Computers and Chemical Engineering 33 (2009) 822837

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

An ontological conceptualization of multiscale models

Aidong Yang , Wolfgang Marquardt
Lehrstuhl fr Prozesstechnik, RWTH Aachen, 52064 Aachen, Germany

a r t i c l e

i n f o

Article history:
Received 8 April 2008
Received in revised form
20 November 2008
Accepted 26 November 2008
Available online 3 December 2008
Keywords:
Multiscale modeling
Conceptualization framework
Ontology
Systems
Classication

a b s t r a c t
Multiscale modeling has recently been widely used in various elds of science and engineering. A few
research efforts have attempted to generalize existing applications mainly by means of classifying multiscale modeling paradigms, to reach a general understanding of multiscale modeling. However, these
existing efforts have been largely inductive and abstracting from the applications the researchers were
familiar with. This has inevitably caused inconsistencies within the results of different researchers. Further, existing work has not provided a precise denition for the fundamental concepts such as scale, which
arguably has been the main source of confusion. In this work, we attempt to address this issue through
developing a coherent conceptualization framework. Instead of starting with existing applications, our
work is built upon a conceptualization of general systems, from which precise denitions of scale and
inter-scale relations are derived. This is further followed by a precise specication of the composition of a
multiscale model and subsequently a classication of multiscale models, all through rigorous denitions.
This conceptualization framework is represented in form of an ontology by means of a set-theoretic language. It reuses part of an existing ontology (the BWW ontology). The key concepts in this framework are
illustrated through two modeling examples judiciously chosen from the eld of chemical engineering. In
comparison with existing results, the key contribution of this framework is not a classication but rather
a generic conceptual foundation for a better understanding of multiscale modeling, which is intended
to be sharable by researchers from different backgrounds. On the other hand, it is demonstrated that a
rigorously dened classication can be derived from this conceptual foundation. It is also shown that
the classications of multiscale modeling proposed in other works can be interpreted by means of the
suggested conceptualization framework in a precise manner.
2008 Elsevier Ltd. All rights reserved.

1. Introduction
In general, a mathematical model of a system is a collection of
relations (usually in the form of mathematical equations) among
variables representing certain properties of the system measured
in particular domains. By solving a model, one can predict values of
some properties of the system given the values of some other properties, or predict distributions of the values of some properties in
particular domains, given certain boundary conditions. To meet the
demands from science and engineering on the depth of revealing
the characteristics of a system, a mathematical model may involve
relations of properties at different scales of the system. For example, in order to predict certain properties of a material, both the
continuum scale as well as the molecular scale might have to be
considered.

Corresponding author at: Current address: Faculty of Engineering and Physical

Sciences (J2), University of Surrey, Guildford GU2 7XH, UK. Tel.: +44 1483 686577;
fax: +44 1483 686581.
E-mail address: a.yang@surrey.ac.uk (A. Yang).
0098-1354/$ see front matter 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2008.11.015

In recent years, multiscale modeling has become a popular topic

in different elds; the targets being modeled range from engineering systems (e.g. chemical reactors) to materials and biological
systems, to name the frequently reported ones. This status is well
reected by several review papers, namely two from the chemical
engineering community by Pantelides (2001) and more recently by
Vlachos (2005) and one from the applied mathematics community
by Givon, Kupferman, and Stuart (2004). A large number of applications from various domains have been referenced in these reviews.
Additionally, there have been special journal issues dedicated to
the topic of multiscale modeling, such as Issues 8 and 9 in Volume
59 of Chemical Engineering Science. The rapidly growing interests
in multiscale modeling have even led to the creation of two new
journals in 2003, namely Multiscale Modeling and Simulation (by
SIAM) and the International Journal on Multiscale Computational
Engineering (by Begell House Publishers). The former was largely
initiated by the applied mathematics and the latter by engineering
science communities.
Accompanying the widespread applications, there have been
several generalization efforts towards a better understanding of
and improvements to the multiscale modeling paradigms in general. For example, E and Engquist (2003a, 2003b) present, from a

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

Nomenclature
aA
a
/A
AB
AB
AB
AB
AB
AB
c=a+b
C: = D
f:A B
P
PQ
S
xy
x 
/ y

x,y

individual a belongs to set A

individual a does not belong to set A
the Cartesian product of sets A and B
the set of objects in set A but not in set B
the set of objects in A and in B
the set of objects in A or in B
set A is part of set B
set A is equal to set B or part of set B
c is the juxtaposition of a and b
C is dened as D
function f maps set A to set B
the set of all properties
the logical and between statements P and Q
the set of all substantial individuals
thing x is part of thing y
thing x is not part of thing y
the ordered pair of x and y

Greek letters

the set of all things

the empty set

mathematical perspective, a so-called heterogeneous multiscale

method which generalizes a variety of applications of multiscale
modeling and computation. They also formulate a general procedure of applying a heterogeneous multiscale method and analyze
the stability and accuracy of numerical solutions following such
a method. Within the chemical engineering community, generalization is pursued through several attempts on classifying existing
multiscale modeling methods. Pantelides (2001) presents a classication on scale integration strategies. The notions introduced
were applied by Bezzo, Macchietto, and Pantelides (2004) in presenting a general formulation of hybrid multizonal/CFD models of
chemical process equipment. The classication was later expanded
by Ingram, Cameron, and Hangos (2004) who identify a number of integrating frameworks and discuss the compatibility issue
in applying them. More recent work on classication has been
reported during the preparation of this manuscript, e.g. by Li, Ge,
Zhang, and Kwauk (2005) and by Vlachos (2005).
The contributions cited above are all generalization-oriented
and have set a starting point towards a sound theory of multiscale modeling, which is important for making methodological
improvements and for fostering new applications. However, these
contributions are largely result of inductive summary on the applications in one or several elds that their authors are familiar
with. Consequently, it has been inevitable that these proposed
frameworks and classications appear to be overlapping and
inconsistent. Further, such inductive studies can hardly touch on
the fundamental questions such as what a scale exactly means.
We would argue that answers to such questions are essential
for resolving inconsistencies with respect to the understanding, classication, and representation of multiscale modeling
paradigms.
Recognizing the limitations of the existing studies, we propose to take a different approach to improve our understanding
of multiscale modeling. Instead of directly abstracting a variety of
multiscale modeling applications, we rstly direct our attention
onto the conceptualization of general systems. This is reasonable
because the target of a multiscale modeling task, whichever application or eld it is originating from, will always be some kind of
system in the sense of general systems theory as introduced by
Bertalanffy (1968). Secondly, we derive, from the basic concepts

823

of systems, rigorous denitions of notions of interest in multiscale modeling, including scale and inter-scale relationships. As the
result of these two steps, a generic conceptual foundation independent of concrete applications and elds is rendered to support the
understanding of various multiscale modeling paradigms. Thirdly
and nally, we make use of this conceptual foundation and look
into the compositions of multiscale models in terms of scales and
inter-scale relations. The differences in multiscale models naturally
lead to different model classes.
In contrast to existing work, such an approach to the understanding of multiscale modeling appears to be more deductive
than inductive. However, we have to admit that our conceptualization framework is still inuenced and perhaps limited by existing
multiscale modeling applications which we are familiar with, particularly those from our own eld, namely Chemical Engineering.
Thus, this contribution should not be regarded as one that offers a
complete and universally valid classication of multiscale modeling paradigms. Instead, our goal is to provide a coherent framework
for conceptualization which (1) renders a generic yet rigorous conceptual foundation for understanding multiscale modeling that
hopefully can be shared by other parties and (2) demonstrates how
multiscale models, existing or new, can be precisely represented
and classied on the basis of such a foundation. In other words, this
framework is intended to be a basis for a common understanding
by people from different communities and to be open for further
extensions.
As indicated above, our approach starts with a conceptualization
of general systems. In Computer Sciences, an explicit specication of a conceptualization of a domain of discourse is called
an ontology (Gruber, 1993). One of the roles an ontology may
assume is clarifying the meaning of concepts in the context of
knowledge representation (cf. Guarino, 1995). This role is very
relevant to this work because the clarication of the meaning
of a scale, among others, is the key to a precise understanding
to multiscale modeling. An ontology that assumes such a role
is usually composed of the denitions of a set of concepts and
their relations (Uschold & Grninger, 1996). In this work, our
approach is implemented essentially by developing such an ontology which starts with a conceptualization of general systems (i.e.
the rst step). For this purpose we have opted to extend and adapt
part of an existing ontology, which was originally formulated by
Bunge (1977, 1979), later adapted by Wand and Weber (1990) and
referred to as the BungeWandWeber (BWW for short) ontological model in the recent literature (e.g. Rosemann & Green, 2002).
This existing ontology is selected as the basis of our own work,
because it gives precise denitions of basic concepts, from which it
derives very logically the concept of a system and its characterizations.
In the rest of the paper, the entire conceptualization framework
for multiscale models is presented through rigorous denitions and
illustrated by examples. In Section 2, two examples from chemical
process modeling are briey introduced. They will be referred to
throughout the paper. In Section 3, the conceptualization framework (i.e. the theory part of the paper) is presented following the
three steps as mentioned above. In Section 4, the conceptualization framework is applied to the two chemical process modeling
examples in a thorough manner. A comparison with existing work
is made in Section 5, where we try to demonstrate how the conceptualization framework can be used to interpret and unify diverse
terminologies and classications by different authors. Thus, Sections 4 and 5 can be viewed as the application part of the paper.
In particular, Section 5 may hopefully convey the idea that our
conceptualization framework provides a sharable foundation of
understanding and is open for extensions. Finally, some concluding remarks are given, including those with regard to the potential
usage of the results of this work.

