E E E Instrumentation and Measurement Technology Conference Budapest, Hungary, May 21-23,2001.

Recent Advances in Process Systems Engineering

Andreas A. Linninger Laboratory for Product and Process Design, Department of Chemical Engineering,

University of Illinois at Chicago, Chicago IL-60607, U.S.A. Phone: +I 312 996-2581, Fax: +I 312 996 -0808

Email: linninge@uic.edu, URL: http://vienna.che.uic.edu

Abstract - A crucial success factor in a technology driven market is coupled with the ability to understand and interpret new tech- nologies expeditiously and convert this knowledge into a competi- tive advantage. In search for procedures to advance corporate know-how faster, mathematical modeling is becoming an indis- pensable tool. The new challenge for systems research is to create a new breed of computer-based technologies for assistance andlor partial automation of the creative modeling process. Meta- modeling is a modeling paradigm for rapid computer-aided model generation of large multi-scale systems in the industrial practice. It proposes model building by means of phenomena-oriented mod- eling languages. Using the domain-specific language concepts, users compose process models by specifying the physical and chemical phenomena in a fully declarative fashion. Automatic interpretation of the high-level concepts of the model application leads to an equivalent set of system equations. Computer-aided model generation enables engineering teams to formulate highly structured process models in short amount of time. It also focuses the modeling effort onto the fundamental principles governing a process model without the need for explicit coding of all constitu- tive and balance equations. This article discusses recent advances and open challenges for computer-aided model generation.

Keywords - modeling and simulation, computer-aided model gen- eration

I. INTRODUCTION

Phenomena-oriented modeling is a research initiative for the computer-aided support and partial automation of the creative task of process modeling. Recent reviews give a comprehen- sive account of the state-of-the-art and new challenges for computer-aided model building [ 1,2]. This presentation fo- cuses on recent trends in computer-aided generation of mathematical process models.

This task belongs to the most challenging phases of the life- cycle of a process model [3]. Enabling technologies to sup- port life-cycle considerations are also analyzed in [4]. Com- puter-aided approaches for improved integration of software components involved in concurrent engineering is discussed in [5].

On-going research effort provide faster algorithms for simu- lation and optimization [6,8], address dynamic simulation of continuous and discrete processes, i.e. hybrid simulation [9,10], and point towards solution approaches in dynamic optimization [ 11,121. Open interfaces and data sharing among commercial and academic flowsheet simulators and physical

property packages are the target of initiatives such as Global CAPE-OPEN [I 3,141.

Several research groups have also developed computer-aided process modeling and simulation tools. The Model.la lan- guage was a pioneering effort towards model generation [15-191. Modkit [19-201 also offers tools for supporting the work flow in model development. The MODELLER project [21] proposed a formal framework for declarative model building. Model generation is the focus of systems research in hardware-software co-design. The multi-graph approach sup- ports the design and simulation of architecture and perform- ance of electric circuits and their functionality [22,23].

Despite recent progress, methodologies for the computer- aided support of the modeling activity itself have evolved slowly. There is still a need for further automation to enhance large-scale process modeling in the industrial practice. The following paragraphs reflect a topic list expressing the ex- pectations of practicing process engineers [24].

Fast equation generation. Commercial equation-oriented modeling and simulation is becoming an industrial routine. However, equation-oriented process modeling is still a labo- rious task requiring the implementation of all constitutive equations by hand. For process models of industrial size and complexity this may amount to encoding several thousands or more equations. This effort is time-consuming and error- prone. Could we generate mathematical artifacts without ex- plicitly editing every single balance equation?

Inclusion of high-level knowledge. Model equations are es- sential to the mathematical solution of process models. The mathematical relations alone correspond to a high level of abstractions ill-suited for conserving model intention, as- sumptions and simplifications. Can new model representa- tions incorporate both phenomenological knowledge and ra- tionale alongside the highly abstract mathematical expres- sions?

Multi-disciplinary approach. Larger and more complex proc- ess models require multi-disciplinary developer teams. Which type of information technology will provide an electronic workbench for distributed collaborative development of large projects?

