Materials Science and Technology 11 (1995) 1046-1051

The Impact Toughness of C–Mn Steel Arc–Welds

– A Bayesian Neural Network Analysis

H. K. D. H. Bhadeshia∗ , D. J. C. MacKay∗ and L.–E. Svensson†

Mathematics and Physical Sciences Group
∗

Darwin College, Silver Street, Cambridge CB3 9EU, U. K

† Central Research Laboratories, ESAB AB
Gothenburg, Sweden

ABSTRACT

Charpy impact toughness data for manual metal arc and submerged arc weld metal samples
have been analysed using a neural network technique within a Bayesian framework. In this, the
toughness can be represented as a general empirical function of variables that are commonly
acknowledged to be important in influencing the properties of steel welds. The method has
limitations due to its empirical character, but it is demonstrated here that it can be used in
such a way that the predicted trends make metallurgical sense. The method has been used to
examine the relative importance of the numerous variables thought to control the toughness
of welds.

INTRODUCTION

Fusion welding is of the greatest importance in the fabrication of engineering structures. One
of the most important requirements for such structures, including all the welded joints, is that
they should be able to resist brittle fracture. The weld deposits therefore have to be “tough”
with a great deal of energy being absorbed by the metal during the process of fracture.
A test used to characterise toughness is the Charpy test, in which a square sectioned,
notched bar is fractured under specified conditions [1]. The energy absorbed during fracture is
taken as a measure of toughness. The Charpy test is empirical in that the data cannot be used
directly in engineering design. It is nevertheless a useful quality control test which is specified
widely in international standards, and in the ranking of samples in research and development
experiments.
The toughness of a steel depends on many variables, and that of a weld on many more
because of the complexity of the welding process. It is not yet possible to predict the Charpy
toughness of a weld with any reliability. The usual approach is to correlate the results against
chosen variables using linear regression analysis [e.g. 2]. These methods are known to be
severely limited in their application.

1
 Therefore, the most important mechanical property for welds has not been rationalised
quantitatively as a function of the complex array of variables associated with welding. However,
it is known from experience, and from the concepts of fracture mechanics, that certain variables
are more important than others in their effect on toughness. The purpose of the work presented
here was to see whether an artificial neural network [3] can be trained to predict weld toughness
as a nonlinear function of these variables, and to see whether the patterns that emerge from
the work emulate metallurgical experience.
In normal regression methods the analysis begins with the prior choice of a relationship
(usually linear) between the output and input variables. A neural network is capable of realising
a greater variety of nonlinear relationships of considerable complexity. Data are presented
to the network in the form of input and output parameters, and the optimum non–linear
relationship is found by minimising a penalized likelihood. The network in effect tries out
many kinds of relationships in its search for an optimum fit. As in regression analysis, the
results then consist of a specification of the function, which in combination with a series of
coefficients (called weights), relates the inputs to the outputs. The search for the optimum
representation can be computer intensive, but once the process is completed (i.e. the network
trained) the estimation of the outputs is very rapid. In spite of its apparent sophistication, the
method is as blind as regression analysis, and neural nets can be susceptible to overfitting.
However, much of this danger can in principle be minimised or eliminated by combining
the neural network approach with sound statistical and metallurgical theory. MacKay [4–
7] has developed a Bayesian framework for neural networks. This framework allows one to
assess quantitatively the relative probabilities of models of different complexity, and to put
quantitative error bars on the predictions of the models. We have applied this work to the
complex problem of predicting weld metal toughness.

