Model Reduction in Chemical Engineering

Case studies applied to process analysis design and operation

Cover design by JM&BD

Model Reduction in Chemical Engineering

Case studies applied to process analysis design and operation

Proefschrift

ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus, prof. ir. K.C.A.M. Luyben,
voorzitter van het College voor Promoties
in het openbaar te verdedigen op maandag 4 juli 2010 om 15:00 uur

door

Bogdan DORNEANU

Inginer diplomat
(Universitatea Politehnica din Bucureti, Roemeni)
geboren te Medgidia (Roemeni)

Dit proefschrift is goedgekeurd door de promotor:

Prof. ir. J. Grievink

Samenstelling promotiecommissie:
Rector Magnificus

Voorzitter

Prof. ir. J. Grievink

Technische Universiteit Delft, promotor

Dr. ir. C.S. Bildea

University Politehnica of Bucharest, copromotor

Prof. dr. ir. P.M.J. van den Hof

Technische Universiteit Delft

Prof. dr. F. Kapteijn

Technische Universiteit Delft

Dr. ir. H.J.M. Kramer

Technische Universiteit Delft

Prof. dr. ir. P.M.M. Bongers

Technische Universiteit Eindhoven / Unilever

Prof. dr. P.D. Iedema

Universiteit van Amsterdam

Dr. ir. P.J.T. Verheijen

Technische Universiteit Delft, reservelid

Keywords: model reduction, process modelling, plantwide control, dynamic optimization,

alkylation, ice cream freezing
An electronic version of this thesis is available from http://www.library.tudelft.nl
ISBN 978-90-8570-771-4
Copyright 2011 by Bogdan Dorneanu
All rights reserved. No part of the material protected by this copyright notice may be
reproduced or utilized in any form or by any means, electronic or mechanical, including
photocopying, recording or by any information storage and retrieval system, without the prior
written permission from the author.
Printed in The Netherlands by CPI Whrmann Print Service

Motto

It can scarcely be denied that the supreme goal of all theory is to make the irreducible basic elements as
simple and as few as possible without having to surrender the adequate representation of a single datum of
experience.
A.

Einstein

A significant part of the research was carried out within the framework of MRTN-CT-2004-512233 (PRISM Towards
Knowledge-Based Processing Systems). The financial support of the European Commission and Unilever R&D Vlaardingen is
gratefully acknowledged.

Table of contents
Summary ............................................................................................................................................... i
Samenvatting ....................................................................................................................................... v
Rezumat............................................................................................................................................... ix
Chapter 1 Introduction .................................................................................................................... 1
1. Role of modelling in chemical engineering ................................................................................... 2
1.1.
Model applications and benefits ....................................................................................... 2
1.2.
Efforts and cost factors for model development and application ...................................... 5
2. Requirements for model use in chemical engineering ................................................................... 5
3. Process models and modelling....................................................................................................... 6
3.1.
Process models ................................................................................................................. 6
3.2.
Model classification ......................................................................................................... 8
3.3.
Model development steps ................................................................................................. 9
3.4.
Dual way of modelling: decomposition and aggregation over detail levels.................... 13
4. Motivation and formulation of the research problem................................................................... 15
5. Research approach and outline of the thesis ................................................................................ 18
6. References ................................................................................................................................... 20
Chapter 2 Process modelling and model reduction ..................................................................... 25
1. Introduction ................................................................................................................................. 26
2. A design view on process modelling ........................................................................................... 27
2.1.
Levels of decomposition in the modelling procedure ..................................................... 28
3. Contextual and structure levels of decomposition ....................................................................... 30
3.1.
Contextual level.............................................................................................................. 30
3.2.
Structure level ................................................................................................................ 31
3.2.1. Process decomposition and physical structure of the model ........................................... 31
3.2.2. Behavioural structure of the model ................................................................................. 36
3.2.3. Model connectivity ......................................................................................................... 37
4. Mathematical, numerical and software implementation levels .................................................... 39
4.1.
Mathematical level: behavioural models of nodes .......................................................... 40
4.1.1. Coordinates ..................................................................................................................... 40
4.1.2. Classification of quantities .............................................................................................. 41
4.1.3. Classification of equations .............................................................................................. 42
4.1.4. Validity and applicability conditions (knowledge constraints) ....................................... 46
4.1.5. Types of operations in model equations .......................................................................... 46
4.1.6. Model classification ........................................................................................................ 46
4.2.
Mathematical level: Connectivity ................................................................................... 47
4.3.
Mathematical level: Aggregation at system level ........................................................... 49
4.4.
Relationship between the structure and the complexity of the model ............................. 51
4.4.1. Number of equations ....................................................................................................... 52
4.4.2. Number of terms in the model equations ........................................................................ 52
4.4.3. Connectivity of the model equations ............................................................................... 53
4.5.
Numerical level .............................................................................................................. 53
4.6.
Software implementation level ....................................................................................... 54
5. Existent model reduction techniques ........................................................................................... 55
5.1.
Approaches to model reduction of process models ........................................................ 55
5.2.
Contextual level.............................................................................................................. 55
5.3.
Structure level ................................................................................................................ 56
5.3.1. Model reduction at physical structure level..................................................................... 56
5.3.2. Model reduction at behavioural structure level ............................................................... 58
5.4.
Mathematical level ......................................................................................................... 59

6.
7.
8.
9.

5.4.1. Model simplification techniques ..................................................................................... 60

5.4.2. Model order-reduction techniques................................................................................... 61
5.5.
System level ................................................................................................................... 62
5.6.
Numerical level .............................................................................................................. 62
Model reduction for chemical engineering applications .............................................................. 63
Selected research problems in model reduction ........................................................................... 64
Conclusions ................................................................................................................................. 66
References ................................................................................................................................... 67

Chapter 3 Application of model reduction to design of plantwide control systems .................. 73

1. Introduction ................................................................................................................................. 74
1.1.
Model reduction for complex processes ......................................................................... 74
1.2.
Design of the plantwide control systems ........................................................................ 74
2. Model reduction........................................................................................................................... 75
2.1.
Model simplification ...................................................................................................... 76
2.1.1. Linearization ................................................................................................................... 76
2.1.2. Approximation of functional expressions ....................................................................... 77
2.1.3. Simplification of physical properties model.................................................................... 77
2.1.4. Simplification of chemical kinetics ................................................................................. 77
2.2.
Model-order reduction .................................................................................................... 77
2.2.1. Linear model-order reduction.......................................................................................... 78
2.2.2. Nonlinear model-order reduction .................................................................................... 79
2.3. Linear model-order reduction by balanced realization ............................................................ 79
3. Model reduction with process knowledge ................................................................................... 81
3.1.
Classical model reduction algorithm for chemical plants ............................................... 81
3.1.1. Obtaining the rigorous model .......................................................................................... 81
3.1.2. Obtaining the linear model .............................................................................................. 82
3.1.3. Obtaining the balanced state space model ....................................................................... 82
3.1.4. Obtaining the reduced-order model ................................................................................. 82
3.1.5. Plantwide control structures ............................................................................................ 83
3.2.
Structure-retaining model reduction algorithm ............................................................... 83
A. Identification of units or groups of integrated units to which local control is applied.......... 84
B. Application of tailored reduction techniques to individual units or groups of aggregated
units .......................................................................................................................................... 84
C. Obtaining the reduced model of the full plant ...................................................................... 85
4. Case study: plantwide control of iso-butane - butene alkylation plant......................................... 87
4.1. Process description and the rigorous model ............................................................................ 87
4.2. The linear model...................................................................................................................... 90
4.3. The balanced model ................................................................................................................ 90
4.4. Exploiting the structure ........................................................................................................... 91
4.4.1. The reactor sub-system ................................................................................................... 91
4.4.2. The distillation columns .................................................................................................. 95
4.5. Assessment of the plantwide control structure ........................................................................ 97
5. Conclusions ............................................................................................................................... 101
6. References ................................................................................................................................. 102
Chapter 4 Dynamic optimization using a reduced process model ............................................ 105
1. Introduction ............................................................................................................................... 106
2. Dynamic optimization ............................................................................................................... 107
2.1.
Numerical methods for dynamic optimization ............................................................. 108
2.2.
Single-shooting dynamic optimization ......................................................................... 109
3. Model reduction for dynamic optimization ............................................................................... 110
3.1.
The rigorous model ...................................................................................................... 111
3.2.
The reduced model ....................................................................................................... 113

4.
5.
6.

3.2.1. The reactor sub-system ................................................................................................. 113

3.2.2. The distillation columns ................................................................................................ 114
3.2.3. Performance of the reduced model ................................................................................ 114
The dynamic optimization ......................................................................................................... 115
Conclusions ............................................................................................................................... 119
References ................................................................................................................................. 120

Chapter 5 Development of a reduced model for ice cream freezing ......................................... 125
1. Introduction ............................................................................................................................... 126
1.1.
Model reduction for complex units and products ......................................................... 126
1.2.
Ice formation in a scraped surface heat exchanger ....................................................... 126
2. Physical aspects and modelling approach .................................................................................. 128
2.1.
Ice cream structure ....................................................................................................... 129
2.1.1. Ice particles ................................................................................................................... 129
2.1.2. Liquid matrix ................................................................................................................ 129
2.1.3. Fat droplets.................................................................................................................... 129
2.1.4. Air bubbles .................................................................................................................... 130
2.2.
Scraped-surface heat exchanger ................................................................................... 130
2.3. Freezing mechanism .............................................................................................................. 131
2.4. Scope and approach to modelling of the ice cream freezing unit .......................................... 132
3. Specification of the 5-D model of the ice cream freezing unit................................................... 135
3.1.
Spatial domains with thermodynamic phases ............................................................... 135
3.2.
The generic equation of conservation for the physical resources ................................. 137
3.3.
Mass ............................................................................................................................. 138
3.3.1. The bulk domain ........................................................................................................... 140
3.3.2. The frozen layer domain ............................................................................................... 140
3.3.3. The equipment related domains and phases .................................................................. 141
3.4.
Momentum ................................................................................................................... 141
3.5.
Energy .......................................................................................................................... 141
3.5.1. The bulk domain ........................................................................................................... 143
3.5.2. The ice particles in the bulk domain ............................................................................. 143
3.5.3. The rotor ....................................................................................................................... 144
3.5.4. The frozen layer domain ............................................................................................... 144
3.5.5. The freezers wall.......................................................................................................... 144
3.5.6. The coolant.................................................................................................................... 144
3.6.
Overview of the reductions in the 5-D model ............................................................... 144
4. Specification of the 3-D model of the ice cream freezing unit................................................... 145
4.1.
Mass ............................................................................................................................. 147
4.1.1. The bulk domain ........................................................................................................... 147
4.1.2. The frozen layer domain ............................................................................................... 150
4.2. Momentum ............................................................................................................................ 153
4.3. Energy ................................................................................................................................... 155
4.3.1. The bulk domain ........................................................................................................... 156
4.3.2. The frozen layer domain ............................................................................................... 161
4.3.3. The freezer wall ............................................................................................................ 163
4.3.4. The coolant.................................................................................................................... 163
4.4. Overview of the model reduction for the conservation equations in the 3-D model.............. 164
5. Constitutive equations for the 3-D model .................................................................................. 165
5.1.
Thermodynamic equations of state ............................................................................... 166
5.1.1. Density of phases .......................................................................................................... 166
5.1.2. Specific heat and enthalpy ............................................................................................ 168
5.1.3. Ice-water-sugar phase equilibrium ................................................................................ 170
5.2.
Fluxes in a spatial direction .......................................................................................... 170

6.
7.
8.
9.

5.2.1. Mass flux in axial direction ........................................................................................... 170

5.2.2. Mass flux along the internal coordinate ........................................................................ 171
5.2.3. Energy flux in axial direction ........................................................................................ 171
5.2.4. Energy flux in radial direction ...................................................................................... 171
5.2.5. Energy flux in radial direction over the frozen layer and the metal wall ....................... 172
5.3.
Rate laws ...................................................................................................................... 172
5.3.1. Mass transfer and ice melting processes ....................................................................... 173
5.3.2. Energy transfer processes .............................................................................................. 182
5.4.
Conductivities and transfer coefficients ....................................................................... 186
5.4.1. Mass transfer coefficients ............................................................................................. 186
5.4.2. Viscosity ....................................................................................................................... 188
5.4.3. Thermal conductivity .................................................................................................... 189
5.4.4. Heat transfer coefficients .............................................................................................. 190
5.5.
Overview of the reductions for the constitutive equations in the 3-D model ................ 192
Summary of the mathematical structure of the 3-D model ........................................................ 193
Input-output structure of the reduced model for model use ....................................................... 197
Conclusions ............................................................................................................................... 199
References ................................................................................................................................. 200

Chapter 6 Numerical application of the reduced model for ice cream freezing ...................... 203
1. Introduction ............................................................................................................................... 204
2. Development of a numerical scheme and implementation ........................................................ 206
2.1.
Discretization of coordinates and mathematical operators ........................................... 206
2.2.
Discretization of model equation for the steady state model ........................................ 209
2.3.
Discretization of the population balance equation for the ice particles......................... 210
2.3.1. Finite volume method ................................................................................................... 211
2.4.
Software environment .................................................................................................. 213
2.5.
Verification of the model implementation .................................................................... 214
3. Model applications .................................................................................................................... 215
3.1.
Model simulation .......................................................................................................... 215
3.1.1. Simulation scenario ....................................................................................................... 215
3.1.2. Simulation results and discussion.................................................................................. 217
3.2.
Parametric sensitivity analysis for validation purposes ................................................ 223
3.2.1. Sensitivity analysis by singular value decomposition ................................................... 224
3.2.2. Results of the parametric sensitivity analysis and discussion........................................ 226
4. Conclusions ............................................................................................................................... 233
5. References ................................................................................................................................. 233
Chapter 7 Conclusions and recommendations ........................................................................... 235
1. Research questions and approach .............................................................................................. 236
2. Conclusion on model reduction approach with test results ........................................................ 237
2.1.
Generation of model reduction options by a decomposition approach ......................... 237
2.2.
Application fo the structure-retaining model reduction approach................................. 238
2.2.1. Model reduction for complex processes with simple products ...................................... 238
2.2.2. Model reduction for simple processes with complex products ...................................... 239
3. Recommendations ..................................................................................................................... 241
3.1.
Model reduction for complex processes with simple products ..................................... 241
3.2.
Model reduction for a single unit with complex products ............................................ 242
3.2.1. Application of the structure-retaining approach ............................................................ 242
3.2.2. Selective model refinement ........................................................................................... 242
3.2.3. Preparing for model validation and application ............................................................ 243
3.3.
Decomposition of the process for model reduction ...................................................... 243
4. References ................................................................................................................................. 245

Chapter 8 - Appendix ...................................................................................................................... 247

1. Summary of differential operations involving the mathematical operators in cylindrical
coordinates ........................................................................................................................................ 248
2. Mass and population conservation equations............................................................................. 249
a. The population balance equations for the dispersed phase .................................................. 249
b. The mass conservation equations for the pseudo-continuous phase .................................... 250
c. The mass conservation equations for the frozen layer domain ............................................ 252
3. Momentum conservation equations ........................................................................................... 253
4. Energy conservation equations .................................................................................................. 254
a. The energy conservation equations for the fluid bulk ......................................................... 254
b. The energy conservation equations for the ice particles in the bulk domain ....................... 255
c. The energy conservation equations for the frozen layer domain ......................................... 257
d. The energy conservation equations for the freezer wall ...................................................... 258
e. The energy conservation equations for the coolant ............................................................. 258
5. Averaging procedure along the radial and angular coordinates ................................................. 259
6. Overview of coordinate-averaged mass conservation equations................................................ 261
a. The averaged conservation equations for the bulk domain.................................................. 261
b. The averaged conservation equation for the frozen layer .................................................... 262
7. Coordinate-averaged conservation equation for dispersed phases ............................................. 263
8. Coordinate-averaged conservation equation for the pseudo-continuous phase .......................... 268
9. Averaging of the conservation equation for the ice phase in the frozen layer............................ 271
10. Expression of the volume flow rate ........................................................................................... 273
11. Physical properties for the simulation of the ice cream freezing model .................................... 276
12. The shrinking core model for the melting of an ice particle ...................................................... 277
13. Mass conservation equations for the bulk domain expressed per unit length of freezer ............ 278
14. Conservation property of the overall amount of water .............................................................. 280
15. Singular vectors for the SVD on the model parameters ............................................................. 282
15.1.
Operational parameters ................................................................................................. 282
15.2.
Input parameters ........................................................................................................... 282
15.3.
Freezer design parameters ............................................................................................ 283
15.4.
Empirical parameters .................................................................................................... 283
16. References ................................................................................................................................. 283
Curriculum Vitae ............................................................................................................................ 285
Publications ...................................................................................................................................... 286
Acknowledgement............................................................................................................................ 287

SUMMARY
Model Reduction in Chemical Engineering
Case studies applied to process analysis, design and operation

During the last decades, mathematical models have become widely used for supporting a
broad range of chemical engineering activities, such as product and process design and development,
process monitoring and control, real time optimization of plant operation or supply chain
management. The use of models for these applications allows for a rapid screening of the design
alternatives, better-informed decisions, low cost and time-to-market of the products and processes,
and finally, a better design and operation. Chapter 1 of this thesis discusses the role and the aspects of
the modelling in the chemical engineering area and the continued need for model reduction.
Tremendous advancements have been made in the past years in the development of numerical
techniques, in computing speed and data storage. However, along with the increase in the speed of
computing, the process models have become more extensive and more complicated. Process models
increasingly cover a wider scope, represent a deeper process integration and intensification and have
more details at much smaller spatial scales. Such rigorous models of complicated systems cannot
always be effectively used for applications such as design and optimization. These inverse
problems require repetitive solution of the model in numerous iterative passes and remain critical in
computing time and use of resources. A reduction of the model complexity is required to make a
model-based solution practical.
Finding such a reduction in a systematic way is not at all obvious and it often demands a
significant effort without a fully satisfactory outcome. In many of the current approaches in systems
engineering order-reduction is applied to a model in its entirety, without preserving the underlying
network structure of the process or its multi-scale decomposition. Retaining the meaningful structural
features of a process in a reduced model is a necessity for practical applications, associated with the
(re-)usability over a longer time span.
The first goal of this work is to explore the landscape and reveal a trade-off between two
opposing and unsatisfactory extremes of performing model reduction:
(a) without any regard for underlying process structure
(b) to account for the process structure up to the smallest level, but forcing to perform model
reduction to very small sub-systems, with issues of overall consistency over time-scales
upon re-aggregation into a process model
The second, more practical goal is to show that such reduced process models, accounting for
relevant process structure can be obtained and are indeed suitable for use in design and control. These
goals are the motivation for the research and lead to the results presented in this thesis.

Summary
The novelty of the thesis is in systematizing and exploiting the essential structural features in
model reduction for a chemical process. The structural features of the chemical process models are
classified in Chapter 2 of the thesis in groups of attributes on several levels of model decomposition:
(a) Contextual level, offering information related to the model goals and applications,
hypothesis to be tested, performance indicators
(b) Structure level, which consists of the information related to the following attributes:

Systemic connectivity (flowsheet, units, streams)

Physical attributes (coordinates, resources, species, phases, etc.)

Behavioural attributes (sources-and-sinks type of phenomena, such as reactions,

transfer, transport, etc.)

(c) Mathematical level, including the mathematical variables and relations (differential &
algebraic equations, domain of applicability etc.) representing the behaviour of the process
(d) Numerical level, covering numerical methods used for solving the model and the model
applications
(e) Software implementation level, having computationally efficient implementations of the
model and solver
This gradual model reduction approach aims first at simplifying the systemic and physical
attributes of the structure level of the chemical process model. Only then additional mathematical
(lower number of equations and simplified terms) and numerical (scheme) reductions are selectively
applied to individual compartments or units. In the following step, the reduced models of the
individual units are connected at system level and the reduced model of the full process is obtained. In
this way, the model reduction procedure is able to preserve the essential structural features of the
process. Moreover, the physical meaning of the variables and equations is kept as much as possible.
The pragmatic criteria chosen for the success of a reduction approach are to develop a model
which: (i) is solved successfully in an acceptable amount of time, (ii) is used effectively for practical
applications, and (iii) retains the topological and physical structure of the process to a sufficient level
of detail
The feasibility and the advantages of the approach are tested for two types of process
applications:
(1) A complex process with a relatively simple (one-phase) product, with the iso-butane
alkylation process as a case study

ii

Summary
(2) A second single operation process with a complex product, exemplified on the freezing
step in ice cream manufacture
For the first type of processes, the simplification acts mostly on the representation of the
systemic and the higher level of the physical structure (units) of the chemical process in the model.
The process is decomposed in units in the process flowsheet and these units are modelled
individually. If required, model reduction is applied to the model of these units as well. The resulted
reduced models of the units are then connected in order to obtain the reduced model of the full
process.
The application and the advantages of the new reduction approach are tested in Chapter 3 for
the assessment of the plantwide control structures for the iso-butane alkylation plant. Chapter 4
discusses the application of the approach when dealing with the dynamic optimization of the plants
operation. The behaviour of this reduced model is compared with a rigorous model of the process.
Moreover, a comparison is done with another reduced model obtained by applying the classical
model order-reduction approach to the rigorous model of the plant as a whole. The application in
Chapter 3 is solved successfully in a very short amount of time, while the dynamic optimization in
Chapter 4 requires a longer computation time, due to initialization issues.
For the second type of processes (complex product formation) the simplification acts on the
lower levels of the physical structure (species, phases, domains, compartments), as well as on the
behavioural level (phenomena).
The development of a dynamic model to support process design and operation for the freezing
step in ice cream manufacture is described in Chapter 5 of this thesis. For this particular application,
the development of a rigorous model covering all the phenomena in detail is conceptually quite
conceivable. Nevertheless, it is currently impossible to solve such a rigorous model because many
high-dimensional laws of conservation will arise and the knowledge regarding the parameters of the
rate phenomena occurring inside the system is still limited. For this reason, a model reduction is
performed; in this case both conceptually during the model development, as well as during the
numerical stage of the modelling procedure. For this application, a reduced model is solved in steady
state (Chapter 6), in an acceptable amount of time, being able to predict the right trends for the model
outputs. In addition, the model is used successfully for a sensitivity analysis of the model parameters
type of problem.
Follow-up questions are related to (a) what is lost in the models predictive accuracy when the
reduction is done by decomposition of the process in entities, (b) how deep a decomposition should
go for a particular application, and (c) whether the models of process units were reduced enough to
achieve the desired goals. The issue of consistent reduction over models of the decomposed elements
and their re-aggregation should also be investigated further. Moreover dynamic model solution,
model validation and optimization-type of applications of the ice cream freezing model are still

iii

Summary
required to demonstrate further the applicability of the reduced model for process and product design
and operation applications.

iv

SAMENVATTING
Model Reduction in Chemical Engineering
Case studies applied to process analysis, design and operation

In de laatste decennia worden wiskundige modellen steeds meer gebruikt ter ondersteuning
van een brede reeks van chemisch technologische activiteiten, zoals product- en procesontwerp en
ontwikkeling, procesbewaking en regeling, real-time optimalisatie van procesvoering en integraal
ketenbeheer. Gebruik van modellen voor deze toepassingen kan leiden tot een snelle screening van
ontwerp alternatieven, beter genformeerde besluiten, lage kosten en korte marktintroductietijden van
producten en processen en, ten slotte, tot betere ontwerpen en procesvoering. Hoofdstuk 1 van dit
proefschrift bespreekt de rol en inhoudelijke aspecten van modelbouw in het chemisch technologische
vakgebied en de aanhoudende behoefte aan modelreductie.
Er zijn geweldige vooruitgangen geboekt in de voorbije jaren in de ontwikkeling van
numerieke technieken, in rekensnelheid en data opslag. Echter, parallel met de toename van
rekensnelheid zijn de procesmodellen omvangrijker en ingewikkelder geworden. Procesmodellen
bestrijken in toenemende mate een breder gebied, representeren diepere procesintegratie en intensificatie en bevatten meer details op een veel kleinere ruimtelijke schaal. Zulke grondige
modellen van ingewikkelde systemen kunnen niet altijd effectief worden gebruikt voor toepassingen
zoals ontwerp en optimalisatie. Deze omgekeerde (inverse) problemen vereisen het herhaald
oplossen van het model in talloze iteraties en blijven zo kritiek in rekentijd en gebruik van andere
hulpmiddelen. Reductie van de rekencomplexiteit van het model is nodig om een modelgebaseerde
oplossing praktisch mogelijk te maken.
Het vinden van een dergelijke reductie op een systematische manier is zeker niet
vanzelfsprekend. Het vereist vaak een aanzienlijke inspanning zonder een volledig bevredigende
uitkomst. Want, bij veel van het huidige aanpakken in systems engineering wordt modelreductie
toegepast op een model in zijn geheel, zonder de onderliggende netwerkstructuur van het proces of
zijn meerschaligheid te behouden. Het bewaren van zinvolle structuuraspecten van een proces in een
gereduceerd model is noodzakelijk voor praktische toepassingen in verband met bruikbaarheid van
het model over een langere periode.
Het eerste doel van dit werk is het landschap te verkennen en een afweging duidelijk te maken
tussen twee tegenovergestelde en onbevredigende uitersten bij het uitvoeren van een modelreductie:
(a) Zonder enig aandacht voor de onderliggende processtructuur
(b) De processtructuur tot op het fijnste niveau meenemen, maar daardoor gedwongen te zijn
de modelreductie uit te voeren op heel kleine subsystemen, met kwesties als overall
consistentie over tijdschalen bij het weer samenvoegen tot een procesmodel.

Samenvatting
Het tweede, meer praktische doel is aan te tonen dat zulke gereduceerde procesmodellen, die
rekening houden met de relevante processtructuur, kunnen worden verkregen en inderdaad geschikt
worden bevonden voor gebruik bij ontwerp en regeling. Deze doelen zijn de motivatie voor dit
onderzoek en leiden tot de resultaten die in dit proefschrift staan.
Het nieuwe van dit proefschrift is gelegen in het systematiseren en benutten van essentile
structuuraspecten bij modelreductie voor een chemisch proces. De structuuraspecten van chemische
procesmodellen worden geclassificeerd in Hoofdstuk 2 van dit proefschrift in termen van groepen van
attributen op verscheidene niveaus van modelcompositie:
(a) Context niveau, dat informatie biedt met betrekking tot het doel en de toepassing van het
model, de hypothese die moet worden getest en indicatoren voor de geleverde prestaties
(b) Structuur niveau, dat informatie verschaft over de volgende kenmerken:
Knopen en verbindingen in het systeem (flowsheet, eenheden en stromen)
Fysische attributen (cordinaten, type behoudsvariabelen, componenten, fasen,
etc.)
Gedragsattributen (bron en putverschijnselen zoals reacties, overdrachten,
transporten, etc.)
(c)

Wiskundig niveau, dat gaat over variabelen en relaties die in de vorm van differentiaal en
algebrasche vergelijkingen het procesgedrag beschrijven, met ongelijkheden als
afbakening van het toepasbaarheid gebied

(d) Numeriek niveau, betreffende numerieke methoden voor het oplossen van het model voor
toepassingen
(e) Software niveau, voor realisatie van een rekenkundig efficinte implementatie van model
en oplossingsmethoden in software
Deze gefaseerde aanpak van modelreductie beoogt eerst de systeemkundige en fysische
attributen op het structuur niveau van het chemische proces en zijn model te vereenvoudigen. Pas
daarna worden selectief mathematische (minder vergelijkingen en vereenvoudigde termen) en
numerieke (algoritmisch) reducties toegepast op individuele bouwstenen in het model. In de
vervolgstap worden de gereduceerde modellen van de afzonderlijke eenheden weer verbonden op
systeem niveau en wordt het gereduceerde model voor het complete proces verkregen. Op deze wijze
is de modelreductieprocedure in staat essentile structuuraspecten van het proces te behouden.
Bovendien blijft de fysische betekenis van variabelen en vergelijkingen zoveel mogelijk bewaard.

vi

Samenvatting
De pragmatisch gekozen criteria voor het succes van een reductieaanpak zijn dat een model
wordt ontwikkeld dat (i) wordt opgelost in een aanvaardbare tijdspanne, (ii) effectief kan worden
gebruikt voor praktische toepassingen en (iii) de topologische en fysische structuur van het proces in
voldoende mate van detail behoudt.
De haalbaarheid en de voordelen van de ontwikkelde aanpak van modelreductie worden
getoetst aan twee types van procestoepassingen:
(1) Een complex proces met relatief simpele (een-fase) producten, met als voorbeeld het isobutaanalkylatie proces
(2) Een enkele proceseenheid met een complex product, met de vriesstap bij het make van
roomijs als voorbeeld
Voor het eerste type proces richt de vereenvoudiging zich in hoofdzaak op de representatie
van het systeem en van de hogere niveaus van de fysische structuur (eenheden) van het chemische
proces in het model. Het proces wordt ontbonden in eenheden in de procesflowsheet en deze
eenheden worden afzonderlijk gemodelleerd. Zonodig, wordt modelreductie toegepast op de modellen
van deze eenheden. De zo verkregen gereduceerde modellen van de eenheden worden dan weer
verbonden tot een gereduceerd model van het gehele proces.
De toepassing en de voordelen van de nieuwe reductieaanpak worden getoetst in Hoofdstuk 3
voor de beoordeling van alternatieve fabrieksbrede regelstructuren voor het iso-butaanalkylatie
proces. Hoofdstuk 4 bespreekt de toepassing van deze aanpak wanneer het dynamische gedrag bij
procesvoering wordt geoptimaliseerd. Het gedrag van het gereduceerde model wordt vergeleken met
dat van een grondiger model van het hele proces. Bovendien wordt een vergelijking gemaakt met het
gedrag van een ander gereduceerd model, dat is verkregen door toepassing van een klassieke
reductiemethode op de orde van een grondig model van het gehele proces. De toepassing in
Hoofdstuk 3 wordt succesvol opgelost in zeer korte tijd, terwijl de dynamische optimalisatie in
Hoofdstuk 4 een langere rekentijd vereist als gevolg van rekenkundige initialisatie kwesties.
Voor het tweede type proces (vorming van een complex product) werkt de vereenvoudiging
op zowel de diepere niveaus van de fysische structuur (chemische componenten, fasen, domeinen,
compartimenten) als de beschrijving van de verschijnselen die het gedrag bepalen.
De ontwikkeling van een dynamisch model ter ondersteuning van ontwerp en procesvoering
voor de vriesstap bij het maken van roomijs wordt beschreven in Hoofdstuk 5 van dit proefschrift.
Voor deze specifieke toepassing is de ontwikkeling van een zeer grondig model dat alle relevante
verschijnselen in detail beschrijft conceptueel zeer wel denkbaar. Niettemin is het thans praktisch
onmogelijk een dergelijk model te op te lossen omdat er veel hoogdimensionale behoudswetten in
zullen voorkomen en kennis van de snelheidswetten van de verschijnselen en hun parameters
beperkend is. Daarom is op voorhand modelreductie nodig; in dit geval zowel conceptueel bij de
modelformulering als bij de numerieke fase van de modelontwikkeling. Voor deze toepassing wordt
in Hoofdstuk 6 het gereduceerde model onder stationaire omstandigheden in voldoend korte tijd

vii

Samenvatting
opgelost, waarbij de juiste fysische trends van de modeloutputs worden verkregen. Bovendien is het
gereduceerde model met succes gebruikt voor een gevoeligheids analyse met betrekking tot bepaling
van parameters met dominante effecten op de outputs.
De vragen met betrekking tot vervolgonderzoek zijn gerelateerd aan welke informatie verloren
gaat in het voorspellende vermogen van een gereduceerd model als een reductie wordt uitgevoerd
door een ontbinding van een proces in zijn eenheden. Aan hoe diep een decompositie moet worden
doorgevoerd voor een specifieke toepassing en aan de diepte van reductie van modellen van
proceseenheden om het gewenste doel te bereiken. De kwestie van een consistente reductie over de
modellen van de ontbonden eenheden en hun samenvoeging moet ook verder onderzocht. Bovendien,
voor het model van de vriesstap voor roomijs is het nodig de mogelijkheden van dynamische
simulatie, model validatie en optimalisatie toepassingen aan te tonen ten behoeve van proces en
product ontwerp en toepassingen bij procesvoering.

viii

REZUMAT
Model Reduction in Chemical Engineering
Case studies applied to process analysis, design and operation

n ultimele decenii, modelele matematice sunt folosite pe scar tot mai larg pentru o gam
extrem de extins de aplicaii n ingineria chimic. Cteva exemple n acest sens sunt proiectarea i
dezvoltarea de procese i produse, monitorizarea i controlul proceselor, optimizarea n timp real a
operrii proceselor sau managementul aprovizionrii. Utilizarea modelelor n aceste situaii permite o
evaluare rapid a alternativelor de proiect, decizii informate, costuri reduse i lansare rapid pe pia a
produselor i proceselor. Toate acestea au ca rezultat o mai bun proiectare i operare a proceselor
chimice. Capitolul 1 al acestei teze trateaz rolul i particularitile folosirii modelelor n ingineria
chimic, precum i necesitatea continu de a le reduce complexitatea.
Un progres extraordinar a fost fcut n ultimii ani n domenii precum dezvoltarea de metode
numerice pentru rezolvarea modelelor, creterea vitezei de calcul i stocarea datelor. O consecin a
acestor evoluii este c modelele proceselor au devenit mult mai detaliate mai complicate. Modelele
au un scop din ce n ce mai larg, descriu procese integrate i intensificate i consider tot mai multe
detalii la scale din ce n ce mai mici. Astfel de modele exacte ale unor sisteme complexe nu pot fi
ntotdeauna utilizate ntr-un mod eficient pentru aplicaii precum proiectarea sau optimizarea de
proces. Aceste probleme inverse necesit rezolvarea n mod repetat a modelului ntr-un numr
mare de iteraii i reprezint nc un punct critic n ceea ce privete folosirea resurselor de calcul.
Reducerea complexitii reprezint un pas necesar n abordarea unei probleme de inginerie chimic
utiliznd modele matematice.
Reducerea la nivel sistemic a unui model nu este ntotdeauna evident, de cele mai multe ori
rezultatul final nefiind satisfctor n ciuda unui efort considerabil. Numeroase metode din ingineria
sistemelor consider modelul ntregului proces, urmrind reducerea numrului de ecuaii, fr a ine
cont de structura procesului sau de modul n care acesta se poate descompune la diverse scale.
Pstrarea caracteristicilor structurale ale procesului este o necesitate pentru utilizarea modelului redus
n aplicaii practice, asociat cu posibilitatea de a (re-)utiliza modelul pe o perioad mai lung de
timp.
Un prim scop al acestei teze este analiza domeniului pentru a dezvlui o imagine de ansamblu
a posibilitilor de reducere a complexitii modelului. Dou extreme au fost identificate n acest caz:
(a) Metode care nu iau n considerare structura intern a procesului
(b) Metode care iau n considerare structura intern a procesului pn la cel mai mic detaliu
i reducnd complexitatea unui numr mare de subsisteme foarte mici. Dezavantajul unei
astfel de abordri este pierderea consecvenei globale de-a lungul scalelor de timp n
momentul reasamblrii subsistemelor pentru a obine modelul ntregului proces

Rezumat
Un al doilea scop mult mai practic este de a arta c astfel de modele cu complexitate redus,
care in cont de structura intern a procesului, pot fi obinute i sunt ntr-adevr potrivite pentru
utilizarea n proiectare i control. Aceste scopuri reprezint motivarea pentru cercetare i a condus la
rezultatele prezentate n aceast tez.
Noutatea acestei teze const n sistematizarea i utilizarea caracteristicilor structurale
fundamentale ale modelelor proceselor chimice n scopul reducerii complexitii acestora. Aceste
caracteristici sunt clasificate n Capitolul 2 n grupuri de atribute, pe diverse nivele de descompunere
a modelului:
(a) Nivelul contextual, care ofer informaii referitoare la scopul i aplicaiile modelului,
ipotezele care sunt verificate, indicatorii de performan
(b) Nivelul structural, compus din informaii legate de urmtoarele caracteristici:

Conectivitate la nivel sistemic (flowsheet, operaii unitare)

Fizice (coordonate, resurse, specii, faze etc.)

Comportamentale (fenomene de generare sau consum de resurse, precum reacii

chimice, transfer, transport etc.)

(c) Nivelul matematic, alctuit din informaiile legate de variabilele i relaiile matematice
(ecuaii difereniale i algebrice, domeniul de aplicabilitate etc.) care definesc comportamentul
procesului
(d) Nivelul numeric, cuprinznd metodele folosite pentru rezolvarea modelului i a aplicaiilor
practice
(e) Nivelul implementrii software a modelului, care include modul de aplicare a algoritmilor
de rezolvare folosind tehnica de calcul
Reducerea complexitii modelului urmrete, ntr-un prim pas, simplificri la nivel structural,
n special al conectivitii sistemului i al structurii fizice, a modelului procesului chimic ce trebuie
modelat. O dat ce acest pas a fost realizat, tehnici de reducere suplimentar a complexitii la nivel
matematic (numr de ecuaii i termeni simplificai), respectiv numeric (metode matematice) sunt
aplicate n mod selectiv modelelor de operaii unitare sau compartimente. La pasul urmtor, modelele
reduse ale operaiilor unitare sunt conectate la nivel sistemic pentru a obine modelul redus al
ntregului proces. n acest fel, procedura de reducere a complexitii poate pstra caracteristicile
structurale fundamentale ale procesului chimic i semnificaia fizic a variabilelor i ecuaiilor.
Din punct de vedere practic, o metod de reducere este performant dac: (i) modelul redus
este rezolvat ntr-un timp acceptabil, (ii) modelul redus poate fi utilizat n mod eficient pentru aplicaii

Rezumat
practice i (iii) modelul redus pstreaz structura topologic i structura fizic a procesului la un nivel
de detaliu suficient de ridicat.
Aplicabilitatea i avantajele acestei proceduri de reducere a complexitii modelelor sunt
testate pe dou tipuri de procese:
(1) Un proces complex cu un produs relativ simplu (monofazic), folosind procesul de alchilare
a izobutanului drept exemplu
(2) Un proces simplu (o singur operaie) cu un produs complex (multifazic), avnd drept
exemplu operaia de congelare parial din procesul de fabricare a ngheatei
Pentru primul tip de proces, simplificarea consider n principal modul de reprezentare a
nivelului structural, n general conectivitatea la nivel sistemic i nivelele superioare (operaii unitare)
ale structurii fizice a procesului chimic de modelat. Procesul este descompus n operaii unitare care
apoi sunt modelate individual. Reducerea complexitii se aplic modelelor individuale, acolo unde
este cazul. Aceste modele reduse sunt apoi combinate pentru a obine modelul redus al ntregului
proces.
Capitolul 3 testeaz modul de aplicare i avantajele procedurii de reducere atunci cnd se
dorete evaluarea structurilor de reglare a instalaiei de alchilare a izobutanului. Capitolul 4 trateaz
modul de aplicare al procedurii de reducere n scopul optimizrii dinamice a procesului. Rspunsul
modelului redus este comparat cu cel al unui model detaliat al procesului i cu un model redus obinut
prin aplicarea procedurilor clasice de reducere a complexitii, care acioneaz asupra modelului
ntregului proces. Aplicaia practic din Capitolul 3 este rezolvat cu succes ntr-un timp de calcul
foarte scurt, n timp ce optimizarea dinamic din Capitolul 4 necesit un timp ceva mai lung, din
cauza unor probleme de iniializare.
n cazul celui de-al doilea tip de proces (formarea unui produs complex) simplificarea
acioneaz la nivelele inferioare ale structurii fizice (specii, faze, domenii, compartimente), precum i
la nivel comportamental (al fenomenelor). Obinerea unui model dinamic redus utilizabil pentru
proiectarea i operarea procesului de congelare parial a ngheatei este descris n Capitolul 5 al
tezei. Pentru acest tip de aplicaie, deducerea unui model riguros al procesului, care s descrie toate
fenomenele n detaliu, este conceptual posibil. Cu toate acestea, rezolvarea unui astfel de model
detaliat nu este posibil n prezent din cauza numeroaselor ecuaii de conservare i a informaiilor
limitate asupra vitezei fenomenelor care au loc n proces. Din aceste motive, reducerea complexitii
se face att la nivel conceptual pe parcursul dezvoltrii modelului, ct i la nivel matematic. Pentru
aceast aplicaie, modelul redus prezice corect variaia variabilelor de ieire ale modelului, n regim
staionar (Capitolul 6), i poate fi rezolvat ntr-un timp relativ scurt. n plus, modelul redus este folosit
cu succes pentru analiza senzitivitii parametrilor modelului.
Rmn de investigat n continuare chestiuni referitoare la: (a) ce se pierde n capacitatea de
prezicere a modelului atunci cnd reducerea complexitii se face prin descompunerea modelului n
uniti individuale, (b) ct de detaliat trebuie realizat aceast descompunere i (c) dac modelele

xi

Rezumat
individuale au fost reduse suficient pentru atingerea scopurilor dorite. Descompunerea modelului n
elementele componente ntr-un mod consistent i re-conectarea acestora trebuie investigate n
continuare. n plus, pentru a demonstra aplicabilitatea la proiectarea i operarea procesului de
congelare parial a ngheatei, sunt necesare rezolvarea n regim dinamic, validarea folosind date
experimentale i aplicarea pentru rezolvarea unor probleme de optimizare.

xii

1
INTRODUCTION

The reduction of chemical process models is investigated in this PhD thesis. In view of the inherent
complexity of model reduction, an exploratory approach is chosen by means of case studies. The motivation for
this research came from the particularities of the model development for chemical engineering applications. This
Introduction sets the stage for the use of models in chemical engineering. It outlines the main benefits models
bring for applications in the chemical engineering area. It then shows there is a continuing need of developing
techniques to model process systems with wider scope, deeper integration, more detailed scales. Process models
are increasingly used in inverse problems implying a need for reduced computation times. The possibilities for
model reduction are presented in the framework of the generic model development steps. The key driver of this
thesis research is a wish to preserve the (topological) structure in model reduction. Modelling is often done at
different scales of resolution. Therefore, a dual way of modelling of process models is highlighted, based on
decomposition/aggregation of structure and reduction/refinement of behaviour. It is from this dual perspective that
the scope of this thesis is introduced together with a statement of the scientific questions that are dealt with in the
thesis. The chapter is concluded with an outline of the thesis.

Chapter 1

1. Role of modelling in chemical engineering

Mathematical modelling has a long history in chemical engineering. In their review of 50
years of mathematics in chemical engineering (Ramkrishna & Amundson, 2004) wrote: The
beginning of this era was marked by the concerted effort of a few to raise the mathematical
consciousness of the profession to think fundamentally about processes. With the advent of
fundamental insights in transport phenomena and reaction kinetics and the introduction of the digital
computers in the 1960s, mathematical modelling in chemical engineering became a pillar in process
development, design and operation.
At first, the use of mathematics aimed at formulating models that could be solved analytically
(Constales, Marin, Nicolis, Van Keer & Yablonsky, 2003). The efforts were focused on
circumventing the limitations of early computers, mainly the lack of speed and storage capacity for
handling very large problems (Ponton, 1995). Mathematics and chemical engineering developed in
parallel, chemical engineering becoming the source of inspiration for new developments in areas like
nonlinear science (Constales, Marin, Nicolis, Van Keer & Yablonsky, 2003). In (Ilievski, 2010), the
advances in numerical algorithms satisfy a law that is similar to the Moores law describing the
advances in chip technology. Moreover, with the increase in computational power, wider ranges of
applications become possible to solve, not only for chemical processes, but also for materials and
products as well (Ramkrishna & Amundson, 2004).
In spite of tremendous advancements in the development of both numerical techniques and in
computing and data storage, the complexity of process models, and hence of the model applications
has risen as well. For this reason, the reduction of the model continues to play an important role in
modelling applications.

1.1. Model applications and benefits

In recent decades, models have become widely used for handling chemical engineering
activities at all levels, during the entire product and process lifecycle, as shown in Table 1.1. The
references presented in this table are primarily chosen for being good illustrations of the model
application in chemical engineering area, and not necessarily for all being state-of-the art reviews.
The use of models for all these applications has several important advantages: saving of time,
efforts and costs; enhanced pace of innovation; deeper understanding of the phenomena that take
place inside the systems; better designs and manufacturing excellence.

Significant savings or

enhanced revenue is achieved during the process lifecycle by using models (Cameron & Ingram,
2008). However, models do not come free. The modelling requires very significant efforts, with
associated costs. It is often an iterative process to reach the specified goals and functions of the
models. In addition, model applications may constrain available resources, such as computing times
and, mainly, human efforts to formulate a problem and analyse the computational results in an
effective and timely manner.

Introduction
Table 1.1. Model applications in chemical engineering activities
Model applications

References
Betlem, Rijnsdorp & Azink, 1998; Heiszwolf, Engelvaart,
van den Eijnden, Kreutzer, Kapteijn & Moulijn, 2001;

Understanding fundamental phenomena

Van den Bergh, Ban, Vlugt & Kapteijn, 2009

Gani, 2004; Eden, Jrgensen, Gani & El-Halwaqi, 2004;

Product development and design

Costa, Moggridge & Saraiva, 2006; Charpentier, 2009

Bongers, 2009; Bongers & Almeida-Rivera, 2009

Product applications

Jaksland, Gani & Lien, 1995; Badyga & Bourne, 1997;

Van Goethem, Barendregt, Grievink, Moulijn &

Process research and development

Verheijen, 2008; Lutze, Gani & Woodley, 2010

Kiparissides, 1996; Saldivar & Ray, 1997; Taylor &
Krishna 2000l; Bermingham, Verheijen & Kramer, 2003;

Process design and engineering

Van Lith, Betlem & Roffel, 2003; Almeida Rivera,

Swinkels & Grievink, 2004; Bildea & Dimian, 2008
Henson, 1998; Qin, 1998; Srensen, Macchietto, Stuart &
Skogestad, 1996; Jockenhvel, Biegler & Wchter, 2003;

Process operations: monitoring, control, real-

Mendez, Cerda, Grossmann, Harjunkoski & Fahl, 2006;

time optimization

Bos, 2006; Martin Rodriguez, Cano & Matzopoulos, 2010

Production planning and scheduling

Mendez, Henning & Cerda, 2000; Pinto, Joly & Moro,

Supply chain management and integration

Applequist, Pekny & Reklaitis, 2000; Chen & Lee, 2004;

2000; Kopanos, Puigjaner & Georgiadis, 2010

Georgiadis, Tsiakis, Longinidis & Sofioglou, 2011

The advantages of modelling in chemical engineering applications will be explained further in

more detail, with a focus on process engineering.
Modelling makes possible the understanding of the fundamental phenomena and their
interactions that take place in the process. The models capture and retain quantitative and qualitative
technological knowledge. Thus, better-informed decisions can be taken at all levels of design and
operation.
The process models are usually developed from fundamental chemical engineering, physics
and chemistry relationships and validated against observed data. The experimental effort that is
required for supporting the selection of unit operations, their scale-up, or the development of process
policies and control systems could increase a lot during the development stage (Marquardt, 1996).
When developing a new process, the knowledge regarding the nature and behaviour of relevant
phenomena that take place inside the system is often limited, hence the possibility of modelling them
from first principles is limited too. As a result, most of these systems are represented by empirical
models based on experimental data. A high number of experiments for obtaining the necessary data
will increase the cost and the time-to-market of the product or process. This number is reduced, for

Chapter 1
example, by using the model for the design of optimal experiment procedures & experimental
conditions for determining its parameters, like for example in (Francheschini & Macchietto, 2008).
Moreover, the use of models accelerates innovation by examining the design space and rapidly
screening the design alternatives. In order to improve customers satisfaction, companies need to
develop new ranges of products, different from what is already on the market. In the same time, the
products are required to meet tighter environmental and safety regulations. In this case, it has to be
decided the best alternative to meet requests for designing new processes and equipment or for
making simple improvements to existing ones. In some cases, processing units or even the whole
plant is redesigned in order to improve the performance (for example, the production of fuels by
Fischer-Tropsch synthesis Dry, 2002).
Process models are also used for support of plant operation in process simulators, in process
state estimation and performance monitoring, and in real time optimizers, as shown in table 1.1.
The focus of the thesis will be on the modelling of process engineering aspects, where process
engineering is taken in a broad sense, covering process development, design and support of
operations. Models for product design and development can be separated into two distinct categories:
performance structure and structure processing. The relationships between product performance,
expressed in quality factors, as appreciated by customers, and product structure remain outside the
purpose of this thesis. Regarding the structure processing group, the formation of the product
structure and composition in a process is seen as a normal part of the processing operation, and thus
covered by process modelling.
In chemical engineering applications, the system is usually equivalent with a chemical
process. A chemical process is defined as a series of successive physical and chemical operations in
which raw materials are transformed into products. The transformations take place in units such as
reactors, distillation columns, heat exchangers, etc. These units form the core operations of the
process. Often the performance of the product is characterised only by properties such as temperature
and chemical composition, which in turn define density, heat capacity, and viscosity. These products
will be called simple products as their performance depends on the thermodynamic state of a single
phase. In other cases, these properties are not sufficient to describe fully the product. For a product
with a complex internal structure, consisting of different components and phases, more information
will be required to define the relationship between the performance, and the structure and
composition. These types of products will be called complex products.
The characteristics of the chemical process and its modelling are influenced by both the
process and the products structures. For this reason, the processes will be classified as:
(1) processes with simple products
(2) processes with complex products
Applications from both classes will be discussed in this thesis.

Introduction

1.2. Efforts and cost factors for model development and application
The development and the application of a model require significant effort. This effort comes
in terms of human efforts and skills, as well as in terms of time required for the model development.
There is always a trade-off between the cost and the expected benefit of the model. Two stages can be
distinguished when trying to evaluate the effort and the costs: (a) the initial model development and
(b) the model applications with maintenance.
The efforts and costs of the initial model development stage are quite visible: it takes time, it
requires good modelling skills and (expensive) human labour time, and there must be an
infrastructure for model implementation (hardware, licensed software, etc.) and computing (in some
cases massive computing is required). These efforts and costs are quite considerable and must be
explicitly allocated upfront by the problem owner.
The efforts and costs in the second stage, the model application and maintenance are less
visible. The most visible is the cost with tuning a model to its applications. However, the most
important cost factor at this stage is in fact a loss factor. For example, potential benefit and generation
of income is lost by having a model that is performing below its target level.
Time is playing also a critical role in two ways:
(1) Required time for the delivery of a planned application
(2) Required computing times for solving a problem. This can be an important factor in realtime applications. However, for off-line applications a slow turn-around time for a model
computation will slow down the project execution and the generation of critical results.

2. Requirements for model use in chemical engineering

The obvious presumption is that the costs of the model development, use and maintenance
will be offset by the benefits of the model applications. The requirements for model development, use
and maintenance are:
a) At system and goal level: Specifications of the boundaries for the system to be modelled and the
kind of problems to be solved by the model, such as simulation and optimization of performance for
specified external scenarios.
b) At network structure level: Graphical and textual representation of the inner network of the
system being modelled (such as a process flowsheet). This network may have a layered structure. For
example, the nodes in the network represent a process unit, which in itself is a sub-network of
interconnected sub-nodes. These sub-nodes can be geometric domains in units.
c) At network behaviour level: Textual representation of the understanding of the relevant
mechanistic structures (co-existing thermodynamic phases), phenomena and behaviour of nodes and
connecting streams. The model must account for the true system. This means that the model should
predict the behaviour of the process accurately enough for the model application goals. A model must

Chapter 1
account for the uncertainty of the experimental observations. It is very important to understand which
properties of the system are modelled accurately and which are treated superficially (Hjalmarsson,
2005). Moreover, how this affects the performance of the resulted model should be clear. The
understanding of the behaviour of the system can be succinctly captured in modelling assumptions.
d) At mathematical level: Representation of the variables, operators and equations, which form a
mathematical model of the system. Alternative model formulations are well possible. There is a tradeoff between model completeness (in terms of coverage of phenomena) and computing efforts for
solving the model. Once several model alternatives are available, the user should be able to select
them in a systematic way (Verheijen, 2003) in order to guarantee the model performance.
e) At numerical level: The mathematical model(s) can be classified in certain categories (for
example, NLA, ODE, DAE, MINLP) for which classes of numerical methods exist. The mathematical
model is transformed to a corresponding numerical version. This transformation often involves the
replacement of the mathematical operators in the model by equivalent numerical operations.
f) At software level: The numerical model and the solution methods are coded in different ways
into different software environments. The implemented model has a clear structure for potential reuse of the model parts.
g) Availability of interfaces to enter problem data into the implemented model and to present
computed results in suitable formats to the model users.
h) Documentation: The implemented model must be well documented. All the assumptions and
decisions, and the model structure are clearly stated, as well the assessment of the models
performance. Though in many cases background decisions are not reported (Cameron & Ingram,
2008), better documentation of the intermediate steps could provide helpful for future model
development. The information on the different versions of the model is easily retrieved in order not to
have to repeat previous work should model improvement or re-use prove necessary. The different
model alternatives usually derive from each other by detailing and aggregation, by changing or
moving scale, or by reduction within the same scale. It is important that these relationships between
model alternatives are clear, to allow the ease of switching between different versions and
simulations.
In the following, the focus will be on developing models to be used for applications with time
critical computations.

3. Process models and modelling

3.1. Process models
A (process) system is defined as a finite region in some space characterized by a set of
quantifiable variables. This space can be a conjunction of several spaces, among which a physical

Introduction
space, a signal space, and a design space. The system that evolves in time is called a dynamic system.
The definition of a dynamic system includes (Katok & Hasselblatt, 1996):
(1) A phase space X - (also called state space), a set of elements (called points) that
represent possible states of the system in any moment in time
(2) Time t -, which can be either continuous (real) or discrete (integer)
(3) A law of evolution g -, which is a rule allowing, when the state of the system at some
moment in time is known, to determine the state of the system at any other moment in
time. The existence of this law is equivalent to the assumption that the process is
deterministic in the past and in the future. It is assumed that the structure of the law of
evolution does not depend on time.
The formal definition of a dynamical system (Katok & Hasselblatt, 1996) is in this case:
A dynamical system is a triple (X, T, G), where X is a set (phase space), T is either integer or
real, and G is a one-parametric group of transformation of X (with discrete time if T is integer).
Moreover, the system is a combination of interconnected entities, integrated to perform a
specific function.
The model of a system will be defined by a set of relations between properties of the system
that are measured in particular domains (Yang & Marquardt, 2009). The relations between the
properties of the system are written in the form of mathematical equations. Moreover, the interactions
between the system itself and the surrounding environment are written in mathematical form as well.
The set of equations (the mathematical model) is the result of the modelling, defined as a goal
oriented activity to describe the fundamental behaviour of real or hypothetical systems. Modelling is
both a conceptual process and a work process, with the goal of delivering a model suitable for an
application goal. During the modelling procedure, the modeller gathers information from the system
of reference for the development of a model suitable for a defined goal.

S12

2
S31

S23
3
2

Figure 1.1.Structure model of a process

Chapter 1
A common way of visualising a process system is as a multi-level network. The process units
are the nodes of the network, while the connecting streams of resources are the network edges (Figure
1.1). Starting from this network view, a structure model of the process is developed. This model
shows which kind of nodes are present and how are they connected (which streams of resources are
connecting the different units in the process). The structure model of the process is represented
visually by the process flowsheet or topologically by the Boolean connectivity. Moreover, each of the
nodes performs a specific function, which defines a behavioural model of the node. The overall
behavioural model of the process is a combination of the behavioural models of the nodes and the
structure/connectivity model. This pattern can repeat itself several times at different levels of detail.
The behavioural model of a node in a process network may be the result of the structure and
behavioural models of sub-nodes at a lower level of structure. The typical levels that can be
considered for a chemical process are presented in Figure 1.2: process system, units, compartments in
units, phase domains (droplets, bubbles, films, etc.) and interfaces, rate phenomena, etc.

Space

enterprise
site

km
plant

m
process units
mm
droplets, particles, films
m
nanoclusters
nm
molecules
ns

ms

min

day week month year

Time

Figure 1.2. Scales in process modelling (After Marquardt, see Grossmann, 2004)

Having defined the structure model for the process, the different laws defining the
mathematical model are easily written and implemented into a software environment that allows for
both the solution and the use of the model by a user. The modelling is often iterative and can lead to
several alternatives of the model for a specific application. The different stages of the model
development will be detailed in Section 3.3 of this chapter.

Introduction
In the following, the focus of the thesis will be on dynamic physical-chemical models with
description of the process units in continuous space and time. The process units will be integrated
through a network structure into a process flowsheet. The spatial scale runs from particle
(micrometers) to plant size (< kilometre), while the time scale runs from fast (milliseconds) to slow
(hours/days), as shown in Figure 1.2.

3.2. Model classification

Several criteria can be defined in order to classify models: the intended use of the model
(model goals); the internal information (physical, chemical, etc.) contents, including scales; the
mathematical features; the computational features of the model.
Models are related to specific types of transformation that take place in the process. On this
basis, one can distinguish:
(1) Physical- chemical models, which cover the transformation of mass, energy, momentum,
electrical charge, exergy, etc.
(2) Economic models, which cover the transformation of capital
(3) Signal/information based models, covering the monitoring and control actions
(4) Safety, reliability, maintenance models that focus on the availability of operational states
(5) Planning and scheduling type of models focused on the allocation and withdrawal of
resources in time and location
Based on understanding the phenomena inside the system, process models are classified in
three distinct categories (Cameron, Wang, Immanuel & Stepanek, 2005):
(1) Mechanistic or fundamental the model is based on the underlying understanding of the
physics and chemistry of the process. In this case, the conservation of resources and the rate
phenomena are derived from first principles.
(2) Empirical or black box the model is based on experimental data, typically in the form of
time series of input and output data.
(3) Grey-box the model is based on a combination of both experimental data and underlying
physics. Usually, the conservation of resources is derived from first principles, while the
rate phenomena are data generated.
Generally, the mechanistic models are preferred since they will be accurate over a much larger
range of conditions than the empirical models (Bequette, 1998). However, the knowledge regarding
the physics and chemistry of the real case does not always allow for the development of a mechanistic
model. It may also happen that developing fundamental models is too expensive and preference is
given to black box or grey-box models when enough empirical data is available. For this reason,

Chapter 1
many models are of either black- or grey-box type. The models presented in this thesis are of the
grey-box type.

3.3. Model development steps

In order to identify the justification and the options for model reduction it is required to
understand how models are being developed. During the modelling procedure, the modeller gathers
information from existing experimental data, literature and experience to define the structure and
behavioural model of the process. The mathematical model is further implemented into a software
environment that allows for use of the model by a user, as presented in Figure 1.3.
The development of the model for a process takes place in several iterative steps (Lohmann &
Marquardt, 1996; Foss, Lohmann & Marquardt, 1997; Hangos & Cameron, 2001; Preisig, 2010) each
of them having as a result one or more model alternatives. For the following discussions, the
development steps considered are, as summarised in Figure 1.4:
a) Model specification and design. During this step, the scope of the model and its physical and
mathematical content are defined. Different models are used for different purposes. It is important
to have the goal defined as clearly as possible before the actual modelling procedures are
performed. It is customary to identify also future scenarios for business or scientific applications.

System of reference
Goal

MODEL
Physical structure

Software
implementation

Functional structure

Data
Mathematical structure

USE

Solution strategy

Literature

Experience

Figure 1.3. Process modelling environment

A reference system is chosen and the system demarcation from the environment is defined. The
functional specifications, such as inputs, outputs and the performance indicators are specified

10

Introduction
according to the goal of the model. The different entities that form the system and the connections
between them are specified. This represents the internal network of the system, or the topology.
Decisions regarding the assumptions of the model (type of resources, scales of resolution for the
system and the sub-systems, time and geometric space, fundamental state, reactions, etc.) are taken
and the behavioural laws (equations) for the different model components are defined. The degrees
of freedom and the policy for the optimal decision-making are identified. A part of the variables is
defined as known in such way that the number of equations is equal to the number of the
unknowns. Moreover, it is investigated if the model is well posed and if it can be solved (Leitold &
Hangos, 2001). The existence of a solution or the multiplicity of solutions is examined (Unger,
Krner & Marquardt, 1995; Bildea & Dimian, 1998; Kiss, Bildea, Dimian & Iedema, 2002). Once
these checks have been done, for cases when dealing with an inverse problem, some of the
prespecified variables can be released to allow embedding of the model in a nonlinear
programming problem (NLP). Lastly, testing scenarios are specified, for testing the models
functionality and applicability. During this step several model alternatives will be developed,
according to the decisions taken. The result of the model specification and design step is the
mathematical model alternative(s). Two mathematical model alternatives differ by the equations of
the model. The different model alternatives of a specific system are related and derived from each
other, either by model expansion and detailing or by model reduction.

Model
specification
and design

Model

Model

Model

implementation

validation

application

Valid
alternative

Applied
alternative

Mathematical
alternative

Numerical
alternative

Equations

Method &
Software

Parameters &
Model fit with data

Results

Figure 1.4. Model development cycle

b) Implementation of the model for numerical solutions. This step includes the selection of the
software environment and of the suitable numerical methods that will be used for solving the model

11

Chapter 1
and using it for applications. The numerical methods should be stable, accurate and fast enough for
the specified model application. Commercial packages, such as differential algebraic equation
solvers, linear algebra packages, large-scale simulators, NLP solvers etc. are usually used for this
step. Often, further manipulation of the model is required in order to make the numerical
implementation of the model possible. The model can be simplified by applying mathematical
manipulations (Marquardt, 2001) such as model reduction, re-arrangement of the model equations
or transformation to alternative representations of the model. Once the model is implemented, the
correctness of the implementation is verified via case studies with known solutions. This implies
comparing the results of the model with experimental data or just experience. Similar to the
previous step, different mathematical model alternatives will lead to different numerical
implementations of the model. Moreover, depending on the choice of the numerical methods and
software environment, the same mathematical alternative could lead to several numerical
alternatives. The result of this step is the numerical model alternative(s). Two numerical model
alternatives differ by the numerical method used for the solution of the model and by the software
environment used for solving the model, while the mathematical model is the same.
c) Validation of the model. In this step, the verification of the numerical implementation of the
model is performed, as well as the validation of the model in the domain of applications. The
model validation involves the following steps: (1) qualitative assertion of correct behaviour at the
domain boundaries or at other asymptotic conditions; (2) tuning of model parameters using
experimental data, with estimation of the errors and covariance; (3) testing of model predictive
power using other independent experimental data; (4) obtaining information on the predictive
uncertainty over the window of validation. The different parameters of the model are determined
from experimental data and the model is tested by comparing its behaviour with known cases.
Model discrimination of the different mathematical and numerical model alternatives is performed
in order to decide which of them performs satisfactorily for the desired application. Different
approaches (Stewart, Henson & Box, 1996; Burke, Duever & Penlidis, 1997; Stewart, Shon & Box,
1998; Asprey & Maccietto, 2002; Lakner, Hangos & Cameron, 2005; Ye, Meyer & Neuman, 2008)
have been developed and reported for applications in the area of process engineering. Limitations
of the model applications are identified during this step as well. The resulting valid model should
fulfil a set of performance specifications; otherwise, the procedure is iterated to one of the previous
steps. A first attempt is rarely successful in obtaining a proper model; hence, the modelling can also
be a time-consuming task. For this reason, it is very important to spend enough time on collecting
all the necessary information to define it properly. In general, empirical models are valid only for a
limited range, depending on the set of data used to obtain the model. The mechanistic models are
valid for a wider range, but their development is far more complex and time-consuming (Cameron,
Wang, Immanuel & Stepanek, 2005). The result of this step is called a valid model alternative(s).
The different valid model alternatives differ by the values and the number of the model parameters

12

Introduction
obtained during the validation. When working with one model alternative, different values of the
model parameters are obtained when applying different methods of estimation, or when using
different data sets to validate the model. For use of several model alternatives, the number and the
meaning of the model parameters may differ from one alternative to the other.
d) Application of the model with documentation. During this step, the model is used in two ways:
(1) For routine computational applications, such as simulation, parameter identification or control.
This can be done in a direct way, by simulation, or in an inverse way, by achieving a
performance and output related target. The direct way implies the specification of all
degrees of freedom, including initial and boundary conditions. The inverse use of the model
implies the use of a performance criterion, which is optimised by varying over degrees
freedom in a model. Such an optimization can be performed with respect to economics,
parameter fitting, data-model reconciliation applications, etc.
(2) The model or parts of the model can be re-used for other types of applications.
After this step, the alternative model & applications are obtained. The applied alternatives of a
specific model differ by the type of results obtained for alternative scenarios.

