Professional Documents
Culture Documents
Design
Jean-Luc Koning
May 28, 1994
CML-R1-TR-94-22
Robotics Institute
Carnegie Mellon University
5000 Forbes Avenue
Pittsburgh, PA 15213-3890
1 Introduction 6
1.1 What. is this Report About? . . . . . . . . . . . . . . . . . . . .. 6
1.2 Quick Survey of the Problem . . . . . . . . . . . . . . . . . . .. 6
1.3 Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . .. 1
1.4 Select.ion and Substitutio~lof Material . . . . . . . . . . . . . . . 8
1.5 Outline of the Report . . . . . . . . . . . . . . . . . . . . . . .. 10
3 Rcpresentation of t h e Charts 16
3.1 Four Possible Representations . . . . . . . . . . . . . . . . . . . . 16
3.1.1 Continuous envelopes . . . . . . . . . . . . . . . . . . . . 16
3.1.2 Discret envelopes . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.3 Rectangles . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 . 1 . 4 Polygons . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.5 Fuzzy contours . . . . . . . . . . . . . . . . . . . . . . . . 2U
3.2 Polygons for Material Domains . . . . . . . . . . . . . . . . . . . 21
3.2.1 Specifying property boundaries . . . . . . . . . . . . . . . 22
3.2.2 Compiling the d a t a . . . . . . . . . . . . . . . . . . . . . . 23
1
4.3.1 Queries for accessing materials classes . . . . . . . . . . . 34
4.3.2 Combining matching degree . . . . . . . . . . . . . . . . 36
4.3.3 Example of material selection . . . . . . . . . . . . . . . . 37
6 Conclusion 44
2
List of Figures
1.1 Sccsaw in a playground . . . . . . . . . . . . . . . . . . . . . . . #
3
A.2 Matcrial selection chart for light fracture-resistant components
(courtesy ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5I
4 . 3 Material selection chart for cheap, stiff c0rnponent.s (courtesy). . 52
-4.4 Material select,ion chart for cheap, strong components (courtesy). 53
A.5 ?daterial select,ion chart for components resistant to corrosion
(courtesy). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4
List of Tables
5
Chapter 1
Introduction
6
material selecsion arc developed and refined. A large part o f cngineeririg de-
sign: however! is case-baTed. Yew designs are created by adapting and reusing
p a r k of previous designs in a new context, or application. Part of the process
of adapting a prior [base) design to a new (target) design is the subst,itntion
of materials in the base, to suit the conditions of the target application. This
involves re-evaluating the reaSons why a particular material was used in the
ha..e. and whether the same or new criteria are relevant in the target. Once
the ncw criteria are established, appropriate materials are selected and suhsti-
tilted. Thc snbslitution step, hcnce, subsumes the selection task. For examplc,
if a honsohold Oven were being adapted for an aerospace application, one would
haw t,o reevaluate the material choices based on strict weight considerations,
possibly replacing heavy steel components by ceramic coated aluminum o n e s
t-
The see-saw ca5e is represented in the database as a frame where the varions
attrihules (dimensions, material types, etc.) as well as their intrinsic reason(s)
are stored (sce Section 4.1).
Consider the target application for the see-saw: a hot and cold water faucet.
Figure 1.2 depicts n faucet that utilizes the see-saw principle. The rotation
7
in t,be vertical plane regulates t,he relative quantity of hot and cold water. The
vertical nroLionof its hinge increass or decreases the combined flow. The Caucct
represents a target, application where the environmental conditions are quite
dilTercnt from that of the see-saw in the playground.
Figure 1.2: Hot and cold water faucet using the see-saw concept.
Although the (see-saw) principle is identical in both cases the material the
faucet, should be made of does not necessarily meet the same rcquirements. On
the onc hand some requirements for rhe see-saw may no longer be valid: and
on t,he other hand some new requirements may crop up. For example, since
the heam is of a much smaller dimension its cost may become less important.
Also, the material may have to be resistant to certain environments if used with
chemical products.
Clearly, thc problem is to come up with the material(s) that meet, the set of
requirements. This calls for typical engineering knowledge and reasoning ability
on materials. Both types of informatlion are provided in the materials selection
charts settled by M. Ashby.
8
Strength L1
I b
I
Density
that do not correspond to any actual material. Secondly, one rarely looks for
materials with a precise value of the combined property but rather Cor materials
within acceptable ranges of values, some of these values being more appropriate
t,han others.
This notion is exemplified in Figure 1.3. A designer may look for materials
t h a t lie between two oblique lines L1 and L z . The area between L1 and La is the
preferred rangc of values for the combined property. The ranges of values next
t,o lines L1 and L.1 (shown by the shaded area on Figure 1.3) represent values
closc 60 the preferred range. Materials in these areas should not be t,otally
rejected although they do not pertain to the set of the preferred materials. This
defines three conceptually different regions: (1) above the upper shaded area or
under the lower shaded area are the materials one is definitely not interested
in, (2) materials between L1 and L2 definitely meet the required combined
property. arid ( 3 ) the materials crossed by one of the shaded are= but ontside
t,he interval [ L l , L z ] are acceptable to a certain extent (they are not totally
rejected). Actually, in this third region, the closer the material t o t,he stripe
[L.1:L?] (the preferred threshold) the better.