VARIABLES

It is possible to choose a set of variables which should, using experience of welding metallurgy,
have an influence of the Charpy toughness. These variables are listed in Table 1, and described
below.
In general, the toughness decreases as the strength increases. This is because plastic
deformation, which is the major energy absorbtion mechanism during fracture, becomes more
difficult as the strength increases. Hence, the yield strength is included as a variable. The
nature of the welding process itself may have a significant effect on toughness. For example, the
submerged arc welding process is quite different from the manual metal arc welding process,

2
 Variable Range Mean Standard
Deviation
Process Submerged Arc
Manual Metal Arc
Yield Strength MPa 347–645 471 12.7
Carbon wt.% 0.029–0.13 0.08 0.004
Silicon wt.% 0.28–1.14 0.49 0.05
Manganese wt.% 0.77–2.50 1.32 0.07
Phosphorus wt.% 0.008–0.028 0.015 0.001
Sulphur wt.% 0.002–0.017 0.010 0.0005
Aluminium wt.% 0.001–0.04 0.014 0.002
Nitrogen p.p.m.w. 26–119 67 4
Oxygen p.p.m.w. 234–821 412 30
Primary Microstructure % 0–91 34 4
Secondary Microstructure % 9–100 66 2
Allotriomorphic Ferrite % 16–62 31 2
Acicular Ferrite % 11–81 55 2
Widmanstätten Ferrite % 0–35 14 2
Temperature K 213–293 259 25
Charpy Toughness J 4–215

Table 1: The variables. The abbreviation p.p.m.w. stands for parts per
million by weight.

leading to the development of different microstructures and variations in impurity content.

However, heat input per se is not included since its effect is via the microstructure [8–12],
which is included in detail in the analysis.
The major solute additions to steels, i.e. C, Mn and Si, have large effects on the trans-
formation behaviour and strength. Impurity elements (P, S, Al, N, O) are included because
of their known tendency to embrittle or because of their importance in the formation of non-
metallic inclusions in welds.
All fusion welding processes involve the deposition of a small amount of molten steel
within a gap between the components to be joined. When the steel solidifies, it welds the
components together. The fusion zone represents both the deposited metal and the parts of

3
the steel component melted during the process, and is a solidification microstructure, often
called the primary microstructure. In practice, the gap between the components to be joined
has to be filled by a sequence of several weld deposits. These multirun welds have a com-
plicated microstructure. The deposition of each successive layer heat–treats the underlying
microstructure. Some of the regions of original primary microstructure are reheated to tem-
peratures high enough to cause the reformation of austenite, which during the cooling part
of the thermal cycle transforms into a different microstructure. Other regions may simply be
tempered by the deposition of subsequent runs. The microstructure of the reheated regions is
called the reheated or secondary microstructure. The fractions of the primary and secondary
microstructures are included as input variables (Table 1).
In addition, the details of the primary microstructure are also included in the list of input
variables, since the phases involved (allotriomorphic, Widmanstätten and acicular ferrite) are
known to have a major influence on the weld properties.
Iron undergoes a ductile–brittle transition as a function of temperature [1]. The flow stress
of iron is sensitive to temperature, the strength increasing as the temperature decreases. At
some critical temperature, it becomes easier to cleave iron without expending much energy.
Below this critical temperature, the iron behaves in a very brittle manner. Hence, the test
temperature is included as an important variable.
All of these input variables should to varying degrees influence the Charpy toughness,
which is the output variable.

EXPERIMENTAL DATA

All of the data used in the analysis are from experiments conducted at the ESAB Central
Research Laboratories [13–15]. These data represent a total of 181 different combinations of
the variables listed in Table 1, from all–weld metal tests in which the joints were deposited to
ISO2560 specification as described elsewhere [11]. In this specification, the joint geometry is
such that there is a minimal dilution of the weld metal, so that experiments can be conducted
on all–weld samples. Hence, the development of welding electrodes is usually carried out using
this joint design. References 13 and 15 contain data on manual metal arc welds where as
Ref. 14 deals with submerged arc welds. All of these welds are typically classified as carbon–
manganese welds, because they do not contain deliberate additions of elements such as nickel,
boron etc.
The manual metal arc welds were deposited in the flat position using the stringer bead
technique, the parent plate thickness being 20 mm. The welding current and voltage used

4
were 190 A and 23 V respectively, the weld consisting of some 27 runs with 3 runs per layer
deposited at a speed of 0.004 m s−1 . The interpass temperature was typically 250 ◦
C. The
electrodes used were 4 mm in diameter.
The submerged arc welds were fabricated to the same geometry, but with a variety of
proprietary wires and flux combinations. The wires were of 4 mm diameter (OK Autrod 12.10,
12.22 & 12.32) and the fluxes included OK Flux 10.61, 10.71 and 10.81. The terms in italics are
trade marks of ESAB AB – there is nothing special about these consumables but is worth noting
that they are all commonly used in welding mild steels. The significant welding parameters
are listed in Table 2; the interpass temperature was about 250 ◦ C in each case.

