Available online at www.sciencedirect.

com

Acta Materialia 56 (2008) 10941105

www.elsevier.com/locate/actamat

Revisiting Hume-Rotherys Rules with articial neural networks

Y.M. Zhang, S. Yang, J.R.G. Evans *
Department of Materials, Queen Mary, University of London, Mile End Road, London E1 4NS, UK

Received 22 September 2007; accepted 3 October 2007

Available online 26 December 2007

Abstract

Hume-Rotherys breadth of knowledge combined with a quest for generality gave him insights into the reasons for solubility in metal-
lic systems that have become known as Hume-Rotherys Rules. Presented with solubility details from similar sets of constitutional dia-
grams, can one expect articial neural networks (ANN), which are blind to the underlying metals physics, to reveal similar or better
correlations? The aim is to test whether it is feasible to predict solid solubility limits using ANN with the parameters that Hume-Rothery
identied. The results indicate that the correlations expected by Hume-Rotherys Rules work best for a certain range of copper or silver
alloy systems. The ANN can predict a value for solubility, which is a renement on the original qualitative duties of Hume-Rotherys
Rules. The best combination of input parameters can also be evaluated by ANN.
2007 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

Keywords: Hume-Rotherys Rules; Articial neural networks; Solubility limit of metals; Backpropagation networks; Binary alloys

1. Introduction
Materials science seeks to understand the causative rela-
tionships between composition, processing, structure and
properties at a level that allows composition and processing
parameters to be selected to provide targeted properties.
Such relationships can be discerned by experiment and, in
a few instances, by predictive theory. They can sometimes
be obtained by molecular modelling. All molecular model-
ling techniques can be classied under three general catego-
ries: (1) ab initio electronic structure calculations, which are
based upon quantum mechanics; (2) semi-empirical meth-
ods, which are also founded upon quantum mechanics,
but which enhance computational speed by using approxi-
mations based upon experiment; (3) molecular mechanics,
an empirical method based on classical physics which is
computationally fast [1].
Another approach is to use correlation methods made
possible by articial neural networks (ANN), which are
nding growing acceptance in many subjects for modelling
*
Corresponding author. Tel.: +44 (0)20 76794689.
E-mail address: j.r.g.evans@qmul.ac.uk (J.R.G. Evans).
complex problems [2]. They are model-free in the sense that
they can process complex inputoutput relationships with-
out an explicit mathematical model [3] and are becoming
popular in materials science in solving problems that are
not suitable for traditional statistical methods [4]. They
can process large amounts of information and mimic bio-
logical systems in learning ability and capability to general-
ize. They can handle non-linearity, imprecise and fuzzy
information and are fault and failure tolerant [5]. Impor-
tantly, they oer the materials scientist compositionally
predictive power in which conventional theory is some-
times lacking because of theoretical complexity. A good
example is their use in predicting dielectric constants from
composition [6].
Various networks have been devised but backpropaga-
tion networks in which the data are forward-fed into the
network without feedback and without same-layer neural
connections are the most widely used [2,7]. The model is
shown schematically in Fig. 1. In such articial systems,
learning is a process of updating an internal representation
of an external system. During learning, the magnitude of
the weightings or synapse strengths is adjusted repetitively
as the network is presented with training data.

