Materials Science & Engineering A 728 (2018) 218–230

Contents lists available at ScienceDirect

Materials Science & Engineering A

journal homepage: www.elsevier.com/locate/msea

A statistical assessment of ductile damage in 304L stainless steel resolved T

using X-ray computed tomography
⁎ ⁎
A.J. Coopera, , O.C.G. Tuckb, , T.L. Burnetta, A.H. Sherrya,c
a
School of Materials, University of Manchester, Oxford Road, M13 9PL, UK
b
National Nuclear Laboratory Limited, Central Laboratory, Sellafield, Cumbria CA20 1PG, UK
c
National Nuclear Laboratory, Birchwood Park, Warrington WA3 6AE, UK

A R T I C LE I N FO A B S T R A C T

Keywords: X-ray computed tomography (XCT) has been shown to reveal the true extent of ductile damage below the
Hot isostatic pressing fracture surface of failed test specimens, which is often significantly underestimated when probed using 2D serial
Stainless steel sectioning techniques and a microscope, since a single plane of material may only exhibit only a handful of
Ductile fracture resolvable voids. In contrast XCT offers the capability to generate large datasets consisting of hundreds, if not
X-ray tomography
thousands, of individually resolvable voids, where each void can be characterised in terms of its distance to
nearest neighbour, distance from the fracture surface, volume, orientation, amongst other parameters, or col-
lectively in the form of void volume fraction. In order to probe the various stages of ductile failure in greater
detail, where it may be possible to directly assign voids to the various stages of ductile failure using their
individual geometric and location characteristics, it is essential that we determine methodologies to analyse such
datasets. In order to give this study a material basis, we have employed the use of statistical methods to char-
acterise differences in ductile void characteristics between two forms of stainless steel of the same grade (304L):
one manufactured by hot isostatic pressing and the other by forging, since it has been shown previously that
these materials display appreciably different fracture toughness properties. Through statistical sampling of the
voids within these materials we aim to shed some light on the differences in mechanistic ductile fracture and the
capabilities of XCT for probing mechanistic failure of materials.

1. Introduction
Fracture of metals can occur via a number of failure modes in-
cluding: catastrophic failure (termed brittle) or via comparatively
controlled tearing (termed ductile). The failure mechanisms are gov-
erned by its metallurgical and microstructural makeup, temperature,
stress state, and loading conditions [1]. Ductile failure occurs via the
nucleation, growth, and coalescence of voids in the material micro-
structure, which nucleate either at the interface between the matrix and
foreign secondary phase particles or via particle cracking [2,3]. It is
generally accepted that these voids grow under the influence of in-
creasing plastic strain and hydrostatic stress [4–7] until their respective
plastic zones interact with those of neighbouring voids; the ligament of
material connecting the two voids fails once it reaches sufficient levels
of strain. During ductile failure ahead of a pre-cracked tip, thousands of
voids form in close proximity to the extending crack. Voids that are
located directly ahead of the crack tip will be subjected to such high
levels of plastic strain that coalescence occurs; voids that experience
coalescence in close proximity to the fracture path through the material
may become part of the fracture surface. In contrast, voids located out
of the direct line of crack propagation will be subjected to lower levels
of plastic strain and as such may only grow a relatively small amount.
While the global fracture path is governed by the plastic zone and ap-
plied stress, the local fracture path is mediated by the statistical dis-
tribution of particles through the microstructure. On average, as dis-
tance increases from the crack tip (or below the fracture surface), one
might encounter almost all forms of void growth, from voids partially
exposed to the fracture surface to voids typically displaying ‘just nu-
cleated’ characteristics at distances far remote from the fracture path,
and this forms the basis for our investigations as we aim to explore the
applicability of X-ray computed tomography (XCT) to probe ductile
failure mechanisms in greater detail.
XCT is an increasingly powerful imaging technique that allows in-
vestigation of the internal structure of both naturally occurring and
engineering materials. XCT is particularly well suited to the study of
ductile damage during material failure, since it allows for the visuali-
zation (and crucially, the measurement) of the voids formed during the
failure process [8–10]. There are few viable methods to characterise

⁎
Corresponding authors.
E-mail addresses: adam.cooper@manchester.ac.uk (A.J. Cooper), olivia.cz.tuck@nnl.co.uk (O.C.G. Tuck).

https://doi.org/10.1016/j.msea.2018.05.036
Received 4 April 2018; Received in revised form 9 May 2018; Accepted 10 May 2018
0921-5093/ Crown Copyright © 2018 Published by Elsevier B.V. All rights reserved.
A.J. Cooper et al. Materials Science & Engineering A 728 (2018) 218–230

Table 1
Elemental composition of HIP (both powder and consolidated form) and Forged 304L.
Grain size Cr Ni Mo Mn Si C O/ppm N/ ppm

304L Spec. (wt%) – 18.5–20.00 9.00–10.00 – < 2.00 < 1.00 < 0.035 200 –
Forged 94 µm 19.40 9.648 0.345 1.654 0.573 0.027 15 817
Powder (wt%) – 19.2 9.44 – 1.37 0.74 0.022 110 –
HIP (wt%) 27 µm 19.5 9.45 0.01 1.33 0.72 0.022 120 840