This type of imprecision can easily be handled by making use of fuzzy set
theory. It would enable the ranking of material classes with respect to their
degree of preference given some property. For example, one could say that
materials class B meets the requirements more adequately than materials class
A arid much more than materials class C. Also, the fuzzy set theory provides a
means t o aggregate several material rankings relative to different requirements
(see Sect,ion 4.3)
9
1.5 Outline of the Report
The rest of the report is organized as follows. Chapter 2 presents the usage of
the materials selection charts. In Chapter 3 several computer representations of
the charts are given and discussed. One of them ir selected to be implemented
i n the system that will actually perform material modifications. Chapter 4 is
lhe key chapter of this document. It introduces the not.ion of justifications
attached to a case and highlight its importance. This same chapter explains
in detail fuzzy orders of magnitude and also focuses on the system capabilities.
Thc queries the syst,ern enables are fully explained and an example of material
selection is shown Chapter 5 gives an overview of the program. T h e purpose
of t h a t chapter is t o e a e the possible re-use of the code written for our system.
Chapter 6 gives some concluding remarks and states the place such a system
has in a case-based design application like CADET. Appendix A shows the
material selection charts that our system uses so far. Appendix 8 s u m s up all
the materials classes and materials properties taken into account by the system
10
Chapter 2
'l'his section presents the material selection problem in mechanical design. For
a more detailed explanation see [Ashby and Jones, 19801, [Ashby, 19911. Only
some aspects relevant to an understanding of the system are presented here.
th.e conceptual dessgn stage; approximate data need for selecting the pos-
sible range of materials.
t k e detailed design. slagr; making use of still higher levels of precision and
detail! such as accurate d a t a from handbooks and data sheet,s provided by
material suppliers.
This paper is only concerned with the first two levels. The purpose of the
system we describe here is to find materials that parts of a device should be
made of and t,hat best fit t h e specifications ror t,he whole artifact. Finding the
exact best material is not that important sirice this choice may still be altered
due to the taking into consideration of parts interactions.
11
2.2 Material Classes
At Ihc conceptual design stage all niatcrials should be considered as potential
candidat,es. They are conventionally c.lassified into nine broad classes: engineer-
ing alloys, engineering polymers, engineering ceramics, engineering composites,
porous ceramics, glasses, woods. elastomers, polymer roams (cf. Section B.1).
Making a select,ion among these material classes is usually driven by neces-
sary constraints s u c h as: i t must operate a t high temperature, must, be cheap,
light, for instance. In other words, materials must satisfy certain individual
properties. At the embodiment design stage, the engineering designer seeks to
identify the design-limiting combination of material properties. One way of ef-
ficiently examining the relationships between the properties of all nine classes
is by referring t o Ashby's material selection charts.
combination).
It is then easy t o pick off the subset of materials that are optimal for each
loading geometry. For any straight-edge laid parallel to the E ? / p = C line, all
the materials which lie on the line will perform equally well as a light plate loaded
in compression; those above are better (they can withstand greater loads), those
below are worse. If the straight edge is translated towards the top left corner
of the diagram while retaining the same slope, the choice narrows. The same
procedure, applied to the rotating cylinder ( E / p ) or rod in bending ( E i l p ) ,
leads t o different equivalences and optimal subsets of materials.
Classes of materials crossed by the same line are equally good with respect
t o the criteria the straight edge represents. The criteria for optimal materials
selection that we are going to use here are summarized in Table 2.1. They all
13
appear on one of the malerial selection charts. The purpose of Table 2.1 is to
display the most irriportant criteria given the mode of loading of the artifact.
Among t.he mechanical and thermal properties, there are Len that are of
primary irnportance, both in characterizing the material, and in engineering
design. In this paper, we have used Ashby’s charts that refer to the following
properties: young’s modulus, density, strength, fracture toughness, cost, and
corrosion resishnce. Thcse charts are explained in Appendix A .
14
Table 2.1: Sample of property-combinations which determine performance in
design (courtesy).
15
Chapter 3
.._....,__
C
With t,his method a chart becomes a two dimensional array whose elements
contain Lhe name(s) of the classes of materials that would be plotted on the
corresponding sqnare. T h e logarithmicscales used for the charts give rise to Lwo
problems. First, if one wants to use the same method with the logarithmic grid
one ends up approximating too roughly the regions showing material classes. See
Figure 3.2.. The squares do not have the same size. Second, one can avoid t,his
drawback by deciding to take into account a linear scale. But this alternative
lcads to storing a huge amount of data. For instance the scale for the strength
property goes from 0.1 to 10,000 which makes 100,000 echelons of size 0.1
(cf. Figure A.1).
