Analysis of toughness of welding alloys for

high strength low alloy shipbuilding steels

E. A. Metzbower, J. J. DeLoach, S. H. Lalam, and H. K. D. H. Bhadeshia

notch test is a quality control test used to select steels that

In previous work by the present authors a model was are resistant to brittle fracture. Knowledge of the Charpy
developed for the estimation of the tensile properties energy at only one temperature is not sufficient to indicate
(strength and ductility) of ferrous alloys intended for the transition curve. The Charpy energy is therefore mea-
the welding of high strength low alloy steels used in the sured at two different temperatures (218 and 251uC);
construction of ships. The method used to model the for steels of the 600 MPa class it is necessary to achieve
properties as a function of a large number of variables values of 81 J at 218uC and 47 J at 251uC. The dynamic
was based on a neural network within a Bayesian tear test is also a measure of resistance to brittle fracture
framework. This method is particularly useful when and it too is measured at two temperatures (21 and 229uC);
attempting to understand complex non-linear phenom- the minimum requirements in this instance are 610 J at
ena where the distribution of data within the input space 21uC and 407 J at 229uC. Satisfying the specifications
is not obvious. In the present work, a similar approach at both temperatures is an important quality control require-
is used to model the toughness (characterised by ment based on experience.
Charpy and dynamic tear tests) of the same alloys. The The correlation between the chemical composition of the
level of noise in the experimental data is perceived to be steel and its toughness is inexact, although for steels and
high, but it has nevertheless been possible to recognise weld metals various indicators abound. Increasing the
reasonable trends and uncertainties when making strength usually decreases the fracture toughness if the
predictions. For example, the toughness shows a non- microstructure remains the same. Carbon can drastically
linear deterioration as the oxygen concentration is change the transition curve in steel, as can manganese.7
increased; this behaviour is expected but can now be Nickel is sometimes known to promote low temperature
expressed quantitatively. STWJ/248 resistance to brittle fracture. Notch toughness is particularly
influenced by oxygen. In wrought steels, this is evidenced by
Dr Metzbower is in the Physical Metallurgy Branch, the lower transition temperatures in steels that are fully
Code 6320, US Naval Research Laboratory, Washing- killed by the addition of silicon and aluminium, as
ton, DC 20375 – 5000, USA, Dr DeLoach is in the US compared with semikilled or rimmed steels.
Navy Surface Warfare Center, Aberdeen, MD, USA,
and Sree Lalam and Professor Bhadeshia (hkdb@cus.
cam.ac.uk) are in the Department of Materials Science DATABASE
and Metallurgy, University of Cambridge, Pembroke The set of experimental data is identical to that used in the
Street, Cambridge CB2 3QZ, UK. Manuscript recei- previous study;1 most of the experimental details have been
ved 24 July 2001; accepted 6 September 2001. described there. The Charpy samples were of standard size
and geometry, i.e. 10 mm square sectioned specimens with
# 2001 IoM Communications Ltd. 2 mm depth notches of radius 0.25 mm and an included
angle of 45u. The rectangular dynamic tear test samples
were 181 mm in length, 38 mm in width, and 15.8 mm in
thickness, with a machined 9.5 mm depth parallel sided
notch of root radius 0.13 mm. The tip root radius after
INTRODUCTION pressing becomes 0.025 mm.
The toughness of weld metals remains a property that is The input variables are the chemical compositions of the
especially difficult to understand and more so to model. as deposited weld beads, either the measured Charpy values
This is because the ability of a metal to absorb energy at 218 or 251uC or the dynamic tear test values at 21 or
during failure depends on a very large number of variables; 229uC, and the cooling rate at 538uC. It should be noted
furthermore, the tests that are frequently used to char- that some of the temperatures stated in the present paper
acterise this property are empirical (although reliable appear unusual because they are conversions from Fahren-
indicators of quality). In a previous study,1 the tensile heit to Celsius (e.g. 538uC is 1000uF). The cooling rate was
properties of a particular class of experimental welding determined from the welding parameters (voltage, current,
alloys were modelled using a neural network method2 – 5 welding speed), preheat, and thickness of the plate.8,9
within a Bayesian framework. However, neither strength Table 1 gives the minimum, maximum, mean, and standard
nor ductility alone is a sufficient indicator of weld perform- deviations of each of the variables.
ance. The toughness must also meet design specifications. Figure 1a and b illustrate the range of the experimental
The purpose of the present work, therefore, was to develop data; the scatter evident on the graphs shows that the
similar models for the toughness of the weld metals, as behaviour expected at the two different test temperatures
characterised by Charpy and dynamic tear tests. may often be different. A further point to note is that many
Both the prior study1 and the present work utilise a data of the data fail the Charpy and tear test requirements
set that originated in research that had the intention of outlined above; it is important in creating models to include
developing welding consumables for joining high strength the results of both successful and unsuccessful experiments.
low alloy steels for ship construction.6 Aside from the temperature effects, both Charpy and
Ferritic steels have a ductile to brittle transition that is a dynamic tear tests are subject to considerable scatter,
function of composition and temperature. The Charpy V particularly in the region of the transition temperature.

