International Journal of Fatigue 160 (2022) 106884

Contents lists available at ScienceDirect

International Journal of Fatigue

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

Defect tolerant fatigue assessment of AM materials: Size effect and

probabilistic prospects
Xiaopeng Niu a, Shun-Peng Zhu a, b, *, Jin-Chao He a, Ding Liao a, c, José A.F.O. Correia c, d,
Filippo Berto e, Qingyuan Wang f, g, *
a
School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China
b
Institute of Electronic and Information Engineering of UESTC in Guangdong, Dongguan 523808, China
c
CONSTRUCT, Faculty of Engineering, University of Porto, Porto 4200-465, Portugal
d
INEGI, Faculty of Engineering, University of Porto, Porto 4200-465, Portugal
e
Department of Mechanical and Industrial Engineering, Norwegian University of Science and Technology, 7491 Trondheim, Norway
f
MOE Key Laboratory of Deep Earth Science and Engineering, College of Architecture and Environment, Sichuan University, Chengdu 610065, China
g
Advanced Research Institute, Chengdu University, Chengdu 610106, China

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

Keywords: Despite the significant advantages such as rapid prototyping of complex structures, engineering application of
Extremum value statistics Additive Manufacturing (AM) technique has been limited due to the weak fatigue performance and lacking of
Size effect fatigue design rules of AM parts. Particularly, this application relies on the improvement of AM processing
Fatigue strength
technique and integrity assessment of AM parts, the latter usually demands costly and time-consuming fatigue
Additive manufacturing
Defect-tolerant assessment
tests, especially for full-scale tests. Accordingly, the size effect and fatigue performance scatter in AM materials
are explored on basis of defect data scanned by CT and experimental data in this work. Firstly, the distribution of
maximum defects of AM materials under size effect is extrapolated via extreme value statistics theory; then a
strength altering factor is proposed to characterize the effect of defects on fatigue scatter and size effect. From
essential structure of fatigue curves, the relationship among defects, fatigue strength and fatigue curves are
quantitatively analysed to extrapolate the probabilistic fatigue curves for AM materials; finally, three series of
AlSi10Mg specimens with different processing parameters and gauge volumes prove the method effectiveness
that the “location” and “scale” of fatigue curves is successfully predicted, and defect-tolerant assessment for AM
parts is performed by employing the predicted fatigue curves.

located on the surface. The type of AM technique and corresponding

1. Introduction processing parameters both alter the defects and microstructure of AM
parts, and therefore influence the fatigue performance [3,4]. To better
Additive manufacturing has attracted increasing attentions due to its understand these relations, Liu et al. [5] reviewed the fatigue properties
lower energy consumption and rapid prototyping of complex parts [1], of AM Ti-6Al-4V alloy under different processing parameters, printing
and it is expected to play an important role in manufacturing industries orientations, and the stress ratio, etc., which clarifies the influence of
in reality. However, the application of AM is still limited due to the such factors on fatigue performance of AM parts.
inherent existence of defects produced during processing, which To optimize AM technology and improve the fatigue performance of
significantly affects the integrity of AM parts. Therefore, characterizing AM parts, not only for the processing parameters but also the influence
the influential defective attribute on fatigue performance of AM parts of defects on fatigue performance of AM parts should been intensively
are essential for widespread application and structural integrity design. studied. Recently, various studies on defect features such as size, loca­
In general, AM parts containing inherent defects such as porosity, tion, shape and density have been carried out [6–8]. Results show that
inclusions, oxides and micro-cracks [2], which often become potential defect size plays a dominant role [9], and the defects can be treated as
crack initiation sites due to stress concentration, especially for those cracks. From this point of view, Murakami-Endo [10] and El-Haddad

* Corresponding authors at: School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China
(S.-P. Zhu). MOE Key Laboratory of Deep Earth Science and Engineering, College of Architecture and Environment, Sichuan University, Chengdu 610065, China (Q.
Wang).
E-mail addresses: [email protected] (S.-P. Zhu), [email protected] (Q. Wang).

https://doi.org/10.1016/j.ijfatigue.2022.106884
Received 2 February 2022; Received in revised form 26 March 2022; Accepted 27 March 2022
Available online 29 March 2022
0142-1123/© 2022 Elsevier Ltd. All rights reserved.
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Nomenclature S0 plane area of polished section

Δσw0 fatigue strength in the absence of defects
√̅̅̅̅̅̅̅̅̅
area effective defect size C1 material parameter depends on defect position
hmean average size of the defects recorded in S0 S stress amplitude
x exceedance over the threshold A, B material constants of Basquin equation
u threshold on defect size ε strain amplitude
γ shape parameters of Generalized Pareto distribution σ ’f fatigue strength coefficient
σ scale parameters of Generalized Pareto/normal ε’f fatigue ductility coefficient
distribution
b fatigue strength exponent
N number of defects exceeding the threshold
c fatigue ductility exponent
T return period of the maximum defect
Nf number of cycles to failure
V prospective material volume
Nd design life
ρ average density of exceedances
E Young’s modulus
xmax,V modal value of the maximum defect
K’ cyclic strength coefficient
G limit state function based on load–life interference
n’ cyclic strain hardening exponent
xp,V p-th percentile of the maximum defect distribution
AM additive manufactured
Vsurf surface volume
HSV highly stressed volume
Vg gauge volume
MLE maximum likelihood method
Vint internal volume
BM block maxima method
VPS control volume in Block maxima approach
POT peak over threshold method
lg , rg length of specimen gauge and radius
HSSV highly stressed surface volume
h thickness of specimen gauge
SIF stress intensity factor
a average experimental radius of the critical defects
LEFM linear elastic fracture mechanics
σc,max maximum local stress
GP generalized Pareto distribution
σc,n% n-th percentile of maximum stress
ϑ, SAF strength altering factor
Δσ w fatigue strength (stress range)
L-PBF laser powder bed fusion
ΔK crack propagation threshold
CV coefficient of variation
Kf notch fatigue factor
EBM electron beam melting
ΔKth,lc long crack propagation threshold DMLS direct metal laser sintering
√̅̅̅̅̅̅̅̅̅
area0 El-Haddad material size parameter

[11] evaluated fatigue strength of castings via fracture mechanics-based statistical analysis of pores, intermetallic particles and grains. Note from
approaches. In references [12,13], a quantitative evaluation method of [21–25] that the defects can be integrated with the damage coupled
fatigue strength was explained and presented based on the parameter constitutive equations, where the influences of damage on material
√̅̅̅̅̅̅̅̅̅
area (square root of the projected area of an inclusion). Later, the properties are evaluated by introducing a damage variable to evaluate
√̅̅̅̅̅̅̅̅̅
parameter area has been widely used for fatigue strength evaluation of the progressive deterioration of material.
AM parts considering similar features of defects. To describe the influ­ As an essential matter of properties transfer from small specimens to
ence of defects on fatigue behavior of AM parts, Romano et al. [14] real components, size effect [26] requires an in-depth study. Extreme
elaborated the competition between surface and internal defects, and value statistics were utilized to extrapolate the distribution of maximum
demonstrated the relation between fatigue performance and defects in defects in AM parts under size effect [27]. In addition, Le et al. [28]
√̅̅̅̅̅̅̅̅̅
AM fatigue by employing parameter area. Recently, by considering the clarified the origin of the observed size effect on fatigue performance of
concomitant effect of defect and notch stress gradient, Benedetti et al. Ti-6Al-4V alloys using laser powder bed fusion (L-PBF). However, these
[15] raised a novel procedure for building the Kitagawa–Takahashi di­ work [27,28] only concentrates on the statistic size effect but ignores
agram of flawed AM material and notched components, holding the other types of size effect and the fatigue performance scatter in AM
advantage of being independent of the size of the critical defect in parts. To summarize, aforementioned analyses focused on studies about
contrast to global approaches but with slightly accuracy decreasing. defect characterization, strength assessment and single size effect for
Defect-based fatigue models are vital for linking fatigue performance AM materials, whereas few studies devoted their efforts to various size
between AM materials and parts. Hu et al. [16] defined an effective effects and fatigue performance scatter in AM materials, which is
initial crack length on basis of the defect radius and the cyclic plastic essential for integrity assessment of AM parts and promote application of
zone, and predicted the fatigue life using the NASA/FLACGRO fatigue AM in engineering practice.
crack growth model. Numerical simulation methods of defects are also In this regard, this work studies the size effect and fatigue perfor­
implemented for fatigue life assessment of AM parts. In order to char­ mance scatter in AM materials via extreme value statistics. Furthermore,
acterize the effects of scale, position and geometry of gas pores on the a defect-tolerant assessment method for AM parts is developed. Section 2
fatigue strength of AM Ti-6Al-4V, a mechanistic approach [17] was revisits the theory of extreme value statistics and provides recent ad­
developed and the Smith-Watson-Topper model was implemented to vances in AM fatigue modeling; in Section 3, a strength altering factor
predict the fatigue life using local stress and strain method. Seeing that (SAF) is proposed to characterize the correlation between fatigue
superiority of numerous analyses, Xie et al. [18] investigated the strengths of AM materials under size effect; later, the physical implica­
competition relationship between lack-of-fusion defects and gas porosity tion of material parameters and geometric features of Basquin, Cof­
through computed tomography scans technology and FE simulations. In fin–Manson–Basquin and Ramberg–Osgood curves are elucidated in
addition, based on the statistical theory, probabilistic models were Section 4, and the extrapolation framework is illustrated; in Section 5,
raised to estimate the fatigue life of AM materials [19,20], in which the three series of specimens with different processing parameters and
possibility of initiating a micro crack was computed through the gauge volumes are used for experimental validation, and the proposed