'These two reamus prompted us to abandon lhis representalion.
3.1.3 Rectangles
'I'he t.wo representalions suggested so far require a great deal of data and com-
put,ation. Their advantage is that they provide accurate information.
For the sake ol simplicity one c.an represents these regions by rectangles
where parallel edges correspond to lower and upper bounds of a class in the
chart.
17
Figure 3.2: Discret representat,ion of a region on a logarithmic grid.
With this representation one faces now a lack of precision. Figure 3.3 high-
lights this drawback. The same representation would be identical for far different
classes of materials. The shaded area does not correspond t o any material of
the class.
......... ........
......... ........
3.1.4 Polygons
The given shape of t,he material class envelopes resemble rounded polygons. One
may t,ake advantage of this fact and approximate them by true polygons where
each edge corresponds to bounds of the properties shown on the chart. In the
case of Figure 2.1 the chart presents five properties: young’s modulus, density,
and t,hree types of stiffness ( E l p , E i j p , E f l p ) . Thus, any class of materials
on the chart would be represented by an IO-edge polygon. Figurc 3.4 shows
18
a n example of such a polygonal rrpresentatlon with only one type of stiffness.
This is R 6-edge polygon, the four other edges are introduced by the other two
guide-lines.
On rishby's charts material class envelopes are naturally either vertical (par-
allel t o the ordinate axis): horizontal (parallel t o the abscissa axis) or slopcd to
the righl (rather parallel to t,he combined property). Therefore choosing these
axis to define the polygons leads to a reasonable approximation of t,he actual
regions. The worst c a x (rigure 3.5) would arise when a class envelope iu normal
to the slant-wise axis. This case doesn't appear on any material selection chart.
Y'his ~ y p cof representation does not have t,he drawback of requiring a huge
amount of data, and also allows a rather fair approximation of t h e regions it
represents.
19
3.1.5 Fuzzy contours
Ashby's charts are not a source of numerical data for precise analysis, they are
approximate. The above snggested represen tation does not take into consider-
ation the imprecision of the charts. Fuzzy set theory (Dubois and Prade, 19881
[Zadch, 19651 can help remedy this problem.
Let us first recall some hasic notions. A set is fuzzy if i t s boundaries are
not precisely delimited. This concept of fuzzy sets generalizes the concept oC
ordinary sets. A fuzzy set A on a universe Cl is characterized by its membership
fimct.ion: p~ : Cl + [0,1]. For all w belonging t o Q , pa(*) is t h e extent t,o which
w belong l o .A. If p n ( w ) = 0 it doesn't, belong to it at all, and if ~ A ( u )= 1: it.
belongs fully. If p n ( w ) E ( 0 , l ) : it belongs more or less to A .
Let us illustrate the difference between crisp and fuzzy boundary represen-
tat.ion with an example drawn from the material selcction domain. where A is
the range of densities for the class of engineering polymers and C2, in this case.
is the set of all possible densities. When A is a crisp set, the representation GC
ils characteristic function looks like on Figure 3.6 (..e suppose that the class of
densities of engineering polymers range from .8Mg/rn3 to 2.0hfg/m3).
" I t - - - - - - - : R
20
T h e set { w E n I p~ > 0} is called the support of A; this is the set of v a l u a
for w thaL belong to A a t least to a certain extent ((0.8,2) in Figure 3 . 5 ) . 'The
set {a E 0 I p~ = 1) is called the core of A; this is the set of values for w that.
cornplet.ely bclong t o .4 ((1.0,1,8) on Figure 3.7).
In a trvo-dimension universe the core and the support of a fuzzy set become
surfaces. This is the case encountered with Ashby's charts since the clnsses of
materials are plotted against two properties at the same time. The membership
function of fuzzy sets looks then like a trapezoid volume (see Figure 3.8).
/
Figure 3.8: Fuzzy representation of Ashby's charts,
21
Here, only a crisp representation by polygons h s been envisaged. Let u s
uoticc that the fuzzy counterpart would not have led t o much more complex
computations. In that case material domain would have been represent,ed by
t,rvo polygons. \t‘e didn’t choose the frizzy representation because Ashby’s chart,s
only give information concerning the contour delimiting the support (there‘s n o
way to infer the core of such sets given the charts). Indeed, no material of a
class may be found outside the envelope of the class.
In the pdygon-type representation, edges of material domains corrcspond
to upper and lower bounds of the properties shown on t h e chart. Encoding the
charts t,hen conies down to a two step process. First. the various bounds must
be defined for any property of a material class. Second, these data must h e
compiled to ease information retrieval.
This way of storing d a t a is not suited for fast access to the ‘best’ material
c l a s . One has to look up every possible class before retricving the right one.
