THE APPLICATION OF NEURAL COMPUTING METHODS

TO THE MODELLlNG OF FATIGUE IN NI-BASE SUPERALLOYS

J.M.Schoolingi and P.A.S.Reed#.

TDept. of Materials Science and Metallurgy, University of Cambridge, U.K.
#Dept. of Engineering Materials, University of Southampton, U.K.

Abstract The fatigue process is sufficiently complicated that, in spite of all

the work which has been undertaken in this area, a comprehensive
The current financial climate is driving a move towards increased fundamental behavioural model is lacking. Hence a reliable
use of computer modelling techniques in alloy design and empirical method for predicting fatigue life, using existing data,
development in order to reduce cost. In this paper the potential for remains a desirable goal.
use of neural computing methods in the prediction of fatigue
resistance in Ni-base superalloys is assessed. Initial work has been The complexity of the fatigue process and the noise associated with
conducted on the Stage II (Paris regime) behaviour, as the literature fatigue test results has meant that even traditional empirical
indicates that this is the simplest region of the fatigue crack growth methods, such as regression analysis, have failed to produce a
curve to predict, with an approximately linear relationship existing sufficiently comprehensive, robust model which accurately predicts
between log(da/dN and log(AK), and the crack growth rates being trends in the data and yet does not model the noise.
principally affected by temperature, Young’s modulus and yield
strength. These three parameters were chosen for initial data Artificial neural networks are powerful computing devices designed
collection and modelling. The predictions made are of fatigue life, to mimic the structure and learning capabilities of the brain,
calculated from the slope and intercept values of the linear portion of consisting of a large number of simple computational elements (or
the log-log fatigue crack growth curve. A test dataset has been nodes) which are extensively interlinked via weighted connections.
successfully predicted along with the trends in the data. The effect Such networks have the ability to learn rather than being
of adding ultimate tensile strength and electron valencies as inputs to programmed, can pick out complex patterns or trends in data, and
the model is assessed. It is shown that validation of models can deal with noisy or irrelevant data points in an input dataset. In
produced against metallurgical experience, and careful construction this paper the possibilities for using such a neural network as a tool
of the database are important conditions for effective use of neural to model fatigue using existing fatigue crack growth data are
network models for fatigue life predictions. evaluated.

Neural Networks

In recent years there has been an increased interest in the ability to Artificial neural networks are computational tools based on the
model the mechanical properties of alloys from compositional and structure and function of the brain [Z]. They are composed of
processing data in order to reduce the cost and time required for simple computational elements (called neurons or nodes) which
alloy development. Increased expectations of engine performance, imitate the most basic function of a biological neuron. These
in terms of speed and range, have led to the need for higher artificial neurons are then connected to others by a series of
performance alloys, while the financial climate has forced a connections broadly analogous to, although much simpler than,
reduction in the cost of new products. those in the brain. The nodes are arranged in layers, with each node
being connected to every node in the adjacent layers (figure 1). The
Good fatigue crack growth resistance is an essential property for simplest node sums N weighted inputs, performs a non-linear
superalloys operating in the high temperature stage of gas turbine function on the sum, and then passes the result to the nodes in the
engines; hence over the last thirty years many studies on the fatigue next layer [3]. Data flow forwards only through the network.
crack growth behaviour of Ni-base superalloys have been
conducted. The general trends in fatigue crack growth behaviour There are two phases in the development of a neural network model
with variations in test conditions (load ratio, temperature and [4]. Initially the network is ‘taught’ using a number of example
environment), microstructure and processing route are well datasets - this process is referred to as training. The training
established. In general such alloys exhibit two stages of long crack datasets consist of a series of inputs paired with the corresponding
fatigue behaviour: Stage I facetted crack growth. occurring along output (the ‘target’ output). An input dataset is applied, the network
crystallographic ( 111 r planes corresponhing to persist&t sli: calculates an output and compares it with the target. The error
bands; and more homogeneous Stage II crack growth resulting in between output and target is calculated, and the weights in the
essentially flat fracture &faces [l]. -Inspection mtervals for turbine network are then adjusted using an algorithm (in the case of the
comuonents are often determined on the basis of Stage II fatigue network used in this work, a ‘back propagation’ algorithm) in order
crack growth rates, assuming an initial crack length Equal to The to improve the output. The data are fed through repeatedly until the
smallest detectable flaw size. Such flaws might be surface network output is deemed sufficiently accurate (the solution has
scratches, coating cracks or casting defects such ai pores. In this ‘converged). After this training the network is tested on a set of
paper Stage II crack growth is concentrated on as the regime of previously unseen data.
most interest with respect to lifing procedures.
Superalloys 1996
Edited by R. D. Kissinger, D. J. Deye, D. L. Anton.
A. D. C&l, M. V. Nathal, T. M. Pollock, and D. A. Woodford
The Minerals, Metals &Materials Society, 1996
409
 Input Hidden output relationships between variables. The ability of networks to
layer layer layer generalise and find patterns in large quantities of often noisy data is
also a major advantage. However the number of sets of training
A data requjred to establish a robust network is dependent on the
number of inuut variables (the ‘dimensionalitv’ of the data).
Previous studies have suggested that the size of ihe training data
should be between three and ten times the input dimensionaIity[5].