368 Science and Technology of Welding and Joining 1999 Vol. 6 No. 6 ISSN 1362 – 1718
 Metzbower et al. Toughness of welding alloys for HSLA steels 369

Table 1 Input and output variables (concentrations are in

wt-% except where shown)

Standard
Variable Minimum Maximum Mean deviation

C 0.001 0.06 0.0307 0.0098

Mn 1.05 3.44 1.4361 0.1915
Si 0.05 0.4 0.2618 0.0555
Cr 0 0.21 0.0664 0.0530
Ni 1.66 5.63 3.1324 0.7800
Mo 0 1.23 0.5048 0.1452
Cu 0 0.48 0.1018 0.0766
S 0.001 0.012 0.0034 0.0019
P 0.001 0.015 0.0041 0.0026
Al 0.001 0.082 0.0066 0.0060
Ti 0.0008 0.3 0.0089 0.0181
Nb 0 0.069 0.0016 0.0041
V 0 0.031 0.0032 0.0042
B 0 0.01 0.0011 0.0021
O, wt-ppm 109 627 216.7565 58.6969
N, wt-ppm 6 135 29.8164 25.2642
Cooling rate, K s21 1.32 76.17 27.1079 23.0104
CVN (218uC), J 8 358 181.9955 57.0641
CVN (251uC), J 3.8 242 143.6876 64.3903
DT (21uC), J 128 2606 1322.4916 538.1394
DT (229uC), J 80 2380 950.2694 583.3663

CVN Charpy V notch; DT dynamic tear. Dynamic tear test data are
generated on welds that did not contain titanium. A more detailed
description of both the data set and the neural network employed
can be found elsewhere.1

Table 2 Number of models in committee and largest

value of additional error sv associated with any
member of committee

Data set Models in committee sv

1 Plots of a measured Charpy V notch energy at 251uC
CVN (218uC) 1 0.055
as function of that measured at 218uC, and b dynamic
CVN (251uC) 6 0.081 tear energy measured at 229uC as function of that
DT (21uC) 6 0.091 measured at 21uC
DT (229uC) 77 0.222
committee versus its size L gives a minimum that defines
the optimum size of the committee. Using this procedure,
Most of the scatter is due to local variations in the the optimum size of the committee for the four experimental
properties of the steel plate or weld, whereas some is due to data sets is given in Table 2.
difficulties in preparing perfectly reproducible notches. The Once the optimum committee is selected, it is retrained on
inhomogeneity of a weld is considerably greater than that of the entire data set without changing the complexity of each
plate steel and the corresponding scatter in experimental model, with the exception of the inevitable, although
values is expected to be greater. relatively small, adjustments to the weights.
Charpy V notch tests
RESULTS AND DISCUSSION Results
To avoid overfitting during the training process, the data The correlation between the experimental Charpy tests at
were divided into parts for training and testing.2 – 6 The 218 and 251uC and the values predicted using the
model is at first produced using only the training data and committee models is shown in Fig. 4. The committees
then tested for its ability to generalise on unseen data evidently perform rather well. It is emphasised that the error
against the remaining (test) data. The performance of the bars plotted in Fig. 4 and all subsequent graphs represent
best models in predicting the training and test data is shown the fitting error, the magnitude of which depends on the
in Fig. 2 for the Charpy tests and Fig. 3 for the dynamic position in the input space. The additional error sv is taken
tear tests. as the largest value of sv for any member in the committee
A committee of models can make more reliable predic- of models. These sv values are given in Table 2. The rather
tions.5 To create an appropriate committee, the best models large perceived noise in the Charpy values at 251uC is not
are ranked using the values of the test errors. Committees surprising given the scatter in the data at the lower test
are then formed by combining the predictions of the best L temperature. Only the fitting error, which is a function of
models, where L~1, 2,… and the size of the committee position in the input space, is plotted on all the diagrams
is given by the value L. A plot of the test error of the included in the present paper.

