Computer-Aided Design 100 (2018) 39–51

Contents lists available at ScienceDirect

Computer-Aided Design
journal homepage: www.elsevier.com/locate/cad

Tolerance synthesis of fastened metal-composite joints based on

probabilistic and worst-case approaches✩
Ramzi Askri a, *, Christophe Bois a , Hervé Wargnier a , Nicolas Gayton b
a
Univ. Bordeaux, I2M, UMR 5295 351 Cours de la Libération, F-33400 Talence, France
b
Sigma-Clermont, Institut Pascal 4 Impasse Blaise Pascal, F-63178 Aubière, France

article info a b s t r a c t
Article history: Variabilities in geometrical and material properties occur systematically after manufacturing and joining
Received 4 November 2017 operations on different parts of an assembly. The behaviour and structural performance of compos-
Accepted 23 February 2018 ite/composite or composite/metal fastened joints are particularly sensitive to some of these variabilities.
Controlling the effect of variabilities by tolerancing uncertain parameters is therefore crucial to avoid the
Keywords:
failure of the structure. However, performing a variability study is generally costly because a large number
Fastened joints
of configurations need to be evaluated, especially when the behaviour model of the system is a numerical
Composite
Finite element one.
Uncertainties This paper presents an approach for tolerance synthesis of uncertain parameters in fastened metal-
Monte Carlo composite joints. The low time-cost of the approach is ensured by using a reduced finite element
Genetic algorithm model of the joints and a strategy to reduce the number of calculations. Both probabilistic and worst-
case approaches for the propagation of uncertainties can be applied through the proposed tolerancing
synthesis. An allowable tolerance value for an uncertain parameter can then be easily calculated by
identifying an analytical law which links tolerance to a structural performance criterion. The robustness of
the proposed approach is illustrated by its application to a 4-bolt metal-composite single-lap joint where
several sources of variability are introduced (i.e. hole-location error, pin/hole clearance, fastener preload).
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction performance and avoid failure, the effect of these uncertainties

Joining structures is an operation both to maintain two or
more parts in position and to transfer loads between them. Both
functions are made possible thanks to various joining technologies,
such as bonding, welding or fastening, with fastening being one
of the most ancient and commonly used joining techniques. This
technology allows the transfer of high loads through mechanical
connections while maintaining a removable assembly and easy
monitoring.
Sizing fastened joints is generally based on deterministic meth-
ods and consists in considering only the nominal geometric and
material properties of parts. Although this approach is widely
used by designers, it remains improvable. In fact, errors in design
parameters (i.e. gap between nominal and actual values) are sys-
tematically obtained after manufacturing. The difference between
targeted and actual properties may therefore give misleading ex-
pectations of the joint behaviour. To ensure the desired structural
✩ This paper has been recommended for acceptance by Michael Yu Wang.
*
Corresponding author.
E-mail addresses: ramzi.askri@u-bordeaux.fr (R. Askri),
christophe.bois@u-bordeaux.fr (C. Bois), herve.wargnier@u-bordeaux.fr
(H. Wargnier), nicolas.gayton@sigma-clermont.fr (N. Gayton).
is generally addressed by applying margins to an allowable load
obtained with standard tests and by requiring tight tolerances to
the process department.
The behaviour of fastened composite joints is particularly sensi-
tive to geometrical variabilities such as fastener-hole clearance and
hole-location errors [1–4]. Contrary to the classical joining process
for metal parts, the axial preload applied to composite parts is
generally low, not only for fear of damaging the material during
tightening but also because of the relaxation phenomenon, which
leads to a decrease in fastener preload with time [5]. This low
preload involves a loss in adherence during loading and therefore
leads to a load transfer mode principally through the pin/hole
contact. However, this sliding phase preceding the pin/hole contact
depends essentially on the distance between contact surfaces.
Having different clearances and hole-location errors consequently
creates an imbalance in the distribution of loads between fasteners
and eventually an overload on some fasteners. This uncontrolled
overload may be the cause of a structural failure. Controlling
these variabilities is therefore a crucial issue to ensure structural
integrity without systematically employing a complex assembly
grid combined with expensive apparatus in order to respect tight
tolerances.

