International Journal of Fatigue 151 (2021) 106317

Contents lists available at ScienceDirect

International Journal of Fatigue

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

The effect of defect population on the anisotropic fatigue resistance of

AlSi10Mg alloy fabricated by laser powder bed fusion
Zhengkai Wu a, Shengchuan Wu a, *, Jianguang Bao b, c, Weijian Qian a, Suleyman Karabal d,
Wei Sun c, Philip J. Withers d
a
State Key Laboratory of Traction Power, Southwest Jiaotong University, Chengdu 610031, China
b
Research Institute of Aero-engine, Beihang University, Beijing 100191, China
c
Faculty of Engineering, University of Nottingham, Nottingham NG7 2RD, UK
d
Henry Royce Institute, Department of Materials, The University of Manchester, Manchester M13 9PL, UK

A R T I C L E I N F O A B S T R A C T

Keywords: Metallurgical defects have a critical influence on the anisotropic fatigue resistance of additively manufactured
Additive manufacturing parts under cyclic loading. Here X-ray computed tomography (CT) has been used to characterise the defect
Internal defects population for laser powder bed fusion processed AlSi10Mg alloy and correlated with the tensile and high cycle
X-ray tomography
fatigue (HCF) properties of specimens loaded both parallel and perpendicular to build direction. Despite similar
Kitagawa-Takahashi diagram
Extreme value statistics
tensile strengths, those tested perpendicular to the build direction exhibit a higher elongation and a higher fa­
tigue strength (114 MPa) than those tested parallel to it (45 MPa). The near surface defects preferentially act as
the fatigue crack initiation site for almost all the tested HCF specimens. The large oblate (pancake-shaped)
defects were found to orient primarily within the build plane giving a larger projected area within this plane
leading to a highly anisotropic fatigue strength. Extreme value statistics were used to predict the likely defect
population in the critical near surface region of fatigue samples based on X-ray CT measurements. Finally, a
fatigue performance assessment diagram considering these extreme value defects was established using the
Kitagawa-Takahashi diagram.

is mainly governed by the largest defect within the surface volume

1. Introduction [11–14], while excluding the collaborative effect of residual stress and
surface quality. Regarding the size of defects, it is traditionally repre­
Manufacturing defects, such as the porosity and lack of fusion (LOF) sented by the effective area of the defect using Murakami’s equation in
defects, are generated during additive manufacturing (AM) process and terms of fracture mechanics [15–17]. Particularly, extreme value sta­
have been believed to play a critical role in determining the fatigue tistics (EVS) has been commonly used to estimate the largest defect of a
resistance of AM fabricated metallic structures [1,2]. Since such defects larger part [11,17,18], in which the predicted value can be taken as the
might not be fully eliminated even with various post treatments such as initial defect size used in fracture mechanics approaches to assess the
thermomechanical methods (stress-relieving, annealing, hot isostatic fatigue performance of AM metallic parts [2,17].
pressing, etc.) and surface treatments (mechanical or electrolytic pol­ Further, the build direction (BD) has been found to impact seriously
ishing, shot-peening, vibratory grinding, etc.) [3–5], the effect of defects on the service resistance of AM materials, such as tensile strength and
on fatigue performance has been of great interest. This is both with an fatigue behaviors [11,19–21]. Samples tested normal to the build di­
view to optimizing AM processes and to improve the life prediction and rection were reported to show higher ductility and fatigue strength
reliability of AM metal parts [6–8]. compared to those tested parallel to it[22,23]. During the bulid process,
Previous studies have indicated that the size, location, and shape of large defects are typically retained between AM layers due to insuffi­
defects are essentially responsible for the significant scatter in high cycle cient fusion. These tend to be prolate (pancake shaped) and are more
fatigue (HCF) data of AM metals [9,10]. Generally, larger defects on or detrimental to the fatigue resistance of specimens tested parallel to the
beneath the surface are more likely to drive fatigue failure as a result of build direction than for one tested normal to cause they have a larger
concentrating high local stress [9,10]. In such case, the fatigue strength projected area in this direction leading to a higher stress concentration

* Corresponding author.
E-mail address: [email protected] (S. Wu).

https://doi.org/10.1016/j.ijfatigue.2021.106317
Received 26 January 2021; Received in revised form 8 April 2021; Accepted 7 May 2021
Available online 12 May 2021
0142-1123/© 2021 Elsevier Ltd. All rights reserved.
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Nomenclature Abbreviations
2D Two-dimensional
Symbols 3D Three-dimensional
√areaeff Murakami’s defect size parameter AM Additive manufacturing
√area Actual size of defect BM Block maxima
a, b, c Principal semi-axes of the ellipsoid C(T) Compact tension
C, D Intrinsic factor without defect and sensitive factor of CIR Crack initiation region
defects EVS Extreme value statistics
d Shortest distance between two adjacent defects FCG Fatigue crack growth
e Elongation H Horizontal direction, i.e. normal to the build direction
f Loading frequency H-T Tensile specimen loaded normal to the build direction
Fw Murakami geometry factor H-C(T) Compact tension specimen loaded normal to the build
h Shortest distance between specimen free surface and defect direction
edge H-HCF High cycle fatigue specimen loaded normal to the build
hs Thickness of surface area direction
Nf Fatigue life HCF High cycle fatigue
R Stress ratio (σmin/σmax) ISIA Improved backward statistical inference approach
RA Reduction in area K-T Kitagawa-Takahashi
Sd Surface area of defect LOF Lack of fusion
V Prospective volume under examination MPB Melt pool boundary
V0 Inspection sub-volume. ML Machine learning
Vd Volume of defect POT Peaks-over-threshold
ΔK Stress intensity factor range SIF Stress intensity factor
ΔKth Crack propagation threshold SEM Scanning electron microscopy
Δσ ω Fatigue limit LPBF Laser powder bed fusion
Δσ ω0 Fatigue limit in the absence of defects V Vertical direction, i.e. parallel to the build direction
θ Angle with respect to the build direction V-T Tensile specimen loaded along the build direction
λ, α Location and scale parameter of Gumbel function V-C(T) Compact tension specimen loaded along the build
σ0.2 Yield strength direction
σb Ultimate tensile strength V-HCF High cycle fatigue specimen loaded along the build
σw Basis fatigue limit under R = − 1 direction
Ψ Sphericity CT Computed tomograph
Ф Equivalent diameter