0-7803-6646-810 1 /$I 0.00 0200 1 I EEE

1665

Consistency and Structural Analysis. Consistent formulation of large-scale process models commands meticulous book- keeping of variables, degrees of freedom and versatility in problem scaling and initialization. Is it possible to design algorithms that can examine consistency, solvability and structural properties of mathematical artifacts?

Few research contributions target at developing solution strategies for the latter more difficult problems. It is a conser- vative yet common opinion to consider modeling an art not amenable to systematic treatment by information scientific approaches. This work aims at challenging some aspects of this point of view.

Motivation.

Computer-based modeling approaches face a dilemma be- tween two extremes: specificity versus generalizy. Specialized modeling languages allow the fast model development for expert users [25-261. They provide a concise vocabulary of high-level language constructs suited for a specific modeling purpose. When applied outside their original scope, modeling becomes cumbersome or impossible for want of desired ex- pressions. Furthermore, rigid modeling languages fail to fully attend to the human factor. Model developers often prefer software environments in which language and functionality can be customized.

On the other hand, the equation-oriented approach permits the description of a wide range of phenomena via abstract mathematical expressions. Clearly, models based on purely mathematical abstractions are generic, but fail to capture valuable features of a process model such as the underlying phenomena, the intention or work progress. In consequence, both monolithic specialized modeling languages as well as generic equation-oriented systems fail to address the needs of new computer-based model generation.

One approach to overcome this dilemma in information mod- eling is entitled meta-modeling [23]. Meta-modeling is a modeling philosophy that supports custom-built modeling languages. Its main conjecture is user adaptability. It encour- ages users to meate their specific modeling language. Each user-defined modeling paradigm constitutes a distinct mod- eling dialect with domain-specific building blocks, semantic rules, and modeling activities. Each language definition also encodes the mapping between high-level language constructs into equivalent mathematical relations. The challenge for such an open environment is to uphold user innovation, while providing generic consistency control mechanisms.

This article gives an overview of new directions in computer- aided model generation based on meta-modeling. The funda- mental design of meta-modeling rooted in four building para- digms will be presented. The discussion will further describe solutions towards (i) data-driven automatic equation genera-

tion, (ii) an information model for representing the modeling activity and (iii) implementation of a consistency mainte- nance mechanism based on agents.

11. META-MODELING ENVIRONMENTS.

A meta-modeling environment is a platform for the ad-hoc definition of formal modeling languages, i.e. modeling dia- lects. Fig. l . offers an overview of a meta-modeling method- ology rooted in just four formal entities: (i) sme for structur- ing knowledge, (ii) pme for modeling the behavior of sme (iii) properties objects for automatic equation generation and (iv) Agents for consistency maintenance. The functionality of each element is discussed next.

Modeling Concepts.

Each modeling dialect needs to provide for modeling con- cepts to structure the modeling world. The proposed system offers two logical entities for structuring process knowledge and defines means for implementing their behavior, i.e. sub- stantial (sme) and phenomenological modeling elements @me).

I I /

r --- -1 ' 1 Phenomena I

A

Fig 1. Axioms of the Meta-modeling environment

Substantial modeling elements (sme):

Sme delineate a separate quantity of matter or information. Thus sme store the data nodes of a process model. The primal purpose of the sme is to introduce the process quantities and their connectivity. Sme can own attributes to improve the inner organization of the embedded knowledge. Attributes are user-defined features belonging to three distinct categories: (i) fundamental properties, (ii) auxiliary properties or (iii) associations. In addition, an actionlist is set of proce- dures for controlling the communication of each sme instance

1666

and its environment. The actions in the actionlist define the messages an instance of a sme is able to interpret.

Phenomenological Modeling Elements (pme):

Structural dependencies alone are insufficient for complex model building. An orthogonal property governing the be- havior of each balance envelope was proposed in the systems literature and its value recognized by Marquardt [27]. The pme ascertain the behavior of their associated sme. More spe- cifically, they provide the knowledge for the specific imple- mentation of balance equations. Thus pme introduce con- straints among the attributes of the underlying sme. This ac- tive task gives them procedural character when compared to the information structuring and storage function of the sme “data” objects.

Equation Generation via Fundamental Properties cfp):

A key objective in meta-modeling aims at consistent genera- tion of balance equations. In order to ensure their consis- tency, balances can only be synthesized for registered funda- mental properties (fp). All fundamental properties are derived from a virtual root class, see Fig. 1. They define basic entities that quantify the state of a modeling object.

