Composite Structures 125 (2015) 596–604

Contents lists available at ScienceDirect

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

Stacking sequence optimisation of variable stiffness laminates with

manufacturing constraints
Daniël M.J. Peeters ⇑, Simon Hesse 1, Mostafa M. Abdalla
Faculty of Aerospace Engineering, Delft University of Technology, Kluyverweg 1, Delft 2629HS, The Netherlands

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

Article history: The fibre paths of variable stiffness laminates are described through the fibre angles at the nodes of a
Available online 21 February 2015 finite element (FE) representation of the structure. An algorithm is presented to optimise the fibre angles
efficiently. To reduce the number of required FE analyses a multi-level approach is used: the exact solu-
Keywords: tion is first approximated in laminate stiffness space. The second level approximation is a Gauss–Newton
Buckling quadratic approximation in fibre angle space. To ensure manufacturability, a steering constraint is intro-
Computational modelling duced: the norm of the gradient of the fibre angle distribution is constrained. Two formulations are pro-
Lay-up (manual/automated)
posed: either the average steering is constrained; or the local element-wise steering is constrained. The
resulting quadratically constrained quadratic optimisation problem is solved using an interior-point
method. It is shown that the local steering constraint performs best, at the cost of increasing the size
of the problem.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction promising, manufacturable, results [5–10]. Also for stiffened plates,

Today, composite materials are frequently used in the aviation
industry and the first composite-dominated planes like the B-787
or A400M are being built. Traditionally, fibres within a layer have
the same orientation, leading to constant stiffness properties.
However, as manufacturing technology evolved, for example auto-
mated fibre placement machines, the fibre orientation of a layer
can be continuously varied leading to varying stiffness properties
that are best tailored for the applied loads. These composites are
called variable stiffness laminates (VSL).
When designing VSL, manufacturability is not always taken into
account [1]. In an early work, the structure was divided into differ-
ent segments in which the fibre angle was optimised separately.
An example of an outcome can be seen in Fig. 1(a) [2]. A similar
approach has been taken in the optimisation of flutter speed for
wings: the angle of a lot of elements is optimised, but no manufac-
turability constraint is taken into account [3]. In another approach,
the change in fibre angle between adjacent layers is taken into
account, but the set of possible angles is restricted to 0°, 45°
and 90°, so the change in fibre angle is still large [4].
To take manufacturability into account, linearly varying fibre
angles were used as can be seen in Fig. 1(b), which has given
⇑ Corresponding author.
E-mail address: D.M.J.Peeters@tudelft.nl (D.M.J. Peeters).
1
Current address: BMW Group, Research and Innovation Centre, Knorrstrasse 147,
D-80788 München (in cooperation with the Technische Universität München).
the use of linearly varying fibre angle per bay has been investigat-
ed, and again it was shown that varying the fibre angles leads to
better performance [11,12]. Direct parametrisation of the tow
paths using Lagrangian polynomials, splines or NURBS (Non-
Uniform Rational B-Splines) has been done as well. This also
showed large, manufacturable, improvements in buckling load,
but the result is dependant of the basis functions you chose to
incorporate [13–15]. Hence, the total potential of VSL is not
exploited due to the pre-specified set of possibilities.
Furthermore, most methods assume the fibres are shifted, meaning
a choice had to be made whether gaps or overlaps were allowed
during manufacturing [16].
Another approach that leads to manufacturable designs is to
align the fibres in the direction of principal stress. This was
shown to reduce stress concentrations, and could also lead to
reduced weight using the tailored fibre placement method
[17,18]. Also using the load paths, or a hybrid combination of
load paths and principal stress direction has been used to design
variable stiffness laminates [19]. Continuous tow shearing is a
new manufacturing method, leading to varying fibre angles with-
out any gaps or overlaps, but with a thickness variation that is
coupled with the change in fibre angle [20,21]. Using a genetic
algorithm, coupled with a pattern-search algorithm, or using
the infinite strip method large improvements were shown to be
possible [22,23]. A more comprehensive review can be found in
Ghiasi et al. [24].

http://dx.doi.org/10.1016/j.compstruct.2015.02.044
0263-8223/Ó 2015 Elsevier Ltd. All rights reserved.
 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604 597

(a) example of the outcome when structure (b) example of a linearly varying fibre
is divided in parts, taken from Hyer and angle, taken from Lopes et. al. [5]
Lee [2]

Fig. 1. 2 Outcomes of previous optimisations.

