IIW Document No. IX-2165-05

MODELLING OF HAZ HARDNESS IN C-Mn PIPELINE STEELS

SUBJECTED TO IN-SERVICE WELDING PROCEDURES

D. Nolan, Z. Sterjovski and D. Dunne

ABSTRACT Welding onto pressurized pipelines that contain flammable fluid, in order to facilitate repairs or branch connections, is a critical procedure with considerable risk to personnel and infrastructure. Limiting the heat input is obviously an important consideration to prevent “burn-through”, but the potential for rapid cooling of the weldment increases its susceptibility to hydrogen assisted cold cracking (HACC). Therefore, one of the more important factors for in-service welding procedure development relates to the increased risk of formation of hard, brittle microstructures in the grain coarsened heat affected zone (GCHAZ) of the weld, microstructures that increase the risk of HACC. The current work has been successful in utilizing dilatometric simulations to derive two new hardness prediction models that more accurately predict hardness in the GCHAZ for typical in-service welding applications than a commercially adopted and widely used hardness prediction algorithm. Although it is acknowledged that further work is required to validate the models for a wider range of in-service welding conditions, the performance of the models demonstrates their potential for developing improved in-service welding procedures. KEYWORDS In-service welding, Heat-affected zone, Hardness prediction, Dilatometry, Neural network model. AUTHOR CONTACTS * Dr David Nolan (corresponding author), Dr Zoran Sterjovski, Faculty of Engineering, Faculty of Engineering, University of Wollongong, University of Wollongong, NSW Australia 2522 NSW Australia 2522 Email: dnolan@uow.edu.au Email: zoran@uow.edu.au Prof Druce Dunne, Faculty of Engineering, University of Wollongong, NSW Australia 2522 Email: druce@uow.edu.au

1

INTRODUCTION In order to facilitate repairs or branching connections, it is often necessary to produce welded joints on in-service pipelines containing highly flammable and pressurized fluids, such as fuel gas and liquid petroleum products. Welding onto in-service pipelines which are pressurized with flowing product can be conducted in a safe and cost-effective manner, however, there are increased risks of pipeline material failure during such procedures. The first, and most immediate, risk is caused by the internal fluid pressure which, when combined with localized heating of the pipeline, can lead to failure of the pipe wall due to reduced material strength. This effect, commonly referred to as “burn-through”, is a developing issue, since there is a growing tendency to use higher strength pipe steels with reduced wall thickness, and the benefits afforded by the high strength steels at lower temperatures are not realized at higher temperatures. The second important risk to consider is the generation of relatively high cooling rates in the weld zone by the fluid flowing inside the pipe. This has two important consequences. Firstly, at higher cooling rates there is a greater prospect of austenite transforming at lower temperatures to bainite and/or martensite, with an attendant increase in hardness. This is particularly important wherever older pipe materials, with higher carbon and alloy contents (high carbon equivalent) are encountered. Secondly, the higher cooling rate leads to higher diffusible hydrogen content retained in the weld zone. Both of these factors increase the risk of hydrogen-assisted cold cracking (HACC), a significant threat to the integrity of a welded joint, particularly as the effect may not be immediately apparent and may take several hours or days to develop. Much of the past research effort pertaining to procedure development and risk management for in-service welding has focused on predicting weld cooling rates and the risk of burn-through [1-5]. While this effort has been fruitful, several other important factors have received insufficient attention, including the quantification of diffusible hydrogen content during typical in-service welding procedures and the effect of material composition and cooling rate on hardness, particularly in the pipe heat-affected zone (HAZ) at the very high cooling rates expected during in-service welding procedures. Hardness prediction algorithms that attempt to predict weld HAZ hardness as a function of composition and cooling rate do exist, but these relationships are derived from studies on steels with a wide range of compositions, using cooling rates that are generally lower than those expected during in-service welding. The current work addressed this matter by conducting a series of in-service welding simulations and measurements of weld HAZ hardness, in order to assess the validity of existing hardness prediction methods. The current work involved both numerical (regressive) and neural network methods to predict HAZ hardness, and the results of these analyses were compared with existing algorithms. An attempt was also made to correlate the output of all models with actual in-service weld (ISW) hardness data. BACKGROUND Hydrogen-assisted cold cracking represents a significant threat to the integrity of pipelines, acting as the source of defects which have the potential to initiate catastrophic failure under normal service conditions [6]. It is generally accepted that for in-service welding, hydrogen cracks are most likely to form at the toe of the multipass fillet weld in the HAZ of the pipe or the fitting, since the outermost weld pass may not entirely be subjected to a tempering effect

2

from subsequent fill passes. HAZ hardness is therefore expected to be maximum in this region. HACC occurs at near ambient temperatures and may be delayed for several hours, or even days, after welding is completed, making detection during normal fabrication operations difficult. There have been a number of general reviews on the phenomenon of hydrogen cracking in steel weldments [7-9] which reflect the general understanding that the main factors determining susceptibility to HACC include; the presence of diffusible hydrogen, high local concentrations of residual or applied stresses, the presence of defects, and a microstructural susceptibility. However, these factors are interdependent and the critical levels of diffusible hydrogen and stress necessary to initiate and/or propagate HACC are specific to the particular welding process, weld configuration and material in question. A number of standards and guidance notes have been developed to provide guidance on processing parameters for the avoidance of HACC in the HAZ [10-16]. However, welding onto pipelines that contain flowing fluid, liquid or gaseous, presents a very specific set of welding conditions with much more severe cooling conditions and a consequently greater risk of the hardness reaching critical levels in the weld zone. Some more specific guidance on the particular case of welding onto in-service pipelines is given in AS 2885.2:2002 [13], the Welding Technology Institute of Australia (WTIA) Technical Note 20 [17] and proprietary software packages hereafter referred to as the Battelle model [18], the PRCI model [19] and the CRC-WS model [20]. However, a considerable gap remains in knowledge pertaining to effects of composition and hydrogen content on the risk of HACC for in-service welding. To prevent HACC in any given application, the diffusible hydrogen content must be lower than the critical level required to initiate and propagate cracking under the prevailing stress and microstructural conditions. The applicable stresses include internal pressure and structural loads, residual stresses from welding, and stress concentrations, such as those associated with welding defects. Since no mechanical or thermal stress relief treatment is carried out, a significant amount of residual stress acting on the weld cannot be avoided and must always be assumed. The first step toward avoiding HACC during in-service pipeline welding is therefore to minimize the hydrogen level by using low-hydrogen electrodes, or a low-hydrogen welding process. However, low hydrogen levels cannot always be guaranteed and the effect of welding conditions (both atmospheric and process-related) on diffusible hydrogen levels is not well quantified [21]. Given the uncertainty over diffusible hydrogen levels, even when using designated low-hydrogen electrodes, it has been considered prudent to develop procedures that minimize the formation of crack-susceptible microstructures as an added assurance against hydrogen cracking. Although significant work has been undertaken in this regard [22], further work is required to confirm the influence of material composition and cooling conditions on the development of crack susceptible microstructures. Hardenability and Hardness Ferritic pipeline steels are prone to transformation hardening in the weld heat-affected zone (HAZ) as a result of the relatively high cooling rates experienced during the weld cycle. This is especially so in the case of in-service welding, where the fluid flowing within the pipeline provides a very efficient heat sink, quenching the austenitised steel at rates of up to 150 °Cs-1 through the transformation range on cooling. The prior austenite grains produced in the pipe

