Mixed Integer Linear Programming formulations of the

stacking sequence and blending optimisation of

composite structures
G. Ntourmasa,b,∗, F. Glockb , F. Daoudb , G. Schuhmacherb , D.
Chronopoulosa , E. Özcanc
a
Institute for Aerospace Technology & The Composites Research Group, The University
of Nottingham, NG7 2RD, UK
b
Stress Methods and Optimisation, Airbus Defence and Space GmbH, 85077 Manching,
Germany
c
Computational Optimisation and Learning Lab, The University of Nottingham, NG7
2RD, UK

Abstract

This manuscript proposes two novel formulations for the manufacturable

stacking sequence retrieval of laminated composite structures. Detailed siz-
ing of composite structures is commonly tackled by a two-stage optimisation
approach, the first stage being a weight minimisation resulting in a contin-
uous thickness and stiffness distribution of the structure. The second stage,
which is the focus of the current paper, aims at retrieving stacking sequences
that match the optimised target thickness and stiffness characteristics of the
first stage, while also fulfilling a set of predefined design and manufactur-
ing rules, with blending being one of the most crucial ones. The problem
of retrieving layered designs that meet all of the prescribed composite rules
is formulated as two Mixed Integer Linear Programming instances which

∗
Corresponding author
Email address: [email protected] (G. Ntourmas)

Preprint submitted to Composite Structures October 21, 2021

mainly differ in how the blending is treated. Using either of the two for-
mulations, mathematical programming algorithms can be employed to solve
the problem to global optimality. Results show that both formulations can
produce manufacturable solutions that fulfil the imposed composite rules.
High-quality solutions can be obtained consistently, while benefiting from
increased design freedom concerning the blending formulation.
Keywords: Integer Programming, Blending, Stacking Sequence
Optimisation

1. Introduction

Composite materials, such as fibre-reinforced plastics, find application

in many high-end products such as aeronautical or automotive structures,
wind turbines, sports equipment and more because they offer reduced weight
and enhanced mechanical characteristics. Aside from this, composites are
generally anisotropic materials, providing the engineer with increased design
freedom when compared to the classic metallic materials. However, the prob-
lem that arises is that with increased design freedom, the complexity of sizing
the structure in detail also increases.
Most high performance aircraft structures are produced by laying unidi-
rectional tapes of polymer pre-impregnated fibres one next to each other to
form a lamina or ply with a specific fibre orientation. More plies are added on
top, in such a way that their orientation and number forms a laminate which
satisfies the stiffness and thickness requirements of the part. Large-scale
structures are discretised in so-called patches across their span to address
the local requirements and therefore the stacking sequence varies between

2
the patches.
During the optimisation process, a large set of constraints must be taken
into account. First of all, design rules regulating the stacking characteristics
of each patch in the structure should be followed. Secondly, to ensure struc-
tural integrity and manufacturability, the stacking sequence of neighbouring
patches needs to be integrated into a design that follows specific rules, re-
ferred to as manufacturing rules, in which blending or continuity of the plies
is of significant importance. Finally, the structure needs to fulfil a large set
of physical constraints, related to strength, buckling, aeroelasticity, damage
tolerance and other requirements.
The problem of composite stacking sequence optimisation has been stud-
ied thoroughly by many researchers over the last decades. A review of dif-
ferent approaches to stacking sequence optimisation can be found in Ghiasi
et al. [15, 14]. Two particularities of the problem dominate the nature of the
proposed optimisation approaches. Firstly, the problem contains mostly inte-
ger design variables but both discrete and continuous constraints. Secondly,
many of the continuous constraints require computationally expensive Finite
Element (FE) model evaluations. Therefore, although metaheuristics have
been widely employed to perform the optimisation [26, 46, 34, 9, 4, 42, 22, 3],
their usage becomes prohibitive for large-scale problems due to the increased
number of optimisation cycles required, combined with the computational
expense of the FE evaluations. Furthermore, gradient-based optimisation
methods [45, 31, 44, 6, 13, 25] are very well suited for the continuous part of
the problem, combined with the solution cost of the FE model, but fail to
efficiently handle the discrete part of the problem.

3
 In an attempt to incorporate the benefits and eliminate the drawbacks
of both gradient and stochastic algorithms, a two-stage optimisation process
has been employed by several researchers, previously. During the first stage,
a weight minimisation is performed [18, 36] using either lamination [19, 32]
or polar parameters [35, 37] to describe the structural stiffness. The result
of this stage is an optimal, continuous distribution of thickness and stiffness
characteristics. The continuous thickness is usually rounded up to the near-
est integer number of layers in order to achieve a manufacturable design.
This discretisation is performed since the thickness of the pre-impregnated
tape that will be used in manufacturing, is decided a priori and is treated as
a constant parameter during the sizing optimisation. This discrete thickness
and the aforementioned stiffness distribution, can be used as targets by the
second stage of the optimisation process which is usually a stochastic algo-
rithm [28, 20, 7, 23] that can handle design and manufacturing rules more
efficiently.
The principal novelty of this work is the formulation of the composite
manufacturing rules as two different Mixed Integer Linear Programming
(MILP) models. Although potential structural engineering applications of
operational research have been identified [27], a complete MILP formulation
of the stacking sequence optimisation has not yet been presented to the best
of the authors’ knowledge with the exception of some composite design rules
which have been previously formulated in such a way [2]. Additionally, Kang
& Blom [24] have worked towards that direction, having however treated
the blending rules in the context of compliance with pre-computed lami-
nates, which is not the case for the present study. Other studies focusing on

4
tree structures, branch and bound algorithms and beam search algorithms
[40, 39, 50, 29, 30, 10, 11], are algorithmic approaches built around the spe-
cific nature of the stacking sequence optimisation problem in question. In
this work, two complete formulations accounting for all composite design
and manufacturing rules which can be used in any optimisation solver able
to handle a MILP problem formulation are presented.
The rest of this paper is structured as follows: In Section 2 composite
guidelines relevant to the stacking sequence retrieval for aeronautical struc-
tures are presented. The two MILP formulations of the blended stacking
sequence optimisation are formulated in Section 3. Results from the two
formulations are demonstrated in Section 4 and the findings of this work are
summarised in Section 5.

2. Composite rules

Besides physical constraints such as buckling, strength, maximum al-

lowable displacements etc., that naturally govern the optimisation problem,
there are also several rules, which are commonly used in the aerospace indus-
try when designing composite structures. These rules serve as guidelines for
manufacturing laminates, which are less prone to high stress concentration
and unwanted mechanical coupling effects. Most of these rules have also been
presented in previous works [21, 5]. Below, the composite rules implemented
and used in this work are summarised.
The composite rules are grouped into two categories, i.e. design and
manufacturing rules. Design rules regulate the stacking sequence of a single
uniform laminate.

5
 1. Symmetry. Symmetric laminates about the mid-plane are commonly
used to avoid bending-extension coupling.

2. Balance. Balanced laminates consist of equal +θ and −θ orientations

(θ , 0o , 90o ), in order to eliminate shear-extension coupling.

3. Minimum percentage. A minimum percentage of all fibre orientations

used in a laminate might be enforced to minimise matrix degradation
and favour a fibre-dominated failure mode instead.

4. Damage tolerance. External plies should never be in the direction of

the main load path. Commonly, a 45o or −45o layer is placed in the
outermost part of the laminate.

5. Contiguity. The maximum grouping of equally oriented layers is limited

in order to minimise interlaminar stresses and ensure a homogeneous
stress distribution.

6. Grouping. Grouping +θ and −θ layers might be applied to reduce

bending-twist coupling.

7. Disorientation. The difference in fibre orientation between two adjacent

layers might be limited to 45o to minimise inter-laminar shear effects.