To exploit the possibilities of VSL fully, a three-step approach exact FE response f is first approximated in terms of the in- and
has been developed. The first step is to find the optimal stiffness out-of-plane stiffness matrices A and D and their reciprocals:
distribution in terms of the lamination parameters. This is dis- ð1Þ
X
cussed in detail in IJsselmuiden [25,26]. The second step is to find f  /m : A1 þ /b : D1 þ wm : A þ wb : D þ c ð2Þ
n
the optimal manufacturable fibre angle distribution, the focus of
this paper [27–29]. The third step is to retrieve the fibre paths, dis- where the : operator represents the Frobenius inner product, mean-
cussed in Blom [30]. A schematic overview of this approach is ing A : B ¼ trðA  BT Þ; the reciprocal and linear approximation terms
shown in Fig. 2. / and w are calculated from a sensitivity analysis [34,35], m denotes
In this paper an algorithm to optimise manufacturable fibre the membrane, b the bending part and n runs over all the nodes.
angle distributions is developed. The stacking sequence at each This approximation is a generalisation of the linear-reciprocal ones
node of a finite element model will be optimised. To ensure used in the convex linearisation method [36]. The approximations
manufacturability, the rate of change in fibre angle between nodes, are convex in stiffness space provided that / P 0, which is always
referred to as steering from here on, is constrained [31,32]. satisfied by construction. For many responses that enjoy homogene-
This paper is organised as follows: first the problem formulation ity properties the free term c equals zero. This way of approximat-
is discussed in Section 2, next the manufacturing constraints are ing structural responses works for stiffness, buckling, strength and
discussed in Section 3. The solution procedure is explained in eigenfrequency problems.
Section 4, followed by the results in Section 5, and finally the con- To optimise the fibre angles, the first approximation is evaluat-
clusion in Section 6. ed by considering the dependence of the stiffness matrices on the
fibre angles. As a function of fibre angles, it no longer has a simple
mathematical form and is not generally convex. Hence, a second
2. Problem formulation level approximation is made:
ð2Þ ð1Þ
In structural optimisation, the minimisation of an objective (e.g., f  f 0 þ g  Dh þ Dh T  H  Dh ð3Þ
weight or compliance) subject to performance constraints (e.g., on ð1Þ
where f 0 denotes the value, g the gradient and H is an approxima-
stresses) is studied. More generally, the worst case response (e.g.,
tion of the Hessian of the first approximation at the approximation
in the case of multiple load cases) may be minimised. Additional
point. g and H can be calculated starting from
constraints, often arising from manufacturing considerations, may
be imposed to guarantee certain properties of the design. f
ð2Þ
ðhÞ ¼ f
ð1Þ
ðsðhÞÞ ð4Þ
In the optimisation of variable stiffness laminates a suitably-de-
fined steering norm 1 is constrained to be less than a maximal where s contains the components of the stiffness matrices A and D.
steering value 1U representing the upper manufacturing limit. Deriving this leads to
More details about the formulation of the steering constraint will ð1Þ ð2Þ ð1Þ
@f @f @f @sa
be given in Section 3. gi ¼ ¼ ¼  ð5Þ
Thus, the following general problem formulation is considered: @hi @hi @sa @hi