824

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

2. Two illustrative examples

Two examples of multiscale modeling of chemical processes are
judiciously chosen from the literature. Example 1 models a system
on different scales to simplify computations, while a physical interpretation of the scales considered in modeling exists in Example 2.
A more precise account on the difference between these two types
of models will be given later in the paper. In the following, these
two examples are briey introduced; details of the mathematical
formulation of the models can be found in Section 4.
2.1. Example 1
The rst example is a so-called hybrid multizonal/computational uid dynamics (CFD) model of chemical process equipment
(Bezzo et al., 2004). As schematically shown in Fig. 1, a piece of
equipment (e.g. a vessel) is rst spatially decomposed into a number of zones. Each zone is characterized by a number of variables
such as temperature, pressure, and so on. Generally, these variables have different values in different zones, to reect the spatial
heterogeneity of the overall equipment. Furthermore, each zone is
spatially decomposed into a number of cells. Each cell is again characterized by a set of variables which often complies with the one
for a zone. For different cells in one zone, these variables may also
have different values.
The reason why the piece of equipment is decomposed on two
different levels, is because some phenomena (e.g. uid dynamics)
must be modeled at the level of cells in order to ensure the precision of modeling. For other phenomena (e.g. chemical or particulate
kinetics), it may be suitable to model them at the level of zones for
reducing computational costs while the expected precision of modeling is still achieved. This modeling approach has been used in e.g.
crystallizers modeling (Kulikov, Briesen, & Marquardt, 2005). It is
worth noting that within this modeling paradigm, the system is of a
continuum nature. The decomposition of the system into different
levels is carried out only for computational reasons.
2.2. Example 2
The second example is a homogeneousheterogeneous chemical reactor model (Vlachos, 1997). As schematically shown in Fig. 2,
the reactor is considered to have two parts: the (homogeneous) bulk
uid phase and the (heterogeneous) solid catalyst surface. The solid
surface is the place where adsorption, reaction, and desorption of
molecules occur. The entire reactor is modeled by adopting a continuum model for the bulk uid phase and the solid surface. In the
uid phase, the reactants of the surface reaction diffuse towards the
solid surface, and the products of the surface reaction diffuse from
the solid surface into the bulk uid phase. The model for the solid
surface relates the mass conservation at the surface to the rates
of the surface kinetics phenomena. To generate important kinet-

Fig. 2. Modeling of a chemical reactor.

ics information as required by an accurate continuum model for

the surface, the surface is further decomposed through a molecular lattice which comprises a number of sites of a particular spatial
layout. This molecular lattice is not a continuum model. Instead,
it is treated by a Monte-Carlo molecular simulation. Note that the
relationship between the continuum model of the surface and its
molecular lattice model is different from that between the model
of a zone and the model of a cell in Example 1; the former involves
different types of laws applied by the two models while the latter
does not.
3. The conceptualization framework
As indicated earlier, the conceptualization framework is developed on the basis of the BWW ontology which has been specied
using a rigorous set-theoretic language by the original authors. This
specication has received some criticism on the lack of understandability (Rosemann & Green, 2002). However, such a language is able
to convey the semantics of the concepts in the ontology with a high
degree of precision. To preserve this advantage, such a language is
still adopted in this current work. Regarding the ontologys scope,
the original work by Bunge (1977, 1979) was intended to represent
general systems for general purposes. The adaptation by Wand and
Weber (1990) was done later for providing an ontological model of
information systems. In this current work, we have so far focused
on a conceptualization that allows us to formulate general multiscale system models. Accordingly, the ontology presented here only
addresses this particular purpose instead of providing a complete
characterization of systems.
This current work is mainly based on the original ontology by
Bunge (1977, 1979). At a few occasions, the work by Wand and
Weber (1990) is reused. For a denition involved in the ontology,
links to the original references (if any) are always given. Further,
wherever the adaptation to the original formulation is signicant,
the denitions are marked with *.
The conceptualization framework is organized according to the
three steps to an understanding of multiscale modeling introduced
in Section 1. Consequently, the framework is composed of three
parts, namely basic concepts of a general system, scales and interscale relations, and multiscale models of systems, as schematically
shown in Fig. 3.
3.1. Basic concepts of a general system

Fig. 1. Modeling of a piece of chemical equipment.

3.1.1. Things, couplings, and laws

In this subsection, basic notions about things, their couplings
and the laws applicable to them are presented. They provide the
basis for dening systems in Section 3.1.2.
To get started, two fundamental concepts are taken for granted
without being rigorously dened: substantial individuals are individuals that exist; properties of any substantial individual are
called substantial properties, or simply properties. Based on these

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

825

Fig. 4. Composition of a thing.

Denition 5 (Composition of a thing). Let x  be a thing. The

composition of x is the set of its parts, i.e. C(x) = {y |y  x}.2
Fig. 3. Three parts of the conceptualization framework. Part 1 and Part 2 (shaded)
together are referred to as the conceptual foundation of the framework.

two concepts, the concept of a thing is introduced as the pair of a

substantial individual and the set of its properties. For example, a
chemical reactor as a thing refers to the individual of the reactor
as well as its properties such as volume, conversion, yield, etc.

Note that, according to such a denition, some parts of a thing

may appear in the composition of the thing together with parts of
their own. For example, if the thing x has the structure shown in
Fig. 4, then C(x) = {a,b,c,d,e,f}.
Now we characterize things in greater detail by representing
their properties and dening their states.

Denition 1 ((B-I 3.1)1 : Thing). Let s S be a substantial individual

and call p P the collection of its (unarized) properties. Then, the
individual together with its properties is called the thing x:

Denition 6 ((B-I 3.6): Functional schema). Let x =

s,p(s) be a thing
of class T . A functional schema xm of x is a certain nonempty set
M together with a nite sequence F of nonpropositional functions
on M, each of which represents a property of T. Briey,

x:=
s, p(s).

xm :=
M, F, where
F =
Fi |Fi is a function on M,

To denote the complete set of things, we introduce

1 i n.

Denition 2 ((B-I 3.2): Totality of things). The totality of things is

called :

Denition 7 ((B-I 3.9): State function, State). Let x be a thing modeled by a functional schema xm =
M,F, and suppose that each
component of the function

:={
s, p(s)|s S p(s) P}.

F =
F1 , F2 , . . . , Fn  :

We now examine the part-of relation between two different

things on the basis of the concept of a juxtaposition of things. A juxtaposition of things x and y is dened as the pair of the juxtaposition
of the substantial individuals represented by x and y and that of the
sets of properties of the two substantial individuals.
Denition 3 ((B-I 3.3): Juxtaposition of things). Let x =
s1 , p(s1 )
and y =
s2 , p(s2 ) be two things. Then the juxtaposition of x and
y is the third thing z = x + y =
s1 + s2 ,p(s1 + s2 ), where + represents
association or juxtaposition.

M V1 V2 Vn

represents a property of x, where Vi , i = 1, . . ., n represents a dimension of M. The Fi , for 1 i n, is called the ith state function of x, F
the total state function for x, and its value
F(m) =
F1 (m), F2 (m), . . . , Fn (m)
for m M is said to represent the state of x at m.

This denition basically states that if no factual change or novel

information arises when associating (or juxtaposing) one thing (e.g.
one uid phase of a multiphase reactor) with another thing (e.g. the
reactor to which the uid phase belongs), then the former is deemed
to be part of the latter.
With this part-of relation between things, the composition of a
thing is dened by means of

To illustrate Denitions 6 and 7, consider a room as a thing. The

room has a number of properties or formally state functions such as
temperature (F1 ), pressure (F2 ), and moisture (F3 ). The value of each
of these state functions is determined by the specic time (as an
instance of V1 ) and the spatial location (as an instance of V2 ) where
the value is measured. The time and spatial location together form
a particular m.
The above two denitions are comparable to some notions
adopted by Klir in the context of general systems (Klir, 1985). In
particular, a state function is equivalent to what is termed as variable by Klir for an operational representation of a property in terms
of a specic measurement or observation. An element m in domain
M is called a support set which represents a concrete backdrop
(e.g. in terms of time, space, etc.) in which a measurement or an
observation on a property is conducted.
To reect the fact that not every arbitrary state of a thing is
necessarily realistic, we further introduce the concept of law:

1
To show precise links to the references including Bunge (1977, 1979) and Wand
and Weber (1990), these references are marked as B-I, B-II, and WW, respectively. The reference number of the referenced item as appearing in the original
reference is appended after these labels. E.g. (B-I 3.1) implies a reference to Denition
3.1 given by Bunge (1977).

2
The composition of a thing may change over time, so does the relation with
other things. Considering this in this denition, the notation C(x) should rather be
C(x,t). However, the argument t is omitted here for clarity. The same simplication
is used also in the denitions of any systems composition and structure in Section
3.1.2.

Thing x is a part of thing y if and only if the juxtaposition of x

and y is equal to y. This is reected by
Denition 4 ((B-I 3.4): Part-of relation). Let x, y  be things. The
x is a part of y iff x + y = y. Symbol for the part-whole relation is .
Its negation is marked by 
/ .

826

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

Denition 8 ((B-I 3.10): Law3 ). Let xm =

M,F be a functional
schema for a thing x. Any restriction on the possible values of the
components of F and any relation among two or more such components is called a law, denoted by l(x). The totality of the laws of x is
dened as
L(x):={li (x),

1 i n}

An example for a law which represents a restriction on the possible value of a state function is the temperature of a room cannot
be lower than 20 . An example for a law which represents a relation between two state function is the pressure of a (closed) room
is proportional to its temperature.
Remark 8. We call a set of things as of a certain type if these things
have the same total state function and the same totality of laws.
According to this notation, the zone and the cell described in
Example 1 (Section 2) are of the same type. Although a zone and
a cell are characterized by a different set of laws in the example,
this has been done due to the computational concerns. From a
physical point of view, any state function applicable to a cell is
always applicable to a zone, and the same is true with respect
to the applicable laws. This is because both a zone and a cell
are of the same continuum nature. On the contrary, a site in the
molecular lattice and the solid surface as a whole, as involved in
Example 2, are of different types: a site has state functions such as
occupation-site function and local coordination (cf. Section 4.2.2
for a detailed explanation of the model of a site), which are not
applicable to the solid surface as a whole. Similar differences exist
also with respect to the set of applicable laws.
Based on the notion of the state of a thing, Bunge (1977) further
introduces the concepts of event (as a change of states) and history
(as a collection of events in time). These concepts in turn allow a
denition of the notion of actions between two things thing x acts
on thing y iff the existence of x changes the history of y. Omitting
these detailed denitions, here we directly come to
Denition 9 ((B-II, p. 64 ) Coupled things). Let x and y be two things.
They are said to be coupled iff x acts on y and/or y acts on x. Such
coupled things are denoted by b(x,y).
For example, a uid phase and a solid phase in a reactor are
considered as two coupled things, if there exists mass transfer (as
an inter-action) between these two phases. Note that two things
can be coupled through multiple actions. For this consideration,
we further introduce
Denition 10 ((B-II, p. 6) Coupling). Let x and y be coupled things.
An individual coupling relation, or simply a coupling between x and
y is denoted by c(x,y). The totality of the couplings between x and y
is denoted by C(x,y): = {ci (x,y), 1 i n}.
According to this denition, the totality of the couplings
between two chemical plants, for example, may include all the
material ows between them; each ow is regarded as one distinct
coupling. In the following, we examine the inuence of couplings
between things on their state functions. It is recognized that some
properties of a thing exist only because of the existence of certain
coupling(s). For example, the amount of material or energy leaving from or entering into a plant arises due to a certain material
or energy exchange with another plant which leads to a coupling
between the two plants. If there is no such material or energy connection between two plants, this property will not be considered.