Typical fp in chemical engineering systems include moles as a unit for the amount of matter, total enthalpy and fugacity as a measure for thermal and chemical potential. Definition of a fundamental property implies intent to exchange or equili- brate it among different sme objects along their connections. Independent of their domain, the full semantic meaning of the new quantity is completely delineated by (i) the attributes of the sme (ii) the knowledge provided by the union of pme and sme and (iii) the consistency maintenance mechanism.

Based on the fundamental property concept, the generation of balance equation can be implemented as a data-driven search over the network of modeling objects. Equations are gener- ated via selecting a particular type of fp from among the available object property slots. A data-driven search engine scans all sme in the network for the selected concept in order to synthesize the appropriate mathematical expressions. The necessary query information can be deduced from the funda- mental properties in the meta-modeling hierarchy as well as the published set of object properties. The actual instantiation of equation objects can thus be automated to a large degree.

The equation generation mechanism is depicted schematically in Fig. 2. The nodes belong to the class of substantial model- ing elements, sme, representing data objects. The arcs be- tween the sme nodes indicate associations to other objects within the model hierarchy. Each sme possesses fundamental properties of particular type symbolized by a particular shape of the property tokens, i.e. rectangle, rhomboid, etc. Complex fundamental properties may be linked to auxiliary properties

via constraints. Association between a phenomenon, pme, and an sme creates additional relations among fundamental and auxiliary properties of the sme. Pme encode the specific information, the fundamental property contributes to the proper balance equation of that particular type.

Automatic balance equations can only be generated for to- kens of type fundamental properties. The actual generation of balance equation entails a search of the object network ac- cording to a specific fp type. Each objects supplies the infor- mation that it contributes to the balance equation. This infor- mation typically entails mathematical expressions that cap- ture physical properties, constitutive equations or other con- text-specific object properties.

It should be noted that this equation “seeking” approach is entirely generic. Hence, the procedure can effectively synthe- size mathematical expressions for any modeling language, which was defined by the concepts outlined above. This pro- cedure also requires no knowledge about the particular se- mantic meaning of the fundamental properties or their corre- sponding balance equations.

Consistency Maintenance by Independent Symbolic Proce- dures -Agents

In the proposed methodology, the pme and sme contribute the specific aspects of process knowledge. Consistent association of sme and their subparts is supervised by Agents. They also

Behavior (pme):

Context-specific Relations among

Fig 2. Schematic of equation generation via data-driven search for fundamental propertied

1667

encode generic rules and restrictions for the fusion of a pme with its underlying sme context. Conflicts among competing pme as well as instructions for consistency maintenance that cannot be decided from within the sme and pme context are delegated to agents. Since they implement autonomous pieces of executable code and they interact with each other via a specific message exchange protocol, we adopted the term agents [28]. Their major task includes (i) the supervision of generic code generation via the association a pme and sme, (ii) enforcing the consistency mechanism of the modeling activity, and (iii) tracking of the modeling history.

The union of data (sme) and behavior @me) is triggered by invoking the user-defined agents available in the actionlist of each sme object. Meta-model designers can freely append new agents to the actionlist of the corresponding sme. Hence available model building actions are readily accessible via sme’s graphical interface. This fully declarative knowledge- building via the activation of custom-defined messages leads to an evolving process model via a sequence of modeling activities.

Tracking the Evolution of the Model - Model history.

The sequence of actions provides a natural vehicle for track- ing the modelers’ decisions. The modelhistory records ac- tions selected in an interactive modeling session in. Fig. 3 depicts the information flow for reporting the modeling his- tory. This feature serves two purposes: (i) Documentation of model activity, and (ii) Animation of the model evolution.

Documentation of model activity involves the persistent stor- age of type and attributes of each action of the user selection. The model-history records all relevant actions selected by the users. Therefore this list contains information equivalent to the knowledge embedded in the object network, sme and pme.

A similar model history approach was used in BatchDesign- Kit (BDK) project [29]. BDK offered an operation-based language for the conceptual design of chemical manufactur- ing recipes. In BDK, each relevant user operations is re- corded on the batch sheet of operational tasks. Hence, the BDK flowsheet and all its units with their associated mass and energy balances can be computed progressively.