3.4. Dual way of modelling: decomposition and aggregation over detail levels
Under certain goals and conditions, for a certain application, different models are obtained
that adequately describe features of the process. This can range from very extensive and fundamental
models with attention to the finest details of phenomena to lumped, input-output type of models, with
a minimum of internal mechanisms represented in the model. The result is a family of models, where
one should be able to distinguish the kinds of relationships that can be obtained between different
models of the same process (Aris, 1994). One way of representing the relationships between the
models of the same process is presented in Figure 1.5. In explaining this Figure, the terminology of
(Aris, 1994) will be followed, which means that the word hypothesis captures modelling goal(s) and
assumptions.
The complexity of a model depends on a series of attributes. Firstly there are physical
attributes (entities), related to the process considered: number of units, number of physical resources
to be considered (mass, energy, momentum, etc.), number of domains, phases, and species, as well as
the number and type of phenomena which need to be taken into account. Secondly, there is a second
group of attributes relating to the mathematical structure of the model (type of equations, number of
variables, and complexity of the problems to be solved, etc.). The attributes that affect the model
complexity will be discussed in more detail in Chapter 2 of the thesis.
Starting from the process , with a set of hypotheses H1 , a first model M 1 is obtained.
For example, the model M 1 is formed by the full set of differential equations obtained from
applying conservation laws for mass, energy and momentum over the system along a spatial

13

Chapter 1
coordinate. Assuming that the model obtained in this way is too complex (for example the momentum
conservation equation is defined by a rigorous Navier-Stokes equation) a second set of hypotheses
( H 3 H 5 ), is required to obtain versions of the model which are easier to solve.
For example, hypothesis H 3 leads to a
steady

state

version

of

the

conservation

H1

equations, while hypotheses H 4 and H 5 simplify

H2

M1

the momentum conservation equations.

The models can be simplified further

H3

(hypotheses H 7 H 9 ) until a desired ratio of

complexity versus accuracy is achieved. The
models M 3 M 8 are derived from model M 1 by
letting certain parameters take on a limiting
value (Aris, 1994). This procedure of obtaining

M3

H4
H6

M4

M2
H5

M5
H8

H7

M6

M7
H9

simple models will be called reduction.

It is possible to derive a second model,
M 2 of

the

process

from

set

of

hypotheses H 2 .

M8
Figure 1.5. Relationships between models
(adapted from Aris, 1994)

This set of hypotheses affects the way the system is defined, for example when the model
M 1 is so complicated (number or complexity of equations, number of parameters, etc.), that it seems

that better insight is gained by drastic simplification. In this way, important features of M 1 are
extracted in a way that allows for easier analysis (Aris, 1994). Once the model M 2 is available, its
complexity can be increased by adding more detail, going in the opposite way to reduction. This
procedure will be called refinement. Once the model reaches a desired level of complexity versus
accuracy, this refinement stops.
In chemical engineering applications, the mathematical modelling of physical systems is
performed by decomposition and aggregation of the system across a hierarchy of appropriately
chosen levels (Klatt & Marquardt, 2009).
The decomposition is performed by breaking the process into smaller entities, which are
modelled individually. Once the model of each entity is available, the original model can be reconstructed by coupling these individual models. This is performed through aggregation.
To conclude, four operations are performed on process models, in sets of two:
Decomposition and Aggregation of STRUCTURE over levels
Refinement and Reduction of detail/complexity of BEHAVIOUR

14

Structure

Introduction

Reduction

Sub-system 0

M1

M1

Aggregation

Decomposition

Refinement

Reduction

Sub-system 1

M2

M2
Refinement

Level of detail
0

Level of detail
1

Behaviour

Figure 1.6. Dual way of modelling of chemical processes

0 = reference level; 1 = reduced/simplified

These four operations are shown in Figure 1.6, where the decomposition and the aggregation
are vertically oriented, while the refinement and the reduction are horizontally presented.
This approach has several advantages. By decomposing the system into smaller entities, the
understanding of the behaviour and the function of each hierarchical level is easier. It is possible to
standardise models of entities (reactors, distillation columns, etc.) which can be easily re-used for
building a different aggregated (complex) model. Once a standardization is available, the modelling
procedures become easier and faster. In the same time, the modelling procedures are flexible.
Commercial software packages, such as Aspen Dynamics or gProms make use of standard model
libraries, which can be easily (re)used and interconnected for developing complex process models.
Lastly, the two previous approaches (refinement and reduction) can be combined when some
parts of the model are reduced and others are refined.
In the following discussion, model reduction approaches will be considered, with focus on the
model attributes.

4. Motivation and formulation of the research problem

Societal demands for better performance of engineering systems pushes for wider integration
at upper scales of systems, while the developments in science and engineering offer opportunities for

15

Chapter 1
high precision and intensification at smallest scales. Thus, chemical engineering systems become
more complicated, having repercussion for the modelling. The models are required to cover a wider
scope and because of the deeper integration between scales in a node of a network and of stronger
interactions between different nodes. In addition, the range of scales is continuously expanding,
covering more details of the phenomena at the micro-scale and supply chain integration at the upper
scale. The conceptual and the mathematical capability to model such complicated systems is not
always matched by the computational power to handle such models in practical applications, like
model validation, design and optimization. Table 1.2 presents some features of the model that makes
them hard to solve.
Model reduction consists of a series of approaches and techniques used for obtaining accurate,
yet computationally less expensive models of process systems.
The state-of-the-art of the model reduction involves a host of diverse methods and techniques,
which can be classified in two distinct categories:

Dedicated conceptual approaches in the physical-mathematical domain. These approaches

are often related to the level of detail of specific phenomena (physical or kinetic lumping,
reduction of the reaction networks, etc.) in chemical engineering systems. Often, the
modelling and the model reduction are focused on one single phenomenon (for example
lumping of species or of reactions).

Numerical, data driven techniques in systems engineering (projection and balancing

methods, etc.)

Table 1.2. Model-inherent complications when solving models

What makes a model harder to solve?

Behaviour

Mathematical operations on functions, like differentiation and integration

Loss of stability
Stiffness by a wide range of scales

Structure

Mathematical perspective

Non-linear terms in equations; e.g. chaotic behaviour by a single logistic equation

High degree of coupling between several equations by some variables, especially

when the coupling induces an irreducible set
Large number of variables and equations

Numerical

Numerical stability

perspective

Convergence
Discrete approximations of continuous mathematical operations

The traditional approaches for reducing the complexity of process models do not always
exploit the structural and the discrete features of the models. The systems engineering model

16

Introduction
reduction techniques consider the model of the process as a whole when attempting to simplify it. The
starting point is a rigorous model of the process. This rigorous model is obtained considering relevant
phenomena that take place in the process at pre-selected ranges of time and spatial scales. In a first
group of approaches, the reduction focuses on the reducing the number of model equations (modelorder reduction techniques), without accounting for the inner structure of the process. Although
significant reduction of the number of equations can be achieved, attention must be paid on the
internal structure of the problem, as well as on retaining the physical meaning of the variables
(Wattanwar, 2010). Often, the differential equations are being transformed into equally hard to solve
sets of non-linear algebraic equations (Van den Bergh, 2005).
The second group of techniques focuses on reducing the complexity of the model equations
(model simplification techniques), without affecting their number. In this case, the physical insight is
retained by the obtained model, as well as the number of equations.
The standard methods for reducing the model complexity are reviewed in Chapter 2 of the
thesis.
Sometimes, the knowledge on the phenomena that take place in the process is limited and it is
not possible to obtain a fully rigorous model. The information on the phenomena or the physical
properties is yet to be obtained. From the first stages of the modelling procedure, the modeller has to
take some decisions on how these properties are modelled. Decisions are being made based on
experience, trial-and-error, etc. on the different attributes of the model, leading to a conceptual
model. Once this model is available, a second step in the model reduction procedure is focused on
reducing the mathematical structure, leaving unexploited the physical insight.
To conclude, the computational efficiency of the models is improved conceptually or/and
numerically. The conceptual model reduction approaches are usually used during the model
specification and design stage. The procedure involves the reduction of the discrete attributes of the
model (species, resources, phases, reactions, domains, coordinates, transfer phenomena, connectivity,
etc.). The number of such discrete structural attributes of a model determines in a multiplicative way
the number of variables and equations. The numerical model reduction techniques come into use as
soon as the mathematical model alternatives are obtained. The procedure involves in this case the
reduction of the number and of the complexity of the functional expressions in the model equations.
The structural aspects of a process model are present in chemical engineering textbooks.
Although model structure is being present to some extent (Aris, 1994; Rice & Do, 1995; Weiss &
Preisig, 2000; Hangos & Cameron, 2001; Yang & Marquardt, 2009), the structuring is not yet treated
satisfactorily in the context of model reduction. Often, the starting point of the modelling is
represented by the equations and the variables as a predefined set.
The overall ambition of this thesis is to take a more fundamental and domain specific
approach than just applying numerical data driven techniques for model order reduction, and to look

17

Chapter 1
at and exploit the underlying chemical engineering structure elements of a model in developing
reduced models as well.
Two extremes of performing model reduction can be considered:
(a) The underlying structure of the process model is not taken into account at all. The numerical
model (order-) reduction techniques are facing a problem of enormous size. Even when results
are obtained, the reduced model has no clear relationship to the underlying flowsheet structure.
When it comes to having an understanding about the reduction criteria, the benefit of this
approach is that these criteria have been consistently applied to the entire model. From the cutoff criteria used for the model reduction, it is clear which modes of behaviour have been ignored
or simplified.
(b) The structure of the process model is taken fully into account up to the smallest level of detail.
The conceptual model reduction is followed by a numerical (order-) reduction and is applied to
each behavioural model in the various nodes and sub-nodes of the network. This approach is
very labour-intensive. Furthermore, the question arises on how to reduce the numerous submodels individually in a consistent way (are the corresponding modes of behaviour retained or
ignored?). There is another problem with respect to the interaction between levels of detail. The
behaviour of the nodes at the smallest level is determined by their imbedding at larger scales. It
is a function of the interaction between the nodes and the emerging conditions. As these are a
priori unknown, it is difficult to tell which phenomena and rate processes are dominant and what
can be neglected.
It is obvious that neither extreme is satisfactory. The key issue is how to find a good balance
between dealing with the structural effects while having a consistent approach to the behavioural
reduction. Answering this issue by sheer theory development is yet too demanding in a general
setting, even for chemical process systems. Therefore, in this thesis an exploratory approach has been
chosen, aiming to identify the interactions in the STRUCTURE-BEHAVIOUR reduction problem by
using specific cases. The main research question is:
Are there ways to develop or adapt model reduction methods to better preserve essential
(spatial) structure elements in chemical engineering models?
Model Reduction in Chemical Engineering Case studies applied to process analysis,
design and operation
The theme is addressed through several research questions:
1)

What are the main structural features of chemical process models that are co-determine the

model size and the computational efforts? Where is the structure being addressed in the
modelling procedures?

18

Introduction
2)

How does the model structure reflect in the reduction approaches?

3)

Is it possible to exploit efficiently these structural features when reducing the complexity of

the process model?

4)

Are the reduced models fit for purpose, and what are the suitable criteria for such fitness?

5. Research approach and outline of the thesis

Finding answers to these research questions is attempted through a case driven approach.
The main topics and contributions of the research are:
a) To identify and systematize the main structural features of process models
b) To identify and systematize the model reduction approaches and techniques, as used for
chemical engineering applications
c) To develop a new, more comprehensive approach that accounts for both the structural and
behavioural features of the process models as well as for their the numerical attributes (number and
structure of the equations), capturing specific contributions of existent techniques
d) To investigate and test the application of these model reduction approaches for common
chemical engineering applications
The focus of the research will not be on developing new numerical model reduction
approaches, but on the conceptual use of the existent techniques. Moreover, automated model
building with reduction approaches is beyond the scope of the thesis.
The first three topics will be addressed in Chapter 2, Process modelling and model reduction,
of the thesis. The network concept with nodes and edges, presented in Figure 1.1, is used for the
process representation, and the main attributes of the process structure are defined. The way the
structure affects the model size and the most important existent reduction approaches are
systematized. A structure-retaining model reduction approach that exploits the essential structural
features of a process is developed in this chapter as well. This approach brings together existent
model reduction techniques and uses them within their specific context and applicability.
The main body of this work, Chapters 3 6, will address the fourth topic. Chapter 3,
Application of model reduction to design of plantwide control systems, describes a new procedure of
applying model reduction that takes into account the inner structure of the process. The idea behind it
is to decompose the model into smaller entities and to apply model reduction to these entities
individually. The resulting reduced models of the entities are connected according to the process
structure in order to obtain the reduced model of the full process. The application of this procedure for
two chemical process applications is presented in Chapters 3 and 4. The assessment of the plantwide
control structures of a complex process with simple structured products is the focus of Chapter 3.
Chapter 4, Dynamic optimization using a reduced process model, presents the design and
implementation of the optimum dynamic operation profile in a large-scale chemical plant. The same

19

Chapter 1
process is considered for both applications: the iso-butane alkylation process, though a different
model is obtained for each application. The starting point of the model reduction procedure is, for
both cases, a rigorous model of the chemical plant.
Chapter 5, Development of a reduced model for ice cream freezing, presents the conceptually
driven development of the reduced model for a process unit with products with a complex structure.
The freezing step in the ice cream manufacture process is the case study for the model reduction. For
this particular application, it is not possible to develop a rigorous model of the process for two
reasons:
(a) The knowledge regarding many of the phenomena that take place inside the system is
limited
(b) The full model with all fundamental conservation laws for all the thermodynamic phases
present will become so complicated that solving it is still far beyond current computational feasibility
for practical applications. The model reduction is performed in this case conceptually, during the
model development.
Chapter 6, Numerical application of the reduced model for ice cream freezing, solves the
steady state version of this reduced ice cream freezing model and tests the numerical feasibility of
applying the model for validation purposes.
Chapter 7, Conclusions and recommendations, summarises the main results and limitation of
the thesis research and recommends directions for future development.

6. References
Almeida-Rivera, C.P., Swinkels, & P.L.J., Grievink, J. (2004). Designing reactive distillation
processes: present and future, Computers and Chemical Engineering 28, 1997
Applequist, G.E., Pekny, J.F., Reklaitis, G.V. (2000). Risk and uncertainty in managing chemical
manufacturing supply chains, Computers and Chemical Engineering 24, 2211
Aris, R. (1994). Mathematical modelling techniques, Dover Publications
Asprey, S.P., & Macchietto, S. (2002). Designing robust optimal dynamic experiments, Journal of
Process Control 12, 545
Badyga, J. & Bourne, J.R. (1997). Interaction between chemical reactions and mixing on various
scales, Chemical Engineering Science 52, 457
Bequette, B.W. (1998). Process dynamics: modelling, analysis and simulation, Prentice Hall, 5
Bermingham, S.K., Verheijen, P.J.T., & Kramer, H.J.M. (2003). Optimal design of solution
crystallization processes with rigorous models, Chemical Engineering Research and Design
81, 893
Betlem, B.H.L., Rijnsdorp, J.E., & Azink, R.F. (1998). Influence of tray hydraulics on tray column
dynamics, Chemical Engineering Science 53, 3991
Bildea, C.S. & Dimian, A.C. (1998) Stability and multiplicity approach to the design of heatintegrated PFR, AIChE Journal 45, 2662

20

Introduction
Bildea, C.S. & Dimian, A.C. (2008). Chemical process design: Computer-aided case studies, WileyVCH, Weinheim
Bongers, P.M.M. (2009). Process and equipment design optimising product properties and attributes,
Computer Aided Chemical Engineering 26, 225
Bongers, P.M.M., Almeida-Rivera, C.P. (2009). Product driven process synthesis methodology,
Computer Aided Chemical Engineering 26, 231
Bos, R. (2006). Monitoring of industrial processes using large scale first principles models, PhD
Thesis, Delft University of Technology, The Netherlands
Burke, A.L., Duever, T.A., Penlidis, A. (1997). Choosing the right model: Case studies on the use of
statistical model discrimination experiments, The Canadian Journal of Chemical
Engineering 75, 422
Cameron, I.T., Wang, F.Y., Immanuel, C.D., Stepanek, F. (2005) Process systems modelling and
applications in granulation: A review, Chemical Engineering Science 60, 3723
Cameron, I.T., Ingram, G.D. (2008). A survey of industrial process modelling across the product and
process lifecycle, Computers and Chemical Engineering 32, 420
Charpentier, J.C. (2009). Perspective on multiscale methodology for product design and engineering,
Computers and Chemical Engineering 33, 936
Chen, C.L., Lee, W.C. (2004). Multi-objective optimization of multi-echelon supply chain networks
with uncertain product demands and prices, Computers and Chemical Engineering 28, 1131
Constales, D., Marin, G.B, Nicolis, G., Van Keer, R., & Yablonsky, G.S. (2003). Mathematics in
chemical kinetics and engineering: An introduction to the special issue dedicated to papers
presented at the MaCKiE-2002 conference, Chemical Engineering Science 58, 4747
Costa, R., Moggridge, G.D., & Saraiva, P.M. (2006). Chemical product engineering: An emerging
paradigm within chemical engineering, AIChE Journal 52, 1976
Dry, M.E. (2002). The Fischer-Tropsch process: 1950 2000, Catalysis Today 71, 227
Eden, M.R., Jrgensen, S.B, Gani, R., & El-Halwaqi, M.M. (2004). A novel framework for
simultaneous separation process and product design, Chemical Engineering and
Processing: Process Intensification 43, 595
Foss, B.A., Lohmann, B., Marquardt, W. (1997). A field study of the industrial modelling process,
Journal of Process Control 19, 153
Franceschini, G., Macchietto, S. (2008). Model based design of experiments for parameter precision:
State of the art, Chemical Engineering Science 63, 4846
Gani, R. (2004).Chemical product design: challenges and opportunities, Computers and Chemical
Engineering 28, 2441
Georgiadis, M.C., Tsiakis, P., Longinidis, P., & Sofioglou, M.K. (2011). Optimal design of supply
chain networks under uncertain transient demand variations, Omega 39, 254
Hangos, K.M., Cameron, I.T. (2001). Process modelling and model analysis, Academic Press

21

Chapter 1
Heiszwolf, J.J., Engelvaart, L.B., van den Eijnden, M.G., Kreutzer, M.T., Kapteijn, F., Moulijn, J.A.
(2001) Hydrodynamic aspects of the monolith loop reactor, Chemical Engineering Science
56, 805
Henson, M.A. (1998). Nonlinear model predictive control: Current status and future directions,
Computers and Chemical Engineering 23, 187
Hjalmarsson, H. (2005). From experiment design to closed-loop control, Automatica 41, 393
Ilievski, Z. (2010). Model order reduction and sensitivity analysis, PhD Thesis, Eindhoven University
of Technology, The Netherlands
Jaksland, C.A., Gani, R., Lien, K.M. (1995). Separation process design and synthesis based on
thermodynamic insights, Chemical Engineering Science 50, 511
Jockenhvel, T., Biegler, L.T., & Wchter, A. (2003). Dynamic optimization of the Tennessee
Eastman process using the OptControlCentre, Computer and Chemical Engineering 27,
1513
Katok, A., & Hasselblatt, B. (1996). Introduction to the modern theory of dynamical systems,
Cambridge University Press
Kiparissides, C. (1996). Polymerization reactor modelling: A review of recent developments and
future directions, Chemical Engineering Science 51, 1637
Kopanos, G.M., Puigjaner, L., & Georgiadis, M.C. (2010). Optimal production scheduling and lotsizing in dairy plants: The yoghurt production line, Industrial and Engineering Chemistry
Research 49, 701
Kiss, A.A., Bildea, C.S., Dimian, A.C., & Iedema, P.D. (2002). State multiplicity in CSTR
separator recycle polymerisation systems, Chemical Engineering Science 57, 535
Klatt, K.U. & Marquardt, W. (2009). Perspectives for process systems engineering Personal views
from academia and industry, Computers and Chemical Engineering 33, 536
Lakner, R., Hangos, K.M., Cameron, I.T. (2005). On minimal models of process systems, Chemical
Engineering Science 60, 1127
Leitold, A., Hangos, K.M. (2001). Structural solvability analysis of dynamic process models,
Computers and Chemical Engineering 25, 1633
Lohmann, B., Marquardt, W. (1996). On the systematization of the process of model development,
Computers and Chemical Engineering 20, S213
Lutze, P., Gani, R. & Woodley, J.M. (2010). Process intensification: A perspective on process
synthesis, Chemical Engineering and Processing 49, 547
Marquardt, W. (2001). Nonlinear model reduction for optimization based control of transient
chemical processes, Proceedings of the 6th international Conference of Chemical
Process Control, AIChe Symp. Ser. 326, Vol. 98, 12
Martin Rodriguez, H., Cano, A., & Matzopoulos, M. (2010). Improve engineering via whole-plant
design optimization, Hydrocarbon Processing 89, 43

22

Introduction
Mendez, C.A., Henning, C.P., & Cerda, J. (2000). Optimal scheduling of batch plants satisfying
multiple product orders with different due-dates, Computers and Chemical Engineering 24,
2223
Mendez, C.A., Cerda, J., Grossmann, I.E., Harjunkoski, I., & Fahl, M. (2006). State-of-the-art review
of optimization methods for short-term scheduling of batch processes, Computers and
Chemical Engineering 30, 913
Pinto, J.M., Joly, M., Moro, L.F.L. (2000). Planning and scheduling models for refinery operations,
Computers and Chemical Engineering 24, 2259
Ponton, J. (1995). Process systems engineering: Halfway through the first century, Chemical
Engineering Science 50, 4045
Preisig, H.A. (2010). Constructing and maintaining proper process models, Computers and Chemical
Engineering 34, 1543
Ramkrishna, D. & Amundson, N.R. (2004). Mathematics in chemical engineering: A 50 year
introspection, AIChE Journal 50, 7
Rice, R.G., Do, D.D. (1995). Applied mathematics and modelling for chemical engineers, John Wiley
& Sons
Qin, S.J. (1998). Control performance monitoring a review and assessment, Computer and
Chemical Engineering 23, 173
Saldivar,E. & Ray, W.H. (1997). Mathematical modelling of emulsion copolymerization reactors:
Experimental validation and application to complex systems, Industrial and Engineering
Chemistry Research 36, 1322
Srensen, E., Macchietto, S., Stuart, G., & Skogestad, S. (1996). Optimal control and on-line
operation of reactive batch distillation, Computers and Chemical Engineering 20, 1491
Stewart, W.E., Henson, T.L., Box, G.E.P. (1996). Model discrimination and criticism with singleresponse data, AIChE Journal 42, 3055
Stewart, W.E., Shon, Y., Box, G.E.P. (1998). Discrimination of fit of multiresponse mechanistic
models, AIChE Journal 44, 1404
Taylor, R., Krishna, R. (2000). Modelling reactive distillation, Chemical Engineering Science 55,
5183
Unger, J., Krner, A., Marquardt, W. (1995). Structural analysis of differential-algebraic equation
systems Theory and applications, Computers and Chemical Engineering 19, 867
Van den Bergh, J. (2005). Model reduction for dynamic real-time optimization of chemical
processes, PhD Thesis, Delft University of Technology, The Netherlands
Van den Bergh, J., Ban, S., Vlugt, T.J.H., & Kapteijn, F. (2009). Modelling the loading dependency
of diffusion in zeolites: The relevant site model, Journal of Physical Chemistry C 113,
17840

23

Chapter 1
Van Goethem, M.W.M., Barendregt, S., Grievink, J., Moulijn, J.A., Verheijen, P.J.T. (2008).
Towards synthesis of an optimal thermal cracking reactor, Chemical Engineering Research
and Design 86, 703
Van Lith, P.F., Betlem, B.H.L., Roffel, B. (2003). Combining prior knowledge with data driven
modelling of a batch distillation column including start-up, Computers and Chemical
Engineering 27, 1021
Verheijen, P.J.T. (2003). Model selection: An overview of practices in chemical engineering,
Computer-Aided Chemical Engineering 16, 85
Wattanwar, S. (2010). Identification of low order models for large scale processes, PhD Thesis,
Eindhoven University of Technology, The Netherlands
Weiss, M., Preisig, H.A. (2000). Structural analysis in the dynamical modelling of chemical
engineering systems, Mathematical and Computer Modelling of Dynamical Systems 6, 325
Yang, A., Marquardt, W. (2009). An ontological conceptualization of multiscale models, Computers
and Chemical Engineering, 33, 822
Ye, M., Meyer, P.D., Neuman, S.P. (2008). On model selection criteria in multimodel analysis, Water
Resources Research 44

24

2
PROCESS MODELLING AND MODEL REDUCTION

The development of chemical process models is performed in a sequence of stages, leading to a specific
internal structure of the model. In this chapter a framework to analyze the model structure, focusing on the
elements that offer opportunities for model reduction is presented. The chapter introduces as well the most
important techniques that are used for model reduction in the chemical engineering area. A novel approach to
perform model reduction is introduced, which is trying to balance the model structural (nodes and edges) and
behavioural (equations) characteristics to obtain a computationally feasible model. This approach takes an
integrative look at all options for model reduction from the outset. It first covers the conceptual phase, where a
mathematical model is developed addressing structural and behavioural features. Then, in the numerical phase,
classical numerically driven (model order-) reduction techniques are applied. The application of this procedure
will be studied in the following chapters of the thesis with the help of case studies that are outlined at the end of
the chapter.

Chapter 2

1. Introduction
A chemical process can be defined as a system that manufactures chemical products from
feeds, supported by utilities and directed by automatic and human control. Any system has a
boundary and an interior. The inside of a process can be seen as a multi-level network. The process
units are the nodes of the network, while the network edges are formed by the streams connecting
these units between themselves and with the outside environment across the boundary. These streams
transport physical resources (species, mass, energy, momentum) or signals. The units effect the
transformations in the states of these resources or signals. The multi-level aspect of the system
appears because many nodes (units) have an internal network structure as well, comprising
interconnected compartments and thermodynamic phases with transfer of resources between them.
Sensors and actuators link the state of the physical resources to signals and reversely.
In this chapter, the term process model relates to a mathematical description of the behaviour
of the internal part of the system, while accounting for the multi-level network structure. The process
model is supposed to have predictive power regarding the evolution of the process behaviour, given
enough conditions at the system boundaries. These boundary conditions reflect the state of the
external environment, to which the process responds. A specification of an adequate set of boundary
will be called a scenario.
Often, model applications in the process engineering area deal with a high dimensionality and
a high degree of coupling between the different nodes of the process (Charpentier & McKenna, 2004;
Vlachos, 2005; Klatt & Marquardt, 2009). This increases the complexity of the models, which
depends not only on the total number of equations that need to be solved, but on the number of terms
in these equations as well. These are determined by both the physical structure (process network) and
the behavioural features (phenomena taking place in the process).
The number of equations and their complexity can be influenced in each model development
step. The model development stage can be sometimes time consuming (Cameron, Wang, Immanuel &
Stepanek, 2005). Predictive accuracy has to be balanced against model complexity and computational
speed, often in iterative modes between model refinements and reductions.
Model reduction is deemed necessary when the computing times for solving a model become
critical for the success of the applications. This is the case in so-called inverse problems, where a
model may be solved very many times to meet target conditions or an optimal performance. Another
case for reduction involves model applications where a model user needs very short turn-around time
of a model simulation because of required fast and frequent interpretation of modelling results (like at
stock exchange trading or in real time control).
Model reduction consists of a set of approaches and techniques that can be used for obtaining
a model that represents the process and its behaviour in a way that is computationally feasible
(Marquardt, 2001; Antoulas, Sorensen & Gugercin, 2000). In this chapter, an inventory of the most
important model reduction techniques available is presented. The procedures will be presented from a

26

Process modeling and model reduction

chemical engineering point of view, identifying the structural and behavioural model elements that
can be subject to reduction.
The chapter is structured in the following way: Section 2 introduces the decomposition of a
physical system for applications in process modelling. Moreover, definitions of the models attributes
are introduced for the contextual and structure levels in Section 3, and for the mathematical,
numerical and software implementation levels in Section 4. The main outcome of this section is a list
of discrete model building attributes that can be used for model reduction or refinement. Section 5
explores existing model reduction techniques and their use for applications in chemical engineering,
taking into account the model decomposition presented in Sections 3 and 4. A structure-retaining
approach to perform model reduction taking into account both the structural and behavioural features
of the process models is presented in Section 6. A case study approach is used in this thesis to assess
the application of the model reduction approach. The selection and motivation of the chosen problems
is elaborated in Section 7, while conclusions are summarised in Section 8.

2. A design view on process modelling

Process modelling is a systematic activity to study, analyse and optimize the behaviour of a
physical system by mathematical and computational means. The first outcome is a mathematical
model whose features can be analysed. The second outcome is a software implementation of the
model that can perform computations for a pre-specified class of operating scenarios. The third
outcome is data obtained from model applications by users in simulations and optimizations.
The inside of a process system is seen as a network of individual entities, which will be
called nodes, connected by edges, as presented in Figure 1.1 in Chapter 1. For chemical
engineering applications, the modelling of the process is performed by developing models for the
individual nodes separately and connecting the resulting models, according to the network structure,
in order to obtain the full model of the process.
In the following, a design view on process modelling and model reduction is introduced. The
underlying philosophy is that the development of a mathematical model for a process is very similar
to an engineering design procedure. A modeller has to make decisions regarding the creation of
structure and behaviour in the model, reflecting the structure and behaviour in the object of
modelling, just as a designer makes decisions regarding the object under design. The engineering
design procedures are often organised in a hierarchical manner, working from an overall perspective
towards fill in of details. This hierarchical decomposition is usually domain specific and takes
advantage of the structure and geometric scales that commonly occur in the class of objects. The
result is a sequence of decision making at increasing levels of detail. One well-known example in
chemical engineering is the hierarchical design method of Douglas (Douglas, 1988) for the conceptual
design of chemical processes. Alternative designs may arise in parallel due to the preferred design
decisions.

27

Chapter 2
In presenting this view, a decomposition of the model development in a sequence of stages is
the common thread. The different levels of decomposition and the different structures obtained during
the modelling procedure are shown. The outcome is an overview of the design decision variables in
model building which can be used for model reduction (or refinement) purposes. Alternative
competitive models may arise due to the decisions taken during the model development.
When reading the following sections, a distinction must be kept in mind with respect to the
decomposition of the modelling procedure (and the model itself) and the structural decomposition of
the object of modelling.

2.1. Levels of decomposition in the modelling procedure

The creation of models and model alternatives is done in the first two stages of the model
development cycle: Model specification and design and Model implementation, shown in Figure 1.4.
Several alternative process models result from the modelling procedure, depending on the
development stage and the models specifications (goal, assumptions, performance, etc.), each with
its own qualities, performance and behaviour.
However, the models of the same process share a common structure, which is extended or
reduced through the modelling process, for a desired set of specifications. This common structure in
process models is related to the generic features of the physical system of reference.
In the following, six sub-levels of decomposition (Figure 2.1) will be discussed: the contextual
level, the structure level, the mathematical level, the aggregation at the system level, the numerical
level and the software implementation level.

Model specification and design

Contextual
level

Structure
level

Mathematical Aggregation
at system
level
level

Model implementation
Numerical
level

Software
level

Figure 2.1. Levels of decomposition of the modelling procedure

For every level of decomposition, a set of decisions regarding the model are being taken and a
representation of the model is obtained. This decomposition of the modelling process is done to
identify the model design options that can be used for model reduction.
At contextual level, information related to the goal and application of the model is contained.
The structure level offers information on the physical structure of the system and the phenomena
(behaviour) considered by the model. The mathematical level includes all the mathematical
relationships describing the structure and the behaviour of the system. At numerical level, decisions

28

Process modeling and model reduction

related to the computational solution strategy of the model are taken, while at software level
information on the implementation of the model in software environment is specified.
An overview of the activities and the outcome for each level of decomposition in Figure 2.1 is
presented in Table 2.1.
Table 2.1. Levels of decomposition of the modelling procedure
Level

Activity in the level

Outcome

1. Define process system to be modelled

systems environment

Input-output

3. Define model goal(s) & application(s)

representation

4. Define hypothesis and scenarios to be tested

5. Specify model inputs & outputs

decomposing the process

Physical
(internal) and

2. Select the phenomena that take place in the

functional

process which are relevant for the intended model

structure

use

Aggregation at

alternative

Structure

1. Derive a suitable network model structure by

Mathematical

6. Specify performance criteria

Mathematical

Model specification and design

Contextual

2. Define the boundaries of the process with the

1. Define the mathematical relationships for the

physical entities behaviour
2. Define the mathematical relationships for the

Mathematical
model structure

connectivity
1. Connect the different model attributes

system level

Network-type
model structure

1. Model analysis with transformations and

scaling to improve solvability
Numerical

solving per node

Solution strategy

3. Model reductions in sub-sets of equations (per

node)
4. System-wide aggregation of the numerical
model (for example, up to multi-scale simulation)
Software

1. Implement the model in a software

environment

Software

Numerical alternative

Model implementation

2. Select the numerical methods/algorithms for

implementation

2. Implement the numerical methods in a

software environment

In Section 3, an inventory of the discrete attributes of the contextual and structure levels is
presented. Section 4 offers an inventory and classification of the attributes at the mathematical level,

29

Chapter 2
as well as the inventory of the model attributes at system, numerical and software implementation
levels. To support the understanding of the decomposition of the modelling process the same
modelling example of a chemical process is presented all along these sections, together with the
exposure of the decomposition approach. For every level detailed in the following, the iso-butane
alkylation process is used as an example.

3. Contextual and structure levels of decomposition

3.1. Contextual level
This level (Figure 2.2) includes the specification of the process system to be modelled, the
goal of the model and the applications it will be used for (dynamics, steady-state simulation, design,
process control, etc.). For each application, the hypothesis to be tested and the performance indicators
are specified. The boundaries of the process system need to be specified, as well as the interaction
between the process system and its environment.
The concept of boundary will be defined as
part of the structure level of decomposition in Section
3.2.1.
The inputs and the outputs are fixed as

Contextual level
Process
system

Goal

Application type

well at this level of decomposition. The inputs

comprise information related to feeds, utilities, mode
of operation, set points for operation, etc. The outputs
include chemical products, energy products (steam,

External
boundary

Scenario
Scenario
Scenario

Scenario
Scenario
Hypothesis

Scenario
Performance
Scenario
indicator

power), waste, emissions, operational and economic

Figure 2.2. Structure of the contextual level

performance indicators, etc.

In the same time, the scenarios for testing are defined. The result of this step is an input-output
type of structure of the process, as shown in Figure 2.3.
The alkylation process is used to produce high-octane blending component for gasoline by
liquid phase conversion of n-butene with excess iso-butane, using acid catalyst. The goal of the
model is the understanding the behaviour of an iso-butane/butene alkylation plant. The desired
product is the iso-octane. The raw materials are the iso-butane and the olefin (1- and 2-butene). The
fresh feed also contains propane and n-butane.
The present commercial routes for alkylation rely on the use of sulphuric acid (H2SO4) or
anhydrous hydrogen fluoride (HF) as catalysts (Hommeltoft, 2001). The process takes place
usually at high pressure, to ensure the liquid phase and low reaction temperatures. For the further
development, the sulphuric acid route is chosen. The alkylation reaction requires a high interfacial
surface area between the acid catalyst and the hydrocarbon. Current commercial alkylation
processes the dispersions are acid-continuous with dispersed hydrocarbon droplets (Albright,
2002).
The plant behaviour will be studied for two scenarios: steady state and dynamic operation.
Two applications are envisaged: the assessment of the plantwide control structures and the design
and implementation of the optimal dynamic operation.
30

Process modeling and model reduction

3.2. Structure level

The simplest way of representing
the process is by using an input output
type of structure (I-O), as shown in Figure

Inputs

Outputs

MODEL

2.3, which is a result of the contextual level

of decomposition. Since usually the process
is seen as a network of interconnected
nodes, the model should be able to represent

Figure 2.3. Input output representation of the process

this as well.
In the same time, each individual node in the process structure performs a specific function,
which can be defined through a mathematical relationship.
The structure level specifies the nodes of the network and the connections between them.
Moreover, the specific functions (behaviour) of each individual node are defined.
Further, the mathematical relationships defining the behaviour of the network nodes, as well
as the connectivity are specified, at mathematical and system level, respectively.
Two sub-levels are identified for the structure of the model:
(1) Physical structure of the model, which relates to the nodes considered and the connections
between them;
(2) Behavioural structure of the model, including the function performed by each individual node.

3.2.1. Process decomposition and physical structure of the model

It is assumed that the process (flowsheet) structure is available. The process decomposition is
performed leading to a reduced/simplified network structure from which the process model can be
built. The physical structure level consists of all the information regarding the nodes in the process
network, as well as the connectivity.
A series of discrete attributes are identified at this level of decomposition. These are specified
according to the way the process network is defined and depending on the type of application which
is considered. When looking at levels in a process flowsheet one commonly distinguishes between:
(0) Process as a system; (1.a) Units in the process (nodes); (1.b) Connections transporting resources
(edges).
Within a unit, the following levels can be distinguished: (2) compartments (for example trays
in the distillation column) and (3) domains (for example the hold-up of process fluids and the heating
agent inside of a heat exchanger).
Inside the domain, one may have: (4) region, (5) phase, (6) species, (7) interface, and (8)
coordinate.
Table 2.2 presents the relations between the different attributes of the model, as considered in
this thesis. Each of the attributes presented in the table are detailed in the following sections.

31

Chapter 2

Phase

Region

Domain

Attribute
Unit

Level

Compartment

Table 2.2. Relations between attributes of the model at structure level

Mass

Resource

Species
Population
Energy
Momentum
Information

Structure

Physical

Unit

Connectivity

Behaviour

Compartment

Domain

Region

Phase

Interface

Coordinate

Sources and sinks

Buffer effect

Stream

Transfer

Transport

Physical

Behaviour

= uses resource; = contains attribute; = involves phenomena; = is connected by

The delimitation between the

Boundary

attributes is done by means of

Phase 1

boundaries and interfaces, as shown

in Figure 2.4. The interface is a
physical

means

of

delimitation

Interface
Boundary
Phase 2

between attributes inside the system,

where the different attributes in the
model come in contact.

Figure 2.4. Boundaries and interfaces

32

Process modeling and model reduction

The interface is defined as a very steep transition region between two thermodynamic phases,
leading to a jump in physical properties.
A boundary represents a conceptual mean of separating among different attributes and
between attributes and the outer world. The boundary is defined in geometric terms and has no
behavioural features. As such it does not influence the number of equations in a model.
For the applications of the iso-butane alkylation plant, the external boundary of the model is
taken as the battery limit of the plant, excluding the utility generation and waste treatment systems
while including all processing operations to turn C3-C4 feeds into alkylates.
Since the process is defined by the transformations it causes to the resources and by means of
these resources, a definition of a resource should be given before discussing the structure in more
detail.
A resource represents any physical or information entity, R , of limited availability that needs
to be used in order to obtain a transformation inside the process.
The most important physical resources considered in modelling of chemical processes are the
mass, energy, momentum. For particular applications, other types of resources can be considered,
such as electrical charge in electrochemical processes. Mass, energy and momentum are treated as
field quantities, being locally continuously distributed over space and time.
The most important information resource is the signal, distributed over time, as used in control
applications.
Some of these resources can be classified into smaller sub-divisions (kinetic energy, potential
energy, work, etc.).
A special, though very relevant in some practical cases, mass resource is showing up as a
population of particulates, being small discrete lumps of matter having an internal state, like mass or
size or energy. Examples are crystals, droplets of liquid dispersed in another phase or gas bubbles in a
liquid phase.
The species represent another group of resources. A species is defined as an ensemble of
identical chemical or biochemical entities that are generated or consumed through a transformation
called reaction.
In the case of the alkylation example, the resources that will be taken into account are the mass,
the energy and the momentum. Since for the reaction the hydrocarbons are dispersed inside the acid,
the hydrocarbon droplets population should be also considered.
Moreover, the utilities: energy (hot, cold), process water, power are available from both external
sources, as well as co-generated in the process.
Alkylation serves to dispose of the C3-C4 cut from the FCC (fluid catalytic cracking) unit by
converting much of this cut into alkylates. The alkylation usually converts iso-butane and n-butenes
into products. Light hydrocarbons such as propane are present as inert. Liquid sulphuric acid is used
as catalyst.

33

Chapter 2

Sulphuric
acid

Sulphuric
acid

ALKYLATION
PLANT

Butene

Reaction

Separation

Light
ends (LE)

iso-Octane
Dodecane

Propane

Others

Propane

Figure 2.5. Input-output diagram for the isobutane alkylation plant

iso-Butane (A)

Butenes(B)

Heavy
ends (R)

iso-Butane

Reactant

Dimethyl
hexanes
(P)

iso-Butane
Butene

Table 2.3. Typical alkylate composition

Type
Species

Product

The main constituents of the alkylation

product are trimethylpentanes (Dimian &
Bildea, 2008). Undesired byproducts, including
light (C3-C5) paraffins and isoparaffins, acidsoluble oils and pseudo-alkylates (mixtures of
C5-C16) are produced in variable amounts,
depending on the employed feed and on the
operating conditions (Albright & Wood, 1997).
The composition of alkylate obtained in the
process depends on the operating conditions,
the number of passes in the reactor.

Catalyst
(C)
Inert (I)

1- & 2Butene

iso-Pentane
2,3-Dimethylbutane
2- & 3-Methylpentane
2,3 & 2,4-Dimethylpentane
2,2,3-Trimethylbutane
2- & 3-Methylhexane
2,2,4-Trimethylpentane
2,2,3-Trimethylpentane
2,3,4-Trimethylpentane
2,3,3-Trimethylpentane
2,5-Dimethylhexane
2,4-Dimethylhexane
2,3-Dimethylhexane
3,4-Dimethylhexane
2,2,5-Trimethylhexane
2,2,4-Trimethylhexane
2,4,4-Trimethylhexane
2,3,5-Trimethylhexane
Sulphuric acid
Water
SO2
Propane

A typical composition (Albright, Spalding, Faunce & Eckert, 1988) is shown in Table 2.3.
The input-output diagram of the alkylation plant is presented in Figure 2.5, showing the main
feeds and the products of the process.
a1) Process unit U represents a part of the process system where a transformation of one
or more resources, R takes place. Each unit has a function (or scope) assigned within the structure of
the system, which is related to the transformation taking place inside the unit. The units can be
classified according to the type of resources are transformed in: (a) physical units, where physical
resources are transformed, and (b) information units, where signals are generated, transformed and
applied.
Once the physical units have been defined, information units are added to the process
flowsheet if required. These units are used for the control of quantities such as reactor feed, pressure,
temperature, concentration of the distillation columns outlet streams, etc.
The STRATCO effluent refrigerated sulphuric acid alkylation unit (Caton, Racen &
Troutman, 2008) is taken as an example. Five major sections are considered, as shown in Figure
2.6. The sections presented in Figure 2.6 do not always represent single process units, but groups of
such units.
(1) Reaction section the reacting hydrocarbons are brought into contact with sulphuric acid
catalyst under controlled conditions
(2) Refrigeration section the heat of reaction is removed & the light hydrocarbons are purged

34

Process modeling and model reduction

(3) Effluent treating section the
free acid, alkyl and di-alkyl sulphates
are removed from the net effluent stream
to avoid downstream corrosion and
fouling
(4) Fractionation section isobutane is recovered for recycle to the
reaction
section
and
remaining
hydrocarbons are separated into the
desired products
(5) Blowdown section spent acid is
degassed, the waste water pH is adjusted
and acid vent steams are neutralized
before being sent off-site

Recycle iso-Butane

Propane

Refrigeration
Acid

Effluent
treating

Alkylation
reaction
iso-Butane

Spent
Acid water

Olefin

Blowdown

Refinery
wastewater
treatment

Fresh
Acid
Spent
Acid Process
water

F
r
a
c
t
i
o
n
a
t
i
o
n

n -Butane

Alkylate

Figure 2.6. Block flow diagram of a STRATCO effluent

refrigerated sulphuric acid alkylation unit (Caton, Racen, &
Troutman, 2008)

a2) Compartment C represents a spatially distinct subdivision of the process unit U that is
separated by a boundary from the rest of the unit.
a3) Domain D represents a spatially distinct subdivision of the geometric space of the
compartment C that is separated by an interface I from another domain. Usually, a domain consists of
a single phase. A second phase may be present, where usually one is continuous, while the other is
dispersed.
To illustrate the concepts of compartment and domain for the alkylation plant a heat
exchanger will be considered.
The
following
Table 2.4. Domains inside a heat exchanger
compartments are identified in this
Domain
case: the tubes, the shell, the wall
Compartment
and the insulation.
Process
Heating
Wall
Insulation
Four domains are present
stream
agent
(Table 2.4): the process stream
Tubes
Liquid
phase 1
whose temperature needs to be
increased, the heating agent, the
Shell
Fluid
exchanger wall and the insulation.
phase 2
Each of these domains is
Wall
Solid
characterised by a different type of
phase 1
phase that is present inside the
Insulation
Solid
domain.
phase 2
a4) Region represents a subdivision of a domain D of a compartment C characterized by
a common characteristic behaviour inside the region. As well as the domains, two distinct regions are
separated by an interface. The domain is a spatial notion, while the region is contained inside a
domain and has certain distinctive behavioural features, for example the transition from laminar to

35

Chapter 2
turbulent flow, or the sudden extinction of reactions. The interface of the region cannot be specified in
terms of a function of a spatial or time coordinates, but it is defined by the passing of an inequality
constraint on a set of state variables. A region may also be defined as a physical boundary layer
within a domain and forming an interface between a wall and the bulk of a fluid.
a5) Phase represents thermodynamic phase as a subdivision of a region or a domain. A
system may be characterized as homogeneous (single phase with 1 ) or heterogeneous
(multiphase with 2 ) according to the number of distinct phases. Two different phases are
separated by an interface I through which transfer of resources may take place. The interface may be
a planar interface (as shown in Figure 2.7b) or a domain structure with regular or irregular islands of
one phase embedded in a surrounding matrix of a different phase, as shown in Figure 2.7a).
The alkylation reaction takes place
in liquid phase, with the hydrocarbons
dispersed in the acid. Two phases are
considered in this case (see Figure 2.7a):
(1) a continuous liquid acid phase and
(2) the dispersed hydrocarbon liquid
droplets. In case of the distillation
columns, two interconnected phases are
considered for each tray (see Figure
2.7b): (1) a liquid and (2) a vapour
phase. Mass and energy are transferred
through the interface between the two
phases.

Liquid
phase

HC

Mass

HC
H2SO4

HC

Interface
Energy

HC

Vapour
phase

Figure 2.7. Schematic representation of phases considered

in the alkylation process: a) Reactor; b) Distillation tray

3.2.2. Behavioural structure of the model

The attributes of the behavioural level correspond to the transformations that take place inside
the nodes considered inside the process network. The following groups can be distinguished in this
case:
b1) Sources and sinks . This group of attributes generates or consumes resources, such as

reactions or birth and death processes in case of particulates. For example, a reaction k is defined as a
mechanism through which a set of species, called reactants, are being transformed into a set of
different species, called products. Depending on the level of decomposition desired, reactions take
place at phase, region, domain, or unit level.
The production of iso-octane involves the liquid-phase reaction of 1-butene with iso-butane un
the presence of a second liquid phase of sulphuric acid. The alkylation reaction is accompanied by
several other acid-catalyzed reactions such as olefin oligomerization, product cracking and product
isomerisation, all of which result in the formation of several inferior products. The mechanism
usually cited for iso-butane alkylation involves a carbenium ion chain mechanism initiated by the
protonation of the olefin (Hommeltoft, 2001). Some basic reactions are (Lee & Harriott, 1977):

36

Process modeling and model reduction

A) Chain initiation:
k1

C4 H 2 SO4 C4 HSO4

(2.1)

k1

B) Main chain:
k2

i C4 C4 C8

(2.2)

k2
k3

C8 i C4 i C4 C8

(2.3)

k3

C) By-product reactions:
k4

C8 C4 C12

(2.4)

k4

k5

C12 C5 C7

(2.5)

k5
k6

C12 C6 C6

(2.6)

k6

k7

C12 C7 C5

(2.7)

k7

b2) Buffer effects (change in hold-up). A resource R can be stored or depleted in a region or

domain. The determination of hold-ups and changes therein are done only in dynamic models
considering time derivatives of hold-ups of resources.

3.2.3. Model connectivity

The final group of attributes offers information related to the way the different entities inside
the process network are connected between themselves and with the environment.
c1) Streams. The stream S represents a moving quantity of resource R that is being

transferred from a process unit to another, as shown in Figure 2.8. The streams are also transferred
between the process and the environment .
Three categories of streams are considered in most of the process application: material, energy
and information streams. In case of the material streams, the decomposition can go even further.
A material stream S can be subdivided into different phases, and species . For the
purpose of this research, the stream is considered just as an interconnection between different entities
of the network. In some cases, within a physical stream, a transformation of the resource R may take
place. If this happens, the stream will be considered as a unit. A stream has two components:
(1) Topological component, which defines the units connected by a particular stream, and the
direction of the stream
(2) Behavioural component, which indicates the amount of resource being moved from one
unit to the other

37

Chapter 2
c2) Transfer. A transfer process is defined as a mechanism through which a resource is

being transported, through a boundary or interface, from one node of the system to the other. Species
mass diffusion between phases and radiation from a hot surface into a fluidic volume are examples of
transfers. Depending on the level of decomposition desired, the transfer takes place at phase, region,
domain, or unit level. In tranfer, the diffusive flow is usually dominant.
c3) Transport. A transport process J is defined as a mechanism through which a resource
R is being transported into, through and out of a distinct region. Depending on the level of

decomposition desired, the transport takes place at phase, region, and domain or unit level.
Physically, transport is achieved by convective and diffusional flows. While the transfer takes place
between phases, for example, the transport is possible inside of a single phase. The convective flow is
usually dominant in transport.

S1R2

B()

U1

S1R1

R1
2

S
U2

S
S

R2
2

R1
3

U3

U4

R1
6

S 4R1
S5R1

U2

U3

S1R2

D2
R2
J1,1

S 2R2

D3

R1
3,4

S 4R1

I
D1

S 2R1

S3R2

S3R1

S3R1

D4

S5R1

Figure 2.8. Model decomposition:

a)

Streams and units; b) Domains and transport; c) Phases and transfers

For the iso-butane alkylation, the streams are visualised in Figure 2.9, which presents an
example of the plants flowsheet from Dimian & Bildea, 2008. A more detailed flowsheet is
presented in Caton, Racen & Troutman, 2008.
Examples of transformations that involve the transfer of a resource for the iso-butane
alkylation plant are presented in Table 2.5. The table is not complete, but offers an idea of the
variety of units required in the process.

38

Process modeling and model reduction

F0

i-C4H10

i-C4H10, C4H8 recycle

F1

LC

F0
C4H8
C3H8

C3H8

FC
FC
PC
FC

PC

LC

PC

LC

X
TC

H2SO4
TC

FEHE

Acid
catalyst
removal

LC
TC

FC

i-C8H18

C3
C8

C4
FC

TC

TC
LC

LC

LC

Heavies
Spent
H2SO4

Figure 2.9. Flowsheet for the iso-butane alkylation plant (Dimian & Bildea, 2008)
Table 2.5. Examples of physical process units for the alkylation plant
Unit
Function
Transformation
Reactor
Heat
exchanger
Flash
Distillation
column
Settler
Coalescer
Mixer
Pump
Compressor
Pipe
PID controller

Alkylation
Feed cooling
Condense depropanizer feed
Remove heat of reaction
Remove propane
Remove iso-butane
Remove n-butane
Remove alkylate
Acid removal

Resource

Reaction

Mass

Heat exchange

Energy

Type of
transformation
Rate process

Transfer

Separation
Mass

Feed water removal

Feed mixing
Reactor inlet stream
Reactor vapour stream
Transport feed & products
Control process operation

Mixing
Pressure
change

Momentum

Transport
Computation

Mass
Information

Transport

4. Mathematical, numerical and software implementation levels

This section aims to present a systematic overview of the structure and attributes of a process
model and its implementation. A discussion of common elements of this structure and certain
attributes can be found in textbooks on process modelling (Hangos & Cameron, 2001) as well. There
it is spread out over several chapters. The purpose of presenting this concise overview is to set the
stage for a systematic subsequent discussion of model reduction options, according to the presented
structure and attributes. Furthermore, the systematic presentation of the model equations plays also a
vital role in Chapter 3, 4 and 5 of this thesis. The mathematical level is split in the behavioural models
of nodes (Section 4.1), connectivity of nodes to a network (Section 4.2) and the aggregated network
model (Section 4.3)

39

Chapter 2

4.1. Mathematical level: behavioural models of nodes

Once the information related to the physical structure level is defined, the result is a networktype of representation of the model similar to the one in Figure 2.8, together with details on the
phenomena that take place in each individual block.
At this point, the mathematical

Mathematical

relations defining the behaviour of the

Coordinate

nodes are written. In many cases, for a

certain type of entity a general equation is

Quantities

written, which is further refined according

to a specific case. The mathematical

Variable

Mathematical
operations

Parameter

Equations

Differential

Algebraic

relations relate closely to the structure of the

model, accounting for the number and the

Domain of
applicability

nature of the model discrete attributes.

Figure 2.10. Structure of the mathematical level

At the behavioural mathematical level, the main attributes that are identified are the model
equations, which can be classified in differential and algebraic, as seen in Figure 2.10.
Within an equation, one can identify the different quantities (variables and parameters), as
well as the mathematical operations (differentiation, integration, summation, power, etc.).
In the following, a classification of the different components of the models equations is
presented.

4.1.1. Coordinates
The coordinate system consists of a set of variables used to determine in a unique way the
position (in time and in space) within units of the process . The most common used coordinates in

product and process modelling are: (a) the time coordinate t ; (b) the spatial coordinate z ; and (c) the

internal coordinate s of a particulate.

Except for the time, the coordinates presented above are usually vectors, containing a number
of components related to the degree of detail used or desired in the decomposition of the process.
There can be a generic set of coordinates for the entire system or the various nodes considered
in the process network may have a different set of coordinates. For a specific process unit, a domain
has its own set of coordinates. For example, in a fixed bed reactor, the fluid phases flowing through
the bed have one set of spatial coordinates and the catalyst particles have another set of spatial
coordinates, while sharing time as a joint coordinate.
The coordinates have two main roles when defining the model. They may occur as variables in
the equations, as in explicit functions in initial and boundary conditions or in rate functions. The
second role is in the definition of the mathematical operators occurring in the model equations, such
as differentiation or integration.

40

Process modeling and model reduction

4.1.2. Classification of quantities
The mathematical structure of the model represents relationships between quantities
describing the system. These quantities are classified in two categories: variables and parameters. The
classification of the quantities does not offer model reduction options.
b1) Variables. This category contains the quantities that vary along the systems coordinates.

The following groups are identified, following the concept of a state space model:
1)

State variables x . Represent quantities indicating the hold-up of a resource in a region or

domain, for a dynamic system. Conservation laws are written for the state variables and the
behaviour of the system is characterised by the evolution of its states along the coordinates. From
the way system variables change with the coordinates, the systems can be classified in two
categories (Bequette, 1998):
i) Lumped systems none of the variables of interest depends on any spatial coordinate
ii) Distributed systems at least one variable of interest depends on the spatial position z or
on an internal coordinate s

2)

Input variables, u . An input variable represents the quantity that is linked and determined by the
external environment of the system considered for the model. The variation of the input variables
affects the value of the state variables. The evolution in time of an input variable must be specified
before a problem can be solved. Inputs are usually chosen based on the knowledge of the system
considered. This group of variables is classified in:
i) Control (manipulated) variables - are (externally) modified in a controlled way as part of a
scenario, in order to achieve desired performance. In this category, one can distinguish
between (a) inflow variables, on which the user does not necessarily have a direct influence,
and (b) operational variables on which the user has a direct influence.
ii) Disturbances - the modification does not take place in a controlled way. This group of
variables may have a stochastic nature.

3)

Output variables y . The output variables represent the quantities derived from the state
variables, which are of direct interest to the model user. The inputs modify the state of the system
that further has an effect on the outputs. The output variables offer information related to the
performance of the model and represent a link to the external environment. They are used for the
extraction of relevant information for the user of the model. Two categories are identified: (a) the
outflow variables, and (b) the quantities that relate to the quality of the (product) output.

41

Chapter 2
4)

Algebraic variables (as co-states related to the constitutive equations). The algebraic (or the
auxiliary) variable v represents a group of quantities derived from the input and state variables.
Within this category one can distinguish:
(a) Thermodynamic state variables v1 (enthalpy, entropy)
(b) Physical property variables v2 (density, specific heat, viscosity, conductivity, diffusion)
(c) Rate of change variables r , represent the rates by which the resources are produced or
consumed inside the system. A typical example of change variables are the chemical reaction
kinetic rates.
(d) Fluxes of resources in transport and transfer rate processes

5)

Algebraic variables related to the connectivity by streams, . This group of variables involves
quantities such as the composition and the flow rate variables of streams connecting process unit.

6)

Boolean or logical variables related to the nodes and edges being active in a flow sheet. These
variables are useful in modelling for process synthesis by means of a topological superstructure.
The nodes and streams have each logical variables, whose value (on/off) indicate whether a node or
a stream is considered to be active in the flowsheet. Such variables are not considered in this thesis.

b2) Parameters. Represent quantities that do not vary along the systems coordinates or with
changing external conditions. These quantities are completely specified in the context of use, in order
to be able to solve the problem mathematically. Parameters are often adjusted to achieve desired
performance of the model. This group is classified in:
i) Design parameters, , representing the sizing and the number of the units
ii) Empirical behavioural parameters, , taking away uncertainties in the model equations by
tuning to experimental data

b3) Range and units. This category offers information related to the lower and upper limits for
the variation of the different quantities. Units of measurement are also required.

4.1.3. Classification of equations

The number and the variety of the attributes at physical structure level influence the number of
variables, as well as the number and the complexity of the equations. The complexity of a set of
equations depends more on the nonlinearity and the number of the terms occurring in an equation, on
the mathematical type of operation within such terms (derivatives, integrals, etc.), and on the coupling
between the equations by jointly occurring variables, than on the number of equations. Three views
on defining the types of equations are possible:
(1) Functional, systems view: input, change, output

42

Process modeling and model reduction

(2) Mathematical view: linear, nonlinear, algebraic, differential, integral
(3) The physical view: conservation laws, constitutive equations for thermodynamic state and
equilibrium, property relations, rate expressions
In the following section, the focus is on the mathematical view in presenting the equations,
while the classification will be done taking into account the physical view. The classification of the
model equations does not offer model reduction options.

c1) Differential equations. This group of equations arises from conservation laws or from
mathematical transformations. The behaviour of the domain is modelled by applying the conservation
principles to the resources of the system. The resource as such is conserved in the domain, although it
may change its amount. The conservation of the resource signifies that what enters the domain should
leave, is transformed into another quantity, or it accumulates in the domain. The mass, energy and
momentum are strictly conserved, while the species amounts may transform.
In general, a conservation equation is written in the form of a resource balance, often called an
equation of change:
accumulation

resource in resource out resource resource (2.8)

of

resource through boundary through boundary generation consumption

For a general input-state-output system, the set of differential equations will be written as:

dx t , z , s
dt

f x t , z, s , z x t , z, s , s x t , z, s , u t , z, s , v t , z, s , ,

(2.9)

Here t is the time variable, z is the spatial coordinate vector, and s is the internal coordinate.
The variable x t , z , s n is the state variable vector, u t , z , s m is the input variable vector,
v(t , z , s ) q is the algebraic (co-state) variable vector, is the empirical parameters vector,

while is the design parameter vector. In the above equation n is the state space dimension. The
function f describes the dynamics of the domain.
In case of distributed systems, a (parabolic or hyperbolic) partial differential equation (PDE) is
obtained in the form of equation (2.9). For lumped systems, the result is an ordinary differential
equation (ODE):
~

d x t
dt

f x t , u t , v t , ,

(2.9.a)

43

Chapter 2
In the following, the focus will be on the physical interpretation of the model equations. When
modelling process systems, the most important quantities for which conservation equations are
written are the mass (for pseudo-continuous phases) or the population (for distributed phases), the
energy, the momentum and the electrical charge. For the purpose of this thesis the Maxwell-type of
equations for electro-magnetic fields are not considered.
1)

Mass balances. When mass is considered as a resource being transformed, the generic balance
equation is written for any of the species present in the domain, in the form of the mass balance
equation (see Appendix 1). The number of independent balances is equal to the number of species
considered.

2)

Population balances. The population balance equation (Appendix 2) is the form the generic
balance equation takes for any of the populations of entities of dispersed phases considered in a
domain, when these entities have a distributed internal state.

3)

Momentum balances. In a similar way, a generic momentum balance can be written for the cases
when the momentum is the resource transformed inside the domain. The generic momentum
balance equation is presented in Appendix 3. The number of the independent balances is equal to
the number of spatial coordinates considered.

4)

Energy balances. In the generic form, the balance equation (see Appendix 4) is written for the
total energy. The total energy is the sum of internal, kinetic and potential energy, all of which are
associated to the mass. Work is a summation of various other energy effects, such as mechanical
work, volume work, electrical work, etc.

c2) Initial and boundary conditions

Together with the differential equations of the domain, additional relations should be specified
to complete the model to define known values of the state variables x on the domains boundaries.
Depending on the type of boundary being used, these additional constraints are classified as follows:
i) Initial conditions. These constraints define the values of the state variables at the lower
boundary of the time axis:
0 f 0 x, u , v, ,

(2.10)

t 0

ii) Boundary conditions. These constraints define the values of the state variables on a specific
geometric boundary of the domain. They appear in three major forms (Hangos & Cameron, 2001): (a)
first type or Dirichlet, (b) second type or Neumann, and (c) third type or Robbins.
0 f BC x, x, u , v, ,

(2.11)

44

Process modeling and model reduction

c3) Algebraic equations for co-states. A third set of equations characterizing the processes
occurring in a domain is represented by relationships expressing thermodynamic equilibrium, reaction
rates, transport rates for heat, mass, momentum, physical properties and so on. Several alternatives
are available for writing these equations for a particular physico-chemical system.
In general, this type of equations is written in the following form:

0 h x t , z , s , u t , z , s , v t , z , s , ,

(2.12)

The function h describes the (algebraic) constitutive equations for the rates, the thermodynamic
equilibrium and the physical properties of the phases.
The algebraic equations for co-states are classified as follows:
1)

Thermodynamics equations of state. The equations of state are thermodynamic equations

between state variables of the system, for a given set of physical conditions (OConnell & Haille,
2005).

2)

Physical property equations. The physical properties of a system are not constants. Processing
condition may vary in a wide range during operation. Moreover, the system consists of a variety of
materials (pure species, mixtures, and phases) in any state (liquid, solid, gas). Information
regarding the physical properties (for example density, viscosity, thermal conductivity, diffusivity,
etc.) of the materials is required to complete the model. Equation (2.12) will be used to write this
type of relations.

3)

Equilibrium equations. These equations are written for systems with chemical, thermal,
mechanical or electrical potential equilibrium.

4)

Rates of transfer, transport and change. The differential equations describing the system (see
Appendix 1-4) consider a number of terms of transfer. This number of terms is determined by the
number of distinct rate phenomena that influence the system . Examples of such terms that appear
in a typical process are sources and sinks like net convection, diffusion (or dispersion), interfacial
transfer of resources, generation (or consumption) in reactions, etc. Relations describing these
types of transfer are derived and added to the model. The complexity of these terms adds to the
overall complexity of the model.
All these terms will be written in the general form expressed in equation (2.12).

c3) Output equations. The relations between the state variables x and the output variables
y are usually defined by algebraic equations. This can be written in a general form as:

y t , z, s g x t , z , s , v t , z , s , ,

(2.13)

45

Chapter 2
Here y (t , z , s ) p is the output variable vector. The function g describes the way in which
the observations are deduced from the state and the algebraic variables. The dimension of the input
vector, m and the dimension of the output variable vector p , are much smaller than n , and
usually m p .
It has to be mentioned that the output equations in (2.13) do not show the internal gradients of
the states with respect to the coordinates t , z , s . Mathematically, this would be consistent with the
state differential equations for a distributed system in equations (2.9). However, in practice, such
gradients are not easy to measure. For this reason, the use of the states is a more adequate choice for
the output equations.
These equations do not always have to be written in an explicit form, but an expression
relating the state variables to the output variables should always be possible to derive.