Current Amps Voltage No. of Beads Welding Speed m s−1

450 30 14–16 0.007
550 30 12 0.007
650 30 10 0.007

Table 2: Welding parameters for the submerged arc welds.

The tests have been carried out over a number of years with the systematic measurement
of all of the important parameters discussed earlier. However, the fraction of primary and
secondary microstructure was not measured for the data in [14]. These fractions were therefore
calculated as described in [9,16].
The routine method for quantitative metallography has been described previously [11],
and the mechanical property measurements were carried out to ISO2560 specification. The
Charpy value used in the analysis represented the mean of between three to five tests.

ANALYSIS

The analysis was conducted using variables normalised between +0.5 and -0.5; this normali-
sation is not essential to the neural net approach but allows a convenient comparison of the
influence of individual input variables on an output. The normalisation procedure is expressed
quantitatively as follows:
x − xmin
xN = − 0.5 (1)
xmax − xmin
where xN is the normalised value of x; xmin and xmax are respectively the minimum and
maximum values of x in the entire dataset (Table 1).
The normalisation is straightforward for all the quantitative variables; however, the weld-
ing process was represented by assigning a value of +0.5 to submerged arc welding, and -0.5

5
to manual metal arc welding.
The network consisted of fourteen input nodes, a number of hidden nodes and an output
node representing the toughness (Fig. 1). The network was trained using a randomly chosen
100 of the examples from a total of 181 available, the remaining 81 examples being used as
‘new’ experiments to test the trained network.

OUTPUT UNIT

HIDDEN UNITS

INPUTS

Fig. 1: A typical network used in the analysis. Only the connections orig-
inating from one input unit are illustrated, and the two bias units are not
illustrated.

Linear functions of the inputs xj are operated on by a hyperbolic tangent transfer function:
!" #
(1) (1)
hi = tanh wij xj + θi (2)
j

so that each input contributes to every hidden unit. The bias is designated θi and is analogous
to the constant that appears in linear regression. The strength of the transfer function is in
each case determined by the weight wij . The transfer to the output y is linear:

" (2)
y= wi hi + θ(2) (3)
i

This specification of the network structure, together with the set of weights is a complete
description of the formula relating to the inputs to the output. The weights are determined
by training the network; the details are described elsewhere [4–7]. The training involves a
minimisation of the regularised sum of squared errors. The term σν used below is the framework
estimate of the noise level of the data.
The complexity of the model is controlled by the number of hidden units (Fig. 2), and the
values of the 16 regularisation constants (σw ), one associated with each input, one for biases
and one for all weights connected to the output.

6
 0.11

0.10
Sigma_nu

0.09

0.08

0.07

0.06

0.05
1 2 3 4 5 6 7 8
Hidden Units

Fig. 2: Variation in σν as a function of the number of hidden units. Several

values are presented for each set of hidden units because the training for each
network was started with a variety of random seeds.

