Chapter 2

There is an extensive literature on materials selection but all textbooks treat the subject in terms of information which must be learned rather than techniques to be acquired. Typically the subject is presented as lists of materials with typical properties and extensive examples of their use [Cra84, Lew90a]. Techniques for examining design problems in order to derive materials selection specifications are rare, as are good formal techniques for assessing the relative suitability of candidate materials. The former are often assumed to be merely the ordinary analytical methods of engineering science and the latter assessment techniques are commonly presented without any distinction between those based on physical models and those which merely systematize the designer's intuition [Die83]. The specialist theme on decision systems later in this book describes these systematic assessment methods in more detail, particularly those which have been built in to commercial materials databases.

This chapter first sets the scene in which materials selection occurs in terms of the processes of engineering design and manufacture and then describes a set of methods based on simple physical reasoning developed by Ashby. The discussion is broadened to generalize from specific recommendations and to take account of aspects which the Ashby method either implicitly assumes or does not address.

There has been increasing interest in the development of databases of materials and materials' properties over the past decade. A number of public and commercial databases have been developed but few apart from the most modest have realistically lived up to the expectations and plans of their constructors. One goal of this book is to attempt to show why the initial promise has been rarely achieved in practice.

Computer-based engineering materials information is vital to the further improvement of design techniques, especially the integration of design with manufacture. While some of the challenges associated with representing Representation:materials information; property data in computerized databanks are now well known, development of complete materials information systems still involves many complex and intractable difficulties. There are many conflicting goals in representing materials information and there is, as yet, no generally useful data model for describing materials data.

Two problems have arisen which a simplistic analysis would dismiss as trivial: first that of attaining agreed ways of identifying materials, and second, ways of relating 'similar' properties. These problems are due partly to the sparsity of real, detailed data.

Materials selection is often considered an activity which takes place at a specific time and place in the course of an engineering design, usually being reassessed from time to time as the design evolves. In all past materials selection systems there has been the implicit assumption that all the information needed to make a materials selection decision is available: information that is unavailable has no effect on the decision. Taking the decision based only on available information can be termed the core problem of materials selection.

On reassessing a selection, previously unknown information which has become available is then taken into account, but studies of the action of selection, what the person actually does, still assume that the decision is taken at a single 'atomic' time during which the information available remains constant. Previous studies work on the basis that the important thing is the procedure for evaluating a given amount of information, either by a constraint system or by a ranking scheme [Neu90].

Actually that is not true. The following two examples, which I hope do not appear too contrived, show why.

Imagine that you are the materials engineer working for a company producing children's stackable play bricks. You are told that children want to be able to put windows in the houses they build, and that you must select a new transparent material for the company to use in a new range of bricks.

The problem is that although you are currently producing a wide range of colours, the polymer grade is the same for all and because it is 'filled' with cheap silica flour, it cannot be made transparent. The same polymer unblended has quite different mechanical properties, in particular it is too flexible in addition to being more expensive. If the surfaces of the bricks are too rubbery then they will mate so that the pull-off force will be much higher than the coefficient of friction suggests.

The task is to find a different polymer which has the same 'stiction' properties as the original so that the lugs and sockets on the new window-bricks make good joints, with the correct pull-off force, with the old bricks and with each other. You consult your materials database but 'stiction' is not listed.

The situation is simply that you have some materials information: a standard friction test against a steel ball-bearing, the modulus, and typical surface roughness after injection moulding, and you must create a plan to use this information to synthesize an estimate of the 'stiction', which is the information you need to make the selection. In addition you will need to check that your estimate is good enough by planning some simple tests (such as getting some samples you have to hand, rubbing them against each other and scratching them with your fingernail) before proceeding to evaluate a lot of data on different polymer grades. Your confidence is not high however, and you will order samples from a number of grades with quite different friction and elasticity trade-offs for more rigourous testing because your ideas about what affects stiction might, after all, have been wrong.

You are the designer of a part consisting of a steel rod bearing against an aluminium alloy casing. The alloy is the material to be selected. The rod is hemispherically-tipped and is held rigid by being made to indent the case on assembly. During service occasional axial overloads of known size on the rod are expected, and the device is intended to indent more deeply, but without moving far enough to dislodge the rod from a fixture at the other end. After one overload another will not occur for a known minimum period, during which time the device will be replaced.

