Applied Soft Computing 30 (2015) 514–528

Contents lists available at ScienceDirect

Applied Soft Computing

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

Application of artificial neural network for predicting crack growth

direction in multiple cracks geometry
Deepayan Gope a , Prakash Chandra Gope b,∗ , Aruna Thakur b , Abhishek Yadav c,1
a
Electrical Engineering Department, Veer Surendra Sai University of Technology (VSSUT), Burla, Odisha, India
b
Mechanical Engineering Department, College of Technology, G. B. Pant University of Agriculture and Technology, Pantnagar 263145, India
c
Electrical Engineering Department, College of Technology, G. B. Pant University of Agriculture and Technology, Pantnagar 263145, India

a r t i c l e i n f o a b s t r a c t

Article history: The objective of this study is to design an efficient artificial neural network (ANN) architecture in order
Received 25 January 2013 to predict the crack growth direction in multiple crack geometry. Nonlinear logistic (sigmoid and tan-
Received in revised form 2 February 2015 gent hyperbolic) and linear activation functions have been used through the one- and two-hidden layer
Accepted 2 February 2015
ANN. 85 tests were conducted on aluminium alloys under different crack positions, defined by crack tip
Available online 11 February 2015
distance, crack offset distance, crack size, and crack inclination with loading axis. The experimental data
set as first degree or second degree were used to train 22 proposed ANN models to predict the output
Keywords:
for new data sets (not included in the training sets). The model results were then compared with the
Artificial neural network
Multiple cracks
experimental data. It was observed that ANN model with combinations of activation functions and two
Crack inclination hidden layers predict the crack initiation direction with good accuracy when higher order input variables
Crack offset distance are presented to the network.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction
Various fracture criteria for cracks subjected to mixed mode loading have been
introduced for the determination of crack initiation direction and critical stress. Grif-
fith [1] introduced a criterion to determine the conditions to initiate the propagation
of a crack. The maximum energy release rate criterion (MERR) [2,3] followed the
Griffith condition and stated that crack growth follows the orientation of maximum
energy release rate. Erdogan and Sih [4] developed the maximum tangential stress
criterion (MTS-criterion) which was one of the first conditions that predicted criti-
cal stress and crack growth orientation. MTS-criterion stated that the crack growth
would occur in the direction of the maximum tangential stress and would take place
when the maximum tangential stress reaches a critical value which only contains
the first mode strength toughness. Due to its simple formulation, MTS-criterion
became one of the most popular criteria in fracture mechanics. Palaniswamy and
Knauss [5] introduced the G-criterion which dealt with a criterion of maximum
energy release rate to determine both the initial crack propagation direction as well
as the conditions of crack instability in terms of fracture stress, crack orientation
angle and crack length. Sih [6] proposed the minimum elastic strain energy den-
sity factor criterion (SED-criterion) which was based directly on the total strain
energy density, i.e. the sum of its distortional and dilation components. The strain
energy density approach has been found to be a powerful tool to assess the static and
fatigue behaviour of notched and unnotched components in structural engineering
[7]. Different SED-based approaches were formulated by many researchers [8–10].
Labeas et al. [11] presented a methodology to predict mixed-mode crack growth and
∗ Corresponding author. Tel.: +91 9411159916; fax: +91 5944 233338.
E-mail addresses: deepayangope@gmail.com (D. Gope), pcgope@rediffmail.com
(P.C. Gope), ar aruna tk@yahoo.co.in (A. Thakur), ayadav.iitk@gmail.com (A. Yadav).
1
Tel.: +91 9760358154.
orientation using the strain energy density theory. Zuo et al. [12] developed a fatigue
crack growth model to predict the lifetime of fatigue crack growth for single and
mixed mode cracks. In Ref. [13], the S-theory is applied to determine crack initia-
tion and direction of cracked T-beams and circumferentially cracked pipes under
combined loading conditions. The strain energy density factor approach was also
used by Balasubramanian and Guha [14] to analyze the effect of weld size of fatigue
crack growth behaviour of shielded metal arc welded cruciform joints failing from
root region.
Recent studies have shown that the coefficients of higher order terms can also
play an important role in the fracture process in notched or cracked structures.
In addition to the singular term, the higher order terms, in particular, the first non-
singular stress term (known as the T stress) have significant effects on the near notch
tip stress field. The T-stress is considered in some studies as an auxiliary parame-
ter for increasing the accuracy of the results for KI . Kim and Cho [15], for instance,
showed that this non-singular term has noticeable effects on the size and shape of
plastic zone near the notch tip. Ayatollahi and colleagues demonstrated that the
first non-singular term have considerable contributions to the stress components
around the notch tip and also on the fracture resistance of notched components
under mode I loading [16–18]. Furthermore, recent studies on cracks show that the
first non-singular stress term in the Williams expansion (n = 2) could significantly
influence the stress field around the crack tip. These effects have been investigated
experimentally by the technique of photoelasticity in both mode I and mode II load-
ing conditions. Ayatollahi and Sedighiani [19] studied the effect of T-stress on the
critical mode I stress intensity factor of brittle and quasi-brittle materials. The min-
imum strain energy density criterion was revisited to take the effect of T-stress
into account. In Ref. [20] the strain energy density criterion was applied to deter-
mine the direction of crack initiation for various biaxial load factors and material
combinations. All these criteria have found applications for predicting the fatigue
crack growth direction in structures with a single crack. However, use of these cri-
teria with caution for multiple cracks, which, in general, have less accuracies than a
single crack and generally inadequate for the analysis of multiple cracks.

