Int J Mech Mater Des (2006) 3:209–222

DOI 10.1007/s10999-007-9024-4

Gear material selection with uncertain and incomplete data.

Material performance indices and decision aid model
A. S. Milani Æ A. Shanian

Received: 18 October 2006 / Accepted: 21 March 2007 / Published online: 10 May 2007

Springer Science+Business Media B.V. 2007

Abstract First, a set of major gear design criteria multi-criteria approach may also be useful in reveal-
are used to develop six material performance indices ing incomparable and/or indifferent alternatives that
for material selection purposes. They are, a surface would not be distinguishable otherwise. To ensure
fatigue limit index, a surface fatigue lifetime index, a rank stability of the chosen materials, the effect of
bending fatigue limit index, a bending fatigue potential uncertainties in designers’ opinions during
lifetime index, an abrasive wear index, and a criteria weighting is introduced by means of a
machinability index. A modified decision matrix in dilation and concentration sensitivity analysis
the presence of data uncertainties and incompleteness process. The main originality lies in resolving a
is then proposed to show the effect of the developed non-compensatory, application-specific, and uncer-
indices. It is shown that using individual material tainty-based multiple criteria material selection prob-
properties and approximations among them may not lem in gear design optimization.
satisfy specific design goals for a specific application.
Next, the ELECTRE III multiple criteria decision aid Keywords Gear design
Material performance
(MCDA) model is applied to rank the best compro- index
Multiple criteria decision making
Imperfect
mised candidate materials, while considering criteria data
ELECTRE III
tradeoffs, designers’ preference information, data
uncertainties and incompleteness. An effort is made
to reconcile mathematically motivated model thresh- 1 Introduction
olds (namely, the indifference, strict preference, and
veto thresholds) with experimentally motivated char- 1.1 Gear material selection
acteristics such as upper and lower limits of measured
material properties. It is shown that the proposed Optimal design of gears requires the consideration of
both material and geometrical parameters (Hofmann
1990; Ognjanovic 1996). From a tradeoff point of
A. S. Milani (&)
view, a choice of stronger material parameters may
Department of Mechanical Engineering, Massachusetts
Institute of Technology, 77, Massachusetts Ave., allow the choice of finer geometrical parameters, and
Cambridge, MA 02139, USA vice versa. An important difference among the two
e-mail: milani@mit.edu types of parameters, however, is that the geometrical
parameters are often varied independently (e.g., the
A. Shanian
Rolls-Royce Canada, 9500 Cote de Liesse, Dorval, QC, face width and diametrical pitch). On the other hand,
Canada H8T 1A2 material parameters can be inherently correlated to

123
210 A. S. Milani, A. Shanian

each other and may not be varied independently, an hardness, HB, is shown in Fig. 1. As seen, no
example of which being the variation of the bending apparent trend may be sought for variation of the first
fatigue limit with the core hardness for some steel two parameters, particularly UTS, with the core
materials (Horimoto et al. 2003). If one allows these hardness; the CS-curve in this figure will be referred
parameters be varied independently in an optimiza- to in a later section. To avoid the difficulty, depend-
tion problem, it may result in infeasible solutions. ing on the application, it may be possible to uncouple
That is, the final choice of material may not be the material selection process from the geometry
possible within available data bases. optimization process (Milani et al. 2005). Once a
When gear material and geometrical parameters good material is captured using a selection model
are optimized simultaneously, it is common to where all properties are accounted for, they can be
assume empirical formulas approximating a relation fixed and geometrical parameters are optimized using
between material parameters (e.g. the bending fatigue conventional continuous techniques, or vice versa. To
limit and the ultimate tensile stress as a function of choose effective materials, however, it is recom-
hardness). As such, the variability of material mended that individual material properties be
parameters is controlled by one or few parameters grouped into a set of performance indices to reflect
(see, e.g., Thompson et al. 2000) and the final choice particular design goals (Ashby et al. 2005). The fact
of material becomes straightforward. is that the individual material properties may interact
When the choice of material is limited to a list of differently in each application.
pre-defined candidates, two difficulties pose a chal-
lenge. First, a discrete optimization process should be 1.2 Material selection models
followed against material parameters. Second, prop-
erties of different candidate materials (alternatives) Ashby et al. (2005) has suggested material property
may not indicate any obvious correlation in the given charts for a wide range of material selection
list. An example of this situation is shown in Table 1. problems. The charts represent two main features:
The goal is to choose best performing material(s) fundamental relationships between material proper-
among nine alternatives. Based on average values of ties and the ability to choose an optimal material
data in Table 1, a plot of the bending fatigue limit and for a particular application based on predefined
the ultimate tensile stress (UTS) with the core performance indices. In a recent work (Guisbiers in

Table 1 Suggested materials and their properties in a gear material selection problema
Materials Material properties ID

Hardness Surface fatigue limit Bending fatigue limit UTS

(MPa) (MPa) (MPa)
Surface (Bhn) Core (Bhn)

Cast iron 200 200 330 100 380 M1

Ductile iron 220 220 460 360 880 M2
S.G. iron 180–300 180–300 480–620 240–440 590–1100 M3
Cast alloy steel 220–320 220–320 560–700 420–450 590–1100 M4
Through hardened alloy steel 220–320 220–320 600–740 500–580 800–1580 M5
Surface hardened alloy steel 519–565 192–265 1160 680 1850 M6
Carburised steel 601–692 256–337 1500 920 2300 M7
Nitrided steel 647–738 256–337 1250 760 1250 M8
Through hardened 160–210 160–210 450–550 420–440 560–710 M9
carbon steel
a
Data are taken form Hofmann (1990) where Vickers hardness values have been converted to Brinell values using conversion tables
in http://www.gordonengland.co.uk/hardness/brinell_conversion_chart.htm

123
Gear material selection with uncertain and incomplete data 211

Fig. 1 Example of 2500 0.9

variation of two material
0.8
parameters and a design

UTS & Bending fatigue limit (MPa)

factor with core hardness 2000 C s = -0.0007 H B + 0.8987
2
0.7
R = 0.9863
0.6
1500
0.5