Similarly, animation of the model evolution supports reitera- tion of each decision in the process modeling activity. The step-by-step simulation of the user-defined modeling activi- ties of the model-history gradually constructs the objects and their behavior on the workspaces. An information model to track all significant decisions in the model building phase is expected to facilitate collaborative model development. Ani- mation of the model evolution will also enable new model developers to grasp large-scale problems and their decisions incrementally.

111. APPLICATIONS VIA META-MODELS

Fig. 4. summarizes all phases of the three-staged approach of TechTool, experimental phenomena-oriented modeling envi- ronment currently under development in the Laboratory for Product and Process Design (LPPD) at the University of Illi- nois at Chicago in collaboration with VAI, Austria.

At the top of the hierarchy, meta-modeling defines the mod- eling language an its semantic rules. Fig. 5. depicts the lean class hierarchy of the simple modeling dialect. It recognizes but one fundamental property h o l e s . As stated earlier, the meta-modeling layer houses the means for the definition of new modeling concepts and features for their refinement or modification.

In the subsequent model building phase, see Fig. 4.-middle, axiomatic concepts defined at the meta-modeling layer can be aggregated or refined. This model building activity leads to a problem representation as a semantic network of object rela- tions and their corresponding mathematical equations. Since equations are generated automatically by means of aggregat- ing high-level phenomena, this approach it is also known as phenomena-oriented modeling.

Finally, the solution layer serves as an interface between model application and simulation experiment, cf. Fig. 4 - bottom. Hence, the mathematical artifacts generated at the phenomenological layer are solved in the solution layer. This entails the structural analysis of the generated mathematical artifacts, i.e. degree of freedom analysis, index analysis, etc. and the selection of an adequate solution strategy. More de- tails on structural analysis and solution are given elsewhere [10,30].

I CONNECT AGENT (UNIT-I, S1)

J / \ \ I 5

Workspace 1/ Model History Agent acts on

I sme and/or vme I I = Unit-1

I 1 OtherActions ..... I

Fig 3. Recording the model evolution and history

1668

,-

solution of differential equations using the ADCEND mod- eling language [32].

I 1 Agents - Modeling concepts

I ------- Flowsheets.

44 GMLCode - I Mathematical Expressions --. Simulation Experiments

Fig. 4. The three layers of the TechTool Modeling Environment

IV. OPEN QUESTIONS AND FUTURE RESEARCH

Multi-scale modeling. An open challenge for advanced mod- eling tools is the support of multi-scale models. This question asks for the integration of models at different length or time scales. As an example consider the incorporation of quantum mechanical computations into the business models of a phar- maceutical company. The challenge exceeds the mere aspects of software integration, which are already being addressed by standardization. The more intricate difficulty lies in sharing information among different levels of abstraction. This task is also known as vertical integration. New research should de- velop mathematical theories to conserve the salient feature of a detailed process model, when incorporated into an ampler context.

Glass-box solver routines. Solver algorithms are usually con- sidered detached black-box codes. Their internal organization is usually not accessible to model developers. When model- ing distributed systems, it becomes clear that the transition of procedural knowledge of a process model and solver strate- gies does not exhibit a clear boundary [31]. Consider grid generation for a finite element methods or interval refinement procedures required in certain classes of dynamic optimiza- tion problems. A smooth transition from procedural process knowledge to pretreatment of mathematical artifacts and customization of the solver would be desirable. Therefore software for assembling and manipulating solvers and proce- dural knowledge are needed. In a futuristic vision, one would wish for tools in which the solver routines could be imple- mented as just another instance of a process model. In the open literature, only a few contributions envision avenues towards algorithms modeling. A rare example discusses the implementation of an explicit Runge Kutta method for the

Consistency Maintenance. The discussion of the meta- modeling environment focused on the support of user defined paradigms. This implies however correctness and consistency of'the languages defined by the designers. An open question relates to formal methods to verify the consistency of user defined languages. This problem has been categorized as dy- namic semantics.