min maxð f 1 ; f 2 ; . . . ; f n Þ Deriving again, the Hessian is found to be

ð1Þ ð1Þ ð1Þ
s:t: f nþ1 ; . . . ; f m 6 0 ð1Þ @2f @2f @sa @sb @f @ 2 sa
Hij ¼ ¼   þ  ð6Þ
12  12U 6 0 @hi @hj @sa @sb @hi @hj @sa @hi @hj
where f 1 up to f n denote structural responses that are optimised Convexity is guaranteed by omitting the underlined part of Eq.
and f nþ1 up to f m are constraints. (6), which is not guaranteed to be positive definite, and using the
Structural responses, such as stiffness and strength, are calcu- Gauss–Newton part which is positive semi-definite. An approxima-
lated using finite element analyses (FEA). Since each FEA is compu- tion has to have equal function and gradient at the approximation
tationally expensive, greater efficiency may be achieved by using point as the approximated function. Hence, using only part of the
structural approximations, reducing the number of FEAs [33]. The Hessian does give a valid approximation.
598 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604
The second level approximations, Eq. (3), are used to construct a
convex sub-problem by replacing the responses in Eq. (1) by the
approximations. The convex sub-problem is solved using an
extended version of the predictor–corrector interior-point method
[37]. More details about the solution procedure are given in
Section 4. Once the solution of the sub-problem is obtained, the
stiffness matrices in the first level approximation are updated
and the second level approximation is updated. This is repeated
until the improvement in the first level approximation is smaller
than a given tolerance or a maximum number of iterations is done.
When either of these criteria is met, an FEA is performed to update
the first level approximation. When the improvement between
successive FEAs becomes smaller than a given tolerance, the opti-
mal solution is found.
To guarantee global convergence, a sufficient condition is to
have conservative approximations: the approximation at any level
should be larger than the previous level at the next iterate. To
achieve this, a strictly convex function of zero value and gradient
at the approximation point, called damping function from here
on, is added [38]. The shape of the damping function is usually cho-
sen to match the type of approximation being used, and is denoted
ð1Þ ð2Þ
d and d for the first and second level approximation. The damp-
ing function is scaled by a damping factor, denoted qð1Þ and qð2Þ . The
damping factor is updated throughout the iterations to guarantee
conservativeness. Details of the damping functions and the damp-
ing factor update procedure are described in Appendix A. The solu-
tion procedure is shown in Fig. 3, and is explained in Algorithm 1.
Algorithm 1. Solution Procedure
1: start from an initial fibre angle distribution.
2: perform an FEA and calculate the sensitivities for the first
ð1Þ
level approximation f .
ð1Þ
3: add the damping function qð1Þ  d to the first level
approximation.
4: calculate the gradient and Hessian for the second level
ð2Þ
approximation f . applied: the average steering of each element is constrained. This
ð2Þ
5: add the damping function qð2Þ  d to the second level
approximation.
6: apply the steering constraint, build the Lagrangian L and
solve the system.
ð1Þ
7: calculate first level approximation f and update
ð2Þ
damping factor of level two d .
8: decide if the new point is accepted: if the first level
ð1Þ
approximation f is improved, the point
is accepted. If the point is not accepted, go back to step 5,
else continue.
9: check whether the first level approximation f has
ð1Þ
converged, or the maximum number of
iterations is reached. If either of these conditions is met,
continue, else return to step 4.
10: perform an FEA and update the damping factor of first
ð1Þ
level approximation d .
11: decide if the new point is accepted: if the FE response has
improved, the point is accepted. If
the point is not accepted, go back to step 3, else continue.
12: if FEA has converged, the optimal fibre angle distribution
is found, else return to step 2.
3. Manufacturing constraints
To ensure a laminate can be manufactured using automated
fibre placement, the rate of change in fibre angles should be
limited. This has two reasons. One, the fibre placement machine
has to be able to lay down the path without tow wrinkling. Two,
the convergence or divergence of fibres should not be excessive
to limit the area of gaps and overlaps. The derivative in the fibre
path direction is the curvature, the derivative perpendicular to
the fibre path indicates the thickness build-up [30]. Thus, the norm
of the gradient of the fibre angles, measuring the steering is con-
strained. This can be done on a global (averaged) level, limiting
the number of gaps and overlaps, or on a local level, preventing
wrinkling. The maximum local steering can be physically interpret-
ed as the inverse of the minimal turning radius. For example: a
maximal local steering of 2 m1 corresponds to a minimal turning
radius of 500 mm. The numerical values of the global steering are
harder to interpret: it denotes the average minimal turning radius
needed, which only gives an indication of the expected number of
gaps and overlaps.
The steering 1 is given by
12 ¼ rh  rh ð7Þ
The average steering can be found using
Z
1
12 ¼ 12 dX ð8Þ
X X
where X is the total area of all elements.
When constraining the global steering, discretizing leads to
2
12 ¼  hT  L  h ð9Þ
X
where L is the standard FEM discretisation of the Laplacian.
Using a global steering constraint has the advantage that only
one constraint per layer is posed. However, since the average steer-
ing of each layer is constrained, the maximum is not taken into
account: hence, manufacturability is not guaranteed, not even if
a low upper bound 1U is used.
To guarantee manufacturability local steering constraints are
is calculated using
2
12 ¼  hT  Le  h ð10Þ
X
where the subscript e denotes the element. Hence, one constraint
per element per layer is posed, significantly increasing the compu-
tational effort required. However, manufacturability is ensured.
4. Solution of the sub-problem
As described in Section 2, a multilevel approximation approach
is used for the solution of the optimisation problem. The crucial
numerical component is the solution of the second level
approximation. The solution of this quadratically constrained
quadratic optimisation is explained in this section. The solution
procedure will be explained for the local steering constraint. The
same formulation works for the global steering constraints, or a
combination there of.
First, the minimax optimisation problem, Eq. (1), is rewritten
using a bound formulation [34]:
min z
s:t: fi  z  e 6 0 ð11Þ
12  12U 6 0
where e is a vector consisting of only ones and zeros: one if the
function is an objective, zero if it is a constraint. The Lagrangian
of the problem is
 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604 599