Cs
0.4
1000
0.3

500 0.2
UTS
Bending fatigue limit 0.1
Surface condition factor
0 0
150 200 250 300 350 400 450 500 550
H B (Bhn)

press), using Ashby’s methodology, a study was Ashby (2000) and Brechet et al. (2001), where a
performed to provide a material design analysis for systematic procedure for multicriteria material selec-
microelectromechaniacal systems (MEMS). The aim tion was described using Ashby‘s charts and possible
of the work was to facilitate the selection of materials new directions were pointed out.
with minimum residual intrinsic stresses. Finally, the Another approach seen in the literature is the use
appropriate materials for the bridge of a MEMS-RF of multiple attribute decision-making (MADM) mod-
switch and a variable capacitor were presented. In els (Pratyyush and Jian-Bo 1998). Sample studies
Matos and Simplı́cio (2006), innovation and sustain- involving the application of MADM or MADM-
ability issues in mechanical design through material related models in material selection can be found in
selection was discussed. A case study to replace poly- Shanian and Savodogo (2006) (for a thermal loaded
vinyl chloride (PVC) for transport applications was conductor), Milani et al. (2005) (for a transmission
also presented. It was concluded that the natural gear), Dehghan-Manshadi et al. (2007) (for a cryo-
materials crock and wood are suitable candidates for genic storage tank used for transportation of liquid
reducing health damage and environmental impacts. nitrogen), and Rao (2006) (for a product operating in
In Ashby et al. (2004), efforts were made to select a high-temperature oxygen-rich environment). In
materials by taking into account a large number of problems dealing with MADM, the goal is often to
design and manufacturing alternatives. Interpretation choose the best performing material(s) from a given
of results, obtained based on different strategies in the list of alternatives which generally show no obvious
design of mechanical components, was presented. In dominance one over another. As a result, it can be
Esawi and Ashby (2004), a computer aided method- said that the solution can be a sub-optimum point on
ology for the selection of a joining procedure was the Pareto frontier. It is worth adding that apart from
introduced. The proposed decision tool allowed the selection applications, there are also MADM models
discrimination of joint geometry, loading condition, that are used for sorting and classification purposes
material, and other attributes, to identify the best (Collette and Siarry 2003).
subset of available process parameters that satisfy
design requirements. Other examples include a Remark In practical material selection, most often it
selection procedure for adhesive materials in (Suárez is impossible to say that a decision is good or bad by
et al. 2003), where different charts were presented to referring to a mathematical model alone. Factors such
solve general cases involving adhesive bonding. as prior experience, analogy, expert’s insight are
Inclusion of cost models within the concept of necessary to study solutions derived from mathemat-
material and process selection is found in studies ical routines before making a final decision. In light
such as (Esawi and Ashby 2003). Examples in a wide of this, multiple criteria/attribute optimization models
range of industrial applications can be found in are sometimes referred to as multiple criteria decision

123
212 A. S. Milani, A. Shanian

aid (MCDA) models, where the aim is to construct 1995; Juvinall and Marshek 1999). The definition of
something which is liable to help the analyst make a material performance indices should therefore be
final decision (Roy 1993). based on relationships characterizing these criteria.
Figure 2 shows examples of a failed gear due to the
1.3 Objectives and organization of the paper bending and surface fatigue failure modes. It is worth
adding that next to the choice of material, in practice,
In a previous work (Milani et al. 2005), the compen- a failure can occur due to factors such as faulty usage,
satory TOPSIS MCDA model (see, e.g., Pratyyush the wrong choice of dimensions, poor manufacturing,
and Jian-Bo 1998) was discussed and employed to etc. (Akinci et al. 2005). Among all, here the choice
solve a gear material selection case study. However, of material is studied.
no performance indices were adapted, nor were
material data uncertainties and/or incompleteness 2.1 Surface fatigue failure
accounted for. This paper, first, aims at developing
a set of six material performance indices that can be Surface fatigue failure is known to be the most
used for more efficient gear material selections. The common type of gear failure (Neale 1995; Fernandes
proposed indices are according to typical gear failure and McDuling 1997). One form of this failure is
criteria discussed throughout Sect. 2. Next, in Sect. 3, called pitting where due to excessive Hertzian stress,
the data from the case study of Milani et al. (2005) is in cyclic loading, relatively smooth-bottomed cavities
used to show the effect of the developed performance appear at or near contact surfaces (Juvinall and
indices in the decision-making process. Uncertainties Marshek 1999). Spalling is another form where areas
and incompleteness in material data are also pointed of the skin flake away due to a continuation of pitting.
out in this section. Section 4 aims at a new For surface-hardened gears, this failure can occur due
application of the non-compensatory ELECTRE III to the formation of sub-surface cracks in the case-
(Élimination et Choix Traduisant la Réalité) (Roy core boundary (also called case exfoliation or case
1993, 1990; Vallée and Zielniewicz 1994; Rogers crushing). Cracks may originate from the surface or a
et al. 2000; Collette and Siarry 2003) approach to subsurface, depending on the contact condition (see
solve the multi-criteria decision problem formed in Fernandes and McDuling 1997 for details).
Sect. 3. ELECTRE III can arrange select alternatives From a material selection point of view, the
(gear materials) into equivalence classes that are surface fatigue mode receives the most attention
completely or partially ordered. Data uncertainties, among decision criteria: if it is not prevented, the
often resulting from materials testing, are incorpo- ensuing pits or the removed pieces from tooth
rated into the model via the definition of criteria
thresholds. It is revealed that the proposed approach
may not only be useful to elicit the best performing
material(s), but also to estimate incomparable and/or
indifferent alternatives. In the same section, the effect
of uncertainties in criteria weighting by means of
dilatation and condensation operations is discussed to
ensure the robustness of the final solution. Finally,
concluding remarks are given in Sect. 5 and possible
extensions for introducing new design indices are
outlined.