1359-6454/$30.00 2007 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.actamat.2007.10.059
 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
Fig. 1. A model of a feed-forward hierarchical articial neural network.
In Arkadan et al.s work [8], the location and shape of
a crack were deduced from measured magnetic eld val-
ues as inputs. Raj et al. [9] used ANN in metalworking
to predict forging load in hot upsetting, cutting forces
in machining and loads in hot extrusion. Guessasma
and Coddet [10] used an ANN to quantify the relation-
ship between automated plasma spraying process param-
eters and microstructural features of aluminiumtitanium
coatings.
Huang et al. [11] predicted the mechanical properties of
a ceramic tool based on materials properties. Malinov and
Sha [12] used ANN for correlation between processing
parameters and properties in titanium alloys, such as fati-
gue life and corrosion resistance. They also have been used
in ceramic casting [13], to interpret ultrasonic non-destruc-
tive testing (NDT) of adhesive joints [14], for modelling the
cold rolling forces [15], to predict continuous-cooling trans-
formation curves in steel from chemical composition [16]
and to predict timetemperature transformation diagrams
for titanium alloys [17].
The potential uses for prediction of properties of mat-
ter are interesting for materials scientists. Homer et al.
[18] used physical properties such as molecular weight,
number of bonds and temperature as input factors to pre-
dict the viscosity, density, heat of vaporization, boiling
point and acentric factors for pure, organic, liquid hydro-
carbons over a wide range of temperatures (Treduced 
0.450.7). A three-layer backpropagation network ANN
was applied to the formulation of BaTiO3-based dielec-
trics and for analysis of the electrical properties of PZT
[19,20].
Since ANN can sometimes provide signicant improve-
ments over linear regression analysis, there is increasing
interest in their applicability to materials science where
the interrelationships are not well understood and amena-
ble to analytical solutions [21]. Such is the case in a survey
of terminal solid solutions in several alloy systems in which
Hume-Rothery devised simple and useful rules on solubil-
ity limits which have become known as Hume-Rotherys
Rules:
(1) The atomic size factor [22]: If the atomic diameters of
the solute and solvent dier by more than 14%, the
solubility is likely to be restricted because the lattice
distortion is too great for substitutional solubility.
A criterion of 15%, which was taken by later writers,
is also tested in this work.
1095
(2) The electrochemical factor [22]: or eect of the peri-
ods of the solvent and solute atoms was quantied
by Cooke and Hume-Rothery [23] after the electro-
chemical factor introduced by Pauling [24]. Strongly
electropositive components are more likely to form
compounds with electronegative components than
to form solid solutions, particularly in the later B
sub-groups.
(3) The relative valency factor [22]: Other factors being
equal, a lower valence metal is more likely to dissolve
one of higher valency than vice versa, i.e., the ten-
dency for two metals to form solid solutions is not
necessarily reciprocal. This has been found to be valid
mainly for alloys of copper, silver or gold combined
with metals of higher valency.
Traditionally, phase diagrams were determined by
(often tedious) experimentation, which nevertheless suf-
fered from the diculty of reaching equilibrium at low
temperatures (typically <0.5Tm) in reasonable timescales
[25]. Hume-Rotherys quest for generality was thus well
founded but has, to an extent, been superseded by compu-
tational calculation of phase diagrams following pioneering
work by Kaufman and Bernstein [26] and Hillert [27] and
others in the 1970s giving rise to such tools as CALPHAD,
which has been widely used, e.g., Refs. [2833].
After the signicant work done by Hume-Rothery
et al. on the prediction of solid solubility in alloys, Dar-
ken and Gurry [34], Chelikowsky [35], Alonso and Simo-
zar [36], Alonso et al. [37] and Zhang and Liao [38] all
contributed in dierent ways to the prediction of solid sol-
ubility in terms of a soluble/insoluble criterion.
The authors aim is to stand at the place where Hume-
Rothery stood, with the added advantage of the ANN
and, based on a large number of constitutional diagrams
and physical parameters of metals provided by handbooks
[39,40], to simulate the process that Hume-Rothery used to
derive his rules. Hume-Rothery considered 60 systems, and
so one faces the problem of limited data. Nevertheless,
there are many situations where materials scientists would
like to benet from ANN in situations where the data set
has inherent limitations. It has been proved that ANN
methods provide an ecient tool for experimental data
analysis even when the database size is small [41]. Further-
more, this study attempts to predict the solubility quantita-
tively rather than produce a classication of soluble/
insoluble. Thereafter, it is attempted to extend the method
to a wider range of silver and copper alloys based on binary
systems and including dierent structures.
One interesting question is whether the parameters
Hume-Rothery used are sucient to determine the solu-
bility. Hume-Rothery amended the rules by introducing
the structural parameter 14 years after the rst presenta-
tion of the rules [42]. In this paper, ANN have been used
to determine the relative importance by deliberately omit-
ting one or two parameters. Another aim is to explore the
eects of the format of the input parameters. The combi-
1096 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
nations of these dierent parameters are tested in this
paper.
2. Types of ANN
There are some more frequently used ANN from which
to select. They have their own characteristics and special
applications [2]:
1. Backpropagation articial neural networks (BPANN).
This type of network is versatile and can be used in
many elds such as data modelling, classication, fore-
casting, control, data and image compression and pat-
tern recognition [43].
2. Hopeld networks. These are two-layer recurrent net-
works, which are ideal for solving optimization prob-
lems [44,45].
3. Adaptive resonance theory (ART) networks. These are
made up of a layer that receives inputs and a layer that
consists of output neurons and can, like Hopeld net-
works, be used for pattern recognition, completion
and classication [2].
4. Kohonen networks. These are two-layer networks, which
convert multidimensional input patterns into lower-
order data [46]. In addition to pattern recognition and
classication, Kohonen networks also can be used for
data compression, i.e., high-dimensional data can be
mapped into lower-dimensional space without losing
content [47].
5. Counterpropagation networks. These networks, which
are developed by Hecht-Nielsen [48,49], are trained by
supervised and unsupervised learning to create a self-
organizing database and can be used for function
approximation and classication [50].
6. Radial basis function (RBF) networks. These are a type
of a multi-layer feed-forward error-backpropagation
networks with three-layers [51]. The selection of RBF
or BPANN depends on the type of problem [52]. RBF
networks train faster than BPANN but are less versatile
and are somewhat slower [53].
The reason for selecting the proposed ANN is now
explained. As Basheer and Hajmeer [2] mentioned, the
decision depends strictly on the problem logistics. The
Kohonen network is required by a clustering problem,
BP or RBF networks can model mapping problems, but
Hopeld networks can only solve some optimization prob-
lems. ANN selection also depends on the type of input
(Boolean, continuous or a mixture of these) and the speed
of the network once it is trained. The initial problem, which
simulates the process that Hume-Rothery used to derive his
rules has soluble/insoluble as output and is a type of clas-
sication problem, so a probabilistic neural network [54,55]
is designed for use. This is a type of radial basis network
suitable for classication problems. In subsequent work,
the problems are all mapping problems, so BPANN are
used.
3. Data collection
The silver and copper alloy solid solubility limits (at.%)
are recorded from Massalski et al. [39], Moatt [56] and
ASM Handbook, vol. 3, Alloy Phase Diagrams [57]. The
physical parameters, radii, valences and electrochemical
factors (electronegativity) of solvent and solute atoms are
taken from Stark and Wallace [40] and from Aylward
and Findlay [58]. The valences of elements, which were
mentioned by Hume-Rothery in 1934, follow his represen-
tation. Radii of Al, Ga, and a-Fe also followed Hume-
Rotherys representations [59]. The structure parameter is
taken from Ref. [57].
The whole data set is used in two ways: (1) the 60-alloy
systems, which were rst mentioned by Hume-Rothery in
1934, are used for training the neural network and testing
whether the Hume-Rotherys Rules work in this range of
alloy systems; (2) all the 408 silver and copper alloy systems
collected are used to repeat the process.
4. Conguration of the neural network
The ANN are constructed, trained and simulated by
MATLAB software, run under the Microsoft Windows
XP operating system on ThinkPad (IBM ThinkPad T40
Model 2373-12H: Intel Pentium M1.3 G, 256 MB, 30
GB, 14.1 in. TFT), Java VM Version: Java 1.4.2 with
Sun Microsystems Inc. Java HotSpotTM Client VM.
4.1. Number of neural network layers
A two-hidden-layer sigmoid/linear network can repre-
sent any functional relationship between inputs and
outputs if the sigmoid layer has enough neurons [60]. A
two-hidden-layer network, with a tan-sigmoid transfer
function in the rst hidden layer and a linear transfer func-
tion in the second hidden layer, is thus adopted. The num-
ber of neurons in the second hidden layer is constrained by
the number of outputs required by the problem. The out-
put in this work is solubility, so there is one neuron in
the second hidden layer.
4.2. Number of neurons in the rst hidden layer
The choice of number of neurons in the rst hidden
layer is up to the designer. The optimum number of neu-
rons in the rst hidden layers may be a function of (1)
input/output vector size, (2) size of training and testing
sub-sets and, more importantly, (3) the problem of non-lin-
earity [2]. The optimum number is found by trial and error
by placing a dierent number of neurons in the rst hidden
layer for the same data set.
4.3. Improving generalization
Overtting occurs during network training if the error
associated with the training set becomes low, but the error
