ISIJ International, Vol. 36 (1 996), NO. 11, pp.

1373-1 382

Bayesian Neural Network Analysis of Fatigue Orack Growth Rate

In Nickel Base Superalloys

Hidetoshi FUJll. D. J. C. MACKAY1)

and H. K. D. H. BHADESHIA1)
Department of Materials Science and Metallurgy, University of Cambridge, PembrokeStreet, Cambridge CB230Z, UK.
1)Darwin College, Mathematics and Physical Sciences Group, Silver Street, Cambridge CB39EU, UK.
(Received on Mayl3. 1996, accepted in final form on September9. 1996)

The fatigue crack growth rate of nickel base superalloys has been modelled using a neural network model
within a Bayesian framework. A 'committee' model was also introduced to increase the accuracy of the
predictions. The rate was modelled as a function of some51 variables, including stress intensity range AK,
log AK, chemical composition, temperature, grain size, heat treatment, frequency, Ioad waveform, atmosphere,
R-ratio, the distinction between short crack growth and long crack growth, sample thickness and yield
strength. The Bayesian method puts error bars on the predicted value of the rate and allows the significance
of each individual factor to be estimated. In addition, it was possible to estimate the isolated effect of
i i
particu lar variables such as the grai n size, which ca nnot n practice be varied ndependentl y. This demonstrates
the ability of the methodto investigate newphenomenain cases where the information cannot be accessed
experimentally.

KEYWORDS:
fatigue cracks; neural network; nickel-base superalloys; modelling.

the estimation of the outputs is very rapid.

1. Introduction
A difficulty with neural network models is that an
Superalloys have been used in aeroengine gas turbines over-flexible network can 'discover' false structure in
for about 50 years. Their weight percentage in engines data, a behaviour knownas 'overfitting'. have used We
has increased over the years to 300/0. The safe service a Bayesian framework4) in order to control this prob-
lives of these components are increasingly becoming lem. This mitigates against excessive complexity and
fatigue limited due to the high operating stress.1) In- permits the calculation of error bars, indicative of the
deed, Iow cycle fatigue performance had becomethe reliability of the prediction.5) In addition, in this study
life limiting factor in over 750/0 of major structural a committee model is introduced to make predictions
components advanced engines by 1975.2) Fatigue in
of more accurate. These approaches, combinedwith metal-
turbine discs arises from the variations in both thermal lurgical theory have been used here to model the
and mechanical stress during the fiight. A
typical loading fatigue crack growth rate in nickel base alloys.
cycle comprises starting up, takeoff and climb, cruising,
landing and shut-down. The highest stresses are ex- 2. Neural Network Framework
perienced in the bore of the disc early in the flight
Figure I shows the structure of the neural network
cycle, generally while it is in the lower temperature range
used in our model. Factors such as chemical composition,
200300'C. Stress in the rim region is lower, but at a grain size, heat treatment and temperature are input from
higher temperature, 500600'C.3)
Thus, the requirements for fatigue crack propagation
the left hand side. To predict the output, that is, the
logarithm of the fatigue crack growth rate, hidden units
are dependent on the environment where the material is
used. Full quantification of fatigue crack growth rate is were used between the input and the outputs so that
essential in both life prediction and in the development more complex relationships could be expressed. The
of new materials. However, the fatigue propagation is
che[nicill
affected by manyfactors including chemical composition,
composition
grain size, heat treatment, temperature, atmosphere,
R-ratio and frequency. Accordingly, the modelling of grilin size- fatigue crzlck
fatigue has to cover a wide range of conditions, and it growth rate
is not
easy to predict the fatigue crack growth rate of
OLltpLlt
an unknown material. In this study, in view of the
complexity of the phenomena,neural network models
were applied in place of the usual regression analysis ~ input
or physical models. Neural networks are capable of
realising a variety of non-linear relationships of con- hidden units
siderable complexity. Oncea network has been trained, Frg. l. Neural network model used in this study.

1373 @1996 ISIJ

 ISIJ International. Vol, 36 (1996), No. 11

transfer function relating the inputs to the ith hidden Water Quench: Oil Quench: Air Cooling :
units is given by
Furnace Cooling = 60 : 30 : 10 : l
(~ w{;)xj+
= tanh~ Oi(1)
hi (1) The shape of load waveformwas defined as shownin
......
Fig. 2. A sinusoidal waveform was defined as Oand a
•

Therelationship betweenthe hidden units and the output triangular waveform as I The '10ading time', during .

is
a linear: which the maximumload is being applied, and the
y=~w{2)hi+0(2) 'unloading during which the minimumload is
time',
..........
applied, were also selected as input factors so that the
The coefficients w and biases Oof thesc equations are triangular load waveform covers also trapezium or
determined in a such way as to minimize the energy square shape waveforms. The grain size was defined by
function, as explained later. Becuase the hyperbolic three factors: two of the three are the maximum grain
function is a non-linear function, a non-linear relation- size and the minimumgrain size in the distribution.
ship can be predicted in this model. The microstructure is sometimes inhomogeneous, con-
taining someregions with a fine grain size and others
3. The Data Base where the size is coarse. The difference between the
Thedata base consists of 1894 combinations of fatigue meangrain size in these two regions in sometimesquoted
crack growth and 51 inputs including the stress intensity and is used here as an input. The data for short-crack
factor range AK, chemical composition, temperature, growth were also included. Short-crack growth was
defined as Oand long-crack growth as I Log AK is
grain size, the condition of heat treatment, frequency, .

10ading condition, atmosphere, R-ratio, Ioad waveform, included as an input factor because it has a metallur-
gical meaning and hence helps to find the optimum
sample thickness and yield strength. A11 these data are
relationship between the input and the output.
from the published literature.3'6~38) Table I shows the
details of the data used for the modelling.
Manyof the nickel base superalloys were given three
4. The Analysis
stepsofheat treatment. Eachheat treatment wasspecified Both the input and output variables were first
nor-
as a combination of temperature, time and cooling rate. malized within the range ~0.5 as follows:
Whenfewer than 3steps of heat treatment were carried x-x~,~
out, the temperature, the duration and cooling rate of xN= -0.5 ......
..........(3)

the remaining steps were set to zero. The cooling rates x~ax ~xmin

were assigned relative magnitudes as follows39): where XN is the normalized value of x, x max is the

Table l. Input variables for the prediction of fatigue crack growth rate.

variable Ran e Mean standard Deviation Valiable Ran e Mean standard Deviation
da/dN,
um 1oxl0~8 - 0,646 Ni, wt% 40- 73 55,34 8.234
tog da/dN, um -8 - -o.1898 CI o.03 -
19 5 14.89 5 127
AK, MPam~1/2 4O3-246 27.16 22,47 co o -
17 5,982 7.552
tog AK, MPam~1/2 o605 - 2,39 l,316 0,3 67 1 Mo o6 - 3.094 1,991
AI
Temperature, K 293 - I 123 660,3 304 5 0.3 - 5,5 1,926 1.732
Ti
Mlntmumgrain size, um 7 - 5000 295,8 l 024 0.8 -
3,52 2.l07 l,098