3
Bunge used the term Law Statement, which is shortened to Law in this paper
for simplicity.
4
This is the page number in which the semantics of this item is discussed in the
original reference.

To distinguish such properties or state functions from the rest, we

introduce the following
Denition 11 (Coupling-induced state function). Let x be a thing
coupled with at least one other thing. Any state function of x is called
a coupling-induced state function, denoted by Fkc , iff it represents a
property of x that occurs only due to the coupling c(x,y) between x
and another thing y. Such a property is denoted as pci . By denition,
pci (p(x|y  b(x, y)) p(x)).
In general, the characteristics of a coupling, manifested via the
coupling-induced state functions of two coupled things, can be
inuenced by the characteristics of the two things independent of
the coupling between them. These characteristics are manifested
via non-coupling-induced state functions of the two things. Such
inuences are expressed via a coupling mechanism law as dened by
Denition 12 (Coupling mechanism law). Let x and y be two coupled things. Let c(x,y) C(x,y) be an individual coupling between x
and y. Then, a coupling mechanism law, denoted as lc (x,y) is dened
as any relation that maps non-coupling-induced state functions of
x and/or y to coupling-induced state functions of x or y induced
by c(x,y). The totality of the coupling mechanism laws related to a
coupling c(x,y) is dened as
Lc (c(x, y)):={lic (x, y), 1 i n}.
The term mechanism is used here, because such a law usually
explains why and how the interaction occurs between the two parties involved in a coupling. Examples of coupling mechanism laws
can be easily found in the studies of transport phenomena (Bird,
Stewart, & Lightfoot, 1960). For instance, the rate of heat transfer (as
a coupling-induced state function) between two objects is proportional to the difference of temperatures (as non-coupling-induced
state functions) of the objects.
While a coupling mechanism law relates coupling-induced state
functions to the characteristics of two coupled things, a topological
connection law on a given coupling relates a coupling-induced state
function of one thing to one of another thing:
Denition 13 (Topological connection law). Let x and y be two coupled things. Let c(x,y) C(x,y) be an individual coupling between x
and y. Let fxc be any coupling-induced state function of x induced
by this coupling and fyc be any coupling-induced state function
of y induced by this coupling. Then, a topological connection law,
denoted as lt (x,y) is dened as any relation between fxc and fyc . The
totality of the topological connection laws on the coupling c(x,y) is
dened as
Lt (c(x, y)):={lit (x, y)), 1 i n}.
Typically, a topological connection law on a coupling equates
a coupling-induced state function of one thing to that of another
thing. Consider for example two chemical plants coupled with a
pipe. The amount of material owing out of plant A through this
pipe is equal to the amount of the material owing into plant B.
The term topological refers to the relations between two connected
things, none of which is part of another one. It is used here in contrast to mereological which refers to relations between one thing
and one or more things which are part of the former. Using the
term topological is appropriate here because a connection law as
dened above is for specifying the link between two parties which
are not part of each other. Otherwise the connection law would be
mereological. The semantics of such laws will be dened later (cf.
Denition 27).
3.1.2. Systems and subsystems
On the basis of the notion of a thing and related concepts, we
dene and characterize systems and subsystems in this subsection.

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

A system is essentially dened as a set of things and their connections. Rigorously, we introduce
Denition 14 ((WW 10): System). Let C be a set of things (i.e. C ),
and let BC = {(x,y)|x,y C b(x,y)}. Let (C,BC ) be a graph, where C is
the set of vertices and BC the set of edges. Then (C,BC ) is a system
denoted by , iff it is a connected graph.
This denition basically states that a system is the collection of
things, any of which must be connected to one or more of things of
the collection. For example, a chemical reactor that has a number of
phases is considered as a system, only when every phase is linked
to at least one other phase in the reactor.
Corollary 14.

Any system  is a thing, i.e.  .

In Section 3.1.1, we have introduced the concept of composition

of a thing through Denition 5. For convenience, we dene this
concept again particularly for a system and also introduce the term
component.
Denition 15* ((B-II 1.2): Systems composition). Let  be a system.
The composition of  is the set of things in :
C():={x|x  x  }.
Remark 15. Any element in the composition of the system ,
denoted by c, is called a component of .
Still using a multiphase reactor as an example, the composition
of this system is the phases of the reactor as they are all the things
that are part of the reactor.
In the original formulation of Bunge (1979), the notion of atomic
composition (or A-composition) is employed, to represent a collection of parts of a thing or a system, if these parts are atomic, i.e. if
they are not intended to be further decomposed according to the
nature of the system studied or to the purpose of the study. According to Bunge, the atomic parts of a system are things of the same
class. This is problematic, because there exist hybrid systems such
as human-machine systems and software-hardware systems which
are composed of atomic parts of different classes of things. To avoid
this problem, we redene the notion of A-composition through the
following:
Denition 16 (A-component). Let  be a system, and c be a component of . c is an atomic component or an A-component of , denoted
by cA , iff c does not contain any other components of , i.e.
/ {x C()|y C(), y =
/ x, y  x}.
cA C()&cA
Denition 17* ((B-II 1.2): A-composition). Let  be a system. The
composition of  in terms of all A-components of  is called the
A-composition of , denoted by CA :
CA ():={cA,i , 1 i nA }.
Denitions 16 and 17 may be illustrated by one of the previous
examples. If the thing x in Fig. 4 is a system  (thus assuming connectivity between the components of x), then b, c, e, and f are the
A-components of x, but b and d are not. The A-composition of x is
CA () = {b,c,e,f}.
Employing the notion of A-component as dened above, the
concepts A-environment and A-structure are dened as follows:
Denition 18 ((B-II 1.2): A-environment). Let  be a system. The Aenvironment of  is the set of all things which are not components
of , but are coupled with some A-components of :
/ C() (y)[y CA ()&b(x, y)]}.
EA ():={x|x

827

Denition 19 ((B-II 1.2): A-structure). Let  be a system. The Astructure of  is the set of couplings among the A-components of ,
and among them and the things in the A-environment of :
SA ():=BA () B A (),
where
BA () = {c(x, y)|x, y CA ()} =
/ ,
B A () = {c(x, y)|x CA () y EA ()}.
Hence, BA () is the set of coupling relations among Acomponents of , and B A () is the set of coupling relations between
 and its A-environment.
To illustrate the denitions of A-environment and A-structure,
consider the computer network system of a university, as a system
composed of inter-connected PCs, and all the external devices connected to particular PCs of the system. If the PCs are not to be further
decomposed within a particular context of investigation, i.e. if they
are considered as A-components of the system, then these external devices are the A-environment of the system. A-structure of the
system is the collection of all the connections between the PCs in
the computer network system and the links between the external
devices and the PCs to which these external devices are connected.
Now we are ready to dene a subsystem by using the concepts
of A-composition, A-environment, and A-structure:
Denition 20 ((B-II 1.6, WW 12): Subsystem). Let  be a system with
A-composition CA (), A-environment EA (), and A-structure SA ().
Then x is a subsystem of , iff x is denoted by x  a system, and
[CA (x) C A ()]
[EA (x) {EA () {C A () CA (x)}}]
[SA (x) SA ()].
The latter conditions state that (i) the A-composition of a subsystem x of system  is part of (or equal to) the A-composition of
, (ii) the A-environment of subsystem x is part of (or equal to)
the summation of the A-environment of  and the part of the Acomposition of  other than the A-composition of this subsystem,
and (iii) the structure of the subsystem x is part of (or equal to) that
of .
Still using the university computer network example, one can
state that a computer cluster of the university is a subsystem of the
entire network. Here, the PCs in the cluster form a subset of the
entire connection of PCs in the entire network. The environment of
the cluster comprises elements from the rest of the network and/or
those from the environment of the entire network. The connections
between the PCs in the cluster are part of the connections between
all the PCs in the network.
3.2. Scales and inter-scale relations
When a system is investigated for a certain purpose, a particular decomposition of it might be taken, either following numerical
principles (e.g. nite volume discretisation in CFD modeling) or
accounting for the well-recognized, natural kinds of the components of the system (e.g. continuum, molecular, and atomic levels
of material). Generally speaking, decomposition of a system can
result in multiple levels of abstraction, each of which is composed
of components of a particular granularity (in space, time, etc.). In the
context of modeling, these levels have been conventionally called
scales, giving rise to the term of multiscale modeling. In this subsection, we attempt to clarify the semantics of levels or scales and
identify typical inter-scale relations.
We rst introduce the concepts of level (or scale) and level hierarchy. Intuitively, a level of a system represents part of a certain

828

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

Corollary 21c. It is possible that a level Vi of a system  only represents part of the system. That is, for xj Vi , j = 1, 2, . . . , m, it is possible

that

CA (xj Vi )

j=1

CA ().

Example 2 of Section 2 provides an illustration for Corollary 21c,

because in that example only the solid surface, but not the uid
phase, is decomposed on Level 1. Consequently, no representation
of the uid phase exists on Level 1 (cf. Fig. 6).
Fig. 5. Level hierarchy of a chemical equipment represented by a hybrid multizonal/CFD model (Example 1).

Remark 21. We call two neighboring levels homogeneous if all the

components in these levels are things of the same type (cf. Remark
8). Otherwise they are called being heterogeneous.
To illustrate these two kinds of neighboring levels, recall the
explanations previously given for Remark 8 referring to the two
examples of Section 2. It has been made evident that the level of
zones and the level of cells are homogeneous neighboring levels, and
the level of molecular lattice sites and that of the continuum solid
surface are heterogeneous neighboring levels.
Denition 21 claries what constitutes a level. In the following,
we further dene the environment and the structure of a level.

Fig. 6. Level hierarchy of a chemical reactor represented by a heterogeneoushomogeneous model (Example 2).

decomposition schema. In a multilevel system, one level is said to

be lower than another one, if any component in the former is
always part of a particular component in the latter. For instance,
in Example 1 introduced in Section 2, the level of cells is lower
than the level of zones, because any cell is always a component of
a particular zone. Rigorously, we introduce
Denition 21 ((B-II 1.8): Level (or scale) and level hierarchy of a
systemce:5 ). Let  be a system. Let V = {Vi |1 i n} be a family of n
/ xj . Then,
nonempty sets of things, Vi C(). Further, xi , xj Vi ,xi 
(i) (Vi < Vj ):=(x)[x Vj (y)(y Vi y C(x))].
(ii) H =
V,< is called a level hierarchy of , and any Vi V is called a
level (or scale) of .
We have used the convention that V1 is the level at the bottom of
the level hierarchy, i.e. the components on this level are not further
decomposed. V2 is the level immediately above V1 . Following this
convention, for 1 i n, Vn becomes the level at the top of the level
hierarchy.
The above denition can be illustrated with the two examples
introduced in Section 2. Both of the two examples adopt a 3-level
hierarchy, as shown in Figs. 5 and 6.
i1

Corollary 21a ((B-II 1.8)). For any Vi , i > 1, x Vi C(x) Vk .

k=1

As an illustration to this corollary, one can see in Fig. 6 that reactor at Level 3 has the following components: solid surface and
uid phase from Level 2, and the sites from Level 1. Besides,
one can also see that Level 2 contains two subsystems of reactor,
which gives an example for the following corollary.
Corollary 21b. Any level Vi of a system  not at the bottom of the
level hierarchy H must contain some subsystem(s) of . That is

Vi V, i > 1 x Vi , x .

