You are on page 1of 18

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220356104

An artificial intelligence-based design tool for thin film composite materials.

Article  in  Applied Artificial Intelligence · July 1992


DOI: 10.1080/08839519208949957 · Source: DBLP

CITATIONS READS
10 224

4 authors, including:

Jon Sticklen Martin C. Hawley


Michigan Technological University Michigan State University
106 PUBLICATIONS   732 CITATIONS    130 PUBLICATIONS   2,074 CITATIONS   

SEE PROFILE SEE PROFILE

John Delong
Goodyear
10 PUBLICATIONS   195 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Web Based Modeling View project

Recent Epidemics and Public Health View project

All content following this page was uploaded by Jon Sticklen on 16 May 2014.

The user has requested enhancement of the downloaded file.


AN ARTIFICIAL INTELLIGENCE-BASED DESIGN T OOL
FOR THIN F ILM COMPOSITE MATERIALS

Jon Sticklen & Ahmed Kamel Martin Hawley & John Delong
AI/KBS Laboratory - CPS Dept Composite Materials Center

Michigan State University


East Lansing, MI 48824-1027

ABSTRACT

Current design approaches for composite materials may typically be categorized as


either (a) dependent on the experience of seasoned designers of manufacturing proto-
cols for composites or (b) dependent on fundamental studies of the materials involved.
We are currently undertaking research to combine these two approaches in an Artifi-
cial Intelligence-based problem solving system. We report research in progress aimed
at automating capability (a) by leveraging a known AI technique: Routine Design. We
have implemented a prototype system in the language of Routine Design (DSPL)
which covers part of the domain of protocol design for thin film epoxy-resin composite
materials. Results encourage further development. We conclude by describing the
overall problem solving architecture we are developing.

Page 1
AN ARTIFICIAL INTELLIGENCE-BASED DESIGN T OOL
FOR THIN F ILM COMPOSITE MATERIALS

1. INTRODUCTION

The emerging field of composite materials offers one pivotal area for the establish-
ment of a revitalized American industrial base. There is a key enabling step for realiz-
ing this potential in which Artificial Intelligence (AI) may prove to be important:
enabling a rapid “specifications to manufacturing” time; i.e. shortening the time be-
tween setting material specification, and successful realization of a material meeting
those specifications. This key step is largely dependent on an ability to capture
existing design knowledge and rapidly modify it in light of altered product specification.
Our ultimate research goals are to first capture fabrication protocols, and second, to
alter the protocols in a principled way to meet changed processed materials specifica-
tions.

In this report, we discuss research in progress towards meeting the first goal. In
particular, we describe part of an compiled level knowledge-based system we have
under development which will automate the design of fiber-reinforced thin film epoxy
resins. Other research has been conducted applying AI in the broad field of composite
materials but that research has addressed different problems than the one we
approach. (See for example [1; 2] ).

We start by outlining our motivation followed by a description of our domain: fiber-


reinforced thin film epoxy resins. We then give a description of the underlying AI tech-
niques used in our application followed by a detailed description of our implemented
system, and a description of test cases we have run. We conclude with a discussion of

Page 2
the enhancements in progress.

2. MOTIVATION

The area of polymer composite materials can be viewed as a modern generalization


of the area of metallurgy. Like metallurgy, the ultimate goal in composites is to
understand properties of physical materials and methods of material fabrication which
enable the creation of materials with desired properties. Metallurgy deals with
properties and fabrication of materials made from metals; polymer composites deal
chiefly with properties and fabrication of materials created largely from the realms of
polymer science and chemical engineering; the area is distinctly multidisciplinary. One
difference between metallurgy and composite materials in general is the basis for the
term “composites.” In metallurgy, typically either metals are reacted chemically to form
a new compound (as in steel) or metals are commingled at the molecular level (as in
metallic alloy formation such as tin). In composite materials, the processed material
typically will retain macroscopic areas resembling the starting materials. For example,
in epoxy-resin carbon fiber composite materials, the carbon reinforcing fiber retains
much of its individual identity after processing is completed.1

A typical chronology for designing a composite material is as follows [2]. First, mac-
roscopic properties which are desired in the completed composite are set. Properties
such as final material tensile modulus, resistance to acids and alkalis, electrical resis-
tance are parametrized. Based on these desired properties, the composite designer
proposes an initial plan for the production of the composite. This plan includes both an
ingredients list for all materials to be initially present, and a preliminary protocol which
states how the initial mixture is to be processed. Next, the composite designer esti-
mates how well the proposed composite design meets the initially stated, desired

1. Metal-matrix composites also exist, but the ideas expressed here are the same.

Page 3
properties. This estimate is typically carried out by actually producing samples of the
composite, then performing laboratory testing to determine properties of interest. Ulti-
mately, a proposed composite design will result in an actual material which can be
subjected to laboratory testing. One goal of composite researchers is to provide better
models for proposed designs in order to limit the number of candidate materials which
must actually be fabricated for testing. Following one round of design proposing, and
matching to specifications, successive rounds of redesign are usually required before
convergence of proposed composite properties to desired properties takes place.