Maxlmum
grain size. um 7- 5000 313,2 1022 Fe o 35 56 - 12,07 12. 12
Difference in grain size between -35 - o -0.8936 5432 c o007 o.06 - o,03865 o.o I176
major phase and minor phase B o o, l - o,01418 o,02604

l st step H:eat Treatment, l 116 - 1578 1321 l03 o zr o 0.35 - oo1907 0.04495
Temperature. K si
o 0,3 l - o,05634 o,08841
Duration, hour 0,5 - 7 2,602 1.685 Nb o 5.35 - 1.968 2402
coohngrate, K/sec -15 - 5 -5,629 3,134 Mn o 0,28 - o.03728 o,07740
2nd step Heat Treatment, o - 1413 955 S 292. 1 Cu o o.06 - 4525 x l0-3 o.o 40 1 l
o o.ol l 8.551 x l0-4 2,438 x l0-3
Temperature. K -

Duration, hour o - 24 12 22 9065 Ca o o,006 - 2 598 l0-4 x 1'22lx l0-3

coohng rate, K/sec -5 - o -3.036 2291 Mg o o.002 - 8659 x l0-5 4,07 l0-4
1x
3rd step Heat Treatment, o I 143
- 869 3 312,0 s o o005 - 3,350 x l0-4 I'03 Ix l0-3

Temperature, K Sn o o,0027- l 169 x l0-4 5A96x ~0-4

Duration, hour o - 24 l0,96 6 179 Pb o o.oo004
- l,732 x l0-6 8, 143x l0-6

cooling rate, K/sec -5 - o -4.459 1,554 Bi o o.ooo0125

- 5.412 x l0-7 2.545x l0-6
Frequency, Hz o,ol - Ioo 2~,47 29,3~ Ag o o.ooool
- 4,329 x ro-7 2036 x 10-6
Loading Time, s o 600 - l 5,27 71 96 w o 6,5 - 0.4628 1,539
Untoading Time, s o 500 - 7439 55 14 Ta o 65 - 0,3935 1.385
Hf o o, I 4.488 x l0-3
Load shape oor I O7355 O441 2 - o.02071
Atrnosphere I x l0-6 - 760 691 4 217,9 Re o3 - o.
1346 0.6213
Y203 o 1,1 o04704 o2226
R-ratio o,05 -
0,8 o 171 0,2 175 -

Short or tong crack growth oor I 0,9 161 0,2774

sample thickness. mm 4A -25 11,39 4,063
Yield stress. MPa 324 - 1690 9119 242 3

C 1996 ISIJ 1374

 ISIJ International, Vol. 36 (1 996). No. 11

Load Shape= O

::I
ri
Load Shape= l C~
~i
~O
(/)

UnloadingTime LoadingTime O
~+1
Fig. 2.
H-~l
Definition of load shape. Numberof hldden units
Fig. 3. Variation of as a function of numberofhidden units,
maximum
value and x~i* isthe minimumvalue of each (r*,

variable of the original data. This normalization is not

Overf itted
essential to the neural network approach but allows a
convenient comparison of the influence of individual
input variables on outputs. Real function
Using the normalized data, the coefficients (weights)
w and O
bias were determined in such a wayas to minimize
the following energy function5):

M(w)= pED+ ~Eo(

* ~(.) .*...
..........(4)
Fig. 4. overestimation of function.

of the model. Thesehyperparameters define the assumed

The minimization was inplemented using a variable
metric optimizer.40) The gradient of M(w) wascomputed
using backpropagation algorithm.4 1)Theenergy function
consists of the error function, EDand regularization E~.
The error function is the sumsquared error as follws:
l y(x~;
ED(w)=
2~( ~ w)
-
t~)2
where {x~,t~} is the data set. x~ represents the inputs
and t~ the targets. The mis a label of the pair. The error
function EDis smallest whenthe model fits the data well,
i,e., wheny(x~;w) is close to t~. The coefficients w and
biases eshownin Eqs. (1) and (2) makeup the parame-
ter vector w. A number of regularizers E~(.) are added
to the data error. These regularizers favour functions
y(x;w) which are smooth functions of x. The simplest reg-
ularization methoduses a single regularizer E~ ~~ w~-
= approximation given in Ref. 42). The M(w) was mini-
Here, however, we have used a slightly more compli-
cated regularization method known as the Automatic
relevance determination model.4) Eachweight is assigned
to a class c depending on which neurons it connects. For
each input, all the weights connecting that input to the
hidden units are in a single class. The hidden units'
biases are in another class, and all the weights from
the hidden units to the outputs are in a final class. E~(*)
is defined to the
sum of the squares of the weights in
class c.42)
l
E~(.)(w)=-~w~
2 ie* •••••
This additional term favours small values of
w and
decreases the tendency of a model to 'overfit' noise in
the data set. The control parameters oc, and together
with the numberof hidden units determine the complexity
Gaussian noise level a~= l/p and the assumedweight
variances a~(.)= 1/0((.)'(T~ is the noise level inferred by
the model. The parameter oc has the effect of
encourag-
ing the weights to decay. Therefore, a high value of
a~
implies that the input parameter concerned explains
a
relatively large amount of the variation in the output.
.... . ...
..(5) Thus, (T~ is regarded as a good expression of the
significance of each input though not of the sensitivity
of the output to that input. The values of the hy-
perparameters are inferred from the data using the
Bayesian methods of Ref. 42). In this method, the
hyperparameters were initialized to values chosen by
the operator, and the weights were set to small random
initial values (Gaussian with
mean Oand standard
deviation O.3). The objective function M(w) was
minimized to a chosen tolerance, then the values of the
hyperparameters were updated using a Bayesian
mized again, starting from the final state of the
previous optimization, and the hyperparameters were
updated again, repeating 8times.
As the numberof hidden units increases, the difference
((T~) between predicted values
and experimental values
decreases monotonically, as shownin Fig. 3. More
com-
plex relations can be modelled with a larger numberof
hidden units. However, the function maythen be over-
fitted, as shown in Fig. 4, because experimental data
always contain errors. In order to reduce overfitting,
••••••••••(6)
the test error (the value of the error, function for
a
non-trained data set) wasmeasured, using 942 randomly
chosen rows of data which were not included in the
training set. Figure 5
shows the change in the test error
p as a function of the numberof hidden units. There is a
minimumat 17 hidden units. The increase in the test

1375 C 1996 ISIJ

 ISIJ International, Vol. 36 (1996), No. 11
1.8 03 e
1.6 ' o298 e
0.296
14
e
12 o29•i ,
~ ~
~ o
O
o . 292
~
~ l ~
Q) ~:~ 0.29 e
~' 0,8
c"
e
Cl,

~H
Q)
e
Q)
~4
o 288
0,6 '
o 286
04-
0.2-
, , , I l o284
o282
O o 28
o 2
O 6 8 10
,1
Io 12 14 16 18 20
Numberof hidden units Numberof models in committee
Fig. 5. Variation of test error as a function of number of Fig. 7. Changein test errorwith the numberofmodels in the
hidden units. committee.