4.1.4. Validity and applicability conditions (knowledge constraints)

The knowledge constraints are usually defined as inequalities, which result from the
translation of the modelling assumptions (Hangos & Cameron, 2001). These inequality constraints
define a set in the space of all variables entering a constitutive equation (rate equation,
thermodynamic equation, equation of state) over which the equations can be reliably applied.
0 j x t , z, s , v t , z, s , ,

(2.14)

4.1.5. Types of operations in model equations

Another set of attributes at mathematical level is represented by the different type of
operations present in the model equations, which are determined by the behaviour of the attributes in
the physical structure.
The operations range from simple algebraic operations such as addition/subtraction , ,

multiplication/division , / , derivation/integration , , power/logarithm ^,log and to more

complex special functions such as Bessel functions, etc.

4.1.6. Model classification

Depending on the type of functional expression on the right-hand side of the set of differential
and algebraic equations, the models are classified in:
i) Linear models contain models for which the functional expressions in equations (2.9)
(2.14) are written as a linear combination of the variables. For a linear model, the model
equations are written in the following way:
46

Process modeling and model reduction

dx t , z , s
dt

A t , z, s x t , z, s B t , z, s u t , z, s C t , z, s v t , z, s

y t , z, s D t , z, s x t , z, s E t , z, s v t , z, s
0

(2.15)

F t , z, s x t , z, s H t , z, s u t , z, s G v t , z, s
ii) Nonlinear models for these models the functional expressions are not written in linear
form. Most of the real systems are of nonlinear nature. Explicit solutions to nonlinear
differential equations can rarely be obtained (Bequette, 1998).

4.2. Mathematical level: Connectivity

The connection between the nodes of the networks (shown in Figure 2.11) is an important
feature of the chemical processes. In the following, a way of accounting for the connectivity between
the different units (attributes) of the model in the model equations will be presented.
The connectivity needs to be accounted for in the process model as well. Two types of streams
are considered in Figure 2.11: (1) streams that connect the unit k to the outside environment, and (2)
streams that connect the unit k to other units in the process network. The first type of streams is
defined by the input variables u , while the second type is defined by the flow of resources being
transported through stream j . This vector of streams j is made up of flow rates of the species,
R

momentum, energy, etc. in this stream.

Outside environment

Stream
j

Unit
k

j 1
R

j 2
R

Unit

Unit
Figure 2.11. Connection of units by streams

47

Unit
k+1

Chapter 2
The stream vectors are ordered in a stream matrix, M :
R

M 1

R
... j

(2.16)

A unit (as a node) will have some gates through which it is connected with streams inside the
network. A new flow vector is introduced, representing the set of flows per node k through its gates g.
There is one flow per gate and there may be several gates for a node. A flow through a gate g of a
node k, is denoted by g , k .
R

These flows will be arranged in a flow matrix:

M
R

R
1,1
R

2,1
...
R
k ,1

1,2
R

2,2
...
R

k ,2
R

... 1, g
1,3

R
R
2,3 ... 2, g

... ... ...

R
R
k ,3 ... k , g
R

(2.17)

The flows must be connected with streams inside the network, and they belong to a node.
For a lumped system, a differential equation, containing the matrix of flows M is written.
R

This matrix is post-multiplied with the unit vector, to obtain a linear addition of flows.
~

d x t
dt

R
f x t , M , u t , v t , ,

(2.18)

For a distributed system, the internal balance equation does not change:
dx t , z , s
dt

f x t , z, s , z x t , z, s , s x t , z, s , u t , z, s , v t , z, s , ,

(2.19)

In this case, the flows appear in the model equations through the boundary conditions:

R
0 f BC x, x, M , ,

(2.20)
B

When the boundary conditions are given at the inlet only, it is necessary to add an additional
set of outlet boundary conditions. The function of this additional set is to relate an outgoing flow,
g , k with the internal state x at the outlet.
R

Obviously, the flow vectors g , k are unknown and must be specified by making a connection
R

with the streams. Each flow through a gate, g , k will be connected with a stream, j . This can be
R

48

Process modeling and model reduction

done by means of a connectivity matrix for a node, Ck . This matrix has entries c that can take the
R

following values:
1

c0
1

out

(2.21)

no connection
in

The connectivity matrix of node k is build up in the following way:

Ck

R
c1,1
R
c
2,1

c R
g ,1

c1,2
c1, j

c2, j

R
R
cg ,2 cg , j
R

(2.22)
k

The row index counts the flows, while the column index counts the streams. Each row and
column of this connectivity matrix have only one non-zero element. A row (of length j ) is associated
with a gate g and the non-zero element in this row selects the stream j that will be connected to the
flow through this gate. A column (of length g and having a single non-zero element) is associated
with a stream and indicates through which gate g of a unit k this stream is passing.
The flow matrix is connected to the stream matrix for each node k:

R
, k

Ck M

M C M

(2.23)

The matrix C

(2.24)

is a column wise matrix, collecting Ck , with k 1,..., K . The matrix C

R

can have only two non-zero elements of opposite signs (-1 for a source and +1 for a sink). The matrix
M is a matrix collecting all the flow matrices M , k , with k 1,..., K .
R

4.3. Mathematical level: Aggregation at system level

The model equations presented in the relations (2.9) (2.14) do not account for the model
connectivity. Since the connectivity is an important property of the chemical process models, as
mentioned in the previous sections, the equations will be rewritten in order to account for this feature.
The result is a set of differential and algebraic equations (PDAEs) shown in Figure 2.12, which
represents both the internal and behavioural structure of the model.
Once the models of the single entities (nodes and streams) are available, the next step in the
model specification and design step (see Figure 1.2 in Chapter 1) is the coupling of these individual

49

Chapter 2
models to obtain the full model (see Figure 2.12) of the process , presented in Figure 2.8. The
coupling is performed by connecting the individual entities through the streams of resources
(material, energy, momentum, information). The output stream of one entity is the input stream of the
following entity, as the flowsheet of the application shows.
At mathematical level, the interconnection is represented as an algebraic equation (equality)
between sub-models variables such as mass flow, mass fraction, specific enthalpy, heat flow,
temperature, entropy, pressure, etc. These equations act as constraint equations for the state variables
and can lead to a high index problem.
Initial & boundary
conditions

Inputs
uk

dxk
f k xk , z xk , s xk , vk , uk , k , k , k
dt
0 hk xk , vk , uk , k , k , k
0 f k ,0 xk , vk , uk , M , k , k , k

t 0

, k

0 f BC ,k xk , vk , M ,k , k , k , k
0 j xk , vk , k , k , k

Outputs

yk g k xk , vk , k , k

k
Physical
properties

Design
parameters

Figure 2.12. Input-output and behavioural structure of the model

For this reason, it is important to know which input variables sets do not result in a high index
problem and how to interconnect the sub-models in such way to ensure that the resulting model does
not have a high index structure. The requirement for an index of at most one is that equation (2.12)
can be solved in implicit form. This is equivalent to the algebraic equation set should have Jacobian
of full rank with respect to the algebraic variables (Hangos & Cameron, 2001).
To detect the cause of the high index problem it is important to know (Soetjahjo, 2006):
a) if the state variables constraints are due to internal constraint equations within the model, or due
to the choice of the input variables, or due to the interconnection of the sub-models
b) which variables appear in state variables constraints
c) how to solve the index problem, either by symbolic index reduction or by special numerical
handling

50

Process modeling and model reduction

In practical terms, avoiding high index problems is equivalent to specifying the variables that
can be manipulated in practice. For example, for the dynamics of the distillation column the reflux
must be specified, and not the composition of the distillate. The occurrence of high index problems
can create a trade-off in model reduction: by simplifying equations by elimination of fast phenomena
and imposing constraints on state variables, a reduction in the number of variables and equations can
be obtained. This potential gain can, however, be removed by the creation of a high-index problem.

4.4. Relationship between model structure and computational complexity

In the previous sections, a classification of the model equations was introduced. It has to be
noted that the classification is useful for understanding the model structure, but does not offer handles
or structural elements to change the number of model equations.
The complexity of the model is determined by the following factors:
(1) Number of model equations
(2) Number of terms in the model equations
(3) Type of operations in the model equations
Table 2.6. Discrete decision attributes at model structure level

Physical structure

Level

(Scalar)
Phenomena

Connectivity
(directional
phenomena)

Attribute

Counter

Resource, R

nR n

Unit, U

nU

Compartment, C

nC

Domain, D

nD

Region,

Phase,

Coordinate, C

nC

Boundary, B

nB

Sources & sinks,

Buffer effect,

Stream, S

nS

Transfer,

Transport, J

kJ

51

Chapter 2
For this reason, a formal counting variable (integer) of the number of attributes is introduced
in Table 2.6. These are discrete decision variables available for making model changes, for either
reduction or refinement.

4.4.1. Number of equations

The total number of variables and equations of the model is determined by the physical
structure of the model. The number of model equations is calculated using a combinatorial product
over a set of counting variables of the relevant attributes.

a)

The number of state equations (differential equations) is determined by the product of the
number of physical attributes and the number of resources considered:

NDE nR n

nU nC nD

n n

b)

(2.25)

The number of algebraic equations is determined by the number of sources and sink terms, the
number of transfer and transport phenomena, the phase and reaction equilibria, the physical and
transport properties considered in the model. In practice the number of algebraic variables is
three to ten times the number of differential equations (Van den Bergh, 2005), depending on the
implementation of the physical properties (as hidden module or as explicit equations in the
model). The number of initial conditions is equal to the number of state variables considered,
while the number of the boundary conditions is equal to the product between the number of state
variables and the number of boundaries.

NAE n n k k J n pp ...

(2.26)

List of symbols

NDE = number of differential equations

NAE = number of algebraic equations
n pp = number of physical and transport properties

4.4.2. Number of terms in the model equations

The number of terms and the type of operations in the model equations are determined by the
behavioural structure (phenomena), as well as the connectivity attributes of the model.
The number of terms in differential equations is determined by the number of coordinates
considered, the number of sources, sinks, transfer, and transport phenomena considered in the model.

52

Process modeling and model reduction

4.4.3. Connectivity of the model equations
The incidence matrix of the model shows which variables occur in which equations. If this
matrix is manipulated by permutations into a lower-block triangular form, the numerical solution
process is decomposed in solving a sequence of smaller sub-problems. Usually, a process has recycles
inside the units (such as counter current connections between compartments) and between units as
well as heat and power integration between the units. The more of these connections arise, the denser
the incidence matrix becomes and prospects on decomposition are lost. Removing weak recycles
between far removed downstream and upstream units in the process from the model can be more than
proportional beneficial in reducing computational efforts.

4.5. Numerical level

The third level of decomposition considered in
the modelling procedure is the numerical level (see
Figure 2.13).
An important step in modelling the process
involves the decisions regarding the numerical methods
that will be available and chosen for solving the model.
Most of the models of chemical processes contain a
large number of nonlinear differential and algebraic
equations (DAEs). The numerical solution of the model
implies transformation, substitutions and other types of
manipulation of the model equations in a form that can

Numerical level
Scenario
Scenario
Scenario
Scenario

Scenario
Algorithm
Domain of
applicability

Scenario
Scenario

Discretization

Settings
Scenario
Scenario

Performance
indicator

Scenario
Scenario
Scenario
Scenario

Scenario
Scenario

Post processing

be dealt with by the respective numerical method or

environment.

Figure 2.13. Structure of numerical level

For example, when dealing with a set of partial differential equations (PDEs), one numerical
approach involves the discretization of the model equations along one or more coordinates and
transforming the equations into a set of ordinary differential equations (ODEs). The transformation
will lead to an increase of the number of ODEs to be solved proportional to the number of discretized
coordinates.
This approach involves also a decision on which coordinate will NOT be discretised and used
as the continuous coordinate in the numerical integration of the resulting ODE set.
An alternative approach in solving (P)DAEs is by application of functional approximations
with minimization of weighted residuals. This leads, among others, to orthogonal collocation methods
on finite elements. Here, decisions are needed regarding the kind of support functions, the distribution
of collocation points along the coordinates and more.
For these reasons, the numerical implementation of the model can be a time consuming task as
well. At this level, information related to the chosen set of suitable numerical solution algorithms, the

53

Chapter 2
settings of the numerical algorithms parameters (tolerances, maximum number of iterations,
discretization steps, domains of applicability, etc.) should be included. Information concerning the
post processing of the results should be defined at this level as well.
Two issues need to be considered when solving a system of DAEs (Rascol, Meyer, Le Lann,
& Prevost, 1998): the high order index and the consistent initial conditions.
The differential index of the DAE system is the minimum number of system differentiation
with respect to time in order to convert it into an explicit set of first order ODEs (Leitold &
Cameron, 2001).
There are three main causes for a high index DAE (Soetjahjo, 2006):
a)

internal state constraints

b)

non causal relation of the chosen input variables of the model

c)

interconnection of the DAE sub-models

Avoiding high index DAE is also a matter of proper modelling, as discussed in Section 4.3.
The calculation of consistent initial values, equation (2.10), is important for applications such
as optimal control or parameter identification problems. For the solution of the DAE system is
required a set of initial conditions (2.10). The input variables are specified and from equation (2.12)
the algebraic variables can be calculated if the model is well defined. Once these variables are known,
the differential equation (2.9) is solved.

4.6. Software implementation level

The final level of decomposition for the modelling process is the software implementation.
Once the strategy of solving the model is clear after the numerical level of decomposition, the
following step is to write the model equations into a software environment. A distinction is kept
between the implementation of the (scaled and discretised) model equations and of the numerical
technique to solve these equations. The latter are often pre-packaged. The writing of the model
equations into the software environment often requires an important amount of work. The variables
and the equations should be written in a certain way, specific to the software.
Depending on the software environment to be used, the model equations are implemented in
two ways:
a)

Open-model format - the model equations are written completely, with their mathematical

operators. Sometimes the equations are rewritten to avoid division (by zero), and transforming
them into multiplications.
b)

Model library format the software environment has a library of models of the different

entities inside the process network available. The user chooses the desired models and connects
them as the topological structure specifies, specifying which are the inputs, the outputs and the
parameters.

54

Process modeling and model reduction

In both cases, the software environment have available a built-in library of solvers for
differential equations, solving linear systems of equations, etc. In some cases (for example in
gProms) tools for other applications such as (dynamic) optimization or parameter estimation are
available as well.

5. Existent model reduction techniques

As seen from the previous sections, every choice, at each level of decomposition will have an
effect on the final model. Moreover, different decisions lead to different resulting models for the same
application.
In practice, the model should be able to preserve the input-output behaviour over specified
windows of inputs and outputs and represent the reality in a not very computationally intense way,
having in the same time a low maintenance cost.

Model reduction is one of the approaches used to obtain such a model. This approach has a
long history in the area of systems and control. Though very extensive, the systems and control
literature related to model reduction focuses to a large extent on the reduction at the mathematical
structure level, more specifically on reducing the number of equations of the system. In addition,
since this is usually done in an aggregated way, it can prove unsuccessful in many cases, as it will be
shown further.

5.1. Approaches to model reduction of process models

Several ways of classifying the model reduction techniques can be envisaged, but for the
further section, the following is considered (Marquardt, 2001):
(1) Model simplification. This category includes all the techniques used to reduce the model

complexity by reducing the complexity of the functional expressions, represented by the functions
f , g , h in the model equations. Model simplification preserves the total number of model equations,

but may diminish the number of terms and their nonlinearity.

(2) Model order-reduction. The techniques in this category replace the original large-scale model

with a model having less number of differential equations. They will operate on a priori given
mathematical models and aim for a (significant) reduction in computational times.
There must be a clear distinction between model order-reduction, the group of techniques that
reduce the size (number of differential equations) of the model, and model reduction, which includes
all methods that can be used for reducing the model complexity (acting on both the number and the
quality of these equations).
Sections 3 and 4 of this section introduce the discrete decision attributes that allow
modification of the model. In the following, the way each of these groups of attributes can be used for
performing model reduction is identified.

55

Chapter 2

5.2. Contextual level

Model reduction is performed from the first steps of the model development cycle. All the
choices taken at the contextual level will affect the outcome of the modelling procedure.
Table 2.7. Typical applications for model use
Application type

Analysis

Optimization

Improvement

Feasibility

Parameter

Variable type

of

of

of

of

estimation

performance

performance

design

new design

Input

Operating conditions

Design parameters

Physical properties
Output

At this level, the real system is reduced to a model with a single application or groups of
applications that will be analysed using the model.
The decisions regarding the application and the goal(s) will have an effect on the input-stateoutput structure of the final model. Depending on the application the model will be used for, different
groups of variables will act as known ( ) and unknown (?) for the model, as shown in Table 2.7. For
the most important applications in process engineering (analysis and optimization of performance,
improvement and feasibility of design), the variables that act as inputs and outputs are defined.
Moreover, the choice to operate in steady- or dynamic state will have an effect on the number
of coordinates considered by the model. This is one of the few options for model reduction at this
stage.

5.3. Structure level

The number of equations depends strongly on the number of discrete attributes at the structure
level, as shown in equations (2.25) and (2.26). In the chemical engineering area, the reduction is
being performed at both the physical, behavioural and the mathematical structure level.
The decisions regarding the physical structure of the system will affect both the number of
model equations and the complexity of the differential and algebraic equations (PDAEs) system that
is obtained.
The existent techniques at structure level are summarised in Table 2.8 together with papers
where each technique is used. The references in Table 2.8 are for exemplification purpose only.

56

Process modeling and model reduction

5.3.1. Model reduction at physical structure level
The physical structure level affects the number of model equations. Starting from the initial
levels of decomposition, model reduction is performed in several ways:

a)

Resources
The number of resources directly affects the number of differential equations that have to be

solved. The reduction in this case is performed by ignoring resources that are not important for the
application considered (Yermakova & Anikeev, 2005). In chemical engineering often, the momentum
balance is replaced by (algebraic equation) information related to the flow regime. For applications
that take place in isothermal conditions, the energy balance is ignored as well.
Table 2.8. Model reduction options in chemical engineering applications at structure and system level
Options
Level

Approach

Example

Attribute
Yermakova &Anikeev,

Resource

Ignore

2005;
Okino & Mavrovouniotis,
1998

Unit
Ranzi, Dente, Goldaing,

Connectivity

Behaviour

Lumping

Domain
Region

Bozzano & Faravelli, 2001;

Henda & Falcioni, 2006;
Dones & Preisig, 2010

Phase
Coordinate

Physical

Structure

Compartment

Steady state assumption

Gerstlauer, Mitrovi, Motz

Integrate/average out

Mercenier & Michel, 1994;

& Gilles, 2001

Symmetry assumption

Transfer

Ignore terms

Transport

Average

Sources and

Lumping

sinks

Time-scale analyis

Housiadas, Klidis &

Tsamopoulos, 2007
Baur, Higler, Taylor &
Krishna, 2000;
Schrefler, 2004

Sensitivity analysis

Okino & Mavrovouniotis,

1998

Structure

Unit

Lumping/Aggregation

Antelo,Banga &Alonso,2008

Connectivity

Stream

Lumping

Altimari & Bildea 2008;

System

Altimari & Bildea, 2009

b)

Species
Although considered as a resource in this thesis, the model reduction of the species is well

developed in the chemical engineering literature, hence a separate section is required. Lumping of

57

Chapter 2
species into pseudo-species is one way of reducing the model at this level. The lumping is performed
using a continuum or a discrete view (Okino & Mavrovouniotis, 1998). In the first approach, a
complex mixture is treated as a continuum. The species are not identified as a specific individual, but
only as a distribution of species. The mixture is described in terms of rate constants, or physical
constants such as boiling point, molecular structure or molecular weight. In the discrete lumping, the
components of the mixture are treated individually. On basis of the reactivity, the species vector size
is reduced to one of lower dimensions. Often the reduction of species is related to the reduction of the
number of reactions, as shown in the next section.

c)

Units, compartments, domain and regions

The number of physical blocks considered in the model has an effect on the number of

equations that have to be solved, as well as the number of boundary conditions that have to be
considered. An example is the separation of a mixture of two components. This can be performed in
two ways: (1) using a distillation column, for which rigorous transfer equations are written for each
try or (2) using a simple separation type of unit for which the equilibrium between the gas and the
liquid phase at operation conditions define the output flows of the unit. Lumping different individual
blocks will decrease the problem size.

d)

Phases
The reduction is performed in this case by lumping distinct, dispersed phases into pseudo-

continuous phases (Henda & Falcioni, 2006). In this way, there is no need for writing balance
equations for each separate phase, separately. The system of PDEs that has to be written for the
dispersed phases is replaced in most cases by a single PDE and a set of algebraic equations defining
the composition of different species inside the pseudo-continuous phase. For example, in the case of
the iso-butane alkylation, the reaction takes place in liquid phase.

e)

Coordinates
Model reduction options emerge in the choice of the coordinate system by assuming certain

geometric symmetries and thus reducing the required number of coordinates. Assuming steady state
instead of dynamic behaviour (Gerstlauer, Mitrovi, Motz & Gilles, 2001) or imposing symmetry
assumptions (Housiadas, Klidis & Tsomopoulos, 2007) or going from Eulerian to Lagrangian
coordinates are techniques used to transform the system from a number of n coordinates to a much
smaller number as well.
Another option is to work with averaged values of functions over a sub-space (Mercenier &
Michel, 1994; Housiadas, Klidis & Tsomopoulos, 2007). By performing averaging over this subspace, the coordinates spanning this sub-space drop out and are replaced by a single scalar (an area or
a volume of the sub-space). This latter reduction option is applied in Chapter 5 of this thesis.

58

Process modeling and model reduction

5.3.2. Model reduction at behavioural structure level
The behavioural structure level affects the number of algebraic equations as well as the
number of terms in the differential equations. At this level, the following reduction approaches are
identified:

a)

Sources & sinks

The use of complex reaction sets will have a strong effect on the complexity of the resulting

model, due to the presence of different time scales and uncertainties regarding the kinetic parameters.
The model reduction techniques are linked in this case with the identification of the key reactions and
the sets of species that give insight into the behaviour of the process.
Three strategies have been pursued in the context of chemical reaction systems: lumping,
sensitivity analysis and time-scale analysis (Okino & Mavrovounitis, 1998). Lumping transforms the
original variables to a lower dimensional vector. The sensitivity analysis is used for the identification
and elimination of the insignificant reactions and species based on their impact on the system, while
the time-scale analysis identifies the different scales over which the species react and the fast-time
scale reactions and species are assumed to be at steady state.

b)

Transfer & transport terms

The reduction is done in this case by neglecting significant terms that have no practical

interest or by averaging (integrating) the equations over suitable scales in time and space. A common
example is the case of the reactive distillation column. Two different approaches are available for
modelling of the unit (Baur, Higler, Taylor & Krishna, 2000): (1) the equilibrium stage model, in
which the vapour and liquid phases are assumed to be in thermodynamics equilibrium, and (2) the
non-equilibrium stage model, in which the finite mass-transfer rates across the vapour-liquid interface
are accounted for.

c)

Derivatives
Depending on the time scales of interest, some derivatives for the states that settle very fast

into a steady state, a pseudo-steady state hypothesis can be applied. On the other hand, for states
changing very slowly relative to other states, a pseudo-constant state can be applied. A similar
approach can be followed for spatial coordinates.

5.4. Mathematical level

Since the equations are the core of any model, it is obvious that an important effort in reducing
the complexity will focus at this level of decomposition.
As mentioned in Section 4.3 of this chapter, a system is represented in mathematical form
as a set of differential and algebraic equations (PDAEs), shown in Figure 2.12. In the following

59

Chapter 2
sections model reduction techniques used to reduce the complexity of this PDAE system will be
discussed.
The existent techniques for model reduction at mathematical lever are summarised in Table
2.9. The references in Table 2.9 are for exemplification purpose only.
Table 2.9. Model reduction options in chemical engineering applications at mathematical and numerical level
Options

Example

Linear

Projection-

Marquardt, 2001; Sandberg &

based

Rantzer, 2004; Antoulas, 2005;

Non-projection

Astolfi, 2007; Wattamwar,

based

Nonlinear

Duysinx & Golinval, 2007

Approximation of

Algebraic equation

Marquardt, 2001

functional expressions
Linearization

Numerical

Weiland & Backx, 2010

Hedengren & Edgar, 2005; Brls,

Linearization

Simplification

Mathematical

Order-reduction

Attribute

Differential equation

Level

Approach

Marquardt, 2001

Approximation of functional
expressions
Nafe & Maas, 2002;

Simplification of chemical

Bhattacharjee, Schwer, Barton &

kinetics

Green, 2003

Simplification of physical

Marquardt, 2001

properties models

Domain of

Focus on domain of interest

Bildea, Dimian & Iedema, 2001

applicability
Discretization

Motz, Mitrovi, Gilles, 2002;

Kasiri & Bashiri, 2010

Numerical

Functional approximation

Seferlis & Grievink, 2001;

Dalaouti & Seferlis, 2006

scheme

Qamar, Elsner, Angelov,

Operator

Warnecke & Seidel-Morgenstern,

approximation

2006; Mesbah, Kramer, Huesman

& Van den Hof, 2009

5.4.1. Model simplification techniques

This is a very popular technique since in many applications, such as dynamic optimization, the
repeated evaluation of the right-hand side of the model described by Figure 2.12 is the main
computational effort. Hence, by reducing the complexity of the functions to be evaluated, an

60

Process modeling and model reduction

important reduction of the costs can be achieved. This is done by techniques such as linearization or
by approximation of the functional expressions, for the differential equations of the model. For the
algebraic equations, techniques such as simplification of chemical kinetics (Nafe & Maas, 2002;
Bhattacharjee, Schwer, Barton & Green, 2003) or of physical properties models (Marquardt, 2001)
are used as well.

5.4.2. Model order-reduction techniques

The simplification of the (algebraic) right-hand side of the model equations is not always
enough to make the model computationally effective. For this reason, methods to reduce the number
of equations in the PDAE system have been developed. Two groups of techniques are often used for
the model-order reduction, depending on the type of the model:
a)

Linear model order-reduction

These techniques are dedicated to linear models and are classified in two categories:

1) projection-based methods, including methods such as balanced realization-based methods

(Antoulas, 2005), proper orthogonal decomposition (POD)-based methods (Wattamwar,
Weiland & Backx, 2010), Krylov-subspace (Phillips, 1998) or momentum matching methods
(Astolfi, 2007)

2) non-projection methods, such as the Hankel optimal model reduction method (Benner,
Quintana-Orti, & Quintana-Orti, 2004), the singular perturbation methods (Cao & Schwartz,
2005) or various optimization-based methods (Sootla, Rantzer & Kotsalis, 2009)
More details on the methods in this category can be found in Chapter 3, Section 2.2.1 of this
thesis.
b)

Nonlinear model order-reduction

Most nonlinear model order-reduction methods are extensions of linear methods for model

reduction (Brls, Duysinx & Golinval, 2007). For example, some reduction techniques apply linear
projection to the nonlinear model as if it were a linear model.
The simplest approach for generating reduced order models for nonlinear systems is based on
linearization of systems nonlinearity and subsequent application of linear model-order reduction
methods. The main drawback of this approach is that the obtained reduced model is valid only locally,
around the initial operating point of the nonlinear system. However, these techniques do not work for
cases when the models have a strongly nonlinear behaviour.
Since these approaches are not always successful, due to the high nonlinearity of the chemical
and biochemical plant models, some particular methods have been developed, for example the use of
neural networks (Padhi & Balakrishnan, 2002), or the use of hybrid models (Vega, Lima & Pinto,
2008; Romijn, zkan, Weiland, Ludlage & Marquardt, 2008).

61

Chapter 2
However, the main issue in working with nonlinear model-order reduction is that the reduced
models obtained can be even more difficult to solve than the full models (Marquardt, 2001).

5.5. System level

The existent techniques are also summarised in Table 2.8, together with papers where each
technique is used. The references in Table 2.8 are for exemplification purpose only.

a)

Units
The lumping of different individual process units, presented in the previous section, is a

technique used at system level as well.

b)

Streams
Different streams can be lumped together into a single stream, as shown in (Bildea, Dimian &

Iedema, 2000) or in (Altimari & Bildea, 2008) where the different recycle streams were coupled as
one big recycle.

5.6. Numerical level

a)

Discretization
The approximation of the set of partial differential equations (PDEs) defining the chemical

process system, by a set of ordinary differential equations is often needed in order to reduce excessive
demand for computation time required by the solution of the systems. The PDEs are solved
numerically using two types of techniques (Motz, Mitrovi & Gilles, 2002):
(1) While keeping the time derivatives, the space coordinates are discretized by methods such as
finite difference, finite element, and orthogonal collocation, to obtain a set of ordinary
differential equations (ODEs). Such schemes are referred to as the method of lines.
(2) Both space and time coordinates are discretized to obtain a set of nonlinear algebraic equations.
The reduced (discrete) system is obtained using different discretization approaches for which
parameters such as time step, grid spacing, node distribution, etc. need to be adequately selected,
taking into account the nature of the application and the physical system under consideration.

b)

Functional approximation
These techniques are trying to transform the set of PDEs into a set of ODEs. The state

variables are rewritten as a set of known support functions i , which are chosen to satisfy as many
conditions (symmetry, boundary) of the problem as it is feasible:
n

a
i

(2.27)

i 1

62

Process modeling and model reduction

For example, in the case of the orthogonal collocation the support functions are orthogonal
polynomials (Finlayson, 1974). The variables ai become the new unknowns of the system.

c)

Operator approximation
These techniques are trying to approximate the operations in the equations in such way that

they are easier to solve. For example, for the differentiation:

x xz k 1 xz k

z
z k 1 z k

(2.28)

In a similar way, the approximation of the integration operator is also possible.

6. Model reduction for chemical engineering applications

As seen in the previous sections, the state-of-the-art of the model reduction approaches
involves a host of diverse methods and techniques that act on both the conceptual, structural level of
the process models, as well as on the numerical level.
Traditionally, the procedure is to apply order-reduction to a model in its entirety (Marquardt,
2001), without preserving the underlying network structure of the process or its-multi scale
decomposition.
Although in many cases significant reduction of the number of equations is achieved, the
benefit is often limited, because the internal structure of the problem is destroyed, the physical
meaning of the variables is lost and there is little or no decrease in the solution time (Van den Bergh,
2005). Attention must be paid on the internal structure of the problem, as well as on retaining the
physical meaning of the variables (Wattanwar, 2010).
For this reason a structure-retaining model reduction approach has been developed. The
approach aims first at simplifying the physical (units, domains, phases, etc.) and the systemic
(connectivity) level of the chemical process model, as well as the behavioural (phenomena) structure,
with the help of the options summarised in Table 2.8. Only then additional mathematical (number of
equations) and numerical scheme reductions are selectively applied to individual compartments or
units, using the options presented in Table 2.9. In the following step, the reduced models of the
individual units are connected at system level and the reduced model of the full process is obtained. In
this way, the reduction procedure is able to preserve the essential structural features of the process.
Moreover, the physical meaning of the variables and equations can be preserved as much as possible.
This signifies that the model M 1 is firstly decomposed on the Structure axis in the sub-models
M 2 to M 4 , as shown in Figure 2.14. The model M 1 is either a rigorous model of the process system

of be studied, or a conceptual version of it, for the cases when the rigorous model cannot be obtained.
In the following step, each individual sub-model is reduced at behavioural level. At this point, the
phenomena that take place inside the system defined by the model, which are relevant for the

63

Chapter 2
application considered, are preserved. Thirdly, the resulting models are further reduced at
mathematical level by applying model-order reduction techniques, if still required. As a result, the
reduced sub-models M 2' to M 4' will be obtained. Finally, these reduced models are connected to obtain

Structure

the reduced model M 1' of the initial process defined by the conceptual model M 1 .

Reduction

M1
M2

Sub-system 0

M4

M1

M3
Refinement

M4
Decomposition

M3
Aggregation

Aggregaation

Decomposition

M2

Reduction

Sub-system 1

M2

M2
Refinement

Sub-system 2

Level of detail
0

Level of detail
1

Behaviour

Figure 2.14. Approaches to model reduction of process models

0 = reference level; 1 = refined level

The following criteria are defined and used in this thesis for assessing the success of the
reduction approach:
(1) The reduced model M 1' retains a satisfactory predictive accuracy
(2) the reduced model M 1' is solved successfully in an acceptable amount of time
(3) the reduced model M 1' is used successfully for practical applications

7. Selected research problems in model reduction

The proposed structure-retaining model reduction approach covers a host of decisions,
involving mainly discrete elements: which attributes, connectivities, terms to reduce or to simplify.
This makes it very hard to develop a synthesis theory for process model reduction with associated
tools to manipulate the discrete model objects, like with mixed integer nonlinear programming

64

Process modeling and model reduction

(MINLP), as for instance in process synthesis. To test the practical feasibility of this approach in an

No consideration

Flowsheet

Unit

Compartment

Phase

Phenomena

Detailed structure

explorative manner, case studies have been performed from which lessons will be learned.

Level
0

Level
1

Level
2

Level
3

Level
i

Level
L-1

Level
L

Cut level
Figure 2.15. Line of exploration for the model structure

Within this set-up of doing case studies, there are generic issues for each process. A first
decision that needs to be taken while performing the model reduction is how far one should go in
decomposing the model structure. An intuitive answer is that the choice for the cut level (Figure 2.15)
is most likely application goal dependent. Scales below the cut level are ignored in the reduced
model.
For processes with relatively simple internal structure of the products (for example singlephase), the cut level should go at least to the unit level. In case of processes with a more complex
internal structure of the products, the cut level must be taken at a lower level (phase, species,
molecular level, etc.). This lower level is required by the presence of the detailed scales inside the
product.
The cut-off criteria for details, size and time between the different attributes should be done in
a consistent way. On how to perform this, a generic answer is difficult to give at this stage. For this
reason, a case study approach will be used in the following chapters to study the application of the
model reduction procedure presented in Section 6.
The selected case studies are not the outcome of a carefully crafted design of experiments to
test and evaluate the model reduction approach in the most effective way. The cases are practically
chosen to match the funding conditions for the thesis project. Nevertheless, the cases represent very
interesting and relevant extremes of the spectrum of chemical processes. Two types of applications
will be studied for this assessment:
(1) complex (multiple unit) processes with relatively simple (one-phase, similar components)
products
(2) simple (single unit) processes with complex (multiphase, multi-component) products

65

Chapter 2
REC

FB 0

MIX1

F1

F8
INERT

HEX2

F6
FA0+FI0

MIX2

F7

HEX1

REACTOR
SPLIT

F3

F2

FIN

COL1
F9

i-OCTANE

F4

P1

COL2
FOUT

P2

F10

F5

COL3

HEAVIES

Figure 2.16. The iso-butane alkylation plant

Figure 2.17. Ice cream structure

For the first type of process, the iso-butane alkylation process will be studied. The details of
the process are presented in Section 3 of this chapter. The process flowsheet (Figure 2.16) consists of
some of the most common process units: chemical reactors, distillation columns, heat exchangers,
mixers, pumps, etc. Another source of complexity is the presence of a recycle of raw materials, which
have to be reused in the process.
Two types of generic inverse applications are considered for the use of the process model: (a)

the assessment of the plantwide control structures, and (b) the dynamic optimization of the plants
operation.
For the second type of process, the freezing step in ice cream manufacture is selected. The
complexity lays in this case not in number of units in the process flowsheet, but in the number of
phases of the product (Figure 2.17), the distributive nature of its product structure as well as the many
interactive phenomena that need to be taken into account.

8. Conclusions
Current numerical reduction approaches focus on the (process) model in its entirety, without
trying to preserve the underlying network structure of the process or its multi-scale decomposition.
Retaining the meaningful structural features in a reduced model of a process is a necessity for longterm practical use of the model, certainly in a practical manufacturing environment where adaptations
in process unit models will occur regulary.
In this chapter, the most important features of the chemical process model structure have been
identified and the opportunities for performing model reduction have been described. A structureretaining model reduction has been presented. The approach aims first at simplifying the physical
(units, domains, phases, etc.) and the systemic (connectivity) level of the chemical process model, as
well as the behavioural (phenomena) structure. Only then additional mathematical (number of
equations) and numerical scheme reductions are selectively applied to individual compartments or
units. In the following aggregation step, the reduced models of the individual units are connected at
system level and the reduced model of the full process is obtained. In this way, the reduction

66

Process modeling and model reduction

procedure maintains the essential structural features of the process and the physical meaning of the
variables is preserved as much as possible.
The reduction must be performed in such way that the resulting model retains the predictive
accuracy of the original up to a satisfactory level. Moreover, the reduced model should be solved
sufficiently fast for practical applications.
The applicability and the effectiveness of the structure-retaining model reduction will be
studied in the following chapters by means of case studies. This approach allows for an easy
identification of the characteristics of applying the reduction procedure for specific cases, which can
be further extrapolated to generic applications.

9. References
Albright, L.F., Spalding, M.A., Faunce, J., & Eckert, R.E. (1988). Alkylation of isobutene with C4
olefins. 3. Two-step process using sulphuric acid as catalyst, Industrial and Engineering

Chemistry Research 27, 391

Albright, L.F. (2002). Alkylation of isobutene with C3-C5 olefins: Feedstock consumption, acid usage,
and alkylate quality for different processes, Industrial and Engineering Chemistry Research
41, 5627
Albright, L.F. & Wood, K.V. (1997). Alkylation of isobutene with C3-C4 olefins: identification and
chemistry of heavy-end production, Industrial and Engineering Chemistry Research 36,
2110
Altimari, P., & Bildea, C.S. (2008). Coupling exothermic and endothermic reactions in plug-flow
reactor separation recycle systems, Industrial and Engineering Chemistry Research 47,
6685
Altimari, P. & Bildea, C.S. (2009). Integrated design and control of plantwide control Systems
coupling exothermic and endothermic reactions, Computers and Chemical Engineering 33,
911
Antelo, L.T., Banga, J.R., & Alonso, A.A. (2008). Hierarchical design of decentralized control
structures for the Tennessee Eastman Process, Computers and Chemical Engineering 32,
1995
Antoulas, A.C. (2005). An overview of approximation methods for large-scale dynamical systems,

Annual Reviews in Control 29, 181

Asprey, S.P., Macchietto, S. (2002). Designing robust optimal dynamic experiments, Journal of

Process Control 12, 545

Astolfi, A. (2007). A new look at model reduction by moment matching for linear systems,

Proceedings of the 46th IEEE Conference on Decision and Control, 4361

Baur, R., Higler, A.P., Taylor, R., & Krishna, R. (2000). Comparison of equilibrium stage and
nonequilibrium stage models for reactive distillation, Chemical Engineering Journal 76, 33

67

Chapter 2
Benner, P., Quintana-Orti, E.S., & Quintana-Orti, G. (2004). Computing optimal Hankel norm
approximations of large-scale systems, Proceedings of the IEEE Conference on Decision

and Control 3, 3078

Bequette, B.W. (1998). Process dynamics: modelling, analysis and simulation, Prentice Hall, 5
Bhattacharjee, B., Schwer, D.A., Barton, P.I., & Green, Jr., W.H.G. (2003). Optimally-

reduced

kinetic models: reaction elimination in large-scale kinetic mechanisms, Combustion and

Flame, 135, 191

Bildea, C.S., Dimian, A.C., & Iedema , P.D. (2000). Nonlinear behaviour of reactor-separator-recycle
Systems, Computers and Chemical Engineering 24, 209
Bildea, C.S., Dimian, A.C., & Iedema, P.D. (2001). Multiplicity and stability approach to the design
of heat-integrated multibed plug flow reactor, Computers and Chemical Engineering 25, 41
Bongers, P.M.M., Almeida-Rivera, C. (2009). Product driven process synthesis methodology,

Computer-Aided Chemical Engineering 26, 231

Brls, O., Duysinx, P., & Golinval, J.P. (2007). The global model parameterization for non-linear
model-order reduction in flexible multibody dynamics, International Journal for Numerical

Methods in Engineering 69, 948

Burke, A.L., Duever, T.A., Penlidis, A. (1997). Choosing the right model: Case studies on the use of
statistical model discrimination experiments, The Canadian Journal of Chemical

Engineering 75, 422

Cameron, I.T., Wang, F.Y., Immanuel, C.D., Stepanek, F. (2005) Process systems modelling and
applications in granulation: A review, Chemical Engineering Science 60, 3723
Cao, L. & Schwartz, H.M. (2005). Reduced-order models for feedback stabilization of linear systems
with a singular perturbation model, Asian Journal of Control 7, 326
Caton, J., Racen, B., Troutman, G. (2008). Introduction to sulphuric acid alkylation unit process
design,http://www2.dupont.com/Clean_Technologies/en_US/assets/downloads/Intro%20to

%20H2 S04%20Unit%20Design.pdf
Charpentier, J.C. & McKenna, T.F. (2004). Managing complex systems: Some trends for the future of
chemical and process engineering, Chemical Engineering Science 59, 1617
Christofides, P.D., Daoutidis, P. (1998). Robust control of hyperbolic PDE systems, Chemical

Engineering Science 53, 85

OConnell, J.P., Haille, J.M. (2005). Thermodynamics: fundamentals and applications, Cambridge

University Press, 72
Dalaouti, N. & Seferlis, P. (2006). A unified modelling framework for the optimal design and
dynamic simulation of staged reactive separation processes, Computers and Chemical

Engineering 30, 1264

Dimian, C.A. & Bildea, C.S. (2008). Chemical Process Design. Computer-aided cases studies, Wiley-

VCH, 261

68

Process modeling and model reduction

Dones, I. & Preisig, H.A. (2010). Model simplification and time-scale assumptions applied to
distillation modelling, Computers and Chemical Engineering 34, 732
Douglas, J.M. (1988). Conceptual design of chemical processes, McGraw-Hill
Finlayson, B. A. (1974). Orthogonal collocation in chemical reaction engineering, Catalysis Review
10, 69
Gerstlauer, A., Mitrovi, A., Motz, S., & Gilles, E.D. (2001). A population model for crystallization
processes using two independent particle properties, Chemical Engineering Science 56,
2553
Gopal, V., & Biegler, L.T. (1998). A successive linear programming approach for initialization and
reinitialization after discontinuities of differential-algebraic equations, SIAM Journal on

Scientific Computing 20, 447

Hangos, K., Cameron, I. (2001). Process modelling and model analysis, Academic Press, 51
Hangos, K., Cameron, I. (2001). A formal representation of assumptions in process modelling,

Computers and Chemical Engineering 25, 237

Hedengren, J.D. & Edgar, T.F. (2005). Order reduction of large-scale DAE models, Computers and

Chemical Engineering 29, 2069

Henda, R. & Falcioni, D.J. (2006). Modelling of heat transfer in a moving packed bed: Case of the
preheater in nickel carbonyl process, Journal of Applied Mechanics 73, 47
Hjalmarsson, H. (2005). From experiment design to closed-loop control, Automatica 41, 393
Hommeltoft, S.I. (2001). Isobutane alkylation. Recent developments and perspectives, Applied

Catalysis A: General 221, 421

Housiadas, K.D., Lidis, G., Tsamopoulos, J. (2007). Two- and three-dimensional instabilities in the
film blowing process, Journal of Non-Newtonian Fluid Mechanics 141, 193
Kasiri, N. & Bashiri, A. (2010). Comparative study of different techniques for numerical reservoir
simulation, Petroleum Science and Technology 28, 494
Kiss, A.A., Bildea, C.S., Dimian, A.C. (2005). Design of recycle systems with parallel and
consecutive reactions by nonlinear analysis, Industrial and Engineering Chemistry

Research 44, 576

Klatt, K.U. & Marquardt, W. (2009). Perspectives for process systems engineering Personal views
from academia and industry, Computers and Chemical Engineering 33, 536
Lakner, R., Hangos, K.M., Cameron, I.T. (2005). On minimal models of process systems, Chemical

Engineering Science 60, 1127

Lee, L. & Harriott, P. (1977). The kinetics of isobutene alkylation in sulphuric acid, Industrial and

Engineering Chemistry Process Design and Development 16, 282

Leitold, A., Hangos, K.M. (2001). Structural solvability analysis of dynamic process models,

Computers and Chemical Engineering 25, 1633

Marquardt, W. (1996). Trends in computer-aided process modelling, Computers and Chemical

Engineering, 20, 591

69

Chapter 2
Marquardt, W. (2001). Nonlinear model reduction for optimization based control of transient
chemical processes, Proceedings of the 6th international Conference of Chemical

Process Control, AIChe Symp. Ser. 326, Vol. 98, 12

Mercenier, J. & Michel, P. (1994). A criterion fro time aggregation in intertemporal dynamic models,

Mathematical Programming 64, 179

Mesbah, A., Kramer, H.J.M., Huesman, A.E.M., Van den Hof, P.M.J. (2009). A control oriented
study on the numerical solution of the population balance equation for crystallization
processes, Chemical Engineering Science 64, 4262
Motz, S., Mitrovi, A., & Gilles, E.D. (2002). Comparison of numerical methods for the simulation of
dispersed phase systems, Chemical Engineering Science 57, 4329
Nafe, J., & Maas, U. (2002). A general algorithm for improving ILDMs, Combustion Theory and

Modelling, 6, 697
Okino, M.S., Mavrovouniotis, M.L. (1998). Simplification of mathematical models of chemical
reaction systems, Chemical Reviews 98, 391
Padhi, R. Balakrishnan, S.N. (2002). Proper orthogonal decomposition based feedback optimal
control synthesis of distributed parameter systems using neural networks, Proceedings of

the American Control Conference 6, 4389

Phillips, J.R. (1998). Model reduction of time-varying linear systems using approximate multipoint
Krylov-subspace projectors, IEEE/ACM International Conference on Computer-Aided

Design, Digest of Technical Papers, 96

Prasad, V., & Wayne Bequette, B. (2003). Nonlinear system identification and model reduction
using artificial neural networks, Computers and Chemical Engineering, 27, 1741
Qamar, S., Elsner, M.P., Angelov, I.A., Warnecke, G., Seidel-Morgenstern, A. (2006). A comparative
study of high-resolution schemes for solving population balances in crystallization,

Computer and Chemical Engineering 30, 1119

Ranzi, E., Dente, M., Goldaniga, A., Bozzano, G., & Faravelli., T. (2001). Lumping procedures in
detailed kinetic modelling of gasification, pyrolysis, partial oxidation and combustion of
hydrocarbon mixtures, Progress in Energy and Combustion Science, 27, 99
Rascol, E., Meyer, M., Le Lann, J.M., & Prevost, M. (1998). Numerical problems encountered in the
simulation of reactive absorption: DAE index reduction and consistent initialisation,

Computers and Chemical Engineering 22, S929

Romijn, R, zkan, L., Weiland, S., Ludlage, J. & Marquardt, W. (2008). A grey-box modelling
approach for the reduction of nonlinear systems, Journal of Process Control 18, 906
Sandberg, H. & Rantzer, A. (2004). Balanced truncation of linear time-varying systems, IEEE

Transactions on Automatic Control, 49, 217

Scherpen, J.M.A. (1994). Balancing for nonlinear systems, PhD Thesis, University of Twente,
Enschede, The Netherlands

70

Process modeling and model reduction

Schrefler, B.A. (2004). Multiphase flow in deforming porous material, International Journal for

Numerical Methods in Engineering 60, 27

Seferlis, P. & Grievink, J. (2001). Optimal design and sensitivity analysis of reactive distillation units

using collocation models, Industrial and Engineering Chemistry Research 40, 1673
Soetjahjo, A.T.M.J. (2006). Mathematical analysis of dynamic process models Index, inputs and
interconnectivity, PhD Thesis, Delft University of Technology, The Netherlands
Sootla, A., Rantzer, A., & Kotsalis, G. (2009). Multivariable optimization-based model reduction,

IEEE Transactions on Automatic Control 54, 2477

Stewart, W.E., Henson, T.L., Box, G.E.P. (1996). Model discrimination and criticism with singleresponse data, AIChE Journal 42, 3055
Stewart, W.E., Shon, Y., Box, G.E.P. (1998). Discrimination of fit of multiresponse mechanistic
models, AIChE Journal 44, 1404
Unger, J., Krner, A., Marquardt, W. (1995). Structural analysis of differential-algebraic equation
systems Theory and applications, Computers and Chemical Engineering 19, 867
Van den Bergh, J. (2005). Model reduction for dynamic real-time optimization of chemical processes,

PhD thesis, Delft University of Technology, The Netherlands

Vlachos, D.G. (2005). A review of multiscale analysis: Examples from systems biology, materials
engineering and other fluid-surface interacting systems, Advances in Chemical Engineering
30, 1
Vega, M.P., Lima, E.L., & Pinto, J.C. (2008). Use of bifurcation analysis for development of
nonlinear models for control applications, Chemical Engineering Science 63, 5129
Wattanwar, S.K. (2010). Identification of low-order models for large-scale processes, PhD Thesis,
Eindhoven University of Technology, The Netherlands
Wattanwar, S.K., Weiland, S., Backx, T. (2010). Identification of low-order parameter-varying
models for large-scale systems, Journal of Process Control 20, 158
Willard, S. (2004). General topology, Dover Publications, 28
Yang, A., Marquardt, W. (2009). An ontological conceptualization of multiscale models, Computers

and Chemical Engineering, 33, 822

Ye, M., Meyer, P.D., Neuman, S.P. (2008). On model selection criteria in multimodel analysis, Water

Resources Research 44
Yermakova, A. & Anikeev, V.I. (2005). Analysis of the CSTR with two-phase flow under phase
equilibrium conditions, Chemical Engineering Science 60, 3199

71

Chapter 2

72

3
APPLICATION OF MODEL REDUCTION TO DESIGN
OF PLANTWIDE CONTROL SYSTEMS

The chapter proposes an approach for obtaining reduced-order models of chemical processes, with
application to the design of plantwide control systems. The approach is based on the inherent structure that exists
in a chemical plant and identifies units or groups of units that determine the steady state and dynamic behaviour of
the plant. Specific reduction techniques with different accuracy are applied to each unit, followed by the coupling
of the reduced models, according to the process flowsheet structure. By means of a case study, the iso-butane butene alkylation plant, it is shown that the approach retains the nonlinearity of the original plant model and
preserves the significance of important model variables. These features make the reduced model a valid option for
the effective assessment of plantwide control structures.

Chapter 3

1. Introduction
1.1. Model reduction for complex processes
The discussion in Chapters 1 and 2 of the thesis showed that, for the case of complex
processes, the reduction procedures focus on the upper levels of the network structure of the process:
process units, compartments, domains.
In this chapter, the focus will be on developing a reduced model for a complex process with a
single-phase product, in order to highlight the characteristics of the model reduction approach when
dealing with this type of processes. The influence of the products internal structure is not important
in this case, and will be neglected. The main attribute of such a model is the many network-like
aspects of the process structure, which features units with a reduced number of distinct compartments
and domains, like trays in distillation columns.
The reduction procedure is illustrated by means of a case study: the iso-butane alkylation
plant. It will be demonstrated that an important decrease of the simulation time is obtained while the
reduced model preserves the behaviour of the original model. Moreover, having reduced models of
the units connected accordingly to process structure is a natural way of simplifying the adaptation to
future plant changes.

1.2. Design of the plantwide control systems

In todays competitive environment, high economical performance of chemical and
biochemical plants is achieved by cost-effective steady-state integrated design and by continuously
responding to the market conditions through dynamic operation. The desired policy of operation is
accomplished by control systems that maintain the steady state or implement the optimal dynamic
behaviour (Engell, 2007).
The synthesis, design and optimization of the plantwide control system require dynamic
models of the chemical / biochemical plant. Often, the models are very complex, and therefore call
for high computational speed and high memory storage capacity. Moreover, the problems are likely to
be ill conditioned, since the applications involve solving an inverse problem. This often results in
significant numerical troubles, unstable convergence and eventually, unreliable results.
Because of numerical and time constraints, it is not always possible to consider models of the
highest complexity. The model must be adapted to the specific purpose (Baldea, Daoutidis & Kumar,
2006) although in many cases it is not evident which simplifications are suitable. Therefore, for
particular simulations, it is necessary to find an appropriate technique to reduce these effects. To
achieve these objectives, the model quality is crucial. The model should predict with good accuracy
the behaviour of the real system it describes. Moreover, some process models that are developed from
experimental observations are far from being highly accurate. For this reason, it is important for the
reduced model to take into account this uncertainty (Nagy and Braatz, 2003, 2007). In the same time,

74

Application of model reduction to design of plantwide control systems

the problems that need to be solved often imply repeated solution of the process model. Thus, it is
important that the complexity of the model be limited, in order to allow solution during a restricted
time. Another requirement for the process models is the low cost of maintenance, which implies that
the effort to adapt the model to future plant changes must be kept as small as possible. Additionally,
in many applications of the chemical industry, the process is highly nonlinear and it is important that
this characteristic be preserved (Nagy, Mahn, Franke & Allgwer, 2007).
Model reduction is one technique that can be used for obtaining such a model (Marquardt,
2001; Antoulas, Sorensen & Gugercin, 2000). In this chapter, a new approach to model reduction,
with applicability to the dynamic models of chemical processes and plantwide control, is presented.
This approach makes use of the inherent flowsheet structure that exists in a chemical plant in the form
of units or groups of units that are connected by material streams. The decomposition mirrors the
decentralization of the control problem. The recommended procedure exploits the knowledge about
the process and applies the model reduction techniques to individual units of the plant, and then
couples these reduced models.
The chapter is structured in the following way: Section 2 explores existing model reduction
techniques and the application of model reduction to chemical plants. Next, the proposed approach to
model reduction is introduced in Section 3. The algorithm is applied to the iso-butane alkylation plant
in Section 4. Finally, conclusions are presented in Section 5.

2. Model reduction
The models of dynamical systems contain a large number of equations, both differential and
algebraic, and, quite often, these equations contain complex functional expressions. Such models can
be written in the following, general form:

dx
f x t , u t ,
dt

y t g x t , u t ,

(3.1)

Here t is the time variable, x t n is the state vector, u t m is the input vector, y (t ) p
the output variable vector, is the parameters vector, and n is the state space dimension. The
dimension of the input vector, m and the dimension of the output variable vector p , are much smaller
than n , and usually m p . The function f describes the dynamics of the system, while the function
g describes the way in which the observations are deduced from the state and the input.

To this set, new algebraic equations are added to the dynamic model to describe the physical
and thermodynamic relations:

0 h x t , u t ,

(3.2)

75

Chapter 3
Model reduction has a long history in the systems and control literature. The main idea of this
technique is to reduce the complexity of the dynamic model, while preserving its input-output
behaviour. The review on model reduction techniques in Marquardt (2001) is recommended. One can
find several ways of classifying the model reduction methods, but probably the best way is the one
presented in the above-cited paper:

Model simplification a technique that preserves the number of equations of the original
model, but reduces the complexity of the functional expressions in the model equations

Model-order reduction a technique that replaces the original large-scale model with a
model having (much) less equations

2.1.

Model simplification
The model simplification techniques are very popular in the area of chemical engineering. In

many applications (for example in dynamic optimization), the repeated evaluation of the right-hand
side of the model described by equation (3.1) is often the main computational effort. Hence, by
reducing the complexity of the functions to be evaluated, the model simplification methods are as
important as the ones that reduce the number of equations.
Among these techniques, one can distinguish the linearization around a nominal operating
point, the approximation of functional expressions and the simplification of chemical kinetics (Nafe
and Maas, 2002) and/or physical properties models. In addition, significant model reduction is also
possible by model lumping (Ranzi, Dente, Goldainga, Bozzano & Faravelli, 2001), in which
thermodynamic phases and chemical components are taken together into pseudo-phases and lumped
species. However, such lumping can increase the complexity of the functional expressions in the
remaining model equations.

2.1.1. Linearization
The linearization around a nominal operating point is one of the most often used methods to
simplify nonlinear models. The main idea of this procedure is to rewrite the system presented in
equation (3.1) under the following form:
~

~
~
d x(t )
A x(t ) B u (t )
dt
~

(3.3)

y (t ) C x(t ) D u (t )

where

x(t ) x(t ) xnom (t )

(3.4)

u (t ) u (t ) unom (t )

(3.5)

y (t ) y (t ) ynom (t )

(3.6)

76

Application of model reduction to design of plantwide control systems

In

the

with A

previous
n n

, B

n m

equations A , B , C
, C

pn

and D

and
p m

are

the

system

matrices,

, while the subscript nom represents the nominal

operating point.
One way of realizing the transformation is by using the Taylor series expansion.
The linearization reduces drastically the computational effort. The simulation is more reliable
since the nonlinear system it is replaced by a linear system, easier to solve. However, the linear
models do not adequately predict the process dynamics for the cases when the system is far away
from the nominal operating point.

2.1.2. Approximation of functional expressions

This technique implies the replacement of a complicated, nonlinear model equation by simpler
functional expressions. This simple functional expression has to approximate the original up to a user
specified tolerance, which leads to a multi-variate nonlinear approximation problem.

2.1.3. Simplification of physical properties model

The most common method of simplifying the physical properties models is the development
of local thermodynamic models. The use of local thermodynamics reduces significantly the
computational effort, while maintaining the accuracy at an acceptable level.
The easiest way of reducing the complexity is to re-write the functional expression of a
property in a polynomial form.

2.1.4. Simplification of chemical kinetics

In many applications, large-scale chemical kinetic models are quite common. The complexity
arises from the large number of reactions and components, and it can be reduced both by eliminating
from the reaction network: a) reactions (Bhattacharjee, Schwer, Barton & Green, 2003), which is a
task of the model simplification, and b) species, which is a special case of model-order reduction. The
use of moments in population balance models, for eliminating the distributed nature of the population
balance, is an example of this technique.

2.2. Model-order reduction

Reducing the complexity of the functional expressions in the model equations is not always
enough to achieve the model reduction. In many cases, the number of the model equations needs to be
reduced also. In the model-order reduction class, two main groups of methods can be distinguished:
linear and nonlinear model-order reduction, depending on the type of the full model.

77

Chapter 3
2.2.1. Linear model-order reduction
The largest group of order-reduction methods applies to the linear systems. The problem can
be stated in the following form:
Given the linear dynamical system (3.3), one should find an approximation
^

^ ~
d x(t ) ^ ^
A x(t ) B u (t )
dt
~

(3.7)

y (t ) C x(t ) D u (t )

with A k k , B k m , C p k , D p m and k n such that the following properties are

satisfied (Antoulas, 2000):

The approximation error is small, and a global error bound can be defined.

System properties, like stability, are preserved.

The procedure is computationally stable and efficient.

The approach has many advantages. First, the linear models are easy to be obtained, for
example through Taylor series expansion, as mentioned in the previous section. In the same time,
there is a considerable large range of treatment techniques for linear models, which spreads from
the backward Euler method to multistep methods of solving the ordinary differential equations
(ODEs) that describe the system. In addition, finally yet importantly, stable and well-understood
numerical linear algebra algorithms allow the analysis, control or simulation of such a system.
Many methods of obtaining the approximation presented in equations (3.7) have been
developed in different fields, including control, fluid dynamics, and chemical engineering. They can
be classified into two categories: (a) projection-based methods and (b) non-projection based methods.
(a) The common feature of the projection-based methods is that they are trying to find such k
and n k dimensional subspaces S1 and S 2 of the state space, S that the reduced system
will result from the projection of the state onto S1 and the residual onto S 2 , and the
elimination of the states in S 2 . In this category, one can mention techniques such as Krylovsubspace or momentum matching methods, balanced realization-based methods or proper
orthogonal decomposition (POD)-based methods (Skogestad & Postlethwaite, 1996;
Antoulas & Sorensen, 2001; Rathinam & Petzold, 2003; Penzl, 2006).
(b)

In the non-projection methods, the state-space of the approximation has no connection with
the unreduced system. The most used techniques in this category are the Hankel optimal

78

Application of model reduction to design of plantwide control systems

model reduction method, the singular perturbation method or various optimization-based
methods (Mkil, 1991; Marquardt, 2001).
Each of the methods above has advantages and drawbacks. For example, the Krylov subspacebased methods can be applied for a very high-order system, but have robustness, stability and
efficiency issues (Bai, 2002). In the case of the Hankel-norm approximation and balanced realization
methods, very low-rank approximations are possible and accurate low-order models will result, but
the solution require dense computations and can be carried out only for low-dimension (few hundreds
equations) models (Antoulas & Sorensen, 2001). In the same time, the accuracy of the result is
guaranteed by theoretical results and can be calculated a priori (Skogestad & Postlethwaite, 1996).
The POD-based methods are not system invariant, the resulting simplification being heavily
dependent on the initial excitation, but can be applied for high-complexity systems (Antoulas &
Sorensen, 2001).

2.2.2. Nonlinear model-order reduction

Since the real world applications require the use of nonlinear models, ways to reduce the
complexity of such models had to be developed as well. Some of them apply linear projections to the
nonlinear model as if it were a linear model (Scherpen, 1994).
Since these approaches are not always successful, due to the high nonlinearity of the chemical
and biochemical plant models, some particular methods have been developed, for example the use of
neural networks (Prasad & Bequette, 2003), or the use of hybrid models (Nagy, Mahn, Franke &
Allgwer, 2007). However, the main issue in working with nonlinear model-order reduction is that
the reduced models obtained can be even more difficult to solve than the full models (Marquardt,
2001).

2.3. Linear model-order reduction by balanced realization

The linear model-order reduction methods are easy to apply and the accuracy of the obtained
reduced models is guaranteed by theoretical results. The linear model is easy to obtain, sometimes
only by means of Taylor series expansion. In the same time, the model is easy to solve and analyze,
by means of stable, well-understood numerical linear algebra algorithms.
In the following, methods for obtaining reduced-order models by balanced realization will be
presented. In this approach, a new state-space description is obtained in such a way that the
controllability and observability gramians are equal and diagonal. The balanced realization can be
easily obtained by simple state similarity transformations, and routines for doing this are available in
commercial software, such as MATLAB.
The most important property of a balanced realization is that after the balancing, each state is
as controllable as it is observable, and the measure of states observability and controllability is given
by its associated Hankel singular value (Skogestad & Postlethwaite, 1996). This property becomes

79

Chapter 3
interesting for the model reduction techniques, which remove the states with little effect on the
systems input-output behaviour.
Considering the balanced linear model, which can be described by equations (3.3), the state

~
~
x
space vector x is partitioned into ~1 , where x2 is the vector of the states to be removed. After

x2
~

partitioning the matrices A , B and C , equation (3.3) is rewritten as follows:

~

~
~
~
d x1 (t )
A11 x1 (t ) A12 x2 (t ) B1 u (t )
dt
~

~
~
~
d x2 (t )
A21 x1 (t ) A22 x2 (t ) B2 u (t )
dt
~

(3.8)

y (t ) C1 x1 (t ) C2 x2 (t ) D u (t )

Two main approaches to reduce the order of the model by balanced realization can be
mentioned in this case:

(a) Balanced truncation removes all the states corresponding to the small Hankel singular
~

values. By eliminating the states with little effect on the systems input-output behaviour ( x2 ),
^

equations (3.8) become equations (3.7), with x(t ) x1 (t ) , A A11 , B B1 , C C1 and D D .

An important property of the truncation is that it retains the system behaviour at high frequency.

(b) Balanced residualization sets to zero the derivatives of the states with little effect on the
systems input-output behaviour. An important property of the residualization is that it preserves
the steady-state gain of the system. In this case, the matrices in equation (3.7) are calculated as
follows:
^

A A11 A12 A221 A21

^

B B1 A12 A221 B2

(3.9)

C C1 C2 A221 A21
^

D D C2 A221 B2
It must be mentioned that an important advantage of using the balanced realization techniques
for obtaining the reduced models is that the error can be calculated a-priori. The infinite norm of the

80

Application of model reduction to design of plantwide control systems

error between the original model and the reduced-order model is twice the sum of the Hankel singular
values of the eliminated states (Skogestad & Postlethwaite, 1996).