[11,23–25]. For laser powder bed fusion (LPBF) processed aluminum population determined by CT and the critical defects determined by
alloys, large LOF defects are usually generated between subsequent post-mortem fractography, the effect of build direction on the fatigue
build layers due to the poor flowability and high reflectivity of the resistance is discussed. Finally a fatigue strength assessment method
powder to high density laser heat source [25,26]. These defects confer a considering defect distribution is proposed by invoking EVS and
high degree of anisotropic fatigue performance [1,27]. Kitagawa-Takahashi (K-T) diagram [2,17,30].
Conventionally, destructive optical microscopy and scanning elec­
tron microscopy (SEM) of cross sections have been employed to char­ 2. Materials and methods
acterize internal defects in two-dimensional (2D) [11,12]. Recently,
high-resolution X-ray computed tomography (CT) has attracted 2.1. Specimen preparation
increasing attention to quantify three-dimensional (3D) features of in­
ternal defects, providing a promising tool for the characterisation of the BLT-S310 (Xi’an Bright Laser Technologies Co., Ltd) AM forming
anisotropy of defects in AM materials [28,29]. equipment was used to build the AlSi10Mg alloy sample block (110 ×
The aim of this paper is to better understand the impact of defects on 55 × 120 mm3). The AlSi10Mg powder particles used here were ranged
the anisotropic fatigue resistance of LPBF AlSi10Mg alloy structures, in from 15 to 45 μm in diameter (based on the sieving method and tested by
this case a simple wall structure. Because the defect population is SEM) and chemical composition is shown in Table 1. The following
dependent on the geometry additively manufactured rather print the optimal LPBF parameters were used: laser power ~ 360–400 W, scan­
test-pieces directly, fatigue test-pieces are extracted from the LPBF wall ning speed ~ 1200–1500 mm/s, hatch spacing ~ 0.13–0.16 mm, using a
both parallel and perpendicular to the build direction. X-ray CT is first layer thickness of 50 μm and a scan strategy employing a 67◦ angle
used to probe the defect population characteristic of the test-pieces prior between successive layers (see Fig. 1(a)). After AM but prior to
to undertaking standard mechanical test experiments to quantify the machining specimens, the printed block was heated to 300 ◦ C for about
monotonic tensile, HCF performance and fatigue crack growth (FCG) 2 h (SR1 condition according to ASTM F3318-18 standard) to relieve
threshold. Based on the anisotropic characteristics of the initial defect residual stress after which was naturally aged for 2 weeks.

Table 1
Chemical composition of the used AlSi10Mg powder.
Alloying element Si Mg Fe Cu Mn Trace elements Al

wt. % 9.0 ~ 11.0 0.4 ~ 0.6 ≤2.0 ≤0.6 ≤0.35 0.25 Balance

2
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 1. Schematics showing (a) the scan strategy for the LPBF process used to manufacture the AlSi10Mg alloy wall and (b) the orientation of the test-pieces relative
to the build direction (BD).

Fig. 1 presents the schematics of LPBF process strategy and the 2.2. Tensile and fatigue tests
specimen extraction along with the coordinate system used in this work.
As illustrated the compact tension (C(T)) samples were cut in two ori­ To obtain tensile properties for the AM build, three uniaxial tensile
entations. Firstly, so that we could assess the crack propagation tests were conducted in each direction at a strain rate of 0.0002 s-1 on an
threshold in the X direction with the Mode I driving force applied MTS-810 Axial/Torsional machine on test-pieces having the design
relative to the vertical build direction (Z), often referred to the Z-X shown in Fig. 2(a). FCG threshold tests were carried out on C(T) test-
configuration, and secondly, the threshold of crack propagation in the pieces (Fig. 2(c)) using an MTS ACUMEN3 testing machine, and crack
build direction (Z) with Mode I driving force applied relative to the advance was detected using a constant-amplitude sinusoidal load with a
horizontal direction (X), often referred to as the X-Z configuration. For stress ratio R = 0.1 and a frequency f = 10 Hz. The threshold stress in­
simplicity these are referred to as V-C(T) and H-C(T) respectively, which tensity factor (SIF) range, ΔKth, was obtained by the load shedding
relates to the mode I loading direction. The tensile (T) and HCF samples method. The SIF range, ΔK, is taken as the ΔKth till the FCG rate of 1 ×
are similarly referred to as V-T, H-T and V-HCF and H-HCF respectively. 10− 6 to 1 × 10− 7 mm/cycles was reached. Uniaxial HCF testing was
The geometries and dimensions of the uniaxial tensile specimens, HCF undertaken using the test-piece design shown in Fig. 2(b) on a QBG-50
specimens, and C(T) specimens were designed according to ISO high frequency testing machine. A constant-amplitude sinusoidal load
6892–1:2009, ISO 1099:2017, and ISO 12108:2012 standards, respec­ was applied with R = 0.1 and f = 85 Hz. Since the stress level used in the
tively (as seen in Fig. 2). All specimens were also ground and polished HCF tests is generally lower than the yield strength, intense plastic
prior to mechanical tests. deformation of the specimen is avoided, no significant rise in

Fig. 2. Specimen geometry and dimensions (Units in mm) for (a) the uniaxial tensile, (b) the HCF, and (c) compact tension (C(T)) test-pieces.

Table 2
Summary of different experiment types, load direction with respect to the build direction and number of specimens test in each case.
Experiments Material types Load direction Specimen name No. of specimens

Monotonic tensile Tensile specimen Vertical V-T 3

Horizontal H-T 3
Cracking threshold C(T) specimen Vertical V-C(T) 3
Horizontal H-C(T) 3
Uniaxial HCF HCF specimen Vertical V-HCF 22
Horizontal H-HCF 19