2
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 1. Block maxima approach based on microscopy of polished sections.

method is then utilized for defect-tolerant assessment of AM parts; ⎧ [ ( ]

finally, conclusions are drawn in Section 6. ⎪
⎨ exp − x − u)− 1γ
− for x > u
G(x) = σ (6)
⎪
⎩
2. Extreme value statistics and AM fatigue modeling 1 for x ≤ u

Extreme value statistics has attracted many attentions in practical

2.1. Generalized extreme value theory
applications due to singular values in extreme conditions that play a
more significant role than average value sometimes. It is worth noting
As a branch of statistics, extreme value statistics span more than a
that engineering design is always based on extremes, because extreme
century of history, which was firstly developed to study hydrology such
values such as load and strength are the critical parameters causing
as flood frequency analysis [29,30], and now quite practical in term of
failure of engineering parts. For example, during the structural design
its implementations to various applications. The main aim of the
process of pressure vessels, ASME Committee focus on the average load
extreme value theory is to predict the statistical probabilities of events
pressure vessel, is also interested in the maximum load and the mini­
that rarely happen by available observed data, which concerns the
mum strength of the material.
investigation of the maximum/minimum value within random
Design values of practical applications are usually obtained through
distributions.
a balance between safety and cost to guarantee a normal operation in
Considering the sample X1 , X2 , ⋯Xn independent and identically
service period and reasonable costs. However, it is difficult to estimate
distributed with common cumulative probability function F, defined as:
extreme capacities and operating conditions of structure due to the ex­
P(Xi ≤ x) = F(x) (1) istence of multi-source uncertainties [34–36] and the lack of available
data. According to this, the usage of safety factors has been a classical
The distribution of the maxima Mn = max{X1 , X 2 , ⋯Xn } is defined as:
solution to the issue, which often lead to conservative design and heavy
P(Mn ≤ x) = P(X1 ≤ x, X2 ≤ x, ⋯Xn ≤ x) = [F(x)]n (2) structures with extra costs. Consequently, the safety factor approach is
not an optimal solution for engineering precise design purpose. Nowa­
where F is usually unknow and n is large in practice, the distribution of days, extreme value statistics have been a well suitable technique
Mn is not decided by the parent distribution but asymptotic property of capable of test cost reduction and information acquisition of extreme
the maximum or the minimum [31,32], which can be expressed via the events.
Generalized Extreme Value (GEV) distribution with a cumulative
probability function as:
{ [ (x − u) ]− 1γ } 2.2. Defects in AM fatigue modeling
G(x) = exp − 1 + γ (3)
σ Extreme value theory develops rapidly and has become a significant
part of probability theory, which has been applied to various practical
where [u, σ , γ] represent the location parameter, scale parameter and
situations [37,38] that is primarily concerned with extremes, like the
shape parameter, respectively. According to different practical problems
prediction of characteristic parameters and maximum defects in fatigue
and cumulative distribution function F [33,27], the shape parameter γ is
modeling under size effect.
determined and GEV is divided into three types of distributions:
About application in qualification analysis of defects, statistical
Gumbel (Type I γ = 0, σ>0):
characterization of material fatigue was presented through a novel
{ [ (x − u) ]}
G(x) = exp − exp − (4) extreme value approach [28], which includes consistent mathematical
σ modeling of initiating defects, fatigue strength and life. To solve the
Fréchet (Type II γ > 0, σ >0): issue that univariate set-up may cause the estimation error of the true
⎧ distribution, a multivariate extreme value analysis [39] based on in­
⎪
⎨ 0 for x ≤ u formation of the largest observations in each control volume is devel­
[ ( ]
G(x) = x − u)− 1γ (5) oped. Considering that a larger defect is most possibly to lead to failure
⎪
⎩ exp − for x > u
σ compared with the smaller one [40], the fatigue strength of materials
will not be influenced by all the defects or an average, but the upper tail
Weibull (Type III γ < 0, σ>0):
of defect size distribution. Consequently, extreme value statistics,
appropriately enough, estimate the occurrence possibility of the
maximum defect influence the fatigue strength rely on the capability
that estimating accurately the tail of the distribution according to a set of

3
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 2. Peaks over threshold approach based on CT scans.

independent data. Then two methods were developed for practical ap­ distribution. In short, POT is competent to derive the distribution of the
plications based on different branches of extreme value statistics the­ maximum defect.
ories, namely block maxima (BM) and Peak Over Threshold (POT), In POT, the exceedances in x larger than a threshold value u are
which focus on measurement of the maximum defects in randomly extracted, which present asymptotic properties [31]. The distribution of
chosen areas (volumes) and the size of defects larger than a threshold the exceedances x can be estimated by a Generalized Pareto (GP) dis­
value, respectively. tribution Wγ,u,σ . The cumulative distribution function of the GP is shown
Murakami [41] employed extreme value statistics to estimate the in Eq. (12).
maximum inclusion in randomly chosen polished section areas using BM ⎧
− 1
sampling. In BM, the whole specimen is divided into several equivalent ⎨ Wγ,u,σ (x) = 1 − 1 + γ⋅x − u γ for γ ∕
⎪
=0
sub-volumes, afterwards the maximum defect size in each sub-volume is σ (12)
⎪
⎩
recorded.
− x−σ u
W0,u,σ (x) = 1 − e for γ→0
In particular, a conventional control volume VPS is defined:
where the shape γ and scale parameters σ can be estimated using MLE,
VPS = S0 ⋅hmean (7) the threshold u can be calculated by the mean excess plot [46]. In case
that the defect size obeys the exponential distribution, the exceedances x
where hmean is the average size of the defects on polished section via 2D
can be described according to Eq. (12) for γ → 0. The MLE of σ can be
analysis,S0 is the area of polished section, as shown in Fig. 1.
expressed as Eq. (13).
Murakami adopted BM and fitted the data with Gumbel distribution,
∑N
wherein return period T of the maximum defect in the prospective ma­ (x − u)
σ = i=1 i
̂ for xi > u (13)
terial volume V is expressed as: N
V where N is the number of exceedances (the defects xi > u); xi is the size
T= (8)
VPS of defects exceeding the threshold u.
The definition of return period T can be found in both ASTM [42] and The return period T of the maximum defect in volume V can be
ESIS [43] publications. This concept originates from geophysical sci­ computed by Eq. (14).
ences where the annual extremes of natural phenomena are concerned, N
which represents average interval between two extreme events T= ⋅V = ρ⋅V (14)
V0
exceeding certain level.
The distribution of the maximum defect in volume V can be where ρ represents the average density of exceedances x in the CT
approximated by: scanned volume V0 .
The modal value of the maximum defect in a volume V can be
GV (x) = [G(x)]T (9)
expressed as:
The modal value (the most frequent value) is (1 − 1
T
)-th percentile for ( )
1
GV (x):
−
xmax,V = Wγ,u,σ 1 − (15)
T
( )
1
xmax,V = GV 1 − ≈ u + σ ⋅In(T) (10) xmax,V = u + σ⋅ln(T) for γ = 0 (16)
T

where the parameters u and σ can be estimated using the maximum where u and σ mainly depend on the processing technology, T calculated
likelihood estimation (MLE). The distribution of the maximum defect in by ρ and V in Eq. (14) are dependent to the processing technology and
volume V can be expressed as: the specimen size, respectively. Three parameters in Eq. (16) reflect the
{ [ (x − x )]} statistic and production size effects, which can be utilized to assess fa­
GV (x) = exp − exp −
max,V
(11) tigue strength of defective materials under size effect.
σ
The p-th percentile of the maximum defect distribution in a pro­
In addition, another method of extreme value statistic by employing spective material volume V is computed through correlating Eqs. (2) and
generalized Pareto distribution, namely POT approach, involves all the (12):
defects larger than the threshold value u, schematically described in [ ]− ( )
Fig. 2. POT, conveniently utilized to the defect evaluation, can filter all (17)
T 1
xp,V = W0,u, σ (x) = u − σ⋅ln 1 − pT
the measurements close to the resolution [27] and no critical informa­
tion is ignored. In the studies [27,44], the POT shows a good perfor­ The above-mentioned methods have been used to predict the
mance for maximum defect prediction. Note from [45] that CT scans is maximum defect distribution and fatigue strength of AM materials with
appropriate for the POT focusing on the upper tail of the defect size different volumes. However, the production size effect [27] and fatigue