Fig. 2. Schematic overview of the three-step approach [25].

Fig. 3. Flowchart of the optimisation.

X  
Lðko ; ke ; h; z; so ; se Þ ¼ z þ ko  f  z  eT þ so where the criteria related to so ; se and ke are given per structural
o
! ! response or element. For numerical stability, the optimality condi-
X X tions with respect to the slacks are multiplied with the slacks:
þ hT  ke  Le  h  12U  ke  jXe j
e e k o  so  l ¼ 0
! ð14Þ
X ke  Xe  se  l  Xe ¼ 0
þ ke  se  jXe j
e
! Linearising and using the Schür complement to remove the
X X slacks from the equation:
l lnðso Þ þ jXe j  lnðse Þ ð12Þ 2 3 2
ko  H þ 2  ke  Le g 2  Le  h 0 3 2 3
o e Dh rh
6 7
where ko and ke denote the Lagrangian multipliers of the structural 6 gT  ksoo 0 eT 7 6
6 Dk o 7
7
6r 7
6 ko 7
6 76 7¼6 7
6 7
responses and the steering constraint of an element; both have to 4 ð2  Le  hÞT 0  sekXe e 0 5 4 Dke 5 4 r kg 5
be non-negative. The slack of the structural responses and of the Dz rz
0 e 0 0
constraints are given by so and se . The homotopy factor is denoted
by l. The optimality criteria are obtained by setting the gradients ð15Þ
of the Lagrangian to zeros: Following the predictor–corrector algorithm [37], this system is
T first solved with the homotopy factor l equal to zero, this is the pre-
rz : 1  ko  e ¼ r z dictor step. Next, the duality gap is determined by
X X
rh : ko  g þ 2  ke  Le  h ¼ rh  T  T
o e dg ¼ sPo  kPo þ sPe  kPe ð16Þ
rko : f  z  eT þ so ¼ r ko
where the superscript P refers to the outcome of the predictor step.
l ð13Þ
rso : ko  ¼ r so l is determined by
so
rke : hT  Le  h  jXe j  12U þ se  jXe j ¼ rke l ¼ b  dg ð17Þ
l  jXe j where b is a constant. Details on determining the homotopy factor
rse : ke  jXe j  ¼ r se
se and the interior-point method can be found in Zillober [37]. This
value for l is used to update the right-hand-side in Eq. (15) and
600 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604

Fig. 4. Graphical representation of the structure, loading and boundary condition

for the first example.