 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
becomes large when new data are presented to the network.
This means that the network has learned the training
examples but is unable to generalize to new situations.
There are two common methods for improving generaliza-
tion: Bayesian Regularization and Early Stopping [60].
Bayesian Regularization tends to provide better generaliza-
tion performance than Early Stopping in training function
approximation networks. As a result, Bayesian Regulariza-
tion is used for improving generalization. This involves
setting the sum of squares of the network errors on the
training set to give the best generalization.
4.4. Partitioning of the database
The generalizing ability of the network depends on the
training database size [2]. Although ANN can be obtained
from a training database of any size, like other empirical
models, generalization of these models outside the model
domain is adversely aected. Since ANN are required to
generalize for the unseen data, data used for training
should be large enough to cover the possible known varia-
tion in the problem domain.
The development of an ANN based on Bayesian Regu-
larization requires partitioning of the parent database into
two sub-sets: training and testing. Currently, there are no
denitive rules for determining the required sizes of the
data sub-sets. Rules of thumb derived from experience
and analogy between ANN and statistical regression exist
[2]. Following the method suggested by Matlab [60], the
sets are picked as equally spaced points throughout the
original data. Ratios between 2:1 and 5:1 are tested and,
based on regression coecient for the testing set, the ratio
4:1 is selected, i.e., partitioning the whole data set into ve
groups, four groups being used for training, while one
group is used for testing. The size of the training set
and testing set are thus determined, but the choice of test-
ing set still plays a crucial role, because the training set
should include all the data belonging to the problem
domain. In this work, the problem domain is not clear,
so, referring to Malinov and Shas work [4], a loop pro-
gram was used to redistribute the database in order to
make the training set cover the problem domain. The dis-
tribution was selected on the basis of regression coecient
R (R = 1 corresponds to perfect correlation), for the test-
ing set. However, where M, the slope of the linear regres-
sion line, is smaller than 0.9, the regression coecient
1097
uration of the neurons of hidden layers, which would
impede the learning process [2]. There are two functions
for scaling the inputs and targets of networks that have
been implemented in the Matlab Neural Network Toolbox:
PREMNMX, which is used to scale inputs and targets so
that they fall in the range [1, 1]; and PRESTD, which
normalizes the inputs and targets so that they will have
zero mean and unity standard deviation. As the transfer
functions employed here are tan-sigmoid transfer function
and linear function, PREMNMX is adopted.
Because the size of the database in this work is small,
it is crucial to make the training set cover the problem
boundary. As in Malinov and Shas work [4], a looped
program is used in order to nd the best combination
of database distribution and number of neurons in the
rst hidden layer. The criteria used to nd the best com-
bination are discussed below.
5. Determination of input parameters
The network input parameters are the physical param-
eters including (1) atomic size parameter, (2) valence
parameter, (3) electrochemical parameter, i.e., electroneg-
ativity, and (4) structure parameter of solvent and solute
atoms, which were not mentioned in 1934, but were
introduced in 1948 [42] concerning the detailed examina-
tion of Vegards law [61] in the case of metallic solid
solutions.
Three dierent expressions of these parameters are used:
1. The raw data that Hume-Rothery used. Details are dis-
cussed below.
2. The original collected values for each parameter of sol-
vent and solute atoms.
3. The original collected parameters are converted into
functionalized values before putting them into the
networks:
(a) For the size factor. The dierence between the
atomic diameters of solvent and solute atoms
divided by the diameter of the solvent atoms is
used.
(b) For the valence factor. These are integers and the
original values are used, leaving the neural network
to decide the relations between valence of solvents
and solutes.
(c) For the electrochemical factor. The dierence
provides an unreliable criterion, and  so the selection was between that of the solvent and solute atoms is
 B  used.
based on u jM  1j 1  R Bmax , where B was the
(d) For the structure parameter. The expressions of the
intercept on the A-axis, and Bmax is the maximum solubil-
ity. The ideal value of this parameter is zero.
4.5. Data normalization
The data should be normalized within a uniform range
(e.g., [0, 1] or [1, 1]) in order to prevent larger numbers
from overriding smaller ones and to prevent premature sat-
structures can be put in terms of numbers 114 for
the Bravais lattices, but this revealed little eect of
the structures. They can also be expressed in three
sets of numbers representing primitive cell dimen-
sions, angles and systems. This allows some
similarities to be explored. The three sets are (1) unit
cell length (a = b = c; a = b 6 c; a 6 b 6 c), (2) axes
angles (a = b = c = 90; a = b = 90, c = 120;
1098 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
a = b = c 6 90; a = c = 90 6 b; a 6 b 6 c 6 90)
and (3) crystal system (simple; base-centred; face-
centred; body-centred).
6. Determination of the output parameters
In Hume-Rotherys Rules, a soluble/insoluble criterion
is described. However, it would be more advantageous to
attempt to predict the original value of the solubility. The
output parameters, which are the solubility limits of each
alloy system, are therefore expressed in two ways:
(1) Follow the specialized criterion: if the solubility of
solute metal in solvent metal exceeds 5 at.% [38,62],
it is said that this solute metal is soluble in the solvent
metal.
(2) Use the original maximum solubility limits of each
alloy system.
7. Results
The data are used in two ways: (1) the rst 60-alloy sys-
tems mentioned by Hume-Rothery in 1934 are used as a
start; (2) the whole 408-alloy systems are then used to test
whether Hume-Rotherys Rules work for copper and silver
alloy systems in general.
7.1. Testing Hume-Rotherys rules within 60-alloy systems
Of the four parameters, the size factor, the electrochem-
ical factor and the relative valency factor were those used
by Hume-Rothery in 1934, so these are used in initial tests
for predicting solubility using the following criteria:
1. If the atomic diameters dier by more than 14%, it
means that the size factor is unfavourable, and the
input number for this parameter is zero, or it is one.
2. If the valency of solvent atom is lower than that of sol-
ute atom, the input number for this parameter is one, or
else the number is zero.
3. If the dierence of the electronegativity of solvent and
solute atom is more than 0.4, mentioned by Darken
and Gurry [34] and Zhang and Liao [38], the input num-
ber for this parameter is zero, or else it is one.
4. If the solubility of solute metal in solvent metal exceed
5 at.%, the output number is one, or else it is zero.
In this case, the problem to be solved is a classication
problem, so a probabilistic neural network [54,55] is
designed for use. This is a radial basis network suitable
for classication problems. The modied criterion of 15%
for the diameter dierence mentioned by Darken and
Gurry [34] is used in the next trial, also with soluble/insol-
uble as output, and both results are listed in Table 1 in
terms of the percentage of all the 60 predictions that are
wrong. Comparing these results, a slight dierence is
found: the percentage error of the predicted results based
Table 1
Testing Hume-Rotherys Rules with 60-alloy systems using his criterion
(14% variation), the later suggestion of 15% and the 15% criterion with
structural identity (same or not)
Choice (%) Error of prediction (%)
Testing Whole data
14 8.3 15
15 8.3 13
Structure parameter 8.3 13
on the experimental results for the whole data set using
the 15% criterion is slightly lower than that for 14%. This
result shows the consistency with the modied criterion
of size factor, although Hume-Rothery stated in his later
work [63]: the 14% dierence does give a better correlation
with solubility data than the commonly accepted 15%.
Using the same approach and including the 15% crite-
rion, the structure parameter is introduced next in terms
of whether the structures of solvents and solutes are the
same (i.e., 1 same, 0 not same). The results are listed in
the last row of Table 1. Comparing these results with the
previous one, there is no improvement in correlation. This
indicates that the structure parameters do not play a very
important role in solubility when the 5 at.% solubility limit
is selected as the threshold. Zhang and Liao [38] com-
mented that taking the 5 at.% threshold at any temperature
is not precise enough, and solubility limits are not accu-
rately predicted when the rules are deployed in this way.
In the next trial, the original values of input parameters
and of output solubilities are used and the structure param-
eter is incorporated using an integer for each of the 14
Bravais lattices. From this trial to the end, all the problems
modelled are mapping problems, so BPANN are adopted
throughout. The separate results for training sets and test-
ing sets are shown in Table 2. For the training set, the
regression coecient R is 0.996, and slope M is 0.984.
However, the prediction for the testing data from the
trained network in this case is very poor (M is 0.193, and
R is 0.383), and this clearly indicates that, although the net-
work trains satisfactorily on the actual values of input
data, it is unable to use these for prediction.
In the next trial, input variables are based on function-
alized values, the structure parameter is omitted, and the
results are shown in Fig. 2, in which the training set (a) is
distinguished from the testing set (b). The values of M, B
and R for the training set are 0.977, 0.23 and 0.993, respec-
tively, and for the testing set they are 0.962, 1.19 and
0.992, respectively. This demonstrates reasonable correla-
tion (the ideal values are M = 1, B = 0 and R = 1).
Table 2
Testing Hume-Rotherys Rules using original parameter values
M B R
Training Testing Training Testing Training Testing
0.984 0.193 0.158 2.75 0.996 0.383
 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105 1099