3

during welding are coarsest in the HAZ near the fusion boundary and it is here that the microstructure is at the greatest risk of HACC. In general, increased rates of cooling lead to higher hardness levels for a given steel composition. Combined with the introduction of diffusible hydrogen through the welding process, and the increased levels of diffusible hydrogen trapped within the weld zone at higher cooling rates, this increase in hardness causes an increase in susceptibility to hydrogen cold cracking. The susceptibility of a given steel composition to hardening in the HAZ as a result of the weld thermal cycle is often predicted by the use of carbon equivalent (CE) formulae. These formulae describe the compositional effect on hardenability, or the tendency to form martensite, and when combined with the effects of plate (or pipe wall) thickness, weld configuration and heat input on cooling time, they can be used in procedures for estimating the preheat conditions necessary to avoid HACC. There are a number of such formulae used in this way as a measure of weldability [23], but perhaps the most widely used is the International Institute of Welding’s CEIIW formula: CEIIW = C + Mn/6 + (Cu+Ni)/15 + (Cr+Mo+V)/5 (1) The CEIIW formula is generally considered to be more relevant for evaluating the weldability of older steel types, such as steels produced prior to thermomechanical controlled processing, with higher carbon contents from 0.15-0.3%. Another widely used formula, Pcm, regards the effect of carbon as much more significant than that of the alloying elements, and is generally considered to be more appropriate for the more modern, carbon-reduced or microalloyed steels: Pcm = C + Si/30 + (Mn+Cu+Cr)/20 + Ni/60 + Mo/15 + V/10 + 5B (2) For example, the AWS method for determining the minimum necessary preheat to avoid HAZ cracking recommends Pcm be used for steels with C<0.11% and CEIIW for steels with C≥0.11% [10]. Another formula, developed by Yurioka et al [24] as a weldability index for a wide variety of steels, incorporates an interactive term for carbon and alloying elements. The CEN formula approaches the values of CEIIW when applied to higher carbon steels and the values of Pcm when applied to lower carbon steels: CEN = C + A(C) {Si/24 + Mn/6 + Cu/15 + Ni/20 + (Cr+Mo+Nb+V)/5 + 5B} (3) where: A(C) = 0.75 + 0.25 tanh [20(C-0.12)] The following correlation between the formulae have been established: CEN = 2 Pcm – 0.092 (where C ≤ 0.17%) (4) and, CEN = CEIIW + 0.12 (where C > 0.17%) (5)

4

The effect of cooling rate is not specifically incorporated as a variable in these CE equations, even though it has a profound effect on microstructural evolution. Further, the cooling rates associated with in-service welding on relatively thin pipe material, with and without preheating, are likely to be at, or beyond, the upper limit of the cooling rates for which the above equations have been developed. Note that the cooling rate parameter, t8/5, is widely used to describe the critical cooling regime from 800-500°C where austenite transforms to ferrite (at lower cooling rates), or bainite and martensite (at progressively higher cooling rates). It is the development of the harder and less tough bainitic and martensitic constituents that is of most concern with respect to HACC in the HAZ. Yurioka [23] has provided an explanation of the effect of increased hardenability (as determined by composition) and cooling rate on HAZ hardness in medium and low carbon steels. This model was developed for the case of low heat input welding, where cooling rates are sufficient to develop high levels of diffusible hydrogen at ambient temperature in the presence of hard microstructures. He claims that the HAZ hardness of a higher carbon steel increases more significantly with increasing hardenability (or an increase in tM, the critical value of the cooling rate (t8/5) required for a fully martensitic structure) than it does for a carbon-reduced steel. Due to the strong influence of carbon on hardenability, the higher carbon steel has a higher value of tM than the lower carbon steel. Moreover, other factors that promote hardenability, such as alloying and increased austenite grain size, will have a stronger effect on raising tM than for lower carbon steels. For low carbon contents, the hardenability is more strongly affected by carbon than the alloying elements. Therefore, as suggested above, the weldability of a higher carbon steel (eg, 0.15-0.3%) is more appropriately evaluated by a formula such as the CEIIW formula, whereas Pcm is more relevant to low carbon steels because carbon is more strongly weighted. The CEN formula was proposed as a compromise which, by incorporating an interactive effect between carbon and alloying elements, approaches the CEIIW determination for higher carbon materials and the Pcm determination for lower carbon materials. The empirical CE formulae above indicate hardenability, but hardenability cannot be related directly to hardness levels, since it only indicates the likelihood of transformation to martensite on cooling. Once a fully martensitic structure is achieved, then the hardness of a fully martensitic HAZ microstructure is determined largely by the carbon content, and can be estimated by the following relationship [25]: HM = 884C (1-0.3C2) + 294 (7) where carbon content is less than 0.8%. Nitrogen is also an interstitial element that can affect hardness of martensite, but Yurioka et al [25] maintain that it can be effectively ignored in the HM formula, since the level of nitrogen in modern structural steels is generally far lower than the carbon content [25]. If the cooling rate does not exceed the critical cooling rate for fully martensitic structure, then the microstructure will be mixed, with both martensitic and bainitic constituents. Although it has been shown [26] that the strength of steel can actually increase for small volume fractions of bainite, before decreasing more markedly with a further increase in bainite at the expense of martensite, as a general rule, it is to be expected that decreasing martensite volume fraction will result in decreasing strength (and hardness).