Table 1
Overview of the results using 1 and 5 iterations of first level approximation.
Fig. 5. Curvature per element when using a global steering constraint of 4.
Steering Only FE Only FE Only FE 5 Level 1 5 Level 1 5 Level 1
Mode 1 Mode 2 FE analyses Mode 1 Mode 2 FE analyses
1 1.3158 1.352 13 1.3356 1.3571 12
2 1.5088 1.5713 8 1.5102 1.5548 8
3 1.6549 1.7302 18 1.653 1.7174 8
limits on global steering. For each limit the problem is solved
4 1.7682 1.8356 19 1.7723 1.8296 15 twice: once with a single approximation level (no iterations of
5 1.8789 1.8837 25 1.8578 1.9134 17 the first approximation), and once with five iterations of the first
approximation allowed. In Table 1 the value of the global steering
constraint used is given in the first column, in the next columns the
normalised optimal buckling loads and number of FEAs when a
the system is solved again, this is the corrector step. Based on this, single level of approximations is used are given. The normalised
all variables are updated. This is repeated until convergence is optimal buckling loads and number of FEAs if the first approxima-
reached. tion is updated five times are given in the last columns. The results
given in Table 1 show that the number of FEAs never increases and
5. Results usually decreases significantly when five iterations of the first
approximation are allowed. The optimal results are not exactly
To demonstrate the optimisation algorithm, consider the cylin- the same using both approaches but the differences are not sig-
drical panel with a hole shown in Fig. 4, in which all relevant nificant. From here on, five iterations of the first level approxima-
dimensions are indicated. All edges are simply supported. The pan- tion are used.
el is subjected to uni-axial compression at the curved edges, which Next, the effect of the steering limits is investigated. In Table 2,
are constrained to remain straight. The material used has Young’s in the first column the upper bound on average steering is given,
moduli E1 of 154 GPa, and E2 of 10.8 GPa, a shear modulus G12 of the next two columns list the normalised buckling loads, in the
4.02 GPa, a Poisson’s ratio m12 of 0.317 and a ply thickness of next four columns the maximum local steering observed in each
0.225 mm. The laminate consists of 16 layers, by choosing it to layer is given, and the final column lists the number of FEAs. The
be balanced and symmetric, four layers are optimised. The model results are as expected: the higher the steering limit, the higher
consists of 864 nodes and 1544 triangular elements. The plate is the buckling load. Further, the first two buckling loads are closer,
optimised for two buckling modes, accounting for possible modal but never identical, for higher steering limits. The most important,
interactions, subject to steering constraints. however, is that the maximum local steering is much higher than
All results are normalised using the buckling load of a quasi- the global limit, and no simple correlation between them is
isotropic (QI) laminate. As a reference, the optimisation is per- observed. These high steering values likely cause manufacturing
formed in lamination parameter space. The results show that com- problems.
pared to a QI laminate, the best constant stiffness laminates (CSL) The steering value per element is shown in Fig. 5 where it is
has a 20.1% increase in buckling load; while using a VSL a 98.3% observed that the maximum steering is reached only in a small
improvement is found. Thus, optimally, steering may cause up to area. The mechanisms for buckling load improvement for these
65% increase in performance compared to the best CSL. configurations are reasonably well understood based on the earlier
First, the effect of the multi-level approximation scheme is work of Ijsselmuiden et al. [39] using lamination parameters. The
studied. The optimisation problem is solved for different upper load is redistributed away from the cut-out towards the supported

Table 2
Overview of the results using the global steering constraint.