It should be noted that although grouping and disorientation design rules

may both be advisable for a structure, their requirements are contradictory
and hence cannot be applied simultaneously.
The manufacturing rules consider the transitioning between laminates
placed in neighbouring patches.

6
1. Continuity/blending. Continuity, also commonly referred to as blend-
ing, ensures manufacturability and structural integrity of the laminated
composite. Various definitions of blending are available in the litera-
ture. Adams et al. [1] used two simplified blending definitions, i.e.
outward (Fig. 1a) and inward (Fig. 1b) blending, in which only outer-
most or innermost plies, respectively, are allowed to be dropped. A ply
drop indicates that a ply stops and does not continue on to the next
patch of the structure. Campen et al. [8] introduced generalised (Fig.
1c and Fig. 1d) and relaxed generalised (Fig. 1e) blending. These two
definitions allow for significantly greater design freedom since they do
not limit the ply drop position to an outer or inner ply. Generalised
blending requires all plies in a thinner patch to continue in the adjacent
thicker panels, while in relaxed generalised blending two neighbouring
panels are considered to be blended if no dropped edges are in physi-
cal contact. In the current work, generalised blending is used instead
of the relaxed generalised blending, as the latter leads to high stress
concentrations.

2. Maximum dropping. The maximum number of plies dropped in transi-

tions between neighbouring laminates is limited to ensure smooth load
distribution throughout the structure.

3. External covering ply. At least one of the outer plies in a laminate is

not allowed to be dropped in order to ensure structural integrity.

4. Internal covering ply. Not more than a certain number of consecu-

tive plies are allowed to be dropped simultaneously so that potential

7
 delamination initiation zones are avoided.

Other manufacturing rules concerning the exact transitions within the taper
area are not considered, because this level of detail is not in the scope of the
current work. It should be mentioned that the taper area refers to the zone
formed by the plies of two neighbouring patches merging into each other.
The full potential of the generalised blending definition is exploited in the
current work (Fig. 1d), in a sense that the set of plies used in the thinnest
patch must not necessarily extend to the rest of the patches in the struc-
ture. This is not the case for other implementations of generalised blending,
such as the Stacking Sequence Tables (SSTs) [21], the Global Shared Layer
Blending [23] or the modified Shepard’s method [33], in which plies present
in the thinnest patch must be continued across all other patches and not only
the neighbouring ones (Fig. 1c). However, it should be noted that recent
studies using the Ply Drop Sequence concept [49], multiple SSTs [51] and a
general resolution scheme [38] have managed to eliminate some or all of the
aforementioned drawbacks, offering increased design freedom.

3. Stacking sequence retrieval

In this study, a two-stage optimisation process for the detailed sizing of

aerospace composite structures is investigated. First, a weight minimisation
of the structure is performed using a gradient-based optimisation algorithm,
the result of which is a continuous thickness and stiffness distribution. A
given structure is manually discretised into patches, which represent areas
that will be considered as individual laminates during the optimisation. This
gradient-based optimisation is part of the multidisciplinary optimisation tool,

8
 (a) Outward blending (b) Inward blending

(c) Generalised blending example 1 (d) Generalised blending example 2

(e) Relaxed generalised blending

Figure 1: Blending examples between 3 patches.

referred to as LAGRANGE [41], developed in-house at Airbus. In the second

stage of the process, which is the focus of this work, the stacking and blending
optimisation problem is formulated as two different MILP models, which are
solved using Gurobi [17]. The resultant stacking sequence, fulfils all design
and manufacturing rules by matching the optimal stiffness characteristics
calculated during the first optimisation stage. These stiffness characteristics
are provided to the second stage of the optimisation in the form of lamination
parameters.

9
3.1. Lamination parameters

Lamination parameters were first introduced by Tsai & Pagano [47].

They can be used to decouple the stacking-sequence-dependent part from
the material-dependent part of a laminate’s stiffness matrix. In the general
case of an anisotropic laminate, 12 lamination parameters and 5 material
parameters fully define the stiffness matrix. The extensional stiffness A,
bending extension coupling stiffness B and bending stiffness D matrices are
formulated as:

   
 A
 11 
 
 1 ξ1A ξ3A 0 0  





 



 U1 


A 1 −ξ1A ξ3A 0 0 

 
  
 


 22 
   
 
U

 
   

 2
   
−ξ3A
   

 A12 
 0 0 1 0  

= h  U3 (1)

A66  0 0 −ξ3A 0 1

    

 
   
 







 

 U4



A  0 ξ2A /2 ξ4A 0 0 
   
 


 16 
   






   U5 
 
A  0 ξ2A /2 −ξ4A 0 0
 
26
   


 B11 

 0 ξ1B ξ3B 0 0  







 U1 
 

B 0 −ξ1B ξ3B 0 0

 
  
 

 22 
   

U

 
   

   2 
−ξ3B
    
B12  2 0 0 0 0
h 
   
 

=   U3 (2)


 B66 

 4 0 0 −ξ3B 0 0

 
 

   







 

U
 
 4


B 0 ξ2B /2 ξ4B 0 0 

 
   


 16 
   
 





   U5 

B  0 ξ2B /2 −ξ4B 0 0
 
26

10
    


 D11


 1 ξ1D ξ3D 0 0  





 



 U1 


D 1 −ξ1D ξ3D 0 0 

 
  
 

 22 
    

U

 
   

   2
−ξ3D
    
D12  3 0 0 1 0
h 
   
 

=   U3 (3)

D 
 12 
0 0 −ξ3D 0 1
  
 66 

   
  






  U4 


D  0 ξ2D /2 ξ4D 0 0 
   
 


 16 
   






   U5 
 
D  0 ξ2D /2 −ξ4D 0 0
 
26

In the above equations, U denotes the material constants and h the thickness
of the laminate. The lamination parameters are defined as:

z
A 1 Z 2i
ξ[1,2,3,4] = [cos(2θ), sin(2θ), cos(4θ), sin(4θ)]dz (4)
h − z2i

z
B 4 Z 2i
ξ[1,2,3,4] = 2 z [cos(2θ), sin(2θ), cos(4θ), sin(4θ)]zdz (5)
h − 2i

z
D 12 Z 2i
ξ[1,2,3,4] = 3 z [cos(2θ), sin(2θ), cos(4θ), sin(4θ)]z 2 dz (6)
h − 2i
where − z2i and zi
2
stand for the distance of the bottom and top surface of the
ith ply with respect to the midplane of the laminate.

3.2. Mixed Integer Linear Programming (MILP) formulation

The input for the stacking sequence calculation is a simple geometric

definition of the patches, together with the optimal number of the total lay-
ers and lamination parameters for each patch. The MILP formulation of
this problem is discussed in this section. Two distinct formulations of the
blended stacking sequence optimisation, namely explicit and implicit, have
been developed. As far as the design rules are concerned, their expression

11
does not fundamentally differ between the two formulations. The distinctive
difference between the two is how blending and ply drops are modelled. Nev-
ertheless, both the explicit and implicit formulation make use of the same
principles of generalised blending described in Section 2. Therefore, the re-
sulting structure from each formulation is expected to be the same if the
same input and constraints are used for each optimisation.
In brief, the explicit formulation considers all patches having the same
number of layers, with that number being equal to the number of layers in
the thickest patch of the structure. Since each patch has a different optimum
number of layers, some of the patches will consist of plies that correspond to
ply-voids as displayed in Fig. 2a. The implicit formulation demonstrated in
Fig. 2b takes into account the fact that each patch has a different, known
number of optimum layers, and defines the appropriate number of layers for
each individual patch. The benefit of the explicit formulation is that the
blending of the structure is rather straightforward, as continuity is ensured
by placing plies of the same orientation one next to the other. The disadvan-
tage of this formulation is that the locations of the ply drops are not known
and therefore there are many possible ‘positions’ one specific ply could have
within the laminate. Therefore, since the order at which plies will appear is
not known, the calculation of the lamination parameters for such a stack is
cumbersome. The advantage of the implicit formulation is that the calcula-
tion of lamination parameters is straightforward as the position of each ply is
clearly defined. The downside of the method is the fact that formulating the
manufacturing rules and most importantly the continuity of the structure is
more challenging.