voids formed during ductile failure in terms of their spatial distribution; materials’ differing fracture toughness properties. Due to the large
one method is to section failed test specimens and image using a high- number of resolvable voids in these test pieces, which all play a part in
resolution microscope such as a scanning electron microscope (SEM). the failure mechanism to a certain degree, we have employed the use of
2D representation of voids can misleading however, since sectioning of statistical models in order to establish relationships between void
material may typically only reveal a handful of voids per slice of ma- characteristics, namely void size, distance to nearest neighbour, and
terial and in addition to this, it is impossible to ascertain an accurate distance from the fracture surface.
measure of void size, with voids appearing smaller than they actually
are since material sectioning will rarely, if ever, split a void directly
2. Experimental
through its mid-point.
3D representation is possible by serial sectioning, either mechani-
2.1. Material and microstructures
cally (e.g. Robo-Met.3D) or through the use of laser ablation [11] or
Plasma focussed ion beam (FIB) [12,13]. Whilst the use of FIB is quite
The material examined was HIP 304L material, supplied by Areva,
widespread, it is limited to small volumes on the order of 50 µm3, this
and forged 304L stainless steel by Creosote Forge et Creusot Mécanique,
route offers the capability to resolve extremely small features and can
Areva (Le Creusot, France). For HIP304L, stainless steel grade 304L
generate crystallographic contrast to see grain structures. Plasma FIB
argon gas atomised powder was heated from ambient temperature to
[14,15] allows the access of increasingly large volumes of material up
1423 K (1150 °C) at a rate of 633 K (360 °C) h−1 and held at 1423 K
to 300 µm3. In contrast, XCT allows for the fast scanning of compara-
(1150 °C) and 104 MPa for a period of 180 min under an inert argon
tively large volumes of material on the order of mm, whilst providing
atmosphere. Post-HIP heat treatment of HIP304L was performed by
microscale resolution, and is most appropriate for this study. XCT al-
heating from room temperature to 1343 K (1070 °C) at 633 K (360 °C)
lows for the scanning of comparatively large volumes of material,
h−1, held for 280 min, and water quenched. Forged 304L pipe was
where determination of void size and location relative to the fracture
subjected to similar heat treatment as the HIP materials (1343 K
surface is governed by the detection limits of the XCT scanner and
(1070 °C), for ca. 250 min) and water quenched.
detector.
The materials’ elemental compositions (wt%) and grain sizes are
We have recently employed the use of XCT to probe the ductile void
tabulated in Table 1; elemental analysis was performed by The Welding
characteristics, and by extension the ductile failure mechanism, of
Institute (TWI, UK) by inert gas fusion technique on an ELTRA ONH-
stainless steel manufactured by forging and hot isostatic pressing (HIP)
2000. Grain size measurements were conducted in accordance with
[16]. Because there has been a substantial amount of work already
ASTM E112-96.[23]
completed looking at comparisons in material behaviour between HIP
The microstructures of the materials studied are shown in Fig. 1.
and forged steels, the reader is directed to [17–19] for mechanistic
HIP 304L displays an appreciable number of fine oxide inclusions in the
insight (microstructures, chemistry, tensile and fracture toughness
microstructure, whereas F304L displayed a relatively cleaner micro-
testing) into the differences between these two materials’ fracture be-
structure. It is these oxide inclusions which are thought to act as in-
haviour and we focus here instead on differences in ductile damage
itiation sites for the formation of voids during ductile failure. The origin
characteristics resolvable using XCT.
and mechanistic effects of these inclusions on fracture have been pre-
HIP is an advanced manufacturing technique which utilizes the
viously reported on in great detail [17,18], and the microstructures are
production and consolidation of metal alloy powder of required
only included here to place some context on the results presented.
chemistry, subjected to high temperature and isostatically controlled
J-R fracture toughness testing was performed on F304L and
pressure, to produce bulk metal that typically displays isotropic mi-
HIP304L, the experimental details and results of which are presented in
crostructure and uniform material properties throughout the bulk
a previous study [24]. Specimens selected from the fracture toughness
component [20,21]. HIP is extremely well suited to the manufacture of
testing for analysis were loaded to an applied J and crack growth of
complex geometric components without the need for additional ma-
742 kJ m−2/0.454 mm and 672 kJ m−2/0.367 mm, for HIP304L and
chining and welding stages and offers the capability to join dissimilar
F304L, respectively. These specimens were selected as they exhibited
materials in a more favourable fashion to conventional arc welding.
the smallest degree of crack growth and loaded to the smallest applied
While HIP is clearly a more advanced form of manufacture over con-
J, in an attempt to assess specimens which were loaded to within the
ventional forging and casting, there remain fundamental questions
Jmax validity limit of 738 kJ m−2, as determined using ASTM
surrounding microstructure [22], material properties, and fracture be-
E1820 [25].
haviour [17] that require thorough mechanistic exploration if HIP is to
Specimens prepared for metallurgical analysis were sectioned,
make an impact in future nuclear manufacturing, which is rightly
mounted, ground, and polished in accordance with the recommended
governed by safety case regulation. This means that new manufacturing
procedures in ASTM practice E3-01 [12]. Electron microscopy was
approaches, regardless of their recognized improved performance and
performed using an FEI Quanta 650 ESEM equipped with a field
reduced costs, must be met with rigorous fundamental metallurgical
emission gun. The SEM was performed under vacuum using a 20 kV
and material assessment if they are to replace more conventional ‘tried
accelerating voltage and a spot size of 4.0 nm, at a working distance of
and tested’ routes to manufacture.
ca. 10 mm.
Herein we have extracted material from the ductile tearing region of
failed fracture toughness test pieces and imaged the cores using XCT so
as to characterise and quantify the level of ductile damage in HIP and 2.2. Tomography specimen preparation
forged variants of 304L stainless steel, in order to shed light on the
Cylindrical core specimens were extracted from failed compact

219
A.J. Cooper et al.
Fig. 1. SEM backscatter electron images showing the microstructure of (a)
F304L and (b) HIP304L, highlighting the increased number of fine oxide in-
clusion in the microstructure of HIP304L compared to F304L.
Fig. 2. EDM sample extraction plan. Three samples were extracted for each specimen within the ductile tearing zone (Left (L), Centre (C) and Right (R)) of a failed CT
half specimen. A fourth sample was extracted behind the fatigue pre-crack away from any ductile damage.
220
Materials Science & Engineering A 728 (2018) 218–230
tension (C(T)) specimens (HIP304L and F304L) using electro-discharge
machining (EDM). The cores measured approximately 0.5 mm in dia-
meter and 20 mm in length. Three cores (L, C, R) were extracted along
the ductile tearing zone of each C(T) specimen, and a single core (A)
was extracted in a region of the machined notch far from any crack
propagation; this was taken in order to detect the presence of any initial
voids within the initial material in order to distinguish between initial
porosity and ductile damage. At the resolution of the XCT scanner,
negligible porosity (defined as less than 2 × 2 × 2 voxels in size) was
detected within the cores extracted from position A. Extraction of the
cores from the ductile tearing region was located as central to the
specimen as reasonably possible, to ensure that the material under in-
vestigation had been subjected to plane strain conditions. The extrac-
tion plan for the cores is shown in Fig. 2.
2.3. 3D X-ray computed tomography
XCT was performed on each core using a Zeiss Versa 500 machine
with a 4X objective, using 100 kV X-rays and 9 W power. 1602 pro-
jections were collected over 360° rotation. An exposure time of 8 s was
used, and an LE3 filter was employed in order to minimise the effects of
ring artefacts. A final voxel size of 0.75 µm was obtained.
Data processing was performed using Avizo® 9.0.0 software. In order
to reduce image noise and improve contrast between voids and the
steel, an edge-preserving smoothing filter (non-local means, Avizo) was
applied to the data, which ensured any contrast differences could be
confidently attributed to the presence of voids in the steel. The im-
mediate surface perimeter of the core was cropped by five voxels to
remove the porous surface as a result of EDM machining of the cores,
which may contribute to the inaccurate identification of voids. Void
segmentation was performed such that any voids in contact with the
fracture surface were assigned to the air surrounding the core. This was
to ensure consistency when segmenting voids close to/in contact with
the fracture surface. Fig. 3 highlights the result of the void segmenta-
tion procedure based on local contrast difference, where the majority of
visible voids are successfully segmented from the surrounding steel.
Void segmentation was performed using Avizo's Top Hat function,
which allows the segmentation of voids from surrounding material
based on differences in local contrast. This was performed by seg-
menting voids out from the surrounding metal in several stages, by
computing a Top Hat image across different kernel sizes: finer voids
were segmented using a kernel size = 3, and larger voids segmented
using a kernel size = 10, since it was challenging to segment all voids
A.J. Cooper et al. Materials Science & Engineering A 728 (2018) 218–230

