Composite Structures 297 (2022) 115979

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Development of aviation industry-oriented methodology for failure

predictions of brittle bonded joints using probabilistic machine learning
Yuval Freed a, b, *, Navid Zobeiry b, Marco Salviato b, c
a
Aviation Group, Israel Aerospace Industries, Ben-Gurion International Airport, Israel
b
Materials Science & Engineering Department, University of Washington, Seattle, USA
c
Aeronautics & Astronautics Department, University of Washington, Seattle, USA

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

Keywords: The bonding assembly concept of cured aerospace composite parts is considered an efficient approach from
Adhesively Bonded Joints almost every perspective. It simplifies the design and provides great opportunities for weight and cost reductions.
Composite Materials However, even with strict quality assurance, such assembly may lead to undetectable low bond-line strengths,
Virtual Crack Closure Technique
usually referred to as “kissing bonds”. As part of a certification effort of bonded composites, the maximum
Machine Learning
Gaussian Process Regression
allowed disbond size has to be determined. In this study, an approach for determination of the residual strength
Mixed Mode Bending of bonded joints is proposed. Virtual crack closure failure parameters of the Loctite EA 9394 paste adhesive are
Failure Predictions determined based on a comprehensive test campaign. Predictions from more than 5,000 finite element simu­
Design Allowables lations were analyzed using a machine learning strategy. The accuracy of the optimal failure parameters was
assessed, and the failure parameters were statistically adjusted to account for predictions variability. The validity
of the proposed approach and the corresponding failure parameters were examined using two classic bonded
joints design concepts, bonded scarf joint and bonded T-joint.

1. Introduction used as a benchmark. In the ACCA (or the X-55 aircraft), advanced
composite materials were used in the fuselage and vertical tail to replace
Composite materials are increasingly used in the aviation industry the original DO-328 J metallic design (approximately 40% of the total
over the last half century, given their high specific stiffness and strength air-vehicle parts were replaced). As discussed in Ref. [7], 90% reduction
ratios. Using composite materials for primary structure introduces in fuselage part count and 80% reduction in the tail part count were
several engineering challenges, such as real time monitoring of fabri­ reported, leading to a significant recurring cost reduction.
cation defects, quality assurance management, sensitivity to environ­ The main drawback of using paste adhesives as an assembly process
mental conditions and impact damages, efficient design, suitable tooling is that bonded joints still require a very strict processing control to
and more [1–5]. From the assembly point of view, attachment of cured ensure sufficient quality for the specific materials and processes used for
composite parts using fasteners requires careful design that accounts for a given structure. The main concerns are issues such as improper surface
various aspects such as layup orientation for suitable fastener bearing preparation, bond-line contamination, high humidity, and insufficient
strength, local pad-ups to avoid knife-edge scenarios, and sufficient control of the cure cycle. These may lead to undetectable low bond-line
access for drilling and fasteners installation, whether performed strengths, usually refer to as “kissing bonds”. This risk was highlighted
manually or automatically. Alternatively, bonding cured composite in 2005 when an Airbus A310 aircraft lost its rudder (made of co-bonded
parts (even large critical structures such as pressure hulls and wing box sandwich structure) during flight due to weak bond-line between the
structures) have the potential to be very effective from almost every substructure and its outer skins [8]. Another relevant accident is a case
perspective. Such assembly process features a low weight design and of a 7 feet separation of a wing skin to main spar bond-line in a Cessna
may lead to a dramatic reduction in assembly effort and related costs Corvalis aircraft during flight test in December 2010. The FAA investi­
compared to bolted joints. Lockheed Martin demonstrated this a few gation revealed quality assurance failures in the manufacture of the
years ago during the development of the Advanced Composite Cargo wings of the damaged airplane [9]. As of yet, no reliable Non Destructive
Aircraft (ACCA) [6], in which a Dornier 328 J (DO-328 J) aircraft was Inspection (NDI) technique is available to adequately ensure that a

* Corresponding author.
E-mail address: [email protected] (Y. Freed).

https://doi.org/10.1016/j.compstruct.2022.115979
Received 3 March 2022; Received in revised form 13 June 2022; Accepted 6 July 2022
Available online 10 July 2022
0263-8223/© 2022 Published by Elsevier Ltd.
Y. Freed et al. Composite Structures 297 (2022) 115979