On the other hand with Ashby’s charts one directly looks a t the correct, region
of the chart. If one wants a high (resp. low) value of the property shown on the
horizontal axis then one focuses on the right-hand (resp. left-hand) side of the
chart without even taking into consideration materials on the left.-hand (resp.
right-hand) side. If one wants a high (resp. low) value of the property shown
on the vertical axis then one focuses on t,he top (resp. bottom) half of the chart
u-ithout even taking into considcration materials on the bottom (resp. top) half.
If one wants a high (rcsp. low) value for t h e combined property one focuscs on
the top left (resp. hott,om right) corner of the chart. Sce Figures 2.1, A . l , A.2,
A.3, A.4.
22
'Ihe tabular d a t a is thcn compiled to speed access.
In ordcr to improve the retrieval of the right material given some charac-
teristics ( e & low stiffness, high strength! etc.) pieces of solution may be pre-
compiled and attached to the threshold values. By pieces of solution we mean
information such as (1) the material classes that appear beforr the threshold,
(2) t,hr ones that appear afler. and (3) the ones that overlap it. These threc
lists are called less, greater and equal respectively. See Table 3.2.
A compilation of this type is made for each property. Looking for mat,eri-
als with a particular characteristic (for instance: density less than . S M g / m 3 ,
strength c q u d to ~ O O M P Qor stiffness gnater t,han 3 x lo3) becomes straight-
forward. If the value being looked for is a threshold value for that property the
answer is given by the corresponding slot (either less, e p 0 l or gmaler). Othcr-
wise, the answer is derived from the three lists of material classes att,ached to
the threshold values just prcceding and just following the value looked for. The
+
iriterscction between the list, greaterlt) and ltss(t 1), where t and t + 1 are
two adjaceiit t,hresholds, givcs the list of all the material classes whose property
+
value belongs to ( t , i I).
T w o default threshold values are added to the table of compiled inCormat,ion
for each property. 'The first threshold gets the smallest possible value, its slot
23
Table 3.2: Compiled data for a particular property P
named ymnter contains all the material classes like t h e slot named less attached
to the last threshold. The obher threshold (the last one) gets the maximum
possible value for the particular property.
One could divide the lists of threshold values in several parts corresponding
to different orders of magnitudes in order to facilitate the access t,o malerials
whose property is in a particular range like low density, high stiffness, medium
strength, etc. This idea is exploited in a slightly different way. A more detailed
discussion iu found in Chapter 4.
24
Chapter 4
Description of the
Materials Selection System
In chapter 1 we have succinctly seen a n example showing the reusing of the
see-saw coucept for a water faucet (see Figure 4.1).
The process of finding the right material Cor the water faucet involves the
following steps:
1 . 1,ook up thc material and material justifications of the base case,
2. J,ook up the features of the target artifact,
3. Deduce (possibly new) mat,erial properties the base component should
possess in the target. context.
4 Infer t,he fuzzy values these propcrties should have.
5,Suggcst the best materials for substitution
In this chapter we shall detail these various steps
25
4.1 Material Information Attached to a Case
4.1.1 Justifications in the retrieved case
In the see-saw cme Cor instance, various pieces of information are provided
regarding the material it is made of.
See-saw in a playground
Material: wood parallel to grain
Inherited specifications:
-1- mode of loading: bending of rods and t u b a
-2- normal load: support a weight of at least 1OU kg
B e h a v i o r a l justifications:
-1- bending resistance:
mechanical reason - high (or better)
to avoid the see-saw t o bend under normal load
-2- torsion resistance:
mechanical reason - medium (or better)
to avoid t,he see-saw t o bsvist under normal load
-3- fracture resistance:
mechanical and safety r e s o n -high (or better)
to avoid the see-saw to break
under normal load
h n c t ional justifications:
aesthetic reason -
-1- UV radiation resistance: at least good
to avoid color fading
-
-2- t,emperature of use:
safety reason
very cold to yery hot
t,o avoid painful skin contact
The inhxrided specifications are deduced from the general specifications given
by the designer. In the example shown in Section 1.3 concerning the see-saw the
description of the artifact mentions that a beam, hinged at its center, should
Yupport, the weight of two people sitting on both extremit,ies. Knowing t,his the
mode of loading is looked up in t h e table of modes of loading. Such a table can
be found in [Ashby! 19911
There are two types of justifications used in the system: behavioral and
functional. A reason is attached t o each justification. nere only three types of
reason are mentioned: mechanical, safety and axsthetic reasons. A degree ol
importance can also be added t,o t,he justifications that would convey the extent
f o which the justification must be t d e n into account in determining the right
rnatcrials class. For instance, one may consider aesthetic reasons less impor-
tant and thercfore focus first on satisfying mechanical and safety constraints.
Aesthetic reasons would become secondary. The introduction of a degree of
importance would lead to a ranking ofjustifications and then lead fo enhancing
the choice between classes of materials.