4
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 3. HSV, Vsurf and HSSV from different samples.

life scatter also caused by defects are ignored. In this study, POT is propagation usually start from surface defects, even if a larger one ap­
adopted to predict the distribution of maximum defects of AM specimens pears inside the material. As most failures originate from the surface of
with different size and process according to Eqs. (16)–(17) and CT data, specimen, the size of the critical defect causing the failure is regarded as
whereas POT is intrinsically about extrapolation [47] and has a better the maximum defect size of surface volume vsurf in Fig. 3, which is
performance on defect analysis comparing with BM. Then the distribu­ defined as an external circular. For smooth specimens, vsurf can be
tion of maximum defects is used to extrapolate fatigue strength in Sec­ determined according to Eq. (18).
tion 3 and probabilistic fatigue curves of AM materials under size effect ( ( )2 )
in Section 4. vsurf = Vg − Vint = π⋅lg ⋅ rg2 − rg − h (18)

where lg and rg are the specimen gauge length and radius, respectively; h
2.3. Prospective material volume computing in POT is the thickness of surface volume [48,49] and is defined as:

In this study, POT is used to determine the defect size characteristics h=

a
(19)
of material volume where crack initiation/propagation is most likely to 0.8
occur. Accurate prediction of defect size requires the reasonable selec­
where a is the measured average radius of the critical defect, the radius a
tion for prospective material volume V. In particular, occurrence of
of the equivalent circle derived from Eq. (20), as shown in Fig. 3.
crack initiation/propagation requires certain local stress that corre­
√̅̅̅̅̅̅̅̅̅
sponding to the micro and macro structures under the same external area
loads. In view of this, the prospective material volume V in Eq. (14) can a= (20)
π
be determined via equivalent surface volume [27] integrating with
highly stressed volume (HSV) [26], named as highly stressed surface In addition, the crack initiation/propagation is more likely to occur
volume (HSSV), see Fig. 3. within the highly stressed regions. A large defect, at the surface region,
It is generally thought that surface defects are much more harmful will not lead to failure if the stress does not exceed corresponding to
with respect to internal ones. In this context, crack initiation and threshold. Consequently, crack initiation/propagation occurs at the
overlapped region of surface and highly stressed volume. The HSV is
defined as the location subjected to more than n-th percentile of
maximum stress in the critical region [50].
σ c,n% = n%⋅σ c,max (21)

where σ c,n% is the minimum stress level that HSV subjected to, σ c,max is
the maximum stress in the target region and n% is an empirical
parameter. For the smooth specimens, notched specimen and part, the
Vsurf , HSV and HSSV are depicted in Fig. 3, it is worth pointing out that
the volume of the materials within the gauge length can be regarded as
the HSV for smooth specimens, consequently Vsurf is equal to VHSSV and
the prospective material volume are expressed uniformly as Vsurf in the
following analysis. For the notched ones, HSSV can be computed via
finite element analysis [26] according to the thickness of surface volume
h defined in Eq. (19).

Fig. 4. Fatigue data of AM specimens with different scales.

5
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

smooth specimens containing defects produced by L-PBF are employed,

thus statistical and production size effects are taken into consideration
for fatigue curves extrapolation through inherent defects distribution
statistics and analysis.

3.2. Strength altering factor for AM fatigue strength assessment

Fatigue strength assessment of material containing defects has been

investigated [11,12], wherein Murakami [13] shows that small non-
propagating cracks are usually presented at the tip of defects under
stress level near the fatigue strength, which is the threshold stress at
which the small cracks do not grow. Kitagawa–Takahashi (K–T) diagram
is usually implemented to depict the relation between defect size and
fatigue strength, as shown in Fig. 5. One widely used method, namely
Murakami & Endo model, was developed to describe this relationship:
/( √̅̅̅̅̅̅̅̅̅)
(22)
1/6
Δσw = C1 (HV + 120) area

where C1 is a parameter corresponding to location of the critical defects,

HV is the Vickers hardness. In particular, if material does not contain a
definite defect, the fatigue strength can be computed via Vickers hard­
ness [58,64,65].
In order to characterize a surface defect with different location and
geometries, an modified El-Haddad model is adopted [14,41], where the
√̅̅̅̅̅̅̅̅̅
crack length can be substituted by area and stress intensity factor (SIF)
Fig. 5. Schematic of Murakami & Endo and modified El-Haddad models: (a) K- threshold is expressed as:
√̅̅̅̅̅̅̅̅
T diagram; (b) SIF threshold versus area. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√̅̅̅̅̅̅̅̅̅
ΔK = Y⋅Δσ π area (23)
3. Fatigue strength assessment of AM materials under size effect
where Y = 0.5 for internal defects and Y = 0.65 for surface defects,
3.1. Size effect in AM fatigue which depends on the defect location.
According to the modified El-Haddad model, the relationship be­
In conventional metal fatigue analysis, fatigue tests of small speci­ tween SIF and defect size can be expressed as:
mens are often carried out to obtain basic fatigue properties and √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√̅̅̅̅̅̅̅̅̅
component fatigue assessment. Due to size effect, however, the fatigue area
ΔK = ΔKth,lc ⋅ √̅̅̅̅̅̅̅̅̅ √̅̅̅̅̅̅̅̅̅ (24)
performance of components is different with that of tested specimens area + area0
even subjecting to similar load. For example, fatigue strength of material
decreases with increasing material volume. Note from [51] that an where ΔKth,lc is the long crack propagation threshold, which can be
√̅̅̅̅̅̅̅̅̅
obvious size effect in AM materials can be also found, which is one of the calculated based on the experimental data according to Eq. (23); area0
main reasons of fatigue tests differ from AM parts. Thus, it is essential to is the El-Haddad parameter defined as:
investigate the size effect in AM fatigue for further engineering ( )2
√̅̅̅̅̅̅̅̅̅ 1 ΔKth,lc
applications. area0 = (25)
π Y⋅Δσw0
Generally, size effects can be classified into four types [26,52] ac­
cording to their sources, namely the statistical, geometrical, production where Δσ w0 is the fatigue strength of material free of defects, which can
and surface size effects. The larger the volume of materials, the higher be estimated as the stress value corresponding to a plastic deformation
probability of larger defects inside it, which normally causes the final of 0.05% in the stable cyclic curve [66].
failure, namely statistical size effect [53–55]. As depicted in Fig. 4, fa­ Substitute Eqs. (23) and (25) into Eq. (24), the fatigue strength can
tigue data of specimens with two scales of AM Ti-6Al-4V alloy is pre­ be expressed as:
sented [28], namely standard specimen and small specimen. It can be √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
found that two S–N curves present an obvious statistical size effect and √̅̅̅̅̅̅̅̅̅
area
the small specimens have higher fatigue strength than the large ones. Δσw = Δσw0 ⋅ √̅̅̅̅̅̅̅̅̅ √0̅̅̅̅̅̅̅̅̅ (26)
area + area0
Geometrical size effect originates from geometric discontinuity,
which depends on the stress gradient [56], and usually work together As depicted in Fig. 5, the shaded area below the threshold line rep­
with statistical size effect. The microstructure, residual stress and defects resents the safety region where small cracks do not grow. In terms of the
are influenced by the AM processing parameters, post processing and fatigue strength in Fig. 5 (a), linear elastic fracture mechanics (LEFM)
surface finish, which causes the difference of fatigue strength, namely can be utilized to estimate fatigue strength in the presence of large de­
production and surface size effects [57]. Large amounts of works on the fects, and the fatigue strength Δσ3w0 is considered as the boundary of
effect of material volume [27,28], AM process parameters [58–60] and applicability of LEFM and elastic plastic fracture mechanics [67]. From
surface finish [61–63] on fatigue performance of AM materials have Fig. 5 (b), the growth rate of ΔKth decreases as the defect size increases,
been carried out. However, few of them evaluated defect-tolerant for ΔKth = ΔKth,lc for a long crack, in which LEFM is applicable.
AM parts by considering the coupling effect of these factors. Based on the above-mentioned approaches, the relation between
Generally, fatigue initiates from geometric discontinuities such as fatigue strength and defect size of material can be explicit. The predic­
micro-cracks, inclusion and lack of fusion, and the failure probability of tion of fatigue strength under size effect can be conducted based on POT
materials and parts increases due to the higher possibility that a larger and the information of defects detected by CT. Furthermore, the relation
defect appears comparing with small specimens. In this study, the of fatigue strengths can be characterized, then to establish the link

6
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 6. Fatigue life of AM materials: (a, c, e) fatigue data; (b, d, f) normalized fatigue data.

between fatigue performance of AM materials under size effect.