O experimental conditionsl3'14) are summarized in the

3rd than 40 MPa
-1
first column in Table 2.). WhenAKis less
ml/2, the difference in the predictions is
very small, but
O -g when AK more than 40MPami/2, the
is difference is
~ 2nd .!'! ~
large. In the latter region, the error bars are also large,
\3
~~ indicating that there are insufficient or imprecise data in
~4 /i;!~;~.'
l Ist that region of input space. The error bars in this case

~ ~~
.A are for 67 "/, confidence.
A
le\,EI

~5 'committee' model wastherefore introduced in order

~:' to see whether more reliable predictions to be made. The
tLOO ~6 methodis as follows:
-1 (1) The individual models are first ranked via their
~7
-8
~ test errors.
(2)A committee of N models is then formed by
4 7 ro ao
m~)
40 70 ioo combining the best Nmodels, where N=l, 2, 3 • • •

AK (MPa
.

Fig. 6. comparison ofbest, second bestand third bestmodd.

The meanprediction of the committee is ~
I N
error over 17 hidden units indicates that the function y_ = N i= yi ~i •• ...
••••••••••(7)

maybe overfitted. However, the increase is very small.

This indicates that the distribution of data is close to the and the associated error in
~ is given by
assumedGaussian distribution and Bayesian modelling
IN l N
~ N~
worked well. In principle, whenthe Bayesian modelling a2 (T~
+ (yi-y) .,........(8)
= Ni=1 i=1
is completely optirnized, and infinite numberof hidden
units could be used without overfitting.43) Figure 7showsthe decrease in test error by combining
models. A committee of seven of the best individual
5. Committee Model models has a minimum test error. This committee
therefore was used to make a number of predictions,
The same data can be modelled in manyways, for
as shownlater. Using the committee model, the agree-
exampleby varying the numberof hidden units or starting
value of (T~. The variety of models thus produced can be ment between experimental data and prediction is very
high for both the training and test data, as can be seen
ranked according to the magnitude of the test error. The
in Fig. 8.
best individual modelwould then have the minimumtest
error. However, it is possible in principle to reduce the 6. Significance of Individual Inputs
test error further by using the average of predictions from

a numberof models, i.e. a committee of models. The neural network model allows us to estimate the
The test errors of the variety of models produced are significance of individual factors in infiuencing the fatigue
not very different from that of the best model, as can be crack growth rate using the value of a*. A high value of
seen in Fig. 5. This indicates that these models should (T~ implies that the input parameter concerned explains
lead to similar predictions. However, for somechoices a relatively large amountof the variation in fatigue crack
of input variables the predictions nevertheless
are growth rate in the data set. Note that it is not an
different because of the limitations data.
in the training indication of the sensitivity of fatigue crack growth rate
Figure6 shows the predictions using the three best for that input parameter.
models for Astroloy at room temperature (The main Figure 9 shows the values of (T~ of the seven best

C 1996 ISIJ 1376

 ISIJ International. Vol, 36 (1996), No. 11

(a)
models. Although somescatter is observed, somestrong
0.4 tendencies are found. For example, it can be found that
Training dataset
10g AKIs clearly more strongly linked to fatigue crack
"r'
':":~:i'
growth rate than AKon its own. This is coincident with
Q)

~,
~ 08 the Paris' Iaw.44) As expected, the fatigue crack growth
,e
rate was found to be sensitive to the frequency, Ioading
>
1:'e) time, atmosphere and yield strength.
0.0
~' ';;'~~a'r'
O
.~ 7. Effect of Individual Inputs
1:, ~~!Jb
a;
h,
D
p~ ~O.~ The neural network permits the effect of each input
~',r• '
to be examined individually, which maybe impossible
to do in practice. Figure 10 shows the estimated fatigue
-0.4
crack growth rate of Astroloy at room temperature; the
experimental data are from the published literature, 13' 14)
-0.4 0.2 0.0 0.2 0.4
summarized in Table 2. The error bar is for 67010
Experimental value confidence the sarne as before. As expected,15) the crack
(b)
growth rate decreases whenthe grain size increases, since
that generally leads to more heterogeneous slip.45) As a
0.4 Test dataset
'~J" reference, although the fatigue crack growth rate seems
::J~::~~i:I~:/1:1::' constant near the threshold region of the 40-50 ptm grain
G;
alloy because of the error bar, the most probable value
~ 02
-1
c5 of the prediction substantially decreases around the
>
l:~G,

0.0
region whenthe AK
is reduced.

~' "'~~:;:'
t~" t'
In practice, changes in the grain size are usually
c,
.H
1:i
.~;'i' T"-
achieved by heat treatment. This was the case for the
a;
~1 experimental data shown in Fig. lO. Of course, heat
p~ ~O.,.D
treatment mayaffect other features within the material.
"~:~~'r"
In fact, In addition to increasing grain size, the heat
-0.4 treatmentl3,14) reduced yield strength from 021 to I
954 MPa. In the neural network model, the grain size
-0.4 -0.2 0.0 0.2 0.4 alone can be changed, as shown in Fig. 11, without
Experimental value altering any of the other inputs. Figure is in this
re- II
Fig. 8. Comparison between predicted and experimental spect proof that an increase in grain size causes a re-
fatigue crack growth rate. duction in the fatigue crack growth rate.
(a) training data set (b) test data set In fact, the crack growth rate increases whenonly the

4
l
3.5 e
O
3
2.5
~
A
~l
~~
2
~O I
~5 ~.i
1.5
A~
1 ..
~1 E~l
'
~
e Jt
[]
o I~ ~e~:'~
0.5 !i~ I ~o A~A e~ ~'^~~fu~~ •5L
~~ !e e~
o
~ i et , ! ' ~~~
.^

~ ~~;OO~:~:~~~O~:I~~5-~~~P'~~:e~/)~~i~~
e"- h o o
~~~~
~~~(i (L' E:: e~ o'H f::~~c\! F::*e'cf) ~~~ ha' (u cT~- a'cr)p::* "'~:s

~~;~ ~~"~ o
h'-'-
~$:~1Se'l ~~~~ ,,$~;.o
a)*'*,:$A:i
,:s~.o ~'
1)~'*c~JP:i G';:~i::~S~::~~
H ~~e0~1~ee~~bQ:~
1~5~S'~(~)(J)
~ (1)~
O~q'
~~~
~~~~ ~~~~ (~~~a~~~
F' f:{~;~ F~ h.-*
o~:{(1)
of~~'1:' i~~
' ' '

h
H .*~
$:::h
OH ~H ~ ~I~~~ ~~1~
~~
~
'~$a~
cL)
*4),:' ~ '~s
'l) :~']
~]
O~:: ~:: ~:~

Fig. 9. Significance of individual factors on fatigue crack growth rate.