The beharioraf jrasdificotiorrs constitute the intrinsic reasons that led to the
rriatcrial choice. They relate t o the function of the artifact itself and thus always
hold. A behavioral characteristic of a see-saw found on a playground is to
26
support the weight of two people sit,ting at Loth ends. The mode of loading
of t,liis a r t i f x t pertains to the class of “bending of rods and tubes”. Such
information can be found in [Ashby, 19911 where a table (see Table 2.1) of the
ten modes of loading relative t,o any mechanical design are given. Each mode
of loading delermines t,he relevant mechanical property/properties the material
should possess. In the see-saw case, bending resistance, torsion resistancc and
fracture resistauce are t,he three important properties. Chapter 2 explains how
their Inathemat,ical expressions are exploited.
The values given t o the properties are conveyed by means of orders of magni-
tude that can he given by the designer either direct,lyor through a user interface.
A user interface could help the designer specify the range of weight the see-saw
should support for instance and the qualitative values for the three types of
resistance could t,hen be derived. An example of how a user interface could he
useful in dekrmining fuzzy qriantities is given in Section 4.2.3.
The funciioaal justij5cations take into consideration the environment of the
artifact. Let us say, we are talking about an outdoor location for the see saw.
If the see-saw is intended to he used outside in very cold weather, one wants to
avoid metal because of possible skin contact. The other functional justificatio~i
relates to an aest,het,icpoint of view. One doesn’t want the color of the see-saw
to fade t,oo quickly when exposed t o the sun’s radiation. These two types of
juslification are directly specified by the designer.
Orders of magnitude are used here rather than actual values This facility
allows designers to work with incomplete or inexact information about an evolu-
ing design concept. They usually don’t have a precise knowledge of the value a
property must get but know roughly the suited range of value. Thesc quanti-
lies can easily be represented as fuzzy intervals [Dubois and Prade, 19881. We
can divide the whole range of possible values in five categories: very low, lorn.
medium. high and very high (see Section 4.2 for a more detailed discussion).
Fnnct,ionaljustifications can also be modified. In case of a water faucet, resis-
tance to ultra-violet radiation is not required, and t,emperatures below freezing
27
are unlikely t o be encountered. Therefore, t,he corresponding justifications at-
tached to the retrieved case can hc discarded. In addition, one may require the
artifact to be resistant. to salt corrosion if. for instance, the faucet is going to
he placed in a marine environment.
h precise value can be given for the material properties. For Example
the statement: “the material density is 5Mg/rn3”. In this statement the
density is precisely known. Houever, sometimes one doesn’t have the
knowledge t o specify such a precise information. In that Case it is more
realistic t,o provide a range of possible values.
The range of values then becomes B fuzzy interval, i.e., an interval with
imprecise boundaries, whose definition is given in Section 3.1.5.
28
4.2.2 Partitioning a set of values
The scales used in Ashby’s material selection charts are logarithmic. but this
does not really pose a problem when one wants to define a non-fuzzy partit,ion
over a universe OC possible values. Figure 4.2 shows the characteristic functions
of the five orders of magnitude for the set of possible material densities in Ihe
case of non-fnzzy orders of magnilude on a logarilhmic scale.
Charact,eristic function: p
very very
high
Densit.y
.I .3 1 3 10 30
Although the materials densities range from 0.1 to 30,tfg/nr3 the ends are
cxt.ended to infinity. The lower (resp. upper) bound of interval very-low (resp.
very-high) is --x (resp. +m). One has the general property that any value
only pertains to one interval at a time:
29
high
( 1 2 3 4 5 6 7 8 9 1 0 1 1 1 2 1 3
With a crisp partition one has the relation p ~ ( z>) U, (i.e., the value z is
A ) for only one order of magnitude A. Thus one is able to state Lhat “1: is A”
or ‘‘z is not. A ” . On thc other hand, the advantage of having a fuzzy partition
over a crisp partition is that it allows to determine that a value z is more A
t,han B , and also t,o quantify to what extent x is A or z is more A than B
~ “ ~ ( ~ ) > ~ B ( z ) , ~ : E R : is
A ,more E B
B E A~ ~than
30
Charperistic function: r
very
high
Density
t
1 3 10
4 Pat-kart-medium
31
/
Figure 4.7: 'Three-dimensional representation of fuzzy orders of magnitude.
32
very
very low low medium high high
A simple algorithm t,o define t,he various orders of magnitude could be:
This last step could be performed fairly simply by answering questions like:
The answer to these two questions, for instance, would provide t h e lower
bound of the core and the support respectively for the order of magnitude high
attached to the materials property stiffness. These two questions are generic
when the t,erms ‘high’ and ‘stiffness‘ are replaced by any order of magnitude
and material property. As a matter of fact these two questions are sufficient to
define all the fuzzy intervals (See Figure 4.3) since:
Only the two exlreme intervals are not regular t,rapezoids. The luwer (resp.
upper) hound of the core and the support of the first (resp. lart) int,erval equal
the minimum (resp. m;utimum) possible value.