2 Material performance indices

Typical failure modes considered in the design of

gear components are (a) surface fatigue (b) root Fig. 2 Example of failed gear due to (a) surface fatigue and
bending fatigue, and (c) wear (Wilson 1997; Neale (b) root bending fatigue (www.tribology-abc.com)

123
Gear material selection with uncertain and incomplete data 213

surfaces can lead to other failure modes such as should ideally yield a higher Sfe and UTS. Of course,
bending fatigue and/or abrasion (Fernandes and in the presence of other criteria, this ideal condition
McDuling 1997). Following reference (Juvinall and may not be achievable because of tradeoffs among
Marshek 1999), the effective (modified) surface them.
fatigue limit of a material, SH, is given by:

SH ¼ Sfe Cli CR (MPa) ð1Þ 2.2 Bending fatigue failure

where Cli is the service life factor and is often chosen Bending fatigue is known to be another major failure
based on the expected number of cycles before failure mode for gears in different standards such as DIN and
occurs (for 107cycles, Cli is taken as unity). CR is a AGMA (Hofmann 1990). Theoretically, when the
reliability factor and is chosen based on the expected induced bending stress on a gear exceeds the standard
reliability level (for 99% reliability, CR is unity). Sfe R. R. Moore bending fatigue limit, S0 n, micro-cracks
is the nominal surface fatigue limit measured in a may nucleate in the root region. If the excessive
laboratory condition (for 107cycle lifetime, 99% loading continues, the nucleated cracks can then
reliability, temperature < 2508F). Design factors such gradually grow and eventually cause full breakage of
as Cli and CR are dimensionless and serve to adjust the tooth (often called final or sudden fracture Neale
the laboratory measured values to actual working 1995). Similar to Eq. (1), after considering a set of
conditions.1 design factors, the effective bending fatigue limit, Sn
The service life factor Cli, next to the number of is given by Wilson (1997) and Juvinall and Marshek
cycles, may also relate to factors such as lubrication, (1999):
material, and residual stresses (Wilson 1997) (for Y
instance, it is known that compressive residual Sn ¼ S0n Ci (Mpa) ð4Þ
stresses can increase the surface fatigue strength 0

(Choi and Liu 2006)). When material properties in a where Sn is the nominal bending fatigue limit and Ci are
pre-defined list of candidates vary substantially in a set of design factors. Among different Ci used in the
their magnitudes, it may not be reasonable to analysis of the bending fatigue mode, the surface
disregard the material dependence of CLi. An esti- factor, CS, is known to be a function of both surface
mation of such a dependence to a property ration, finish (see Bayoumi and Abdellatif 1995 for a detailed
namely UTS Sfe , is shown in Appendix A. The ration
analysis) and material hardness HB (Wilson 1997,
connects to the slope of the material Basquin type S- Juvinall and Marshek 1999). With an increase of HB,
N curve. As a result, for SH to be optimal in Eq. (1), CS diminishes, and a lower Sn value according to Eq.
two material-related performance indices should be (4) is estimated. Therefore, to maximize Sn, a third
maximized: material-related performance index is defined as:

g1 ¼ Sfe (MPa) ð2Þ g3 ¼ S0n CS ðHB Þ(MPa) ð5Þ

UTS Similar to the approach used in Thompson et al.
g2 ¼ ð3Þ
Sfe (2000), here it is considered that HB = Hcore. Also
CS(HB) is approximated by a straight line (Fig. 1)
Considering Eq. (3) alone, it is noted that an optimal from the available test data in Juvinall and Marshek
value of g2 does not necessarily mean a higher value (1999) for a machined finish. Hence, Eq. (5) can be
of Sfe and UTS. On the other hand, an optimum value re-written as
of the material index in Eq. (2) requires a higher Sfe.
Hence, multi-criteria optimization of the two indices g3 ¼ S0n ð0:0007HB þ 0:8987Þ(MPa) ð6Þ

1
Concerning the longer service lifetime of the teeth
According to the AGMA standards, there are other design under the bending fatigue mode (discussed in
factors that can be considered for specific conditions. For
instance, a hardness ratio factor may be used when pinion is Appendix A), a fourth material performance index
substantially harder than gear (Juvinall and Marshek 1999) can be added as

123
214 A. S. Milani, A. Shanian

UTS Remark Generally speaking, a good lubrication

g4 ¼ ð7Þ condition can reduce abrasion by providing a thicker
S0n
oil film that keeps fine particles from scratching the
An optimal value of g3 requires a higher S0 n and a surface. Good maintenance, such as cleaning the gear
lower Hcore; whereas an optimal value of g4 implies a case and oil supply pipes, or the use of air cleaners
higher UTS and dust covers, can also prevent wear failure
S0n ration. Similar to the case of surface
fatigue mode, ideally, the relation among these two considerably (Neale 1995). On average, in gear
criteria should force the decision model to choose mechanisms wear failure is seen less frequently than
materials with both higher UTS and S0 n. fatigue failure modes (Akinci et al. 2005). As a result,
the abrasion criterion may be assigned a lower
2.3 Abrasive wear failure (abrasion) importance weight in a multi-criteria material selec-
tion model.
In some gear mechanisms such as in gearboxes,
abrasion is occasionally a cause of failure (Akinci
et al. 2005). It happens when foreign gritty materials
fall into the gear unit and rub against the tooth 2.4 Machinablity
surface. The ensuing grooves are normally formed on
the tooth flanks and in the direction of sliding Although machinability is often considered as a
surfaces (Neale 1995). It is important that wear not be secondary material property, it may be worth includ-
mixed with the Hertzian surface fatigue discussed in ing in a gear material selection problem as a second-
Sect. 2.1, as in one two members are in sliding level decision criterion. The total cost of a gear
contact and in the other they are pressed over each consists of both manufacturing and material costs
other. Based on the analytical wear equation, the rate (Hofmann 1990). For the manufacturing cost, the
of wear depth is given by Juvinall and Marshek easier the material is to machine, the lower the cost.
(1999): Therefore, between two materials of nearly indiffer-
ent fatigue and abrasion performance indices, a
 gpv decision model should be able to choose the one
d¼ (mm/s) ð8Þ
HSurface with better machinability. The criterion can be
characterized using the following relationship (Ju-
where g is a dimensionless wear coefficient and vinall and Marshek 1999).
depends on the wear type, lubrication condition, and
metallurgical compatibility of contact surfaces. For 1150kt
two surfaces of identical materials g remains V¼ ð1  Ar Þ1=2 ð10Þ
HB
unchanged; p is the surface interface pressure
(MPa); v is the sliding velocity (mm/s) and depends where V is the cutting velocity (ft/min), kt is the
on the gear speed; HSurface is the surface hardness thermal conductivity (Btu/(h ft 8F)), HB is the Brinell
(Bhn). According to Eq. (8), the harder the surface, hardness, and Ar is the area reduction at fracture.
the lower the wear rate, i.e., the higher the wear Therefore, concerning machinability, the material-
resistance. From a material selection point of view, related performance index is
Eq. (8) introduces a fifth material-related
performance index, namely, g6 ¼ Hcore (Bhn) ð11Þ

which needs to be minimized according to Eq. (10).