5

Perhaps the most widely adopted empirical method for prediction of HAZ hardness in pipeline steels is the one developed by Yurioka et al [25]. This relationship is said to be reliable for transformable steels with the following composition range: C<0.3%, Ni<5%, Cr<1%. The algorithm, hereafter referred to as the Yurioka model, is defined as: HVmax = 442C + 99CEII + 206 + (402C – 90CEII + 80)arctan(x) (8) where: x = (log(t8/5) – 2.30CEI – 1.35CEIII + 0.882) / (1.15CEI – 0.673CEIII – 0.601) (9) CEI = C + Si/24 + Mn/6 + Cu/15 + Ni/12 + Cr/8 + Mo/4 + ∆H (10) CEII = C + Si/24 + Mn/5 + Cu/10 + Ni/18 + Cr/5 + Mo/2.5 + V/5 + Nb/3 (11) CEIII = C + Mn/3.6 + Cu/20 + Ni/9 + Cr/5 + Mo/4 (12) Note that ∆H is a term introduced to account for the strong hardening effect of boron, such that; ∆H = 0 when B ≤ 1ppm, ∆H = 1.5 (0.02-N) when B ≤ 2ppm, ∆H = 3.0 (0.02-N) when B ≤ 3ppm, and ∆H = 4.5 (0.02-N) when B ≥ 4ppm, While this relationship was developed using materials with a wide range of compositions and cooling rates (t8/5 from 3 to 100 seconds), it has been adopted as the basic algorithm for hardness prediction in a commercially-applied, software-based, procedure development tool for in-service welding applications [19]. The current work seeks to review the application of this algorithm to in-service welding applications, and to develop an improved predictive system based on empirical evidence derived from materials and conditions more closely approximating those used for in-service welding applications. It compares the output from the Yurioka model, and from numerical and neural network models developed from data derived from dilatometer simulations. It also compares the output of all three models against a set of measured results from actual welding trials conducted under in-service welding conditions. EXPERIMENTAL Materials and Procedures The experimental work was based on a series of thermal cycle simulations conducted in a high quench rate dilatometer, that were designed to generate the cooling conditions experienced by the grain-coarsened HAZ during in-service welding procedures. The simulations were carried out using a Theta dilatometer, wherein samples with 5 mm outside diameter, 3.5 mm inside diameter and 10 mm in length were held between two quartz tubes and heated to 1400 °C by a water-cooled induction coil under vacuum (see Figure 1). Using helium quench gas, it was possible to achieve t8/5 cooling times of between 2 and 10 seconds,

6

the range of cooling rates typical for in-service welding applications. A type 'S' thermocouple was spot welded to the sample to monitor and control the temperature of the sample during the thermal cycling.

quenching gas

thermocouple

induction coil

specimen

quenching gas outlet

quartz tube holder

LVDT

10 mm

5 mm 0.75 mm

Figure 1: A schematic illustration of the experimental setup for the dilatometer, including the dimensions of the cylindrical test specimen. The various Australian and USA pipeline and tap fitting steels investigated in the current work are listed in Table 1. The table shows the compositional analysis and calculated CE values for each material. Note that with the analytical methods available, it was not possible to measure boron contents of less than 0.0003%. As no boron was detected in the analysis, boron content was therefore assumed to be nil for calculations made in applying the Yurioka model. Standard procedures for metallographic preparation were used to prepare cross-sections for optical microscopy. Polished sections were etched in 4% Nital solution to reveal microstructures and images were captured in a digital format on a Leica optical microscope. Hardness of materials was determined using a Vickers hardness tester, at a load of 5 kg for the materials prior to heat treatment and 10 kg for material following thermal cycle simulation. Hardness measurements are reported as an average of five results taken randomly in the heat treated specimens. Hardness testing was carried out in accordance to AS1817.1-2003, Metallic materials - Vickers hardness test - Test methods (ISO 6507-1:1997, MOD). Modelling the effects of composition and cooling rate on the CGHAZ hardness An empirical model was developed for predicting the HAZ hardness using the dilatometer data and simple regression methods to approximate relationships of best fit between hardness and carbon equivalent. The relationships for each of the cooling rate conditions were then combined to form a hardness prediction model with carbon equivalent and cooling rate as variables.

7

Table 1: Pipe and tap fitting steel data. Steels A1 to A10 are pipe materials typical of Australian applications. The remaining steels are from North America, where the TAPf is a fitting materials. Compositions are given in weight percentage.

Material A1 A2 A3 A5 A6 A8 A10 AGA-J TAPp P6 TAPf% C 0.1700 0.0800 0.0700 0.1100 0.2400 0.1000 0.1000 0.2400 0.1300 0.2900 0.1300% Si 0.2500 0.2600 0.1300 0.1300 0.0500 0.1500 0.2400 0.0300 0.1700 0.0050 0.3900

% Mn 1.2300 1.6300 1.3500 0.4400 0.9500 1.1500 1.4200 1.1900 1.3800 0.9600 1.4300% Cu 0.0090 0.0010 0.0240 0.0330 0.0190 0.0140 0.0210 0.0100 0.0980 0.0400 0.2800% Cr 0.0260 0.0220 0.0240 0.0120 0.0200 0.0190 0.3100 0.0440 0.2100 0.0300 0.1000% Ni 0.0090 0.0200 0.0200 0.0230 0.0190 0.0180 0.0170 0.0140 0.0190 0.0380 0.1000% Mo 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050 0.0330% V 0.0050 0.0600 0.0050 0.0050 0.0050 0.0050 0.0400 0.0530 0.0430 0.0050 0.0050% B <0.0003 <0.0003 <0.0003 <0.0003 <0.0003 <0.0003 <0.0003 <0.0003 <0.0003 <0.0003 <0.0003% P 0.0080 0.0160 0.0200 0.0110 0.0090 0.0150 0.0140 0.0150 0.0150 0.0110 0.0080% S 0.0050 0.0050 0.0050 0.0060 0.0190 0.0080 0.0050 0.0180 0.0050 0.0240 0.0050% Ti 0.0050 0.0050 0.0210 0.0210 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050 0.0050% Al 0.0210 0.0130 0.0250 0.0290 0.0050 0.0390 0.0300 0.0050 0.0300 0.0080 0.0330% Nb 0.0460 0.0530 0.0400 0.0270 0.0050 0.0250 0.0400 0.0260 0.0400 0.0050 0.0390% O 0.0012 0.0007 0.0024 0.0068 0.0078 0.0033 0.0019 0.0091 0.0020 0.0033 0.0017% N 0.0072 0.0018 0.0056 0.0089 0.0036 0.0050 0.0077 0.0033 0.0054 0.0079 0.0087

CEIIW 0.38 0.37 0.30 0.19 0.41 0.30 0.41 0.46 0.41 0.46 0.42Pcm 0.25 0.18 0.15 0.14 0.29 0.17 0.20 0.31 0.23 0.35 0.24CEN 0.39 0.26 0.21 0.18 0.41 0.24 0.32 0.47 0.38 0.47 0.38

OD (mm) 508 406 406 324 273 324 864 508 1219 508 N/AThick (mm) 8.6 8.6 7.8 5.8 6.6 5.2 13.8 6.4 14.3 7.9 N/A

Grade (API 5L) X60 X70 X60 X42 X42 X42 X65 X52 X70 X52 A537*

A second predictive model was developed using Neuralworks Professional II/Plus software to build a back-propagation neural network model. The input variables and ranges used for the model are shown in Table 2. Table 2: Data ranges (maximum value – minimum value) and median values used in the sensitivity analysis. Compositions are given in weight percentage.