Denition 22 (L-environment). Let  be a system with levels

V = (Vi |1 i n). The environment of Vi , or an L-environment, is
dened as the set of things which are not components in Vi but
are coupled with one or more components in Vi :
/ Vi &[y Vi , b(x, y)]}.
EL (Vi ):={x |x
Denition 23 (L-structure). Let  be a system with levels
V = (Vi |1 i n). The L-structure of Vi is dened as the set of couplings among the components of Vi , and among them and the things
in the L-environment of Vi :
SL (Vi ):=BL (Vi ) B L (Vi ),
where
BL (Vi ) = {c(x, y)|x, y Vi } =
/ ,
B L (Vi ) = {c(x, y)|x Vi y EL (Vi )}.
To illustrate the above two denitions, consider a two-level reactor system where a reactor (at Level 2) is composed of a vessel and
a temperature control jacket (both at Level 1). The L-environment
for Level 1 may include the air surrounding the jacket outside the
reactor (as a heat sink) to which the jacket loses heat. This heat
sink should not be regarded as part of the L-environment of Level 2,
because it is directly coupled with the jacket at Level 1 of the system.
The L-structure of Level 1 of this system, according to Denition 23,
is the set of couplings including the (internal) heat transfer between
the jacket and the vessel, as well as that between the jacket and the
external heat sink.
In the context of multiscale modeling, it is of particular interest
to investigate the relation between a component at a certain level
and its immediate composition, as dened by
Denition 24 (Immediate composition). Let  be a system with levels V = (Vk |1 k n). The immediate composition of a component x on
level Vi , i > 1, denoted by Cim (x), is dened as a set of components
on levels preceding Vi such that, each component in this set does
not belong to the composition of any component located in a level
below Vi . That is
/ C(z)]}
C im (x|x Vi , 1 < i n):={y C(x) (z Vk , k < i)[y

5
Bunge (1979) uses the term level structure instead of level hierarchy. We
have changed the term here to avoid possible terminological confusion between
this concept and L-structure introduced by Denition 23.

Remark 24. An element in the immediate composition of a component c is called an immediate component of c.

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

As an example, solid surface and uid phase in Fig. 6 are both

immediate components of reactor, however the sites (at Level
1) are not.
A model of a system with multiple levels often takes into account
the relations between a component of a system and its immediate
components. In particular, many multiscale modeling applications
are motivated by the possibility of better modeling properties of a
component at a higher level of abstraction through relating them
to some properties of its immediate components. In the following,
we distinguish between two kinds of relations.
Denition 25 (Aggregation law). Let  be a system with levels
V = (Vk |1 k n). Let x be a component on level Vi , i > 1 with an
immediate composition Cim (x). An aggregation law on x, denoted
as la (x), is a relation that maps one or more state functions of all
the components in Cim (x) to a state function of x. The totality of
aggregation laws of x is dened as La = {lia , 1 i n}.
Remark 25. We call an aggregation law an averaging law, if it
represents the averaging (with respect to the dimension(s)6 along
which the system is decomposed) of one state function of all components in the immediate composition of x (i.e. Cim (x)) to obtain
a state function of x. An averaging law usually exists between two
homogeneous neighboring levels (cf. Remark 21).
Denition 26 (Disaggregation law). Let  be a system with levels V = (Vk |1 k n). Let x be a component on level Vi , i > 1 with an
immediate composition Cim (x). A disaggregation law on x, denoted as
ld (x), is a relation that maps a state function of x to a state function
of all the components in Cim (x). The totality of disaggregation laws
of x is dened as Ld = {lid , 1 i n}.
To illustrate the above two denitions, the two examples of Section 2 are used again. As an aggregation law, a state function of a
zone at Level 2 in Example 1 (cf. Fig. 5) which represents a property
such as the turbulent energy dissipation rate is determined by the
same property of all CFD cells in this zone, as appearing at Level
1. In this particular case, the value of the property of the zone can
be obtained by spatially averaging the values of this property of
all the cells. As a disaggregation law, state functions that represent
the temperature, chemical composition, etc. of a cell at Level 1 are
determined by those of the entire zone at Level 2 which contains
this cell. As one of the modeling alternatives, the values of these
properties of each cell can simply be set equal to those of the entire
zone.
In Example 2, within its three-level hierarchy, some state functions that represent the properties of the surface at Level 2, such as
the desorption rate of a certain chemical species, can be determined
by the micro-conguration of the lattice, i.e. by how every site that
appears in Level 1 is occupied. The micro-conguration of the lattice is obtained through Monte Carlo simulation. This essentially
implies an aggregation law. Further, the adsorption, desorption, or
reaction transition probability of each site at Level 1, as an important
state function of the site involved in the Monte Carlo simulation, is
dependent of the temperature of the site. This temperature is set to
be equal to the temperature of the entire surface which appears on
Level 2. The relation between the temperature of a site and that of
the surface gives another instance of disaggregation law.
In addition to aggregation laws and disaggregation laws which
primarily govern non-coupling induced state functions, we still need
to set up the connection between the coupling induced state functions of a component and some of its immediate components. For
this purpose, we introduce

6
A dimension along which a system may be decomposed is represented as Vi in
Denition 7.

829

Fig. 7. Coupling between two zones.

Denition 27 (Mereological connection law). Let  be a system with

levels V = (Vk |1 k n). Let x be a component on level Vi , i > 1 with
an immediate composition Cim (x). A mereological connection law on
x, denoted by lm (x), is a relation between a coupling-induced state
function of x induced by a coupling with y Vi or z EL (Vi ), and one
or more coupling-induced state functions of a set of components
U Cim (x), each of which is induced by a coupling between u U
and v C im (y) or w C(z). The totality of mereological connection
laws of x is dened as Lm :={lim , 1 i n}.
To illustrate Denition 27, let us rst examine the coupling
between two zones in Fig. 2. Fig. 7 represents a (mass transfer) coupling between Zone A and Zone B. Now consider one of the two
zones, e.g. Zone A. This zone has two coupling-induced state functions Fout and Fin , representing the rate of the mass ow leaving this
zone and entering Zone B, and vice versa. In Fig. 8, it is shown that at
Level 1 (i.e. the level of cells), each cell at the boundary of Zone A is
coupled to one cell at the corresponding boundary of Zone B. Such
a coupling introduces two coupling-induced state functions of this
cell, fout and fin . As an instance of the mereological connection law,
there exists a (equality) relation between Fout (Fin ) and the sum of
fout (fin ) of all cells at the boundary of Zone A which is connected to
Zone B.
In some simpler cases, an input or output of a component at
a higher level is directly assigned to one of its immediate components. In such a case, a mereological connection law simply
represents this assignment. Fig. 9 illustrates an example in chemical plant simulation. The variables of a material stream (all being
coupling-induced state functions), which ows from Plant A to
Plant B as part of the simulated owsheet when viewed at the plant
level (as opposed to the unit level), eventually need to be assigned
to those of a material stream entering a certain process unit within
the destination plant (i.e. Plant B), in order to set up the coupling
between these two plants at the unit level.
3.3. Multiscale models of systems (Part 3)
Notions introduced in Sections 3.1 and 3.2 have paved the way
for precisely characterizing a multiscale model of a system. In this
subsection, we rst dene the concept of a single-scale system
model. Thereafter, the concept of a multiscale model and its classication are introduced. Note that only denitions of these models
are given here; one should refer to Section 4 for examples of these
models.
A single-scale system model is a collection of the laws of all components at a certain level of observation and the laws on couplings
among these components. Rigorously, we introduce

Fig. 8. Couplings between cells at the boundaries.

830

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

Denition 28 (Single-scale model). Let  be a system and V a level

of . For any component x on level V, let L(x) be the totality of laws
on x. Let SL (V) be the L-structure of level V. Let Lt (t) and Lc (r) be
the totality of the topological connection laws and that of coupling
mechanism laws of a coupling r SL (V), respectively. The single-scale
model of  for level V, denoted as Ms (,V), is dened as follows:
M s (, V ):={L(x)|x V } {Lt (r)|r SL (V )} {Lc (r)|r SL (V )}.
A multiscale model is a collection of single-scale models as
well as laws characterizing inter-scale relations of the system being
modeled. Rigorously, we introduce
Denition 29 (Multiscale model). Let  be a system with levels
V = (Vk |1 k n,n > 1). For any component z on level Vi , 1 < i n, let
La (z),Ld (z), and Lm (z) be the totality of its aggregation, disaggregation, and mereological connection laws, respectively. Thus, the
multiscale model of , denoted as Mm (), is dened as follows:
n

M m ():={M s (, Vk ), Vk V } {La (z)|z Vk }

k=2

Denition 30 (Scale-collecting model). Let  be a system with levels V = {Vk |1 k n,n > 1}. Let Mm () be a multiscale model of .
For any component z on level Vi , 1 < i n, let La (z), Ld (z), and Lm (z)
be the totality of its aggregation, disaggregation, and mereological
connection laws, respectively. Mm () is a scale-collecting model, iff
n

k=2

k=2

k=2

{La (z)|z Vk } {Ld (z)|z Vk } {Lm (z)|z Vk } = .

This denition simply states that a scale-collecting model is one

where there is neither an aggregation law, a disaggregation law, nor
a mereological connection law on any level that has the possibility
of possessing such a law (i.e. any level other than Level 1).
Denition 31 (Scale-connecting model). Let  be a system with levels V = {Vk |1 k n,n > 1}. Let Mm () be a multiscale model of .
For any component z on level Vi , 1 < i n, let La (z), Ld (z), and Lm (z)
be the totality of its aggregation, disaggregation, and mereological
connection laws, respectively. Mm () is a scale-connecting model, iff
{Lm (z)|z Vk } =
/

k=2

k=2

k=2

k=2

Reecting how closely the different levels of a system are linked,

multiscale models can be further classied into three types7 . In one
extreme case, a scale-collecting model simply collects a number of
single-scale models without involving any laws on the inter-scale
connections. In another extreme case, a scale-integrating model
involves at least aggregation or disaggregation laws. In between,
a scale-connecting model involves some mereological connection
law(s) but does not have any aggregation or disaggregation laws.
Rigorously, we introduce

Denition 32 (Scale-integrating model). Let  be a system with levels V = {Vk |1 k n,n > 1}. Let Mm () be a multiscale model of . For
any component z on level Vi , 1 < i n, let La (z), Ld (z) be the totality
of its aggregation and disaggregation laws, respectively. Mm () is a
scale-integrating model, iff
/ .
{La (z)|z Vk } {Ld (z)|z Vk } =

{L (z)|z Vk } {L (z)|z Vk }
k=2

Fig. 9. An example of simple mereological connection laws.

and

k=2

k=2

{La (z)|z Vk } {Ld (z)|z Vk } = .