https://doi.org/10.1016/j.cad.2018.02.008
0010-4485/© 2018 Elsevier Ltd. All rights reserved.
40 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
Analysing the effect of geometric and material variabilities
on joint performance requires an efficient behaviour model. This
physical model should be able to model the main phenomena
required to estimate structural performance while enabling a low
calculation time in order to perform the propagation of uncer-
tainties. Ensuring these conditions becomes complicated with the
presence of complex part geometries and non-linear phenomena
as contact. This is the main problem which makes it difficult to
produce such a variability analysis.
Most variability analyses to be found in the literature are there-
fore performed under the assumption of rigid (not deformable)
parts. In this case, designers generally try to calculate optimal
tolerances to ensure geometric constraints as mountability and
position of parts after joining [6–11]. However, when the main de-
sign criteria depend on load distribution and are managed by part
deformation, a structural analysis, including material behaviour,
should be employed. With the increase in computation capacity in
the last decades, there are now a few studies dealing with the effect
of variabilities on complex structures with the help of structural
analysis. Söderberg et al. [12] studied the effect of geometrical
variabilities on the resistance of a composite wing box during
assembly based on the Monte Carlo method. However, concerning
fastened structures, only sensitivity studies, limited to the analysis
of a few parameter values, can be found. For example, McCarthy
et al. and Gray et al. [13–15] have examined the effect of bolt-
hole clearance and bolt-torque on load distribution and failure
in single- and multi-bolt composite joints by performing ten or
so simulations. However, to obtain statistical quantities, such as
the failure probability associated with the tolerances of uncertain
parameters, hundreds of configurations should be simulated.
In order to ensure the feasibility of such an analysis of uncer-
tainties in fastened joints, Askri et al. [16] recently proposed a
reduced fastener model based on Multi-Connected Rigid Surfaces
(MCRS). This model is able to take relevant phenomena into ac-
count while reducing the calculation time by about 80% compared
to a 3-D model made of solid finite elements.
This paper therefore aims at evaluating the ability of this model
to perform the propagation of uncertainties required for the de-
velopment of a tolerance synthesis. The use of different tech-
niques to propagate uncertainties, such as the Monte Carlo method,
raises several issues even if the model selected agrees with pre-
diction quality and computation time requirements. Simulating
many joint configurations needed to perform the tolerance anal-
ysis requires a robust reduced model especially in terms of model
building and convergence of non-linear calculations. In addition,
the number of configurations needed to fulfil the propagation of
uncertainties should be studied according to the desired precision,
the total calculation time and the number of uncertain parameters.
To determine the optimal tolerance regarding the targeted me-
chanical performance, several uncertainty propagations consider-
ing different values of tolerance needs to be performed. Depending
on the method used, calculation time can be prohibitive.
In this paper, a general description of the proposed tolerancing
approach is first presented to show the different tools and methods
used and the interaction between them. These tools and methods
are then detailed through a case study where only one source of
uncertainties is considered (i.e. hole-location errors). In the last
section, the use of the approach to design a joined structure in an
uncertain context is discussed, taking into account several sources
of uncertainties (i.e. hole-location errors, pin/hole clearance and
fastener preload).
2. Proposed approach
Computation time is the main obstacle to the propagation of
uncertainties in fastened structures. This cost is due to the time
required to evaluate one configuration defined by a set of input
parameters but also to the total number of calculations needed to
simulate the effect of uncertainties.
In this section, we first show how the calculation time for a
single configuration could be decreased using a reduced behaviour
model of the joint. Next, we focus on the definition of a tolerancing
approach based on the reduced model. The implementation of the
proposed tolerancing approach in a design process is then detailed.
2.1. Joint behaviour model
Concerning the cost of a single calculation, the calculation time
depends mainly on the model selected for the joined structure.
Due to the complexity of the physical phenomena related to joint
behaviour and the need for an accurate modelling of these phe-
nomena, the use of a solid 3-D finite element model is generally
preferred. However, due to the large number of degrees of free-
dom generated by this type of model, calculation time becomes
prohibitive as soon as the number of fasteners increases.
Several reduced joint models can be found in the literature. The
most popular one, which is generally used by engineers to model
large fastened structures, is based on a representation of fastener
geometry and behaviour by a connector element linking the two
parts of the joined structure [17,18]. Despite its low computation
time compared to 3-D models, the use of this simplified model is
generally restricted to deterministic studies. Indeed, the behaviour
law of the connector element, which includes many phenomena
such as contact, clearance recovery, fastener preload, depends on
geometric, material and contact parameters. Whatever method is
used to identify the behaviour law, from experimental tests [17]
or from local 3-D models [18], the law should be re-identified if a
parameter in the joint changes. This modelling approach becomes
very costly and therefore it is unsuitable for the propagation of
uncertainties.
To find a compromise between the ability of the model to deal
with variabilities and the calculation time, Askri et al. [16,19] have
developed a reduced finite element model based on structural
elements and shell theory. Thanks to the explicit modelling of
the fastener and its interaction with parts, the joint model does
not need a re-identification of any parameter when variability is
introduced. Thus, the model proposed by Askri et al. is well adapted
to the propagation of main uncertainties in a fastened joint such as
clearance or fastener preload.
The construction of the MCRS fastener model is based on a phys-
ical approach where deformation modes are studied in order to
define the relevant simplified geometry and kinematic behaviour.
As a result of this study 4 rigid surfaces are defined, representing
contact surfaces in a bolt: 2 surfaces for the head (HS1 and HS2 )
and 2 surfaces for the pin (PS1 and PS2 ), as shown in Fig. 1. Two
rotational degrees of freedom are considered between head and
pin surfaces to take into account the rotation of heads due to the
location of bending deformations in the junction between head
and pin. Three translational degrees of freedom are also introduced
in the connection between the two pin surfaces to take axial and
transverse deformations into account. Each degree of freedom is
associated to an equivalent stiffness by defining three connectors.
These equivalent stiffnesses are identified using a 3-D solid ele-
ment simulation through an approach based on both equivalent
displacement versus resultant forces response and elastic strain
energy.
The reduced model has demonstrated a high capacity to take
relevant phenomena into account and to provide interesting me-
chanical quantities with a low calculation time compared to a
traditional 3-D solid elements model. The saving in calculation
time is around 80%. Note that error related to modelling is not
taken into account in this paper, but it could be integrated as an ad-
ditional source of uncertainties, as proposed in [20]. Error related
 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
Fig. 1. Kinematic behaviour of the MCRS fastener model [16].
to modelling can be estimated by comparing results obtained with
the reduced model to results obtained with experiments or a 3-D
solid elements model.
The MCRS model of fastened structures is used for the propaga-
tion of uncertainties in the tolerancing approach presented in the
following section.
2.2. Tolerance synthesis
The tolerancing method is illustrated in Fig. 2 and detailed
below. The behaviour model can be assimilated to a function
where the input is one or several uncertain parameters gathered
in variable X and the output is a Mechanical Performance Indicator
variable (MPI) of the structure noted Y :
MCRS joint model
X :−−−−−−−−→ Y .
Fig. 2. Proposed approaches for the tolerance synthesis.
41
Note that the domain of variation of the uncertain parameter
which has to be tolerated is defined by a given tolerance interval
noted T .
Using the behaviour model, it is possible to evaluate the re-
sponse of a single configuration corresponding to a single value
for each uncertain parameter. When dealing with the variability
problem, the result of the simulation of a set of configurations is
therefore a set of MPI values. The target reliability level, noted r,
represents the probability of obtaining a product which complies
with a given MPI value, noted Y :
r = Prob Y (X ) ≤ Y .
( )
(2)
Note that in the proposed approach, Eq. (2) must be reversed
since the reliability level r is an input parameter imposed by speci-
fications. Thus, the propagation of uncertainties aims at obtaining a
relation between a tolerance value T and an MPI value Y calculated
for a given reliability level r:
propagation of uncertainties
(T , r ) :−−−−−−−−−−−−−−→ Y . (3)
The reliability level generally depends on the industrial applica-
tion and the risk that can be assumed. For example, in the automo-
tive industry, to reduce production costs, obtaining a small fraction