12
(a) Schematic representation of the explicit formu- (b) Schematic representation of the implicit formu-
lation. lation.

Figure 2: An illustrative example of the same blended stack for two neighbouring patches
as represented by the explicit and implicit formulation. Each colour represents a different
fibre orientation, ply-voids are illustrated using hatched plies and arrows indicate ply
continuity.

3.2.1. Explicit formulation

Various design variables are introduced in order to formulate the blended
stacking sequence problem. Each set of design variables, i.e. x, y, b, c, d, z, u,
g, f contains many individual members. The number of members for each set
depends on the usage of the design variable and different indexes are used to
distinguish between the individual members of each set of design variables.
Index i ∈ {1, 2, ..., I} is used to denote the exact layer in a specific patch
j ∈ {1, 2, ..., J}. It is important to mention that the range of i is constant for
all patches j, even if most of the patches are expected to have less layers than
I which is the number of layers for the thickest patch in the structure. Index
θ ∈ {1, 2, ..., Θ} denotes the different available fibre orientations. Finally,
k ∈ {1, 2, 3, 4} distinguishes between the lamination parameter which is of
interest. A summary of the indexes employed to differentiate between the
members of each design variable set is given in Table 1.
The objective is to minimise the deviation from the optimal values of the

13
Table 1: Summary of indexes used to denote the various design variables in the explicit
formulation.

Name Description
i Layer
j Patch
θ Fibre orientation
k Lamination parameter

lamination parameters as formulated below:

J
wkA,B,D |ξkj
A,B,D A,B,D
X
min − (ξkj )optimal | ∀k (7)
j=1
A,B,D
where ξkj ∈ [−1, 1] are the lamination parameters calculated for a spe-
A,B,D
cific stack, whereas, (ξkj )optimal are the optimal parameters delivered by
the gradient-based optimisation. Weight coefficients wkA,B,D may be used
to emphasise specific parts of the objective function. Two new sets of de-
sign variables need to be introduced to assist with the linear formulation of
A,B,D A,B,D
the objective function. The first one is gkj ∈ [0, 1], defined as gkj =
A,B,D A,B,D
|ξkj − (ξkj )optimal |. The objective function is not linear in its current
form, because of the absolute value, therefore, a simple linearisation needs
A,B,D
to be performed. The additional set of design variables fkj ∈ {0, 1}, is
also introduced to enable the linearisation described by the following set of
equations

14
 A,B,D A,B,D A,B,D
ξkj − (ξkj )optimal ≤ fkj ∀j, k
A,B,D A,B,D A,B,D
(ξkj )optimal − ξkj ≤ 1 − fkj ∀j, k
A,B,D A,B,D A,B,D
ξkj − (ξkj )optimal ≤ gkj ∀j, k
(8)
A,B,D A,B,D A,B,D
(ξkj )optimal − ξkj ≤ gkj ∀j, k
A,B,D A,B,D A,B,D A,B,D
gkj ≤ ξkj − (ξkj )optimal + 2(1 − fkj ) ∀j, k
A,B,D A,B,D A,B,D A,B,D
gkj ≤ (ξkj )optimal − ξkj + 2fkj ∀j, k
To make the formulation of the problem possible, the binary design vari-
ables xijθ ∈ {0, 1} defining the orientation of a specific layer within a certain
patch are introduced. More specifically, every layer consists of Θ design vari-
ables, each one representing whether the corresponding fibre orientation is
used or not. Each layer may only have one orientation, or none, in the case
the ply is dropped. Therefore, the following feasibility constraints need to
be defined.

Θ
X
xijθ ≤ 1 ∀i, j (9)
θ=1

Additionally, the number of layers in a patch must be equal to the optimum

discretised thickness (nj ) which originates from the continuous optimisation
and remains fixed throughout the discrete optimisation.

I X
Θ
xijθ = nj
X
∀j (10)
i=1 θ=1

In many cases, laminates are designed to be symmetric. Whenever sym-

metry is required, the following constraints need to be incorporated in the
optimisation model.

15
 xijθ = x(I−i+1)jθ ∀i{1, 2, ..., I/2}, j, θ (11)

Balanced laminates can be extracted from the optimisation by including the

following constraints.

I
X I
X
xijθ1 = xijθ2 ∀i, j, θ1 = −θ2 , {0, 90} (12)
i=1 i=1

The minimum percentage (p) design rule is implemented as:

I
xijθ ≥ pnj
X
∀j, θ (13)
i=1

The damage tolerance rule is demonstrated below for the case that the outer-
most ply is equal to 45o . This layer also has to be continuous over all patches
according to the external covering ply manufacturing rule. The following
constraints fulfill both requirements.

x1jθ1 = 1 ∀j, θ1 = 45 (14)

The grouping design rule for the case of symmetric laminates is formulated
as:

xijθ1 ≤ xijθ2 ∀i ∈ {1, 2, ..., I/2}, j, θ1 = −θ2 , {0, 90} (15)

For non symmetric laminates, the i range needs to be modified accordingly.

The contiguity and disorientation constraints cannot be easily implemented
in the explicit formulation and are hence only going to be presented for the
implicit formulation in the next section.

16
 In order to define the continuity manufacturing constraint between neigh-
bouring patches, further design variables need to be introduced. A set of
them is yij ∈ {0, 1}, defining the presence or absence of a specific ply within
a patch. The continuity constraint is expressed as:

xij2 θ = xij1 θ yij2 ∀i, j, θ. (16)

Because the constraint presented above is non-linear, it needs to be linearised

in order to be handled by a MILP solver. The linearisation constraints for
the product of two binary variables [12] are now introduced.

xij2 θ ≤ xij1 θ ∀i, j, θ

xij2 θ ≤ yij2 ∀i, j, θ (17)

xij2 θ ≥ xij1 θ + yij2 − 1 ∀i, j, θ

In the above equation, patches j1 and j2 denote two neighbouring patches
within the structure. An additional constraint which is not necessary, but
has been observed to assist the convergence of the optimisation, is used to
define the number of ply drops per patch based on the known number of
layers for each patch.

I
yij = nj
X
∀j (18)
i=1

Before moving on to the definition of the rest of the constraints, it is worth

formulating the lamination parameter values in a way that serves the specific
approach to the problem. The formal definition of lamination parameters has
already been presented in the previous section (equations 4-6).

