difficult and sometimes even impossible task. It is activated. coli cultivations has recently been described [7]. parameter estimation with a limited set of experimental data leads to unreliable results while multiple solutions exist. namely overflow. the model developed by Lin does not consider variation of cell response caused by the physiological state of the bacteria.2. coli comprises of 50 kinetic rate equations and 46 equations that simulate gene expression in the corresponding pathways [182]. The consideration of alarmones contribute to a more sensitive prediction capacity and a simulation of dynamic cell behavior [181]. coli cultivations in research and production. Still these reduced versions at heuristic level. For this reason. 8. in this section an alternative modeling approach is proposed considering three main cell states.1 DIVISION OF PHYSIOLOGICAL STATES In an effort to create a tractable model able to describe substrate and oxygen uptake capacities of the cells. although there is a great field of application for models for description of E. which has a major impact on the process performance [178]. Hence. For example. The alarmone guanosine tetraphosphate (ppGpp) acts as a main modulator of gene transcription [179]. the simulation of the impact of ppGpp and other nucleotides on the cell‟s regulation is of great interest for predicting the physiological state of the cell and consider it for modeling [126. Among recent strategies to overcome these problems are the application of simplified kinetics for expressing the biochemical reaction rate equations. The integration of signal molecules can serve as bridge between the regulation on the metabolome and the transcriptome. need to be continuously fitted to new conditions and do not offer nor inside of process conditions. 184]. 87 . when the cell starves as in largescale fed-batch cultivations. e. coli. Other approaches to reduce the complexity of complex models included temporal decomposition and establishment of pools which describe the physiologic status of the cell with very few parameters [185]. However. although precision of prediction could not be fully maintained. 180].g. This leads to the problem that for a wider application range of the model. the model reduction maintained the dynamic simulation capacity of the model for the starvation response of E. For the equation system see Appendix A.An approach to Mechanism Recognition for model based analysis of Biological Systems M. with linearlogarithmic (so called lin-log) approaches [183. coli. nor of the physiological state of the bacteria. Therefore. Also the successful application of piecewise-linear approximations for a multilevel considering model for E. This model considers time dependent uptake capacities over time and thus enables a better fit of experimental glucose data. Despite the important advances achieved. Here. their application is very restricted due to the afore mentioned limitations. Nicolas Cruz Bournazou transcription and enzyme activity [177]. Models included the general stringent response in E. Lin [91] proposes a model where cellular response dynamics are lumped into two variables.

A clear difference of the expression of enzymes in the glycolysis and the TCA cycle depending on the state of the culture was determined. MR enables a closer analysis of the dynamics of the model and the correctness of its structure. Veloso [190] developed a software sensors framework to monitor E. namely: Biomass concentration Substrate concentration Acetate concentration Oxygen concentrations in the offgas Carbon dioxide concentrations in the offgas 88 . This information can be used to account for altered enzyme concentration at different cultivation conditions (leading to different physiological conditions inside the cell). The expression of it does not change at altered substrate uptake [159]. Three submodels. Acetate is then converted to acetyl-CoA via the acetyl-CoA synthase (Acs) [159. each one describing one of the aforementioned physiological states of the cell. coli fermentations. coli cultivations. Despite of this. 186]. The simplicity of the models applied. coli cultivations have proven the dynamic development of transcription. 159. The expression of Acs is strongly reduced (more than 30-fold) at high substrate uptake rates (at the onset of acetic acid accumulation). 187]. Recent studies in chemostat and accelerostat E. This indicates that the pathway provides acetate accumulation at lower substrate uptake rates.Mechanism Recognition in Escherichia coli cultivations growth on substrate limitation and starvation (under which the regulatory interaction of the alarmone ppGpp is active). although acetate is produced. Similar to the concept of MR. the robustness and a better evaluation of the changing states of the process are potential contributions of MR to the description of E. Over expression of this pathway has shown that acetate accumulation could be reduced [188]. Besides. The low Km value of Acs of 200 µM supports this hypothesis. By this means it is possible to reduce the number of parameters and increase model flexibility and model identifiability [189]. For example. A further advantage of this approach is the achievement of a reliable mechanistic recognition of non measurable parameters. are to be developed in an effort to increase prediction accuracy while maintaining model simplicity. He applied an asymptotic observer based on a simplified model with five state variables. the usual conversion of accumulated acetate via the acetate kinase to acetyl-phosphate is characterized by a value of the catalytic enzyme of 7-10 mM [158]. acetate was shown to be synthesized already at low carbon supply via the pyruvate oxidase (Bpox) from pyruvic acid [156]. protein expression and carbon conversion at different dilution rates [155.

μ3 = 0) regime B: . 𝑘𝑖 are the yield (stoichiometric) coefficients. 𝐶𝑇𝑅 is the carbon dioxide transfer 𝑚 rate from liquid to gas phase and 𝑂𝑇𝑅 is the oxygen transfer rate from gas to liquid phase. Veloso divides the process in four possible phases. 𝑂.simultaneous oxidative and overflow growth on glucose (μ1 . μ2 = 0) regime D: . glucose. the growth rate is considered to be constant for each regime. μ1 = μ2 = 0) - - This proposition allows building observable models with a reduced number of measured state variables. Nicolas Cruz Bournazou 0 𝐹𝑖𝑛 ∗ 𝑆𝑖𝑛 𝑊 𝑚 0 𝑂𝑇𝑅 −𝐶𝑇𝑅 (8. Finally. Assuming a piecewise constant growth rate simplifies the model to the extent where observability is obtained.simultaneous oxidative growth on acetate and glucose (μ1 . and dissolved carbon dioxide concentrations. 𝐹𝑖𝑛 and 𝑆𝑖𝑛 are the substrate feed rate and the glucose concentration in the feeding solution. Although theoretically applicable for process monitoring. μ2 = μ3 = 0) regime C: .1) where 𝑋. 𝐴𝑐. From a physical point of view. respectively. The results obtained cannot be translated to a first principle model to extend state of information. respectively. and 𝐶 represent biomass. The model proposed for monitoring is simple and has no physical basis nor offers information about the system.An approach to Mechanism Recognition for model based analysis of Biological Systems combined in the form: 1 𝑋 −𝑘1 𝑑 𝑆 𝐴𝑐 = 0 𝑑𝑡 𝑂 −𝑘5 𝐶 𝑘8 1 −𝑘2 𝑘3 −𝑘6 𝑘9 1 𝜇1 0 −𝑘4 ∗ 𝜇2 ∗ 𝑋 − 𝐷 ∗ 𝜇3 −𝑘7 𝑘10 𝑋 𝑆 𝐴𝑐 + 𝑂 𝐶 M. Similar to the MR approach.oxidative growth on acetate (μ3 > 0.μ2 > 0. and μ3 are the specific growth rates. μ3 > 0. the proposed method has some important drawbacks. 𝑆. 𝜇2 . D is the dilution rate and 𝑊 is the culture medium weight. The model is not continuous (discontinuity hinders gradient calculation in the switching point). 𝜇1 . 89 . dissolved oxygen. acetate. Veloso divides the batch process to enable its description with simpler models.oxidative growth on glucose (μ1 > 0. named regimes: regime A: .

This supposes important process knowledge impossible to obtain in a real process.Mechanism Recognition in Escherichia coli cultivations - The regimes have to be known beforehand. Three models compete in different intervals of the process. substrate concentration and cellular metabolic activity. In order to overcome the above mentioned disadvantages.1 GENERAL MODEL The model for the description of E. Although models able to describe individual batch and fedbatch fermentations exist. To overcome this drawback. acetate formation. Starvation model (ST). Fermentation processes are characterized by its dynamic behavior described by parameters such as growth rate. coli K12 fed-batch fermentations was extracted from Lin [91]. 8. substrate consumption and cell growth are described with high accuracy. the data needed to fit the models are reduced and a standardization of the model to be applied in different process states is enabled. The candidate models are: Overflow metabolism model (OF). with a relatively simple kinetic model. MR is applied to achieve an efficient mechanistic modeling and simulation of E. the MR approach is proposed. A series of experiments with glucose pulses were carried out to determine the dynamics of the uptake capacity of glucose during fermentation. 8. Growth on substrate Limitation model (SL). Moreover.3 MODELING ESCHERICHIA COLI BATCH FERMENTATIONS WITH MECHANISM RECOGNITION In this section. coli batch fermentations and the simulation framework is validated against experimental data.3. Using an adequate model sequence. This model was originally built in an effort to describe all physiological and regulatory effects. which cause uptake capacity variation of substrate and oxygen. 90 . diverse models are used at various regimes enabling not only a better description of the process. they become unreliable when applied to different fermentation conditions. but also an improved understanding of non measurable characteristics.

This model successfully describes the variations of the uptake capacities of substrate and oxygen of determined fermentations. the enzymes for glucose uptake and for oxygen uptake to biomass [91]. Nicolas Cruz Bournazou To describe these variations.1 Biomass Concentration 5 Substrate Concentration 15 [g/l] 5 0 5 10 15 [g/l] 0 10 5 10 DOT 15 1 Acetate Concentration 150 [mMol/l] 0.05 0 ratio of oxygen consumption [] 0 5 t [h] 10 15 [] 5 t [h] 10 15 Figure 8. 91 .5 0 5 10 15 [%] 100 50 0 5 10 15 ratio of substrate consumption 0.Biomass X [g/l] 14 12 10 8 6 4 2 0 0 1 2 3 4 5 time [h] 6 7 8 9 10 A X S Figure 8. coli under different process conditions.) fitted to experimental batch cultivation data. Acetate A [mMol/l].2: Complex model (Lin et al. The behavior of the six principal variables: biomass.An approach to Mechanism Recognition for model based analysis of Biological Systems M.1 0. and oxygen uptake are depicted in Figure 8. acetate. Substrate S [g/l].05 0. Lin creates two fictitious enzymes in which all physiological and regulatory effects are put together. substrate.1: Integration of the kinetic model proposed by Lin [91] Still. DOT substrate uptake. the model fails to describe the extremely different behavior of E.

The expression of both enzymes. is maximal and constant.3) in order to overcome local minima. The results of the parameter estimation against real data are depicted in Figure 8. for substrate and oxygen uptake capacity.3. Lin [91]summarizes two ratelimiting steps contributing to overflow of acetate in aerobic glucose-based cultures of E. The optimization was carried out with a stochastic optimizer PSO (section 5. Moreover. Acetate production is related to growth inhibition [153.3. 195] The general model is reduced to limit its description capacity to the specific conditions of the overflow metabolism (OF): Substrate concentration is always higher than the minimum amount needed for maintenance. Each submodel describes one of the three main states of the cell: overflow metabolism Figure 8. and starvation Figure 8. 191] and to reduction of recombinant protein production [192. Acetate consumption is considered. 193]. 8. the model fails to describe different experiments.Mechanism Recognition in Escherichia coli cultivations In addition.4. Bacteria consume substrate and oxygen to its maximum possible under consideration of substrate and product limitation. coli: in electron transport system [151] rate-limiting steps in the in the TCA cycle [194. MODEL FOR DESCRIPTION OF THE OVERFLOW METABOLISM Drawbacks of acetate production are widely known and have been reported in various contributions. 92 .5. growth under substrate limitation Figure 8. the model contains discontinues equations which represents a major obstacle for robust simulation and optimization.2.2 SUBMODELS FOR DIVIDING METABOLIC STATES A reduction of the model in three submodels makes it possible to overcome the previously mentioned drawbacks improving description accuracy and robustness. since bacteria also consume acetate in overflow conditions where high acetate concentration is present in the medium.

An approach to Mechanism Recognition for model based analysis of Biological Systems

M. Nicolas Cruz Bournazou

15

Biomass Concentration 5

Substrate Concentration

[g/l]

5 0 0 5 10 15

[g/l]
0 0

10

5 DOT

10

15

1

Acetate Concentration

150

[mMol/l]

0.5 0 0 5 10 15

[%]

100 50 0 0 5 10 15

ratio of substrate consumption 0.05

0.1 0.05 0 0

ratio of oxygen consumption

[]

0

0

5 t [h]

10

15

[]

5 t [h]

10

15

Overflow

Figure 8.3: Comparison between the complex model (dots) vs. the overflow submodel (lines) initializing in four different intervals.

In Figure 8.3 it can be seen that the submodel describes the initial stage of the process with high precision. Four intervals were selected arbitrarily to test the description capacity of the model in different stages of the process. The prediction accuracy is drastically reduced as the concentration of substrate drops. This suggests that the submodel is adequate for the identification of the bacterial growth with overflow metabolism. For the equation system see Appendix B. MODEL FOR DESCRIPTION OF GROWTH UNDER SUBSTRATE LIMITATION Many fermentation strategies as well as metabolic engineering approaches have been applied to reduce the production of acetate [191, 196-202]. It is commonly accepted, that optimal conditions for cultivation are obtained under (moderate) substrate limitation. By these means acetate accumulation is minimized and biomass to substrate yield maximized, while maintaining a relatively high growth rate [155]. A topology of the regulatory network of carbon limitation has been published by Hardiman [181] on which it is possible to estimate to which extend substrate limitation is suitable for application and when starvation and stringent responses are initiated. Following assumptions were made to create the substrate limitation model: Substrate concentration is always higher than the minimum amount needed for maintenance. There is no limiting capacity for oxygen consumption.

93

Mechanism Recognition in Escherichia coli cultivations

15

Biomass Concentration 5

Substrate Concentration

[g/l]

5 0 0 5 10 15

[g/l]
0 0

10

5 DOT

10

15

1

Acetate Concentration

150

[mMol/l]

0.5 0 0 5 10 15

[%]

100 50 0 0 5 10 15

ratio of substrate consumption 0.05

0.1 0.05 0 0

ratio of oxygen consumption

[]

0

0

5 t [h]

10

15

[]

5 t [h]

10

15

Sub. Limit.

Figure 8.4: Comparison between the complex model (dots) vs. the substrate limiting submodel (lines) initializing in four different intervals.

Contrary to the previous submodel, the submodel for substrate limitation is only able to describe cell behavior at very low substrate concentrations Figure 8.4. The model predicts an extremely fast growth in overflow conditions, because it lacks uptake capacity limitation at overflow metabolism responsible for acetate synthesis and excretion. The detailed equation system is presented in Appendix C. MODEL FOR DESCRIPTION OF CELL STARVATION Finally the description of the starvation stage is also considered. The cell starvation is to be avoided at all cost in industrial E. coli cultivations. Still large scale reactors inherently imply special concentration gradients in the medium exposing bacteria to undesired conditions. Extreme substrate limitation may trigger the phosphoenol pyruvate glyoxylate (ppG) shunt. Once the stringent response is activated, reactivating bacterial growth requires long periods under substrate excess. To reduce the models it was assumed that: substrate uptake is lower than required by the cell, hence no energy limitation is required enzyme rate are minimal and constant

94

An approach to Mechanism Recognition for model based analysis of Biological Systems

M. Nicolas Cruz Bournazou

15

Biomass Concentration 5

Substrate Concentration

[g/l]

5 0 0 5 10 15

[g/l]
0 0

10

5 DOT

10

15

1

Acetate Concentration

150

[mMol/l]

0.5 0 0 5 10 15

[%]

100 50 0 0 5 10 15

ratio of substrate consumption 0.05

0.1 0.05 0 0

ratio of oxygen consumption

[]

0

0

5 t [h]

10

15

[]

5 t [h]

10

15

Starvation

Figure 8.5: Comparison between the complex model (dots) vs. the cell starvation submodel (lines) initializing in four different intervals.

