Measurement of fracture toughness in weld metal: results of a European round robin

TWI Bulletin, March/April 1995

After graduating from Cambridge University and obtaining her doctorate from Sheffield University, Isabel Hadley worked as a materials/structural integrity engineer in a number of industries, including steel research, offshore and nuclear power. In 1992, she joined TWI's Engineering Department (now Structural Integrity Department) where she is responsible for the management of projects concerning the integrity of welded structures, especially pressure vessels and pipelines.

Accurate and reproducible measurement of fracture properties, especially in welded joints, is recognised to be one of the cornerstones of structural integrity technology. Isabel Hadley reports.

TWI has been closely involved in the development of fracture mechanics test methods since the 1960s, particularly in the measurement and application of CTOD (crack tip opening displacement), and has recently published the results of an extensive European round robin on fracture toughness of weld metal, ^[1] the results of which are summarised in this report.

In principle, the measurement of fracture toughness ( e.g. critical J, CTOD or K _Ic) in weld metal is very similar to the procedure used for parent materials, formalised in the latest edition of the relevant British Standard ^[2] . Two problems are inherent in the testing of weld metals, however:

• The presence of welding residual stress in as-welded joints is likely to lead to uneven growth of the fatigue pre-crack.
• Test results tend to be highly scattered relative to those of parent material, not least because of the variable microstructure sampled in a weldment. This introduces problems in setting (and/or meeting) sensible specification limits for fracture toughness.

In order to overcome the first problem, a procedure known as local compression has been developed, in which the as-welded specimen is compressed ahead of the machined notch before fatigue pre-cracking, Fig.1. This redistributes welding residual stress and allows the fatigue crack to grow evenly. Regarding the second problem, it was believed that an extensive round robin exercise would be useful in establishing sound statistical data on which decisions could be based.

For this reason, a project was launched (partly sponsored by the European Coal and Steel Community, ECSC) in which a dozen laboratories across the European Union would carry out a series of identical tests on a standard weldment. The overall results were examined at TWI in order to determine the variability between statistics from different laboratories, and also to quantify the scatter associated with a large number of tests.

Standard weldments were made at TWI in a 50mm thick steel plate, using a submerged-arc welding (SAW) procedure; a cross-section through one of the welds is shown, Fig.2. Tests were carried out at two different temperatures (-60 and -20° C), and using three different test geometries:

1. A standard rectangular section (B x 2B) deeply-notched geometry, with the notch sampling the whole of the plate thickness, thus sampling a mixture of weld microstructures (also known as the 'through-thickness notched' geometry).
2. A deeply-notched square section (B x B) geometry in which the notch sampled the same type of weld metal microstructure through the entire specimen thickness (also known as the 'surface-notched' geometry).
3. A shallow-notched square section (B x B) geometry, with notch orientation similar to above, but with the notch sampling the final pass of the weldment. Ignoring any effects of microstructure, the shallow-notched geometry is often associated with higher values of fracture toughness (relative to deeply-notched specimens) because of reduced constraint effects. No standard test procedure yet exists for this type of geometry.

The difference between the three geometries is clearly seen in Fig.3 and 4. Results were analysed separately for each temperature and geometry; however, the detailed results from geometry 1 above only will be described below, because of its practical importance as a standard test geometry. With 12 laboratories participating and testing each geometry in groups of five specimens, a total of 60 datapoints was available for each geometry and temperature tested. Two types of statistical analysis were then carried out on the data:

• All 60 datapoints (for a single geometry and test temperature) were pooled, and the distribution described in statistical terms.
• The results from individual laboratories were compared against each other, to determine whether or not the test was fully reproducible.

An example of the results obtained is shown in Fig.5 using an anonymous reporting format in which the laboratories are labelled A-N. A logarithmic y-axis is used; this is often the best way to present fracture toughness data, since the results tend to be highly scattered. Figure 5 also gives some qualitative indications of the interlaboratory scatter; for example, it is immediately clear that lab J produced consistently lower results than lab E. This aspect of the work will be discussed in more detail later.

Results from all laboratories were next pooled in order to find a statistical description of the distribution of fracture toughness. The best-fit to the data was found to be a log-normal distribution, as shown in Fig.6 (showing results of tests at -60 and -20°C on B x 2B specimens). As expected, fracture toughness was higher at -20°C (where the weld metal properties are in the ductile-brittle transition region) than at -60°C, where lower-shelf behaviour (brittle fracture without significant ductile crack growth) was observed in almost all specimens. A further difference was noted between surface-notched and through-thickness notched specimens tested at the same temperature; an example is shown in Fig.7, where it is clear that the toughness of the surface-notched specimens (sampling only columnar weld metal) was on average higher than that of the through-thickness notched specimens (sampling a mixture of different microstructures).

The grouping of all results in a single log-normal distribution (as in Fig.6 and 7) is useful in quantifying scatter and confirming trends, such as the effect of specimen geometry and temperature on fracture toughness. However, it says nothing about interlaboratory scatter. An important aspect of testing must be that two laboratories in different countries, using the same procedure, should be able to produce statistically indistinguishable results. In order to investigate this aspect of the work, non-parametric statistical methods were used. ^[3] The statistical technique used (the Kruskall-Wallis test) has a number of distinct advantages:

• It is suitable for small samples of results.
• It makes no assumptions about the type of distribution from which the results are drawn.
• It can be used for semi-quantitative or qualitative data ( eg customer satisfaction, product quality) as well as for conventional quantitative data.