17
 I X
4
A,B,D
skθ aA,B,D
X
ξkj = ij xijθ ∀j, k (19)
i=1 θ

Coefficients skθ take into consideration the influence of the trigonometric

terms, depending on which θ design variable is used out of xijθ . Additionally,
coefficients aA,B,D
ij depend on whether A, B or D lamination parameters are
considered and on the position of the ply within the laminate. Due to the fact
that ply drops are included in the problem formulation as design variables,
the distance of each ply from the mid-surface of the laminate is not known a
priori. Another set of design variables bij ∈ {0, 1, 2, ..., M1 } is used for that
purpose, where the upper bound is M1 = i + 1 The definition of bij is as
follows:

i
X
bij = ymj ∀i, j (20)
m=1

In summary, the aforementioned coefficients aA,B,D

ij depend on the bij design
variables and are defined as:

1
aA
ij = ∀i, j (21)
nj

4 2nj + 2
aB
ij = − b ij + ∀i, j (22)
(nj )2 (nj )2

12 12nj + 12 3(nj )3 + 6nj + 4

aD
ij
2
= j 3 (bij ) − bij + ∀i, j (23)
(n ) (nj )3 (nj )3
The term (bij )2 in equation 23 is not linear and therefore needs to be lin-
earised. Due to the fact that the linearisation of a product of two bounded

18
integer variables is cumbersome and involves piecewise linearisations, a trick
is employed to calculate cij ∈ {0, 1, 2, ..., (M1 )2 } as:

i
(bij )2 = cij = (
X
2dmj ) − bij ∀i, j (24)
m=1

In the above equation, a new set of design variables dij ∈ {0, 1, 2, ..., M1 }
was introduced. These are defined as dij = bij yij and once again need to be
linearised. The linearisation of the product of two variables, one of which
is binary and the other an integer with an upper bound of M1 and a lower
bound of 0, is defined by the following four inequalities [16].

dij ≤ M1 yij ∀i, j

dij ≤ bij ∀i, j

(25)
dij ≥ bij − (1 − yij )M1 ∀i, j

dij ≥ 0 ∀i, j

Finally, due to the fact that the product of aB,D

ij xijθ appears in equation 19,

two new sets of design variables uijθ ∈ {0, 1, 2, ..., M1 }, zijθ ∈ {0, 1, 2, ..., (M1 )2 }
need to be introduced and are defined as uijθ = bij xijθ and zijθ = cij xijθ . The
linearisation of these design variables follows the principles of equation 25.
The lamination parameter values can eventually be expressed as:

I X
4
B
X 4 2nj + 2
(ξkj )= skθ (− uijθ + xijθ ) ∀k, j (26)
i=1 θ=1 (nj )2 (nj )2

I X
4
D
X 12 12nj + 12 3(nj )3 + 6nj + 4
(ξkj )= skθ ( zijθ − uijθ + xijθ ) ∀k, j
i=1 θ=1 (nj )3 (nj )3 (nj )3
(27)

19
 The internal covering ply manufacturing constraint is formulated using
the bij design variables as:

b(i+K)j1 − b(i+K)j2 − (bij1 − bij1 ) ≤ K − 1 (28)

where j1 and j2 denote two neighbouring patches and K the number of

maximum consecutive ply drops. The maximum dropping manufacturing
rule is explicitly satisfied due to the fixed number of layers provided as an
input to the optimisation and therefore does not require further constraints.

3.2.2. Implicit formulation

Similarly to the explicit formulation, various design variables need to
be introduced. Each set of design variables, i.e. x, r, t, g, f , contains many
individual members. Once again, index i ∈ {1, 2, ..., Ij } is used to denote
the exact layer in a specific patch j ∈ {1, 2, ..., J}. The range of i is not
constant for all patches j, but rather takes a maximum value of Ij according
to the patch of interest. Index h ∈ {1, 2, ..., H} accounts for all interfaces
between neighbouring patches and index p ∈ {1, 2, ..., Ph } takes into account
the different blending combinations between neighbouring patches. A more
detailed explanation of these two indexes will be presented in the following
paragraphs. Finally, indices θ and k follow the definition of the explicit
formulation and denote fibre orientations and lamination parameter which is
of interest respectively.
The formulation of the objective function is exactly the same as the one
presented in equation 7 and the same goes for the set of design variables
D D
gkj ∈ [0, 1] and fkj ∈ {0, 1}. The definition of the xijθ ∈ {0, 1} set of
design variables follows the same logic with the exception that in the implicit

20
formulation, each patch is assigned with as many Ij layers as needed. All
design rules (equations 11-15) are formulated similarly with the only thing
changing being the number of maximum layers per patch Ij . An additional
minor alternation applies for equation 9 in which the inequality sign needs
to be replaced by the equality one, since the design variables xijθ in this
approach cannot correspond to a ply drop.
Additionally, the two design rules which have not been formulated for
the explicit approach can now be implemented. The contiguity constraint
limits the number of consecutive layers having the same fibre orientation to
a maximum of N . In a mathematical formulation this is expressed as:

xijθ + x(i+1)jθ + . . . + x(i+N )jθ ≤ N ∀i ∈ {1, 2, ..., Ij − N }, j, θ (29)

The disorientation constraint is expressed as:

xijθ1 + x(i+1)jθ2 ≤ 1 ∀i, j, |θ1 − θ2 | ≥ 45 (30)

The calculation of lamination parameter values is now trivial. Equation

19 is once again used except this time, coefficients aB,D
ij can be easily pre-
calculated due to the fact that ply drops have already been taken into account
by using an exact number (Ij ) of xijθ design variables for each patch.
The part of this second approach that becomes cumbersome is formulating
the continuity between neighbouring patches. A new set of binary design
variables tihp ∈ {0, 1} is introduced and is used to regulate and keep track
of whether a specific ply within a patch will be blended with one of the p
possible neighbouring plies. The number of possibilities p per ply depends on

21
the interface h examined and can easily be determined during the setup of
the optimisation problem. A representation of the different plies that could
blend between two patches is given in Fig. 3a. More specifically, for any of
the interfaces h, the maximum number of possibilities per ply (Ph ) can be
calculated as:

Ph = |nj1 − nj2 |+1 (31)

In the above equation, patches j1 and j2 denote the two patches which con-
stitute an interface h, with j1 being the thickest of the two, in case there
is a difference in thickness between them. The number of interfaces H is
easily calculated prior to setting up the optimisation problem based on the
arrangement of the patches. In Fig. 3b, an example of the interfaces between
a specific patch geometry is given.

(a) Knowing the number of ply drops between two neigh- (b) The interfaces between patches are
bouring patches, the different blending possibilities can be marked with red ticks for the given patch
determined for each ply independently. geometry.

Figure 3: Illustrative examples showing the usage of indexes p and h.

Finally, a new set of design variables rihθp ∈ {0, 1} is defined as rihθp =

x(i+p−1)j1 θ tihp . Because the definition of rihθp is non-linear, a standard lin-
earisation needs to be performed. This linearisation of the product of two

22
binary variables is similar to the one presented in equation 17.
The continuity manufacturing rule is then defined as:

Ph
X
xij2 θ = rihθp (32)
p

This constraint alone is not enough to ensure manufacturable composite

patches. Three further sets of constraints need to be defined to ensure the
blending between individual layers is achieved in a meaningful manner. First
of all, a left hand-side ply must blend with only one of the right hand-
side plies. References to left and right hand-side plies follow the example
illustrated in Fig. 3a.

Ph
X
t(i−p+1)h(Ph −p+1) ≤ 1 ∀i, h (33)
p

Secondly, a right hand-side ply must blend with exactly one of the left hand-
side plies.

Ph
X
tihp = 1 ∀i, h (34)
p

Thirdly, in order to avoid unmanufacturable designs that lead to plies inter-

secting (e.g. right hand-side 1 ply matched with left hand-side 2 and left
hand-side 2 matched with right hand-side 1) the following constraints are
defined.

Ph
X Ph
X
t(i−1)hp (Ph − p + 1) ≤ tihp (Ph − p + 1) ∀i, h (35)
p p

Finally, the internal covering ply rule is formulated as:

23
 j1X
+N X
Ph
t(i−p+1)h(Ph −p+1) ≤ 1 ∀i, h (36)
j1 p

The presented set of equations is implemented using the Python interface of

Gurobi.

