ABSTRACT The Optimal Selection of Engineering Entities

D. Cebon and M.F. Ashby

Technical Report CUED/C–EDC/TR 59, November 1997, ISSN 0963-5432.

Cambridge University Engineering DepartmentTrumpington St, Cambridge, CB2 1PZ

November 1997

ABSTRACT

Various principles govern the successful selection of engineering 'entities', by which we meanmaterials, manufacturing processes, components or assembled products. The principles aredescribed here, with examples. There are two fundamental steps in any selection exercise:'Screening', to identify viable candidates, and use of 'Supporting Information' to enable a finalchoice. Data sources must be appropriately structured to support these steps. The paper discussesthese and other basic issues of database design, including data taxonomy, rules for the choice ofattributes (fields), the relationships between entities (records), and data checking and estimatingprocedures. The examples are drawn mainly from the selection of materials, but the principles aregeneral and apply to a wide range of selection problems.

1

1.INTRODUCTION

The act of selection is central to engineering design. The choice may be of a material (a metal, aceramic, a polymer, a composite…), a manufacturing process (casting, forging, injectionmoulding…); a shape (I-section, tube, rolled-hollow section…); a component (a bearing, anactuator, a battery…); or even of a complete product (a bicycle, for instance…). Regardless ofthe entity to be chosen, certain fundamental requirements must be met if the selection process is tobe systematic and effective, and the outcome is to be optimal in some sense.

These requirements are the subject of this paper. It deals with electronic information sources, thestructure of the selection process, database requirements and validation. The examples are drawnfrom the selection of materials and components, but the principles are general, and apply equally toa wide range of selection problems, among them, those listed above.

2.Selection Strategies and Database Structure

Successful selection of standard entities in engineering involves two main steps which we here callscreening and supporting information (Figure 1). The data requirements, selection strategies andsearching methods in the two steps differ fundamentally. Although this paper is mainly concernedwith computerised selection, the screening and supporting information steps are characteristic of allengineering selection activities – even if they are performed on paper, using handbooks ormanufacturers’ catalogues.

2.1Database Taxonomy and Attributes

Rational selection starts with a definition of the kingdom of entities from which the choice is to bemade. Figure 2 illustrates how the kingdoms of Materials , of Manufacturing Processes and ofcomponents (Bearings for illustration), can be subdivided into families, classes, subclasses andmembers. Each member is characterised by a set of attributes which include: identifyinginformation; numeric data, text, and graphical information. Take materials as an example (Cebonand Ashby, 1992). Its kingdom contains the family 'Metals' which in turn contains the class'Aluminium alloys', the sub-class '5000 series' and finally the particular member 'Alloy 5083 inthe H2 heat treatment condition'. This material, and every other member of the materialskingdom, is characterised by a set of attributes: its density, moduli, strength, thermal, electrical andchemical properties. Attributes are a key part of selection. They are discussed more fully below.Other kingdoms of engineering entities have similar hierarchical structures. Figure 2b shows thatfor manufacturing processes (Esawi and Ashby, 1997), and Figure 2c that for one type ofcomponent – bearings (Harmer, 1996).

2

An essential feature of a well-ordered hierarchical classification is that the properties of its memberslie within ranges defined by their sub-class; the properties of all sub-classes lie within rangesdefined by their class; and so on up the tree, as shown in Figure 3. In this way it is possible tocreate a 'generic' record at any level of the tree, which has properties that span the range of allentities below it in the hierarchy. This facilitates development of a comprehensive top levelscreening database, in which a limited number of classes or sub-classes can be used to represent allmembers of the kingdom. This feature also facilitates data validation and checking, as will beexplained later.

ScreeningGenericSelectionFamily CFamily BFamily AProp 1 1.2 - 3.4Prop 2 2.0 - 4.0Prop 3 TProp 4 F Prop 5 AProp 6 EFamily-specificSelectionEntity A3Entity A2Entity A1Prop 1 1.3 - 1.4Prop 2 2.4 - 2.5Prop 3 TProp 4 F Prop 5 AProp 6 ESupporting InformationSearching engine and indexBooksLeafletsManuf'sDatabasesEntity 123Prop 1Prop 2Prop 3InternetCD–ROMsFig. 1Structure for selection of 'standard' engineering entities.