Input Variable Maximum Minimum Median

%C 0.29 0.07 0.18 %Si 0.39 0.005 0.1975

%Mn 1.63 0.44 1.035 %Cu 0.28 0.001 0.1405 %Cr 0.31 0.012 0.161 %Ni 0.1 0.009 0.0545 %Mo 0.033 0.005 0.019 %V 0.06 0.005 0.0325 %P 0.02 0.008 0.014 %Si 0.024 0.005 0.0145 %Ti 0.021 0.005 0.013 %Al 0.039 0.005 0.022 %Nb 0.053 0.005 0.029 t8/5 (s) 10 2 6

Original HV 234 159 197

8

The input variables can be combined into two major areas: one characterizing the original pipe material (chemical composition and hardness); and another quantifying the HAZ simulation conditions (t8/5). Three data files were used in the modelling process, namely the training file, testing file and the sensitivity file. There were 63 cases in the data file that were representative of the overall data, and 15 cases in the testing file that were excluded from the training file. Sensitivity data was used to further validate the model and to interrogate the model to try to gain a physical understanding of the predicted trends. The initial step in building the model is to set the number of processing elements in the input layer (see Figure 2). Individual processing elements in the input layer correspond to an input variable, and the processing element in the output layer corresponds to HAZ hardness. The role of the hidden layer, which is a layer that contains a systematically determined number of processing elements, is to help establish the relationships and inter-relationships between processing elements in the input and hidden layers that lead to the best estimates of output. The latter stages in building the model include defining other parameters such as selecting the learn rule (the mathematical method the program uses to correlate all of the processing elements in all layers), learning coefficients (values assigned to each layer and the overall model) and transfer mode (transferring the weighted sum of the processing elements or neurons into an output prediction) as listed in Table 3.

Output layer

Hidden layer

Input layer

Bias

Figure 2: An illustration of the structure of the artificial neural network program structure used to develop a hardness prediction model. Table 3: Key neural network model parameters for back-propagation.

Learn Rule Delta Rule

Learn Sequence Random

Learn Cycles 500000

Transfer Mode Hyperbolic Tangent

Input Data Range

(0,1)

Output Data Range

(0.2,1)

Epoch Size 63

Learning Coefficients

0.80, 0.85

Momentum 0.35

9

The learn rule used in the model is the Delta rule and the error in the output layer is the difference between the desired output and the actual output. The error calculated for the model was the root mean square (RMS). This error, transformed by the derivative of the transfer function, is “back-propagated” to prior layers where it is accumulated. This back-propagated and transformed error becomes the error term for that prior layer. The process of back-propagating the errors continues until the input layer is reached. The network-type back-propogation derives its name from this method of calculating the error term [27]. The actual weight update equations for the Delta rule are: (13) ijijiijij mCxeCww ×+××+= 21

'

ijijij wwm −= ' (14)

where C1 and C2 are learning coefficients. The weights are changed in proportion to the error (e), and the input to that connection (x). The momentum term (m) is used to help smooth out the weight changes. Also, the subscripts i and j refer to the number of the processing element and the number input variable, respectively. The term wi is the initial weight vector for the ith processing element in the layer, and wi’ is the weight vector after it has been updated by the learning rule [28]. The result of the weighted sum is then transferred into an output via a hyperbolic tangent function [28]. ( ) ( ) ( )zzzz eeeezf −− +−= / (15)

Random weightings are assigned to each processing element as an arbitrary starting point in the training process and the weightings are progressively altered on exposure to numerous repetitions of training examples. Hence, the neural network is in the process of learning and constant adjustment of weightings between processing elements leads to a reduction in the RMS error of the overall training data. In this model, 500,000 cycles at an epoch size of 63 were required to minimize and stabilize the RMS error. Certain indicators are set up during the training process to ensure the model is learning properly and avoiding incorrect learning patterns that may result, for example, from a saturation of weightings. The model reaches its optimum when the RMS error has stabilized. The model is verified against cases in the testing data file, which are independent of cases in the training file. The predicted results are plotted versus the experimental (or actual) results. In this paper, the sensitivity for carbon content, cooling time (t8/5) and and parent material hardness are considered. Sensitivity is represented graphically and when conducting a sensitivity analysis of a certain input variable the remaining input variables are set at their median values (Table 2). Sensitivity measures the response of output (HAZ hardness) across the entire range of an individual input variable, and is represented by ∆H.

minmax ii HVHVH −=∆ (16) where HVi max is the hardness corresponding to the maximum input value for a particular input variable while the remaining input variables are set at their median values, and HVi min is the

10

hardness corresponding to the minimum input value for a particular input variable while the remaining input variables are set at their median values. The output from the above models and the Yurioka model were compared to actual HAZ hardness data from previous investigations by Painter and Sabapathy [29]. These data were obtained from simulated welding trials on steels A1, A2 and A8, using conditions typical of in-service welding procedures. The measured cooling rates and compositions from these trials were used as input into the three hardness prediction models in order to compare their accuracy. RESULTS The microstructure and hardness of each steel prior to the weld thermal simulation are presented in Figure 3.

Figure 3: Photomicrographs showing the microstructures and mean hardness values (HV, 5 kg) for each of the starting materials. Micron bars represent 20 µm.

11

The microstructure of the pipeline materials typically comprises of a mixture of ferrite and pearlite, with coarse grained ferrite and higher proportion of pearlite for older normalized steels with higher carbon content, and fine ferrite with low proportion of pearlite for more modern low carbon steels. Microstructural banding due to rolling during the steelmaking process is also evident in some of the steels, particularly for the A1, A10, AGA-J, TAPp and TAPf materials. The hardness values measured for each of the heat-treated materials are presented graphically as a function of cooling time (t8/5) in Figure 4, and the microstructures of the post heat treated steels at t8/5 cooling times of 2, 6 and 10 seconds are shown in Figures 5 and 6.

(a)

100150200250300350400450500550

0 2 4 6 8 10 12

Cooling time (t8/5)

Har

dnes

s (H

V)

A-1A-2A-3A-5A-6A-8A-10

(b)

100

150

200

250

300

350

400

450

500

550

0 2 4 6 8 10 12

Cooling time (t8/5)

Har

dnes

s (H

V)

AGA-JTAPpP6TAPf

Figure 4: Graphs showing the mean hardness of simulated thermal cycle samples of each material as a function of cooling time, t8/5, for (a) Australian and (b) North American steels.

12

Figure 5: Photomicrographs showing the microstructures and mean hardness values for each of the Australian pipe steels following thermal cycling with cooling times (t8/5) of 2, 6 and 10 seconds. Micron bars represent 20 µm.