Fig. 3. Slices of a F304L core showing (a) the contrast difference between voids and steel, and (b) the segmentation of voids using Avizo's Top Hat segmentation
procedure with highlighted voids in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.).

accurately using a single set of Top Hat parameters. In all cases the
kernel shape was cubic, and the neighbourhood was 26 voxels. The
‘contrast boundary’, which governs the actual size of segmented voids,
was controlled manually by adjusting the masking control parameter in
Avizo which controls the greyscale threshold value of the masked
voxels. This was performed manually and is therefore quite user de-
pendent; an appropriate level of masking was defined by comparing
unmasked and masked sections side-by-side, as shown in Fig. 3, and
varying the level of masking such that an optimum compromise was
attained between the capture of smaller voids and larger voids. Because
of the application of a non-local means filter to the data, which sig-
nificantly reduced the amount of noise in the reconstructions, the re-
solved voids exhibited an appreciable local and intense contrast dif-
ference, and because of this the ‘contrast boundary’ was relatively
sharp. Since the contrast boundary is measured over a length scale, the
effect is more pronounced for smaller voids than larger voids. Finally,
segmented voids were filtered in terms of volume, where any voids
measuring less than 2 × 2 × 2 voxels were removed from the analysis
since this was deemed beyond the resolution of the instrument. Given
that the final voxel size was 0.75 µm, this equates to voids smaller than
3.375 µm3 not being included in any analyses.
In order to quantify size, shape, and nearest neighbour distances of
voids, the volume, Feret shape, and spatial coordinates each void were
measured.
Two of the forged 304L cores are shown in Fig. 4, with the fracture Fig. 4. 3D rendered EDM cores extracted from specimen F304L showing a side
surface at the top of the core. Ductile damage, highlighted blue, can be view (fracture surface at the top) of the reconstructed XCT data. Internal voids
seen to extend below the fracture surface. In a similar fashion, two of are shown in blue. (For interpretation of the references to color in this figure
the HIP304L cores are presented in Fig. 5. Three cores from each ma- legend, the reader is referred to the web version of this article.).
terial were extracted from the ductile tearing zones: Only two are
presented for clarity, but similar levels of damage were visible between whilst this is a reasonable approximation further work in this area could
each set of cores. It is these voids that have been subjected to statistical provide further insights. The distance of a void from the fracture surface
analyses detailed in the following sections. The cores were extracted via was calculated from the top surface of the void to the position of the tip
EDM by two different suppliers; one supplier preferred to extract cores of the core parallel to the z-axis which is parallel to the applied stress
using a sectioning approach while the other supplier adopted looping and the long axis of the sample cores. It is acknowledged that definition
the EDM wire in order to extract cylindrical cores, which explains the of the fracture surface requires careful consideration, and the authors
subtly different geometries of the HIP and Forged cores. Despite this, all have recently completed studies on various approaches to measure void
cores exhibited cross sectional thicknesses of ca. 0.5 mm, and the dif- positions relative to a local uneven fracture surface, through which it
fering geometries do not affect the resolution of the XCT scanner. was determined that the current approach was most appropriate for the
materials studied herein. The distance between neighbouring voids was
2.4. Statistical analyses calculated from the surface of one void to the surface of another along
the line constructed through the neighbouring voids’ respective centre
The statistical analyses were all performed using the statistical points.
package R [26]. In the following analyses, all measured voids were Variables which have been extracted from Avizo® 9.0.0 include:
assumed to be spherical in order to simplify void radii calculations,