of non-compliant parts could be tolerated. However, in the nuclear
or spatial fields, all products must comply with specifications. The
reliability level r is decisive for the choice of the approach required
for the propagation of uncertainties. In fact, as shown in Fig. 2,
two approaches could be applied: the probabilistic approach if a
fraction of non-compliant products are accepted (r < 1), or the
worst-case approach when all products must be compliant.
For the probabilistic approach, Y can be evaluated by approxi-
mating the output distribution of the MPI Y according to the distri-
butions of uncertain parameters. The difficulty in this approach is
to create a set of samples of uncertain parameter values represent-
ing the input distributions. This approach generally needs a large
number of samples to ensure the accuracy of results. The challenge
is to minimise sampling size and so reduce calculation time, while
ensuring the method’s robustness. Several sampling methods exist
in the literature. One of those most frequently applied with proba-
bilistic approaches is the Monte Carlo method (MC). This method is
based on random sampling, and is generally chosen as it is simple
(1) to implement and can deal with complex problems (non-linearity,
42 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
Fig. 3. Summary flow-chart for the design of fastened joints in an uncertain context.
multi-dimensional, etc.). However, a large number of samples
are generally needed in order to ensure a high confidence level.
Alternative methods based on pseudo-random sampling, such as
the Quasi-Monte Carlo (QMC), ensure a better representation of
the domain of variation of uncertain parameters. These methods
are based on low-discrepancy sequences. Several sequences are
presented in the literature, for example by Halton [21], Hammer-
sley [22] or Sobol [23]. Examples of the application of the QMC
method in structural reliability analysis can be found in [24–26].
Results of these studies show a good match between QMC and
structural reliability problems. Nevertheless, MC sampling is more
convenient for analysing the convergence of the method since
for each sampling size a new QMC sequence has to be generated
and evaluated which therefore leads to a considerable increase in
calculation time. In this study, both MC and Sobol sequences have
been used as a sampling method.
Finding worst-cases could be likened to an optimisation prob-
lem. An optimisation method is therefore needed. Several methods
and algorithms exist in the literature and the choice depends
essentially on the nature of the optimisation problem and param-
eters. In this study, the direct problem represented by the function
Y (X ) is composed of numerous local maximums because of the
high coupling between input variables and the non-linear response
of the system. Finding the global maximum value robustly by
performing a small number of calculations is therefore a challenge.
Genetic Algorithms (GA) are widely used to resolve multidisci-
plinary optimisation problems thanks to their ability to adapt
to discontinuous, non-differentiable and multi-model problems.
Examples of studies of design optimisation of composite structures
using GA can be found in [27,28]. In this work, a Genetic Algorithm
was selected to find worst-cases.
The final objective of the proposed tolerancing approach is to
find the allowable tolerance value, noted Ta , corresponding to a
required structural performance which can be translated by an
allowable MPI Ya .
To find Ta , we propose to build an analytical law which links the
tolerance variable T to the MPI variable Y :
T = fr Y .
( )
(4)
The analytical law is built for a given reliability level and using
the suitable propagation of uncertainties approach. The identifi-
cation of parameters of the law
( is based on a limited number of
couples of discrete variables Yi , Ti . Note that obtaining several
)
couples Yi , Ti is fully parallelisable, thanks to the independence
( )
of input parameter values.
The allowable tolerance value Ta can then be calculated by
simply injecting Ya into the analytical law, as shown in Fig. 2:
Ta = fr (Ya ) . (5)
2.3. Implementing the proposed tolerancing approach in a design
process
In this section, we show how to implement the proposed toler-
ancing synthesis in a design process. The main issue is to choose
design parameters and tolerances in accordance with the process
capabilities, in order to find the best compromise between me-
chanical performance and production cost. The implementation
of the proposed tolerancing method in a design process is repre-
sented by the flow chart presented in Fig. 3.
Starting from design specifications, an initial design and sizing
of the studied structure is proposed. Nominal design parameters
are chosen so that the MPI given by this configuration is greater
than the allowable one (Ya ). This initial margin will be absorbed
by the effect of uncertainties on the MPI. The better the choice of
initial margin, the fewer iterations are required to obtain a suitable
solution. The choice of the initial margin during this pre-design
step could be based on the engineer’s experience or on fastened
joint design rules [29].
 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
Fig. 4. Geometry of reference joint.
When the process is fixed, it is then not necessary to evaluate a
set of tolerance values in order to identify the analytical law. To see
if the joint complies with specifications for the available process,
it is sufficient to evaluate the tolerance value Tp associated to the
process. When the process and its capabilities can be modified
or adjusted, n tolerance values of the uncertain parameter to be
toleranced are then selected. This number may vary depending
on the desired accuracy for the analytical law. The designer then
chooses the suitable propagation approach (probabilistic or worst-
case) depending on target reliability level r. The distribution form
and associated parameters of uncertain variables can be defined
from a statistical study based on an experimental investigation on
a batch of parts or process simulations. Note that for a worst-case
approach, only the ranges of variation of uncertain variables are
required.
By considering the allowable MPI Ya , the designer can then
determine an allowable tolerance value Ta using the analytical law.
The next step consists in comparing the allowable tolerance
Ta with the process tolerance limits Tp that can be obtained with
different processes or assembly grids. Depending on process limits
and costs, design parameters can be optimised by resizing the
joint. In the case where a process satisfies cost restriction and
allows a decrease in Ta , we can reduce the mass by proposing
an undersizing design. In this way, the joint performance in the
nominal configuration decreases, but the reliability requirements
are maintained. Conversely, if we have to increase Ta to match a
low-cost process, an oversizing design should be proposed. More
broadly, in a pre-study step where several process methods or
assembly grids are worth considering, it will be possible to build a
Pareto curve which links production costs (translated by tolerance
values) with structural performance (translated by mass calculated
with design parameters, for example).
3. Case study
In this section, a case study defined by a joint configuration, un-
certain parameters and a Mechanical Performance Indicator (MPI)
is proposed to illustrate how the tolerancing approach works. As
a first step, only hole-location errors are considered as a source
of uncertainties in order to show how probabilistic and worst-
case approaches are implemented. The study is generalised by
considering other sources of uncertainties in Section 7.
43
3.1. Reference joint
To illustrate the application of the proposed tolerance approach,
a case study is selected which is representative of real assemblies
in terms of behaviour and the effect of uncertainties. At the same
time, the chosen reference joint is as simple as possible in terms of
geometry and number of fasteners, in order to reduce the calcula-
tion time.
The reference joint considered here is a single-lap metal-
composite joint with 4 bolts, as shown in Fig. 4. The composite
adherent is made from laminated plies of unidirectional carbon fi-
bres and thermoset matrix IMA/M21-12K. The stacking sequence is
[90/−45/0/+45/0/−45/0/+45/900.5 ]s with ply thickness of
0.184 mm. The second adherent is made from aluminium alloy
7075. Bolts are composed of a titanium alloy screw and a steel nut.
The 4 bolts are fixed with an initial axial preload of 1850 N. All
material properties are listed in Table 1.
A finite element model of the reference joint is created in
Abaqus based on an MCRS fastener model and 3-D shell ele-
ments [16]. Boundary conditions are applied to two reference
points linked to each adherent extremity, as shown in Fig. 5. On
both reference points, rotation along Z is permitted. An external
tensile load of 30 kN is applied to the composite adherent. Exper-
imental tests show that this load is sufficient to initiate bearing
damage, which is generally the failure criteria selected to size
fastened joints. An offset of 2 mm from the (XZ) symmetrical plane
is applied to both reference points in order to generate an initial
imbalance in the distribution of loads between fastener rows and
thus extend the relevancy of this study.
3.2. Selection of sources of uncertainties
Three sources of uncertainties are considered in this paper:
hole-location error, clearance variation and fastener preload devi-
ation. These uncertainties are chosen because of their considerable
effect on the load distribution between fasteners which may signif-
icantly affect the mechanical performance of the structure [30–32].
It is worth noting that the tolerances applied to these parameters
directly drive the manufacturing cost of the joined structure.
Hole-location errors are mainly due to the process grid se-
lected to assemble the parts and the precision of drilling machines.
Tolerance definition of the hole-location error is based on ISO
44 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51