As expected, the model describing cell starvation is not able to predict the behavior at any stage of the process Figure 8.5. This indicates that the process does not reach starvation conditions. For the equation system see Appendix D. The structure of the submodels is shown in Table 8.1. The identifiability of the submodels will be determined by the parameter boundaries and the consistency of the general structure indicates the following. This again is key to accurate regime recognition. Unlike the process constant parameters, the regime-wise constant parameters are optimized independently in each regime. The parameters estimated during the recognition process are shown in Table 8.1
Table 8.1: Parameters considered for the model fit

PCPs (general structure, section 6.3.2): 𝑞𝐸𝑜𝑑 𝑞𝑂𝑚𝑎𝑥 𝑞𝑆𝑚𝑎𝑥 𝑞𝐸𝑠𝑑 specific death rates for the uptake enzyme of oxygen [g/(gl)] Maximal specific oxygen uptake rate [g/(gh)] Maximal specific substrate uptake rate [g/(gh)] specific death rates for the uptake enzyme of substrate [g/(gl)]

95

001 g/l Glucose was autoclaved separately to avoid Maillard-reactions.1 STRAIN AND CULTURE CONDITIONS Glucose Yeast -extrakt K2HPO4 KH2PO4 (NH4)2SO4 MgSO4 Na2EDTA*7H2O NaCl FeSO4*7H2O 10 g/l 5 g/l 3 g/l 1.5 g/l 1. Germany. Karlsruhe.4 MATERIAL AND METHODS (All chemicals in this study were either purchased by Sigma Aldrich GmbH. The flasks were incubated at 37°C and 150 rpm on a longitudinal shaker.01 G/l 0.Mechanism Recognition in Escherichia coli cultivations RWPs: 𝑌𝑋𝑆𝑜𝑥 𝑌𝐸𝑂𝑋 𝑌𝐸𝑆𝑋 𝑌𝑋𝐴 𝑌𝑋𝑆𝑜𝑓 Yield of the enzyme for oxidative energyproduction from the biomass growth [-] Yield of the enzyme for substrate uptake from the biomass growth [-] Yield of the enzyme for respirationfrom the biomass growth [-] Yield of the enzyme for acetate from the biomass growth [-] Yield of the enzyme for overflow from the biomass growth [-] 8. Germany. if not otherwise stated.1 g/l 0. coli K12 W3110 picked from a frozenstock culture.) 8. 96 . 1000 ml Erlenmeyer flasks were filled with 250 ml of medium.25g/l 0.037 g/l 0. The medium was inoculated with E. or Carl Roth KG. Munich.4.

97 . Both. which is controlling the heating and cooling devices. The stirrer was working at 700 rpm and polyethyleneglycol 2000 to 3000 was added in small amounts ( 1ml) to reduce foam whenever needed. The reactor was sterilized prior to cultivation at 120°C and 1 bar overpressure for 20 min. The level of pH was controlled to pH 7. the gas inlet and the exhaust gas were steril-filtered with autoclavable cellulose filters. Switzerland) was used for all fermentations Figure 8.0 with NaOH 30% w/w during the acetate production phase and HNO3 10% w/w during the acetate consumption phase.6. Nicolas Cruz Bournazou A reactor KLF2000 (Bioengineering AG. the reactor was sparged with air at a rate of 1 vvm.2 ONLINE ANALYSIS A sterilizable temperature detector Pt100 is connected to an electronic measuring and control unit.4. To insure sufficient oxygen supply.An approach to Mechanism Recognition for model based analysis of Biological Systems BATCH BIOREACTOR EXPERIMENTS M. 8. Wald. Figure 8. coli batch cultivation [203] The inlet gas filter was autoclaved separately while the outlet gas filter was autoclaved together with the vessel by a small steam flow.6: Bioreactor KL2000 at E.

3 Where: VF . Hamburg. CO2 concentration in the exhaust gas is measured by a non-depressive infrared photometer with a selective optical-pneumatic receiver UNOR 4 N (Maihak AG.4 8. O2 concentration in the exhaust gas is determined by an OXYGOR 6 N (Maihak AG. Because of the implementation of a smaller cuvette with a thickness of 1.volumetric gas flow m3/h QO 2 . Mettlach. By this means.6 mm instead of the usual 10 mm. the device takes a sample and measures it every 15 seconds.Mechanism Recognition in Escherichia coli cultivations An Ingold single-rod measuring cell is used to measure the pH of the medium.7 was also developed by EloSystemsGbR for biomass online monitoring. The data of EloCheck was not used for further analysis because of small deviation due to antifoam agent which caused temporary occlusion in the cuvette. The concentration of the dissolved oxygen in the medium is detected with an amperometric (polarographic) Ingold pO2-electrode (Mettler-Toledo Inc. Germany). 98 . Germany).oxygen uptake rate [mol/lh] QCO2 carbon dioxide production rate [mol/lh]  YO2 -mole fraction of oxygen in the incoming gas phase [ ] .4 8.. Germany) device following the magneto-pneumatic measuring principle.fluid volume [l]  VG . The oxygen uptake was calculated based on the exhaust gas analysis as follows: QO 2     VG (YO2  YO2 ) VF * 22.mole fraction of oxygen in the exhaust gas [ ] YO2  ONLINE-MEASUREMENT OF THE OPTICAL DENSITY WITH ELOCHECK EloCheck Figure 8. EloCheck needs no sample dilution which makes it possible to reintroduce the measured sample into the reactor. EloCheck is able to measure the optical density of the sample through a bypass system [204].2 The carbon dioxide production rate was calculated as: QCO2     VG (YCO2  YCO2 ) VF * 22. Hamburg. Both biomass curves from EloCheck and offline analysis were compared and a high similarity on the run of the curves was evident.

The extinction of NadH is measured at λ = 340 nm. Wiesbaden.” were filled with 1 ml sample. per 1.4. During this chemical conversion NadH. is accumulated: The amount of NadH is proportional to the concentration of the glucose. About 3 ml were used for biomass determination. Germany).2 to 0. The biomass containing sample was diluted until the extinction was in the range of the Lambert-Beer law (0.000 rpm. A calibration curve consisting of 99 . The supernatant was separated through centrifugation for 10 min at 13. the glucose was the substrate for the reaction which is catalyzed by the hexokinase and the glucose6-phosphat-dehydrogenase at presence of Adenosintriphosphat (ATP) and Nicotinamide Adenine Dinucleotide (NAD+. NadH). In this test.000 pcs. One flextube supernatant was used for the glucose determination and the second one was stored at –20°C for later acetate determination. Nicolas Cruz Bournazou Figure 8.398*extinction [nm]-0.5).039 ENZYMATIC DETERMINATION OF GLUCOSE Glucose determination was performed with the liquiUVmono Test (Human Geselschaft für Biochemical und Diagnostica mgH.3 OFFLINE ANALYSIS SAMPLE PROCEDURE About 5 ml of crude extract were withdrawn from the reactor every 30 min.An approach to Mechanism Recognition for model based analysis of Biological Systems M. The determination of the extinction was performed at λ = 600 nm.5 ml.7: EloCheck 8. West Germany). Two Eppendorf “Flex-Tubes® 1. The equation for the linear regression between biomass and extinction was: Biomass [g/l] = 0. BIOMASS DETECTION Biomass was determined through optical absorption measurement with a Zeiss Spektral Photometer (PM2A.

𝐶𝐺𝑙𝑢𝑐 𝑛 +1 − 𝐶𝐺𝑙𝑢 𝑐𝑛 𝑡𝑛 +1 − 𝑡𝑛 = 𝐶𝐵𝑖𝑜 𝑛+1 − 𝐶𝐵𝑖𝑜 𝑛 𝐶𝐵𝑖𝑜 𝑛 + 2 𝑞𝐺𝑙𝑢𝑐 where: 𝑞𝐺𝑙𝑢𝑐 8.10148261035.8 1 Figure 8.776*extinction [nm] – 0.8.4 0. The specific substrate uptake rate was obtained by dividing the difference off the acetate concentration between two samples.6 extintion [nm] 0. the calibration curve for the first experiment is shown: Concentration [g/l] = 2. The determination is realized on a light absorbance wave length of λ = 340 nm..4 = specific glucose uptake rate [h-1] 𝐶𝐺𝑙𝑢 𝑐𝑛 = glucose concentration [g/l] 𝐶𝐵𝑖𝑜 𝑛 = Biomass concentration [g/l] 𝑡𝑛 = time [h] ENZYMATIC DETECTION OF ACETIC ACID Acetic acid determination was realized with the UV method. Fullerton.5 2 1. by the biomass at this time.8. which is based on the reaction from acetic acid to acetyl-CoA [205] (ROCHE Test Kits Cat.014 [g/l] 3 concentration [g/l] 2. Germany). No.2 0. R_BIOFARM AG.5 0 0 0. Darmstadt. In Figure 8.Mechanism Recognition in Escherichia coli cultivations duplicate measurements at five different concentrations was made prior to each fermentation. Calibration curve for glucose determination Optical measurement was realized by a Beckman DU640 spectrophotometer (Beckman Coulter Inc. 100 . USA).5 1 0.

Calibration curve of acetate The concentration of acetate was calculated with the following formula obtained out of the calibration curve: Acetate concnetration = 2.2 Figure 8.15 0.412*Absorbance – 0.05 0 0.1 Conc Acetate (g/l) 0. The calibration curve was plotted with the data obtained out of the supplied control solution at five different concentrations Figure 8. by the biomass at this time. Conc vs Abs 0. Fullerton.9.An approach to Mechanism Recognition for model based analysis of Biological Systems M. Nicolas Cruz Bournazou Samples were taken every 30 min.3 0. USA).25 Abs 0.35 0.2 0. Optical measurement was realized by a Beckman DU640 spectrophotometer (Beckman Coulter Inc. 𝐶𝐴𝑐𝑛 +1 − 𝐶𝐴𝑐𝑛 𝑡𝑛+1 − 𝑡𝑛 = 𝐶𝐵𝑖𝑜𝑛 +1 − 𝐶𝐵𝑖𝑜𝑛 𝐶𝐵𝑖𝑜 𝑛 + 2 𝑞𝐴𝑐 where: 𝑞𝐴𝑐 𝐶𝐴𝑐𝑛 8.9.1 0.05 0. centrifuged and stored at -20°C for no more than three weeks.15 0.0186 [g/l] The specific acetate production rate was obtained by dividing the difference off the acetate concentration between two samples..5 = specific acetate production rate [mMol/(g*h)] = acetate concentration [mMol/l] 𝐶𝐵𝑖𝑜𝑛 =Biomass concentration [g/l] 𝑡𝑛 = time [h] 101 .

0.2.0) absolute ethanol 40mM EDTA (pH8.2mM MTT 16.10: Mechanism of the reactions involved in the assay The cycling assay Figure 8.0) 4.6 mM phenazineethosulfate (PES). Briefly. Cellular debris was removed by centrifugation for 5 min at 15.000 rpm.01–0. a 1 ml sample was pipetted into a microcentrifuge tube. The absorbance at λ = 570 nm was recorded for 10 min.Mechanism Recognition in Escherichia coli cultivations DETERMINATION OF INTRACELLULAR NADH CONCENTRATION The intracellular NADH concentration was measured by a cycling assay according to Nisselbaum and Green [206] modified by Bernofsky and Swan [207] and improved by San and Bennett [208]. MTT) and twice the volume of 16. After 10 min of incubation in a 50°C bath. previously incubated at 30°C. The assay was calibrated with 0.1 M Bicine (pH 8.6 ml reagent mixture.2 M NaOH were added for cell lysis.05 mM standard solution of NAD+. The following volumes were added to 1 ml cuvettes: 50 ml neutralized extract.0 M bicine buffer (pH8.1 M HCl while vortexing. The reaction was started by adding 50 ml of alcohol dehydrogenase (ADHII isolated from rabbit) of a concentration of 500 or 100 U/ml in 0.10 was performed using a reagent mixture consisting of equal volumes of 1. The centrifugation was performed for 10 min at 13.6 ml solution A 102 .5-dimethylthiazol-2-yl]-2.6 mM PES Volume 650 µl 650 µl 650 µl 650 µl 650 µl 1300 µl 1 ml cuvettes were filled with the following mixture: 50 µl extract 0.2 mM 3-[4.3 ml water and 0. 40 mM EDTA (pH8. 300 µl of 0. Table 8. Figure 8. 4.0 M bicine buffer (pH8.0) buffer. Composition of solution A Solution 1.0). absolute ethanol.000 rpm. samples where cooled to 0° C and neutralized by carefully adding 300 µl of 0.5-diphenyltetrazolium bromide (thiazolyl blue.0).

0) buffer) The biomass yield was calculated with the following equation: 𝑌𝑥/𝑠 = where: 𝑌𝑥/𝑠 = Biomass yield [-] 𝐶𝐵𝑖𝑜𝑛 +1 − 𝐶𝐵𝑖𝑜𝑛 𝐶𝐺𝑙𝑢𝑐 𝑛 +1 − 𝐶𝐺𝑙𝑢 𝑐𝑛 8. In the case of the biomass curve.1 M Bicine (pH 8. Acetate samples were taken every 30 min. The acetate productionrate was obtained after dividing the difference of acetate between two samples and dividing the result through the time between both samples.3 ml water To start the reaction: M. Ac  Ac2  Ac1   t2  t1  t 8. oscillations at the beginning of the curve were also attenuated per hand.7 Finally.An approach to Mechanism Recognition for model based analysis of Biological Systems 0.4 DATA TREATMENT The data of every pair of fermentations with the same conditions were considered for calculating a fit-curve. This procedure was also applied to all other parameters which were converted into specific rates.4.6 𝐶𝐺𝑙𝑢 𝑐𝑛 = glucose concentration [g/l] 𝐶𝐵𝑖𝑜𝑛 =Biomass concentration [g/l] 8. The data of every pair of fermentations with the same conditions were considered for calculating a fit-curve whenever possible. Samples that were considered as outliers were manually excluded. the acetate building-rate was divided with the biomass to obtain the specific acetate building-rate. 103 . The fit curve was calculated in Matlab with a 7th grade polynomial. Nicolas Cruz Bournazou 50 µl of ADH (500 or 100 U/ml in 0.