221
A.J. Cooper et al.
Fig. 5. 3D rendered EDM cores extracted from specimen HIP304L showing a
side view (fracture surface at the top) of the reconstructed XCT data. Internal
voids are shown in blue. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.).
void volume, 3D Feret shape, and the orientation to the z-axis, θ.
Further information and definitions of these variables can be found in
the Avizo documentation. A void's 3D Feret shape was calculated by the
Avizo default program using the ratio of a void's Feret diameters,
measured normal to one another [27]. Feret diameters of a shape are
measured as the distance between multiple pairs of parallel tangents
(bounding boxes) to the shape. This is sometimes referred to as the
“calliper” measure as it uses the same methodology and aids visuali-
zation of the method. Feret shape is defined as values greater than or
equal to 1, where 1 is a sphere. A value of 1 does not indicate perfect
sphericity, only that the lengths of the two normal diameters are equal;
there may be more variation at different positions of parallel tangents
not measured. In general, orientation to the z-axis is defined over 360°,
i.e. [− 180°, 180°] where 0° indicates the orientation is parallel to the z-
axis. Here, orientation is redefined to the range [− 90°, 90°] since there
Fig. 6. Histograms representing distance from the fracture surface for each of the six cores: Forged cores (left) and HIP cores (right).
222
Materials Science & Engineering A 728 (2018) 218–230
will be symmetry about the orientations, for example, an orientation of
180° is the same as 0°; both are parallel to the z-axis.
3. Results
3.1. Exploratory analysis
Fig. 6 shows the frequency of voids as a function of distance below
the fracture surface for each of the six cores, represented as histograms
with 100 µm intervals. Firstly, it is notable that the three HIP304L cores
contain a significantly larger number of voids than the three F304L
cores, with 1265, 1637, and 1049, and 549, 210, and 268 voids, re-
spectively. On average, ca. 4X more voids were identified in the HIP
samples compared to the forged samples. For all of the HIP304L cores
there is a clear decreasing trend in number of voids as a function of
distance from the fracture surface. The lower values very close to the
fracture surface (up to 200 µm) are attributed to the uneven topography
of the fracture surfaces and were removed as part of the segmentation
procedure, whereby any voids in any contact with the fracture surface
were assigned to the air surrounding the core. This procedure will
therefore undoubtedly significantly reduce the amount of measurable
damage, when expressed in terms of void volume fraction, within close
proximity to the fracture surface. This trend is less obvious for the
F304L cores believed to be a result of (a) the cores exhibiting sig-
nificantly fewer voids, and (b) a consequence of the heavily deformed
nature of the F304L fracture surfaces and poor definition of the fracture
surface. This will be explored in greater detail later.
Fig. 7 shows the distance to nearest neighbour (DTNN) between
voids for each of the six cores, represented as histograms with 10 µm
intervals. As previously, there are significantly more voids in the HIP
cores and it is therefore intuitive that the HIP cores should exhibit a
greater number of voids with smaller nearest neighbour distances than
observed in the Forged cores. The fewer voids in the Forged material
appears to have little bearing on masking of the observed trends, with
both materials exhibiting a clear skew in the DTNN distribution towards
the shorter distance region of the histograms.
In order to probe whether the observed skew in DTNN distribution
is related to the distance of the voids from the fracture surface, DTNN
has been plotted as a function of distance from the fracture surface
(DFFS) for both materials (Fig. 8). The HIP cores exhibit a visible trend
whereby the spatial distribution of voids becomes finer as distance
approaches the fracture surface, indicating a concentrated region of
damage in material ca. 400 µm below the fracture surface, and sug-
gesting that voids become more widely dispersed as their distance from
the fracture surface increases. This trend is less clear for the F304L
cores, thought to be a result of fewer resolvable voids in the analysis.
However one of the cores (a) does indicate that there is a slight trend
showing more closely spaced voids within a region of ca. 500 µm to the
fracture surface. These data suggest that there is fracture ‘process zone’
A.J. Cooper et al.
Fig. 7. Histograms of distance to nearest neighbour void for each of the six cores: Forged cores (left) and HIP cores (right).
where damage is concentrated, in which the average void growth is
significantly larger than in the remaining volume of the cores.
3.2. Model fitting and validation
Normal linear regression models [28] were employed in order to
qualitatively identify any relationships between distance from the
fracture surface (DFFS), distance to nearest neighbouring void (DTNN),
void volume (VV), void orientation normal to the z-axis, θ, and Feret
shape. This method fits a model using a least squares method and re-
quires certain assumptions to be met, i.e. a normally distributed re-
sponse variable (the variable being modelled, also termed the depen-
dent variable) and several independent variables (the variables that
make up the model for the response). Transformations of the dependent
variable [29] can be used to meet the assumptions of normal linear
regression. This employs a transformation, such as natural logarithm,
square root, or inversion to ensure a normal distribution of the de-
pendent variable. The details of any transformations are provided when
appropriate.
Fig. 8. Plots of distance to nearest neighbour against distance from the fracture surface for the six cores: Forged cores (a-c) and HIP cores (d-f).
223
Materials Science & Engineering A 728 (2018) 218–230
The models were fitted and validated in R. Model validation is
employed to assess the appropriateness of each of the normal linear
regression [29] models, and was performed through the use of residual
plots, normality plots, and R2 values [28]. For a given valid model, it is
expected that the residuals are randomly scattered, with approximately
95% of points residing between the ± 2 lines; the normality plots
should fit a straight line. R2 is a measure of variability explained by the
fitted model and is often represented as a percentage; a low value of R2
does not necessarily imply a poor model, rather a model that could still
be improved using additional information. R2 can also be used to
compare models for improved fit.
Models have been fitted for a) distance to nearest neighbouring void
(DTNN), b) distance from fracture surface (DFFS) and c) void volume
(VV).
In all models, an indicator variable is included in order to differ-
entiate between the three cores in each of the two types of steel, since
there may be some measure of variation in each core: A, B and C denote
the cores in the Forged steel and D, E and F denote the cores in the HIP
steel. In all cases at least one of the samples is significantly different
A.J. Cooper et al.
from the base core, showing that the indicator is required for a good fit
and there is variation between cores.
3.2.1. Distance to nearest neighbour (DTNN) model
Fitting a linear regression model for DTNN required a 4th root
transformation of the response variable. The distance from the fracture
surface, Feret shape, orientation to the z-axis volume terms were all
found to be significant. Interpretation is challenging with a transformed
response variable, and therefore quantification, in terms of un-
transformed distance to nearest neighbour void, is not possible. The
qualitative interpretations are:
1. The distance from fracture surface has a positive effect on the dis-
tance to nearest neighbour, suggesting that voids further away from
the fracture surface become more widely dispersed, exhibiting