bonded assembly has retained its full strength [10–11], although some to 2.0 mm, with only minor reduction of the shear strength with respect
progress was reported recently [12]. Therefore, the current certification to the adhesive thickness.
standards [13–16] require the investigation of the maximum disbonds of As can be understood from the studies cited above, the correlation
each bonded joint. This should be consistent with its capability to between the adhesive thickness and its strength (or fracture mechanics
withstand expected limit loads. The determination of maximum disbond properties) are not consistent. It appears that such correlation strongly
size must be conducted by analysis, tests, or both. Disbonds that are depends upon whether the adhesive is brittle in the conditions the test
larger than the allowed size must be prevented via design features such was performed. For brittle adhesives, the fracture process zone is small
as rivets. with respect to the specimens’ dimensions; hence, relatively small en­
Considering the foregoing issues, the need for further research on ergy is dissipated due to micro-cracks at the vicinity of the disbond tip,
adhesively bonded composite joints is widely acknowledged as an and the dependence upon the bond-line thickness is weak. The chosen
essential step towards future certification of bonded structures in the test protocol is also an important parameter governing the bond-line
aviation industry. There are numerous studies that investigate both film thickness dependence. For single lap joints, for example, where sec­
and paste adhesive bond-lines characteristics (note that paste adhesive is ondary bending is inherently induced upon loading, as the bond-line
mainly used for co-bonded composite parts). Balzani et al. [17] and thickness is larger, the adhesive shear strength is reduced. For other
Floros et al. [18] for example, characterized the fracture behavior of testing protocols (such as scarf joints), this phenomenon is mitigated.
composite bonded joints subjected to mode I, mode II and mixed mode The objective of this study is to propose a new industry-oriented
loading conditions using mechanical testing and numerical simulations. methodology to determine failure parameters to predict the residual
While Balzani et al. [17] focused on film adhesive, Floros et al. [18] strength of adhesively bonded joints. Such approach can be adopted by
examined both film and paste adhesives (EA 9396 [19]). In these the aviation industry if the failure criterion is relatively simple, does not
studies, cohesive zone models were used to obtain energy-based failure require large amount of failure parameters to be calibrated and is
criteria. In the study of Gunawardana [20], DCB and Mixed Mode already implemented in commercial software, so failure predictions can
Bending (MMB) specimens were tested to obtain the fracture energies be made to large scale assemblies. The proposed methodology includes
for pure mode I and mode mixity conditions of bonded specimens with implementation of machine learning strategies to analyze a great
paste adhesives. The failure criterion obtained from these tests was used amount of physical test data and simulation data to calibrate the failure
to predict failures of single lap joints. Khan et al. [21] used unidirec­ parameters of the Virtual Crack Closure Technique (VCCT). The VCCT,
tional composite panels as adherents with thin EA 9394 adhesive [22] first introduced by Rybicki and Kanninen [37], has been widely studied
bonding them into a double lap joint specimen. Finite element simula­ during the past four decades for the simulation of brittle materials such
tions were used to predict failures as a function of the lay-up. These were as the adhesive investigated in this work. It is currently implemented in
identified mainly as adhesive failures. Energy based failure criterion was commercial finite element software such as Abaqus and NASTRAN.
proposed by Ducept et al.[23] for unidirectional glass/epoxy adhesively VCCT is also routinely used in the aerospace industry for failure pre­
bonded joints. In that study, the End Notched Flexure (ENF) specimen dictions of composite structures. Calibration of the VCCT method using
along with DCB and MMB specimens were used. Acoustic emission machine learning strategy allows the designer to determine the
measurements were used to identify the failure onset. McFall [29] maximum allowed disbond size for each assembly. As previously dis­
studied lap joints bond line characteristics focusing on strengthening cussed, knowledge of the maximum allowed disbond size is an important
mechanisms of multi-axial loads. Jiang et al. [33] used the modified step towards certification of bonded composites in commercial products.
Arcan’s test method and a stress-based failure envelope was proposed. Section 2 of this paper briefly presents the testing campaign con­
The agreement of the proposed model and experimental measurements ducted. To this end, unidirectional composite panels were co-bonded
was demonstrated in this study. using a Loctite EA 9394 paste adhesive. The Mixed-Mode Bending
Several studies were focused on the effect of the adhesive thickness (MMB) testing fixture was used to obtain mixed mode characteristics of
on the joint strength. Tomblin and his co-workers [24–25] tested both the bonded composite. In Section 3, the VCCT method and the Gaussian
thin and thick lap joint specimens. It was concluded that as the bond-line Process Regression that was chosen as the probabilistic machine
thickness increases, the strength properties of the joint decreases. This is learning method in this study are explained. Calibration of the VCCT
mainly contributed by a secondary bending effect of the lap-joint testing failure parameters using machine learning regression is described in
scheme that is amplified as the overall thickness of the specimen in­ Section 4. To account for the variability between the predictions and the
creases. Similar conclusions were drawn by Figueiredo et al. [26]. test data, the failure parameters are statistically reduced, to ensure
Scarborough [27] investigated a bonded scarf joint under static and sufficient reliability of the failure predictions. This process is discussed
fatigue loads. For such specimen, uniform shear distribution is reached in Section 5. The performance of the calibrated failure parameters is
along the bond-line, without introducing a significant secondary evaluated and validated with respect to classical aviation bonded joints
bending. The shear strength obtained in tests did not show a correlation such as bonded T-joints and scarf joints, as discussed in Section 6 of this
with adhesive thickness. A different approach was proposed by Carraro paper. Concluding remarks are drawn in Section 7.
et al. [28]. In their study, several testing protocols were used to inves­
tigate failures of bonded composites. It was shown, that for mode II the 2. Experimental setup
fracture energy increases with the adhesive thickness. The energy is
dissipated in the form of sub-critical cracking in the vicinity of the dis­ The test campaign presented here is focused on determination of
bond tip of a quasi-ductile adhesive, in which the fracture process zone is failure characteristics of adhesively bonded joints under mixed mode
relatively large (as opposed to brittle adhesives). The effect of adhesive loading conditions. The Mixed Mode Bending (MMB) testing protocol
thickness was also investigated by Davies et al. [30], when two [36] was used. Note that the ASTM standard [36] is valid for de­
aluminum plates were bonded using a paste adhesive. Significant re­ laminations in laminated composites, and not for bonded composites.
ductions of the tensile strength and small reduction of mixed mode and Nevertheless, it was used for the cases studied here, acknowledging that
shear strengths were reported as the thickness increases. In a larger some inaccuracies may be introduced in determining the fracture en­
scale, a bonded wing torsion box was tested by Weisberg et al. [34] using ergies as proposed in the ASTM standard. The MMB testing framework is
various adhesive thicknesses. Minor reduction of the strength with basically a combination of the Double Cantilever Beam (DCB) test used
respect to an increased bond-line thickness was reported. A similar to obtain mode I failure characteristics, and the End Notch Flexure (ENF)
conclusion was drawn in the research of Kwon et al. [35]. In this method that is used to characterize mode II failures. The mode mixity is
research, single lap joint tests were conducted to obtain the shear reached by changing the position of the MMB lever, as is schematically
strength of EA 9394. The bond line thickness was modified from 0.5 mm shown in Fig. 1.

2
Y. Freed et al. Composite Structures 297 (2022) 115979

Fig. 1. MMB test fixture. The mode mixity is controlled by changing the loading position (dimension c in this sketch).