3

KingdomFamilyCeramicsGlassesMetalsPolymersElastomersCompositesClassSteelsCu alloysAl alloysTi-alloysNi-alloysZn-alloysSub-class1000 Series2000 Series3000 Series4000 Series5000 Series6000 Series7000 Series8000 SeriesMember5005-05005-H45005-H65083-05083-H25083-H45154-05154-H2...AttributesDensityCostModulusStrength..Max use tempTherm. conductivityElect. resistivityFormingAvailable forms ....Materials(a)

KingdomFamilyDeformationMouldingPowderCastingMachiningCompositeMolecularSpecialClassInvestmentFull mouldShellSandDieSqueezeCeramic mouldPermanent mouldMemberSand 1Sand 2Sand 3Sand 4Sand 5AttributesSize rangeRoughnessToleranceAspect RatioEconomic batch sizeCapital costTooling costMaterial classShape class ...Processes(b)

KingdomFamilyHydrodynamicRubbing PlainPorous MetalRolling ElementHydrostaticMagneticKnife EdgesPivotsClassThrust - BallThrust - RollerRadial - BallRadial - RollerCombined - BallComined - RollerSub-classCylindricalSphericalNeedleMemberType 1Type 2Type 3Type 4Type 5AttributesCostMassInner DiameterOuter DiameterDynamic load ratingStatic load ratingFatigue load limitMounting ...Bearings(c)

Fig. 2

Examples of the taxonomy of various databases:

(a) materials; (b) manufacturing processes; (c) bearings.

4

Property 2ClassSub-classCheckinglimitsFamilyChecking limitsMembersProperty 1

Fig. 3Schematic relationship between database hierarchy and property (attribute) ranges.

2.2Screening, and the Screening Database

Unbiased selection can only be achieved by considering all the entities in a kingdom to be viablecandidates until these are shown to be otherwise. The first step in identifying viable candidates isthat of screening (Figure 1, top half). The screening step can be performed effectively using acomputerised database containing the attributes of the entities under consideration – materialproperties, or the characteristics of processes, or data for standard components. The database forscreening, however, must be structured to give it certain characteristics if it is to work effectively.In particular, it must be comprehensive, and complete, and the attributes it contains must beuniversal. The meanings of these terms in the context of selection databases are described next.

Comprehensiveness

The term comprehensive means that the database contains all families and classes of entities in thekingdom of interest. A materials database should therefore contain all families of materials in the'materials kingdom' (metals, polymers, ceramics, natural materials and composites etc.) (Ashby,1992). A database of bearings should contain all families of bearings (plain, ball, roller, gas,magnetic..), and so on. If it fails to do this, one or more important families of entities will beoverlooked completely, thereby ruling them out of the selection by default.

It can be tempting for an engineer to assume that the best solution will be the one implied byprevious experience – and therefore to use a non-comprehensive data source – for example, asuppliers catalogue containing only rolling element bearings. This inhibits innovation by inviting

5

a non-optimal selection, due to lack of familiarity or experience with alternative solutions: a plainbearing or a hydrostatic or magnetic bearing might be better suited to the particular application.Thus a non-comprehensive database, with only a limited range of families or classes, is poorlysuited to first-level screening, though it may serve for second-level screening, described in amoment.

Universality

In a screening database each entity appears as a record, the fields of which contain data for itsattributes. Effective screening requires a database which is tightly structured. It is essential tostore the same set of information for each entity. This constrains the choice of attributes, in a first-level screening database, to those which are common to all entities in the database. That is, theyare universal within the kingdom. Decisions about the taxonomy of the database are stronglylinked to the choice of attributes.