The other way to perform step 3 of the algorithm is to specify what would
be thc crisp thresholds of the orders of magnitude and let the interface compute
their fuzzy houndaries via a multiplicative coefficient as explain in Section 4.2.2.
That’s what is actually performed by our system.
33
4.3 Combining Matching Degrees for Materi-
als Class Selection
4.3.1 Queries for accessing materials classes
The fuss? sets based representation in the system, support t,he following types
of queries t,o the materials knowledge base:
34
p7,. )f,,(I<)= 1
property
scale
pa(1i) = ;
*
property
scale
Figure 4.9: Matching degree hetween a material clms I< and the order of rrrag-
nit,ude A of some property.
of magnitude at the same time. Figure 4.10 shows in detail this aspect,
where 4 . B and C are three different orders of magnitude and R is the
property range for the materials class K .
Figure 4.10: Matching degree between a material class I< and the order of
maguilude A of some other property.
Thc answer to this query is a list of orders of magnitude where each ofthem
is ranked according to its computed degree of matching. 'The example
given on Figure 4.1U would lead to t,he answer { ( B , l ) , ( A , $ ) , ( C ,+)}.
4. Gzven a design code: what am the relevant criteria one should take into
cunsideration?
The answer to such a query is a set of criteria that are optimal Cor materials
selection of an artifact whose mode of loading is known. For example, for
a light cylinder with internal pressure the t,hree important criteria are
stiffness of type f , ductile strength of type and brit,tle strenglh of
type y. This type of d a h is sine qua non to determine which materials
selection charts t o use. The answer to this hype of query depends on the
mode of loading of the artifact. It is given by Table 2.1
4.3.2 Combining matching degrees
Once the important criteria that should be taken into considerat,ion have been
found, the system looks for the materials classes that best fit each of t,hem.
For instance, in the example given in Section 4.1.2 three crireria t,urn out to
be important: bending resist,ance, torsion resistance and fracture resist,ance.
In t h a t caSe Ihe system retrieves (1) materials classes that possess at le& a
medinm bending resistance: ( 2 ) materials classes that possess at least a medium
Lorsion resistance and (3) materials classes thal possess at least a low fracture
resistance. A s seen in Section 4.3.1 this gives rise to three lists where materials
c.lasses are ranked according to t,he extent to which they match the property
requirement.
Aggregating these three lists can be performed in various ways. The st,rateg?
should however meet intuitive characteristics. Some of them are dpscribed here
after. See [Dubois and Prade, 19851 for a more general analysis on aggregation
fnnctions.
' ere q ( K ) is the matching degree of the materials class I< with the
\%h
requirement, d (an example of requirement could be a very high stiflriess),
z ( K ) is the aggregated matching degree, f is the aggregation function.
2 . The antonym characteristic should also be true. The materials class that
completely mismatches all the requirements should get 0 as the aggregated
matching degree:
VZI(K),.. . , t , ( K ) E [ O , 11%3 z ( K )
such that z ( K ) = f(rl(K), ....z n ( ( K ) )
36
5. II‘ at lcast one single matching degree increases then lhe global (aggre-
gated) matching degree should not decrease.
6. f should he continuous:
f(Zl(K), , . . , z j ( K ) : ., ., tn(l<))
f ( z l ( I i ). . . . ,z n ( K ) , w l : . _ _wn)
, = ,min niax(wi,zi(I<)),where,min uii =0
,=I ...n * = I . . n.
37
E. Alloys
E. Polymers
E. Ceramics
P. Ceramics
E. Composites 1 1 1 1
Woods 1 1 1 1
Elastoniers 0 0 0 0
0 0 0 0
completely meet the requirement of a t least a medium torsion resistance; it,s de-
gree (1/3) therefore belongs to (0, l). The third chart guides for light fracture-
resistant c.omponent,s. The degrees found in each column are aggregated via
the operator min. This conveys a n intersection between the classes of materials
found from each chart. T h e classes that best meet the requirements get the
degree 1. Engineering composites and woods are the t,wo preferred classes. 0
means the corresponding class is definitely rejected. One can see that t h e clavv
or engineering alloys is not totally rejected (degree 4 / 5 > 0).
Let us siippose one intends to manufacture the artifact in great number.
One may want to add a property concerning its cost while designing it. From
the chart that guides the selection for cheap stiff components (cf. Section A.4),
it. appears that the two best classes of materials are poroiis ceramics and en-
gineering alloys and t o a lesser extent woods (degree 1/2), If this property is
equally important t,o the three previous on=: the class of woods gets a grade
of 1/2 (mini 1/2,1)) while engineering alloys category gets 4/5 (min{ 1.4/5})
and thns becomes the preferred class of material. If all the properties are not
equally important, one may assign t o them some weight and even some fuzzy
weight conveying their degree of importance. The aggregation operat,ion then
would become a weighted min [Dubois and h a d e , 19881.