Table 1
Elastic properties of adherents and bolt materials.
E11 [GPa] E22 [GPa] E33 [GPa] G12 [GPa] G13 [GPa] G23 [GPa] ν12 ν13 ν23
Unidirectional ply IMA/M21-12K
143 7.89 7.89 3.92 3.92 2.76 0.33 0.33 0.43
E [GPa] ν
Aluminium alloy 7075
71.4 0.3
E [GPa] ν
Titanium alloy (Ti-6Al-4V)
110 0.3
E [GPa] ν
Carbon steel
210 0.3

Fig. 5. Finite element model of reference joint.

Fig. 6. Geometrical deviation : (a) Definition of hole-location tolerance, (b) Generation of models with clearance and hole-location error.

standard [33]. In this study, form and orientation defects of holes under cyclic loads due to fretting and fatigue damage. These failure
are not considered. The nominal hole position is therefore defined modes could be taken into account by adding relevant criteria to
by the position of a point on the hole axis, noted Ci . The index i the MPI. In this study, these failure modes are not considered.
corresponds to the numeration of fasteners defined in Fig. 4. The For preload deviation, noted p, several studies highlighted un-
centre of a hole in the deviated position, noted Ci′ , is obtained by a certainties related to the introduction of the fastener preload. The
−→ gap between real and target preload implied by a torqueing op-
translation with a vector ⃗ti = Ci Ci′ , as shown in Fig. 6a. Coordinates
eration is essentially due to variabilities in material properties and
of
( the vector
−
→ −
→ ) ⃗ti are expressed in a polar local coordinate system friction behaviour of fasteners [34,35]. In addition, a loss in preload
Ci , X , Y . occurs during the joint life cycle. This phenomenon is mainly due
⎧ ( ) ⎫ to loosening and relaxation phenomena. The loosening in threaded
⎨θi = X⃗ , t⃗i
⎪ −π ≤ θi ≤ π ⎪
⎬ fasteners is generally caused by dynamic or cyclic loads [36,37].
t⃗i =   T
. (6) Viscoelastic behaviour of materials and contact surfaces is the main
⎩ti = t⃗i 
⎪ 0 ≤ ti ≤
⎪
⎭ mechanism which produces preload relaxation. This phenomenon
2 is amplified in the case of laminated composites because of the
Drilling of holes in each part is subject to variation due to several presence of a polymer matrix in the material. Several studies
phenomena, such as tool bit wear or tool vibrations. Several studies showed the evolution of the loss in preload in relation to clamping
show that a large clearance has a negative effect on joint behaviour duration and temperature [38].
 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
Fig. 7. Density of probability distribution of hole-location errors.
From a modelling point of view, the generation of clearance
deviation and hole-location errors in the models consists in trans-
lating the set of nodes located at the hole border following the
−→
translation vector ⃗ti = Ci Ci′ , as illustrated in Fig. 6b. This approach
needs either a coarse mesh at the hole border or the translation
of several circles of nodes in order to avoid excessive element
distortion. Due to the low amplitude of deviations, it was not nec-
essary in this study to translate several circles. For preload, given
that the fastener preload is controlled by an initial overclosure
between fastener head surfaces and parts, varying this parameter
in the model consists in changing the initial overclosure. A previous
study has shown a proportional relation between overclosure and
preload which makes the installation of the targeted preload easy.
Note that the resolution of the contact law in joint models is
very much dependent on pin/hole clearance and contact pressure
under fastener heads. Abaqus proposes several ways to control the
contact resolution, such as clearance/pressure law or contact sta-
bilisation. Hard contact law was selected in order to avoid the defi-
nition of non-physical parameters. To ensure a robust convergence
for all joint configurations, whatever the deviation amplitudes of
uncertain parameters, the authors performed a study to optimise
contact stabilisation parameters.
3.3. Mechanical performance indicator
A Mechanical Performance Indicator (MPI) based on maximal
bolt-load is defined in this section. Loads transmitted by each bolt
are useful information to apply dimensioning criteria related to
fastened joints, such as bearing damage or shear failure of fasten-
ers [30,39–44]. Maximal bolt-load, Fmax , used to define the MPI, is
calculated from the resulting forces Fi transmitted in each bolt:
Fmax = sup (Fi ) .
i=1,n
The MPI, noted Y , is calculated as the relative variation of the
maximal transmitted load Fmax in the configuration, with errors
compared to the maximal load given by the nominal configuration
0
(zero-clearance and no hole-location errors) Fmax :
0
Fmax − Fmax
Y = 0
. (8)
Fmax
45
Fig. 8. Sobol sequence of 1000 points representing hole-location errors with T =
0.1 mm.
4. Probabilistic approach
4.1. Sampling method
The distribution of the hole-location error is assumed to be
uniform in this section, as shown in Fig. 7. The Sobol sequence
is used here as the QMC sampling method to generate uniform
distribution of this uncertainty. Fig. 8 shows an example of 1000
points representing hole-location errors in one hole, with a loca-
tion tolerance T = 0.1 mm using the Sobol sequence.
4.2. Results of the propagation of uncertainties
Results of the propagation of uncertainties from the QMC
method with 5000 points are discussed in this section. This number
of samples corresponds to a coefficient of variation close to 5% with
a probability of failure of 0.1 (r = 0.9). Fig. 9 presents respectively
the distribution of probability density and the cumulative proba-
bility of the MPI for T = 0.1 mm.
It is interesting first to note the significant increase in the
MPI with the presence of hole-location error. For example, the
MPI in 25% of tested configurations is higher than 10%. This pro-
portion shows the negative impact of hole-location error on the
behaviour of fastened composite joints. The opposite effect can
also be seen with the presence of a small fraction (around 10%) of
configurations, providing negative values for the MPI. In fact, load
distribution between fasteners in a nominal joint (without hole-
location errors) is initially unbalanced due to the offset of external
load. The introduction of hole-location errors leads to a rebalance
and, in some cases the resulting load distribution therefore gives
a lower maximal bolt-load than that given by the nominal joint,
(7) hence the negative values of the MPI.
4.3. Convergence of QMC
Even if the QMC method theoretically ensures uniform sam-
pling, the accuracy of results after the propagation of uncertainties
varies with the number of samples. Choosing an optimal number
of samples could be useful to control the calculation time. In order
to analyse this variation, different sampling sizes are tested, with
46 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51

Fig. 10. Evolution of the precision criteria as a function of QMC sampling size.