Best Linear Fit: A = (0.977) T + (0.23) Best Linear Fit: A = (0.962) T + (-1.19)
100 100

80 80

Predicted from NN (A)

Predicted from NN (A)

60 60

40 40

20 20

0 R = 0.993 0 R = 0.992

-20 -20
0 50 100 0 50 100
Experimental (T) Experimental (T)

c Best Linear Fit: A = (0.966) T + (0.0162)

100

80
Predicted from NN (A)

60 Data Points
Best Linear Fit
40
A=T
20

0 R = 0.992

-20
0 50 100
Experimental (T)

Fig. 2. Prediction of solubility using three functionalized parameters for the 60-alloy system data set: atomic size, valency and electronegativity. (a)
Training set; (b) testing set; (c) whole set.

The functionalized structural parameter described above experimental value. The problem of (1) is that a zero mean
is then incorporated in place of the Bravais lattice number, error can be obtained from very large deviations from the
and the results are shown in Fig. 3. Comparing this with line, and the problem with (3) is that, when many experi-
Fig. 2, the dierence is not great, nor can it be said that mental values are zero or close to zero, the percentage is
one is superior to the other. This could imply that the struc- innite or very high, respectively. Thus (2) provides the
ture parameter does not play a very important role and, best criterion and, furthermore, the standard deviation of
indeed, Hume-Rothery did not include it in 1934. this modulus of error gives a measure of spread and, hence,
There are several ways to evaluate the performance of if large, indicates that the error is not systematic. So in
neural network predictions. The rst and simplest is based assessing the correlation, two parameters are used: the cor-
on the value of the linear regression coecient R for the relation coecient R and the average absolute deviation
plot of predicted vs experimental output. A problem occurs between theory and prediction (mean modulus of error).
when R is low (<0.9), and the slope M is close to unity or The two are plotted in Fig. 4 for all data sets and show a
vice versa. Under these circumstances, slope M, intercept good correlation at high R: at R = 1, the mean modulus
B, and R can be combined to give one parameter u as of error is zero.
dened above, which should be as close as possible to zero. These criteria are compared in the rst two data rows of
This has the advantage of providing a single value that can Table 3 for the testing set and whole set of the 60-alloy sys-
be used as a criterion for parameter selection in a looped tem from plots of predicted solubility against experimental
optimization program. However, in this composite param- solubility. The rst thing to notice about this table is that
eter, the contribution of each of M, B and R is treated as the four ways of assessing the accuracy of prediction (test-
equal, whereas a weighting might be preferable. An alter- ing set) concur. As the linear regression coecient
native method is to consider the mean error of the pre- decreases, parameter u increases much more dramatically
dicted value from the experimental value. There are three and can be regarded as a more sensitive indicator for this
ways in which this error can be calculated: (1) the mean reason. Also, a simple calculation of the mean deviation
true error having the same unit as target values; (2) the of the predicted values from the best t line (mean modulus
mean modulus of error again having the same unit as tar- of error) gives an estimate of the accuracy of prediction.
get; and (3) the percentage error (as modulus) based on the This follows the trend of increasing u and reduced R.
1100 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105

Best Linear Fit: A = (0.961) T + (0.417) Best Linear Fit: A = (0.923) T + (6.67)
100 100

80 80

Predicted from NN (A)

Predicted from NN (A)

60 60

40 40

20 20

0 0
R = 0.985 R = 0.976
-20 -20
0 50 100 0 50 100
Experimental (T) Experimental (T)

c Best Linear Fit: A = (0.956) T + (1.61)

100

80
Predicted from NN (A)

Data Points
60 Best Linear Fit
A=T
40

20

0
R = 0.975
-20
0 50 100
Experimental (T)

Fig. 3. Prediction of solubility using four functionalized parameters for the 60-alloy systems: atomic size, valence, electronegativity and structure. (a)
Training set; (b) testing set; (c) whole set.