38
Chapter 5
The whole code has been written in Common Lisp [Steel Jr., 1990] on a V A X
station.
T h e program is divided in two parts, One containirrg all the d a t a definitions
that will constitute the knowledge base and the other part contains all the
functions t h a t create and access the knowledge base.
39
unit: the appropriate unil for measuring the property (e.g., M g / m ” ) .
thresholds: the crisp thresholds t,hat,lead to dcfine the fuzzy ones? cf. Scc-
tion 4.2.2
’I‘he value this last slot gets is an instance of structure of type thresholds
chat contains the four crisp thresholds t-1, t-2, t - 3 and t-4 t,hat. separates
between the five fuzzy orders of magnitude, plus the two boundaries for the
properly called t-min and t-max. For instance, the property density would gct
the thresholds:
t-min 0.1
t-1 0.3
t-2 1
t-3 3
t-4 10
t-max 30
Once all the materials properties have been defined a special structure called
properties is created with the name of the defined properties as its slot names.
This structure is used to define the materials classes.
The root structure is a generic structure with only one slot that will capture
an instance of the structurc properties where its slot names are the names
of the materials propert,ies defined (see Section 5.1.1). Each of these slots are
instanLiated with a pair (structure range) of values representing t h e lower and
upper bound of that property for the corresponding materials class. Fix exam-
ple, the struclure engineering-polymers will look like:
40
Engineering Polymers
Density
lower bound: -
upper bound: -
Fracture Toughness
lower bound: -
upper bound: -
A l l tile materials classes are stored under a global variable. This variable
IS called materials-sequence*. It is a vector of the structures defining t h e
various materials classes.
4 stiffness: the type of stiffness that should be Laken into account given the
mode of loading like E/p, E * / p , etc.
41
m-greater: list of materials classes whose property range is situated at
least partially after “changcvalue”.
5.2 Program
5.2.1 Data coherence verification and knowledge base ini-
tialization
A lot of information tnay be stored in the d a t a base. Currently we take into
consideration 9 materials classes and 24 properties (see Appendix D ) . For each
property of a materials class two values are provided (the lower and upper
bounds). When these two values are defined one wants t o make sure that the
ripper bound is greater than the lower bound. This type of checking is performed
because inistakes are easy to do when typing in the data.
The second type of coherence checking deals with the dcfinition of the orders
of magnitude. One wants to avoid the situation shown on Figure 4.4. As said in
Section 4.2.2 the boundaries of the core and the support of the fiizzy orders of
magnitude are 20% greater or lower than the crisp bounds. If a and b are two
p
successive crisp thresholds one checks that b - is always greater than a + $
so that the interval [a + g, k]
t - representing the core of the fuzzy order of
magnirude is non empty (see Figure 5.1).
Initializing the data base ronsists in filling thc structure which contains the
lists of change points sequences (Le., *change-points-sequences+), onc per
material property. In the beginning the sequences are empty vcctors. For each
42
lower bound or upper bound of a property an instance of the change-point
structure is rreated with this value as the change point value. T h e slot m-equal
is hound to a list containing the material the value refers to. Once all the
change point,s have been generated the sequence is sorted in increasing order of
thc \ d u e ? then Ihe slots m-less, m-equal and m-greater are updated according
t,o the malerials preceding, overlapping or following the change point value. Two
change-point structures are added at bot,h ends of each sequence. The first
one gets only its slot la-greater updated and the last one only its slot m-less.
If two different change points appear t o have the same change-point value in a
scquence than they arc merged into one.
43
Chapter 6
Conclusion
The approach presented here lies on the concept of materials sclection charts
created by Ashby [Ashby, 19911. AII implementation of these charts has been
suggested as well as a way t o exploit them via queries based on fuzzy orders of
magnit.ude. This work has led to the development ofan autonomous system that
can be hooked up t o more general systems. The initial purpose of this syslem
was to be implemented in bhe framework of the casebased design project CADET
[Sycara et al., 19921.
Casebased design stems from the general theme of case-based problem solv-
ing whose philosophy is l o make use of pas1 experiences in solving new prohlems,
exploiting its memories instead of relying solely on a base of proredures or rules.
I n other words, a casebased systcm is not only an expert or multi-expert system
but also a system that takes into account similar cases already encount,ered.
The CADETsystem has been conceived for the field of mechanical design, its
intent is to play the role of a designer’s assistant by retrieving and re-using pre-
vious succasful designs while avoiding previous failures such as poor material8
or high cost.
The process of case-based design consists of several steps that are not ner-
essarily strictly applied in t,he order given here; some may overlap.