1 CONDITIONS FOR PROPER PROCESS DESCRIPTION WITH MR MR was tested against batch cultivation with E. which also represents the optimization with the largest number of parameters to fit. PROCESS HAS TO HAVE MORE THAN ONE REGIME As previously discussed. Therefore it has to be proven that MR allows using complex models applied to real experimental data. One could argue that.Mechanism Recognition in Escherichia coli cultivations 8. overflow metabolism. coli K12 W3110 bacteria. In other words it is possible to prove that the structure of the complex models describes the process dynamics completely and no uncertainties or unknowns are present besides white noise (stochastic error). However. 8. MR achieves an accurate detection of the change of regime and finds consistent parameter combinations for both models (OF and SL).6 that the objective of the MR approach is to enable industrial application of complex models. The submodels created in the previous section showed good performance when compared to simulated data.5 EXPERIMENTAL VALIDATION First we compare submodel performance against a simulated data set. Nevertheless. This indicates that it is possible to find combinations of simple models with similar dynamics to more complex versions.5. the highest accuracy prediction of these artificially generated observations is clearly obtained with the complex model. 104 . coli is known to present three basic regimes. It has been clearly stated in section 1. growth under substrate limitation and starvation. DYNAMIC PROCESS The batch process is characterized by changes in the state of the system throughout the complete run. it can be assured that the structure of the model is globally correct (no systematic error). the effort of submodel building and MR programming does not justify its application against simulated data. To ensure successful recognition of the different regimes during the process. Certainty of optimal structure for the complex model allows a direct analysis of process phenomena and the precise detection of changing regimes. It is essential to test the performance of MR against data sets obtained in real experiments. the cultivation of E. MR offers a simplified basis for simulation and optimization. Since the complex model used for submodel building is the same model used for data generation. the system is required to fulfill following conditions. also in these cases.

where the parameters of the optimizer were set to maximize the probability of global convergence. LENGTH OF EACH TIME REGIME ASSURES MODEL THE MINIMAL DISTINGUISHABILITY - Based on the variance of the data. Nevertheless. The dynamics of substrate consumption and the reaction capacity of the bacteria assure that the regime following overflow conditions is growth under substrate limitation. 8. Nicolas Cruz Bournazou DETAILED MODEL OF THE GENERAL PROCESS AVAILABLE OF THE FORM: DAE system index zero or one no discontinuities known initial conditions The model applied to describe this process fulfils all the above mentioned conditions PARAMETER BOUNDARIES HAVE TO ASSURE GLOBAL CONVERGENCE WITH GRADIENT BASED OPTIMIZERS The parameter set was tested for global optimality with stochastic optimization based on PSO [209]. it can be assured that the period between overflow regime and substrate limitation is long enough to guarantee identifiability and detection of the switching to substrate limitation regime. the form of the general structure and the conditions of the process. following conditions are to be considered: 105 .5.An approach to Mechanism Recognition for model based analysis of Biological Systems M.2 CONDITIONS FOR ACCURATE SWITCHING POINT DETECTION The above mentioned assumptions assure a successful model description of the process using MR and a rigorous model with small systematic error. to allow a proper detection of the switching points in the process. it is assured that the initial regime of the process is growth under overflow conditions (after a short lag phase). THE SEQUENCE OF THE REGIMES IS KNOWN. INITIAL REGIME IS KNOWN Because of the high concentration of substrate at the beginning of the process.

AVAILABILITY OF AT LEAST ONE OBSERVATION REFLECTING REGIME CHANGE - At this stage (model validation) offline measurements are fitted and interpolated to obtain “online” measurements. In addition. It is worth reminding that the process of MR is strongly dependent of the capability of the complex model to mirror the real process. real time process monitoring is not considered at this stage. The results depicted in Figure 8. that the process offers enough information to detect the difference in bacterial behavior under both regimes. From this it may be supposed. However. 8. DEFINITIVENESS OF THE CHANGING BEHAVIOR - The distinguishability test proves that.54 mMol/l.14 suggest that the cultivation showed a typical E. 37°C Experiments G1 and G2 (replicate) were realized at fermentation temperature of 37°C with a precultivation of 16h. due to the nonlinearity of the model and unknown disturbances of the system. Again. SATISFACTORY STATE OF INFORMATION - The experiment was carried out in duplicate to assure reproducibility and a proper form of the variance-covariance matrix. the submodels are distinguishable. once the submodels have shown to describe the different regimes with precision and allow switching detection. the results obtained by the identifiability and distinguishability test cannot guarantee that the afore mentioned conditions are fulfilled. the physical foundation of the complex model offers a basis for evaluation of the dynamics of the process.5. coli batch cultivation and good reproducibility when compared against its replicate (G2). 16H. Nevertheless. The acetate concentration reached a maximal concentration at 23. During MR the data set of one experiment was utilized. Uptake of 106 .Mechanism Recognition in Escherichia coli cultivations KNOWLEDGE OF THE NORMAL BEHAVIOR OF THE SYSTEM - The typical behavior of the cell in each regime is known. identifiability and distinguishability tests should be repeated considering online measurements exclusively. at least for the simulated experiments. The measurements obtained from experiment G1 were selected for MR validation.3 DATA SET BATCH FERMENTATIONS.11-Figure 8.

An approach to Mechanism Recognition for model based analysis of Biological Systems

M. Nicolas Cruz Bournazou

acetate started at t = 5.68 h (341min). The specific acetate production rate fitted curve had its peak at t = 94 min with 9.9 mMol/(g*h). Acetate uptake started when the glucose concentration reached 0.5g/l. Glucose was completely metabolized at t = 6.7 h (400 min). The maximal glucose uptake rate reached 4.24 [g/l/h] at t = 1.8 h (109 min). The biomass curve showed the typical shape of growth in batch fermentation with E. coli. The lag-phase endured 2 h (120 min). Exponential growth phase was present until t = 6.6h (397 min) when maximum acetate concentration was measured. This point was also the onset of the stationary phase. The biomass yield had a maximal peak at t = 3h (184 min) at a value of 0.33 g/g. Specific intracellular NAD+ concentration reached 0.007 mMol/g at t = 4.1h (247 min). The respiration coefficient had its maximum at the same time. The respiration coefficient reached a minimum at t = 1.7 h (100 min) when the specific acetic acid production rate was maximal.

12 10

4,0 3,5 3,0

Glucose[g/l]

8 6 4 2 0 0 Glucosefit[g/l] D.Biofit[g/l] 200 time [min] GlucoseG1[g/l] BioG1[g/l] 400 GlucoseG2[g/l] D.BioG2[g/l] 600

2,5 2,0 1,5 1,0 0,5 0,0

Figure 8.11: Experimental results batch experiment G1. Part I: Dry biomass and glucose concentrations

107

Dry Biomass[g/l]

Mechanism Recognition in Escherichia coli cultivations

dAc/dt/D.Bio[mMol/gh]; dAc/dt[mMol/lh]

20 15 10 5 0 -5 0 200 dAc/dt/D.Biofit[mMol/g] dAc/dt/D.BioG2[mMol/g] Acfit[mMol/l] time [min] 400 dAc/dt/D.BioG1[mMol/gh] dAc/dtfit[mMol/lh] AcG1[mMol/l] 600 20

15

10

5

0

Figure 8.12: Experimental results batch experiment G1. Part II: Specific concentration of acetic acid

50 45

1,5 1,3 1,1 0,7 0,5 0,3 0,1 -0,1 -0,3 -0,5 0 qO2fit[mMol/gh] qCO2fit[mMol/gh] RQfit[-] 200 time [min] qO2G1[mMol/gh] qCO2G1[mMol/gh] 400 600 qO2G2[mMol/gh] qCO2G2[mMol/gh] RQfit [-] 0,9

qO2, qCO2 [mMol/gh]

40 35 30 25 20 15 10 5 0

Figure 8.13: Experimental results batch experiment G1. Part III: Outgas concentrations

108

Acetic acid [mMol/l]

An approach to Mechanism Recognition for model based analysis of Biological Systems
0,007 0,006

M. Nicolas Cruz Bournazou
60 50 40 30

NadH/D.Bio[mMol/l]

0,005 0,004 0,003 0,002 0,001 0,000 0 200 NadH/D.Biofit[mMol/g] Ac/D.BioG1[mMol/g] time [min] 400 Ac/Biofit[mMol/g] Ac/D.BioG2[mMol/g] 600

20 10 0

Figure 8.14: Experimental results batch experiment G1. Part IV: Metabolite concentration

8.5.4 RECOGNITION OF OVERFLOW AND SUBSTRATE
LIMITATION REGIMES
MR was applied considering the following measured variables: Extracellular acetate concentration (mMol/l) Extracellular substrate concentration (g/l) Biomass concentration (g/l)

INITIAL INTERVAL As previously discussed, the initial interval is characterized by overflow glucose uptake. The submodel adapted for this regime (OF), has a higher identifiability than the substrate limitation model. This is mainly because the submodel considers the substrate and oxygen uptake enzymes to be constant. MR INITIALIZATION Identifiability was obtained after 16 interpolated measurement points same as distinguishability. This permitted initiation of MR after 2.3 h. DETECTION OF SWITCHING POINTS Detection of the switch between overflow and substrate limitation regime was carried out with the method for fault detection through parameter identification. The constant threshold was set equal to a residual value of 0.3.

109

Ac/D.Bio[mMol/g]

Acetate A [mMol/l]. the physical evidence suggests the opposite. SWITCHING INTERVAL The experimental results confirmed that considering an instant switch of regime allows an accurate description of the process.15: OverFlow submodel fitted against experimental data. the program showed no problem to find an accurate solution.Mechanism Recognition in Escherichia coli cultivations INITIAL CONDITIONS OF THE INTERVAL K+1 Due to assumption of instant switch of the regime. fails to describe the initial phase of the cultivation Figure 8.16. need to be lower drastically. it is clear that the submodel lacks the capacity to describe the process after hour 7. like Yield from substrate to biomass for example. the simulation with the estimated parameters is depicted in Figure 8. EXPERIMENTAL DATA The submodel based on overflow metabolism assumption can fit the data obtained from the experiments with some accuracy. the submodel. Nonetheless.15. Still. Still. that the submodel can describe the first phase of the process with high accuracy. 110 . Contrary to the overflow submodel. which considers that all substrate can be consumed efficiently through the trcyclyc acid cycle. parameters.5. Still. it can be clearly seen. Besides. The effects of this “mistaken” assumption are buffered due to the estimation of the initial values of the substrate limitation regime (k+1).5 SIMULATIONS VS. In addition.5. Since the bacteria are not set to any real limitations for substrate uptake.5 approximately. which is reflected in the wrong description of substrate concentration form hour 5. the model consider an unreal growth and substrate consumption. no acetate is produced and the cultivation stops at hour 7. 8. Substrate S [g/l]. the initial values of the non measured variables had to be included in the parameter estimation problem.Biomass X [g/l] 10 8 6 OverFlow A X S 4 2 0 0 1 2 3 4 5 time [h] 6 7 8 9 10 Figure 8.

5. Substrate S [g/l]. Acetate A [mMol/l].An approach to Mechanism Recognition for model based analysis of Biological Systems SL 10 M.Biomass X [g/l] 10 8 6 4 2 0 Starvation A X S 0 1 2 3 4 5 time [h] 6 7 8 9 10 Figure 8. Finally. the starvation submodel was also tested Figure 8.6 RESULTS MR was applied considering the following measured variables: Extracellular acetate concentration (mMol/l) Extracellular substrate concentration (g/l) Biomass concentration (g/l) 111 . Substrate S [g/l].17.16: Submodel for the description of growth under substrate limitation fitted against experimental data.Biomass X [g/l] 8 X S 6 4 2 0 0 1 2 3 4 5 time [h] 6 7 8 9 10 Figure 8. 8.17: Starvation condition described by the corresponding submodel fitted against experimental data. This submodel has no possibility to describe the fermentation. Nicolas Cruz Bournazou A Acetate A [mMol/l]. The optimizer maximizes the yield coefficients to the parameter boundaries without achieving to describe the cultivation.

18. coli batch cultivations.8h (348 min). due to the decrease the process constant parameters in the general structure. Acetate A [mMol/l]. This results are consistent with the evidence shown by metabolite accumulation and acetate uptake t = 5. and the highly relaxed boundaries required.7 (341 min). 10 5 0 4 20 15 6 8 12 14 t[h] Batch fermentation 10 16 18 20 C(X) [gr/l] 10 5 0 0 2 4 6 8 10 t [h] 12 14 16 18 20 112 . MR detected the initiation of the regime of growth under substrate limitation at t = 5. Nevertheless.18: Experimental validation of the MR approach. this set of submodels cannot guarantee adequate recognition of fed-batch processes. MR enables the application of Lins model for the description of E. Substrate S [g/l].Biomass X [g/l] Batch Fermentation G1 10 8 6 4 2 0 0 1 2 3 4 5 time [h] 6 7 8 9 10 Figure 8.Mechanism Recognition in Escherichia coli cultivations The following model sequence was detected by MR Figure 8. A-Criteron 20 15 5 Par 4 Par 3 Par A-Crit.

the main carbon metabolism including the TCA as a generator of major branch point intermediates and products should be included. hence altering the concentration of enzymes of submetabolisms in a different way. it represents a major drawback for fed-batch applications.An approach to Mechanism Recognition for model based analysis of Biological Systems M.19. The reason for this is a likely change in the protein expression in the organism. When the model had been applied to fed-batch processes where switches of the glucose availability per cell were performed. would reduce the systematic error of the present model. standard deviation of 5% in all measurements To test the identifiability of the submodels with the new highly relaxed boundaries and the reduced general structure we calculate the FIM and evaluate the A-criterion in relation to the number of measurements. A complex model with a proper description capacity of the dynamic system is required. Nicolas Cruz Bournazou Figure 8. Also. MR is able to report whether the metabolism is dominated by overflow (acetate production) or by conversion of metabolites in the citric acid cycle. 8. Although this condition can be assured in batch cultivations. An improvement of the model considering at least a part of the regulatory behavior of the cell in a model which is based on kinetic rate equations. Second. an indicator of the current cell state is obtained. The first advantage showed by the application of MR is that the model sequence allows a reduction on the number of parameters to be fitted. However. the acetate synthesis should be modeled in detail including the excretion to the environment and the assimilation. the model has limitations in its application at processes characterized by a highly dynamic change in the cell physiology and genetic expression faster than in batch processes and in substrate-limited fed batch processes operated at rather constant conditions.19: Identifiability test considering white noise.6 CONCLUSIONS MR proved to be able to transform the model originally proposed by Lin into one suitable for parameter estimation and its application in batch processes. Since the intracellular acetate accumulation is regarded as being crucial and a reason for a change in the carbon 113 . A new model is required to build a set of submodels with global optimal structure and higher identifiability. The quality of the measurements was simulated with white noise (standard deviation 5%). The model should include the glycolysis as the different substrate supply is one characteristic of the different phases. the submodels were not suitable to reflect the time course of fermentation parameters (data not shown). The frequency considered for data recollection was 1 min with interpolated data set. The identifiability is reduced as we add parameters to the regime variable structure. The results show that the minimal period required to obtain a parameter estimation with less than 10% average standard deviation in all the parameters is more than 4 hours Figure 8. Finally. The yield coefficients (splits to different submetabolims) are changing.

the accumulation of intermediates. the catabolite repressor (Cra) . approaches in literature are rarely covering the complete main carbon metabolism including the submetabolisms mentioned previously. which in most cases cannot be measured accurately. is not considered. 189].7 FUTURE WORK So far. This assumption is suitable to reduce the set of parameters and to contribute to the fact that many regulatory patterns of the cells are still not understood and cannot be related to single reactions. these models are often applied at continuous cultivations of fed batch processes characterized by a high degree of stability concerning the environment of the cells in the bioreactor.Mechanism Recognition in Escherichia coli cultivations distribution in the metabolism. It is believed that the combination of the consideration of regulatory patterns. model reduction and mechanism recognition can contribute to the great need of process related models that are able to enhance the degree of control. Further regulatory mechanisms were included as the interaction of the cyclic AMP receptor protein. 114 . Again the reduction to create a new set of submodels is required. coli batch process. The model was applied for transient phases of an E. The information that can be gained from such methods seems to be very suitable to integrate them as activation or repression factors to submodels. the pentose-phosphate pathway. the glyoxylate shunt. A recent paper [210] describes the model fusion for the glycolysis. Many attempts are made to establish more robust on line and at line methods which are based on optical methods (on line microscopy) for the determination of cell size and shape. Each submodel is regarded as having several rate limiting reaction steps depending on the substrate and product concentrations present. and also electrooptical methods for measuring the polarizability of cells [172. the tricarboxylic acid cycle. Transient changes cannot be predicted but only in a very narrow range close to the conditions of parameter estimation. Another own approach combined a model of glycolysis [175] and the TCA cycle[211] and a module describing acetate synthesis and excretion and reconversion. the magnitude of these factors can be assumed separately for each phase of the process while reducing the number or rate equations needed to include all possible ways of interaction at transient cell stages. 8. So far. Also. the pyruvate dehydrogenase repressor and the repressor for the operon of acetate kinase. Besides the model-based contribution. This is introduced as a general description factor for each submodel. In combination with MR. but are characterized by a similar degree and change of gene expression. any suitable model should aim to predict the dynamics of acetate synthesis. also the integration of cell physiology measurements that are representing multiple parameters of the cells can be suitable to establish reduced models that increase control possibilities.