3.3. Problem decomposition strategy

Both formulations presented in the previous sections solve the stacking

and blending optimisation problem considering all patches and constraints
simultaneously. In this section, a decomposition strategy is presented for the
implicit formulation of the optimisation problem. The reason for choosing
implicit instead of explicit is because the strategy can be applied with greater
freedom to any patch in the structure. This is due to the lack of ply-voids in
the implicit formulation.
The flowchart of the decomposition strategy is presented in Fig. 4. The
user needs to define a path or a set of paths containing all of the patches
in the structure. Starting from a specific patch, one neighbouring patch is
added at a time, until all patches have been included within the path. For
each of the paths defined by the user, the first sub-problem to be optimised
only includes the first two patches in the path. These two patches are op-
timised only considering design rules applied on them and manufacturing
rules applied on their interface. Once the optimisation of this sub-problem
is converged, the next neighbouring patch defined in the path is included
in the optimisation sub-problem. This time, the stacking sequences of the
previously optimised patches are fixed by using equality constraints. If the
optimisation is carried out successfully, the patch which was added last is

24
fixed and the next neighbouring patch is added in the structure along with
all the relevant design and manufacturing rules. In cases where the newly
added patch borders two or more of the pre-existing patches whose stacking
sequence is fixed, the optimisation problem might become infeasible. Conse-
quently, the fixation constraints applied to the patch being the latest addi-
tion, are removed. This operation is repeated until a feasible solution can be
retrieved by the optimisation. Once all of the patches have been introduced
in the optimisation sub-problem, the procedure is applied for the next user-
defined path. The decomposition method based on the path which resulted
to the stacking sequence with the lowest objective function value, is used to
initialise the complete optimisation problem which considers all constraints
simultaneously, without the stacking sequence of any of the patches being
fixed.
Even if all possible paths are considered in this decomposition strategy,
there is no guarantee that the optimal solution will be obtained in the end.
A way to improve the quality of this local minimum would be to add two or
more patches in each sub-problem at a time, instead of one. This increased
overlapping allows the optimisation process to overcome the issue of getting
stuck at local minima, since stiffness requirements of more ‘free’ patches
would be considered in the objective function of each sub-problem. The
drawback of this strategy would be the increased computational time for
each sub-problem, since the number of plies free to change and the number
of interfaces between the patches increases. It should also be noted that
the quality of the solution achieved by the decomposition strategy highly
depends on the target stiffnesses provided by the gradient-based optimisation

25
 Problem decomposition

Define paths

Optimise 2
neighboring patches

Fix their stacking

sequence

Add next neighboring

patch

Optimise sub-problem
No
No
No Remove fixation
Optimisation
constraints in last fixed
feasible
patch.
Yes

All patches added Perform full

optimisation
Yes

Yes
All paths explored Initialise solution

Figure 4: Flowchart of the optimisation using decomposition developed for the implicit
formulation.

algorithm applied in the first stage. The accuracy of the formulation of the
composite rules, especially the formulation of the blending rule in the first
stage, highly impacts how close the resultant solution obtained from the
decomposition strategy will be to the the global optimum.

26
4. Numerical case studies

Initially, the explicit and implicit formulations are studied using some
illustrative cases and then a well studied benchmark problem from the scien-
tific literature is used to assess the proposed discrete optimisation approach.
Moreover, the MILP formulations are also compared against two previous
approaches.

4.1. Explicit and implicit formulations

Both explicit and implicit formulations for discrete optimisation were able
to produce fully blended composite designs satisfying all specified design and
blending rules. In this section, we first present the results applying the ap-
proaches to some methodology demonstration problems. More specifically,
blended stacking sequences that fulfil a prescribed set of design and manufac-
turing rules are manually defined and the accordant lamination parameters,
which are used as target values during the optimisation, are calculated for
each individual patch. Therefore the optimal objective function value for this
set of problem instances is known a priori and is equal to zero. All design
and manufacturing rules described in Section 2 are used, with the exception
of contiguity, and disorientation which is omitted because grouping is used
instead. The fibre orientations for the demonstration cases are limited to
0o , 90o , ±45o . The demonstration problem instances are comprised of a dif-
ferent number of patches (2 or 4), maximum number of layers in the thickest
patch (40 or 80) and number of ply drops between adjacent patches (4 or
8). In all demonstration cases, patches are placed sequentially one after the
other.

27
Table 2: Comparison of the MILP size for selected demo problems between the explicit
and implicit formulation.

Description Explicit formulation Implicit formulation

Patches Layers Ply drops Constr. Design var. Non-zeros (%) Constr. Design var. Non-zeros (%)
2 40 4 868 368 1.3 911 393 1.1
2 40 8 868 368 1.3 1146 492 0.9
2 80 8 1852 772 0.6 2745 1163 0.3
4 80 8 5071 1990 0.2 6823 2789 0.1

In Table 2, the final MILP problem size in terms of the number of de-
sign variables, constraints and percentage of non-zero elements in the MILP
optimisation matrix is presented for a subset of the cases. The final MILP
size does not directly correspond to the size one would theoretically calculate
given the problem definition, because Gurobi, like other commercial solvers,
applies pre-solve algorithms that might reduce the size of the problem. It can
be observed that the size grows significantly for large instances, as expected.
Moreover, using the implicit formulation results in slightly larger problem
instances. Finally, it is worth noting that an increased number of ply drops
increases the problem size of the implicit formulation, while the size of the
explicit formulation stays the same. This is due to the fact that the design
variables and therefore the constraints of the implicit formulation depend on
the number of blending possibilities between neighbouring patches, which is
analogous to the number of ply drops.
In Fig. 5, the runtimes of the two approaches for all the different demon-
stration problem instances are presented. The individual runtimes are de-
picted using half circles of logarithmically scaled radiuses. More specifically,
the maximum radius in Fig. 5 represents a runtime of 442s while the smallest

28
one 0.8s. All runs have been performed on a PC with an Intel Core i5-8250U
@1.60GHz (4 cores, 8 threads) CPU. The algorithm is stopped if an objective
value less than 10−2 ×(number of patches) is found.

Explicit run time (log scale)

Implicit run time (log scale)

Number of ply drops

6

80 s
yer
60 f la
2 3 4 40 b er o
Number of patches N um

Figure 5: Illustration of run-times for the implicit and explicit formulations for the initial
set of problem instances as the number of patches, layers and ply drops change.

Results presented in Fig. 5, highlight the expected increase in run-time

as the number of layers and patches increase. The implicit approach is better
than the explicit approach for small sized problems, however, as the problem
size increases the explicit formulation is either comparable or better than
the implicit one in terms of time, as expected. Finally, the runtime of the
implicit formulation strongly depends on the number of ply drops between
adjacent patches, with more ply drops resulting in higher runtimes, while the
explicit formulation does not necessarily depend on the number of ply drops.
This can be explained by the fact that more design variables and constraints
need to be introduced in the implicit formulation to account for more ply

29
drops.
Fig. 6 illustrates the 18 panel horseshoe benchmark problem. Nx and
Ny force resultants are given in lbf/in (×175.1 to convert into N/m). The
modulus of elasticity across the fibre direction is E1 = 141GPa and E2 =
9.03GPa across the transverse direction, the shear modulus is G12 = 4.27GPa
and the principal Poisson’s ratio is ν12 = 0.32. The ply thickness is 0.191mm.
18 in. 20 in.

1 2 3 4 5

12 in.
Nx = -270 Nx = -250 Nx = -210
Ny = -325 Ny = -200 Ny = -100
24 in.