σw ’
As show in Fig. 6 (a), the original fatigue data of AM Ti-6Al-4V with ϑ= (27)
σw
electron beam melting (EBM) and direct metal laser sintering (DMLS),
respectively. and Fig. 6 (b) represents the normalized fatigue data where σ w ’ and σ w are the fatigue strength corresponding to prospective
calculated via Eq. (23) [68]. Fig. 6 (c–f) show the data for AM Ni-based volume V and CT scanned volume V0 in Eq. (14), which can be computed
superalloy 718 and AM AlSi10Mg with different processes [69,70]. It by integrating the Murakami & Endo or El-Haddad model with Eqs. (16)
can be found that two groups of original fatigue data have an obvious and (17). SAF is considered as a parameter that characterizes the
difference in Fig. 6 (a, c, e), while the normalized fatigue data are almost disparity of fatigue performance caused by defect size at fracture origin,
overlapped in Fig. 6 (b, d, f). This can be interpreted as that the defect thus not only size effect but also the fatigue scatter caused by defect can
size at fracture origin is equivalent to the same value. be characterized, which is introduced to the commonly used fitting
Based on the comparison between fatigue data and normalized fa­ equations for probabilistic fatigue curves extrapolation under size effect
tigue data in Fig. 6, the strength σ w can be considered as the critical in next section.
parameter to characterize the correlation between fatigue strength of
AM materials with different defect distributions, then a SAF is proposed
for size effect modeling in AM fatigue and defined as:

7
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 8. Curves fitting of Coffin–Manson–Basquin equation.

Fig. 7. Curves fitting of Basquin equation.

4. Probabilistic fatigue modelling for AM materials under size

effect

The commonly used equations depicting S–N curves are usually

employed to fit the curves of fatigue data, which characterize the fatigue
curve trend and shape rather than predicting fatigue lives. The material
parameters obtained by curve fitting procedure depends on specimen
shape, scale and material, but the procedure lacks of discussion on
mechanics of cracks influence the S–N curve trend. Therefore, a material
parameters extrapolation method based on fracture mechanics theory
and defect data is developed in this study.
Actually, S–N curves of specimens containing defects with different
sizes are closely related [64]. If the S–N curve fitted by one group of
fatigue data is determined, P–S–N curves of other groups can be drawn
through the analysis in combination with the statistics of extremal de­
Fig. 9. Curves fitting of Ramberg–Osgood equation.
fects, material parameters and geometric features of fatigue curves, such
as slope, knee point and transition point, wherein fatigue strength Δσw is
the critical parameter to combine the S–N curves of different groups. The ( )B
S = ϑ⋅A Nf (29)
proposed SAF is utilized to derive the P–S–N curves via the Basquin
equation, subsequently it is generalized to the extrapolation of P–ε–N
4.2. Size effect in strain-life curves
curves and cyclic stress-strain curves.

Coffin and Manson [75,76] independently raised an empirical for­

4.1. Size effect in stress-life curves mula, which describes the relation between the cycles of crack initiation
and the amplitude of plastic strain. Together with the Basquin equation
The fatigue behavior is systematically investigated by Wöhler [71], [72], the well-known Coffin–Manson–Basquin equation is raised by
then based on the characteristics of Wöhler equation, Basquin [72] Morrow [77], as shown in Eq. (30).
raised the following exponential equation:
σ’f ( )b ( )c
( )B
S = A Nf (28) ε= 2Nf + ε’f 2Nf (30)
E

where S represents the stress amplitude, Nf represents the number of where the total strain amplitude ε is divided into elastic and plastic
cycles to failure, A and B are material constants. components: σ’f and b denote the fatigue strength coefficient and fatigue
The concept of fatigue strength is unknown until Stromeyer [73] strength exponent, respectively; ε’f and c respectively denote the fatigue
carried out a series of fatigue tests on different materials to prove the
ductility coefficient and fatigue ductility exponent, respectively; E de­
existence of fatigue strength. As depicted in Fig. 7, Basquin curve, fa­
notes the Young’s modulus.
tigue strength and knee point [74] of specimens are closely relevant, any
As depicted in Fig. 8, the transition point is defined as the intersec­
two of them are known and the S–N curve can be drawn. The S–N curves
tion of elastic and plastic lines, which can be determined based on
for specimens with different defect sizes can be associated through
material parameters in Eq. (30). In the left region of transition point, the
combining material constant in Eq. (28) with geometric features of fa­
plastic component is considered to be dominant; conversely, the elastic
tigue curves.
component is dominant. Actually, ε-N and S–N curves corresponding to
The slope of the Basquin curve and cycles of knee points have a slight
low and high cycles fatigue can be linked, and the fatigue yield strength
change for different defect sizes [64]. The quantitative analysis of Bas­
represents the stress level of intersection. Since σ ’f corresponding to fa­
quin curves correlation of different specimens have not been systemat­
ically studied; therefore, the slope of the Basquin curve is regarded as a tigue strength is integrated with SAF to extrapolate this coefficient of
constant value in this study. The knee points are determined as the other groups with different initial defect sizes. The SAF is utilized to
intersection of Basquin curve and the fatigue strength. If the “first” fatigue ductility coefficient to avoid abnormal intersection of upper and
Basquin curve and the connection between the fatigue strength of lower curves for plastic response. Extrapolation equation can be written
specimens with different defect distributions are determined, then the as:
P–S–N curves of other groups can be determined by Eq. (29).

8
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 10. The flowchart of probabilistic fatigue curves prediction.

[σ ’ ( )b ( )c ] where the total strain amplitude ε is also divided into elastic and plastic
(31)
f
ε = ϑ⋅ 2Nf + ε’f 2Nf components: S and ε respectively denote the stress and strain amplitude;
E
K’ and n’ respectively denote the cyclic strength coefficient and cyclic
strain hardening exponent.
4.3. Size effect in cyclic stress-strain curves
Note from Dowling’s research [79], through correlating the Cof­
Ramberg and Osgood [78] raised an equation to describe the relation fin–Manson–Basquin with the Ramberg–Osgood equations, K’ and n’ can
between stress and strain in materials subjected to cyclic fatigue loading. be calculated by:
Generally, the stable hysteresis loops are collected to draw the stress- σ ’f
strain curve via Ramberg–Osgood equation, as shown in Fig. 9 and Eq. K ’ = ( )n’ (33)
(32). ε’f
( ) 1’
S S n
ε= + (32) n’ =
b
(34)
E K’ c
The extrapolation equation of Ramberg–Osgood can be derived ac­
cording to Eqs. (31)–(33), which is expressed as:

σ ( σ ) 1’
(35)
n
ε= + n’ )
E K ’ ⋅ϑ(1−

4.4. Computational framework for probabilistic fatigue curves

Following the POT in Section 2, SAF in Section 3, physical implica­

tion of material parameters as well as geometric features of fatigue
curves mentioned in Section 4, an extrapolation framework for proba­
bilistic fatigue curves of specimens containing different maximum defect
sizes is illustrated in Fig. 10. Three series of AlSi10Mg specimens
(denoted as P1, P2 and P3, respectively) with varying processing pa­
rameters are employed in this procedure.