The test specimen used in the MMB fixture is a standard DCB spec­ were tested, adding up to a total number of 105 specimens. The test
imen, as shown in Fig. 2. The specimens were assembled from unidi­ matrix is shown Table 1.
rectional composite adherents made of Hexply 913 pre-impregnated In this table, ta is the bond line thickness, tt is the overall specimen
resin system with standard modulus fibers. The adherents’ layup was thickness, a is the disbond length, d is the composite substrate’s thick­
[0◦ ]16, with ply thickness of 0.125 mm. The adherents were cured at ness, 2L is the length between the specimen’s supports, Lt is the overall
125 ◦ C (257◦ F) before bonded together using Loctite EA 9394. The specimen length and b is the specimen width. These dimensions are also
curing of the bonded specimens was conducted for five days at 25 ◦ C. A shown in Fig. 2. Out of the 105 test specimens, 88 were used for the data
vacuum bag was used to remove air voids and improve bonding. The analysis. The remainder specimens were disqualified mainly due to
specimens dimensions are 2L = 110 mm, d = 2 mm, and ta = 1, 2 and 3 failures reported at the piano hinge attachments (due to poor bonding),
mm. The initial disbond length, a, was measured for each specimen, and and not at the bond-line itself. The material properties used in this study
is generally spanned between 10 and 35 mm. The test specimens were are given in Table 2.
mounted in the MMB test rig and the mode mixity was reached by Load-displacement curves were recorded during tests. These curves
defining the lever position, c, as required. All tests were conducted in showed a very brittle behavior, manifested by severe snap-back insta­
ambient conditions, under displacement-controlled loading. In these bility (see e.g. Bažant [31], Bazant et al. [32]). This is clear by observing
tests, the disbond initiated a cohesive failure upon loading, followed by the vertical drop of the load right after the peak which agrees with the
adhesive failure at its interface with the composite adherent, as may be data reported by Khan et al. [21], where bonded double lap joint
observed in Fig. 3. Similar failure trend was reported in Khan et al. [21]. specimens were tested. This behavior is different from standard
Overall, 105 test specimens were fabricated. The specimens were delamination, in which fiber bridging and other near-tip mechanisms
divided into three groups with respect to the adhesive thicknesses (1, 2 lead to energy dissipation with stable softening region after reaching the
and 3 mm). Each group was further divided into seven different load maximum load. An example of the recorded load–displacement curves is
combinations (mode mixities). For every combination, five specimens shown in Fig. 4 for specimens with adhesive thickness of t = 1 mm. Note

Fig. 2. MMB test specimen.

3
Y. Freed et al. Composite Structures 297 (2022) 115979

Fig. 3. Test specimen mounted in the MMB test fixture under loading.

such as Abaqus, ANSYS and NASTRAN. This makes it particularly

Table 1
attractive for industrial applications. In addition, it requires a pre­
Test matrix (all dimensions are in mm).
defined definition of the crack path, which is known for bond-lines in
Group ta tt Lt 2L d B Lt-2L Number of specimens assembled components. In addition, the test data investigated in this
A 1 5 140 110 2 25 30 7 × 5 = 35 study is based on testing at room temperature. For such conditions, the
B 2 6 7 × 5 = 35 adhesive is characterized by linear, elastic isotropic performance and a
C 3 7 7 × 5 = 35
very brittle behavior. Hence the linear elastic fracture mechanics
Total 105
assumption that is the basis of the VCCT approach is valid.
This method also has limitations. It is highly sensitive to mesh size,
and an empirical interaction criterion is required to model mixed-mode
Table 2
cracking scenarios. In our study, plenty of test data is available and can
Mechanical Properties.
be used for correlation of such mixed-mode interaction law.
Composite Substrates According to the VCCT, the crack propagation is obtained using
Ply thickness [mm] 0.13 Irwin’s crack closure integral by calculating the product of the tractions
Et11[GPa] 110.00
ahead of the crack tip and the crack faces relative displacements. The
Et22[GPa] 9.51 numerical representation of Irwin’s crack closure integral for a crack of
G12 [GPa] 5.93 length a + Δa was proposed by Rybicki and Kanninen [37] as.
ν12 [GPa] 0.34
EA 9394 Paste Adhesive 1 1
GI = F22 Δu12 , GII = F21 Δu11 (1)
Tensile Modulus [MPa] 4,237 2Δa 2Δa
Shear Modulus [MPa] 1,461
where Δa is a virtual crack extension. The parameters in equation (1) are
described in Fig. 6.
that the displacement in this graph is not the disbond faces opening The main advantage of the VCCT method is its simplicity. The energy
displacement, but the test fixture loading point displacement. release rate can be easily obtained directly from the FEM, as is shown in
Fig. 6. For the current study, the interaction between mode I and II (for
3. Failure criterion and machine learning strategy two-dimensional problems) is described using NASA’s method [47], as.

The coupon test results described in the previous section were [

GT
]( )n− ≥1 (2)
simulated using the finite element method. The failure load was pre­
1
GIc + (GIIc − GIc ) GGIIT GGIIcT
dicted using the Virtual Crack Closure Technique (VCCT), with various
combinations of failure parameters used for each individual simulation. for two-dimensional problems, with GT is the sum of the energy
The optimization of such large simulation database was conducted using release rate components in mode I and II.
a machine learning strategy named the Gaussian Process Regression It is worth mentioning here again that VCCT has been shown to
(GPR). Both methods were chosen after a thorough trade-off study, that provide excellent predictions in the case of brittle materials such as the
accounts for several aspects such as complexity, accuracy, availability in high-performance adhesive investigated in this work. For other mate­
commercial software, and past experience with such methods in the rials featuring a non-negligible Fracture Process Zone (FPZ), quasi-
aviation industry. The VCCT and GPR methods are briefly described in brittle computational models should be used instead as discussed by
this Section. Bazant et al. [32].

3.1. Virtual crack closure Technique (VCCT) 3.2. Gaussian process regression (GPR)

The VCCT method was first introduced in the late 1970′ s by Rybicki The numerical dataset in this study was analyzed using a Machine
and Kanninen [37]. This method is known for its simplicity, and its Learning (ML) strategy to determine the optimal VCCT failure parame­
implementation in finite element framework was studied intensively for ters that best fit the test data. ML is increasingly used for damage cali­
both 2D and 3D models by many researchers [38–42], including proper bration of composites. While standard calibration of such failure
representations for bi-materials [43,44] and mode mixity [40,45,46]. parameters requires time-consuming multidimensional data reduction
For the current study, the VCCT was chosen for several reasons. First, processes, ML methods provide an efficient strategy for such calibra­
it is already implemented in many commercial finite element software tions, especially if the failure mechanism can be described using a

4
Y. Freed et al. Composite Structures 297 (2022) 115979

Fig. 4. Load-displacement curves for specimens with adhesive thickness of t = 1 mm and various mode mixities GII/Gtot.