Nx = -700 Nx = -375
Ny = -400 Ny = -360
6 7 8
Nx = -305 Nx = -290 Nx = -600
Ny = -360 Ny = -195 Ny = -480

9 10

Nx = -1100 Nx = -900
Ny = -600 Ny = -400

11 12 13 14 15
Nx = -330 Nx = -190 Nx = -300
Ny = -330 Ny = -205 Ny = -610
Nx = -375 Nx = -400
Ny = -525 Ny = -320
16 17 18
Nx = -815 Nx = -320 Nx = -300
Ny = -1000 Ny = -180 Ny = -410

Figure 6: Definition of the 18 panel horseshoe blending problem [43].

In Fig. 7 the convergence plots of the two different MILP formulations

are given for the horseshoe problem [43]. The ‘best known objective’ in Fig.
7 corresponds to the smallest objective function value of a solution satisfying
the constraints of the optimisation, while ‘best possible objective’ refers to
the minimum discovered bound for the value of the objective function which

30
is known due to the the branch and bound based algorithm being employed.
Optimality of the solution is proven when the value of the best possible
objective is equal to that of the best known objective. The optimisation is
stopped after 600s because proving optimality becomes time consuming. As
observed earlier, the implicit formulation outperforms the explicit one, since
the maximum number of layers is rather small for this problem, as shown in
Table 3. This is not necessarily the case for larger problems for which the
explicit approach can converge to better solutions in less computational time.
The disadvantage of the explicit formulation though is the lack of providing
a good theoretical bound for the objective (Fig. 7), which in turn results in
bigger computational times until proving convergence.

Explicit and implicit formulation, convergence comparison

Implicit formulation, best possible objective
12
Implicit formulation, best known objective
Explicit formulation, best possible objective
10 Explicit formulation, best known objective
Objective function value

0 100 200 300 400 500 600

Time (s)

Figure 7: Convergence comparison between the explicit and implicit formulation for the
18 panel horseshoe problem.

Next, a comparison between the convergence of the decomposition strat-

egy and the non-decomposed implicit formulation is performed for the 18

31
 Table 3: Number of layers for the patches of the 18-panel horseshoe problem.

Patch 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Layers 34 28 22 20 16 22 20 26 38 36 30 28 22 20 26 32 20 22

panel horseshoe problem. Three different paths, representing patch sequences,

are defined to test the decomposition strategy as shown in Table 4. The first
path begins from panel 5 in the top right of the horseshoe pattern and moves
counterclockwise to patch number 18 in the bottom right following a specific
pattern. The second path is built up starting from the thinnest patch in
the structure and by gradually adding the thinnest possible neighbour each
time. The third path is constructed by starting from the thickest patch in
the structure and by choosing the thickest neighbour to add each time.

Table 4: Paths containing the sequence of patches used for the optimisation using decom-
position of the 18 panel horseshoe problem.

Path Patch sequence

1 [5,8,7,4,3,6,2,1,10,9,11,12,13,14,15,16,17,18]
2 [5,4,3,7,8,6,2,1,10,9,12,11,13,14,17,18,15,16]
3 [9,10,1,2,3,6,4,7,8,5,11,12,16,13,14,17,15,18]

Fig. 8 illustrates the progress plot for the objective value using the ap-
proaches with and without decomposition, where the three triangles mark the
objective value of each local minimum discovered by using the three different
paths. The one used to initialise the optimisation without the decomposition
is that of the best found solution which corresponds to the second path. For
the specific manufacturing rules used, the average run-time of each individ-

32
ual run for the different paths is approximately 10 seconds, including the
overhead of setting up each of the many sub-problems.

Decomposition strategy convergence

No decomposition
13 Decomposition, 3 paths
Path 1
12 Path 2
Objective function value

Path 3
11

10

0 50 100 150 200 250 300

Time (s)

Figure 8: Convergence comparison between the implicit formulation and the decomposition
strategy for the 18 panel horseshoe problem.

The fact that the path which resulted to the best solution of the optimisa-
tion with the decomposition is the one starting from the thinnest patch and
moving to the thinnest neighbour is not incidental. This can be attributed
to the fact that in thinner patches, the placement of a layer influences the
stiffness more compared to a thicker panel. Therefore, allowing the thinner
patches to be optimised first can prevent the decomposition from getting
stuck at a local minimum favouring some of the thickest patches. Finally,
as for the actual convergence, no significant speed up is observed for the ex-
amined case. The optimisation with no decomposition actually converges to
what appears to be a global minimum faster than the optimisation approach
using decomposition. However, the benefit of using decomposition becomes

33
more evident as the size of a problem grows which is the case for realistic
aircraft structures.
To demonstrate this effect, the loads applied to the original formulation
of the horseshoe panel problem are multiplied by a factor of 8. The material
and dimensions of the panels remain the same. After performing the first
optimisation stage again, this increase in loads leads to a doubling of the pan-
els’ thicknesses. Fig. 9 depicts the improvement of the value of the objective
function with respect to time, for the modified horseshoe problem when ap-
plying the optimisation with and without the decomposition technique. For
this bigger optimisation problem, applying the optimisation with the decom-
position yields a much better solution in a considerably shorter time frame.
A second observation is that each of the 3 different paths now requires an
average of approximately 30 seconds to optimise compared to 10 in the orig-
inal problem definition. This is expected, since the number of layers have
doubled and the size of the optimisation problem has increased accordingly.
As the number of layers increases, the design space of the laminate becomes
wider. A wider design space also means that the current thinnest patch in
the structure does not influence the quality of the decomposition that much.
Indeed, in the this modified problem, the path which results to the lowest
objective function is the third one which starts from the thickest patch in the
structure. What is more, the deviation between the objective function val-
ues of the three paths is now considerably smaller compared to the original.
Finally, because the layers increase, the potential to match the same target
stiffness increases which leads to a lower global optimum for this modified
problem.

34
 Decomposition strategy convergence modified problem

12

11
Objective function value

10 No decomposition
Decomposition, 3 paths
9
Path 1
Path 2
8
Path 3
7

0 200 400 600 800 1000 1200

Time (s)

Figure 9: Convergence comparison between the implicit formulation and the decomposi-
tion strategy for the modified 18 panel horseshoe problem in which the applied loads are
multiplied by a factor of 8.

Overall, both the explicit and implicit formulations provide stacking se-
quences that are fully compliant with respect to the specified design and
manufacturing rules. Based on the observations and experience gained up
to now, the explicit formulation can outperform the implicit one for large
industrial problems. However, the explicit formulation still lacks the im-
plementation of the disorientation and contiguity constraint as well as the
decomposition technique.

4.2. Comparison against relevant work

Liu et al. [29] introduced a logic-based method which is a combination of

the branch and bound and global layerwise techniques, developed to retrieve
stacking sequences that optimally match some given lamination parameters

35
for single patches. Table 5 compares the results obtained from the presented
methodology against the published results for various problems introduced
by Liu et al. [29]. Each example uses a different set of design rules when
matching the lamination parameters. In Liu et al. [29] two runtimes are
provided for each instance; one is the time when the optimum solution (t. to
opt.) is found and the other one is the time (Total t.) until the search of
the tree is completed and hence optimality is proven. The time shown for
the implicit MILP formulation of this work is the time spent until optimality
is proven. In the overall, using a mathematical programming algorithm to
achieve the same task is shown to outperform the logic-based method. This
is due to the fact that a search of the entire tree is not necessary to prove
optimality and because more sophisticated techniques such as presolving,
cutting planes and heuristics are used to assist the optimisation convergence.
It should also be noted that for example 1 in Table 5, an objective function
with a lower value was calculated using Gurobi. The provided runtimes for
the current work have been produced using 1 thread on a 1.6GHz computer
compared to a 4GHz computer used in the work of Liu et al. [29].
Finally, the open source genetic algorithm developed by Vicente [48] is
used to optimise the horseshoe problem described in the previous section.
The genetic algorithm uses the Stacking Sequence Table encoding. All de-
sign and manufacturing rules are used, except for the grouping rule, because
the disorientation design rule is activated instead. The settings used for the
runs of the genetic algorithm are summarised in Table 6. The convergence
history of various genetic algorithm runs and of the implicit formulation are
presented in Fig. 10. All algorithm runs have been performed on the same