g5 ¼ HSurface (Bhn) ð9Þ
Note that for case-hardened gears, the gear surface is
which needs to be maximized. For more precise hardened through a heat treatment process after the
calculations, a ratio of surface hardness to elastic gear is cut (Hofmann 1990). Hence the cost of
modulus may be used in Eq. (9) (Juvinall and machining depends on the parent material hardness
Marshek 1999). (Hcore).

123
Gear material selection with uncertain and incomplete data 215

3 Effect of performance indices: a modified performance index g2 versus the corresponding

decision matrix individual material properties are different. Both the
UTS and Sfe criteria prefer material M7 whereas the
The six performance indices, gi; (i = 1,2,...,6) pre- performance index suggests M3 or M2. The reason is
sented in Sect. 2, are applied to the data given in that M7 delivers a larger k exponent (see Eq. (A4)) in
Table 1. In doing so, the upper and lower limit of the Basquin crack initiation curve and thus, based on
each data point is found by assuming the best and Eq. (A5), it is not in the favor of better service life. In
worst scenarios. For example, for the index g4 in Eq. other words, for the same service lifetime, M7 has a
(7), the lower value is found when S0 n is maximum lower Cli. Some other criteria such as g1 and g3 are in
and UTS is minimum. The results are shown in conflict with g2 and they prefer M7 due to its superior
Table 2 (weighting factors in this Table will be cyclic load bearing capacity, etc. Finally, it is worth
referred to later). mentioning that even within the same type of
From a decision-making point of view, Table 2 criterion (a martial property or a performance index),
represents a (modified) decision matrix where alter- using the lower limit, average, or upper limit value
natives are the nine materials and criteria are the six can yield different ranking orders. For instance,
performance indices, whereas in Table 1 (i.e., the according to Fig. 4, the average and lower limit
original matrix) the criteria were the five individual values of g3 suggest that M4 should be preferred over
material properties. It is important to notice that for M3, but the upper limit values of the same index
some data values, e.g., Sfe for material M2, uncer- suggest that they are almost indifferent. In the next
tainty information is not given and, thus, they are section, a multiple criteria pseudo-fuzzy material
considered as incomplete data. selection model is used to accommodate such inter-
Next, to show an example of the effect of and intra- criteria decision conflicts and uncertainties.
performance indices, the ranking of alternatives
based on one performance index (g2) from Table 2
and its individual components (UTS and Sfe) from
Table 1 are compared in Fig. 3. The heights of bars 4 An application of ELECTRE III
are proportional to the alternative ranks: the higher
the bar height, the higher the material rank. As can be Here the goal is to rank the nine material alternatives
seen in Fig. 3, the best ranked materials using the of Table 2 based on the six criteria and their

Table 2 Modified decision matrix with material performance indices and data uncertainties
Weighting factors Material IDs (alternatives)

0.1720 0.0050 0.4260 0.2920 Total 0.102

0.0605 0.0415
Performance indices (criteria)

g5 = HSurface (Bhn) g6 = Hcore (Bhn) g1 = Sfe (MPa) g3 = S0 nCS (MPa) g2 ¼ UTS

Sfe g4 ¼ UTS
S0n
+  + + + +

200 200 330 76 1.15 3.80 M1

220 220 460 268 1.91 2.44 M2
180–300 180–300 480–620 165–340 0.95–2.92 1.34–4.58 M3
220–320 220–320 560–700 283–335 0.84–1.96 1.31–2.62 M4
220–320 220–320 600–740 337–432 1.08–2.63 1.38–3.16 M5
519–565 192–265 1160 485–520 1.59 2.72 M6
601–692 256–337 1500 610–662 1.53 2.50 M7
647–738 256–337 1250 504–542 1.00 1.64 M8
160–210 160–210 450–550 316–346 1.02–1.58 1.27–1.69 M9

123
216 A. S. Milani, A. Shanian

Fig. 3 Ranking of 2.5 2500

alternatives with respect to g2 UTS Sfe
the material performance 2 2000
index g2 and its components

Sfe , UTS (MPa)

Sfe and UTS (all values are
averaged) 1.5 1500

g2
1 1000

0.5 500

0 0
M1 M2 M3 M4 M5 M6 M7 M8 M9
Material ID

Fig. 4 Ranking of 700

alternatives with respect to
600
the lower limit, average, Lower limit Average Upper limit
and upper limit of g3 (notice 500
data incompleteness for
400
(MPa)

materials M1 and M2)

g3

300

200

100

0
M1 M2 M3 M4 M5 M6 M7 M8 M9
Material ID

weighting factors. In Table 2, plus and minus signs it is assumed that data uncertainties are purely due to
for criteria denote their directions (the higher, the random errors and, thus, the expected lower bound of
better, or the lower, the better) based on the errors is zero. If an analyst is aware of any biased
discussions of Sect. 2. The weights are taken from errors during experimental procedures, he/she may
Milani et al. (2005) and are derived based on sets of assign a non-zero indifference threshold according to
recommendations from gear design specialists and an the estimated biased error. For the strict preference
entropy method. To solve the problem, the non- thresholds, the standard deviation of each material
compensatory ELECTRE III methodology is em- performance value is calculated. Then, for the jth
ployed. It is non-compensatory in the sense that a column, the minimum standard deviation of the
very poor performance of a material with respect to a column is assigned as pj. This is to say that one
criterion cannot be offset by its good performance in material should be strictly preferred over another
other criteria. Mathematical considerations of this only if the difference between their performances is
method in the context of material selection are given at least greater than a minimal measurement error.
in Appendix B. Finally, the veto thresholds are chosen to be the same
The nominal value for each material performance as the strict preference thresholds (V j: vj = pj). This is
is found by averaging corresponding upper and lower to say that in the given material selection, no
limits (assuming a uniform uncertainty distribution). fuzziness for the veto condition (described in Appen-
For cases where an upper and lower limit was not dix B) is allowed and it should be a binary condition;
available, the given value was considered as nominal. see also Eq. (B2) for vj = pj. More precisely, if the
The indifference criteria thresholds are assigned to be difference between two material performances with
zero (V j: qj = 0). This is to say that, with respect to a respect to a certain criterion, gj(Mk)  gj (Mi), is
criterion, none of the given materials can be larger than the observed standard deviation, then the
interchanged for the application. Also note that here veto condition is fully activated, i.e., dj(Mi,Mk) = 1,