Neural networks in materials science

It has been demonstrated in the literature that neural networks may
be used with some successto model material properties and material
behaviour [6, 7, 8, 91. While it is acknowledged that material
behaviour is best understood by carrying out experimental
urogrammes, it is not always uossible to describe the behaviour in
direction of data flow ierr& of a simple mathematical expression, and hence quantitative
modelling of behaviour is difficult, and will become more so as
modem materials are further developed with increasingly complex
Figure 1 - Schematic diagram of a neural network behaviour.

It is possible to vary the number of nodes in the hidden layer of the Feedforward neural networks can be very useful in picking out
network (the ‘architecture’), and care must be taken when deciding patterns of behaviour and property relationships from a quantity of
the network architecture, as it is possible to overmodel or experimentally produced data, and their ability to generalise can
undermodel the data. If too complex a network is chosen, then the make them useful in predicting the behaviour of a potential new
training data will appear to be excellently modelled, however the material before it is made. Work on the strength properties of Ni-
network will be modelling the noise in the data as well as the trends, base superalloys [6] has shown that trends in behaviour with
so that when the model is-tested it will not be robust enough to cope varying composition, temperature and material condition can be
with the new data presented to it. On the other hand, if too simple a modelled, and within error limits absolute values of strength may be
network is chosen it will fail to model the trends in enough detail, estimated.
and predictions will again be very inaccurate. Examples of this type
of behaviour are shown in figure 2. In this work, therefore, the This paper describes work carried out on the modelling of the
network is trained over a number of different architectures and the Stage II fatigue properties of such alloys using a supervised
results are compared. feedforward neural network as described above.

The datasets
A number of input variables have been considered for presentation
to the network. These are: temperature, yield stress, Young’s
modulus, ultimate tensile strength and NV number.
Sixtv-four sets of inuut data were gathered from a Rolls-Rovce
database [27], for wb&h fatigue testsare conducted at a frequency
of 0.25 Hz using a l-l-l-l trapezoidal wave form. The network

Cc). a) undermodelling - too few

nodes in hidden layer
b) good model -trends
trained using 34 sets and tested itself on the remaining thirty sets.
Before presentation to the network the data for each input variable
was normalised between -0.5 and +0.5, according to the equation:

picked out

A.
c) overmodelling - too
many nodes in hidden layer
where ni is the normalised value of datapoint i and xi, xmM and Xmin
are the actual values of datapoint i, and the maximum and minimum
valued datapoints. This is to prevent a variable from swamping the
network simply by virtue of having a large absolute value rather than
as a result of the effect of its variation on the result.
Figure 2 - the effect of number of nodes on modelling accuracy
Each set of input data is presented to the network along with a
A point to note is that a neural network learns from experience, and corresponding expected output. In this case the chosen output was a
hence while it may interpolate between data with some confidence. ‘life’ calculated from crack-C and crack-n values where crack-n is
it cannot accurately extrapolate into regions for which it has no the slope of the log dafdN vs. log AK fatigue curve in the Stage II
information, and any attempts at such predictions should be treated
with extreme caution. region and crack-C is the intercept of this line on the da/dN axis<see
figure 3). All fatigue data were uroduced at a load ratio R=O.1. and
The main advantage of neural networks over conventional I&es were calculated using a prediction program for a semi-elliptical
regression analysis techniques is that the network finds an optimum crack [IO]. These ‘life’ values were also normahsed.
solution without the need to specify the relationships or the form of

410
 random number generator which is started using a ‘seed’. This
seed can have any positive value, and using different seeds may
produce slightly different models. In order to ensure that all
potential model types are generated the network is trained using a
number of seed values for each architecture.