Fig. 2 shows that the inferred noise level decreases monotonically as the number of hidden
units increases. However, the complexity of the model also increases with the number of hidden
units. A high degree of complexity may not be justified, and in an extreme case, the model
may in a meaningless way attempt to fit the noise in the experimental data. MacKay [4–7]
has made a detailed study of this problem and has defined a quantity (the evidence) which
comments on the probability of a model. In circumstances where two models give similar
results over the known dataset, the more probable model would be predicted to be that which
is simpler; this simple model would have a higher value of ‘evidence’. The evidence framework
was used to control the regularisation constants and σν . The number of hidden units was set
by examining performance on test data. A combination of Bayesian and pragmatic statistical
techniques were therefore used to control the model complexity. Four hidden units were found
to give a reasonable level of complexity to represent the variations in toughness as a function
of the input variables. Larger numbers of hidden units did not give significantly lower values
of σν ; indeed, the test set error goes through a minimum at four hidden units (Fig. 3).
The optimum parameters for one trained network are presented in Table 3; this listing
would be required in order to reproduce the predictions described, though not the error bars.
The levels of agreement for the training and test datasets are illustrated in Fig. 4, which show
good prediction in both instances. It should be emphasized that the test data were not included

7
 0.7

0.6
Test Error

0.5

0.4

0.3
1 2 3 4 5 6 7 8
Hidden Units

Fig. 3: The test error as a function of the number of hidden units.

in deriving the weights given in Table 3 (except to choose the solution displayed), so that the
good fit is established to work well over the range of data included in the analysis.

8
 Training
dataset

Test
dataset

Fig. 4: Plot of the estimated versus measured toughness; (a) training dataset;
(b) test dataset.

USE OF THE MODEL

We now examine the metallurgical significance of the results. We attempt predictions out of
the range of the experimental data used during training, and examine some aspects which
cannot be studied experimentally.
Fig. 5 illustrates the significance (σw ) of each of the input variables, as perceived by the
neural network, in influencing the toughness of the weld. The process clearly has a large
intrinsic effect, which complies with experience in that submerged arc welds are in general
of a lower quality than manual metal arc welds. Note that this is a process effect which is

9
independent of all the other variables listed. The yield strength has a large effect and that
is well established [1]. It is also widely believed, as seen in Fig. 5, that acicular ferrite has a
large effect on the toughness. Nitrogen has a large effect, as is well established experimentally
[10,17–12]. Oxygen influences welds in both beneficial and harmful ways, e.g. by helping the
nucleation of acicular ferrite or contributing to fracture by nucleating oxides.
It is surprising at first sight that carbon has such a small effect, but what the results really
demonstrate is that the influence of carbon comes in via the strength and microstructure.
Phosphorus and sulphur have only a small effect; the toughness measured was in the as–
welded condition whereas many of the classical embrittlement effects manifest themselves in
the stress–relieved condition. It is also possible that the effects of P and S are higher at
strength levels larger than encountered here. All of the welding consumables are commercially
used so that they are not expected to be embrittlement prone. Elements such as Mn and Si
do not feature greatly presumably because their effect comes in via microstructure. Fig. 5
also shows a relatively small effect of temperature on toughness, but it should be noted that
the temperature range considered is only 80 ◦
C (Table 1), and that a part of the effect of
temperature is to alter the yield strength, which is identified by the model to be one of the
important variables.

0.8
Sigma_W

0.6

0.4
0.2

0.0
Aluminium

Primary
Yield

Sulphur
Silicon

Nitrogen

Acicular
Oxygen

Allotriomorphic
Carbon

Phosphorus

Temperature
Process

Manganese

Fig. 5: Bar chart showing a measure of the model–perceived significance of

each of the input variables in influencing toughness.

The model can be used to estimate the toughness if all of the inputs listed in Table 1

10
are available. The amount of work required to accumulate these inputs is not trivial, but
the situation can be ameliorated. A physical model [11,12,16] based on phase transformation
theory can be used to predict the values of all the inputs from a knowledge of just the chemical
composition and a choice of welding conditions. This was done particularly to examine the
effects of carbon and manganese on weld toughness, given that a lot of work on these lines has
already been reported in the literature.
Fig. 6a shows data generated using the neural network but with all the inputs other than
manganese calculated using our weld model [11,12,16]. In all cases, the calculated inputs are for
manual metal arc welds with 180 A, 34 V, a welding speed of 0.004 m s−1 , interpass temperature
200 ◦ C, ISO2569 weld geometry. The manganese variations are for a basic composition