Fig. 5. Methodology to obtain VCCT failure parameters.

physical model (such as VCCT); thus, limiting the failure parameters and
uncertainties. For composite materials, various successful ML models
have been developed in recent years for fracture toughness predictions
[48], compressive damage [49], assessment of required shims and gaps
during assemble of composite materials [50], detection of defects during
automated fiber placement process [51], damage detection, character­
ization and simulations [52–54] and simulation of the curing process
[55]. Gu et al. [61] employed neural network ML analyses to predict
failure of bonded composite single lap joints. Their training data was
obtained using finite element models of bonded lap joints with various
geometries and material properties, and the optimal failure predictions
as obtained using ML was validated with respect to test data.
In this study, the Gaussian Process Regression (GPR) [56] is used as a
ML strategy. The GPR is a nonparametric Bayesian approach to regres­
Fig. 6. Nodal point forces and relative displacements at the vicinity of the crack sion. While making broad assumptions on the probabilistic distribution
tip used for the VCCT method. of data, GPR offers several distinct benefits including ability to provide
uncertainty quantification on the predictions, based on available test

5
Y. Freed et al. Composite Structures 297 (2022) 115979

data. The regression performed by the GPR method strongly depends on

Unlike most popular machine learning algorithms that learn exact the chosen kernel and its parameters. The kernel is a matrix that de­
values for every parameter in a function, the Bayesian approach utilize a scribes the relations between different observation points and between
probability distribution over all possible values. The Bayesian approach predicted points and observation points. It can be reasonably assumed
specifies a prior distribution and relocates probabilities based on evi­ that if two points are very close to each other, they share similar char­
dence (i.e. “observed” data, or “test” data) using the Bayes’ theorem. acteristics, and vice versa. The kernel function should represent this
The updated distribution, called the posterior distribution, incorporates assumption. A very popular kernel is the Radial Basis Function (RBF),
information from both the prior distribution and the dataset. The pro­ that can be generally described as.
cedure of developing the posterior distribution is also called ’training’. [ ( )2 ]
Once the posterior is trained based on the test data (“observed” data), ( )
kRBF xi , xj = σ 2 exp −
xi − xj
(3)
predictions of any other point can be made based on fitting (also called 2l2
’predictive posterior distribution’). It is important to note that the GPR
algorithm provides mean values and a scatter of data, based on as­ where σ 2 is the variance, xi and xj are adjacent points. It can be easily
sumptions on Gaussian distribution of variabilities, as well as an explicit shown that the assumption discussed above is met by this kernel. The
covariance function (i.e. Kernel). Understanding the obtained scatter parameter l in eq. (3) is a length parameter that represents how fast the
may assist in decisions of which additional test point (i.e., observed RBF function reaches zero. As the length parameter increases, the RBF
data) may assist in increasing the regression accuracy. In this study, the function reaches zero faster. In other words, with large length param­
GPR algorithm included in python package scikit-learn [60] was used. eter, the correlation between adjacent observation point is weaker, and

Fig. 7. Machine learning assessment of the validity of (a) failure load with respect to initial disbond length a and lever position c, (b) failure load with respect to
initial disbond length and the mode mixity, (c) fracture energy with respect to initial disbond length a and lever position c and (d) fracture energy with respect to the
initial disbond length and the mode mixity.

6
Y. Freed et al. Composite Structures 297 (2022) 115979

the prediction function becomes smoother. The default value of the were spanned from 0.02 N/mm to 0.50 N/mm for the fracture energies,
length parameter is l = 1. and 1 – 3 for the mode mixity parameter. The fracture energies values
Finally, this study also used the WhiteKernel algorithm since the test were chosen so that they cover the energy release rate of all 88 test
data is rather noisy. For such case, the machine learning predictions are specimens as obtained using ASTM D6671. This process resulted in
divided into the standard predictions (as obtained using the RBF kernel) 5,632 different models, differ from each other by the failure parameters
and the noise (controlled by the WhiteKernel). and the individual geometrical properties of each specimen. The
( ) ( ) breakdown of the failure parameters used for analyses is given in
k xi , xj = kRBF xi , xj + σ 2e δij (4)
Table 3. The specimens were loaded using a displacement-controlled
loading in the FEM, and the load step was adjusted between 0.02 and
where σ 2e is the noise level, with 1 as its default value.
5x10-6 of the maximum load to ensure proper convergence. With such
parameters, typical execution time of each analysis was of the order of
4. Virtual crack closure technique failure parameters several minutes.
calibrations The optimal failure parameters were based on minimization of the
comparisons between test results and predictions for all combinations of
The process conducted in this study includes several steps. First, test VCCT failure parameters. Therefore, global minimization of a cost
data from 88 specimens was recorded and evaluated. Once the validity function was performed to calibrate VCCT parameters:
of the test data is verified, multiple finite element models that span the ⃒ ⃒
expected domain of the VCCT failure parameters are constructed. GPR ⃒ Predicted Failure Load⃒⃒
min ⃒⃒1 − (3)
machine learning analysis is then conducted to obtain the optimized Test Failure Load ⃒
VCCT failure parameters among more than 5,000 finite element simu­
The predicted failure load was taken at the onset of failure. Note that
lations. The scatter of the predictions is evaluated via statistical ana­
the ML analysis was conducted for multi-dimensional space that in­
lyses. Finally, the validity of the failure parameters is examined in light
cludes the mode I and mode II fracture energies, the mixed mode
of several benchmark problems. Fig. 5 outlines the methodology used
parameter n and the ratio between predictions and test results |1-pre­
here. The last two stages of this process are discussed in the following
diction/test|.
sections.
The optimal VCCT failure parameters are listed in Table 4. As
As a first step, the GPR probabilistic machine learning approach was
observed in this table and also in Fig. 9, the failure surface is generally
employed to evaluate the validity of the test data and identify trends.
ill-posed. In other words, many combinations of GIc, GIIc and n can
Fig. 7(a) and (b) present strong correlation between the failure load and
provide very similar values of |1 – prediction/test|. In general, it can be
key test parameters such as initial disbond length and mode mixity. The
argued that the failure predictions are less sensitive to GIIc and n than to
GPR fitting accuracy was 93.7% and 94.3%, respectively. On the other
GIc. It is also observed that the fracture energies are quite similar for all
hand, poor correlation is reported with respect to the fracture energy
thicknesses. For simplicity, the failure predictions of all 88 test speci­
(only 16% and 15%), as calculated per ASTM 6671D (see Fig. 7(c) and
mens were pooled together to obtain VCCT failure parameters regardless
(d)). This implies that while the test results in terms of failure loads and
of the adhesive thicknesses. This data is highly important from design
specimen characteristics are valid, the equations proposed in the ASTM
point of view, since a single set of failure parameters can be used for
standard for calculation of the fracture energies are not valid for the
structural assemblies with variable adhesive thickness. The drawback of
adhesively bonded composites. This is not surprising recalling that the
such simplification is that the variability of the failure predictions may
ASTM standard [36] is limited to delamination in composite materials
and is used here for bonded composites since no other standard is
available in hand. Table 3
Next, the VCCT failure parameters were calibrated. To this end, Breakdown of FE simulations parameters to obtain VCCT failure parameters.
Finite Element Models (FEMs) were constructed for all 88 test specimens Adhesive Number of GIc [N/ GIIc [N/ n Number of FE
using the Abaqus software. At the vicinity of the disbond tip, four-noded thickness Specimens mm] mm] Simulations
[mm]
quad-elements (CPS4R) were used, with typical element size of 0.1 mm.
The element size was determined based on a convergence study. The 1 31 0.02, 0.02, 1, 1.5, 1,984
meshes included approximately 8,000 – 15,000 elements, depending on 2 23 0.10, 0.10, 2.0, 1,472
3 34 0.25, 0.25, 3.0 2,176
the adhesive thickness. An example of a typical finite element mesh is 0.50 0.50
shown in Fig. 8. For each FEM, the failure parameters GIc, GIIc and n