http://dx.doi.org/10.1016/j.asoc.2015.02.003
1568-4946/© 2015 Elsevier B.V. All rights reserved.
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528
Nomenclature
a crack length
r correlation coefficient
t student t distribution
H crack offset distance
S crack tip distance
T first non-singular stress term, T stress
KI mode I stress intensity factor
KII mode II stress intensity factor
˛ crack inclination angle
 learning parameter
0 crack initiation angle
 Poisson’s ratio
The interaction between multiple cracks has a major influence on crack growth
behaviour. This influence is particularly significant in stress corrosion cracking (SCC)
[21], welding [22,23], lifetime estimation of plant components, etc. as relatively large
number of cracks are initiated due to environmental effects. The interaction between
the multiple cracks can be studied either analytically or experimentally. One can find
number of studies in the literature using various analytical and numerical methods
to treat multiple crack problems. Reviews of various methods are given in Burton
and Phoenix [24] and TerMaath [25]. Development of analytical methods to calculate
the stresses around multiple interacting cracks in elastic solids has been an impor-
tant but difficult research problem for many decades. Such crack interaction effects
are an important aspect in the ageing of such structures as aircraft fuselages and
marine hulls, as well as in tracking damage evolution in poly-crystalline and mul-
tiphase materials in general. In an array of cracks, the strengths of the interactions
are strongly dependent on their local geometry, placement, and orientation and
especially on the proximity of crack tips relative to the crack lengths. Interactions
that do not adversely affect stiffness may greatly reduce the strength and fatigue
life of a structure, and vice versa. In analytical studies on multiple crack geometry,
the stress intensity factors are either evaluated from the governing equations avail-
able for mixed mode cracks [21] or by finite element analysis [26–29]. The presence
of neighbouring cracks in multiple crack system affects the growth direction. This
growth direction depends upon the relative crack positions. As there is no standard
relation to compute the crack growth direction based on the relative crack positions,
the growth direction is computed either by using strain energy density (SED) cri-
terion [8–10] or maximum tangential stress (MTS) criterion [4,26]. In many studies
the crack growth direction is assumed to possess a self similar manner throughout
the growth simulation [27]. As the experimental assessments of crack growth direc-
tion and propagation in multiple crack have limited literature, the prediction made
by any of the available criterion differs significantly from the actual crack growth
direction and hence on the component’s life. In case of multiple crack, the major
challenge is finding a local parameter which controls the growth of an individual
crack due to mechanically applied loads. Multiple crack problems are composed
of cracks that have disparate characteristics and therefore are not intuitively pre-
dictable. The level of uncertainty can be high; in some cases, high enough to defy
common sense. The second challenge is finding the global parameter which con-
trols the interrelationship between the growth of the various cracks. Every physical
problem in fracture mechanics is a multiple cracks problem, or a problem involv-
ing both major and minor cracks. A common approach in the literature is to treat a
multiple crack effect as a single crack problem which is not true and away from the
reality.
Multiple cracks can interact and coalesce in different ways, depending on their
spatial resolution, applied stresses, and the component geometry. Kamaya [27] and
Liu et al. [30] have suggested that this phenomenon may influence the fatigue
life of a component due to the potential of cracks to interact and accelerate the
growth rate compared to cracks that grew individually and independently. The
trend of operating aircraft up to or even exceeding their design lives has raised
concerns on the airframe structural integrity due to fatigue failures, which often
involve multiple cracks [31–34]. Experiments and numerical studies conducted
by Iida [35], Iida and Kawahara [36], Wang et al. [37], Frise and Bell [38], Leek
and Howard [39], and Kamaya [40], which were mainly on long crack coalescence
showed that the effects of interaction amongst closely spaced cracks can signifi-
cantly influence the rate of coalescence, and hence, accelerating the crack growth
rates. This will in turn potentially affect the overall fatigue life of engineering struc-
tures. Multiple interacting crack problems have also been treated in Helsing [41]
and Helsing and Jonsson [42] for rectangular plates, the former using fast multi
pole techniques and the later using special 1/2-power weight functions reflect-
ing the singular stress fields at crack tips. Neither paper deals specifically with
crack kinks or branches and the former does not compute stress intensity fac-
tors for crack tips, but focuses rather on overall stiffness reduction in the plate.
Kachanov [43] presented a useful review of various approaches and the formidable
515
Table 1
Mechanical properties.
Sl. no. Yield Ultimate E (GPa) Hardness
strength strength (B-scale)
(MPa) (MPa)
Mean 99.50 106.50 70.00 51.50
SD 5.00 7.55 3.00 2.88
SD: standard deviation.
difficulties that arise in attempting to deal with closely interacting cracks and their
coalescence.
ANN is a new class of computational intelligence system and is useful to handle
various complex problems with a capacity to learn by examples. ANN has proved
to be a powerful and versatile computational tool in the application of a number
of engineering fields [44–49]. In recent years, ANN has also been introduced in the
field of fracture mechanics in order to predict fatigue life [50–56] of metals and
composite. Artymaik et al. [50] expressed that, with the use of ANN method, fatigue
life can be estimated on the basis of material and stress parameters. Genel [51]
used ANN method to predict strain-life fatigue properties of steels on the basis
of tensile tests. He has shown that the fatigue strength coefficient and fatigue
ductility coefficient can be predicted using ANN with high accuracy of approx-
imately 99% and 98%, respectively. Bucar et al. used ANN approach to describe
the scatter of S–N curves. Venkatesh and Rack [52] trained a feed-forward ANN
to predict the elevated temperature creep-fatigue behaviour of Ni-based alloy. Ple-
une and Chopra [53] also trained a feed forward ANN to predict the fatigue life of
carbon and low alloy steels for specified sets of loading and environmental condi-
tions. Recently Pujol et al. [54] studied the fatigue life prediction under step-stress
conditions, where the cumulative distribution function for the failure of compo-
nents was implemented by means of a neural network. The Kolmogorov–Smirnov
test was applied to compare the results of the new approach to those obtained
with the lognormal distribution function. Bucar et al. [55] used neural network
approach to describe the scatter of S–N curves. Mohanty et al. [57] have predicted
the fatigue crack propagation life of 7020 T7 and 2024 T3 aluminium alloys under
the influence of load ratio by using artificial neural network. Zarrabi and Tsang [58]
presented a new and relatively simple criterion for the mixed-modes I and II frac-
ture that uses an artificial neural network database to obtain the pertinent fracture
toughness.
The crack growth under multiple cracks is no longer in mode I condition. The
crack growth rate is highly influenced by the presence of neighbouring cracks
and the crack also grows in a non-self similar manner as seen in mixed mode
cracks. Although there are a number of criteria for the mixed-mode fracture, none
of them are universally accepted and applied to the practical problems. These
criteria are not suitable for multiple crack configurations. No empirical or math-
ematical relations is available to predict the crack initiation and growth direction
for multiple crack configuration from the crack offset distance, crack tip distance,
crack sizes, ratio of plate thickness to crack size, etc. Hence, other alternate meth-
ods such as artificial neural network and finite element method seem to be the
best alternatives for such studies. However, to the best of authors’ knowledge, no
literature related to prediction of growth direction using ANN technique under
multiple crack geometry is available till date. It is very important to find the
growth direction under the influence of multiple cracks to predict the life of a
component.
In this study, ANN has been used to predict crack growth direction in multiple
crack geometry under static loading. The novelty of the current method is its direct
applicability to multiple cracks since the ANN is tested and validated for the problem.
An original and crucial feature of the work is the evaluation of 22 ANN models with
different ANN architectures to evaluate the crack growth direction, the area where
no single analytical methods proposed so far is suitable for the prediction of the
crack growth direction under the influence of second or more cracks. The current
paper shows the suitability of ANN method to incorporate the challenges on the
predictive capabilities involving interactions between two cracks at different offset
distances, crack tip distance and crack size ratio. For the most challenging cases, the
presented problem is solved in a short time in MATLAB with a notebook PC. In this
study the effect of degree of input variables, number of hidden layers, homogeneous
and hybrid combination of activation functions on the performance of the network
to predict the crack initiation angle has been studied. The predicted results have
been compared with the experimental data. It is observed that the results are in
good agreement with the experimental findings.
2. Material and specimen geometry
In the present investigation the commercially available alu-
minium alloy is used. The sheet thickness of aluminium alloy is
1.0 mm. The mechanical properties are given in Table 1. Rectangu-
lar pieces of required size were cut from 1.0 mm thin aluminium
sheet. Length of each specimen was kept 200 mm. The width of
516 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528
edge crack specimen was taken as 60 mm while the width of the
central crack geometry specimens was selected so as to obtain the
distance from the outer crack tip to the edge of the specimen equal
to or slightly more than the crack tip distance. Mechanical slits were
made at the required position (Fig. 1) by means of a jewellery saw
of 0.1 mm thick blade. For central cracks, at the marked positions
on the plate, two holes of diameter 2.0 mm were made by drilling
and then a crack was formed by cutting with the jewellery saw. All
inclined cracks were made at an angle of 60◦ to the loading axis (ver-
tical axis, which is also the rolling direction of the plate used in this
investigation). The mechanical cracks so obtained are pre-cracked
under fatigue loading with very low constant amplitude loading.
The lengths of the fatigue pre-cracks were kept below 0.5 mm. The
study of crack growth direction under static loading was conducted
with two edge cracks on one side (A1), single edge crack on both
sides (A2), two central cracks (C1), and two central inclined cracks
(C2). The specimen details for above type of crack configurations
are shown in Fig. 1.
Fig. 1. Specimen geometry. Dimensions are in mm and not to the scale.
3. Selection of crack geometry parameters
It has already been discussed in Section 1 that the intensity
of the crack interaction and the effect on the crack growth and
crack initiation direction depends on several specimen geometry
parameters such as crack tip distance S, crack offset distance H, and
relative crack size a2 /a1 . The simulation work of Kamaya et al. [59]
by using S-version of finite element method revealed that when
the ratio of crack offset distance to crack length H/a is small and
a2 /a1 is large, the interaction between the two cracks is significant,
whereas, it had no effect in the opposite case. The combination
of H, S and crack length a1 and a2 considered by Kamaya et al.
[59] were 2.5 ≤ S ≤ 20 mm, 2.5 ≤ H ≤ 20 mm and 1 ≤ a2 /a1 ≤ 0.2. It
was also reported in their study that there was no interaction
between the two cracks for a2 < 0.6a1 . Thus, when the difference
in the crack size is large enough, it is not necessary to account
the interaction in the growth prediction, and the single crack can
be used for analysis. The growth, interaction and coalescence of
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528 517

Table 2
Relative positions of multiple crack geometry and crack initiation direction used in this study.

Test no. Specimen ID Initial crack size (mm) Initial crack inclination Relative position (mm) Crack initiation
from loading axis (◦ ) direction (◦ )