1377 C 1996 ISIJ

 ISIJ International, Vol, 36 (1996), No. 11

Table 2. Main experimental condition inputted for the -3

prediction of Fig. 10.

l 1-13 um 40-5O um /
1377K, 4h, 1423K, 4h,
G)
0>)
4 ll //
923K, 24h,
l033K, 8h,
923K, 24h,
1033K, 8h,
~~-5 YS = 954Ma
!1-,

293K 293K ~ YS = 1021MPa

Air Air ~;
Tl -6
O.
1 Ol \ce

40HZ 40HZ 1:f

tO
i 021MPa 954MPa O -7
AC: air cooling

-3 -8
4 6 B 10 i
20 40
AK (MPa m~)
O -4 .~; Fig. 12. Effect ofyield strength alone on fatigue crack growth
~ rate in Astroloy

~~~5 -3
~ 11-13
um //:
~;
1:' -6 40-50
a; -4
\ce um e'>1
d;
~:f

ttOO

-7 \~-5 f
~
~;
-B ~:,
-6 11-13 ~~:

4 6 8 10 i
gO 40
\ca
um
AK (MPa m~)
~:f
' //+
ttO
40-50 Ilm
Fig. lO. Effect of grain size on fatigue crack growth rate in O -7 ll
Astroloy.

-3 -8
4 6 B Io I
ao 40
AK (MPa m~)
,D -4 '// Fig. 13. Effectofheat treatmentalone on fatiguecrack growth
C,h
rate in Astroloy.

\~-5 Thus, when the yield strength is reduced, the fatigue

~ ll-18 ,~m
crack growth should increase. This relation indicates that
~; the effect of yield strength alone, shown in Fig. 12, is
1:,
-6 ./
40-50 /Im
\ce reasonably predicted.
~::f
The case where only the heat treatment is changed
~O
O -7 without altering the grain size or yield strength is
illustrated in Fig. 13. Thus, any changes in the crack
growth should indicate the presence of other factors
-8 affecting fatigue crack growth. There is
4 6 8 ro 1
ao 40 a clear change
in the near-threshold region which is the part most
AK (MPa m~)
sensitive to microstructure.
Fig, Il. Effect ofgrain size alone on fatigue crack growth rate
in Astroloy. The effect shown in Fig. 13 might therefore be due
to the higher solution treatment reducing coarse y'
value the yield strength is changedfrom I 021 to 954 MPa
(Fig. 12). In the Paris regime, it is considered that fatigue
particles which tend to concentrate stress.13) Our model
does not include direct microstructure because such data
behaviour is dependent on crack-tip strain range, or the
have not been reported frequently. This would be a good
range of crack opening displacement Adper cycle.46)
area for future work.
AK2
Ad=QCF*yE "'
..........(9)
8. Other Predictions
Figure 14 showsthe rate of the short-crack growth. It
where
stress
Q constant and
is

(though, monotonic
(T.y should be the cyclic yield
(Ty is often used as an ap- can be seen that the slope of the curve of short-crack
proximation). 3)
Accordingly growth is very small and the short-crack growth rate is
da AK2 muchhigher than the long-crack growth when AKis
dN
ac Qa*yE ' '
..........(10) small.47) These predictions are also in accordance with

our experience.

C 1996 ISIJ 1378

 ISIJ International, Vol. 36 (1 996), No. 11

3 /
-3

l~
l~a) -4 l' a, -4
rl

\~
~ short orack
C,h

923K i
~-5 ---~ -
-';~~-"/'~/ \~-5
~ "~:-
~
~; ~: 93K
'e -6 tong crack
~:, -6
\ \ccf

V
cd
1:' ~'
too t~Oo

-7 _7

-B -8
4 6 8 iO 1
20 40 4 6 8 10 l
20 40
AK (MPa m~) AK (MPa m~)
Fig. 14. Prediction of short crack growth rate in Astroloy, Fig. 16. Temperaturedependenceof fatigue crack growth rate
in CMSX-4.
-3
-3
O. Iwt%R~.. .*•""

~
/ /
e) -4
e)>, e) -4
~~ -5
Oh
~~
~'71

~ ~-5 / 7718 + 0'1wt%Al

~; ~ !!
1:, -6 roloy ~;
\ce ~:, -6
~::f \,O
t~O 1:f
.../
O -7 t~8
O -7

-8 y
4 6 8 ro 20 40 -8
*
AK (MPa m~) 4 6 8 lO 1
20 40
Fig. 15. Prediction of effect of addition of o,1 wt"/. Re, AK (MPa m~)
Fig. 17. Effect of addition of O,1 wt'/, Al to Inconel 718 on
One can also predict the fatigue crack rate under fatigue crack growth.
unusual conditions. The effect of adding 0.1 wt"/. to Re
is illustrated in Fig. 15. Since the error bars are large -3
because of the absence of experimental data, it cannot rli
~;
be concluded that Re really affects fatigue crack growth.
This is a good example of the safety of the predictions
,D -4 ...~ r
Oh
718
madeby the model, in that the error bars are large when ~-

the model is uncertain. \~-5

Using our model, the fatigue crack growth rate for ~ h
kinds of superalloy can be predicted. For example,
many ~;
~1 -6
718 + 0.2wt%Nb
Figs. 16-18 show the results for CMSX-4 and Inconel
718, respectively. Figure 16 shows the effect of temper-
ature in CMSX-4.The main experimental conditions
\,e
~:f

tOO
/
/
inputs are frequency: 10Hz; Ioad shape: sinusoidal;
atmosphere: air; R-ratio: O.1; single crystal and con-
-7
/
ventional heat treatment values.19) Whenthe tempera- 84 6 8 Io 1
20 40
ture is raised, the fatigue crack growth increases. The AK (MPa m~)
fatigue crack growth is determined by the cyclic-plas- 18. Effect of addition of 0,2wt"/, Nb to Inconel 718 on