This denition states that a scale-connecting model is one where

there are one or more mereological connection laws but no aggregation law or disaggregation law.

7
This classication is essentially based on the three types of inter-scale laws
dened earlier, namely the mereological connection law, the aggregation law, and the
disaggregation law. This classication would be complete, if there were no inter-scale
laws other than these three. However, we consider the validity of this assumption
as an open issue. See the concluding remarks in Section 6 for further discussions of
this matter.

This denition states that a scale-integrating model is one where

there is at least one aggregation law or one disaggregation law.
In the practice of systems modeling, it is rare to develop scalecollecting models; this concept is introduced here mainly for the
purpose of completeness. One example of scale-connecting models is the owsheet-partition-unit schema frequently employed in
chemical process simulation, where different scales are introduced
and connected for carrying out a particular simulation algorithm
(cf. the second example of mereological connection law, illustrated
in Fig. 9). The most interesting and complex multiscale models are
scale-integrating models: it is the existence of the aggregation/
disaggregation laws that brings the most prominent value of considering different scales.
4. Exemplary applications of the conceptualization
framework
In this section, a detailed analysis of the two illustrative examples introduced in Section 2 is presented to give a more complete
explanation on the composition of scale-integrating models. All the
concepts dened in the preceding section will be highlighted in
italic subsequently. When necessary, back references to the numbers of the denitions or remarks are also given. As will be shown
through these two examples, adjustments to the models formulated
by their original authors may be required to properly re-express
them using the current conceptualization framework; these adjustments will be briey discussed at the end of this section.
4.1. Example 1: Hybrid multizonal/CFD model of chemical
equipment
As shown earlier, the model originally formulated by Bezzo et al.
(2004) represents a piece of process equipment as three levels (cf.
Figs. 1 and 5). In the following, we will take a closer look at Level
2 and Level 1 in order to analyze the composition of this hybrid
model according to the conceptualization framework presented in
this work. Unless stated otherwise, all the modeling treatments are
based on Bezzo et al. (2004), where further explanations regarding
the model may be found.
4.1.1. The multizonal model of Level 2
On this level, the system of equipment is decomposed into a
number of (internal) zones; one zone can be connected to one
or more other zones. A coupling between two zones, standing for

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

material transport, is represented through an interface that links

two ports8 , each of which belongs to one of the two connected
zones. In addition, the connection between the equipment and
the environment of the equipment is represented through some
so-called environment zones, which essentially play the role
of couplings between the (internal) zones and the things in its
environment. In the following, the set of ports of an arbitrary zone
z is denoted by Pz . The sets of internal and external zones on this
level are denoted by ZI and ZE , respectively.
4.1.1.1. Laws of individual (internal) zones. Laws applicable to one
(internal) zone typically include the conservations of mass and
energy, chemical reaction kinetics, and other types of kinetics,
depending on the phenomena to be captured in the zone. The collection of these laws can be represented by a differential-algebraic
model of the following form:
FzI

yz ,

dyz
, {xp , xpin , Fp , Fpin , p Pz }, z , z , Vz
dt

= 0,

z Z I , (1)

where yz is a set of internal zone variables (e.g. mass and energy

holdups, compositions, temperature, etc.), xp denotes the vector of
intensive properties (e.g. temperature, pressure) that any material
exiting zone z at port p would have, xpin denotes the vector of intensive properties of the material entering zone z at port p, Fp is the
outlet mass ow at port p, Fpin is the inlet mass ow at port p, az
denotes a set of parameters that are related to uid ow and/or mixing phenomena, and which somehow affect the behaviour of zone z
(e.g. the turbulent energy dissipation rate which may affect the rate
of nucleation in a crystallisation process),  z is a set of uid properties (e.g. density, viscosity, etc.) and/or parameters that affect such
uid properties, and Vz is the volume of zone z which is assumed
to be a given constant.
Applying our conceptualization framework, all the above variables, properties, and parameters are state functions (Denition 7)
of this zone. Especially, variables appearing within the bracket in
Eq. (1) are essentially coupling-induced state functions (Denition
11).
4.1.1.2. Laws of couplings on Level 2. In the original model, an environment zone model is introduced to represent certain relations
between the environment and the internal zones, which is of the
following form:
FzE (yz , u, {xp , xpin , Fp , Fpin , p Pz }) = 0,

z Z E .

(2)

Here the variables yz and u represent the connections between

the environment and this environment zone, and the variables in
the bracket represent connections between this environment zone
and some internal zones. As mentioned above, such an environment zone can be replaced by one or more couplings between the
environment and one or more internal zones. A topological connection law (Denition 13) of the following form can be applied to each
of such couplings:
lt (vcE , {xz , xzin , Fz , Fzin }) = 0,

(3)

vcE

where
denotes all state functions of the environment induced by
this coupling, and the variables in the bracket are state functions of
this (internal) zone induced by this coupling. The collection of such
laws of all couplings with the environment is equivalent to (2).
In addition to the couplings with the environment, there are still
couplings between internal zones. For such a coupling, topological

8
Applying the concepts dened in this paper, one can consider a port as a collection of state functions of a thing induced by a particular coupling between this thing
and another thing.

831

connection laws in the form of identity equations are applied,

Fpin = Fp
Fpin = Fp

xpin = xp
xpin = xp

(p, p ) I,

(4)

where (p,p ) represents one interface between ports p and p , each

belonging to one zone involved in the coupling.
4.1.2. The (discretised) CFD model of Level 1
This level is composed of a number of cells (denoted by C) and
their connections resulting from the discretization of the CFD model
of the equipment.
4.1.2.1. Laws of individual cells. One CFD cell is characterized by
state functions such as volume (Vc ), pressure (Pc ), intensive properties ( c ) which are determined by the Level 2 (i.e. multizonal)
model (e.g. temperature and composition), uid velocity at its centre (vc ), hydrodynamic quantities c (e.g. the turbulent energy
dissipation rate), and nally the physical properties such as density (c ) and viscosity (c ). The CFD model of one cell is typically
composed of laws on mass/momentum conservations, as well as
physical property relations for density and viscosity of the following
form:
c = (Pc , c ),

c = (Pc , c ),

c C.

(5)

The solution of all the laws of one cell provides values of quantities such as vc and c , which are needed for setting up connections
between neighboring cells (cf. Section 4.1.2.2) and between Levels
1 and 2 (cf. Section 4.1.3.1).
4.1.2.2. Laws of couplings on Level 1. In the original model, two adjacent cells (c,c ) share a face f; the set of faces in the CFD model is
denoted as F. Faces at the boundaries of the domain of interest are
also denoted as (c,c ). This treatment implies that a real cell c is
connected to c which can be either another real cell or a topological boundary cell. Here, a face is essentially a coupling between
two cells, through which momentum and mass transport occurs.
For such a coupling, coupling-induced state functions of the involved
cells are the uid velocity and the corresponding mass ow-rates
at the faces; the latter can be derived from the former through a
law of this cell. Comparable to (4), identity relations (as topological
connection laws) should exist between the above coupling-induced
state functions of two coupled cells.
Besides, the solution of the CFD model gives the uid velocity at the centre of each cell. To obtain the uid velocity at a face
between two cells, certain interpolation can be applied. According to Denition 12, this interpolation relation can be viewed as a
coupling mechanism law, as it relates a coupling-induced state function (the uid velocity at the face) to a non-coupling-induced state
function (the uid velocity at the center of a cell).9
4.1.3. Inter-scale laws
Now we come to the part of the original model that addresses
the connection between the above two (essentially homogeneous
neighboring) levels (Remark 21). The notation Cz is introduced to
denote the set of cells on Level 1 comprising the zone z on Level 2.
We will show in the following that all the three types of inter-scale
laws are present in the original model.

9
The law statement in this case does not really reveal the mechanism of this
coupling; it is rather merely a manipulation of the result of solving the CFD model.

832

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

4.1.3.1. Aggregation laws. One rather obvious aggregation law

(Denition 25) relates the volume of one zone and to those of the
cells in the zone:
Vz =

Vc

(6)

c Cz

Another aggregation law, which is more signicant, derives a

volume-averaged property of one zone, such as its energy dissipation rate, from the properties of the cells in the zone:

c Vc

c Cz

z =

(7)

Vz

According to Remark 25, Eq. (7) is an averaging law.

4.1.3.2. Disaggregation laws. Disaggregation laws (Denition 26) are
introduced in this model to relate the properties determined by the
Level 2 (the multizonal) model (i.e.  z in Eq. (1)) to the properties
needed by the CFD model of a cell in this zone (i.e.  c in Eq. (5)). The
original model presents a simplest (identity) relation:
c = z ,

c Cz

(8)

4.1.3.3. Mereological connection laws. Here, mereological connection

laws (Denition 27) are introduced to relate coupling-induced state
functions of cells on Level 1 with those of the zones on Level 2. In
particular, mass ow-rates across the faces of cells which make up
an interface (p,p ) between ports p and p of two adjacent zones (this
set of faces being denoted as Fpp ) are related to the mass ow-rates
across the interface (or coupling) between these two zones. These
mereological connection laws are
Fpp =

max(Fcc , 0),

(c,c  )

Fp p =

max(F cc , 0),

(9)

(c,c  )

where (c, c  ) Fpp , Fcc is the rate of the mass ow from cell c to
its adjacent cell c , Fpp Fpin = Fp , Fpp Fp = Fpin , (cf. Eq. (4)). By
Eq. (9), both Fpp and Fp p are non-negative. In particular Fpp sums
up the rates of the non-negative ows from cells (denoted by c)
belonging to port p to those (denoted by c ) belonging to port p ,
while Fp p sums up the rates of the non-negative ows from cells
belonging to port p to those belonging to port p.
4.2. Example 2: Homogeneousheterogeneous reactor model
As shown earlier in Fig. 6, the example originally formulated by
Vlachos (1997) involves three levels of the reactor being modeled.
As in the previous example, we will take a closer look at Level 2 and
Level 1, and analyze the composition of this reactor model according
to the conceptualization framework presented in this work. Again,
unless stated otherwise, all the modeling treatments are based on
Vlachos (1997), where further explanations regarding the model
may be found.
4.2.1. The continuum model of Level 2
On this level, the system reactor is decomposed into two subsystems, namely the uid phase and the solid surface. A coupling exists
between the two parts to represent the mass transfer between
them.
4.2.1.1. Laws of the uid phase. The model of the uid phase is composed by applying laws of mass conservation and mass diffusion
(Ficks rst law in particular). Concretely, it includes the following
equations (assuming no reaction occurs in the uid phase):
CA

=
t
x

CA
x

(10a)

CA = CAL
D

at x = L,

CA
= rAout rAin
x

(10b)
at x = 0.