to mean error is, in all but one case, greater than unity.
These trends in the assessment criteria are consistent for
both the testing set and the whole set.
The best predictive results for the network are obtained
using the functionalized values of atomic size, valence and
electronegativity to predict the original values of solubility
for the 60-alloy data set used by Hume-Rothery himself,
and the data are plotted in Fig. 2. Inclusion of the struc-
tural factor using the parameter described above weakens
the predictive power of the network (Fig. 3). The reason
for this slightly counter-intuitive nding is that crystallo-
graphic compatibility is likely to become more important
at higher solubility levels, being essential for continuous
solubility. However, the majority of data are at the low sol-
ubility end, where substitutional atoms are at a low coordi-
nation number. Another reason is that the number used to
represent structure actually conceals crystallographic simi-
larities, as discussed in more detail below, and there is not
enough training data for the network to establish these sim-
ilarities by itself. The structure parameter is used to assess
Fig. 4. The correlation between R-values and mean modulus of error. the criterion for solubility that the same crystal structure
for the two elements favours a wide solubility range [64].
The standard deviation for this error is an indicator that This makes it a type of classication problem, not com-
the error is random rather than systematic and, if so, the pletely the same as a mapping problem, and it could be
standard deviation is expected to increase with the mean, argued that including it in this type of network is
as it does, in fact. The ratio of standard deviation of error inappropriate.
 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105 1101

Table 3
Comparison of criteria for predicting solubility using dierent combinations of parameter groups
Conditionsa Test set Whole set
R u Mean modulus of SD of modulus of R u Mean modulus of SD of modulus of
error (at.%) error (at.%) error (at.%) error (at.%)
Size, valence, electronegativity 0.992 0.0579 2.46 3.21 0.992 0.0422 1.65 1.94
(60 alloys)
Size, valence, electronegativity, 0.976 0.168 6.98 4.58 0.975 0.0851 3.21 3.21
structure (60 alloys)
Size, valence, electronegativity 0.695 0.662 7.01 14.1 0.768 0.631 6.30 12.7
(408 alloys)
a
Using functionalized parameters.

7.2. Testing Hume-Rotherys rules with the 408-alloy a less ambiguous estimate of the accuracy of prediction.
systems The mean error of the prediction (testing set) increases by
a factor of 3, and the linear regression coecient drops well
From the results for the 60-alloy systems, it is clear that below 0.9. The mean error for the testing set and the whole
using the three functionalized values of parameters pro- set becomes closer, showing that this set does not train
vides better results, so the same approach is adopted for well, whereas for the 60-alloy set, the whole-set errors are
testing the 408-alloy systems. This represents a nearly much lower than the testing-set errors. It is an inevitable
exhaustive set of known silver and copper alloys. The conclusion that the wider application of the rules intro-
results, shown in the last row of Table 3 and plotted in duces diculties, some of which are discussed below.
Fig. 5, use the same format of inputs as those used for
the 60-alloys set. When this method (omitting structural 7.3. Relative importance of the rules
parameter) is applied to the larger 408-alloys data set, the
regression coecient is low (<0.9), and the comparison It is interesting to enquire which of the four parameters,
between dierent regression coecient values has less i.e., atomic size, valence, structure and electronegativity, is
meaning. Calculation of the mean modulus of error gives the most inuential parameter, assuming that they are

Best Linear Fit: A = (0.623) T + (3.01) Best Linear Fit: A = (0.672) T + (2.9)
100 150

80
Predicted from NN (A)

Predicted from NN (A)

100
60

40 50

20
0
0 R = 0.79 R = 0.695

-20 -50
0 50 100 0 50 100
Experimental (T) Experimental (T)

c Best Linear Fit: A = (0.631) T + (2.99)

150
Predicted from NN (A)

100

50 Data Points
Best Linear Fit
A=T
0
R = 0.768

-50
0 50 100
Experimental (T)

Fig. 5. Prediction of solubility using three functionalized parameters for the 408-alloy systems: atomic size, valency and electronegativity: (a) Training set,
(b) testing set, (c) whole set.
1102 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105

independent of each other. The importance of the struc-
tural parameter has been tested and found not to play a
very important overall role, although of course it does
inuence the possibility of continuous solubility.
The relative importance of size factor, valence and elec-
tronegativity is compared in Table 4. Using the same pro-
cedure (functionalized parameters including structure), the
network is run with one parameter omitted at a time on the
set of 60 systems.
In general, mean error (data columns 3 and 7) varies
inversely with regression coecient (data columns 1 and
5), and the standard deviation of error is between 1.1 and
1.8 times higher than the mean error. Using the mean error
of the testing set as our main criterion for accuracy of pre-
diction, the parameters atomic size, valence and electroneg-
ativity provide the strongest prediction of solubility and, of
these, atomic size has the strongest eect because, when it is
omitted, the error is highest (data row 2). Electronegativity
appears to have a stronger inuence than valence (data rows
3 and 4). In fact, these parameters are not wholly indepen-
dent of each other. As mentioned by Hume-Rothery, they
are related, and their interplay makes the determination
of solubility very dicult [22]. As a result, determining the
relative importance of each parameter is not easy; it can
only be said descriptively that the atomic size and electro-
negativity play more important roles than the valence and
structural parameters.
In the next stage, pairs of parameters are selected to pre-
dict solubility: (1) atomic size and valence factors; (2)
atomic size and electronegativity factors; (3) atomic size
and structural factors; (4) valence and electronegativity
factors; (5) valence and structural factors; and (6) structure
and electronegativity factors. They are shown in Table 5.
The rst thing to notice is that most of the mean errors
are increased compared with the three-input tests reported
in Table 4. The correlation coecient for the testing set is
generally higher than that for the whole set, because the
partitioning procedure described above selects minimum
u for the testing set as criteria rather than for the training
set. An ideal procedure would be to nd the correlation for
both sets for each partition and select the distribution that
gives the closest and highest R-values, as described by Mal-
inov and Sha [4]. When the correlation is poor, however, as
for the eects of valence and structure, the value of R has
little meaning. Table 5 conrms the deductions from the
three-parameter tests that atomic size has the strongest
eect on solubility, and the structural parameter the least
eect. However, some ambiguity attends the relative roles
of electronegativity and valence, which are reversed in this
assessment of ranking. Pearson [65] states that when one

Table 4
Comparison of criteria for predicting solubility using dierent combinations of three parameters
Conditionsa Test set Whole set
R u Mean modulus of SD of modulus of R u Mean modulus of SD of modulus of
error (at.%) error (at.%) error (at.%) error (at.%)
Size, valence, electronegativity 0.992 0.0579 2.46 3.21 0.992 0.0422 1.65 1.94
(60 alloys)
Valence, electronegativity, 0.867 0.308 8.19 11.7 0.924 0.197 4.20 6.34
structure (60 alloys)
Size, structure, electronegativity 0.93 0.365 3.17 4.19 0.968 0.142 3.46 3.66
(60 alloys)
Size, valence, structure (60 0.569 0.477 7.73 13.8 0.761 0.613 7.07 11.0
alloys)
a
Using functionalized parameters.