Consider, as an example, constructing a database of linear force actuators. The kingdom containsmany families (hydraulic, pneumatic, piezoelectric, electrodynamic, magnetostrictive,...) each withclasses, sub-classes and members. The 'universal attribute set' would consist of properties like'maximum force', 'stroke', 'resolution', 'power consumption', 'bandwidth', and 'size', which arerelevant to all actuators, whatever the family. They can be used to select families, classes or sub-classes from a 'generic' database of actuators (Huber, Fleck and Ashby, 1997). Class-specificattributes such as 'maximum fluid flow rate', or 'servo-valve current', which are relevant only tohydraulic actuators, are not included. If a 'generic' actuators database contained these 'family-specific' attributes, and a selection was based on them, some families (eg electrodynamic orpiezoelectric) would fail the selection by default, because they would not posses data for thoseattributes. This restricts the screening attributes to those which are universal within the database.Note, however, that data which is not universal at one level may become so at a lower level in thedatabase hierarchy. Effective screening is often best performed by having two levels of databases,as shown in the top half of figure 1. The first-level is based on the entire kingdom – it contains allfamilies, and attributes that are common to them all. The second-level is based on 'family-specific'databases, each describing a single family and containing the additional attributes peculiar to thatfamily. The first-level screening isolates one or more families. The second level explores thesefamilies only, giving greater resolution. Thus the first-level selection of actuators is based onattributes like maximum force, stroke, etc. When one or more suitable candidates has been found,a 'family-specific' database is used to narrow the selection to one or more particular entities(maximum flow rate, for example, or valve current in a hydraulic actuator database).

6

Completeness

It is important that all records in a screening database are complete, that is, that they contain no'holes' where a datum is missing. Holes can cause entities to fail selections by default becausethey have no value for an attribute on which a search is based. It is, of course, possible to insertdummy values in the holes, chosen to ensure that entities pass rather than fail the search, but this,too, can be misleading. A better solution is to use estimating procedures or bounding methods tobracket the probable value of each missing datum, flagging these to distinguish them from 'real'data. This procedure maximises the effectiveness of the search. Data estimation procedures arediscussed further in Section 3.5.

Selection Strategies for Screening

Screening is performed by linking the technical and economic requirements of the design with theattribute profiles stored in the screening database(s), using attribute limits and optimisingindices. Attribute limits are simple limits placed on the values of certain attributes stored in thedatabase. For example, \"the service temperature of a candidate material must be greater than250oChermoplastics\" or \"the bearing must have an outerdiameter of less than 20 mm\". It does not, however, provide any level of performanceoptimisation.

Optimising indices are derived from a 'cost function' or 'objective function' which describes someaspect of the performance of the entity. Optimisation is achieved by ranking the entities in order ofperformance and then selecting the top few candidates. In the case of materials selection, analysesof component performance lead to 'material performance indices' which are combinations ofmaterial properties that characterise performance in a given application (Ashby, 1991, 1992). Forexample, the specific stiffness E/ρ , or the specific strength σf/ρ (E is the Young's modulus, σf isthe failure strength and ρ is the density). The materials with the largest values of these indicesperform best when used for a light, stiff tie rod, or a light, strong tie rod respectively. There aremany other indices, each associated with maximising some aspect of performance, or minimisingcost (Ashby and Cebon, 1995; Cebon and Ashby, 1994). Similar optimising indices are found inother engineering selection problems (see, for example (Huber, Fleck and Ashby, 1997), (Harmer,Weaver and Wallace, 1996) and (Weaver and Ashby, 1996)).

To summarise: attribute limits isolate candidates which are capable of doing the job; optimisingindices identify those among them which can do the job well. The top performers which satisfy allof the screening criteria are then short-listed.

7

2.3Supporting Information and Data Sources

The result of the screening step is a short-list of candidate entities which satisfy the quantifiablerequirements of the design. To proceed further it is necessary to obtain a detailed profile of eachcandidate: ie its supporting information (Figure 1, lower box).

The data requirements for supporting information differ greatly from those for the screening step.It is not necessary to screen all entities in the kingdom, only to find necessary additionalinformation about the few candidates that have already been identified by the screening step.Typically, it is non-quantifiable information which is sought. For materials this might bephotographs of microstructures or case studies of corrosion protection in a particular environment,or details of availability and pricing. For manufacturing processes it might be examples ofproducts made by a candidate process. For bearings it might be fitting details or examples ofapplications. Supporting information helps narrow the short list to a final choice, allowing adefinitive match to be made between design requirements and entity attributes of the entities.

Data Format and Sources

There are no requirements of universality, comprehensiveness or completeness in supportinginformation. Instead the requirement is for relevance, regardless of format. The information maybe in the form of text, tables, graphs, photographs, computer programs, or video clips, (see forexample (Cebon et al., 1994)). It can be large in quantity and detailed, specific and precise innature. The most common medium for such data is manufacturers' leaflets and catalogues,information which, increasingly, is becoming available in electronic form – in databases, on CD–ROMS, and on the Internet. Some sources of further information for materials are listed inAppendix A. Typically, a database of supporting information contains data of the following types:

• Reference information about the entity and others like it;

• Examples of the use of an entity, or of engineering systems which contain it;• Design guidelines and standards relevant to the use of the entity;• Case studies and worked examples;

• References to other sources of information, including scientific and engineering literature,and suppliers information.