V. CONCLUSIONS AND SIGNIFICANCE:

An overview of developments in process model generation via a meta-modeling approach was given. Meta-modeling allows developers to freely modify, augment or restrict their custom-built design languages. It enables engineers to gener- ate mathematical representations of the systems they study without specifically typing the detailed system equations.

A specific meta-modeling methodology composed of only four axiomatic building blocks was discussed: (i) sme for structuring knowledge, (ii) pme for modeling the behavior of sme (iii) properties objects €or automatic equation generation and (iv) Agents for consistency maintenance.

A generic equation generation approach was proposed. It roots in a data-driven search for registered fundamental prop- erties. This mechanism requires no information beyond the axioms expressed at the meta-modeling level. Its generic na- ture suggests its suitability as a automatic equation generation mechanism.

An information model to represent significant decisions in model development was outlined. The suggested approach also facilitates a step-by-step re-iteration of all phases of the model evolution, thus enhancing opportunities for collabora- tive model development.

ACKNOWLEDGMENT

Partial financial support for this work by VAI, Austria, is gratefully acknowledged.

REFERENCES

[l]

[2]

Marquardt, W., "Perspectives On Life Cycle Modeling", AIChE Symp. Series, 323, Vol(96), pp 192-214,2000. Van Schijndel I. and Pistikopoulos, S; "Towards the Integration of Process Design, Process Control and Process Operability: Current Status and Future Trends", AIChE Symp. Series, 323, Vol (96), pp 99- 112.2000.

1669

Hackenberg, J, Krobb, C.; Schopfer, G; Von Wedel, L. Wyes, J; and W. Marquardt. “A Repository-based Environment for Life-cyle modleing and Simulation”, Proc. JSPS International Workshop on Safety-Assured Operation and Concurrent Engineering, Sponsored by Japan Society for the Promotion of Science, pp C1-C10, December 3-5, 2000 Yokohama, Japan. Batres, R.”Challenges in Concurrent Process Engineering”, Proc. JSPS International Workshop on Safety-Assured Operation and Concurrent Engineering, Sponsored by Japan Society for the Promotion of Science, pp B33-B-B45, December 3-52000 Yokohama, Japan. Batres, R. and Naka, Y.; “Process Plant Ontologies based on Multi- dimensional Framework, AIChE Symp. Series, 323, Vol (96), pp 433- 437,2000. Wright, S. “Algorithms for Linear and Non-lionear Programming”, AIChE Symp. Series, 323, Vol(96), pp 58-69,2000. Biegler, L, Alkaya, D., Anselmo, K, “Mulit-solver modeling for Proc- ess Simulation and Optimization”, AIChE Symp. Series, 323, Vol (96).

Grossmann, I; and Hooker, J; “Logic-based Approaches for Mixed Integer Programming Models and Their Applications in Process Syn- thesis”, AIChE Symp. Series, 323, Vol(96), pp 70-83,2000. Park, T. and P. I. Barton, (1996) State Event Location in Differential- Algebraic Models, ACM Trans. on Modeling and Simulation, Vol. 6. No. 2, PD. 137-165.

pp 125-137,2000,

[lo] Bahl, V: and A. A. Linninger; “Hybrid Simulation of Continuous Dis- crete Systems”, Proc. ESCAPE-10, Florence, pp 163 - 168, S. Piemcci (Ed.), Elsevier, Amsterdam, 2000.

[ l l ] Galan, S. and Barton, P.;”Dynamic optimization of Hybrid Systems”, Comp.Chem. Eng., Vol. 22, S183-S190,1998.

[12] Abel, 0. and Maryuardt, W; “Scenario-Integrated Modeling and Opti- mization of Dynamic Systems”, Aiche J,. Vol 46, No4, pp 803-823, 2000.

[13] Global Cape Open, Global CAPE-OPEN Consortium, 2000; http://www.global-cape-oDen.org/, 2000.

[14] Braunschweig, B.; Pantelids, C.; Britt, H; Sama, S.; Open Software Architectures For Process Modeling: Current Status and Future Per- spectives, ”, AIChE Symp. Series, 323, Vol(96). pp 221-235,2000.

[15] Stephanopoulos, G.; Henning, G.; Leone, H.; (1990a) MODEL.LA. A Modeling language for process engineering II - Multifaceted Modeling of process systems, Comp. Chem. Eng., Vol. 14, No. 8, pp 847 - 869.