Fe − 0.07C − 0.5Si − Mn − 0.005S − 0.0025O − 0.0005N − 0.025Al

. It is interesting that the toughness at relatively high temperatures decreases as the man-
ganese concentration is increased. This upper shelf region involves ductile failure, and an
increase in strength leads to a reduction in the ductile fracture energy. The calculated yield
strength [10,12,16] increases from 403–539 MPa as the Mn concentration is changed from 0.5-
2.0 wt.%. However, the cleavage toughness at low temperatures clearly increases with Mn up
to a concentration of 1.5 wt.%. This is because the calculated [10,12,16] acicular ferrite content
increases from 35-67% when Mn is changed over the range illustrated. The low temperature
toughness for 2 wt.% Mn is nevertheless lower than that for the 1.5 wt.% alloy presumably
because the increased acicular ferrite content is not sufficient to compensate for the increased
strength. Indeed, an optimum manganese concentration of about 1.5 wt.% has been reported
to achieve the best toughness in manual metal arc welds of the type discussed here [22,23].
Fig. 6b shows similar data for carbon (the only difference being that the Mn concentration
is fixed at 1 wt.%.). The explanation is identical to that for the Mn data.
For welds similar to the carbon series, but with the carbon concentration fixed at 0.07
wt.%, the oxygen concentration alone was varied to a range well outside of the training dataset.
These results are presented in Fig. 7 along with the ±1 standard deviation predicted error bars.
It is clear that any attempt to extrapolate beyond the dataset on which the model is based
gives predictions which are not terribly useful. The fact that the toughness increases with
oxygen at low concentrations is strange since the acicular ferrite content (and indeed all the
other inputs) are kept constant. An increase in the oxygen content alone should lead to a
deterioration in toughness because of the tendency for non–metallic oxide particles to initiate
fracture.

11
 Fig. 6: (a) Variation in the normalised toughness as a function of the man-
ganese concentration. (b) Variation in the normalised toughness as a function
of the carbon concentration.

Finally, it is possible using the model to examine effects which cannot easily be produced
experimentally. It has frequently been argued that acicular ferrite is a better microstructure
than Widmanstätten ferrite, because the former with its less organised arrangement of ferrite
plates has a greater capacity to deflect cracks. This was tested for a manual metal arc weld
containing 0.07 wt.% carbon but of otherwise identical composition to the carbon series of welds
(Fig. 6). The allotriomorphic ferrite fraction was set to zero and all inputs except acicular ferrite
and Widmanstätten ferrite were varied in a complementary fashion. The results (Fig. 8) are
exciting – they demonstrate that increased acicular ferrite leads to an improvement of cleavage

12
 233 K

Fig. 7: Variation in the normalised toughness as a function of the oxygen

concentration. Oxygen is varied here without changing any of the other inputs.
The oxygen concentration in the training data was in the range 234–821 p.p.m.

toughness but not of the upper shelf energy – the latter is not expected to change since the
strengths of acicular and Widmanstätten ferrite are virtually identical [24].

0.0
Calculated Normalised
Charpy Toughness

-0.2
30 % Acicular
50 % Acicular
70 % Acicular
-0.4 90 % Acicular

-0.6
200 220 240 260 280 300
Temperature / K

Fig. 8: Variation in the normalised toughness as a function of the acicular

ferrite/Widmanstätten ferrite content, everything else being kept constant.

CONCLUSIONS

An artifical neural network has been used to rationalise Charpy impact toughness data on

13
manual metal arc and submerged arc steel weld deposits. The analysis is empirical but after
appropriate training, is found to reliably reproduce known metallurgical experience.
The method is useful in that the optimised network summarises knowledge in a quantita-
tive manner and can be retrained as new data become available.

ACKNOWLEDGEMENTS

We are grateful to the Education and Research Committee of Darwin College for fostering an
environment in which interdisciplinary learning flourishes.

14
 REFERENCES

1 . J. F. Knott: Fundamentals of Fracture Mechanics Butterworths, London, (1973) 8.