Fig. 8. Typical finite element mesh used for VCCT failure parameters calibration. Total number of elements for this specimen: 8485 elements. Note that the lever is
simulated assuming a rigid body motion.

7
Y. Freed et al. Composite Structures 297 (2022) 115979

Table 4 be increased; this should be acknowledged and addressed to ensure

Summary of VCCT failure parameters. sufficient reliability of the predictions. This topic is further discussed
Adhesive thickness [mm] GIc [N/mm] GIIc [N/mm] n and demonstrated in Section 5 of this paper.
Fig. 10 presents the failure predictions versus test results for all 88
1 0.40 0.32 1.94
2 0.32 0.39 2.27 test specimens. The different colors represent the different adhesive
3 0.33 0.29 3.00 thicknesses. The x-axis denotes the mode mixity GII/Gtot, with GII/Gtot =
All Samples 0.33 0.37 3.00 0 as pure mode I, and GII/Gtot = 1 as pure mode II. The failure param­
eters obtained using GPR for all thicknesses combined were used for the
predictions (GIc = 0.33 N/mm, GIIc = 0.37 N/mm, n = 3). As can be seen
from this graph, the predictions are generally very accurate with respect

Fig. 9. 3D view of GPR analysis (mean values) for failure predictions at adhesive thickness of (a) t = 1 mm, (b) t = 2 mm and (c) t = 3 mm.

8
Y. Freed et al. Composite Structures 297 (2022) 115979

reliability is required (i.e., 99% reliability with 95% confidence level).

The B-Basis values based on the Prediction/test Weibull fitting are
presented in Table 6. From engineering point of view, once Prediction/
Test greater than 1.0, the predictions are not conservative. Therefore,
the upper B-Basis bound is the parameter that is of interest to meet the
required reliability and confidence level. Note that the B-Basis limits
should be measured with respect to the average value (these are basi­
cally the scatter bands around the average value). For example, for ad­
hesive thickness of t = 1 mm, the lower bound represents a deviation of
(1.08 – 0.70) = 0.38, which is larger than that of t = 2 mm (1.00 – 0.66
= 0.34) in which its shape factor is larger.
Once the B-Basis upper limits are obtained, a safety factor can be
defined. As previously discussed, the upper B-Basis bound represents a
Fig. 10. Comparison between predictions and test results for all 88 specimens. value of Prediction/Test that bounds at least 90% of the Prediction/Test
The predictions/test values are shown as a function of the mode mixity GII/Gtot. values, all with 95% confidence level. Understanding the role of this
value, a safety factor can be obtained as (B-Basis Upper Value / Average
to the test data with an average of Prediction/Test = 1.02, but large Value) when the average Prediction/Test is below 1.0, or simply the B-
scatter is reported (standard deviation of 15%). This scatter is addressed Basis Upper Value when the average Prediction/Test value is greater
in a statistical analysis presented in the next section. No specific trend is than 1.0. The safety factors obtained using this method are given in
observed with respect to the mode mixity or the adhesive thickness. Table 6.
An alternative approach for determination of safety factors is to
5. Statistical analysis of VCCT failure parameters evaluate failure parameters that will lead to failure predictions which
are conservative for at least 90% of the test articles population with
As presented in Fig. 10, while the predictions are generally accurate confidence level of 95%. This can be easily achieved by performance of
with respect to the test data for all 88 test specimens, large scatter is machine learning analyses but with different cost function. Instead of
associated with the usage of the VCCT failure parameters. This obser­ the standard cost function of |1.0 - prediction/test|, we can seek for the
vation is also valid for failure predictions of each individual adhesive following cost function: |1.0/(Upper B-Basis Limit) - Prediction/test|.
thickness (i.e., adhesive thickness of 1, 2 or 3 mm) or when all test re­ Such choice of ML regressing goal shifts the entire prediction/test
sults are pooled together, as presented in Table 5. To ensure sufficient domain (as presented in Fig. 11 for example) to ensure that at least 90%
reliability and proper confidence in the failure predictions, statistical of the predictions will be conservative with respect to the test results
analysis was performed. (with confidence level of 95%). Performance of such analysis provides
The values of Prediction/Test were fitted assuming a Weibull dis­ the B-Basis VCCT parameters as listed in Table 7. An example of heat and
tribution. The choice of a Weibull distribution over normal distribution contour maps of the optimized B-Basis GIc and GIIc at n = 2.10 for ad­
is guided by Brot [57] and Brot [58], in which it was clearly concluded hesive thickness of t = 2 mm is shown in Fig. 12.
that the Weibull distribution produces more conservative correlations
than the Log-Normal distribution. Note that the GPR machine learning 6. Validation
approach used in this study assumes normal distributions for the pa­
rameters used for regression. However, the statistical analysis presented Two benchmark problems were chosen for validation of the proposed
here is only used to assess the reliability of the VCCT failure parameters, failure parameters. A bonded Scarf joint is widely used in the aviation
and not related directly to the ML analysis. Fig. 11 presents an example industry, especially for design of bonded repairs. The bonded T-joint is a
of how the Prediction/Test values are fitted to Weibull distribution. very popular design for wing torsion box primary structure (see for
Table 6 presents the relevant statistical parameters obtained from the example, Weisberg et al. [34] and Dharmawan et al. [59]).
Weibull distribution curve fitting. As can be seen from this table, the Test data for a bonded Scarf joint with EA 9394 used as an adhesive is
Prediction/test values fit very well assuming Weibull distributions (R2 is presented by Scarborough [27]. To this end, the scarf angle was taken as
very close to unity for all thicknesses). In average, the predictions are not 10◦ , with 2024-T3 aluminum plates used as adherents. The adhesive
entirely conservative with respect to the test results (average Prediction/ thickness is 0.2 mm, with adherents’ thickness of 6.35 mm. The scarf
Test greater than 1.0 for most of the specimens). This can also be joint was simulated using finite element model and the failure was
observed from Fig. 11, in which most of the Prediction/Test values are predicted using the VCCT failure parameters. The finite element mesh is
above 1.0. The Weibull shape parameter is an indication of how the presented in Fig. 13. Average VCCT failure parameters were used. The
Prediction/Test values are scattered. As the scatter is reduced (with average shear strength reported by Scarborough [27] was 4,918.2 psi
respect to the average value), the shape parameter increases. (33.91 MPa). The finite element simulations predicted shear strength of
Next, sufficient reliability and confidence level should be chosen. To 4401.7 psi (30.35 MPa) using the VCCT average failure parameters. This
this end, a B-Basis value (i.e., 90% reliability with 95% confidence level) is within a reasonable (and conservative) error of 11% with respect to
was chosen. Such reliability is used in the aviation industry for multi- the test result.
load path design, which is generally considered as a good design prac­ Comparison of failure predictions and test results of a bonded T-joint
tice. If the structural detail is considered as a single load-path, A-Basis design are based on IAI test data. The tests were conducted for T-joint
bonded using EA 9396 adhesive, which is very similar to EA 9394 in its
Table 5 strength characteristics. Although the test results presented here were
Summary of predictions accuracy – VCCT predictions. adjusted with respect to the shear strengths of both adhesives, this
Adhesive thickness Number of Average Prediction/ Std. validation should be considered as qualitative and not quantitative.
[mm] Specimens Test Deviation To this end, a bonded T-joint subjected to tensile loads with different
1 31 1.08 0.17 corner radii was modeled using Abaqus software. Two corner radii were
2 23 1.00 0.15 examined, 3 and 5 mm. The relevant dimensions of the T-joints are
3 34 0.99 0.14 shown in Fig. 14(a). Based on tests conducted at IAI, for such bonded
All thicknesses 88 1.02 0.15 joint subjected to tensile load, the failure is oriented at 45◦ with respect
combined
to the loading direction, as shown in Fig. 14(b). Finite element mesh was