The network finds the optimum solution by minimising a penalised

likelihood, in effect trying out a number of solutions to find the best
relationship. As well as calculating the weights for the connections
the network calculates a number of other parameters:

c nu - a measure of the noise allowed in the networks prediction

of the data. Initially this is set at a fairly large value, and the
program modifies the inferred noise level as it develops the
model to more accurately fit the data. The value of this parameter
Figure 3 - Schematic to show crack-C and crack-n varies with the number of nodes, generally decreasing to a limit
as the number of nodes increases and a more complex model is
When choosing a dataset to train the network on it is important to formed such that the relationship between inputs and outputs is
try and ensure that the data are spread evenly over the areas of input better modelled. The value of crnuis a fraction of the range of the
space which are of interest. Otherwise the network may be
attempting to extrapolate from well defined regions of input space output data, e.g. q,“- -0.2 means that the inferred noise in the
into regions about which nothing is known, and predictions may be data is 20% of the range of the target output dataset.
unreliable. The range of values of input and output data is shown in
figure 4. training energy - a measure of the error in predicting the training
dataset targets (this usually follows a similar pattern to qiu).

test energy - a measure of the error in predicting the previously

Training and testing the network unseen test dataset. This initially decreases with increasing
number of nodes to a point, but if overmodelling starts to occur,
The network used is based on a Bayesian statistical framework, and the network has effectively ‘learnt’ the training dataset, then
which allows the probability of a model being a true representation it will be unable to cope with noise in the test set and test energy
of the data to be assessed[ll]. Initial weights are set using a will start to increase.

Temperature Yield stress Young’s Modulus

lo ill 15 L

0 200 400 600 800 1000 500 900 1100 140 160 180200 220 240 260
Range (“C) Rz\e (MPadm) Range (GPa)
(4 (b) (cl

Life (cycles)
UTS (MPa)
25,

5 5

0 0
0 .5
Range Range (cycles)
(4 (4 (0
Figure 4 - graphs to show the range of the input and output data

411
 relevance - this is a value which is produced for each input to the
network, and is a measure of how important that input is in the
0.18t + + + + + + *
model produced. A highly important input would have a 0.16 - +
relevance of the order of 1, whereas an input which is perceived 0.14 - Simple model
to be irrelevant would be assigned a relevance of the order of
10e3or less.
0.12 -
Energy values are calculated from the following equation: g 0.1 - +

energy=ix(o-t)’ (2) 0.08 - +

L

0.06 - Complex model

where o is the output value and t is the target value.
0.04 - +
In order to be useful as a cost-saving design tool, the number of + + + +
inputs to the network should ideally be kept to the minimum 0.02 I I I I I I I
possible. For this reason it was decided to ‘start small’ and present
the network with few variables, gradually adding other possibly 12 3 4 5 6 7 8 9
important variables to see whether they had a positive or detrimental No. of nodes
effect on the predictive ability of the nktwork.~Another advantage of
this approach is that it keeps the input dimensionality low, thus Figure 5a) - Graph of noise vs. no. of nodes for the simple dataset
requiring less data to produce a robust model.
0.6

0.4
The basic database
A review of literature on Stage II Paris regime fatigue crack growth 0.2
and crack tip opening displacement (CTOD) models for crack
growth indicates that the most important properties affecting fatigue
crack growth are likely to be yield strength and Young’s modulus
[12, 13, 14, 15, 16, 17, 18, 191. CTOD theory suggests the
following form of equation for fatigue crack growth rate: -0.2