The first two features above are, of course, highly relevant to the study of weld metal fracture toughness. Each laboratory contributed only five results to each of the 60-strong distributions shown earlier; hence parametric statistical techniques ( eg standard error of difference, T test) would not have been applicable. Non-parametric tests are carried out not on the measurements themselves but on their ranks. Hence the 60 results from all 12 laboratories and for a single geometry and temperature were pooled and arranged in order of CTOD, as in Fig.6 and 7.

A rank (i) was then assigned to each result, starting at i=1 for the lowest value of toughness and increasing to i=60 for the toughest specimen. A marker was also associated with each rank to show which laboratory produced it. Clearly, if the results from each laboratory are indistinguishable, we would expect results from, say, laboratory A to be scattered throughout the distribution. Conversely, if there are systematic differences between laboratories, then one laboratory's results would tend to be clustered at one end of the distribution.

We have already seen from Fig.5 that there seems to be interlaboratory differences between laboratories E and J, but the Kruskall-Wallis test confirmed this in a quantitative way and showed that there were indeed statistically significant differences between various laboratories. Interestingly, the 'outlier' laboratories were not always the same. Hence, one laboratory's results might be outliers for the tests at 60°C, whilst being scattered evenly over the distribution of tests at -20°C. This suggested that it was not the case that one laboratory that consistently failed to carry out the test correctly (perhaps because of inexperience or systematic equipment calibration errors). This may point to there being an important, but as yet uncontrolled factor affecting the reproducibility of the test results.

Another notable aspect of the project results was that a high percentage (typically 60-70%) were formally 'invalid' according to the test procedure. ^[2] The most usual cause of 'invalid' results was excessive curvature of the fatigue crack front (see Fig.8 for an example). This illustrates one of the problems in testing weld metals; the local compression procedure is intended to relax welding residual stresses, but the exact level of compression required to do so is rather a matter of trial and error, and it would seem that the example shown in Fig.8 was 'over-compressed', leading to a highly bowed crack front. However, the crack shape assessment criteria used were based on a procedure ^[2] which is specific to parent metals and may therefore be excessively stringent for application to weld metal tests.

Clearly, a test procedure which cannot be complied with in practice will lose its reputation. Consequently, it would be necessary either to relax the procedure or to learn how to control the factors which are causing the 'invalid' results to be produced. For this reason, the 'invalid' weld metal test results were studied in order to determine whether they had any systematic effect on the fracture toughness measured, and also to determine whether the 'invalid' results could explain the significant lab-to-lab variability of test results. Two examples of this part of the work are illustrated in Fig.9 and 10, which show the measured CTOD as a function of crack tip curvature for B x 2B specimens at -60 and -20°C. The present test procedure for parent metals would treat all specimens for which curvature exceeded 10% as 'invalid', yet crack front curvature of up to 30% was reported.

Figure 9 shows a slight increase in CTOD as the crack front curvature increases, but Fig.10 shows no overall trend. Examination of all four distributions of fracture toughness showed that there was one positive ( eg Fig.9) and one negative correlation between CTOD and crack front curvature, plus two cases ( eg Fig.10) in which no overall trend was seen. It would, therefore, appear that there was no consistent relationship between 'invalid' results and the measured CTOD. Furthermore, it was shown that the exclusion of all 'invalid' data from the results did not affect the mean value of fracture toughness measured. Neither did it remove the striking interlaboratory variations in results, except in one case (surface-notched specimens tested at -20°C were shown to be homogeneous from lab to lab after 'invalid' data were removed, but this was not the case for the other three sets of data examined). As shown in Fig.11, removal of 'invalid' data simply reduced the number of data points available in the distribution, without having a statistically significant effect on it.

It is, of course, difficult to generalise from the data shown, about what factors influence the outcome of a fracture toughness test, and what a suitable definition of 'invalid' behaviour might be. Other sources of 'invalid' data (such as insufficient fatigue crack growth from the machined notch) were also studied as part of this project and results are reported elsewhere; ^[1] however, the emphasis in this article has been on crack front bowing because of its predominance in the experiments described. The indications so far are that:

• The current definition of 'invalid' behaviour (with respect to crack front bowing) given by BS 7448: Part 1, is so stringent that it excludes some 60-70% of results obtained in this study, without apparently affecting the value of CTOD measured.
• Interlaboratory variation in results is pronounced and persists even when only 'valid' data are considered.

Because of this, a number of recommendations have been made to the working party currently preparing BS 7448: Part 2. Firstly, it is recommended that some of the validity criteria should be relaxed. Secondly, new collaborative research has been proposed, which is aimed at improving interlaboratory scatter and defining more closely, the parameters required to make a reproducible test. This will be put forward to the EU Framework IV programme later this year.

References

Originally published in TWI Bulletin, March 1995