Secondly. Finally. The submodels created. As long as the physical background of the 115 . Firstly. a model for the description of E. the importance of considering the state of during model building is stated. if a model is simplified properly based on a systematic analysis and process know how. MR is able to detect depletion of organic matter allowing a better control of the SBR process for WWTPs increasing process efficiency and water quality. Nicolas Cruz Bournazou 9 CONCLUSIONS AND OUTLOOK 9. an additional advantage of process description with multiple submodels is a better understanding of the system.An approach to Mechanism Recognition for model based analysis of Biological Systems M. In both examples the advances of MR are clearly stated offering a systematic methodology for an improved study of the process and its mathematical representation. coli cultivation is analyzed and compared against experimental data. The approach to Mechanism Recognition represents a step forward in what can be considered the right direction toward the future of efficient modeling. In the first case. Attention is directed to make models useful and not as accurate or complex as possible. present a significantly higher identifiability and allow a better analysis of the complex model as well as the cultivation. it is shown how model simplification may lead to more efficient process description and better understanding of the system. it is shown that. a better understanding as well as a more accurate description of the system are enabled. This work proposes some insight to essential questions in biotechnology and chemical engineering: How accurate does a model need to be? How can we take advantage of the limitations of the model? What is the simplest manner to describe dynamic processes? Although a general answer to these questions is still not at hand. some important advances can be concluded from the results presented in this manuscript. In the second case.1 CONCLUSIONS The principal achievement of this thesis is the proposal of an alternative approach to analysis and implementation of complex rigorous models in processes with restricted state of information. Two case studies prove that the use of deep system understanding as the basis for development of simplified models for description of the process offers important advantages in comparison to model building through direct analysis of data correlations and heuristics. Moreover. accurate detection of organic matter depletion with standard measuring methods is enabled.

The method is still in a developing stage and further work is required to maximize efficiency of MR.2 OUTLOOK The advantages of MR for small and large scale problems have been presented and proved in two case studies. which enhance this collaboration. In a sense. 9. and ultimately process monitoring in an efficient and systematic manner. In order to create a complete and efficient MR program able to facilitate modeling. These simplified versions of the complex models offer a better process understanding and an increased applicability for engineering purposes than their complex versions. offers a handful of techniques to study the dynamics of models and increase its tractability. This represents a major drawback in MR application. In fact. Because of the nature of this project. Building of submodels. search for possible manners to perform it. Very sophisticated mathematical methods are available in literature. significant effort is required to build these reduced versions of the complex model. it is the scope of this thesis to motivate further research in this field.1 GENERAL THEORY FOR SUBMODEL GENERATION Model reduction theory. much time needs to be invested in the translation of codes and experimental information to create a common modeling environment. is not only time consuming but misses the software tools required for the scientists of different fields to collaborate in the creation of proper versions. special models can be developed to detect non desired conditions which cannot be measured directly.2. and find examples for its application. its real success can be quantified by the quality and extension of the outlook. 9. as it stands. 116 . structure analysis. further research is required. e. As shown in both case studies presented in this thesis.Conclusions and outlook submodels is understood. parameter estimation. The essence of this project was to test a new hypothesis. This characteristic can be exploited to detect non-measurable states of the process. Nevertheless. both case studies presented in this work prove that model reduction techniques in combination with process understanding may give rise to simpler but accurate submodels.g. an analysis of the description capacity of the submodels is also an indicator of the phenomenon dominating the process. Still. design of experiments. Despite important achievements in software tools. the real objective of this work may be understood as developing a realistic proposition of possible manners to develop efficient methods to MR. MOSAIC and SBPD toolbox.

9. for MR the difference between the quality of the data set obtained in laboratory experiments in comparison to considerably reduced.2. resulting not only in large efforts. Still. a simple but robust method taken from the field of model-based fault identification was applied to detect the time point were the switch of regimes takes place. global optimizers for nonlinear dynamic systems require large computer expenses and can by no means be applied for online purposes. 9. Still.2. Heuristics and scientist intuition are still the key to the reduction of such process. As previously stated. This information is very important to create an optimal set of submodels and is not consider in the state of the art methods for model reduction. e.2 SWITCHING POINT IDENTIFICATION At this early development stage. stochastic optimization methods cannot assure convergence to global optimality. in nonlinear systems under dynamic conditions a general approach to reduce large nonlinear systems with reasonable computer expenses is still missing. should also be tested since they offer alternative approaches for efficient parameter estimation and accurate regime detection in processes with longer time spans.An approach to Mechanism Recognition for model based analysis of Biological Systems M. In this work a stochastic optimization method was applied in an effort to avoid local minima and increase the probability of global solution. Furthermore. accurately detecting the change of regimes. two main aspects of model reduction require further development to reduce the effort of submodel building: 1) Model reduction of dynamic nonlinear systems. Nevertheless. 2) Model reduction based on the state of information. Use of global optimizers is essential in model structure analysis. Present approaches for model reduction do not take the state of information of the process into account.3 GLOBAL OPTIMIZATION In the case of highly nonlinear models. but also in suboptimal and locally restricted solutions with no general validity. design of 117 . The parameter identification method showed to fulfill the requirements at this stage of development. this stage offers large optimization potential. global optimizers should be implemented at the early stages of MR. model structure analysis and parameter estimation require proof of global convergence. A study of modern techniques and a proper adaptation for the case of MR is required to offer an efficient switching point identification approach. Nicolas Cruz Bournazou In addition. Advanced methods of model predictive control.g. At its current development stage. moving horizon.

Still. 118 . offers a great field of application. An example is E. From this. In particular.4 ONLINE MONITORING Finally.Conclusions and outlook experiments. The present approach has been developed taking into account limitations of online applications in all stages. indirect measurements in combination with the correct set of submodels enable the detection of a change in the state of the cells optimizing feed strategies to maximize cell density and recombinant protein production. but the data obtained does not offer direct information about the state of the “optimal” regime.2. It is only after knowing the global best solution for each case that proper conclusion can be made. coli cultivations where optimal growth is known to occur under substrate limitation where the state of the cell cannot be measured directly. it can be assumed that MR can be applied to industrial processes where a certain regime must be maintained to assure optimal process conditions. 9. and model distinguishability studies. which cannot be detected by standard measurements. detection of quasi steady state regimes. this work proves the great potential on the application of MR to enhance monitoring of none measurable variables in dynamic processes.

Nicolas Cruz Bournazou 10 Appendix A. COLI 𝑑𝑉 = 𝐹 𝑑𝑡 𝑑𝑋 𝐹 = − + 𝜇 𝑋 𝑑𝑡 𝑉 𝑑𝑆 𝐹 = − 𝑆 − 𝑆𝑖 − 𝑞𝑆 ∗ 𝑋 𝑑𝑡 𝑉 𝑑𝐷𝑂𝑇 = 𝑘𝐿𝑎 𝐷𝑂𝑇 ∗ − 𝐷𝑂𝑇 − 𝑞𝑂 ∗ 𝑋 ∗ 𝐻 𝑑𝑡 𝑑𝐴𝑐 𝐹 = − 𝐴 + 𝑞𝐴𝑐 𝑝 − 𝑞𝐴𝑐 𝑐 𝑋 𝑑𝑡 𝑉 𝑑𝑟𝑆 = 𝑌𝐸𝑆 /𝑋 ∗ 𝜇 − 𝑟𝑆 (𝑞𝐸𝑆𝑑 + 𝜇) 𝑑𝑡 𝑑𝑟𝑂 = 𝑌𝐸𝑂 /𝑋 ∗ 𝜇 − 𝑟𝑂 (𝑞𝐸𝑂𝑑 + 𝜇) 𝑑𝑡 The initial values 𝑟𝑆0 = 𝑐𝑆 ∗ 𝑟𝑂0 = 𝑐𝑂 ∗ 𝜇𝑚𝑎𝑥 𝑞𝐸𝑆𝑑 + 𝜇𝑚𝑎𝑥 𝜇𝑚𝑎𝑥 𝑞𝐸𝑂𝑑 + 𝜇𝑚𝑎𝑥 A8 A1 MODEL AS PUBLISHED BY LIN [91] A2 A3 A4 A5 A6 A7 A9 A 10 𝜇𝑚𝑎𝑥 = 𝑞𝑆𝑚𝑎𝑥 − 𝑞𝑚 𝑌𝑥/𝑆𝑜𝑥 Calculation of the subsidiary variables: 𝑞𝑆𝑐𝑎𝑝 = 𝑞𝑆𝑚𝑎𝑥 𝑞𝑂𝑐𝑎𝑝 = 𝑞𝑂𝑚𝑎𝑥 𝑟𝑆 𝑟𝑆0 𝑟𝑂 𝑟𝑂0 A 11 A 12 119 .An approach to Mechanism Recognition for model based analysis of Biological Systems M. KINETIC MODEL FOR E.

Appendix 𝑞𝑆 = 𝑞𝑆𝑐𝑎𝑝 𝑆 𝑆 + 𝐾𝑆 A 13 A 14 A 15 A 16 A 17 A 18 𝑞𝑆𝑂𝑥 = 𝑞𝑆 𝑞𝑚𝑟 = min[𝑞𝑚 . then 𝑞𝑚𝑟 = 𝑞𝑆𝑜𝑥 = 𝑞𝑆𝑜𝑥𝑒𝑛 and 𝑞𝑆𝑜𝑥𝑎𝑛 = 0 𝑞𝑆𝑜𝑓 = 𝑞𝑆 − 𝑞𝑆𝑜𝑥 𝑞𝐴𝐶 = 𝑞𝐴𝑐𝑚𝑎𝑥 𝐴𝑐 𝐴𝑐 + 𝐾𝐴𝑐 A 22 A 23 A 24 A 25 A 26 A 27 𝑞𝑆𝑜𝑓𝑎𝑛 = 𝑞𝑆𝑜𝑓 ∗ 𝑌𝑋/𝑆𝑜𝑓 ∗ 𝐶𝑆 𝑞𝑆𝑜𝑓𝑒𝑛 = 𝑞𝑆𝑜𝑓 − 𝑞𝑆𝑜𝑓𝑎𝑛 𝑞𝐴𝑐𝑎𝑛 = 𝑞𝐴𝑐 ∗ 𝑌𝑋/𝐴 ∗ 𝐶𝐴𝑐 𝑞𝐴𝑐𝑒𝑛 = 𝑞𝐴𝑐 −𝑞𝐴𝑐𝑎𝑛 Finally. whenever 𝑞𝑂𝑆 > 𝑞𝑂𝑐𝑎𝑝 we change the values to 𝑞𝑂𝑆 = 𝑞𝑂𝑐𝑎𝑝 and recalculate: 𝑞𝑆𝑜𝑥𝑒𝑛 = 𝑞𝑆𝑜𝑥 = 𝑞𝑂𝑐𝑎𝑝 𝑌𝑂/𝑆 A 19 𝑞𝑚𝑟 ∗ 𝐶𝑆 ∗ 𝑌𝑋/𝑆_𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑒𝑛 𝐶𝑆 ∗ 𝑌𝑂/𝑆𝑜𝑥 − 1 A 20 A 21 𝑞𝑆𝑜𝑥𝑎𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑒 𝑛 Because of the recalculations. the physical limitation of the real maintenance coefficient has to be checked once more. 𝑞𝐴𝑐𝑒𝑛 = 𝑞𝑂𝑐𝑎𝑝 − 𝑞𝑂𝑆 /𝑌𝑂𝐴 and: 𝑞𝐴𝑐𝑐 = 𝑞𝐴𝑐𝑒𝑛 1 − 𝑌𝑋/𝐴 ∗ 𝐶𝐴𝑐 A 28 120 . If 𝑞𝑚𝑟 > 𝑞𝑆𝑜𝑥 . 𝑞𝑆𝑜𝑥 ] 𝑞𝑆𝑜𝑥𝑎𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑥/𝑆𝑜𝑥 ∗ 𝐶𝑆 𝑞𝑆𝑜𝑥𝑒𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑎𝑛 𝑞𝑂𝑆 = 𝑞𝑆𝑜𝑥 𝑒𝑛 ∗ 𝑌𝑂/𝑆 To avoid a larger 𝑞𝑂𝑆 than allowed by the maximal oxygen uptake capacity of the bacteria. if 𝑞𝐴𝑐𝑒𝑛 > 𝑞𝑂𝑐𝑎𝑝 − 𝑞𝑂𝑆 /𝑌𝑂𝐴 then.

000 L g-1] saturation constant [g L-1] volumetric oxygen transfer coefficient [h-1] specific acetate production of consumption rate [g g-1 h-1] 𝑞𝑖 specific degradation rate [g g-1 h-1] 𝑞𝑂 specific oxygen uptake rate [g g-1 h-1] 𝑞𝑆 specific substrate (glucose) uptake rate [g g-1h-1] 𝑟𝑂 . Nicolas Cruz Bournazou A 29 A 30 A 31 A 32 𝜇 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑋/𝑆𝑜𝑥 + 𝑞𝑆𝑜𝑓 ∗ 𝑌𝑋/𝑆𝑜𝑓 + 𝑞𝐴𝑐 𝑐 ∗ 𝑌𝑋/𝐴𝑐 Nomenclature Subscripts 0 𝐴𝑐 𝑎𝑛 𝑐 𝑐𝑎𝑝 𝑑 𝐸 𝑒𝑛 Initial concentration acetate anabolic consumption capacity degradation enzyme energetic 𝜇 𝐴𝑐 𝐷𝑂𝑇 𝐸𝑆 . 𝐸𝑂 [gL−1] 𝐹 𝐻 𝐾 𝑘𝐿𝑎 𝑞𝐴𝑐 specific growth rate [h-1 ] acetate concentration [g L-1] dissolved oxygen tension [%] concentration of the enzymes substrate/oxygen consumption feed flow rate [L h-1] constant derived from the Henry‟s law [14.An approach to Mechanism Recognition for model based analysis of Biological Systems 𝑞𝐴𝑐𝑎𝑛 = 𝑞𝐴𝑐 −𝑞𝐴𝑐𝑎𝑛 𝑞𝐴𝑝 = 𝑞𝑆𝑜𝑓𝑒𝑛 ∗ 𝑌 𝑐 /𝑆 𝐴 𝑞𝑂 = 𝑞𝑂𝑆 +𝑞𝐴𝑐𝑒𝑛 ∗ 𝑌𝑂/𝐴𝑐 M. 𝑟𝑆 ratio of enzyme for oxygen or substrate consumption per biomass 𝑆 substrate (glucose) concentration [g L-1] 𝑡 cultivation time [h] 𝑉 fermenter volume [L] 𝑋 biomass concentration [g L-1] 𝑌 yield [g g-1] 121 .