2 . O. M. Akselsen and φ. Grong: Materials Science and Engineering A A159 (1992) 187–192.

3 . D. E. Rumelhart, G. E. Hinton and R. J. Williams: Nature (London) 323 (1986) 533–536.

4 . D. J. C. MacKay: Neural Computation 4 (1992) 415–447.

5 . D. J. C. MacKay: Neural Computation 4 (1992) 448–472.

6 . D. J. C. MacKay: Network (1994) review article, in press.

7 . D. J. C. MacKay: Maximum Entropy and Bayesian Methods ed. G. Heidbreder, Santa Barbara,
published by Kluwer, Dordrecht, (1994) in press.

8. H. K. D. H. Bhadeshia: Recent Trends in Welding Science and Technology eds. S. A. David

and J. M. Vitek, ASM International, Materials Park, OH, USA (1989) 189–198.

9 . H. K. D. H. Bhadeshia and L.–E. Svensson: Mathematical Modelling of Weld Phenomena, eds.

H. Cerjak and K. E. Easterling, Institute of Materials, London (1993) 109–182.

10 . L.–E. Svensson: Control of Microstructures and Properties in Steel Arc Welds CRC Press, Ann
Arbor, U.S.A. (1994.)

11 . H. K. D. H. Bhadeshia, L.–E. Svensson and B. Gretoft: Acta Metallurgica 33 (1985) 1271–1283.

12 . H. K. D. H. Bhadeshia: Materials Science and Technology 8 (1992) 123–133.

13 . ESAB AB: Internal Report RR 82011 (1982.)

14 . ESAB AB: Internal Report RC 81033 (1981.)

15 . ESAB AB: Internal Report CML 85014 (1985.)

16 . L.–E. Svensson, B. Gretoft, A. A. B. Sugden and H. K. D. H. Bhadeshia: Computer Technology

in Welding Welding Institute, Abington, U. K., paper 24 (1988.)

17 . T. W. Lau, M. M. Sadowsky, T. H. North and G. C. Weatherly: Welding Metallurgy of

Structural Steels ed. J. Y. Koo, Met. Soc. AIME, Warrendale, PA, USA, (1987) 349–366.

18 . R. B. Oldland: Australian Welding Research (Dec. 1985) 31–43.

19 . B. Ahlblom, U. Bergstrom, N. E. Hannerz and I. Werlefors: WI conference on Residuals and

Impurities in Steels Welding Insititute, Abington, Cambridge, U. K., paper 49 (1986.)

20 . φ. Grong, A. O. Kluken and B. Bjornbakk: Joining and Materials 1 (1988) 164–169.

21 . H. Y. Liou, C. M. Liao and S. C. Wang: China Steel Technical Report No. 6 (1992) 30–37.

22 . L.–E. Svensson and B. Gretoft: American Welding Journal 69 (1990) 454s–461s.

23 . G. M. Evans: American Welding Journal 59 (1980) 67s–75s.

15
24 . A. A. B. Sugden and H. K. D. H. Bhadeshia: Metallurgical Transactions A 19A (1988) 1597–
1602.