larger nearest neighbour distances.
2. Feret shape has a negative effect; voids become less spherical as they
located closer together.
3. Orientation to the z-axis, θ, has an effect, so orientation and distance
to nearest neighbour are related
4. Using Core A from the Forged steel as the baseline, core C is very
similar and core B is slightly different. The HIP cores, D, E and F,
however all display a negative effect when compared to the forged
steel baseline core showing that voids in the HIP cores are closer
together.
The residual and normality plots for the DTNN model are given in
Fig. 9. The residual plots, (a) and (b), show a sufficiently random scatter
and the normality plot (c) shows a reasonable fit to the linear line. The
R2 value for this model is 35.3% indicating that there remains an ap-
preciable degree of variation in the data that cannot be solely explained
using these variables.
3.2.2. Distance from fracture surface (DFFS) model
Fitting a linear regression model for DFFS requires a cube root
transformation of the response variable. In this model, the significant
terms were found to be distance to nearest neighbour void and Feret
shape.
The interpretation, again without quantification as a result of the
transformation, is:
1. Distance to nearest neighbour void was found to have a positive
effect on the distance from fracture surface, indicating that voids are
Fig. 9. Residual plots (observed minus predicted values against order and predicted values) and normality plot (a graphical representation of the distribution of the
data compared to a Normal distribution) for the distance to nearest neighbour fitted model.
224
Materials Science & Engineering A 728 (2018) 218–230
found closer together as distance approaches the fracture surface.
2. Feret shape has a negative effect on distance from the fracture
surface, thus suggesting more spherical voids will be located at re-
gions remote from the fracture surface.
3. Orientation to the z-axis, θ, has no effect, suggesting that the or-
ientation of voids does not change as distance from the fracture
surface is varied.
4. Using Core A from the Forged steel as the baseline, cores B and C,
have little discernible differences in comparison. The HIP cores, D, E
and F, again display a negative effect in comparison, indicating that
on average, the voids in the HIP cores are located closer to the
fracture surface than those in the Forged cores.
5. Void Volume was not found to be significant, suggesting that there is
no discernible relationship between the volume of voids and the
distance from the fracture surface.
Fig. 10 shows the residual (a, b) and normality (c) plots for this
model. The residual plots both show strong undesirable patterning: the
residuals vs. index plot show a decreasing trend within each of the six
cores and the residuals vs. fitted values plot indicates either a de-
creasing variance or general decreasing trend. This is a sign that the
fitted model is missing one or more important terms. Decreasing var-
iance in the plots of residuals against DTNN and Feret shape (Fig. 11),
indicate that higher order terms may be required. Square terms were
added in an attempt to improve the model but this was found to have
little improvement to the residual plots. The R2 value for this model is
31.6% showing that there remains a large amount of variation in that
data that has not been explained by these variables.
Overall, this analysis reveals the significant terms for DFFS but the
interpretation is only speculative as the residuals indicate important
terms may be missing from the model. One factor that may explain this
lies in the difficulty in defining the fracture surface: it is clear that from
the 3D reconstructed cores definition of the fracture surface is chal-
lenging, since it extends into the core. This is especially true for the
F304L cores, where the fracture surface can be seen to extend into the
core over a region of ca. 600 µm. More details are provided on this later
in the discussion.
3.2.3. Void Volume (VV) model
Fitting the VV model using a log transformation of the response
variable revealed that distance to nearest neighbour and Feret shape
were significant terms.
The interpretations, again without quantification due to the
A.J. Cooper et al.
Fig. 10. Residual (a, b) and normality (c) plots for the distance from fracture surface fitted model.
Fig. 11. Residuals against (a) distance to nearest neighbour void (DTNN) and (b) Feret shape.
transformation, are:
1. Distance to nearest neighbour void has a negative effect; voids with
smaller inter-particle spacing have larger volume.
2. Feret shape has a positive effect; voids with a smaller volume are
more spherical.
3. Orientation to the z-axis, θ, is not significant showing that the vo-
lume of the voids does not change with their orientation.
4. Again, the three HIP cores, D, E and F, are negative in comparison to
the baseline Forged core, and the remaining Forged cores (B, C) are
not significantly different to A; on average the voids in the HIP cores
have smaller volumes than the voids in the Forged cores.
The residual and normality plots for this model are shown in Fig. 12
and show a good degree of random scatter. The R2 value for this model
is only 5.6% showing that there remains appreciable variation in that
data that has not been fully explained by these variables.
Fig. 13 shows the effect of orientation with distance from the frac-
ture surface. Here it is possible to see almost the full range of orienta-
tions at all distances below the fracture surface and helps to illustrate
how orientation is not found to be significant in the model for distance
from the fracture surface.
3.3. Differences in void characteristics between HIP and Forged 304L
While the aforementioned model fitting investigations reveal the
inter-relations between a variety of different void characteristics, in-
teresting for the purpose of probing a general ductile fracture
225
Materials Science & Engineering A 728 (2018) 218–230
mechanism, it is unable to provide quantitative comparisons between
the two materials employed here as a result of the necessary transfor-
mations imposed during the analyses. In order to reveal quantitative
differences in void characteristics between the materials, a different
approach was therefore necessary. Two sample t-tests are often used to
test if the means of two populations are the same [30] and hence can be
employed usefully here to probe whether any of the parameters differ
significantly between Forged and HIP cores, which may help to provide
further evidence of the differences in mechanistic ductile failure. Since
the cores contained different numbers of voids and their variances were
not the same, a modification of the t-test, Welch's T-test [31] was em-
ployed. Welch's two sample t-test is applied automatically in R and
gives a p-value that is used to interpret the result: a p-value less than
0.05 shows that the means of the two samples are statistically different
at the 95% confidence level whereas a value above 0.05 can only show
that there is not enough evidence to prove that the means are different
at the 95% level.
The results of the Welch's two sample T-test on all variables (DTNN,
DFFS, VV, Feret shape and orientation to the z-axis) are found to be
statistically significant, and there is an appreciable difference in the
measured parameters for the two types of steel. The mean values for the
HIP304L and F304L cores, the difference in means (Forged - HIP) and
the 95% confidence intervals of the difference are given in Table 2,
rounded as appropriate. Since the use of Standard Deviation to describe
data spread is only really appropriate for symmetrical distributions,
standard deviations have not been reported here, however the data
distributions can be noted qualitatively in Figs. 14 and 15.
It is evident that appreciable differences lie in the mean DTNN,
A.J. Cooper et al. Materials Science & Engineering A 728 (2018) 218–230