a1 a2 ˛1 ˛2 H S 
1 2 3 4 5 6 7 8 9

1 A11 10.0 10.0 90 90 3.0 0.0 6.0

2 A12 12.0 12.0 90 90 10.0 0.0 5.0
3 A13 5.0 5.0 90 30 13.0 0.0 15.0
4 A14 7.5 5.0 90 45 13.0 0.0 10.0
5 A15 7.0 7.0 90 80 14.0 0.0 11.0
6 A21 4.0 4.5 90 90 2.0 0.0 2.5
7 A22 4.5 4.5 90 90 4.0 0.0 7.5
8 A23 4.0 4.5 90 90 6.0 0.0 7.0
9 A24 4.5 4.5 90 90 10.0 0.0 4.5
10 A25 28.0 27.0 90 90 2.0 2.0 16.5
11 A26 26.0 26.0 90 90 2.0 4.0 21.0
12 A27 25.0 24.0 90 90 2.0 7.0 18.5
13 A28 24.0 24.0 90 90 2.0 8.0 10.5
14 A29 23.0 22.0 90 90 2.0 11.0 7.0
15 A210 21.0 21.0 90 90 2.0 14.0 11.0
16 A211 27.0 26.0 90 90 4.0 2.0 52.0
17 A212 25.0 26.0 90 90 4.0 4.0 36.0
18 A213 25.0 25.0 90 90 4.0 6.0 34.5
19 A214 24.0 25.0 90 90 4.0 7.0 25.25
20 A215 22.0 23.0 90 90 4.0 11.0 15.5
21 A216 12.5 12.5 90 90 2.0 35.0 7.5
22 A217 12.5 12.5 90 90 4.0 35.0 9.5
23 A218 12.5 12.5 90 90 6.0 35.0 11.0
24 A219 12.5 12.5 90 90 10.0 35.0 20.5
25 A220 12.5 12.5 90 90 20.0 35.0 31.0
26 A221 17.5 17.5 90 90 2.0 25.0 8.5
27 A222 17.5 17.5 90 90 4.0 25.0 10.5
28 A223 17.5 17.5 90 90 6.0 25.0 15.5
29 A224 17.5 17.5 90 90 10.0 25.0 20.5
30 A225 17.5 17.5 90 90 20.0 25.0 36.0
31 C11 4.20 0.0 45 45 0.0 0.0 41.0
32 C12 4.0 0.0 60 60 0.0 0.0 45.0
33 C13 4.0 0.0 75 75 0.0 0.0 16.5
34 C14 4.0 0.0 30 30 0.0 0.0 50.5
35 C15 4.0 0.0 60 60 0.0 0.0 36.0
36 C21 4.0 4.0 90 90 4.0 16.0 12.5
37 C22 4.0 4.0 90 90 4.0 16.0 16.5
38 C23 4.0 4.0 90 90 4.0 16.0 8.5
39 C24 4.0 4.0 90 90 10.0 16.0 15.5
40 C25 4.0 4.0 90 90 8.0 16.0 14.5
41 C26 4.0 4.0 90 90 8.0 16.0 15.75
42 C27 4.0 4.0 90 90 6.0 16.0 11.5
43 C28 4.0 4.0 90 90 12.0 8.0 40.5
44 C29 4.0 4.0 90 90 12.0 14.0 18.0
45 C210 4.0 4.0 90 90 12.0 16.0 15.5
46 C211 4.0 4.0 90 90 12.0 20.0 14.5
47 C212 4.0 4.0 90 90 5.0 20.0 17.5
48 C213 4.0 4.0 90 90 14.0 20.0 14.0
49 C214 4.0 4.0 90 90 10.0 12.0 23.0
50 C215 4.0 4.0 90 90 10.0 20.0 17.0
51 C216 4.0 4.0 90 90 10.0 8.0 24.0
52 C217 4.0 4.0 90 90 10.0 14.0 15.0
53 C31 4.0 4.0 60 60 4.0 24.0 24.0
54 C32 4.0 4.0 60 60 6.0 24.0 12.75
55 C33 4.0 4.0 60 60 8.0 24.0 8.0
56 C34 4.0 4.0 60 60 10.0 24.0 13.0
57 C35 4.0 4.0 60 60 14.0 24.0 11.0
58 C37 4.0 4.0 60 60 4.0 12.0 39.5
59 C38 4.0 4.0 60 60 6.0 12.0 36.0
60 C39 4.0 4.0 60 60 8.0 12.0 29.0
61 C310 4.0 4.0 60 60 10.0 12.0 17.0
62 C311 4.0 4.0 60 60 24.0 12.0 19.0
63 C312 4.0 4.0 60 60 24.0 6.0 37.75
64 C313 4.0 4.0 60 60 24.0 8.0 33.5
65 C314 4.0 4.0 60 60 24.0 10.0 46.0
66 C315 4.0 4.0 60 60 24.0 16.0 19.5
67 C316 4.0 4.0 60 60 12.0 6.0 25.0
68 C317 4.0 4.0 60 60 12.0 11.0 30.5
69 C318 4.0 4.0 60 60 12.0 12.0 33.5
70 C319 4.0 4.0 60 60 12.0 14.0 44.25
71 C320 4.0 4.0 60 60 12.0 16.0 14.5
518 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528

Table 3
Relative positions of multiple crack geometry and crack initiation direction used in this study for validation of ANN.

Test no. Specimen ID Initial crack size (mm) Initial crack inclination Relative position (mm) Crack initiation
from loading axis (◦ ) direction (◦ )

a1 a2 ˛1 ˛2 H S 
1 2 3 4 5 6 7 8 9

1 A16 3.0 3.0 90 90 2.0 0.0 11.0

2 A17 3.0 3.0 90 90 3.0 0.0 8.0
3 A18 3.0 3.0 90 90 3.5 0.0 7.0
4 A19 3.0 3.0 90 90 6.0 0.0 12.0
5 A120 3.0 3.0 90 90 2.5 0.0 10.0
6 C111 4.0 4.0 90 90 4.0 16.0 17.0
7 C112 4.0 4.0 90 90 10.0 10.0 17.0
8 C113 4.0 4.0 90 90 10.0 8.0 25.0
9 C114 4.0 4.0 90 90 12.0 8.0 22.0
10 C16 4.0 4.0 90 90 0.0 0.0 42.0
11 C321 4.0 4.0 60 60 6.0 24.0 15.0
12 C322 4.0 4.0 60 60 12.0 24.0 30.0
13 C323 4.0 4.0 60 60 14.0 6.0 20.0
14 C324 4.0 4.0 60 60 12.0 10.0 32.0

two coplanar short cracks of varying lengths in AA7050-T7451

aluminium alloy were studied under low amplitude cyclic load-
ing by Tan and Chen [60]. It was reported that as the ratio of the
minor crack to major crack length a2 /a1 approached unity, the
influence of interaction on the overall crack shape development
increased to a maximum value and interaction was insignificant
when the crack length ratio a2 /a1 ≤ 0.25. The interaction of multi-
ple cracks in a finite plate studied by Yan and Miao [61] by using
the hybrid displacement discontinuity method indicates that the
crack interaction is insignificant beyond H/a = 4 for all investigated
cases of H/a between 0.1 and 7. In the previous work of Gope and
Thakur [62] on the multiple crack interaction, it was shown that
the crack initiation angle remains less than 10◦ for 0 < S ≤ 12 mm
and H < 2 mm and it exceeds 30◦ for H > 12 mm. The crack initiation
angle remains between 10◦ and 30◦ for 2 mm ≤ H ≤ 12 mm. As men-
tioned in [59–62] the significant crack interaction effect is seen for
H/a <1 and 0 ≤ S/a <1 or H/a ≥1 and S/a ≥1 for equal crack size. Thus, Fig. 2. Illustration of crack initiation angle computation from the fractured speci-
in the present investigation the magnitude of S and H are selected men using CAD software.
appropriately such that 14 < H/a ≤ 0 and 14 < S/a ≤ 0. Thus 85 sets
of test data are designed to include all the combination of H/a and
there is a neighbouring crack. It was shown that the growth direc-
S/a to investigate the significant and insignificant crack interaction
tion is highly influenced by the crack position parameters defined
zone. The number of test samples considered in different zones of
as crack offset distance, crack tip distance and relative crack size.
H/a and S/a for different specimen geometry A1 , A2 , C1 , and C2 along
From the wide range of crack statistics considered in this study few
with the crack initiation angles are presented in Tables 2 and 3.
photographs of the fractured specimens are shown in Figs. 3–7.
The initial condition of S and H are mentioned on the photographs.
4. Test procedure The fractured specimen with two edge cracks on one side of the
plate for crack sizes a1 = a2 = 3.0 mm and H/a <1 are shown in Fig. 3.
The specimen was fixed in the 100 kN servo hydraulic univer- Fig. 4 shows the photographs for H/a ≥ 1 and S/a=0. The effect of the
sal testing machine (ADMET, USA) in static mode and load was presence of neighbourhood crack on growth behaviour and frac-
applied monotonically at 0.01 mm/s crosshead speed. The crack tured pattern is different for H/a <1 and H/a ≥1. It is seen that when
initiation and the stable crack growth during static loading was the cracks are separated with a very small distance (H < a) both
monitored by a travelling microscope (model RTM-500) and CCD the cracks propagate at an angle so as to meet each other at a dis-
camera (Model Sony Cyber Shot 8 M.P. DH-7) attached to another tance equal to or less than the crack length. Thereafter, it propagates
computer. The loading was continued till the specimen fractured. almost perpendicular to the loading axis as a single crack. When the
The crack initiation angle  0 was then measured from the frac- crack offset distance H is further increased and H/a ≥ 1, it is seen that
tured specimen and used in training validation and testing of ANN only one crack propagates in self similar manner and other crack
architecture. The procedure of measurement of the  0 using the becomes dormant. The crack growth direction for two edge cracks
CAD software is illustrated in Fig. 2. The results are presented in on one side with initial conditions a1 /a2 = 1, H/a ≥ 1 and S/a = 0 is
Tables 2 and 3. found to be less than 10◦ . According to MTS [4,19], SED [15–18],
DET [36] criterion, the crack extension angle  0 remains less than
5. Crack initiation direction 10◦ for KI /KII >10. This indicates that crack grows mostly under the
influence of mode I and the effect of mode II stress intensity factor
It is well established that a crack under mode I loading propa- is about 1/10th of mode I and hence the effect of mode II is negligi-
gates in self similar manner where as under mixed mode loading ble as compared to mode I. Hence, it can be said that for H/a ≥ 1 and
it propagates in non-self similar manner. Recently, it was studied S/a = 0, there is a little effect of the presence of neighbourhood crack
by the authors [62] that a crack changes its growth direction when on the growth direction as the crack grows almost perpendicular
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528
Fig. 3. Crack path for two edge crack specimens (on one side) for H/a <1.
Fig. 4. Crack path for different one sided edge crack specimen configurations for
H/a ≥ 1.
to the loading axis, similar to mode I crack growth. However, when
two edge cracks are very close to each other and H/a < 1, the crack
growth direction is highly influenced by the neighbourhood cracks.
It is seen that both the cracks bend towards each other and hence,
the growth direction is increased with respect to the previous case
of H/a ≥ 1. Under such condition it is found that the crack initiation
direction remained more than 10◦ . This indicates that when two
cracks are very close to each other the state of stress produced by
one crack is affected by the state of stress of neighbourhood crack.
Hence, the resultant state of stress is under mixed mode condition
and crack grows under the influence of mixed mode condition.
Fig. 5 shows photographs of the fractured specimen having
two central cracks with different crack configurations H/a ≥ 1 and
S/a > 1. The crack deviation of adjacent crack tips before coalescence
is seen in these figures. In addition to this effect, the propagation
of the crack from outer tips is almost perpendicular to the loading
direction or the crack propagation direction is less than 10◦ . It is
519
Fig. 5. Crack path for different central crack specimen configurations for S/H ≤ 1.
also seen that the progress of the crack from the outer tips is more
as compared to the two inner crack tips. The above behaviour is
seen in all cases when the crack offset distance H/a ≥ 1 and S/H ≤ 1.
No such deviation of adjacent crack tip is seen for other cases as
shown in Fig. 6 for H/a > 1 and S/H = 1. Under such crack statistics
H/a > 1 and S/H > 1, the crack tips of the two facing cracks change
their progress direction so that they approach each other and finally
crack coalescence occurs. It is seen that the fracture of the speci-
men occurred from the outer edge. This indicates that the growth
rate of outer and inner crack tips is also different and depends upon
the H and S values. Fractured photographs show that the crack coa-
lescence phenomena occurred when 1.25 ≤ H/a ≤ 2.5 and S/a > 1 are
satisfied whereas failure occurs because of the progress of the crack
from outer tips for H/a > 2.5. This illustrates that S and H as well as
the edge distance from the outer crack tip plays a major role in
the fracture process in multiple crack geometry. Only the ratio of
S/H is not sufficient to characterize the fracture process or crack
interaction in multiple central cracks. The results shown in Fig. 5
for two central mode I crack indicates that the crack growth from
the inner crack tips is restricted by the stress shielding effect. The
figure indicates that the growth from the outer crack tip is more as
compared to inner tip. This is because of crack shielding. From the
appearance of the crack configuration, it can be said that the accel-
eration effect seems to exceed the shielding effect. This highlights
that while making any simulation studies for such crack statistics
entire crack growth process should be considered.
520 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528