Fig. 9. Results of the propagation of hole-location errors with T = 0.1 mm using

5000 samples based on QMC.

n ranging from 100 to 5000 samples. Fig. 10 shows the evolution

according to sampling size of the relative deviation calculated with
a reliability level of 0.9 and defined by:
Yn − Y5000
∆Y = . (9)
Y5000
The deviation tends to stabilise as the number of samples in-
creases. ⏐∆Y ⏐ is lower than 2% starting from 1000 samples. Note
⏐ ⏐
that this value is coherent regarding the values retained for other
studies dealing with structural reliability analysis [6]. This number
Fig. 11. Evolution of the maximal MPI as a function of GA generation for T =
of samples is thus used for all the results presented in this paper. 0.1 mm.

5. Worst-case approach

5.2. Convergence of genetic algorithm

5.1. Genetic algorithm
The Genetic Algorithm, used here to find worst-cases, is a
meta-heuristic evolutionary optimisation method. Inspired by the
process of natural evolution, GA uses several notions related to
biology. The individual is defined by the uncertain geometric pa-
rameter. In our case, the individual contains one chromosome
which is composed of 2n genes representing hole-location error
parameters. The chromosome is therefore formatted as follows:
t1 θ1 t2 θ2 . . . tn θn .
Concerning GA parameters (initial population size, maximal
number of generations and parameters of genetic operators), a
relevant choice of these parameters is essential in order to ensure
a quick convergence and accurate solution. The optimisation of
GA parameters can be performed by a parametric study [45]. As
this is a very time-consuming study to carry out, the choice of GA
parameters used here is based on recommendations and values
from the literature [45,46] (see Table 2).
The DEAP computation framework developed by Fortin et al.
[47] was integrated into the Python script used to generate finite
element models which include uncertainties. An analysis of the
convergence of the GA is proposed in this section. The tolerance
value for location error considered here is 0.1 mm.
Due to the stochastic character of the GA, the algorithm is
run 3 times independently. Fig. 11 shows the MPI of worst-case
configuration for each generation (Y = max (Y )), and a relative
stabilisation can be seen, starting from the 6th to 10th generations.
Regardless of the non-linear character of the problem and high
coupling between input parameters, the similarity between the
3 curves and the small difference between the proposed optimal
solutions shows the capacity of the GA to adapt to the optimisation
problem studied here. The gap between maximal values produced
by the worst-case approach (29%–30%) and by the reliability-based
approach (25%–26%) shows the advantage of using an optimisa-
tion method to ensure high levels of reliability. The relative gap

Table 2
GA parameters.
Initial population size Maximal number of generations Selection method Crossover probability Mutation probability
100 15 Tournament 0.85 0.01
 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51 47

Fig. 12. Hole-location error parameters and bolt-loads of a worst-case.

between converged value Y obtained with the three runs is close

to 3%. It is also worth noting that this gap has no significant effect
on the analytical tolerancing law.