The advantages of such computerised sources are that: the person(s) who compile the databaseoften improve the consistency of the data; the information can be updated regularly and easily; theprograms can be quicker and more convenient to use than handbooks; and they can provide linksbetween attributes (property data) and other information. Currently, however, they also havecertain disadvantages: each disk uses different terminology, standards and format; they are oftenclass-specific, with manufacturer’s 'spin' on the information; data on the disks often lags behindprinted publications; selection facilities can be rudimentary; and false conclusions can be drawn ifthe database and searching algorithm are badly structured.

8

The Internet contains an expanding spectrum of information sources. Some, particularly those onthe 'World-wide Web' (WWW), contain data for materials, for processes, for components and forproducts. The information is provided by standards organisations, trade associations, learnedsocieties, universities, and individual manufacturers or suppliers, who provide information abouttheir specific product ranges.

The main advantages of the WWW include the direct and timely access to supplier and economicinformation; powerful links between sites; and the uniform, machine-independent, programinterface. The main disadvantages are that it: can be difficult and time consuming to findinformation; there are no uniform standards, designations or units; selection strategies differfrom site to site; and there is no formal control over the quality of the information.

Search Strategies for Supporting Information

Because the data is in 'free' format, the search strategies for supporting information differcompletely from the numerical optimisation procedures that are best for screening. Withoutscreening, the candidate-pool is enormous; there is an ocean of supporting information, anddipping into this gives no help with selection. But once viable candidates have been identified byscreening, it is only necessary for the supporting information system to provide a mechanism foraccessing information about these few entities. This is likely to involve 'tagging' each item ofsupporting of information with the identifiers of all entities for which it is relevant. The simplestsearching approach is to use an index (as in a printed book), or a keyword list, or computerised fulltext search, as implemented in many hyper-media systems – the Encyclopaedia Britannica on CD,for instance, is an example of such a Supporting Information system. It is useful if you knowwhat you are looking for, but overwhelming in its detail if you do not.

An example of a Supporting Information source for materials is given in (Cebon et al., 1994).This disk can be used in conjunction with the Cambridge Materials Selector (CMS) to perform thescreening and supporting information steps for selection of copper-based alloys.

3.The Screening Database: Contents, Structure and Validation

This section considers some further issues of database design for the screening step.

3.1Screening Data Types

A variety of different basic data types can be used in a screening database. The choice of data type(text, Boolean, discrete, numeric, etc) depends on the availability of data, and the precisionrequired.

9

The least structured form of data is generally text. Although text can be searched easily andflexibly, the lack of constraints on its format means that the entries are incomplete, and so cannotbe used for reliable screening. An example will illustrate this: the uses of materials, stored in a textfield (eg 'fasteners, gears, shafts...') can be searched for a particular use, but unless all possibleuses are stored for all materials (an impossible task), such a selection rejects materials for whichthe database entry does not contain the required use. The proper use of text fields is therefore as asource of supporting information.

Boolean fields, for which the data entry is 'True/False' or 'Yes/No', can be used in two differentways. They are essential for properties which can only take one of two alternative values, forexample whether a foam has open-cells or closed-cells. They can also be used profitably fordescribing attributes where precise data is not available. Examples are ways of shaping a material('cast, forge, machine, etc'...); or material capabilities of a process ('aluminium alloys, copperalloys, etc'); or bearing types ('plain' or 'ball' or 'roller'...), each carrying a 'Yes/No' entry.The data entry in discrete fields can take one of three or more possible discrete values. Anexample which is common to many materials data sources is the characterisation of corrosionresistance to particular environments. These might be rated on the five-point discrete scale: 'A, B,C, D, E'; where 'A' means very good, and 'E' means very poor.