Fig.
tic deformation at the crack tip. Plastic deformation has fatigue crack growth.

a strong temperature dependency; the higher the temper-

ature, the easier the plastic deformation. Accordingly, Inconel 718 is y" strengthened material.49) It is
0.1.21,22)
it is reasonable that fatigue crack growth is easier and also knownthat after ageing this alloy contains a small
the growth rate faster at a higher temperature.48) amount of spherical y'.50) Nb is knownto increase the
Figures 17 and 18 show the effect of addition of y" phase and Al increases y'. As shownin these figures,
A1 and Nb. The main experimental inputs are tem- Nb slightlytends to decrease the fatigue crack growth
perature: 923 K, grain size: 30 mm; conventional heat though Al does not have an significant effect.
rate,
treatment; frequency: 20 Hz; atmosphere: air; R-ratio: However, at this time, it was assumedthat the yield

1379 C 1996 ISIJ

 ISIJ International. Vol. 36 (1 996), No. 11
strength was constant even after the addition of these 374.
elements because appropriate data were unavailable. If 15) J.Gaydaand V. Miner: Metall. Trans., 14A (1983), 2301,
the yield strength increases due to the addition of alloying l 6) M. Marchionni and G. Onofrio: Advancein Fracture
F. Gabrielli.

elements, the fatigue crack growth would decrease, as Research, Vol. 2, ed. by K Salama. K. Ravi-Chandar, D. M. R.
Taplin and P. RamaRaos. PergamonPress. Houston, Texas,
mentioned before. (1989), 149. I
17) J. L. Yuen, P. Royand W. D. Nix: Metal/. Trans. A, 15A (1984),
9. Conclusion l 769 .

18) K. Sadanandaand P. Shahinian: J. Eng. Mate,'. Tech., 100 (1978),

(1) A neural network in a Bayesian framework was 38 1.

found to be an effective methodof estimating the fatigue 19) A. Sengupta and S. K. Putatunda: Sc,'. Metall. Mater., 31 (1994),
crack growth rate of a large variety of nickel base l 163.
20) K. Sadanandaand P. Shahinian: Met. Tech., (1982), 18.
superalloys. A
committee of models is found to give 21) M. Clavel and A. Pineau: Meta!!. T,'ans., 9A (1978), 471.
greater reliability than the best individual model. 22) M. Clavel, C. Levaillant and A. Pineau: Creep-FatigueEnviron-
(2) Theuse ofacommittee appeared to reduce the test ment-Interactions, (1980), 24.
3
error by about o/o comparedwith the best individual 23) F. Gabrielli and R. M. Pelloux: Metal/. Trans. A, 13A(1982), 1083.
network. Also, the error bars on the commiittee's pre- 24) F. Gabrielli. G. Vimercati and V. Lupinc: Environmental Effects
dictions are expected to be more reliable. on High Temperature Fatigue Crack GrowthBehaviour of Nickel
Base Superalloys, Mechanical Behavior of Materials-V, Parga-
(3) The model can be used to examine the effect of
monPress, China, (1987).
each variable in isolation. As a result, it was confirm- 25) P. Shahinian: Met. Tech., 5 (1978), 372.
ed that logAK is more strongly linked to the fatigue 26) K. Sadanandaand P. Shahinian: Eng. Fract. Mech., Il 1979), 73.
(
crack growth rate than to AK, as expected from the Paris 27)
28)
K. Sadanandaand P. Shahinian: J. Mater. Sci,, 13 1978), 2347.
K. Sadanandaand P. Shahinian: Creep-Fatigue-Environment
(
law.
Interactions, ed, by R M. Pelloux and N. S. Stoloffs, Metal Soc.
(4) Similarly, it was possible to determine the effect
of AIME, Milwaukee, Wisconsin, (1980), 86.
of grain size alone. It is confirmed that an increase in 29) S. P. Lynch. T. C. Radtke. B. J. Wicks and R. T, Byrnes: Fatigue
the grain size should lead to a decrease in the fatigue Fracl. Eng. Mater. Struct., 17 (1994), 297.
crack growth rate. 30) J. S. Cromptonand J.
W. Martin: Meta!!. T,'ans, A, 15A (1984),
1711.
(5) Using the neural network, it waspossible to reveal
31) L. A. Jamesand W. J. Mills: Eng. F,'acl. Mecll., 5(1985), 797.
the effect of heat treatment alone on the crack growth
32) R. A. Venables, M. A. Hicks and J. E. King: Proc. Int. Symp.
rate in the near-threshold regime, which is sensitive to
on Fatigue Crack Growth Threshold Concepts, ed. by D.
microstructure. This demonstrates the ability of this Davidsons. AIME, (1983), 341.
method to reveal new phenomenain
cases where ex-
33) R. A. Venables and J. E. King: Fatigue 84, (1984), 1371.
periments cannot be designed to study each variable in 34) M. A. Hicks and J. A. King: Inl. J. Fatigue, 5 (1983), 67,
35) J. L. Yuen, C. G. Schmidt
isolation. and P. Roy: Fatigue Fract. Eng.
Mate,'. St,'uci.,
8(1985), 65.
36) J. C. Healy. L. Grabowski and C. J. Beevers: Fatigue Fract.
Acknowledgements