Maximum global Buckling Buckling Maximum local Maximum local Maximum local Maximum local Number of
steering load 1 load 2 steering layer 1 steering layer 2 steering layer 3 steering layer 4 FE analyses
1 1.3366 1.3571 3.2411 2.2651 2.3260 2.3922 12
2 1.5102 1.5548 4.3266 4.6783 5.4617 5.4648 8
3 1.653 1.7174 7.9049 8.9784 7.6136 7.7473 8
4 1.7723 1.8296 12.6603 13.8561 10.2546 10.7625 15
5 1.8578 1.9134 16.1786 23.2913 13.7939 14.7003 17
 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604 601

Table 3
Overview of the results using the local steering constraint.

Maximum global steering Buckling load 1 Buckling load 2 Upper bound local steering Buckling load 1 Buckling load 2 Difference local vs global (%)
1 1.3366 1.3571 3.2411 1.5725 1.6326 +17.6
2 1.5102 1.5548 5.4648 1.7348 1.8001 +14.9
3 1.6530 1.7174 8.9784 1.8526 1.8758 +12.1
4 1.7723 1.8296 13.8561 1.9290 1.9439 +8.8
5 1.8578 1.9134 23.2913 2.0098 2.0164 +8.2

(a) fibre paths of layer 1 and 2 (b) fibre paths of layer 3 and 4
(outer layers)

(c) fibre paths of layer 5 and 6 (d) fibre paths of layer 7 and 8 (at
symmetry plane)
Fig. 6. Fibre paths optimised using local steering constraints.

edges. The most critical area is between the supports (parallel to

the load) and the cut-out. That is where most of the steering is con-
centrated as shown in the figure.
To compare local and global steering constraints, the maximum
local steering found for the global constraint is set as upper bound
on the local steering. From a manufacturing point of view this is
realistic: if this steering can be laid down locally, it is possible at
all places so constraining steering at other places to be lower is
unnecessary. In Table 3, the global steering constraint is shown
in the first column, the optimal buckling loads are given in the next
two columns, the fourth column contains the local steering limits,
the optimal buckling load found using the local constraints is given
in the next two columns; finally, in the last column the buckling
loads found using the local and global steering constraints are
compared.
In this case, the structure is optimised with the upper limit on
local steering set to the observed maximum steering in the globally
constrained optimisation. The local steering cannot, by construction,
exceed the value observed in the globally constrained optimisation.
The global constraint is not applied. For that reason, the average
steering of the panel may exceed the upper bound previously set
in the globally constrained problem. Thus the optimiser has more Fig. 7. Curvature per element when using a local steering constraint of 3.2411.
602 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604

Table 4
Overview of the results using only global and both global and local constraints.

Maximum global Buckling Buckling Maximum local Buckling Buckling Upper bound local Difference local and global vs only
steering load 1 load 2 steering load 1 load 2 steering global (%)
1 1.3366 1.3571 3.2411 1.3388 1.3602 2 +0.2
2 1.5102 1.5548 5.4648 1.5126 1.5559 4 +0.2
3 1.6530 1.7174 8.9784 1.6614 1.7221 6 +0.5
4 1.7723 1.8296 13.8561 1.7731 1.8464 8 +0.05
5 1.8578 1.9134 23.2913 1.8552 1.8858 10 0.2