the progress of the inner crack tips becomes dormant due to the
stress shielding effect caused by the presence of a second crack.
This kind of crack behaviour is seen for S < 12 mm or S/H < 1.0. The
progress of crack tip for (a) H = 12 mm, S = 12 mm, ˛ = 60◦ and (b)
H = 12 mm, S = 14 mm, ˛ = 60◦ are shown in Fig. 7(g and h), respec-
tively. It is observed from Fig. 7(g and h) that inner crack tips of two
facing cracks progress in a direction so as to approach each other.
The outer crack tip progresses perpendicular to the loading direc-
tion. This type of crack behaviour is seen for all inclined cracks with
S/H ≥ 1.
The variations of the crack initiation direction with the crack
statistics and stress intensity factors or mode mixity are shown
in Figs. 8 and 9. The crack initiation angle is determined from the
fractured specimens. Fig. 8 demonstrates the intensity of the crack
interaction in terms of crack deviation and crack geometry param-
eters. Fig. 8 reveals that it is difficult to establish any correlation
between the crack initiation angle and crack geometry parame-
ters. The predictive uncertainty that exists with one of the mostly
used fracture criterion (MTS criterion) is shown in Fig. 8. Particle
filled or fibre reinforced composite material is one of the example
where multiple crack sites exist and crack grows from multiple sites
during mechanical loading. The crack initiation angle as a function
of mode mixity Me for composite material containing 20 wt% wal-
nut shell particle and 10 wt% coconut fibre filled in epoxy matrix is
shown in Fig. 9. The mode mixity is defined as:
2
K 
Me =
I
arctan (1)
 KII
The predicted results included in Figs. 8 and 9 are briefly
described here. According to maximum tangential stress (MTS) cri-
terion [4], the crack extension angle in terms of stress intensity
factors is given as
  
Fig. 6. Crack path for central crack specimen configurations for S/H >1. 1 KI
 K 2
−1 I
0 = 2 tan + +8 (2)
4 KII KII
It is seen that the crack initiation direction remained between
10◦ and 20◦ for H/a ≥ 1 and 1 ≤ S/H ≤ 4. The crack initiation direction According to strain energy density (SED) criterion [8–10], the
became more than 20◦ when S/H < 1. growth direction is obtained from the solution of Eq. (3).
Fig. 7(a and b) shows the photographs of two inclined crack, KI KII
˛ = 30◦ and 60◦ . It shows that growth is under the influence of mode (−41,2 sin 0 + sin 20 ) + (41,2 sin 0 − 3 sin 20 )
KII KI
I and mode II stress condition and hence, it deviates from the ini-
tial crack direction. Fig. 7(c) shows two inclined central cracks for + (−41,2 cos 0 + 4 cos 20 ) = 0 (3)
˛ = 60◦ and S/H ≥ 4. Under the crack statistics S/H ≥ 4 it is seen that where 1 = 3 −4 for plane strain case and 2 = (3 − )/(1 + ) for
failure is due to meeting of the two inner crack tips and crack grows plane stress case.
almost in the direction of crack i.e. about 30◦ to the loading axis. According to generalized maximum tangential stress (GMTS)
However, when the crack offset distance H is increased and S/H < 4 criterion [63] the crack initiation direction can be obtained by solv-
in Fig. 7(d), the outer crack tip is deflected and the growth direction ing Eq. (4).
of the farther tip turned towards the loading direction. The crack
growth behaviour for S ≤ H is shown in Fig. 7(e and f). It is seen that 16T  0
[KI sin 0 + KII (3 cos 0 − 1)] − 2rc cos 0 sin =0 (4)
outer crack tip progresses in a non-self similar manner and there is 3 2
no growth from the inner crack tips. where rc is considered equal to the crack tip damage zone. It is given
The results shown for two inclined cracks indicates that with by Taylor [64] as
smaller value of H < 10 mm and S/H ≥ 4, the growth is along the
1
 K 2
IC
crack direction both from inner and outer crack tip. This may be rc = (5)
because of the line joining between inner crack tips makes an angle 2 u
almost greater than 70◦ with the loading direction. It is experimen-  u is the ultimate tensile strength. KIC is the critical mode I stress
tally shown [MTS, etc. criterion] that an inclined crack with about intensity factor or material fracture toughness.
70◦ crack angle propagates at an angle less than 15◦ . It means the Eqs. (2)–(4) show that  0 depends upon stress intensity factors
effect of mode II stress intensity factor is negligible. The same mech- KI and KII . Figs. 8 and 9 show that none of the existing criteria is
anism may exist under such crack statistics and growth direction suitable to predict the crack initiation direction within a reason-
is less affected. When value of H is increased and S/H is kept less able accuracy. The accurate prediction of crack initiation direction
than 4, the deflected crack growth seen from the outer crack tips is important to study the growth rate under multiple crack geom-
may be due to the existence of mixed mode condition. etry and to estimate the components’ life. The difference between
Fig. 7(e and f) shows the stress shielding effect for the case experimental and predicted results shown in Figs. 8 and 9 is due
H = 12 mm, S = 6 mm, ˛ = 60◦ and S = 11 mm, H = 12 mm, ˛ = 60◦ . to the dependence of stress intensity factors on the crack geome-
Fig. 7(e and f) shows that the outer tip of the cracks progress and try parameters. In multiple crack geometry, it was demonstrated
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528
Fig. 7. Crack paths for different central inclined crack specimen configurations. Initial crack conditions are (a) ˛ = 30◦ and (b–h) ˛ = 60◦ .
in the earlier studies [62,64] that both KI and KII and hence, the
crack initiation direction depend on the crack position parameters
S and H (known as crack spacing parameters) and orientation. Fig. 8
shows that crack initiation angle  0 decreases as S/H increases and
beyond S/H > 3,  0 became less than 10◦ . In terms of stress inten-
sity factors KI and KII ,  0 becomes less than 10◦ when KI /KII > 10.
This indicates that for a given crack offset distance H as the crack
tip distance S increases (i.e. cracks are away from each other) the
interaction effect diminishes and crack grows under the influence
of mode I with negligible effect of mode II. In case of central crack
when H and S are relatively small, both the crack behaves like a
coalesced crack and the progress direction is less affected due to
the presence of other cracks or neighbourhood crack. It is also seen
that for larger value of offset distance H (i.e. distance between two
cracks is higher) the interaction also minimizes. It means when
cracks are separated through a larger distance the state of stress
produced ahead of the crack tip remains unaffected and thus dimin-
ishing nature of crack initiation angle is seen. This shows that the
521
interaction of multiple crack geometry can be well related through
the crack initiation angle, crack tip distance and crack offset dis-
tance. This has its own importance also because crack tip factors
during the crack growth test are easy to compute as compared to
stress intensity factors. Similar results are revealed in the work of
Pook [65]. Pook [65] demonstrated that the stress intensity factors
of closely spaced regular arrays of edge cracks depend primarily on
the crack spacing and are approximately proportional to the square
root of crack spacing for crack length to spacing ratios down to 0.16.
For smaller spacing ratios the stress intensity factor depends on
crack length rather than crack spacing. The general features associ-
ated with crack interaction and coalescence and the conditions for
coalescence in various forms are reviewed by Wang et al. [66]. To
the best of authors’ knowledge, the study focusing to the prediction
of crack initiation direction under multiple crack configurations is
very limited and hence, it is difficult to predict the crack growth
direction using the above mentioned criterion with respect to crack
geometry parameters S and H. Hence, in many studies with multiple
522 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528