9
Y. Freed et al. Composite Structures 297 (2022) 115979

Fig. 11. Weibull distribution fitting of Prediction/Test values for all 88 test specimens. The predictions assumed VCCT fracture parameters of GIc = 0.33 N/mm, GIIc
= 0.37 N/mm, n = 3 obtained by pooling test results from all adhesive thicknesses.

Table 6
Weibull distribution based on predictions/test data values – virtual crack closure technique.
Adhesive thickness [mm] Number of Specimens Average Prediction/Test Std. Deviation R2 Weibull Shape Parameter B-Basis Safety Factor

Upper Lower

1 31 1.08 0.18 0.98 7.21 1.39 0.70 1.39

2 23 1.00 0.15 0.94 8.14 1.25 0.66 1.25
3 34 0.99 0.14 0.97 8.32 1.22 0.68 1.23
All thicknesses combined 88 1.02 0.15 0.96 8.00 1.25 0.76 1.25

Table 7
Average and B-Basis VCCT failure parameters.
Adhesive thickness [mm] B-Basis Upper Limit Average GIc [N/mm] B-Basis GIc [N/mm] Average GIIc [N/mm] B-Basis GIIc [N/mm] Average n B-Basis n

1 1.39 0.40 0.26 0.32 0.17 1.94 2.14

2 1.25 0.32 0.24 0.39 0.25 2.27 2.10
3 1.22 0.33 0.26 0.29 0.18 3.00 3.00
All thicknesses combined 1.25 0.33 0.18 0.37 0.26 3.00 1.00

constructed using Abaqus, and the VCCT criterion was calculated along part of the standard assembly process is a very efficient and promising
the 45◦ crack path starting from the stress concentration at the interface concept. However, even with very strict quality assurance protocols, this
between the adherent and the adhesive. Here too, average VCCT failure process may induce weak bonds that may compromise the structural
parameters were used. Fig. 15 presents a comparison between failure integrity of the joint. A major step towards certification of bonded
predictions and test results. Recalling that this validation problem composite structures is the development of a reliable and robust meth­
should be considered as qualitative, it appears that the VCCT predictions odology to obtain the maximum disbond size that is allowed in the
present a reasonable agreement with the test results (8% and 24% error structural assembly without reducing the structural integrity of the
for R = 3 and 5 mm, respectively). bonded joint below its design goal. Once such disbond size is determined
at critical joints, arresting features (i.e., rivets) can be used as part of the
7. Conclusions design to ensure that such disbond will not propagate before it is
detected during routine maintenance checks.
Using paste adhesives for co-bonding of cured composite parts as In this study, a comprehensive test data of 88 specimens subjected to

10
Y. Freed et al. Composite Structures 297 (2022) 115979

Fig. 12. An example of the GPR output for optimization of B-Basis VCCT failure parameters at n = 2.10. (a) contour map and (b) heat map. Note that the cost
function used here is |1.0/(Upper B-Basis Limit) - Prediction/test|.

Fig. 13. Finite element model of a bonded scarf joint (scarf angle is 10◦ ).

complex loading schemes was analyzed. The Virtual Crack Closure the data that was produced by the finite element analyses (5,632 sim­
Technique (VCCT) was chosen as a mixed mode failure criterion. This ulations), machine learning strategy was used. The Gaussian Process
method was chosen mainly due to its simplicity and the ability to Regression (GPR) method was chosen, due to its ability to consider noise
implement it in complex engineering problems quite quickly. The test and variabilities in high-dimensional data sets, and ease of
data was used to correlate the failure parameters of the VCCT. To handle implementation.

11
Y. Freed et al. Composite Structures 297 (2022) 115979

Fig. 14. (a) T-joint configuration used for validation. The adhesive radius R changes between 3 and 15 mm. (b) typical adhesive failure at approximately 45◦ upon
tensile load.

Fig. 15. Failure predictions vs. test results of a bonded T-joint subjected to tensile loads with respect to the corner radii.