Numerical data can be stored as a range of values (eg '1.0 to 1.5') or as a single point value (eg'2.0'). All engineering data has some associated tolerance or range of reliability, and so it is properto store most numerical attributes in a screening database in range format. Sometimes the range islarge: the range of yield strengths of the 'generic' material 'Wrought aluminium alloys', forexample, is 40 MPa to 500 MPa. More usually it is small: Young's modulus for the Magnesiumalloy 'ZM21' is 43.4 GPa to 43.5 GPa.

There are some cases where the property range (tolerance) is small, for example some dimensionsof standard components. In these cases there is a trade-off between the benefit of having thetolerance available in the selection process, and the cost of storing two numbers (a range) insteadof one number (a point) for each record. This trade-off is more important for databases with largernumbers of records.

Finally, it is possible to store images in a database (eg scanned drawings, graphs or photographsor animations). Like text fields, their utility during the selection process is one of providingsupporting information rather than for screening.

3.2Choice of Fields

The choice of fields must satisfy the three criteria of comparability, measurability anddiscrimination.

10

Comparability

The importance of storing a common set of universal attributes in the screening database wasmentioned in section 2.1. It is necessary to choose these attributes so as to ensure comparabilityof entities. A difficulty arises when the kingdom contains many families: it is that, for reasons ofhistory and convenience, a single attribute may be measured in several different ways. Consider,for example the hardness of materials. The hardness of metals is measured on the Rockwell,Vickers or Brinell hardness scales; that of elastomers is measured on the Shore or Durometer scale;that for minerals on the Moh scale. They are measured with different testing equipment, reported indifferent units, and there are no simple relationships between them. However, since all materialhave a hardness, it is an important attribute in a comprehensive ('generic') materials database.This creates the necessity to choose a common equivalent hardness scale – the Vickers scale, say –for all materials, and to calculate or estimate suitable values for those materials that are not usuallymeasured on that scale. Family-specific or class-specific databases can contain attributes which arenot universal to the kingdom, as long as they are common to all members of the family or class.Thus a database of elastomers could contain the Shore hardness - and this could be used to screenat the family or class-specific level. Similar problems are found in databases of manufacturingprocesses and of components, and can be resolved in a similar way.

It is sometimes useful to create universal attributes which summarise various types of behaviourinto a single quantity. An example of this is the 'maximum service temperature' of a material,defined as \"the highest temperature at which the material can reasonably be used in a load-bearingcomponent, without oxidation, chemical change, or excessive creep becoming a problem\" (Cebonand Ashby, 1994).

Measurability

There are standard ways of measuring many engineering properties: density, modulus, thermalconduction and dielectric properties are examples. There are other properties - somemicrostructural, others relating to corrosion in exceptional environments or to tribologicalbehaviour, which one might like to use for screening, but their use in this function is precluded bylack of data and the near-impossibility of making any sensible estimate to replace them.

Discrimination

The choice of attributes for screening must allow discrimination between entities. Thus density,modulus and strength are discriminating attributes when the kingdom is that of materials; but ifthe kingdom is restricted to the family of carbon steels, density and modulus cease to bediscriminating because all carbon steels have almost the same values of these properties. Strength,hardness, and magnetic coercive force (a property which is non-universal for the greater kingdomof materials but universal for carbon steels) remain discriminating in the family-specific case.

11

3.3Relational Structures

A cardinal rule of data-base structure is that there be no redundant data. Many aspects ofcomputerised selection can be performed adequately with the data stored in a 'flat file' (ie a table),in which the columns of the table contain the fields (properties) and the rows contain the records(entities) - rather like a spreadsheet. For some database functions, however, flat files areinadequate and a 'relational' structure is considerably better. Consider, as an example, the'suppliers' field in a class-specific materials database of (say) aluminium alloys. This databasemay contain some hundreds of different alloys, most of which are manufactured by a small numberof companies (Alcan, Alcoa, AluSwiss, Pechiney, Norsk Hydro etc). In a flat file it is necessaryto store the name and address details of at least one of the companies in every record, with some ofthe suppliers appearing in several different records. This would be a violation of the redundancyrule. Were one supplier to change telephone number, for example, then every instance of thisnumber must be updated, with potential for error and omission.

A relational database structure overcomes this problem. The simplest relational database containstwo data 'tables'. In this example, one table would contain material properties, the other wouldcontain information about the suppliers, as shown schematically in Figure 4. Each supplier wouldappear just once in the suppliers table. The database management software would create andmaintain links between this supplier and every material that it supplies. In a relational database thelinks are as important as the attributes (in fact they can be considered to be attributes) and theyrequire a software system which handles them flexibly and automatically.