10 18 Back-propagation networks are in fact the powerful networks

S/H [62]
which refer to a multi-layered, feed-forward perceptions trained
[62] with an error back-propagation algorithm (error minimization
15
8 Expt (Central crack) technique). The architecture of a simple back propagation ANN
Expt (Edge crack) is a collection of nodes distributed over a layer of input nodes,
K I/KII 12 one or more layers of hidden neurons, and a layer of output neu-
MTS Criterion
6 rons. Neurons in each layer are interconnected to subsequent layer

KI/KII
neurons with links, each of which carries a weight that describes
S/H

9
the strength of that connection. In this investigation, both the
4
homogeneous and the hybrid combinations of sigmoid, tangent
6
hyperbolic, and linear activation functions were used with one
S/H
or two hidden layer network architecture to model the neuron
2
3 activity. Twenty-two such combinations have been used; Table 4.
Inputs are propagated forward through each layer of the network to
emerge as outputs. The errors between those outputs and the target
0 0
(desired output) are then propagated backward through the net-
0 10 20 30 40 50
work and then connection weights are adjusted so as to minimize
Crack initiation angle, deg the sum square error. It is calculated as:

1 
P K
Fig. 8. Effect of crack tip and crack offset distances S and H on the variation of crack
initiation angle with the predicted ones. E= (Dpk − Opk )2 (6)
2
i=1 k=1

crack geometry configurations, the growth direction in the simu-
lation has been assumed as  0 = 0 [27]. To overcome this difficulty,
the aim of the present investigation is focused on the development
of ANN model to predict the crack growth direction under multiple
crack geometry.
6. Artificial neural network approach
where Dpk is the desired output and Opk is the predicted output for
the pth pattern and kth output. During training, an ANN is presented
with data for several times, which is referred to as training cycles.
After each cycle, the error between the ANN output (predicted) and
the desired values are propagated backward so as to adjust the
weight in a manner mathematically guaranteed to converge [30].
The learning algorithm used here for the weights is

wij (t + 1) = wij (t) + wij (t) (7)

Artificial neural network (ANN) is an intelligent tool with paral-
lel computational capability. It can perform nonlinear mapping in
short duration. Once neural network is trained, it provides accept-
able recommendations in a short time. The concept of an artificial
neural network is inspired from the working of the human brain.
Human brain is made of hundreds of billions of processing ele-
ments, known as neurons. The brain acts as a highly complex
nonlinear parallel computer. Similarly an artificial neural net-
work is a massive parallel distributed processor with neurons as
processing units. An artificial neural network acquires the knowl-
edge during the training process by correcting the corresponding
weights so as to minimize an error function. One of the main advan-
tages of this technique is that there is no need to make a prior
assumption about the concerned material behaviour, making it
superior to other conventional prediction techniques.
with its correction
∂E
wij (t + 1) = −˛ + ˇ wij (t) (8)
∂wij
where ˛ is the learning rate, ˇ is the momentum coefficient, and
t is the current step. In the present investigation, it is found that
for ˇ = 1, the oscillation and convergence is appreciable. Detailed
derivation of the back propagation algorithm has been covered in
literature extensively [30,31].
Fig. 10 illustrated one of the ANN structure out of 22 ANN archi-
tectures presented in Table 4 with one input layer of six nodes,
one hidden layer of nine nodes and one output layer of one node.
Each layer has different number of neurons. The functional rela-
tionship between input and output variables can be established

90
SCAB (Biocomposite)
80 SCAB (Epoxy)
Crack initiation angle (-θ0), deg

MTS, T*= 0
70
MTS, T*= 2.5
60 MTS, T*= 4

50

40

30

20

10

0
0 0.2 0.4 0.6 0.8 1 1.2
e
M

Fig. 9. Effect of mode mixity Me and T stress on the crack initiation angle and predicted capability by MTS criterion.
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528 523

Table 4
Combination of different activation functions.

Number Code Degree of input data 1 hidden layer 2 hidden layers Output layer Type

1 t-t-1 First Tangent hyperbolic – Tangent hyperbolic Homogeneous

2 t-s-1 First Tangent hyperbolic – Sigmoid Hybrid
3 s-t-1 First Sigmoid – Tangent hyperbolic Hybrid
4 t –l-1 First Tangent hyperbolic – Linear Hybrid
5 s-l-1 First Sigmoid – Linear Hybrid
6 l-l-1 First Linear – Linear Homogeneous
7 t-t-2 Second Tangent hyperbolic – Tangent hyperbolic Homogeneous
8 t-s-2 Second Tangent hyperbolic – Sigmoid Hybrid
9 s-t-2 Second Sigmoid – Tangent hyperbolic Hybrid
10 t-t-t-1 First Tangent hyperbolic Tangent hyperbolic Tangent hyperbolic Homogeneous
11 t-t-s-1 First Tangent hyperbolic Tangent hyperbolic Sigmoid Hybrid
12 t-s-s-1 First Tangent hyperbolic Sigmoid Sigmoid Hybrid
13 s-s-s-1 First Sigmoid Sigmoid Sigmoid Homogeneous
14 t-t-t-2 Second Tangent hyperbolic Tangent hyperbolic Tangent hyperbolic Homogeneous
15 t-t-s-2 Second Tangent hyperbolic Tangent hyperbolic Sigmoid Hybrid
16 t-s-s-2 Second Tangent hyperbolic Sigmoid Sigmoid Hybrid
17 s-s-s-2 Second Sigmoid Sigmoid Sigmoid Homogeneous
18 l-l-l-2 Second Linear Linear Linear Homogeneous
19 s-s-1 First Sigmoid – Sigmoid Homogeneous
20 s-s-2 Second Sigmoid – Sigmoid Homogeneous
21 s-s-t-1 First Sigmoid Sigmoid Tangent hyperbolic Hybrid
22 s-s-t-2 Second Sigmoid Sigmoid Tangent hyperbolic Hybrid

selection of a proper structure is difficult and the inherent order of

Fig. 10. Artificial neural network architecture.
by considering enough neurons in the hidden layers. However,
too many neurons in the hidden layer may cause over fitting.
Because of over fitting, network yields larger error for the new
data even if the error is very small for training data. Also, larger
the network, more complex it is. There are several methods for
improving the network to provide an adequate fit. Maghami and
Sparks [67] have proposed ANN method for rapid and efficient
controls, dynamic design and analysis for flexible space systems.
The proposed method of Maghami and Sparks [67] is suitable for
applications wherein an arbitrary large source of training data can
be generated. However, in the present investigation, the data has
been generated by performing a set of experiments on a real sys-
tem. As the limited data set in the present investigation is available,
the method proposed by Maghami and Sparks can be tested in
this application in a future work, once a huge set of data is gener-
ated. The authors of [68] have also presented an effective method
for modular neural network. The relationship in our data is eas-
ily trainable in the form of the proposed networks as shown in
the results with various structures of the networks. However, the
proposed method can also be tested while using modular neural
network for this application as a future work. Wilson and Tuffs [69]
have suggested an efficient algorithm for the applications where
nonlinearity in the dataset is high and therefore requires a large
structure. Moreover, if the dimensionality and number of sam-
ples required for efficient training are also high, the trial-and-error
method requires significant wastage of computational efforts. In
our application, trial-and-error method is suitable as the training
time and flops required for each structure are not very high. There-
fore, the network structure and their operational parameters in
the present investigation have been selected mostly by using trial-
and-error method. The experimentation has been performed in a
sufficiently wide range. The comparison among various structures
has been made and presented in the paper in the form of tables
and graphs. Fig. 11 compares the predicted crack initiation angle
with the experimental values (normalized) for training data set,
validation data set and testing data set. In Fig. 11 out of 22 ANN
architecture used in this investigation, the detailed results of [10-
9-1] ANN architecture with learning rate 0.8 and sigmoid activation
function have been used. The training was stopped after 200 train-
ing cycles. However, in Fig. 11(a) the variation of error is shown up
to 500 training cycles to show its effect on the MSE. The correlation
coefficient obtained for validation and testing data set are 0.9687
and 0.9721 respectively. Similar runs are made for other ANN archi-
tectures shown in Table 4 and results are presented and analyzed
in subsequent sections.
6.1. Selection of inputs
For crack growth analysis with multiple cracks, there is a wide
range of crack position parameters that could be used as inputs.
These parameters include crack size, crack tip distances, crack offset
distance, crack inclination with the loading axis, material proper-
ties such as yield strength and modulus of elasticity and specimen
geometry such as crack length to plate width ratio or crack length
to plate thickness ratio etc. Properly selected inputs can be of
great help to reach a satisfactory predictive quantity with a rel-
atively small training data set. Table 5 shows the root mean square
error (RMSE) and correlation coefficient obtained for each set of
input both for training, validation and testing for the combina-
tion number 19 (s-s-2). It is very clear from the table that all six
input parameters (left and right crack size, a1 and a2 , vertical off-
set distance (H), crack tip distance (S), and crack inclination angles
˛1 and ˛2 are implicitly or explicitly needed to obtain acceptable
524 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528

Fig. 11. (a) Variation of MSE with number of Epochs, (b) comparison of normalized training data set, (c) comparison of validation data set, (d) comparison of testing data set.
[10-9-1] network used with second order input variables and sigmoid activation function.

Table 5
Effect of number of input parameters on network architecture.