5.3. Analysis of worst cases found by GA

The 3 worst-case solutions proposed by the 3 runs performed

with the GA have several points in common. These concern the
orientation and the amplitude of hole-locations. Input and output
−→
values of one of the 3 worst-cases is shown in Fig. 12 where Fi
−
→
represents bolt-load and ti = (ti , θi ) the deviated hole-location.
Note that the difference between the sum of bolt-loads in the
X-direction and the global load applied to the joint is essentially
due to friction forces.
Knowing that the amplitude of a hole-location error could take
a value of between 0 and T /2, this parameter reaches the maximal
value of T /2 in all holes. On the other hand, the hole where
the bolt-load is maximal is translated to the opposite side of the
external load, while the other 3 holes are translated in the same
direction as the external load. There is a physical explanation for
this observation. When a hole is shifted to the opposite direction of
the external load, the distance between pin/hole contact surfaces
is reduced. The sliding distance resulting from clearance recovery
becomes smaller and the load transmitted by the bolt increases.
This bolt-overload increases when the sliding distance is maximal Fig. 13. Evolution of bolt-loads as a function of global load.
in the other bolts. To illustrate this phenomenon, curves in Fig. 13
show the evolution of bolt-loads as a function of the global load
applied to the joint. The graph shows the suppression of the sliding
phase in bolt 3 (bolt under greatest load) and the increase in sliding To do this, the QMC method and GA were applied to 3 tolerance
displacement on other bolts. Contact in bolt #4 is established when values for hole-location error: 0.1 mm, 0.2 mm and 0.3 mm. These
bearing deformation in bolt #3 becomes great enough to recover values were chosen to be coherent with industrial practices.
clearance in bolt #4. Results of the propagation of uncertainties are presented( in)
Fig. 14. Points correspond to different couples of parameters Yi , Ti
Using this type of analysis it may be possible to find the worst-
case directly without the application of an optimisation method where Yi is calculated with the QMC method using two reliability
or limiting the space of variation of uncertain parameters in order levels (r = 0.9 and r = 0.99) and GA for r = 1.
to reduce the calculation time. However, finding the worst-case In this case study, where only one source of variability is in-
intuitively in the case of an assembly with a large number of troduced, the relation between T and Y seems to evolve quasi-
fasteners and complex loading and geometry is impractical. In this linearly. Knowing that a tolerance value is equal to zero, the MPI
case, it is unsafe to reduce the space of variation of uncertain is also theoretically equal to zero because this case corresponds
parameters. Geometric methods developed for the optimisation of to the nominal configuration. Consequently, the natural law to be
geometric tolerance [10,11] could be eventually used to determine used here is a proportional function:
critical configurations in order to reduce the variation domain of
T = ar Y . (10)
uncertain parameters. The determination of the worst-case with
GA would therefore be less time consuming. The single unknown parameter here is the slope ar which de-
pends on reliability level. The analytical function shown in Fig. 14
6. Establishing the analytical tolerancing law is identified using the least square method from the results of the
propagation of uncertainties. The values obtained for ar are listed
In this section, methods and tools developed for the propaga- in Table 3.
tion of uncertainties are used to construct the analytical toleranc- As expected, the slope tends to decrease when the reliability
ing law. level becomes higher. This shows the effect of reliability level on
48 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
Fig. 14. Results of the propagation of uncertainties and analytical tolerancing
models using QMC and GA.
Table 3
Parameters of analytical tolerancing model.
r 0.90 0.99 1 arbitrarily. All distributions are truncated at ±3σ in order to be
ar [mm] 6.9410−2 5.0710−2 3.4810−2
manufacturing cost: the lower the tolerance value, the higher the
cost [48]. The analytical tolerancing law allows the designer to
determine the allowable hole-location tolerance Ta to be consid-
ered in order to respect the reliability level required for a given
allowable MPI Ya :
Ta = ar Ya . (11)
7. Design of fastened joints in an uncertain context
In previous sections, the tolerancing approach was applied to
tolerance hole-location errors without any other sources of un-
certainties. In order to demonstrate the ability of the approach to
deal with an industrial problem, a case with several sources of
uncertainties and the consideration of non-uniform distributions
for all uncertain parameters is addressed. The objective remains
the identification of an analytical tolerancing law which links hole-
location tolerance to the same MPI used in the previous sections.
7.1. Quantification of uncertainties
First, we propose to focus on the domain of variation of each
uncertainty (minimal and maximal values). For the radial clear-
ance uncertainty, the allowable value should be higher than hole-
location tolerance, in order to ensure the mounting of all fasteners
(cmin = T /2). The maximal radial clearance cmax is assumed equal
to the sum of minimal radial clearance cmin and half of the sum
of tolerance intervals in hole diameters in both adherents and in
pin diameter. The tolerance interval for each diameter is defined
by a grade of accuracy selected from aeronautical practices. For
an aluminium hole, an accuracy grade of 8, corresponding to a
tolerance interval equal to 22 µm, is retained. For a composite
plate, where the drilling quality is generally worse than for metallic
materials, an accuracy grade of 11 is selected, corresponding to
Table 4
Limits of variation of uncertain parameters.
Uncertain parameter Variable Minimal value Maximal value
Amplitude (t) 0 T /2
Hole-location error
Orientation (θ ) −π π
Radial clearance c T /2 T /2 + 62 µm
Axial bolt-preload p 3000 N 6000 N
a tolerance interval of 90 µm. Concerning bolts, the tolerance
interval for pin diameter, as mentioned in the technical sheet for an
NAS bolt with a pin diameter of 6.35 mm, is 12 µm. The tolerance
interval for radial clearance uncertainty is then 62 µm.
Concerning fastener preload, designers generally tend to apply
the highest allowable tightening. This maximal preload is lim-
ited by the out-of-plane strength of the composite laminate in
order to avoid material damage during the tightening operation.
In addition, uncertainty introduced by the torqueing operation and
preload relaxation implies a significantly lower effective (in ser-
vice) preload than the maximal preload tolerated by the composite
material. Thus, in accordance with aeronautical practices [49–51],
a maximal preload pmax of 6000 N is chosen. From references [5,38],
we assume that axial preload may attempt a minimal value pmin
equal to half of the maximal preload, giving pmin = 3000 N.
Once domains of variation are defined, then a distribution can
be attributed for each uncertain parameter. Here, normal distribu-
tion defined by standard deviation σ and mean value µ was chosen
coherent with physical finite limits of the uncertain parameters.
These limits correspond to minimal and maximal values of the
considered uncertainties defined previously. Table 4 recapitulates
different minimal and maximal values of uncertain parameters.
7.2. Results of the reliability-based approach
The Monte Carlo method is used here as a sampling method
for the probabilistic approach in order to verify the convergence of
the method without increasing the calculation time. The number of
samples used for the propagation of clearance, preload and hole-
location uncertainties is 1000. Three tolerance values on hole-
location error are considered:0.1, 0.2 and 0.3 mm. A fourth case
where only clearance and preload uncertainties are propagated is
also considered in order to study the effect of the two uncertainties
without hole-location errors (T = 0).
Results of the propagation of uncertainties are shown in Fig. 15.
We note that in the case where T = 0, the MPI is higher than
20% with the presence of clearance and preload uncertainties
only. The effect of clearance on the MPI is similar to the effect
of hole-location error. In fact, varying pin–hole clearance affects
the sliding distance before radial contact is established and hence
the distribution of load between fasteners. The MPI is then at its
maximum when clearance in the bolt which transmits the maximal
load is minimal and clearance in the others is maximal. Concerning
preload uncertainty, knowing that the flow of force in the joint
is transmitted by friction stress and through the radial contact
between pins and holes, the axial preload mainly affects the part
of the flow of forces which are transferred by friction. The greater
the sum of bolt preloads, the lower the bolt-loads, generating a
decrease in the MPI whatever the distribution of preloads between
bolts.
We also note that the difference between maximal values of
the MPI from the MC method and the MPI of worst-cases obtained
by GA is greater in this case study compared to the previous one.
In fact, introducing a greater number of uncertain parameters
involves a larger domain of variation and then a lower probability
to sample a configuration which leads to maximise the MPI. This
 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51 49

Fig. 15. Results of the propagation of uncertainties based on probabilistic and worst-case approaches.

Table 5
Parameters of analytical tolerancing models.
r 0.90 0.99 1
ar [mm] 0.245 0.239 0.185
br [mm] −0.554 −0.624 −0.618

probability decreases when considering normal distributions for

the uncertain parameters instead of uniform ones. In fact, we
showed in Section 5.3 that worst-cases correspond to configura-
tions where uncertain parameters take the values of the limits of
the domain of variation. Thus when considering normal distribu-
tions, the probability of sampling values close to the limits of the
domain of variation is then lower. This observation highlights once
again the relevance of using the worst-case approach when the
probability of failure required is close to zero.

7.3. Identifying the analytical tolerancing law

Results from the propagation of uncertainties based on prob-

abilistic and worst-case approaches are presented in Fig. 16. Th
e relation between the MPI and tolerance value appears to be
non-linear, contrary to the case tested previously where only Fig. 16. Results of the propagation of uncertainties and analytical tolerancing
hole-location errors were considered. To describe this relation, a models using MC and GA.

logarithmic function with 2 parameters is chosen to define the

analytical tolerancing law:
8. Conclusion
T = ar Ln Y + br .
( )
(12)