123
Gear material selection with uncertain and incomplete data 217

and the outranking of Mi ? Mk should be rejected; weights. Hypothetically, when a ? ?, the largest
otherwise there is no veto. weight is one and the rest tend to zero. In the light of
Figure 5 summarizes results of the descending and these conditions, a may be considered as a ‘contrast
ascending distillation procedures (see also Appendix intensity’ parameter (Zeleny 1976).
B) and the ensuing final ranking graph. It is seen in Two a values (2 and 0.5) were initially chosen to
Fig. 5c that a candidate material has a high final rank apply a dilation and concentration sensitivity analy-
if it has a high rank in both ascending and descending sis, respectively, for the current material selection
distillation procedures (Collette and Siarry 2003). problem (see Table 3). It is important to note that
Material #6 seems to be the best choice and material dilation and concentration operators do not change
#7 is considered as a secondary option. In each graph, the order of criteria preferences, providing original
an arrow indicates a dominance relation. Any two weights are normalized between zero and one. The
alternatives that are disconnected (share no arrow) ELECTRE III solution method was then repeated for
are incomparable. For instance, according to Fig. 5c, a = .5 and results are shown in Figure 6. For a = 2
Materials #3 and 4 are incomparable; however they results were found identical to the original case (i.e.,
both have a low rank. Finally, indifferent alternatives a = 1) in Fig. 5. Once Figs. 5 and 6 are compared,
are those that appear on a same node of the graph. materials 6 and 7 prove to have the most stable and
highest ranks in all cases. No change in the latter
conclusion could be made for the larger range of a
4.1 Uncertainty in weighting factors: a dilation [ (0,4) checked.
and concentration process

M6 M6
Subjective weighting factors obtained from different
designers’ opinions may have uncertainties of their M6 M7
M7 M7
own, particularly if the number of designers con-
M9
sulted with is not large enough. Therefore, it is M8 M9
worthwhile to make sure that the obtained solutions M9 M8
M8
are robust against potential errors in weight values.
M5
To simulate perturbation of original weights in M3
M5 M5
Table 2, a method of dilation and concentration is M4
adapted from Zeleny (1976). In this method, the M2 M3 M4
original set (w1 ,w2 ,...,wn) is iteratively changed by M3
two operations as follows. M1 M2
M2 M4
Dilation: ðw1 ; w2 ; . . . ; wn Þ is replaced by M1
M1
ðw1 ; w2 ; . . . ; wn Þa ; a[1
(a) (b) (c)
And,
Fig. 5 Results of (a) descending distillation, (b) ascending
distillation, and (c) final ranking
Concentration: ðw1 ; w2 ; . . . ; wn Þ is replaced by
ðw1 ; w2 ; . . . ; wn Þa ; 0<a<1
Table 3 Original, dilated, and concentrated weighting factors
Observe that the concentration operation leads to the
Case a Weighting factors set
increasing equalization of weights; they are hypo-
thetically the same when a ? 0. The condition a = 0 Dilation 2 (0.0295, 0.000025, 0.1814, 0.0852,
is only meaningful when there is no sufficient 0.0036, 0.0017)
evidence to prioritize one criterion against another. Original 1 (0.1720, 0.0050, 0.4260, 0.2920,
On the other hand, the dilation operation lessens the 0.0605, 0.0415)
larger values less and the smaller values more and, Concentration 0.5 (0.4147, 0.0707, 0.6526, 0.5403,
0.2459, 0.2037)
therefore, it provides increased differences among

123
218 A. S. Milani, A. Shanian

Single criterion decision-making tests are performed

M6 M6 M6
M7 using both the performance indices and the individual
M9 material properties. The results reveal that (a) using
M7 M7
the performance indices vs. individual material
M8
M5 M9 properties as decision criteria can yield notably
M9
M5 different solutions, (b) different criteria can yield
M5 M8 different preferred materials, and (c) the lower and
M8
M3 upper limits within the same criterion occasionally
M4 M3
M4 introduce indecisions in the preferred solution.
M2
M2 Therefore, for an effective gear material selection, a
M3
M1
multi-criteria decision model may be used where all
M1 material performance indices and their uncertainties
M2
M4 are accounted for simultaneously.
M1 Based on the results of a chosen sample gear
(a) (b) (c) material selection problem with data uncertainties
and incompleteness, it seems that the ELECTRE III
Fig. 6 Results of (a) descending distillation, (b) ascending multi-criteria decision model is capable of ranking
distillation, and (c) final ranking, when a concentration candidate materials from the best to the worst while
operation with the contrast intensity of 0.5 on original weighs estimating their incomparabilities and/or indifferenc-
is applied
es. The use of fuzzy outranking relations and pseudo
criteria in ELECTRE III can particularly be useful to
account for material data uncertainties. Finally, it is
5 Conclusions and recommendations
shown that a dilation and concentration process can
be used to imitate potential uncertainties in defining
Material selection in multi-objective gear design
weighting factors (e.g., due to hesitation in designers
problems may be uncoupled from geometry optimi-
minds) by varying a contrast intensity parameter, a. A
zation by applying a discrete decision model and a
reliable material selection should encourage candi-
continuous optimization of geometrical parameters.
dates with both the highest and the most stable ranks
For the material selection part, a set of performance
against all sources of uncertainties. The top ranked
indices may be adapted given a particular application.
candidates should not be deterred with a reasonable
The advantage of these indices is in combining
change of a. Further study for more sophisticated
individual material properties into groups to mimic
modeling of random and biased errors due to non-
specific goals of that particular application. In this
standard materials testing or designers’ subjective
study, five material performance indices are devel-
judgments, and also study of robustness of the
oped from major gear failure criteria and the corre-
solution method against potential decision abnormal-
sponding crack initiation slope in Basquin type S-N
ities may be worthwhile.
curves. A secondary material property, machinability,
Other potential extensions. Although this study
is also considered as a sixth criterion. The six indices
was confined to surface fatigue, bending fatigue, and
can be summarized as follows.
abrasion failure modes, there are other modes that
• Surface fatigue limit index: g1 = Sfe (MPa) may be worth considering. A relatively new gear
• Surface fatigue lifetime index: g2 ¼ UTS
Sfe failure mode is the Tooth Interior Fatigue Fracture
• Bending fatigue limit index (for a machined (TIFF) (MackAldener and Olsson 2003). The failure
finish): g3 = S0 n (0.0007HB + 0.8987) (MPa) mechanism in this mode is fundamentally different
• Bending fatigue lifetime index: g4 ¼ UTSS0n from the bending and surface fatigue modes. In TIFF,
• Abrasive wear index: g5 = HSurface (Bhn) a crack can initiate and propagate approximately
• Machinibility index: g6 = Hcore (Bhn) from mid-height of the tooth and slightly below the