3. Model reduction with process knowledge

As seen in the previous section, many reduction techniques have been developed, in order to
obtain models that:

predict the behaviour of the plant with good accuracy

have limited complexity, to allow repeated solution during a restricted time

Although significant reduction of the number of equations can be achieved, the benefit is often

limited, because the structure of the problem is destroyed, the physical meaning of the model
variables is lost and there is little or no decrease in the solution time (Van den Bergh, 2005).
Moreover, especially in the case of nonlinear model-order reduction, the obtained reduced model is
often difficult to solve.
Another desired quality of the reduced model is the adaptability to future plant changes. If in
an industrial application, one of the units in the plant flowsheet is replaced, a new reduced model of
the plant has to be obtained.

3.1. Classical model reduction algorithm for chemical plants

In the following sections, the focus will be on obtaining reduced-order models that can be
used to determine the plantwide control structure of the plant. The main steps of a classical approach
to model reduction will be presented, together with the most common issues that appear when
applying the procedure.

3.1.1. Obtaining the rigorous model

The starting point of the approach is a rigorous dynamic model of the plant. This dynamic
model is obtained using a commercial package, such as Aspen Dynamics or gPROMS.
Usually, a basic control of inventory at unit level is included. In the same time, depending on
the package that is used, some model reduction is already present, such as the use of flow-driven
simulation instead of the rigorous pressure-driven mode, the local thermodynamic models, the
simplified chemical kinetics or the instantaneous models for mixers, pumps, valves or heat
exchangers. Very often, the user is able to insert own equations and data to describe parts of the plant.
Under these conditions, the plant model still contains thousands of differential and algebraic
equations, with initial conditions derived from a nominal steady-state operating point. The size of the
model can lead to numerical difficulties when trying to find a solution. When this happens, the search
for the cause of the difficulties and for a remedy is a time-demanding task. Finding a solution of the
dynamic model implies solving numerically a system of differential and algebraic equations (DAEs).

81

Chapter 3
The numerical algorithms often require an initial guess for the algebraic variables. Depending on the
problem, finding the best starting point for the algorithms can also be a challenging task.

3.1.2. Obtaining the linear model

The linear model in state-space formulation is obtained easily with the help of the dynamic
simulators discussed above. Determining the stability of the model is also an easy task. All it takes is
to calculate the eigenvalues of the system matrix, obtained after the linearization. However, for
unstable systems, the origin of the instability is often very difficult to identify.
It has to be stressed that many chemical and biochemical plants have a strongly nonlinear
behaviour. Some of the most common features are the high parametric sensitivity and the state
multiplicity. These effects are enhanced by coupling the units through heat integration or material
recycles. For this reason, the linear models are reliable near the linearization point, but their accuracy
is poor for large disturbances.

3.1.3. Obtaining the balanced state space model

The balanced realization methods for model reduction are now a well-developed and relatively
simple group of techniques.
For linear systems, the approach requires only matrix computation. In the same time, the error
bounds between the original and the reduced model can be easily and a-priori calculated with the help
of the induced Hankel norm (Skogestad & Postlethwaite, 1996).
For stable systems, the balancing of the state space is straightforward. If the system is
unstable, the stable part is isolated, balanced, and added back to the unstable part of the model.
For very large models, the algorithm often fails due to the ill conditioning of the Lyapunov
equation solved while calculating the gramians (Mullhaupt & Riedel, 2004).

3.1.4. Obtaining the reduced-order model

In the next step, the reduced-order model is obtained from the balanced space model.
Reduction of the system is achieved by retaining only certain states in the representation. This
operation is equivalent to defining a certain subspace within the state space. Small Hankel singular
values in the balanced realization indicate state variables with little contribution to the input-output
behaviour.
The reduced-order model is obtained by equating to zero these variables (truncation) or their
time-derivatives (residualization). Truncation is more accurate in representing the initial part of the
dynamic response, but residualization preserves the steady-state gain. If the dynamics of the systems
is high-order, no significant reduction can be achieved.

82

Application of model reduction to design of plantwide control systems

3.1.5. Plantwide control structures
The term plantwide control refers to all the structural and strategic decisions that have to be
taken in order to accomplish a desired policy of operation. Two approaches are often used for the
assessment of the plantwide control structures (Skogestad & Larsson, 1998):
(a) Several alternatives for the plantwide control structure are evaluated, for a given design
of the plant. The main disadvantage of this approach is the large number of alternatives to be
generated and assessed. In the same time, it is difficult to have a precise problem definition.
This can lead to several solutions for the same plantwide control problem (Larsson &
Skogestad, 2000).
(b) The plantwide control and the design are carried on in the same time. This approach
(Bansal, Perkins, Pistikopoulos, Ross & van Schijndel, 2000; Schweiger & Floudas, 1997)
involves, very often, solving multi-objective optimization problems, based on full nonlinear
dynamic model of the process.
The design of the plantwide control structures also brings some particularities when model
reduction is considered. Although the goal is a control structure for the whole plant, many control
loops are local to certain units. For example, instabilities arising from heat-integration are solved by
manipulating local heat duties; in distillation, composition of product streams is controlled by reflux
rate or reboiler duty.
From a plantwide viewpoint, the design of the control structure is mainly concerned with the
inventory of reactants, products, impurities and by-products (Bildea & Dimian, 2003). The solution is
much simpler by excluding the local control loops from the analysis. This can be done by including
these local control loops into the reduced model of the unit.
Both approaches need, however, a dynamic model of the plant. In the same time, this model
has to be solved many times until a solution of the plantwide control problem is reached. If the model
is solved fast, the control structure alternatives are easily evaluated.

3.2. Structure-retaining model reduction algorithm

After presenting the classical method to obtain the plantwide control structure of the plant
using the model reduction and the main difficulties that can be encountered, a new approach is further
introduced. This approach uses the knowledge about the structure that exists in a chemical plant in the
form of units that are connected by material streams, which mirrors the decentralization of the control
problem (Vasbinder & Hoo, 2003). The recommended procedure is to apply model-order reduction
and/or model simplification to individual units, and then to couple these reduced models. The result
of this coupling is a reduced model of the full plant. The effectiveness of the approach will be proven
by means of a case study: the iso-butane butene alkylation plant.

83

Chapter 3
To obtain the reduced model of the plant, the following steps have to be considered:

A. Identification of units or groups of integrated units to which local control is applied

There are several ways to split the plant flowsheet into units. For example, the flowsheet is
divided into units according to the equipment type: mixers, pumps, reactors, heat exchangers,
distillation columns etc. For the distillation columns, an even more detailed decomposition can be
considered: reboiler, trays, and condenser. Another idea is to split the flowsheet considering the main
operations: mixing section, reaction section, separation section etc.
Therefore, the number of individual blocks to be further analyzed depends on the criteria used
for flowsheet decomposition. For the case of a very fine decomposition, a large number of blocks are
obtained, which lead to a large increase of the time allocated for individual block analysis. When the
decomposition is too raw and the number of blocks is too small, the added value of the approach is
insignificant.
One way of achieving the flowsheet decomposition is by using the concept of controlled group
of units (CGU, Naka, Lu & Takiyama, 1997). The concept of CGU represents an independently
operable part of the flowsheet, which provides an inventory control function to maintain it at a certain
steady state for some time. Each CGU has an internal structure comprised of a number of processing
units and control loops. The decomposition into CGUs should be done in such way that the dynamic
interactions between them are minimized (Naka Lu & Takiyama, 1997, Seki & Naka, 2006). To
achieve this, the connection between the parts of the plant through mass and energy streams are
considered. Then decomposition is done by cutting streams of material, while units connected by an
energy stream are kept together.

B. Application of tailored reduction techniques to individual units or groups of

aggregated units
Each of the models of the units has smaller size compared to the original full model.
Therefore, simple methods such as the reduction of the balanced state space model (which are not
suitable for large-scale models) are easy to apply and to implement numerically for the purpose of the
model reduction.
Moreover, since the error bounds between the original and the reduced-order model can be
calculated a priori (see above), the cut-off criteria for the smaller Hankel singular values for different
process units is defined in a consistent way. For the units with a strongly nonlinear behaviour, the use
of model simplification or nonlinear model-order techniques can be a very good solution. In addition,
due to the smaller size of the unit model, the chances of obtaining a reduced model that is difficult to
solve decrease significantly.

84

Application of model reduction to design of plantwide control systems

C. Obtaining the reduced model of the full plant
The reduced model of the full plant is obtained by connecting the reduced models of the
individual units. The connection is realized through streams characterized by the flowrates of each
chemical species, temperature and pressure. To assess the quality of the reduced-order model, various
disturbances are applied on one of the feed streams of the unit or plant. The responses in the outlet
stream are recorded and compared further with the responses predicted by the full-order model, for
the same disturbances.
Figure 3.1 introduces the algorithm of model reduction with process knowledge. The steps
that need to be performed in order to obtain the reduced model are:
(1) A rigorous dynamic (nonlinear) model of the plant is developed. This step is performed
using several commercial packages, as mentioned in Section 3.1.
(2) Identify the CGUs in the plant flowsheet. The concept of CGUs is presented above.
(3) Achieve the decomposition of the plant flowsheet. This decomposition should cut only
material streams.
(4) The rigorous dynamic model of each CGU is developed using commercial packages. At this
point, control structures for the CGU are added, if necessary.
(5) Proceed to the model reduction of each CGU. The following sub-steps have to be
performed:
a.

Obtain a linear model of the CGU. As mentioned above, many commercial

b.

The balanced realization of the linear model of the CGU is obtained.

c.

The model reduction of the balanced model is performed.

packages can perform this task easily.

(6) Check if the resulting reduced models of the CGUs are accurate enough. This can be
achieved by comparing the behaviour predicted by the reduced and the rigorous models of
each CGU, in the time domain or in the frequency domain (or both). The comparison can be
graphical or statistical. If the resulting reduced model is not accurate enough, another
technique should be used.
(7) If the reduced models of the CGUs are accurate, the full model of the plant is reconstructed
by coupling together the reduced model of the CGUs.
(8) Check if the obtained full model of the plant is accurate. This is done by validation against
the original nonlinear model. If the full model is not accurate enough, return to
a.

Step (5): A different model reduction of the CGUs is attempted;

b.

Step (3): The balanced realization of the linear model of the CGU is obtained.

c.

Step (2): Different CGUs are identified.

(9) If the full model is accurate enough, the algorithm stops.

85

Chapter 3
Start
Base case
(Full plantwide model)

Identification
of
CGUs
Plantwide model
decomposition
Base case
(CGU model)

Individual unit
control structure

Model reduction
of CGU
First iteration
Subsequent iterations
Linearization
of CGU

Other model reduction

technique
(case dependant)

Balanced realization
of CGU

Model reduction of
CGU balanced model

Reduced model
(CGU)

NO

Accurate ?
(comparison
in the time
domain)

YES

Plantwide
control structure

Plantwide model
reconstruction

Reduced model
(Full plant)

NO

Accurate ?
(comparison
in the time
domain)

YES

Stop

Figure 3.1.The algorithm of the structure retaining model reduction approach

86

Application of model reduction to design of plantwide control systems

In the following section, the effectiveness of the approach will be proven by means of a case
study.

4. Case study: plantwide control of iso-butane - butene alkylation

plant
The alkylation of iso-butane with butene is a widely used method for producing high-octane
blending component for gasoline. In industrial practice, the modern processes use the sulphuric acid
as a catalyst. Although the chemistry of alkylation is very complex, for our purposes, the following
reactions will be used for the overall chemistry:
C4 H 8 i C4 H10 i C8 H18

(3.10)

C4 H 8 i C8 H18 C12 H 26

(3.11)

This particular case study has been chosen because the plant structure includes the most
common units that can be seen in the industry: heat integrated reactor, distillation columns, mixers,
pumps, recycles. The structure of the plant flowsheet will be presented in the following, while a
detailed description of the process can be found in the book of Dimian and Bildea (2008).
In the classical approach to model reduction, in the control community, the reduced model is
developed for controller tuning. In the following, a reduced model of the alkylation process will be
developed, to be used for the analysis of plantwide control structures. Several different plantwide
control structures are proposed and the behaviour of the reduced model of the plant is assessed. This
is done for deciding the control strategy, and not the controller tuning, which means that steady state
behaviour characteristics, such as sensitivity, are important. Moreover, the plantwide control
structures do not control the different units in the plant flowsheet (like the reactor, or the distillation
columns), but the way the reactants are introduced into the process. Once the reduced model of the
plant is obtained, the plantwide control structure is added and the behaviour of the model is assessed.

4.1. Process description and the rigorous model

The plant produces 26 kton/year iso-octane, with a selectivity of 85% for a yearly run of 8000
hours. The raw materials are the butene and the iso-butane. The butene stream has a large amount of
propane, which is an inert for the reaction, as impurities (10%). The main product is the iso-octane.
Since the selectivity of the process is not 100%, some heavy products (modelled as dodecane) are
formed.

87

Chapter 3
iso-Butane
Butene

iso-Butane
iso-Octane
Butene

Separation

Reaction

Dodecane

Propane

Propane
Figure 3.2. The basic iso-butane alkylation plant flow scheme
REC

iso-Butane
FB 0

MIX1

F1

Propane

F8

INERT

HEX2

Butene
+
Propane

FA0+FI0

F6

MIX2

F7

HEX1

REACTOR
SPLIT

F3

F2

FIN

COL1

iso-Octane

F9

i-OCTANE

F4

P1

COL2
FOUT

P2

F10

F5

COL3

HEAVIES

Dodecane
Figure 3.3. The iso - butane butene alkylation plant flowsheet

The process consists of two basic operations (Figure 3.2). The main operation is the alkylation
of iso-butane with butene to iso-octane and dodecane, which occurs in the reaction section. The two
reactants are converted into products at low temperature, in a reactor modelled as a CSTR. The
reaction volume is 4 m3 and the reaction pressure is 8 bar. In practice, the temperature ranges between
0 and -10 0C. In this study, the reaction temperature is chosen to be -5 0C. The cooling is achieved in

88

Application of model reduction to design of plantwide control systems

an external heat exchanger. For heat integration purposes, the reactor outlet stream acts as coolant to
the reactor inlet stream, as seen in Figure 3.3.
Table 3.1. Distillation column design specifications
Column

Units

specifications

Column 1

Column 2

Column 3

(COL1)

(COL2)

(COL3)

29

24

Number of stages

28

Feed stage

18

10

11

Reflux ratio

110

0.1

0.2

Distillate to feed ratio

0.017

0.95

0.91

Reboiler duty

MW

6.65

3.8

0.09

Condenser duty

MW

-6.28

-3.54

-0.33

After reaction, the main product is separated from the reactor outlet stream. The separation is
done by distillation, through the direct sequence. The inert is removed in the first column, while the
reactants, having similar volatilities, are removed next.
In this way, only one recycle stream appears in the flowsheet. The third column separates the
product from the heavies. A summary of the distillation columns design specifications is presented in
Table 3.1.
Table 3.2. Stream summary for the iso-butane alkylation plant
Stream

Temperature

Pressure

0C

bar

FB0

25.0

FA0 + FI0

25.0

Mass flow

kg
s
A

8.0

0.580

8.0

0.550

0.108

FIN

2.3

8.0

1.378

10.662

0.001

Traces

0.110

FOUT

-5.0

8.0

10.409

126.461

11.564

1.734

1.379

F8

48.9

9.0

0.833

10.117

0.925

0.139

0.110

INERT

26.8

8.0

0.001

0.054

Traces

Traces

0.108

F9

61.1

9.0

0.832

10.063

0.925

0.139

0.002

REC

57.0

8.0

0.828

10.054

0.001

Traces

0.002

F10

194.3

8.3

0.004

0.008

0.924

0.139

Traces

IOCTANE

93.5

1.2

0.004

0.008

0.923

0.00

Traces

HEAVIES

228.2

1.5

Traces

Traces

0.001

0.139

Traces

89

Chapter 3
Once the plant structure has been decided, the next step is the equipment design and sizing.
This was achieved using Aspen Plus. A summary of the main streams composition is presented in
Table 3.2.
After obtaining a steady-state model in Aspen Plus, a dynamic model of the iso-butane
alkylation plant is developed in Aspen Dynamics. Aspen Dynamics creates a very basic plantwide
control structure consisting of a push flow control scheme on the raw material feeds, pressure control
and where required, level control. However, this basic control structure does not involve any quality
control loops. This basic dynamic model will be called the open-loop model and will be analyzed
further.

Reactor temperature / [C]

0.025

0.05

0.075

-5

-10

-15

Bottom stream mass flow

/ [kg/h]

50000

45000

40000

Time / [h]

0.2

0.4

0.6

Time / [h]

Figure 3.4.Instability of the open-loop plant:

a) The temperature of the reactor; b) The mass flow of the first column bottom stream

The simulation of the open-loop plant shows that the plant is unstable. Starting from the
nominal (steady-) state, the reactor temperature drops fast and all flow rates dramatically increase
(Figure 3.4). Soon, one level control loop reaches its limits, and overflow occurs.

4.2. The linear model

The original plantwide model is linearized around the nominal operating point (steady state).
This step is done using the linearization toolbox provided by Aspen Dynamics. The system matrix
A has three positive eigenvalues, with unclear origin. The instability could be determined by the

heat-integrated reactor, which presents a strongly nonlinear behaviour, as it will be shown later.
Another source of instability could be the reactor-separation-recycle structure. It is also possible that
one or more distillation columns present multiple steady states, phenomenon that is accompanied by
instability (Jacobsen & Skogestad, 1995). An experienced modeller will consider other possible
causes of the plants instability. The issue of the three positive eigenvalues needs to be analyzed
further.

90

Application of model reduction to design of plantwide control systems

4.3. The balanced model

Obtaining the balanced model using MATLAB fails: the algorithm complains about illconditioned matrices. A balanced realization of the linear model of the original plant, which has 554
states, cannot be obtained. The cause is not easy to be determined.
As seen above, the use of reduced-order models, in a classical way, for designing the
plantwide control seems to be hopeless. It looks difficult to develop a reduced model for the isobutane alkylation plant using this approach. For the assessment of the plantwide control structure,
other methods of obtaining the reduced model have to be used. The following steps will make use of
the knowledge about the process while achieving the model reduction.

4.4. Exploiting the structure

For obtaining the plantwide control structure, the flowsheet of the iso-butane alkylation plant
is decomposed into units and groups of units to which local control is applied. The decomposition
does not cut the connection between units that are heat integrated.
The plant model is split into the following units: mixers, reactor (together with the heat
integration around it), distillation columns, and pumps (Figure 3.3). The mixers and the pumps are
assumed instantaneous, so they are not interesting from the point of view of the model reduction. For
the models of each unit, the molar flowrates of each component, the temperature and the pressure for
the inlet stream are considered as inputs, while the molar flowrates of each component, the
temperature and the pressure of the outlet stream/streams are considered as outputs. Inputs that are
usually considered for control, such as reflux rate or reboiler duties, are already assigned to local
control loops. Therefore, from a plantwide control perspective, the set points of these local control
loops are available as manipulated variables.
The model reduction techniques will be applied further to the reactor and each distillation
column, individually.

4.4.1. The reactor sub-system

The analysis of the reactor subsystem shows instability. However, the unit can be easily
stabilized by adding a temperature controller manipulating the cooling duty. After the stabilization,
the unit is linearized around the steady-state point. The linear model that is obtained has 15 states, and
the agreement with the nonlinear model is good, as seen in Figure 3.5.a.
Further, the reduction of the linear balanced model is attempted. The reduced-order model has
only seven states, but the agreement with the full nonlinear model is questionable (Figure 3.5.a). It
can be concluded that, for this case, no significant order-reduction at high accuracy can be achieved.
A second approach is considered in order to reduce the size of the reactor system model,
preserving in the same time its nonlinearity. This reduction is based on model simplification.

91

Chapter 3

(Deviation from the steady state)

P flow rate /[kmol/h]

0.6

Reduced model 7 states

Linear full model 15 states
Nonlinear full model 15 states

0.3

0
0

0.5

Time /[h]

P flow rate / [kmol/h]

32
Simplified model

Nonlinear full model

30

28
0

0.15

0.3

Time / [h]
Figure 3.5. Model reduction of the reactor sub-system: iso-Octane molar flow in the outlet stream:
a) Time response of the nonlinear, linear and reduced model for a step of 1 kmol/h on the butene in the feed; b)
Time response of the nonlinear and simplified model for a step of 5 kmol/h on the butene in the feed.

A new dynamic model is written, consisting of five component balances, and considering
constant temperature and physical properties:

92

Application of model reduction to design of plantwide control systems

dn
( A)
FIN( A) FOUT
V r1 V r2
dt
A

dn
(B)
FIN( B ) FOUT
V r1
dt
B

dn
(P)
FIN( P ) FOUT
V r1 V r2
dt
P

(3.12)

dn
(R)
FIN( R ) FOUT
V r2
dt
R

dn
(I )
FIN( I ) FOUT
dt
I

FIN( i ) = the molar flow of component i in stream FIN , kmole

h

where:

(i )
FOUT
= the molar flow of component i in stream FOUT , kmole
h

n = the number of moles of component i , [ kmole]

i

i A, B, P, R, I

V = the volume occupied by the reaction mixture, [ m3 ]

rj = the rate of reaction j , kmole 3
m h

j = 1, 2
t = time, h

The volume occupied by the reaction mixture is calculated as a function of the number of
moles of component i as:

V n v
i

(3.13)

3
i

v = the molar volume of component i , m

kmole

where:

The values for the molar volumes are presented in Table 3.3.
The reaction rates are assumed proportional to the reactant concentrations:
r1 k1 c c
A

r2 k2 c c
P

(3.14)

(3.15)

93

Chapter 3
Table 3.3. The molecular volumes of component i
Component

Units

0.082

Propane (I)
iso-Butane (A)
1-Butene (B)

Molecular volume

kmole
3

0.089
0.098

iso-Octane (P)

0.158

Dodecane (R)

0.222

i
c = the molar concentration of component i , kmole 3
m

where:

k j = the rate constant of reaction j , m

kmole h

The rate constant of reaction j it is assumed to be following an Arrhenius-type dependence on

temperature:

k j k j ,0 e

Ea , j

(3.16)

R T

The values for the pre-exponential factors, k j ,0 and the activation energies, Ea , j are taken
from (Mahajanam, Zheng & Douglas, 2001) and are presented in Table 3.4.
Table 3.4. Kinetic parameters for reactions j
Component

Pre-exponential factor

k j ,0
Activation energy

Ea , j

Units

Reaction

Reaction

m3

kmole s

1.62 109

4.16 1012

kJ

kmole

6.5 104

8.1 104

The molar concentration of component i is determined as:

c
i

n
V
i

(3.17)

94

Application of model reduction to design of plantwide control systems

3

k j ,0 = the pre-exponential factor for reaction j , m

kmole h

with:

Ea , j = the activation energy of reaction j , kJ

kmole

The molar flow of component i in the outlet flow is calculated in the following way:
c
i

(i )
FOUT
FOUT

(3.18)

c
i

In practice, a constant temperature inside the reactor is achieved by the use of a temperature
controller. In this way, there is no need for the energy balance equation, and one differential variable
becomes an algebraic one. Moreover, the equations for determining the temperature-dependent
physical properties of the components are no longer necessary.
The control of the heat-integrated reactor consists of a temperature controller, manipulating
the cooling duty, and a level controller, manipulating the outlet flow.
One extra differential equation is added to the model in order to ensure level control on the
reactors outlet flow. A PI controller is considered for this purpose. This extra equation is defined as
follows:
t

1
FOUT FOUT ,0 K t d
t
i 0

(4.19)

FOUT = the molar flow of the reactor outlet stream, kmole

h

where:

FOUT = the bias, kmole

h

K = the gain, kmole 3

m h

= the error between the measured and the desired value, m3

ti = the integral time, h

With the error between the measured and the desired value is determined as:

V V
*

where:

(4.20)
V * = the specified reactor volume, [ m3 ]
V = the measured reactor volume, [ m3 ]

95

Chapter 3
As seen in the Figure 3.5.b, the agreement with the AspenDynamics nonlinear model is
excellent.

4.4.2. The distillation columns

Unexpectedly, the first and the third distillation column are also unstable. It should be
remarked that the dynamic model specifies, realistically, the reflux on a mass basis, in contrast to the
steady-state simulation, which uses moles.
Realizing

the

large

25

Molar flow /[kmol/h]

relative volatilities and the very

different molar weights, the
existence

of

multiple

or

unstable steady states becomes

a possibility (Jacobsen and
Skogestad,

1995).

switching

to

specifications

Indeed,

12.5

stabilizes

The reflux specified

as mass flow

mole-based

columns (Figure 3.6).

The reflux specified

as molar flow

20

40

Time / [h]

the

Figure 3.6.The instability of first column: Distillate stream molar flow

However, in the following the distillation columns will be modelled with mass-based reflux
specification. Temperature controllers will be used to obtain a closed-loop stable system. Therefore,
the control of the three distillation columns consists of temperature control both in the top and in the
bottom of the columns, in addition to the inventory control loops (pressure, levels in the reflux drum
and column sump).
As well as for the reactor system, the model of each distillation column is linearized, and then
the reduced-order model of the balanced realization is obtained.
For the distillation columns, a significant order-reduction is achieved with excellent accuracy,
as seen in Figure 3.7.a, where results are presented for the first distillation column. From a nonlinear
model with about 200 states, the reduced-order model is linear and has around 20 states.
Table 3.5.Model reduction results
Unit

Model reduction technique

Full nonlinear model

Reduced model

CSTR

Model simplification

15 states

5 states

185 states

20 states

191 states

25 states

163 states

22 states

150

30

COL1
COL2

Model-order reduction

COL3
Simulation time, (s)

A summary of the model reduction results is presented in Table 3.5.

96

Application of model reduction to design of plantwide control systems

(Deviation from the steady state)

P flow rate / [kmol/h]

Reduced model 20 states

Nonlinear full model 185 states

2.5

0
0

0.15

0.3

Time / [h]

(Deviation from the steady state)

P flow rate / [kmol/h]

Second order
+
dead time model
Nonlinear full model 185 states

2.5

0
0

0.15

0.3

Time / [h]
Figure 3.7. Model reduction for COL1: iso-octane molar flow in the bottom stream:
a)

Time response for the full nonlinear and reduced model for a step of magnitude 5 kmol/h (~10%) on the feed
stream molar flow; b) Time response for the full nonlinear and the small order model for a step of
magnitude 5 kmol/h (~10%) on the feed stream molar flow

However, from the appearance of the time response of the full nonlinear model (Figure 3.7.a.),
it should be remarked that the behaviour can be well approximated by a model of even lower order,
such as a second-order plus dead-time model with parameters estimated with the help of the system
identification toolbox provided by MATLAB.

97

Chapter 3
Figure 3.7.b. compares the predictions of the nonlinear and the second-order and dead-time
models of the first distillation column, considering the iso-octane molar flowrate in the bottom stream
as output. The agreement between the two models is excellent.
For the further steps in the assessment of the plantwide control structures, the order-reduced
models obtained from the balanced realization of the linear models will be used.

4.5. Assessment of the plantwide control structure

At this point, three different models are available:
(a) The original nonlinear model (R)
(b) The reduced-order linear model (M1)
(c) The reduced-order nonlinear model (M2)
The original nonlinear model, R, is developed in AspenDynamics and will be taken as
reference. Model M1 contains the reduced-order linear model of each unit in the flowsheet, while
model M2 contains the reduced-order linear models of the distillation columns and the simplified,
nonlinear model of the reactor.
Different plantwide control structures are considered and evaluated. In the following, two of
them will be discussed (Figure 3.8).
In control structure CS1, the fresh feeds of both reactants are fixed. This has the advantage of
setting directly the production rate and the product distribution. The disadvantage of this control
structure is that is highly sensitive to disturbances, as it will be shown later.
In control structure CS2, the butene fresh feed is fixed, but the iso-butane is brought into the
process on inventory control. This eliminates the disadvantage of the previous control structure.
As the two control structures presented above differ only in the way the reactants are being
brought into the process, the reduced-order models of the reaction and separation section can be easily
reused.
In order to assess the quality of the reduced models M1 and M2 and their applicability to
plantwide control design, the control structures CS1 and CS2 are applied and the predictions of the
full and reduced models are compared.
When the butene feed is decreased by 10%, the full model predicts that control structure CS1
is not able to control the plant: the amount of butene in the process is not enough to consume the isobutane. The reactant is accumulating in the plant and the recycle flow is continuously increasing.
While simulating the full nonlinear model R, because of the high increase of the recycle flow, the
Aspen Dynamics simulation crashes and has to be stopped. This is also the reason for the smaller
amount of data available for this model. The nonlinear model M2 is correctly predicting the high
sensitivity to disturbances of this control structure, as shown in Figure 3.9. However, the linear model
M1 wrongly indicates that a new steady state is reached. In other words, having a linear reduced-order
for the reactor does not preserve the systems behaviour.

98

Application of model reduction to design of plantwide control systems

FB0

F3
LC

FC

F1

FC

FA0

FP

LC

Reaction
section

Separation
section
FR

CS1
FB0
F3
F1
LC

FC

FC

FA0

FP

LC

Reaction
section

Separation
section
FR

CS2
Figure 3.8.Control structures for the alkylation plant

It can be concluded that the reduced-order model must be able to preserve the nonlinear
behaviour of the full model.
When the same disturbance is applied to the control structure CS2, a new steady state is
reached in a relatively short time for the nonlinear model R and M2, as seen in Figure 3.10. The
excellent accuracy of the nonlinear reduced-order model, M2, is obvious. In the same time, an
important reduction of the computation time, from less than 3 minutes to about 30 seconds, should be
also remarked, as shown in Table 3.5. This reduction of the computation time could result from the
use of linear reduced-order models for the distillation columns. The linearization has as a result the

99

Chapter 3
reduction of the computational effort, as discussed above. In addition, the use of a simple model for
the reactor, which consists only of five mass balance equations, has an important contribution to this

100
(Deviation from the steady state)

Recycle molar flow / [kmol/h]

reduction of the computation time.

CS1

M2
R

75
50

M1

25
0
0

-25

Time / [h]
Figure 3.9.Dynamic simulation results, for 10% decrease of the fresh butene flowrate: Recycle flowrate.

(Deviation from the steady state)

P molar flow / [kmol/h]

0.5

CS2
0

R
-1

M2

-2.5

Time / [h]
Figure 3.10.Dynamic simulation results, for 10% decrease of the fresh butene flowrate: Production rate.
(a)

The reduced model M1 does not predict the sensitivity of control structure CS1, while M2 does
(b)

The reduced model M2 correctly predicts the behaviour of control structure CS2

100

Application of model reduction to design of plantwide control systems

From the results presented in Figure 3.9 and Figure 3.10 it can be concluded that the control
structure CS2 performs better than the control structure CS1 for the case presented above. This
control structure is recommended for use further.
In the discussion above, the way the rigorous, the linear and the reduced models were obtained
has been discussed. Since for the purpose several tools were used, in the following the way these tools
were connected in order to achieve the goal will be summarised.
The rigorous model was obtained using the Aspen Engineering Suite. Aspen Plus and
Aspen Dynamics were used for obtaining the steady state and the dynamic model of the alkylation
plant, respectively. The linear models were determined with the help of the linearization toolbox in
Aspen Dynamics. The balanced realization of the state space model and the model reduction were
performed in MATLAB, as well as the parameter identification for the reduced models of the
distillation columns. Finally, the simplified reactor model and the complete reduced model of the
alkylation plant were obtained using SIMULINK.
The connection between these tools was done through the matrices of the state space
representation of the system (the alkylation plant). The output of Aspen Dynamics linearization tool
is the set of matrices presented in equation (3.3). With the help of these matrices, the linear system is
very easy to be implemented in MATLAB, as a state space model. The model reduction performed
in MATLAB to this state space returns another set of smaller size matrices, which are also easy to
turn into a state space model. SIMULINK is a dynamic simulation package that allows the user to
specify a block diagram representation of a dynamic process. As an extension to MATLAB, the state
space models obtained previously are fully compatible.
Another point that needs to be discussed is the time needed for the development of the
reduced-order model using the structure-retaining approach, assuming that a full nonlinear model is
already available. It is estimated that an experienced modeller will develop the reduced-order model
of the alkylation plant is not more than two days. However, obtaining a nonlinear reduced model of a
unit can be more time consuming.

5. Conclusions
The chapter proposes and demonstrates the advantage of considering the structure of a process
flowsheet when developing reduced-order models to be used during the design of the plantwide
control system.
The recommended procedure is to apply the model reduction to the individual units or groups
of units of the plant, and then to connect together the reduced models. The reduction for the
individual units consisted in a linear model-order reduction of the distillation column models and a
simplification of the integrated reactor model.
Some advantages of applying this procedure have been highlighted by means of a case study:

101

Chapter 3

The procedure is flexible. Tailored model reduction techniques are applied to different plant
units. In the case presented above, the reactor model was reduced by model simplification,
while linear model-order reduction of the balanced realization was used for the distillation
columns. The dynamics of the mixers, pumps and the heat exchangers was neglected.

The solution time is significantly reduced, by factor 5.

The nonlinearity of the model can be preserved. This feature is highly desirable since many
units of chemical and biochemical plants have a strongly nonlinear behaviour, such as state
multiplicity or parametric sensitivity, behaviour that is enhanced by the coupling through
heat integration and/or material recycles. It was shown that considering a linear model for
the reactor in the reduced model of the plant did not preserve essential behavioural features
of the system. After applying a disturbance for the control structure CS1, the linear reducedorder model did not predict the system sensitivity, in contrast to the predictions of the fullorder and the reduced-order nonlinear models.

The maintenance and adaptation to future plant changes is facilitated by the modularity of
the reduced model. Once one of the units in the plant flowsheet needs to be replaced, its
reduced model can be easily inserted in the reduced model of the plant.

The modularity of the reduced model is also useful for the case when a rigorous model for
one of the plant units is not available. However, if a simplified model of the unit is available
and accurate enough, it can be easily inserted in the reduced model of the whole plant.
After demonstrating the advantages of using this model reduction technique for the assessment

of the plantwide control structure of a chemical plant, the next step is to use it for other applications in
the chemical industry. For example, dynamic optimization of plant operation is another time
consuming task for which the advantage of using reduced models is obvious.

6. References
Antoulas, A.C., & Sorensen, D.C. (2001). Approximation of large-scale dynamical systems: An
overview,

Technical

report,

http://www-ece.rice.edu/~aca/mtns00.pdf

(last

visited

29.09.2008)
Antoulas, A.C., Sorensen, D.C., & Gugercin, S. (2000). A survey of model reduction

methods for

large-scale systems. Structured Matrices in Operator Theory, Numerical Analysis,

Control,

Signal

and

Image

Processing,

Contemporary

Mathematics,

AMS

Publications 280, 193

Bai, Z. (2002). Krylov subspace techniques for reduced-order modelling of large-scale dynamical
systems, Applied Numerical Mathematics 43, 9
Baldea, M., Daoutidis, P., & Kumar, A. (2006). Dynamics and control of integrated networks
with purge streams, AIChE Journal 52, 1460

102

Application of model reduction to design of plantwide control systems

Bansal, V., Perkins, J.D., Pistikopoulos, E.N., Ross, R., & van Schijndel, J.M.G. (2000).
Simultaneous design and control optimization under uncertainty, Computers and
Chemical Engineering 24, 261

Bhattacharjee, B., Schwer, D.A., Barton, P.I., & Green, Jr., W.H.G. (2003). Optimally-

reduced

kinetic models: reaction elimination in large-scale kinetic mechanisms, Combustion and

Flame 135, 191

Bildea, C.S., & Dimian, A.C. (2003). Fixing flow rates in recycle systems: Luybens rule revisited
Industrial and Engineering Chemistry Research 42, 4578

Dimian, A.C., & Bildea, C.S. (2008). Chemical Process Design, Computer-Aided Case Studies,
Weinheim: Wiley-VCH, 261

Engell, S. (2007). Feedback control for optimal process operation, Journal of Process Control 17,
203
Jacobsen, E.W., & Skogestad, S. (1995). Multiple steady-states and instability in distillation.
Implications for operation and control, Industrial and Engineering Chemistry Research 34,
4395
Larsson, T., & Skogestad, S. (2000). Plantwide control A review and a new design procedure,
Modelling, identification and control 21, 209

Mahajanam, R.V., Zheng, A., & Douglas, J.M. (2001). A shortcut method for controlled variable
selection and its application to the butane alkylation process, Industrial and Engineering
Chemistry Research 40, 3208

Mkil, P.M. (1991). On identification of stable systems and optimal approximation, Automatica
27, 663
Marquardt, W. (2001). Nonlinear model reduction for optimization based control of transient
chemical processes, Proceedings of the 6th international Conference of Chemical
Process Control, AIChe Symp. Ser. 326, Vol. 98, 12

Mullhaupt, A.P., & Riedel, K.S. (2004). Exponential condition number of solutions of the
discrete Lyapunov equation, IEEE Transactions on Signal Processing 52, 1257
Nafe, J., & Maas, U. (2002). A general algorithm for improving ILDMs, Combustion Theory and
Modelling 6, 697

Nagy, Z.K., & Braatz, R.D. (2003). Robust nonlinear model predictive control of batch processes,
AIChE Journal 49, 1776

Nagy, Z.K., & Braatz, R.D. (2007). Distributional uncertainty analysis using power series
and polynomial chaos expansions, Journal of Process Control 17, 229
Nagy, Z.K., Mahn, B., Franke, R., & Allgwer, F. (2007). Evaluation study of an efficient
output feedback nonlinear model predictive control for temperature tracking in an
industrial batch reactor, Control Engineering Practice 15, 839
Naka, Y., Lu, M.L., & Takiyama, H. (1997). Operational design for start-up of chemical processes,
Computers and Chemical Engineering 21, 997

103

Chapter 3
Penzl, T. (2006). Algorithms for model reduction of large dynamical systems, Linear
Algebra and its Applications 415, 322

Prasad, V., & Wayne Bequette, B. (2003). Nonlinear system identification and model reduction
using artificial neural networks, Computers and Chemical Engineering 27, 1741
Ranzi, E., Dente, M., Goldaniga, A., Bozzano, G., & Faravelli., T. (2001). Lumping procedures in
detailed kinetic modeling of gasification, pyrolysis, partial oxidation and combustion of
hydrocarbon mixtures, Progress in Energy and Combustion Science 27, 99
Rathinam, M, & Petzold, L. (2003). A new look at proper orthogonal decomposition, SIAM
Journal on Numerical Analysis 41, 1893

Scherpen, J.M.A. (1994). Balancing for nonlinear systems, PhD Thesis, University of Twente,
Enschede, The Netherlands
Schweiger, C.A., & Floudas, C.A. (1997). Integration of design and control: Optimization
with dynamic models, Optimal Control: Theory, algorithms and applications, Hager,
W.W., Pardalso, P.M. (Eds.), Kluwer Academic Publishers
Seki, H., & Naka, Y. (2006). A hierarchical controller design for a reactor/separator system
with recycle, Industrial and Engineering Chemistry Research 45, 6518
Skogestad, S., & Larsson, T. (1998). A review of plantwide control, Technical report,
http://www.nt.ntnu.no/users/skoge/publications/1998/plantwide_review1/plantwide_review
1.pdf (last visited 29.09.2008)

Skogestad, S., & Postlethwaite, I. (1996).

Multivariable feedback control, Chichester:John

Wiley and Sons, 449

Van den Bergh, J. (2005). Model reduction for dynamic real-time optimization of chemical
processes, PhD Thesis, Delft University of Technology, The Netherlands
Vasbinder, E.M., & Hoo, K.A. (2003). Decision-Base Approach to Plantwide Control Structure
Synthesis, Industrial and Engineering Chemistry Research 42, 4586

104

4
DYNAMIC OPTIMIZATION USING A REDUCED
PROCESS MODEL

The chapter explores the application of the structure-retaining model reduction approach, presented in
Chapter 3, in the context of dynamic optimization of chemical plants operation. The focus is on the derivation and
use of reduced models for the design and implementation of optimal dynamic operation in large-scale chemical
plants. A case study is used to show the application of the model reduction approach, similar to the previous
chapter.

Chapter 4

1. Introduction
As mentioned in Chapter 3, the development of control systems that maintain the steady state
of a chemical plant is one approach to achieve the economical performances in todays competitive
environment. However, this is not the only answer to the problem. Another approach is responding
continuously to the market conditions through dynamic operation. This requires that the control
systems are able to implement the optimal dynamic behaviour.
Dynamic optimization problems for chemical processes can be found in a wide variety of
applications.
One typical example is the assessment of the trajectories for the optimal operation of batch
and semi-batch reactors (Srinivasan, Bonvin & Palanki, 2003; Srinivasan, Bonvin, Visser & Palanki,
2003; Mohan, Rao, Prasad, Krishna, Rao & Sarma, 2005; Arpornwichanop, Kittisupakorn & Mujtaba,
2005; Srinivasan & Bonvin, 2007; Simon, Introvigne, Fischer & Hungerbuehler, 2008).
Other applications include continuous processes in transient regimes, such as
-

Start-up (Arellano-Garcia, Carmona & Wozny, 2008; Sonntag, Su, Stursberg & Engell,
2008; Haugwitz, Akesson & Hagander, 2009)

Shut-down or grade transitions (Chatzidoukas, Perkins, Pistikopoulos & Kiparissides,

2003; Flores-Tlacuahuac, Biegler & Saldivar-Guerra, 2006; Prata, Oldenburg, Kroll &
Marquardt, 2008)

Process scheduling (Nystrom, Harjunkoski & Kroll, 2006; Terrazas-Moreno, FloresTlacuahuac & Grossmann, 2007; Li & Ierapetritou, 2007; Chatzidoukas, Pistikopoulos &
Kiparissides, 2009).

Another class of applications deals with

-

Interaction between process design and the control systems design (Mohideen, Perkins &
Pistikopoulos, 1996; Bansal, Perins & Pistikopoulos, 2002; Konda, Rangaiah & Lim,
2006; Malcolm, Polan, Zhang, Ogunnaike & Linninger, 2007; Exler, Antelo, Egea,
Alonso & Banga, 2008)

Model predictive control (Engell, 2007; Nagy, 2007; Grossmann, Erdem, Morari,
Amanullah, Mazzotti & Morbidelli, 2008)

Process safety studies (Abel, Helbig, Marquardt, Zwick & Daszkowski, 2000; Xia, Zhao
& Jia, 2008; Miri, Tsoukalas, Bakalis, Pistikopoulos, Rustem & Fryer, 2008; Simon,
Kencse & Hungerbuehler, 2009)

An important group of applications includes system identification and estimation (Abu-elzeet, Roberts & Becerra, 2002; Tsai & Wang, 2005; Nagy, Fujiwara, Woo & Braatz, 2008).
Obtaining a solution for such applications is still a difficult task when taking into account the
fact that very often the dynamic optimization problem involves large-scale models. Just like in the
case of the plantwide control, the quality of the model is crucial for achieving the objective: the model

106

Dynamic optimization using a reduced process model

must represent the plant behaviour with good accuracy, but the complexity must be limited because
both applications require repeated solution during limited time.
The structure-retaining model reduction approach will be used in this chapter for the
derivation of the optimal profiles for the dynamic operation of a chemical plant. The chapter is
structured in the following way: Section 2 presents an overview of the dynamic optimization problem
and the main solution techniques. In Section 3, the reduction procedure is applied for the iso-butane
alkylation plant for obtaining the reduced model to be used for the solution of the dynamic
optimization problem. The results of the dynamic optimization are presented in Section 4. Finally,
conclusions are presented in Section 5.

2. Dynamic optimization
In practice, a chemical process is modelled dynamically using differential-algebraic equations
(DAEs) as shown in Chapter 2, Section 4 of the thesis. In this formulation, the differential equations
are used to describe the dynamic behaviour of the system, for example mass and energy balances,
while the algebraic equations describe the physical and thermodynamic relations. The objective of the
dynamic optimization is to determine, for a dynamic system, the time profiles for a set of decision
variables (pressure, temperature, flowrate, heat duty, etc.) that optimise a given performance criterion,
subject to specified constraints (safety, environmental and operating constraints). This task can be
performed in time or along a spatial parameter. Both time and the spatial parameter can be considered
continuous or discrete.
The dynamic optimization problems can be formulated in many different ways, under
different assumptions. For the application in the following section, it will be stated as follows:
tf

min x t , u t , z t , t , dt
u ( t ), t f ,

(4.1)

s.t.

f x t , x t , u t , z t , 0

(4.2)

g x t , u t , z t , 0

(4.3)

xmin t x t xmin t

(4.4)

umin t u t umax t

(4.5)

zmin t z t zmax t

(4.6)

min max

(4.7)

x 0 x0

(4.8)

107

Chapter 4
In this formulation, x(t ) are the state (dependent) variables, u (t ) are the control (independent)
variables, z (t ) are algebraic variables and are model parameters. The dynamic model of the
process is represented by differential-algebraic equations. The equations (4.2) and (4.3) define such a
system. Equations (4.4), (4.5) and (4.6) are the path constraints on the state, control and algebraic
variables, respectively, while equation (4.8) represents the initial condition on the state variables.

is a scalar objective function at the final time, t f .

There are three classes of dynamic optimization problems:
(a) Parameter optimization
The goal of the parameter optimization is to find a parameter vector that minimizes the
objective function (Geromel, Bernussou & Peres, 1994; Uhl, 2007). The parameter vector can
contain both model parameters and/or initial states of the model.
(b) Parameter estimation
The objective of the parameter estimation problem is to fit a model to existing experimental
measurements. Since a perfect fit is not possible to obtain, the problem finds the best fit according to
a criterion penalizing the deviation between the model predictions and the measured data (Tholudur &
Ramirez, 1996; Banga, Versyck & Van Impe, 2002; Rodriguez-Fernandez, Mendes & Banga, J.R.,
2004; Michalik, Hannemann & Marquardt, 2009).
(c) Optimal control
The goal is in this case finding the vector of the control variables u (t ) that minimizes the
objective function, (Mohideen, Perkins & Pistikopoulos, 1996; Carrasco & Banga, 1997; Esposito
& Floudas 2000; Labadie, 2004). For this type of problems, the parameter vector is fixed.

2.1. Numerical methods for dynamic optimization

Currently, the dynamic optimization problems are solved using a variational approach, by
using different strategies that apply non-linear programming (NLP) solvers (Biegler, 2007) or by
dynamic programming (Cervantes & Biegler, 2001).
The numerical methods for the solution of the dynamic optimization problems are usually
classified in:
(a) Indirect (variational) approaches
The indirect (classical) approaches are based on the transformation of the original problem
into a two-point boundary problem using the necessary conditions of Pontryagin (Biegler, 2007). For
problems without inequality constraints, the optimality conditions can be formulated as a set of

108

Dynamic optimization using a reduced process model

DAEs (Biegler, Cervantes & Wchter, 2002). The resulting problem is difficult to solve when
handling active constraints (Banga, Balsa-Canto, Moles & Alonso, 2005; Biegler, 2007).
(b) Direct approaches
The methods based on direct approaches transform the original dynamic optimization problem
into a nonlinear programming problem (NLP). These methods can be separated into two groups:
1)

Sequential methods

The techniques in this group are also known as control vector parameterization. They decompose
the whole system into the control and state spaces. The strategy is to discretize the control
variables, u (t ) obtaining a relatively small non-linear programming (NLP) problem. The control
variables are represented as piecewise polynomials and the optimization is performed with
respect to the polynomial coefficients (Schlegel, Stockmann, Binder & Marquardt, 2005). The
process model is integrated with a DAE solver at each iteration (Biegler, Cervantes & Wchter,
2002), and the polynomial coefficients are updated. This process can become time consuming
for large-scale problems. Moreover, it is well known that these approaches have properties of
single shooting methods and cannot handle open loop instability (Biegler, 2007). For this group
of methods it is possible to incorporate path constraints, but these are handled approximately,
within the limits of the control parameterization (Kameswaran & Biegler, 2006).
2)

Simultaneous methods

These methods are also known as complete parameterization or direct transcription. In this case,
both the control u (t ) and the state variables x(t ) are discretized. The result is that the size of
the NLP problem is much larger, and requires special solution strategies (Cervantes, Wchter,
Ttnc & Biegler, 2000; Cervantes, Tonelli, Brandolin, Bandoni & Biegler, 2002). As a result,
these methods directly couple the solution of the DAE system with the optimization problem; the
DAE system is solved only once, at the optimal point (Biegler, Cervantes & Wchter, 2002).
There are mainly two different approaches to discretize the state variables explicitly: multiple
shooting and collocation on finite elements (Biegler & Grossmann, 2004; Biegler, 2007).

2.2. The sequential method for dynamic optimization

As mentioned in the previous section, in a sequential method the evolution in time of the
control

variables

is

parameterized

by

finite

number

of

(numerical)

parameters,

1 , 2 ,..., N and 1 , 2 ,..., N :

u t t , k

for

k 1 t k

109

(4.9)

Chapter 4

Initial
guess

,
0

Optimizer

i , xi

i , x i i , x i
,

Gradients
calculation

NLP solver

Sensitivities

Objective function
Point constraints

State variables
xi

Control variables
u ( i , i )
DAE
Solver

Figure 4.1. Dynamic optimization algorithm

Given fixed values of parameters, the objective function of the dynamic optimization problem
is evaluated simply by integrating the dynamic system. The parameters are then updated by the
optimization algorithm and the procedure is repeated, as shown in Figure 4.1.
Since the parameters determine the control profiles used to compute the values of the state
variables, the objective function can be written as:

x t, u t , , u t , y t, u t , , t,

(4.10)

This transforms the infinite dimensional optimization problem into a finite dimensional
problem.
Often, the objective function can be written as an additional differential equation in the
DAE system, by differentiating equation (4.1):

x t , u t , z t , t ,
t

(4.11)

3. Model reduction for dynamic optimization

The advantages of the structure-retaining model reduction approach presented in Chapters 2
and 3 of this thesis make it a good candidate for use in the assessment of the optimal dynamic profiles
for the operation of a chemical plant. The method uses the knowledge regarding the structure that
exists in a chemical plant in the form of units that are connected by material streams. By applying
model-order reduction and/or model simplification to individual units, and then coupling these

110

Dynamic optimization using a reduced process model

reduced models to obtain the reduced model of the full plant, several advantages can be envisaged.
For example, the physical meaning of the variables is kept when the model is broken into smaller
parts, which are connected together. Moreover, due to the flexibility of the procedure, the nonlinearity
of the resulting model is kept.
In the following section, the application of this approach will be used for solving a dynamic
optimization problem. This will be done by means of a case study: the iso-butane alkylation plant.
The model reduction is used for modifying the model of the plant in such way that the desired
behaviour, set by the objective of the dynamic optimization problem, is obtained. Using this reduced
model it will be shown that the dynamic optimization problem can be solved, and the optimum
profiles for the control variables can be determined.

3.1. The rigorous model of the iso-butane alkylation plant

For the purpose of this research, the same process presented in Chapter 3 will be used. The
rigorous model developed in Aspen Dynamics, consisting of a CSTR and three distillation columns
(as shown in the flowsheet in Figure 4.2) will be the basis for the assessment of the optimum dynamic
operation profiles.
REC

iso-Butane
FB 0

MIX1

F1

Propane

F8

INERT

HEX2

Butene
+
Propane

FA0+FI0

F6

MIX2

F7

HEX1

REACTOR
SPLIT

F3

F2

FIN

COL1

iso-Octane

F9

i-OCTANE

F4

P1

COL2
FOUT

P2

F10

F5

COL3

HEAVIES

Dodecane
Figure 4.2. The iso - butane butene alkylation plant flowsheet

For the plantwide control, the option that proved insensitive to disturbances during the
assessment of the plantwide control structure is chosen. In this case, the fresh feed of butene is fixed,
while the iso-butane is brought into the process on flow control, as shown in Figure 4.3. When a

111

Chapter 4
disturbance is applied to one of the fresh feeds, a new steady state is reached in a relatively short time
when this plantwide control is used, as demonstrated in Chapter 3 of this thesis. This is an important
characteristic when solving the dynamic optimization problem. Local control is also present: the
reactor is operated at constant volume and temperature. The reaction temperature is controlled using
the coolant flowrate for this purpose.

FB0
F3
F1
LC

FC

FC

FA0

FP

LC

Reaction
section

Separation
section
FR

CS2
Figure 4.3.Control structure for the alkylation plant

For the distillation columns (Figure 4.4), in

addition to the inventory control loops (pressure, levels
in the reflux drum and the column sump), top and
PC

bottom concentrations are controlled as well.

LC

According to the plantwide control strategy, a

C
O
L

new production rate can be achieved by changing the

butene feed. Moreover, an increase of the recycle flow
is necessary in order to keep the selectivity at a high
value. An increase in the flow rates will lead to an
increase in the energy consumption for the separation
of the different components.

TC

TC

3
LC

The changes in flow rate can be done step-like,

the magnitude of the steps being calculated using the
steady state model of the plant, or in a more efficient
way, by dynamic optimization.

Figure 4.4.Control structure for the

distillation columns

112

Dynamic optimization using a reduced process model

The objective of the dynamic optimization problem should be stated before the model
reduction is performed, in order to be able to choose the right variables to be kept in the reduced
model:

Increase the plant production by 20% while minimizing the energy consumption in
the distillation columns during the transient period. Moreover, a new steady state must be
ensured.

3.2. The reduced model

As seen in the previous section, the full nonlinear model is developed using Aspen
Dynamics. In order to obtain the reduced model, the same procedure presented in Chapter 3 is used.
The plant flowsheet (Figure 4.2) is split into controlled groups of units (CGUs). The following
blocks are obtained:
(a) The reactor (CSTR) + the heat exchanger around it;
(b) Three distillation columns (COL1, COL2 and COL3)
(c) Mixing vessels
(d) Pumps
Since the mixers and the pumps are considered instantaneous (no dynamics), they are not
interesting for the purpose of the model reduction.
For this particular case, the reduced model will be developed using gProms instead of
MATLAB. There are two reasons for this choice:
(1) gProms has already implemented a toolbox for solving steady-state and dynamic
optimizations problems. For the dynamic optimization, two methods can be used: (a) the singleshooting and (b) the multiple-shooting method.
(2) The time needed to solve the reduced model using gProms is much smaller compared to
MATLAB, as it will be shown in Section 3.2.3 of this chapter. This advantage will drastically
reduce the computational time needed for solving the dynamic optimization problem.
The reduced model of the iso-butane alkylation plant that will be used for the dynamic
optimization application consists of the reduced models of the different CGUs, together with the
plantwide control loops shown in Figure 4.3.
In the next sections, the procedure for reducing the individual CGUs is presented, with
emphasis on the particularities derived for the type of application considered (dynamic optimization).

113

Chapter 4
3.2.1. The reactor sub-system
As seen in Chapter 3 of the thesis, the reactor has a strong nonlinear behaviour. For this
reason, model simplification is used to reduce this unit. The simplified dynamic model presented in
Chapter 3, Section 4.4.1, consisting of five component balances, and assuming constant temperature
and physical properties is implemented in gProms.
There is a good agreement between the simplified model developed in gProms and the
AspenDynamics nonlinear model, as shown in Figure 4.5.

(Deviation from the steady state)

P flowrate / [kmol/h]

5
Simplified model

4
Nonlinear full model

3
2
1
0
0

0.5

1.5

Time, / [h]
Figure 4.5. Model reduction of the reactor sub-system:
Time response of the nonlinear and simplified model for a step of +20% on the butene in the feed

3.2.2. The distillation columns

For the distillation columns, linear model-order reduction techniques will be used. The linear
model is obtained using AspenDynamics, similarly to the procedure presented in Chapter 3 of the
thesis. Some modifications to the previous model are done in order to be able to use the resulting
models for the purpose of the dynamic optimization problem. The reboiler duty and the reflux ratio
are considered as output variables of the linear model, together with the outlet streams flow rates.
These modifications have an effect on the number of states the resulting models have once the
reduction procedure is applied, as presented in Table 4.1. The balanced realization of the linear
models of the distillation columns is performed in MATLAB and the resulting balanced models are
reduced. The reduced models of the three distillation columns are further implemented in gProms.

114

Dynamic optimization using a reduced process model

3.2.3. Performance of the reduced model
After all the reduced models of the individual controlled group of units (CGUs) are available,
these models are connected in order to obtain the reduced model of the full iso-butane alkylation
plant. An issue one has to deal while connecting the individual reduced models is the initialization. In
gProms, the model needs good starting points in order to converge. This issue is solved by taking
advantage exactly of the decentralization of the control problem. Having separate reduced models for
each unit in the plant flowsheet allows the individual models to be run separately in order to get good
starting values for the reduced model of the full plant.
Table 4.1. Model reduction results
Unit

Model reduction technique

Full nonlinear model

Reduced model
MATLAB

CSTR

Model simplification

COL1
COL2

Model-order reduction

COL3

gProms

15 states

5 states

6 states

185 states

20 states

25 states

191 states

25 states

29 states

163 states

22 states

17 states

150

30

Simulation time, (s)

The outcome of the model reduction procedure is presented in Table 4.1 together with the time
needed for simulating five hours of dynamic behaviour, under similar conditions. These results are
compared with the reduced model obtained with MATLAB.

4. The dynamic optimization

The reduced model for the iso-butane alkylation plant has been obtained and implemented in
the previous Section. The dynamic optimization problem presented in equations (4.1) (4.8) is
further implemented in gProms and solved using the single-shooting method.
The objective function to be minimised is the sum of the reboiler duties in the three distillation
columns. Two control variables are considered:
(a) The flowrate of the fresh feed of butene ( FA0 )
(b) The flowrate of the first mixers outlet stream ( F1 )
These two variables are on flow control in the plantwide control structure CS2, which is
presented in Figure 4.3. A continuous piecewise-constant control is chosen for these two variables in
the implementation of the dynamic optimization problem.
The optimizer is required to determine the optimum control profiles that ensure the 20%
increase in production, together with the time needed to reach this state. Once this goal is achieved,

115

Chapter 4
the optimizer must ensure that a new steady state is reached and that the production is kept constant
for a fixed amount of time.
The two control variables are discretized into twenty-five time intervals. The size of the first
twenty intervals is free, while for the last five intervals a fixed length is imposed to a value of one
hour each. Moreover, a selectivity constraint is imposed on the process, in order to maintain the
formation of the product obtained in the secondary reaction at a low value. The selectivity is
constrained to a value above 80%. A list of constraints is presented in Table 4.2.
Table 4.2. Constraints for the dynamic optimization problem
Variable

Type

Unit

Inequality

Selectivity
Production rate

end-point

Selectivity
Production rate

Interior
point

Flowrate of second
mixer inlet stream

Lower

Upper

bound

bound

80

100

kmole
h

34.7

35.0

80

100

kmole
h

25

45

kmole
h

20

100

In a first attempt, the full nonlinear model is implemented in AspenDynamics. The software
offers the possibility to perform dynamic optimization using a specialized toolbox. However, this trial
was unsuccessful. The solver stuck shortly after the beginning of the simulation and no solution could
be found.
Subsequently, the solution of the problem using gProms (version 3.01) is attempted. All the
constraints mentioned above are introduced as inequality type constraints in the gProms dynamic
optimization toolbox. The different gProms solvers used for the solution of the dynamic
optimization problem are listed in Table 4.3.
Table 4.3. gProms solvers used for the solution of the dynamic optimization problem
Solver

Type

Algorithm

DOSolver

CVP_SS

Control vector parametrization single-shooting

DASolver

DASOLV

Variable time step/variable order backward differentiation formulae

LASolver

MA48

Direct LU-factorization algorithm for large, sparse, asymmetric

systems of linear equations

NLSolver

BDNLSOL

Nonlinear sets of equations rearranged to block triangular form

116

Dynamic optimization using a reduced process model

The optimum control profiles of the control variables, shown in Figure 4.6, are obtained after
several time-consuming, trial-and-error iterations. The values of the constrained variables are listed in

Fresh butene molar flow /

[kmol/h]

Table 4.4.

50
47

Initial profile
Optimum profile

44
41
38

35
0

10

15

20

25

Recycle molar flow / [kmol/h]

Time / [h]
930
Optimum profile

870
810
750
Initial profile

690
0

10

15

20

25

Time / [h]
Figure 4.6. Optimum control profile for: (a) the fresh butene flow rate and (b) the recycle flow rate

The solution was obtained after a number of about 150 manual iterations, in which the user
inputs initial estimates for the control variables and the time intervals. The initial profiles for the
control variables (Figure 4.3) do not satisfy all the constraints, hence the optimizer recalculates them.

117

Chapter 4
Table 4.4. Values of constrained variables
Interval

Time

Type

Production

Selectivity

rate

Flowrate inlet
second mixer

kmole
h

kmole
h

29.13

82.76

44.02

0.95

29.09

82.56

44.06

1.91

29.00

81.37

44.46

2.86

29.25

81.17

44.86

3.81

29.58

81.18

45.26

4.76

29.91

81.21

45.65

5.72

30.29

80.84

46.29

6.67

30.73

80.90

46.81

7.62

31.13

80.86

47.32

8.58

31.52

80.81

47.83

10

9.53

31.91

80.75

48.33

11

10.48

32.28

80.67

48.83

12

11.43

32.64

80.57

49.33

13

12.39

33.00

80.48

49.83

14

13.34

33.54

80.37

50.32

15

14.29

33.73

80.30

50.86

16

15.25

34.11

80.28

51.32

17

16.20

34.47

80.17

51.82

18

17.153

34.72

80.54

51.93

19

18.11

34.82

80.61

52.02

20

19.06

34.88

80.47

52.16

21

20.06

34.90

80.41

52.23

22

21.06

34.95

80.52

52.23

23

22.0588

34.96

80.55

52.23

24

23.06

34.96

80.56

52.23

Interior point

118

Dynamic optimization using a reduced process model

The optimizer is attempting to solve the problem, but eventually stops at a point which is not a
solution. The user takes the results of this optimization step, using them as new initial estimates of the
dynamic optimization problem and the procedure is repeated until the solution is found. Each manual
iteration step required about two hours.
The advantage of having a computationally effective reduced model at this point is obvious,
since the model, in its reduced version, does not converge easily. One cause of the increased
simulation time is the discretization choice, which introduces discontinuities. Another cause might be
in the solver. However, since the gProms dynamic optimization toolbox is using predefined solvers,
this issue cannot be investigated properly.
Further, the optimum profiles of the control variables obtained during the dynamic
optimization procedure were implemented intro AspenDynamics. The agreement between the

Production rate / [kmol/h]

responses of the nonlinear and the reduced models is excellent, as seen in Figure 4.7.

45
40

Full nonlinear model

35
Reduced model

30
25
20
0

10

15

20

25

Time / [h]
Figure 4.7. Comparison between the response of the full and reduced model after implementation of the optimum
control profiles

The results in Figure 4.7 show that the transition time is quite long, as expected when the
objective of the dynamic optimization problem was set. From an initial guess of about six hours, the
optimum solution led to a transition of about 24 hours. To determine the cause of this behavior a
study of the systems time constant should be performed. This study should be done before trying to
solve the dynamic optimization problem, in order to get better initial guesses for the control variables,
and reduce the optimization time accordingly.

119

Chapter 4

5. Conclusions
The chapter demonstrates the advantage of considering the structure of a process flowsheet
when developing reduced-order models to be used during the derivation of the optimum profiles for
the dynamic operation of a chemical plant. The structure-retaining model reduction approach acts on
the upper levels of the physical structure of the model: process units, compartments, domains. The
individual units have been reduced in the following way: linear model-order reduction techniques
have been applied for the distillation columns models, while the heat integrated reactor has been
reduced by model simplification.
Similar advantages of applying this procedure for the assessment of the plantwide control
structures are present for the dynamic optimization of the chemical plants operation:

The procedure is flexible. Tailored model reduction techniques are applied to different plant
units. In the case presented above, the reactor model was reduced by model simplification,
while linear model-order reduction of the balanced realization was used for the distillation
columns. The dynamics of the mixers, pumps and the heat exchangers was neglected.

The solution time is significantly reduced, by factor 75. However, the derivation of the
optimal profiles proved to be a time consuming task. This can be improved by determining
the time constant of the process a priori or by using a different algorithm/tool to perform the
dynamic optimization.
The rigorous dynamic model of the iso-butane alkylation plant could not be used for obtaining

a solution of the dynamic optimization problem.