Eng. Mate,'. St,"ucl., 15 (1992), 309.
Theauthors wish to acknowledgethe financial support 37) B. A. Cowles, D. L. Sims. J. R. Warren and R. V. Miner, Jr,:
of the Engineering and Physical Sciences Research T,'ans. ASME,102 (1980), 356.

Council and of Rolls-Royce plc. Theywould like to thank 38) B. Lawless, S. D. Antolovich. C. Bathias and B. Boursier:
Professor C. J. Humphreysfor the provision of research Fracture: Interaction of Microstructure, Mechanisms and
facilities.
Mechanics, ed. by J. M. Wells and J. D. Landess. Metall. Soc.
AIME, Los Angels. California, (1984), 285.
39) H. P. Degischer. W. Hein, H. Strecker, W. Wagnerand R. P.
Wahi: Z. Metallkd,, 78 (1987), 237.
REFERENCES 40) W. H. Press, S. A. Teukolsky, W. T. Vetterling and B. P. Flannery:
l) G. W. Meetham:Metall, Mater. Technol., 14 (1982), 387. N~lmerical Recipes in C, 2nd Ed., Cambridge University Press,
2) T. Nicholas and J. M. Larsen: Fatigue: Environment and 418.
Temperature Effects, ed. by J. J, Burke and V. Weisss, 27th 41) D. E. Rumelhart. G. E. Hinton and R. J. Williams: Nature, 323
SagamoreArmy Mater. Res. Conf., Plenum Press, NewYork. (1986), 533.
(1980), 353.
42) D. J. C. MacKay: Neura/ Computation,
4 (1992), 448.
3) J. E. King: Mater. Sci. Technol., 3(1987), 750. 43) R. M, Neal: Radford Thesis 'Bayesian Learning for Neural
4) D. J. C. MacKay: Network.• Computation in Neural Systems, 6 Networks', Dept. of ComputerScience, Univ. of Toronto, (1 995).
(1995), 469.
44) P. C. Paris and F. E. Erdogan: J. Basic Eng. (T,'ans. ASME.
5) D. J. MacKay:Neura/ Computation, 4 (1992), 415.
C.
D), 85 (1963), 528.
J. E. King, R. A. Venables and
6) M. A. Hicks: 6th Int. Conf. on 45) B. A. Lerch, Jayaraman: Mate,'. Sci. Eng., 66 (1984), 151.
Fracture (ICF6), ed. by S. R. Valluri, D. M. R. Taplin, P, Rama 46) J, F. Knott: Fundamental Mechanics, Butterworths, London,
Rao, J. F. Knott and R. Dubeys, NewDehli, India, (1984), 2081. (1973), 240.
7) L. Grabowski and J. E. King: Fatigue Fract. Eng. Mater. Struct.,
15 (1992), 595.
47) S. Suresh and R. O. Ritchie: Int. Met. Rev., 6(1984), 445.
48) M. Klesnil and P. Lukas: Fatigue of Metallic Materials, Elsevier,
8) M. R. Winstone, K. M. Nikbin and G. A. Webster: J. Mater. Czechoslovakia, (1980), 130.
Sci., 20 (1985), 2471
. 49) D. F. Paulonis, J. M. Oblak and D. S. Duvall: Trans. Am. Soc.
9) J. L. Yuenand P. Roy: Scr, Metal/.,19 (1985), 17. Met., 62 (1969), 61 l.
lO) S. P. Lynch, T, C. Radtke, B. J. Wicks and R. T. Byrnes: Fatigue 50) R. Cozar and A. Pineau: Metall. Trans., 4A (1973), 47.
Fract. Eng. Mater, Struct., 17 (1994), 313.
l l) H. Ghonem and D. Zheng: Mater. Sci. Eng., A150(1992), 151.
12) M. Clavel and A. Pineau: Metall, Trans. A, 9A (1978), 471. Appendix
13) J. E. King: Met. Sci., 16 (1982), 345.

l 4) C. W. Brown, J. E. King and M. A. Hicks: Met. Sci,, 18 (1984), Values of coefficient (weight) and bias of the functions

C 1996 ISIJ 1380

 ISIJ International, Vol. 36 (1996), No. 11
el(1) I~'(1) (1)
of the best one model (17 hidden units). Equations (1), , rvl,1
'
Wl,51,
(2) and these values allow us to makeprediction of fatigue
crack growth. The data arranged in a continuous l](1) (1) (I )
17, wl7,l
tl . . l/'

horizontal sequence in the following order: " 17,S 1

O(2) w(12). (2)
. .
W17,

- I .4 1962 -0.01 61546 4, 21 23

1 0.257~;0~ -i ,X4335 l O')1 i)
OI 046O4 -O 7*)7..41
-o,c 1794 89 -0.0004677 .0.0859964 OI I 12s~ 1 14
0.0 34U {) i,*)L)~. 72L) [) Ol 2X502 -{)
! }7t;,Ol
0.4x7949 0.96g457 0.32747 -0.01 44 s5 -O O135372 t] (1
: '1 52L)~. O00772629 -O C)()03 421

-0,0237001 0.0021 1448 O00524809 O00377307 -0.0622692 (i l,~f)30 17 -0.0006022 O. ' 5:S9 73

o ,oOo 9
~ 82 -0.263953 -O .
0038223 *0.02443 52 -O 61 7]7 -C} :33S839_ O0205018 -0.55116g6
-0.467561 -O.675345 -O.539403 -0.2 1087 l -O.
12C]5*) -(1 2;$0655 -0.29 1287 O O7~57L),;
O.21 2356 -0.13 1'145 -O.O1 5359 1 -a.7 18597
0.585487 0.00250017 5.66266 -0.1 52835 -0.407749 -() 2r)7247 -u 040•_)3 4l () Iy 1687
-n.05~055 I~3 4
O0155749 1
-O. OO 802 -O.056 856 1 -0.062904~ 0.0327606 -{1 (tq~ S52'_, ()

.65668 -O. 1091 45 OO108458 0.0460059 -(} []1)95

41J o237 17l
{,
O{]~ ~O147
l .06604 *1

0.0266368 0.0445047 0.00678326 -O O151 224 -O.O0023_ 33 -u (]187507 -(, OOI2744 -()
1 54~. l
-0.00463 4 ~ 0.25 529 1 0,0 10923_ 4 0.0 1221 64 -0.010793S O(]t]58264 l O00534393 -O.{~O~X04}',

-0,0058925 -0.0127843 -0,0084024 0.0'_683 14 -O 034523'-, -r] 140S17 1

O. 80683 O.OX3!64~.~
-0.295983 -0,26804 O. 0062423
l O.81 4157
0.819681 -0.0198176 -O.0308398 -0.552437 0,0329082 ().