Appendix - 𝑖 𝑚 𝑚𝑎𝑥 𝑚𝑟 𝑂 𝑜𝑓 𝑜𝑥 𝑝 𝑆 𝑋 inlet concentration maintenance maximum real maintenance oxygen overflow oxidative production substrate (glucose) biomass 122 .

An approach to Mechanism Recognition for model based analysis of Biological Systems M. Nicolas Cruz Bournazou The nomenclature is identical to the complete Lin model (appendix A) 𝑑𝑉 = 𝐹 𝑑𝑡 𝑑𝑋 𝐹 = − + 𝜇 𝑋 𝑑𝑡 𝑉 𝑑𝑆 𝐹 = − 𝑆 − 𝑆𝑖 − 𝑞𝑆 ∗ 𝑋 𝑑𝑡 𝑉 𝑑𝐷𝑂𝑇 = 𝑘𝐿𝑎 𝐷𝑂𝑇 ∗ − 𝐷𝑂𝑇 − 𝑞𝑂 ∗ 𝑋 ∗ 𝐻 𝑑𝑡 𝑑𝐴 𝐹 = − 𝐴𝑐 + 𝑞𝐴𝑐 𝑝 − 𝑞𝐴𝑐 𝑐 𝑋 𝑑𝑡 𝑉 𝑟𝑆 = 𝑐𝑆 ∗ 𝑟𝑂 = 𝑐𝑂 ∗ The initial values 𝜇𝑚𝑎𝑥 = 𝑞𝑆𝑚𝑎𝑥 − 𝑞𝑚 𝑌𝑥/𝑆𝑜𝑥 CALCULATION OF THE SUBSIDIARY VARIABLES 𝑞𝑆𝑐𝑎𝑝 = 𝑞𝑆𝑚𝑎𝑥 𝑞𝑂𝑐𝑎𝑝 = 𝑞𝑂𝑚𝑎𝑥 𝑞𝑆 = 𝑞𝑆𝑐𝑎𝑝 𝑟𝑆 𝑟𝑆0 𝑟𝑂 𝑟𝑂0 𝜇𝑚𝑎𝑥 𝑞𝐸𝑆𝑑 + 𝜇𝑚𝑎𝑥 𝜇𝑚𝑎𝑥 𝑞𝐸𝑂𝑑 + 𝜇𝑚𝑎𝑥 B. we assume: 𝑞𝑂𝑆 = 𝑞𝑂𝑐𝑎𝑝 : 𝑞𝑆𝑜𝑥𝑒𝑛 = 𝑞 𝑂 𝑐𝑎𝑝 𝑌𝑂/𝑆 B 14 123 . OVERFLOW SUBMODEL B1 B2 B3 B4 B5 B6 B7 B8 B9 B 10 𝑆 𝑆 + 𝐾𝑆 B 11 B 12 B 13 𝑞𝑆𝑂𝑥 = 𝑞𝑆 𝑞𝑚𝑟 = 𝑞𝑚 Because of the overflow condition.

Appendix 𝑞𝑆𝑜𝑥 = 𝑞𝑚𝑟 ∗ 𝐶𝑆 ∗ 𝑌𝑋/𝑆_𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑒𝑛 𝐶𝑆 ∗ 𝑌𝑂/𝑆𝑜𝑥 − 1 B 15 B 16 B 17 B 18 B 19 B 20 B 21 B 22 B 23 𝑞𝑆𝑜𝑥𝑎𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑒𝑛 𝑞𝑆𝑜𝑓 = 𝑞𝑆 − 𝑞𝑆𝑜𝑥 𝑞𝐴𝑐 𝑐 = 𝑞𝐴𝑐 𝑐 𝑚𝑎𝑥 𝐴𝑐 𝐴𝑐 + 𝐾𝐴𝑐 𝑞𝑆𝑜𝑓𝑎𝑛 = 𝑞𝑆𝑜𝑓 ∗ 𝑌𝑋/𝑆𝑜𝑓 ∗ 𝐶𝑆 𝑞𝑆𝑜𝑓𝑒𝑛 = 𝑞𝑆𝑜𝑓 − 𝑞𝑆𝑜𝑓𝑎𝑛 𝑞𝐴𝑐 𝑐𝑎𝑛 = 𝑞𝐴𝑐 𝑐 ∗ 𝑌𝑋/𝐴𝑐 ∗ 𝐶𝐴𝑐 𝑞𝐴𝑐 𝑐𝑒𝑛 = 𝑞𝐴𝑐 𝑐 −𝑞𝐴𝑐 𝑐𝑎𝑛 𝜇 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑋/𝑆𝑜𝑥 + 𝑞𝑆𝑜𝑓 ∗ 𝑌𝑋/𝑆𝑜𝑓 + 𝑞𝐴𝑐 𝑐 ∗ 𝑌𝑋/𝐴𝑐 124 .

SUBSTRATE LIMITATION SUBMODEL The nomenclature is identical to the complete Lin model (appendix A) 𝑑𝑉 = 𝐹 𝑑𝑡 𝑑𝑋 𝐹 = − + 𝜇 𝑋 𝑑𝑡 𝑉 𝑑𝑆 𝐹 = − 𝑆 − 𝑆𝑖 − 𝑞𝑆 ∗ 𝑋 𝑑𝑡 𝑉 𝑑𝐷𝑂𝑇 = 𝑘𝐿𝑎 𝐷𝑂𝑇 ∗ − 𝐷𝑂𝑇 − 𝑞𝑂 ∗ 𝑋 ∗ 𝐻 𝑑𝑡 𝑑𝐴𝑐 𝐹 = − 𝐴𝑐 + 𝑞𝐴𝑐 𝑝 − 𝑞𝐴𝑐 𝑐 𝑋 𝑑𝑡 𝑉 𝑑𝑟𝑆 = 𝑌𝐸𝑆 /𝑋 ∗ 𝜇 − 𝑟𝑆 (𝑞𝐸𝑆𝑑 + 𝜇) 𝑑𝑡 𝑑𝑟𝑂 = 𝑌𝐸𝑂 /𝑋 ∗ 𝜇 − 𝑟𝑂 (𝑞𝐸𝑂𝑑 + 𝜇) 𝑑𝑡 The initial values 𝑟𝑆0 = 𝑐𝑆 ∗ 𝑟𝑂0 = 𝑐𝑂 ∗ 𝜇𝑚𝑎𝑥 𝑞𝐸𝑆𝑑 + 𝜇𝑚𝑎𝑥 𝜇𝑚𝑎𝑥 𝑞𝐸𝑂𝑑 + 𝜇𝑚𝑎𝑥 C1 C2 C3 C4 C5 C6 C7 C8 C9 C 10 𝜇𝑚𝑎𝑥 = 𝑞𝑆𝑚𝑎𝑥 − 𝑞𝑚 𝑌𝑥/𝑆𝑜𝑥 Calculation of the subsidiary variables: 𝑞𝑆𝑐𝑎𝑝 = 𝑞𝑆𝑚𝑎𝑥 𝑞𝑂𝑐𝑎𝑝 = 𝑞𝑂𝑚𝑎𝑥 𝑟𝑆 𝑟𝑆0 𝑟𝑂 𝑟𝑂0 C 11 C 12 125 . Nicolas Cruz Bournazou C.An approach to Mechanism Recognition for model based analysis of Biological Systems M.

Appendix 𝑞𝑆 = 𝑞𝑆𝑐𝑎𝑝 𝑆 𝑆 + 𝐾𝑆 C 13 C 14 C 15 C 16 C 17 C 18 C 19 C 20 C 21 C 22 C 23 C 24 C 25 C 26 C 27 𝑞𝑆𝑂𝑥 = 𝑞𝑆 𝑞𝑚𝑟 = 𝑞𝑚 𝑞𝑆𝑜𝑥𝑎𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑥/𝑆𝑜𝑥 ∗ 𝐶𝑆 𝑞𝑆𝑜𝑥𝑒𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑎𝑛 𝑞𝑂𝑆 = 𝑞𝑆𝑜𝑥𝑒𝑛 ∗ 𝑌𝑂/𝑆 𝑞𝑆𝑜𝑓 = 0 𝑞𝐴𝑐 𝐶 = 𝑞𝐴𝑐𝑐𝑚𝑎𝑥 𝐴𝑐 𝐴𝑐 + 𝐾𝐴𝑐 𝑞𝑆𝑜𝑓𝑎𝑛 = 0 𝑞𝑆𝑜𝑓𝑒𝑛 = 0 𝑞𝐴𝑐 𝑐𝑎𝑛 = 𝑞𝐴𝑐 𝑐 ∗ 𝑌𝑋/𝐴𝑐 ∗ 𝐶𝐴𝑐 𝑞𝐴𝑐 𝑐𝑒𝑛 = 𝑞𝐴𝑐 𝑐 −𝑞𝐴𝑐 𝑐𝑎𝑛 𝑞𝐴𝑐 𝑝 = 0 𝑞𝑂 = 𝑞𝑂𝑆 +𝑞𝐴𝑐 𝑐𝑒𝑛 ∗ 𝑌𝑂/𝐴𝑐 𝜇 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑋/𝑆𝑜𝑥 + 𝑞𝑆𝑜𝑓 ∗ 𝑌𝑋/𝑆𝑜𝑓 + 𝑞𝐴𝑐 𝑐 ∗ 𝑌𝑋/𝐴𝑐 126 .

An approach to Mechanism Recognition for model based analysis of Biological Systems M. STARVATION SUBMODEL 𝑑𝑉 = 𝐹 𝑑𝑡 𝑑𝑋 𝐹 = − + 𝜇 𝑋 𝑑𝑡 𝑉 𝑑𝑆 𝐹 = − 𝑆 − 𝑆𝑖 − 𝑞𝑆 ∗ 𝑋 𝑑𝑡 𝑉 𝑑𝐷𝑂𝑇 = 𝑘𝐿𝑎 𝐷𝑂𝑇 ∗ − 𝐷𝑂𝑇 − 𝑞𝑂 ∗ 𝑋 ∗ 𝐻 𝑑𝑡 𝑑𝐴𝑐 𝐹 = − 𝐴𝑐 + 𝑞𝐴𝑐 𝑝 − 𝑞𝐴𝑐 𝑐 𝑋 𝑑𝑡 𝑉 𝑟𝑆 : 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑 𝑟𝑂 : 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑 The initial values 𝜇𝑚𝑎𝑥 = 𝑞𝑆𝑚𝑎𝑥 − 𝑞𝑚 𝑌𝑥/𝑆𝑜𝑥 Calculation of the subsidiary variables: 𝑞𝑆𝑐𝑎𝑝 = 𝑞𝑆𝑚𝑎𝑥 𝑞𝑂𝑐𝑎𝑝 = 𝑞𝑂𝑚𝑎𝑥 𝑞𝑆 = 𝑞𝑆𝑐𝑎𝑝 𝑟𝑆 𝑟𝑆0 𝑟𝑂 𝑟𝑂0 The nomenclature is identical to the complete Lin model (appendix A) D1 D2 D3 D4 D5 D6 D7 D8 D9 D 10 𝑆 𝑆 + 𝐾𝑆 D 11 D 12 D 13 D 14 D 15 𝑞𝑆𝑂𝑥 = 𝑞𝑆 𝑞𝑚𝑟 = 𝑞𝑚 𝑞𝑆𝑜𝑥𝑎𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑥/𝑆𝑜𝑥 ∗ 𝐶𝑆 𝑞𝑆𝑜𝑥𝑒𝑛 = 𝑞𝑂𝑐𝑎𝑝 𝑌𝑂/𝑆 127 . Nicolas Cruz Bournazou D.

Appendix 𝑞𝑆𝑜𝑥 = 𝑞𝑚𝑟 ∗ 𝐶𝑆 ∗ 𝑌𝑋/𝑆_𝑜𝑥 − 𝑞𝑆𝑜𝑥𝑒𝑛 𝐶𝑆 ∗ 𝑌𝑂/𝑆𝑜𝑥 − 1 D 16 D 17 D 18 D 19 D 20 D 21 D 22 D 23 D 24 𝑞𝑆𝑜𝑥𝑎𝑛 = 𝑞𝑆𝑜𝑥 − 𝑞𝑆𝑜𝑥 𝑒𝑛 𝑞𝑆𝑜𝑓 = 0 𝑞𝐴𝑐 𝑐 = 0 𝑞𝑆𝑜𝑓𝑎𝑛 = 0 𝑞𝑆𝑜𝑓𝑒𝑛 = 0 𝑞𝐴𝑐 𝑐𝑎𝑛 = 0 𝑞𝐴𝑐 𝑐𝑒𝑛 = 0 𝜇 = 𝑞𝑆𝑜𝑥 − 𝑞𝑚𝑟 𝑌𝑋/𝑆𝑜𝑥 + 𝑞𝑆𝑜𝑓 ∗ 𝑌𝑋/𝑆𝑜𝑓 + 𝑞𝐴𝑐 𝑐 ∗ 𝑌𝑋/𝐴𝑐 128 .

Lyberatos. S. H. coli Fed-Batch-Kultivierungen M.0047 Neuer Ansatz zur Mechanismenerkennung in E.201050496 ASM3 extended for two-step nitrification-denitrification: a model reduction for Sequencing Batch Reactors M. N. C.. Arellano-Garcia.1002/cite. Nicolas Cruz Bournazou BIBLIOGRAPHY PERSONAL PUBLICATIONS Electrooptical monitoring of cell polarizability and cell size in aerobic Escherichia coli batch cultivations Stefan Junne. Nicolas Cruz-Bournazou. Arellano-Garcia. Issue 9.An approach to Mechanism Recognition for model based analysis of Biological Systems M. N. Cruz Bournazou M. J. N. G. Lyberatos Prof.DOI: 10.N. 935-942.201050496 Modellreduktion des erweiterten ASM3-Modells für die zweistufige Nitrifikation M. Cruz Bournazou.. Wozny Computer Aided Chemical Engineering Volume 27. 2009. 291–306 A Novel Approach to Mechanism Recognition in Escherichia Coli Fed-Batch Fermentations M. G.1007/s10295-010-0742-5 From data to mechanistic-based models M. . Arellano-Garcia. Cruz Bournazou. Issue 9. H. Chemie Ingenieur Technik Volume 82. Schöneberger.-Ing. N. September. Cruz Bournazou H. Alexander Angersbach and Peter Götz Journal of Industrial Microbiology & Biotechnology Volume 37. Kravaris Prof. M. Arellano-Garcia G.. 30. Neubauer and G. 2010.C. G. P. Wozny Chemical and Process Engineering 2009. DOI: 10. Schöneberger. Kravaris Journal of Chemical Technology and Biotechnology (in press: JCTB 11-0602) 129 . 2010. C.3182/20100707-3-BE-2012. Cruz Bournazou. Junne. G. page 1540. N. September. S.1002/cite. Neubauer Chemie Ingenieur Technik Volume 82. Cruz-Bournazou. Wozny. page 1503. Kravaris Computer Applications in Biotechnology Volume11 Part 1. Arellano-Garcia Dr. G.DOI: 10. G. Junne. Number 9. Wozny. P. 10. Lyberatos. J. G. Sc. Wozny Prof. Pages 651-656 10th Model reduction of the ASM3 extended for two-step nitrification M. H. Arellano-Garcia. C. Wozny. H. H.