Table 5
Comparison of criteria for predicting solubility using dierent combinations of two parameters
Conditionsa Test set Whole set
R u Mean modulus of SD of modulus of R u Mean modulus of SD of modulus of
error (at.%) error (at.%) error (at.%) error (at.%)
Size, valence (60 alloys) 0.852 0.470 4.47 2.50 0.496 1.36 9.52 14.3
Size, electronegativity (60 0.679 0.860 6.99 4.83 0.495 1.36 10.3 13.8
alloys)
Size, structure (60 alloys) 0.675 0.889 10.2 6.42 0.441 1.50 10.7 14.4
Valence, electronegativity 0.91 0.153 7.31 10.1 0.925 0.184 4.54 6.06
(60 alloys)
Valence, structure (60 0.459 1.02 12.4 11.9 0.662 0.886 9.82 11.3
alloys)
Structure, electronegativity 0.607 1.30 11.3 21.1 0.524 1.35 9.00 14.5
(60 alloys)
a
Using functionalized parameters.
 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
component in a binary alloy is very electropositive relative
to the other, there is a strong tendency for them to form
compounds of considerable stability in which valence rules
are satised. Such alloys are said to exhibit a strong elec-
trochemical factor and this is the strongest eect in deter-
mining the constitution of alloys, and one which dominates
all other eects such as energy band or geometrical factors.
8. Discussion
It is important to recognize that there are four factors
that limit the predictive capability of the networks: (1)
imperfections in the network conguration, which the
authors have attempted to minimize through design; (2)
paucity of learning data, which has been discussed above;
(3) the generality of Hume-Rotherys Rules which were
conceived as guidelines; and (4) the fact, recognized by
Hume-Rothery and co-authors, that the available data
are subject to inexactitudes.
8.1. The validity of ANN models
It can be seen that, if the parameters are selected appro-
priately, as shown in Figs. 2 and 3, the prediction of the
solid solubility limit by the ANN is reasonably consistent
with Hume-Rotherys Rules. The ANN, as a method, can
be treated as feasible, although it cannot be relied on den-
itively, and others have reiterated this. It may be regarded
as a useful tool for cautious use in materials science, but the
choice of right ANN plays a critical role in its success, espe-
cially when the data set is restricted, as is often the case in
materials science.
8.2. The eect of number of layers
Basheer and Hajmeer [2] indicate that the choice of the
number of hidden layers and the number of neurons in the
hidden layers are among the most important choices in
ANN design. It is often claimed that, in most function
approximation problems, one hidden layer is sucient to
approximate continuous functions [48,66]; two hidden lay-
ers must generally be necessary for learning functions with
discontinuities [67]. In this work, the type of function is not
clear. Also the neural network users guide [60] suggested
that a two-hidden-layer sigmoid/linear network can repre-
sent any function of input/output relationship. On these
bases and looking at the results produced from this work,
it can be seen that the choice of two-hidden layer network
is a sensible one.
8.3. The eect of size of layer
The choice of size of the rst hidden layer is critical in
the ANN design. There are several rules of thumb available
in the literature relating hidden layer size to the number of
nodes in input (NINP) and output (NOUT) layers [48,6771].
1103
As Basheer and Hajmeer [2] suggested, the most popular
way to nd the optimal number of hidden nodes is by trial
and error with one of those rules as a starting point. How-
ever, facing exotic problems with high non-linearity and
hysteresis such as are shown in Basheers work [66,72],
these rules of thumb may need to be abandoned. There
is some value in beginning with a small number of hidden
nodes and building up iteratively to attain the accuracy
required. This method is adopted in this work through
implementation of the program.
8.4. The reliability of input parameters
Hume-Rothery et al. [22] themselves made it clear that
the exact atomic diameter of an element is always dicult
to dene. Their denition of atomic diameter, as given by
the nearest-neighbour distance in a crystal of the pure
metal, was used here but the radius of an atom is probably
aected by coordination number. Except for the heavy ele-
ments, elements of the B sub-groups tend to crystallize with
coordination number 8N, where N is the group to which
the element belongs. This is due to the partly covalent nat-
ure of the forces in these crystals and, except in Group IV B
(diamond structure), results in the atoms having two sets of
neighbours at dierent distances in the crystal. Cottrell [73]
suggests that the concept of a characteristic size, which sug-
gests hard spheres butted together, is doubtful. Allocating a
single atomic diameter for each element, independent of its
environment, and valences of solvent and solute is too sim-
plistic an approach [62]. Furthermore, within the 408-alloy
systems, the metallic radius of some elements could not be
found, and the covalent radius was used instead. These fac-
tors contribute to the errors for the prediction of solid sol-
ubility limit and are to be distinguished from the intrinsic
weaknesses of the ANN.
An early discovery by Hume-Rothery was that a metal
of lower valence is more likely to dissolve one of higher
valence than vice versa. However, more detailed examina-
tion has not conrmed this. For example, silver dissolves
about 20% aluminium, but aluminium dissolves about
24% silver. For high valence, covalently bonded compo-
nents, the relative valence factor applies. For example,
copper dissolves about 11% of silicon, which behaves as
a four-valent metal in forming CuSi electron phase
alloys, but the solubility of copper in covalently bonded
silicon is negligible [73]. As a result, although Hume-
Rothery [62] accepted that it is still a general principle
that the solubility in the element of lower valency is of
greater extent when dealing with alloys of univalent met-
als copper, silver and gold with metals of higher valency,
in its general form, this principle must be treated with
caution.