For the range of thicknesses investigated and the brittle adhesive Acknowledgement:
investigated in this work, the dependency of the optimized failure pa­
rameters with the adhesive thickness was found to be weak, so a single The experimental data used for this research was obtained by Israel
set of failure parameters can be used to predict failures. This assumption Aerospace Industries (IAI) as part of the OPTICOMS R&D program in the
was evaluated by obtaining the prediction/test values for all 88 speci­ framework of Clean-Sky II European R&D program.
mens using a single set of failure parameters. The average values were
close to unity, but with relatively large standard deviation. To overcome References:
this difficulty, a statistical analysis was performed, correlating the pre­
diction/test ratios using Weibull distribution. Safety factors were ob­ [1] Composite Material Handbook CMH-17, CMF-17 Organization. Composite
Materials Handbook Volume-3 (CMH-17-3).
tained from this statistical analysis, and B-Basis VCCT failure parameters [2] Wang B, Zhong S, Lee TL, Fancey KS, Mi J. Non-destructive testing and evaluation
were obtained. of composite materials/structures: A state-of-the-art review. Advances in
The validity of the failure parameters was evaluated with respect to mechanical engineering. 2020 Apr;12(4):1687814020913761.
[3] Li Y, Li N, Gao J. Tooling design and microwave curing technologies for the
bonded T-joint and bonded scarf joint test data. In general, the pre­ manufacturing of fiber-reinforced polymer composites in aerospace applications.
dictions of both methods were found to be relatively accurate and Int J Adv Manuf Technol 2014 Jan;70(1):591–606.
conservative, especially when considering the simplicity of the failure [4] Towsyfyan H, Biguri A, Boardman R, Blumensath T. Successes and challenges in
non-destructive testing of aircraft composite structures. Chin J Aeronaut 2020 Mar
criterion chosen for this study. 1;33(3):771–91.
[5] Hagnell MK, Kumaraswamy S, Nyman T, Åkermo M. From aviation to automotive-
Declaration of Competing Interest a study on material selection and its implication on cost and weight efficient
structural composite and sandwich designs. Heliyon 2020;6(3):e03716.
[6] Zelinski P. Advanced composite cargo aircraft proves large structure practicality,
The authors declare that they have no known competing financial high performance. Composites 2010.
interests or personal relationships that could have appeared to influence [7] Lambert DB. Composite aircraft life cycle cost estimating model, M.Sc. Thesis, Air-
the work reported in this paper. force Institute of Technology (AFIT), AFIT-GCA-ENV-11-M02 (https://scholar.afit.
edu/etd/1535).