What is required to make the selection is accurate knowledge of the hardness of the candidate alloys, or rather the incremental hardness between the assembly load and the overload. The conditions most resemble the Rockwell test (but with different loads), but the indenter is spherical as in the Brinell test (but with a different radius). The accuracy required is sufficient that the small distinctions between these data become important. How are the candidate alloys to be assessed ?

A sensible course of action is to look up all the types of hardness number available for the alloys in your databases, and to look for correlations between them and other property data which would appear to be relevant, such as yield stress and work-hardening rate. (Different types of indentation geometry 'see' the material differently so that while under-aged and over-aged alloy may have the same Vickers hardness, the Brinell hardness may be different.) You might arrange to have some alloys, for which data is missing, tested and then you might look for materials of particular Rockwell hardness, provided the Brinell hardness fell in a certain range relative to it, but not if the work hardening rate were more than a certain level. Eventually you will need to have a jig made up of exactly the same geometry as the part being designed in order to test the candidate materials that you preselect from the mechanical properties data.

From these two examples, it can be seen that the information needed to make a materials selection decision is never all available at the same time. Most of the decisions made are those concerning what more information to get next, and what cost and time penalty is acceptable for which information, rather than selection decisions themselves. The decisions require reasoning about materials data available and how these relate, through fundamental knowledge of materials behaviour, to the properties required for service and manufacture.

We have to consider the entire course of the design of an engineering component. An individual component made from a single material with a specified set of required behaviours is taken as the basic item on which materials selection techniques are practised. (Of course one possible outcome of any selection method should always be that a single material is not appropriate.)

The design of a component passes through many hands [Sta89]. Typically it is initially designed on a CAD (drafting) system and then passed to an numerical analyst who meshes the geometry and runs a series of finite element analyses (FEA) to assess its fitness for the use to which it will be put, e.g. stress and temperature analyses, creep, thermal expansion, and perhaps some attempt to predict fatigue life. The analyst passes the design back to the synthesist (the designer) in an iterative loop which refines the geometry and modifies materials selection. When the design becomes stable the complete set of component drawings is passed to the production (manufacturing) engineer who assesses the part for ease of manufacture and ideally makes only fine tuning adjustments to the geometry and selected material (see Figure 2.1). Finally maintenance issues and the environmental factors of disposal and recycling may be considered.

The current trend is touse software to attempt to bring as many as possible of the downstream assessment activities 'up front' to the designer's hands, a process termed concurrent design or simultaneous engineering. The attempt will be incomplete even in scope for decades yet, but every additional aspect covered, however small, either reduces the time taken or increases the quality of the design and so is worth striving for.

Materials selection is not an activity in a design which can be considered in isolation of the downstream activities; it is not possible to write an independent software 'materials selection tool' and to plug it into existing CAD/CAM software packages. (Only some unusual, particular aspects of materials selection could be so treated: selection for primary function only, or selection for toxic waste disposal for example.)

The task ahead is to identify and to produce materials information for all the life cycle aspects of the component and in forms appropriate for use in design in general and appropriate for use

Materials information is required for periodic structural assessment of critical components such as pressure vessels and aircraft wings (e.g. British Standard PD6493). Structural assessment has a great deal in common with design but because it is applied to pre-existing types of plant its requirements are much more specific even than that required for detailed design. Data is required for the actual material that a component is made from, not merely good candidate materials.

Some materials information will be embedded in the constitutive equations which non-linear finite element packages manipulate, either for function (creep, deflection etc.) or manufacture (injection moulding flow, rolling or forging forming). Some will be explicit: density, stiffness, price etc. Much will be non-numerical, such as the compatibility of adjacent materials for corrosion susceptibility (which would depend on a topological description of the components).

The implication is that materials engineering research would be most directly useful to engineering design if researchers could orient their work to produce the kinds of material behaviour descriptions which could be 'slotted in' to work together with CAD/CAM software. The tasks of reinterpreting and restating research findings into machine processable forms are likely to grow in importance. Unfortunately, by default these tasks are left to numerical analysts with little metallurgical background since the great majority of materials science graduates have insufficient mathematical or computer science training to cope.

A small minority of commercial materials databases are oriented towards experimental test results but the majority are aimed at supporting the engineering designer. It has been observed that there is a significant tendency in engineering organizations to treat materials data as something which is or should be available free or for minimal cost. Because of this attitude almost no numerical databases (or 'databanks', the term used when certain management controls are put on the database) have been able to show a profit [Vve87]. Subsidiary reasons for poor use by engineers are difficulty of access and lack of understanding [Rum87] and, more commonly, simple inappropriateness [Amm88]. Designers require a stable context and so prefer to work with unchanging information. They thus distrust data which is automatically updated; whether computerized or on paper.