19933 -O O~ 15338 0.00X 1464,-
0.00451~_52 -0.0009641 O.0987821 1
-O. 501 89 0,0330407 -O.()524076 -0.0504001 O.OV22511
4.3 165 l -0.48 1592 -0,636487 -O .O 139726 0,01 1738 -O.ij403 776 -O O06 504 1 0.007 16756
-0.0282924 0.0884676 -0.0089834 0.0093 71 36 1
O, 66277 •O
l
OXO536 -O 59 13
OOO -O 4X3786
-O.0098449 -0.2576 18 0.0093 ~197 I
-O. 00328 -0.2'_)*)807 -r] lOgl()5 -0.01 51873 -O IL)9142
-o,
159894 -0.257499 -0.192879 1
O. 8638S 0,0645236 Ol},~6,i9 13 0.0907889 O.()(,42O72

0.155679 0.0390409 -0.240381 -0.852896

1
O. 561 34 -O.0032•-72 4.00973 -0.0275688 -0,02423 99 [I :~7g8t.' l 0.073 S56 -O 'IX-158 [
03 18523 Orj20925? -O t)X4(~638
O,o 195447 -0.00072 0.0446812 O0456867 -O V 1016lc} ()

2.03949 -1 .54607 0.2476~_ 1 -O O178264 15~,

O,rj:4 O[}it~S274X (J.013 1925 O ! ~S! (15[,

i}X9 1
-0.01 8~_567 0.05867 0.00702303 -O 00400_ O10623l -(] g53 -t).O008272 -U :*)(1L,~l

-0,024051 0.2741 13 O.00402062 -O 04x2637 -O,

171577 -(~ [}S94047 -O O•_92625 () 5246
-O I 1,
'_72 -O.
18879 -0.148521 -0.0026538 OOI_1.9257 f) f)?.1 7895 {J 02•46895 -O.(]]f!6944
0.0372302 0.5S2946
a,
194307 0,14~_ 173
-0,449032 O.O IOQ.593 -5 .26 169 -0,04953 o4 O.2226 1 Lt ~13447 -O 025~, 162 -CI
127719
-o,0295733 -0.001 1472 0,0499055 O0373259 O0277843 u ()3 19O14 -() .OO 19S•_3 -O {] ~7()
17
I ,0579*
2.1 141 3 -0.0667491 O022225i -0.000()1 1~ n(1266705 0.0049S569 -O (;
i XOOIy
-o,o 161592 -0.141771 -0.00027 17 -O OOI3464 O0038S176 •(, 140566 -O OOI7689 O(11'.,1XO~
-O O08505X l*]XOX:~
0.00 116966 O 24 2466
.
O.OI09801 -O 036 1si 3 O.
I 17476 -O [1(.g4675 -()

-o.0950246 125897
-O. -O, ll)61
15 -O 0177107 O006469~_ (1 (, [)
IX65O.~ -O C060164 O ti ~7L)~ l
0.4433 51 O.098 1107 V.0784435 O.
I 0354S
-o.C'_1 804
3 CLQ 50633 'l
-0.608918 -0.1 3773S 0.20265 5 -
!] 4V648y 1162 17
-(),O -o,
I~ s57L)
o,0141651 O. OOOI76,_7 0.0650056 0.05q 1715 .O l:]476X'_)6 -[] i03665 0.016878 () ' _'.74D4

- I .C734 3.68006 0.0892576 .0.0093 18i 1

O.O9 XX79. (1 L'
156584 OO00736]7 0.0t)4~~ I(n
-o,f*'006725 *0.139716 0.00467952 -0.00(.2 118 -0.0 115122 CJ t)L)~6472 O,,U033 128 Ol:~552'i
-o,0038647 -0.248675 -0.0 157045 O.a I '209 13 -0.253276 -{) l 195 0.02.423. 8 -O._'*~7617

-0,1935:~x *0,276131 -0.2•_2298 -0.06 1S5•_6 -0.0~_73 86() -O f)301 5Xq -l J.0314462 -O 045n.:~68

0.2463 i 'i
O.
O146 86 -0.0324843 O091 471 5
-0.09853?8 O00277~3S 2!~89:};- 0056329 -O 7X~7Iy
O438924 ().0227281 -1 ,02684 {) iJ

o,ol 15347 {).O0059105 0.0766978 O05060t)2 0.04 19074 !} OIV] ~ 0.06~_ 25g _rl rl
I 11;5(11
50
a.'_)4
11 1
.3 Ol
15 -O.20007 O0370475 -O 022131 5 ti f]
: ~.~3 5(~t) l]
0576677 -iJ l)1
;~';7(13
o.O0088 18 0.0521372 -0.002694•_ -O 0072765 1 15
-O. 5~2 (]
{! ~~~(T~~ -[1
5
f)OO683 () 1)
I sX~7~
oOI08923 1
O, 4826 -0.003~ 59 -0.09085x8 -O 2[)(,OO 1 t* '
irIXX ~ l'}
024245l -i) i~ I 147
-o,15614 -0.21 6724 O,
177241 -O 0270045 O.O2X~X55 *1 52622 f)
17683+
. -1_)
;:~t](~i.)2
O32098 0.012069•_ 0.0897762 a46. 9998
l 301 87 -0.0084515 3.5945 0.32o672 O805743 fi ~~45 ,.* {1
OI(]S2f,6 -tl
~174fl3
-o 009190~
,
r).O005 1102 O 3248
3. 0.04024 l -O OOX2465 t) i;
1
~{) 96~ ()
17O223
f) -O i iL'I5,4
o. 25
7193 -O,206896 -0.3248';4 OO0065957 0,06] 41 14 V, i )[ ~(16 X0,4 I iJ 03 731 ~,,_ OOO:~]6118
0,003671 72 -,~ 035~763 -0,0056234 -(J.0052304 -O ,O 563 46l _{i ~451 11 -6 78E-05 ~575l
() :~

o,00986222 O.
131,-207 -0.0019856 -O.
I I O205 -O,
I033 i9 -O L]762208 -O Ol 72481 -O.()~)(~:~594

-O 0877709 -0.l097g6 -0,094q. 689 1

-O. 43702 -O O~.77~38 () (1
114475 OOO935145 -t].
IX5~51
-0.4 1
-O.~_53•-74 -0.171996 O ,O 592553 006
-0.5498 1 2 -0.0290427 -0.289947 0.53 1306 l .02075 -(] i2086X O(}OS55005 -O ~]JOX7
-o.O009301 *O.OO049 0.0628477 0.001 451 '_)6 -O
OI~888y [~ f)
I042 17 089 1514
-L~
O C)(JJ~;
I i7
L,

l 3482~* *1.~1218 O297815 034886

-O -O 0573755 -()
! =' 21~.~4S I 041~~76l
-i O.(1
1~Xt)t,~
*O ,( J14 2 ')~27 0.0641 172 -0.00 21 81 1 O005 19207 O.C]05~)006 -(1 [,~t;S~27 Of)0068 113 -V ! ~,~581 l
-o,ool 73 73 -0.582812 O.CI0065073 a 040706(, -o ~2774(~ (1
116a9~ (]
0294663 •(i ~()l~71

-
o.
170969 -0.24*~)] fJl -O 198705 O. 009457
77 Oo'_,057 74 ()
I!v4 75(}_•.
OI 12789 -O t) j7.~O 14
-o,00565,44 •0.0422 77 1 -0.0351079 O. 8f,679
1
l 157](•, -0.0025694 4.c)59 33 -o r,65*1 96 -O 55.49,) [! ~(;785:~ (].03
18201 -{)O;*)4 '_4 1
-o ,o 23 555x 0.0002849 1
O. 43 52~ o.07*:;4 ~Y) I -O.023 5~73 fl [1*,.~~. 73S OO0621995 1:~055
O.
1
,'J
61 7925 *2.4096S -0.3489 93 1
o.O 531~73 0.0166L)77 C r.,'_)'_)91 OL) 0334517
-t) -O (J[)~6677

oOo~al 4h 0.0?S784~ -0.001557 -O OO;?~,~~ ao1 595t)4 -tl i4Sj2X 648E-05 O.()2~ O5 19