(1) The CT data of P1, P2, P3 and POT are used to predict the dis­
tribution of maximum defects in Vsurf . Then the SAF 1-2 are
calculated via the Murakami & Endo or El-Haddad model, and
Fig. 11. The Kitagawa–Takahashi diagram incorporating SAF. SAF 3 is with respect to fatigue strength scatter caused by

9
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Table 2
Parameters of the defects from CT data and predicted defect sizes.
Process u ρ(defects/mm3 ) σ xmax,V x1% x99%
(mm) (mm) (mm) (mm) (mm)

P1 0.150 0.383 0.049 0.297 0.227 0.522

P2 0.100 0.315 0.021 0.165 0.135 0.261
Fig. 12. Geometry of the cylindrical specimens. P3 0.100 0.010 0.014 0.073 – –

Table 1 Table 3
Main dimensions of specimens. Fatigue strength assessment and calculation results of SAF.
Process Gauge length Gauge diameter Grip diameter Vsurf Method Process σxmax,V σx1% σx99% ϑ2− x1% ϑ2− x99%
(mm) (mm) (mm) (mm3) (MPa) (MPa) (MPa)
P1 13 5.1 11.5 52.10 Murakami & P1 203.071 212.314 184.859 1.046 0.910
P2 16 6.0 12.0 68.75 Endo P2 256.224 265.036 237.342 1.034 0.926
P3 11 5.5 10.0 14.63 model
Modified El- P1 183.517 200.435 148.711 1.092 0.810
Haddad P2 227.641 243.505 192.377 1.070 0.845
microstructure, residual stresses and human errors, etc., which is model
characterized via Kitagawa–Takahashi diagram of P3. Meanwhile
the fitting S–N curve of P3 is regarded as the Baseline 1. SAF 1-3
and Baseline 1 are both as the input for step (2). and gauge volume are employed for method verification. The geometry
(2) The modal S–N curves (maximum probability S–N curve) of P1 and dimensions of the specimens are shown in Fig. 12 and Table 2,
and P2 can be drawn based on the SAF 1 and Baseline 1 according respectively. All specimens are produced using an EOS M400 powder-
to Eq. (29), which is regarded as the Baseline 2. Finally, the bed machine, and main differences are the inert gas recirculation sys­
P–S–N curves can be determined through the similar derivation tem, the layer thickness and the platform preheating temperature. More
procedure based on the SAF 2-3 and Baseline 2. According to Eqs. details about the material and processing techniques can be found in
(31) and (35), the P–ε–N curves and probabilistic Ram­ [27,44].
berg–Osgood curves can be obtained by repeating the above In Table 1, the Vsurf of P1, P2 and P3 are listed out, which are
procedure. considered as the analysis object considering that failures are more
likely to start from the surface, the size of the critical defect in a spec­
It is worth pointing out that SAF 1 and SAF 2-3 reflect the size effect imen can be regarded as the of the maximum defect size within Vsurf . The
and fatigue strength scatter in AM materials, respectively. As shown in Vsurf can be derived according to Eqs. (18)–(20).
Fig. 11, SAF 1 represents the difference of modal values (three series of The defect information in Vg of P1, P2 and P3 using CT scans are
AlSi10Mg specimens) under size effect; SAF 2 characterizes the Kita­ given in Table 2. According to Eqs. (16) and (17) and CT scans, the
gawa–Takahashi diagram between blue dotted lines reflects the defect modal value, 1% and 99% of the maximum defect distribution in pro­
size scatter of the specimens of interest; the difference between experi­ spective material volume Vsurf are computed and presented in Table 2.
mental data and modified El-Haddad model reflects the fatigue strength Moreover, fatigue strength and the SAF corresponding to modal value,
scatter caused by microstructure, residual stresses and human errors, 1% and 99% cumulative distribution of defect sizes are listed in Table 3.
etc., which are difficult to quantify separately and thus coupled into one And ϑ3 represents fatigue strength variability of approximately 11%.
term SAF 3, expressed as Eq. (36). The Basquin parameters of P3 is fitted through least square method, and
⎧ the average fatigue strength is 270 MPa [70]. All the fatigue experiments
[Δσi − Δσ ] [
Δσ w − Δσjw
]
⎪
⎪
⎪ max w w
+ max were carried at R = − 1 [82]. Predicted P–S–N, P–ε–N curves and
⎪
⎪
⎪
⎪ Δσ w Δσ w
for σx ’n% < σx ’max,VPi (i=1,2) probabilistic Ramberg-Osgood curves are shown in Figs. 13–16, which
⎨1 −
ϑ3 =
2 are derived based on the material constants fitted by fatigue data, listed
⎪
⎪ [Δσi − Δσ ] [
Δσ w − Δσjw
]
in Table 4.
⎪
⎪ max w w
+ max
⎪
⎪ Δσ w Δσ w Stress-life plots depicted in the Fig. 13 have shown a good agreement
⎪
⎩1 + for σx ’n% > σx ’max,VPi (i=1,2)
2 of the both two methods, where most experiment points fall within the
(36) range of S–N curves with survival rate of 1% and 99%. Comparing with
Murakami & Endo model, El-Haddad model provides more conservative
j
where Δσiw and Δσ w represent the fatigue strength of i, j-st experimental predictions. It is quite different with the conclusion in [80], because the
point of P3, which are above and below the corresponding to Δσ w on relative value SAF is employed based on extrapolation method in this
Kitagawa diagram, respectively. study. The use of SAF calculated by Murakami & Endo model over­
Note from [80] that the application of the Murakami & Endo model is estimates the fatigue strength of P1, and a larger predicted life scatter
conservative, and Beretta et al. [81] found that modified El-Haddad than experimental data is presented, which is caused by the lack of data
model can provide more accurate predictions by investigating the fa­ points. The predicted P–S–N curves through probabilistic control vol­
tigue limit of high strength steel. In this study, SAFs are computed by ume method is shown in Fig. 14, which integrates the highly stressed
Murakami & Endo model and El-Haddad model, separately, and the volume (surface) with Weibull weakest-link method [83,84]. The
extrapolated results are demonstrated and compared. probabilistic control volume method focuses on the statistical size effect,
thus it is invalid, as shown in Fig. 14 (a), for P1 and P3 produced via
5. Model validation and comparison different processing parameters. For P2 and P3 with similar processing
parameters in Fig. 14 (b), wherein statistical size effect plays a main role
5.1. Experiments and discussions and all experiment points fall in the predicted ranges. However, an
excessive life dispersion and abnormal 50% survival fatigue curve are
To demonstrate the effectiveness of the extrapolation approach, presented due to a pure statistical method. By comparison, the compu­
three series of AlSi10Mg specimens with varying processing parameters tational framework in this study have more accurate prediction of life

10
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 13. P–S–N curves of AlSi10Mg alloy: (a–b) Murakami & Endo model, (c–d) modified El-Haddad model.

Fig. 14. P–S–N curves of AlSi10Mg alloy predicted by probabilistic control volume method [80].

dispersion and extensive application by taking advantage of defect in­ points and the slope of S–N curve is regarded as a constant value. In
formation and employing both theories of statistics and fracture addition, there is a complex non-linear relationship between defect size
mechanics. and logarithmic fatigue life, which depends on both maximum defect
In Fig. 15, two methods slightly underestimate the strain-controlled size distribution and crack propagation model.
fatigue life scatter of P1, but perfectly predict P–ε–N curves of P2. In this study, the use of POT extrapolates the maximum defect size
Looking at Fig. 16, The probabilistic Ramberg–Osgood curves cover the from a large volume (CT scanned volume) to a small one (Vsurf ). It can
most of experiment points. Though not all the probabilistic fatigue also be applied to extrapolation from a small volume to a large one,
curves are perfectly described, the predicted results and the extrapola­ which can significantly reduce the test cost and provide an actual fatigue
tion methods are still essential for fatigue performance assessment of AM performance for structure reliability and safety design. In the following
materials under size effect. In addition, it can also be found that the section, the extrapolated method, from a small volume to a large one,
modal curves are distinguished from the mean fatigue curves, because are utilized to the defect-tolerant assessment for AM parts.
the maximum defect distribution is approximated by a GP distribution
that presents skewness and some details are ignored. For instance, knee

11
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 15. P–ε–N curves of AlSi10Mg alloy: (a–b) Murakami & Endo model, (c–d) modified El-Haddad model.