13

Figure 6: Photomicrographs showing the microstructures and mean hardness values for each of the North American pipe and fitting steels following thermal cycling with cooling times (t8/5) of 2, 6 and 10 seconds. Micron bars represent 20 µm. Development of the numerical model The relationships between carbon equivalent and measured hardness of the grain-coarsened HAZ for each cooling rate are shown in Figures 7-11. The equations representing lines of best fit for each data set, and the relevant R2 values are also presented. The relationships between carbon equivalent and hardness for the Pcm and CEN data sets are best represented by simple logarithmic equations, whereas the relationships between hardness and CEIIW appear to be best approximated by simple exponential equations. While the equation for hardness versus Pcm has the best fit for the highest cooling rate condition, and the same can be said for the hardness versus CEIIW equation for the lowest cooling rate, the equations relating hardness to CEN give the most consistent predictions of measured hardnesses across the range of cooling rates investigated. Therefore, an empirical relationship based on CEN has been developed in the current work, rather than one based on the alternative carbon equivalent parameter CEIIW

14

and Pcm. The logarithmic relationships describing the lines of best fit for CEN and hardness, and the relevant R2 values are given in Table 4.

Cooling time, t8/5 = 2 sec

y = 259.59Ln(CEN) + 673.04, where R2 = 0.934750

150

250

350

450

550

0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

Carbon Equivalent (%)

Har

dnes

s (H

V)

CE IIW

Pcm

CEN

Log. (CEN)

Figure 7: Graph showing hardness as a function of carbon equivalent for t8/5 value of 2 seconds.

Cooling time, t8/5 = 4 sec

y = 261.46Ln(CEN) + 659.98, where R2 = 0.937150

150

250

350

450

550

0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

Carbon Equivalent (%)

Har

dnes

s (H

V)

CE IIW

Pcm

CEN

Log. (CEN)

Figure 8: Graph showing hardness as a function of carbon equivalent for t8/5 value of 4 seconds.

15

Cooling time, t8/5 = 6 sec

y = 256.04Ln(CEN) + 643.29, where R2 = 0.960350

150

250

350

450

550

0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

Carbon Equivalent (%)

Har

dnes

s (H

V)

CE IIW

Pcm

CEN

Log. (CEN)

Figure 9: Graph showing hardness as a function of carbon equivalent for t8/5 value of 6 seconds.

Cooling time, t8/5 = 8 sec

y = 254.45Ln(CEN) + 635.57, where R2 = 0.961450

150

250

350

450

550

0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

Carbon Equivalent (%)

Har

dnes

s (H

V)

CE IIW

Pcm

CEN

Log. (CEN)

Figure 10: Graph showing hardness as a function of carbon equivalent for t8/5 value of 8 seconds.

16

Cooling time, t8/5 = 10 sec

y = 213.43Ln(CEN) + 549.36, where R2 = 0.8282

50

150

250

350

450

550

0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

Carbon Equivalent (%)

Har

dnes

s (H

V)

CE IIW

Pcm

CEN

Log. (CEN)

Figure 11: Graph showing hardness as a function of carbon equivalent for t8/5 value of 10 seconds. Table 4: Formulae for predicting hardness as a function of CEN at each cooling rate.

Cooling rate Equation R 2

(t8/5)

2 s HV CE N = 259.6 ln(CE N) + 673 0.93

4 s HV CE N = 261.5 ln(CE N) + 660 0.94

6 s HV CE N = 256.0 ln(CE N) + 642 0.96

8 s HV CE N = 254.5 ln(CE N) + 636 0.96

10 s HV CE N = 213.4 ln(CE N) + 549 0.83

The equations for HVCEN (Table 4) have the form aln(CEN)+b where the coefficients a and b are functions of t8/5.The equation developed from these logarithmic relationships in order to predict CGHAZ hardness based on CEN and t8/5 between 2-10 seconds is given by: HVCEN= (-1.3893x2+11.681x+239.58) ln(CEN) + (- 2.4286x2+15.543x+645.6) (17) where x is t8/5 (s), and the two parabolic terms are best-fit relationships for the coefficients of the HVCEN equations in Table 4 as functions of x.

17

This regression model is simpler than the Yurioka model (Equation 8) insofar as it is based on a single carbon equivalent parameter, and the effect of C and alloy composition in general is taken into account through this parameter. It is subsequently referred to as the CEN model. Development of the neural network model Figures 12 and 13 show graphs of measured HAZ hardness versus predicted HAZ hardness for the training data and testing data, respectively. There is no overlap between the two sets of data and the predictions that occur in both instances result in a significantly low RMS error (the square root of the sum of errors between the predicted and experimental hardness for each data case).

R2 = 0.98130

100

200

300

400

500

600

0 100 200 300 400 500 600

Experimental HAZ VHN

Pred

icte

d H

AZ

har

dnes

s (H

V)

Figure 12: Comparison of measured and predicted HAZ hardness values for the neural network model training data (RMS error is 18.0 HV 10).

R 2 = 0 .98250

100

200

300

400

500

600

0 100 200 300 400 500 600

E xperim enta l H A Z V H N

Pred

icte

d H

AZ

har

dnes

s (H

V)

Figure 13: Comparison of measured and predicted HAZ hardness values for the neural network model testing data (RMS error is 17.9 HV 10).

18

The sensitivity of the key input variables is shown in Figures 14 to 16. Figure 14 shows the effect of changes across the range of carbon contents on predicted HAZ hardness, and Figures 15 and 16, respectively, show the effect of changes across the range of original hardness and t8/5 (s) on predicted HAZ hardness.

200

250

300

350

400

450

500

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35C (wt %)

Pred

icte

d H

AZ

hard

ness

(H

V) ∆Η

Figure 14: Neural network model sensitivity data showing the predicted HAZ hardness across the range of carbon contents used in the model.

200

250

300

350

400

450

500

125 150 175 200 225 250

Orginal PM VHN

Pred

icte

d H

AZ

hard

ness

(HV) ∆H

Figure 15: Neural network model sensitivity data showing the predicted HAZ hardness across the range of original PM hardness values used in the model.

19

200

250

300

350

400

450

500

0 2 4 6 8 10 12

t8/5 (s)

Pred

icte

d H

AZ

hard

ness

(HV) ∆Η

Figure 16: Neural network model sensitivity data showing the predicted HAZ hardness across the range of t8/5 cooling times used in the model. Comparison of the predictive models The Yurioka model appears to consistently underestimate the hardness arising from the dilatometer thermal cycles, whereas the CEN model gives a much more reliable prediction of hardness (see Figure 17).