1. Development of u Functional Description. At the simplest level, the de-
sired artifact. c a l ~be viewed as a black box which takes cert,ain inputs and
produces desired outputs. The functiori of the black box is described by
qualitalive relations explainiug how the inputs and outputs are related.
‘The system’s job is to realize an artifact which will convert the inputs into
the desired outputs.
2. Retrieval of Cases. A set of design cases (or case parts) bearing similarity
l o a given collection of features are accessed and retrieved. Retrieval is
performed using not only the existing features of t.he input specification
and also behavior description.
44
3. Dcl;elopment of a Synthesis Strategy. A synthesk strategy is a description
of how the various c a e s and case pieces will fit together to yield a working
design.
4. Physzcal Synthesis. Realization of the synthesis strategy at a physical
Icvcl. This is a difficult problem since undesirable inleractions among case
parts may occur. In addition, since it k very rare to retricve cases that
m a d l y match the design specifications, cases and case pieces must he
physically modified before actual synthesis.
So far. CADET implements the first four steps indicated above. Verification
and debugging is currently left to the human designer.
Once a cave has been retrieved, it, may need some adaptation to become part
of the whole artifact. This adaptat,iou phase can be seen as twofold. First, the
case itself may need to be modified to suit the features of the output specifica-
tions. ‘lhis s k p only brings into p h y the intrinsic characteristics of the case.
Secondly, due to the interact,ions between the (possibly already modified) case
and its new environment, other adaptations may be needed. This second step
intervenes at a lower level of abstraction right before the actual synthesis.
The tcrm adaptation has a broad meaning. In this report we have focused on
material adaptation and more precisely on the one that first takes place. Mate-
rial modification due t o interaction with ot,her devices has not been considered
here.
CADET’S architecture [Sycara et al., 19911 is centered around a design black-
board. This blackboard is used to maintain information about the various de-
sign alternatives that are being actually considered at any given time. The
materials selection system is consulted during case selection to check material
requirements and during case adaptation.
45
Acknowledgements
1 wish to acknowledge:
Katia Sycara and D. Kavinchandra who have contributed many ideas and
substantial time to the creation of the material selection system prescnt,ed
here.
Mark Fox who accepted me as Visiting Scientist at the Center for Inte-
grated Manufacturing Decision Systems.
46
Appendix A
h,lat,erials offering t,he greatest stiffness-to-weight, ratio lie towards t,he upper
left corner.
47
1
2. uy’J p = C : criterion for plastic bending, torsion of beams and shafts,
1
3. u$ J p = C: criterion for plastic bending of plates
Materials offering the greatest strength-to-weight ratio lie towards the upper
left comer.
48
A.3 Materials Selection Chart for Light Fkacture-
Resistant Components
O n Figure A.2 the property of fracture toughness, whose symbol is lirc. is
plotkd against density. The three guide lines are given by:
1. u y / C ~ =
p C : criterion for axial tension of ties,
49
2. o;/C,p = C: criterion for bending, torsion or buckling of beams: shafts
and columns,
50
Figure 11.2: Material selection chart for light fracture-resistant components
(courtesy).
51
Figure A.3: Material selection chart for cheap, stiff components (courtesy).
52
RELATIVE COST PER UNIT VOLUME C,p (Mg/rn’)
Figure A.4: Material selection chart for cheap, strong components (courtesy)
53
Figure A.5: Material selection chart for components resistant to corrosion (cour-
tesy).
54
Appendix B
This appendix summarizes the materials classes and the mechanical and en-
gincering properties that arc currently exploited by the syst,em (taken from
[Ashby. 19911).
55
B.2 Material Properties
Derivity P
Young’s Modulus E
Strength “f
Fracture Toughness I~IC
Stiffness E l p , E i l p , E: J p
Ductile Strength “ J l P ! “:/P, u:/P
Brittle Strengt,h I i r c l p , x , ~ c l p >ii,&Jp
Relative Cost CR
Resistance Lo Aerated Water
Resistance to Salt, Water
Resistance to Strong Acids
Resistance to Strong Alkalis
Resistance to Organic Solvent
Resistance to U-V Radiation
56
Bibliography
[Sycara et al.: 19921 Sycara, K. P., Navinchandra, D . , Guttal, R... Koning, J.-L..
arid Karaqimhan, S. (1992). Cadet: A case-based synthesis tool for engineering
design. Iniernaidonal Journal of E'xppert Systems, 4(2):157-188.
[Sycara ct al.: 1991) Sycara, K. P., Navinchandra, 0..and Yarasimhao. S.
(1991). A transformational approach to case based synthesis. A I EDAIII.
5 ~ ) .
[Ladeh! I S G ] Zadeh, L. (1965). Fuzzy sets. fnformalion and C o n t d , 8:338-
353.
[Z,eInankow: 1989) Zemankwa, M . (1989). Filip: A fuzzy intelligent informa-
tion syst,em with learning capabilities. Information System, 14(6):473-436.
58