123
Gear material selection with uncertain and incomplete data 219

case–core boundary and it is more critical in idler
k

ri
gears than in single stage gears (MackAldener and Ni ¼ ND ðA2Þ
rD
Olsson2003). Another well-known failure mode is
tooth plastic deformation where a sudden heavy static where k is a material constant (the smaller the k, the
or cyclic load causes material yielding and subse- longer the crack initiation phase). ND and rD
quently permanent deformation (Akinci et al. 2005). correspond to the number of cycles and the stress
However, this type of failure is seen to be one of the level at the endurance limit point. Equation (A2)
least encountered modes in practice (Akinci et al. applies to both the surface and bending fatigue
2005), partially because a work-hardened skin tends modes. For simplicity, the subsequent derivations are
to develop during the plastic deformation (Neale followed for the surface fatigue mode; for the
1995). The consideration of material fracture tough- bending fatigue mode, a similar derivation applies
ness, elastic moduli, electric and thermal properties by replacing the corresponding parameters.
are among other potential extensions to this study. The Basquin type crack initiation curve passes
through two material points: (Nfe, Sfe) and (NUTS,
Acknowledgments The first author would like to
acknowledge the financial support of the Natural Sciences UTS); Nfe is the nominal number of cycles corre-
and Engineering Research Council of Canada. sponding to Sfe (normally 107cycles (Juvinall and
Marshek 1999)); NUTS is the number of cycles at
which the tooth fails under a cyclic load with an
Appendix A amplitude equal to UTS. Using the first material point
in Eq. (A2) yields
Estimation of Cli dependence on material and
k
number of cycles ri
Ni ¼ Nfe ; Ni  Nfe ðA3Þ
Sfe
An ultimate gear failure in service is announced
when (1) one or more teeth have completely broken And, using the second point in Eq. (A3) gives
away, or (2) the gear unit has been damaged so  
badly such that the vibration and noise levels are not log NNUTS
fe

acceptable anymore (Neale 1995). Once a crack on k¼  ; k[1 ðA4Þ

the root or surface of a tooth is initiated, a gear may log UTS
Sfe
still continue working for a few more cycles until
the final breakage occurs. For a given stress Finally, assuming Ni & N (total number of cycles)
magnitude and material, the total number of cycles and recalling the definition of Cli as the service
before final bending fatigue (Jelaska et al. 2003) or lifetime factor that adjusts the allowable stress, we
surface fatigue (Choi and Liu 2006) failure occurs is can write ri & Cli Sfe. Comparing with Eq. (A3) it
defined: follows

N ¼ Ni þ Np ðA1Þ
1=k
Ni
Cli  ; Cli  1 ðA5Þ
where Ni and Np are the number of cycles required for Nfe
the crack initiation phase and the crack propagation
phase, respectively. Most of the tooth fatigue life is If test data for an individual Sfe and UTS value are
dedicated to the first phase (Jelaska et al. 2003; Choi not available, a linear approximation is often made as
and Liu 2006) (i.e., Ni >> Np), and therefore it is Sfe = 0.5UTS (Juvinall and Marshek 1999). In that
important to choose materials with higher resistance case, k in Eq. (A4) reduces to a constant and becomes
to crack initiation. Neglecting plastic strains in the independent of material. Subsequently, Eq. (A5)
high cycle fatigue (HCF) region, and assuming the takes the form Cli = a(Ni)b where a, b are two
Basquin type S-N law for the crack initiation stage, Ni constants (see Wilson 1997 for a similar form for
for a given stress magnitude of ri is estimated by steel gears). In this study, however, the life factor Cli
Jelaska et al. (2003), Choi and Liu (2006) is allowed to vary with the individual material