A powerful feature of links is that they work in both directions. Thus it is possible to use theexample system as a materials selector to select materials that are made by particular suppliers (eg“select all aluminium alloys made by US suppliers”). Alternatively it could be used as a suppliersselector to select the suppliers that manufacture particular materials (eg “select all US suppliers whomake aluminium alloys”).

12

Materials DatabaseMATERIAL 1MechanicalSuppliers DatabaseSUPPLIER 1NameAddressTelephoneSUPPLIER 2NameAddressTelephoneSUPPLIER 3NameAddressTelephoneThermalSuppliersSupplier 1Supplier 3Supplier 5MATERIAL 2MechanicalThermalSuppliersSupplier 1Supplier 2Supplier 4SUPPLIER 4NameAddressTelephoneSUPPLIER 5NameAddressTelephoneFig. 4Relational structure of material and supplier data tables.

3.4Data Validation: Range Checks and Attribute Correlations

The value of a database depends on its accuracy and its completeness – in short, on its quality.One way of maintaining or enhancing quality is to subject the data to validating procedures. Therange-checks and physically based correlations, described below, provide powerful tools for doingthis. The same procedures fill a second function: that of providing estimates for missing data.This is essential when direct measurements are not available (Ashby, 1997).

Range Checks

Trivial though it may seem, it is valuable, when creating a database, to tabulate the known range ofvalues of each attribute for each class in the database, and to check whether each new datum lieswithin the appropriate range. A convenient way of presenting the information is as a table in whicha low (L) and a high (H) value are stored, identified by the family or class. An example, listingYoung's modulus, E, for the generic material families, is shown in Appendix B. Range checks,usually based on experience or current practice, can be devised for almost any attribute of almostany entity. The use of the ranges for property checking is obvious: any new datum should lie inits proper range; if it does not, it should be checked. Why bother with such low-level stuff?Because in computations involving engineering entities, the most common error is the use of an

13

attribute-value which is expressed in the wrong units, or is, for less obvious reasons, in error byone or more orders of magnitude (slipped decimal point, for instance). Class-specific range checkscatch errors of this sort.

Correlations Between Values for Attributes

Materials which are stiff have high melting points. Solids with low densities have high specificheats. Metals with high thermal conductivities have high electrical conductivities. These rules-of-thumb describe correlations between two or more material properties which can be expressed morequantitatively as limits for the values of dimensionless property groups. They take the form:

CL(1)

(or larger groupings), where P1, P2, P3 are material properties, n and m are simple powers(usually -1, -1/2, 1/2 or 1), and CL and CH are dimensionless constants – the lower and upperlimits between which the values of the property-group lies. An example is the relationship betweenexpansion coefficient, α (units: K-1), and the melting point, Tm (units: K):

CL≤αTm≤CH

(2)

Values for the dimensionless limits CL and CH for this group are listed in Appendix B for anumber of material classes. Where they exist, such correlations permit checks and estimates whichare much more discriminating and precise than are the simple range checks.

There are other useful relationships between attributes which arise from the mechanics,thermodynamics or economics of a particular application. These can also be used profitably fordata checking. Examples are:

• the upper and lower bounds on the properties of composites, which can be used toestimate or check the properties of a composite from the properties of its constituent parts;• the local buckling of structural sections which can be used to estimate or check themaximum 'shape factor' of a structural section (Ashby, 1991);

• the buckling load of a shaft which can be used to check the maximum force of a linearactuator;

• the burst limit of bearings which can be used to check their maximum rated speeds;• the torque of a permanent magnet motor which is limited by its size and the remanance ofthe material of its magnet;

• the economic batch size of a manufacturing process which is related to the capital cost ofthe processing equipment.

14

Automated Checking Procedures

The hierarchical structure of entities (section 2.1) and relationships between the properties can beused profitably to automate aspects of data checking. The scheme that is used in the developmentof materials databases for the Cambridge Materials Selector (CMS) (Cebon and Ashby, 1994)involves two independent stages as shown in figure 5. In the first stage, the range of everyproperty of each member of the database is checked against the ranges stored for the 'generic' sub-class above it in the database hierarchy (see figure 3). Similarly the sub-class entities are checkedagainst the ranges stored for the classes above them, and so-on. In the second stage of checking,the property correlations described above are checked, using the same hierarchical procedure.