(10c)

Here CA is the concentration of A, x is the spatial coordinate, L is

the boundary of the uid phase on the opposite of its interface with
solid surface, CAL is the concentration of A at L, D is the diffusivity,
rAout is the rate of transporting A out of the uid phase due to the
adsorption at the solid surface, rAin is the rate of transporting A into
the uid phase due to the desorption at the solid surface. Clearly,
rAin and rAout are two coupling-induced state functions (Denition 11)
of the uid phase (see also Section 4.2.1.3).
4.2.1.2. Laws of the solid surface. The model of the solid surface is
composed by applying laws of mass conservation of the surface
species AS as well as kinetics laws of adsorption, desorption, and
surface reaction. Concretely, it includes
CAS
= ra rd rrs ,
t

(11a)

ra = ka CA0 (1 A ),

(11b)

rd = kd f (A , T, wAA ),

(11c)

rrs = kr CAS .

(11d)

Here CAS is the concentration of AS, CA0 is the concentration of

A in the uid phase at the boundary adjacent to the surface (which
can be viewed as a coupling-induced state function of the surface),
ra , rd , rrs are rates of adsorption, desorption, and surface reaction
respectively, ka , kd , krs are temperature-dependent rate constants
of adsorption, desorption, and surface reaction respectively, wAA is
the energy of inter-adsorbate interations, and  A is the coverage
dened as
A =

CAS
CT

(11e)

where CT is the total concentration of sites on the surface.

To determine the function f in Eq. (11c), one may employ a meaneld model, which assumes the atoms are uniformly distributed on
the surface and share the same number of neighbors. This treatment
does not require further modeling at the non-continuum level, but it
is not able to count for imperfections such as defects and impurities
on the surface. It is well known that adequately addressing these
factors can be very important for providing a reliable description
of the surface. To tackle this problem, this function can take the
following form, which employs the information available from the
modeling at the non-continuum level:
f = f (j, wAA , T, qj ,

j = 0, 1, . . . c).

(12)

Here c is the coordination of the surface (e.g. c = 4 for a square lattice representing the (1 0 0) surface), qj is the fraction (over all sites)
of the occupied sites with local coordination equal to j. As we will
show later in Section 4.2.3.1, the value of qj can be determined by
aggregating certain state functions of the sites in the lattice modeled
on Level 1.
4.2.1.3. Laws of the coupling. The coupling between the uid phase
and the solid surface is characterized by the following topological
connection laws (Denition 13):
CA|x=0 = CA0 ,

(13a)

rAin

(13b)

= rd ,

rAout

= ra .

(13c)

Eq. (13a) equates the concentration of A at the boundary in the

uid phase (CA|x=0 ) with the concentration of A referred to by the
surface model (CA0 ) (cf. Eq. (11b)). Eq. (13b) equates (i) the rate of

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

transporting A into the uid phase due to desorption (rAin ) to (ii) the
desorption rate referred to by the surface model (rd )(cf. Eq. (11c)).
Eq. (13c) sets up the same relation for rates pertaining to adsorption.
It is worth noting that, according to Denition 13, a topological connection law relates only coupling induced state function of
two coupled things. Now consider the variables involved in Eqs.
(13a)(13c) which have just been termed a set of topological connection laws. One can see that within the set of individual variables,
only rAin , rAout (cf. Section 4.2.1.1) and CA0 (cf. Section 4.2.1.2) have
been previously identied as coupling-induced state functions. All
the other variables are not coupling-induced because they exist
regardless of whether there is a coupling between the uid phase
and the solid surface or not. This problem can be resolved if the
following, uniformed way of formulating models is applied:
One uses non-coupling induced state functions of a thing only
within laws of the thing.
One introduces, when needed, additional coupling-induced state
functions which have identity relations with some of the noncoupling-induced state functions.
One uses only coupling-induced state functions in topological connection laws.
In fact, a treatment of this kind is frequently applied in chemical process modeling, where the composition of a material stream
leaving a chemical process unit (as a coupling-induced state function)
is set to be identical to the composition of the material in the unit
(as a non-coupling-induced state function). This is exactly what the
second point is about. These non-coupling-induced state functions
are usually used for modeling the phenomena within the material
body, corresponding to the rst point. These coupling-induced state
functions are usually used for setting up the connections with other
units, corresponding to the third point.
Regarding the current reactor modeling example, one can
introduce (corresponding to the second point) additional couplinginduced state functions for CA|x=0 of the uid phase and rd /ra of the
surface, by

CA|x=0
= CA|x=0 ,

(13a1)

rd

(13b1)

ra

= rd ,
= ra .

(13c1)

CA|x=0
, rd ,

ra

Here
and
are simply a coupling-induced copy of
their non coupling-induced counterpart; they should not exist if no
coupling is to be modeled. This is comparable to the introduction of
material stream variables which would not be needed if a chemical
process unit does not have any connections to other things. After
this treatment, Eqs. (13a)(13c) may be rewritten, following the
third point, as

CA|x=0
= CA0 ,

(13a2)


rAin = rd,

(13b2)

rAout

ra .

833

reaction. A transition probability denotes the probability for a particular type of events (i.e. adsorption, desorption, or reaction) to
occur at a site in the lattice, at a particular point in time. The laws
of a site relate the transition probabilities to other state functions,
which take the following forms:
ai = a (Ti , CA0,i , i ),

(14a)

di = d (Ti , CA0,i , i , lci ),

(14b)

ri = r (Ti , i ).

(14c)

Besides, the local coordination (lci ) of a site is related to the

occupation-site function of its neighboring sites k :
lci =

k .

(14d)

k=1

Here c is the number of its neighboring sites. {k } can be considered as coupling-induced state functions of this site. Each of
these coupling-induced state functions is, unsurprisingly, equal to the
occupation-site function of its corresponding neighbor, which is a
(non-coupling-induced) state function of the neighbor. This equality
relation can thus be viewed as a (simple) coupling mechanism law
(Denition 12) of the coupling between this site and the neighbor.
4.2.3. Inter-scale laws
Now we turn to the part of the original model that addresses
the connection between the above two (heterogeneous neighboring)
levels. We will show in the following that aggregation and disaggregation laws (Denitions 25 and 26) are present in the original
model.
4.2.3.1. Aggregation laws. As mentioned earlier, the aim of the noncontinuum modeling of Level 1 is to provide information needed for
determining the state function qj of the surface at Level 2, which is
the fraction (of all sites) of the occupied sites with local coordination
equal to j. This has been done through Monte-Carlo (MC) simulation. The MC simulation essentially solves the Master Equation
(Reese, Raimondeau, & Vlachos, 2001)

dP
[W P W P ],
=
dt

(15)

where P is the probability of the surface being in conguration ,

W is the transition probability per unit time of the surface changing from conguration to . Since Eq. (15) relates state functions of
the entire surface instead of individual sites, it should be regarded as
a (stochastic) law of the surface on Level 2. However, this law is not
actually included in the reactor model, because it cannot be solved
directly given the large number of possible congurations. Instead,
a MC simulation is employed, the execution of which essentially
provides the following relation between the state function qj of the
surface on Level 2 and a set of state functions of all sites on Level 1,
qj = f MC (k , ak , dk , rk ,

k = 1, 2, . . . , NC),

j = 0, 1, . . . , c,

(13c2)

(16)

This set of toplological connection laws now contains couplinginduced state functions only.

where k ,  ak ,  dk ,  rk are the same as in Eqs. (14), NC is the number of the sites in the lattice, c is the coordination of the lattice
same as used in Eq. (12). According to Denition 25, Eq. (16) can be
considered as a set of aggregation laws.

4.2.2. The non-continuum model of Level 1

On this level, the solid surface is depicted by a (large) number
of sites in a lattice. A site i in the lattice can be characterized by
state functions including the temperature (Ti ), the concentration
of A at the uid phase boundary adjacent to the site (CA0,i ), the
occupation-site function (i ), the local coordination (lci ), and nally
the transition probabilities ( ai ,  di , and  ri ) which are respectively of three micro-processes, namely adsorption, desorption, and

4.2.3.2. Disaggregation laws. From Eqs. (14a) to (14c) we can see

that characterizing one site requires information such as temperature and the concentration of A at the uid phase boundary adjacent
to the site. These state functions are determined by the links to the
corresponding state functions of the entire surface. In this particular model, the state functions of one site are considered equal to

834

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

those of the surface. This implies that a set of disaggregation laws

are employed in form of identity equations.
4.3. Discussion
The above two subsections have shown that the conceptualization framework for multiscale modeling is applicable to two
different examples. However, some adjustment to the original
model was required in each case. For example, in the rst model (cf.
Section 4.1.1.2), the concept of environment zone introduced by
the original author was dropped, and the part of model addressed by
this concept was re-expressed in the form of topological connection
laws that govern the couplings between the environment and internal zones of the equipment. In the second model (cf. Section 4.2.1.3),
several coupling-induced state functions added to the original set of
variables, in order to retain the standard form of topological connection laws as dened in the conceptualization framework. Generally
speaking, the model elements being adjusted in the two examples
were not inappropriate in their original form as far as these specic modeling applications are concerned. However, these original
formulations appear to have been adopted in a rather case-specic
or ad-hoc manner. After the adjustments, it was demonstrated that
the same elements could be formulated merely using the concepts
generically dened in the conceptualization framework.
5. Interpreting and comparing with some other
generalization work
The conceptualization framework developed in this work has
been intended to help to clarify fundamental concepts such as
scale and to present generic compositions of various kinds of multiscale models. In this section, we try to relate our conceptualization
framework to the some generalization-oriented work by other
researchers.
5.1. The classication of scale integration strategies by
Pantelides (2001)
In Pantelides (2001), a distinction is made between scale decoupling which deals with different scales of a system separately
without linking them together, and scale integration which refers
to strategies of modeling where different scales are connected.
According to the way these scales are connected, scale integration
is further classied into the following:
Serial integration. According to Pantelides (2001), this is a strategy in which a ner scale model is employed to provide parameters
or data needed by a coarser scale model. To generalize, one can say
that serial integration connects models at different scales which can
be solved sequentially. Applying our conceptualization framework,
serial integration results in (i) scale-connecting models (Denition
31) or (ii) scale-integrating models (Denition 32) with either aggregation law(s) (Denition 25) or disaggregation law(s) (Denition 26)
(but not both). More rigorously, we introduce
Denition D1 (Serial-integration model). Let  be a system with
levels V = {Vk |1 k n,n > 1}. Let Mm (,t) be a multiscale model of .
For any component z on level Vi , 1 <i n, let La (z), Ld (z), and Lm (z)
be the totality of its aggregation, disaggregation, and mereological
connection laws, respectively. Mm () is a serial-integration model, iff

The rst condition points to the case of a scale-connecting model.