For solving the dynamic optimization problem using the reduced model a manual approach
must be followed to reach convergence. Several time-consuming optimization steps are performed, in
which the solver does not reach a solution. The best available point available is used as initial
estimates for the following dynamic optimization step, until a solution is found. Although the use of
gProms did lead to a solution, the difficult procedure required showed that the dynamic optimization
was not trivial.

6. References
Abel, O., Helbig, A., Marquardt, W., Zwick, H., & Daszkowski, T. (2000). Productivity optimization
of an industrial semi-batch polymerization reactor under safety constraints, Journal of Process
Control 10, 351
Abu-el-zeet, Z.H., Roberts, P.D., Becerra, V.M. (2002). Enhancing model predictive control using
dynamic data reconciliation, AIChE Journal 48, 324

120

Dynamic optimization using a reduced process model

Arellano-Garcia, H., Carmona, I., & Wozny, G. (2008). A new operation mode for reactive batch
distillation in middle-vessel columns: start-up and operation, Computers and Chemical
Engineering 32, 161
Arpornwichanop, A., Kittisupakorn, A, & Mujtaba, I.M. (2005). On-line dynamic optimization and
control strategy for improving the performance of batch reactors, Chemical Engineering and
Processing 44, 101
Banga, J.R., Balsa-Canto, E., Moles, C.G., & Alonso, A.A. (2005). Dynamic optimization of
bioprocesses: Efficient and robust numerical strategies, Journal of Biotechnology 117, 407
Banga, J.R., Versyck, K.J., & Van Impe, J.F. (2002). Computation of optimal identification
experiments for nonlinear dynamic process models: a stochastic global optimization approach,
Industrial & Engineering Chemistry Research 41, 2425
Bansal, V., Perkins, J.D., Pistikopoulos, E.N. (2002). A case study in simultaneous design and control
using rigorous, mixed-integer dynamic optimization models, Industrial and Engineering
Chemistry Research 41, 760
Biegler, L.T. (2007). An overview of simultaneous strategies for dynamic optimization, Chemical
Engineering and Processing 46, 1043
Biegler, L.T., Cervantes, A.M., Wchter, A. (2002). Advances in simultaneous strategies for dynamic
process optimization, Chemical Engineering Science 57, 575
Biegler, L.T., Grossmann, I. E. (2004). Retrospective on optimization, Computers and Chemical
Engineering 28, 1169
Carrrasco, E.F., Banga, J.R. (1997). Dynamic optimization of batch reactors using adaptive stochastic
algorithms, Industrial & Engineering Chemistry Research 36, 2252
Cervantes, A, Biegler, L.T. (2001). Optimization strategies for dynamic systems, in Encyclopaedia of
Optimization, Floudas, C.A. & Pardalos, E.M. (Eds.), Kluwer Academic Publishers, Vol. 4,
216
Cervantes, A., Tonelli, S., Brandolin, A., Bandoni, A., Biegler, L.T. (2002). Large-scale dynamic
optimization of a low density polyethylene plant, Computers and Chemical Engineering 26,
227
Cervantes, A.M., Wchter, A., Ttnc, R.H., Biegler, L.T. (2000). A reduced space interior point
strategi for optimization of

differential algebraic systems, Computers and Chemical

Engineering 24, 39
Chatzidoukas, C., Perkins, J.D., Pistikopoulos, E.N., & Kiparissides, C. (2003). Optimal grade
transition and selection of closed-loop controllers in a gas-phase olefin polymerization
fluidized be reactor, Chemical Engineering Science 58, 3643
Chatzidoukas, C., Pistikopoulos, S., Kiparissides, C. (2009). A hierarchical optimization approach to
optimal production scheduling in an industrial continuous olefin polymerization reactor,
Macromolecular Reaction Engineering 3, 36
Engell, S. (2007). Feedback control for optimal process operation, Journal of Process Control 17, 203

121

Chapter 4
Esposito, W.R., Floudas, C.A. (2000). Deterministic global optimization in nonlinear optimal control
problems, Journal of Global Optimization 17, 97
Exler, O., Antelo, L.T., Egea, J.A., Alonso, A.A., & Banga, J.R. (2008). A tabu search-based
algorithm for mixed-integer nonlinear problems and its application to integrated process and
control system design, Computers and Chemical Engineering 32, 1877
Flores-Tlacuahuac, A., Biegler, L.T., & Saldivar-Guerra, E. (2006). Optimal grade transitions in the
high-impact polystyrene polymerization process, Industrial and Engineering Chemistry
Research 45, 6175
Geromel, J.C., Bernussou, J., & Peres, P.L.D. (1994). Decentralized control through parameter space
optimization, Automatica 30, 1565
Grosmann, C., Erdem, G., Morari, M., Amanullah, M., Mazzotti, M., & Morbidelli, M. (2008). Cycle
to cycle optimizing control of simulated moving beds, AIChE Journal 54, 194
Haugwitz, S., Akesson, J., & Hagander, P. (2009). Dynamic start-up optimization of a plate reactor
with uncertainties, Journal of Process Control 19, 686
Kameswaran, S., & Biegler, L.T. (2006). Simultaneous dynamic optimization strategies: Recent
advances and challenges, Computers and Chemical Engineering 30, 1560
Konda, N.V.S.N.M., Rangaiah, G.P. Lim, D.K.H. (2006). Optimal process design and effective
plantwide control of industrial processes by a simulation-base heuristic approach, Industrial
and Engineering Chemistry Research 45, 5955
Labadie, J.W. (2004). Optimal operation of multireservoir systems: State-of-the-art review, Journal of
Water Resources Planning and Management-ASCE 130, 93
Li, Z. & Ierapetritou, M.G. (2007). Process scheduling under uncertainty using multiparametric
programming, AIChE Journal 53, 3183
Malcolm, A., Polan, J., Zhang, L., Ogunnaike, B.A., & Linninger, A.A. (2007). Integrating systems
design and control using dynamic flexibility analysis, AIChE Journal 53, 2048
Michalik, C., Hannemann, R., & Marquardt, W. (2009). Incremental single-shooting A robust
method for the estimation of parameters in dynamical systems, Computers and Chemical
Engineering 33, 1298
Miri, T., Tsoukalas, A., Bakalis, S., Pistikopoulos, E.N., Rustem, B., & Fryer, P.J. (2008). Global
optimization of process conditions in batch thermal sterilization of food, Journal of Food
Engineering 87, 485
Mohan, S.V., Rao, N.C., Prasad, K.K., Krishna, P.M., Rao, R.S., & Sarma, P.N. (2005). Anaerobic
treatment of complex chemical wastewater in a sequencing batch biofilm reactor: Process
optimization and evaluation of factor interactions using the Taguchi dynamic DOE
methodology, Biotechnology and Bioengineering 90, 732
Mohideen, M.J., Perkins, J.D., Pistikopoulos, E.N. (1996). Optimal design of dynamic systems under
uncertainty, AIChE Journal 42, 2251

122

Dynamic optimization using a reduced process model

Nagy, Z.K. (2007). Model based control of a yeast fermentation bioreactor using optimally designed
artificial neural networks, Chemical Engineering Journal 127, 95
Nagy, Z.K., Fujiwara, M., Woo, X.Y., & Braatz, R.D. (2008). Determination of the kinetic
parameters for the crystallization of paracetamol from water using metastable zone width
experiments, Industrial and Engineering Chemistry Research 47, 1245
Nystrom, R.H., Harjunkoski, & Kroll, A. (2006). Production optimization for continuously operated
processes with optimal operation and scheduling of multiple units, Computers and Chemical
Engineering 30, 392
Prata, A., Oldenburg, J., Kroll, A., & Marquardt, W. (2008). Integrated scheduling and dynamic
optimization of grade transitions for a continuous polymerization reactor, Computers and
Chemical Engineering 32, 463
Rodriguez-Fernandez, M., Mendes, P., & Banga, J.R. (2004). A hybrid approach for efficient and
robust parameter estimation in biochemical pathways, Biosystems 83, 248
Schlegel, M, Stockmann, K., Binder, T., & Marquardt, W. (2005). Dynamic optimization using
adaptive control vector parameterization, Computers and Chemical Engineering 29, 1731
Simon, L.L., introvigne, M., Fischer, U., Hungerbuehler, K. (2008). Batch reactor optimization under
liquid swelling safety constraint, Chemical Engineering Science 63, 770
Simon, L.L., Kencse, H. Hungerbuehler, K. (2009). Optimal rectification column, reboiler vessel,
connection pipe selection and optimal control of batch distillation considering hydraulic
limitations, Chemical Engineering and Processing 48, 938
Sonntag, C., Su, W.J., Stursberg, O., & Engell, S. (2008). Optimized start-up control of an industrialscale evaporation system with hybrid dynamics, Control Engineering Practice 16, 976
Srinivasan, B., Palanki, S., & Bonvin, D. (2003). Dynamic optimizaton of batch processes: I.
Characterization of the nominal solution, Computers and Chemical Engineering 27, 1
Srinivasan, B., Bonvin, D., Visser, E., Palanki, S. (2003). Dynamic optimizaton of batch processes: II.
Role of measurements in handling uncertainty, Computers and Chemical Engineering 27, 27
Srinivasan, B. & Bonvin, D. (2007). Real-time optimization of batch processes by tracking the
necessary conditions of optimality, Industrial and Engineering Chemistry Research 46, 492
Terrazas-Moreno, S., Flores-Tlacuahuac, A., Grossmann, I.E. (2007). Simultaneous cyclic scheduling
and optimal control of polymerization reactors, AIChE Journal 53, 2301
Tholudur, A., & Ramirez, W.F. (1996). Optimization of fed-batch bioreactors using neural network
parameter function models, Biotechnology Progress 12, 302
Tsai, K.Y. & Wang, F.S. (2005). Evolutionary optimization with data collocation for reverse
engineering of biological networks, Bioinformatics 21, 1180
Uhl, T. (2007). The inverse identification problem and its technical application, Archive of Applied
Mechanics 77, 325

123

Chapter 4
Xia, L., Zhao, Q.C., & Jia, Q.S. (2008). A structure property of optimal policies for maintenance
problems with safety-critical components, IEEE Transactions on Automation Science and
Engineering 5, 519

124

5
DEVELOPMENT OF A REDUCED MODEL FOR ICE
CREAM FREEZING
The focus of this chapter is on the development of a reduced model for a process unit with a complex
product: the freezing step in ice cream manufacturing process. Ice cream consists of a liquid matrix in which air
bubbles, fat globules and ice crystals are dispersed. The presence of multiple thermodynamic phases with
changing size distributions for the dispersed phases and the rotational-axial movement of the ice cream mixture
will make the dynamic modelling of the freezing step very complicated. A reduced dynamic model must be
suitable to support process design and operation in a computationally effective way. The main physical
phenomena accounted for in the model are: (a) the axial convective transport of mass; (b) the radial outflow of
heat at the coolant wall to the refrigerant; (c) the growth of the frozen ice layer at the wall; (d) the periodic
removal of the ice by scraping and the melting of the ice particle population in the bulk liquid; and (e) the cooling
of the ice cream mix.
The development of the reduced model is obtained in two steps. (A) From a conceptually conceivable,
truly comprehensive model of the process, a relatively detailed model with five coordinates is developed. This 5-D
model is still computationally too intensive for use in inverse problems. (B) Further reductions are performed,
primarily aiming at reducing the number of coordinates of the model, going from a 5-D to a 3-D model. The
conservation equations for mass, energy and momentum per thermodynamic phase and population balances are
complemented with rate equations for the relevant physical phenomena, as well as phase equilibrium conditions
and thermodynamic equations of state. The rate laws are kept deliberately simple by coupling a flux to only one
dominant driving force and allowing for experimentally adjustable rate parameters. The resulting 3-D model
consists of several partial differential equations, with the appropriate number of initial and boundary conditions,
and coupled with many nonlinear algebraic equations.
For the use of the reduced model in simulation and optimization applications, it is practical to have a
complementary overview of systems input-output. To this end, the model inputs have been classified into feed
and operational conditions, design and physical parameters, while the outputs relate to the product quality and
process performance.

Chapter 5

1. Introduction
1.1. Model reduction for complex units and products
As seen in Chapter 3 and 4 of the thesis, the model reduction in case of the complex processes
with simple, single-phase products focuses on the upper levels of the network structure of the process.
However, many processes deal with more complex, multiphase products. The complexity of the
process lies in these cases in both the structure of the process and the product. Hence the model
reduction can be performed at both process (units, compartments, etc.) and product (phases, species,
etc.) scales.
In this chapter, the focus will be on developing a reduced model for a single process unit with
a complex product. In this way, the characteristics of the model reduction approach when dealing
with multiphase products will be highlighted. The influence of the full process flowsheet structure
will be neglected at this point.
The main distinction of such a unit with a complex structured product is a reduction in the
network-like aspects of the process structure, as it features only the unit possibly having internal
distinct compartments and domains, like trays in distillation columns. Moreover, more spatial
coordinates inside the unit need to be considered, along which the behaviour and the product structure
is distributed. The product structure considers more dispersed phases and populations having internal
states and a host of transfer phenomena occur between these component phases and populations.
The model reduction can be performed in two stages:
a)

a conceptual reduction in the model development stage

b)

a numerical reduction of the mathematical model of the process

This chapter will focus on the options that can be applied for a conceptual model reduction in
the development stage.
Similar to the previous chapters a case study will be used to outline the model reduction
approach and its advantages. The development of a dynamic model for the freezing step in ice cream
manufacture will be presented. The model is intended to be a tool for the design and operation of the
freezing of ice cream. For the model to be of a real help in product design and processing studies, the
model must be suitable for use in an inverse model, hence the need to ensure the solution of the model
is computationally feasible.

1.2. Ice formation in a scraped surface heat exchanger

The freezing of the ice cream mix is one of the most important operations in making the ice
cream, for upon it depend the quality, palatability, and yield of the finished product (Arbuckle, 1986).
However, understanding the behaviour of the material inside the freezer is still incomplete (Duffy,
Wilson & Lee, 2007). Ice cream (Figure 5.1) is a complex colloidal system comprised, in frozen state,

126

Development of a reduced model for ice cream freezing

of ice particles, air bubbles, partially coalesced fat globules, all in discrete phases, surrounded by an
unfrozen continuous matrix of sugars, proteins, salts, polysaccharides and water.
The freezing process can be divided into two parts: (a) a dynamic freezing step, during which
the mix is quickly frozen while under agitation to incorporate air, and (b) a hardening step, that
freezes the remaining water, without agitation. The first phase of the freezing accounts for freezing of
33-67% of the water depending on the drawing temperature; the hardening process may account for
an additional 23-57%, depending on the drawing temperature (Arbuckle, 1986).
In dynamic freezing, the most important changes in the material streams are the formation of
ice particles, incorporation of air (overrun), formation of small air cells, and destabilization of the fat
emulsion. Each of these steps is critical to production of high quality ice cream with the desired
physical characteristics and must occur whether the freezing is done in a batch or a continuous
process. However, differences in formation of these structural elements may occur based on the type
of freezer used and the type (and amount) of ingredients used in the mix. The product of the dynamic
freezing is soft frozen and flowable, but the degree of stiffness varies with formula, process, freezer
design, overrun and temperature.
The air content / air bubble size and the ice particle size affect the texture of ice cream. For
this reason, these properties could be used as quality indicators. Since these are distributed properties
and direct measurement of these variables inside the freezer barrel is difficult, if not impossible to
achieve, are specified in practice. The feed intake, the processing conditions and equipment design
parameters are adjustable to meet the target output variables.

Figure 5.1. Ice cream structure

The steps that need to be followed in order to obtain the model are:

127

Chapter 5
a)

A detailed model of the ice cream freezing step is developed. The fundamental phenomena are
conceptually considered. Conservation equations for the mass, energy and momentum are
considered, taking into account the complexity of the product. Information related to the phase
equilibrium conditions, equations of state, and the rate equations are added to the model.
Simplifications are introduced in a judicious way, and the variables and the equations are written
on a rigorous coordinate set.

b)

Having too many coordinates for practical use (computation times and effort) of the model, the
coordinate set is reduced. Averaged values are taken for the variables over the domains of the
eliminated coordinates. Moreover, reduction of description of phenomena that essentially play
out in a 3-D geometric space, like the momentum, is needed. Projection on a lower dimensional
space is carried out.

c)

Since the knowledge about the mathematical structure and the parameters of the rate equations is
limited, these rate equations need to be determined with insight from non-equilibrium
thermodynamics (NET) and using experimental data. In a first attempt, each rate equation is
made dependent on a single driving force in a linear way, while most of the rate parameters are
obtained from literature data. The resulting parameters are not representing the reality in full
measure, but the model is able to predict the right trends (for example the order of magnitude).
The obtained reduced model is more sophisticated than any preceding model, yet it is reduced

with the conceptually conceivable rigorous model of the ice cream freezing unit.
The chapter is structured in the following way: Section 2 introduces the physical phenomena
that take place inside the freezer. The detailed (5-D) model, with respect to the conservation
equations, is introduced in Section 3. Section 4 presents the reduced (3-D) model, while the equations
for the rate laws are presented in Section 5. The 3-D reduced model of the freezing unit is
summarized in Section 6. The numerical implementation of the model is presented in Section 7.
Section 9 discusses some applications for future model use with some numerical results, while
conclusions are presented in Section 10 of this chapter.

2. Physical aspects and modelling approach

In the following section, the ice cream ingredients and structure are discussed in more detail.
The features of the freezing equipment, as well as the mechanism are presented in order to create the
perspective for the development of the conceptual model of the freezer unit. The end of this section
discusses the objective and the approach of the model reduction for the ice cream freezing.

128

Development of a reduced model for ice cream freezing

2.1. Ice cream structure

The ice cream microstructure (Figure 5.1) has a strong effect on the sensory manifestation. A
typical ice cream consists of about 30% ice, 50% air, 5% fat and 15% matrix (sugar solution) by
volume.
This means that all three states of matter are contained in its microstructure: solid ice and fat,
liquid sugar solution and gas (Clarke, 2004). These main ingredients provide the sensory properties of
the final product: ice gives cooling, fat provides creaminess, air gives lightness and softness, sugar
provides sweetness, while the flavours enhance the taste.
However, not only the right amount of components gives the desired properties of the ice
cream, but also the way these elements are brought together in the microstructure.
In the following section, the influence of the ice particles, the liquid matrix, the fat droplets
and the air bubbles on the ice cream structure will be discussed in more detail.

2.1.1. Ice particles

Ice particles form a discrete phase inside the ice cream structure. The amount and size of the
particles have a strong effect on the quality of the product. Formulations that lead to numerous, small,
discrete ice particles also lead to enhanced smoothness in texture. Ice crystal formation in ice cream is
generally by secondary nucleation in the scraped surface freezer (Goff, 2002). The freezer is the only
place in the entire ice cream manufacturing and distribution chain where ice is nucleated. In all other
stages of the manufacturing and distribution process any change in temperature results in an increase
(colder temperatures) or decrease (warmer temperatures) in the size of existing particles, or possibly
the complete disappearance of smaller, less thermodynamically stable particles. This implies that the
number of particles in the product after storage and distribution will always be less than the number
formed in the freezer. As the tendency is always for the larger particles to grow at the expense of
smaller ones, it is important to form a large number of small particles in the freezer.

2.1.2. Liquid matrix

The matrix is a concentrated solution of sugars, stabilizers and milk proteins in water. As the
premix becomes freeze-concentrated, components that are either dissolved or dispersed in water are
increasingly brought together as temperature decreases and water is removed from solution. Two of
the important hydrocolloid components include the polysaccharides and the casein micelles.
Typically, at the end of the freezing process, 50% of the water in the ice cream mix is frozen.

2.1.3. Fat droplets

The fat phase exists in the form of tiny globules that have been formed in the homogenizer.
The proteins act as emulsifiers and give the fat emulsion its needed stability. When the mix is subject
to the whipping action of the freezing barrel, the fat emulsion begins to break down partially and the

129

Chapter 5
fat globules begin to flocculate or destabilize. The air bubbles beaten into the mix are being stabilized
by this partially coalesced fat. If emulsifiers were not added, the fat globules would have so much
ability to resist the coalescing, that the air bubbles would not be properly stabilized and the ice cream
would not have the same smooth texture (Eisner, Wildmoser & Windhab, 2005; Koxholt, Eisenmann
& Hinrichs, 2001).

2.1.4. Air bubbles

The air content in the ice cream structure is defined as overrun. The overrun, expressed as
percentage, is generally defined as the volume of gas to the volume of liquid (Clarke, 2004)
The main role of the air is to make the ice cream soft. Air is distributed in the form of
numerous small air bubbles of size range 20-50 m (Goff, 1997). The gas phase volume varies greatly
from a high of more than 50% to a low of 10-15%. Air is incorporated either through a lengthy
whipping process (batch freezers) or as it is drawn into the mix by vacuum (older continuous freezers)
or is injected under pressure (modern continuous freezers). The incorporation of air alone, or the
shearing action alone, independent of freezing, are not sufficient to cause the high degree of fat
destabilization that occurs when ice crystallization and air incorporation occur simultaneously
(Marshall, Goff & Hartel, 2003).

2.2. Scraped-surface heat exchanger

The dynamic freezing of ice cream is performed in a scraped surface heat exchanger (SSHE).
Ice cream freezers consist of a cylindrical barrel typically 0.2 m in diameter and 1m long. However,
the size may vary for the freezers designed for different production rates (Clarke, 2004). The mix is
pumped along with air into the SSHE (Figure 5.2) and the action of the rotor inside the tube blends
the air into the matrix. The freezing medium in the jacket, typically ammonia, freezes the water into a
thin layer of ice on the barrel wall from which finely dispersed ice particles are continuously scraped
and incorporated into the ice cream (Russell, Burmester, & Winch, 1997).
The operation of the freezer is controlled by several variables. The pressure of the refrigerant
sets the temperature at which it evaporates, hence the wall temperature (typically -30 0C). The mix
and air inflow and the ice cream outflow rates determine the residence times, which can vary from 0.4
to 2 minutes, the overrun, the pressure inside the barrel (typically 5 atm) and the throughput (which
can be as much as 3000 l/h in an industrial freezer). All of these factors, together with the rotation
speed of the rotor, which is typically 200 rpm, determine the outlet temperature of the ice cream, as
well as the size distribution of the air bubbles and ice particles.
The greatest advantage of using SSHE is the constant removal of the stagnant film near the
exchanger wall and the subsequent increase in the heat transfer coefficients, and reduction in fouling
at the wall (Rao & Hartel, 2006).
The time between scrapes is typically 0.1 s (Clarke, 2004), thus a very thin layer of ice can
form before it is removed by the movement of the scraper blades. The number of scraper blades may

130

Development of a reduced model for ice cream freezing

vary and is typically in the range of 1 to 4. A typical SSHE cross-sectional configuration is shown in
Figure 5.3.
The small ice particles that are scraped off the freezers wall are dispersed into the liquid mix
by the beating of the dasher. There are large temperature gradients inside the barrel, both radial
(colder at the wall to warmer at the centre) and axial (warmer at the inlet to colder towards the outlet).

Figure 5.2. Scraped-surface heat exchanger for ice cream

Figure 5.3. Cross-sectional diagram of a

freezing

typical scraped-surface heat exchanger

Therefore, the dispersed ice particles into the warmer mix at the centre of the barrel, they melt
and cool it down; near the inlet the particles all melt, but one-third of the way through the barrel the
mix becomes cold enough for the ice particles to survive (Clarke, 2004). The freezing rates inside the
barrel can vary from 5 to 27 0C/min, respectively.
The two primary flow directions of the fluid are flow in the angular direction due to rotation
and flow in the axial direction due to the imposed pressure gradient. The velocity components of the
two flows form a helical path for the product through the SSHE. A complete mixing in the angular
direction is desired to maximize the efficiency of heat transfer and ensure uniform thermal treatment.
At the same time, the mixing in the axial direction should be minimized because the axial mixing
increases the residence time distribution of the product in the SSHE (Wang, Walton & McCarthy,
1998).
Ice cream behaves as a non-Newtonian material. Moreover, both convection and dissipation of
heat can be significant inside the SSHE. Due to these many interacting physical phenomena, coupled
with the fact that the geometry is complicated, the overall behaviour inside the SSHE becomes very
complex.

2.3. Freezing mechanism

In a recent study, (Qin, Chen, & Farid, 2004) shows that, during the ice formation on a
subcooled surface, the ice appears primarily at some spots and forms separate ice islands. These
islands grow larger until the subcooled surface was frozen-over by a thin film of ice. During this
period, the thickness of these ice patches was almost unchanged. For this reason, (Qin, Chen, & Farid,
2004) concluded that this is an indication of the fact that the growth of ice along the cooling surface is
much faster than the growth rate in the normal direction. It was also concluded that the formation of
this film characterizes an intermediate, transitional stage between the ice nucleation and the normal

131

Chapter 5
growth of the ice layer. After this transitional stage, the growth of the ice layer starts. For ice layers
growing from aqueous solutions, concentration gradients of solutes are formed near the growing
front. In this case, the coupling between the heat and the mass transfer should be taken into account
(Qin, Chen, Ramachandra & Free, 2005).
In (Marshall, Goff & Hartel, 2003) the following mechanism of describing the ice formation
and development in a scraped-surface heat exchanger is presented:
(a) Because of the low temperatures and the rapid freezing process, the ice layer forms at the
cylinder wall. This layer is not fully crystallized, but is a slush layer. This slush contains abundant
small ice particles, in addition to a concentrated solute (sugars, salts, etc.) and probably has a slightly
higher density than the original mix.
(b) The scraper blade removes this slush layer from the barrel wall and mixes it with the warmer
mix remaining in the centre of the barrel.
(c) The cold slush melts as it cools the warmer mix. In a continuous freezer, this occurs near the
entry point of the mix into the freezer barrel as the freezing process begins.
(d) Once the bulk of the mix is cooled sufficiently, some of the ice particles from the slush begin to
survive. These ice particles grow into rounded disk-shaped particles based on the conditions (heat
transfer and mixing) in the barrel.
(e) As ice content increases during the freezing process (or along the length of the freezer in a
continuous process), the new ice layer scraped from the wall primarily contributes to the growth of
the ice particles formed earlier in the process, although some nuclei from the slush layer survive and
grow ending up in the product.
(f)

As more and more water is turned into ice, the remaining fluid becomes increasingly viscous

since the formation ice is taking up essentially pure water (other molecules are excluded from ice as
particles form).

2.4. Scope and approach to modelling of the ice cream freezing unit
The modelling of the freezing unit is at its beginnings. Only recently, a transition from a trialand-error approach for the development of the unit to a more rigorous approach, where in the centre
of the design is a model based on fundamental knowledge, has been started. However, the changes are
done in small steps.
In (Bongers, 2006) the freezer unit was modelled in a simplified way. The barrel was
considered as a series of well-mixed stages. Heat, mass and momentum equations were developed
considering the phases as pseudo-continuous. Details regarding the distribution of the air bubbles and
ice particles within the ice cream structure were not taken into account.
In other models, the distribution of ice particles or air bubbles is present only partially. For
example, in (Aldazabal, Martin-Meizoso, Martinez-Esnaola & Farr, 2006) only the ice particles
distribution is considered and the growth of these particles is modelled, while in (Duffy, Wilson &
Lee, 2007) the modelling of the isothermal fluid flow inside the freezer unit is attempted.

132

Development of a reduced model for ice cream freezing

Other models focused on the heat (Sun, Pyle, Fitt, Please, Baines & Hall-Taylor, 2004) or
momentum (Yataghene, Pruvost, Fayoll & Legrand, 2008) inside a scraped surface heat exchanger,
using a generic non-Newtonian fluid.
Table 5.1. Models of the ice cream freezing step
Model attributes

Scope of model

Sun et al,

Aldazabal

Bongers,

Duffy et al,

Yataghene

2004

et al, 2006

2006

2007

et al, 2008

Heat

Ice growth

Heat transfer

Fluid flow

CFD

transfer in

in SSHE

in SSHE

in SSHE

Distributed

No

Yes (Ice)

No

No

No

Domains

Continuous

Fluid

Liquid

Fluid

Fluid

Fluid

Dispersed

Ice

Ice + Air + Fat

Phases

Components

3+

Dynamic

No

No

Yes

No

No

Spatial

equations

Differential

Coordinates

(conservation)

analysis of
flow pattern

SSHE

Internal

Total

Mass

Yes

Partial

Yes

Yes

Yes

Population

No

Partial

No

No

No

Fluid

No

Yes

No

No

Energy

domain
Momentum

Yes

No

Yes

Yes

Yes

Algebraic

Rate laws

Yes

Growth only

Yes

Yes

Yes

equations

Physical

Yes

Yes

Yes

Yes

Yes

properties

An overview of the main characteristics of these models is presented in Table 5.1. The models
which refer to the modelling of the ice cream freezing have the scope in bold, while the rest refer to
models of non-Newtonian fluids in a scraped surface heat exchanger.
To conclude, there is not yet available a model of the freezing unit that combines all the
physical effects present:
-

axial mass transport with spiralling flow and mixing

mass, heat and momentum transfers between phases

non-Newtonian flow behaviour

evolution of size distributions of the dispersed phases, such as ice crystals and air bubbles

scrapping of the solid ice layer from the inner heat exchanger surface, etc

133

Chapter 5
The ambition of this research is to combine some main effects into a new, more
comprehensive model. Yet, the resulted model remains a reduced model with respect to the
conceptually conceivable complete model, covering in principle all relevant coordinates, phases,
species and phenomena.
In the previous sections, information regarding the physical aspects of the ice cream freezing
has been presented. The focus in the following is on the conceptual phase of the model formulation
and reduction for ice cream freezing in a SSHE.
Applying the modelling and model reduction procedure (summarized in Table 2.1 and Figure
2.13, respectively, in Chapter 2 of this thesis) for the ice cream freezing will result in several models.
An overview of the different types of models and in which component of the model the
simplifications were implemented, for each case, is presented in Figure 5.4.
Firstly, there is the conceptual model of the ice cream freezing, as the scientists perception
of physics of the process. This contains all the available information regarding the phenomena and
their interactions in the unit. This model is the mental model reference from which reductions are
going to be made.

6+
Thermodynamic
phases

5
Thermodynamic
phases
5 Spatial
coordinates
3 Spatial
coordinates

5
Thermodynamic
phases

Dire
ctio
n

of R
e

duc
tion

Equations
of
change
Equations
of
change

Phase
equilibrium

3-D
model

4
Spatial domains
4
Spatial domains

Phase
equilibrium

Physical
properties

Physical
properties

Transfer
rate
laws

5-D
model

Transfer
rate
laws

Conceptual
model

Figure 5.4. Ice cream modelling approach

Secondly, there is the 5-D model. This model considers five coordinates (the time and four
spatial coordinates), four spatial domains (as shown in Figure 5.5) and five thermodynamic phases
inside the freezer. The conservation equations are written in a rigorous way. This model includes also

134

Development of a reduced model for ice cream freezing

rigorous information regarding the phase equilibrium and all the rate law equations. Solving a model
with five coordinates (time and spatial) requires a huge computational effort. Moreover, the
knowledge about the mathematical structure and the parameters of the detailed, distributed rate
equations is limited.
For these reasons, a simplification is required in order to make the solving of the model
possible. Firstly, the number of coordinates is reduced by averaging over one or more coordinates.
The averaging procedure is explained later. Moreover, simplifications are introduced in the fluid
flow equations, in the rate laws, as well as in the physical properties models. The result of this
procedure is the 3-D model.
The 5-D and the 3-D models are further detailed in the following sections.

3. Specification of the 5-D model of the ice cream freezing unit

Once the process system to be modelled has been defined, the suitable network structure, as
well as the mathematical relationships for the behaviour of the different entities should be specified.
The following sections discuss the 5-D model of the ice cream freezing unit, which is derived from
the conceptual model.
The first step in developing a chemical engineering model is to set the conservation equations.
For this, all the fundamental physical resources and phenomena are conceptually considered. Thus,
the model includes equations for the mass, energy and momentum for all the phases and domains
considered.
The conservation equations are based on five coordinates are considered: the time, t , the axial
coordinate, z , the radial coordinate, r , the angular coordinate, , and the internal coordinate of the
distributed phase, s . For this reason, this version of the model of the ice cream freezing step will be
called the 5-D model.
This 5-D model is not developed in full detail in this section. Only the conservation equations
are covered. Understanding these conservation equations is a pre-requisite for developing the more
practical reduced 3-D model. In the description of the conservation equations of the 5-D model in this
section, the emphasis is on making proper physical assumptions. To avoid pages of extensive
mathematical equations, the resulting mathematical equations are presented in the Appendices.

3.1. Spatial domains with thermodynamic phases

Taking into account the physical aspects presented in Section 2 and considering the
intermediate stage in ice formation instantaneous, the phases at the process side are assumed split
between two domains (see Figure 5.5). The two domains at the process side are forming the barrel
compartment, although in reality, the barrel is separated by the blades in a number of domains, which
are not considered at this point:

135

Chapter 5

Figure 5.5. Model domains

1)

The frozen layer domain

The water that is exposed to the freezer wall crystallizes instantaneously and forms a frozen
layer (FL) of immobile (solid) ice. In this layer, the ice crystallizes only by primary nucleation
due to fast freezing. The ice is subsequently removed from this layer by the scraper blades. The
rate of ice formation is determined by the cooling rate and the heat transfer. The rate of change
for the frozen layer is determined by the movement of the scraper blades. In reality, all the
components are present in this layer. However, for the purpose of this research, we will assume
that only water and ice are predominant in this layer.

2)

The bulk domain

This domain is represented by the moving fluid ice cream. The ice scraped from the frozen layer
as a size distributed population of small crystalline particles - is transported into this domain by
the movement of the scraper blades. In the same time, a part of the ice coming from the frozen
layer is melting into the bulk (B). The ice can crystallize in this compartment by secondary
nucleation. However, this is not considered in our research, because the bulk fluid is assumed too
warm for that purpose. The bulk consists of ice particles, air bubbles and fat globules dispersed
in a pseudo-continuous and pseudo-homogeneous fluid matrix of unfrozen water, sugars,
proteins, etc. In order not to make the model much too complex, the following phases will be
considered as pseudo-components of the fluid matrix in the next chapters: water, sugar, fat
and other components.

136

Development of a reduced model for ice cream freezing

Table 5.2. Thermodynamic phases considered in the model
Phases

Domains
Process side
Bulk

Frozen

layer

layer

Ice

Air

Matrix

Equipment related
Wall

Coolant

Rotor

Solid metal

Boiling liquid

= dispersed phase; = continuous phase

The model will consider two types of phases: the phases at the process side and the equipment
related phases, which are shown in Table 5.2. Each phase contains at least one component. The
phases at process side are: (a) the ice particles, as a size distributed population, (b) the fluid matrix,
and (c) the air phase in the bulk domain and (c) the solid layer of ice in the frozen layer domain.
Although the fat is a separate phase inside the ice cream structure, in the following it will be assumed
as being lumped inside the liquid matrix phase. In addition to these process related phases, some
equipment related phases must be accounted for the energy and momentum balance: the solid phases
(not deformable) the freezer wall and the rotor, and the fluid coolant medium boiling ammonia.

3.2. The generic equation of conservation for the physical resources

A pseudo-continuous approach is taken in which the entire space between the barrel wall and
the rotor space is assigned to the bulk fluid and the frozen layer. The rotating scraper blades are
ignored in the modelling. The effects of these rotating blades are accounted for the equations of
change for mass, energy and momentum. A correction for the effective volume available to the fluid
phases must be made, after subtracting the volume taken by the blades from the void volume.
The coordinates for the domains considered in the model are presented in Table 5.3.
The conservation equations of a physical resource can be written in differential form, in the
following general way (adapted from Hangos & Cameron, 2001):
R
R
R
R
J q
t

(5.1)

The divergence of J is given by equation (7.5) in Appendix 1.

R

137

Chapter 5
Table 5.3. Coordinates considered in the model
Domains
Coordinate

Process side
Bulk layer

Equipment related

Frozen layer

Wall

R wall r Rc

0 R0

0 z L

Axial
Spatial

Rotor

0t

Time

Radial

Coolant

R0 r R FL

R FL r R

0 2

Angular
Internal

R r R wall

NA

i
0 s i smax

Initial conditions and a number of boundary conditions for each coordinate considered must
be specified (Hoffman, 2001) together with equation (5.1).
List of symbols

= specific quantity of resource

= flux of resource

q = terms related to the sources and sinks of property

R

R = terms related to the transfer terms from other phases

L = freezer length, m

R0 = internal radius of the freezer, m

R = external radius of the freezer, m

FL = the thickness of the frozen layer, m

smax
= upper bound for the internal coordinate, m or kg
i

wall = wall thickness, m

3.3. Mass
The conservation equations for the mass are written taking into account the complexity of the
processing. Information related to the phase equilibrium conditions, equations of state, and the rate
equations are considered. Simplifications are introduced in a judicious way, and the variables and the
equations are written on a rigorous coordinate set.
The following assumptions are considered when writing the conservation equations for the
mass:

138

Development of a reduced model for ice cream freezing

1. In the bulk domain, only two phases are considered: the ice (particles) phase and an aggregated
liquid matrix phase. This liquid matrix phase is assumed pseudo-continuous. The mixture of the
matrix phase, the solid ice particles and the air phase is known as ice cream.
2. The ice phase is considered as a size distributed population of particles, dispersed in the
matrix phase. The state of a particle is characterized by internal state variables. Only one internal
state variable will be considered: the diameter or equivalently, the mass of a particle. The age and the
temperature of a particle could be considered as the second and the third internal state variable,
respectively, but will not be taken into account (e.g., for the purpose of calculating the residence time
distribution of the ice particles leaving the freezer).
3. The aerated liquid phase is an aggregated one, containing the continuous liquid matrix and
dispersed air bubbles.
4. The air phase is normally dispersed as a population of bubbles of different sizes inside this
liquid phase in a real physical system. In this model, the air phase will represented as a separate,
lumped phase, being treated as a pseudo-component inside the liquid matrix.
5. The liquid matrix contains water and some dissoluted organic components, such as fat, sugar,
proteins, etc. Only the following components are considered in the model: water, fat, and
sugar. The rest of the components, depending on the ice cream recipe, will be lumped into one
single component known as other. In total, the model assumes a number of five (pseudo)components.
6. The frozen layer domain contains only solid ice. The ice in the frozen layer it is assumed
immobile in axial and angular direction. The thickness of the frozen layer in radial direction varies
periodically due to growth by freezing in between two successive scrapings.
7. The temperature of the frozen layer is considered lower than the temperature inside the bulk
domain at the same axial position.
8. Ice crystallization does not occur inside the bulk domain. Only the melting of particles is
considered. No nucleation, attrition or agglomeration is considered.
9. There are some mass-related transfers between the frozen layer and the bulk domain. The
removal of the ice from the frozen layer and its transfer into the bulk is due to the scraping action.
Water from the matrix phase and dispersed ice particles in the bulk domain are (radially)
transferred to the frozen layer.
10. Though the coolant exerts its action by (partial) evaporation, the total flow remains constant
along the axial direction and the mass balance is not of interest.
In the previous assumptions, different phases were presented, which will also be used in the
development of the models equations. An overview of their composition is presented in Table 5.4.

139

Chapter 5
3.3.1. The bulk domain
The mass conservation equations for the bulk domain are written for the dispersed phases (ice
particles), as well as for the pseudo-continuous phases (liquid matrix, air).
Table 5.4. Nomenclature and composition of the phases considered in the model
Phases

Components
Ice

Water

Air

Sugar

Fat

Other

Ice
Matrix
Air
Aerated liquid
Ice cream
Frozen layer

= dispersed phase; = continuous phase

(a) The conservation equations for the dispersed phases for the bulk domain are written in the form
(Ramkrishna, 2000) of the population balance equation (7.9), in Appendix 2.
(b) The conservation equations for the mass of the pseudo-continuous phases in the bulk domain are
written in the form of the mass balance equation (7.16), in Appendix 2.

3.3.2. The frozen layer domain

The starting point for the derivation of the mass conservation equations for the frozen layer is
the mass balance equation (7.21), in Appendix 2. Since the frozen layer consists practically of a solid
layer of pure ice, the concentration of ice inside this layer is constant.
The only thing changing during

FL

the freezing is the position of the

a

boundary of the frozen layer, through

the layers thickness. The thickness of

t

the frozen layer modifies periodically

ts

through two actions, shown in Figure

Figure 5.6. Periodical modification of the thickness of the frozen

5.6:

layer

(a) the crystallization of water on the freezers wall during the time between two successive scrapes,
ts

(b) the removal of the thin layer of ice obtained by the movement of the scraper blades at ts

140

Development of a reduced model for ice cream freezing

3.3.3. The equipment related domains and phases
The mass conservation equations for the equipment related domains phases (rotor, wall and
coolant) are not required, since no transformation of this resource is considered by the model in these
phases.

3.4. Momentum
In a similar way to the mass conservation equations, the momentum conservation is written on
the same rigorous coordinate set. It is assumed that the momentum of the phases inside the ice cream
mixture are added up and lumped into the momentum of the ice cream mixture as a whole. The
momentum equations are written in the compact form of a substantial derivative balance equation for
the 3-D vectorial momentum, shown in equation (7.26), in Appendix 3.
The consequence of this lumping is that information on the momentum of the individual
phases is lost. To remedy for this loss the following assumption is made:
11. The ice and aerated liquid phases in the bulk layer are assumed to have locally equal
velocities for each of the directions (axial, radial, and angular).

3.5. Energy
Some additional assumptions need to be taken into account when writing the energy
conservation equations:
12. The aerated liquid and the ice phases are in thermodynamic equilibrium for the given
composition of the matrix phase at an equilibrium temperature. Below this temperature the ice
particles form. Above this temperature they melt.
13. The liquid ice equilibrium temperature is lower than the aerated liquid temperature due to a
melting point depression due to the sugars in the liquid.
14. Due to external cooling, the frozen ice layer has internally lower temperature than the aerated
liquid in the bulk domain.
15. After breaking away from the frozen layer, the ice particles in the bulk domain heat up rapidly to
the temperature of the particles existing in the bulk domain. This assumption of rapid heating of the
ice particles is justified by the relatively high thermal conductivity inside the ice crystal, relative to
the heat conductivity in the bulk fluid and the heat transfer rate to the crystals.
16. The temperature of the inner part of the ice particle is uniform.
17. The ice particles floating in the bulk will all have the same temperature, which is lower than the
aerated liquid phase temperature. This temperature is taken equal to the ice-water-sugar phase
equilibrium.
18. The heat transfer from the aerated liquid phase to ice particles is used for the supply of energy
for a phase change between solid (ice) and liquid water.

141

Chapter 5
19. The temperature at the interface between the frozen layer and the bulk will be equal to the
equilibrium temperature of the ice in the bulk domain. From there, a (small) temperature gradient
over the frozen layer exists towards a lower temperature at the interface with the inner side of the
metal wall due to cooling by the coolant at the outer side of the freezer wall.
20. The metal wall will have inside lower temperatures than the frozen layer, while they have equal
temperature at their interface.
21. Fast dynamic storage effects in the metal wall will be ignored by virtue of high thermal
conductivity, in such way that a thermal pseudo-steady state flux is maintained in radial direction.
22. At the interface with the coolant, the metal wall takes the (lower) temperature of the coolant.
23. The coolant has a constant temperature all over the entire external surface of the metal wall,
which is assumed to be maintained by phase equilibrium between liquid and vapour phase of the
coolant.
24. A pseudo-steady state assumption is applied to the energy balance of the rotor. Dynamic storage
effects are ignored. The rotor has internally a uniform temperature in radial and angular directions due
to a high conductivity. The rotor temperature can vary in axial direction.
25. The axial rotor temperature is set locally equal to the temperature of the aerated liquid phase at
the interface, with no heat transfer between rotor and fluid. This latter condition is physically a
simplification, but thought to have a minor effect to the amount of work done by the rotor on the
fluid.
26. The energy content is expressed by means of enthalpy, accounting for volume work.
Under these assumptions of non-equilibrium between some of the domains and phases, five
energy balances will be required. These energy balances are coupled by heat transfer terms,
expressing the heat content of mass being transferred between phases as well as the direct conductive
thermal heat transfer. The energy balances will be written for:
(a) The aerated liquid phase (assumed to have the same temperature).
(b) The (melting of) ice particles in the fluid bulk
(c) The frozen layer, applying the average thickness over the periodic scraper rotation
(d) The metal wall (pseudo steady state assumption leading to a flux expression)
(e) The coolant (to get the cumulative cooling duty)
The coupling between these balances is done in the following way:
a)

0/1: Work done by the rotor/scraper is absorbed by the fluid in the bulk

b)

1/2: Heat flux by heat transfer from the fluid phase to the colder floating ice particles and the

energy flux of melt water flow from the melting ice particles into the fluid bulk
c)

2/3: Heat flux from the relatively warm bulk fluid to the frozen layer and, in opposite

direction, the energy flux of the scraped ice particles from frozen layer to the bulk
d)

3/4: Heat transfer flux from the frozen layer to the metal wall

142

Development of a reduced model for ice cream freezing

e)

4/5: Heat transfer flux from the metal wall to the axially flowing coolant
The driving forces considered for the heat transfer are presented in Table 5.5. The energy

transfers between the frozen layer and the floating ice particles and from the floating ice particles to
the fluid in the bulk are both coupled with mass transfer processes.
The scraping rate and the

Table 5.5. Type of transfers between the phases

Conduction

Convection

Phase 2

Rotor

(averaged) flux of ice particles

Ice

from the frozen layer into the

Air

FL

frozen layer extracts heat by

FL

Ice

freezing water into ice at the

FL

Wall

interface of the frozen layer and

Wall

Coolant

(equilibrium)

Work

cooling rate co-determine the

Phase 1

bulk. The heat flux through the

the

bulk

between

two

successive scrapings.
The heat transfer from the bulk fluid to the ice particles determines the rate of melting of the
ice particles and the flux of cold water into the bulk.
The energy balances are written in the following sections in terms of enthalpy. Relations to
link the enthalpy and the temperature (typical quantity used in practical applications) are described in
Section 5 of this chapter.

3.5.1. The bulk domain

The matrix phase and the air phase are assumed to be a pseudo-continuous phase, the aerated
liquid phase.
The conservation equation for the energy of the aerated liquid bulk domain is written in the
form of the energy balance equation (7.34), in Appendix 4.

3.5.2. The ice particles in the bulk domain

It should be mentioned the fact that there are two types of ice particles in the bulk: the fresh
particles coming from the frozen layer by scraping, and the ice particles already present in the bulk. It
is assumed that the heating of the ice particles coming from the frozen layer inside the bulk domain is
very rapid. Once in bulk, the ice particles temperature starts to change due to the difference in
temperature between the particles and the liquid bulk, and eventually melt. Since the ice particles
have different sizes, the heating and the melting will take place with different rates depending on the
sizes and internal temperature of the particles. This means that the temperature should be considered
as a new internal state of an ice particle. When modelling the ice particle population as a whole a new

143

Chapter 5
internal coordinate should be considered while writing the population balance equations for the ice
particles: the thermal coordinate, .
i

The introduction of one extra coordinate will increase the dimensionality of the
i

population balance equations to 6-D. This will increase the model complexity beyond the possibility
to validate experimentally such an extension. For this reason, the inner temperature of a particle will
not be considered as an additional inner state variable.
In the following, it is assumed that all ice particles have the same temperature while they are
melting. The temperature has a value between the temperature of the frozen layer, which is the
temperature of the fresh particles scraped from the wall, and the temperature of the ice-matrix
equilibrium. The ice particles do not warm up inside.
Moreover, we assume that the ice particles in the bulk are one pseudo-continuous phase. The
mass conservation equation is derived in this case from the population balance for the ice particles, by
integrating over all the particle sizes. The energy conservation equation can be written in a similar
way with equation (7.43) in Appendix 4.

3.5.3. The rotor

The energy conservation equation for the rotor is not required since it is assumed no heat
transfer interaction with the aerated liquid phase. The mechanical work term has been transferred to
the energy balance of the liquid phase.

3.5.4. The frozen layer domain

The energy conservation equation for the frozen layer is written in the form of the energy
balance equation (7.48) in Appendix 4.

3.5.5. The freezers wall

The energy conservation equation for the freezers wall is written in the form of the energy
balance equation (7.53) in Appendix 4.

3.5.6. The coolant

The energy conservation equation for the coolant is written in the form of the energy balance
equation (7.57) in Appendix 4.

3.6. Overview of the reductions in the 5-D model

The following main reductions have been performed with respect to the mass balances in the
5-D model:

144

Development of a reduced model for ice cream freezing

1)

lumping of distributed phase: air phase, normally a population of bubbles, is treated as a

pseudo-continuous phase
2)

lumping of phases: air with the liquid phase; fat with the liquid phase

3)

lumping of components in one pseudo-component: all the other components of the ice cream

mix, such as proteins, stabilizers, emulsifiers, etc. in others; different types of sweeteners present in
the mix, in sugar
4)

lumping of components in the liquid phase into water, fat, sugar, air, other

5)

single component (ice) in the frozen layer

6)

no crystallization phenomena considered in the ice cream fluid, except for scraping and melting
With respect to the momentum balance:

7)

lumped momentum balance over liquid and ice phases in the bulk domain
The reductions with respect to the energy balance in the 5-D model consist of:

8)

eliminating the rotor energy balance

9)

ignoring the internal energy distribution in the ice particles by assuming uniform temperature

10) applying steady state assumption to the energy balance of the metal wall
11) assuming constant temperature for the coolant
While even this 5-D model of the ice cream freezing unit is the result of some significant
reductions, it is computationally too demanding for routine use in inverse problems. Therefore,
further reductions will take place, primarily aiming at reducing the number of coordinates of the
model, going from a 5-D to a 3-D model.

4. Specification of the 3-D model of the ice cream freezing unit

In the previous section, the conservation equations are written on a rigorous coordinate set and
are able to accommodate rate terms for any fundamental phenomena. However, having five
coordinates is not practical for solving the model in a repetitive mode in inverse problems. For this
reason, the set of coordinates will be reduced. This is done by:

Eliminating the angular and radial coordinates. The reason for eliminating these coordinates is
that it offers major gains in model reduction, while fair approximations can be made for the
flow and mixing events in the domains of these coordinates. Radial and angular flows and
mixing are very fast due to the rapid rotation of the scraping blades and thought to be
instantaneous

Integrating out of the derivative terms with respect to each eliminated variable over its
respective domains

Averaging the variables over the domains of the eliminated coordinates. The averaging
procedure is detailed in Appendix 5.

145

Chapter 5
The resulting reduced model will have only three coordinates: the time t , the axial position z ,
and the internal coordinate of the particles s . For this reason, this reduced model will be called the
i

3-D model.
The chosen elimination of the radial and angular coordinates induces the following additional
modelling assumptions. The counting of the assumptions is continued from Section 3.
27. Cylinder symmetry is imposed on the freezer, covering the external metal wall, the frozen layer
of ice against the inner surface of this wall, and the bulk domain in which the aerated liquid phase is
moving forward.
28. Drastic simplification of hydrodynamics of the ice cream is considered. The rotation time of the
scraper is very short relatively to the time the aerated liquid phase is transported through the freezer
in axial direction. Due to the fast rotation of the scraper, the ice cream will exhibit complex spiralling
movements and mixing in radial and angular directions with slow net transport in axial direction. This
complex hydrodynamics will be simplified, knowing the effective mixing time is small.
Consequently, the ice cream is practically well mixed within a cross-sectional area, perpendicular to
the axial direction. Therefore, all process variables in the bulk domain will be averaged over such a
cross-sectional area, while they will vary with axial position and time.
29. The axial convective velocities of the ice and aerated liquid phases in the bulk are averaged
over the cross-sectional area. The velocities are taken proportional to the rotational speed of the
scraper, which provides for the momentum.
30. Axial dispersion is not considered for the time being, although in practice the rotational
movement by a scraper will induce some dispersion. The mathematical and computational
convenience of this assumption lies in avoiding second order spatial derivatives with two-point
boundary values, demanding iterative approaches to satisfy such boundary conditions.
31. The effects of change in density of the bulk mixture on the axial velocity can be accounted for by
assuming the total ice cream mass flow is considered constant through the freezer. When the density
of the mixture reduces with the increased ice content, the axial velocity must go up.
32. The event of scraping of the ice from the frozen layer by a scraping blade happens almost
instantaneously; it is very short relative to the time interval for growth of the ice layer by freezing
between two successive scrapings.
33. Work done by the rotor/scraper is absorbed by the fluid in the bulk.
34. The temperature and pressure inside the bulk domain are assumed to be radially uniform, but
varying in axial direction.
35. In view of the small temperatures in the freezer (-10 to +10 0C) the temperature dependency of
some of the physical properties (density, specific heat, conductivity) of pure components is neglected.
By assuming that the perfect mixing takes place within the time interval of rotation of a
scraping blade, it is not relevant to consider dynamics of events on a time scale below this rotation

146

Development of a reduced model for ice cream freezing

time within the 3-D model. The rotation time provides for the lower bound on a relevant time
difference.

4.1. Mass
In the following sections, the result of the averaging with respect to the radial and angular
coordinates of the conservation equation in the bulk and in the frozen layer will be presented. The
averaged terms of the conservation equations are based on the lemmas presented in Appendix 5. An
overview of the resulting equations is presented in Appendix 6.

4.1.1. The bulk domain

The starting point of the averaging procedure is the equation of change for either a distributed
or a pseudo-continuous phase. The conservation equation is multiplied with r and integrated from
r R0 to r R FL and 0 to 2 , and the averaging procedures presented in Appendix 5

are applied. The resulting conservation equations apply over the radial (cross-sectional) areas of the
spatial domains (rotor, bulk, frozen layer, wall) in the freezer.
The complete derivation of the equation for the ice phase can be found in Appendix 7. When
assuming the distributed phase as being pseudo-continuous, the terms in the conservation equation of
the dispersed phase are related to their pseudo-continuous equivalents. The complete derivation of
this equation for the liquid phase can be found in Appendix 8.
Unavoidably, due to the presence of many phases, spatial domains, chemical species, and
coordinates, the symbolic notation of the physical variables becomes quite complicated. In the list of
symbols below, a superscript refers to a thermodynamic phase, while a subscript or a string of
subscripts refer to a resource (mass, energy), a spatial domain (B, FL), a chemical species ( j ) and a
spatial coordinate ( z , s ).
In case of a resource being transferred between two domains, the direction of transfer is
described by an arrow. For example B FL refers to a transfer from the bulk to the frozen layer
domain.
The scalar quantities, such as concentrations, do not have a directional index as a subscript, in
contrast to the vectorial ones, like fluxes, where a directional index must be added. The time
dependency of the variables is assumed implicitly and is not expressed as such. The overhead bar
refers to a quantity that has been averaged over a cross-sectional area (radially averaged).

For example j m , B , j , z is a cross-sectional averaged flux ( j ) of mass ( m ) of phase ( ) in

domain ( B ) of species j at position z .
The first term on the left-hand side of the conservation equations accounts for the variation in
time of the volume-specific quantity of resource (concentration of particles for the dispersed phase,

147

Chapter 5
and concentration of the pseudo-continuous phases). The second and third terms account for the flux
of particles (for the dispersed phases) or mass (for the pseudo-continuous phases) in the direction of
the principal coordinates considered.
The terms on the right-hand side of the conservation equations account for the different
sources and sinks and the transfer to and from different domains or phases. For this particular case,
the source term refers to the melting of ice particles in the bulk in the case of the pseudo-continuous
matrix, while the transfer terms refer to the transfer of particles and mass to and from the frozen layer.

a)

The balance equation for the dispersed and distributed phases

The ice particles in the bulk are treated as a dispersed phase, with size as a distributed phase. The air
bubbles could also be treated in this way, but for the purpose of this research, the air is assumed as a
pseudo-component inside the continuous phase.



n B j B , z j B , s

S B
S

B B D
R t , B FL B
B R t , FL B

AB
AB
t
z
s

(5.2)

With the initial condition:

The initial distribution of particles is defined by:

n B (0, z , s ) n B ,0 z , s

(5.3)

And the boundary conditions:

1)

In axial direction
~

At the entrance of the freezer, the flow of particles is defined by a given flow function F B , z :
~

F B, z t, s
j B , z AB

z 0

2)

(5.4)

Along the internal coordinate

For the maximum particle size, it is considered the particle flux vanishes (Ramkrishna, 2000):

j B,s

(5.5)

smax

The functions n B ,0 and F B , z must be specified as model inputs by the model user.
The concentration of the particle phase (as a pseudo-continuous phase) is determined using the
following equation:

n B a ds c B

(5.6)

148

Development of a reduced model for ice cream freezing

List of symbols

n B = the radially averaged concentration of dispersed particles in the bulk, # 3

m

j B , z = the radially averaged flux of particles in axial direction z in the bulk, # 2

m s

j B , s = the radially averaged flux of particles along the internal coordinate s in the bulk, # 2
m s

u z = the radially averaged velocity in axial direction, m

s

rs

= the radially averaged growth/melting rate, m

s

B B = the radially averaged birth of particles in the bulk domain term, # 3

m s

D B = the radially averaged death of particles in the bulk domain term, # 3

m s

R t , FL B = the radially averaged transfer term (flux) from the frozen layer to the bulk, # 2
m s

R t , B FL = the radially averaged transfer term (flux) from the bulk to the frozen layer, # 2
m s
~

F B , z = flow of particles with size s in phase entering the bulk at z 0 , #

s

n B ,0 = nucleation rate of particles in phase in the bulk, # 3

m
AB = cross-sectional area of the bulk, m 2
S B = circumference of the bulk area, m

b)

The balance equation for the pseudo-continuous phases

The liquid matrix and the air phase are treated as a pseudo-continuous phase.


B c B , j j m, B , j , z


SB

G R
t , B FL , j
B, j
B
B
t
z
AB

(5.7)

With the initial condition:

The concentration of the pseudo-continuous phase is defined by:

c B , j 0, z c B , j ,0 z

(5.8)

149

Chapter 5
And the boundary condition:
~

At the entrance of the freezer, the flow is defined by a given entrance flow function F B , z , j ,0 :
~

F B , z , j ,0 t
j m , B , j , z AB

z 0

(5.9)

List of symbols

c B , j = the radially averaged concentration of component j in the bulk, kg 3

m

B = the radially averaged volume fraction of phase in the bulk,

j m , B , j , z = the radially averaged mass flux of component j in axial direction in the bulk, kg 2
m s

R t , B FL , j = the radially averaged transfer term (flux) of component j from the bulk to the frozen
layer, kg 2
m s

G B , j = the radially averaged rate of change of component j in the bulk, kg 2

m s
~

F B , z , j ,0 = mass flow of component j entering the bulk at location z 0 , kg

s

4.1.2. The frozen layer domain

The starting point is the equation of change for mass of the ice phase in the frozen layer. The
frozen layer contains only ice and this ice is treated as a continuous phase. Any change in the mass of
ice per unit area of metal wall is expressed in a change of thickness of the ice layer. It is physically
the only degree of freedom because the ice density is more or less a constant (a weak function of
temperature).
The frozen layer is subject to a periodic scraping operation, as shown in Figure 5.6. In between
two successive scrapings, the growth of the layer by freezing occurs. Due to the thickness of the
scraper blades relative to the circumferential length along the freezing wall, the duration of the
scraping action at any radial position is very short relative to the time interval between two scrapings.
The time interval between two successive scrapings is divided between the freezing and the
scraping intervals.
The assumption that the frozen layer consists only of solid ice leads to the concentration of the
ice particles inside the frozen layer being equal to the density of ice:
i

c FL t , z
i

(5.10)

150

Development of a reduced model for ice cream freezing

After the averaging procedure in Appendix 6 is applied, the following equation is obtained for
the variation of the ice layer thickness during the freezing interval:
i

FL t , z R t , m , B FL t , z R t , B FL , w t , z

i
t
FL

(5.11)

The right hand side terms represent the ice and water fluxes to the frozen layer. The derivation
of this equation can be found in Appendix 9.
With the initial and final conditions:
At the beginning of a freezing interval, the thickness of the layer is at its minimum, because a
scraping has just been completed:
At

t 0

FL FL , min z

(5.12)

At the end of a freezing interval, the thickness of the frozen layer is at its maximum:
At

t ts

FL FL , max z

(5.13)

The moment of the scrapping is known, co-determining the maximum thickness that can be achieved.
The thickness of the frozen layer is an unknown variable.
List of symbols

FL , min = the minimum thickness of the frozen layer, m

FL , max = the maximum thickness of the frozen layer, m
i

R t , m , B FL = transfer term of ice (mass) from the bulk to the frozen layer, kg 2
m s
l

R t , B FL , w = transfer term of water from the bulk to the frozen layer, kg 2

m s
i

A rigorous equation for calculating the fluxes R t , m , B FL and R t , B FL , w will be presented in

Section 5 of this chapter. Considering the short time intervals between two successive scrapings and
the separation of the freezing event and the scraping action, it is fair to assume that the respective
fluxes are constant over such short time intervals. Hence, the corresponding dynamic models per subinterval can be solved analytically.

151

Chapter 5
(A) The dynamic model for the freezing of the frozen layer
Equation shows how the thickness of the frozen layer grows due to the incoming fluxes of
water and ice from the bulk. This growth happens over a very short time interval 0, ts . The value of

the time ts is inversely proportional to the frequency of scraping, r :

1

ts

(5.14)

r*

Where this frequency is proportional to the rotational frequency and number of blades:

r* r Nblades

(5.15)

Invoking that for the time interval 0, ts the fluxes of water and ice coming from the frozen
layer are constant, equation (5.11) is rewritten:
i

d FL t , z

R t , m, B FL z R t , B FL , w z
i
FL

dt

(5.16)

Equation (5.16) is integrated analytically to obtain an expression for the maximum thickness
of the frozen layer:
FL , max

d FL t , z

FL ,min

ts

R t , m, B FL z R t , B FL , w z
i
FL

t0

dt

(5.17)

Hence, the maximum thickness of the frozen layer at the end of the freezing sub-interval is:
i

FL , max FL , min

R t , m, B FL z R t , B FL , w z
i
FL

t s t0

(5.18)

(B) The dynamic model for the scraping of the frozen layer
At time ts , the actual scraping of the frozen layer occurs. This process happens almost
instantly, over an amount of time much shorter than the growth time interval. The thickness of the
frozen layer is decreased from the maximum thickness up to a minimum thickness, FL , min .
For this case, the thickness of the frozen layer varies in the following way:
FL t , z
t

R t , m , FL B z

(5.19)

i
FL

152

Development of a reduced model for ice cream freezing

At the beginning of the scraping interval, the thickness of the layer is at its minimum, because
the layer had the time to grow during the freezing interval. At the end of the scraping interval, the
thickness is at its minimum, because a scraping has just been completed.
Using the initial thickness just before scraping and the final one, after the scraping, the
integration leads to:
FL ,min

FL ,max

d FL t , z

i
FL

t0

dt

(5.20)

t s t0

(5.21)

Rt, m , FL B z
i

FL , max FL , min

Rt, m , FL B z

ts

i
FL

R t , m , FL B = transfer term of ice (mass) from the frozen layer to the bulk, kg 2
m s
This rate of scraping of ice is an unknown flux. In principle, it could be derived from an
analysis of the effects of the mechanical forces breaking up the ice layer into fragments. However,
this is a very complicated matter. There is an easier way of deriving this rate by making use of the
strictly periodic nature of the scraping. The net effect of the growth by freezing is equal to the net
effect of the scraping, returning the thickness of the ice layer back to its initial thickness. An
expression for the transfer term of ice from the frozen layer to the bulk as the sum of the transfer
terms for the water and ice coming from the bulk to the frozen layer by equating (5.18) and (5.21):
i
l
i
t t
R t , m , FL B t , z R t , m , B FL z R t , B FL , w z s 0

t s t0

(5.22)

The consequence of this equation is that expressions for these freezing rates must be derived
from thermodynamic and kinetic considerations and will be presented in Section 5.
4.2. Momentum

A momentum flux of a mass stream with the velocity u , mass density and pressure P is
given by P u 2 . Due to friction losses, the momentum flux will decrease, leading to a relation
between pressure and the velocity changes. As the equipment is often designed and operated to keep
the mass flux constant, the velocity is considered virtually constant and the pressure change is the key
variable of interest in equipment engineering models. The reason for this is that mechanical work is
required to overcome this pressure change in attaining a target mass flux. The objective of this subsection is to deliver a reduced momentum balance that establishes a relation between the mass flux
and the pressure drop.