The valencies of transition metals are variable and com-
plex and have been analysed by Hume-Rothery et al. [74]
and Cockayne and Raynor [75]. As suggested by Cottrell
[73], due to the valency complication caused by partly lled
d shells, the transition metal alloys generally do not follow
1104 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
the rule. Gschneider [76] modied the relative valence rule
so that the solubility is low when a metal in which d orbi-
tals strongly inuence the valence behaviour is alloyed with
a simple sp metal, but that the solubility is likely to be bet-
ter in the d metal than the reverse.
The electronegativity rule needs a scale, such as that
given by Mullikan, based on the equation v 12 I A,
where I is the ionization energy, A is electron anity,
and v is Mullikan electronegativity. When divided by 2.8,
this scale matches the empirical scale of Pauling reasonably
well. In the case of transition metals, as emphasized by
Watson and Bennett [77], the partly lled d states of tran-
sition metals at energies near the Fermi energy inuence
electronegativity. Watson and Bennett presented an elec-
tronegativity scale for transition metals that matched Paul-
ings scale, and could be scaled by 2.8 to bring it to
Mullikans scale of v values. Most importantly, Li and
Xue [78] have mentioned that the although electronegativ-
ity is often treated as an invariant property of an atom, as
in Paulings scale, it actually depends on the chemical envi-
ronment of the atom, e.g. valence state and coordination
number. The electronegativity values adopted in this pro-
ject are based on Paulings work, so the above eects are
not entirely taken into account.
The method adopted for expressing structure parame-
ter has some limitations. First, the expressions used to
distinguish dierent crystal structures can conceal simi-
larities. For unit cell length, a = b = c and a = b 6 c
are distinguished but can have considerable similarity.
Secondly, from this expression, the face centred cubic
(fcc) and the hexagonal close packed (hcp) systems are
expressed as quite distinct sets, but there are some simi-
larities between these two structures. They are both close
packed systems, and stacking faults can blur the dier-
ence. Indeed, the Cu and Zn systems demonstrate high
solubility, even though one component, Zn, is hcp and
the Cu is fcc. The Ag and a-Li system is a similar case.
Thirdly, there are several complex structural systems that
cannot be distinguished from other systems by using this
expression, such as a-Mn, whose structure is cI58, and b-
Mn, whose structure is cP20. These all aect the ability
of the structure parameter to contribute to predicting
the solubility.
8.5. The generality of Hume-Rotherys rules
Hume-Rothery and co-workers state: In general, the
solubility limit is mainly determined by these factors, and
it is their interplay that makes the results so complex
[22]. For the 60-alloy systems mentioned by Hume-Rothery
in 1934, and using some of the parameter values that can be
found in Hume-Rotherys paper or his book and others
that follow his representations, the results for prediction
of the solid solubility limit are satisfactory.
However, from theory as analysed by others in later
work [38,73,79], and also from attempts to predict the solid
solubility limit of the 408-alloy systems, it can be said that
Hume-Rotherys Rules work properly in a certain range of
alloy systems, but cannot be treated as general principles.
Also, it needs to be said that, despite using Hume-Roth-
erys Rules, one cannot predict the solid solubility limits
accurately. However, these rules are still useful guidelines
for judging the solubility of alloy systems.
9. Conclusions
ANN oer materials scientists a relatively new tool for
examining their data with the intention of making predic-
tions while theory is still too opaque to be predictive. It
is often the case in materials science that data sets are lim-
ited, either inherently, because of the limits imposed by the
number of elements, or extrinsically, because of the high
cost of experimentation. This study has taken one of the
cornerstones of physical metallurgy and adopted ANN
for predicting the solid solubility limit of alloy systems
based on Hume-Rotherys Rules. Application of a two-hid-
den-layer backpropagation network with functionalized
input parameter values for dierent classes and numbers
of alloy systems, indicates that: (1) ANN is a useful tool
for dealing with forecasting problems or mapping prob-
lems in materials science; (2) Hume-Rotherys general prin-
ciples work well in several alloy systems, such that the
ANN can be used to estimate solid solubility. When
the 60-alloy systems used by Hume-Rothery are tested,
the rules work very well, as demonstrated by the ANN cor-
relation. The wider application of the rules to a set of 408
silver and copper alloys is less successful, but this is consis-
tent with the inherent simplication of the rules which are
already documented.
References
[1] Dorsett H, White A. Overview of molecular modeling and ab initio
molecular orbital methods suitable for use with energetic materi-
als. Salisbury: Aeronautical and Maritime Research Laboratory;
2000. p. 3.
[2] Basheer IA, Hajmeer M. J Microbiol Methods 2000;43:3.
[3] Fausett LV. Fundamentals of neural networks: Architectures, algo-
rithms and applications. Englewood Clis, NJ: Prentice-Hall; 1994.
[4] Malinov S, Sha W. Comput Mater Sci 2003;28:179.
[5] Jain AK, Mao JC, Mohiuddin KM. Computer 1996;29:31.
[6] Scott DJ, Coveney PV, Kilner JA, Rossiny JCH, Alford NMcN. J
Eur Ceram Soc 2007;27:4425.
[7] Rumelhart DE, Hinton GE, Williams RJ. Learning internal
representation by error propagation. In: Rumelhart DE, Mcclle-
land JL, editors. Parallel distributed processing: Exploration in the
microstructure of cognition, vol. 1. Cambridge, MA: MIT Press;
1986.
[8] Arkadan AA, Chen Y, Subramaniam S, Hoole SRH. IEEE Trans
Magn 1995;31:1984.
[9] Raj KH, Sharma RS, Srivastava S, Patvardhan C. Int J Mach Tools
Manufact 2000;40:851.
[10] Guessasma S, Coddet C. Acta Mater 2004;52:5157.
[11] Huang CZ, Zhang L, He L, Sun J, Fang B, Zou B, et al. J Mater