da AK2 -0.4
dlv OccrsE
where da/dN is the crack growth rate, AK is the stress intensity -0.6
factor range, ors is the yield stress and E is the Young’s modulus -0.6 -0.4 -0.2 0 0.2 0.4 0.6
[17, 181. The literature also suggests that test temperature is an Target
important factor affecting fatigue crack growth rates [20, 211.
Hence initial modelling used an input dataset consisting of
Figure 5b) - Graph of network output vs. target output for a type-B
temperature, o,, and E. model
The network was trained for 2-8 hidden nodes. For each number of The second set of models on the other hand display a decreasing
nodes seeds of 0,20,40, 60 and 99 were used. noise level, reaching a minimum of about 0.034 at four hidden
nodes, indicating the formation of a more complex model. Training
A plot of uriU vs. number of nodes is shown in figure Sa. Two and test energies again followed a similar pattern, with values of
different types of model are generated by the network, depending training energy around 0.015 and test energy around 0.03. When
on the initial random weights (determined by the ‘seed’). The first the relevance values of the input variables for the two different types
of model were compared it was found that the simple models all had
type tend to have relatively high onu values (0.18, or 18% of the
very low relevance for Young’s modulus (e.g. 1*10-4,.c.f. 0.6 for
range of the output data), this value being independent of the yield stress) whereas the more complex models attributed a much
number of nodes. This indicates that an over-simple model is being higher relevance to Young’s modulus (e.g.1.4, c.f. 2.2 for yield
produced which is just as accurately described with two nodes as stress), more in accordance with CTOD theon, and experimental
with five or seven: Training and lest energies for these models evidence. A graph of predicted output vs. iarget output for a
followed a similar pattern and were also high, with values of about comulex model is shown in figure 5b. It can be seen that
0.5 in both cases. This type of model will be referred to as the predictions are good, with most oflthe predictions lying very close
simple model. to the line for output = target.

412
 When predictions of varying yield stress are examined, however, an
unexpected result is obtained, as life is predicted to decrease rapidly
as yield stress increases from 700-800 MPa, followed by a slight
it!: increase to 1000 MPa and then a further gentle decrease in life
8 (figure 6b). CTOD theory on the other hand would indicate an
0 .‘a increase in life with increasing yield stress in a similar manner to
. q that seen in the Young’s modulus predictions. There are two
4 possible reasons for this apparent anomaly. It is possible that it
q
may be explained by the fact that these predictions were made for a
i”, temperature of 600 “C, and the data contain only two alloys with a
yield stress less than 800 MPa at this temperature, both of which
-0.6 - have a Young’s modulus of about 170 GPa, rather than the
195 GPa used for prediction. Hence attempting to predict life in
the data range used may effectively be an extrapolation into an
unknown area. While such extrapolations may not necessarily be
totally misleading, this example illustrates that care is needed if
attempting them. It also indicates the importance of constructing,
where possible, a database which would cover the possible range of
200 300 400 500 600 700 800 data of interest if neural network modelling is to be used as a design
tool. That is not to say however that a neural network could not be
T ("Cl used to reduce experimentation - if an unexpected result is found,
1 then one or two experiments in the area of interest may suffice to
test the proposed model, without an entire test matrix being
(b) required. The results could then be added to the training dataset.

Another possible explanation is revealed by further examination of

the input data. If the data for life are plotted versus yield stress for
temperatures between 550 “C and 650 “C (figure 7) it can be seen
that there is in fact a reduction in life with increasing yield stress at
low yield stress ranges. The points which cause this are all points
for one particular alloy in the database which is a casting alloy, as
opposed to a wrought alloy, and hence has a very different
microstructure compared to the other alloys in the database.
Examination of the actual test data revealed that the tests had been
conducted on specimens containing only two grains, and the cracks
had deflected at high angles from the normal to the principal stress,
giving low apparent crack growth rates. The existence of only two
grains in the specimens and the high deflection angle of the cracks
indicate that the cracks may have propagated in Stage I, where
CTOD models do not apply, as slip is inhomogeneous, being
600 700 800 900 1000 1100 1200 concentrated in intense bands. The low crack growth rates may
yield stress (MPa) have occurred as a result of crack closure [21] and/or shielding due
to crack deflection [26], both of which reduce the effective stress
Figure 6 - Predictions of life generated using the simple database for intensity at the crack tip. This result indicates the importance of
(a) varying temperature and (b) varying yield stress careful construction of the database to cover the problem of interest,
without including spurious data.
It is important to test the physical validity of a neural network
model, to ensure that the network really has converged, and has not
found a local minimum in the data. Thought experiments were 5 104 y
conducted using a model dataset in which the input variables took
Data for casting alloy
on the median value of the original dataset (T=600 “C, 0ys=950 4 104 0
MPa, E=195 GPa) and then each variable in turn was varied in 0
order to see what trends the neural network had found. The best six
complex models, ranked on test error, were chosen and the best
simple model was used for comparison. The prediction of life with 33 104 /s
K
variation in temperature is shown in figure 6a. The prediction s
follows the expected trend of decreasing life with increasing
temperature [21], with the effect becoming more marked as
temperature increases. This can be understood both in terms of the
decrease in yield stress (typically about 10% for a given alloy over
this temperature range) and also the likely change in deformation
mechanism. At higher temperatures a transition from planar to
wavy slip is expected, with increased cross slip occurring and slip
becoming inherently less reversible, resulting in more damage
accumulation per cycle. The simple model produces a much simpler 600 700 800 900 1000 1100 1200
linear trend, indicative of undermodelling. The prediction of life yield stress (MPadm)
variation with Young’s modulus for the complex models shows a
roughly linear increase in life with increasing Young’s modulus, in Figure 7 - Life versus yield stress for all data in the temperature
accordance with what might be expected from CTOD theory and range 550-650 “C.
from experiment [22], while the simple models show life to be
independent of E, as would be expected from the low relevance
assigned to it.