3
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

temperature was observed during the HCF tests. The runout limit was set equivalent sphere, i.e. one having the same volume as the defect, and Ψ
to 1 × 107 cycles. After tests, post-mortem fractographic analysis of the denotes the ratio of the surface area of a sphere (with a volume equal to
HCF test-pieces was conducted on a TESCAN MIRA3 SEM. The experi­ the defect volume) to that of the defect applying the following equations
ment and specimen matrix are listed in Table 2. respectively:
√̅̅̅̅̅̅̅̅
3 6Vd
2.3. X-ray computed tomography (CT) Φ= (1)
π
Prior to the HCF tests, all the HCF specimens were scanned to π 1/3 (6Vd )2/3
determine the defect population by X-ray CT at the Henry Moseley X-ray Ψ= (2)
Sd
Imaging Facility at The University of Manchester. A Nikon Metrology
225/320 kV Custom Bay machine was used to scan the 15 mm gauge where Vd is the defect volume and Sd is the defect surface area. Both the
length. A voxel size of 9.0 μm was selected allowing the whole gauge V-HCF and H-HCF specimens contain characteristic lack of fusion (LOF)
region of the specimens to be imaged [28]. Scanning was performed at and gas porosity manufacturing defects; the former being larger and
an 80 kV accelerating voltage, a 62 μA current and a molybdenum more pancake shaped, the latter smaller and more spherical in common
reflection target. 5013 projections were collected with exposure times of with previous observations [9,23].
1000 ms. The 2D projection images were reconstructed via the filtered Regarding the distribution of defect sizes, Fig. 4 indicates that some
back projection algorithm for 2D slices, which were further recon­ very large LOF defects are present up to 1 mm in longest dimension
structed into 3D topography of defects by AVIZO 9.1® software and having an equivalent diameter about 300 μm (see also the LOF defects
quantitative data about defects were analysed. It should be pointed out shown in the insets in Fig. 3). The proportion of defects with an equiv­
that a single voxel can be affected by noise in the X-ray CT data. Hence, alent diameter smaller than 80 μm is more than 90%, showing that small
the smallest distinguishable 3D defect that can be resolved with confi­ (gas porosity) defects dominate the population. Those defects having an
dence is generally taken as about 3 times the pixel size [29]. Accord­ equivalent diameter larger than 160 μm make up less than 1% of the
ingly, only segmented defects with an equivalent diameter greater than population. As shown in Fig. 5, there is a trend for the larger (LOF)
27 μm were included in detailed quantitative analyses. defects to be more irregular and oblate and the smaller (gas porosity)
defects more spherical; when the equivalent diameter is greater than
3. Results 160 μm, the sphericity of the defects tends to below 0.6.
As was expected, the distributions of equivalent diameter (see Fig. 4)
3.1. Defect characterization and the sphericity (see Fig. 5) of the manufacturing defects for the
samples cut in the different directions are essentially the same since they
Fig. 3 shows representative 3D volume renderings of the defects are excised from the same block material. However, it is clear from Fig. 3
within the gauge volume of the test-pieces. It is evident that the LPBF that the larger (LOF) defects are essentially pancake shaped and mostly
AlSi10Mg material contains numerous defects varying in size and shape, lie in the plane of the build. As a result, they project very different
and the distribution of defects is relatively uniform. The effective profiles in the plane perpendicular to the tensile axis for the V-HCF and
diameter (Ф) and the sphericity (Ψ ) can be employed to quantify the H-HCF test-pieces (see Fig. 6 ).
defects size and shape [9,24,23]. Here Ф denotes the diameter of the

Fig. 3. 3D Renderings of the defects for the HCF specimens cut parallel and perpendicular to the build direction with 4 characteristic defects shown at higher
magnification, the false-colour coding is used only to improve clarity.

4
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 4. Distributions of equivalent diameter and the cumulative frequency curve of defects in (a) a V-HCF and (b) a H-HCF test-piece.

Fig. 5. Variation in sphericity as a function of equivalent diameter for the defects in (a) a V-HCF and (b) a H-HCF test-piece.

Fig. 6. All defects projected onto the plane along the loading direction of a V-HCF and a H-HCF specimen, in which the false-colour coding is used only to
improve clarity.

In order to quantify their orientation, each defect is modelled as an angle of the defects. Fig. 7(b) shows the projected area of the same defect
equivalent ellipsoid having principal semi axes a, b, c [24,31]. The co­ at different orientation angles. By definition, the projected area of the
ordinate system for the model is shown in Fig. 7(a). The angle, θ, defect is the largest when θ = 0◦ and decreases with increasing θ.
characterising the angle of the minimum semi-axis of the ellipsoid, c, The orientation angle θ as a function of defect size is plotted in Fig. 8
with respect to the loading direction, is used to define the orientation accordingly. A strong dependence of defect orientation angle θ on build

5
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 7. An equivalent ellipsoid model of the larger ‘pancake’ defects: (a) the coordinate system for the ellipsoid model: z, corresponds to the axial loading direction of
the sample; a, b and c are the three principal semi-axes of the ellipsoid (representing the defect) respectively; the angle, θ, characterising axis c of the ellipsoid
compared to the loading direction of a given specimen; (b) spatial configuration of defect at different orientation angles.

Fig. 8. Defect orientation angle vs. defect size for the V-HCF (red) and H-HCF Fig. 9. Tensile stress–strain responses for samples tested along the build di­
(blue) test-pieces. rection (V (red)) and perpendicular to it (H (blue)).

direction can be observed. Defects of large size (e.g., >160 μm) clearly and plotted in Fig. 10 as S-N curves. It is clearly observed that the V-HCF
show preferred orientation tending towards θ =0◦ for the V-HCF and θ specimens demonstrate significantly lower fatigue strength than the H-
=90◦ for the H-HCF test-pieces. HCF specimens. Fatigue life data in both directions show extensive
scatter, particularly for the H-HCF specimens. In order to give a con­
3.2. Mechanical properties servative and reliable assessment of the fatigue performance, fatigue
P–S–N curves under the confidence of 95% are further plotted in Fig. 10
The engineering stress–strain curves of vertical and horizontal using the improved backward statistical inference approach (ISIA)
specimens are plotted in Fig. 9 respectively and uniaxial tensile prop­ recently proposed by S.C. Wu [35]. Many studies have shown that the
erties are listed in Table 3. It is evident that while the strengths for the variations in the size, shape, location, and distribution of process-
two test orientations are similar, the elongation for specimens tested in induced defects are the main reasons for the large scatter in fatigue
the H direction is about twice that for those in the V direction, in line life data [9,36–38]. Further investigation to the fatigue fractured sur­
with previous results under the same build conditions [23,32–34]. faces was carried out to analyze the effect of defects on fracture
mechanism.
3.3. High cycle fatigue properties Fractography of post-mortem HCF specimens are shown in Fig. 11
exhibiting a fatigue crack initiation region (CIR), a stable crack
The HCF test results in different directions are summarized in Table 4