Inputs Number of input variables Output Training RMSE Testing RMSE Training correlation coefficient Testing correlation
coefficient

a1 , a2 , H, S, ˛1 and ˛2 6 0 0.015 0.048 0.9207 0.9511

a1 /a2 , S/H, ˛1 /˛2 3 0 0.073 0.247 0.7611 −0.913
a1 /a2 , S/H 2 0 0.094 0.203 0.6841 0.0599
S/H 1 0 0.118 0.241 0.5741 −0.2015

crack initiation direction. When these parameters are used in non-
dimensional form (a1 /a2 , S/H, ˛1 /˛2 ), the RMS error is quite high
and is about 4 times than that with six inputs taken as a1 , a2 , H, S,
˛1 and ˛2 . The networks used had one hidden layer with 9 neurons
(combination no 19, s-s-2). Learning rate parameter was taken as
0.9). All data were normalized between 0 and 1. A probable justi-
fication could be that, that for zero value of the parameter which
is in numerator, the ratio is taken as zero irrespective of the mag-
nitude of denominator. The other reason may be that only three
values were supplied when ratio in non-dimensional form was used
instead of six values for each output. Hence, in this investigation dif-
ferent order of input variables (xin , i = 1, 2, . . ., 6 and n = 1 or 2 or
3) have been used and the predictive quality is compared with first
order, second order and third order input parameters. The results
show that inclusion of second order input variables improve the
predictive quality by 20–30% as compared to first order, but third
or higher order input data sets yield poor results.
6.2. Number of training cycle
Number of training cycle plays an important role for over
fitting. Early stopping or over training may cause higher error.
The stopping criterion for the training process is based upon
the root-mean-square-error (RMSE) for the validation data set
which was entirely different from training data set (used for
computing the gradients, Jacobian Matrix, and Hessian Matrix
during optimization process for weight refinement) as well as
testing dataset (used to test the trained network after com-
pletion of the training process). However, the results for the
validation data set were not depicted graphically in order to
avoid ambiguity as this data set has the only role of stopping
the training process at appropriate time and thus reducing the
chances of memorization. However, the various forms of results
for the validation data set are shown, wherever feasible and felt
necessary.
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528 525

0.35 0.5 100

19 (training) 19 (testing)
RMSE
0.3 20 (training) 20 (testing)
0.4 PRED (%) 80
13 (training) 13 (testing)
0.25
10 (training) 10 (testing)
0.3 60

PRED (%)
0.2
RMSE

RMSE
0.15 0.2 40

0.1
0.1 20
0.05
0 0
0
0 2 4 6 8 10 12 14 16 18 20 22
0 250 500 750 1000 1250 1500 1750 2000 2250 2500
Epochs Combination number

Fig. 12. Dependence of root mean square error (RMSE) on numbers of iterations for Fig. 14. Variation of root mean square error (RMSE) and PRED (%) with number
different combinations of activation function. of neurons in first and second hidden layers for 0.8 learning rate and 200 epochs.
Optimum number of neurons used are either [10-8-1], [8-10-1] or [8-1].

0.2 1

6.4. Optimization of hidden layer and number of neurons

0.9
0.15
As per universal function approximate theorem, if less number
Correlation coefficient (r)

RMSE (Training) RMSE ( Testing)

of neurons are required in the hidden layer for better approx-
0.8
imation then only one hidden layer can be used. Multi-layer
RMSE

0.1 RMSE(Validation) r (Training)

r ( Testing) r(Validation)
0.05
0 0.5
0 0.2 0.4 0.6 0.8
Learning rate
Fig. 13. Variation of root means square error (RMSE) and correlation coefficient
with learning rate for eight numbers of neurons in a hidden layer for training data
set, validation data set and testing dataset (s-s-1 model).
Fig. 12 shows the variation of root-mean-square-error (RMSE)
with number of iterations for the ANN architecture with one and
two hidden layers. It can be seen from the results that most of the
ANN architecture with 100–200 training cycles yield reasonable
accuracies and about 70–80% of the predicted results are within 5%
error. However, with higher training cycles, in some cases the RMSE
is found to be lower, but the predictive quality measured with PRED
(ε) is found to be less than 70%. The prediction capability (PRED) is
calculated from the relative error. It is defined as the ratio of data
points with error less than or equal to ε to the total number of data
points. ε used in this investigation is 5%.
6.3. Learning rate
The quality of prediction also depends on the learning rate
parameter, as it influences the output. Higher learning rate parame-
ter indicates an ANN with better output approximation capabilities.
Learning rate parameter has been selected for each structure by
starting with a low value and increasing it in small steps till the
speed of convergence is increased without entering the process
into an oscillatory mode. Effect of momentum parameter was not
observed in a significant manner as there might be no local min-
ima in most of the situations. The variation of RMSE and correlation
coefficient with learning rate coefficient corresponding to 200 iter-
ations, one hidden layer (s-s-1), and nine neurons in hidden layer is
shown in Fig. 13. Figure shows that with learning rate for 0.6–0.8,
high predictive quality can be achieved with RMSE less than 0.015.
feed-forward neural networks are believed to be capable of rep-
0.7 resenting any functional relationship between input and output, if
there are enough neurons in the hidden layers. However, too many
0.6
neurons in the hidden layers may cause over-fitting. The error on
the training set is driven to a very small value due to the powerful
ANN learning process, but when new data is presented to the net-
work, the error is large. The network has memorized the training
1 1.2
examples, but it has not learned to generalize to new situations.
Clearly, when over-fitting occurs, the network will not general-
ize well. The ideal method for improving network generalization
is to use a network that is just large enough to provide an ade-
quate fit. The larger a network is, more complex are the functions
the network can represent. In the present investigation one and
two hidden layered ANN structure were evaluated. For the one
hidden layered network structure, combination of eleven differ-
ent activation functions were tested, and the performance of each
network for validation and test data were evaluated on the basis
of root-mean square-error, correlation coefficient, and prediction
capability. It is seen that though the correlation coefficient and
RMSE are reasonably good for first and second order input data but
less than only 50% of the test data were predicted within 5% error
for first order and second order input variables with single hidden
layer. For the two hidden layer network structures, eleven combi-
nations of different activation functions were trained and tested. On
the basis of their test performance indices, more than 70% data were
predicted within an error of 5% by all eleven-combination models
with two hidden layers except from linear activation function. This
concludes that, in case of single hidden layered structure, second
order input variables yield better network performance, whereas,
for double hidden layers ANN architecture, use of first or second
order input variables are equally good.
The performance of all 22 networks were checked by changing
the neurons in the hidden layer for single hidden layered structure
and number of neurons in first and second hidden layer for two
hidden layered architecture. In order to decide the optimum hid-
den neurons, different ANN architecture were trained and tested for
number of neurons ranging from 6 to 14. On the basis of high gener-
alization capability and best prediction accuracy, in testing, Fig. 14
is drawn. It is found that most of the two hidden layered combina-
tions yield better results for [8-10-1] or [10-8-1]. The number in the
square bracket shows the number of hidden neurons in the hidden
526 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528

1 1
t-t-t-1

Predicted, ANN ( Normalized)

0.9 t-t-t-2
0.8
Correlation coefficient

t-t-1
t-t-2
0.8
s-s-s-1
0.6
s-s-s-2
0.7 s-s-1
0.4 s-s-2
0.6

0.5 Training 0.2

Testing
0.4 0
0 2 4 6 8 10 12 14 16 18 20 22 0 0.2 0.4 0.6 0.8 1
Combination number
Experimental ( Normalized)

Fig. 15. Performance of different networks.

Fig. 16. Comparison between predicted and experimental crack initiation angle  0
(testing data set) for different combination of activation functions.
layers and output layer. Best performance was seen from single
hidden layered network with [8-1] hidden neurons. The stability Table 6
of all 22 networks was analyzed by repeating the same network Statistical parameters of predicted results.
ten times with randomly initialized weight and biased parameters.
ANN code Predicted life ratio MLR SD
From the best results among the ten runs of each network, Fig. 15
shows the performance results of all 22 networks for [10-8-1] and Max. Min.
[8-1] neurons in hidden layers with 200 epochs. t-t-t-1 0.8762 1.3979 1.018 0.0969
t-t-t-2 0.7645 1.5074 1.065 0.2042
t-t-1 0.7645 1.5074 1.065 0.2042
7. Results and discussion
t-t-2 0.8421 1.4161 1.022 0.1166
s-s-s-1 1.0 1.40 1.027 0.0933
The purpose of the present study is to investigate applicabil- s-s-s-2 0.9204 1.40 1.025 0.0983
ity of ANN in predicting the crack initiation and growth direction s-s-1 0.6298 2.6814 1.065 0.4110
under multiple crack geometry using crack position parameters. s-s-2 0.9511 1.6290 1.079 0.1704