Barton. John Wiley and Sons. Computational systems biology. Marquardt. P. Balsa-Canto. 2. Computers & Chemical Engineering. Computer Aided Chemical Engineering. Peterson. 8. New York. 1989.X.T. H. Baldazzi. Applied engineering in agriculture (USA). Jones. Donev. Res.L. Choe. 18. Linninger. Li. 22. Hjalmarsson. University of London. Bonvin. Kitano.C. 4(1): p. 1987.P. 2009.O.. 1-13.. 1691-1724. and A. and J. Brendel.S.J. 13(1-2): p.Bibliography REFERENCES 1. 11. P.E. 24. H. Computers & chemical engineering. Wright. Automatica.L. Rippin. 57-71.org/ViewItem. Systems Biology: A Brief Overview. Floudas. Y. M.R. Frank. Atkinson. 2000. 9. F. Empirical model building and response surfaces. 73-82. A. Optimal Catheter Placement for Chemotherapy. 16.I.T. 26(3): p. Optimum Experimental Designs. 20. 3. Banga. E. Luyben. de Jong. 2002. Juditsky. Guidance for Industry Process Validation: General Principles and Practice. 455-492. 110(3): p. 2003: p.. Sindhwani.K. 1998: Kluwer Academic Publishers. P. and A.. Society for Industrial Mathematics.. 1995. Benveniste.. Esposito.F. Ropers. 34(9): p. Glorennec. Q. 39(5): p. 459-474. 2006. R.W. 2011. 1992: Oxford University Press. Automatica. W. Nonlinear Modeling: Advanced Black-Box Techniques.J. 2158-2161. and J.T. and C. Kitano. D.R. Nonlinear controllers for trajectory tracking in batch processes. Incremental identification of kinetic models for homogeneous reaction systems. 420(6912): p. Rigorous dynamic models of distillation columns.R.J.. Inverse Problem Theory and Methods for Model Parameter Estimation. Zhang. 14. 21. Batch process systems engineering: a retrospective and prospective review. An iterative identification procedure for dynamic modeling of biochemical networks. 13(4): p. 85–104. J. IEEE IEEE/ACM Transactions on Computational Biology and Bioinformatics. http://www. Welch. 2010.L.A.. P. O. 1543-1555. DFG-Jahresbericht 2010. Carrier. Nonlinear Black-Box Modeling in System-Identification-A Unified Overview. Model Reduction Using Piecewise-Linear Approximations Preserves Dynamic Properties of the Carbon Starvation Response in Escherichia coli.html DFG. BASF erhöht Forschungsausgaben ed. D.ftd. Nature. 1291-1310. and W. and A. 17: p. 61(16): p. 23. and W. 9(1): p. Draper. and N. 19.A. 1998.. Schonlau. 25. 5. 15. USA. 1987.Y.H. M. B. Eng. Gep. 10.gov. fda. http://www. Technometrics. 1990. 28: p. BMC Systems Biology. and W. Industrial & Engineering Chemistry Research. Computers & chemical engineering. 2005. 130 . D. C. p.M. Dynamic Model Development. 4. Sjoberg. 1993. A.. G.R.A. 2002. V. Ivanchenko. Model Selection: an overview of practices in chemical engineering. 12. C. 1967. The modelling and simulation of combined discrete/continuous processes... A. 26(10): p. H. Kravaris. 1991. Preisig.. Vandewalle. M.. 971-976. Delyon. Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some new results. Discrimination among mechanistic models. 1662-1664. B. L.aspx?ItemId=108368&CultureCode=de. Ind. Global optimization for the parameter estimation of differential-algebraic systems.. ed. 1992. D. 2011. 13. and H. 31(12): p. and W.alphagalileo. 206-210. Batch type transesterification process for winter rape oil. H. Chem.de/unternehmen/industrie/:investitionen-beim-chemiekonzern-basferhoeht-forschungsausgaben/60007679. Deutschland.A. R. Alonso. Constructing and maintaining proper process models. A. and D. Verheijen. Auld. 2010. 6. Feldman.A. Journal of Global Optimization. Box. J. Lueshen. FTD. E. Ljung. 17. Chemical Engineering Science. N. 5404-5420. 223-228. Tarantola. Korus. Efficient Global Optimization of Expensive BlackBox Functions. 11. D.. 7.N. Hill. Suykens. PhD diss.. 2010. and J.A.

30(2): p. Schoneberger. 35. Chinese Journal of Chemical Engineering.A. D... Green.M.. Schmidt. H. V. T. 2007.1016/j. Quevedo Casin.10. 41.D. and G. 216223. 42.W. Qualitative Process Theory. 32. R. 1988. Husar. P. Barz. 44. De Lira. and T.. K. 192(1): p. Bar-Shalom.008. Wozny. Blom. Oxberry.. V. P. Arellano-Garcia. 11(1).H. G. P. Herold. 15-123. and G. 1994. T. De Lira Ramirez. and W.I. A. Barton. H. H. Berleant. Ochs. 1998. IEEE Transactions on. and W. 95(2): p. The Journal of Physical Chemistry A. Utkin. 35 (11): p. J. 24(6): p. 12 Kuntsche. Interacting multiple model methods in target tracking: a survey. W. Escobet. Qualitative process theory: twelve years after. and G. MOSAIC a webbased modeling environment for co de generation. A.I. and B. Optimally-reduced kinetic models: reaction elimination in large-scale kinetic mechanisms. 191-208. Aerospace and Electronic Systems. and V.2010. A. S. Wozny. Dano.. Barz. Biotreatment. Optimal automatic reaction and species elimination in kinetic mechanisms. B. 780-783.. atp edition. Krishnamurthy. Überwachung von Anlagenteilen: Neues Werkzeug ermöglicht Statusermittlung.2257-2273. 155(1-2): p. and C. IEEE Transactions on. Computer Applications in Biotechnology. Wozny. 33(8): p. 2008. and Y. Combustion and Flame.doi. IEEE Transactions on. 131 . Kraus. 45.. S. 147197. 39. 291-306. and J.J. Modelbased fault diagnosis in PEM fuel cell systems. http://dx. D. Wozny. Wu. 34.J. T. The interacting multiple model algorithm for systems with Markovian switching coefficients.P. Artificial Intelligence. 22(2): p. and W. Qualitative Algebra and Graph Theory Methods for Dynamic Trend Analysis of Continuous System. Puig. J. analysis. S. 110(3): p. 33. Arellano-Garcia. K. An efficient sparse approach to sensitivity generation for large-scale dynamic optimization. IEEE Transactions on.. Quevedo. King. Zhang. C. Forbus. 2010. 2011. Bar-Shalom. 971-976. A. and M. S.. 37. Dynamical modelling. M. 1997. Automatic Control. Kuipers. 31. Automatic Control. Mazor. Serra. Heine. J. D. 215-255. Cedersund.. 28. Forbus. 85-168. 613-624. Bhattacharjee. and G.. Complexity reduction of biochemical rate expressions. Hotop. 40. E. and M. Y. N. Global dynamic optimization for parameter estimation in chemical kinetics. H. Green. Schwer. Variable structure systems with sliding modes. monitoring and control design for nonlinear bioprocesses. V. Bioinformatics.. A Novel Approach to Automated Mechanism Recognition based on Model Discrimination. S. Green. Wang. Chemical and Process Engineering Selected full texts. Averbuch. 135(3): p.J. Dayan.A. 49(3): p. Computers and Chemical Engineering.. Mitsos. Journal of Power Sources. Arellano-Garcia. J.An approach to Mechanism Recognition for model based analysis of Biological Systems 26. Computers & Chemical Engineering. Signal Processing. Doucet. M. H. 103-123. Perrier. 24(1-3): p. T. 2001. Artificial Intelligence. and R. 2010. 27. Gordon. R.. Riera.. S. Taylor. 2009. 46. 2011. D. 30. Cruz-Bournazou. Qualitative and quantitative simulation: bridging the gap. 848. 2010: IEEE Computer Society. Ross. From data to mechanistic-based models. 1997: p. 19(2): p. 59(1): p. and J.F. 36. 212-222. 29. 2008. 43. An Automated Approach to Build Process Models by Detecting Biological Phenomena in (fed-) batch Experiments. 308-315. Madsen..I. 2009.compchemeng. Combustion and Flame. 38. G. 1977. J. Schöneberger. 118–132. Nicolas Cruz Bournazou Singer.. 2010.org/10. LPV modelbased fault diagnosis using relative fault sensitivity signature approach in a PEM fuel cell.D. Barton.H. 1984.B. Kuntsche. and H. Puig Cayuela. Particle filters for state estimation of jump Markov linear systems. 2003. Arellano-Garcia. 34(1): p. Dochain. S. Artificial Intelligence in Perspective. 2006. Chemical Engineering Transactions. Barton. Downstream Processing and Modelling. A. M.H. Feroldi.P.

1249-1266. Berlin: s.J. Discrimination and goodness of fit of multiresponse mechanistic models. Marquardt. Chemical Engineering Journal. J. M. 4732. 14: p. L. and R. 254 S. 59. Stamatelatou. Systems biology: a brief overview. 88: p. 2009. 35. Box. Reverse engineering of biological complexity.. H. F. C. Capodaglio. 9(3): p. Streichert. and Schering (Berlin West). Advances in metabolic control analysis.G. Rev. Vol. and G. Guthke. Invariant manifold methods for metabolic model reduction. Raduly. 70. F. Statistische Methoden in der experimentellen Forschung Kolloquium. BMC systems biology. AIChE journal.S. 72. Hady. 2010. M. J. Science.W. 53. 86103. Toepfer. 261-282. 1998. and G. Trevino-Quintanilla. Computers & chemical engineering. Shon. 60. 44(6): p. 2001. T.. and L. and S. 14(1): p. W. 1988. Almeida. E. K. Y. 150(2-3): p. 2006. Fischer. Dochain. N. 58. Kremling. 2002. Marquardt. M. 801-818. Voit. Application to anaerobic digestion. 56. ACM. 2(1): p. 1662. Allgöwer. H. and R. 54. Fraser. and F.. Analysis of Approximately Lumpable System. Mavrovouniotis. and J. Lambeck. Bayer. 28: p. Silvert.. Vinga.P. 2001. Hecker. E. Statistics as a catalyst to learning by scientific method part II-a discussion.E. Biotechnology Progress. Delgado. 1.C. Fraser. 71. 685-690. and G. Computer-aided web-based application to modular plant design. Chaos. J. L. Vaccari. 69. 2003. B.P. A Benchmark for Methods in Reverse Engineering and Model Discrimination: Problem Formulation and Solutions. Hassis.E. Kuo. 67. 59(13): p. S. E. 124-133. S. 1773-1785. 48. Kaufmann. A. Liao. Heine. Fischer. Simplification of mathematical models of chemical reaction systems.E.n. 2004: p. King. 96(1): p. 221-233. Lumping Analysis in Monomolecular Reaction Systems. 2009. 196-206. Biosystems..R. 1998. R.A. The steady state and equilibrium approximations: A geometrical picture. 2002.Bibliography 47. Schuster. 3(9): p. and J. Chemical Engineering Science. C. and W. K. Bardow. 11: p. Gadkar. Csete. 63. 52. J.. A. An invariant manifold approach for CSTR model reduction in the presence of multi-step biochemical reaction schemes. 13(8): p. 64..G. Parameter optimization in S-system models.. 295(5560): p. Gene regulatory network inference: Data integration in dynamic models--A review. 1996: Springer Us. 1991. The regulation of cellular systems. The Journal of Chemical Physics. Incremental and simultaneous identification of reaction kinetics: methods and comparison. 55. Roussel. State and parameter estimation in chemical and biochemical processes: a tutorial. 28(8): p. 98(2): p. and S.D... Bailey. Freyre-Gonzalez. Stewart.. I. 66. 57. S. Industrial & Engineering chemistry fundamentals. Mathematical modeling: a chemical engineer's perspective.J. 2004. Sauter. 2673-2684. 62. 1989/90 und 1990/91. S.. 1664. J. Box.. 11(1). Doyle. W. Lyberatos.. Journal of process control. and J. 462-475. 51. M. Syrou.E. 2008. Journal of Quality Technology. 61. Vasconcelos. 1993. M.. and W. 51. 8(1): p. 391-408. G. Bullinger. Science. Chou. Computer Applications in Biotechnology. Wozny. Mathematical modeling and analysis in biochemical engineering: past accomplishments and future opportunities. and M. Analyzing Regulatory Networks in Bacteria Nature Educations. Kravaris. D.L. 30(3): p. Towards integrated information models for data and documents. Berlin Kurzfassungen von Vorträgen der Wintersemester 1988/89.C. 1999: Academic Pr. Comparing mathematical models on the problem of network inference. 1999. Simplification of wastewater treatment plant models using empirical modelling techniques. Heinrich. and D. E. 2004. A. Aris. 132 . 16-29. Chem. 68. 50. Young Researchers 2004. 295(5560): p.J. Violet. 2010.A. A. 1969. Biotechnology progress. S. 31(1): p. Genome Research. A Software Supported Compartmental Modeling Approach for Paenibacillus Polymyxa.C. Spieth. Doyle. 2004. Vilela. R. Van Someren. Wei. International Journal of General Systems. 24. 1404-1412.. 2010. 65. and E. N. Gilles. 8-20.E. T. Kitano. Weiss.C. and J. 49. Modelling as a discipline. B. Okino. Computer Aided Chemical Engineering. 1998.