The data used in the selection of a material for a component being designed, and how that data is used, depends greatly on the stage which the design process has reached. Designs are not produced in complete detail before they are analysed in any way: the whole process described in Figure 2.1 is carried out several times at increasing levels of detail.

These are often termed the conceptual, the embodiment and the detailed stages of design [Fre85, Pah84, Hub88, Ash89a]. At the conceptual stage the analyses for function and manufacturability are typically carried out in the designer's head; for embodiment there will be consultation and after the detail stage a complete product model (drawings) will be produced. These stages too have information feedback which modifies previous assumptions.

Figure 2.2 illustrates the situation, it describes a set of processes 'orthogonal' to those of Figure 2.1. Creative design usually involves learning whilst producing trial designs which lends a special character to the feedback information. The entire process can be considered as contributing to an evolving and increasingly detailed specification. Conceptual design, and the embodiment design of some types of component, can be done purely on a basis of the material class, but all detail design, and much embodiment design, require more detailed information to be used in more specific ways [Ben89]. The implications for the usefulness of different types of materials database are profound.

With enough commitment and effort, and a careful choice of the properties to be stored, it is possible, though still not easy, to construct a database of properties for a hundred or so materials classes which is complete. A complete database is one where there is a value for every property for every 'material'. In this case each 'material' is a materials class.

It has been estimated that there is a choice of something like 80,000 different materials that can be used in mechanical design. These are commonly grouped into five simple categories: metals, polymers, ceramics, elastomers and composites, but all competent engineers have a good feel for something like a hundred distinct classes of material, such as thermoplastic, potting-resin, stainless steel, brass, hardboard, cork, glass fibre-wadding etc. Experienced design engineers will be familiar with several dozen materials in each of several of these classes.

It is possible to construct complete materials class databases for two reasons. First, for each class there is a large number of different materials and it is only necessary to find one of them for which a particular property has been measured. Second, the level of precision of any property is low because of the inevitable spread among different members of the class, so it is possible, indeed necessary, to merge distinct properties into a property class at the same time. For example, when storing the hardness property of the class of brasses it is only necessary to find a few brasses for which the hardness has been measured, not all of them, and in addition it is possible to ignore the slight differences between Vickers, Brinell and Rockwell hardness testing methods.

If no experimental data can be found, then at the level of a material class a materials engineer would have no difficulty in estimating the missing value and putting it in the database with an appropriate annotation.

If a more detailed database is attempted with perhaps a tripling of the number of classes from 100 to 300, it is not only three times harder to find experimental or standard values, it also becomes necessary to distinguish or to explicitly convert from one type of property to another. This occurs because at this increased level of detail the differences become important between, say, Vickers and Brinell hardness, or double-cantilever beam and double-torsion measures of fracture toughness. This increase in the level of distinguishability necessary is the root cause of problems of both designation (identification) and 'similar' property conversion (see below).

It is observed that databases holding data about individual materials (leaving aside for the moment the precise meaning of 'individual') rather than classes are extremely sparse. This is because the complete data set simply does not exist and has never been measured. Contrary to commonly held belief, complete and detailed material property information is simply not readily available and there is an extensive literature which supports this point [Bam89, CEC86, Gra86, Kau88, Krö87b, McC87, Owe87, Sar89a, Sar90a, Swi85, Vve87, Wes86].

If a database is complete, then simple and powerful indexing and search techniques are possible. Sparse databases are far less easy to use effectively. This is not just inconvenience: null data introduces quite fundamental difficulties in the types of operations possible with relational databases [Dat90].

The strength of the set of techniques associated with the Ashby selection charts is their foundation in the physics of the design problem being addressed.

This new set of methods for materials selection includes a memorable graphical method of displaying materials on a log-log plot of one property against another (see Figure 2.3). Ashby has devised a methodology for analysing simple problems such that the merit indices which determine the optimal materials properties can be found even if the design problem itself is not completely specified [Ash90, Ceb91]. It can thus be performed early in the design process, during conceptual design, rather than having to wait until a fully-detailed stress calculation has been made.