6
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Table 3
Tensile properties and long fatigue crack growth threshold for the two test orientations.
Orientation Ultimate tensile strength σb /MPa Yield strength σ0.2 /MPa Elongation e/% Area reduction RA /% ΔKth, MPa⋅m1/2

Vertical 273 ± 3.3 181 ± 1.6 6.4 ± 0.6 24 ± 5.2 1.2 ± 0.16
Horizontal 273 ± 2.9 185 ± 5.8 12.1 ± 0.4 36 ± 3.7 1.2 ± 0.12

Table 4
Fatigue test results of LPBF AlSi10Mg run in load-control mode at R = 0.1.
V-HCF Specimens H-HCF Specimens

Specimen No. σmax, MPa Δσ, MPa Nf, Cycles Specimen No. σmax, MPa Δσ, MPa Nf, Cycles

V-HCF04 140 126 46,100 H-HCF19 200 180 125000

V-HCF09 140 126 56,200 H-HCF18 200 180 152000
V-HCF02 140 126 47,100 H-HCF09 200 180 34800
V-HCF22 140 126 69,100 H-HCF05 180 162 155000
V-HCF07 140 126 67,900 H-HCF14 180 162 99000
V-HCF12 100 90 384,100 H-HCF11 180 162 55800
V-HCF13 100 90 316,100 H-HCF13 160 144 120000
V-HCF19 100 90 180,200 H-HCF02 160 144 365400
V-HCF06 100 90 124,800 H-HCF01 160 144 354700
V-HCF21 100 90 188,700 H-HCF06 147 132.3 466400
V-HCF20 80 72 687,800 H-HCF10 147 132.3 244300
V-HCF01 80 72 674,700 H-HCF15 147 132.3 265000
V-HCF18 80 72 648,700 H-HCF17 140 126 345000
V-HCF11 80 72 1,364,800 H-HCF12 140 126 693400
V-HCF05 80 72 669,800 H-HCF16 140 126 835600
V-HCF03 60 54 1,917,300 H-HCF04 147 132.3 >10,000,000
V-HCF14 60 54 3,350,000 H-HCF07 133 119.7 >10,000,000
V-HCF17 60 54 1,726,300 H-HCF03 133 119.7 >10,000,000
V-HCF10 60 54 >10,000,000 H-HCF08 127 114.3 >10,000,000
V-HCF15 60 54 >10,000,000
V-HCF08 50 45 >10,000,000
V-HCF16 50 45 >10,000,000

propagation region and a fast fracture region. In agreement with pre­ 4.2. Anisotropic fatigue behavior
vious observations [9,38], the critical defects (CIR) were found to lie
either at, or just beneath the surface for all specimens. These were used It is noteworthy that, as seen in Fig. 9, the stress strain responses in
to determine the initial flaw sizes for the fatigue modelling. the two test directions are essentially the same below 200 MPa which is
Fig. 12 shows some examples of the LOF defects identified as the CIR the highest stress level used in the HCF tests. By contrast the fatigue
for the samples tested parallel (V) or perpendicular (H) to the build behaviors shown in Fig. 12 are completely different. This reflects the
direction. Significant differences in the shape and size of the defects are tendency for the large LOF defects to lie in build plane (θ=0◦ ). In this
evident: the critical defects are relatively shallow and disc-like for the V- regard, the equivalent size of a 3D defect can be expressed using the
HCF specimens, but deep and narrow for the H-HCF specimens. In this √areaeff parameter widely employed in fracture mechanics [16,37].
study, the variation in defect size and shape is responsible for the scatter Fig. 14(a) outlines our approach to define the defect size as a func­
in HCF life data in accordance with previous observations [9,38]. The tion of its location with respect to the specimen surface or another
scatter is more pronounced for the H-HCF specimens due to the wide neighboring defect. Note that some studies have argued that the initial
variation in the angle of the critical defect to the specimen surface. fatigue crack growth for irregular cracks and defects starts from the
deepest concave corner due to the higher SIF at that point [16,23,37]. As
4. Discussion

4.1. Anisotropic mechanical behavior

To understand the fracture behavior and correlate it to the aniso­

tropic mechanical properties reported in Section 3.2, fracture surfaces of
LPBF AlSi10Mg specimens tested in the two directions relative to the
build direction were investigated by SEM (see Fig. 13). The fracture
surfaces for the V-T specimens are relatively flat and decorated with LOF
defects, while the H-T specimens exhibit greater topographic variation
with fewer LOF defects evident. This difference reflects the much lower
ductility in the former where the fracture seems to be dominated by the
presence of the LOF defects, whereas in the latter exhibits a conventional
cup-cone failure with shear lips around the periphery and void nucle­
ation and growth at the centre indicative of a ductile failure with the
LOF defects playing less of a role in failure. Recent studies have shown
that the coarser cells and lower hardness at the melt pool boundary
(MPB) can lead to tearing along the MPBs [32–34]. It may be that a high
fraction of MPBs contribute to the low elongation of V-T specimens.

Fig. 10. HCF P–S–N curves of H-HCF (blue) and V-HCF (red) specimens.

7
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 11. Representative fracture surfaces for (a) and (b) V-HCF test-pieces and (c) and (d) for H-HCF test-pieces.