413
 A Type-ii
q Low E
1.5 0 relevance
. Type-i
-0.2 q 8
A q
A 0.
-0.3 A u
A 0
A q

600 700 800 900 1000 1100 1200 200 300 400 500 600 700 800
Yield stress (MPa) T (“C)
4
-0.25

-0.3

a
.s -0.4
1
2 -0.45

-0.5
J . , Typ,e-i

160 170 180 190 200 210 220 230 1100 1200 1300 1400 1500 1600
4 E (GPa) UTS (MPa)
4
Figure 8 - Predictions of life generated using a database containing UTS for (a) varying yield stress
(b) varying temperature (c) varying Young’s modulus (d) varying UT.9

Refining the data base an empirical relationship between ‘hardenability’ and the ratio of
ultimate tensile strength (UTS) to yield stress. Materials with a ratio
The model produced so far has been somewhat simplistic, with only greater than 1.4 are proposed to cyclically harden, those with a ratio
three input variables used, and while some success has been less than 1.2 to cyclically soften, and those with a ratio between 1.2
obtained it was thought useful to add further parameters to see and 1.4 to be cyclically stable. Therefore UTS was added to the
whether network performance improved. However, collecting data database as an input to the network to provide a measure of
can be problematic as in order to use a given dataset it is necessary deformation characteristics.
to have values for all the inputs. It would have been informative to
look at grain size and other microstructural parameters, as Stage II The addition of UTS as an input produced three types of low noise
fatigue is expected to be microstructurally sensitive to some extent models, which had comparable noise levels and training and test
[20]. Unfortunately, microstructural information is difficult to energies to those produced previously. In this instance a number of
obtain-from reports, papers and standards to a sufficient degree of the models (about half) assigned a low relevance to E. Of those that
accuracy to be useful in the network. Hence material property data did not, again the best six models were chosen, ranked by test
and such data as are calculable from thermodynamic phase error, and were then compared with a model with low relevance for
calculation programs such as MTDATA are used as inputs, as these E. When predictions of variation in life with yield stress are made
am more readily available/determinable. (using a mean UTS value of 1360 MPa), it may be seen that the
models fall broadly into two groups, those that predict a large drop
in life with yield stress increasing from 700-800 MPa (type-i) and a
group which indicates a more linear dependence of life on yield
Accountine for cvclic hardening. stress (type-ii) (figure8a). Above 800 MPa the two sets of models
agree reasonably well. While the type-ii models at first seem
In equation 3, the yield stress in the denominator is that at the crack promising, examination of predictions for temperature and Young’s
tip. This may or may not be similar to the bulk yield stress, as modulus calculation cast doubt on their validity (figure Sb &c). The
repeated yielding of material in tension and compression may cause prediction of temperature dependence shows a different form from
cyclic hardening or softening. Hence the cyclic yield stress may previous models, and indicates a lessening of the effect of
differ from the bulk yield stress. Data for cyclic yield stress are not temperature as temperature increases, whereas observations on
readily available. However Manson and Hirschberg [23] proposed increased dislocation mobility with temperature would indicate an