Proposing fabrication plans for composites may engage problem solving strategies
from AI. In our system described below, Routine Design [3; 4] is used as the theoretical
underpinning to capture the compiled level knowledge a composites fabricator uses to
design new fabrication protocols based on past experience.

3. DOMAIN OF THE EXAMPLE

Composite materials can be grouped into three major classes: polymer matrix, ce-
ramic matrix, and metal matrix composites. In each of these types of materials, the
matrix material is typically combined with a reinforcement such as a fiber or
particulate in such a way as to achieve specified properties. We take as our problem
domain the first class, composites fabricated using polymer matrices. Polymer
composites can be further classified according to the type of polymer used, either
thermoset or thermoplastic, which largely determines the conditions employed in the
processing of the material. The example used in this work focuses on a particular
thermoset system, an epoxy resin matrix with fiber reinforcement, and more
specifically on materials with a thin filament epoxy resin matrix and fiber
reinforcement.

Page 4
4. ROUTINE DESIGN: CAPTURING COMPILED P ROCESS PLANNING

The theory on which we built our system is “Routine Design” [3; 4; 5] . Although Rou-
tine Design was initially developed for application in the realm of design problem solv-
ing for engineered artifacts, it has also been used to capture planning knowledge [5] ,
and it is this use of Routine Design on which we built for the generation of plans for
composite material fabrication. As noted in [5], when approached from a formal point
of view, design and planning are similar in that both are synthesis activities aimed at
producing a “product” which can achieve some goal.

4.1. INTUITIONS UNDERLYING ROUTINE DESIGN

The basic intuition underlying Routine Design is that in order to capture device de-
sign knowledge, the epistemic level terms of the approach should correspond to terms
that are “natural” in the domain being modeled. This intuition of Routine Design is
shared across the broader class of Generic Tasks as first suggested by Chandraseka-
ran and his colleagues [6; 7; 8] . A consequence that follows from this intuition is the use
of hierarchical structures of design specialists to perform design, each responsible for
designing a particular part of the overall plan. Hierarchies are used not because the
design is intrinsically hierarchical, but because hierarchical decomposition is a typical
means utilized to manage complexity of understanding how to fabricate a complex de-
vice.

4.2. INFORMATION PROCESSING T ASK OF ROUTINE DESIGN

The information processing task for planning is graphically depicted in Figure 1. The

Page 5
input is a set of planning constraints, and the output should be a full set of specifica-
tions for the required plan. The information processing task can be summarized as
follows:

• working on a problem that has been done many times before, each time with
different but similar requirements, until the problem solving knowledge has been
compiled into a form that allows efficient solution of the problem, and

• design proceeds with each sub-problem by selecting from previously known


sets of well-understood alternatives.

Design Design Design


Constraints Problem Specifications
Solver

Figure 1: Information Processing Task of Routine Design

4.3. KNOWLEDGE REPRESENTATION IN ROUTINE DESIGN

The general structure of a Routine Design problem solver is shown in Figure 2. A


design problem solver consists of a collection of design specialists. Each specialist is
responsible for choosing a small part of the overall design. A part of the decision mak-
ing carried by each specialist is to determine (locally) which of a number of plans to
carry out. S1 inFigure 2 has two such plans to choose between. Generally each spe-
cialist chooses just one of its plans. The actions that constitute a plan include doing a
calculation for a local value, satisfying a local constraint, and requesting another spe-
cialist to refine the current plan. For example, the left plan in S1 invokes the S2 spe-

Page 6
Specialist S1
Plan Plan •••

Planning Decision Planning Decision


Constraint Constraint
Call Specialist S2 Call Specialist S2, S3

Specialist S2 Specialist S3
Plan Plan ••• Plan Plan •••

Planning Decision Planning Decision Planning Decision Planning Decision


Constraint Constraint Constraint Constraint

Figure 2: General Structure of a design specialist.

cialist. If a plan fails then alternate plans are tried. If part of a plan fails then an attempt
is made to redesign the part of the plan that caused the failure. Potential causes of
failure (i.e., where to try to fix a plan) is precompiled into the specialist.