Parameters are identified using the least-square method (see
Table 5).
From an industrial point of view, the analytical law makes it
possible to evaluate the design margin required to counter the
effect of variabilities. For example, to switch from an assembly
grid based on a single drilling operation per fastener (T close to
0.02 mm) to an assembly grid based on interchangeability of parts
(T close to 0.1 mm), the design margin should absorb an increase
in Y from 1.18 to 1.22 for a reliability of 0.99, and from 1.32 to 1.5
for a reliability of 1.
In this paper, a method is proposed to control uncertainties
in fastened metal-composite joints by determining appropriate
tolerances. The main problem that makes such an analysis difficult
to carry out is the calculation time required for the number of
configurations to be simulated. The solution proposed here is to
use a Multi-Connected Rigid Surfaces joint model. This model has
previously shown its ability to take into account preponderant
physical phenomena and provide essential mechanical quantities
while keeping to a reasonable calculation time. Two approaches
are considered for the propagation of uncertainties: a probabilistic
50 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
approach based on Monte Carlo or Quasi-Monte Carlo sampling
methods and a worst-case approach using a Genetic Algorithm.
The general tolerancing approach was applied to a 4-bolt single-
lap aluminium-composite joint for the tolerance of hole-location
errors. The results of propagation of uncertainties have shown that
the relation between tolerance value and Mechanical Performance
Indicator can be represented by a proportional function when
no other sources of uncertainties are involved, and a logarithmic
function when uncertainties on clearance and preload are added.
We also propose a way to integrate this formalisation between
mechanical performance and uncertainties in a sizing loop in order
to find a relevant compromise between performance and cost.
Several future studies can be highlighted to complete some
aspects of this work. First, the method should be extended to other
types of variability resulting from orientation or form errors and
material defects. Errors related to finite element modelling could
be also taken into account to make more robust the application
of the proposed tolerance synthesis in an industrial context. In
addition, the capability of the MCRS model could be used to deal
with other mechanical performance criteria, such as net cross-
section failure, bearing failure or fretting and fatigue damage. In the
same way, due to its significant influence on load distribution, the
non-linear behaviour of material resulting from bearing damage
or time-dependent phenomena should be considered to extend
the domain of validity of the proposed approach. Finally, using
such a method in an industrial context requires specific software
to automate the generation of FE models and the uncertainty
propagation processing due to repetitive tasks.
References
[1] Hyer MW, Klang EC, Cooper DE. The effects of pin elasticity, clearance, and
friction on the stresses in a pin-loaded orthotropic plate. J Compos Mater
1987;21:190–206. http://dx.doi.org/10.1177/002199838702100301.
[2] Kelly G, Hallström S. Bearing strength of carbon fibre/epoxy laminates: Effects
of bolt-hole clearance. Compos Part B: Eng 2004;35:331–43.
[3] Al-Nassar YN, Khurshid H, Arif AFM. The effect of clearance and pre-tension on
the performance of a bolted-joint using 3D FEA. Arab J Sci Eng 2012;37:749–63.
http://dx.doi.org/10.1007/s13369-012-0191-8.
[4] McCarthy MA, McCarthy CT, Padhi GS. A simple method for determining the
effects of bolt–hole clearance on load distribution in single-column multi-bolt
composite joints. Compos Struct 2006;73:78–87. http://dx.doi.org/10.1016/j.
compstruct.2005.01.028.
[5] Caccese V, Mewer R, Vel SS. Detection of bolt load loss in hybrid compos-
ite/metal bolted connections. Eng Struct 2004;26:895–906. http://dx.doi.org/
10.1016/j.engstruct.2004.02.008.
[6] Bruyère J, Dantan J-Y, Bigot R, Martin P. Statistical tolerance analysis of bevel
gear by tooth contact analysis and Monte Carlo simulation. Mech Mach Theory
2007;42:1326–51. http://dx.doi.org/10.1016/j.mechmachtheory.2006.11.003.
[7] Dantan J-Y, Qureshi A-J. Worst-case and statistical tolerance analysis based
on quantified constraint satisfaction problems and Monte Carlo simula-
tion. Comput Aided Des 2009;41:1–12. http://dx.doi.org/10.1016/j.cad.2008.
11.003.
[8] Nigam SD, Turner JU. Review of statistical approaches to tolerance analysis.
Comput Aided Des 1995;27:6–15. http://dx.doi.org/10.1016/0010-4485(95)9
0748-5.
[9] Dumas A, Dantan J-Y, Gayton N. Impact of a behavior model linearization
strategy on the tolerance analysis of over-constrained mechanisms. Comput
Aided Des 2015;62:152–63. http://dx.doi.org/10.1016/j.cad.2014.11.002.
[10] Homri L, Teissandier D, Ballu A. Tolerance analysis by polytopes: Taking
into account degrees of freedom with cap half-spaces. Comput Aided Des
2015;62:112–30. http://dx.doi.org/10.1016/j.cad.2014.11.005.
[11] Xu S, Keyser J. Geometric computation and optimization on tolerance dimen-
sioning. Comput Aided Des 2014;46:129–37. http://dx.doi.org/10.1016/j.cad.
2013.08.025.
[12] Söderberg R, Wärmefjord K, Lindkvist L. Variation simulation of stress during
assembly of composite parts. CIRP Ann—Manuf Technol 2015;64:17–20. http:
//dx.doi.org/10.1016/j.cirp.2015.04.048.
[13] McCarthy CT, McCarthy MA. Three-dimensional finite element analysis of
single-bolt, single-lap composite bolted joints: Part II—-effects of bolt-hole
clearance. Compos Struct 2005;71:159–75. http://dx.doi.org/10.1016/j.comp
struct.2004.09.023.
[14] Gray PJ, McCarthy CT. A global bolted joint model for finite element analysis
of load distributions in multi-bolt composite joints. Composites B 2010;41:
317–25. http://dx.doi.org/10.1016/j.compositesb.2010.03.001.
[15] McCarthy CT, Gray PJ. An analytical model for the prediction of load dis-
tribution in highly torqued multi-bolt composite joints. Compos Struct
2011;93:287–98.
[16] Askri R, Bois C, Wargnier H, Lecomte J. A reduced fastener model using multi-
connected rigid Surfaces for the prediction of both local stress field and load
distribution between fasteners. Finite Elem Anal Des 2016;110:32–42. http://
dx.doi.org/10.1016/j.finel.2015.11.004.