Fig. 12. Residual (a, b) and normality (c) plots for the volume fitted model.

DFFS, and VV between F304L and HIP304L. These differences are
consistent with previous observations into the microstructure and
fracture toughness values: we have shown that this particular heat of
HIP304L displays a 40% reduction in fracture toughness (J0.2BL) of the
forged 304L [32,33]. F304L exhibits a DTNN that is approximately
double that of HIP304L, as well as a mean DFFS that is approximately
double that of HIP304L. It was shown previously that DTNN and DFFS
are related, however we have not investigated whether there is any
significance in this ratio difference between HIP304L and F304L. It is
assumed that analysis and correlation of several materials would be
necessary in order to probe this relationship, however it is intuitive that
there is such a relationship between these two materials. It is thought
that the HIP304L's reduction in DTNN by ca. 50% is related to its re-
duction in fracture toughness (40%) and the similar orders of magni-
tude obtained here are not coincidental.
F304L exhibits fewer initiating particles in the microstructure, and
thus fewer voids during failure, which permits the austenite matrix to
undergo larger levels of plastic deformation before resulting in local
failure. Because of this, the mean DTNN would be expected to be sig-
nificantly larger than that of the HIP304L. By extension, the mean DFFS
would also be expected to be larger for F304L than HIP304L, since the
HIP material fails at lower levels of plastic strain and thus voids located
comparatively far away from the fracture surface may not grow suffi-
ciently large to be resolved using XCT with the current settings, or in

Fig. 13. Distance from fracture surface against orientation to the z-axis, adjusted to the range [− 90°,90°] for all Forged and HIP cores: Forged cores (a-c) and HIP
cores (d-f).

226
A.J. Cooper et al. Materials Science & Engineering A 728 (2018) 218–230

Table 2
Summary of the mean differences between Forged and HIP cores throughout the entire core.
Measure Forged mean HIP mean Lower 95% confidence interval Difference in mean p-values Upper 95% confidence interval

–16
DTNN (μm) 38.3 18.7 17.7 19.6 < 2.2 × 10 21.5
DFFS (μm) 587 319 246 268 < 2.2 × 10–16 292
VV (μm3) 817 246 397 572 < 1.85 × 10−10 746
Feret shape 2.04 1.81 0.178 0.23 < 2.2 × 10–16 0.278
Orientation, θ (°) 9.16 2.81 2.801 6.35 0.00046 9.897

some cases, may not even nucleate at all since they are not subjected to
sufficiently large plastic strains. Furthermore, F304L exhibits a mean
VV more than three times that of HIP340L, again related to the distance
between neighbouring voids: HIP304L fails before voids are able to
grow to such large sizes. These relationships are thus believed to be a
product of the particle spacing of the oxide inclusions in the micro-
structure.
For a distribution representation of the measured differences in void
characteristics between the two materials, Fig. 14 shows the density
distribution plots for each of the variables over their respective range of
values, where density was determined in R by standard smoothing of
the peaks in corresponding histogram plots for the purpose of clarity.
Since the ductile fracture process zone (i.e. region that contains the
most consistent distribution of voids) appears to be within 600 µm from
the tip of the cores, the data within this region was isolated and re-
analysed in order to determine whether or not the fracture ‘process
zone’ exhibits more noticeable differences in void characteristics which
might further explain differences in failure mechanism. Restricting
analysis to this region of the cores revealed even greater differences
between the two materials, except for orientation, and these are sum-
marized in Table 3. For orientation, this can be interpreted as Forged
and HIP cores exhibiting voids with comparable orientations, or both
materials indicate an equally varied spread of orientations within the
first 600 µm.
Fig. 15 shows density distribution plots for each of the variables in
the first 600 µm from the fracture surface, providing a graphical re-
presentation of the differences between the Forged and HIP cores.
4. Discussion
Through using X-ray CT to visualise the extent of ductile damage in
fractured compact tension test specimens, various damage character-
istics can be extracted and analysed enabling the development of our
mechanistic understanding of ductile failure. There are several void
characteristics that are inter-related to one another, which vary in a
predictable manner when subjected to sufficiently large levels of plastic
strain. The work here indicates that as distance approaches the fracture
surface, akin to traversing up the plastic strain gradient in the vicinity
of the crack tip, voids appear closer together (DTNN) and, by extension,
also increase in size (VV).
Void nucleation is governed not only by the applied load at an in-
itiating particle, but also by either the interfacial energy between the
particle and surrounding matrix or particle strength. This energy, which
according to Griffiths must be greater than the surface free energy
created through formation of a void (or crack) [1], is naturally related
to the chemistry of the interfacial bond as well as the particle size. As
such, void nucleation occurs over a distribution of stresses, which re-
sults in some voids nucleating at particles sooner than others, while
others may not nucleate at all if not subjected the necessary conditions.
This effect is more noticeable at distances remote from the crack where
the plastic strain field is reduced, and is reflected by an increase in
distance to nearest neighbour (Fig. 8) and in the frequency of voids
from the fracture surface (Fig. 6). In other words, as plastic strain de-
creases, the proportion of voids subjected to the levels of strain required
to permit either their growth over sufficiently large distances to result

Fig. 14. Density plots of (a) DTNN, (b) DFFS, (c) VV, (d) Feret shape, and (e) orientation to z-axis, for all Forged and HIP cores.

227
A.J. Cooper et al. Materials Science & Engineering A 728 (2018) 218–230

Fig. 15. Density plots of (a) DTNN, (b) DFFS, (c) VV, (d) Feret shape, and (e) orientation to z-axis, for all Forged and HIP cores, with the analysis restricted to the first
600 µm of material below the fracture surface.