S. H. Industrial & Engineering Chemistry Research. Optimal determination of steric mass action model parameters for [beta]-lactoglobulin using static batch experiments. Goldaniga. E. Macchietto. Dente.G. 2001. Engell. 2010. Kevrekidis. S. K. and D.H.K. Bock. 245-279.W. Nagy. 1(4): p. pyrolysis. 75. Diehl. Churchill. partial oxidation and combustion of hydrocarbon mixtures. Data-driven calibration of penalties for least-squares regression. SIAM review. S. 1989. 2001. M. Massart. 99. Slemrod. Wozny..A. 2008. A. and G. Monte Carlo simulation in statistical physics: an introduction: Springer. 785-802. G. 2007.P. V. P. Willett. R. M. Franceschini. Abstracting the ODE semantics of rule-based models: Exact and automatic model reduction. Data Mining: Practical machine learning tools and techniques. 46(1): p. and C. Nonlinear Dynamics. Küpper. 2008. H. Syrou. Nicolas Cruz Bournazou 80.. G. and T. and P. 81. Lin. 82. Coarse-Grained Multiscale Computation: Enabling Mocroscopic Simulators to Perform System-Level Analysis. 92. in Ruprecht-Karls-Universitaet Heidelberg. Dokoumetzidis. Statistical tools for optimal dynamic model building. M. Biostatistics.M. Fontana. O.. P. Feret. Faravelli. and M.O. Kerr. Chemical Engineering Journal. C. Y. W. Efficient moving horizon state and parameter estimation for SMB processes. C. Heermann. The quasi-steady-state assumption: a case study in perturbation. Kazantzis. 31(3): p. Optimization of biological nutrient removal in a SBR using simulation-based iterative dynamic programming. T. 183-201. Binder. Schloder.. and G. and E. Lumping procedures in detailed kinetic modeling of gasification. Löffler. 10: p. Macchietto. 2009.. Xu. BMC Bioinformatics. 85. 27(1): p. L. 3(1): p.J. 2009.M. Krivine. Computers and Chemical Engineering. J.K. I. Neubauer. Journal of Process Control. Numerische Methoden fuer Optimale Versuchsplanungsprobleme bei nichtlinearen DAE-Modellen. Ranzi. 2003. 74. and S. Schlöder. 87. T. and S. Harmer.. and S. Determination of the maximum specific uptake capacities for glucose and oxygen in glucose-limited fed-batch cultivations of Escherichia coli. 2000. Kostina. 40-51.An approach to Mechanism Recognition for model based analysis of Biological Systems 73. G. 84. 2(2): p. 2009. M. Enfors. 2010. Progress in energy and combustion science. Franceschini. 93. Korkel.. R. 91. H. Proper lumping in systems biology models.G. and L. 94. 2009. 347-57. 83.. H. V. 76. 220-232.Y.. B. Asprey. Iterative approach to model identification of biological networks. and P. Reduction of very large reaction mechanisms using methods based on simulation error minimization. Macchietto. Gear. Berger. 77. 19(5): p. and S. and J. M. Witten. Frank.. Numerical Methods for Optimal Control Problems in Design of Robust Optimal Experiments for Nonlinear Dynamic Processes. J. Comparison of data reduction techniques based on the performance of SVM-type classifiers: IEEE. S. B. Experimental design for gene expression microarrays. Turányi. 63 (19): p 4846-4872. 86.G. R. Biotechnol Bioeng. Gunawan. Kevrekidid. Doyle Iii. 156(2): p.. Arlot. 11-19.R. 446-477. Validation of a model for biodiesel production through model-based experiment design. Azam. Segel. Arellano-Garcia. Model-based design of experiments for parameter precision: State of the art. 24(2-7): p. Gadkar. A. Hyman. 79.G. Danos. and T. 6(155). Ghoshal. K. 73(5): p. 2002. 2009. Yoo. Aarons. Barz. 88. and S. 89.H. 90. Lee. A new model reduction method for nonlinear dynamical systems.G. 183-194. Combustion and Flame. Communications in Mathematical Sciences. 78. and J. 2005. 2001. J. and F.P. 1261-1267. N. 417-428. Journal of Chromatography A. 715-762. 139(1): p. Bozzano. I.B. Runborg. S.. E. The Journal of Machine Learning Research. 133 . Equation-Free. and L. A. Systems Biology. 2005: Morgan Kaufmann.W. C. Bock. C. Kravaris. Georgescu. Theodoropoulos. and I. D. IET.P. Korkel. 59(1): p.A. Chemical Engineering Science. Kim. Mathiszik.

C. Nahrstedt.Bibliography 95. 80(1 2): p. 111-123. Wozny.u. 5425-5435. Revista Mexicana de Ingeniería Química. Logsdon. Holmstrom. Kostina. J. Technometrics. 24: p. Model selection.K. Chemie Ingenieur Technik. Biegler. 35(2): p. J. The optimal design of dynamic experiments. 18(9): p. 1995.C. Boyd. Drews. Maximin designs for exponential growth models and heteroscedastic polynomial models. J. 107. P. R. P. Pant. 5(1): p. 99.A.P.M. Park.P. AIChE Journal. Quantitative description of Listeria monocytogenes inactivation kinetics with temperature and water activity as the influencing factors. Geeraerd.. 2008. W. 29(2): p. and Y. 1628-1639. Computer Aided Chemical Engineering. H. and K.H.E. design of experiments in non-linear situations. Journal of Process Control. 35-47.S..W. A. I. 236(1-3): p. 2005. C. 2006. 103.G. 2005. Ann. 1061-1074. and F. 109. 62(18-20): p. Lucas. G. Hunter. Applied Microbiology and Biotechnology. and J. and A.L. 46(1-2): p. Valdramidis. and S. Linhart. Desalination. 77-90. Lim. 1986: John Wiley & Sons. Kraume. 1965. Chemical Engineering Science. Journal of Theoretical Biology. 64(5): p.S. and G. 2009. 76(1): p. 110. 96. Espie. and E. K. Vaca.E. 108.E. 112. J. and L. Zavala. Gaze. 2007. A. and B. 117.H. Monroy-Loperena.B. 1999. K. V. and H. Drews. Kinetic models analysis.T. R. Ko.P. and G. T. model prediction and methodological validation on dynamic data. 114.. Application of collocation techniques to the design of petlyuk distillation columns.R.. and A. Filtration zur Partikelabtrennung bei der Wasserreinigung. 37(5): p. 100(1-3): p. L. Van Impe. Schoneberger. MOSAIC: ein Modellierungsserver für die Verfahrenstechnik. C. Biometrika. A. Arellano-Garcia. Approximate Gauss-Newton Method for Robust Parameter Estimation. Arora. 1990. Zucchini. Macchietto. 134 . O.. A fast moving horizon estimation algorithm based on nonlinear programming sensitivity. Experimental design for optimal parameter estimation of an enzyme kinetic process based on the analysis of the Fisher information matrix. Model-based recognition of fouling mechanisms in membrane bioreactors. IEEE Transactions on. 104. 98. 100. Performance of some SQP algorithms on structural design problems.K. The TOMLAB Optimization Environment in Matlab 1. Gysemans. 43(5): p.. and W. J. Jimenez-Gutierrez. and S.. Tseng. 116. 1323-1333. J. G. 224-233. 111. 1989. A. A physics-based MOSFET noise model for circuit simulators. Hung. and M. H. 2007. Vocks. Improving the Efficiency of Membrane Bioreactors by a Novel Model-based Control of Membrane Filtration. 105. Drews. Binder. p.F. Optimal dynamic experiments for bioreactor model discrimination.K.J. Buzzi-Ferraris. Shaw. Electron Devices.V. 876-884. Statistics for experimenters. Felix Oliver Lindner.A. G. Imhof. S. 113. Impe. Cooney. K. Reiner. Schöneberger. 1989.. Accurate solution of differential-algebraic optimization problems. Statist. D. Chemical Engineering Science. Box. P.. 307-323. 2009. G. 391-402. G. Hu. Gimbel. 2001. and W..F.J. Designs for discriminating between two rival models. M. 106. Box.D. McDonald. Quantifying microbial lag phenomena due to a sudden rise in temperature: a systematic macroscopic study. Kraume. 118. M. Schaller. 2005. 2008.L. 79-88. Swinnen. 531-532. Hunter. 121-127. 2006. PAMM. 10(1): p.. Nutrients removal in MBRs for municipal wastewater treatment. International Journal of Food Microbiology. 97. Manenti. Cheng. Laird.T. Panglisch. 115.K.M. Hunter. Kraume. K. U.. Journal of Food Engineering. 85-96. A. 561-576.A. Biegler. M. M.M. and J. Kondjoyan. Arellano-Garcia. Schaller. H. 826-837. M. Bracklow. and K. Industrial & engineering chemistry research. A.. A.. 238(1): p. and R. 345. Patel. Zerry. J. H. Wozny. V. Experimental study and mechanistic kinetic modeling for selective production of hydrogen via catalytic steam reforming of methanol. 2006. 223-229. 1978: Wiley New York. Thanedar. 2009. 101. 102. 28(11): p. J. 1986.S. Hitzmann. 1959.G. and L. 7(3): p.. Bernaerts.

and M. M. Mino..An approach to Mechanism Recognition for model based analysis of Biological Systems 119. Cruz Bournazou. S. 2000: Intl Water Assn. Luccarini. Activated sludge model No. G. Gray. Wozny. Veenstra. Kampschreur. Chemical Engineering Research and Design.W. 1155. Vanrolleghem. 130. 2002. Journal of Hydroinformatics.. 35(1): p. 462-474.C. Kraume. van Loosdrecht. G.L. M. G.N. 1864-1876. and D. IEEE Transactions on. Gujer. G. Goronszy. 2011: p. Henze. 28(4): p. Journal of Chemical Technology and Biotechnology 2012. Office of Water Washington. Isermann. and M... P. 15-38. Lind. Bragadin.. 83(3): p. 35. M. Yuceer. 2005: Intl Water Assn. 1997. Gernaey. 25(5): p. 1992. T. Rieger. and M. M. 183-193. Frank.B. 134. Automatic Control. G. and A. 1997. I.. S. Marais. EPA 832-F-99-073. 138. 123. 31(2): p. 4(1): p.H. Observability and controllability of piecewise affine and hybrid systems. 763-783. 10. Control vector parameterization approach in optimization of alternating aerobic-anoxic systems. Sottara. Wentzel. and C. 2009. Optimal Control Applications and Methods. Observability analysis of piece-wise constant systems.F. Kaelin. Evaluation of an ASM 1 model calibration procedure on a municipal-industrial wastewater treatment plant. Mello. K. Takacs. 39(1): p. P. Gernaey. Wilderer. Henze. L. Kravaris. 2002. Supervision.. van Loosdrecht. Activated sludge wastewater treatment plant modelling and simulation: state of the art. Bemporad. Siegrist. M. Velmurugan. and S.J. G. Matsuo.. 4. M. Montali.L. Water Environment Research. R. N. Mino.. Goshen-Meskin. 573-584. D. Comparison of fixed and variable threshold RAIM algorithms.A. M. Mechanism and design of sequencing batch reactors for nutrient removal. Wastewater Technology Fact Sheet Sequencing Batch Reactors. I. 45(10): p. B. Journal of process control. 58(6): p. Ferrari-Trecate. The activated sludge model no. Case study: SBR plant. W. and M. 648-660. fault-detection and diagnosis methods–a short introduction. R.A. Drews. Utilization of SBR technology for wastewater treatment: an overview. ASM3 extended for two-step nitrification-denitrification: a model reduction for Sequencing Batch Reactors.. 1995. Navigation. 11-18. Gernaey. Survey of robust residual generation and evaluation methods in observerbased fault detection systems. Colombini. 1999.C. 1999. Water science and technology: a journal of the International Association on Water Pollution Research. and M. Aerospace and Electronic Systems. Mata-Alvarez. 1990: Oxford University Press. 128. Design and physical features of sequencing batch reactors. Sturza..V. Ketchum. 133. IEEE Transactions on.C.M.C.N. W. K. No. 2004. Environmental Modelling & Software. 41(23): p. W. Irvine. Chem. Activated sludge: theory and practice. A. 125. FaultDiagnosis Applications. DOI: 137. H. and G. 2008. P. Nicolas Cruz Bournazou 131. and D. S. A.. 121. D. and J. Orhon. Water Science & Technology. M. Sin.A. Henze. 11-45. Winter 1988-89.. Petersen.Y. Van Loosdrecht. Mancini. Modelling nitrite in wastewater treatment systems: a discussion of different modelling concepts. and I. Model-Based Design of Sequencing Batch Reactor for Removal of Biodegradable Organics and Nitrogen. 136. 276-284. B. 135. Berber. 2000.. 126. M. Ding. Morari. and J. 82(5): p. 139.V. and R. Ind. 403-424. Vol. Clarkson. Sequencing batch reactor technology. 7(6): p. Lyberatos. Henze. H. 2: biological phosphorus removal. and X. 1056-1067. Balku. Theory..M. Gujer. 30(6): p. 2005. Water Science and Technology. Eng. M. K. Artan. 5539-5553. Water science and technology. Jørgensen. D. 120. 124. Environmental Modelling and Software. T. 127. Arellano-Garcia.3694. Process improvement by application of membrane bioreactors. Wett. Formal verification of wastewater treatment processes using events detected from continuous signals by means of artificial neural networks. M. T. Mace. 3. Bar-Itzhack. Res. 135 . and P. L. Brown.1002/jctb. 122. M. L. N. 1-11. 129. 2010. M. 19(9): p. M.K. 132.