123
220 A. S. Milani, A. Shanian

Log The goal is to select the best alternative(s) given

σi
Cli = the performance values of each alternative with
S fe
respect to each criterion (normally given as an m ·
UTS 2 n decision matrix) and the corresponding weights. In
S fe2
order to account for possible scatter (uncertainties) in
UTS1 the decision matrix values, a set of threshold values
S fe1 should be added to the above nominal list. Thus, the
actual problem furthermore consists of
−1/ k 2
• n strict preference thresholds, pj; j = 1,...,n,
− 1/ k1 reflecting the upper limit of uncertainty in each
1
Log
criterion value (e.g., due to random errors, etc.);
NUTS
1
Ni • n indifference preference thresholds, qj; j = 1,...,n,
N fe N fe reflecting the lower limit of uncertainty in each
criterion value;
Fig. A1 Schematic of material and cycle dependence of a life
• n veto thresholds, vj; j = 1,...,n, each of which
factor assuming Basquin type S-N curves (1 and 2 subscripts
refer to two different designs) reflecting a maximum allowable limit for a veto
condition (explained later).
Fro each criterion, qj  pj  vj. To solve the actual
parameters Sfe and UTS via the general definition of k
problem using ELECTRE III (Roy 1993, 1990;
in Eq. (A4). This is because in the proposed material
Vallée and Zielniewicz 1994; Rogers et al. 2000;
selection problem, alternatives are of different UTS Sfe Collette and Siarry 2003), a pseudo-criterion defini-
values (Table 2). For a given service lifetime Ni, the
tion instead of a classical definition (fixed values)
higher the UTS
Sfe ration, the higher the service life factor should be used using the indifference and strict
Cli (Fig. A1), and the higher the effective (modified)
preference thresholds. To understand this more
endurance limit. Also from Fig. A1, for a given
closely, let us consider a simple single-criterion
material, the lower the service lifetime required, the
example in Fig. B1. Each material has one indiffer-
higher the Cli chosen.
ence zone (here a band) with a range of (q, q)
through which no preference relation may be made.
Remark The derived ration UTS Sfe may also be used in There also exists a strict preference zone, (p, p),
a different way: given an Srfei level in Eq. (A3), a
beyond which a strong preference can be marked.
higher UTS
Sfe refers to a longer crack initiation phase Now based on the difference in the positions of two
(see also Fig. A1 for a constant horizontal line). The
given alternatives such as Mi, Mk on the criterion
quantity Srfei  1 may be interpreted as the excessive
axes, one can infer one of three cases as follows. One
stress percentage, e.g., due to an unexpected over-
is considered indifferent to another, there is a weak
load, etc.
preference of one to another, or there exists a strong
preference of one over another. Just to generalize, for
biobjetive problems each zone on the criteria space
Appendix B
becomes a rectangle, for 3D problems a cuboid, and
eventually an n-dimensional hyper cuboid. Note that
ELECTRE III
with respect to a given criterion, threshold values can
change with alternatives, or for simplicity they may
The typical decision making problem considered
be assumed to be independent of alternatives (e.g., in
consists of
Fig. B1, p = p0 ; q = q0 ).
• m alternatives (material candidates) Mi; i = 1,...,m; Following the above criteria definition, for each
• n design criteria (material performance indices)gj; pair of materials such as (Mi, Mk), a fuzzy outranking
j = 1,...,n; relation can be built in ELECTRE III methodology.
• n weighting factors (designer’s preference infor- This is done by introducing a credibility degree dik
mation) xj; j = 1,...,n. which represents the credibility of the outranking

123
Gear material selection with uncertain and incomplete data 221

Fig. B1 Schematic of
preference relations via a
pseudo-criteria definition
(cases shown are for gj (Mk)
 gj (Mi) > 0 and gj is to be
maximized)

relation of Mi ? Mk (says that material Mi outranks For an ideal alternative, M*i , one expects cj(M*i ,
material Mk). In turn, the credibility degree is defined Mk) = 1 and dj(M*i ,Mk) = 0 for all k; k = i. Following
by using the concordance cj (Mi, Mk) and discordance the solution process, the cj(j = 1,2,...,n) indices are
dj (Mi, Mk) indices with respect to the performance aggregated in a global concordance index as:
index gj. The concordance index allows the solution
Pn
mechanism to verify if Mi is at least as good as Mk j¼1 xj cj ðMi ; Mk Þ
with respect to the jth design criterion. On the other Cik ¼ Pn ðB3Þ
j¼1 xj
hand, the discordance index verifies if there exists a
very high opposition (veto) to the outranking relation
Finally, the credibility degree for the pair of
of Mi ? Mk. Thus, materials with very weak
materials is calculated by:
performances in one or more design criteria are
always in danger of being vetoed against their Y 1  dj ðMi ; Mk Þ
outranking relations. The concordance index in- dik ¼ Cik : ðB4Þ
1  Cik
creases linearly from its minimal value when gj j2F

(Mk)  gj (Mi) passes the strict preference threshold 

pj and reaches to its maximum value when gj where F is defined as F ¼ jj 2 F; dj ðMi ; Mk Þ[Cik g
(Mk)  gj (Mi) is within the indifference zone and F  F  f1; 2; . . .; ng. It is worth pointing out
identified by qj. To summarize, that in Eq. (B4) when at least one dj(Mi, Mk) = 1,
dkl = 0. The latter clearly constitutes the non-
compensatory aspect of the ELECTRE approach via
8
< 0 ;pj <gj ðMk Þ gj ðMi Þ the veto condition. It says, if material Mi as compared
gðMi Þþpj gðMk Þ to Mk is very poor in only one particular design
cj ðMi ;Mk Þ ¼ pj qj ;qj <gj ðMk Þ  gj ðMi Þ  pj
: criterion (i.e., vj < gj(Mk)  gj(Mi)), it will not be
1 ;gj ðMk Þ gj ðMi Þ  qj
preferred over Mk, regardless of its performance in
ðB1Þ
other criteria. The measure of such a veto condition is
the threshold value, vj, and should be pre-defined by a
Note that the presented relation uses the conven-
careful analyst.
tion that gj needs to be maximized. Similarly, the After calculating all credibility degrees for each
discordance index increases linearly from its minimal
and every pair of materials with respect to all criteria,
value when gj(Mk)  gj(Mi) is below the strict
fuzzy outranking relations (Roy 1993) can be estab-
preference threshold pj and reaches to its maximum lished, this time between pairs of credibility degrees.
value when gj (Mk)  gj(Mi) outreaches the veto
The final classification of alternatives is then based
threshold vj:
on building two complete preorders such as Z1 and Z2
through descending and ascending distillation proce-
8 dures. Recall that a preorder is a relation (here the
< 0 ;gj ðMk Þ gj ðMi Þ<pj
gðMk Þpj gðMi Þ material dominance) that is reflexive and transitive
dj ðMi ;Mk Þ ¼ vj pj ;pj  gj ðMk Þ gj ðMi Þ  vj
: but not anti-symmetric (i.e., indifferent solutions are
1 ;vj <gj ðMk Þ  gj ðMi Þ
allowed). A preorder is complete when no incompa-
ðB2Þ rability between two solutions is allowed (Collette