Table 3 Standard set of inputs used to observe trends (wt-%, except where indicated)

C Si Mn P S Cr Mo Ni Al B Cu N, wt-ppm Nb Ti V O, wt-ppm

0.036 0.31 1.72 0.001 0.004 0.03 0.59 2.63 0.005 0 0.061 8 0 0.007 0.001 181

Science and Technology of Welding and Joining 2001 Vol. 6 No. 6

370 Metzbower et al. Toughness of welding alloys for HSLA steels

2 Predicted values using best models versus measured 3 Predicted values using best models versus measured
values for Charpy V notch tests at a,b 218 and c,d values for dynamic tear tests at a,b 21 and c,d 229uC
251uC for training and test data: all data are normalised for training and test data: all data are normalised

Science and Technology of Welding and Joining 2001 Vol. 6 No. 6

 Metzbower et al. Toughness of welding alloys for HSLA steels 371

(a)

(b)

(c)
4 Comparison of predicted and measured Charpy V notch
values for as deposited weld metal tested at a 218 and
b 251uC: calculations were carried out using committees
of models (error bars represent ¡1s values, where s is
standard deviation)

Predictions
The most effective method of examining the performance of
the models is to use them to make predictions regarding
trends as a function of each of the inputs, for a reasonable
selection of baseline inputs. The set of baseline inputs is
given in Table 3. The cooling rate was set at 61.2 K s21.
These inputs were selected to represent a weld metal of
particular interest in the context of ship construction. As
pointed out in Ref. 1, an alloy such as this should have a
mixed microstructure of bainite and low carbon martensite, (d)
given the combination of high manganese and nickel
concentrations and the presence of molybdenum.
Figure 5 shows example plots of the effects on the Charpy
values of varying the carbon, manganese, and nickel
concentrations and the cooling rate. All the plots have
the same vertical scale. It is particularly interesting that all
these factors cause a substantial increase in strength1 but
have a different effect on the Charpy values. Figure 5 shows
that to achieve the combined strength and toughness in the
weld, limits must be placed on the concentration of different
elements; for example, contrary to the general assumption,
the presence of nickel in the range 2 – 4 wt-% does not lead
to an improvement in toughness in the present context.
Varying the molybdenum concentration from 0 to 1 wt-%
did not affect the Charpy toughness values significantly.
Figure 6 shows example plots of the effects of varying the
concentrations of nitrogen, oxygen, aluminium, and tita- 5 Variations of Charpy V notch values at 218 and 251uC,
nium. Nitrogen is expected to increase the strength of the calculated using committee models, as functions of
weld metal and thus decrease its toughness, so the reason a carbon, b manganese, and c nickel concentrations and
for the initial increase is not understood. By contrast, those d cooling rate: error bars represent ¡1s values

Science and Technology of Welding and Joining 2001 Vol. 6 No. 6

372 Metzbower et al. Toughness of welding alloys for HSLA steels

(a)

(b)

(c)
7 Comparison of predicted and measured dynamic tear
values for as deposited weld metal tested at a 21 and
b 229uC: calculations were carried out using commit-
tee models (error bars represent ¡1s values)

elements that form oxides or carbides show a significant

deterioration in toughness and ductility as their concentra-
tions are increased. It can be concluded that in the present
class of welds, the main role of the inclusions is to nucleate
voids or brittle fracture, rather than leading to improve-
ments in microstructure via their ability to nucleate phase
transformations.
It is noteworthy that in Fig. 6c and d, the effects of
aluminium and titanium can be seen to be rather small,
again consistent with the hypothesis that the inclusions are
(d) not a key feature of microstructural development.

Dynamic tear tests

Results
The correlation between the experimental measurements
and data predicted using the committee of models is shown
in Fig. 7. The committees evidently perform rather well. It is
emphasised again that the error bars that are plotted here
represent the fitting error, the magnitude of which depends
on the position in the input space. The additional error sv is
given in Table 2 and is taken as the largest value of sv for
any member in the committee of models. The rather large
perceived noise in the dynamic tear values is not surprising
given the scatter in the data. Only the fitting error, which is
a function of position in the input space, is plotted on all the
diagrams included in the present paper.
6 Variations of Charpy V notch values at 218 and 251uC,
predicted using committee models, as functions of a nitro- Predictions
gen, b oxygen, c aluminium, and d titanium concentra- The set of baseline inputs for the predictions is given in
tions: error bars represent ¡1s values Table 3. Setting the cooling rate at 61.2 K s21 should lead

Science and Technology of Welding and Joining 2001 Vol. 6 No. 6

 Metzbower et al. Toughness of welding alloys for HSLA steels 373

(a)

(a)

(b) (b)

(c) (c)

(d)
(d)