153

Chapter 5
Due to the complex rotational movements in axial direction and the non-Newtonian flow
behaviour of the ice cream, a major shortcut is taken in modelling the momentum conservation:
radially symmetric axial convective flow of the bulk is driven by a pressure drop over the equipment.
In (Bongers, 2006) the flow is expressed as a function of the pressure drop, P using the
theory of (Frederickson & Bird, 1958). This concerns the analytical solution of the equation of motion
for the steady state axial flow of and incompressible, non-Newtonian fluid in a long cylindrical
annulus.
The following assumptions are considered in this derivation by (Frederickson & Bird, 1958):
36. It is assumed that the flow is isothermal. This assumption would imply that the viscous
dissipation term in the energy balance equation is negligible or compensated by heat removal.
37. The fluid is incompressible ( const. ).
38. Steady state flow is assumed.
39. The flow is laminar.
40. The freezer is sufficiently long to neglect the end effects.
41. It is assumed that the local shear stress depends on the local shear rate by a power law model.
Under the assumptions of incompressible fluid, steady state laminar flow, the z -component of
the equation of motion in cylindrical coordinates will be:
1 r B , rz
P
i l a

gz
B
z
r
r

(5.23)

This equation is valid over the entire annular region for any kind of fluid.
If

P
is assumed constant:
z

P P0 PL

(5.24)

Equation (5.23) becomes:

1 d r B , rz P
i l a

B
gz
r
dr
L

(5.25)

For the power law model, the local shear stress depends on the local shear rate by:

B , rz K Bi l a

u z
r

n 1

u z
r

(5.26)

154

Development of a reduced model for ice cream freezing

The volume rate of flow is determined by integrating the velocity distribution over the annular
domain:
1

v,iB l a 2 R 2 u z d

(5.27)

The resulting equation relating mass flow rate and pressure drop is derived from (Frederickson
& Bird, 1958). The derivation is detailed in Appendix 10.
1

i l a

i l a

P
n
R
n R0
R

1
L 2 K i l a 2n 1
R

B
3

2 n 1
n

(5.28)

The equation counts two variables: pressure drop and mass flow. In order to be able to solve
the equation, the mass flow is specified, while the pressure drop is determined using equation (5.28).
List of symbols
P = the pressure drop, Pa
P0 = the static pressure at freezers inlet, Pa
PL = the static pressure at freezers outlet, Pa
R0 = the inner radius of the annulus, m
R = the outer radius of the annulus, m

= the dimensionless radial coordinate,

K B

i l a

= the consistency of the ice cream (power law), Pa s n

n = flow behaviour index of the ice cream (power law),

i l a = mass flow, kg s

i l a = density, kg

m3

u z = average velocity, m
s

B , rz = local shear stress, Pa

4.3. Energy
Reduced energy balances will be presented for the bulk, the frozen layer, the freezer wall and
the coolant. The energy balance for the rotor is ignored on the grounds of assumptions 25 and 26,
assuming the rotor is at any axial position in local thermal equilibrium with the liquid in the bulk

155

Chapter 5
domain (having the same temperature). The averaging of the energy equations is performed similarly
to the averaging of the mass conservation equations: multiply the conservation equation with r and
integrate from r R0 to r R FL and 0 to 2 , and then apply the averaging procedures
presented in Appendix 5.

4.3.1. The bulk domain

In case of the bulk, the energy conservation equations are written for the aerated liquid
phase and for the ice particles in the bulk. As the two phases have different temperature with heat
transfer from the liquid to the ice particles, two energy balances will be developed.
a)

Single ice particle

The particles melting rate, required for the ice particles population balance, is obtained from
the assumption that the melting rate and the conductive heat transfer rate are balanced around a single
ice particle, as shown in Figure 5.7.
The incoming heat is used for melting the ice into water and for heating up the inner part of a
particle, rising its inner temperature in this way. This heat balance can be turned into a rate of change
of the ice particles coordinate by means of some model reductions.

qi, T

TB

l a

TB
i

q i, melt
Figure 5.7. Representation of the energy transfer terms around an ice particle

The internal diffusion of heat from the boundaries to the inner part by thermal conductivity is
much higher than the transfer rate of heat from the bulk fluid to the crystal. All heat coming into the
crystal is first used for heating the crystal to the melting point and then used for the melting of the
surface layer. Only a small fraction of the total incoming heat that propagates inwards is needed for
heating up the inner core of a crystal to the equilibrium temperature as compared to the required heat
of melting. The time it takes to heat up a crystal to its melting point is also very short relative to the
melting time itself.

156

Development of a reduced model for ice cream freezing

The melting rate of the ice particles is determined by the available heat flow coming from the
bulk. It is assumed that the dissolution process of water molecules is almost instantaneous (not
limiting) once the heat activation is present.

qi, melt qi, T

(5.29)

List of symbols

qi, melt = heat flow due to melting of ice particles, W

qi, T = heat flow towards the ice particle, due to temperature difference between the ice particle and
the liquid bulk, W
b)

The ice phase

The consequence of the thermal modelling assumption 17 is that all the ice particles, fresh and
old, have all the same temperature. By summation (integration) over the energy content of all
particles, the energy balance over the ice phase in the bulk domain is written in the following form:
i i i i
B c B h B TB

i
i i
e B , z

ln B c B
i
i

DP

E B FL q T , B i
i
Dt
z
ln TB

(5.30)

An expression for determined the energy flux in axial direction, e B , z is defined in Section 5.2.3 of
this chapter.
With the initial condition:
The initial enthalpy profile for the ice phase is defined considering that the initial temperature of the
mix is known:
TB

TB ,0 ( z )

t 0

(5.31a)

This is equivalent to:

i

h B TB
i

h B ,0 ( z )

(5.28 b)

t 0

And the boundary condition:

At the entrance of the freezer, the temperature of the mix is known:
TB
i

TB , z 0 (t )
i

z 0

(5.32)

157

Chapter 5
The relation used to determine the enthalpy from the temperature is defined in Section 5.1.2 of this
chapter.
The mass flux related heat transfer effects are determined from the following equation:
i

E B FL

SB
AB

i
i
i
i
R t , m, B FL h B R t , m , FL B h FL

(5.33)

The heat of melting is incorporated in the definition of the enthalpy, as shown in Section 5.1.2
i

of this chapter. The melting heat of the ice particles and the thermal heat transfer term q T , B i in
equation (5.30) are much more important than the heat effects associated with the mass transfer
i

between the ice phase in the bulk and the frozen layer, E B FL . Rate laws for these terms will be
presented in Section 5.5.2 of this chapter.
c)

The liquid phase

l a l a l a l a
l a
TB
B c B h B
e B , z l
l
i

E B FL , w S B q
T , B FL q T , B i
t
z
AB

l a l a

ln B c B

DP

Q v , B Q scrap , B
l a
Dt
ln TB

(5.34)

With the initial condition:

The initial temperature profile for the liquid is known:
TB

l a

TB ,0 ( z )
l a

t 0

(5.35a)

This is equivalent to:

l a

hB

l a

T
l a

h B ,0 ( z )

(5.32 b)

t 0

And the boundary conditions:

At the entrance of the freezer, the temperature profile is known:
TB

l a

TB , z 0 (t )
la

z 0

(5.36)

The enthalpy profile is determined using the equation defined in Section 5.1.2 of this chapter.

158

Development of a reduced model for ice cream freezing

In this case, the heat transfer rate related to mass transfer effects is determined from the following
equation:
l

E B FL , w

SB
AB

l
l
R t , B FL , w h B , w

(5.37)

Equations for the transfer terms that appear in energy balances (5.30) and (5.34), as well as
expressions to relate the enthalpy with the temperature must be added to the model. These are
presented in Section 5 of this chapter.
List of symbols
i

h B = the enthalpy of the ice phase in the bulk, J

kg
TB = the temperature of the ice phase in the bulk, K
i

e B , z = the energy flux in axial direction, W 2

m
i

E B FL = the heat effects associated to the mass transfer between the ice phase in the bulk and the ice

phase in the frozen layer, W 3

m
i

R t , m , FL B = the radially averaged transfer term (flux) from the frozen layer to the bulk, kg 2
m s
i

R t , m , B FL = the radially averaged transfer term (flux) from the bulk to the frozen layer, kg 2
m s
i

h FL = the enthalpy of the ice phase in the frozen layer, J

kg

i
qmelt
= the heat generation term due to the melting of the ice particle, W 3
m
i

q T , B i = the conductive heat transfer term due to the difference in temperature between the ice

particle and the equilibrium temperature, W 3

m
i

R m , B = the radially averaged melting term for the ice in the bulk domain kg 3
m s

hcryst = the enthalpy of crystallization, J

kg
i a

hB

= the enthalpy of the liquid phase in the bulk, J

kg

159

Chapter 5
TB

l a

= the temperature of the liquid phase in the bulk, K

E B FL , w = the heat effects associated to the heat transfer between the liquid phase in the bulk and the

ice phase in the frozen layer, W 3

m
l

q T ,B FL = the conductive heat transfer term due to the difference in temperature between the liquid

phase in the bulk and the frozen layer, W 3

m

Q scrap , B = averaged work done by scraping the ice layer into the bulk domain, W 3
m

P = radially averaged pressure in the bulk, Pa

The energy balances for the two phases in the bulk domain, represented by equations (5.30)
and (5.34) include several heat exchange rates on which there is an uncertainty on the actual form of
the exchange equations defining them. To eliminate these terms for the solution of the energy
balance, these balances are added to form the overall energy balance per unit volume of the bulk
domain:
i l a
i l a
H B
e B , z
i
l
l

E B FL E B FL , w S B q
W B
T , B FL
AB
t
z

(5.38)

With:
i l a

HB

l a

H B TB H B
i

T
l a

(5.39)

The ice particles in the bulk and the liquid phase have different temperatures, one for each phase.
i

i i

H B B c B h B TB
l a

HB

l a

l a l a

cB

hB

(5.40)

T
la

(5.41)

Having one energy balance for two phases, in order to obtain two temperatures another
condition must be imposed. The temperature of the ice particles is taken to the ice-water-sugar phase
equilibrium temperature in the bulk domain, according to assumption 17:
TB TB , e
i

(5.42)

160

Development of a reduced model for ice cream freezing

The temperature of the liquid phase can be determined in this case from the overall energy
balance over the bulk domain in equation (5.38).
List of symbols
i l a

HB

eB , z

i l a

= the total enthalpy of the phases in the bulk domain, J

kg

= energy flux in axial direction, W 2

m

TB , e = the equilibrium temperature in the bulk domain, K

W B = the work done in the bulk, W 3

m

4.3.2. The frozen layer domain

While the radial coordinate has been eliminated for the bulk region, it cannot be directly
removed in the frozen layer, as radial heat conduction from bulk to the freezer wall through the frozen
layer is essential for heat removal. The objective of this sub-section is to provide an analytical
equation for the heat flux by conduction in axial direction as function of temperature at both
interfaces of the frozen layer for the time-averaged thickness of the ice layer (time average over
growth and scraping intervals). This heat flux expression is obtained by analytically solving a pseudosteady state energy balance in the radial direction.
The averaging procedures presented in Appendix 5, applied for the frozen layer have to take
into account the periodical modification of the frozen layers thickness.
In the following, the frozen layer will be treated as a cylindrical wall, with a thickness equal to
the average thickness of the frozen layer. The terms containing the velocity in the axial direction can
be discarded, since the only relevant contribution is the contribution of the heat conduction in radial
direction.
i i

FL
h FL r q
r , FL

0
t
r

(5.43)

With the initial condition:

The initial temperature profile is known:
TFL
i

TFL ,0 ( z )
i

t 0

(5.44)

The enthalpy profile is determined using the equation defined in Section 5.1.2 of this chapter.
And the boundary conditions:

161

Chapter 5
The cumulative heat flux at the inner boundary with the bulk is defined as the sum of the mass
flux related energy streams and a conductive heat transfer term from the bulk liquid to the frozen
layer:
i

q r , FL

R FL

E B FL E B FL , w q T ,B FL

(5.45)

Assuming a pseudo-steady state energy balance over the frozen layer, the heat flux at the freezer wall
is given by integration of:

r q r , FL

0
r

(5.46)

In addition:

q r , FL
i

dT
dr

(5.47)

With the boundary conditions:

At r R FL

T TFL

(5.48)

At r R

T Twall

(5.49)

When assuming that the thermal conductivity of the frozen layer is constant, the following expression
is obtained:

TFL Twall R FL q r , FL
i

ln R
R FL

R FL

(5.50)

List of symbols
TFL = the temperature of ice in the frozen layer, K
i

Twall = the temperature of the wall at the product side, K

i = the thermal conductivity of the ice, W m K

FL = the thickness of the frozen layer, m
R = the radius of the freezer, m

162

Development of a reduced model for ice cream freezing

4.3.3. The freezer wall
The objective of this sub-section is to provide an analytical equation for the heat conduction in
axial direction as function of temperature at both interfaces of the freezer wall, using a pseudo-steady
state energy balance in radial direction over the wall. The terms containing the velocity in the axial
direction can be discarded, since the only relevant contribution is the contribution of the heat
conduction in radial direction.
metal metal

wall
h wall r q
r , wall

0
t
r

(5.51)

The temperature difference over the freezer wall Twall Twall , coolant is determined from solving
the energy balance over the freezer wall in radial direction:
qr , wall

wall

dTwall
dr

(5.52)

With the boundary conditions:

At r R

Tdomain Twall

(5.53)

At r R wall

Tdomain Twall , coolant

(5.54)

When assuming that the thermal conductivity of the wall is constant:

Twall Twall , coolant R FL q r , FL

ln

R FL

R wall

wall

(5.55)

List of symbols
Twall , coolant = the temperature of the wall at the coolant side, K

wall = the thermal conductivity of the metal wall, W m K

FL = the thickness of the metal wall, m

4.3.4. The coolant

The conservation equation for the coolant is written in the following way:

NH

3 hc Tc Fc hc Tc

q
T , wall coolant
t
z

163

(5.56)

Chapter 5
The enthalpy of the boiling coolant is determined from the enthalpies of the liquid and vapours phases
from the following equation:

hc Tc xc hc

Tc 1 xc hcl Tc

(5.57)

Assuming that the temperature of the coolant is constant and pseudo-steady state, the conservation is
rewritten as:

Fc hc Tc

xc
q T , wall coolant
z

(5.58)

With the boundary condition:

At z 0

xc xc ,0

(5.59)

List of symbols
Tc = temperature of the coolant, K

NH = density of ammonia, kg
3

m3

hc = enthalpy of the coolant, J

kg
Fc = coolant flow rate, kg
s
xc = mass fraction of vapours,
v
l
hc , hc = enthalpy of the vapour and liquid coolant, J
kg

hc = enthalpy of vaporization for the coolant, J

kg

4.4. Overview of the model reduction for the conservation equations in the 3-D
model
A first reduction of conservation equations of the 5-D model was the reduction of the number
of coordinates from five to three by averaging with respect to the radial and angular coordinates for
the mass, population and energy balances, using the averaging procedures defined in Appendix 5 of
this thesis.
The axial velocities of all the phases inside the ice cream mix are equal. The effects of change
in density of the mixture are accounted in the axial velocity.
All process variables are constant over the cross-sectional area, while varying with the axial
position and time. Axial dispersion is not considered.

164

Development of a reduced model for ice cream freezing

The fluxes to and from the frozen layer are assumed to be constant during the time between
two successive scrapings, but varying with axial position and time.
The momentum balance is written in a reduced form, as a relationship between the mass flux
and the pressure drop.
The ice particles in the bulk have the same temperature, equal to the phase equilibrium
temperature of the ice-water-sugar.
The energy balance for the phases in the bulk domain is written for the liquid ice cream,
consisting of matrix, air and ice particles.
When writing the energy balance for the frozen layer, this layer is considered as a solid wall,
with a thickness equal to the average thickness of the layer.
Pseudo-steady state is assumed when writing the energy balance over the frozen layer, freezer
wall and the coolant.

5. Constitutive equations for the 3-D model

The differential equations defining the conservation of resources during the ice cream freezing
process, presented in the previous sections, contain variables within the various terms that have not
yet been specified or related to each other. The number of variables present in the conservation
equations exceeds the number of these equations; hence, some additional equations are required to
close the mathematical model of the process. These are the so-called constitutive equations.
These relations are presented in the order of their relationship to the various terms in the
conservation equations:
(a) Dynamic derivative terms: the hold-up of a resource (mass, energy, ...)
(b) Convective derivative term with a flux
(c) Rates of the various kinetic phenomena, such as ice particle melting rate, mass and heat
transfer effects, scraping rate, work done by the rotor, viscous dissipation etc.
The following constitutive equations will be presented:
1)

thermodynamic equations of state to obtain the density, the specific heat and the enthalpy

2)

the fluxes of convective transport of mass, energy and momentum

3)

the rate laws for conductive phenomena, with associated transport coefficients (diffusion
coefficients, thermal conductivity and viscosity) per phase and the associated transfer
coefficients (heat transfer, momentum) over boundary layers in a fluid phase; the solid phases
have no boundary layer effects

4)

rate laws for transfer of mass, energy and momentum between phases and for sources and
sinks within a phase
The laws defining these terms will be presented in the following sections, together with the

thermodynamic equations of state. An overview of all the constitutive equations is presented in Table

165

Chapter 5
5.6. For the simplicity of notation, for the radially averaged terms in the conservation equations the
dash will be dropped.

5.1. Thermodynamic equations of state

The thermodynamic equations of state will be presented for the mass, energy, transport.

5.1.1. Density of phases

a) Density of the liquid phase

The density of the air and liquid mix (the aerated liquid phase) is determined as a function
of the densities of the components in the following way:
1
mix

domain

x , j
domain
j
j

(5.60)

The density of the components of the matrix is assumed to have a constant value during the
freezing process. The values of the densities are presented in Table 7.2, in Appendix 11.
The density of the air is determined using the ideal gas law, at the initial temperature and
pressure of the ice cream mix:
a

domain

M a Pdomain
R Tdomain

(5.61)

List of symbols
mix
domain
= density of the mix, kg

m3

xdomain
, j = mass fraction of component j ,

j = density of component j , kg

m3

Pdomain = pressure inside the domain, Pa

Tdomain = temperature inside the domain, K

R = specific gas constant, J

kg K

M a = molecular weight of air, kg

kmole

166

Development of a reduced model for ice cream freezing

Table 5.6. Constitutive equations for the ice cream freezing model
Type

Reference

Counter

of

phase

phase

Mass

Momentum

Energy

constant

constant

constant

Air

Eq. (5.61)

Metal

constant

Coolant

constant

Mixing rule

Eq. (5.60)

Ice (particle)

Eq. (5.70) & (5.72)

Liquid

Eq. (5.71)

Air

Eq. (5.71)

constant
constant

Eq. (5.63)
Enthalpy

constant
Specific heat

Ice
Liquid
Density

equation of state

Flux

Thermodynamic

law

Eq. (5.65)

Eq. (5.66)
Eq. (5.73)

Eq. (5.74)

Solid ice (FL)

Metal
constant
constant

constant

Liquid

constant

constant

Coolant

Viscosity

Diffusivity

Air

Conductivity

Ice

Metal

Eq.

Mixing rule

constant
constant
constant
Eq (5.135) & (5.137)

(5.133)

coefficient

Transfer

Conductive transport

Coolant

Ice (particle)

Eq. (5.125)

Liquid

Eq. (5.129)

Eq. (5.139)

Eq. (5.138)

Ice cream
mix

Eq. (5.140)

Solid ice (FL)

Freezer wall

Eq. (5.140)

Coolant

Eq. (5.142)
Aerated

Eq. (5.109)

Eq. (5.116)

Eq. (5.81)

Eq. (5.116)

Eq. (5.90)

Eq. (5.116)

liquid
Ice (particle)

Solid ice

rate

Phase transfer

(FL)
Aerated

Solid ice

liquid

(FL)
Eq. (5.120) & Eq. (5.124)

Rotor
Ice
Solid ice (FL)

(particle)

Metal

Coolant

Eq. (5.95)

Eq. (5.116)

Rotor

167

Chapter 5
5.1.2. Specific heat and enthalpy
The specific heat for the various components and phases is assumed constant. The values for
the components of the ice cream mix are presented in Table 7.2 in Appendix 11.
The enthalpy of a specific phase is determined using an arbitrary zero reference for the
enthalpy. This zero reference enthalpy is determined for a reference temperature equal to 298 K and
liquid phase:
Tref 298

(5.62)

The relationship between the enthalpy and the temperature, when a phase transition is present:
2
1
hdomain
T hdomain
Tref

1 2

c p ,1j T dT h

1 2

Tref

1 2

c p , 2j T dT

(5.63)

T 1 2

For the case of the phase transition between the liquid water and ice, the enthalpy of the phase
change (enthalpy of crystallization) is taken at the normal melting temperature at 1 bar:
T

1 2

273.15

(5.64)

The value of this enthalpy is presented in Table 7.2, in Appendix 11.

For the species for which the phase transition is not present, the enthalpy is determined from
the following equation:
T

hdomain
, j T hdomain , j Tref

c T dT

(5.65)

p, j

Tref

The enthalpy of a mixture is determined using a weight additive model, by adding the
enthalpy contribution of each pure component:

hdomain
xdomain
, j hdomain , j
mix

(5.66)

In equation (5.66), any non-ideal interactions have been ignored.

List of symbols
Tref = the reference temperature, K

cp , j = intrinsic heat capacity of component j in phase , J

kg K

1 ,2 = phases
T

1 2

= temperature of the phase change, K

168

Development of a reduced model for ice cream freezing

h

1 2

= enthalpy of the phase change, J

kg

J
hdomain
, j = enthalpy of a pure component j in phase ,
kg

a) Enthalpy of the dispersed phase

In case of the dispersed phases, the averaged enthalpy is determined considering the

contribution of each particle, hn , domain

, at the temperature of a single particle, Tn, domain
. The use of a

single particles temperature is not practical from a computational point of view. This implies the
introduction of a new coordinate in the population balance for the dispersed phase, as discussed in
Section 3.5.2 of this chapter.
According to assumption 17, the ice particles have all the same temperature. For this reason,
the averaged enthalpy of the dispersed phase will be determined using the enthalpy calculated at the

:
average temperature of the particles, Tdomain

hdomain

hn , domain

Tdomain mmax

n m dm

domain

domain

m

(5.67)

min

List of symbols

hn , domain
= enthalpy of a single ice particle, J
kg

Tn, domain
= temperature of a single ice particle, K

Tdomain
= average temperature of ice particles, K

hdomain
= enthalpy of the ice phase, J
kg phase

ndomain
= the ice particle density, # 3
m

m = particles mass, kg
#

domain
= volume fraction of phase ,

= density of phase , kg

m3

169

Chapter 5
5.1.3. Ice-water-sugar phase equilibrium
The particles melt as long as the following condition is respected:
TB z TB , e

(5.68)

The value of the equilibrium temperature of the bulk is determined using:

TB , e Tm
i

1860 xs , B

(5.69)

M s 1 xs , B

42. The sucrose is chosen to represent the different types of sugars present in the ice cream mix.
From the look of the freezing curve for most of the mix compositions, it can be concluded that
in the temperature range where the SSHE operates, only about one third to one-half of the water is
present in solid state. A good source for calculating initial freezing temperatures and freezing curves
of mixes can be found in (Leighton, 1926), (Bradley & Smith, 1983) and (Bradley, 1986).
List of symbols
TB = the temperature of the bulk phases, K

TB , e = the equilibrium temperature of the liquid phase inside the bulk, K

Tm = the melting point of ice, K
i

M s = molecular weight of sugar, kg

kmole
xs , B = sugar mass fraction in the bulk,

5.2. Fluxes in a spatial direction

A flux involves an intra phase transport of resources.

5.2.1. Mass flux in axial direction

The mass and population balance equations contain convective derivative terms with a flux. In
the following, relations for computing these fluxes are defined.

a)

The population balance equation for the dispersed and distributed phases (ice particles in

the bulk)
The flux of particles along the axial coordinate is calculated using:

170

Development of a reduced model for ice cream freezing

jB , z uB , z nB

b)

(5.70)

The mass balance equation for the pseudo-continuous phases (liquid matrix in the bulk, air)
The mass flux of component along the axial coordinate is calculated from the following

equation:
jm , B , j , z u B , z B cB , j

(5.71)

5.2.2. Mass flux along the internal coordinate

The population balance equation contains a flux along the internal coordinate, which is
determined from the following equation:
jB , s rs, B nB

(5.72)

5.2.3. Energy flux in axial direction

The mass flux in axial direction and the energy flux can be easily related using the following
equation:
eB , z jm , B , z hB

(5.73)

List of symbols
uB , z = velocity in the bulk domain, m
s

rs, B = growth/melting rate of ice particles in the bulk domain, m

s

eB , z = energy flux in axial direction, W 2

m

5.2.4. Energy flux in radial direction

For two of the domains (the frozen layer and the freezer wall), the radial coordinate is still
important from the point of view of the energy balance. These two domains are treated as composite
cylindrical walls (Bird, Stewart & Lightfoot, 2002). The conductive flux in radial direction is
determined using Fouriers law:
qr , domain

dTdomain
dr

(5.74)

The boundary conditions must be defined for both the frozen layer and the freezer wall:
a) The frozen layer
At r R FL

Tdomain TFL( i )

(5.75)

171

Chapter 5
At r R

Tdomain Twall

(5.76)

At r R

Tdomain Twall

(5.77)

At r R wall

Tdomain Twall , coolant

(5.78)

b) The freezer wall

List of symbols
qr , domain = radial energy flux in the domain, W 2
m

= heat conductivity of the phase , W m K

TFL( i ) = temperature of the frozen layer, K
Twall = temperature of the wall at the product side, K

Twall , coolant = temperature of the wall at the coolant side, K

5.2.5. Energy flux in radial direction over the frozen layer and the metal wall
An equation for the temperature drop TB Tcoolant is the outcome of solving the energy balance
i

over the frozen layer and the metal wall in radial direction.
The heat transfer coefficient for these walls is determined assuming that this layer behaves
like a cylindrical solid wall.
The heat flux in radial direction will be determined from the following equation:

q0 U 0 TB Tcoolant
i

(5.79)

The expression for the overall heat transfer coefficient for cylindrical solid walls is presented
in Section 5.4.4 of this chapter.
List of symbols
q0 = heat flux in radial direction, based on the inner surface of the freezer, W 2
m
U 0 = overall heat transfer coefficient based on the inner surface of the freezer, W 2
m K
Tcoolant = temperature of the coolant, K

5.3. Rate laws

A rate law involves a transfer of a physical resource between phases or domains, as well as
point processes, such as sources and sinks within a phase. The knowledge about the mathematical

172

Development of a reduced model for ice cream freezing

structure and the parameters of the rate equations presented in the previous sections is limited.
Rigorous equations can be written for few of these rates. Even if rigorous equations could be written,
the imposed drastic simplification of the fluid flow behaviour requires approximations of the laws. In
the next sections, a pragmatic, linear structure of the rate laws is proposed. This structure must be
consistent with insights from non-equilibrium thermodynamics (NET) and needs to be validated using
experimental data.
The rate laws for melting, mass and heat transfer are all structured in the same fundamental
way: the rate is proportional to a single thermodynamic driving force. This driving force is
mechanically enhanced or diminished by the scraper rotation. The transfer coefficient is built up from
a nominal transfer coefficient and an empirical adjustment parameter (fudge factor). The nominal
transfer rate coefficient represents a theoretical, state-of-the-art estimate. The empirical adjustment
factor is needed for fine-tuning to experimental data.
Rate Crate f r krate Fd

(5.80)

List of symbols
Crate = the empirical adjustment parameter (fudge factor)

f r = the effect of the scraper rotation on the transfer rate

r = rotor speed, rps

krate = the nominal transfer rate coefficient (a state-of-the-art estimate from the literature)
Fd = the thermodynamic driving force, e.g., arising from differences between phases

In the next sections, the specific rate laws are introduced, covering changes in mass, particle
distributions, momentum and energy, in that order. After completing the description of the rate laws
the determination of the nominal transfer rate coefficients is discussed.

5.3.1. Mass transfer and ice melting processes

The mass transfer events involve exchanges of water and ice between two phases and
domains: water in the bulk domain, ice particles in the bulk domain and solid ice in the frozen layer.
The rate laws for the mass transfer terms will be detailed in the following order:

ice particles transported from the bulk domain to the frozen layer

components in the liquid phase of the bulk domain transported to the frozen layer

scraping of ice from the frozen layer into the bulk domain

The different mass transfer terms for ice and water are shown in Figure 5.8. An overview of
all the mass related transfer terms is presented in Table 5.7, at the end of this section.

173

Chapter 5

Figure 5.8. Transfer terms for the: (a) ice and (b) water

a)

Transfer of ice particles from the bulk to the frozen layer

It is possible that a small fraction of the ice particles from the bulk is transferred to the frozen
layer and is integrated into the frozen layer. This process of ice particle transfer will happen only if
there are enough particles in the bulk and a favourable temperature difference between the ice
particles in the bulk and the colder surface of the (growing) frozen layer to which particles may stick
and freeze.
The flux in this case will be proportional to the concentration of ice particles in the bulk
domain:

Rt, B FL t , z , s Ca k B f1 r nB t , z , s
i

(5.81)

This easily measurable force, number of particles, is chosen over the fundamental
thermodynamic driving force, the chemical potential difference between the ice particles and the
frozen layer. An additional assumption is that the ice particles, irrespective of their size, will have an
equal chance of being frozen into the frozen layer. For this reason, the transfer coefficients are
independent of the size s in the rate equation. Hence, the total mass transfer rate is obtained by
i

integration over all sizes and is written as:

Rt, m , B FL t , z Ca k B f1 r B cB t , z
i

(5.82)

Where the concentration of the ice particles in the bulk domain, cB is given by
i

cB nB a ds
i

(5.83)

In addition, the action of the rotating scraper will have an influence on the rate of transfer. The
rotor action is accounted for in the term f1 r . This term has the following qualitative properties:

174

Development of a reduced model for ice cream freezing

When the rotor speed is zero, the spontaneous transfer of ice particles from the bulk to the frozen

layer is extremely unlikely and will be ignored

When the rotor speed is very high, the ice has too high a momentum to stick on the wall

There is a value of the rotor speed, r , max 0, for which the transfer rate of ice from the

frozen layer to the bulk is a maximum

The conditions for maximum are stated as follows:
f1
r

r r ,max

f1
r

r r ,max

(Ascending slope)

(5.84)
0

(Descending slope)

The function meeting the conditions above is pragmatically postulated as a Weibull

distribution-type of function, due to the flexible properties of this distribution:
r

1 e
r
f1 r
0

for r 0
for r 0

(5.85)

and are adjustable parameters. However, they must be chosen such that the function

is positive and that there is a maximum in this function.

(5.86)

It is noted that under these parametric conditions, the Weibull function (5.85) satisfies the
boundary conditions (5.84). The maximum rotational speed is determined from:

r , max 1

for 1

(5.87)

Transfer of component j from the bulk to the frozen layer

b)

Since it is assumed that the frozen layer contains only ice, the transfer of sugar, fat, Air
and other components from the bulk to the frozen layer is not possible. For this reason, for these
components, the rate vanishes:
Rt, B FL , j Cb k B f 2 r B cB , j t , z 0
l

due to the fact that:

(5.88)

k L ,j 0
l

(5.89)

175

Chapter 5
For the transfer of water, the driving force is a difference between its actual local
concentration in the bulk and the phase equilibrium concentration.
Rt, B FL , w t , z Cb k B ,w f 2 r cB ,w t , z cB ,w, e TB , e

(5.90)

The term f 2 r takes into account the rotor action. It is assumed that when there is no rotor
action at all, the transfer rate of water to the frozen layer surface is negligible. For sufficiently high
rotor action, the mass transfer limitation at the bulk side completely vanishes. Therefore, this term has
the following properties:
For

r 0

f 2 r 0

(5.91)

For

f 2 r 1

(5.92)

Several equations can be postulated for the f 2 r term. For the purpose of this research, the
following function will be used:
k
f 2 r r
1 k r

(5.93)

n 1

(5.94)

List of symbols
Cb = the fudge factor,
l
k B ,w = mass transfer coefficient for water, m
s

cB ,w, e = equilibrium concentration of water at the interface between the frozen layer and the bulk,
l

kg
3
m
c)

Transfer of ice from the frozen layer to the bulk by scraping action

For the transfer term of ice particles coming from the frozen layer the following two cases are
considered:
(1) The distribution of the ice particles coming from the frozen layer is physically the result of
a merger of two size distributions: (a) one of the ice crystals coming from the bulk into the frozen
layer, and (b) one of the fresh ice formed by crystallization of the water coming into the frozen layer
from the bulk. In this case, complex phenomena need to be taken into account: within the frozen
layer, the nucleation, growth, agglomeration of the newly formed ice particles resulted from the water
crystallization, as well as growth, agglomeration of the ice particles that come from the bulk.

176

Development of a reduced model for ice cream freezing

Covering this implies the set-up of a new population balance equation for the ice phase in the frozen
layer. Moreover, the history as reflected by their attained states - of the ice particles must be
considered: the particles coming from the bulk have a different history and states than the ones that
are newly formed. This will complicate the frozen layer model even more.
For this reason, a second, simpler approach will be considered:
(2) Any ice formed by crystallization of either the water or the ice coming from the bulk forms
a compact uniform layer. Upon scraping, where the movements of the blades crack the solid ice layer,
a completely new distribution of ice crystal sizes is formed. This implies that any information on the
origin of the ice in the frozen layer is erased. Moreover, local processes such as nucleation, growth
and agglomeration of ice particles in the frozen layer can be neglected. By the scraping of the blades,
a new distribution of ice particles is formed, entering the bulk domain. This approach is captured in
the following assumption:
43. The thickness of the frozen layer and the rotational speed of the scraper determine the particle
size distribution of the ice particles being scraped away. This distribution removed from the frozen
layer is independent of the size distribution of the ice particles inside the bulk. Due to the lack of
detailed fundamental information, a qualitatively suitable empirical distribution function is postulated
for the size of the scraped ice particles coming from the frozen layer.
Knowing a size distribution function one can relate that to the mass transfer rate. When
specifying a mass transfer rate a time-averaged value will be taken over the time interval of the
growth of the frozen layer and the scraping, smoothing the periodic dynamic character. The mass
transfer term of ice from the frozen layer to the bulk (assuming the ice particle population is a pseudocontinuous phase in the size coordinate) is defined as:

Rt, m , FL B t , z Rt, FL
a s ds
B t , z, s
i

(5.95)

When assuming that the frozen layer is in short term stationary state, the sum of the incoming
and outgoing mass fluxes over the scrapping period must vanish.

t t
i
i
l
Rt, m , FL B t , z Rt, m , B FL t , z Rt, B FL , w t , z s 0

t t
s

(5.96)

The (number) flux of ice particles with size s from the frozen layer to the bulk is
i

determined as a fraction of the total mass flux in the following way:

177

Chapter 5

Rt, m , FL B t , z norm
s
i

Rt, FL

B t , z, s

(5.97)

m s
i

m s v s
i

(5.98)

44. The ice particles are assumed to have a spherical shape. Since the real shape of the ice crystal is
very irregular, the use of a shape factor could be considered to calculate the volume of the ice particle,
v .
i

It is necessary to know this number flux, Rt, FL

B t, z, s
i

because it alters the size

s is the normalised size distribution

distribution of the ice particles in the bulk. The function norm
i

of the ice particles removed by the scraping blades. This latter function is unknown and it is very
difficult to get a theoretical prediction on basis of the mechanics of the scraping process. Therefore,
an empirical function will be used:

45. The normalised distribution function norm

s
i

for the size of scraped away ice particles is

chosen to be a probabilistic Weibull-type of distribution, with the scale parameter equal to and the
shape parameter equal to .
The probability density function of the Weibull distribution for the independent coordinate x
is given by:
x s

x s 1 e
i

norm s
0

for x s 0

(5.99)

for x s 0

0
0

(5.100)

This independent coordinate x is related to the size coordinate s . There is a normalisation condition
in the size coordinate:

i
norm

s ds 1

(5.101)

s smax
i

(5.102)

178

Development of a reduced model for ice cream freezing

The relation between x and s must be specified. The independent coordinate x runs

] in
from [0, ) . It is a function of the particle size, s , which may cover a smaller finite range (0, smax
i

. The precise relationship

physical reality. For this reason, x must correspond to s smax
i

between x and s is open to choice and will depend on the actual physical situation.
i

List of symbols
= scale parameter of the Weibull distribution

= shape parameter of the Weibull distribution

x s = the normalised coordinate of the particle size distribution leaving the frozen layer
i
m = the ice particle mass, kg
#
3
i
v = the volume of the ice particle, m
#
2
i
a = the surface area of an ice particle, m
#

Transfer of component j from the frozen layer to the bulk

d)

Only ice is transferred from the frozen layer to the bulk. The value of the transfer term for all
other components j from the frozen layer to the bulk is zero, because of a zero concentration in the
frozen layer:

cFL , j t , z 0

Rt, FL
B , j C f k FL , j f 2 r cFL , j t , z 0
l

(5.103)

(5.104)

Rate of change of ice particle population by generation ( BB ) and removal ( DB )

i

e)

According to assumption 8, there is no crystallization or nucleation of the ice particles inside the bulk
domain. Moreover, due to the same assumption, the death of particles due to agglomeration or
attrition is ignored. For this reason, the rates of change of ice particles by generation and removal in
the bulk are zero:
BB 0

(5.105)

DB 0

(5.106)

179

Chapter 5
f)

Rate of change of ice particles due to melting in the bulk domain

The rate of change of the ice particles size coordinate is expressed by rs . This rate of change
i

determines the flux of melting particles along the size coordinate:

jB , s rs, B nB

(5.107)

The derivative term of this flux with respect to the particles coordinate s is represented in
i

the population balance by a new algebraic variable:

i

jB ,s t , z , s
i

g n ,B

(5.108)

By integrating of this local rate of change over the particle coordinate one can obtain the mass
transfer rate due to the joint effect of all melting ice particles, Gm ,B , which is equal to the mass
i

transfer of liquid water due to melting of ice particles in the bulk, GB ,w .

l

The specification of the internal coordinate of an ice particle will lead to determining the
transfer rate due to the melting of the ice particles:
The mass transfer rate due to the joint effect of all melting ice particles in the bulk is
determined as follows:

i
i

j t , z , s
B,s
l
i
i
i
i
i
GB ,w g n ,B a ds
a
ds

s
0
0

(5.109)

List of symbols
i
g n ,B = number rate of local change of ice due to melting of the ice particles, # 3
m s
i
Gm ,B = mass rate of local change of ice due to melting of ice particles, kg 3
m s
l
GB ,w = mass rate of local change of water due to melting of ice particles, kg 3
m s

The mass transfer rate due to melting can be determined from the sequence: rs , jB ,s , Gn ,B ,
i

GB ,w . A constitutive equation must be specified for the rate of change of the ice particles coordinate,
l

rs .
i

The starting point for the derivation of such a constitutive equation is the heat balance around
an ice particle, shown in Figure 5.7.

180

Development of a reduced model for ice cream freezing

The heat flow towards the ice particle is calculated from the following equation, determined
using the shrinking core model described in Appendix 12:

qi, T 8 i s i TB TBi

(5.110)

The heat flux through the particle surface is determined using:

q T
i

q , T

(5.111)

a
i

The energy supply to all the ice particles the following relation is used:

qmelt
q T nB ds
i

(5.112)

The melting rate for a single ice particle can be written as:

rs t , z , s
i

q T
hcryst
i

(5.113)

Table 5.7. Overview of the rate laws for the mass related events
Rate

Fudge

Rotor action

Nominal transfer

Driving

factor

effect

coefficient

force

Transfer of ice from the FL to the B

f1 r

r*

Transfer of ice from the B to the FL

Ca

f 2 r

k Bi

nBi

Transfer of water from the B to the

Cb

f 3 r

k Bl,w

cBl,w cBl,w,e

Ice melting in the B

Cc

kri

TB TB ,e

Water production by

Cc

kri

TB TB ,e

FL

ice melting in the B

The rate of local change of ice due to melting is determined from the following expression:
i

Gm , B g n , B a
0

j i t , z , s i

B,s
i
ds
a
ds

s
0

(5.114)

The rate of local change of water due to ice melting is:

l

GB , w Gm ,B
i

(5.115)

181

Chapter 5
List of symbols
i
q T = heat flux due to temperature difference between the ice particles and the liquid bulk, W 2
m

i
qmelt
=heat flux due to melting of ice particles in the bulk, W 2
m

hcryst = the enthalpy of crystallization, J

kg
g)

Rate of change of water due to ice melting in the bulk

The mass transfer rate of water due to the ice melting in the bulk domain is determined from
the expression for GB ,w in equation (5.109).
l

5.3.2. Energy transfer processes

Energy is expressed by means of enthalpy in this model. In the following sections, the rate
laws for the energy transfer terms are detailed. The different energy transfer terms are shown in
Figure 5.9.
Transfer due to a transfer of mass, q j

a)

The enthalpy flux transferred between two phases or domains due to a transfer of mass is
determined by its enthalpy content:
q j ,D D t , z Rt, m, D1 D2 hD 1

(5.116)

List of symbols
q j ,D D = the enthalpy flux due to mass transfer of phase between the domains D1 and D2 ,

W
m 2

Rt, m, D1 D2 = the mass flux of phase transferred between the domains D1 and D2 , kg 2
m s

hD = the enthalpy of phase in the domain D , J

kg

Heat term due to melting of the ice particles, qmelt

i

b)

The heat term due to the melting of ice particles in the bulk is determined from the mass
flux of ice from melting and the heat of crystallization in the following way:

qmelt
Gm ,B hcryst
i

(5.117)

182

Development of a reduced model for ice cream freezing

Figure 5.9. Energy transfer terms

It has to be noted that this term does not show up explicitly in the enthalpy balance for the
bulk phase. This is because the phase change effects have been accounted for in the definition of the
enthalpy per phase, sharing a common reference point for the definition of the enthalpy (Section 5.1.2
of this chapter).
c)

Transfer due to temperature difference, qT

The heat flux due to a temperature difference is determined using the heat transfer coefficient
as the nominal transfer coefficient and the temperature difference as driving force:

q T t , z U domain
T

(5.118)

Equation (5.118) is applied for calculating the heat transfer terms presented in the first column of
Table 5.8.
List of symbols

U domain
= total heat transfer coefficient, W 2
m K

U B

l a

= total heat transfer coefficient from the aerated fluid bulk phase to frozen layer, W 2
m K

183

Chapter 5
U B ,i = total heat transfer coefficient between the ice particles and the fluid bulk, W 2
m K
i
U FL
= total heat transfer coefficient between the frozen layer and the freezer wall, W 2
m K

U 0 = total heat transfer coefficient between the bulk and the coolant, W 2
m K
Table 5.8. Transfer terms due to temperature differences
Transfer term

Nominal transfer coefficient

q T

l a
B TFL

UB

Driving force

FL

TB TFLi

q T T

i
U FL

TFLi Twall

qT T
B
B ,e

U B ,i

TB TB ,e

qTwall Tc

U0

Twall Tcoolant

FL

wall

d)

l a

Heat generation due to mechanical work in the bulk

The heat generation due to mechanical work in the bulk WB it is assumed to come from two
main sources (Bongers, 2006): (1) the viscous dissipation, due to the fluid flow around the rotor Qv , B ;
and (2) the scraping friction between the blades and the barrel wall Qscrap , B :
WB t , z Qv , B t , z Qscrap , B t , z

(5.119)

According to (Bongers, 2006), this dissipation of mechanical energy accounts for as much as
50% of the heat removed by the refrigerant.
In the following, rate laws for these two terms will be postulated.
d1) Heat generation due to the viscous generation in the bulk domain
The energy balance over the ice cream mixture in the bulk domain accounts for viscous
dissipation. In (Bongers, 2006) the viscous dissipation term is equal to the product of the shear stress
and the shear rate.

Qv , B t , z Cv B t , z B t , z

(5.120)

The shear stress is calculated using a power law-type equation (Bird, Stewart & Lightfoot,
2002), considering steady-state shear flow:

B , rz Bi l a B

(5.121)

184

Development of a reduced model for ice cream freezing

Where:

u z
r

(5.122)

An averaged shear rate for the barrel domain is used as an approximation:

B C r

2 D
D D0

(5.123)

List of symbols
WB = the heat generation due to mechanical work, per unit volume in the barrel, W 3
m

Qv , B = the viscous dissipation term, W 3

m

B , rz = the shear stress, Pa

Bi l a = the viscosity of the ice cream, Pa s

B = the shear rate, 1 s

Cv , C = fudge factor
D = outer diameter of the freezer, m
D0 = inner diameter of the freezer, m

d2) Heat due to the scraping of the ice layer at the freezer wall
For the scraping friction, the work done by the scraper is a function of the flux of ice being
scraped from the frozen layer at the wall and the energy it takes to brake solid ice into particles. The
following equation will be used to determine this term:
Qscrap , B c1

S B i
i
Rt , m, FL B mech
AB

(5.124)

represents the energy required to break one kilogram

The intrinsic ice breaking energy mech
i

of ice into particles. The factor

SB
is needed to convert from area (of incoming ice flux) to unit
AB

volume in the barrel.

List of symbols
c1 = fudge factor,

185

Chapter 5
i
mech
= intrinsic ice breaking energy, J
kg
i
Rt, m , FL B = flux of ice being scraped from the frozen layer at the wall, kg 2
m s

5.4. Conductivities and transfer coefficients

A warning is in order before discussing the way the conductivities and the transfer coefficients
are determined. By neglecting in the model reduction effort the details of the fluid flow, the price to
pay is a severe sacrifice with respect to the rigour of estimation of the transfer coefficients. The
classical way of estimating transfer coefficients by means of dimensionless numbers between various
forces and the use of physical engineering correlations, has become nearly impossible due to lack of
information on local flow fields and Reynolds numbers. In addition, the multiphase nature of the ice
cream in the bulk domain and its non-Newtonian behaviour are hampering factors in finding good
predictions.
Therefore, the equations for transfer coefficients to be presented in the following must be seen
as tentative ones, representing tendencies. Care has been taken to have the proper physical asymptotes
for extreme conditions.

5.4.1. Mass transfer coefficients

a)

Ice particles from the bulk to the frozen layer

The overall mass transfer coefficient for the ice particles in the bulk is determined from two
resistances to transfer in series: a resistance in the liquid phase side ( 1/ k L ) and one for the
i

integration at the surface of the frozen layer ( 1/ kT ):

i

k B
i

k S kT
i

(5.125)

k S kT

The mass transfer coefficient of the ice particles is taken constant, with the value presented in
Table 7.2, in Appendix 11.
The resistance at the surface gets smaller if the freezing potential is bigger. Therefore, the
transfer coefficient at the side of the frozen layer surface is taken proportional to a freezing potential
expressed as a scaled temperature difference ( T ):
i

kT kT ,0 T

i

(5.126)

Moreover, the freezing potential is expressed as a scaled temperature difference:

186

Development of a reduced model for ice cream freezing

Ti

TFL TB , e
i

(5.127)

TB , e
i

a 1

(5.128)
In view of the principle of considering linear driving forces dependency, it is recommended to

take the exponent a equal to one.

List of symbols
i
k B = overall mass transfer coefficient for the ice particles in the bulk, m
s
i
k S = mass transfer coefficient of the ice particles, m
s
i
kT = transfer coefficient due to temperature difference between the frozen layer and the bulk, m
s

Ti = the freezing potential,

b)

Water from the bulk to the frozen layer

A similar mechanism as for the ice particles is put in place for the transfer of water. The
overall transfer coefficient for the water from the bulk to the frozen layer is determined from two
resistances to transfer in series: a resistance at the liquid phase side ( 1/ k L ,w ) and one at the frozen
l

layer surface ( 1/ kT ,w ). The latter resistance gets smaller if the freezing potential is bigger.
l

Therefore, the transfer coefficient at the side of the frozen layer surface is taken proportional to a
freezing potential expressed as a scaled temperature difference ( T ,w ).
l

k B ,w
l

k L ,w kT ,w
l

(5.129)

k L ,w kT ,w
l

As there is plenty of water at the bulk side, if there is enough rotor action, the mass transfer
coefficient k L ,w becomes very large relative to kT ,w and it cancels out. At the surface of the frozen
l

layer, the water must turn into ice with a certain rate coefficient:
k B ,w kT ,FL , w kT ,FL , w,0 T

l

aw

(5.130)

Similarly, the freezing potential is expressed as a scaled temperature difference:

Ti

TFL TB , e
i

(5.131)

TB , e
i

187

Chapter 5
aw 1

(5.132)
In view of the principle to consider linear driving force dependency, it is recommended to take

the exponent aw equal to one.

List of symbols
l
k B ,w = overall mass transfer coefficient for the water in the bulk, m
s
l
k L ,w = mass transfer coefficient of the water, m
s
l
kT ,w = transfer coefficient due to temperature difference between the frozen layer and the bulk, m
s

Ti = the freezing potential,

5.4.2. Viscosity
The rheology of ice cream is very complex: it depends on the number, size and shape of the
suspended ice, fat and air particles, the concentration of the sugars, protein and polysaccharides, and
the temperature. Most of these factors are changing significantly during the manufacturing process.
The ice, fat and air particles are forming, the concentration of the solution is increasing and the
temperature decreases. As a result, the viscosity of ice cream increases by several orders of magnitude
(Clarke, 2004).
The ice cream mix, which is a solution of sugars and stabilizers and a suspension of fat
droplets, is a shear-thinning liquid, obeying a power-law type equation. As the mix freezes and air is
inserted inside the structure, its viscosity increases both because the temperature is decreased and
because ice particles and air bubbles are formed as well. At -5 0C (the temperature at which it
normally leaves the factory freezer) ice cream is a viscoelastic, shear-thinning fluid. Like the mix, it
obeys the power-law equation, but with different values of the rate law parameters. As its temperature
is lowered, it becomes more solid-like.
The following power-law type equation will be used for determining the viscosity of the
aerated ice cream mix:

Bi l a K Bi l a T

n 1

(5.133)

The terms K B

i l a

and n in equation (5.133) are rheological constants, characterizing the

fluid. Newtonian fluids are a special case of equation (5.133) in which K B

i l a

The consistency K B

i l a

i l a

and n 1 .

is calculated as a function of the temperature from the following equation:

188

Development of a reduced model for ice cream freezing

K B

i l a

B
C 1 xv a
T

(5.134)

List of symbols

Bi l a = non-Newtonian viscosity of the ice cream mix, Pa s

xv = volume fraction of air,
a

A, B, C = constants, determined experimentally

The values for the parameters used to calculate the viscosity are presented in Table 7.1, in
Appendix 11.

5.4.3. Thermal conductivity

The values for the thermal conductivities of the pure components and phases are assumed
constant and the values considered are presented in Table 7.2 in Appendix 11.
Two main factors affect the apparent thermal conductivity of the frozen ice cream:
temperature due to its strong influence on the ice fraction and the ice cream apparent density due to
the greater influence of the air insulation effect (Cogn, Andrieu, Laurent, Besson & Nocquet, 2003).
Thermal conductivity of the un-aerated mix

1)

The thermal conductivity of the liquid phase is modelled using the intrinsic thermal
conductivity values of each pure component and their volume fractions. The thermal conductivity of
the mix is determined as a function of the thermal conductivity of the pure components, using the
parallel model (Renaud, Briery, Andrieu & Laurent, 1992):

mix

domain
xv , domain
, j j
j

xv , domain
, j xdomain , j

(5.135)

mix
domain
j

(5.136)

List of symbols
mix

domain
= thermal conductivity of the un-aerated mix, W
mK

xv , domain
, j = volume fraction of component j ,

xdomain
, j = mass fraction of component j ,

j = thermal conductivity of component j , W m K

189

Chapter 5
2)

Thermal conductivity of the aerated mix

As the air intrinsic thermal conductivity is very low, the air induced porosity has a large

influence on the apparent thermal conductivity of commercial frozen ice creams. Since the
experimental observations show that the air phase was dispersed as quite spherical bubbles inside the
structure (Goff, 1997; Caillet, Cogn, Andrieu, Laurent & Rivoire, 2003), the thermal conductivity is
modelled using the Maxwells relationship (Cogn, Andrieu, Laurent, Besson & Nocquet, 2003):
mix

domain domain

2domain
2 domain

mix

mix , a

mix

2domain

mix

mix
domain

(5.137)

List of symbols
mix , a

domain
= thermal conductivity of the aerated mix, W
mK

a = the air volume fraction,

a = thermal conductivity of air, W m K

5.4.4. Heat transfer coefficients

1)

Heat transfer coefficient of the ice cream mix

It is assumed that the heat transfer coefficient is based on the properties of the overall (multi-

phase) ice cream mixture, without taking into account the contributions of the individual phases in the
mixture.
The heat transfer coefficient of the ice cream mix is estimated using an empirical correlation,
based on the penetration theory (Bongers, 2006):

Bl a 2 l a cpl a N blades D Bl a

(5.138)

This expression for the heat transfer coefficient is valid for laminar flow or for the transition
region between laminar and turbulent flow.
In equation (5.138) is a correction factor accounting for the axial dispersion and the
incomplete temperature equalization and mixing. The correction factor is a function of the Pclet
number (Trommelen, Beek & Van de Westelaken, 1971).
List of symbols
= empirical correction factor,

i l a = density of the fluid bulk, kg

m3

190

Development of a reduced model for ice cream freezing

Bi l a = thermal conductivity of the fluid bulk, W m K
2)

Heat transfer coefficient at the liquid side of the ice particles

The expression for the heat transfer coefficient for an ice particle is determined by applying

the shrinking core model for the melting of the ice particle:

ice

(5.139)

The derivation of equation (5.139) is described in Appendix 12.

3)

Heat transfer coefficient of the walls

For determining the heat transfer coefficient of the freezer wall and the frozen layer, the

following equation will be used:

1
wall

rin ln rout / rin

wall

4)

(5.140)

Overall heat transfer coefficient of cylindrical walls

For determining the overall heat transfer coefficient of cylindrical walls the following

expression is defined, based on the inner surface (Bird, Stewart & Lightfoot, 2002):
n
ln ri / ri 1
1
1
1

r0 U 0 r0 0 i 1 i 1,i
rn n

(5.141)

The equation is defined for heat conduction through a laminated tube with a fluid at temperature
Ta inside the tube and temperature Tb outside, and a number of n regions.

5)

Heat transfer coefficient of the coolant

The coolant is assumed to have a high, constant heat transfer coefficient, corresponding to the boiling
regime.

c 12000

(5.142)

List of symbol

wall = heat conductivity of the wall, W m K

rin , rout = inner and outer radius of the wall, m

c = heat transfer coefficient for the coolant, W

m 2 K

191

Chapter 5
U 0 = overall heat transfer coefficient based on the inner surface, W 2
m K

0 = heat transfer coefficient of the fluid with the temperature Ta , W

m 2 K

n = heat transfer coefficient of the fluid with the temperature Tb , W

m 2 K

i 1,i = thermal conductivity of the region i 1, i , W m K

r0 = inner radius of the tube, m
rn = outer radius of the tube, m
ri = the radius of the region i , m

5.5. Overview of the reductions for the constitutive equations in the 3-D model
At the constitutive equations level, the following reductions were considered:
1.

The density of air is calculated using the ideal gas law.

2.

The sucrose is chosen to represent the different types of sugars present in the ice cream mix.

3.

The enthalpy of the ice particles dispersed in the bulk is determined considering the contribution

of every particle at the temperature of the ice phase in the bulk.

4.

The viscosity of the aerated ice cream mix is determined using a power-law type of equation.

5.

The properties (density, specific heat, viscosity) of the other pure components are considered

constant along the freezer.

6.

The ice particles are assumed to have a spherical shape.

7.

The transfer coefficients at the side of the frozen layer surface are taken proportional to a

freezing potential expressed as a scaled temperature difference ( T ).

i

8.

The heat transfer coefficient of the ice cream mix is estimated using an empirical correlation,

based on the penetration theory.

9.

The frozen layer and the freezer wall are treated as composite cylindrical walls for the heat

balance equations.
10. The heat transfer coefficient of the coolant is considered constant along the freezer length.
11. The rate laws for melting, mass and heat transfer are all structured in the same fundamental way:
the rate is proportional to a single thermodynamic driving force. This driving force is mechanically
enhanced or diminished by the scraper rotation. The transfer coefficient is built up from a nominal
transfer coefficient and an empirical adjustment parameter.
12. The distribution of the ice particles removed from the frozen layer is independent of the size
distribution of the ice particles inside the bulk. Due to the lack of detailed empirical information, a
Weibull-type distribution is assumed for the scraped ice particles removed from the frozen layer.

192

Development of a reduced model for ice cream freezing

13. The rate of change of ice particles due to melting in the bulk is determined assuming that the
incoming heat from the aerated bulk liquid is used only for the melting.
14. The heat generation due to the viscous dissipation is determined using a power-law type of
equation for the shear stress, under steady state shear flow assumptions.
15. The scraping friction is determined using a power-law type of function based on the amount of
ice being scraped from the frozen layer at the wall.

6. Summary of the mathematical structure of the 3-D model

An overview of the physical structure of the reduced model for the ice cream freezing is
presented in Table 5.9. Based on this physical structure, the mathematical model consists of a number
of seven partial differential equations (PDEs):
a)

Five mass conservation equations for the five species in the pseudo-continuous matrix phase

b)

One population conservation equation for the ice particle in the bulk

c)

One energy conservation equation for the fluid in the bulk, consisting of matrix, ice particles and
air
To this set of partial differential equations, a number of seven initial conditions and nine

boundary conditions are added.

Moreover, a number of eleven nonlinear algebraic equations (NLAEs) are related to the
various transfer terms in the balance equations:
a)

One algebraic equation to determine the thickness of the frozen layer

b)

One algebraic equation for the transfer rate of water from the bulk to the frozen layer

c)

Two algebraic equations for the transfer rate of ice from and to the frozen layer

d)

One algebraic equation for the melting rate of ice particles in the bulk

e)

One algebraic equation for the generation of water due to ice melting in the bulk

f)

One algebraic equation for the heat transfer rate from the bulk to the coolant

g)

Two algebraic equations for the viscous dissipation and work in the bulk

h)

One algebraic equations for the thermodynamic equilibrium ice-water-sugar

i)

One algebraic equation to determine the pressure drop

j)

Physical properties of the different pure species considered in the model: six species x three

Finally, some more algebraic equations are present to determine:

properties (density, specific heat, conductivity) = eighteen equations for the pure component
k)

Mixing rules for the three physical properties above three equations

l)

The viscosity of the ice cream mix one equation

m) Enthalpy of the different phases: two phases (aerated liquid and ice phase in the bulk) x one
equation = two equations
n)

Mass fluxes in axial direction one equation x six species = six equations

193

Chapter 5
o)

Mass fluxes along the radial coordinate one equation x one species = one equation

p)

Energy flux in axial direction one equation x one pseudo-component = one equation

q)

Mass transfer coefficients for the liquid and the ice particles in the bulk = two equations

r)

Heat transfer coefficients for the ice cream mix, liquid side of the ice particles, frozen layer,
freezer wall and the coolant five equations

Table 5.9. Overview of the ice cream freezing model

3-D (time, axial and internal coordinate)
Water

Bulk (B)

Sugar

Fat

Other

Transfer

Mechanis

(differential)

Mass

Convection
FL

Generation
Continuity of fluxes

BC
Frozen layer (FL)

Ice

Transfer

Mechanism

B
Known initial thickness of layer

IC
Bulk (B)

Ice

Mechanis

(differential)

Momentum

Transfer

B & FL

Consumption

IC

Known inlet distribution of particles

BC

Known distribution of maximum size particles

Pressure drop driven radially symmetric axial convective flow

(algebraic)
Bulk (B)

Mechanism

(differential)

Energy

Conservation

Population

Convection

BC

Convection
Generation

Viscous dissipation, scraping friction, melting

Transfer

FL

Work
Known inlet temperature

Frozen layer (FL)

(algebraic)
Freezer wall

(algebraic)
Coolant

(algebraic)

194

Air

Development of a reduced model for ice cream freezing

3-D (time, axial and internal coordinate)
Density
Enthalpy

Physical property

Viscosity

variations

Conductivity
Ice cream mix

Heat transfer

Frozen layer

coefficients

Freezer wall

Mass transfer

Ice particles

coefficients

Water

Constitutive equations (algebraic)

Coolant

Water from B to FL

Mass

Water due to melting of ice in B

Ice from B to FL
Ice from FL to B

Rates of transfer

Population

Melting of ice in B
Heat transfer due to transfer of mass
Heat transfer due to a

B to FL

temperature difference

FL to Wall
Wall to Coolant

Energy

Heat transfer due to melting of ice particles

Heat generation due to

Viscous dissipation

mechanical work

Scraping friction

Thermodynamic equilibrium ice-water-sugar

s)

Rotor contribution function two equations

t)

Distribution of the ice particles being scraped from the frozen layer one equation
The total number of equations in the nonlinear set is seven differential equations, with seven

initial and nine boundary conditions, and fifty-four nonlinear algebraic equations. The initial and
boundary conditions are actually degrees of freedom in the use of the model and must be specified
when solving the model.
There are also numerous parameters in the model, which can be classified as:

Physical parameters (e.g., in physical correlations)

Design parameters (e.g., diameters, length of the freezer)

Operational parameters (e.g, rotational speed)

The reduced model does not contain uncertainty descriptions of the model parameters or the
constitutive equations.

195

Chapter 5
Table 5.10. Conservation equations for the 3-D model
Resource

Phase

Conservation equations

Initial

Boundary conditions

j j

the bulk

nB
t

BB DB

Frozen

B, z

B,s

s
S
S

Rt , FL B B Rt,B FL B
AB
AB
z

B cB, j

Matrix

Mass & Population

Ice particles in

condition

nB (0, z , s )
_

nB ,0 z , s

1)

cB, j 0, z

cB , j z

B GB, j B Rt,B FL, j

SB
AB

FL,max FL,min

End of freezing interval

Momentum

Ice

P
R

i l a
L 2 KB

i l a i l a R3

cream
in bulk

H Bi l a
t
EBl FL ,w

i l a
B

i
B FL

S B l
q
WB
AB T , B FL

Energy

i
FL
z

AB

smax

z 0

F B , z , j ,0 t
AB

t s t0

t s t0

2 n 1
n

TBi l a t ,0 TB,0i l a

TBi,0,lz a

R FL

Twall Twall ,coolant R FL qr ,FL

Fc hc Tc

Freezer

Coolant

Rt,im , FL B t , z

TBil a 0, z

TFLi Twall R FL qr , FL

wall

jm, B , j , z

n
R
n

1 0

R
2n 1

Frozen
layer

z 0

i
FL
z

FL,max FL,min

Bulk
domain

B,z

F B , z t , s

Rt,im ,B FL t , z , s Rt,lB FL ,w z

layer
End of scraping interval

2) jB,s

m,B , j , z

xc
qT , wall coolant
z

PSSA

R FL

ln R

R FL

i
ln

R wall

wall

xc

z 0

xc ,0

PSSA = pseudo-steady state assumption

A remark is in order on the wellposedness and solvability of the model. The model is well
posed in the sense that the number of variables, derivatives of variables and equations match and form
a closed system (no rigorous proof offered on paper; the hard empirical verification is in the
implementation of the model for numerical solution).
An overview of the most important equations in the mathematical structure of the reduced
model for the ice cream freezing step is presented in Table 5.10.
Furthermore, it is noticed that the model equations are continuous in the variables and at least
once differentiable, which favours the existence of a solution. Whether there can be multiple solutions

196

Development of a reduced model for ice cream freezing

cannot be a priori excluded due to some non-linearity in the algebraic equations for the rate laws.
Multiplicity of solutions and the possible occurrence of a high differential index (>1) in this PDAE set
are points of attention when developing the numerical solution procedure.