a result, the effective size of the defect √areaeff may be more repre­ alloy and that differences in microstructure do not significantly affect
sentative of the tendency for crack initiation than the actual size of the HCF results in the two directions. It also suggests that specimens
defect √area. Taking this approach some of the observed defects are having a SIF close to 1.2 MPa √m (see Table 3) tend towards infinite life
considered in Fig. 14(b). While only four fractured specimens are suggesting that ΔKth is of this order.
illustrated, almost all tests (18 out of 18 failed V-HCF specimens and 14
out of 15 failed H-HCF specimens) displayed similar features originating
from surface, or sub-surface defects. 4.3. Fatigue strength prediction
Taking the √areaeff as the critical defect size, the original fatigue S–N
curve of Fig.10 can be redrawn as Fig. 15(a), showing a strong corre­ 4.3.1. Extreme value statistics of defects
lation between fatigue life and critical defect size for both test-piece As discussed above, there is conclusive evidence that fatigue failure
orientations. In addition, it is clearly seen that the effective defect initiated from the larger defects, particularly when they are on or near
sizes for the V-HCF specimens are significantly larger than those of H- the free surface. Consequently, it is important to be able to estimate the
HCF specimens despite containing the same defect population because likely life on the basis of the defect population and to estimate the
they project a larger area onto the fracture plane. This correlates with chance of failure for a given fatigue loading condition in terms of the
the inferior fatigue lives suggests that it is the larger effective defect size probability of the sample containing a critical defect of a given size. In
(√areaeff) necessarily leads to their lower fatigue strength and shorter order to predict the size of the largest defects from a 3D defect popu­
fatigue life. In order to investigate if this effect can fully explain the lation, two types of sampling strategy based on the EVS can be adopted:
difference between the fatigue behaviour of the V-HCF and H-HCF test- block maxima (BM) and peaks-over-threshold (POT) [11,12]. Here, it is
pieces, all the data is brought together in a single plot in Fig. 15(b) using preferable to perform a fatigue strength prediction based on the
the stress intensity factor (SIF), ΔK [9,38]. By comparing the results maximum defect size near the surface. The BM approach is probably the
from Fig. 15(a) and (b), it can be seen that the fatigue data in different most common and simplest sampling strategy as an EVS. It consists of
directions broadly come together when the effective defect size is taken dividing the specimen into a number of equivalent sub-volumes and
into account. Therefore, it is reasonable to conclude that the defect size then selecting the maximum value in each sub-volume [17,18]. These
√areaeff governs the anisotropic fatigue resistance of LPBF AlSi10Mg extreme values of the largest defect inside each of the sub-volumes can
be expressed using a Gumbel function [39]:

8
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 12. Magnification SEM images of representative critical defects in fracture surfaces for (a), (b), (c) V-HCF specimens and (d), (e), (f) H-HCF specimens.

Fig. 13. Fracture surfaces for (a) V-T and (b) H-T test-piece.

9
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 14. (a) Methodology for calculating the effective defect size √areaeff (represented by the red line) of irregularly shaped defects where h is the shortest distance
between the specimen free surface and the defect edge, hs is the surface depth, and d is the shortest distance between two adjacent defects; and (b) magnified SEM
images of 4 typical critical defects on three V-HCF and one H-HCF fracture surfaces. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

G(z) = exp[− exp(− (z − λ)/α] (3) specimen and a H-HCF specimen respectively. The location and scale
parameters, λ and α in Eq. (4), were determined using the least square
where G(z) is the probability that the defect size is no more than the method based on the linear relationship between (− ln(− ln G(√areaeff)))
maximum value z, and λ and α are the location and scale parameter, and √areaeff, as shown in Fig. 16(a). It is thus found that the maximum
respectively. √areaeff within the ten sub-volumes both from H-HCF and V-HCF
The maximum defects in all sub-volumes are collected and then ar­ specimens can be well described by the Gumbel distribution. Further, G
ranged in ascending order. The extreme defect size √areaeff in the i-th (√areaeff) is normally set to 95% (i.e. a 5% probability that the defect
sub-volume (i = 1, 2, 3, …, n) can be expressed as √areaeffi and √areaeff1 will be larger than this), and then the maximum √areaeff is determined
≤ √areaeff2 ≤ ⋯ ≤ √areaeffn. The cumulative probability of the as 472 μm for H-HCF specimen and 1022 μm for V-HCF specimen in the
maximum defect √areaeffi in the i-th sub-volume can be estimated using: gauge volume respectively (as seen in Fig. 16(b)).
( √̅̅̅̅̅̅̅̅̅ ) ( ( √̅̅̅̅̅̅̅̅̅ )/ ] Note that the estimated results above are the defect sizes when the
G area effi = i/(n + 1) = exp[ − exp − areaeffi − λ α (4)
theoretical probability G(√areaeff) = 95%. In order to infer the likeli­
√̅̅̅̅̅̅̅̅̅ [ ( ( ( √̅̅̅̅̅̅̅̅̅ ))] hood of a critical defect of a given size in our sample, the gauge volume
areaeffi = α⋅ − ln − ln G area +λ (5)
effi
needs to be considered. In fact, because all the critical defects are found
in within a near surface region, it is the volume of this region that should
where √areaeff is the square root of the area of the equivalent ellipsoid
be considered. To estimate the largest defect size in a certain volume of
perpendicular to the stress direction obtained from X-ray CT data (see
LPBF AlSi10Mg alloy, the return period T of the maximum defect in a
Fig. 7). Here the largest defects √areaeffi (i = 1, 2, 3, …, 10) from ten
prospective material volume can be calculated as [39,40]:
sub-volumes (30 mm3) was extracted from X-ray tomograph of a V-HCF

10
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

Fig. 15. (a) S–N curve including only those HCF specimens failed from defects,
(b) SIF range as a function of fatigue life considering the defect size (The size
and colour of the markers indicate the effective critical defect size √areaeff
measured from the fracture surface).
Fig. 16. Distribution of defect sizes √areaeff in two directions: (a) linear fitting
T = V/V0 (6) for λ and α of Gumbel function for the V-HCF (red circles) and H-VCF (blue
triangles) and (b) prediction of the maximum defect size under certain proba­
where V0 is the inspection sub-volume; V is the prospective volume bility for the gauge volume.
under examination. Then the cumulative probability is:
( √̅̅̅̅̅̅̅̅̅ ) ( ( √̅̅̅̅̅̅̅̅̅ )/ ]
G area effV = 1 − 1/T = exp[ − exp − areaeffV − λ α (7)

The maximum defect (√areaeffV) in a certain material volume is

estimated to be:
√̅̅̅̅̅̅̅̅̅
areaeffV = − ln( − ln(T − 1)/T)⋅α + λ (8)
Based on this analysis, the maximum defect in the near-surface vol­
ume of HCF specimens can be predicted. According to the defect location
results from the post-mortem fractography of the HCF specimens, the
surface depth hs is set at 1200 μm, giving near-surface gauge volume
(Vsurface) of 215 mm3. Fig. 17 shows the predicted curves for the
maximum defect size as a function of volume for testing parallel and
perpendicular to the build direction. As one would expect from Fig. 17,
for a given volume the projected area of the largest defect is essentially
twice as large for loading along the build direction compared to
perpendicular to it.
For our test-pieces (Vsurface= 215 mm3), the predicted maximum
defect under certain volume sizes 785 μm and 365 μm for V-HCF and H-
HCF specimens respectively. It is evident that 15 out of 18 V-HCF Fig. 17. Predicted maximum defect size with probability of 95% from the X-ray
specimens and 14 out of 15 H-HCF specimens are under the predicted CT measurements as a function of volume alongside the critical defect sizes
curve. The predicted reliability reaches 83% and 93% for V-HCF and H- observed by fractography for the H-HCF and V-HCF test-pieces.
HCF samples respectively. It is noteworthy that all the three defects
larger than the predicted size in V-HCF specimens are multiple defects
(as seen Fig. 14(b)). In the current X-ray imaging data-based prediction
method, the effect of combined defects has not been considered causing