Merit indices are a well-established concept in materials selection [Die83] though devising appropriate indices for specific design problems is often an intellectual challenge [Lew90a]. They are best described by an example: for a simple beam in bending, where the design is deflection limited and a beam of minimum weight is required, materials which have the same value of E/r^² (the merit index) will be equally good. If the problem involves a plate rather than a beam then the merit index is E/r^³ (where E is the modulus and r the density).

On a log-log plot with the appropriate axes, materials that are equally good lie on straight lines. In Figure 2.3 the better materials lie to the upper left of the diagram. Thus for light, stiff beams the best materials are (crack-free) ceramics, engineering composites or possibly wood, but for plates they are clearly wood.

The visual impression is sufficiently useful that a sequence of manual selections made on a number of diagrams on paper serves to perform a complex selection for conceptual design without requiring calculation or computerized databases. This is possible because each diagram contains all the required information in an easily accessible form.

This log-log display of material properties is used as part of the user interface to a computerized materials database and selection system: the Engineering Materials Selector (EMS) currently under development by Cebon and Ashby in the Cambridge University Engineering Department [Ceb91].

The full power of the method becomes apparent when the charts can be produced quickly and easily with axes which are not just predefined properties but new combinations of properties (merit indices) which accurately model a specific design problem. Thus a chart of l/a (thermal conductivity divided by heat capacity) plotted against (Kc^{^2/3})/r (fracture toughness to the two-thirds power divided by density) would compare thermal distortion with the brittle strength of a rod in bending, a not-unlikely combination for the study of thermal shock induced fracture.

There are two interesting aspects of the Ashby method that are critical but not obvious, though they are documented to some extent in more complete descriptions of the technique [Ash89a].

First, the diagrams are most useful for selection at the conceptual stage of design because of the reliance on complete data being available for every property, for every material. The sparseness of real data implies that data from several closely related materials can be, and must be, merged as a materials class to get a complete set. This implies that the method only works for those properties for which it is easy to identify classes of materials which have similar property values. This is true for the properties such as stiffness and thermal expansion, but largely false for properties such as corrosion or wear resistance. These capricious properties are discussed in more detail later.

Second, shortlists of materials produced by each chart are independent. Each chart fully encapsulates the specification of one of the behaviours required by the finished component; if it does not then a modification of the merit index to use an additional or different property can always ensure that it becomes so. There is seldom any connection between design specifications [Fre85] so trade-offs between one chart and another have no physical interpretation and a fall-back must be made to the techniques of decision analysis (see Chapter 5).

Thus the Ashby methods reduce the dimensionality of the problem by combining several properties into a smaller number of merit indices and the charts provide an excellent human interface for software design aids. They are complete in data, so they give accurate guidance to the types of material that must be investigated more specifically and they are based on physical models so they give guidance to the types of properties that must be used to evaluate those more specific materials.

A lingering concern is whether most designers have been properly trained to, or are fully capable of, abstracting the physical essence from their design problems to derive correct merit indices (even some text book authors make mistakes [Lew90a]) and, if not, how a library of pre-packaged physical models (with associated merit indices) could best be presented or remembered. A simple list of merit indices and boundary conditions is not sufficient [Ash89a, Ash90]. The majority of design problems produce a set of merit indices with differing boundary conditions. Some must pass a pass/fail criterion whereas others should just be as high as possible, with no threshold value [Fre85, Hub88]. These boundary conditions are just as much a part of the physical model as the merit indices themselves but are harder to present simply.

There are five main techniques for materials selection which are used today in addition to the new Ashby/Cebon methods. These are first discussed briefly, then a comparison is made.

Most industry, quite rightly, uses the evolutionary technique for its day-to-day adjustments to changing conditions, usually driven by the batch-to-batch variations from materials suppliers. For slightly greater changes they will attempt to reason logically but qualitatively about materials and properties. This involves arguing that what is required is something like the current material except that it must also have some extra property, looking for the most similar material which has that property and then redesigning the component or adjusting the manufacturing process so that the new property can be used. A software aid which reasons in this manner is under development [Dem88, Zuc89].

The need to become more effective and efficient in manipulating large quantities of materials data is pushing many data suppliers to computerize. Once the data is in a database user-friendly software can (in principle) be used to entice a potential user into believing that as well as holding the data required for material selection, the programming will also take away much of the pain of looking at the data as well.