The second condition refers to a scale-integrating model (by the rst
part of this condition) which however has either only aggregation
laws or only disaggregation laws (indicated by the second part of
this condition).
Parallel integration. According to Pantelides (2001), this is a
strategy which involves the simultaneous use of descriptions at
different scales applied to the same computational domain. The
results of one description form inputs to the other, and vice versa.
This is exactly the case of a scale-integrating model which has both
aggregation law(s) and disaggregation law(s). More rigorously, we
can derive
Denition D2 (Parallel-integration model). Let  be a system with
levels V = {Vk |1 k n,n > 1}. Let Mm () be a multiscale model of
. For any component z on level Vi , 1 < i n, let La (z) and Ld (z) be
the totality of its aggregation and disaggregation laws, respectively.
Mm () is a parallel-integration model, iff
n

{La (z)|z Vk } =
/

k=2

(i)

(ii)

Denition D3 (Simultaneous-integration model). Let  be a system

with levels V = {Vk |1 k n,n > 1}. For any component x on level Vk , let
L(x) be the totality of laws on x. Let SL (Vk ,t) be the L-structure of level
Vk . Let Lt (r) and Lc (r) be the totality of the topological connection laws
and that of coupling mechanism laws of a coupling r SL (Vk ), respectively. For any component z on level Vi , 1 < i n, let La (z), Ld (z), and
Lm (z) be the totality of its aggregation, disaggregation, and mereological connection laws, respectively. Let Mm () be a multiscale model
of . Mm () is a simultaneous-integration model, iff
{L(x)|x V1 } {Lt (r)|r SL (V1 )} {Lc (r)|r SL (V1 )} =
/
n

and

{L(x)|x Vk } {L (r)|r SL (Vk , t)} {L (r)|r SL (Vk , t)} = and

k=2

k=2

k=2

= .

k=2

k=2

{L (z)|z Vk } {L (z)|z Vk } {Lm (z)|z Vk } =

/

{Ld (z)|z Vk } =
/ .

k=2

Hierarchical integration. According to Pantelides (2001), this is a

strategy in which a ner scale model is formally embedded within
the higher-scale model to represent a set of relations among macroscopic quantities occurring in the latter. In serial integration and
parallel integration introduced earlier, an implicit assumption has
been that the models of different scales are solved or evaluated individually before serial or iterative connections between them are
made. In comparison, hierarchical integration stresses the direct
embedding of a ner scale model into a coarser scale model; the
entire system model is then solved as a whole. Thus, this integration strategy is mainly concerned with the solution (as opposed to
the composition) of a multiscale model. With regard to the composition of the model particularly in terms of the relations between
different modeling scales, the model solved using this strategy can
be either a serial-integration model or a parallel-integration model.
Thus, no additional concept is derived for this strategy.
Simultaneous integration. According to Pantelides (2001), this
is a strategy in which the higher-scale model is formed completely from ner scale descriptions. Applying concepts previously
dened in this paper, the result of simultaneous integration is
essentially a multiscale model in which only the lowest level has
laws; state functions of all other levels are determined only through
inter-scale laws. More rigorously, we introduce

n
{La (z)|z Vk } {Ld (z)|z Vk } =
k=2
k=2
n
and {Lm (z)|z Vk } =
/ , or
k=2
n
n
{La (z)|z Vk } {Ld (z)|z Vk } =
/ and
k=2
k=2
n
n
{La (z)|z Vk } = or {Ld (z)|z Vk }
k=2
k=2

and

k=2

The rst condition states that intra-level laws exist at the lowest
level of the system. The second condition states that the levels other
than the lowest one of the system are not modeled with any intralevel laws. Finally, the third condition states that there must be
some inter-level laws dened at the levels where intra-level laws
do not exist.

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

domain integration framework generically and rigorously, we can

derive:

5.2. The classication of integration frameworks by Ingram et al.

(2004)
Ingram et al. (2004) further expand the classication of
Pantelides (2001), mainly based on an analysis on the relations
between the regions of the system domain addressed by models of different scales. In general, a system domain can be viewed as
a dimension along which the system under investigation is decomposed, which can typically be temporal or spatial. In the work of
Ingram et al. (2004), multiscale models of process systems are particularly classied according to the spatial relationship between the
balance volumes that comprise the model.
Their classication includes serial, parallel, and simultaneous
integrations. Serial integration is further classied into simplication, transformation, and one-way coupling. The former two
treatments eventually result in a single-scale model by making a
micro-scale model part of a macro-scale model after either reduction or transformation. The case of one-way coupling, where there is
only one-way information ow from the micro-scale to the macroscale, essentially refers to serial integration model discussed earlier
(Denition D1). Simultaneous integration is equivalent to what
has been dened in Denition D3. Parallel integration has a different meaning from the term introduced by Pantelides (2001): it
now refers to the case where multiple models are applied to the
same system domain; each of them can be a multiscale model
of any kind. Certain links exist between the models in parallel.
Since this parallel integration framework primarily connects models which are not of different scales of one system, it is not analyzed
here.
There are other two types of integration frameworks introduced
by Ingram et al. (2004):
Multi-domain integration. This is a framework where the
micro-scale and macro-scale parts of a model occupy adjacent,
non-overlapping parts of the system domain. That means, more
generally, different parts of a system are modeled at different scales.
Viewing it in terms of levels of a system as formulated in this current
paper, it can be seen as the case where the model is composed of
(sub)models of different levels; each level is only partially modeled.
To explain this, the multi-domain model of the catalytic packedbed reactor given by Ingram et al. (2004) is analyzed as follows. The
entire model has been interpreted by Ingram et al. as a combination
of a macro-scale model of the liquid phase and micro-scale models
of catalytic pellets (cf. Fig. 10a). However, it can also be viewed as
a result of rst decomposing the reactor into the uid phase and
the catalyst phase and then decomposing the catalyst phase into
pellets (cf. Fig. 10b). Various laws are applied to the uid phase, to
the coupling between the uid phase and the catalyst phase, and to
the pellets. Additionally, the inter-scale relations are introduced to
connect the pellets with the catalyst phase.
This view can be manifested to a large extent by just looking at
the mass conservation equation of component A in the liquid phase,
as part of the model for that phase:
CA
u0 CA
km a
=

(CA CAP |r=Rp )

b x
b
t

(17)

The second term on the right hand side, if denoted by one

quantity, is essentially a coupling-induced state function. Furthermore, this term presents a coupling mechanism law that
characterizes the coupling between the liquid phase and the catalyst phase. In addition, CAP |r=Rp presents the link between the
catalyst phase and individual pellets.10 To characterize the multi-

10

835

This link is essentially an aggregation law: the concentration of A in the catalyst

phase at the interface with the uid phase is related to the concentration of A at the
surface of each pellet. In this reactor model, all pellets are geometrically identical,

Denition D4 (Multi-domain integration model). Let  be a system

with levels V = {Vk |1 k n,n > 1}. For any component x on level Vk ,
let L(x) be the totality of laws on x. Let C(x) be the composition of
x. Let Mm () be a multiscale model of . Mm () is a multi-domain
integration model, iff all the following statements hold:
(1) Vs Vk , k > 1, {L(x)|x Vs } = and {L(x)|x (Vk
/
Vs )} =
x Vk , k > 1, if L(x) = , then
(2)
y C(x), L(y) =
/
x V1 , if (y, x C(y) and L(y) = ), then
(3)
.
L(x) =
/
The rst condition states that at a certain level above the bottom,
some parts are modeled with laws, while no laws are applied to
the rest of this level. The second condition states, for any part at a
certain level above the bottom which is not modeled, there must
be some components of it (located at levels lower than this level)
which are modeled with laws. The third and last condition states,
if a certain part of a system has never been modeled at any level
above the bottom, it must be modeled with laws at the bottom level.
The combination of these three conditions ensures that the model
includes the models of all of its parts which may be developed not
at a single level (or scale).
If we restrict a multi-domain integration model to one in which
the parts not modeled at one level are always modeled at the level
immediately below it, then the condition for such a model can be
simplied as follows:

Vs Vk ,

k > 1,

{L(x)|x Vs } =

and

{L(y)|y (Vk Vs )} =
/ , and x, z, x Vs ,

z C im (x),

L(z) =
/ .

Embedded integration. This is a framework where the

macroscale model spans the system domain while the microscale
model describes only a portion of that domain. Applying this
framework to the example of catalytic packed-bed reactor yields
a model where the reactor as a whole is described by averaging the
uid phase and the catalyst phase, but where the process chemistry are supplied by a detailed modeling of the pellets. Viewing
this example from the perspective of system levels, one can see
that the reactor system is actually decomposed by the same level
hierarchy as in the multi-domain integration (cf. Fig. 10b). However,
the applications of various laws are different in the two integration
frameworks, as revealed in Table 1, where X indicates that there
are laws actually applied.
To give a generic and rigorous description of models generated
by this integration framework, we introduce
Denition D5 (Embedded integration model). Let  be a system with
levels V = {Vk |1 k n,n > 1}. For any component x on level Vk , let L(x)
be the totality of laws on x, C(x) be the composition of x. Let Mm ()
be a multiscale model of . Mm () is an embedded integration model,
iff:

x Vk , k > 1,
(1) L(x) =
/ , and
(2) y C(x), L(y) =
/ ,

and z C(x), L(z) = .

so that those pellets at the same axial position have the same surface concentration.
Otherwise, certain averaging treatment on different surface concentrations of all the
pellets at the same intersection might have to be involved in the aggregation law,
giving rise to a typical averaging law.