4.4. EXAMPLE USING ROUTINE DESIGN FOR PLANNING

To illustrate Routine Design, we will focus on a specific example [5]. The domain of
the example is air force mission planning. The planning process involves the
assignment of resources to various tasks. The resources involved are aircraft and their
stores located at airbases across the theater of operations. This example is limited to
a single type of mission, the Offensive Counter-Air (OCA) mission. An OCA mission is
an air strike directed against an enemy airbase. Figure 3 shows the hierarchy of
design/planning specialists for the mission planner.

Page 7
OCA

base aircraft

••• F-111 F-4 A-10

Figure 3: The hierarchy of specialists for the mission planner

The design activity begins when the OCA specialist is requested to plan a mission.
The OCA specialist contains a single design plan which first requests the base spe-
cialist to determine a base, and then requests the aircraft specialist to determine (and
configure) an appropriate aircraft for the mission. The base specialist selects a base
from a list of candidate bases geographically near the target. The aircraft specialist
uses considerations of threat types and weather conditions at the target to select an
appropriate aircraft for the mission. The aircraft specialist then runs its plan sponsors,
corresponding to the three different types of aircraft in turn in an attempt to select an
aircraft type. The corresponding specialist for the selected aircraft is then called for to
configure the aircraft.

The main point about this simple fragmentary example is that the framework of
Routine Design eases the conceptual burden of producing a plan for an often solved
problem by decomposing the overall plan into a number of more easily handled sub-
steps. Routine Design has been demonstrated to gives a reasonable framework for
such decompositions.

Page 8
5. ROUTINE DESIGN S YSTEM FOR THE D ESIGN OF E POXY RESINS

5.1. PROBLEM DECOMPOSITION AND REPRESENTATION

Figure 4 shows the Routine Design decomposition for the task of designing a thin
film, epoxy-resin composite with fiber reinforcement. It is important to realize at the
outset that this decomposition includes both material selection, and the selection of
fabrication protocols.

The top level specialist, CompositeMaterial, has one plan which first calls on the
Matrix specialist to select a suitable matrix to achieve the required properties. It then
calls on the Fiber specialist to select an appropriate fiber. Eventually, it calls on
CureConditions to design the appropriate cure conditions given the chosen materials.
Composite Material

Matrix
Fiber CureConditions

Epoxy Curing Agent


CarbonFiber GlassFiber

Amine Anhydrides Dicyandiamide

AliphaticAmine AromaticAmine

Figure 4: DSPL Specialist Hierarchy

The Matrix specialist also has one plan. This plan first calls on the Epoxy specialist
to select an appropriate type of epoxy.2 The Matrix specialist then calls on the
CuringAgent specialist to select a suitable curing agent.

2. Currently there is only one choice in our system: diglycidyl ether of bisphenol A (DGEBA).

Page 9
The CuringAgent specialist has three alternative plans from which to choose. Based
on the properties required from the composite material being designed (mainly the
required usage and the thermal properties represented in the glass transition tempera-
ture) this specialist chooses one of its three plans. This plan in turn calls on the Amine
specialist, the Anhydrides specialist or the Dicyandiamides specialist in order to select
the appropriate material to be used as the curing agent. The Amine specialist in turn
has two alternative plans from which to choose according to the desired usage of the
final product. These plans in turn call on either of the AromaticAmines specialist or the
AliphaticAmines specialist.

The Fiber specialist similarly has two design plans to choose from. This choice is
done based on the required tensile properties represented in the tensile modulus of
the required product. These two plans in turn select either the CarbonFiber specialist
or the GlassFiber specialist. These specialists are responsible for selecting the
appropriate type of carbon, or glass respectively. Finally, the CureConditions specialist
has one plan that is responsible for setting the appropriate cure conditions
(temperature, pressure, and time) based on the selected materials.

Again, it is important to emphasize that the DSPL problem solver we have devel-
oped for this small testbed problem depends on compiled knowledge.

5.2. SAMPLE PROBLEM

Consider a design in which we desire a composite material that has a glass


transition temperature of at least 200°C and a tensile modulus of 50 GPa., and which
will be used in an outdoor structure application.

Our system starts by calling on the top level specialist, CompositeMaterial. This
specialist calls on the Matrix specialist, which in turn calls on the Epoxy specialist

Page 10
which chooses DGEBA by default. The Matrix specialist then calls on the CuringAgent
specialist. CuringAgent has embedded knowledge that amines are a preferable choice
for civil engineering applications, and hence will make the appropriate selections.The
Amines specialist now examines the requirements and finds that aromatic amines are
more suited for civil engineering applications. Now AromaticAmines is called upon and
uses the required glass transition temperature to select diaminodiphenylsulfone (DDS)
which has a glass transition temperature of 220°C.