5.2. Defect-tolerant assessment of AM parts

∏
3 [ ( )b ( )c ]
S= ϑk ⋅ σ ’f 2Nf + Eε’f 2Nf (38)
In conventional fatigue design rules, due to the lack of sufficient data k=1
for size effect and scatter analysis, safety factors are commonly used to
Supposing that fatigue life/strength follows a specific distribution, ϑ1
consider these factors. However, it can’t provide a quantitative char­
and ϑ2 ⋅ϑ3 (ϑ2 represents ϑ2− xn% in this section) describe the “location”
acterization of safety and reliability, and often causes conservative
and “scale” characteristics, respectively, which jointly affect the mate­
design. Nowadays, higher reliabilities are required for complex struc­
rial’s fatigue strength.
tures, methods for defect-tolerant assessment are expected, especially
In order to characterize the P–S–N curves varying with the pro­
for AM parts that containing inherent defects.
spective volume, take fatigue data of P2 and modified El-Haddad model
A progress from small specimen to parts requires a comprehensive
as an example, the three-dimension S–N response surfaces are shown in
consideration on size effect, notch effect [52] and service condition
Fig. 18. The size effect is not obvious compared to the fatigue life varies
(load, temperature, etc.), where several uncertainties [36,85,86,90]
with stress level, wherein the stress difference is 80 MPa when Nf = 103 .
should be further taken into consideration and modelled as variables in
Random load is assumed to follow normal distribution. Considering
fatigue design. As depicted in Fig. 17, defects size and loads are taken as
that the stress range of μ ± 3σ covers 99.73% of the normal distribution
random variables in the procedure of safe-life curve deriving. In strain-
and the lower fatigue curve is that we care about, thus the curve under
controlled fatigue tests, before plastic deformation, S is equal to the
stress level μ + 3σ is consider as the safe-life curve, as depicted from
actual stress amplitude; after that, S becomes a fictitious value, greater
point a to point c in Fig. 17, the stress at point c can be computed by
than the true stress [87]. The modal S–N curve of AM parts was derived
using stress at point a and coefficient of variation (CV):
by the extreme value statistics theory and SAF. Then, the lower
boundary curve considering the existence of defect size scatter can be Sa
Sc = (39)
determined via load-life interference model [88], the limit state function (1 + 3CV)
G is defined as:
Notch features, including holes, grooves, fillets and shoulders, are a
G = Nf − Nd (37) vital geometry with widespread applications in engineering practice.
These geometrical discontinuities are source of stress concentrations and
where Nf is the number of cycles to failure, which is obtained by fatigue play important roles during safety design. Considering the notch geo­
tests; Nd denotes the designed life. When the Nf is higher than designed metric features, fatigue strength range of notched specimen can be
life Nd , the structure is deemed safe. Therefore, the n% survival S–N expressed as:
curve can be as the safe curve under reliability n% under constant
Δσ w
amplitude load, which is expressed as: Δσ = (40)
Kf

12
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

Fig. 16. Probabilistic cyclic stress–strain curves of AlSi10Mg alloy: (a–b) Murakami & Endo model, (c–d) modified El-Haddad model.

Table 4
Summary of material constants for P3.
Process Basquin Coffin–Manson–Basquin Ramberg–Osgood

A (MPa) B σ’f (MPa) ε’f (MPa) b c K’ (MPa) n’

P3 2113.49 − 0.158 401.8 101.7 − 0.073 − 1.287 321 0.056

Fig. 18. P–S–N curves versus prospective volume.

Fig. 17. The safe-life curve deriving for AM parts under size effect.

13
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

and service conditions, Eqs. (41) and (42) provide a reference for the
safe-life and defect-tolerant design, wherein the accuracy relies on
determination of Kf , appropriate characterization of the critical defect
and quantitative analysis of multi-source uncertainties.

6. Conclusions

This paper studies and quantifies the relationship among defects,

fatigue strength and fatigue curves by means of extreme value statistics
theory and Kitagawa–Takahashi diagram. Furthermore, a framework for
probabilistic fatigue curves prediction is established. Three series of
specimens with different processing parameters and gauge volumes are
employed for method validation and comparison. Finally, the predicted
method is further utilized to defect-tolerant assessment of AM parts. Key
conclusions of this work are summarized as below:
Fig. 19. Defect-tolerant assessment strategy of AM parts.
(1) A strength altering factor is proposed to extrapolate fatigue
curves for specimens containing different defects (originating
from process and volume difference) by integrating with Basquin,
Coffin–Manson–Basquin and Ramberg–Osgood equations. The
results indicate that fatigue curves of specimens with different
initial defect sizes are closely related. Based on CT scans and
certain quantity of fatigue data, fatigue curves of AM materials
under size effect can be predicted.
(2) The fatigue scatter of AM materials can be characterized using
strength altering factor combining with the extreme value sta­
tistics. Results show that the predicted probabilistic fatigue
curves cover most of fatigue test points, which provide a refer­
ence to structure reliability and safety design of AM parts under
fatigue scatter.
(3) Based on the extrapolation method, multiple curves of specimens
with different process (production size effect) and volume (sta­
tistic size effect) can be predicted. Only CT scans and a few
experimental data are required, wherein both the costs and
period of experimental tests are decreased.
Fig. 20. Kitagawa–Takahashi diagram describing the lower boundary on the (4) A general framework considering the influences of size effect,
fatigue strength. notch effect and fatigue scatter, etc. on reliability and defect-
tolerant of AM parts is developed; the probability of AM parts
where Δσ w is the fatigue strength range of smooth specimens, Kf is the remains in operation without failure is integrated into safe-life
notch fatigue factor [89]. Based on Eqs. (38)–(40), the safe-life curve of assessment; the allowable defect size of AM parts can be deter­
AM parts can be drawn by: mined. It is extrapolation-based analysis and better takes
∏3 [ ( )b ( )c ] advantage of available data compared to the conventional
’
k=1 ϑk ⋅ σ f 2Nf + Eε’f 2Nf empirical method, also provides a new way for safe-life and
S= (41) defect-tolerant assessment of AM parts.
Kf (1 + 3CV)

Different from the use of empirical safety factor, several factors such
as the size effect, defect size scatter, load fluctuation and notch effect, Declaration of Competing Interest
etc. are integrated into safe-life curve, which satisfies the requirements
of design while avoiding unnecessary safety margin. Furthermore, more The authors declare that they have no known competing financial
design variables and parameters can be considered to derive a reason­ interests or personal relationships that could have appeared to influence
able safe-life curve for parts via this strategy. the work reported in this paper.
In case fatigue failure of AM parts, the safe-life curve is employed to
evaluate defect-tolerant of AM parts. As depicted in Fig. 19, size effect Acknowledgement
and defect size scatter influence the allowable defect size and fatigue
strength of AM parts. Nevertheless, the safety margin with respect to Financial support of the National Natural Science Foundation of
geometric features, load fluctuation and microstructure, etc. calls for a China (No. 11972110), Sichuan Science and Technology Program (No.
Kitagawa–Takahashi diagram of AM part (lower boundary), which can 2022JDJQ0024), and Guangdong Basic and Applied Basic Research
be characterized by parameters in Eq. (41) and allowable defect size can Foundation (No. 2021B1515140030) are acknowledged.
be computed by:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ References
√̅̅̅̅̅̅̅̅̅
ϑ3 ⋅Δσ w0 area
Δσ w = ⋅ √̅̅̅̅̅̅̅̅̅ √0̅̅̅̅̅̅̅̅̅ (42) [1] Liu QC, Elambasseril J, Sun SJ, Leary M, Brandt M, Sharp PK. The effect of
Kf (1 + 3CV) area + area0
manufacturing defects on the fatigue behaviour of Ti-6Al-4V specimens fabricated
using selective laser melting. Adv Mater Res 2014;891–892:1519–24.
For P1 and P2 (Kf = 1, CV = 0), the estimated Kitagawa–Takahashi [2] Kim FH, Moylan SP. Literature Review of Metal Additive Manufacturing Defects.
diagram obtained by Eq. (42) is depicted in Fig. 20, which well describes NIST Adv Manuf Ser 2018;100–16:1–17.
the lower boundary on the fatigue strength. For more complex structures [3] Sanaei N, Fatemi A. Defects in additive manufactured metals and their effect on
fatigue performance: A state-of-the-art review. Prog Mater Sci 2021;117:100724.