100

200

300

400

500

600

100 200 300 400 500 600

Predicted Hardness (HV)

Mea

sure

d H

ardn

ess (

HV

)

HV YURHV CEN

1:1

Figure 17: Comparison of the predictions of the Yurioka [32] and CEN models for the simulated weld thermal cycle data. In turn, the neural network model provides even higher accuracy in predicting hardness resulting from the dilatometer experiments (see Figure 18). However, it should be noted that

20

the Yurioka model (equation 8) was developed from data acquired from actual welds produced by Tekken testing where as the currently developed models (CEN and ANN) are based on thermal simulations. Therefore, the actual heating and cooling cycles are unlikely to be exactly compared in the two cases.

100

200

300

400

500

600

100 200 300 400 500 600

Predicted Hardness (HV)

Mea

sure

d H

ardn

ess (

HV

)

HV CENHV ANN

Linear ( )

Figure 18: Comparison of the ouput of the CEN and the ANN models for the simulated weld thermal cycle data. The work thus far has succeeded in deriving hardness prediction algorithms that more accurately predict hardness for the conditions and materials studied in the dilatometry work. However, the question remains, how well do these algorithms predict the hardness of actual welds? In Figure 19, the measured and predicted results for CGHAZ hardness are presented for selected Australian pipeline materials A1, A2 and A8. It can be seen that the Yurioka model consistently underestimates the actual weld HAZ hardness. The neural network model (ANN) consistently overestimates the hardness, but the differential is much smaller than for the Yurioka model, and the predictions are close accurate for the A2 and A8 materials. The CEN regression model provides the closest estimates of hardness for the A1 and A2 materials. DISCUSSION The in-service welding of pipeline and tap fitting materials is a critical procedure, involving considerable risk to personnel and infrastructure. As such, a comprehensive appreciation of microstructure and hardness development in the pipe steel is of paramount importance. The current work seeks to enhance the methodologies currently available for predicting maximum HAZ hardness in pipeline steels typical of in-service welding applications in the oil and gas industry.

21

0

50

100

150

200

250

300

350

400

450

A1 A2 A8

Material

Har

dnes

s (H

V)

HV meas (AGL)

HV ANN

HV CEN

HV YUR

Figure 19: The measured hardness, HVmeas, from actual welding trials [29] on A1, A2 and A8 pipeline materials (t8/5 cooling times of 4.3, 3.9 and 3 seconds, respectively) compared to the hardness predicted for the same cooling conditions and compositions by the artificial neural network model (HVNN), the CEN model (HVCEN) and the Yurioka model (HVYUR). While this study provides a valuable analysis of material behaviour under such conditions, it is important to recognize the limitations of the experimental work. For example, the dilatometer is limited to a maximum heating rate of ~50 °Cs-1, which is much slower than the heating rate of an actual weld HAZ and this limitation has implications for the extent of grain growth on re-austenitisation of the steels. The actual heating time expected in a weld thermal cycle over the temperature range from 1000 °C to 1400 °C is likely to be in the order of 1 second [30], whereas the dilatometer takes approximately 8 seconds to achieve heating over this temperature range. The implication here is that there is likely to be greater austenite grain growth during the dilatometer heating cycle than would be expected from typical welding procedures. Further, the current simulations are likely to result in more complete dissolution of alloying elements in the reaustenitised structure, resulting in a harder martensite on cooling. If dissolution is complete then increasing austenite grain size is not likely to further increase the hardness of a fully martensitic structure. However, greater austenite grain size will result in higher hardenability, that is, the capacity to form martensite at lower cooling rates. In this way, larger austenite grains could lead to a greater volume fraction of martensite, and therefore higher hardness at any given t8/5 compared to the CGHAZ structure developed during actual welding. In any case, this will mean that the results will tend to be conservative with respect to hardness prediction, providing data that overestimates hardness in the GCHAZ. This is a difficult effect to quantify, and the authors realise that further work is required to properly investigate this matter. For example, it may be possible to upgrade the dilatometer equipment to increase the capacity for heating and thereby increase the heating rate so that it more closely approximates that of real weld applications. A related problem is that austenite grain coarsening in an actual weld takes place in a steep gradient and it has been suggested that grain growth in the GCHAZ is impeded by the grain size gradient extending

22

into the grain refined HAZ [31]. Thermal simulations in dilatometers and other simulators produce a relatively uniform peak temperature zone with resultant grain sizes that are typically higher than in the actual weld GCHAZ. Another important source of experimental error lies in the simulation of the cooling rate. In real applications, cooling is not linear but exponential, with higher cooling rate at elevated temperatures, reducing to lower cooling rates at lower temperatures. The current work was based on a linear cooling rate across the whole cooling range, such that for a given t8/5 value the cooling rate through the martensitic transformation range is likely to be much higher than for actual welding, underestimating the amount of auto-tempering and resulting in an over estimate of material hardness for a given cooling condition. Again, such a scenario would be expected to result in a conservative estimate of cracking susceptibility. The authors propose to also review the cooling cycle and determine whether it is possible to more accurately simulate the cooling cycle of a weld on the dilatometer equipment. Hardness Values and Microstructures The results show a tendency toward a consistent decrease in hardness with increasing t8/5. Sharper decreases in hardness, such as the marked decrease for the A6 material from t8/5 of 8 to 10 seconds in Figure 4(a), can be attributed to the onset of diffusional transformation products, that is, the structure is no longer fully martensitic. Hardness appears to decrease with increasing t8/5, regardless of whether the structure remains martensitic or never was, as illustrated by A1 and A5 materials in Figures 4(a). In the case of martensitic structures, the decrease in hardness can be accounted for by more pronounced auto-tempering effects as cooling rate decreases through the martensitic transformation range. For structures containing diffusional transformation products, the softening with increasing t8/5 is probably also partially related to auto-tempering effects, as well as to coarsening of the bainitic ferrite structures. For the cooling rates examined, the microstructures were predominantly low carbon martensite, or bainitic ferrite, or mixtures thereof. The martensite is expected to be supersaturated with carbon and prone to auto-tempering because of the extended heating time. Bainitic ferrite is also expected to form with residual carbon supersaturation, although carbon partitioning results in islands of high carbon austenite that can transform at lower temperatures to bainite or martensite. Both the bainitic ferrite and the residual “phase” are susceptible to auto-tempering as cooling rate decreases. The microstructures shown in Figures 5 and 6 appear to correlate well with the hardness measurements. Fine, martensitic plate structures are apparent for high CE materials and short t8/5 times and development of bainitic ferrite structures with increasing t8/5, associated with progressive softening. The A6 and A10 materials are good examples of such a transition, with predominantly martensitic structures apparent for t8/5 values of 2 and 6 seconds, but the development of grain boundary ferrite and bainitic ferrite at t8/5 of 10 seconds (see Figure 5). Relationship between Carbon Equivalent and Hardness – CEN Model The CE formulae indicate the influence of composition on hardenability, that is, the likelihood of formation of martensite on cooling. Increasing alloy content results in an increase in CE values, and this is generally taken as an indication that a greater proportion of martensite will be produced for a given cooling rate, and that a higher hardness level will result for a given cooling rate. The relationships between CE and measured hardness for given cooling rates are shown in Figures 7-11.