Fig. 5The scheme for checking material properties in CMS databases.

3.5Approximations, Estimates and Non-Existence of Data

The need for completeness, ie the absence of holes in records for screening has already beenmentioned. There are three ways of dealing with holes.

15

Approximations

An attribute may not strictly apply for the particular entity. The 'modulus of rupture' σMOR is onesuch property in material properties database. It describes the bending stiffness of a brittlematerial, like glass, and is familiar to ceramic engineers (who measure strength in this way) but notto metallurgists (who use other measures of strength). An engineer seeking a material with anadequate modulus of rupture really wants one with an adequate bending strength - so a practicalsolution is to fill the modulus of rupture field in the metals part of the database with a comparableproperty: the yield strength. Then metals with adequate bending strength will pass selectionsbased on the modulus of rupture.

Estimates

In some cases the value of an attribute may not have been measured, or may not be in the publicdomain. Then it is necessary, for completeness, to estimate a value. Some material properties fallinto this category. The estimation of attributes is an activity in which much is done but little iswritten. For material properties, usefully accurate estimation techniques exist. These are based onphysical relationships between properties (Ashby, 1997). As discussed in Section 3.4, they are ofprimary help both for estimating and checking attributes.

Process attributes and attributes of components are harder to estimate. When the entity is part of agroup with which it has much in common (its 'siblings'), weighted interpolation can be used toestimate the missing attributes of one entity when those of its siblings are known (Bassetti, Brechetand Ashby, 1997).

It is essential that approximations and estimates in the database are flagged in some way, so thatif the material is selected on the basis of one of these properties, the engineer knows to seek furtherinformation to check the estimated data.

Non-existence of Data

There are some (rare) attributes which need to be stored in a screening database, but are notcommon to all entities in that database. They are therefore exceptions to the universality rule.These are best designated as 'not-applicable' in the database. They can be recognised by the factthat it would be acceptable for entities which have a 'not-applicable' entry to fail a selection basedon that property.

Consider, for example, selecting a dielectric material for a capacitor, on the basis of its dielectricconstant. The dielectric constant is an electrical property which is important to a 'generic'materials database, even though it is relevant and measurable only for insulators; for conductors itis non-applicable. An engineer seeking a dielectric would not wish to consider a conductor, so it is

16

acceptable for all conductors in a generic database to have 'not applicable' entries for dielectricconstant, and to fail selections based on it.

4.CONCLUSIONS

Optimal selection of standard engineering entities, whether performed on paper or a computerscreen, is best achieved by a two-step process: a 'Screening' step followed by a step of'Supporting Information'. In the first step the kingdom of entities is reduced to a small candidatelist which meets the attribute limits and maximises optimising indices (Section 2.2). The secondstep retrieves detailed contextual information about each of the candidates (Section 2.3) and enablesa final choice to be made.(i)

A database suitable for screening should have the following characteristics:

• It should be comprehensive – contain all general classes of entities in the 'kingdom' ofinterest.

• The attributes it contains should be universal – common to all of the entities in thedatabase. The attributes should further satisfy the requirements of comparability,measurability and discrimination.

• It should be complete – have no holes or gaps without any data. This can be achieved bythe use of approximations and estimates to fill the holes.

• It should have a relational structure (or similar), to minimise data redundancy.

• It should, where possible, exploit a hierarchical taxonomy, so as to facilitate datachecking between layers of the structure.

• Range checks and physically-based relationships between the attributes should be used toimplement automatic data checking procedures.

(ii)

The supporting information system can have information stored in any format. The onlyrequirement is that items of information should be 'tagged' according to the identifiers ofrecords in the screening database. Once a particular entity has been isolated by thescreening process, all information about it can be retrieved rapidly from the supportinginformation system.

Taken together, the two-step selection process is capable of selecting an 'optimal' entity from akingdom containing many thousands of entities. Successful examples exist in the fields ofmaterials (Cebon and Ashby, 1992), processes (Esawi and Ashby, 1997), section shapes (Weaverand Ashby, 1996) and components (Harmer, 1996; Harmer, Weaver and Wallace, 1996; Huber,Fleck and Ashby, 1997). The procedure appears to have generality, and is presented here as aguide for structuring selection systems in engineering.