[16] Stephanopoulos, G.; Henning, G.; Leone, H.; (1990b) MODEL.LA. A Modeling language for process engineering I - The formal framework, Comp. Chem. Eng., Vol. 14, No. 8, pp 813 - 846.

[17] Bieszczad, J., A. Koulouris and G. Stephanopoulos (2000). MODEL.LA: A Phenomena-Based Modeling Environment For Com- puter-Aided Process Design. MChE Symposium Series No. 323, 96, 438-441

PME Q) > .I Y

[18] Bieszczad, J., A Framework for the Language and Logic of Computer- Aided Phenomena-Based Process Modeling. Ph.D. thesis, MIT, Febru-

[I91 Bogusch, R. and Marquardt, W; “A formal representation of process model equations”, Comp. Chem. Eng., 21,10,1105-1115,1997,

[20] Lohmann, B. (1998). Ansaetze zur Unterstuetzung des Arbeitsablaufes bei der rechnerbasierten Modellierung verfahrenstechnischer Prozesse, PhD Thesis, D82, RWTH Aachen.

[21] Preisig, H. and M. Westenveele, “MODELLER - A computer tool suporting large-scale process modeling”, Paper 243b, AIChE Annual Meeting, Nov 12 - 17, Los Angeles, CA, 2000.

1221 Karsai G., Sztipanovits J. , Ledeczi A., Moore M.: Model-Integrated System Development: Models, Architecture and Process, 21st Annual Intemational Computer Software and Application Conference (COMPSAC), pp. 176-181, Bethesda, MD, August, 1997.

[23] Sztipanovits J., Karsai G., Biegl C., Bapty T., Ledeczi A.. Malloy D.: “MULTIGRAPH: Architecture for Model-Integrated Computing, Pro- ceedings of the Intemational Conference on Engineering of Complex Computer Systems“, pp. 361-368, Ft. Lauderdale,FL, November 6, 1995.

[24] Linninger, A; “Computer-Aided Evolutionary Model Development”, Dagstuhl Conference on Adaptation and Evolution in Embedded In- formation Systems, Nov. 2 - 6, ‘98, Schloss Dagstuhl, Saarland, Ger- many, 1998.

[25] Kettenis, D;”COSMOS: A simulation language for continuous, discrete and combined models”, Simulation, 58:1, pp 32-41, 1992.

[26] Mattson, S.; Elmyuist, H, and Otter, M; “Physical Systems Modeling with Modelica”, Control. Eng. Practice, Vol6,4, 501-505, 1998.

[27] Marquardt, W., “Dynamic Process Simulation- Recent progress and Future Challenges”, Proc. CPC IV, Texas, pp. 131- 180, USA, 1991.

[28] Russel, S. and Norvig, P. Artificial Intelligence - A Modern Approach, Prentice-Hall. New Jersey, 1995.

[29] Linninger, A. A. and G. Stephanopoulos; “A Natural Language Ap- proach for the Design of Batch Operating Procedures“, Informatica, Special Issue on Natural Language Processing and Multi-Agent Sys- tem, Vol22,4, pp 423 - 434, ISSN 0250-5596.1998.

[30] Chowdhry, S. and Linninger, A., Computer Aided Formulation of Dynamic Systems, Paper 219c, AIChE Annual Meeting, Oct. 31-Nov. 5, Dallas, TX, 1999.

[31] Oh, M. and Pantelides, C; “A modelling and simulation language for combined lumped and distributed parameter systems,” Comp. Chem. Eng. 20,6-7,611-633. 1998.

[32] Piela, P., Epperly, T., Westerberg, K. and Westerberg, A.; (1991). ASCEND. An Object-Oriented Computer Environment for Modeling and Analysis: The Modeling Language. Comp. Chem. Eng., Vol. 15, No. 1, pp 53-72.

ary 2000.

i

-E--- r/ “““ (from Newpackage

_-_I___--- --

MATERIAL Property: #MOLES: TYPE: [MOLES] IMPLEMENTION: as SumOfCOMPOUNDS

Fig 5 . Class Hierarchy for a simple modeling dialect

1670