836

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

Fig. 10. Decomposition of the reactor. (a) Two-level decomposition and (b) Three-level decomposition.
Table 1
Applications of various laws in the reactor example.
Type of integration
framework

Multi-domain
Embedded

Items in the reactor system (cf. Fig. 10b) to which laws may apply
Reactor as a
whole

Fluid phase

Catalyst phase

Pellet

Coupling between uid

phase and catalyst phases
(i.e. two bulk phases)

Inter-scale connection
between reactor and
two bulk phases

Inter-scale connection
between catalyst phase
and pellets

X
X

X
X

X
X

5.3. The heterogeneous multiscale methods by E and Engquist

(2003a, 2003b)
E and Engquist (2003a, 2003b) generalize a variety of multiscale
modeling and computational methods and call them heterogeneous multiscale methods. Intuitively, a mathematical model
developed using a heterogeneous multiscale method applies different physics for different scales or levels of the system considered.
Using the concepts dened in our ontology, such a model is a scaleintegration model of a system involving heterogeneous neighboring
levels.
E and Engquist propose a general procedure of applying such
a method, which comprise three steps, namely reconstruction,
microscale model evaluation, and compression. In the reconstruction step, certain microscopic scale properties are determined
using some macroscopic scale properties before the microscopic
model can be evaluated. In the compression step, some macroscopic scale properties are obtained through processing certain
microscopic scale properties after the microscopic model has been
evaluated. Applying concepts dened in our ontology, the reconstruction step should be done essentially according to certain
disaggregation laws, while certain aggregation laws are for the compression step.
E and Engquist make the distinction between two types of
multiscale modeling problems. Type A refers to those in which a
macroscopic description is known but ceases to be valid in a localized region in space and/or time. Type B refers to problems in
which some macroscopic properties (called missing data) cannot be made available by the macroscopic model and therefore
one has to make use of microscopic models. This is typically the
case where some part of the macroscopic model (e.g. a constitutive
relation) is unknown or very difcult to obtain. Such a distinction can be expressed with denitions derived previously in this
paper: A model of Type A is a multi-domain integration model (cf.
Denition D4); a model of Type B is a parallel-integration model (cf.
Denition D2). Of course, to preserve what is originally meant by
E and Engquist, any model must be of a system involving heterogeneous neighboring levels.
6. Concluding remarks
Multiscale modeling has recently become an interesting topic
in various elds. The intended contribution of this paper is the

formulation of a conceptualization framework for improving the

understanding of multiscale modeling paradigms. The key element
in this framework is a clarication of the nature of a scale in the
multiscale modeling context. This framework has been developed
by following a three-step approach. The rst step, i.e. the starting
point of this approach, is a conceptualization of general systems.
Based on this conceptualization, the second step derives rigorous
denitions for precisely characterizing scales and inter-scale relationships. In the third step, the composition of a general multiscale
model is introduced, from which a number of multiscale model
types are derived, all on the basis of the conceptual foundation built
by the two previous steps. This conceptualization framework is presented in form of an ontology, which reuses part of the existing
BWW ontology (Bunge, 1977, 1979; Wand & Weber, 1990). In comparison with the BWW ontology, the major additions introduced in
the ontology include the notions of laws on intra-level couplings,
the environment and structure of a level, the laws on inter-scale (or
inter-level) relations, and nally the notion of a multiscale model
and its classication. The key concepts in the conceptualization
framework are illustrated by two rather complete multiscale modeling examples. Furthermore, the new framework is compared with
the results of some existing clarication attempts in multiscale
modeling. We demonstrate that the model types generalized by
the previous work are all special cases of the general multiscale
model dened in our conceptualization framework. Furthermore,
all of the previous model types can be rigorously dened using the
concepts offered by the conceptual foundation (i.e. Parts 1 and 2)
of our framework.
Different from the existing efforts towards the generalization
of various multiscale modeling paradigms, the approach we have
taken does not directly attempt to make an abstraction of some
existing applications, but rather starts with dening basic concepts
of general systems. If the existing efforts are more of an inductive
nature (i.e. one which infers of a generalized conclusion on multiscale modeling paradigms from particular instances of multiscale
modeling applications), our approach tends to be more deductive
(i.e. one which concludes about multiscale modeling paradigms
by following necessarily from general premises on systems, laws,
scales, etc.). We argue that the advantage of this approach is to
be able to provide in the rst place a common conceptual ground,
which can potentially be shared by multiscale modeling researchers
from different backgrounds in the course of reaching a sound
understanding to multiscale modeling. On the other hand, this

A. Yang, W. Marquardt / Computers and Chemical Engineering 33 (2009) 822837

conceptualization framework must be regarded as open. Thus, any

existing and future applications from different elds may bring
extensions to this framework, particularly to the types of interscale laws (in Part 2) and the composition and the classication of
multiscale models (in Part 3). We emphasize that any extension is
expected to be performed in a way similar to the current framework,
i.e. every new element should have a precisely dened semantics
and be logically related to the existing elements in the framework.
So far we have concentrated on the role of the conceptualization framework in contributing to the understanding to multiscale
modeling. In addition, the authors also envisage the potential of
this conceptualization framework in the area of computer-aided
modeling of systems, particularly with phenomena-based methods.
In this area, research is directed to the development of computer
tools which can automatically carry out part of the tasks involved in
building a mathematical model, by generating mathematical equations based on the physical description of the modeling target as
supplied by the human modeller (Marquardt, 1995). One difculty,
with respect to the reusability of such a tool, is that the tool is usually capable of modeling only the type of systems for which specic
domain knowledge has been incorporated in its implementation.
To address this problem, an ontology-based modeling approach has
been proposed (Yang, Morbach, & Marquardt, 2004), which allows
a tool to implement only a generic modeling logic while the domain
knowledge required for a particular modeling task is supplied separately to the tool as its input. Ontologies play very important
roles in this approach, because the domain knowledge is represented through a domain ontology, while a meta-ontology, as an
abstraction of different domain ontologies, forms the basis for the
generic modeling logic that the tool implements. It has been initially
demonstrated that such a tool is capable of processing correctly the
physical description of a concrete system formulated according to a
domain ontology, which itself instantiates the meta ontology (Yang
et al., 2004). In connection to this initiative, the conceptualization
framework reported here has the potential of being applied as a
meta ontology, based on which a computer-aided modeling tool
may be implemented such that multiscale modeling is supported by
the tool. To this end, it will become necessary to specify the ontologies (including the meta ontology) using a computer-processable
language such as OWL (McGuinness & van Harmelen, 2004). This
will also make the ontologies more accessible by the users through
ontology browsers. This has been the case of OntoCAPE which, as a
formal ontology for computer-aided process engineering, is already
represented using OWL (Morbach, Yang, & Marquardt, 2007). Ongoing research and development in these directions will be reported
elsewhere (Yang, 2008).
Acknowledgements
This work has been supported in part by the DFG in the Collaborative Research Center IMPROVE (SFB 476) and by the EC in the
COGents project (IST-2001-34431).

837

References
Bertalanffy, L. V. (1968). General system theory: Foundations, development, applications. New York: George Braziller Inc.
Bezzo, F., Macchietto, S., & Pantelides, C. C. (2004). A general methodology for hybrid
multizonal/CFD models. Part I. Theoretical framework. Computers and Chemical
Engineering, 28, 501511.
Bird, R. B., Stewart, W. E., & Lightfoot, E. N. (1960). Transport phenomena. New York:
John Wiley and Sons.
Bunge, M. (1977). Treatise on basic philosophy, Vol. 3. Ontology I: The furniture of the
World. Boston: Reidel.
Bunge, M. (1979). Treatise on basic philosophy, Vol. 4. Ontology II: A World of systems.
Boston: Reidel.
E, W., & Engquist, B. (2003a). The heterogeneous multiscale methods. Communications in Mathematical Sciences, 1, 87133.
E, W., & Engquist, B. (2003b). Multiscale modeling and computation. Notices of the
AMS, 50(9), 10621070.
Givon, D., Kupferman, R., & Stuart, A. (2004). Extracting macroscopic dynamics:
Model problems and algorithms. Nonlinearity, 17, R55R127.
Gruber. (1993). A translation approach to portable ontology specications. Knowledge Acquisitions, 5(3), 199220.
Guarino, N. (1995). Formal ontology, conceptual analysis and knowledge representation. International Journal of Human-Computers Studies, 43(56), 625640.
Ingram, G. D., Cameron, I. T., & Hangos, K. M. (2004). Classication and analysis of
integrating frameworks in multiscale modelling. Chemical Engineering Science,
59, 21712187.
Klir, G. J. (1985). Architecture of systems problem solving. London and New York:
Plenum Press.
Kulikov, V., Briesen, H., & Marquardt, W. (2005). Scale integration for the coupled
simulation of crystallization and uid dynamics. Chemical Engineering Research
and Design, 83(A6), 706717.
Li, J., Ge, W., Zhang, E. J., & Kwauk, M. (2005). Multi-scale compromise and multilevel correlation in complex systems. Chemical Engineering Research and Design,
83(A6), 574582.
Marquardt, W. (1995). Trends in computer-aided process modelling. Computers and
Chemical Engineering, 20(6/7), 591609.
McGuinness, D. L., & van Harmelen, F. (2004). OWL web ontology language overview.
W3C Recommendation. Available online at http://www.w3.org/TR/owl-features/.
Accessed 15 November 2008.
Morbach, J., Yang, A., & Marquardt, W. (2007). OntoCAPEA large-scale ontology
for chemical process engineering. Journal of Engineering Applications of Articial
Intelligence, 20(2), 147161.
Pantelides, C. C. (2001). New challenges and opportunities for process modelling. In
R. Gani & S. Bay-Jorgensen (Eds.), Proceedings of ESCAPE 11 conference (pp. 1526).
Elsevier.
Reese, J. S., Raimondeau, S., & Vlachos, D. G. (2001). Monte Carlo algorithms for
complex surface reaction mechanisms: Efciency and accuracy. Journal of Computational Physics, 173, 302321.
Rosemann, M., & Green, P. (2002). Developing a meta model for the
BungeWandWeber ontological constructs. Information Systems, 27, 75
91.
Uschold, M., & Grninger, M. (1996). Ontologies: Principles, methods and applications. Knowledge Engineering Review, 11(2), 93155.
Vlachos, D. G. (1997). Multiscale integration hybrid algorithms for homogeneous
heterogeneous reactors. AIChE Journal, 43(11), 30313041.
Vlachos, D. G. (2005). A review of multiscale analysis: Examples from systems
biology, materials engineering, and other uid-surface interacting systems.
Advanced Chemical Engineering, 30, 1.
Wand, Y., & Weber, R. (1990). An ontological model of an information system. IEEE
Transactions on Software Engineering, 16(11), 12811291.
Yang, A. (2008). From a generic idea to a generic set of tools: towards computeraided multiscale modeling. In Tenth intenational symposium on process systems
engineering (abstract)
Yang, A., Morbach, J., & Marquardt, W. (2004). From conceptualization to model generation: the roles of ontologies in process modeling. In C. A. Floudas & R. Agrarwal
(Eds.), Proceedings of FOCAPD (pp. 591594).