123
222 A. S. Milani, A. Shanian

and Siarry 2003). The final preorder Z is built by Horimoto, M., Matsumoto, H., Makino, T., Murai, N., Orita,
intersecting the two complete preorders, Z = Z1 \ Z2. K., Arimi, Y., Fujikawa, S., Nishino, T.: Effect of core
hardness and case depth on low-cycle-impact-fatigue
Normally, descending and ascending distillations do property in carburized steel. J. Soc. Mater. Sci. Jpn.
not result in the same ranking of alternatives and thus 52(11), 1318–1324 (2003)
the final preorder is partial (incompatibilities are Jelaska, D., Glodež, S., Podrug, S.: Numerical modeling of the
allowed). Conversely speaking, two candidate mate- crack propagation path at gear tooth root. In Proceedings
of the ASME/AGMA International Power Transmission
rials may be incomparable in the final decision if their and Gearing Conference, Chicago, ASME Paper
ranks in the descending and ascending procedures are DETC2003/PTG-48026 (2003)
different. Indifferent materials are marked if their Juvinall, R.C., Marshek, K.M.: Fundamentals of Machine
ranks in both procedures are identical (see, e.g., Component Design, Wiley, New York (1999)
MackAldener, M., Olsson, M.: Tooth Interior Fatigue Fracture-
Rogers et al. 2000; Collette and Siarry 2003 for more computational and material aspects. Int. J. Fatigue 23,
details). The challenge is greater when the total 329–340 (2003)
number of votes and vetoes for two or more Matos, M.J., Simplı́cio, M.H.: Innovation and sustainability in
alternatives are similar. mechanical design through materials selection. Mater.
Des. 27(1), 74–78 (2006)
Milani, A.S., Shanian, A., Madoliat, R., Nemes, J.A.: The ef-
fect of normalization norms in multiple attribute decision
References making models: a case study in gear material selection.
Struct. Multidisciplinary Optim. 29, 312–318 (2005)
Akinci, I., Yilmaz, D., Canakci, M.: Failure of a rotary tiller Neale, M.J.: Component Failures, Maintenance and Repair—A
spur gear. Eng. Fail. Anal. 12, 400–404 (2005) Tribology Handbook. Butterworth Heinemann, Boston
Ashby, M.F.: Multi-objective optimization in material design (1995)
and selection. Acta Mater. 48(1), 359–369 (2000) Ognjanovic, M.: Decisions in gear train transmission design.
Ashby, M.F.: Materials Selection in Mechanical Design. Res. Eng. Des. 8, 178–187 (1996)
Elsevier, Boston (2005) Pratyyush, S., Jian-Bo, Y.: Multiple Criteria Decision Support
Ashby, M.F., Bréchet, Y.J.M., Cebon, D., Salvo, L.: Selection in Engineering Design. Springer-Verlag, Berlin (1998)
strategies for materials and processes. Mater. Des. 25(1), Rao, R.V.: A material selection model using graph theory and
51–67 (2004) matrix approach. Mater. Sci. Eng. A (431), 248–255
Bayoumi, M.R., Abdellatif, A.K.: Effect of surface finish on (2006)
fatigue strength. Eng. Fract. Mech. 51(5), 861–870 (1995) Rogers, M., Bruen, M., Maystre, L.: Electre and Decision
Brechet, Y., Bassetti, D., Landru, D., Salvo, L.: Challenges in Support. Kluwer Academic Publishers, London (2000)
materials and process selection. Prog. Mater. Sci. 46(3–4), Roy, B.: Decision aid and decision making. In: Bana e Costa,
407–428 (2001) C.A. (ed.) Readings in Multiple Criteria Decision Aid.
Choi, Y., Liu, C.R.: Rolling contact fatigue life of finish hard Springer-Verlag, Heidelberg (1990)
machined surfaces. Part 1. Model development. Wear Roy, B.: Aide multicritère à la décision, Méthodes et Cas.
261(5–6), 492–499 (2006) Economica, Paris (1993)
Collette, Y., Siarry, P.: Multiobjective Optimization. Springer- Shanian, A., Savadogo, O.: A material selection model based
Verlag, New York (2003) on the concept of multiple attribute decision making.
Dehghan-Manshadi, B., Mahmudi, H., Abedian, A., Mahmudi, Mater. Des. 27(4), 329–337 (2006)
R.: A novel method for materials selection in mechanical Suárez, J.C., Diez de Ulzúrrun, I., Biezma, M.V., Ruiz Román,
design: Combination of non-linear normalization and a J.M., Martı́nez, M.A., del Real, J.C., López, F.: Case
modified digital logic method. Mater. Des. 28(1), 8–15 studies in adhesives selection. J. Mater. Process. Technol.
(2007) 143, 219–224 (2003)
Esawi, A.M.K., Ashby, M.F.: Cost estimates to guide pre- Thompson, D.F., Gupta, S., Shukla, A.: Tradeoff analysis in
selection of processes. Mater. Des. 24(8), 605–616 (2003) minimum volume design of multi-stage spur gear reduc-
Esawi, A.M.K., Ashby, M.F.: Computer-based selection of tion units. Mech Mach Theory 35, 609–627 (2000)
joining processes: methods, software and case studies. Vallée, D., Zielniewicz, P. : ELECTRE III-IV, version 3.x –
Mater. Des. 25(7), 555–564 (2004) Aspects méthodologiques, Université de Paris-Dauphine,
Fernandes, P.J.L., McDuling, C.: Surface contact fatigue fail- Document du LAMSADE no 85, 1994
ures in gears. Eng. Failure Anal. 4(2), 99–107 (1997) Wilson, C.E.: Computer Integrated Machine Design. Prentice-
Guisbiers et al. (in press). Guisbiers, G., Overschelde, O.V., Hall, London (1997)
Wautelet, M.: Materials selection for thin films for radio Zeleny, M.: The theory of the displaced ideal. In: Zeleny, M.
frequency microelectromechanical systems. Mater. Des. (ed.) Lecture Notes in Economics and Mathematical
Hofmann, D.: Design and stress analysis of spur and helical Systems. Springer-Verlag, Heidelberg (1976)
gears, Teaching Pack on Gear Technology. British Gear
Association, UK (1990)

123