85 tests under different crack statistics were conducted and used

in the optimization of ANN architecture. Out of 85 data sets 50 data
sets were used for training, 14 data sets were used for validation
and remaining 21 data sets were used for testing of the optimized
ANN architecture.
As can be observed from Figs. 14 and 15, the tangent hyper-
bolic activation function and sigmoid function are superior than
linear one for the fracture mechanics based crack growth direction
application. Irrespective of number of hidden layers, both activa-
tion functions, i.e. tangent hyperbolic and sigmoid, predict the crack
initiation direction with correlation coefficient greater than 0.97.
However, in the homogeneous group, sigmoid activation function
is found to be superior than tan hyperbolic function for both type
of input data, i.e. first order as [xi ; i = 1, 2, . . ., 6] or second order [xi2 ;
i = 1, 2, . . ., 6]. In hybrid group, the predicted quality drops drasti-
cally when linear activation function is included in the analysis. In
hybrid group, it is also seen that when sigmoid activation function
in the output layer along with tangent hyperbolic in first or second
hidden layer is used (t-t-s) or (t-s), or vice versa, it gives almost
similar performance parameters and prediction results.
Fig. 16 shows the predicted crack initiation direction (normal-
ized) from the different neural network versus experimental crack
initiation angle for testing data sets.
The quality indicator of ANN architecture is its property of gen-
eralization. Fig. 16 reveals that the highest predictive quality is
obtained from the two-layered network with sigmoid activation
function and second order input data. The mean life ratio, MLR
(ratio of experimental to predicted) and associated standard devi-
ation along with minimum and maximum life ratio for testing data
set are shown in Table 6 for selected networks. The statistical signif-
icance tests were conducted between MLR obtained from different
ANN architectures. The statistical significance test for difference
of means between t-t-t-1 and t-t-t-2 shows that student t value
is 0.9299. On the basis of a two-tailed test at a 0.01 level of sig-
nificance, the tabulated t value corresponding to 99.7% confidence
level and 40 degree of freedom is ±2.70. As the computed t value
is within the tabulated range, it can be concluded on the basis of
two-tailed test at a 0.01 level of significance that there is essentially
no difference between the two MLR obtained from t-t-t-1 and t-t-
t-2. Similar conclusions were obtained from other networks with
homogeneous and hybrid activity functions except linear combi-
nation. The student’s t test result from correlation coefficient (r)
of the test data sets are found to be 16.039 against the tabulated
value of 2.86 for 99.5% confidence level and 19 degree of freedom
for s-s-s-2 network architecture. Since t > t0.995,19 , it indicates that
the correlation coefficient is significantly greater than zero at a 0.01
level. Similar tests were conducted for other combinations also. Out
of all combinations, network with pure linear activation function
fails to qualify the statistical significance test. Hence, to establish
the best combination of activation function, hidden layer and type
of input data, another performance parameter, PRED is included
in the analysis along with RMSE and correlation coefficient. PRED
data presented in Fig. 14 shows that less than 50% data are pre-
dicted within an error of 5% with single layer network and first
order input data sets. However, the percentage is improved to about
72% with second order input data. With two-layered network, the
predictive quality ranges between 80 and 90% except linear acti-
vation function. On the basis of analysis of predictive quality of all
22 networks, s-s-s-2 is found to be superior amongst all other net-
works investigated in the work and the ANN follows very close to
the target data with very high overall quality of prediction. All the
networks used in this work are universal-function-approximators.
In our experimentation also, these structures are able to map the
relationship between input and output effectively and efficiently.
Instantaneous or online training has also been performed along
with the batch learning. It has been found that these network struc-
tures are capable to learn with instantaneous or online process of
weight refinement also with the same effectiveness. It is, therefore,
evident from this numerical experimentation that the proposed
methodology is appropriate for online applications also.
 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528
8. Conclusions
The present study shows that ANN can be used to predict the
crack initiation direction from the multiple crack geometry param-
eters. The best prediction quality of neural network is observed
with sigmoid activation function with two hidden layer and sec-
ond order input parameters. A network with the tangent hyperbolic
function in the hidden layer and the sigmoid function in the output
layer or vice versa may also be one of the good choices among other
homogeneous networks. The combinations of sigmoid or tan hyper-
bolic with linear activation functions gave poor accuracy and hence,
are not suitable. For the single or two-hidden layered networks, sec-
ond order input variables shows reasonable improvement in the
predictive quality of the network as compared to first order input
variables.
References
[1] A.A. Griffith, The phenomena of rupture and flow in solids, Philos. Trans. R. Soc.
A 221 (1921) 136–163.
[2] M.A. Hussain, S.L. Pu, J. Underwood, Strain energy release rate for a crack under
combined mode I and mode II, Fract. Anal. ASTM STP 560 (1974) 2–27.
[3] C.H. Wu, Fracture under combined loads by maximum-energy-release-rate cri-
terion, J. Appl. Mech. 45 (1978) 553–558.
[4] F. Erdogan, G.C. Sih, On the crack extension in plates under plane loading and
transverse shear, J. Bas. Eng. 85 (1963) 517–519.
[5] K. Palaniswamy, W.G. Knauss, Propagation of a crack under general, in-plane
tension, Int. J. Fract. 8 (1972) 114.
[6] G.C. Sih, Some basic problems in fracture mechanics and new concepts, Eng.
Fract. Mech. 5 (1973) 312–365.
[7] F. Berto, P. Lazzarin, A review of the volume-based strain energy density
approach applied to V-notches and welded structures, Theor. Appl. Fract. Mech.
52 (2009) 183–194.
[8] G.C. Sih, A special theory of crack propagation methods of analysis and solution
of crack problems, in: G.C. Sih (Ed.), Mechanics of Fracture, vol. 1, Noondhoff
Int. Pub., Leyden, Amsterdam, 1973, pp. XXI–XLV.
[9] G.C. Sih, Discussion on some observations on Sih’s strain energy density
approach for fracture prediction, Int. J. Fract. 10 (1974) 279–284.
[10] G.C. Sih, B.M. Barthelemy, Mixed mode fatigue crack growth predictions, Eng.
Fract. Mech. 13 (1980) 439–451.
[11] G. Labeas, T. Kermanidis, Stress multiaxiality factor for crack growth prediction
using the strain energy density theory, Theor. Appl. Fract. Mech. 45 (2006)
100–107.
[12] J.Z. Zuo, A.T. Kermanidis, S.G. Pantelakis, Strain energy density prediction of
fatigue crack growth from hole of aging aircraft structures, Theor. Appl. Fract.
Mech. 38 (2002) 37–51.
[13] L. Nobile, C. Carloni, M. Nobile, Strain energy density prediction of crack initi-
ation and direction in cracked T-beams and pipes, Theor. Appl. Fract. Mech. 41
(2004) 137–145.
[14] V. Balasubramanian, B. Guha, Fatigue life prediction of welded cruciform joints
using strain energy density factor approach, Theor. Appl. Fract. Mech. 34 (2000)
85–92.
[15] J.K. Kim, S.B. Cho, Effect of second non-singular term of mode I near the tip of
a V notched crack, Fatigue Fract. Eng. Mater. Struct. 2 (2009) 346–356.
[16] M.R. Ayatollahi, M. Nejati, Determination of NSIFs and coefficients of higher
order terms for sharp notches using finite element method, Int. J. Mech. Sci. 53
(2011) 164–177.
[17] M.R. Ayatollahi, M. Dehghany, On T-stresses near V-notches, Int. J. Fract. 165
(2010) 121–126.
[18] M.R. Ayatollahi, M. Nejati, Experimental evaluation of stress field around the
sharp notches using photoelasticity, Mater. Des. 32 (2011) 561–569.
[19] M.R. Ayatollahi, K. Sedighiani, Mode I fracture initiation in limestone by strain
energy density criterion, Theor. Appl. Fract. Mech. 57 (2012) 14–18.
[20] C.P. Spyropoulos, Crack initiation direction from interface of bonded dissimilar
media, Theor. Appl. Fract. Mech. 39 (2003) 99–105.
[21] M. Kamaya, N. Totsuka, Influence of interaction between multiple cracks on
stress corrosion crack propagation, Corros. Sci. 44 (2002) 2333–2352.
[22] A.P. Cisilino, M.H. Aliabadi, Three dimensional BEM analysis for fatigue crack
growth in welded components, Int. J. Press. Vessels Pip. 70 (1997) 135–144.
[23] C. Wessel, A. Cislino, O. Santi, J. Otegui, M. Chapetti, Numerical and experimental
determination of three-dimensional multiple crack growth in fatigue, Theor.
Appl. Fract. Mech. 35 (2001) 47–58.
[24] J.K. Burton, S.L. Phoenix, Superposition method for calculating singular stress
fields at kinks, branches and tips in multiple crack arrays, Int. J. Fract. 102 (2000)
99–139.
[25] S.C. TerMaath, A Two-dimensional Analytical Technique for Studying Fracture