414
increasing effect of temperature as demonstrated by type-i. Addition of NV to the database again produces two types of
Dependence of life on Young’s modulus for type-ii models is the complex model, similar in form to those described above (e.g.
reverse of that predicted by CTOD theory and experimental figure 9, Nv=2.27). The relevance attributed to NV is quite high,
observation. and life is predicted to increase roughly linearly with NV by both
models (figure 10). The difference in life is not predicted to be very
The dependence of life on UTS is shown in figure 8d. The type-i large as the range of NV values in the database is quite small. This
models uredict an evident but modest deuendence of life on UTS. indicates that, although there is an effect of NV on fatigue
indicating some importance of UTS,&although it seems less resistance, the benefits obtained by trying to alter NV within the Ni-
important than temperature and E. The type-ii models however based alloy system would not be great as this class of alloys tends
indicate a strong dependence of life on UTS below about 1250 to exhibit a restricted range of NV values.
MPa, above which value the two types of model agree reasonably
well. When the values of predicted life are examined it is noted that
below 1250 MPa the values fall well outside the range of the dataset
(-0.75 as opposed to a limit on the data of -0.5). Again it is
possible that the model is trying to extrapolate too far from the
known database, and so such predictions should be regarded with
care, as in the case of the yield stress predictions described above.
Examination of the database reveals that at 600 “C the alloys with
low UTS also have very low Young’s modulus (about 170 GPa as
opposed to the value of 195 GPa used for predictions).
Of the three models, type-i best reflects the actual trends in the data.
The existence of two complex models with low noise and test
energy indicates that within the experimental noise in the data there
is more than one way to mathematically model the data. However,
only one of these models makes physical sense. This difference in
prediction by two sets of models which are ostensibly similar
shows the importance of thoroughly examining a neural network
model in the light of metallurgical knowledge.
205 210 215 220 225 230 235 240 245
NV
The effect of allov instabilitv
Figure 10 - Prediction of variation of life with NV
It has been observed that some alloys which are unstable to the
formation of the detrimental sigma phase have an increased
resistance to fatigue crack propagation [24]. It is possible to
express this instability in terms of an electron valency number (NV) Summarv and Conclusions
of the matrix, which can be readily calculated from alloy
composition [25]. This parameter, NV, was used as an input to the A possible limitation on use of neural networks as predictive tools
network in order to investigate this dependence. NV has the for design purposes is the amount of data required - large numbers
advantage, along with yield stress, of being a means of entering of inputs require large, complete datasets which are difficult to
compositional information without needing to enter the composition compile. However success has been achieved in modelling
itself (which would vastly increase the input dimensionality and Stage II fatigue crack growth behaviour in Ni-based superalloys.
hence the number of datasets needed to produce a robust model). A basic dataset containing temperature, yield stress and Young’s
modulus models trends adequately, while refining the database to
0 include ultimate tensile strength improves the performance.
n.
-0.1 0.
The importance of careful construction of a database has been
00.
stressed, in order as far as possible to cover areas OFinput space
that may be interesting, as extrapolative predictions must be treated
-0.2 0 with caution. Experimental results that may be of use in a database
0 . n
a? 0 0 should be carefully logged to ensure that relevant inputs, such as
3 . microstructural information, are recorded accurately. It has been
z -0.3 0
0. l!i seen that it is important to validate any neural network model
produced to ensure that its predictions are in line with experience.
B -0.4 o*P**
r2 It has been shown that it is possible to use a neural network to
-0.5 investigate and model trends in fatigue crack growth behaviour with
0 UTS type-i ,” variation in material properties, based either on proposed
0 UTS type-ii mechanisms or on observed empirical trends. A trend for increased
-0.6 . Nv type-i 0 Stage II fatigue life with increased instability to sigma-phase
. formation (indicated by electron valency number, NV) has been
8 NV type-ii .
-0.7 shown to exist, although the observed effect is small due to the
small range of NV in Ni-based superalloys.
200 300 400 500 600 700 800
T ("Cl While further work on validation of such networks is required
before they can be used in design, if such work is done neural
Figure 9 - comparison of predictions using data-base containing networks may prove to be of great benefit in future alloy design
UTS and database with NV added programmes.

415
 Acknowlednements 15 J.N.Hall, J.W.Jones & A.K.Sachdev, “Particle size, volume
fraction and matrix strength effects on fatigue behaviour and
Professor C.J.Humphries is thanked for provision of research particle fracture in 2124 aluminium-SiCp composites”, &&
facilities at the University of Cambridge. Rolls-Royce plc are Sci. & Erie., Al83 (1994), ~~69-80
thanked for financial support for this work and for the provision of
data. Thanks are also due to Joy Jones of the University of 16 A.J.McEvily, “Current aspects of fatigue”, Metal Sci., 1977,
Cambridge for helpful discussion and advice. ~274