in their resolution in XCT is reduced, or, their nucleation is prevented
altogether. As a result, a range of void sizes are observed in failed
specimens, which as we have seen typically increase in size as distance
approaches the fracture surface. In contrast, at distances local to the
crack tip, the stress distribution required to nucleate voids is dominated
by the high levels of plastic strain present in the surrounding matrix,
and therefore the majority of particles in the microstructure initiate
voids within this region, reflected by the increased number of voids and
reduced nearest neighbour distances.
It has been shown previously that equivalently graded HIP and
Forged stainless steel exhibits clear differences in fracture and impact
toughness [18,32], as well as the number of oxide particles in the mi-
crostructure [17], which act as initiation sites for the nucleation and
growth of voids during ductile failure. Whilst the spatial distribution of
these initiating particles varies significantly between HIP and Forged
304L, the flow behaviour of the materials has been previously reported
as being comparable, exhibiting similar strain hardening behaviour
indicating that the rate of void growth should be comparable for spe-
cimens tested under equivalent loading conditions. The rate of void
growth can be estimated using various void growth models, such as
those proposed by Rice and Tracey [5] and McClintock [4] amongst
others, and indeed Rice and Tracey calculations have been shown to
glean comparable results for both HIP and Forged 304L [32]. In order to
highlight differences in the ductile failure mechanism, damage below
the fracture surfaces has been quantified ‘globally’ in terms of void
volume fraction [16]. Significantly, little differences in VVF were ob-
served between the two materials, despite their appreciable differences
fracture toughness initiation and resistance to ductile tearing, sug-
gesting that the ductile fracture mechanism of these two different steels
is not controlled by the volume fraction of voids below the fracture
surface, but is instead dominated by the inter-particle distance of nu-
cleating voids. This is perhaps intuitive, since the total volume fraction
of fewer, larger voids (as seen in the forged specimens) would be ex-
pected to equate to the total volume fraction of more, finer voids, at
void coalescence and fracture initiation, if said voids are assumed to
grow with comparable growth rates, highlighted using the schematic in
Fig. 16.
Indeed, the coalescence stage of failure will occur at different levels
of plastic strain for two comparable materials exhibiting different inter-
particle spacing, but the VVF of ductile damage below the fracture
surface at each material's respective fracture initiation, providing the
materials exhibit similar strain hardening behaviour, should be com-
parable. Significantly, if negligible differences between VVF data are
discernible between the two materials, then significant differences be-
tween the void size, distribution, and frequency might be expected in
order to explain the differences in fracture and impact toughness pre-
viously reported. The data herein has revealed that appreciable differ-
ences are indeed present between HIP and forged 304L when con-
sidering average void volume and nearest neighbour distance, and these
differences govern the fracture toughness properties. HIP304L displays
a marked reduction in average nearest neighbour distance than F304L,
and this is the case when analysing all of the voids throughout the core
(reduction of ca. 50%, Table 2) as well as only the voids considered to
be within the fracture process zone (reduction of ca. 48%, Table 3). This

Table 3
Summary of the mean differences between Forged and HIP cores within the first 600 µm below the fracture surface.
Measure Forged mean HIP mean Lower 95% confidence interval Difference in mean p-values Upper 95% confidence interval

–16
DTNN (μm) 29.9 15.4 12.6 14.5 < 2.2 × 10 16.4
DFFS (μm) 328 232 80 96 < 2.2 × 10–16 111
VV (μm3) 992 231 522 470 < 7.1 × 10−10 999
Feret shape 2.228 1.814 0.344 0.414 < 2.2 × 10–16 0.483
Orientation, θ (°) 6.546 3.664 − 1.640 2.882 0.2103 7.441