The CompositeMaterial specialist then calls on the Fiber specialist to select the
most appropriate fiber. This specialist examines the required tensile modulus and
because it is not a high value, it selects glass fiber as appropriate. Next, GlassFiber
chooses E-glass as the type of fiber to be used because it has a tensile modulus of 52
GPa.

The CompositeMaterial specialist will then call on the CureConditions specialist


which inspects the materials selected to this point, and selects a cure cycle of 1 hour
at 150°C followed by 3 hours at 220°C at atmospheric pressure.

Now, if we changed the initial requirements to require the designed composite to


have a tensile modulus of 100 GPa. instead of 50 GPa. while keeping the other
requirements constant, we will notice that the design process takes the same actions
as before until it comes to the Fiber specialist to select a fiber. At this stage, we notice
that the Glass design plan is ruled out because of the higher tensile modulus. The
CarbonFiber specialist is then called on, which chooses AS4-Carbon because it has a
tensile modulus of 145 GPa.

If we now relax the requirement on the glass transition temperature to be 150°C


instead of 200°C and rerun our system, we find that the AromaticAmine specialist now
chooses meta-phenylendiamine (MPDA) which has a glass transition temperature of

Page 11
160°C. This choice is based on a precompiled knowledge that MPDA is the preferred
choice of aromatic amines. The real-life reason for this choice is that MPDA has a
relatively low cost compared to DDS which has a higher glass transition temperature
and which was chosen in the first case. This choice of another curing agent further
affects the decision made by the CureConditions specialist which selects a cure cycle
of 2 hours at 125°C followed by 2 hours at 175°C at atmospheric pressure.

6. DISCUSSION

The testbed system described above is our initial step in implementing a compre-
hensive problem solving architecture to automate the task of material selection and
fabrication protocol design for composite materials. The example we discussed is
straightforward from a composite materials viewpoint. However, the straightforward
nature of the problem solution is strongly facilitated by the DSPL framework we
developed. We are encouraged by our initial results.

As a first extension, we are currently broadening the coverage of the current DSPL
implemented problem solver. Because of the modular nature of a DSPL representa-
tion, additional compiled knowledge will be straightforward to add. We have under de-
velopment now additions to the described DSPL problem solver that include: coverage
over more epoxies, and more fiber types.

The compiled-level DSPL problem solver is only one part of our proposed architec-
ture. Because the area of composite materials is a relatively new one, in which com-
piled expertise is not always available, a Routine Design problem solver cannot by it-
self be assured of producing reliable results. The reason is familiar; the compiled level
problem solver may encounter novel design requirements. To deal with such novel
requirements (which cannot be successfully addressed by using compiled level

Page 12
knowledge only) we will add a second major component to our approach: a qualitative
device model for both the proto-composite material, and for the finished composite
material.

For this next phase of our project, we will use an emerging Model-Based Reasoning
methodology for the representation of the qualitative model: Functional Reasoning [9;
[10; 11]. From an AI viewpoint, the expected integration of Routine Design and Func-
tional Reasoning will be similar to that already performed by Sticklen and
Chandrasekaran in which Hierarchical Classification and Functional Reasoning were
integrated into one diagnostic problem solver [12; 13]. The analogy to the previous re-
search will be that one action of the compiled level unit (Routine Design) will be to
focus problem solving of the deep level unit (Functional Reasoner) in the same sense
that compiled level classification problem solving played the role of focuser for func-
tional reasoning previously. The proposed interaction is graphically shown in Figure 5.

• • • • • •Task: with stated


Routine constraints,
Designer parametrize a
design template

proposed
change

result of
proposed
change

• • • • • •Task: with given starting


Functional conditions, derive
Reasoner consequences on
device

Figure 5: Interacting Routine Designer and Functional Reasoner

Page 13
There are two distinct areas of scientific benefit we expect as a result of the re-
search we have begun: benefits in composite materials, and benefits in AI. Completion
of our research track will have manifold benefits in the area of composite materials.
These benefits fall in the general categories of design and research assistance,
technology transfer, and education. As a design tool, the composite materials
Knowledge Based System we will develop will provide a means of material design and
cost optimization and provide a framework for further exploration into process
optimization. Such systems will suggest development of new materials and suggest
new directions for chemistry and process development. The knowledge captured by
the system will promote technology transfer, especially from the performance driven
aerospace applications in which significant technological breakthroughs are achieved,
to cost driven applications such as in the consumer industry. In education, use of our
framework will facilitate training of students and composite industry professionals in
the interdisciplinary area of composite materials, where exposure to all of the
necessary contributing fields is often inadequate.