23

Of the three CE formulae considered, CEN appears to give the best correlation with hardness for the examined range of cooling conditions from t8/5 of 2-8 seconds, and although this correlation is less reliable at t8/5 of 10 seconds, it provides the most robust approximation over the full range of cooling rates considered. The Pcm formula provides the closest correlation with measured hardness values for the highest cooling rate condition (t8/5 = 2 s). However, as t8/5 increases, the Pcm formula becomes less reliable and the CEN formula becomes more accurate in predicting the hardness levels. The CEIIW formula provides a reasonable indication of hardness, especially at lower cooling rates, but it is generally less reliable than the other formulae. Further, although such a relationship may be acceptable where carbon equivalent is lower than ~0.5, the predicted exponential increase in hardness with increasing carbon equivalent beyond this level would be unrealistic. The finding that the use of CEN gave the best predictions of hardness is consistent with the CEN formula being developed to account for steels with a wide range of carbon contents, from the more modern low carbon steels (eg, <0.1%C) to the older types that predate thermomechanical controlled processing and contain higher levels of carbon (eg, up to 0.29%C in the current work). The CEN formula for hardness developed in the current work is simpler than Yurioka’s formula which involves three CE formula variants and the effect of carbon content being considered both through direct terms and through the CE formulae. Neural Network Model A back-propagation artificial neural network was also successfully used to predict the simulated HAZ hardness values of a number of pipeline and tap fitting steels with respect to a number of input parameters. The 15 input parameters for this model included alloy composition, cooling time (t8/5) and parent material hardness (see Table 2). The levels of agreement for the training data and test data sets are shown in Figures 12 and 13. The errors of the predictions were quite low, with the RMS error in the training data being 18.0 HV10 and 17.9 HV10 in the test data. A sensitivity analysis on selected input variables was also carried out to validate the model and observe the response of the output data (HAZ hardness) over the input range (Figures 14 to 16). Positive values of sensitivity (∆H) indicate an increase in HAZ hardness with an increasing value of the input variable being considered. Negative values of ∆H predict a decrease in HAZ hardness with an increasing value of the input variable. Comparing the effect of changes across the range of the input variable on HAZ hardness values, it is evident that HAZ hardness is most sensitive to carbon content (∆H=188), followed by original parent material hardness (∆H=53) and t8/5 (∆H=-47). Carbon has long been reported to have a strong positive influence on hardness and strength. The carbon content values used in the artificial neural network model ranged from 0.07 to 0.29 wt%, and the predicted hardness variation, ∆H, across this range was 188 (Figure 14). The predicted effect of carbon on hardness in the model is confirmed by the work of Irvine and Pickering [32]. Their work showed a significant increase in hardness with increasing carbon content of various steels that have martensitic and bainitic microstructures, which are the dominant microstructures of the simulated HAZ samples in the current work. Increasing carbon content strongly increases the hardenability of steel [33], thus reducing the capacity to form softer transformation products (such as high temperature bainite and ferrite) and

24

promoting the formation of a predominantly martensitic structure, which is significantly harder than a bainitic or ferritic structure. In fully martensitic structures hardness increases with increasing carbon content mainly due to the increased concentration of carbon in solution [34]. The original (as-received) parent material hardness was incorporated into the model to further define the influence of material characteristics on the output (HAZ hardness). Original hardness was selected because it is a property that relates quantitatively to composition, microstructure, grain size and segregation effects. The hardness of normalised low carbon steels will depend overwhelmingly on carbon content (and hence volume fraction of pearlite) and is therefore expected to correlate with the hardness of martensite produced in the HAZ during ISW. Figure 15 shows that the predicted HAZ hardness increases with increasing original hardness (∆H = 53). Increasing t8/5 (s) from 2 s to 10 s (or decreasing cooling rate from 150°C/s to 30°C/s) results in a ∆H value of –47 (see Figure 16). Hence, reducing the cooling rate (increasing cooling time) results in a significant drop in hardness, although it is not as significant as the effect of decreasing carbon content over the ranges considered. The reduction in hardness as a result of decreasing cooling rate is due predominantly to martensitic microstructures undergoing auto-tempering (if a predominantly martensitic structure is formed); or the formation of other transformation products (such as bainite and/or ferrite) that have lower hardness values. Bainitic and ferritic transformation products may also contain localized regions of martensite in the microstructure, especially at shorter cooling times. Comparison of the models The accuracies of the CEN and ANN models examined in this current work were compared to the Yurioka model for the dilatometer data. It is to be expected that the current models will provide more accurate predictions, since they have been developed on the data being used to “test” them. Indeed, the CEN model has shown to be significantly more accurate than the Yurioka model (see Figure 17). It is also interesting to note that the neural network model provides an even greater accuracy than the CEN model (see Figure 18). As mentioned previously, the Yurioka model was developed using data from the actual welds and therefore its relatively low performance for simulated weld data may not be surprising. Therefore, the performance of the three models was examined against results obtained from actual welding trials on the A1, A2 and A8 pipeline steels conducted by Painter and Sabapathy [29]. The results of this comparison are presented in Figure 19. The Yurioka model significantly underestimates HAZ hardness for all three materials. This is consistent with the application of the Yurioka model to the dilatometer simulated weld thermal cycle data, where a consistent underestimation of hardness was also observed. The two models developed in the current work over-estimate hardness for the A1 steel, with the CEN model providing the most accurate prediction. For the A2 and A8 steels, the neural network model slightly overestimates the hardness while the empirical model slightly underestimates hardness. The significant overestimation of hardness for the A1 steel may be attributable to the heat treatment conditions used in the dilatometry work, which promoted excessive grain coarsening of the prior austenite grains and perhaps more extensive dissolution of carbon in austenite. The A2 and A8 materials had lower carbon contents and so the effects of excessive grain coarsening and/or carbon dissolution on hardness levels are expected to be less significant, as the ANN and CEN models predict.