11
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

0 indicates the impact without defects and D = 0.53 for Al alloys [42].
Considering the stress ratio, a value of Δσ w0 = 188 MPa was obtained
through the Walker equation when R = 0.1 [43].
Fig. 18 shows the improved K-T diagram based on El-Haddad
equation reporting the testing points of the specimens in the different
build directions. It is evident that all failures (except for two V-HCF
specimens) lie above the El-Haddad curve, and the farther from the
curve, the shorter the fatigue life of the specimens.
The benefit of the K-T diagram is that it provides a quantitative
relationship between the defect size and the fatigue limit, so that the
fatigue limit can be predicted in terms of the defect size in the material
[2,17,30]. Considering the probability distribution of defects sizes
√areaeff by EVS (see Fig.16(b)), a fatigue strength assessment approach
expressed by probabilistic K-T diagram is shown in Fig. 18. Accordingly,
the fatigue limit of the LPBF AlSi10Mg alloy in different build directions
under different failure probabilities is predicted taking the mean (50%)
and 95% estimates of the critical defect size obtained from the EVS
Fig. 18. Probabilistic Kitagawa-Takahashi diagram of the material described
by the El-Haddad formulation considering the expected defect distribution (The
approach. The results in Table 5 show that the prediction error for the V-
colour of the markers indicates the fatigue life). HCF specimen is within 10%, while for the H-HCF specimen is over-
conservative by about 35%.
As to why the H-HCF fatigue limits are underpredicted this maybe
the lower predicted reliability of V-HCF specimens than that of V-HCF
because only the defect size and its anisotropy have been included when
specimens.
evaluating their effects on the fatigue performance of LPBF AlSi10Mg
alloy, while the defect shape and location (see Fig. 12(d) – (f)) are not
4.3.2. Kitagawa-Takahashi diagram
considered in this analysis. Recently, as a data-driven method, machine
The aim of defect estimation is to assess the fatigue performance of a
learning (ML) has shown great potential in the field of critical defect
given volume of LPBF material volume on the basis of the largest
prediction and fatigue life prediction [44–46]. By using the ML method,
probable defect. As discussed in Section 4.2, the defects can be consid­
our studies have explored the comprehensive influence of defect loca­
ered as short cracks having an equivalent size equal to their projected
tion, size, and morphology on the fatigue life of LPBF Ti-6Al-4V alloy
area perpendicular to the loading direction. Thus, their impact on the
[44]. Furthermore, the microstructural inhomogeneity and processing
fatigue limit can be predicted by adopting the K-T diagram approach
residual stress in different build directions have been excluded so that
[2,17,30].
the defect size and anisotropy can be clearly separated from complex
The improved K-T diagram described using El-Haddad model based
factors. It follows that more detailed fatigue resistance analysis based on
on the √areaeff parameter can be further depicted as below [2,41]:
the crack initiation and propagation in relation to the stress level, de­
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√̅̅̅̅̅̅̅̅̅
areaeff0 fects features (size, shape, location, and distribution, etc.), microstruc­
Δσw = Δσw0 √̅̅̅̅̅̅̅̅̅ √̅̅̅̅̅̅̅̅̅ (9) ture inhomogeneity and residual stress is required and is the basis of our
areaeff0 + areaeff
on-going work.
( )2
√̅̅̅̅̅̅̅̅̅ 1 ΔKth
areaeff0 = (10) 5. Conclusions
π Fw Δσw0
Tensile, HCF and FCG threshold test campaigns were carried out to
where Fw = 0.65 is the Murakami geometry factor for near-surface de­
investigate the fatigue resistance influenced by the defects inside LPBF
fects. ΔKth = 1.2 MPa⋅m1/2 is the long crack growth threshold (see
AlSi10Mg alloy in both vertical and parallel to build direction. X-ray CT
Table 3). It is noteworthy that Δσw0 represents the fatigue limit of a
was adopted to analyse in three dimensions the size, shape, and location
given volume without defects, and it is a purely theoretical estimation
of the defects. The main conclusions are summarized as follows:
for an ideal material. Pang et al. [42] proposed an empirical model to
estimate the fatigue limit based on the microstructure and tensile
(1) A significant difference was observed in the mechanical proper­
properties as:
ties in test parallel and perpendicular to the build direction. When
σ w = Cσ b − D⋅σ2b (11) loaded perpendicular to the build direction the elongation is
~12%, about twice that for tests parallel to it, while the ultimate
where σ w is the basis fatigue limit under R = − 1; C is intrinsic factor and tensile strength is similar value (~273 MPa) in both directions.
D is fatigue sensitivity coefficient to assumed defects. Here C =

Table 5
Fatigue limit predictions (from Fig. 18) alongside the experimental results (from Fig. 10) under different failure probabilities.
Orientation Predicted values Experimental results Deviation

G(√areaeff) √areaeff/μm Δσw /MPa P Δσw /MPa

Vertical 50% 439 49 50% 45 +8.9%

95% 1022 33 95% 32 +3.1%
Horizontal 50% 215 71 50% 114 − 37.7%
95% 472 48 95% 75 − 36.0%