In AI, successful completion of this research will provide a tangible advance. One
current trend in AI is to attempt trouble shooting and other types of reasoning from so-
called first principles [14; 15] . Previously, we have demonstrated an alternative in the
area of diagnostic medicine [9]: that robust problem solving can be achieved by com-
bining deep level problem solving (to handle novel situations) with compiled level
problem solving (to achieve efficient problem solving). The work we report here is the
first phase of similarly motivated work to develop cooperative problem solving be-
tween deep and compiled level units in the area of composite material fabrication pro-
tocol generation.

Page 14
7. ACKNOWLEDGEMENTS

Our work is strongly interdisciplinary between Artificial Intelligence and Composite


Materials. In that, we are fortunate to have on the MSU campus a strong research
center broadly exploring the area of composite materials.

Work reported here was supported directly by the Research Excellence Fund, State
of Michigan. Dr. Sticklen was also supported for this research as a Fellow by the
Ameritech Corporation. In general, parallel and synergistic research is underway in the
KBS Laboratory of MSU which is supported by the McDonnell Douglas Corporation.
Equipment support for the AI/KBS Laboratory, MSU is in part from Apple Computer.

8. REFERENCES

[1] LeClair, Steven R.; Abrams, Frances L.; Matejka, Richard F. Qualitative
Process Automation: Self-Directed Manufacture of Composite Materials. AI
EDAM; 1989. 3(2): 125-136.

[2] Venkatasubramanian, V, Lee, Young, & Gryte, Carl G. (1987) Design of Poly-
mer Composites: A Knowledge-Based Framework. Paper 97e, AIChE Annual
Meeting, New York, November, 1987.

[3] Brown, D.C. (1984) Expert Systems for Design Problem-Solving using De-
sign Refinement with Plan Selection and Redesign. Ph.D. dissertation.
Computer Science Department, Ohio State University.

[4] Brown, David C. and Chandrasekaran, B. (1986) Knowledge and Control for a
Mechanical Design Expert System. IEEE Expert. July, 1986. Pp 92-100.

Page 15
[5] Chandrasekaran, B., Josephson, J., Keuneke, Anne, & Herman, David. (1989)
An Approach to Routine Planning. International Journal of Man-Machine
Studies. (pp. 377-398)

[6] Chandrasekaran, B. (1983). Towards a Taxonomy of Problem-Solving Types.


AI Magazine. 4 (pp. 9-17).

[7] Chandrasekaran, B. (1985). Generic Tasks in Knowledge-Based Reasoning:


Characterizing and Designing Expert Systems at the “Right” Level of
Abstraction. Proceedings of The IEEE Second Annual Conference on
Artificial Intelligence Applications.

[8] Chandrasekaran, B. (1986). Generic Tasks in Knowledge-Based Reasoning:


High-Level Building Blocks for Expert System Design. IEEE Expert. (pp. 23-
30).

[9] Sticklen, Jon, Chandrasekaran, B. & Bond, W. (1989) Distributed Causal


Reasoning. International Journal of Man-Machine Studies. 1989.

[10] Sembugamoorthy, V., & Chandrasekaran, B.Functional Representation of


Devices and Compilation of Diagnostic Problem-Solving Systems. J. Kolodner,
& C. Reisbeck (ed), Experience, Memory, and Learning. Lawrence Erlbaum
Associates. 1986.

[11] Sticklen, Jon, & Chandrasekaran, B. Use Of Deep Level Reasoning in Medical
Diagnosis. In Proc. of The Expert Systems in Government Symposium.
McLean, Virginia. 1985

[12] Sticklen, Jon & Chandrasekaran, B. Integrating Classification-Based Compiled


Level Reasoning with Function-Based Deep Level Reasoning. Applied
Artificial Intelligence: 3(2). 1989.

[13] Sticklen, Jon. MDX2: An Integrated Medical Diagnostic System. Ph.D.


dissertation. Columbus, OH: Ohio State University. 1987.

Page 16
[14] Davis, R. Diagnostic Reasoning Based on Structure and Behavior. Artificial
Intelligence. 24 (pp. 347-410). 1984.

[15] Reiter, R. A Theory of Diagnosis from First Principles. Artificial Intelligence.


32 (pp. 57-95). 1987.

Page 17

View publication stats

You might also like