-o 004090~. 0.30~b32 -0.0020354 v 05Ox.1.:~ -O 0742Oi 4 (} t)~ 141 15 OO29312
,1 -O O(,:~6448
-o,o~9'7287 -O .OS~_6333 -0.061 4629 OO39~.0f,l OOOIL)~(J3•~ -i) ~(]21~~ -O 169576 O()~7~S64
0.45y68 -0.07•_3947 0.019050_ -O 022674~
o,
137609 -0.00408 12 5.
14098 -0.3 1'~~:} l -0.7~~1X_~ {]
21 l *r,O:~ r]
020u252 C) •4i)
I(J7 l

1381 C 1996 ISIJ

 ISIJ International, Vol. 36 (1996), No. 11
*0.003 881 5 -0.0007176 O. 068105 -0.0731455 0.0252t)3 3 (),fJ ~59S7q O04467 -O {J;~:~v Il
1.24723 0.256464 -0.241518 0.0228675 -0.0577381 Oi]2 197(, -1)0322g17 O.f)
I ~5r,74
0.0224943 0.05961 09 0.00388392 -0.0230036 -O 055335l Oti(13443 3 ()
OOOI7498 -O 24y794
484l
O, 0061 *O.0408 79 1 *0.0320709 -0.02526 18 -0.067863 6 -O (, ~33569 O0060419 -0.0 ~L)5984
*O.049194 *0,0755708 -O.0579398 -0.0016536 -0.03 15945 L)
i06005 -O. 1331 55 O.O(164205
*0.07923 O,
I 07823 -0.0 826431 0.80528
-O.
I03268 -0.0046591 -4.93577 0.064334 O.
199896 (;
25582 ()
O122491 O~Og244
-O.O 180949 -O.OOI 443 1 0,196136 0.0377392 -0.0026402 OO751392 ().0523 123 -0.0 XI0566
-0.990539 0.90651 1 1
-O. 71 54 -0.00093 1 9 -0.0935835 O()525959 i~
00996061 -O.
f)O6519
-0.0214085 0.030909 -0.0045416 -0.0042301 -0.0256392 -() "561 56 -{~.001 Ol 18 O!XXS18
-0.0010563 1
-O. 601 52 0.0129229 0.065821 8 0.331 48 3 Ol*)405g ().OI
16979 O~*,X72g
0.256384 OL36088 0.292048 0.0834913 0.01 1301 ('
1X2732
{1
iI.O 154206 0.0~ J~51 l
0.0434354 0.081 51 88 O.OI 17528 l .07729
0.3 14774 -0.040034 -O.381057 -O 897832 -O.2~*,672 -(J 1~4073 l)
O181756 -O ;~~328
-0.0121481 -0.0001603 O.0878642 -a I16777 -0.001 65 14 O,(;
i 2g55~ i} 00426682 -O (J~46424
0.34135 0.303801 *O,
I19733 -O 021 9934 O0663369 -O r)t)28603 l, 00513059 -O (,lJ2X 77 1
*O,029 1006 -0.247727 -0.0025908 OCJIOI 432 1
O. 91 g93 -(J 107235 (]
O0025784 .() [)~.'866

0.00366581 -0,0296207 0.00449987 0.061 8769 0.147288 rJ OL,501 66 OO1293 13 Oi.~3 803
I 17377
O. O,
159999 0.13 1126 -O.
1 348S
~_
-0.09O784 -O IS7052 -O 255689 O()~~ I078
O468245 *0.0149445 -O.
122632 -0.932455
-0.962353 0,0]20859 *3,52892 O224686 0.031 7055 fl 53007 (:1.07307'_)4
OC}~,. I05X
0.00831608 0.00040014 -0.200926 O16616 -0.0108734 -(] 83 56
(]{]1 -O . OIS03 94 O.(]-17L)X36
-2.27475 0.496536 -O,
II1228 O00753608 O.U514y3~ [: I)29527 t, 'J3670:] I -O (il]
IL)851
-0.012557 -O.
123928 O. 00788
1Ol 0.00279065 -0,0052t)75 (, 141~01 I} OOO q93l O.[ It;'
J2~26
-0.0136052 -0.229956 0.00624093 -O.cJ207847 0.20*12 17 {, f)6S77 -l]
OIqL)77~ {} Ihifl61
O.
152024 0.224609 1
O, 77026 -O,
I 11855 -0.03. f:)2492 -[,
I IS6974*i l) OO33~'.53 -V.(i I;t]725
-0.417461 -0.001 4194 O,02277 19 -0.34x89s
-0.51 9534
1.OI 953 -0.008236 5.31535 l
.0947~ [j ~5432! i)
042B326 () ~UX{]4;
-O .020C316 0.001 6621 1 -0.0648407 1
o.1 855 0.0065,)94 V.!!~77847 -o.05 17l -U 125S63
19443.
-1
. 12325 -O.
174255 0.009624gI 0,03 04073 O.()
i 220 11 (' a0176161 :~ L)Xl~-O5

9
-O ,O 04 1 14 -0.0860648 *0.00 56 17 6
-O.003 149 O0481_95 l OO~87323 -o.OOI 6565 -O OL'73863
O.O198439 O ,03 174 58 0.0275842 0.0775849 -0.056 639 1 -O 0307539 O,O 149039 -0.0~ 14144
-0.0447774 -0.0603983 -0.0504416 -0.0682791 0.00345~38 U.O()603354 -o O 147936 -O. 1~561
.

0,274506 0,280638 O.0184508 O3941 07

O.5897 0.0001 61 44 0.00532556 0.043709 -O.20 1114 -O 367367 OOl 59L)58 -t~ *~68~s
-0.03 764 11 O.bD098011 -O,
118534 -0.008083 -0.00g7X8~. O f320~. []~.* {:,
006049S6 -O ()7,,5:~ j
3,531 5•_ *0,61 4506 1 1
-O, 33 73
O0559532 -0,01. 70XLS
1 !6627-
-i:} (~ -6 Iyl~-05
4 _,,or~-o ~
-0.0041979 0,270673 *0.0004865 -0.004936 -O 013:3X84 [, :r,6]S~ i } i)006:~6 17 .(;
~~47u7
0.0145388 1
O. 32198 -0.0121099 0.009 ~9955 1~,4~.~
-0.3 [i !'64454 l O05i ~y53 -O 278s57
-O 233885 *0.343638 -0.27lg35 1
a. 81 542 0.02665 1l -1! It)0644 -O,
1155X3 f.}
l x j %92
-0.223744 0.0147336 0.022098 -rJ.
.902705
1.OI 009 -0.01 16936 5.22962 -0,509488 0.00086X73 22~X ll {) O:~4
1165 Oh27565
-O.0062006 -0.0009626 -O.0322381 -o rJ62991 0.053 ~3 1 5 -fl 54S6 tli~3 !~.O680096 OOi48~S3
2.24933 -1.3 997 O .044 87 1 6 -O .00643 56 O 021 4:96 -O (,:~~. 4a9
3 -O 027~161 -O Ot]73 881
-0.0136734 -0.091 9141 5
-O .007 1 82 O.O I'_5 108 -0,0368662 -O tl2 12276 () OOOI3078 -O.C}~XS 58 1
0.00060467 0.30398 -0.0107303 O031 91 83 -CI,
126684 -C) (J73507 f~ ,OO30
166 -O I1641 l
-OI 0216
, 1
-O. 3587 -O.
I 14262 0.04 14296 0.04903 87 () C)y
16036 O119095 -O O*) 14238
O504224
.
0,043995 -0.0076497 0.2857 19
-0,674524 6.82259 5,96923 8.572R. 4 771311 ~ :s3 (]t)2 -(• l(~183 ~J3y55
-7,22549 -7. 81 ~4 7.224 l5 3.86769 6 45772 7O2(~t, 6.L)324 l 7 35 t,c]]

-lO.l016 5.79936

C 1996 ISIJ 1382