The simplest database functions, usually available as standard with current business database packages, include the capability of filtering the list of materials to extract only those where properties exceed or equal some minimum target: e.g. strength greater than 200 MPa and sea water corrosion equal to A or B (on a scale of A to F). Facilities currently available in many modern personal computer packages include using calculated combinations of properties to filter or to rank an extracted list, although of course such options have always been available if explicitly programmed.

The only formal technique for materials selection usually taught to students is the weighted objective function method [Die83], one of the class of decision analysis techniques. The required properties, strength, density etc. (x) are assigned a relative importance or weight (w) for the task in hand.

The data are reviewed for all the candidate materials and those materials with the best values for the more important properties are selected. The calculation and weights define the objective function to be maximized. The favourite type and most common is linear addition (Equation 2.1).

The problems of normalization, fairness and consistency are considerable and subtle (described fully in Chapter 5 on decision techniques), but this is the primary user interface method of the important polymers database PLASCAMS and is likely to be the only formal materials selection technique known to most engineers.

Expert systems for selection contain pre-recorded decisions of the materials best suited to particular circumstances and information and procedures to help a user to find the recorded situation which is the best match for current circumstances. They also synthesize new material suitabilities by deduction from the elementary information they contain [Har87, Lee89]. This distinguishes them from database systems which contain data and provide procedures for the user to arrange, order and classify the data in any way that seems appropriate (see Chapter 7 on knowledge-based systems). In practice any useful knowledge-based expert system must contain or access a substantial database of materials, properties, typical uses and known hazards, and many database systems calculate data that they do not explicitly store (such as merit indices), so the distinction between the two is often not clear [Har85, Kel86, Kai88, Hay89].

A primary requirement of any technique for materials selection is incremental stability. The behaviour must be such that refinements of users' input lead to refinements in the system's shortlist, not complete changes of sequence, otherwise it will not be trusted. We have already seen that there are immense difficulties in persuading engineers to improve their use of materials information so this is not a trivial matter.

The formal methods for materials selection fall into two classes, those based on multiple constraints with shortlists and those based on objective functions and metrics [Neu90]. The constraint methods are incrementally stable whereas those based on objective functions are not.

The target property and Ashby/Cebon methods are constraint-based techniques since each target property or merit index yields a separate shortlist of materials. If the same material is at the top of all shortlists then the task is completed, but that is unlikely.

The usual technique is to consider in turn all the materials on all the shortlists, starting with those that appear on all of them. The danger exists that there is some material in the database which has just missed inclusion on one or more shortlists, but which otherwise has quite good properties; so it is instructive to review which materials passed and failed which property constraints [Gra86, Ben89, Ceb91]. (Other methods exist based on the concept of dominated alternatives and these are discussed further in Chapter 5.)

If the inclusion levels are set too narrowly then there is a higher likelihood that a useful material may be missed completely, whereas if they are too broad (and if there are many different shortlists) then the total number to be checked becomes overwhelming. The method of setting constraints by visually moving a line over a graphical summary of the data, as used with the Ashby charts, is a powerful technique for ensuring that the right balance is struck between the two extremes.

The weighted objective function technique described above involves finding the materials furthest from the origin in the direction in which the function increases fastest in an n-dimensional space, where each axis represents one of the material properties (or a compound property such as a merit index);.

It can be seen that 'furthest' means that a metric is required for the space. This metric is determined by the weights that the user puts in to the system which determine how much a 'distance' along one property is worth in terms of 'distances' along the others. A value to the designer of one property compared with another is required.

If any weight is changed the whole set of objective function contours tilts so that a different shortlist of best materials appears. This incrementally unstable behaviour can be unnerving but it arises because no material is ever rejected, it is just moved up and down the ranking of all the materials.

Some constraint selection techniques work by including materials and some by exclusion;. These two techniques are identical except where some data is missing, estimated or uncertain, which is of course all the time.

The humblest approach, and the one also taken by the PAL adhesives selector expert system [Lee89], is to exclude a material only when there is definite evidence that it is unsuitable. This accords with the view that materials should remain candidates pending new information (see below). If there is no great need to reduce the size of a shortlist then excluding materials based on an estimate of their properties would seem to be dangerous, However this is the way in which most selection systems that use estimated data operate: they treat it the same way as real data as far as shortlisting is concerned (although of course it is labelled as estimated). These systems are thus neither entirely inclusive nor exclusive in their approach.