14
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

[4] Benedetti M, Dallago M, Santus C. Statistical significance of notch fatigue [35] Zhu SP, Liu Q, Peng W, Zhang XC. Computational-experimental approaches for
prognoses based on the strain-energy–density method: Application to fatigue reliability assessment of turbine bladed disks. Int J Mech Sci 2018;142-143:
conventionally and additively manufactured materials. Theor Appl Fract Mech 502–17.
2020;109:102720. [36] Niu XP, Wang RZ, Liao D, Zhu SP, Zhang XC, Keshtegar B. Probabilistic modeling of
[5] Liu S, Shin YC. Additive manufacturing of Ti6Al4V alloy: A review. Mater Des uncertainties in fatigue reliability analysis of turbine bladed disks. Int J Fatigue
2019;164:107552. 2021;142:105912.
[6] Pellizzari M, AlMangour B, Benedetti M, Furlani S, Grzesiak D, Deirmina F. Effects [37] Donegan SP, Tucker JC, Rollett AD, Barmak K, Groeber M. Extreme value analysis
of building direction and defect sensitivity on the fatigue behavior of additively of tail departure from log-normality in experimental and simulated grain size
manufactured H13 tool steel. Theor Appl Fract Mech 2020;108:102634. distributions. Acta Mater 2013;61(15):5595–604.
[7] Benedetti M, Fontanari V, Bandini M, Zanini F, Carmignato S. Low- and high-cycle [38] Gu T, Stopka KS, Xu C, McDowell DL. Prediction of maximum fatigue indicator
fatigue resistance of Ti-6Al-4V ELI additively manufactured via selective laser parameters for duplex Ti–6Al–4V using extreme value theory. Acta Mater 2020;
melting: Mean stress and defect sensitivity. Int J Fatigue 2018;107:96–109. 188:504–16.
[8] Beretta S, Anderson C, Murakami Y. Extreme value models for the assessment of [39] Schmiedt AB, Dickert HH, Bleck W, Kamps U. Evaluation of maximum non-metallic
steels containing multiple types of inclusion. Acta Mater 2006;54(8):2277–89. inclusion sizes in engineering steels by fitting a generalized extreme value
[9] Beretta S, Romano S. A comparison of fatigue strength sensitivity to defects for distribution based on vectors of largest observations. Acta Mater 2015;95:1–9.
materials manufactured by AM or traditional processes. Int J Fatigue 2017;94: [40] Murakami Y, Beretta S. Small defects and inhomogeneities in fatigue strength:
178–91. Experiments, models and statistical implications. Extremes 1999;2(2):123–47.
[10] Murakami Y, Kodama S, Konuma S. Quantitative evaluation of effects of non- [41] Murakami Y. Inclusion rating by statistics of extreme values and its application to
metallic inclusions on fatigue strength of high strength steels. I: Basic fatigue fatigue strength prediction and quality control of materials. J Res Nat Inst Stand
mechanism and evaluation of correlation between the fatigue fracture stress and Technol 1994;99(4):345. https://doi.org/10.6028/jres.099.032.
the size and location of non-metallic inclusions. Int J Fatigue 1989;11(5):291–8. [42] ASTM-E2283-08. Standard practice for extreme value analysis of nonmetallic
[11] El Haddad MH, Topper TH, Smith KN. Prediction of non propagating cracks. Eng inclusions in steel and other microstructural features. ASTM Designation E 2283-
Fract Mech 1979;11(3):573–84. 03. ASTM Int.; 2010. p. 1305–12.
[12] Murakami Y, Usuki H. Quantitative evaluation of effects of non-metallic inclusions [43] ESIS-TC20. Technical recommendations for the extreme value analysis of data on
on fatigue strength of high strength steels. II: Fatigue limit evaluation based on large non-metallic inclusions in steels. GKSS, Geesthacht, Germany, European
statistics for extreme values of inclusion size. Int J Fatigue 1989;11(5):299–307. Structural Integrity Society; 2003. ISSN 1616-2129.
[13] Murakami Y. Effects of small defects and nonmetallic inclusions on the fatigue [44] Romano S, Abel A, Gumpinger J, Brandão AD, Beretta S. Quality control of
strength of metals. Mech Corros Propert Ser Key Eng Mater 1991;51-52:37–42. AlSi10Mg produced by SLM: Metallography versus CT scans for critical defect size
[14] Romano S, Brückner-Foit A, Brandão A, Gumpinger J, Ghidini T, Beretta S. Fatigue assessment. Addit Manuf 2019;28:394–405.
properties of AlSi10Mg obtained by additive manufacturing: Defect-based [45] Nourian-Avval A, Fatemi A. Characterization and analysis of porosities in high
modelling and prediction of fatigue strength. Eng Fract Mech 2018;187:165–89. pressure die cast aluminum by using metallography, x-ray radiography, and micro-
[15] Benedetti M, Santus C. Building the Kitagawa-Takahashi diagram of flawed computed tomography. Materials 2020;13(14):3068.
materials and components using an optimized V-notched cylindrical specimen. Eng [46] Geman S, Geman D. Stochastic Relaxation, Gibbs Distributions, and the Bayesian
Fract Mech 2020;224:106810. Restoration of Images. IEEE Trans Pattern Anal Mach Intell 1984;PAMI-6(6):
[16] Hu YN, Wu S, Withers PJ, Zhang J, Bao HYX, Fu YN, Kang GZ. The effect of 721–41.
manufacturing defects on the fatigue life of selective laser melted Ti-6Al-4V [47] Atkinson HV, Shi G. Characterization of inclusions in clean steels: A review
structures. Mater Des 2020;192:108708. including the statistics of extremes methods. Prog Mater Sci 2003;48(5):457–520.
[17] Biswal R, Syed AK, Zhang X. Assessment of the effect of isolated porosity defects on [48] Kitagawa H, Takahashi S. Applicability of fracture mechanics to very small cracks
the fatigue performance of additive manufactured titanium alloy. Addit Manuf or the cracks in the early stage. In: Proc 2nd Int Conf. Mech. Behaviour of Mater -
2018;23:433–42. ICM2; 1976. p. 627–31.
[18] Xie C, Wu S, Yu Y, Zhang H, Hu Y, Zhang M, Wang G. Defect-correlated fatigue [49] Murakami Y. Metal fatigue: Effects of small defects and nonmetallic inclusions.
resistance of additively manufactured Al-Mg4.5Mn alloy with in situ micro-rolling. Oxford: Elsevier; 2002.
J Mater Process Technol 2021;291:117039. [50] Mohammad M, Abdullah S, Jamaludin N, Innayatullah O. Predicting the fatigue
[19] Haridas RS, Thapliyal S, Agrawal P, Mishra RS. Defect-based probabilistic fatigue life of the SAE 1045 steel using an empirical Weibull-based model associated to
life estimation model for an additively manufactured aluminum alloy. Mater Sci acoustic emission parameters. Mater Des 2014;54:1039–48.
Eng, A 2020;798:140082. [51] Fatemi A, et al. Fatigue behaviour of additive manufactured materials: An
[20] Kapoor R, Rao VSH, Mishra RS, Baumann JA, Grant G. Probabilistic fatigue life overview of some recent experimental studies on Ti-6Al-4V considering various
prediction model for alloys with defects: Applied to A206. Acta Mater 2011;59(9): processing and loading direction effects. Fatigue Fract Eng Mater Struct 2019;42
3447–62. (5):991–1009.
[21] Wang X, Meng Q, Hu W. Continuum damage mechanics-based model for the [52] Liao D, Zhu SP, Keshtegar B, Qian G, Wang Q. Probabilistic framework for fatigue
fatigue analysis of welded joints considering the effects of size and position of inner life assessment of notched components under size effects. Int J Mech Sci 2020;181:
pores. Int J Fatigue 2020;139:105749. 105685.
[22] Lee CD, Shin KS. Constitutive prediction of the defect susceptibility of tensile [53] Bažant ZP. Size effect on structural strength: A review. Arch Appl Mech 1999;69
properties to microporosity variation in A356 aluminum alloy. Mater Sci Eng, A (9–10):703–25.
2014;599:223–32. [54] Kloos KH, Buch A, Zankov D. Pure Geometrical Size Effect in fatigue tests with
[23] Zhan Z. Experiments and numerical simulations for the fatigue behavior of a novel constant stress amplitude and in programme tests. Materialwiss Werkstofftech
TA2-TA15 titanium alloy fabricated by laser melting deposition. Int J Fatigue 1981;12(2):40–50.
2019;121:20–9. [55] Alava MJ, Nukala PKVV, Zapperi S. Size effects in statistical fracture. J Phys D Appl
[24] Pei C, Zeng W, Yuan H. A damage evolution model based on micro-structural Phys 2009;42(21):214012.
characteristics for an additive manufactured superalloy under monotonic and [56] Makkonen M. Notch size effects in the fatigue limit of steel. Int J Fatigue 2002;25
cyclic loading conditions. Int J Fatigue 2020;131:105279. (1):17–26.
[25] Wan H, Wang Q, Jia C, Zhang Z. Multi-scale damage mechanics method for fatigue [57] Ai Y, et al. Probabilistic modeling of fatigue life distribution and size effect of
life prediction of additive manufacture structures of Ti-6Al-4V. Mater Sci Eng, A components with random defects. Int J Fatigue 2019;126:165–73.
2016;669:269–78. [58] Tonelli L, Liverani E, Valli G, Fortunato A, Ceschini L. Effects of powders and
[26] Zhu SP, Ai Y, Liao D, Correia JAFO, De Jesus AMP, Wang Q. Recent advances on process parameters on density and hardness of A357 aluminum alloy fabricated by
size effect in metal fatigue under defects: a review. Int J Fract 2021. https://doi. selective laser melting. Int J Adv Manuf Technol 2020;106(1–2):371–83.
org/10.1007/s10704-021-00526-x. In press. [59] Sing SL, Wiria FE, Yeong WY. Selective laser melting of titanium alloy with 50 wt%
[27] Romano S, Brandão A, Gumpinger J, Gschweitl M, Beretta S. Qualification of AM tantalum: Effect of laser process parameters on part quality. Int J Refract Metal
parts: Extreme value statistics applied to tomographic measurements. Mater Des Hard Mater 2018;77:120–7.
2017;131:32–48. [60] Pegues JW, et al. Fatigue of additive manufactured Ti-6Al-4V, Part I: The effects of
[28] Le VD, Pessard E, Morel F, Prigent S. Fatigue behaviour of additively manufactured powder feedstock, manufacturing, and post-process conditions on the resulting
Ti-6Al-4V alloy: The role of defects on scatter and statistical size effect. Int J microstructure and defects. Int J Fatigue 2020;132:105358.
Fatigue 2020;140:105811. [61] Vayssette B, Saintier N, Brugger C, El May M, Pessard E. Numerical modelling of
[29] Makkonen L, Tikanmäki M. An improved method of extreme value analysis. surface roughness effect on the fatigue behavior of Ti-6Al-4V obtained by additive
J Hydrol X 2019;2:100012. manufacturing. Int J Fatigue 2019;123:180–95.
[30] Castillo E, Hadi A, Sarabia JM, Balakrishnan N. Extreme Value and Related Models [62] Hamidi Nasab M, Romano S, Gastaldi D, Beretta S, Vedani M. Combined effect of
with Applications in Engineering and Science. Int Stat Inst 2005:368. surface anomalies and volumetric defects on fatigue assessment of AlSi7Mg
[31] Beretta S. More than 25 years of extreme value statistics for defects: Fundamentals, fabricated via laser powder bed fusion. Addit Manuf 2020;34:100918.
historical developments, recent applications. Int J Fatigue 2021;151:106407. [63] Nakatani M, Masuo H, Tanaka Y, Murakami Y. Effect of Surface Roughness on
[32] L. Makkonen, M. Tikanmäki, Statistical evaluation of extreme ice loads, 2019, p. 5. Fatigue Strength of Ti-6Al-4V Alloy Manufactured by Additive Manufacturing.
[33] Stopka KS, Gu T, McDowell DL. Effects of algorithmic simulation parameters on the Procedia Struct Integrity 2019;19:294–301.
prediction of extreme value fatigue indicator parameters in duplex Ti-6Al-4V. Int J [64] Murakami Y, Takagi T, Wada K, Matsunaga H. Essential structure of S-N curve:
Fatigue 2020;141:105865. Prediction of fatigue life and fatigue limit of defective materials and nature of
[34] Niu XP, Zhu SP, He JC, Ai Y, Shi K, Zhang L. Fatigue reliability design and scatter. Int J Fatigue 2021;146:106138.
assessment of reactor pressure vessel structures: Concepts and validation. Int J [65] Nishijima S. Statistical Analysis of Fatigue Test Data. J Soc Mater Sci Jpn 1980;29
Fatigue 2021;153:106524. (316):24–9.