7. Input-output structure of the reduced model for model use

The previous sections presented the physical and behavioural structure of a reduced model
according to the outline in Chapter 2 of this thesis. In spite of the reduction efforts, the ice cream
freezer model is still very complex with respect to its internal structure.
When using a model of a process unit for computations, having this view on the internal
structure can be rather hampering by its richness of details. Therefore, it is systems engineering
practice to develop a complementary view at a higher level of abstraction for the use of the model.
This view is presented as an information input-output diagram (Figure 5.10).
Operational
parameters
Model
assumptions &
uncertainties

ok

FREEZER
MODEL

Inputs
uk

Outputs

7 DEs + 54 NLAEs
7 I.C. + 9 B.C.

yk

ek
Empirical
parameters

Design
parameters

Figure 5.10. Input-output structure of the model

Depending on the objectives of the model, the variables in the input-output structure may play
a different role, as detailed in the following sections. The choice for the different groups of
parameters and variables for the ice cream freezing model is detailed in Chapter 6 of the thesis.
a)

Validation of the model

The freezer model contains many unknown adjustable parameters, which, in a model
parameter estimation and validation effort, can be statistically estimated from experimental data on
inputs and outputs. Figure 5.11 shows this information flow.
After having done parameter estimation based on one set of experimental data, a model
validation step can be done by means of predictions using another set of independent measured data.
197

Chapter 5
Operational
parameters
Model
assumptions &
uncertainties

ok

Outputs

Inputs

FREEZER
MODEL

uk

yk

ek
Empirical

Design
parameters

? parameters

Figure 5.11. Specifications for model parameter estimation and validation

b)

Analysis of the freezer performance

The analysis of the freezer performance signifies the simulation of the model outputs given the
inputs, operational conditions, and design parameters. This application is depicted in Figure 5.12. The
output variables are connected with quality criteria of the final product and with the process
performance. Examples of product quality variables are softness, creaminess and cooling sensation.
Relations that link the output variables to the product quality exist (Bongers, 2008; 2009), but are
beyond the purpose of this thesis.
c)

Optimization of the freezer operation

For this application, depicted in Figure 5.13, the model is able to determine the optimal
operating conditions given the outputs, the inputs and the design parameters. Depending on the
objectives of the model, the variables in the input-output structure may play a different role, as
detailed in the following sections.
Operational
parameters

Operational
parameters

Model
assumptions &
uncertainties

ok

Model
assumptions &
uncertainties

ok

Outputs

Inputs

FREEZER
MODEL

uk

k
Design
parameters

Outputs

Inputs

yk

FREEZER
MODEL

uk

ek
Empirical
parameters

yk

ek

Design
parameters

Empirical
parameters

Figure 5.12. Specifications for analysis of the

Figure 5.13. Specifications for the optimization of the

freezer performance

freezer operation

198

Development of a reduced model for ice cream freezing

Operational
parameters

ok

FREEZER
MODEL

Outputs

Inputs

FREEZER
MODEL

uk

yk

ek

Design
parameters

Model
assumptions &
uncertainties

ok

Outputs

Inputs
uk

Operational
parameters

Model
assumptions &
uncertainties

Empirical
parameters

Figure 5.14. Specifications for improvement of

Design
parameters

yk

ek
Empirical
parameters

Figure 5.15. Specifications for product design

freezer design

d)

Improvement of the freezer design

The improvement of the freezer design signifies that the model is able to calculate the optimal
design parameters given the outputs, the inputs and the operating conditions, as shown in Figure 5.14.
e)

Product design

This application implies that the model is able to predict the optimal design parameters and
the optimal operating conditions of the freezer, given the outputs and the inputs, as shown in Figure
5.15.
It should be noted that the use of the freezer model for optimization purposes require some
additional modelling by specifying:

A proper objective function, fit for its goal

Inequality constraints specifying the operational window, equipment integrity

window

Knowledge constraints which bound the domain of validity of the empirical rate
laws, and which must avoid the improper use of such rate laws outside tested areas

The optimization studies with the freezer model are kept beyond the scope of this thesis for a
practical reason of time constraints on the thesis research effort.

8. Conclusions
The development of a reduced 3-D model of the ice cream freezing step has been presented in
this chapter. The structure-retaining model reduction approach acts more on the contextual and the
lower levels of the physical structure of the model (domains, phases, species, etc.), and less on the

199

Chapter 5
numerical level of the process model. The starting point for the model reduction procedure is not an
existing complete, rigorous model of the freezing unit, but a conceptual one under design.
The model considers six species: (a) liquid water, (b) ice, (c) sugar, (d) fat, (e) air, and (f)
other components. The ice is assumed distributed inside a fluid matrix composed of the other species;
hence, two phases are present in the unit. The unit is decomposed in the following compartments: (a)
the barrel, with two domains, the bulk and the frozen layer, (b) the freezer wall and (c) the coolant.
The equations of conservation are written for the following resources: (a) mass, for the
continuous phases, (b) population, for the distributed ice phase, (c) energy, and (d) momentum; three
coordinates are considered: time, axial position and internal coordinate of the ice particles.
Regarding the phenomena that take place in the unit:
a)

Mass: transport along the freezer, transfer between the barrel domains for the water

b)

Population: transport along the freezer, transfer between the barrel domains, and melting

for the ice particles in the bulk

c)

Energy: transport along the freezer, transfer from the fluid bulk to the scraped ice

particles and to the coolant through the barrel wall

d)

Momentum: pressure drop driven radially symmetric axial convective flow

Having formulated the model equations, and presented an input-output view of the use of the
reduced model, the next step is the numerical implementation of the model in a suitable software
environment and its use for practical computational applications. The numerical implementation does
not consist only in writing the model equations in the format required by the software, but also in a
verification of the correctness of the equations and their implementation. This is done in Chapter 6 of
this thesis. The numerical solution method, as well as the application of the ice cream freezing model
for a parametric sensitivity type of analysis will be discussed in Chapter 6 as well.

9. References
Aldazabal, J., Martin-Meizoso, A., Martinez-Esnaola, J.M., & Farr, R. (2006). Deterministic model
for ice cream solidification, Computational Materials Science 38, 9
Arbuckle, W.S. (1986). Ice cream, AVI Publishing, 4th edition
Bird, B.R., Stewart, W.E., & Lightfoot, E.N. (2002). Transport phenomena, John Wiley & Sons
Bongers, P.M.M. (2006). A heat transfer model of a scraped surface heat exchanger for ice cream,
Proceedings of the 16th European Symposium on Computer Aided Process Engineering
and 9th International Symposium on Process Systems Engineering, 539
Bongers,

P.M.M.

(2008).

Model

of

the

product

properties

for

process

synthesis,

Proceedings of the 18th European Symposium on Computer Aided Process Engineering, 55

Bongers, P.M.M. (2009). Process and equipment design optimising product properties and attributes,
Proceedings of the 19th European Symposium on Computer Aided Process Engineering,
225

200

Development of a reduced model for ice cream freezing

Bradley, R.L. Jr. and Smith, K. (1983). Finding the freezing point of frozen desserts, Dairy Record
84, 114
Bradley, R.L. Jr. (1986). Plotting freezing curves of frozen desserts, Dairy Record 85, 86
Caillet, A., Cogn, C., Andrieu, J., Laurent, P., Rivoire. (2003). A. Characterization of ice cream
structure by direct microscopy. Influence of freezing paramters, Lebensmittel-Wissenschaft
und-Technologie 36, 743
Clarke, C. (2004). The science of ice cream, Royal Society of Chemistry
Cogn, C., Andrieu, J., Laurent, P., Besson, P., Nocquet, J. (2003). Experimental data and modelling
of thermal properties of ice cream, Journal of Food Engineering 58, 331
Duffy, B.R., Wilson, S.K., Lee, M.E.M. (2007). A mathematical model of fluid flow in a scrapedsurface heat exchanger, Journal of Engineering Mathematics 57, 381
Eisner, M.D., Wildmoser, H., Windhab, E.J. (2005) Air cell microstructuring in a high viscous ice
cream matrix, Colloids and Surfaces A: Physicochem. Eng. Aspects 263, 390
Fredrickson, A.G., Bird, R. B. (1958). Non-Newtonian Flow in Annuli, Industrial and Engineering
Chemistry 50, 347
De Goede, R., and de Jong, E.J. (1993). Heat transfer properties of a scraped-surface heat exchanger
in the turbulent flow regime, Chemical Engineering Science 48, 1393
Goff, H.D. (1997). Colloidal aspects of ice cream A review, Int. Dairy Journal 7, 363
Goff, H.D. (2002). Formation and stabilization of structure in ice-cream and related products, Current
Opinion in Colloid and Interface Science 7, 432
Goff,H.D.

Theoretical

aspects

of

the

freezing

process,

(http://

www.foodsci.uoguelph.ca/dairyedu/freeztheor.html)
Hangos, K.M., Cameron, I.T. (2001). Process modelling and model analysis, Academic Press
Hoffman, J. D. (2001). Numerical methods for engineers and scientists, 2nd Edition, Marcel Dekker,
Inc., New York
Koxholt, M.M.R., Eisenmann, B., Hinrichs, J. (2001). Effect of the fat globule sizes on the meltdown
of ice cream, Journal of Dairy Science 84, 31
Leighton, A. (1926). On the calculation of the freezing point of ice-cream mixes and of the quantities
of ice separated during the freezing process, Journal of Dairy Science 10, 300
Marshall, R.T., Goff, H.D., Hartel, R.W. (2003). Ice Cream, 6th Edition, Kluywer Academic/Plenum
Publishers, New York
Qin, F.G.F., Chen, X.D., Farid, M.M. (2004). Growth kinetics of ice films spreading on a subcooled
solid surface, Separation and Purification Technology 39, 109
Qin, F., Chen, X.D., Ramachandra, S., Free, K. (2005). Heat transfer and power consumption in a
scraped-surface heat exchanger while freezing aqueous solutions, Separation and
Purification Technology 48, 150
Ramkrishna, D. (2000). Population balances: Theory and applications to particulate systems in
engineering, Academic Press

201

Chapter 5
Rao, C.S., Hartel, R.W. (2006). Scraped surface heat exchangers, Critical Review in Food Science
and Nutrition 46, 207
Renaud, T., Briery, P., Andrieu, J., & Laurent, M. (1992). Thermal properties of model foods in
the frozen state, Journal of Food Engineering 51, 83
Russell, A. B., Burmester, S.S.H., & Winch, P.J. (1997). Characterization of shear thinning flow
within a scraped surface heat exchanger, Trans IChemE 75, 191
Sun, K.H., Pyle, D.L., Fitt, A.D., Please, C.P., Baines, M.J., & Hall-Taylor, N. (2004). Numerical
study of 2D heat transfer in a scraped surface heat exchanger, Computers & Fluids 33, 869
Trommelen, A.M., Beek, W.J., & Van de Westelaken, H.C. (1971). A mechanism for heat transfer in
a Votator-type scraped-surface heat exchanger, Chemical Engineering Science 26, 1987
Wang, W., Walton, J.W., McCarthy, K.L. (1998). Flow profiles of power law fluids in scraped
surface heat exchanger geometry using MRI, Journal of Food Process Engineering 22, 11
Yataghene, M., Pruvost, J., Fayolle, F., & Legrand, J. (2008). CFD analysis of the flow pattern and
local shear rate in a scraped surface heat exchanger, Chemical Engineering and
Processing 47, 1550

202

6
NUMERICAL APPLICATION OF THE REDUCED
MODEL FOR ICE CREAM FREEZING

This chapter deals with a numerical implementation and application of the ice cream freezing model. The
3-D model is reduced to its steady state version (2-D). The numerical implementation and verification of this

model in MATLAB is presented. The nonlinear set of differential and algebraic equations is discretized using a
finite volume numerical approach, which offers a good trade-off between accuracy and speed. The predictive
capabilities of the model are assessed on basis of a base case scenario, which is further used as a starting point for
the generation of parametric sensitivity information for identifying and screening of parameters having a dominant
effect on the model outputs.

Chapter 6

1. Introduction
The model development procedure presented in Chapter 1 of this thesis (summarised in Figure
1.4) consists of four main steps: (a) the model specification and design; (b) the model
implementation; (c) the model validation and (d) the model application.

Model
specification
and design

Model

Model

Model

implementation

validation

application

Valid
alternative

Applied
alternative

Mathematical
alternative

Numerical
alternative

Equations

Method &
Software

Parameters &
Model fit with data

Results

Figure 6.1. Model development cycle

In the following sections, the model development procedure is continued with an application
of model implementation and verification, as shown in Figure 6.1, to the ice cream freezing model.
These steps are continued with the model validation. This step consists of: a) the design and execution
of an experimental programme for collecting informative data; b) model parameter estimation; and c)
validation of these estimates by means of model predictions of process outputs against measured data,
independent from the model fitting data set. Due to the significant effort required, this step will not be
treated fully. Only the analysis of the parametric sensitivity of the model will be treated in this
chapter.
A reduced dynamic model of the freezing unit has been developed in Chapter 5 of this thesis.
Based on this physical structure, the mathematical model consists of a number of seven hyperbolic
partial differential equations (PDEs) and ten coupled nonlinear algebraic equations (NLAEs),
written for three coordinates: time, axial position and internal coordinate of the ice particles in the
bulk. To this set of partial and differential equations (PDAEs), algebraic equations to determine the
physical properties of the different species considered in the model, as well for the matrix are taken
into account as well. An overview of the physical structure of the ice cream freezing model is
presented in Table 5.9, in Chapter 5, while the conservation equations are summarised in Table 5.10.
The resulting input-output structure of the model is presented in Figure 6.2.

204

Numerical application of the reduced model for ice cream freezing

Operational
parameters
Model
assumptions &
uncertainties

ok

FREEZER
MODEL

Inputs
uk

Outputs

7 DEs + 54 NLAEs
7 I.C. + 9 B.C.

yk

ek
Empirical
parameters

Design
parameters

Figure 6.2. Input-output structure of the model

3 D model

time t ; position z ; particle size s

Variables : x states; v algebraic; p parameters

Equations :

FPDE x t , x, z x, s x, v; p 0
FNAE x, v; p 0

Boundary conditions :

BC x, v; u t 0

Initial conditions :

IC x; x0 z , s 0

(6.1)

Obtaining and interpreting the numerical solutions of the 3-D model summarised in equations
(6.1) offers a significant challenge. For this reason, it is preferred to start in a simpler way, by
considering the steady state version of the model first. The result is a 2-D model, for which a
numerical solution scheme will be developed in the following sections. The viability of this scheme
will be shown by means of a single case study. A more extensive application of the model for further
improvement of the design and operation of the freezer is, as interesting as it may be, beyond the
scope of this thesis.
Section 2 of this chapter presents the numerical implementation of the steady state version of
the reduced (3-D) model. The selection of the numerical methods and the software environment used
for solving the model are briefly introduced, together with the discretization of the model equations.

205

Chapter 6
Moreover, a verification of the model implementation, for asserting the correct behaviour of the
numerical implementation for some extreme cases (all rate parameters are assumed zero, the bulk
domain consists of only liquid water, etc.) is described in this section as well. Pilot results for solving
the steady state model are presented. The use of the model for inverse problems is not considered yet,
just its direct use in simulations. In Section 3, the first stage of the model validation step is discussed:
the analysis of the parametric sensitivity of the model. A sensitivity analysis based on a singular value
decomposition approach is used. The conclusions are presented in Section 4.

2. Development of a numerical scheme and implementation

Having formulated all the model equations, the next step is the implementation of the steady
state model in a software environment. It is clear that an analytical solution is beyond practical reach;
hence, a numerical solution of the steady state model is attempted.
Within the set of model equations, the population balance for the ice particles is the toughest
to solve. There are several numerical techniques available for solving population balance equations.
Recently, comparative studies have been published on the accuracy of these models (Qamar, Elsner,
Angelov, Warnecke & Seidel-Morgenstern, 2006; Mesbah, Kramer, Huesman & Van den Hof, 2009).
It has been shown that high order finite volume methods are well capable of dealing with population
balances. These schemes are simple to implement and have relatively low computational
requirements, as compared to other methods.
The total number of equations in the nonlinear set is seven differential equations, with seven
initial and nine boundary conditions, and fifty-four nonlinear algebraic equations. The initial and
boundary conditions are actually degrees of freedom in the use of the model and must be specified
when solving the model. These nonlinear partial differential and algebraic equations (PDAEs) are
approximated in MATLAB by a set of ordinary differential (ODEs) and algebraic equations (AEs),
as explained in Section 3.1.3 in Chapter 2 of this thesis. This set of ODEs is solved using ODE
solvers available in MATLAB.
Although a state-of-the-art dynamic PDAE solver, such as gProms or Jacobian, might have
been more efficient for the solution, the potential transferability of the model to industrial partners
who do not have access to such solvers lead to the decision to use MATLAB as the software
environment for the numerical implementation.

2.1. Discretization of coordinates and mathematical operators

Three coordinates are considered in the domain of the original 3-D model of the ice cream
freezer:
(a) the time coordinate, t
(b) the spatial coordinate, z
(c) the internal coordinate of the ice particles, s
i

206

Numerical application of the reduced model for ice cream freezing

The discretization of the coordinates in the domain involves a few aspects:
(a) the lower bound of the domain
(b) the upper bound of the domain
(c) the discretization resolution
(d) direction of the discretization
The discretization resolution has a strong effect on the accuracy of the solution. The smaller
the spacing, the more accurate the solution becomes. However, a high number of intervals will have
an effect on the increase of the computational time. The direction of the discretization has an effect on
the stability of the solution.
The span of the coordinates is determined in the following, being defined as the ratio between
the upper and the lower bounds of the coordinates. These bounds are estimated for the required
domain of operation and design. Time will also be considered for comparison with spatial position
and particle size.
For the time coordinate, t :

The ratio between the maximum and the minimum value of the time variable is calculated.
The maximum value is considered the operation time, which usually has values of the order of weeks.
The minimum value of the time is the time the particles grow.

max
operation time
106

2 108
min time particle grows 10

(6.2)

For the spatial coordinate, z :

Similarly, a ratio is calculated for the spatial coordinate. The maximum value is the freezer
length, which is in the order of meters. The minimum value considered is the maximum particle size,
since no distinction is desired between the ice crystals.
zmax
freezer length
10

105
zmin biggest crystal size 104

(6.3)

For the internal coordinate, s :

i

smax
i

smin

biggest crystal size 104

103
smallest crystal size 107

(6.4)

Since the ratio of the time coordinate is the highest of these three, it is concluded that this
coordinate is preferred as the continuous coordinate, when a numerical integrator for differential
equations is used. The other two coordinates can be discretized over their domain. If time drops out,

207

Chapter 6
the spatial coordinate comes next into consideration as the continuous coordinate for a numerical
integrator for differential equations.
a)

The time coordinate, t

Under the steady state assumption, the time coordinate does not appear in the model.
b)

The spatial coordinate, z

In the case of the spatial coordinate, the lower bound is equal to zero, while the upper bound is
equal to the length of the freezer, L .
zmin 0

(6.5)

zmax L

(6.6)

For given lower and upper bounds, the simplest spacing is equidistant, but sometimes it may be
more convenient to have a smaller spacing (higher resolution) in certain parts of the domain, and
a larger spacing in other parts. The spacing should not go below 105 L . Under the steady state

assumption, the spatial coordinate is not discretized and kept as a continuous coordinate.
The internal coordinate, s
i

c)

The lower bound should be equal to or smaller than the critical nucleus size and equal to or
larger than zero. The upper bound should be chosen such that the concentration of particles at
this point is assumed zero. Since it is assumed that the ice is being formed only in the frozen
layer, the upper bound for the internal coordinate is:

smax
2 5 FL
i

(6.7)

For given lower and upper bounds, the simplest spacing is equidistant, but sometimes it may be
more convenient to have a smaller spacing (higher resolution) in certain parts of the domain, and
a larger spacing in other parts. Depending on the employed crystallization kinetics, the use of a
crystal size domain with non-equidistant node spacing is favourable.
Moreover, whether a numerical method is appropriate for the reduction of the population balance
equations (PBEs) to a set of differential and algebraic equations (DAEs) is foremost
determined by the dominant crystallization mechanisms (Bermingham, 2003).
If crystal growth occurs, the PBE has a convective character. In this case, the PBE is classified as
a hyperbolic PDE. Hyperbolic PDEs are best solved with an upwind finite difference method or
a finite volume method (Bermingham, 2003). For crystal melt, a similar procedure is used,
except that the convective character in now directed downwind along the internal (size)
coordinate.

208

Numerical application of the reduced model for ice cream freezing

Although first order upwind finite difference and finite volume methods do not suffer from
numerical instabilities when used correctly, they do invariably exhibit numerical diffusion. As a
result, acceptable solutions require relatively fine grids (Bermingham, 2003).
Additional assumptions on the numerics are
considered:
The size distributed population balance of the ice

1.

phase is discretized for nc classes of particles (from

smin
to smax
)
i

The ice particles in class k 1 are melting into

2.

Figure 6.3. The melting of the particles in

class k

particles in class k and water (Figure 6.3).

For the ice particles size, discretization is performed in the following way:
i

sk sk 1 sk
i

(6.8)

In the following study, the discretization is assumed equidistant. The discretization step is
calculated in the following way:
sk
i

smax
smin
nc 1
i

(6.9)

nc = the number of classes of particles considered,

2.2. Discretization of model equations for the steady state model

Considering the discretization of the ice particles size, and assuming steady state, the
conservation equations for the mass are re-written in the form presented in Table 6.1.
When going from the continuous to the discerte coordinate, one needs only to show how it
affects the mathematical operators, such as partial derivative and integral. The algebraic operations
( , , , /, etc.) are not affected as these take place in a single point, independent of the
discretization.
The transfer terms for the mass related are calculated using the equations in Table 6.2. The
discretization of the rate of change of the ice particles in the bulk will be discussed in more detail in
the following section.

209

Chapter 6
Table 6.1. Mass conservation equations for the bulk continuous and discrete form
Phase

Continuous & dynamic

jB , z

Ice

i
n,B

Rt,BFL

Matrix

G
l
B

Bl Rt,lB FL,w

Component

jB , k

SB
AB

i
n , B ,k

l
B,w

jBl,z ,w

SB
S
Rt,iB FL ,k B
AB
AB

SB
AB

Rt,lB FL ,w

l
B,w

SB
AB

jBl,j , z

z jBl,j , z 0

j in Matrix

Rt,iFL
B ,k

SB
AB

jml,B ,w, z

Water in

Rt,FL B

Discrete + Steady state

i

Table 6.2. Overview of the rate laws for the mass related events continuous and discrete form
Rate

Continuous

Discrete equation

equation

Transfer of ice fromFL to B, Rt,iFL

B

i
Rt,im ,FLB norm

i
Rt,im ,FLB norm
,k

m i

mki

Transfer of ice from B to FL, Rt,iB FL

Ca k Bi f1 r nBi

Ca k Bi f1 r nBi,k

Transfer of water from B to FL, Rt,lB FL ,w

Cb k Bl,w f 2 r cBl,w cBl,w,e

Rate of melting of ice in the B,

i g ni,B a i ds

Gm , B

Rate of change of water due to ice melting in

i g ni,B a i ds

the B, GB ,w

i g ni,B ,k ski
k

i g ni,B ,k ski
k

2.3. Discretization of the population balance equation for the ice particles
The model is solved numerically considering a steady-state assumption. This assumption
transforms the mass and energy conservation partial differential equations in a set of ordinary
differential equations (ODEs) with respect to the spatial coordinate. However, the population balance
equation for the ice particles in the bulk remains a partial differential equation. In order to transform it
into an ODE with respect to the spatial coordinate, the finite volume method will be used for the
gridding of the size coordinate. The population balance will be solved under the assumption of
persistent melting of the ice particles, giving rise to a hyperbolic PDE, where the characteristic is
incoming at the upper side of the internal coordinate domain.

210

Numerical application of the reduced model for ice cream freezing

2.3.1. Finite volume method
The first step in applying the finite volume method (FVM) is the definition of a grid for the
size domain. The grid is shown in Figure 6.4. The space between the grid points forms the so-called
control volumes. The number densities are computed at the geometric centres of these control
volumes, which are the computational nodes.

Figure 6.4. Grid for the finite volume method

Hence, the discretized number density corresponds to the average number density in each
control volume, for example:
nB , k z
i

1
s
i

i
k 12

k1

nB s , z ds
i

(6.10)

The partial derivative with respect to the particle size must be evaluated at the computational
nodes. The calculation of the derivative requires the values for the number densities and the melting
rates at the grid points. The derivative is calculated using:
g n ,B

jB ,s
i

sk

sk

rs nB
i

r n
i

i
B

sk

k1

rs nB
2

sk 1 sk 1
2

k1

(6.11)

The melting rate at the grid point is obtained directly from the kinetic rate expressions.
However, the number densities cannot be evaluated at these grid points. The values for the number
densities can estimated by several interpolation techniques (Mesbah, Kramer, Huesman & Van den
Hof, 2009) such as the first-order upwind or higher-order interpolation schemes.
2.3.1.1. First-order upwind interpolation

The values of the number density at the grid points are simply approximated by the values at
the computational nodes. The sign of the melting term determines which point to choose. The
computational nodes for the interpolation need to be upstream of the grid points:

211

Chapter 6
i

nB , k 1

rs, B , k 1 0
i

nB i,k 1
i

nB , k

k 2...N
k 2...N

for

rs, B , k 1 0
i

(6.12)

nB i,k
i
nB ,k 1 i

2
nB , k 1

rs , B , k 1 0

k 1...N

rs, B , k 1 0

k 1...N 1

for

(6.13)

For melting of the ice particles, equation (6.11) is approximated as follows:

gn, B

jB ,s
i

sk

s
i

rs nB
i

nB ,k 1 rs

sk

rs

k1

nB ,k

k1

k 1...N 1

(6.14)

sk

The boundary condition must be implemented as well. A hypothetical computational node

located at N 1 accounts for the upper limit in the particle size domain. The number density is put to
zero at this nonexistent point, therefore the derivative at the computational node N becomes:
i

gn, B

sk

rs nB
i

0 rs

nB ,N

N1

(6.15)

s
i

sN

The implementation of the first order approximation of the finite volume method is outlined in
more detail in (Bermingham, 2003). It has been shown that the first order upwind interpolation
schemes suffer from numerical diffusion. This effect can be surpassed by using a larger number of
grid points or by implementing higher order interpolation formulae.
2.3.1.2. Higher order interpolation

A higher order scheme is obtained by piecewise polynomial interpolation of the space

between the computational nodes.
The approximation at the grid point sk 1 is written as:
i

nB ,k 1 nB ,k
i

1
1
i
i
i
i
nB ,k 1 nB ,k
nB ,k nB ,k 1
4
4

(6.16)

where is a parameter in the range 1,1

For 1 a second-order, accurate fully one-sided upwind scheme is obtained; while for

1 a standard, second accurate central scheme is obtained. For other values of the parameter, a
combination of the two schemes is obtained (Mesbah, Kramer, Huesman & Van den Hof, 2009).

212

Numerical application of the reduced model for ice cream freezing

Equation (6.16) is valid for the situation of growth. For melting, the expression should be
mirrored at the particular grid points sk 1 with k 1...N :
i

nB ,k 1 nB ,k 1
i

nB ,k 1 nB ,k
i

1
1
i
i
i
i
nB ,k nB ,k 1
nB ,k 1 nB ,k 2
4
4

1
1
i
i
i
i
nB ,k 1 nB ,k
nB ,k nB ,k 1
4
4

(6.17)

(6.18)

Two choices for the parameter are often encountered in literature: 1 3 and 1
(Qamar, Elsner, Angelov, Warnecke, & Seidel-Morgenstern, 2006). It has been shown that for

1 3 schemes linear equations become third order accurate, while for other cases the equations
become second order accurate.
Moreover, a flux limiter is implemented into equations (6.17) (6.18), which is able to limit
the spatial derivatives near sharp changes in the solution domain, therefore avoiding oscillatory
behaviour of the solution (Qamar, Elsner, Angelov, Warnecke & Seidel-Morgenstern, 2006).

2.4. Software environment

The differential equations of the ice cream freezer model have been discretized using the
procedure presented in Section 2.2 and 2.3 of this chapter, under the steady state assumption. In this
way, the PDAE system has been transformed into a set of nonlinear ODEs. It appears that the
algebraic equations can be rewritten in an explicit form, with the exception of the equations for
calculating the temperature difference between the frozen layer and the wall, and the wall and the
coolant.
2 D model

position z ; particle size s

Variables : xs states; vs algebraic; p parameters

Equations :

ODE ' s :

xs z ODE xs , vs ; p

NLAE ' s :

vs

NAE x, v; p

for

s S grid

(6.19)

Input as Boundary condition : xs z0 u z0 , s

Output variables :

y zL output xs L

The numerical solution of the discretized steady state ice cream freezing model presented in
equations (6.19) is performed using MATLAB.

213

Chapter 6
The resulting set of ODEs is solved using the built-in MATLAB ode solver ode15s, which
is a multistep, variable order solver based on the numerical differentiation formulas. At each step, the
solver estimates the local error in the i th component of the solution. This error must satisfy the
following condition (Matlab, 2008):

i max RelTol abs y i , AbsTol i

(6.20)

List of symbols

i = local error in the i th component of the solution

RelTol = specified relative tolerance of the solver
AbsTol = specified absolute tolerance of the solver

y i = the i th component of the solution

The relative tolerance of the solver, RelTol is set to1.0e 05 , while the absolute tolerance is
set to 1.0e 06 for all the variables, with the exception of the frozen layers thickness ( FL ), for
which the tolerance is1.0e 09 . In this way, it is ensured that the tolerances are small compared to
the values of the state variables.

2.5. Verification of the model implementation

Before proceeding with the model applications, the correctness of the numerical
implementation of the model is verified.
A first verification is performed by checking the conservation property of the overall amount
of water. In this way, the correctness of the mass conservation equations is assessed. This step is
performed in Appendix 14.
The second verification relates to the model implementation and it is done by testing the
behaviour of the steady state model in an extreme case: all the rate parameters are zero. The values
for the rest of the model parameters used in this step are shown in Tables 6.3 - 6.5 and in Appendix
11. The results for the verification step when all the mass transfer rates parameters are considered
zero are presented in Figure 6.5. The results show, obviously, that there is no transfer of mass
between the liquid bulk and the frozen layer. No water is being frozen, and no ice is being scraped
from the wall. The composition of the ice cream mix remains unchanged along the freezer.
Figure 6.6 presents the temperature profiles for the case when all the mass transfer rates
parameters are assumed zero, as well as the parameters for all the heat transfer terms. In this case, all
the temperatures remained unchanged along the freezer, as expected. Similarly, the pressure drop
vanishes when friction losses disappear by zero viscosity.

214

Numerical application of the reduced model for ice cream freezing

280

Water

Temperature, T / [K]

Mass flow, F / [kg/s]

0.06

0.04

0.02

Ice

0
0

0.1

0.2

0.3

0.4

0.5

0.6

Bulk

270
260
250
Coolant

240
0.7

0.1

Axial position, z / [m]

0.2

0.3

0.4

0.5

0.6

0.7

Axial position, z / [m]

Figure 6.5. Verification of the model

Figure 6.6. Verification of the model implementation:

implementation: Mass flows in the bulk domain

Temperatures

To prove rigorously that each of the rate laws has been implemented correctly by means of
comparisons with analytical solutions for simplified cases is a rather cumbersome effort and has been
skipped. Consequently, this has to be checked later against physical meaningfulness of the simulated
cases, such as correct physical trends and steady states.

3. Model applications
In the following sections, a procedure for using the steady state model of the ice cream
freezing for simulation and validation purposes will be presented.

3.1. Model simulation

Once the model equations are available and implemented in a software environment, and the
numerical implementation is verified, the model can be used for different types of applications such
as simulation or validation. During this step, the capability of the model to represent the physical
system must be assessed. This is done by model simulation and comparing model predictions with
actual experimental data.

3.1.1. Simulation scenario

The model is used as a simulation tool for reproducing physically plausible scenarios by
means of case based simulations. In the following section, the base case scenario, which will be used
to analyze the parameric sensitivity of the model, is defined.
Given the operating conditions (Table 6.3), the inputs (Table 6.4), the physical properties
(Appendix 11), the design parameters (Table 6.5), the empirical parameters (Table 6.6), the outputs of
the model are analysed.

215

Chapter 6
Table 6.3. Overview of the operating conditions variables for the base case
Name
Inlet temperature

Symbol

Unit

TB ,0

Value
278.15

Inlet pressure

P0

Pa

5 105

Feed flow rate of the mix

Finic

l
h

280

Temperature of the coolant

Tc

248.15

Rotor speed

rpm

240

Table 6.4. Overview of the input variables for the base

Table 6.5. Overview of design parameters for the base

case

case
Symbol

Unit

xBi,0

Value
0.000

Name
Freezer length

Symbol
L

Unit

Ice

Value
0.690

Water

xB l,w,0

0.645

0.204

Sugar

xBl,s ,0

0.170

D0

0.180

Fat

xB l,f ,0

0.080

0.01

xB ,o ,0

wall

Other

0.105

Overrun

Ovr

100

blade

104

Type of sugar
(molar weight)

Ms

kg

kmole

342.3

Inner freezer
diameter
Rotor
diameter
Freezer wall
thickness
Scraper blades
thickness
Number of
scraping blades

N blades

Inlet mass fraction

Name

Table 6.6. Overview of the empirical parameters for the base case
Name

Symbol

Unit

Ca

Value
1

Cb

Cc

2.5

5.0

Parameters for the work done by the

scraper

c1
c2

J
kg

Parameter for the viscous dissipation

3.5

Scale

40 106

Shape

4.0

107.57

29565

Fudge factors
for mass transfer
terms

Transfer of ice from B

to FL
Transfer of water from
B to FL
Ice melting in B

Parameter for the Nusselt number

Parameters of the
Weibull distribution

Power law parameters

Rotor action function parameters

Minimum thickness of the frozen layer

216

hcryst

3.0

0.65

103

3.0

FL,min

105

Numerical application of the reduced model for ice cream freezing

Table 6.7. Overview of the numerical control parameters for the base case
Name
Number of discretization points

Symbol

Unit

nc

Value
200

Maximum particle size considered

i
smax

100 106

The output variables are given in Table 6.8.

Table 6.8. Overview of the output variables for the base case
Name
Outlet temperature of the mix

Symbol

Unit

ic
Tout

Mass fraction of water in the product

ic
xout
,w

Mass fraction of ice in the product

ic

xout ,i

Mean diameter of the volume weighted distribution of ice

particles
Mean diameter of the surface area weighted distribution of
ice particles
Pressure drop

d 4,3

i
d3,2

Pa

Total amount of heat removed

Qtot

kW

Production rate

ic
Fout

l / h

3.1.2. Simulation results and discussion

The results are obtained for the steady state model of the ice cream freezing process, by using
the first order upwind approximation technique described in Section 6.2.1. This method is not as
accurate as the higher-order methods, but has the advantage of being fast. For the purpose of this
study, the results are not required to be very accurate, since the model is required to follow only the
expected trends of the variables. For applications where the behaviour of the freezer must be followed
closely by the model, an increase in accuracy will be required.
An option to increase the accuracy of the first

Table 6.9. Simulation times for different levels of

accuracy

for

the

first

order

upwind

approximation for particle size

order upwind approximation is to use a higher

number of discretization points for the ice particle

Number of
discretization
points, nc

Simulation
time, t [s]

100
200
500
1000

6
15
96
555

size. This will have an effect on time required for the

model solution, as seen from the results in Table 6.9.
Figure 6.7 presents the simulation results of
the size distribution of the ice particles in the bulk
domain, for different numbers of discretization
points.

The use of a higher number of points will have an effect on the solution, but in the same time,
there is an increase in the simulation time. The relative error between the simulation with 100 and

217

Chapter 6
1000 discretization points for the internal coordinate of the ice particles is less than 1.5%, while the
increase in computation time is of factor 100.

3.6

Particle size distribution,

n / [#/m4]

Particle size distribution,

n / [#/m4]

4
3

nc = 1000
nc = 500

3.55

nc = 200

3.5

nc = 100

3.45
3.4

30

32

34

36

38

Particle size, s / [m]

1
0
0

20

40

60

80

100

Particle size, s / [m]

Figure 6.7. Simulation results for different levels of accuracy of the first order upwind approximation:
Particle size distribution of the ice particles in the bulk domain

For this reason, a number of 200 discretization points for the size coordinate are used in
obtaining the results presented in the following sections. The simulation results are displayed for 120
integrator stepping points in the axial direction. The computation time is ~ 15 seconds.
The size distribution of the

40

Particle size distribution,

n / [#/m4]

ice particles coming from the

frozen layer to the bulk has an
influence

on

the

final

size

distribution of the ice particles, as

shown in Figure 6.8 and Table
6.10.
For
presented

the
in

application

the

24
16
8

layer

to

the

bulk

75

100

particles coming from the frozen layer in the bulk: Particle size

considered, with the value of the

distribution, 4 .

50

Figure 6.8. Simulation results for different size distributions of the ice

is

shape parameter for the Weibull

25

Particle size, s / [m]

sections, a normal distribution of

frozen

following

the ice particles coming from the

32

distribution of ice particles in the bulk at the freezers exit

(a)

no particle melting (dashed line); (b) particle melting (solid line).

represents the shape parameter of the Weibull distribution

218

Numerical application of the reduced model for ice cream freezing

The flow and temperature profiles are plotted along the axial direction of the heat exchanger
in Figures 6.9 and 6.10 respectively. The results obtained for the output variables are presented in
Table 6.10. The mass flow profiles in Figure 6.9 show that the ice does not survive inside the aerated
liquid bulk at the beginning of the freezer. It takes some time until the aerated liquid phase in the
bulk, which is warm upon entrance, cools down and reaches the equilibrium temperature of the icewater-sugar phase equilibrium.
This signifies that all the ice
Mass flow, F / [kg/s]

coming from the frozen layer during

this period melts completely, and the
temperature of the bulk liquid is
decreased

until

the

equilibrium

temperature. Once this temperature is

reached, the ice particles coming from
liquid

phase,

and

Water + Ice

0.04

Ice

0.02

Water

0
0

the frozen layer survive inside the

aerated

0.06

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Axial position, z / [m]

the

Figure 6.9. Mass flow profiles along the freezer (axial direction): ice

concentration of ice inside the bulk

and water

domain increases.

Table 6.10. Overview of the output variables for the ice cream freezing model for different size distributions of the
ice particles coming from the frozen layer to the bulk
Output

Outlet temperature of the

mix
Mass fraction of water in
the product
Mass fraction of ice in the
product
Mean diameter of the
volume weighted
distribution of ice particles
Mean diameter of the
surface area weighted
distribution of ice particles
Pressure drop
Total amount of heat
removed

Symbol

Unit

Size distribution of ice particles

from the FL to the B
Exponential,
Normal,

ic

Tout

269.1

269.02

ic
xout
,w

0.2218

0.2222

ic
xout
,i

0.4196

0.4192

i
d 4,3

71.1 106

41.6 106

i
d3,2

64.5 106

39.2 106

Pa

2.79 104

2.79 104

Qtot

kW

14.50

14.50

This behaviour is better visible in the temperature profiles shown in Figure 6.10. At the
beginning of the freezer, the aerated liquid phase in the bulk temperature decreases fast until the
temperature of the ice-water-sugar phase equilibrium.

219

Chapter 6
Figure 6.11 presents the concentration profiles of the ice cream mix along the freezer. Since
for all the components of the mix, with the exception of the ice and water, no transfer is considered,
the concentration profiles of these components remain constant along the tube. Moreover, the sum of
the mass fractions is equal to one, as expected, as seen in Figure 6.11.

Temperature, T / [K]

280

Bulk
Ice in Bulk

273

Frozen layer

266
259

Freezer wall

252

Coolant

245
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Axial position, z / [m]

Figure 6.10. Temperature profiles along the freezer (axial direction)

Mass fraction, z / [ ]

1
0.8

Ice
Water
Sugar
Fat
Other
Air
Total

0.6
0.4
0.2
0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Axial position, z / [m]

Figure 6.11. Mass fraction profiles along the freezer (axial direction)

The meltdown of the ice particles is visualised using the particles size distribution plots,
shown in Figures 6.12 and 6.13. In this figures, two types of processes are presented:

220

Numerical application of the reduced model for ice cream freezing

Particle size distribution,

n / [# / m4]

40
z = 100 %
32
z = 75 %
24
z = 50 %

16
8

z = 25 %

z=0%
0

25

50

75

100

Particle size, s / [m]

Figure 6.12. Ice particle size distribution along the freezer axial direction, z (a) with no particle melting (dashed
line) and (b) with particle melting (solid line)

Cumulative
distribution function

1
0.75
z = 100%

0.5
No melting

z = 25%

0.25
0

25

50

75

100

Particle size, s [m]

Figure 6.13. Ice particle cumulative distribution along the freezer axial direction, z (a) with no particle melting
(dashed line) and (b) with particle melting (solid lines)

Firstly, the melting of the ice particles is neglected (the dashed lines). The ice particle
distribution grows along the axial direction of the freezer and the shape remains unaffected. The only
phenomenon that changes the magnitude of the ice distribution is the transfer to and from the frozen

221

Chapter 6
layer. The number of ice particles steadily increases, while the shape corresponds with the size
distribution of the scraped away ice particles.
Secondly, the melting of the ice particles during axial transport is considered (the solid lines).
In this case, both effects are visible: there is an increase of the total amount of ice particles as the fluid
moves along the tube, due to the transfer from the frozen layer, and, in the same time, the size
distribution shifts towards smaller sizes, due to melting.

Rates for mass related

processes, Rm / [kg/m3s]

25

Transfer of water
from B to FL

20
15

Transfer of ice
from FL to B

10

Melting of ice in B

Transfer of ice
from B to FL

0
0

0.2

0.4

0.6

Axial position, z / [m]

Figure 6.14. Rates for the mass transfer related processes along the freezer axial direction, z
Water and ice phases in the bulk domain

In Figures 6.14 and 6.15, the variation of some of the mass and energy related processes are
presented.
From Figure 6.14 it can be seen that the amount of ice transferred from the bulk to the frozen
layer is practically negligible compared to the amount of water in the same direction. Due to the
steady state assumption, the amount of ice coming from the frozen layer into the bulk domain is equal
to the sum of the water and ice being transferred from the bulk.
In the same time, it can be noticed that at the beginning of the freezer, all the ice coming from
the frozen layer melts completely. Only once the equilibrium temperature is reached the ice particles
start to survive, and the rate of melting decreases.
Concerning Figure 6.15, it is interesting to note that the heat due to scraping diminishes when
moving along the axial coordinate, because the frozen layer diminishes in thickness due to decreasing
liquid water content available for freezing.

222

Numerical application of the reduced model for ice cream freezing

Rates for energy related

processes, E / [kW/m3]

10000
7500

Work
in the Bulk

Heat due
to scraping

5000
Heat transfer due to
temperature difference
between B & FL

2500

Viscous
dissipation

0
0

0.2

0.4

0.6

Axial position, z / [m]

Figure 6.15. Rates for the energy transfer related processes along the freezer axial direction, z
Aerated liquid phase in the bulk domain

3.2. Parametric sensitivity analysis for validation purposes

Collecting the data required to validate the model can be resource intensive. Furthermore,
poorly planned experiments will cause a greater loss of time and resources and will yield to little
useful information. For this reason, it is important to develop a strategy to perform the model
validation in a systematic way in order to maximise the information obtained from each experiment
and to minimise the number of analyses, the cost of materials and the required time. This is achieved
by design of experiments.
The design of experiments is a technique that aims to obtain the maximum information from
an experimental apparatus. There are different ways of making a design of experiments. When models
of an experimentally investigated system exist, these can be used in devising experiments that will
yield the most informative data, in a statistical sense, for use in parameter estimation and model
validation (Franceschini & Macchietto, 2008). For example, a priori information on system behaviour
(as captured by the model) is used before an experiment takes place. The experimental results will
hopefully offer new information for updating the model. Thus, before designing the experiments it is
important to determine how the behaviour of the system depends on the values of the different
parameters that characterize the system. The analysis of how a system responds to changes in the
parameters us called parametric sensitivity (Varma, Morbidelli & Wu, 1999).

223

Chapter 6
A sensitivity analysis is useful for understanding the relationships between the different
parameters of the model and the model performance. This relationship is quantified by a sensitivity
value. If a small change in a parameter value causes a large change in the output variables then these
outputs are said to be sensitive to that parameter. For this reason, the parameter will have to be
determined very accurately.
Two types of sensitivities can be defined (Varma, Morbidelli & Wu, 1999):

a) Local sensitivity provide information on the effect of a small change in each input parameter,
around a fixed nominal value, on each output or performance (dependent) variable. In this way,
the directionality of the parameters impact on the performance will be determined.

b) Global sensitivity provide information on the effect of large (real-life) variations of all
parameters on the output or performance (dependent) variables. In this way, the magnitude of the
parameters impact on the performance will be determined.
The (local) sensitivity is defined in the following way:

S jm

y j

(6.21)

wm

These sensitivities are scale dependent.

List of symbols
y = output (performance) variable

w = parameter

3.2.1. Sensitivity analysis by singular value decomposition

When the dependency of multiple outputs with respect to multiple parameters is considered,
one obtains a matrix filled with local sensitivities. Several questions arise at this point. What should
be done with such a matrix? What is the dominant pattern of influence? A tool for analysing a
sensitivity matrix of the model parameters is the singular value decomposition (SVD). It is used for
the identification of the dominant parameters, for example those parameters whose modifications
cause the largest changes into the system. Moreover, this approach offers information related to the
direction of the model outputs as well (Seferlis & Grievink, 2001).
The principle is to apply a SVD to the local sensitivity matrix. The local sensitivity matrix S L
is calculated by applying small variations on the model parameters:

224

Numerical application of the reduced model for ice cream freezing

y1
y1
w w
m
1

SL

y j
y j

w
wm
1

(6.22)

The calculation of the singular values is scale-dependent (Seferlis & Grievink, 2001),
therefore a scaling is needed in order to be able to compare the perturbations:
y j
wm

ws y j

y s wm

(6.23)

The SVD rewrites the scaled sensitivity matrix in the following form:
S U V T
s
L

(6.24)

with
U T U V T V Ir

(6.25)

diag 12 , 22 ,..., r2

(6.26)

where
r min j , m

(6.27)

List of symbols
y s = scaling factor on the model performance variables

ws = scaling factor on the model parameters

S Ls = scaled sensitivity matrix
= singular values matrix

U = matrix of the orthonormal input basis vector directions for S Ls

V = matrix of the orthonormal output basis vector directions for S Ls

r2 = singular values
An algorithm is defined for performing the singular value decomposition. The following steps
need to be followed:
1. Select the relevant parameters, w for which an analysis should be performed. They can be grouped
in categories such as operational, process design, physical, numerical control parameters, etc.

225

Chapter 6
2. Select the relevant performance variables, y .
3. Define a base case with fixed parameters values, w0 and calculate the values of the performance
variables, y0 . For the ice cream model, the base case is defined in Section 6.2.2.
4. Perform sensitivity analysis by making small local (linear) variations (0.5-1.0%) in the parameters
of the model. The components of the sensitivities matrix are scaled with the nominal values of the
performance variables and the parameters for the base case scenario ( y0 and w0 , respectively).
5. Apply SVD to the local sensitivity matrix and determine the number of dominant singular values.
Find the directionality of impact (matrices U and V for the dominant singular values) on the
performance variables.
6. Find the magnitude of the impact on the performance variables by making global, directional
variations (10-20%) in the parameters of the model. The dominant directions determined during
Step 5 will be chosen for this global study.
In the following section, the focus is on the generation of local parametric sensitivity
information for identifying and screening the dominant parameters. Accurate values of these
dominant parameters must be determined from experimental data obtained from a model-based design
of experiments.

3.2.2. Results of the parametric sensitivity analysis and discussion

A perturbation of 0.5% is applied on the model parameters and the effects are analyzed for
each category of parameters separately. The results of the SVD and the dominant effects are presented
in the following sections, for a number of four groups of parameters: (a) the operational parameters;
(b) the input parameters; (c) the freezer design parameters; and (d) the empirical parameters.
Table 6.11. Output variables for the parametric sensitivity analysis of the ice cream freezing model
Output

Symbol

Outlet temperature of the mix

ic
Tout

Mass fraction of water in the product

ic
xout
,w

Mass fraction of ice in the product

ic
xout
,i

Mean diameter of the volume weighted distribution of ice

particles
Mean diameter of the surface area weighted distribution of
ice particles
Pressure drop
Total amount of heat removed

i
d 4,3

226

i
d3,2

P
Qtot

Numerical application of the reduced model for ice cream freezing

The results will be presented in the form of the singular vectors for each group of parameters.
The input singular vector u corresponds to the largest singular value and indicates the input direction
that causes the largest modification in the performance. The corresponding output singular vector

v indicates the output direction in which the inputs are most effective.
Table 6.11 shows the output variables that are considered for the sensitivity analysis.
3.2.2.1. Operational parameters

The operational parameters considered for this study are presented in Table 6.12.
From the singular values presented in Figure 6.16 it is concluded that there is one dominant
direction for the operational parameters group. The computation time is ~ 60 seconds.
Table

6.12.

Operational

parameters

for

the

parametric sensitivity analysis of the ice cream

100

freezing model
Symbol

Inlet temperature of the

TB ,0

75

Singular values

Input

mix
Coolant temperature

Tc

Inlet mass flow of the

Finic

25

mix

Figure 6.16. The singular values for SVD on the

TB ,0

operational parameters of the model

Rotational speed

50

Finic

Tc

0.75

i
d3,2

Qtot

0.75

0.5

0.5

0.25

0.25

ic
ic
ic
xout
d 4,3
Tout
xout
,i
,w
i

-0.25

-0.25

-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.17. The singular vectors for the operational parameters group:
a) Input singular vector; b) Output singular vector

The singular vectors are presented in Figure 6.17. These singular vectors for the dominant
directions are listed in Table 7.4, in Appendix 15. From the results in Figure 6.17, it can be
concluded that the most sensitive perturbation of parameter combinations is an increase in the coolant

227

Chapter 6
temperature, coupled with an increase of the inlet mass flow of ice cream mix and rotational speed.
This will determine a decrease of the outlet temperature of the product and the pressure drop along
the freezer, as well as an increase of the water concentration in the final product.
It can be concluded that the set of dominant parameters to be considered in the validation
experiments for the operational parameters group are the coolant temperature, the inlet mass flow of
ice cream mix and the rotational speed.
3.2.2.2. Input parameters

The input parameters considered for this study are listed in Table 6.13.
The singular value decomposition reveals two dominant directions (Figure 6.18) for the input
parameters group. The singular vectors are presented in Figures 6.19 - 6.20 and listed in Table 7.5 in
Appendix 15. The computation time is ~ 95 seconds.
Table 6.13. Input parameters for the parametric
sensitivity analysis of the ice cream freezing model
Input

Symbol

Inlet water mass fraction

xB ,w,0

Inlet sugar mass fraction

xB ,s ,0

Inlet fat mass fraction

xB , f ,0

Inlet mass fraction of other

xBl,o ,0

2.5

Singular values

2
1.5
1
0.5
0

components
Overrun

Ovr

Type of sugar (molar weight)

Ms

l
xB l,w,0 xB , f ,0

xBl,o ,0

xB l,s ,0

Ovr

Figure 6.18. The singular values for SVD on the input

parameters of the model

Ms

i
d 3,2

0.75

0.75
0.5

0.5

0.25

0.25

ic
i
ic
ic
d 4,3
xout
Tout
xout
,i
,w

-0.25

-0.25

-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.19. The first set of singular vectors for the operational parameters group:
a) Input singular vector; b) Output singular vector

228

Qtot

Numerical application of the reduced model for ice cream freezing

The first direction of change shown in Figure 6.19 indicates that the most sensitive
perturbation in the input parameters is an increase of the concentrations of water and fat in the inlet
mix will have an effect on the decrease of the output temperature of the product and an increase of the
ice formation. Moreover, a decrease of the pressure drop along the freezer and of the total amount of
heat removed is also predicted.

l
xBl,w,0 xB , f ,0

xBl,o ,0

xBl,s ,0

Ovr

Ms

i
d 3,2

Qtot

0.75

0.75
0.5

0.5

0.25

0.25

ic
i
ic
ic
d 4,3
xout
Tout
xout
,i
,w

-0.25

-0.25
-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.20. The second set of singular vectors for the operational parameters group:
a) Input singular vector; b) Output singular vector

The second direction of change in Figure 6.20 indicates that an increase of the amount of
water, coupled with decrease of the amount of air in the inlet mix will decrease the outlet temperature
and the ice content in the final product. The type of sugar will also have a strong effect on these
variables as well, as shown in Figure 6.20.
The validation experiments for the input parameters group must be performed in two
directions. Firstly, a set of parameters consisting of the inlet mass fractions of water and fat, as well as
the type of sugars should be considered. Secondly, to the inlet mass fraction of water and the type of
sugars, the overrun must be added to get the most significant results.
3.2.2.3. Freezer design parameters

Table 6.14 lists the freezer design parameters considered for the sensitivity analysis.
The singular values (Figure 6.21) reveal two dominant directions for the freezer design
parameters group. The singular vectors are shown in Figures 6.22 - 6.23 and listed in Table 7.6 in
Appendix 15. The computation time is ~ 85 seconds.
The first set of singular vectors in Figure 6.22 indicate that an increase of the freezer diameter
will lead to an increase of the outlet temperature, a decrease of the water concentration coupled with
an increase of the ice concentration in the final product and a decrease of the pressure drop along the
freezer. Moreover, an increase of the total amount of heat removed is indicated as well.
The second set of singular vectors, in Figure 6.23, indicate that the decrease of the freezer
length, coupled with a decrease of the rotor diameter and an increase of the freezer wall thickness will

229

Chapter 6
have an effect of decreasing the outlet temperature of the mix. The ice formation, the pressure drop
along the freezer and the total amount of heat removed inside the freezer are reduced as well under
these conditions.
Table 6.14. Freezer design parameters for the
parametric sensitivity analysis of the ice cream
14

freezing model

12

Symbol

Freezer length

Inner freezer diameter

Rotor diameter

D0

Singular values

Input

10
8
6
4
2

Freezer wall thickness

wall

Scraper blades thickness

blade

Figure 6.21. The singular values for SVD on the freezer

design parameters of the model

D0

wall

blade

0.75

ic
i
ic
ic
d 4,3
xout
Tout
xout
,i
,w

i
d 3,2

Qtot

0.75

0.5

0.5

0.25

0.25

-0.25

-0.25

-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.22. The first set of singular vectors for the freezer design parameter group:
a) Input singular vector; b) Output singular vector

D0

wall

blade

0.75

i
d 3,2

0.75

0.5

0.5

0.25

0.25

ic
i
ic
ic
d 4,3
xout
Tout
xout
,i
,w

-0.25

-0.25

-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.23. The second set of singular vectors for the freezer design parameter group:
a) Input singular vector; b) Output singular vector

230

Qtot

Numerical application of the reduced model for ice cream freezing

Similarly, to the input parameters group, the freezer design parameters group validation
experiments should be focused on two directions. Firstly, the change in freezer diameter, indicated by
the first set of singular vectors must be analyzed. Secondly, validation experiments must be
performed considering as set of dominant parameters the freezer length, the rotor diameter, the
freezers wall and the blade thickness.
3.2.2.4. Empirical parameters

Finally, the empirical parameters used for this study are listed in Table 6.15.
It appears from the singular values presented in Figure 6.24 that there are two dominant
directions. The singular vectors for the dominant directions are shown in Figures 6.25 - 6.26 and
listed in Table 7.7 in Appendix 15. The computation time is ~ 190 seconds.
Table 6.15. Empirical parameters for the parametric sensitivity analysis of the ice cream freezing model
Input
Fudge factors
for mass transfer
terms

Symbol
Transfer of ice from B
to FL
Transfer of water from
B to FL
Ice melting in B

Parameter for the Nusselt number

Parameter for the viscous dissipation
Parameters for the work done by the
scraper

Ca
Cb
Cc

C
c1
c2

Scale
Shape
Rotor action function parameters

Minimum thickness of the frozen layer

FL,min

Parameters of the
Weibull distribution

The first set of singular vectors

25

indicates that the increase of the mass transfer

20

Singular values

processes between the frozen layer and the

bulk domain ( Ca and Cb ) will have an effect
of decreasing the outlet temperature of the
mix and the pressure drop along the freezer. A

15
10
5

decrease of the ice formation is also

predicted, due to the increase in the rotor

speed action ( r ).

Figure 6.24. The singular values for SVD on the

empirical parameters of the model

231

Chapter 6
The second set of singular vectors shows that an increase of the ice particle size coming from
the frozen layer ( ), coupled with the increase of the rotor action will have a strong effect on the
increase of the size of the ice particles in the final product.

Ca Cb Cc C c1 c2 n k FL ,min

0.75

ic
i
ic
ic
d 4,3
xout
Tout
xout
,i
,w

i
d 3,2

Qtot

Qtot

0.75

0.5

0.5

0.25

0.25

-0.25

-0.25

-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.25. The first set of singular vectors for the empirical parameter group:
a) Input singular vector; b) Output singular vector

Ca Cb Cc C c1 c2 n k FL ,min

0.75

ic
i
ic
ic
d 4,3
xout
Tout
xout
,i
,w

i
d 3,2

0.75

0.5

0.5

0.25

0.25

-0.25

-0.25

-0.5

-0.5

-0.75

-0.75

-1

-1

Figure 6.26. The second set of singular vectors for the empirical parameter group:
a) Input singular vector; b) Output singular vector

The singular vectors obtained for the empirical parameters group show that the validation
experiments must be performed in two directions as well. The first direction consists of the
parameters related to the mass transfer of ice and water to the frozen layer, the parameters of the rotor
action function, the parameters of the viscous dissipation term, and the scale of the Weibull
distribution function. The second direction relates to the parameters of the rotor action function as
well, together with the shape of the Weibull distribution function, and the minimum thickness of the
frozen layer.

232

Numerical application of the reduced model for ice cream freezing

4. Conclusions
The steady state version (2-D) of the reduced 3-D model of the ice cream freezing step
developed in Chapter 5 of this thesis has been implemented numerically in MATLAB. The model
reduction acted in this case at numerical level of the process model, more specifically on the
approximation of the functions and operators. The equations were discretized along the internal
coordinate of the ice particles using a first order upwind approximation. A verification of the
correctness of the numerical implementation of the model has been performed using extreme cases
with analytical solutions (the transfer rates were set to zero).
Moreover, the model has been used for some simulation scenarios for predicting the correct
trends of the outputs. These results have been also used successfully for generating parametric
sensitivity data for the identification and screening of dominant parameters of the model. Accurate
values of the dominant parameters must be determined from experimental data obtained from a
model-based design of experiments. The computation times for the solution of this steady state model
are short enough, certainly when compared to the time it takes to analyse and interpret the results.

5. References
Bermingham, S.K. (2003). A Design Procedure and Predictive Models for Solution Crystallization
Processes, PhD Thesis, Delft University of Technology, The Netherlands
Franceschini, G. & Macchietto, S. (2008). Model-based design of experiments for parameter
precision: State of the art, Chemical Engineering Science 63, 4846,
Matlab (2008), User Manual, version 7.7.0.471 (R2008b)
Mesbah, A., Kramer, H.J.M., Huesman, A.E.M., & Van den Hof, P.M.J. (2009). A control oriented
study on the numerical solution of the population balance equation for crystallization
processes, Chemical Engineering Science 64, 4262
Qamar, S., Elsner, M.P., Angelov, I.A., Warnecke, G., & Seidel-Morgenstern, A. (2006). A
comparative study of high-resolution schemes for solving population balances in
crystallization, Computers and Chemical Engineering 30, 1119
Seferlis, P. & Grievink, J. (2001). Optimal design and sensitivity analysis of reactive distillation units
using collocation models, Industrial and Engineering Chemistry Research 40, 1673
Varma, A., Morbidelli, M. & Wu, H. (1999). Parameteric sensitivity in chemical systems, Cambridge

University Press

233

Chapter 6

234

7
CONCLUSIONS AND RECOMMENDATIONS

This thesis covers the formulation and testing of an approach for the reduction of process models used for
chemical engineering applications. In this concluding chapter, the key contributions of this research are
highlighted being (partial) answers to the research questions posed at the beginning of this thesis. These key
contributions are the formulation of a reduction procedure, which is able to preserve the essential structural
features of a process, and the testing of this procedure on two realistic process modelling case studies.
Additionally, various paths for further research are discussed, together with some recommendations for future
study. These recommendations refer to the application of the model reduction procedure in a consistent and
generic way, and to the further development of the reduced models developed during the case studies.

Chapter 7

1. Research questions and approach

Modelling is an important tool in understanding the behaviour of a given system, as well as
for making decisions during the product and process lifecycle. For many applications, the model has
to be able to predict the behaviour of the system with sufficient accuracy, in a computationally fast
way to meet demands of an efficient work process. For this latter reason, reduction of the model
complexity is often required. Although a technique used in many areas of engineering, modelling and
model reduction is still based on the expertise of individuals (Cameron & Ingram, 2009), obtained by
experience. The need for a more systematic approach remains.
Existent model reduction techniques from the system and control area in chemical and process
engineering can be applied to a complete process (unit) model, but often achieve only modest results.
For example, the gain achieved in computing time is not as large as needed and this gain is offset by
loss of structural process information in the reduced model. On the other hand, specific existent
chemical engineering approaches to model simplification often select one critical model attribute,
such as lumping of species of reactions, but leave the remainder of a model as it is. Would it be
possible to find a more balanced approach between the systemic reduction without paying much
regard for the underlying (physical) structure and the chemical engineering approach of focusing on a
single model attribute and ignoring the systems view?
Therefore, the following questions are addressed in this thesis:
1.

What are the main structural features of chemical process models that co-determine the model
size and the computational efforts? Where is the structure being addressed in the modelling
procedures?

2.

How does the model structure reflect in the reduction approaches?

3.

Is it possible to exploit efficiently these structural features when reducing the complexity of the
process model?

4.

Are the reduced models fit for purpose?

These research questions have been tackled by:

(a) Formulating a structure-retaining model reduction approach, which acts as an integrating

framework for selectively applying existent reduction techniques to a process to be modelled;
(b) This approach is tested on two realistic process modelling case studies.
The main conclusions of this research are presented in Section 2, while recommendations for
further work are given in Section 3.

236

Conclusions and recommendations

2. Conclusions on model reduction approach with test results

2.1. Generation of model reduction options by a decomposition approach
The most common way of approaching the modelling for chemical engineering applications is
to decompose the system into individual entities that perform a transformation of resources, using the
network concept with nodes and edges. The process nodes are connected by material or information
edges, depending on the type of resources they transport. The following levels of decomposition for
the modelling procedure have been identified:
(a) Contextual level, which includes all the decisions related to the specification of the model
and the applications it will be used for
(b) Structure level, consisting of all the information referring to the physical entities
composing the system, the phenomena that take place inside the individual entities, as well as the
connections between them
(c) Mathematical level, composed of information related to the mathematical relations
defining the behaviour of the process
(d) Numerical level, which contains the decisions for the numerical solution method and
algorithm
(e) Software implementation level, which contains the decisions for the way of
implementation of the model and numerical algorithm
At each level of decomposition, a series of attributes have been identified that have an effect
on the model complexity. These attributes range from physical and behavioural entities composing
the process system (units, domains, phases, etc.) to numerical schemes that are used to solve the
process model. All have an effect on the complexity of the final software-implemented model. The
model complexity depends on the number of equations that need to be solved, the number of terms in
the model equations, as well as on the mathematical operations these equations contain and the
coupling between the equations. For these reasons, each of the attributes offers handles for
performing model reduction.
A structure-retaining model reduction is proposed. This approach has two stages:
decomposition and aggregation. Decomposition is applied in the first stage to: (i) the nodes in the
process network, (ii) the resources flowing through the edges and being processed in the node
according to the behavioural model for each node. The model reduction within a node can be
performed at all respective levels of the modelling procedure:
(1) Conceptual:

by changing model functions and goals

(2) Structure:

by simplifying the physical (compartments, domains, phases,

phenomena, etc.) and behavioural (rate equations) structure of the
model

237

Chapter 7
(3) Mathematical: by reducing the number and the complexity of the model equations
(4) Numerical:

by reducing the numerical accuracy and order of approximations of the

mathematical operations

(5) Software:

by efficient coding of model equations for fast execution

In the second stage, aggregation takes place: the overall process model is constructed from the
reduced models of the nodes using the network structure. Starting from a network-based
decomposition the reduction procedure is able to preserve the essential structural features of the
process. Moreover, the physical meaning of the variables and equations is preserved as much as
possible.
This approach has both conceptual and numerical elements, which makes it hard to assess its
fitness for purpose on theoretical grounds only. Therefore, making test applications with this
approach is seen as a practical way to evaluate its performance. The pragmatic criteria for success of
the structure-retaining model reduction approach are that one is able to develop a model that:
- is solved successfully in an acceptable amount of time
- retains sufficient predictive accuracy for practical applications
- retains the topological and physical structure of the process to a sufficient level of detail

2.2. Applications of the structure-retaining model reduction approach

The application of the method has been investigated for two types of processes:
(A) A complex (multi-unit) process with simple (single phase) product: an iso-butane
alkylation plant
(B) A simple (single unit) process with complex (multiphase) product: the freezing step in an
ice cream manufacture process

2.2.1. Model reduction for complex processes with simple products

For the iso-butane alkylation plant, a rigorous dynamic model of the process is possible to
develop, since the laws of conservation describing the model are not too difficult to implement and
solve. In the same time, information regarding all the process rate phenomena is available. Moreover,
the solution of the rigorous model is obtained in a satisfactory amount of time for off-line simulation
applications, using Aspen Dynamics. The complexity of this rigorous model lies in its nonlinear
behaviour, determined by the reactor-separation-recycle structure in the plant flowsheet.
The use of this rigorous model for two practical applications is tried. These two applications
are: (i) the assessment of the plantwide control structures and (ii) the dynamic optimization of the
plants operation. When the solution of these two problems is attempted, the rigorous model, works
for the plantwide control problem, but fails when the solution of the dynamic optimization problem is
sought. This requires the development of a reduced model of the plant.