Process Technol 2002;129:399.
[12] Malinov S, Sha W. Mater Sci Eng A 2004;365:202.
[13] Martinez SE, Smith AE, Bidanda B. J Intell Manuf 1994;5:277.
[14] Bork U, Challis RE. Meas Sci Technol 1995;6:72.
 Y.M. Zhang et al. / Acta Materialia 56 (2008) 10941105
[15] Larkiola J, Myllykoski P, Nylander J, Korhonen AS. J Mater Process
Technol 1996;60:381.
[16] Gavard L, Bhadeshia H, MacKay DJC, Suzuki S. Mater Sci Technol
1996;12:453.
[17] Malinov S, Sha W, Guo Z. Mater Sci Eng A 2000;283:1.
[18] Homer J, Generalis SC, Robson JH. Phys Chem Chem Phys
1999;1:4075.
[19] Guo D, Wang YL, Nan C, Li LT, Xia JT. Sens Actuators A
2002;102:93.
[20] Cai K, Xia JT, Li LT, Gui ZL. Comput Mater Sci 2005;34:166.
[21] Schooling JM, Brown M, Reed PAS. Mater Sci Eng A 1999;260:222.
[22] Hume-Rothery W, Mabbott GW, Channel-Evans KM. Philos Trans
Soc A 1934;233:1.
[23] Cooke CJ, Hume-Rothery W. J Less Common Met 1966;10:52.
[24] Pauling L. The nature of the chemical bond and the structure of
molecules and crystals: An introduction to modern structural
chemistry. Ithaca, NY: Cornell University Press; 1960. p. 8895.
[25] Ohtani H, Ishida K. Thermochim Acta 1998;314:69.
[26] Kaufman L, Bernstein H. Computer calculation of phase diagrams:
with special reference to refractory metals. New York: Academic
Press; 1970.
[27] Hillert M. Calculations of phase equilibria. In: ASM, editor.
American Society for Metals Seminar on Phase Transforma-
tions. Metals Park (OH): ASM International; 1968. p. 181218.
[28] Lim SS, Rossiter PL, Tibballs JE. Calphad 1995;19:131.
[29] Yang J, Silk NJ, Watson A, Bryant AW, Chart TG, Argent BB.
Calphad 1995;19:415.
[30] Fries SG, Ansara I, Lukas HL. J Alloys Compd 2001;320:228.
[31] Ohnuma I, Fujita Y, Mitsui H, Ishikawa K, Kainuma R, Ishida K.
Acta Mater 2000;48:3113.
[32] Du Z, Yang H, Li C. J Alloys Compd 2000;297:185.
[33] Liu ZK, Zhong Y, Schlom DG, Xi XX, Li Q. Calphad 2001;25:299.
[34] Darken LS, Gurry RW. Physical chemistry of metals. New
York: McGraw-Hill; 1953.
[35] Chelikowsky JR. Phys Rev B 1979;19:686.
[36] Alonso JA, Simozar S. Phys Rev B 1980;22:5583.
[37] Alonso JA, Lopez JM, Simozar S, Girifalco LA. Acta Metall
1982;30:105.
[38] Zhang BW, Liao SZ. Shanghai Met 1999;21:3.
[39] Massalski TB, Murray JL, Bennett LH, Baker H, editors. Binary
alloy phase diagrams. Metals Park (OH): American Society for
Metals; 1986. p. 187. pp. 90882.
[40] Stark JG, Wallace HG, editors. Chemistry data book. London: Mur-
ray; 1982. p. 24. pp. 279.
[41] Guessasma S, Montavon G, Coddet C. Neural Networks, design of
experiments and other optimizations methodologies to quantify
parameter dependence of atmospheric plasma spraying. In: Marple
R, Moreau C, editors. Proceedings of the Thermal Spray 2003:
Advancing the science and applying the technology. Materials Park
(OH): ASM International; 2003. p. 39.
[42] Axon HJ, Hume-Rothery W. Proc Roy Soc A 1948;193:1.
[43] Hassoun MH. Fundamentals of articial neural networks. Cam-
bridge, MA: MIT Press; 1995.
[44] Hopeld JJ. Proc Natl Acad Sci USA-Biol Sci 1984;81:3088.
[45] Hopeld JJ, Tank DW. Science 1986;233:625.
[46] Kohonen T. Self-organization and associative memory. Ber-
lin: Springer; 1989.
[47] Zupan J, Gasteiger J. Anal Chim Acta 1991;248:1.
1105
[48] Hecht-Nielsen R. Neurocomputing. Reading, MA: Addison-Wesley;
1990.
[49] Hecht-Nielsen R. Neural Networks 1988;1:131.
[50] Zupan J, Gasteiger J. Neural networks for chemists: An introduc-
tion. New York: VCH-Weinheim; 1993.
[51] Schalko RJ. Articial neural networks. London: McGraw-Hill;
1997.
[52] Pal SK, Srimani PK. Computer 1996;29:24.
[53] Attoh-Okine NO, Basheer IA, Chen D-H. Use of articial neural
networks in geomechanical and pavement systems. Washing-
ton: Transportation Research Board, National Research Council;
1999. p. 5.
[54] Specht DF. Neural Networks 1990;3:109.
[55] Vicino F. Substance Use Misuse 1998;33:335.
[56] Moatt WG, editor. The handbook of binary phase diagrams. Sche-
nectady (NY): General Electric Co.; 1977. Sections AgAl to AgZr,
CuAl to CuZr.
[57] ASM handbook, vol. 3, Alloy phase diagrams. Metals Park (OH):
ASM International; 1992.
[58] Aylward GH, Findlay T, editors. SI chemical data. New York: Wiley;
1998. p. 613.
[59] Hume-Rothery W. Elements of structural metallurgy. London: Insti-
tute of Metals; 1961. p. 10710.
[60] Matlab. http://www.mathworks.com/access/helpdesk/help/pdf_doc/
nnet/nnet.pdf/.
[61] Vegard L. Z Phys 1921;5:17.
[62] Hume-Rothery W, Smallman RE, Haworth CW. The structure of
metals and alloys. London: Metals and Metallurgy Trust of the
Institute of Metals and the Institution of Metallurgists; 1969.
[63] Hume-Rothery W. Acta Metall 1966;14:17.
[64] Wyatt OH, Dew-Hughes D. Metals, ceramics and polymers: An
introduction to the structure and properties of engineering materi-
als. London: Cambridge University Press; 1974. p. 42.
[65] Pearson WB. The crystal chemistry and physics of metals and
alloys. New York: WileyInterscience; 1972. p. 68.
[66] Basheer IA. Comput Aided Civil Infrastruct Eng 2000;15:440.
[67] Masters T. Practical neural network recipes in C++. Boston: Aca-
demic Press; 1993.
[68] Widrow B, Lehr MA. Proc IEEE 1990;78:1415.
[69] Upadhaya B, Eryureka E. Neural Technol 1992;97:170.
[70] Lachtermacher G, Fuller JD. J Forecast 1995;14:381.
[71] Jadid MN, Fairbairn DR. Eng Appl Artif Intell 1996;9:309.
[72] Basheer IA. Neuromechanistic-based modeling and simulation of
constitutive behavior of ne-grained soils. PhD thesis, Kansas State
University, Manhattan; 1998. 435p.
[73] Cottrell A. Concepts in the electron theory of alloys. London: IOM
Communications; 1998. p. 567. pp. 72, 92.
[74] Hume-Rothery W, Irving HM, Williams RJP. Proc Roy Soc A
1951;208:431.
[75] Cockayne B, Raynor GV. Proc Roy Soc A 1961;261:175.
[76] Gschneider KA. L.S. Darkens contributions to the theory of alloy
formation and where we are today. In: Bennett LH, editor. Theory of
alloy phase formation. Warrendale: The Metallurgical Society of
AIME; 1980. p. 134.
[77] Watson RE, Bennett LH. Phys Rev B 1978;18:6439.
[78] Li KY, Xue DF. J Phys Chem A 2006;110:11332.
[79] Miedema AR. J Less Common Met 1973;32:117.