and J. 64(5): p. Zawada. SIAM review. 149. S. Extension of ASM3 for two-step nitrification and denitrification and its calibration and validation with batch tests and pilot scale data. J. Akin. and U. A. 3569-3576. and C. 2000. Kaelin. Kayser. Lahtvee. Stuetz. Busby. M. 143: p. Renneberg. and W. J. 147.. 60-65. 2875. 1999. Turk. 1680-1692. sensitivity analysis. and optimization of hybrid systems. C.R. 2873-2878. 1999. 1999. and D. 207-213. 153.W. W.. Alkylation of Benzene with Isopropanol over SAPO-5: A Kinetic Study.N. Hecht. On-line monitoring equipment for wastewater treatment processes: state of the art. 40(8): p.. 142.C. Simel.S. 135(11): p. 154.J. J. P. Axelsson. and D. On-Line Detection of Acetate Formation in Escherichia coli Cultures Using Dissolved Oxygen Responses to Feed Transients. 353-63. Adamberg. and W. 256-289. A. M. 4173. F. Ugurlu. R.R. G.. E. Lee. Lyberatos.N. Barton. P. Metabolic flux analysis of Escherichia coli in glucose-limited continuous culture. 161.. 12-50. Burgess. 2007. W. Microbiology. The Acetate Switch [Review]. Adaptive optimization of a nitrifying sequencing batch reactor. Biosensor for rapid phosphate monitoring in a sequencing batch reactor (SBR) system. Shiloach. 1997. 47(2): p. Vilu. and A.M. L. Rieger. 1-34. V. Hagander.M. 2001. and J. J. 150. K.K. P. Applied and environmental microbiology. Upadhyayula. Eugster. ACM Transactions on Modeling and Computer Simulation (TOMACS). Russell. 141. 2005. On-line monitoring of wastewater quality: a review. C.W. Regulation of acetyl coenzyme A synthetase in Escherichia coli.. 182(15): p. and G.N. Bourgeois. 2005. Specific growth rate dependent transcriptome profiling of Escherichia coli K12 MG1655 in accelerostat cultures. M. A. P. D. 2005. 33(17): p. and R. K. Reichelt.A. 693. Microbiol Mol Biol Rev. 86: p. and J. Kierzenka. S. J. Siegrist. Bungay. A. Journal of chemical technology and biotechnology. 144. S. Vanrolleghem. 1175-1180. Barford. Wolfe. Luli. J. Biotechnology and Bioengineering. 152. Impact of dissolved oxygen concentration on acetate accumulation and physiology of E. 156. 151. Growth-rate-dependent metabolic efficiency at steady state. Kornaros. Microbiology. 145. Maintaining Rapid Growth in Moderate-Density Escherichia coli Fermentations. Canadian Journal of Civil Engineering/Revue Canadienne de Genie Civil. 19(3): p. and R. 2008. Humphries. 1990.J. 56(4): p.. 1004. evaluating transcription levels of key genes at different dissolved oxygen conditions. Mak. 337-348. Comparison of growth. Kumari. The Canadian Journal of Chemical Engineering. Lee. Solving index-I DAEs in MATLAB and Simulink. 157. Chan. Swartz. Hovel-Miner. 143. El-Mansi.B. 21(1): p. Journal of bacteriology. Wolfe. Monitoring and control of biological nutrient removal in a sequencing batch reactor. Metab Eng. R.E. 590-598. E.I. E. 76(4): p.. Weber. 69(1): p. 233-237..A. Modeling. 7(5-6): p.Bibliography 140. M.M. Abdel-Hamid. Abner.S. Control of carbon flux to acetate excretion during growth of Escherichia coli in batch and continuous cultures. S. Pyruvate oxidase contributes to the aerobic growth efficiency of Escherichia coli. and A. Diez-Gonzalez. 41(3): p. Water and Environment Journal. and T. Nahku. Guest.H.K. 148.T. 13(6): p. J. Journal of general microbiology. Rinas. L. Browning. 538-552. 2003. Shampine. 151(3): p. 600-605.J. Attwood. and H. Akesson.S.F. and R. 147(6): p. Operating strategies for variable flow sequencing batch reactors.E. 158. 155. Manser. Erm. 2009. Katsogiannis. 160.M.. coli BL21. K. Tocaj. and A. acetate production. 2005. Journal of biotechnology. and acetate inhibition of Escherichia coli strains in batch and fed-batch fermentations.. 136 . 145(1): p. S. Rottermann. and J. Water Research. simulation. 1986. 146. 1-8.P. Karlsson. The ability of Escherichia coli O157:H7 to decrease its intracellular pH and resist the toxicity of acetic acid. Strohl. Water Research. 89(4): p. I.J. Preliminary assessment of a shortcut in nitrogen removal from wastewater. D. 159. 12(4): p.F.W. G. Valgepea. 1989. Beatty. 2001. Stephenson. Automation in Water Quality Monitoring.M. 2003. 407-415. Holms. Mavinic.. 43(6): p. Biotech Bioeng. K. 1483. Microbiology. M. 2005. B. 2010. Phue.. Process Biochemistry. Biosensors and Bioelectronics. and J. O. 2002.

Liddell. Scarff. Cruz Bournazou. V. C. Physiological Stress Responses in Bioprocesses. 67(2): p. and W. Merten. C.. Recombinant protein production with prokaryotic and eukaryotic cells: a comparative view on host physiology: selected articles from the meeting of the EFB Section on Microbial Physiology.. 1922. Applied mathematics and computation. Hardiman. P. Biotechnology and Bioengineering. Biotechnology progress. 257-278. Closed-loop control of fed-batch cultures of recombinant Escherichia coli using on-line HPLC. Magnusson.W. and P.R. M. 79(1): p. 1-8. Curr Opin Microbiol. 2009. 63-74. K. 169. Reuss. M. L. 1998. and R.S. Biotechnology and Bioengineering 1994. G.A. and A. S. D. Nebe-Von-Caron. and D. Journal of microbiological methods. 167. 2000. 12(2): p. M. Junne. Rossi. 174. 46(3): p. Dynamic modeling of the central carbon metabolism of Escherichia coli. O. J. 170. 1996. 178.J.. M.R. 255-263. 197-223. 137 . Y. Tamburini.A. M. 2004.. 2002. The application of multi-parameter flow cytometry to monitor individual microbial cell physiological state. Semmering. 171. F. Harvey. Trends Microbiol. and S. Matteuzzi. 2006. N. Influence of controlled glucose oscillations on a fed-batch process of recombinant Escherichia coli.L. F. Farewell. Lemuth. Diabetes Care. 181. S. L. A. C. 176. Enhanced production of L-(+)-lactic acid in chemostat by Lactobacillus casei DSM 20011 using ion-exchange resins and cross-flow filtration in a fully automated pilot plant controlled via NIR. Electrooptical monitoring of cell polarizability and cell size in aerobic Escherichia coli batch cultivations. Assessment of in line near infrared spectroscopy for continuous monitoring of fermentation processes. Noisommit-Rizzi. 166. 359-374. Angersbach.O. A. Hewitt.J. L. J. 2001: Springer Netherlands. and C. Cord-Ruwisch. Chassagnole. C. B. 26(1): p. and M.. Journal of biotechnology. Cytometry Part A. Scheper. 164. J. H. 2010: p. 236-42.W. and P. Online detection of feed demand in high cell density cultures of Escherichia coli by measurement of changes in dissolved oxygen transients in complex media Biotechnology and Bioengineering. Lindemann. 3-16. 2005. S. Mauch. Thornhill. Thomas.. and G. Siemann-Herzberg. Evaluating the accuracy of continuous glucose-monitoring sensors. 170-6. and B. Near infrared spectroscopy for bioprocess monitoring and control: current status and future trends. C. Hengge. Journal of Industrial Microbiology and Biotechnology. 17-39.. S. Want. 2004.T. Austria.C. Amaretti. Studies related to antibody fragment (Fab) production in Escherichia coli W3110 fed batch fermentation processes using multiparameter flow cytometry.M. 2003. ppGpp: a global regulator in Escherichia coli.. 172. A. Nystrom. 177. 147-156. N. D. Gonder-Frederick. McNeil.. 175-193. 175. 42(1): p. 165.J.. Shapiro. Multi wavelength fluorescence lifetime spectroscopy: a new approach to the study of endogenous fluorescence in living cells and tissues. 44(7): p. K. 2007. Vaccari.J.N. 173. 2000. 2000.P.. O. 14(1): p. R. T. Biotechnology progress...U.A. 148-154. and T. Neubauer. E. Kovatchev. 79(1): p. Bacterial nucleotide-based second messengers. 819829 Chorvat Jr. Wang. 5th-8th October 2000. 27-37. and M. 2004: p. B. E. 85( 4): p.. 75(2): p. Götz. 2009. M. and T. Reuss. J Biotechnol. 1816-1821. Cox. M. 2009. Marose. 6(3): p. Biotechnology and bioengineering. Journal of Biotechnology. Schmid. Whiffin. 2000. H. A modified collocation method for solving differential-algebraic equations.. 422-433 Turner.An approach to Mechanism Recognition for model based analysis of Biological Systems 162. Arnold. S. 27(8): p. Andersson. Lin.. 168. Kara. Neubauer. Yang. A. 13(5): p. Pesavento.G. 180. Clarke. Laser Physics Letters. Gonzalez-Vara. Topology of the global regulatory network of carbon limitation in Escherichia coli. Critical reviews in biotechnology.Y. 19(6): p. Impact of plasmid presence and induction on cellular responses in fed batch cultures of Escherichia coli.. M. Keller. L.M. Chorvatova. 163. 116(3): p. 179. Microbial analysis at the single-cell level: tasks and techniques. Nicolas Cruz Bournazou Tosi. 53-73. Hewitt. Enfors. 132(4): p. Two Dimensional Fluorescence Spectroscopy: A New Tool for On Line Bioprocess Monitoring.S..

R. Biosystems. Schmid. Domach. Bioprocess and biosystems engineering. 1990. 44(8): p. Cruz Bournazou. 197. 273(41): p. A. and R. 1296. R. V. Shimizu. Dawson. Lin. coli fermentations with software sensors. Acetyl-CoA synthetase overexpression in Escherichia coli demonstrates more efficient acetate assimilation and lower acetate accumulation: a potential tool in metabolic engineering. A. P. and K. Monitoring of fed-batch E. Multi-omics approach to study the growth efficiency and amino acid metabolism in Lactococcus lactis at various specific growth rates. M. Lahtvee. Effect of modified glucose uptake using genetic engineering techniques on high-level recombinant protein production in escherichia coli dense cultures. Rocha. Neubauer. S. Srinivasan. S. R. Homans. San. Bennett. H. Biotechnology and bioengineering. and excretion of acetate. 1990.C. Computers & chemical engineering. Computer Aided Chemical Engineering. De La Puente.N. H.A. 381-388. P.H. Costa. 27-44. 28: p.O. Dynamic optimization of batch processes:: II. Bonvin. 27: p. the gene which encodes acetyl coenzyme A synthetase in Escherichia coli. D. and B. Arellano-Garcia. Van Dien. 2009. K. Chou. Escherichia-Coli Growth Dynamics . R. A. Jung.A. Biotechnology and bioengineering. Palanki.. 2878.B. Bennett. 1990. 870-874. Usuda. Ferreira.. 191. M. Biotechnology and bioengineering. 31(2): p. MacDonald. Schöneberger. 195. A Novel Approach to Mechanism Recognition in Escherichia Coli Fed-Batch Fermentations. 640... Visser. Ferreira. and A.. and K. Journal of bacteriology. Chou. Ben-Bassat.. S. 1998. Arike.M. S.L. and A. 952-960. 651-656. and R. Applied and environmental microbiology. H.N. Y.C. 185.. Biotechnology and Bioengineering. and K. Topology of the Na+/Proline Transporter ofEscherichia coli. Field. N. Nishio. Carlsen. Machado.T. Aller.Bibliography 182. Cloning. 2003. 177(10): p. Rocha. 192. acetate. 56(5): p. Improved expression of human interleukin-2 in high-cell-density fermentor cultures of Escherichia coli K-12 by a phosphotransacetylase mutant. Biochemistry. H. Iwatani. K. Matsui. 4158-4168.Y. The central metabolic pathways of Escherichia coli: relationship between flux and control at a branch point. Wozny. Jensen. Bennett. A. San. Simple constrained-optimization view of acetate overflow in E. I.H. 102-116. W. Neway. L. Joshi. and K. Leifker.. Räbenhagen. San.J. Effects of medium quality on the expression of human interleukin-2 at high cell density in fermentor cultures of Escherichia coli K-12. 194.J. Effect of modulated glucose uptake on high-level recombinant protein production in a dense Escherichia coli culture.A. Tebbe. 644-647. Tishel. 187. Production of recombinant human growth hormone in Escherichia coli: expression of different precursors and physiological effects of glucose. Kumari. 1998.Y. M. and functional expression of acs.. G. D. Nahku. Kageyama.. 2010. 2010. Bauer. 190.O. 11078-11082. 199. 183.N.J. N. Journal of Biological Chemistry. Donohue-Rolfe. Microb Cell Fact. 2009. Iwahata. Applied and environmental microbiology. S. 2011. Holms. 732-738. Eisenbach. 196.Y.S. E. K. Miyano. Wolfe. Junne. Majewski. E. N. K. 71(6): p. 1-11. 10(6): p. H. Quick. 189. 69-105. J. Veloso. Hybrid dynamic modeling of Escherichia coli central metabolic network combining Michaelis-Menten and approximate kinetic equations. 150-7.C. 1988. 56(3): p. Biotechnology progress. C.W. 186. B.N.A. Tholema.. Current topics in cellular regulation. Adamberg.. G. C. Imaizumi. 100(2): p. and S. 138 . 35(7): p. and J. Neway. 26400. Solution structure of the complex between the B-subunit homopentamer of verotoxin VT-1 from Escherichia coli and the trisaccharide moiety of globotriaosylceramide. 2006. 188. and G. Dynamic modeling of Escherichia coli metabolic and regulatory systems for amino-acid production. Vilu. 36(1): p. R. R. 184. 32(3): p.O. I. 27(1): p. and E. Palsson. M. Shimbo. Y. efficiency of conversion to biomass. 1994. and E. 147(1): p. 1990. and M. Applied microbiology and biotechnology. G. Castro. 1995. characterization. and salts. K.H.M. 10(12): p. J Biotechnol. Role of measurements in handling uncertainty.. 193. 17-30. and S. 37(31): p. coli.C. D. and J. 1994. S. 1986. 198.. H.a 3-Pool Biochemically Based Description.

M. Bolivar. Y. B. Yoshida.S. 182-192. Nicolas Cruz Bournazou 203. K.. V. and M. and J. 2010. Imaizumi. Seki. Miyano. and O. and M. Green. Ramirez. 53(2): p. Biotechnology and bioengineering. Journal of structural biology. coli cultivations. and S. Shimbo. 4(2): p.N. D. Bunin. 210. Analytical biochemistry. San Martin.Y. 2011. Vadali. Vrahatis. 205.. 201. M. Utility of an Escherichia coli strain engineered in the substrate uptake system for improved culture performance at high glucose and cell concentrations: An alternative to fed―batch cultures. Studienarbeit. 2002. Carazo. Konstantinov. and I.N. G. N. and T. 212-217. 1(2): p. 99(4): p. and K. Journal of Biotechnology. Bennett. 207. 2006. Continuous On-line Electrooptical Monitoring of Physiological State of Bacterial Cultures in Biotechnological Processes. Kinetic modeling of tricarboxylic acid cycle and glyoxylate bypass in Mycobacterium tuberculosis.C.E. Nishio. S. 235-306. 27(2): p.P. 204... N.J. Sariyar.. Dammerova. 1973. modifizierte Stärke-Betimmung des Acetylgehaltes-Enzymatisches Verfahren. Kageyama. Matsui. A structural model for the Escherichia coli DnaB helicase based on electron microscopy data. A balanced DO-stat and its application to the control of acetic acid excretion by recombinant Escherichia coli.M. F. 202.J. 1969. Van Dien. Parsopoulos. Recent approaches to global optimization problems through particle swarm optimization. 209. Metabolic engineering. S. 1995. K. Iwahata. 452-458. Kishimoto. Rudolph. Theor Biol Med Model. S. Bernofsky. and V. Angersbach. 211.T. Design of Experiments and parameter Estimation of a kinetic model for E. 139 . 3: p. G. Caspeta. Swan. Stamford. K. 27. A. Blackwood. 17-30. 893-901. An improved cycling assay for nicotinamide adenine dinucleotide. 206. Iwatani. Berrios-Rivera. M. 2002. J. N. A. Y. 167-176. L. Y. Metabolic engineering through cofactor manipulation and its effects on metabolic flux redistribution in Escherichia coli. A simple ultramicro method for determination of pyridine nucleotides in tissues 1. K. N.J.R. 36(7): p. H.B. and K. T. Ghosh.An approach to Mechanism Recognition for model based analysis of Biological Systems 200. 1995. Horton. Dixon. C.. Gosset. Nisselbaum. Lara. 2008.K. San. Biotechnology and bioengineering. R. 208. 114(3): p.V. Usuda. E. F.E. and its application to assessment of drug targets. 147(1): p. Lemoine. 1990. Dynamic modeling of Escherichia coli metabolic and regulatory systems for amino-acid production. A. Yang. 750-758.T. Natural computing. Singh. DIN_EN_ISO_11213. A. Analytical biochemistry.

Master your semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master your semester with Scribd & The New York Times

Cancel anytime.