36
Table 5: Single patch optimisation results, comparison against the work of Liu et al. [29].

Results from [29] MILP Implicit formulation

Example t. to opt.(s) Total t.(s) Obj. t.(s) Obj.
1 4.77 3600+ 0.0806 3.83 0.0794
2 0.52 7.10 0.1729 0.13 0.1729
10 3.76 3600+ 0.0892 1.69 0.0892
11 0.23 5.95 0.0984 0.16 0.0984
13 0.19 0.90 0.3828 0.14 0.3828
14 0.97 1.59 0.3776 0.18 0.3776
15 6.40 9.06 0.1120 0.10 0.1120

1.6GHz CPU using only one core. The implicit formulation converges to a
much better solution. This can be attributed to two reasons. First of all,
the blending formulation in this work allows for more design freedom, as
explained in section 2, which leads to a smaller feasible objective function.
Secondly, the mathematical programming algorithm used to solve the prob-
lem formulation, as presented in the current work, can consistently lead to
the optimal value of the objective function, while the genetic algorithm gets
trapped in different local minima, based on the contextual random number
generator seed. However, the genetic algorithm is able to provide decent
feasible solutions in less time.

5. Conclusions

Both formulations can be used in conjunction with mathematical pro-

gramming algorithms and are able to retrieve laminated composite structures

37
 Table 6: Settings used in the genetic algorithm of Vicente [48].

Parameter Value
Population size 200
Maximum number of generations 1000
Elitism percentage 4%
Probability of mutation 15%
Probability of crossover 75%

MILP formulation and genetic algorithm, convergence comparison

20.0
Implicit formulation, best possible objective
17.5 Implicit formulation, best known objective
Genetic algorithm, different runs
15.0
Objective function value

12.5

10.0

7.5

5.0

2.5

0.0
0 100 200 300 400 500 600 700 800 900
Time (s)

Figure 10: Convergence comparison between the implicit formulation and a guide-based
genetic algorithm [48] for the 18 panel horseshoe problem.

which fulfil a set of prescribed design and manufacturing rules. The objec-
tive function of this optimisation involves the minimisation of the absolute
difference between the stiffness of the computed design and a set of target
stiffness characteristics provided by the first stage of the optimisation.
The derived formulations, namely explicit and implicit, mainly differ in

38
the way blending is treated. The explicit formulation offers a much more
direct approach to the representation of blending, which however increases
the complexity of defining the objective as a linear function of the available
design variables. The derivation of blending constraints for the implicit for-
mulation is more complex, however other composite rules and the objective
function can be easily derived. Finally, a decomposition technique has been
developed for the implicit formulation to assist with the discovery of good
local optima in a much shorter time frame.
The results presented show that both formulations are able to tackle a
widely used benchmark in a reasonable time frame. The explicit formula-
tion is able to tackle problems of larger size in a shorter time. However,
the utilisation of the decomposition technique on the implicit formulation is
very promising for applications in industrial composite parts of much larger
scales. Compared to an open-source genetic algorithm, the current formu-
lations lead to better discrete stacks in a given time frame. This can be
partly attributed to the fact that the current formulations offer more design
freedom than most available blending representations and partly to mathe-
matical programming algorithms, offering a more robust convergence towards
the global optimum compared to heuristics. Comparison of the current MILP
formulation against a custom-made Branch and Bound algorithm focusing on
the stacking sequence optimisation of single patches, shows that optimality
of the solution can be proven faster for all examined cases.
The present paper offers two novel formulations of the thoroughly exam-
ined discrete blending and stacking sequence optimisation problem. Future
work may involve the incorporation of manufacturing cost requirements and

39
objectives in the formulation. The explicit formulation might be better suited
for such an addition, but the drawbacks of its implementation must first be
dealt with. Improvements on the quality of the solution retrieved by the de-
composition technique can also be achieved by overlapping more patches for
each sub-problem at an extra computational cost, or by exploring more ad-
vanced techniques of removing fixation constraints is the case a sub-problem
becomes infeasible.

Acknowledgments

This project has received funding from the European Union’s Horizon 2020
research and innovation programme under the Marie Skłodowska-Curie grant
agreement No 764650.

6. Declaration of Interests

The authors declare that they have no known competing financial inter-
ests or personal relationships that could have appeared to influence the work
reported in this paper.

7. Data availability

The data that support the findings of this study are available on request
from the corresponding author.

References

[1] Adams, D. B., Watson, L. T., Gürdal, Z., & Anderson-Cook, C. M.

(2004). Genetic algorithm optimization and blending of composite lam-

40
 inates by locally reducing laminate thickness. Advances in Engineering
Software, 35 , 35 – 43.

[2] Allaire, G., & Delgado, G. (2016). Stacking sequence and shape opti-
mization of laminated composite plates via a level-set method. Journal
of the Mechanics and Physics of Solids, 97 , 168–196.

[3] An, H., Chen, S., & Huang, H. (2018). Stacking sequence optimization
and blending design of laminated composite structures. Structural and
Multidisciplinary Optimization, .

[4] Aymerich, F., & Serra, M. (2008). Optimization of laminate stacking

sequence for maximum buckling load using the ant colony optimization
(aco) metaheuristic. Composites Part A: Applied Science and Manufac-
turing, 39 , 262–272.

[5] Bermell-Garcia, P., Verhagen, W. J., Astwood, S., Krishnamurthy, K.,

Johnson, J. L., Ruiz, D., Scott, G., & Curran, R. (2012). A frame-
work for management of knowledge-based engineering applications as
software services: Enabling personalization and codification. Advanced
Engineering Informatics, 26 , 219 – 230.

[6] Bruyneel, M. (2011). Sfp—a new parameterization based on shape func-

tions for optimal material selection: application to conventional compos-
ite plies. Structural and Multidisciplinary Optimization, 43 , 17–27.

[7] Bruyneel, M., Zein, S., & Grihon, S. (2014). A two-step optimization
approach for the optimal design of composite structures, including geo-

41
 metric non-linear behavior, design rules and manufacturing constraints.
In 4th Aircraft Structural Design Conference.

[8] Campen, J. V., Seresta, O., Abdalla, M., & Gürdal, Z. (2012). General
blending definitions for stacking sequence design of composite laminate
structures. In 49th AIAA/ASME/ASCE/AHS/ASC Structures, Struc-
tural Dynamics, and Materials Conference.

[9] Chang, N., Wang, W., Yang, W., & Wang, J. (2010). Ply stacking
sequence optimization of composite laminate by permutation discrete
particle swarm optimization. Structural and Multidisciplinary Optimiza-
tion, 41 , 179–187.

[10] Fedon, N., Weaver, P. M., Pirrera, A., & Macquart, T. (2021). A method
using beam search to design the lay-ups of composite laminates with
many plies. Composites Part C: Open Access, 4 , 100072.

[11] Fedon, N., Weaver, P. M., Pirrera, A., & Macquart, T. (2021). A repair
algorithm for composite laminates to satisfy lay-up design guidelines.
Composite Structures, 259 , 113448.

[12] Floudas, C. A. (1995). Nonlinear and Mixed-Integer Optimization: Fun-

damental and Applications. New York: Oxford University Press.

[13] Gao, T., Zhang, W., & Duysinx, P. (2012). A bi-value coding param-
eterization scheme for the discrete optimal orientation design of the
composite laminate. International Journal for Numerical Methods in
Engineering, 91 , 98–114.