238

Conclusions and recommendations

Two model reduction approaches are applied and their results are compared with outcomes of
the rigorous model:
(a) A reduced model is obtained using the traditional model reduction approach. Model
order-reduction is attempted on the full rigorous model. Indeed, a reduced model is obtained which,
however, wrongly approximates the rigorous models behaviour.
(b) The structure-retaining model reduction approach has been applied for obtaining the
reduced model of the process. The conceptual stage of the model reduction acts at the systemic and
the higher levels of the physical structure (units) of the chemical plant, for this type of processes.
The process model is decomposed in smaller models of each unit in the process flowsheet. Then,
numerical model reduction is performed for each unit or groups of units individually. The
distillation columns are reduced by balanced linear model-order reduction, eliminating the states
for which the Hankel singular values are small. The model of the reactor is reduced by model
simplification, considering only the component balances and constant temperature and physical
properties. The obtained reduced models are connected according to the flowsheet in the next stage,
to obtain the reduced model of the full process.
The reduced model has been implemented in MATLAB / SIMULINK for the assessment
of the plantwide control structures, and in gProms for the dynamic optimization of the plants
operation. The simulation results of the reduced models were accurately approximating the
behaviour of the rigorous nonlinear model of the iso-butane alkylation plant.
Due to the differences in handling the models, improvement of the solution time of the
reduced model has been observed in gProms, when compared to MATLAB / SIMULINK. The
two reduced models have been solving considering the same operating scenario and using the same
hardware capabilities.
The two applications using the structure-retaining reduced model of the alkylation plant were
successful: the reduced model reflected the process flowsheet structure, and it could be solved at the
expense of a small sacrifice in predictive accuracy. The result is not yet 100% satisfactory as the
dynamic optimization problem took a (too) long time to solve. The cause of this slow convergence
can be due to an inadequate model structure and inherent to the nature of the optimization problem.

2.2.2. Model reduction for simple processes with complex products

Ice cream is a complex product and its formation by freezing was the second test problem.
The development of a rigorous model for the freezer is conceptually conceivable, but practically
infeasible. The model will become very complex, because the formation of the very involved internal
structural features of the ice cream has to be properly included in the model. However, it is currently
impossible to formulate such a comprehensive model because of the limited quantitative knowledge

239

Chapter 7
regarding the rate phenomena that occur inside the system, let alone to solve it when having to deal
with laws of conservation of an exceptional high dimensionality ( 5 coordinates).
For these reasons, an essential reduction is performed in this case both conceptually and
numerically, during the stages of the modelling procedure, yielding a reduced 3-D model. The
purpose of this reduced model is the application to parameter estimation, simulation of operational
scenarios and design optimizations.
For this type of processes, the conceptual stage of the structure-retaining model reduction
approach acts on the lower levels of the physical structure (species, phases, domains, compartments,
coordinates), as well as on the behavioural level (phenomena) of the process model. The numerical
stage focused less on reducing the number of model equations, and more on simplifying the different
terms in these equations.
The model considers mass, momentum and energy as resources that are transformed inside the
unit. Five coordinates are considered initially, which are further reduced to three. Moreover, the
actual product structure consists of five different phases, with six species. The ice particles are
modelled as a population with a distributed size, while all other phases (including air bubble and fat
globules) are assumed pseudo-continuous and homogeneous. The momentum conservation has been
simplified to an expression relating the mass flow with the pressure drop. The model assumes that all
the ice particles in the bulk have the same temperature and the energy conservation equation for the
bulk is written for the liquid ice cream phase (matrix, air, ice particles). The frozen layer is treated as
a solid wall and pseudo-steady state is assumed when writing the energy balances over the frozen
layer, the freezer wall and the coolant. Further simplification has been introduced in the expression of
the rate laws, which are assumed dependent only of a single driving force. The obtained reduced
model has seven partial differential equations, with seven initial conditions and nine boundary
conditions, and fifty-four nonlinear algebraic equations.
The steady state version of this model was successfully implemented in MATLAB, and
verified. The numerical implementation of the model equations approximates the population balance
for the ice particles in the bulk using a first order finite volume method. The verification implied the
testing of the conservation properties of the model and of proper physical limits for extreme
conditions. The accuracy of the verified reduced model prediction could not be fully tested since data
from an experimental programme was not ready and its acquisition was beyond the purpose of the
thesis. However, it has been ascertained for a base case scenario that the model predicts the right
trends for the output variables, using the order-of-magnitude estimates for the value of the empirical
parameters present in the model.
Being able to predict proper trends, the model has been used to determine the relative
importance of the model parameters and their influence on the model outputs. This parametric
sensitivity is a first stage of the model validation step in preparation to a design of experiments for
parameter estimation and validation.

240

Conclusions and recommendations

To conclude, the resulting reduced model of the ice cream freezer retains important features of
the physical structure of the process and its steady state version is solved fast enough for simulation
and parametric sensitivity-type of applications. Since the accuracy of the reduced model has not been
verified against experimental data, this criterion for the success of the model reduction procedure still
remains to be checked.

3. Recommendations
The following section is structured in an inverse order relative to the conclusion section. First,
the recommendations for the particular cases are discussed, to offer some perspective on the specific
cases. Since for the process of type A, the rigorous model is available, the recommendations will
focus mainly on aspects of the structure-retaining model reduction approach, while for the process of
type B, a second group of recommendations will focus on model development and refinement for the
ice cream freezing step. The decomposition of the process modelling for model reduction is discussed
next, for improvements towards a more generic structure-retaining model reduction approach.

3.1. Model reduction for complex processes with simple products

The application of the model reduction strategy that accounts for the internal structure of the
process in the case of the iso-butane alkylation plant demonstrated the advantages of the approach,
and, in the same time, it raised some more questions.
The dynamic optimization of the plant operation has been solved with some difficulty,
requiring trial-and-error iterations for the initial values of the initial profiles of the control variables.
One reason might be that the structure decomposition is application dependent and, for this particular
case, was not done up to a satisfactory level. The optimality of the decomposition strategy needs to be
investigated further. The relation to the formulation of an optimization problem is also part of this
effort. For example, which parts of the model have significant impact on the objective function should
be investigated. Another reason for the difficulty of solving the dynamic optimization problem might
not be in the model, but in the solvers used in the gProms or Aspen Dynamics optimization
toolbox. The investigation of this issue cannot be easily performed because access to the code is not
always possible. Moreover, it could be that a combination of both the model and the numerical
solvers is behind the difficult solution of the dynamic optimization problem, which would make the
investigation even harder.
Another issue that needs to be investigated further is how to ensure that the decomposition of
the different individual units is performed in a consistent way. The cut-off criteria for the model
decomposition focused on eliminating the least important states in an individual unit, without
accounting explicitly for the effects at the significant scale of the full process. It should be ensured
that the decomposition of the different individual entities in the model structure is done in such way
to maintain the significant (time) scales of the full process.

241

Chapter 7
Moreover, the case discussed in Chapters 3 and 4 of the thesis has a relatively simple
flowsheet structure, although meaningful for the purpose of the research. The application of the
procedure for more complex process (with more recycle loops, for example) could offer extra insight
into the benefits of the structure-retaining model reduction approach.

3.2. Model reduction for a single unit with complex products

3.2.1. Application of the structure-retaining approach
Similar to the processes with a complicated flowsheet structure, in the case of the ice cream
freezing model the issue of setting a consistent criteria for cut-off when applying model
decomposition must be further investigated. The significant time and spatial scales of the important
phenomena that take place inside the freezer must be kept in the reduced model, and the
decomposition should be done accordingly.

3.2.2. Selective model refinement

Several selective refinements are recommended in order to make the reduced model of the
freezing step more meaningful as a tool for product design type of applications.
Firstly, the complex movement of the ice cream mix inside the freezer barrel resembles a
helical flow. Redefining the spatial coordinate of the unit accordingly would allow for a more
rigorous approach. The axial coordinate can be replaced by the introduction of a helical coordinate,
often used for modelling of extruders (Yu & Hu, 1997). The simplification of the momentum
conservation in the form of the pressure drop mass flow dependency does not fully account for the
complex movements that take place inside the freezer.
Secondly, it is possible that, in some of the cases, the dependence of the rate laws on a single
driving force is not sufficient to describe effectively the processes that take place inside the freezer.
The dependence of the rate laws on more driving forces must be investigated further. In the same
time, the possibility that a specific rate does not depend in a linear way on the dominant driving force
must be considered as well. Rather than inserting up-front all generic dependencies on driving forces
in the model, dedicated experiments on lab-scale must reveal which dependencies are the dominant
ones and only these are included in the model. The structure-retaining model reduction approach
allows for the different phenomena that take part in the process to be kept separated and distinct. In
this way, the change in the dependence of the rate laws on one or more driving forces, due to ongoing
experimental testing, can be done without having to repeat previous steps in the model development.
Thirdly, the final product qualities are dependent not only on the distribution of ice particles
inside the microstructure, but also on the distribution of the air bubbles and fat droplets. Adding such
distributions with corresponding internal coordinates to the model will lead to an increase in model
complexity, since more phases will be considered, with corresponding conservation equations, rate
laws for the relevant physical phenomena, physical properties, etc. However, it may be possible to use

242

Conclusions and recommendations

only one generic internal size coordinate for all distributed phases (ice, air, fat) by use of proper
scaling factors per phase, thus limiting the increase in complexity.

3.2.3. Preparing for model validation and applications

In order to have a powerful tool for applications, and extension from a steady state to the
dynamic version of this reduced model has to be developed, verified and validated. Once the
validated dynamic model is available, it can be used for applications such as simulations of operating
scenarios, control studies, dynamic optimizations, etc.
Another important point is that the accuracy of the model predictions depends on the
numerical methods and tools that are used for its solution. For solving the partial differential
equations, more sophisticated methods such as higher order finite volume methods or orthogonal
collocation could be used if deemed necessary. The accuracy of the solutions will increase, but
together with that, the computation time of the model solution will go up as well. This can be
surpassed to some extent by use of state-of-the-art (dynamic) PDAE solvers such as gProms, which
behave better than MATLAB.
A first step towards the model validation has been performed and described in Chapter 6 of
this thesis: identification of parameter sets with a high impact on the model outputs. The next step in
the model development procedure is to develop and perform experiments in order to generate data for
validation. Based on the importance and influence of the parameters on the model outputs, as
determined from the parametric sensitivity analysis, a design of experiments must be performed in
order to get suitable data for parameter estimation. Once the model parameters are obtained, the
model can be validated against a second independent set of experimental data. This validation will
allow the model to be used further for practical applications, such as product design, optimization of
process operation, etc.
The experimental procedure requires further research as well. In the past experimental set-ups,
it is common to have quantitative information on the product structure at the end of the freezer, hence
the final product composition, including information on the population distribution. This was because
past models did not account for the entire range of phenomena that take place inside the unit.
Dedicated experiments for measuring specific product structure variables along the freezing process
will improve the quality of the final validated model, by offering information on the phenomena
accounted for in the reduced 3-D model.

3.3. Decomposition of the process for model reduction

From the discussion in the previous sections, it can be concluded that the decomposition of the
process for model reduction depends on the intended model application. For different types of
applications and processes, the decomposition and the model reduction approach acts at different
levels of detail.

243

Chapter 7
Although the case studies presented in this thesis offered options on how to perform the
reduction for particular processes, there is still an issue on how to decompose the model in an
optimum way for generic applications and to set criteria for cut-off of time and spatial scales in a
consistent way over all the individual units. The break down of the model structure (Figure 7.1) can
be done at different levels of detail, as shown in Chapter 2 of the thesis. The performance of the
model clearly depends on the cut level in this line of exploration for the model structure. The choice
of this level is critical: too high involves the modelling of many useless features; too low may lead to
useless results. The focus in this thesis was on attributes of the structure. Further, features of the

No consideration

Flowsheet

Unit

Compartment

Phase

Phenomena

Detailed structure

dynamic behaviour, or of the control system should be investigated as well.

Level
0

Level
1

Level
2

Level
3

Level
i

Level
L-1

Level
L

Cut level
Figure 7.1. Line of exploration for the model structure

The approach used in the development of the models presented in this thesis proved to be a
challenge in itself. The two extremes that were considered were:
(a) The rigorous modelling approach. In this case, the model is developed in full detail, taking
into account all the structure and behaviour components present in the system, right from the
beginning of the model development. Often, this may prove to be time consuming, since the initial
level of complexity can lead to difficulties in model solving from an early stage.
(b) The simple to complex approach. The models developed using this approach are
considering initially only partial and relevant structure and behaviour components of the system, in
order to gain knowledge on the various internals of the process. Once this goal is achieved, more
detail is added to the initial case, in order to reach a desired level of model complexity.
There should always be a trade-off between these two approaches in order to be able to
develop a model, which is suitable for practical use.

244

Conclusions and recommendations

The success of the structure-retaining approach depends largely on two factors. First, the
experience of the model developer has an effect on the different choices during the model
development. Based on the user skills, and the clarity of the model goals, the reduced model takes
into account all the important phenomena, without being too complex for solving. Second, it is
important that adequate tools (numerical methods, software, and hardware) are available for solving
the model. The reduced model should be accurate enough to be used in practical applications, but
sometimes the available tools are not sufficient to ensure it due to the high complexity still present in
the models.
Taking into account all these factors, a testing programme has to be developed for the
decomposition of process for modelling in an optimum way. A similar procedure to the design of
experiments could be followed, in order to choose meaningful processes and applications that will
offer maximum of information on the reduction approach. This information should include how
detailed the physical structure will be to achieve the modelling goal for a generic application.
Moreover, consistent treatment of the different individual nodes in the process network should be
ensured, such that the overall time and spatial scales retain their significance.
In many cases, the different software environments available offer several model reduction
options that can be used when implementing process models. Aspen Dynamics, for example has a
linearization toolbox, while MATLAB offers order model reduction for the balanced realization of
linear models. The implementation of automated model reduction tools in such software
environments, with options of the most common existent (order-) reduction techniques available
(linearization, model-order reduction, etc.), will be beneficial for the development time of future
reduced process models, as well as for the use of the structure-retaining model reduction approach.
Once these model reduction options are available in various software environments, new
techniques of model synthesis can be investigated. The model can be formulated as a superstructure
over a field of (reduced) model building blocks and this superstructure offers options on how to
connect the different blocks in a way that satisfies constraints on behaviour and meets targets of an
objective function for model performance.
The application of the structure-retaining model reduction approach to processes with smaller
scales (such as nanoclusters, molecular scale, etc.), where the behaviour is different from the
processes at macroscales, or for networks in other engineering areas (electrical, mechanical, etc.),
where the structure is different from a chemical process, could be investigated as well.

4. References
Cameron, I.T. & Ingram, G.D. (2009). A survey of industrial process modelling across the product
and process lifecycle, Computers and Chemical Engineering 32, 420
Yu, Q., Hu, G.H. (1997). Development of a helical coordinate system and its application to analysis
of polymer flow in screw extruders. Part I. The balance equations in a helical coordinate
system, Journal of Non-Newtonian Fluid Mechanics 69, 155

245

Chapter 7

246

8
APPENDIX

Appendix

1. Summary of differential operations involving the mathematical

operators in cylindrical coordinates
For writing the different conservation equations in this thesis, as well as in the following
sections of this appendix, different mathematical operators were used. Using a cylindrical system of
coordinates, the mathematical operators have the following form (Bird, Stewart & Lightfoot, 2002):

(8.1)

1
r r
r r

(8.2)

1

r

(8.3)

(8.4)

1 r r 1 z z

r
r
r
z s

(8.5)

ur
r

1 ur u
u
rz r
r
r

r
z
u
u
1 u
u
r r z
r
r
r
z
u z
1 u z
u z
zr
z
zz

r
r
z

: u rr

(8.6)

The substantial derivative for a scalar field:

Dx x
x u x
x

ur
uz
Dt t
r r
z

(8.7)

The substantial derivative for a vector field:

D z z

ur z z u z z
Dt
t
r
r
z
D r r
r u r
r 2

ur

uz

Dt
t
r
r
z
r
D
z u
r

ur

uz

Dt
t
r
r
z
r

248

(8.8)

Appendix

2. Mass and population conservation equations

a. The population balance equations for the dispersed phase

The equations written in this form are used for the ice particles in the bulk. They could be
used for the air bubbles if they would be considered as a dispersed phase. The conservation equations
in the bulk domain for the dispersed phases are written in the form of the population balance
equation, in phase :

nB z u z nB r ur nB u nB s rs nB BB DB (8.9)
t

With the initial condition:

Initially, there is an initial concentration of dispersed particles inside the bulk domain:
At

nB nB ,0 z , r , , s

t0

(8.10)

The boundary conditions:

For each one of the coordinates, one boundary condition is sufficient, since the partial
differential equations are of hyperbolic-type and first order (Hoffman, 2001). The boundary
conditions should be defined upstream of the flow direction.
a)

In axial direction

Continuity of flows:
At

b)

(8.11)

(8.12)

u n t , z , r , , s
jB , z ,0 t , r , , s
B
z
z 0

z0

In radial direction

The radial inflow is assumed non-zero at r R .

At

c)

u n t , z , r , , s
jB , r,0 t , z , , s
B
r
r 0

rR

In angular direction

Periodic continuity:
At

d)

t , r 2

t, r 0

(8.13)

Along the internal coordinate

The boundary condition for the particle size is put at the upper range of the particle size due to
the size reduction by melting of the particles.
At

s smax

nB t , z , r , , s

nB , s t , z , r ,

s smax

249

max

(8.14)

Appendix
In case of particle growth, the boundary condition for the particle size is put at the lower range
of the particle size.
List of symbols

nB = concentration of particles, # 3
m s

= phase
t = time, s

u = phase velocity, m
s

z = axial coordinate, m

r = radial coordinate, m

= angular coordinate, rad

s = internal coordinate, m or kg

rs = the rate of growth/melting along the internal coordinate s , m or kg

s
s

nB ,0 = initial concentration of ice particles, # 3

m

jB , z ,0 = initial flux of particles in axial direction, # 2

m s

jB , r,0 = initial flux of particles in radial direction, # 2

m s

BB = birth rate of particles in the bulk, # 3

m s

DB = death rate of particles, # 3

m s

b. The mass conservation equations for the pseudo-continuous phase

The equations written in this form are used for the components of the matrix phase, as well as
for the air phase in the bulk domain. They could be used for the ice particles in the bulk if they would
be considered as a pseudo-continuous phase. The equations of change for the mass of the pseudo-

continuous phases in the bulk domain are written in the following form, per specie j in phase :

B cB , j

u z B cB , j r ur B cB , j u B cB , j B RB , j (8.15)

250

Appendix
With the initial condition:
An initial concentration of component j in phase is considered:
t0

At

cB , j cB , j ,0 z , r ,

(8.16)

The boundary conditions:

a)

In axial direction

Continuity of fluxes:
z0

At

b)

u z B cB, j
jm , B , z , j ,0 t , r ,

z 0

(8.17)

In radial direction
Radial inflow is non-zero at r R .
rR

At

c)

ur B cB, j
jm , B , r , j ,0 t , z,

rR

(8.18)

(8.19)

In angular direction

Periodic continuity:

At

It is assumed that the density of the continuous phase is calculated as:

B, j B cB, j

(8.20)

List of symbols

cB , j = concentration of component j in the bulk, kg phase 3

m phase

B = the volume fraction occupied by phase in the space between the rotor axis and the freezer
wall,

u = bulk phase velocity, m

s

RB , j = rate of change of component j in the bulk, kg 3

m s

cB , j ,0 = initial concentration of component j in the bulk, kg 3

m

jm , B , z , j ,0 = initial mass flux of component j in axial direction, kg 2

m s

jm , B , r , j ,0 = initial mass flux of component in radial direction, kg 2

m s

251

Appendix
j = component of the phase
c. The mass conservation equations for the frozen layer domain

The conservation equation for the frozen layer is written in the following way:

cFL
i

u z cFL
r ur cFL
u cFL
RFL
i

(8.21)

With the initial condition:

An initial concentration of ice is considered inside the frozen layer:
At

t0

cFL
cFL
,0 z , r ,
i

(8.22)

The boundary conditions:

a)

In axial direction

Continuity of fluxes
At

b)

z0

i
i
u z i cFL
jm ,FL , z ,0 t , r ,

z 0

(8.23)

In radial direction
The radial inflow is non-zero at r R FL .

At

r R FL

c)

In angular direction

i
i
ur i cFL
jFL
, r ,0 t , z ,

r R FL

(8.24)

(8.25)

Periodic continuity:
At

u
i

List of symbols
i
cFL
= concentration of ice in the frozen layer, kg 3
m
i
RFL
= rate of change of the ice in the frozen layer, kg 3
m s

i
kg
cFL
3
,0 = initial concentration of ice in the frozen layer,
m
i
jm ,FL , z ,0 = initial mass flux of ice in axial direction, kg 2
m s
i
jm ,FL , r ,0 = initial mass flux of ice in radial direction, kg 2
m s

252

Appendix

3. Momentum conservation equations

In a similar way with the mass conservation equations, the momentum conservation is written
on the rigorous coordinate set. It is assumed that all the phases inside the bulk domain are lumped into
a single phase, the ice cream.
From (Bird, Stewart & Lightfoot, 2002), the equations of motion in terms of , applied to the
bulk domain have the following form:

D B

i l a

Dt

Bi l a g

(8.26)

The equations are written without assuming that B is symmetric.

With the initial condition:
An initial velocity profile is assumed inside the bulk:
At

t0

u t 0 u0

(8.27)

The boundary conditions:

a)

In axial direction

The outlet pressure of the product is known:

At
b)

zL

P PL

(8.28)

In radial direction

At

r R FL

ur 0

(8.29)

At

r 0

ur 0

(8.30)

c)

In angular direction

Periodic continuity:
At

(8.31)

List of symbols
P = pressure, Pa

PL = the pressure at freezers outlet, Pa

Bi l a = the density of the bulk phase, kg

m3

u = velocity, m
s

B = the shear stress, Pa

g = gravitational acceleration, m 2
s

253

Appendix

4. Energy conservation equations

The energy conservation equation has the general form:
D h
Dt

ln Dp
q
: u

ln T P Dt

(8.32)

Equation (8.32) represents the equation of energy written in terms of flux, q (Bird, Stewart &
Lightfoot, 2002). This equation will be detailed for each of the phases considered in the following
sections, for cylindrical coordinates.
a. The energy conservation equations for the fluid bulk

The liquid phase and the air phase are assumed to be a pseudo-continuous phase, the aerated
fluid bulk phase. It is assumed that the density of the aerated fluid is calculated in the following
form:

Bl a Bl a cBl a

(8.33)

The equation of change for the energy of the aerated fluid bulk phase becomes:

D B

l a

cB

l a

hB

l a

Dt

q ln

l a
B

cB
ln TB

l a

DP
: u QT , B WB

P Dt

(8.34)

With the initial condition:

An initial enthalpy profile is assumed inside the bulk:
At

t0

hB

l a

hB ,0 ( z , r , )

(8.35)

la

t 0

The boundary conditions:

a)

In axial direction

At

z0

la
u z hB l a
u h

z 0 z B z 0

(8.36)

At

zL

TB
z 0

zL

(8.37)

b)

In radial direction

The energy at the interface goes entirely to the frozen layer:

At

r R FL

l a TB
QT , B
B r

r R FL

(8.38)

At

r 0

hB
r

(8.39)

l a

0
r 0

254

Appendix
c)

In angular direction

Periodic continuity:
l a

At

hB

q , B

(8.40)

List of symbols

Bl a = the volume fraction occupied by the aerated fluid phase in the space between the rotor axis
and the freezer wall,
cB

l a

= concentration of the aerated fluid phase in the bulk, kg phase 3

m
phase

Bl a = the density of the aerated fluid phase in the bulk, kg 3

m
hB

l a

= the specific enthalpy of the aerated fluid phase, J

kg

qB = the conductive heat flux vector, W 2

m
TB = the temperature inside the bulk, K

: u = the viscous dissipation term,

W
m3

u = the velocity, m
s
QT , B = energy flow due to a difference in temperature between the bulk and the other phases,
W
m3

WB t , z = work done in the bulk layer, W 3

m
DP
= the substantial time derivative of the pressure
Dt

Bl a = the thermal conductivity, W m K

b. The energy conservation equations for the ice particles in the bulk domain

It is assumed that the heating of the ice particles coming from the frozen layer inside the bulk
domain is very rapid. The temperature changes in axial direction and in time. Once in bulk, the ice
particles temperature starts to change due to the difference in temperature between the particles and

255

Appendix
the liquid bulk, and eventually melt. Since the ice particles have different sizes, the heating and the
melting will take place with different velocities for each type of particles. This means that a new
internal coordinate should be considered while writing the population balance equations for the ice
particles: the thermal coordinate, .
i

The population balance for the ice particles in the bulk will be rewritten in such way that the
concentration of the particles is a function of two internal variables: the internal coordinate, s and a
i

thermal coordinate, :
i

i
i
i
i
i
i
i
nB z u z nB r ur nB u nB
t

si rs nB i r nB

(8.41)

DB

When the particle diameter is assumed as internal coordinate, the thermal coordinate is the
particles temperature. For the case of the particles mass considered as internal coordinate, the
particles specific enthalpy is taken as thermal coordinate. The introduction of one extra variable
i

will increase the models complexity. For this reason, the inner temperature of a particle will not be
considered as an additional inner state variable.
It is assumed that all the ice particles have the same temperature until they completely melt.
This temperature is the equilibrium temperature. The energy coming from the bulk is used only for
the warming of the surface and the melting. The ice particles do not warm up inside. Moreover, the
ice particles in the bulk are assumed as one pseudo-continuous phase. The mass conservation
equation is derived in this case from the population balance for the ice particles, by integrating over
all the particle sizes. The energy conservation equation can be written in a similar way with equation
(8.34):

D B cB hB
i

Dt

i
T , B

(8.42)

With the initial condition:

An initial enthalpy profile is assumed inside the bulk:
At

t0

hB
i

hB ,0 ( z , r , )
i

t 0

(8.43)

The boundary conditions:

a)

In axial direction

At

z0

b)

i
u z hB i
u h

z 0 z B z 0

In angular direction

Periodic continuity:

256

(8.44)

Appendix
hB

At

q , B

(8.45)
0

List of symbols
i
Q T , B = heat term due to difference in temperature between the ice particles and the bulk, W 3
m
i
hB = the specific enthalpy of the ice particles in the bulk, J
kg

c. The energy conservation equations for the frozen layer domain

The energy conservation equation in terms of q (8.32) written for the frozen layer:

D FL
hFL
i

Dt

FL

(8.46)

With the initial condition:

An initial enthalpy profile is assumed inside the bulk:
At

t0

hFL
i

hFL
,0 ( z , r , )
i

t 0

(8.47)

The boundary conditions:

a) In axial direction

At

z0

i TFL i
FL
Qin
z z 0

(8.48)

b) In radial direction

The energy at the interface goes entirely to the frozen layer:

At

r R FL

i TFL i
QT , B
FL

r r R

FL

(8.49)

At

rR

i TFL i
FL
QT , wall
r r R

(8.50)

List of symbols
i
FL
= the volume fraction occupied by the ice phase,

i
cFL
= concentration of the ice phase in the frozen layer, kg phase 3
m phase

257

Appendix
i
hFL
= the specific enthalpy of the ice phase in the frozen layer, J
kg

QT , B = energy due to a difference in temperature between the bulk and the other phases, W 3
m
i
Q T , FL = heat term due to difference in temperature between the frozen layer and the wall, W 3
m

FL
= the thermal conductivity of the ice phase inside the frozen layer, W
mK

d. The energy conservation equations for the freezer wall

The energy conservation for the freezers wall is written under pseudo steady state
assumption, leading to a flux expression.
r qr , wall
r

(8.51)

With the initial condition:

An initial enthalpy profile is assumed inside the bulk:
At

rR

qr , wall

rR

QT , wall

(8.52)

The boundary conditions:

At

r R wall

qr , wall

r R wall

QT , c

(8.53)

The heat flux is evaluated using Fouriers law of heat conduction:

qr , wall wall r

dTwall
dr

(8.54)

List of symbols
qr , wall = heat flux through the freezer wall, W 2
m

wall = the heat conductivity of the wall, W m K

e. The energy conservation equations for the coolant

The coolant is assumed to have a high heat transfer coefficient. Newtons law of cooling is
used to relate the normal heat flux component to the temperature difference between the freezer wall
and the coolant:
qcoolant U wall , coolant Twall , coolant Tcoolant

(8.55)

258

Appendix
List of symbols
U wall , coolant = heat transfer coefficient at the coolant side, W 2
m K
Twall , coolant = temperature of the freezer wall at the coolant side, K
Tcoolant = temperature of the coolant, K

5. Averaging procedure along the radial and angular coordinates

The following lemmas are introduced for performing the averaging along the radial and
angular coordinates:
a)

Averaging with respect to

For a function f r , , the averaging over the angular coordinate is performed in the
following way:
1

f r , d f r

(8.56)

List of symbols
f r , =function

f = angularly averaged function

b)

Averaging with respect to r

For a function f r , , the averaging over the radial coordinate is performed in the following
way:
R

2
r f r , dr f r
2
2
R1 R0 R0

(8.57)

List of symbols

f r = radially averaged function

R0 = lower bound of the radius, m
R1 = upper bound of the radius, m

c)

Averaging with respect to r and

259

Appendix
For a function f r , , the averaging over the radial and angular coordinate is performed in
the following way:

R12 R02 f r

R1 2

r f r , d dr

(8.58)

R0 0

When looking at the PBE-ice in equation (8.9) one can see:

(a) Terms in which there is no derivative operation with respect to the radial and angular
coordinates (e.g., in the derivative terms with respect to time, axial position and size and the
algebraic source and sink expressions). The averaging procedure over radial and angular
coordinates is essentially identical for all cases.
(b) Terms in which there is a derivative operation with respect to the radial coordinate.
(c) Terms in which there is a derivative operation with respect to the angular coordinate.
For each of the cases above, the averaging will be performed as follows:
(a) No derivative operation with respect to the radial and angular coordinates

In this case, equation (8.57) is applied.

R12 R02 T r

R1 2

r T r , d dr

(8.59)

R0 0

(b) Derivative operation with respect to the radial coordinate

In this case, the integration with respect to the radial coordinate is performed first, followed by
the averaging with respect to the angular coordinate. The resulting averaged term is:

2 R1 jR1 , R0 jR0 ,

R1 1 r j
0 R r r r dr d
0

(8.60)

Often, there is no transport at the boundary with the rotor ( r R0 ):

jR0 , 0

(8.61)

Hence, the averaged term:

2 R1 jR1 ,

R1 1 r j
0 R r r r dr d
0

(8.62)

List of symbols

260

Appendix
T = transfer term, unit 3
m s

j = flux of component, unit 2

m s

(c) Derivative operation with respect to the angular coordinate

In this case, the integration with respect to the angular coordinate is performed first:
2

j
2
d j r , 0 j r , 2 j r ,0

(8.63)

If periodic continuity is assumed, the two terms in equation (8.63) cancel each other:
2

j
d 0

(8.64)

6. Overview of coordinate-averaged mass conservation equations

a. The averaged conservation equations for the bulk domain

The averaging reduced the number of coordinates for the bulk domain to the following:
(1) Time:

0t

(2) Axial coordinate:

0 z L

(3) Crystal size coordinate:

0 s smax
min L, R R0
i

The resulting conservation equations inside the bulk become, written considering the average
terms:
_i

nB AB

j i A j i A

z B, z
B
B
s B,s

(8.65)

BB AB Rt, FL
B S B DB AB Rt , B FL S B
i

l l

B cB , j AB

j l A R l
S
R

z
m
B
j
z
B
t
B
FL
j
B
B

,
,
,
,
,
, j AB

a a

B cB AB

ja A 0
z
m
, B , z B
t

(8.66)

(8.67)

With the initial conditions:

261

Appendix
_
i

_
i

nB (0, z , s ) nB ,0 z , s

At t 0 ,

(8.68)

cB , j 0, z cB , j z

(8.69)

cB 0, z cB ,0 z

(8.70)

The boundary conditions:

a)

In axial direction

Continuity of flows:
i

jB , z AB

At z 0

FB t , s
i

(8.71)

z 0

jm , B , j , z AB

FB t , s
i

(8.72)

z 0

a
FB ,0 t
jm , B , z AB

z 0

b)

(8.73)

Along the internal coordinate

The boundary condition for the crystal size is put at the upper range of ice crystal size due to

the size reduction by the melting of the particles

_
i

nB t , z , smax
nB ,si t , z

At s smax
i

(8.74)

max

For the following sections, the averaged terms will be denoted without the dash.
b. The averaged conservation equation for the frozen layer

The averaging reduced the number of coordinates for the frozen layer domain to the
following:
0t

(1) Time:

The resulting conservation equations inside the frozen layer become, written considering the
average terms is a differential equation showing the variation of the thickness of the frozen layer:
FL t , z
t

Rt, m , B FL t , z , s Rt, B FL , w t , z
l

(8.75)

cFL
t, z
i

With the initial condition:

An initial concentration of ice is considered inside the frozen layer:

262

Appendix
At

t 0

FL FL ,0 z

(8.76)

At

t ts

FL FL ,0 z

(8.77)

7. Coordinate-averaged conservation equation for dispersed phases

The starting point is the equation of change for dispersed phases in equation (8.9). This
equation is multiplied with r and integrated from r R0 to r R FL and 0 to 2 . For
each term of equation (8.9), the averaging procedures presented in Appendix 5 are then applied.

nB

The accumulation term

a)
2 R1

r dr d

nB

n
t

0 R0

2 R1

t

B

R0

r dr d

(8.78)

The upper bound is defined as:

R1 R FL

(8.79)

Applying the lemma (8.57):

2 R

1
nB r dr d nB
t 0 R0
t

FL

R02

(8.80)

It is defined that:
N B

z, s nB z, s R FL

R02 AB nB

(8.81)

The cross-sectional area of the bulk domain is calculated using:

AB R FL R02
2

(8.82)

Taking into account the definitions above, equation (8.78) is rewritten in the following way:
2 R1

nB
t

0 R0

b)

nB AB

N B

r dr d
t
t

(8.83)

The convective term in axial direction z u z nB

263

Appendix
2 R1

u z nB

0 R0

2 R1

r dr d

2 R1

r dr d

jB , z

0 R0

(8.84)

r dr d

B, z

0 R0

When applying the lemma (8.57):

jB , z r dr d

z 0 R0
2 R1

z j R

B, z

FL

R02

(8.85)

The radially and angularly averaged flux of particles should be written in the following form:

jB , z u z nB

(8.86)

In this case, two possibilities are considered:

The radially averaged number of particles, nB ,1 , is defined in the following way:

i)

nB ,1

1 1
nB r r dr
AB R0

(8.87)

An averaged velocity is defined as:

u z ,1

jB , z

(8.88)

nB ,1

List of symbols

N B = number of dispersed particles, #

m

jB , z = radially and angularly averaged flux of particles, # 2

m s

nB ,1 = radially averaged number of particles, # 3

m

u z ,1 = radially averaged velocity, m

s

ii)

The radially averaged velocity, u z ,2 , is defined in the following way:

264

Appendix
R

1 1
u z r r dr
AB R0

u z ,2

(8.89)

It is defined that:

jB , z

nB ,2

(8.90)

u z ,2

List of symbols

nB ,2 = radially averaged number of particles, # 3

m

u z ,2 = radially averaged velocity, m

s

Note: The radially averaged flux of particles in axial direction, jB , z , cannot be calculated using :

jB ,z nB ,1 u z ,2 nB ,2 u z ,1

(8.91)

The convective term in radial direction r ur nB

c)

2 R1
1 r u r nB
1 r jB , r
r dr d
r dr d
0 R r
r
r
0 R0 r
0

2 R1

(8.92)

When applying equation ((8.60):

R ,

1 r jB , r


0 R r r r dr d 2 R FL jB , R FL , R0 jB , R0 ,
0

2 R1

0
1 r jB , r

0 R r
r
0

2 R1

r dr d S

j
S
j
0
B
B , R FL ,
B , R0 ,

(8.93)

(8.94)

The circumference of the freezer is determined in the following way:

S B 2 R FL

(8.95)

S0 2 R0

(8.96)

265

Appendix

The convective term in angular direction u nB

d)

2 R
1 u nB
1 jB ,

dr

0 R r
0 R r r dr d

0
0

2 R

(8.97)

There are complex and fast exchanges between flows in the ice cream in radial and angular
directions. This happens due to the rapid rotations of the scraper. These exchanges are assumed to
lead to strong mixing in a cross-sectional plane, perpendicular to the axial transport direction. This
results into a uniform process condition within this plane. Therefore, the averaging is done over both
directions simultaneously, while accounting for the exchanges with the frozen layer and the no-flux
condition at the inner surface with the rotator.
For the convective term in radial and angular direction, the following simplification will be
used:

2 R
1 r ur nB
1 u nB

dr

r dr d

0 R r
0 R r
r

0
0

2 R

(8.98)

Rt, FL B S B Rt, B FL S B

The generation term si rs nB

e)

rs nB

2 R

0 R0

2 R

B,s

r dr d

2 R

nB r dr d

0 R0

(8.99)

r dr d

0 R0

When applying the lemma (8.57):

2 R1

jB , s r dr d

0 R0

,s
B

FL

R02

(8.100)

The radially and angularly averaged flux of particles along the internal coordinate should be
written in the following form:

jB , s rs nB

(8.101)

266

Appendix
List of symbols

rs = radially averaged growth rate, m or kg

s
s

The radially averaged number of particles nB is defined as:

1 1
nB r r dr
AB R0

nB

(8.102)

It is defined that:

jB , s

rs

(8.103)

nB

Hence, the convective term in the axial direction is calculated as follows:

2 R1

rs nB

0 R0

jB , s AB

r dr d
z

(8.104)

The transfer terms BB and DB

f)

Applying the lemma (8.57):

2 R1

r dr d RB AB

(8.105)

0 R0

RB BB , DB

(8.106)

Taking into account these changes, the conservation equation for the dispersed phase in the
bulk is rewritten using the radially and angularly averaged terms in the following way:
_

nB AB

j A j A

i B , s

z
B
z
B
B
,
s

(8.107)

BB AB Rt, FL B S B DB AB Rt, B FL S B

Where the fluxes are calculated using:

267

Appendix

jB , z u z nB

(8.108)

jB , s rs nB

(8.109)

This is the outcome of the averaging procedure for the dispersed phase in the bulk domain.

8. Coordinate-averaged conservation equation for the pseudocontinuous phase

The starting point is the equation of change for mass of the pseudo-continuous phase in the
bulk domain in equation (8.15). This equation is multiplied with r and integrated from r R0 to
r R FL and 0 to 2 . For each term of equation (8.15), the averaging procedures

presented in Appendix 5 are then applied.

a)
2 R1

B cB , j

The accumulation term

B cB , j

0 R0

r dr d

t
2 R1

t
0

R0

cB , j r dr d

(8.110)

Applying the lemma (8.57):

2 R

1
B cB , j r dr d B cB , j
t 0 R0
t

u z B cB , j

0 R0

FL

R02

The convective term in axial direction z u z B cB , j

b)
2 R1

2 R1

r dr d

2 R1

B jm , B , z , j

0 R0

(8.111)

r dr d

(8.112)

jm , B , z , j r dr d

0 R0

jm , B , z , j u z cB , j

(8.113)

When applying the lemma (8.57):

2 R

1
B jm, B , z , j r dr d B jm , B , z , j
z 0 R0
z

268

FL

R02

(8.114)

Appendix
The radially and angularly averaged flux of component j should be written in the following
form:

jm , B , z , j u z cB , j

(8.115)

The radially averaged concentration for the component j in the bulk, cB , j is defined as:

i)

1 1
cB , j r r dr
AB R0

cB , j ,1

(8.116)

The velocity is defined as:

jm , B , z , j

u z ,1

(8.117)

cB , j ,1

List of symbols

jm , B , z , j = radially and angularly averaged flux of component j , kg 2

m s
_

cB , j ,1 = radially averaged concentration for component j , kg 3

m

u z ,1 = radially averaged velocity, m

s

The radially averaged velocity, u z ,2 is defined in the following way:

ii)

u z ,2

u z r r dr
AB R0

(8.118)

The concentration is defined as:

_

cB , j ,2

jm , B , z , j

(8.119)

u z ,2

List of symbols
_

cB , j ,2 = radially averaged concentration for component j , kg 3

m

269

Appendix

u z ,2 = radially averaged velocity, m

s

Note: The radially averaged mass flux of component j in axial direction, jm , B , z , j cannot be

calculated using:

jm , B , z , j cB , j ,1 u z ,2 cB , j ,2 u z ,1

(8.120)

The convective term in radial direction r ur B cB , j

c)

1 r u r B cB , j

r dr d
0 R r
r
0

2 R1

(8.121)

1 r B jm , B , r , j

r dr d
0 R r
r
0

2 R1

When applying equation ((8.60):

1 r B jm , B , r , j

r dr d
0 R r
r
0

2 R1

(8.122)

2 R FL B jm, B , R FL , , j R0 B jm , B , R0 , , j

1 r B jm , B , r , j

r dr d
0 R r
r
0

2 R1

(8.123)

S j

j
B B m , B , R FL , , j 0 B m, B , R0 , , j

The convective term in angular direction u B cB , j

d)

2 R1
1 u B cB , j
1 B jm , B , , j
r dr d
r dr d
0 R r

0 R0 r
0

2 R1

(8.124)

In a similar way with the ice phase, the averaging in both radial and angular direction for the
liquid phase is done simultaneously. The exchanges with the frozen layer and the no-flux condition
at the inner surface with the rotator are taken into account.

270

Appendix
For the convective terms in radial and angular direction the following simplification will be
used:

1 r u r B cB , j

r dr d Rt, B FL , j S B
0 R r
r
0

2 R1

(8.125)

The generation term, RB , j

e)

Applying the lemma (8.57):

2 R1

RB , j r dr d RB , j AB

(8.126)

0 R0

Taking into account the changes, the conservation equation for the pseudo-continuous phase
in the bulk will be:

l
l
AB B cB , j
_

u l l cl A R l

S
R

z z
B
B, j
B
t , B FL , j
B
B , j AB
t

(8.127)

This is the outcome of the averaging procedure for the pseudo-continuous phase in the bulk
domain.

9. Averaging of the conservation equation for the ice phase in the

frozen layer
The starting point is the equation of change for mass of the ice phase in the frozen layer domain in
equation (8.21). Since the frozen layer contains only ice, the following procedure is proposed for
averaging out r and :
The mass of the frozen layer per unit length of freezer is calculated in the following way:
i
FL

FL R 2 R FL t , z

ci
FL

(8.128)

Hence, the concentration of ice inside the frozen layer is determined from the following equation:

cFL

mFL
FL
i

(8.129)

2
R 2 R FL t , z

271

Appendix
i
mFL
FL
.
The first term of equation (8.21) is rewritten as
t

It is assumed there is no movement along the axial coordinate:

j

i
m , FL , z

(8.130)

The boundary of the frozen layer is moving along the radial coordinate.

The convective term in radial direction r ur cFL

1 r ur cFL

r dr d
0 R r
r
FL

(8.131)

2 R u R cFL R FL u R FL cFL

The velocity at the freezer wall is zero:

uR 0
i

(8.132)
When applying equation (8.60):

1 r ur cFL
i
r dr d 2 R FL Rt, m , B FL
0 R r
r
FL

(8.133)

The convective term in radial direction jm ,FL , :

R
1 u cFL
i
i

r dr d u cFL
0 R r

R FL
FL

u cFL
i

dr

(8.134)

The variation in time of the mass of the frozen layer:

mFL
FL
i
l

S B Rt, m , B FL S B Rt, B FL , w 0
t
i

(8.135)

Or

2
i

R 2 R FL t , z cFL

i
l
S B Rt, m , B FL S B Rt, B FL , w
t

272

(8.136)

Appendix

is constant.
The concentration of ice inside the frozen layer cFL
i

But:

2 cFL

R FL
t

S B Rt, m , B FL S B Rt, B FL , w
i

(8.137)

From which results the conservation equation for the ice phase inside the frozen layer:
FL t , z
t

Rt, m , B FL t , z , s Rt, B FL , w t , z
l

(8.138)

cFL
t, z
i

i
mFL
= the mass of the frozen layer, kg
m

With the initial condition:

An initial concentration of ice is considered inside the frozen layer:
At

t0

FL FL ,0 z

(8.139)

At

t ts

FL FL ,0 z

(8.140)

10. Expression of the volume flow rate

Under the assumptions of incompressible fluid, steady state laminar flow, the z -component of
the equation of motion in cylindrical coordinates will be:
1 r B , rz
P
i l a

gz
B
z
r
r

(8.141)

This equation is valid over the entire annular region for any kind of fluid.
If

P
is assumed constant:
z

P P0 PL

(8.142)

Equation (8.141) becomes:

1 d r B , rz P
i l a

B
gz
r
dr
L

(8.143)

(Fredrickson & Bird, 1958) integrate equation (8.143) to:

273

Appendix
B , rz

2
F CI R
r

2
r

(8.144)

Where:
P
i l a
B
gz
L

(8.145)

List of symbols
F = the sum of forces per unit volume

CI = the constant of integration

R = the outer radius of the annulus

The radial distance CI R represents the position at which B , rz 0 .

Equation (8.144) is the starting point for the derivation of an expression relating the flow rate
to the pressure drop for laminar flow in an annulus.
For the power law model, the local shear stress depends on the local shear rate by:

B , rz K Bi l a

n 1

u z
r

u z
r

(8.146)

Equations (8.144) and (8.146) are combined and integrated to give the velocity distribution:
1

F R n CI2
n

d
uz R

i
l
a

2 K

0
B
1

for

0 CI

(8.147)

for

CI 1

(8.148)

F R n 1
CI2 n

uz R

d
2 K i l a

B
Where:

r
R

(8.149)

R0
R

(8.150)

List of symbols
= the dimensionless radial coordinate,
R0 = the inner radius of the annulus, m

274

Appendix
The boundary conditions:
At 0

uz 0

(8.151)

At 1

uz 0

(8.152)

Both equations (8.147) and (8.148) must give the same value of the velocity at CI . The
constant of integration is determined as a function of and

1
from the equality of the two
n

expressions for the velocity in equations (8.147) and (8.148):

1

CI

1
C2
n

C2 n
I d C I d
I
0

(8.153)

This is the determining equation for CI , which is a function of 0 and

1
.
n

The volume rate of flow is determined by integrating the velocity distribution over the annular
region:
1

v,iB l a 2 R 2 u z d

(8.154)

Knowing CI , equation (8.154) is easily integrated by Taylor series expansion to determine the
volume rate of flow.
Equations (8.147) and (8.148) are substituted in equation (8.154) and one integration step is
performed to give:
1

i l a

v , B

F R n 1 2
R
C 2
2 K i l a I
B

0
3

2
I

1
1
n

n d

1
1
n

n d

(8.155)

1
2 n 1

n
CI2 1 n 0 n CI2 2
2n 1

2 n 1
n

(8.156)

Replacing in equation (8.154):

1

i l a

v , B

F R n

n
R

C 2 2
2 K i l a 2n 1 I

2 n 1
n

1
n

CI2 0 2

2 n 1
n

The expression of the flow is easily calculated once CI has been determined.

275

(8.157)

Appendix

11. Physical properties for the simulation of the ice cream freezing
model
Table 8.1. Overview of physical properties variables for the ice cream freezing model
Name

Symbol
Ice

Unit

Value
916.7

kg
m3

Water

kg
3
m

999.8

Sugar

kg
3
m

1600

Fat

kg
m3

900

Other

kg
3
m

1996.6

Ice

m K

2.2

Water

m K

0.6

Bulk

m K

0.5

Wall

wall

m K

70

coolant

W 2

m K

12000

Ice

C pi

kg K

2000

Water

C pw

kg K

4200

Sugar

C ps

kg K

1100

Fat

C p f

kg K

2100

Other

C po

kg K

1160

Air

C pa

kg K

1000

Ice in the B

kSi

m
s

106

kTi,0

m
s

k Bl,w

m
s

104

Latent heat of crystallization

hcryst

J
kg

334 103

Standard melting temperature of

water

Tm

273.15

Density

Thermal
conductivity

Heat transfer coefficient of

coolant

Specific heat

Mass transfer
coefficients

Ice in the B due

to temperature
difference
between B & FL
Water from B to
FL

276

Appendix

12. The shrinking core model for the melting of an ice particle
The ice particle is submerged in a fluid with a higher temperature. The ice is surrounded by a
stagnant (film) layer, while the rest of the fluid is turbulent.
From the pseudo-steady state energy balance around a spherical particle, the following result
is obtained:
2 T
r
0
r r

(8.158)

The boundary conditions for equation (8.158) are defined as:

At the interface between the particle and the fluid, r s
i

a)
T

T
i

r si

At the interface between the fluid film and the fluid, r s

i

b)
T

(8.159)

r si

(8.160)

By integration of equation (8.158) the temperature profile around a particle are obtained:

1 1 s

f
T T s r
i

T T
i

s
i

(8.161)

The heat flow towards the surface of an ice particle is calculated as:

q a i q''

r s i

2
dT
i
f
4 s
dr

i
r s

(8.162)

After differentiating equation (8.162) and substituting the result in equation (8.161), the
following expression is obtained for the heat flow towards a single ice particle:

q 4

s
i

T T
i

(8.163)

It is common in modelling of crystallization processes to use a size dependent film layer:

si

(8.164)

When applying this size dependent layer the following equation is obtained:

q 8 f s i T T i

(8.165)

277

Appendix
The melting rate for a single particle, determined under the assumption that all the heat is
directed towards the particle meltdown, is determined from:
rs
i

(8.166)

hcryst

List of symbols
T r = the temperature, K
r = the radial coordinate for a particle, m
T = the temperature of the ice particle, K
i

T = the temperature of the fluid, K

f

= the thickness of the fluid film, m

q = heat flow towards a single particle, W
i
a = surface area of a spherical particle, m 2

f = thermal conductivity of the fluid, W m K

i
rs = melting rate for a single particle, kg
s

hcryst = enthalpy of crystallization, J

kg

13. Mass conservation equations for the bulk domain expressed per unit
length of freezer
Considering the definitions:
_
i

N B nB AB
i

(8.167)

B , j B cB , j AB
l

(8.168)

B B cB AB
a

(8.169)

The equations of change are written in the following form:

278

Appendix

u N

N B
i

B , j
l

i
B

s i

_
_
_

i i
i
i
i
i
rs N B BB AB Rt , FL B S B DB AB Rt , B FL S B

u R
z

i
z

_
l
z

l
B, j

l
t , B FL , j

S B RB ,j AB

(8.170)

u 0

B
a

_
a

a
B

With the initial conditions:

N B (0, z , s ) N B ,0 z , s
i

B , j 0, z B , j ,0 z

At t 0

(8.171)

B 0, z B ,0 z
a

The boundary conditions:

a)

In axial direction

Continuity of fluxes:
~
i i
i
i
FB t , s
u z N B

z 0

At z 0

_ l l
l
FB , j ,0 t
u z B , j

z 0r

(8.172)

_a a
a
FB ,0 t
u z B

z 0
b)

Along the internal coordinate

The boundary condition for the crystal size is put at the upper range of ice crystal size due to

the size reduction by the melting of the particles

At s smax
i

N B t , z , smax N B ,smax t , z

(8.173)

N B ,smax t , z 0

(8.174)

List of symbols
i
N B ,smax = nucleation rate, #
m
l
B , j = concentration of component j in the liquid phase in the bulk, kg
m
a
B = concentration of air in the bulk, kg
m

279

Appendix

14. Conservation property of the overall amount of water

Water is present as:
(1) liquid water in the bulk domain
(2) ice particles in the bulk domain
(3) solid ice in the frozen layer
The total amount of water does not change when there is no chemical reaction involving
water.
FB t , z FB , w t , z FB t ,0 FB , w t ,0 FB ,0 t
i

(8.175)

The amount of water (in solid and liquid phase) should be conserved if it is assumed there is
no accumulation in the frozen layer. For the following, the conservation equations will be used in the
form expressed per unit length of freezer, as defined in Appendix 13.
The mass of an ice particle is considered as an internal coordinate, s m .
i

The conservation equations are written in the following form:

u N
i

NB

B ,w
l

i
B

u
z

i
z

_
l
z

l
B,w

s i

_
_
_

i i
i
i
i
i
rs N B BB AB Rt , FL B S B DB AB Rt , B FL S B

l
l
l
Rt , FL B , w S B Rt , B FL , w S B RB , w AB

(8.176)

(8.177)

The concentration of ice particles is determined in the following way:

i dm i
dm i
i
i
i
i
i
i
B N B dm N B i ds i N B ds

ds
ds
0
0
0

(8.178)

The following relation is introduced:

dm
i

ds

1
1
#

(8.179)

The integration of equation (8.176) for the s - coordinate, term by term:

ds
N ds dm

(8.180)

i
i i
i
i
i ds
u
N
ds
u

B
0 z B
i
z
dm

(8.181)

i
B

i
B

ds r
i

i
i
i
i
N B ds rs B

(8.182)

280

Appendix

rs B 0

For s smax
i

(8.183)

It is assumed that particles with the size s 0 cannot appear, hence:

i

rs

si 0

(8.184)

The rate of change of ice in the bulk should be equal with the rate of change of the water in
the liquid phase in the bulk:

i
B

B
0

d i i
i
l
rs N B ds RB , w 0
i
ds
0

ds Dm ,B ds
i

(8.185)

The frozen layer it is assumed to have a constant hold-up at each location. The particle fluxes
in and out the frozen layer must be equal:

i
t , FL B

ds Rt, B FL ds Rt, FL
B , w Rt , B FL , w 0

(8.186)

By adding the equations (8.176) and (8.177):

i
i ds i
i
ds
l
i
l
l
B i B ,w u z B i u z B ,w
dm
dm

0
t
z

(8.187)

After taking into account equation (8.175):

B B ,w
i

i
i
l
l
u z B u z B ,w
0

z

(8.188)

Assuming steady state and integrating for the length of the freezer:

i
l
l
u
u

B
z B,w
0
dz 0

i
z

(8.189)

After the integration:

u z B

u z B

t,z

u z B ,w

t ,0

u z B ,w

t,z

(8.190)

t ,0

Considering the definitions of the flows:

i
z

u B

FB ,0

(8.191)

t ,0

u z B ,w

FB , w ,0

(8.192)

t ,0

u z B

FB

(8.193)

t,z

281

Appendix

u z B ,w

FB , w

(8.194)

t,z

Equation (8.175) is obtained.

15. Singular vectors for the SVD on the model parameters

15.1.

Operational parameters

Table 8.2. Singular vectors for SVD on the operational parameters of the model
Input

Output

u1

v1

P0

-2.2e-15

-0.44232

ic
Tout

TB ,0

0.086819

0.506063

ic
xout
,w

-0.26831

ic
xout
,i

-0.02409

i
d 4,3

-0.0303

i
d3,2

-0.6749

-0.13886

Qtot

Tc

0.552294

Fin ic

0.573117

15.2.

Scaled singular vectors

0.599142

Input parameters

Table 8.3. Singular vectors for SVD on the input parameters of the model
Input

Scaled singular vectors

Output

u1

u2

v1

v2

xBl,w,0

0.700519

0.324062

-0.41762

-0.36569

ic
Tout

xBl,s ,0

0.072802

0.188483

0.38732

-0.23308

ic
xout
,i

xBl,f ,0

0. 604682

0.098289

-0.16648

0.837098

ic
xout
,w

xBl,o ,0

0.000188

0.00049

0.036041

-0.07859

i
d 4,3

Ovr

-0.02235

-0.39619

0.043712

-0.09815

i
d3,2

Ms

0.371258

-0.83236

-0.75834

-0.19433

0.263771

-0.24008

Qtot

282

Appendix

15.3.

Freezer design parameters

Table 8.4. Singular vectors for SVD on the freezer design parameters of the model
Input

15.4.

Scaled singular vectors

Output

u1

u2

v1

v2

-0.08129

-0.60258

0.425611

-0.47097

ic
Tout

0.986003

0.057741

-0.41509

0.467772

ic
xout
,w

D0

0.073692

-0.64664

0.220077

-0.24801

ic
xout
,i

0.013769

0.009756

i
d 4,3

wall

0.088767

-0.3282

0.017164

0.011687

i
d3,2

-0.74512

-0.66639

blade

0.088767

-0.3282

0.206008

-0.23142

Qtot

Empirical parameters

Table 8.5. Singular vectors for SVD on the empirical parameters of the model
Input

Scaled singular vectors

Output

u1

u2

v1

v2

Ca

0.46419

-0.04306

-0.42951

-0.00524

ic
Tout

Cb

0.494268

-0.07836

Cc

0.000224

5.44e-06

0.438453

0.000994

ic
xout
,w

-0.04759

0.008085

-0.00041

-0.00043

-0.23251

-0.00053

ic
xout
,i

c1

0.231014

-0.02116

c2

0.464195

-0.04309

-0.01294

0.622092

i
d 4,3

0.026641

0.553067

0.106593

0.43135

-0.01671

0.782451

i
d3,2

00.494268

-0.07836

-0.72244

-0.02683

0.026641

0.553067

-0.21627

0.00492

Qtot

FL , min

0.109704

0.430891

16. References
Bird, B.R., Stewart, W.E., Lightfoot, E.N. (2002). Transport phenomena, John Wiley & Sons
Fredrickson, A.G., Bird, R. B. (1958). Non-Newtonian Flow in Annuli, Industrial and Engineering
Chemistry 50, 347

283

Appendix
Hoffman, J. D. (2001). Numerical methods for engineers and scientists, 2nd Edition, Marcel Dekker,
Inc., New York

284

Curriculum Vitae

Curriculum Vitae
Bogdan Dorneanu was born in Medgidia, Romania, on the
st

21 of July 1981.
In 2000 he completed his high-school studies and
graduated from the Nicolae Blcescu Highschool (Medgidia)
majoring in Mathematics and Physics. In the same year he started
his university education at University Politehnica of Bucharest,
Romania and in 2005 he received a degree in Chemical
Engineering.
During his university studies he spent a 3-month internship at the Laboratoire de Gnie des
Procds des Solides Diviss, at Ecole des Mines dAlbi-Carmaux, France, working on the
experimental and modelling of the pyrolysis of cellulose.
In January 2006 he was appointed as PhD student for the Product and Process Engineering
group at DelftChemTech, Faculty of Applied Sciences, at Delft University of Technology under the
supervision of Prof. Johan Grievink and Dr. Costin Sorin Bildea. His research was carried on under a
Marie Curie fellowship in the framework of the PRISM project.
During the PhD period, Bogdan attended several advanced courses on Plantwide Control,
Separation Processes, Optimization and Control, Product and Process Design and Modelling,
Computational Fluid Dynamics and Process Intesification in United Kingdom, Czech Republic,
Spain, Denmark and The Netherlands. He also spent three months as a visiting researcher at the Joef
Stefan Institute, in Ljubljana, Slovenia, working on the improvement of a vinyl acetate
polymerization reactor model.
The results of Bogdans research have been preesented locally and internationally in CACEsymposia and congresses (NPS6 10, ESCAPE 17 21, NL-GUTS). For the last four years, he was
additionally involved in the supervision of undergraduate students during their graduations projects.
Presently he works as a Post Doc researcher at the Joint Research Centre of the European
Commission, at the Institute for the Protection and Security of the Citizen, the Security Technology
Assessment Unit, in Ispra, Italy, on the modelling of critical infrastructures failures.

285

Publications

Publications
Journals and conference proceedings
1.

Dorneanu, B., Bildea, C.S., Grievink, J. On the application of model reduction to plantwide
control, Computer Aided Chemical Engineering, Vol. 24, 2007, 823

2.

Dorneanu, B., Bildea, C.S., Grievink, J. Model reduction techniques for dynamic optimization of
chemical plants operation, Computer Aided Chemical Engineering, Vol. 25, 2008, 337

3.

Dorneanu, B., Bildea, C.S., Grievink, J. On the application of model reduction to plantwide
control, Computers and Chemical Engineering, Vol. 33 (3), 2009, 699

4.

Dorneanu, B., Bildea, C.S., Grievink, J., Bongers, P.M. A first-principles model for the freezing
step in ice cream manufacture, Computer-Aided Chemical Engineering, Vol. 26, 2009, 171

5.

Dorneanu, B., Bildea, C.S., Grievink, J., Bongers, P.M. A reduced model for the freezing step in
ice cream manufacture, Computer-Aided Chemical Engineering, Vol. 28, 2010, 265

6.

Dorneanu, B., Bildea, C.S., Grievink, J. Process modeling and model reduction for chemical
engineering applications, 21st European Symposium on Computer-Aided Process Engineering,
Greece, May 2011

286

Acknowledgement

Acknowledgement
This thesis is the result of my research during the period of 2006 to 2010 at the Delft
University of Technology, on knowledge-based tools for process systems engineering. At this time, I
wish to express my gratitude to all the people who have contributed to this thesis, although
unfortunately I am able to mention just a few.
Firstly, I would like to thank my two supervisors, Johan Grievink, and Sorin Bildea for their
support during this research project. Your support, advice and encouragement ensured the success of
the work and the completion of this book. Although you often asked me to be careful not to be pushed
in two opposite directions due to your different ways of thinking, I have realized that actually the path
was somewhere in the middle Now, every Tuesday afternoon it will be somewhat empty without
our discussions and the green tea.
Secondly, I would like to thank all my colleagues and students at the Product and Process
Engineering Group. My first (and part-time) office colleagues Marco and Peter: your presence was
often greatly missed. It follows a long list of people that helped creating a special (and more social)
atmosphere in the Proeffabriek: David, Malte, Renske, David, Nasim, Pietro, Bashar, Ivan, as well as
the members of the staff: Mark, Mojgan, Caroline, Ruud, Henk. Listing all the memories of the
coffee breaks, barbecues, Sinterklaas or Christmas parties would require another thesis, not to
mention the conferences, courses or workshops. I look happily back at the days at the office and I
thank you all!
Un mulumesc Olgi, care s-a lsat convins s aplice la Delft pentru a ajuta la creterea
numrului de romni din Olanda. Ajut enorm cnd e cineva cu care te nelegi ntr-un mediu nou.
Discuiile de pe mess, telefon i prnz au fost mereu extreme de interesante. Ca s nu mai vorbim
de micul dejun de la Scallys, duminca la 10:00.
Apoi, un mulumesc prinilor mei. Cei care au fost permanent alturi de mine i care, n
acelai timp, mi-au permis s mi gsesc drumul de unul singur. V mulumesc!
Finally, Jo, thank you for being next to me and for your understanding during the last year of
my thesis work!

Bogdan

289

290