[17] Gray PJ, McCarthy CT. A highly efficient user-defined finite element for load
distribution analysis of large-scale bolted composite structures. Compos Sci
Technol 2011;71:1517–27.
[18] Kapidžić Z, Nilsson L, Ansell H. Finite element modeling of mechanically
fastened composite-aluminum joints in aircraft structures. Compos Struct
2014;109:198–210. http://dx.doi.org/10.1016/j.compstruct.2013.10.056.
[19] Askri R. Approche fiabiliste pour le tolérancement des assemblages par fixation
de structures composite-métal (Ph.D. thesis), University of Bordeaux; 2016.
[20] Oberkampf WL, DeLand SM, Rutherford BM, Diegert KV, Alvin KF. Error and
uncertainty in modeling and simulation. Reliab Eng Syst Saf 2002;75:333–57.
http://dx.doi.org/10.1016/S0951-8320(01)00120-X.
[21] Halton JH. Algorithm 247: Radical-inverse quasi-random point sequence.
Commun ACM 1964;7:701–2. http://dx.doi.org/10.1145/355588.365104.
[22] Hammersley JM, Handscomb DC. Random, pseudorandom, and quasirandom
numbers. In: Monte Carlo methods. Springer Netherlands; 1964. p. 25–42.
http://dx.doi.org/10.1007/978-94-009-5819-7__3.
[23] Sobol’ IM. On the distribution of points in a cube and the approximate evalua-
tion of integrals. USSR Comput Math Math Phys 1967;7:86–112. http://dx.doi
.org/10.1016/0041-5553(67)90144-9.
[24] Dai H, Wang W. Quasi-Monte Carlo method for structural reliability analysis.
Hangkong Xuebao/Acta Aeronaut Astron Sin 2009;30:666–71.
[25] Dai H, Wang W. Application of low-discrepancy sampling method in structural
reliability analysis. Struct Saf 2009;31:55–64. http://dx.doi.org/10.1016/j.stru
safe.2008.03.001.
[26] Zhang H, Dai H, Beer M, Wang W. Structural reliability analysis on the basis
of small samples: An interval quasi-Monte Carlo method. Mech Syst Signal
Process 2013;37:137–51. http://dx.doi.org/10.1016/j.ymssp.2012.03.001.
[27] Gomes HM, Awruch AM, Lopes PAM. Reliability based optimization of lam-
inated composite structures using genetic algorithms and Artificial Neural
Networks. Struct Saf 2011;33:186–95. http://dx.doi.org/10.1016/j.strusafe.20
11.03.001.
[28] Narayana Naik G, Gopalakrishnan S, Ganguli R. Design optimization of com-
posites using genetic algorithms and failure mechanism based failure crite-
rion. Compos Struct 2008;83:354–67. http://dx.doi.org/10.1016/j.compstruct.
2007.05.005.
[29] Kulak GL, Fisher JW, Struik JHA. Guide to design criteria for bolted and riveted
joints. American Institute of Steel Construction Inc; 1987.
[30] McCarthy MA, Lawlor VP, Stanley WF, McCarthy CT. Bolt-hole clearance ef-
fects and strength criteria in single-bolt, single-lap, composite bolted joints.
Compos Sci Technol 2002;62:1415–31.
[31] Askri R, Bois C, Wargnier H. Effect of hole-location error on the strength of
fastened multi-material joints. Procedia CIRP 2016;43:292–6. http://dx.doi.or
g/10.1016/j.procir.2016.02.040.
[32] Lecomte J, Bois C, Wargnier H, Wahl J-C. An analytical model for the prediction
of load distribution in multi-bolt composite joints including hole-location
errors. Compos Struct 2014;117:354–61. http://dx.doi.org/10.1016/j.compstr
uct.2014.06.040.
[33] ISO 1101, Geometrical product specifications (GPS) —Geometrical tolerancing
—Tolerances of form, orientation, location and run-out, 2012.
[34] NASA, Criteria for preloaded bolts, Lyndon B. Johnson Space Center, 1998.
[35] Seibel A, Japing A, Schlattmann J. Uncertainty analysis of the coefficients of
friction during the tightening process of bolted joints. J Uncertain Anal Appl
2014;2:21. http://dx.doi.org/10.1186/s40467-014-0021-5.
[36] Ibrahim RA, Pettit CL. Uncertainties and dynamic problems of bolted joints and
other fasteners. J Sound Vib 2005;279:857–936. http://dx.doi.org/10.1016/j.js
v.2003.11.064.
[37] Thoppul SD, Gibson RF, Ibrahim RA. Analytical and experimental characteriza-
tion of the effects of vibration on relaxation in composite bolted joints 2006,
p. 1081–92.
[38] Caccese V, Berube KA, Fernandez. J. Daniel Melo M, Kabche JP. Influence
of stress relaxation on clamp-up force in hybrid composite-to-metal bolted
joints. Compos Struct 2009;89:285–93.
[39] McCarthy CT, McCarthy MA, Lawlor VP. Progressive damage analysis of multi-
bolt composite joints with variable bolt-hole clearances. Compos Part B: Eng
2005;36:290–305.
[40] Egan B, McCarthy MA, Frizzell RM, Gray PJ, McCarthy CT. Modelling bearing
failure in countersunk composite joints under quasi-static loading using 3D
 R. Askri et al. / Computer-Aided Design 100 (2018) 39–51
explicit finite element analysis. Compos Struct 2014;108:963–77. http://dx.
doi.org/10.1016/j.compstruct.2013.10.033.
[41] Ireman T, Ranvik T, Eriksson I. On damage development in mechanically
fastened composite laminates. Compos Struct 2000;49:151–71. http://dx.doi.
org/10.1016/S0263-8223(99)00130-0.
[42] Matthews FL, Roshan AA, Phillips LN. The bolt bearing strength of glass/carbon
hybrid composites. Composites 1982;13:225–7. http://dx.doi.org/10.1016/
0010-4361(82)90003-9.
[43] Wang H-S, Hung C-L, Chang F-K. Bearing failure of bolted composite joints. Part
I: Experimental Characterization. J Compos Mater 1996;30:1284–313.
[44] Le Goff E, Bois C, Wargnier H. A progressive intra- and inter-laminar damage
model to predict the effect of out-of-plane confinement on pin-bearing be-
haviour of laminated composites. J Compos Mater 2017;51.
[45] Le Riche R, Haftka RT. Improved genetic algorithm for minimum thickness
composite laminate design. Compos Eng 1995;5:143–61. http://dx.doi.org/10.
1016/0961-9526(95)90710-S.
51
[46] Forouraghi B. Worst-case tolerance design and quality assurance via genetic
algorithms. J Optim Theory Appl 2002;113:251–68. http://dx.doi.org/10.1023/
A:1014826824323.
[47] Fortin F-A, Rainville F-MD, Gardner M-A, Parizeau M, Gagné C. DEAP: Evolu-
tionary algorithms made easy. J Mach Learn Res 2012;13:2171–5.
[48] Chase KW, Parkinson AR. A survey of research in the application of tolerance
analysis to the design of mechanical assemblies. Res Eng Des 1991;3:23–37.
http://dx.doi.org/10.1007/BF01580066.
[49] Zhao Y. Torque limit for bolted joint for composites part A: TTTC properties of
laminated composites. NASA; 2002.
[50] Thomas FP. Experimental observations for determining the maximum torque
values to apply to composite components mechanically joined with fasteners.
Alabama: Marshall Space Flight Center; 2006.
[51] Helsehurst RB. Design and analysis of structural joints with composite mate-
rials. DEStech Publications; 2013.