in Brittle Materials Containing Interacting Kinked and Branched Cracks (Ph.D.
dissertation), Cornell University, 2000.
[26] X. Yan, Multiple crack fatigue growth modeling by displacement discontinuity
method with crack-tip elements, Appl. Math. Modell. 30 (2006) 489–508.
527
[27] M. Kamaya, A crack growth evaluation method for interacting multiple cracks,
JSME Int. J. Ser. A. 46 (2003) 15–23.
[28] Masayuki Kamaya, Growth evaluation of multiple interacting surface cracks.
Part I: Experiments and simulation of coalesced crack, Eng. Fract. Mech. 75
(2008) 1336–1349.
[29] M. Kamaya, Growth evaluation of multiple interacting surface cracks. Part II:
Growth evaluation of parallel cracks, Eng. Fract. Mech. 75 (2008) 1350–1366.
[30] Q. Liu, C.H. Wang, S.A. Barter, P.K. Sharp, Study of interaction and coalescence
of multiple small surface cracks in a high strength aluminium alloy, in: 11th
International Conference on Fracture, Turin, Italy, 2005.
[31] R. Jones, S.N. Atluri, S. Pitt, J.F. Williams, Developments in the analysis of inter-
acting cracks, J. Eng. Fail. Assess. 2 (1995) 307–320.
[32] J. Schijve, Multi-site damage in aircraft fuselage structures, Fatigue Fract. Eng.
Mater. Struct. 18 (1995) 329–344.
[33] C.F. Tiffany, J.P. Gallagher, C.A. Babish, Threats to aircraft structural safety,
including a compendium of selected structural accidents/incidents, in: Aero-
nautical Systems Center (ASC/EN), Engineering Directorate, Wright–Patterson
Air Force Base, Air Force Material Command, United States Air Force, OH, 2010,
pp. 167.
[34] S. Pitt, R. Jones, Multiple-site and widespread fatigue damage in aging aircraft,
Eng. Fail. Anal. 4 (1997) 237–257.
[35] K. Iida, Shapes and coalescence of surface fatigue cracks, in: ICF Interna-
tional Symposium on Fracture Mechanics, Beijing, China, 1983, pp. 679–
693.
[36] K. Iida, M. Kawahara, Propagation and coalescence of fatigue cracks initi-
ated from collinear or parallel adjacent surface flaws, in: Fourth International
Conference on Pressure Vessel Technology, 1st ed., Mechanical Engineering
Publications for the Institution of Mechanical Engineers, London, 1978, pp.
257–263.
[37] X. Wang, M. Modarres, P. Hoffman, Analysis of crack interactions at adjacent
holes and onset of multi-site fatigue damage in aging airframes, Int. J. Fract.
156 (2009) 155–163.
[38] P.R. Frise, R. Bell, Fatigue crack growth and coalescence at notches, Int. J. Fatigue
14 (1992) 51–56.
[39] T.H. Leek, I.C. Howard, An examination of methods of assessing interacting
surface cracks by comparison with experimental data, Int. J. Press. Vessels Pip.
68 (1996) 181–201.
[40] M. Kamaya, Growth evaluation of multiple interacting surface cracks. Part I:
experiments and simulation of coalesced crack, Eng. Fract. Mech. 75 (2008)
1336–1349.
[41] J. Helsing, Fast and accurate numerical solution to an elastostatic problem
involving ten thousand randomly oriented cracks, Int. J. Fract. 100 (2000)
321–327.
[42] J. Helsing, A. Jonsson, A seventh order accurate and stable algorithm for the
computation of stress inside cracked rectangular domains, J. Multiscale Com-
put. Eng. 2 (2004) 47–68.
[43] M. Kachanov, On the problems of crack interactions and crack coalescence, Int.
J. Fract. 120 (2004) 537–543.
[44] H. Al-Nashash, Y. Al-Assaf, B. Lvov, W. Mansoor, Laser speckle for materials clas-
sification utilizing wavelets and neural networks image processing techniques,
J. Mater. Eval. 59 (2001) 1072–1078.
[45] F. Aymerich, M. Serra, Prediction of fatigue strength of composite laminates by
means of neural networks, Key Eng. Mater. 144 (1998) 231–240.
[46] Y. Cheng, W.L. Huang, C.Y. Zhou, Artificial neural network technology for the
data processing of on-line corrosion fatigue crack growth monitoring, Int. J.
Press. Vessels Pip. 76 (1999) 113–116.
[47] S. Dinda, D. Kujawski, Correlation and prediction of fatigue crack growth for
different R-ratios using Kmax and K+ parameters, Eng. Fract. Mech. 71 (2004)
1779–1790.
[48] R.M.V. Pidaparti, M.J. Palakal, Neural network approach to fatigue-crack-
growth predictions under aircraft spectrum loadings, J. Aircr. 32 (4) (1995)
825–831.
[49] K. Sadananda, A.K. Vasudevan, R.L. Holtz, E.U. Lee, Analysis of overload effects
and related phenomenon, Int. J. Fatigue 21 (1999) 233–246.
[50] P. Artymaik, L. Bukowski, J. Feliks, S. Narberhaus, H. Zenner, Determination of
S–N curves with the application of artificial neural networks, Fatigue Fract. Eng.
Mater. Struct. 22 (1999) 723–728.
[51] K. Genel, Application of artificial neural network for predicting strain-life
fatigue properties of steels on the basis of tensile tests, Int. J. Fatigue 26 (2004)
1027–1035.
[52] V. Venkatesh, H.J. Rack, A neural network approach to elevated temperature
creep-fatigue life prediction, Int. J. Fatigue 21 (1999) 225–234.
[53] T.T. Pleune, O.K. Chora, Using artificial neural networks to predict the fatigue
life of carbon and low-alloy steels, Nucl. Eng. Des. 197 (2000) 1–12.
[54] J.C.F. Pujol, J.M.A. Pinto, A neural network approach to fatigue life prediction,
Int. J. Fatigue 33 (3) (2011) 313–322.
[55] T. Bucar, M. Nagode, M. Fajdiga, A neural network approach to describing the
scatter of S–N curves, Int. J. Fatigue 28 (2006) 311–323.
[56] Z. Zhang, K. Friedrich, Artificial neural network s applied to polymer compos-
ites: a review, Compos. Sci. Technol. 63 (2003) 2029–2044.
[57] J.R. Mohanty, B.B. Verma, D.R.K. Parhi, P.K. Ray, Application of artificial neural
network for predicting fatigue crack propagation life of aluminum alloys, Arch.
Comput. Mater. Sci. Surf. Eng. 1 (3) (2009) 133–138.
[58] K. Zarrabi, K.H. Tsang, An elastic mixed-modes I and II fracture criterion using
an artificial neural network database, in: 5th Australasian Cong. on Appl. Mech.,
ACAM 2007, 10–12 December 2007, Brisbane, Australia, 2007.
528 D. Gope et al. / Applied Soft Computing 30 (2015) 514–528
[59] M. Kamaya, E. Miyokawa, M. Kikuchi, Growth prediction of two interact-
ing surface cracks of dissimilar sizes, Eng. Fract. Mech. 77 (2010) 3120–
3131.
[60] J.T. Tan, B.K. Chen, A new method for modelling the coalescence and growth
of two coplanar short cracks of varying lengths in AA7050-T7451 aluminium
alloy, Int. J. Fatigue 49 (2013) 73–80.
[61] X. Yan, C. Miao, Interaction of multiple cracks in a rectangular plate, Appl. Math.
Modell. 36 (2012) 5727–5740.
[62] P.C. Gope, T. Aruna, Experimental investigation of crack growth direction and
crack growth rate in multiple cracks, Fatigue Fract. Eng. Mater. Struct. 34 (2011)
804–815.
[63] M.L. Williams, On the stress distribution at the base of a stationary crack, J.
Appl. Mech. 24 (1957) 109–114.
[64] D. Gope, P. Chandra Gope, A. Thakur, Influence of crack tip and crack offset
distance on crack interaction and growth direction in multiple cracks, Int. J.
Struct. Integr. 4 (2013) 321–348.
[65] L.P. Pook, Stress intensity factor expressions for regular crack arrays in
pressurised cylinders, Fatigue Fract. Eng. Mater. Struct. 13 (1990) 135–
143.
[66] Y.Z. Wang, J.D. Atkinson, R. Akid, R.N. Parkins, Crack interaction, coalescence
and mixed mode fracture mechanics, Fatigue Fract. Eng. Mater. Struct. 19 (1996)
427–439.
[67] P.G. Maghami, D.W. Sparks Jr., Design of neural networks for fast convergence
and accuracy: dynamics and control, IEEE Trans. Neural Netw. 11 (1) (2000)
113–123.
[68] L. Yong, X. Yao, Evolving modular neural networks which generalize well, in:
IEEE Int. Conf. Evolutionary Computation, 1997, pp. 605–610.
[69] E. Wilson, D.W. Tuffs, Multilayer perceptron design algorithm, in: Proc. of
the 1994 IEEE Workshop on Neural Networks for Signal Processing, 1994, pp.
61–68.
Deepayan Gope is a graduate scholar in Electrical Engineering at Veer Surendra Sai
University of Technology (VSSUT), Burla, Odisha, India.
Prakash Chandra Gope received the B.Sc. Eng. degree in Mechanical Engineering
from Sambalpur University, Odisha, India, and the M.Sc. Eng. and Ph.D. degree in
Machine Design and Mechanical Engineering from National Institute of Technology,
Jamshedpur. He is currently a Professor of Mechanical Engineering at College of
Technology, G. B. Pant University of Agriculture and Technology, Pantnagar, India.
Since his joining to G. B. Pant University, he has supervised the research of over
20 graduate students and 5 Doctoral students. He has published over 83 research
papers. He is also the recipient of Sir Rajender Nath Mookherjee award for his best
paper published in Journal of Institution of Engineers, India in 2004.
Aruna Thakur received the B.E. degree from U.P. Technical University, Lucknow,
India and M. Tech degree from G. B. Pant University of Agricultura and Technology,
Pantnagar, India in 2009. She worked as Assistant Professor in the Department of
Mechanical Engineering Chouksey Engineering College Bilaspur, Chatis Garh India
for period of one year in 2011. She is currently a Ph.D. scholar in National Institute
of Technology, Rourkela, Odisha, India.
Abhishek Yadav received his Master degree in Electrical Engineering from Indian
Institute of Technology, Kanpur. Currently he is with Electrical Engineering Depart-
ment of College of Technology, Pantnagar, India as Assistant Professor. His current
research areas are artificial network and power system and controls.