12
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

(2) The fatigue strength showed significant anisotropy being just (~ References
45 MPa) in the build direction but (~ 114 MPa) perpendicular to
it. All the fatigue failures were the result of initiation from critical [1] Sanaei N, Fatemi A. Defects in additive manufactured metals and their effect on
fatigue performance: a state-of-the-art review. Prog Mater Sci 2020;116:100724.
(LOF) defects located near the free surface. The effective defect [2] Beretta S, Romano S. A comparison of fatigue strength sensitivity to defects for
size √areaeff adjusted to take into account the proximity to the materials manufactured by AM or traditional processes. Int J Fatigue 2017;94(2):
surface or other defects is largely able to explain the anisotropy 178–91.
[3] Uzan NE, Ramati S, Shneck R, Frage N, Yeheskel O. On the effect of shot-peening
bringing the stress intensity factor vs life curves onto a single line. on fatigue resistance of AlSi10Mg specimens fabricated by additive manufacturing
(3) Based on the 3D X-ray CT analysis small (spherical) gas porosity using selective laser melting (AM-SLM). Addit Manuf 2018;21:458–64.
defects and larger pancake shaped LOF defects were observed. [4] Zhang CC, Zhu HH, Liao HL, Cheng Y, Hu ZH, Zeng XY. Effect of heat treatments on
fatigue property of selective laser melting AlSi10Mg. Int J Fatigue 2018;116:
The later are much more important from a fatigue performance 513–22.
viewpoint and occur predominantly between the AM layers. [5] Tammas-Williams S, Withers PJ, Todd I, Prangnell PB. Porosity regrowth during
Their shapes were approximated by oblate ellipsoid and the heat treatment of hot isostatically pressed additively manufactured titanium
components. Scripta Mater 2016;122:72–6.
preferred orientation quantified in terms of the alignment of the
[6] DebRoy T, Mukherjee T, Milewski JO, Elmer JW, Ribic B, Blecher JJ, et al.
smallest principal semi-axis to the build direction. This enabled Scientific, technological and economic issues in metal printing and their solutions.
the projected area of the defect in the plane perpendicular to the Nat Mater 2019;18:1026–32.
loading axis to be calculated. This was found to be significantly [7] Daniewicz SR, Shamsaei N. An introduction to the fatigue and fracture behavior of
additive manufactured parts. Int J Fatigue 2017;94:167.
larger for testing parallel to the build direction than perpendic­ [8] Yap CY, Chua KC, Dong ZL, Liu ZH, Zhang DQ, et al. Review of selective laser
ular to it. melting: Materials and applications. Appl Phys Rev 2015;2(4):041101.
(4) Extreme value statistics has been used to predict the effective size [9] Hu YN, Wu SC, Withers PJ, Zhang J, Bao HYX, Fu YN, et al. The effect of
manufacturing defects on the fatigue life of selective laser melted Ti-6Al-4V
(√areaeff) of the largest defects on the basis of X-ray CT mea­ structures. Mater Des 2020;192:108708.
surements. Focusing on the critical near surface region this [10] Serrano-Munoz I, Buffiere J-Y, Mokso R, Verdu C, Nadot Y. Location, location &
approach has been used to predict the largest defect likely to be size: defects close to surfaces dominate fatigue crack initiation. Sci Rep 2017;7:
45239.
observed in this region without having to X-ray CT scan each [11] Tang M, Pistorius PC. Fatigue life prediction for AlSi10Mg components produced
sample. The 95% predicted bound for the largest defect size is by selective laser melting. Int J Fatigue 2019;125:479–90.
greater 83% and 93% of the critical defects determined by frac­ [12] Romano S, Abel A, Gumpinger J, Brandão AD, Beretta S. Quality control of
AlSi10Mg produced by SLM: Metallography versus CT scans for critical defect size
tography for the V-HCF and H-HCF specimens, respectively. assessment. Addit Manuf 2019;28:394–405.
(5) The relationship between the defect size √areaeff and fatigue [13] Benedetti M, Fontanari V, Bandini M, Zanini F, Carmignato S. Low- and high-cycle
limit has been established using the probabilistic K-T diagram. fatigue resistance of Ti-6Al-4V ELI additively manufactured via selective laser
melting: Mean stress and defect sensitivity. Int J Fatigue 2018;107:96–109.
Based on the expected defect size √areaeff obtained by extreme
[14] Anderson CW, Shi G, Atkinson HV, Sellars CM, Yates JR. Interrelationship between
value statistics, the K-T diagram is shown to give good predictions statistical methods for estimating the size of the maximum inclusion in clean steels.
of the fatigue strength for the LPBF AlSi10Mg alloy test-pieces in Acta Mater 2003;51(8):2331–43.
the two measurement directions. [15] Yadollahi A, Mahtabi MJ, Khalili A, Doude HR, Newman JC. Fatigue life prediction
of additively manufactured material: Effects of surface roughness, defect size, and
shape. Fatigue Fract Eng Mater Struct 2018;41(7):1602–14.
This work has focused on the influence of the large LOF defects [16] Murakami Y. Metal fatigue: effects of small defects and nonmetallic inclusions.
which can be significant for many AM methods. The use of X-ray CT to Oxford: Elsevier Science Ltd; 2002.
[17] Romano S, Brandão A, Gumpinger J, Gschweitl M, Beretta S. Qualification of AM
establish defect populations and to use this as the basis for predicting the parts: extreme value statistics applied to tomographic measurements. Mater Des
fatigue properties is shown to be a good way of estimating fatigue life 2017;131:32–48.
and could be used to predict the life of AM parts without the need for [18] Hu YN, Wu SC, Wu ZK, Zhong XL, Ahmed S, Karabal S, et al. A new approach to
correlate the defect population with the fatigue life of selective laser melted Ti-6Al-
very large AM processing trials accompanied by fatigue testing pro­ 4V alloy. Int J Fatigue 2020;136:105584.
grams. Instead, X-ray CT could be used to help refine the AM process to [19] Xie Y, Gao M, Wang F, Zhang C, Hao K, Wang H, et al. Anisotropy of fatigue crack
minimise the size and optimise the location of large defects before growth in wire arc additive manufactured Ti-6Al-4V. Mater Sci Eng A 2018;709:
265–9.
proceeding to more modest fatigue testing programs. [20] Lu JX, Chang L, Wang J, Sang LJ, Wu SK, Zhang YF. In-situ investigation of the
anisotropic mechanical properties of laser direct metal deposition Ti6Al4V alloy.
Declaration of Competing Interest Mater Sci Eng A 2018;712:199–205.
[21] Xie C, Wu SC, Yu YK, Zhang HO, Hu YN, Zhang MB, et al. Defect-correlated fatigue
resistance of additively manufactured Al-Mg4.5Mn alloy with in situ micro-rolling.
The authors declare that they have no known competing financial J Mater Process Technol 2021;291:117039.
interests or personal relationships that could have appeared to influence [22] Suryawanshi J, Prashanth KG, Scudino S, Eckert J, Prakash O, Ramamurty U.
the work reported in this paper. Simultaneous enhancements of strength and toughness in an Al-12Si alloy
synthesized using selective laser melting. Acta Mater 2016;115:285–94.
[23] Larrosa NO, Wang W, Read N, Loretto MH, Evans C, Carr J, et al. Linking
Acknowledgements microstructure and processing defects to mechanical properties of selectively laser
melted AlSi10Mg alloy. Theor Appl Fract Mec 2019;98:123–33.
[24] Nadot Y, Nadot-Martin C, Kan WH, Boufadene S, Foley M, Cairney J, et al.
The authors acknowledge supports from the Joint Fund of Large- Predicting the fatigue life of an AlSi10Mg alloy manufactured via selective laser
scale Scientific Facility of the National Natural Science Foundation of melting by using data from computed tomography. Addit Manuf 2019;32:100899.
China (U2032121). S.C. Wu as an honorary staff professor is also [25] Maskery I, Aboulkhair NT, Corfield MR, Tuck C, Clare AT, Leach RK, et al.
Quantification and characterisation of porosity in selectively laser melted Al-Si-
grateful for his visiting position at the Henry Royce Institute for 10Mg using X-ray computed tomography. Mater Charact 2016;111:193–204.
Advanced Materials funded by EPSRC through EP/R00661X. P.J. [26] Awd M, Siddique S, Johannsen J, Emmelmann C, Walther F. Very high-cycle
Withers is grateful for funding through an ERC Advanced Grant (COR­ fatigue properties and microstructural damage mechanisms of selective laser
melted AlSi10Mg alloy. Int J Fatigue 2019;124:55–69.
REL-CT) agreement number 695638. The X-ray facilities were made
[27] Read N, Wang W, Essa K, Attallah MM. Selective laser melting of AlSi10Mg alloy:
available through the Henry Moseley X-ray Imaging Facility, through process optimisation and mechanical properties development. Mater Des 2015;65:
grants EP/F007906/1, EP/F001452/1, EP/I02249X, EP/ M010619/1, 417–24.
[28] Du Plessis A, Yadroitsev I, Yadroitsava I, Le Roux SG. X-ray microcomputed
EP/F028431/1 and EP/M022498/1, EP/T02593X/1.
tomography in additive manufacturing: a review of the current technology and
applications. 3D Print Addit Manuf 2018;5:227–47.