Any system which works even partially in the inclusive manner is implicitly assuming that it contains all the materials and properties which are relevant to selection decisions, since uncertain or absent information leads to materials not appearing as candidates on the final shortlist. This leads to scale-up problems as databases increase their depth and scope while attempting to retain the same selection aiding mechanisms. As databases get bigger and more detailed the proportion of missing information gets higher; thus techniques which work well with small, complete databases based on generic properties and materials classes can be expected to run into difficulties if the database is expanded and enters a regime where other, perhaps more complex, techniques are appropriate.

It is likely that large, future systems will not only have several mechanisms for generating shortlists, but several mechanisms for combining shortlists too.

The type of thinking described in the two examples at the beginning of this chapter (toy bricks and indented aluminium) is different from the Ashby materials selection methods because good physical models for the properties required often do not exist, hence merit indices cannot be derived. An engineer could attempt to set weights on properties thought to be important, but selection on that basis would only be a tiny part of the whole selection process.

It is clear that there is much more to materials selection than the core problem of using some algorithm to pick winners based on current information. Most of the important activity actually occurs in the 'grey lines' of Figures 2.1 and 2.2 which represent feedback, and not in the labelled boxes at all. Only one current project takes this viewpoint: the SPLINTER system [Dem88, Zuc89] makes it a fundamental principle that information is always incomplete and that different properties are not properly defined with respect to each other. It aims to support a strategy which ensures semantic consistency while reasoning about increasingly detailed descriptions of categories of materials.

It is the balance between the value of unavailable information, the probability of it having certain values and the cost of obtaining it which form the crux of the problem. This is the classic problem of the value of uncertain information; techniques for estimating it and for rationally planning actions based on it are covered in basic textbooks on decision theory [Wat87, Daw88, Nef90]. So the principal action of materials selection is a cost/benefit decision using uncertain information and not the traditional core problem of selection among alternatives based on precise information. If the cost to obtain a more accurate parameter value is too high in comparison with the benefit that arises from the higher accuracy then that parameter value is adequate for the design at that stage..

One way for a designer to plan how to get different information between one decision point and the next is to consider the cost to the designer of each course of action. In the real world all information on which a decision is based has a cost associated with it, a cost best expressed as hours which a designer would have to spend ferreting it out. Current mental knowledge is free, books and tables require the effort of going to a bookshelf or a library and perhaps the time to use an index and then to transform the information found into a form useful for the design in hand, e.g. ksi to MPa, or hardness to yield-strength. Further effort is required to learn to use dial-up or CD-ROM databases. No single cost measure is adequate, the designer must also balance the costs of elapsed time (if someone else is sent to do the job), and cash in the case of buying the data or paying for some tests to be made. All unavailable data thus appears in the decision process as a set of costs (in three currencies: personal time, elapsed time and money) and accuracies (of the data available at each cost).

This view of materials selection leads to a clear conclusion: rational design relies on estimates of the costs of the appropriate materials data. These estimates are probably more important than the data itself. If the designer has an accurate view of them then the work can be planned in the optimum manner but if there are only inaccurate estimates then his design will be worse than that of a more cost-aware engineer, even if the latter has worse quality data to hand. Reducing the costs of materials data means that individual designs are cheaper and quicker, but increasing the accuracy of the cost estimates means that design is more rational in the long run. It also enables designers to do well what they do best: balance a number of competing and cooperating factors.

There are no computerized or manual databases which provide estimates for the costs of obtaining materials property data of different types. This is a serious impediment to the development of software tools to aid the important part of materials selection, the planning of the information gathering, rather than the easier process of making decisions based on available information.

Teaching expertise in materials selection to apprentice designers currently concentrates on making them aware of what information is available where. The view presented here indicates that teaching the cost and reliability of different information sources is equally important.