8 Variations of dynamic tear values at 21 and 229uC 9 Variations of dynamic tear values at 21 and 229uC
predicted using committee models, as functions of a predicted using committee models, as functions of a
nickel, b molybdenum, and c nitrogen concentrations carbon, b manganese, c oxygen, and d aluminium con-
and d cooling rate: error bars represent ¡1s values centrations: error bars represent ¡1s values

Science and Technology of Welding and Joining 2001 Vol. 6 No. 6

374 Metzbower et al. Toughness of welding alloys for HSLA steels

to a microstructure consisting of a mixture of bainite and models. The error bars become very large, indicating that
low carbon martensite, given the combination of high the models frequently lack the knowledge to make reliable
manganese and nickel concentrations and the presence of predictions given the experimental data.
molybdenum.1
Figure 8 shows plots of the effects on the dynamic tear
values of varying the nickel, molybdenum, and nitrogen CONCLUSIONS
concentrations and the cooling rate. It is particularly A neural network method within a Bayesian framework has
interesting that all of these factors cause a substantial been used to model experimental data for both the Charpy
increase in strength1 but have a different effect on the toughness values at 218 and 251uC and the dynamic tear
dynamic tear values. These plots show the contrary effects test values at 21 and 229uC for ferritic steel weld metals
of strength1 and toughness. All the plots have the same appropriate for the welding of high strength low alloy steels.
vertical scale. The inputs for the neural network were the concentrations
In contrast, carbon and manganese, which both increase of sixteen chemical elements and the weld cooling rate at
the strength,1 show a different effect on the dynamic tear 538uC.
values (Fig. 9). At 21uC carbon increases the dynamic tear The Charpy model behaves well and can often be used to
value, whereas at 229uC carbon shows the expected make reliable predictions. By contrast, the dynamic tear
behaviour of decreasing the dynamic tear values as the data do not seem to be sufficiently comprehensive to give
carbon content is increased. Figure 9 indicates that reliable extrapolation behaviour, the predictions being
manganese, a strengthening element, must be limited to a associated with large uncertainties. This information in
surprisingly small range of values for the weld to surpass the itself is valuable and would not have been revealed without
minimum levels required. This plot shows that to achieve a the Bayesian framework.
desirable combination of strength and toughness in the
weld, limits must be placed on the concentration of different
elements; these limits are ordinarily problematic to assess ACKNOWLEDGEMENTS
given the complex effects of the individual elements. Funding support from the Office of Naval Research
However, models such as those presented here allow the (Arlington, VA, USA) and from the Cambridge Common-
exploration of the input space to determine the optimum wealth Trust is gratefully acknowledged.
combination of inputs.
Figure 9c and d show the effects of oxygen and
REFERENCES
aluminium on the dynamic tear values. As expected from
1. e. a. metzbower, j. j. deloach jr, s. h. lalam, and h. k. d. h.
the presence of oxide, increasing the oxygen content bhadeshia: Sci. Technol. Weld. Joining, 2001, 6, (2), 116 – 124.
decreases the dynamic tear values. Conversely, aluminium 2. d. j. c. mackay: Neural Comput., 1992, 4, 415.
can be either beneficial or harmful depending on the test 3. d. j. c. mackay: Neural Comput., 1992, 4, 448.
temperature. Over the range of values calculated, the 4. d. j. c. mackay: Neural Comput., 1992, 4, 698.
present requirements are satisfied for all aluminium 5. d. j. c. mackay: in ‘Mathematical modelling of weld pheno-
contents. mena III’, (ed. H. Cerjak and H. K. D. H. Bhadeshia), 359 –
The addition of chromium from zero to 0.3 wt-% reduces 389; 1997, London, The Institute of Materials.
the dynamic tear values at both temperatures from appro- 6. h. k. d. h. bhadeshia: ISIJ Int., 1999, 39, 966 – 979.
ximately 1500 J to less than 1000 J. In Ref. 1, increasing the 7. j. j. deloach jr: in ‘Welding and weld automation in
shipbuilding’, (ed. R. DeNale), 85 – 104; 1996, Warrendale,
amount of chromium was found to have very little effect on PA, TMS.
the yield strength of the weld metal. 8. g. e. dieter: in ‘Mechanical metallurgy’, 492 – 497; 1976, New
One important feature to be noted from all the data York, McGraw-Hill.
presented in Figs. 8 and 9 is that unlike the Charpy models, 9. l. p. conner (ed.): in ‘Welding handbook’, 8th edn, Vol. 1,
there are great uncertainties in extrapolating the ductile tear 75 – 79; 1987, Miami, FL, American Welding Society.

Science and Technology of Welding and Joining 2001 Vol. 6 No. 6