13
Z. Wu et al. International Journal of Fatigue 151 (2021) 106317

[29] Withers PJ, Bouman C, Carmignato S, Cnudde V, Grimaldi D, Hagen CK, et al. X-ray [38] Tammas-Williams S, Withers PJ, Todd I, Prangnell PB. The influence of porosity on
computed tomography. Nat Rev Methods Primers 2021;1:18. fatigue crack initiation in additively manufactured titanium components. Sci Rep
[30] Tenkamp J, Koch A, Knorre S, Krupp U, Michels W, Walther F. Defect-correlated 2017;119:160–72.
fatigue assessment of A356–T6 aluminum cast alloy using computed tomography [39] Beretta S, Murakami Y. Statistical analysis of defects for fatigue strength prediction
based Kitagawa-Takahashi diagrams. Int J Fatigue 2018;108:25–34. and quality control of materials. Fatigue Fract Eng Mater Struct 1998;21(9):
[31] Vincent M, Nadot-Martin C, Nadot Y, Dragon A. Fatigue from defect under 1049–65.
multiaxial loading: defect stress gradient (DSG) approach using ellipsoidal [40] Wu ZK, Wu SC, Zhang J, Song Z, Hu YN, Kang GZ, et al. Defect induced fatigue
equivalent inclusion method. Int J Fatigue 2014;59:176–87. behaviors of selective laser melted Ti-6Al-4V via synchrotron radiation X-ray
[32] Tang M, Pistorius PC. Anisotropic mechanical behavior of AlSi10Mg parts tomography. Acta Metall Sin 2019;55(7):811–20.
produced by selective laser melting. JOM 2017;69(3):516–22. [41] El Haddad MH, Topper TH, Smith KN. Prediction of non propagating cracks. Eng
[33] Xiong ZH, Liu SL, Li SF, Shi Y, Yang YF, Misra RDK. Role of melt pool boundary Fract Mech 1979;11(3):573–84.
condition in determining the mechanical properties of selective laser melting [42] Pang JC, Li SX, Wang ZG, Zhang ZF. General relation between tensile strength and
AlSi10Mg alloy. Mater Sci Eng A 2019;740:148–56. fatigue strength of metallic materials. Mater Sci Eng A 2013;564:331–41.
[34] Rosenthal I, Stern A, Frage N. Strain rate sensitivity and fracture mechanism of [43] Dowling NE, Calhoun CA, Arcari A. Mean stress effects in stress-life fatigue and the
AlSi10Mg parts produced by Selective Laser Melting. Mater Sci Eng A 2017;682: Walker equation. Fatigue Fract Eng Mater Struct 2009;32(3):163–79.
509–17. [44] Bao HYX, Wu SC, Wu ZK, Kang GZ, Peng X, Withers PJ. A machine-learning fatigue
[35] Li CH, Wu SC, Zhang JY, Xie LY, Zhang YH. Determination of the fatigue P-S-N life prediction approach of additively manufactured metals. Eng Fract Mech 2021;
curves – A critical review and improved backward statistical inference method. Int 242:107508.
J Fatigue 2020;139:105789. [45] Zhang XC, Gong JG, Xuan FZ. A deep learning based life prediction method for
[36] Biswal R, Zhang X, Syed AK, Awd M, Ding J, Walther F, et al. Criticality of porosity components under creep, fatigue and creep-fatigue conditions. Int J Fatigue 2021;
defects on the fatigue performance of wire + arc additive manufactured titanium 148:106236.
alloy. Int J Fatigue 2019;122:208–17. [46] Zhan ZX, Li H. Machine learning based fatigue life prediction with effects of
[37] Domfang Ngnekou JN, Nadot Y, Henaff G, Nicolai J, Kan WH, et al. Fatigue additive manufacturing process parameters for printed SS 316L. Int J Fatigue 2021;
properties of AlSi10Mg produced by additive layer manufacturing. Int J Fatigue 142:105941.
2019;119:160–72.

14