It is interesting that the costs of different types of material parameter are weakly dependent on the material and strongly dependent on the property. Knowing which is which, that hardness data is easy but that creep data is expensive, then assumes a greater importance in teaching. Better cost estimates would result if:

Reducing the cost of property data is always helpful, if not crucial. On this model of behaviour we can see several ways of doing this, but this aspect is already well known. Cheaper searching for materials data can be aided by:

So a set of materials data as presented by some future database might give the result shown in Table 2.1 if it did not have the data to hand. Any package sophisticated enough to do this cost estimation could probably provide the first estimated value automatically, and without further cost. For example, the strain-hardening rate of a particular aluminium alloy 6021-T6 can be estimated by the following procedures:

fast, very rough	Take the numerical average values for all the other 602x alloys in your databook.
quick, rough	Examine the specification of the 6021 alloy, consult a metallurgist as to the hardening mechanisms and determine the precipitation and solution hardening effects. Pick another alloy of similar expected behaviour and and use its strain hardening value.
1-5 hours	Do a literature search using indexes (Citations Index) for academic papers which used 6021 and see if any measured its strain-hardening rate. [No certainty of getting data, but fairly sure to get some if it actually exists.]
2-20 hours	Use a database, arrange subscription, connect personal computer and modem, log in to a remote system etc. If no result try another database. [No certainty of getting a result even if it does exist somewhere.]
3-10 hours	Do a series of ball-indentation tests and use the Tabor correlation of the variation of hardness with load to get an estimate of the strain-hardening index. [Certain to give a result.]
30+ hours	Make up some tensile specimens and measure a stress-strain curve. [Certain to give the best answer.]

A software system which could provide such estimates could be a useful addition to traditional materials property databases. It would have to be a knowledge-based system but requiring only a very limited type of knowledge (see Chapter 7).

The foregoing arguments in this chapter have tacitly assumed that there does exist universal agreement on what is meant by a material and what is meant by a property. Unfortunately this is not so, which makes the procedures of materials selection, and more generally the production of any kind of engineering materials database, more difficult than might be expected. This is the designation problem due to levels of distinguishability referred to earlier.

The identification of a material is a definition of exclusion based on a particular set of properties at a particular date and time. This definition might seem obscure, but it comes about because the only use of identification is to distinguish one material from another, not to identify it absolutely because that is impossible. These counter-intuitive arguments will now be explained in more detail.

For particular material properties it is possible to ensure greater accuracy by specifying that the same property be measured by two different methods, e.g. density by x-ray lattice parameter determination as well as by Archimedes method, or Young's modulus by ultrasonic means as well as by extension. However in both those cases there are real systematic differences between the two measurement methods because they are measuring slightly different properties (x-rays do not take vacancies, dislocations or porosity into account, ultrasonics do not 'see' anelasticity caused by dislocation movement).

The whole idea of materials having independently identifiable properties is a convenient fiction, in practice there are only specimens and measurements. Different groups of people have different ideas of how experimental measurements should be abstracted into designated 'properties' and how sets of shared property values should be used to designate 'materials'.

A 'single' material is defined differently depending upon the aspects of the material most important at the time. For example, for the purpose of recycling soft-drink cans it is only necessary to distinguish 'aluminium' from 'steel' (most easily by testing for ferromagnetic properties). However when aluminium alloys are welded it is necessary to distinguish 'heat-treatable' from 'solution-strengthened' alloy types in order to avoid poor quality in the heat affected zone. Much more precise alloy specifications are required when planning a manufacturing process, such as making soft-drinks cans. Here not just the alloy composition and heat treatment must be specified, but also the precise sequence of heating and rolling and even the directions in which the sheet has been rolled. These increasingly detailed designations are often hierarchical, but not always: if another set of properties is introduced, such as corrodability, then a complex lattice of relationships results.

A complication is that the very concept of 'materials data' is inappropriate for advanced materials whose microstructure is created in the same operation as the final component, e.g. ceramics and most varieties of composites [Krö87a, Wu91]. There is no 'material' which can be abstracted from the results of tests on differently shaped specimens, except with respect to simple scalar properties such as density or water-content.

Few databases for materials selection have lived up to the expectations and plans of their constructors .

Development of complete materials information systems still involves many complex and intractable difficulties due largely to two things. First, a lack of what might be called a reference model or planning model for describing formally how materials data, information and knowledge are generated from test results and applied to design problems in different industries. Second, a lack of people trained in useable techniques (where such techniques exist) which can define and describe representations for data, information, knowledge and concepts which are necessary if a reference model is to be useful.

Future databases will continue to fail to integrate with engineering design, and the problems of materials identification and property definition will persist until ways of structuring and relating materials information in databases have been developed and agreed as a basis on which to proceed. That is the subject of the next chapter.

property	value	accuracy	cost
modulus	100 GPa
yield	200 MPa
hardness	600 MPa
ductility	n/a	(20%)	£20
ductility	n/a	(10%)	£200
ductility	n/a	(3%)	£1000