17

5.ACKNOWLEDGEMENTS

We wish to thank Mr KM Wallace and Dr PJ Clarkson, past and present Directors of theCambridge University Engineering Design Centre and the Engineering and Physical SciencesResearch Council for their support of research into selection principles.

6.REFERENCES

Ashby, MF. (1991). 'Materials and shape.' Acta Metall. 39: 1025–1039.

Ashby, MF. (1992). Materials selection in mechanical design. Oxford, Butterworth Heinemann.Ashby, MF. (1997). 'Checks and estimates for material properties.' to appear in Proc Roy Soc A.Ashby, MF and Cebon, D. (1995). Case studies in materials selection. Cambridge, Granta Design Limited.Bassetti, D, Brechet, Y and Ashby, MF. (1997). 'Estimates for material properties: The method of multiplecorrelations.' to appear in Proc. R. Soc.

Cebon, D and Ashby, MF. (1992). 'Computer–aided materials selection for mechanical design.' Metals andMaterials. 8(1): 25-30.

Cebon, D and Ashby, MF. (1994). Cambridge Materials Selector User’s Manual, Version 2.0. Cambridge, UK,Granta Design Limited.

Cebon, D, Hope, MBH, Charlton, C and Ashby, MFA. (1994). 'Megabytes on Coppers.' CD–ROM. GrantaDesign Ltd and the Copper Development Association.

Esawi, AMK and Ashby, MF. (1997). 'Computer-based selection of manufacturing processes.' CambridgeUniversity Engineering Department, Rept No. CUED/C-EDC/TR50.

Harmer, QJ. (1996). 'Selection charts for rolling element bearings.' Cambridge University Engineering Department,Rept No. CUED/C-EDC/TR47.

Harmer, QJ, Weaver, PM and Wallace, KM. (1996). 'Design-led component selection.' Submitted to CAD.Huber, JE, Fleck, NA and Ashby, MF. (1997). 'The selection of mechanical actuators based on performance indices.'Proc R. Soc. 453: 2185-2205.

Weaver, PM and Ashby, MF. (1996). 'The optimal selection of material and shape.' Engineering Design. 7(2).

7.APPENDICES

Appendix A

There are many computerised sources of information about materials. Some of these provideselection facilities, but few satisfy the strict requirements necessary for the screening process,described in section 2. They can all be used as sources of supporting information. They includegeneral packages like 'MaterialSpec' by Autodesk; 'M-Vision' by PDA; and the CambridgeMaterials Selector (CMS), described above. There are also many class-specific packages,generally produced by material manufacturers or trade associations. Notable among theses are: the'Plascams' polymers database, by the RAPRA; the 'Aluselect' aluminiums selector, by Euroal; the'Campus' discs, featuring products of various European Polymer manufacturers; and informationdiscs for the UK metals trade associations by the Engineering Information Company.

18

Appendix B: Examples Range Checks and Correlation Checks

Ranges of Young's Modulus E for Broad Material Classes

MATERIAL CLASSAll SolidsMetals: ferrousMetals: non-ferrousFine Ceramics*Glasses

Polymers: thermoplasticPolymers: thermosetsPolymers: elastomersPolymeric foams

Composites: metal-matrixComposites: polymer-matrixWoods: parallel to grain

Woods: perpendicular to grain

EL (GPa)0.00001

704.691470.12.50.00050.00001

812.51.80.1

EH (GPa)

10002205701000834.1100.121802403418

* Fine ceramics are dense, monolithic ceramics such as SiC, Al2O3, ZrO2 etc.

Limits for the Group αTm and αTg for Broad Material Classes*

Correlation* CL < αTm< CHAll SolidsMetals: ferrousMetals: non-ferrousFine Ceramics*Glasses

Polymers: thermoplasticPolymers: thermosetsPolymers: elastomersPolymeric foams

Composites: metal-matrixComposites: polymer-matrixWoods: parallel to grain

Woods: perpendicular to grain

CL (x 10-3)

0.113260.318113516100.126

CH (x 10-3)

562721243354156372010417

*For amorphous solids the melting point Tm is replaced by the glass temperature Tg.

19