228
A.J. Cooper et al.
Fig. 16. Schematic showing how VVF should not be significantly affected by
inter-particle spacing of void-initiating particles for HIP304L (top) and F304L
(bottom).
notable difference in distance to nearest neighbour is clearly related to
the volume fraction of initiating particles in the microstructures of the
two materials, with HIP304L displaying a greater fraction of oxide in-
clusions in the microstructure (Fig. 1 and [17,18]). In addition, voids in
the F304L cores are able to grow on average to much larger volumes
than in HIP304L, again when analysing voids over the entire core (in-
crease of ca. 233%, Table 2) as well as restricting the analysis to only
voids considered to be located within the fracture process zone (in-
crease of ca. 329%, Table 3). This marked difference is a result of the
size and volume fraction of initiating particles in the microstructure;
voids are able to grow to significantly larger sizes in F304L than
HIP304L because there are fewer voids overall in the plastically de-
forming austenite matrix, thereby permitting extension of the plastic
zone to larger volumes of material surrounding the crack. Alternatively,
the reason that HIP304L displays such smaller voids is that these voids
are not provided with the opportunity to grow to such volumes due to
the local ductile crack growth as a result of a marked reduction in void
nearest neighbour distance in close proximity to the crack tip.
The DFFS model, together with any data presented as a function of
distance from the fracture surface, consistently reveals clearer trends
when applied to the HIP specimens than the forged specimens. This
may partly be associated with the fewer number of resolvable voids in
the F304L cores in comparison to the HIP cores, thereby providing in-
sufficient data to glean statistically meaningful conclusions (this argu-
ment could be applied when considering the data presented in Fig. 6 in
the absence of additional information). However, we believe the less
convincing trends associated with F304L DFFS data can be attributed to
the vastly different fracture surface topography between the two ma-
terials. Because of the high levels of plastic deformation observed sur-
rounding the crack tip of F304L (Fig. 4), the fracture surface can be seen
to extend into the core by ca. 600 µm. A consequence of this is that
voids which are located in close proximity to the ‘true’ fracture surface
may in fact be lying nearly 600 µm from the tip of the core: the point
229
Materials Science & Engineering A 728 (2018) 218–230
which has been taken to be the reference point for the fracture surface,
and given that this distance is approximately a third of the total length
of the core, introduces a significant degree of uncertainty in these
analyses. This phenomenon is likely to explain why the density plots in
Fig. 14 and Fig. 15 suggest a relatively even distribution of voids as a
function of distance below the forged fracture surfaces when compared
to the distributions gleaned from the HIP specimens, which show a
clear increase in density as distance approaches the fracture surface.
Because of this issue regarding fracture surface determination of the
forged cores, we believe that the difference in mean distance from
fracture surface between the two materials has little value in this par-
ticular work and is not a useful interpretation for characterising dif-
ferences in damage characteristics.
On average, the voids in the HIP cores are slightly more spherical in
nature than in the Forged cores (Feret shape: 1.81 cf. 2.04, Table 2),
when considering voids throughout the entire cores. This difference in
Feret shape may not appear to be large, but the T-tests returned p-va-
lues on the order of 10–16 indicating that the differences between the
materials were statistically significant. This observation is clearer when
considering only the fracture process zone (Feret shape: 1.814 cf. 2.228,
Table 3), due to the voids in the forged fracture process zone becoming
less spherical.
Finally, there appears to be no detectable effect of distance from the
fracture surface on void z-axis orientation, θ. The model for distance
from the fracture surface does not find orientation to be significant and
when looking at Fig. 13 it is possible to see the full range of θ values at
each point along the core length. When comparing the void orientation
characteristics for HIP and forged materials, the void orientation den-
sity distribution was found not to vary significantly between the two
materials, with both materials containing voids exhibiting orientations
across the full range of θ values relative to the cores’ z-orientation. This
was found to be the case irrespective of whether voids throughout the
entire core were analysed (Fig. 14) or if analysis was restricted in the
fracture process zone (Fig. 15). Because of these little differences in the
distributions, as well as seemingly little preference to specific orienta-
tions with respect to the cores, it is thought that the measured differ-
ences in mean orientation (Tables 2 and 3) between the two materials
do not provide a useful insight into void growth behaviour in their
current state.
5. Conclusions
Herein we have used XCT to probe the ductile damage that occurs
during fracture toughness testing of two forms of 304L stainless steel.
Due to the large datasets generated from the analysis technique, we
have employed the use of statistical modelling to probe the relation-
ships between various void characteristics. These have been used to
explain differences in mechanistic ductile failure between the two
materials. The data highlight relationships between void volume, dis-
tance to nearest neighbour, and distance from the fracture surface, and
the following main conclusions can be drawn:
1. For both materials studied, voids are located closer together and
increase in volume as distance approaches the fracture surface.
2. HIP304L displays significantly more (ca. 4X) damage (when ex-
pressed as number of voids) when compared to F304L.
3. HIP304L displays significantly reduced nearest neighbour distances
between voids below the fracture surface than F304L and average
void volume. These relationships are able to collectively explain
HIP304L's lower fracture toughness when compared to F304L.
Acknowledgements
The authors would like to thank the UK EPSRC and National Nuclear
Laboratory for funding this research project. The X-ray CT experiments
were undertaken within the Henry Mosley X-ray Imaging Facility
A.J. Cooper et al.
(HMXIF) at the University of Manchester, established with funding from
EPSRC through grants EP/F007906/1, EP/F001452/1 and EP/
I02249X/1.
Data availability
The raw/processed data required to reproduce these findings cannot
be shared at this time as the data also forms part of an on-going study.
References
[1] T.L. Anderson, T.L. Anderson, Fracture Mechanics: Fundamentals and Applications,
third edition, Taylor & Francis, FL, USA, 2005.
[2] Z. Chen, C. Butcher, Micromechanics Modelling of Ductile Fracture, Springer,
Netherlands, 2013.
[3] P.F. Thomason, Ductile Fracture of Metals, Pergamon Press, Oxford, England, 1990.
[4] F.A. McClintock, J. Appl. Mech. 35 (1968) 363–371.
[5] J.R. Rice, D.M. Tracey, J. Mech. Phys. Solids 17 (1969) 201–217.
[6] V. Tvergaard, Int. J. Fract. 18 (1982) 237–252.
[7] V. Tvergaard, A. Needleman, Acta Metall. 32 (1984) 157–169.
[8] M. Daly, T.L. Burnett, E.J. Pickering, O.C.G. Tuck, F. Léonard, R. Kelley,
P.J. Withers, A.H. Sherry, Acta Mater. 130 (2017) 56–68.
[9] M. Daly, F. Leonard, A.H. Sherry, ASME 2014 Pressure Vessels and Piping
Conference, American Society of Mechanical Engineers, 2014, pp. V005T011A007-
V005T011A007.
[10] A. Hosokawa, D.S. Wilkinson, J. Kang, M. Kobayashi, H. Toda, Int. J. Fract. 181
(2013) 51–66.
[11] M.P. Echlin, M. Straw, S. Randolph, J. Filevich, T.M. Pollock, Mater. Charact. 100
(2015) 1–12.
[12] M.D. Uchic, L. Holzer, B.J. Inkson, E.L. Principe, P. Munroe, MRS Bull. 32 (2007)
408–416.
[13] M.D. Uchic, M.A. Groeber, A.D. Rollett, JOM 63 (2011) 25–29.
230
Materials Science & Engineering A 728 (2018) 218–230
[14] J. Lindsay, T.L. Burnett, J. Goulden, P. Frankel, A. Garner, B. Winiarski,
P.J. Withers, Microsc. Microanal. 23 (2017) 284–285.
[15] X.L. Zhong, P.J. Withers, X. Zhang, S.B. Lyon, T.L. Burnett, X. Zhou, M.G. Burke,
Microsc. Microanal. 23 (2017) 286–287.
[16] A.J. Cooper, O.G.C. Tuck, T.L. Burnett, A.H. Sherry, Springer International
Publishing, Cham, 2018, pp. 737–754.
[17] A.J. Cooper, N.I. Cooper, J. Dhers, A.H. Sherry, Metall. Mater. Trans. A 47 (2016)
4467–4475.
[18] A.J. Cooper, N.I. Cooper, A. Bell, J. Dhers, A.H. Sherry, Metall. Mater. Trans. A 46
(2015) 5126–5138.
[19] A.J. Cooper, W.J. Brayshaw, A.H. Sherry, Metall. Mater. Trans. A (2018).
[20] H.V. Atkinson, S. Davies, Metall. Mater. Trans. A: Phys. Metall. Mater. Sci. 31
(2000) 2981–3000.
[21] C. Barre, Adv. Mater. Process. 155 (1999) 47–48.
[22] S. Irukuvarghula, H. Hassanin, C. Cayron, M.M. Attallah, D. Stewart, M. Preuss,
Acta Mater. 133 (2017) 269–281.
[23] ASTM E112-96, Standard Test Methods for Determining Average Grain Size, ASTM
International, West Conshohocken, PA, 1996.
[24] A.J. Cooper, J. Dhers, A.H. Sherry, 2016, V06AT06A012.
[25] ASTM E1820, Standard Test Method for Measurement of Fracture Toughness, ASTM
International, West Conshohocken, PA, 2003.
[26] R Development Core Team, R Foundation for Statistical Computing, 2017.
[27] M.H. G, Particle Size Measurements: Fundamentals, Practice, Quality, Springer,
Netherlands, 2009.
[28] S.G. Park, M.C. Kim, B.S. Lee, D.M. Wee, J. Korean Inst. Met. Mater. 46 (2008)
771–779.
[29] G. A, H. J, Data Analysis using Regression and Multilevel/Hierarchical Models,
Cambridge University Press, New York, USA, 2009.
[30] E.A. C, W.W. A, Statistical Analysis Quick Reference Guidebook, Sage Publications,
California, USA, 2007.
[31] W.B. L, Biometrika 34 (1947) 28–35.
[32] A.J. Cooper, R.J. Smith, A.H. Sherry, Metall. Mater. Trans. A 48 (2017) 2207–2221.
[33] A.J. Cooper, W.J. Brayshaw, A.H. Sherry, Metall. Mater. Trans. A 49 (2018)
811–816.