16
 APPENDIX

Table 3 contains the values for the weights obtained after completing the training of the
network. These data can be used in combination with Table 1 and equations 1–3 in order to
use the network to make predictions of weld metal toughness.
 Description Abbreviation weight
(1)
Bias to hidden unit 1 θ1 -0.139995
(1)
Bias to hidden unit 2 θ2 -0.248802
(1)
Bias to hidden unit 3 θ3 0.589258
(1)
Bias to hidden unit 4 θ4 -0.080140
Bias to output unit θ(2) -1.40271
(1)
Process to hidden unit 1 w1,1 -0.953784
(1)
Process to hidden unit 2 w2,1 0.455700
(1)
Process to hidden unit 3 w3,1 0.010017
(1)
Process to hidden unit 4 w4,1 0.355382
(1)
Yield strength to hidden unit 1 w1,2 0.599265
(1)
Yield strength to hidden unit 2 w2,2 -0.396398
(1)
Yield strength to hidden unit 3 w3,2 -0.195261
(1)
Yield strength to hidden unit 4 w4,2 -0.051569
(1)
Carbon to hidden unit 1 w1,3 0.003035
(1)
Carbon to hidden unit 2 w2,3 0.000652
(1)
Carbon to hidden unit 3 w3,3 0.002913
(1)
Carbon to hidden unit 4 w4,3 0.004131
(1)
Silicon to hidden unit 1 w1,4 -0.052339
(1)
Silicon to hidden unit 2 w2,4 0.069965
(1)
Silicon to hidden unit 3 w3,4 -0.088386
(1)
Silicon to hidden unit 4 w4,4 0.039075
(1)
Manganese to hidden unit 1 w1,5 -0.009951
(1)
Manganese to hidden unit 2 w2,5 -0.006226
(1)
Manganese to hidden unit 3 w3,5 -0.010586
(1)
Manganese to hidden unit 4 w4,5 -0.008486
(1)
Phosphorus to hidden unit 1 w1,6 -0.108883
(1)
Phosphorus to hidden unit 2 w2,6 0.070553
(1)
Phosphorus to hidden unit 3 w3,6 0.029519
(1)
Phosphorus to hidden unit 4 w4,6 0.047509
(1)
Sulphur to hidden unit 1 w1,7 -0.004894
(1)
Sulphur to hidden unit 2 w2,7 -0.008146
(1)
Sulphur to hidden unit 3 w3,7 -0.006305
(1)
Sulphur to hidden unit 4 w4,7 -0.009485

Table 3a: The weights describing the trained network

 Description Abbreviation weight
(1)
Aluminium to hidden unit 1 w1,8 0.003158
(1)
Aluminium to hidden unit 2 w2,8 0.004146
(1)
Aluminium to hidden unit 3 w3,8 0.001416
(1)
Aluminium to hidden unit 4 w4,8 0.005775
(1)
Nitrogen to hidden unit 1 w1,9 0.305801
(1)
Nitrogen to hidden unit 2 w2,9 0.123011
(1)
Nitrogen to hidden unit 3 w3,9 -0.533206
(1)
Nitrogen to hidden unit 4 w4,9 0.028576
(1)
Oxygen to hidden unit 1 w1,10 -0.008650
(1)
Oxygen to hidden unit 2 w2,10 0.051781
(1)
Oxygen to hidden unit 3 w3,10 0.333115
(1)
Oxygen to hidden unit 4 w4,10 -0.204586
(1)
Primary Mic. to hidden unit 1 w1,11 0.012150
(1)
Primary Mic. to hidden unit 2 w2,11 -0.020293
(1)
Primary Mic. to hidden unit 3 w3,11 -0.044580
(1)
Primary Mic. to hidden unit 4 w4,11 -0.008752
(1)
Allotriomorphic to hidden unit 1 w1,12 0.025770
(1)
Allotriomorphic to hidden unit 2 w2,12 -0.018466
(1)
Allotriomorphic to hidden unit 3 w3,12 0.016131
(1)
Allotriomorphic to hidden unit 4 w4,12 0.028425
(1)
Acicular to hidden unit 1 w1,13 -0.074019
(1)
Acicular to hidden unit 2 w2,13 0.623779
(1)
Acicular to hidden unit 3 w3,13 0.655171
(1)
Acicular to hidden unit 4 w4,13 -0.748358
(1)
Temperature to hidden unit 1 w1,14 -0.077889
(1)
Temperature to hidden unit 2 w2,14 -0.077859
(1)
Temperature to hidden unit 3 w3,14 0.138927
(1)
Temperature to hidden unit 4 w4,14 0.084853
(2)
Hidden unit 1 to output unit w1 7.474170
(2)
Hidden unit 2 to output unit w2 6.714920
(2)
Hidden unit 3 to output unit w3 7.727480
(2)
Hidden unit 4 to output unit w4 8.835680

Table 3b: The weights describing the trained network