25

It would be prudent to extend this evaluation of the predictive models to other examples of in-service welds, and the authors intend to pursue this matter. However, on the basis of the current results it can be said that the relationships derived from the current work more accurately predict the measured hardness than the commercially applied Yurioka model. Additionally, the ANN model consistently predicts slightly higher hardness values than those measured and therefore has the advantage of being conservative in that it overestimates the GCHAZ hardness. Moreover, both of the models developed in the current work are clearly more appropriate for in-service welding applications than the Yurioka model. CONCLUSIONS The current work has been successful in utilizing dilatometric simulations of the weld thermal cycles to derive two new hardness prediction models that more accurately predict hardness in the grain-coarsened HAZ for typical in-service welding conditions. These models are shown to provide more accurate estimates of hardness than a widely used commercial hardness prediction algorithm. Further, the ANN model has the advantage, from a practical standpoint, of making conservative predictions, that is, slightly overestimating the actual GCHAZ hardness. It is acknowledged that further work is required to validate the models for a wider range of in-service welding applications, but the work provides a useful demonstration of the potential for improved in-service welding procedures. ACKNOWLEDGEMENTS The current work was conducted as part of a project sponsored by the Cooperative Research Centre for Welded Structures (CRC-WS) and its core partners, the Australian Pipeline Industry Association (APIA) and the Welding Technology Institute of Australia (WTIA). CRC-WS was established and is supported by the Australian Government Cooperative Research Centres Program. Additional in-kind support from Bluescope Steel is gratefully acknowledged. The authors would also like to thank Bill Bruce from the Edison Welding Institute (USA) for providing steels from North American applications for inclusion in the work. It should also be noted that this paper has previously been submitted in similar form for publication in the journal Science and Technology of Welding and Joining. REFERENCES 1. B. Phelps, B.A. Cassie and N.H. Evans, “Welding onto live natural gas pipelines”,

Metal Construction, August, 1976, pp. 350-354. 2. D.J. Hicks, “Guideline for welding on pressurized pipe”, Pipeline and Gas Journal,

March, 1983, pp.17-19. 3. J. Kiefer and R. Fischer, “Repair and hot tap welding on pressurized pipelines”, Proc. of

the 11th Annual Energy Sources Technology Conference, New Orleans, 1988, Publ. by ASME, New York, PD-Vol.14, 1987, pp. 1-10.

4. P.N. Sabapathy, M.A. Wahab and M. Painter, “The prediction of burn-through during in-service welding of gas pipelines”, International Journal of Pressure Vessels and Piping, 77, 2000, pp.669-677.

26

5. M. Painter, “In-service welding of thin-walled, high strength gas pipelines: CRC-WS research”, Proc. of the 1st International Conference on Welding Onto In-Service Petroleum Gas and Liquid Pipelines, Publ. by the Welding Technology Institute of Australia, Wollongong, March, 2000, Paper 21.

6. W. Bruce and J. McHaney, “Lessons to be learned from past in-service welding incidents”, Proc. of the 1st International Conference on Welding Onto In-Service Petroleum Gas and Liquid Pipelines, Publ. by the Welding Technology Institute of Australia, Wollongong, March, 2000, Paper 8.

7. N. Yurioka and H. Suzuki, “Hydrogen assisted cracking in C-Mn and low alloy steel weldments”, International Materials Reviews, Vol.35, No.4, 1990, p.217-249.

8. J.L. Davidson, “Hydrogen-induced cracking of low carbon, low alloy steel weldments”, Materials Forum, 19, 1995, p.35-51.

9. D. Dunne, “A review of the theoretical and experimental background of hydrogen assisted cold cracking of steel weldments”, Proc. of the 1st International Conference on Weld Metal Hydrogen Cracking in Pipeline Girth Welds, Publ. by the Welding Technology Institute of Australia, Wollongong, Australia, March, 1999, p.3-1.

10. AWS D1.1-86, Structural welding code – Steel, AWS, Miami, Florida, 1986. 11. BS5135-1984, Process of arc welding of carbon and carbon-manganese steels, British

Standard Institution, 1984. 12. AS/NZS 1554.1:2000, Structural steel welding – Welding of structural steel, Standards

Australia, 2000. 13. AS 2885.2:2002, Pipelines – Gas and liquid petroleum, Part 2: Welding, Standards

Australia, 2002. 14. WTIA Technical Note 1, Weldability of steel, Publ. by the Welding Technology

Institute of Australia, 1996. 15. D. Uwer and H. Hokne, “Determination of suitable minimum preheating temperature for

the cold-crack-free welding of steels”, IIW Document IX-1631-91, International Institute of Welding, 1991.

16. N. Yurioka and T. Kasuya, Welding in the World, 35, 1995, p.327. 17. WTIA Technical Note 20, Repair of steel pipelines, Publ. by the Welding Technology

Institute of Australia, 1994. 18. M. Cola, W. Bruce, J. Kiefner, R. Fischer, T. Bubenik and D. Jones, “Development of a

simplified weld cooling rate model for in-service gas pipelines”, Pipeline Research Council International Inc., 1991.

19. Thermal Analysis Model for Hot-Tap Welding, Version 4.2, Pipeline Research Council International Inc., 2002.

20. M.J. Painter and V. Tyagi, “Mechanized in-service welding and software development”, CRC-WS Project 00-95 Final Report, Cooperative Research Centre for Welded Structures, Wollongong, Australia, October, 2002..

21. D. Nolan, W. Bruce, P. Grace and D. Dunne, “Weldability issues for in-service pipeline welding”, Proc. of the International Conference on Pipeline Repairs and In-Service Welding, Wollongong, 2003, Published by the Welding Technology Institute of Australia, 2003.

22. W.A. Bruce, “Selecting an appropriate procedure for welding onto in-service pipelines”, Proc. of the International Conference on Pipeline Repairs, Wollongong, Publ. by the Welding Technology Institute of Australia, March, 2001.

23. N. Yurioka, “Physical metallurgy of steel weldability”, ISIJ International, Vol.41, No.6, 2001, pp.566-570.

24. N. Yurioka, H. Suzuki and S. Ohshita, Welding Journal, 62(6), 1983, pp.147-153.

27

28

25. N. Yurioka, M. Okumura, T. Kasuya and H.J. Cotton, “Prediction of HAZ Hardness of Transformable Steels”, Metal Construction, April, 1987, pp.217R-223R.

26. Y. Tomita and K. Okabayashi, Metallurgical Transactions A, 14A, 1983, p.485. 27. NeuralWare, Neural Computing Guide, Technical Publications Group, Pittsburgh, USA,

2003. 28. NeuralWare, Advanced Reference Guide, Technical Publications Group, Pittsburgh,

USA, 2003. 29. M.J. Painter and P. Sabapathy, “In-Service Welding of gas Pipelines”, CRC-WS Project

96:34 Final Report, Cooperative Research Centre for Welded Structures, Wollongong, Australia, 2000.

30. J. F. Lancaster, Metallurgy of Welding, 6th Ed., Abington Publishing, Cambridge, 1999. 31. S. Mishra and T. DebRoy, Acta Mater., Vol. 52, 2004, p.1183. 32. K. Irvine and F. Pickering, Journal of the Iron and Steel Institute, 187, 1957, p.292. 33. R. Honeycombe, Steels: Microstructures and Properties, Edward Arnold, 1981, p. 131. 34. D. Askeland, Ed., The Science and Engineering of Materials, 2nd Ed., Chapman and

Hall, 1990, p.263.