12
Y. Freed et al. Composite Structures 297 (2022) 115979

[8] Aviation Investigation Report A05F0047, Transportation Safety Board of Canada, [36] Astm. 6671D, Standard test method for mixed mode I-mode II interlaminar fracture
2005 (https://www.tsb.gc.ca/eng/rapports-reports/aviation/2005/a05f0047/a05f toughness of unidirectional fiber reinforced polymer matrix. Composites 2006.
0047.html). [37] Rybicki EF, Kanninen MF. A finite element calculation of stress intensity factors by
[9] FAA Press Release, FAA Proposes $2.4 Million Civil Penalty Against Cessna a modified crack closure integral. Eng Fract Mech 1977 Jan 1;9(4):931–8.
Aircraft, September 22, 2011. [38] Sethuraman R, Maiti SK. Finite element based computation of strain energy release
[10] da Silva, LFM., Öchsner, A. and Adams, RD. eds., 2011. Handbook of adhesion rate by modified crack closure integral. Eng Fract Mech 1988 Jan 1;30(2):227–31.
technology (Vol. 1, p. 1543). Heidelberg: Springer. [39] Narayana KB, Dattaguru B. Certain aspects related to computation by modified
[11] Baker A, Gunnion AJ, Wang J, Chang P. Advances in the proof test for certification crack closure integral (MCCI). Eng Fract Mech 1996 Sep 1;55(2):335–9.
of bonded repairs–Increasing the technology readiness level. Int J Adhes Adhes [40] Bittencourt TN, Barry A, Ingraffea AR. Comparison of mixed-mode stress-intensity
2016 Jan;1(64):128–41. factors obtained through displacement correlation, J-integral formulation, and
[12] Laser bond line inspection becomes reality, CompositeWorld, 2.8.2022. htt modified crack-closure integral. ASTM Spec Tech Publ 1992.
ps://www.compositesworld.com/articles/laser-bondline-inspection-becomes-rea [41] Beuth JL. Separation of crack extension modes in orthotropic delamination models.
lity. Int J Fract 1996 Dec;77(4):305–21.
[13] EASA CM-S-005. Bonded repair size limits in accordance with CS-23, CS-25, CS-27, [42] Oneida EK, van der Meulen MC, Ingraffea AR. Method for calculating G, GI, and GII
CS-29 and AMC 20-29. to simulate crack growth in 2D, multiple-material structures. Eng Fract Mech 2015
[14] EASA CM-S-010. Composite materials - the safe design and use of monocoque May;1(140):106–26.
sandwich structures in principal structural element applications. [43] Banks-Sills L, Farkash E. A note on the virtual crack closure technique for a
[15] Ashforth C, Ilcewicz L. Certification of bonded aircraft structure and repairs. NATO bimaterial interface crack. Int J Fract 2016 Oct;201(2):171–80.
STO-MP-AVT-266, 2018. https://doi. org/10.14339/STO-MP-AVT-266-06-PDF; [44] Farkash E, Banks-Sills L. Virtual crack closure technique for an interface crack
2018. between two transversely isotropic materials. Int J Fract 2017 Jun;205(2):
[16] NATO - STANAG 4671, Unmanned Aircraft Systems Airworthiness Requirements 189–202.
(USAR). [45] Krueger R. Development and application of benchmark examples for mixed-mode
[17] Balzani C, Wagner W, Wilckens D, Degenhardt R, Büsing S, Reimerdes HG. I/II quasi-static delamination propagation predictions. 2012 Apr 1.
Adhesive joints in composite laminates—a combined numerical/experimental [46] Mabson GE, De Carvalho NV, Krueger RO. VCCT with Progressive Nodal Release
estimate of critical energy release rates. Int J Adhes Adhes 2012 Jan;1(32):23–38. for Simulating Mixed-Mode Delamination: Formulation, Algorithmic
[18] Floros IS, Tserpes KI, Löbel T. Mode-I, mode-II and mixed-mode I+ II fracture Improvements and Implications. In: American Society for Composites Technical
behavior of composite bonded joints: Experimental characterization and numerical Conference 2018 Sep 24 (No. NF1676L-29357).
simulation. Compos B Eng 2015 Sep;1(78):459–68. [47] Reeder JR. 3-D Mixed Mode Delamination Fracture Criteria-An Experimentalist’s
[19] Henkel Loctite EA 9396 datasheet, https://www.henkel-adhesives.com/vn/en/pr Perspective. In: American Society for composites 21st annual technical conference
oduct/industrial-adhesives/loctite_ea_9396_aero.html. 2006 Jan 1.
[20] Gunawardana S. Prediction of failure initiation of adhesively bonded joints using [48] Zobeiry N, Reiner J, Vaziri R. Theory-guided machine learning for damage
mixed-mode fracture data. Wichita State University; 2005. characterization of composites. Compos Struct 2020 Aug;15(246):112407.
[21] Khan MA, Aglietti GS, Crocombe AD, Viquerat AD, Hamar CO. Development of [49] Reiner J, Vaziri R, Zobeiry N. Machine learning assisted characterisation and
design allowables for the design of composite bonded double-lap joints in simulation of compressive damage in composite laminates. Compos Struct 2021
aerospace applications. Int J Adhes Adhes 2018 Apr;1(82):221–32. Oct;1(273):114290.
[22] Henkel Loctite EA 9394 datasheet, https://www.henkel-adhesives.com/us/en/ [50] Manohar K, Hogan T, Buttrick J, Banerjee AG, Kutz JN, Brunton SL. Predicting
product/industrial-adhesives/loctite_ea_9394.html. shim gaps in aircraft assembly with machine learning and sparse sensing. J Manuf
[23] Ducept FD, Davies P, Gamby D. Mixed mode failure criteria for a glass/epoxy Syst 2018 Jul;1(48):87–95.
composite and an adhesively bonded composite/composite joint. Int J Adhes Adhes [51] Sacco C, Radwan AB, Anderson A, Harik R, Gregory E. Machine learning in
2000 Jan 1;20(3):233–44. composites manufacturing: a case study of automated fiber placement inspection.
[24] Tomblin JS, Yang CC, Harter P. Investigation of thick bondline adhesive joints. Compos Struct 2020 Oct;15(250):112514.
National Technical Information Service (NTIS): FInal report; 2001. [52] Sause MG, Schmitt S, Kalafat S. Failure load prediction for fiber-reinforced
[25] Tomblin JS, Seneviratne W, Escobar P, Yap YK, Harter P. Adhesive behavior in composites based on acoustic emission. Compos Sci Technol 2018 Aug;18(164):
aircraft applications. National Institute for Aviation Research (NIAR): Wichita State 24–33.
University; 2003. [53] Patel DK, Parthasarathy T, Przybyla C. Predicting the effects of microstructure on
[26] Figueiredo JC, Campilho RD, Marques EA, Machado JJ, da Silva LF. Adhesive matrix crack initiation in fiber reinforced ceramic matrix composites via machine
thickness influence on the shear fracture toughness measurements of adhesive learning. Compos Struct 2020 Mar;15(236):111702.
joints. Int J Adhes Adhes 2018 Jun;1(83):15–23. [54] Califano A, Chandarana N, Grassia L, D’Amore A, Soutis C. Damage detection in
[27] Scarborough HL. The investigation of toughness and its role in fatigue performance composites by artificial neural networks trained by using in situ distributed strains.
of adhesively bonded joints (Doctoral dissertation, Washington State University). Appl Compos Mater 2020 Oct;27(5):657–71.
[28] Carraro PA, Meneghetti G, Quaresimin M, Ricotta M. Crack propagation analysis in [55] Zobeiry N, Humfeld KD. A physics-informed machine learning approach for solving
composite bonded joints under mixed-mode (I+ II) static and fatigue loading: heat transfer equation in advanced manufacturing and engineering applications.
experimental investigation and phenomenological modelling. J Adhes Sci Technol Eng Appl Artif Intell 2021 May;1(101):104232.
2013 Jun 1;27(11):1179–96. [56] Rasmussen CE, Williams CKI. Gaussian processes for Machine Learning.
[29] McFall BD. The Effects of Multi-Axial Loading on Adhesive Joints. The Ohio State Cambridge, MA: The MIT Press; 2005.
University; 2018. [57] Brot A. Weibull or Log-Normal Distribution to Characterize Fatigue Life
[30] Davies P, Sohier L, Cognard J-Y, Bourmaud A, Choqueuse D, Rinnert E, et al. Scatter–Which Is More Suitable?. In: International Committee on Aeronautical
Influence of adhesive bond line thickness on joint strength. Int J Adhes Adhes Fatigue 2019 Jun 2 (pp. 551-561). Springer, Cham.
2009;29(7):724–36. [58] Brot A. Three Faces of Aeronautical Fatigue. In: International Committee on
[31] Bažant ZP. Snapback instability at crack ligament tearing and its implication for Aeronautical Fatigue 2017, Nagoya, Japan.
fracture micromechanics. Cem Concr Res 1987 Nov 1;17(6):951–67. [59] Dharmawan F, Li HC, Herszberg I, John S. Applicability of the crack tip element
[32] Bažant ZP, Le JL, Salviato M. Quasibrittle fracture mechanics and size effect: a first analysis for damage prediction of composite T-joints. Compos Struct 2008 Nov 1;
course. Oxford University Press; 2021 Nov 12. 86(1–3):61–8.
[33] Jiang X, Qiang X, Kolstein H, Bijlaard F. Analysis on adhesively-bonded joints of [60] Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M,
FRP-steel composite bridge under combined loading: arcan test study and Prettenhofer P, Weiss R, Dubourg V, Vanderplas J. Scikit-learn: Machine learning
numerical modeling. Polymers 2016 Jan;8(1):18. in Python. the Journal of machine Learning research. 2011 Nov 1;12:2825-30.
[34] Weissberg V, Green A, Mey-Paz H. Towards a fastenerless all composite wing. [61] Gu Z, Liu Y, Hughes DJ, Ye J, Hou X. A parametric study of adhesive bonded joints
In27th Congress of the International Council of the Aeronautical Sciences 2010, with composite material using black-box and grey-box machine learning methods:
ICAS 2010 2010 Sep (Vol. 3, pp. 2443-2452). Deep neuron networks and genetic programming. Composites Part B: Engineering.
[35] Kwon JS, Choi DG, Park JS, Lee SY. Adhesive bonded composite laminate double 2021 Jul 15;217:108894.
lap joint and progressive failure analysis, 18th european conference on composite
materials. Athens: Greece; 2018.

13