42
[14] Ghiasi, H., Fayazbakhsh, K., Pasini, D., & Lessard, L. (2010). Opti-
mum stacking sequence design of composite materials part ii: Variable
stiffness design. Composite Structures, 93 , 1–13.

[15] Ghiasi, H., Pasini, D., & Lessard, L. (2009). Optimum stacking se-
quence design of composite materials part i: Constant stiffness design.
Composite Structures, 90 , 1–11.

[16] Glover, F. (1975). Improved linear integer programming formulations of

nonlinear integer problems. Management Science, 22 , 455–460.

[17] Gurobi Optimization, L. (2018). Gurobi Optimizer Reference Manual.

[18] Herencia, J., Weaver, P., & Friswell, M. (2008). Initial sizing optimi-
sation of anisotropic composite panels with t-shaped stiffeners. Thin-
Walled Structures, 46 , 399–412.

[19] IJsselmuiden, S. (2011). Optimal Design of Variable Stiffness Composite

Structures using Lamination Parameters. Ph.D. thesis TU Delft.

[20] IJsselmuiden, S., Abdalla, M., Seresta, O., & Gürdal, Z. (2009). Multi-
step blended stacking sequence design of panel assemblies with buckling
constraints. Composites Part B: Engineering, 40 , 329–336.

[21] Irisarri, F.-X., Lasseigne, A., Leroy, F.-H., & Le Riche, R. (2014). Op-
timal design of laminated composite structures with ply drops using
stacking sequence tables. Composite Structures, 107 , 559–569.

[22] Javidrad, F., Nazari, M., & Javidrad, H. (2018). Optimum stacking

43
 sequence design of laminates using a hybrid pso-sa method. Composite
Structures, 185 , 607–618.

[23] Jing, Z., Fan, X., & Sun, Q. (2014). Global shared-layer blending method
for stacking sequence optimization design and blending of composite
structures. Composites Part B: Engineering, 69 , 181–190.

[24] Kang, L. S., & Blom, A. W. (2016). Optimization of ply orientations

for multi-layer composite parts. US Patent 2017/0228473 A1.

[25] Kiyono, C., Silva, E., & Reddy, J. (2017). A novel fiber optimization
method based on normal distribution function with continuously varying
fiber path. Composite Structures, 160 , 503 – 515.

[26] Le Riche, R., & Haftka, R. (1993). Optimization of laminate stacking

sequence for buckling load maximization by genetic algorithm. AIAA
Journal, 31 , 951–956.

[27] Levary, R. R. (1994). Engineering applications of operations research.

European Journal of Operational Research, 72 , 32 – 42.

[28] Liu, D., Toropov, V., Zhou, M., Barton, D., & Querin, O.
(2010). Optimization of blended composite wing panels using
smeared stiffness technique and lamination parameters. In 51st
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and
Materials Conference.

[29] Liu, X., Featherston, C. A., & Kennedy, D. (2019). Two-level layup opti-
mization of composite laminate using lamination parameters. Composite
Structures, 211 , 337 – 350.

44
[30] Liu, X., Featherston, C. A., & Kennedy, D. (2020). Buckling opti-
mization of blended composite structures using lamination parameters.
Thin-Walled Structures, 154 , 106861.

[31] Lund, E., & Stegmann, J. (2005). On structural optimization of compos-

ite shell structures using a discrete constitutive parametrization. Wind
Energy, 8 , 109–124.

[32] Macquart, T., Bordogna, M., Lancelot, P., & De Breuker, R. (2016).
Derivation and application of blending constraints in lamination pa-
rameter space for composite optimisation. Composite Structures, 135 ,
224–235.

[33] Meddaikar, Y., Irisarri, F.-X., & Abdalla, M. (2017). Laminate op-
timization of blended composite structures using a modified shepard’s
method and stacking sequence tables. Structural and Multidisciplinary
Optimization, 55 , 535–546.

[34] Montemurro, M., Koutsawa, Y., Belouettar, S., Vincenti, A., & Van-
nucci, P. (2012). Design of damping properties of hybrid laminates
through a global optimisation strategy. Composite Structures, 94 , 3309–
3320.

[35] Montemurro, M., Pagani, A., Fiordilino, G., Pailhès, J., & Carrera, E.
(2018). A general multi-scale two-level optimisation strategy for design-
ing composite stiffened panels. Composite Structures, 201 , 968–979.

[36] Montemurro, M., Vincenti, A., & Vannucci, P. (2012). A two-level pro-
cedure for the global optimum design of composite modular structures—

45
 application to the design of an aircraft wing part i. Journal of Optimiza-
tion Theory and Applications, 155 , 1–23.

[37] Panettieri, E., Montemurro, M., & Catapano, A. (2019). Blending con-
straints for composite laminates in polar parameters space. Composites
Part B: Engineering, 168 , 448 – 457.

[38] Picchi Scardaoni, M., Montemurro, M., Panettieri, E., & Catapano, A.
(2020). New blending constraints and a stack-recovery strategy for the
multi-scale design of composite laminates. Structural and Multidisci-
plinary Optimization, .

[39] Sanz-Corretge, J. (2017). A procedure to design optimum composite

plates using implicit decision trees. Structural and Multidisciplinary
Optimization, 56 , 1169–1183.

[40] Sanz-Corretge, J., & Echeverría, M. (2018). A novel technique for the
design of hybrid composite laminates based on dynamic programming
and dynamic tree trimming. Structural and Multidisciplinary Optimiza-
tion, 57 , 1507–1521.

[41] Schuhmacher, G., Daoud, F., Petersson, O., & Wagner, M. (2012).
Multidisciplinary airframe design optimisation. In 28th International
Congress of the Aeronautical Sciences (pp. 44–56). volume 1.

[42] Seresta, O., Gürdal, Z., Adams, D., & Watson, L. (2007). Optimal design
of composite wing structures with blended laminates. Composites Part
B: Engineering, 38 , 469–480.

46
[43] Soremekun, G., Gürdal, Z., Kassapoglou, C., & Toni, D. (2002). Stack-
ing sequence blending of multiple composite laminates using genetic al-
gorithms. Composite Structures, 56 , 53–62.

[44] Sørensen, S. N., & Lund, E. (2013). Topology and thickness optimization
of laminated composites including manufacturing constraints. Structural
and Multidisciplinary Optimization, 48 , 249–265.

[45] Stegmann, J., & Lund, E. (2005). Discrete material optimization of

general composite shell structures. International Journal for Numerical
Methods in Engineering, 62 , 2009–2027.

[46] Todoroki, A., & Haftka, R. (1998). Stacking sequence optimization

by a genetic algorithm with a new recessive gene like repair strategy.
Composites Part B: Engineering, 29 , 277–285.

[47] Tsai, S. W., & Pagano, N. J. (1968). Invariant properties of composite

materials. In Composite Materials Workshop.

[48] Vicente, F. S. (2019). Stacking Sequence Retrieval of Large Composite

Structures in Bi-step Optimization Strategies Using Mechanical Con-
straints. Master’s thesis TU Delft the Netherlands.

[49] Yang, J., Song, B., Zhong, X., & Jin, P. (2016). Optimal design of
blended composite laminate structures using ply drop sequence. Com-
posite Structures, 135 , 30–37.

[50] Zein, S., Basso, P., & Grihon, S. (2014). A constraint satisfaction pro-
gramming approach for computing manufacturable stacking sequences.
Computers and Structures, 136 , 56–63.

47
[51] Zeng, J., Huang, Z., Chen, Y., Liu, W., & Chu, S. (2019). A simu-
lated annealing approach for optimizing composite structures blended
with multiple stacking sequence tables. Structural and Multidisciplinary
Optimization, 60 .

48