15
X. Niu et al. International Journal of Fatigue 160 (2022) 106884

[66] Miller KJ. The short crack problem. Fatigue Fract Eng Mater Struct 1982;5(3): [79] Dowling NE. Mechanical behavior of materials: engineering methods for
223–32. deformation, fracture, and fatigue. 2nd ed. NJ: Prentice-Hall; 1998.
[67] Farahmand B, Aliabadi M. Fracture mechanics of metals, composites, welds, and [80] Zhang J, Li H, Yang B, Wu B, Zhu S. Fatigue properties and fatigue strength
bolted joints: Application of LEFM, EPFM, and FMDM theory. Appl Mech Rev 2002; evaluation of railway axle steel: Effect of micro-shot peening and artificial defect.
55(4). Int J Fatigue 2020;132:105379.
[68] Murakami Y, Masuo H, Tanaka Y, Nakatani M. Defect Analysis for Additively [81] Beretta S, Ghidini A, Lombardo F. Fracture mechanics and scale effects in the
Manufactured Materials in Fatigue from the Viewpoint of Quality Control and fatigue of railway axles. Eng Fract Mech 2005;72(2):195–208.
Statistics of Extremes. Procedia Struct Integrity 2019;19:113–22. [82] Romano S, Patriarca L, Foletti S, Beretta S. LCF behaviour and a comprehensive life
[69] Yamashita Y, Murakami T, Mihara R, Okada M, Murakami Y. Defect analysis and prediction model for AlSi10Mg obtained by SLM. Int J Fatigue 2018;117:47–62.
fatigue design basis for Ni-based superalloy 718 manufactured by selective laser [83] Sun C, Song Q. A method for predicting the effects of specimen geometry and
melting. Int J Fatigue 2018;117:485–95. loading condition on fatigue strength. Metals 2018;8(10):811.
[70] Romano S, Miccoli S, Beretta S. A new FE post-processor for probabilistic fatigue [84] Härkegrd G, Halleraker G. Assessment of methods for prediction of notch and size
assessment in the presence of defects and its application to AM parts. Int J Fatigue effects at the fatigue limit based on test data by Böhm and Magin. Int J Fatigue
2019;125:324–41. 2010;32(10):1701–9.
[71] Burhan I, Kim HS. S–N curve models for composite materials characterisation: An [85] Li XQ, Song LK, Bai GC. Recent advances in reliability analysis of aeroengine rotor
evaluative review. J Compos Sci 2018;2(3):38. system: a review. Int J Struct Integr 2022;13(1):1–29.
[72] Basquin OH. The exponential law of endurance tests. Am Soc Test Mater Proc 1910; [86] Narayanan G. Probabilistic fatigue model for cast alloys of aero engine
10:625–30. applications. Int J Struct Integr 2020;12(3):454–69.
[73] “The determination of fatigue limits under alternating stress conditions. Proc R Soc [87] Chopra OK, Shack WJ. Margins for ASME Code fatigue design curve - Effects of
Lond Ser A Contain Pap Math Phys Character 1914;90(620):411–25. surface finish and material variability. In: American Society of Mechanical
[74] Oppermann RH. Prevention of the failure of metals under repeated stress. Engineers, Pressure Vessels and Piping Division (Publication) PVP; 2003. p. 71–85.
J Franklin Inst 1941;232(3):293. [88] Zhu SP, Foletti S, Beretta S. Probabilistic framework for multiaxial LCF assessment
[75] Manson SS. A complex subject-some simple approximations. Exp Mech 1965: under material variability. Int J Fatigue 2017;103:371–85.
193–226. [89] Hertel O, Vormwald M. Statistical and geometrical size effects in notched members
[76] Coffin LF. A study of the effects of cyclic thermal stresses on a ductile metal. Trans based on weakest-link and short-crack modelling. Eng Fract Mech 2012;95:72–83.
Am Soc Mech Eng 1954;76:931–50. [90] Liu X, Liu J, Wang H, Yang X. Prediction and evaluation of fatigue life considering
[77] Morrow JD. Cyclic plastic strain energy and fatigue of metals. Internal Friction material parameters distribution characteristic. Int J Struct Integr 2022;13(2):
Damping Cyclic Plasticity 1965:45–87. 309–26.
[78] Ramberg W, Osgood WR. Description of stress-strain curves by three parameters.
National Advisory Committee for Aeronautics, p. Technical Note No. 902; 1943.

16