17 R.J.Donahue, H.McI.Clark, P.Atanmo, R.Kumble &

A.J.McEvily, “Crack opening displacement and the rate of
fatigue crack growth”, I. J. of Frac. Mech., 8 (1972), pp209-
219
S.R.Holdsworth & W.Hoffelner, “Fracture mechanics and 18 D.Broek, Elementan, Enzineerinz Fracture Mechanics, Chapter
crack growth in fatigue”, Hiah Temuerature Alloys for Gas 9,Klewer Academic, 1989
Turbines, R.Brunetaud et al eds., DRiedel(1982), ~345-368
19 J.F.Knott, “Models of fatigue crack growth”, Fatigue Crack
P.D.Wasserman, Neural Comnutina Theorv and Practice, Growth - 30 Years of Proaress, R.A.Smith ed., Pergamon
Chapter 1, Van Nostrand Reinhold, 1989 1984, ~~31-52

R.P.Lippman, “An introduction to computing with neural nets”, 20 S.D.Antolovich & N.Jayaraman, “The effect of microstructure
IEEE Acoustics. Sueech and Signal Processing Maaazine, 4 (2) on the fatigue behaviour of Ni-base superalloys”, Fatigue -
(1987), ~~4-22 - Environment and temuerature effects, J.J.Burke & V.Weiss
eds., Plenum 1983, ~~119-144
J.A.Powell, “Preface: Neural networks, neuro-fuzzy and other
learning systems for engineering application and research”, 21 J.E.King, “Fatigue crack propagation in nickel-base superalloys
Neural Networks, Neuro-Fuzzy and Other Lear&z Systems - effects of microstructure, load ratio and temperature”, Mat.
for Enaineering Anulications and Research Conference, Institute Sci. & Tech., 3 (1987), ~~750-764
of Civil Engineers, London (1994), pp.i-iv
22 M.O.Spiedel in High temnerature materials in gas turbines,
T.J.Stonham, “Neural networks, how they work, their M.O.Spiedel & P.R.Sahm eds., Elsevier 1974, ~~207-251
strengths and weaknesses”, Neural Networks. Neuro-Fuzzv
and Other Learnine Svstems for Engineering Annlications and 23 S.S.Manson & M.H.Hirschberg, “Fatigue behaviour in strain
Research Conference, Institute of Civil Engineers, London cycling in the low- and intermediate-cycle range”, Fatigue - an
(1994), ppl-7 interdiscinlinarv anoroach, J.J.Burke, N.L.Reed & V.Weiss
eds., Syracuse University Press 1964, ~~133-178
J.Jones, “Neural network modelling of the tensile properties of
Ni-base superalloys” (CPGS dissertation, University of 24 A.James, Personal communication, Rolls-Royce, 1995
Cambridge, 1994)
35
-- J.P.Michalisin & D.L.Pasquire, “Phase transformations in Ni-
H.K.D.H.Bhadeshia, D.J.C.Mackay & L.-E.Svensson, base superalloys”, Proc. Conf. on Structural Stabilitv in
“Impact toughness of C-Mn steel arc welds - Bayesian neural Superallovs, Seven Springs, PA, Sept 1968, ~~134-170
network analysis”, Mat. Sci. and Tech, 11 (1995), pp1046-
1051 26 S.Suresh, Fatigue of Materials, Camb. Uni. Press (1991),
~~255-258
A.Mukherjee, S.Schmauder, M.Rtihle, “Artificial neural
networks for the prediction of mechanical behaviour of metal 27 Rolls-Royce design database
matrix composites”, Acta Metall. Mater., 43 (1995), 11,
~~4083-409 1
U.Bork, R.E.Challis, “Nondestructive evaluation of the
adhesive fillet size in a T-peel joint using ultrasonic lamb waves
and a linear network for data discrimination”, Meas. Sci. &
Tech., 6 (1995), ~~72-84

10 T.M.Edmunds, “Fatigue crack propagation program SA16”

(Rolls-Royce in-house package)
11 D.J.C.MacKay, Biaback User Manual, 1994
12 L.A.James & W.J.Mills, “Effect of heat-treatment and heat-to-
heat variations in the fatigue-crack growth response of Alloy
718”, Enc. Fract. Mech., 22 (1983), pp797-817

13 C.Laird & G.C.Smith, “Crack propagation in high stress

fatigue”, Phil. Ma%,, 7 (1962), ~~847-857
14 K.Tanaka, T.Hoshide & NSakai, “Mechanics of fatigue crack
propagation by crack-tip plastic blunting”, Enz. Fract. Mech.,
1 Q 11984‘1 n8Wi

416