and maximal local steering when only the global constraint is

Fig. 8. Curvature per element when using a local steering constraint of 8 and a
global constraint of 4.
freedom to allow larger regions to attain higher curvatures. This
relaxation of the constraint is translated ultimately into an improve-
ment in performance.
For the upper bound on local steering set to 3.24, the top view of
the fibre paths is shown in Fig. 6, where both the layer and its bal-
anced counterpart are shown. The steering per element of layer 2 is
shown in Fig. 7. Compared to Fig. 5, the maximum steering is used
over a larger area: the potential of the machine and material is
used more effectively.
Global and local steering constraints can be posed simultane-
ously such that the local constraints avoid tow wrinkling, while
the global constraints limit the area of gaps and overlaps. To quan-
tify the effect of local constraints on the optimal performance, the
local constraint is set to double the global constraint. In Table 4 the
upper bound on global steering constraint is given in the first col-
umn, the next three columns contain the optimal buckling loads
applied; in the next two columns the optimal buckling loads and
maximum local steering when the local steering is constrained to
be double the global constraint are given, in the last column the
performance of the optimised structures found with both global
and local and only global constraints are compared.
From the results in Table 4, it is clear that adding local steering
constraints has almost no effect on the performance, while the
maximum local steering is considerably decreased. In some cases
the performance is even slightly increased when applying local
constraints: the gradient-based optimisation follows a different
path, ending up in another optimum. This example clearly demon-
strates the strength of the local steering constraint: (almost) the
same performance can be obtained while having precise control
over the maximum local steering. Comparing the steering per ele-
ment for this case, shown in Fig. 8, with the steering distribution
without local constraint, shown in Fig. 5, it can be seen the steering
is more spread out.
As a second example, a flat plate without cut-outs, loaded in
shear is optimised. The left and right edge are constrained to
remain straight, the upper and lower edge are simply supported.
The plate is square with side length of 0.5 m. The same material
as in the previous example is used. The laminate is 16 layers thick,
by assuming it is balanced and symmetric, four layers are opti-
mised. The mesh is 10  10 nodes; this rather coarse mesh is suf-
ficient for this simple example. The plate is optimised for
maximum buckling load, considering the first two buckling modes,
while the stiffness has to be at least the QI stiffness.
The local steering constraint was always active during the opti-
misation. In Table 5 the first column consists of the local steering
constraint that was used, in the second, third and fourth column
the normalised optimal buckling loads and stiffness are given, in
the next four columns the global steering per layer is given and
in the final column the number of FEAs is given. Looking at the
results in Table 5, it can be seen the stiffness constraint is always
active and the buckling improvement gets larger for higher local
steering. The results were found using an initial starting point of
45° on all layers. The ⁄ with local steering of 5 and 7 denotes a dif-
ferent initial point was used. The fibre paths for a steering con-
straint of 5 are shown in Fig. 9.

Table 5
Overview of the results using the local steering constraint on the second example.

Upper bound local Buckling Buckling Stiffness Global steering Global steering Global steering Global steering Number of FE
steering load 1 load 2 layer1 layer 2 layer 3 layer 4 analyses
1 1.4634 1.5700 1.0002 0.9028 0.9244 0.9200 0.9208 6
3 1.9831 2.0724 1.0023 2.6069 2.7129 2.7085 2.7125 8
5 2.0163 2.0517 1.0032 3.3131 3.9222 4.1879 3.9190 7
7 2.0953 2.1072 1.0005 3.6864 4.7147 5.3417 4.8093 7
9 2.1074 2.1213 1.0015 3.9911 5.5425 5.2029 6.4330 10
15 2.1798 2.2136 1.0022 6.7201 8.3499 8.5434 9.0048 8
20 2.2161 2.2449 1.0017 7.0214 8.9703 8.9088 10.1307 9
 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604 603

(a) fibre paths of layer 1 and 2 (b) fibre paths of layer 3 and 4
(outer layers)

(c) fibre paths of layer 5 and 6 (d) fibre paths of layer 7 and 8 (at
symmetry plane)
Fig. 9. Fibre paths optimised using local steering constraint of 5 for plate under shear load.

6. Conclusion Acknowledgements

A method was developed to optimise the stacking sequences This work is supported by the CANAL (CreAting Non-conventionAl
of VSLs including manufacturing constraints. The stacking Laminates) Project, part of the European Union Seventh Framework
sequences are defined at the FE nodes using a vector of fibre Program.
angles. This parametrisation in terms of fibre angles results
directly in the information needed to determine the fibre paths Appendix A. Details of the damping function
and hence to manufacture the laminates. Additional steering
constraints are imposed on the norm of the gradient to assure For the first approximation the damping shape is chosen to be
the smoothness, and, hence, the manufacturability of the fibre [25]:
angle distributions. The method offers true stacking sequence ð1Þ
X
d ¼ A0 : A1 þ D0 : D1 þ A1 1
0 : A þ D0 : D  12 ð18Þ
optimisation.
n
To reduce computational time a multi-level approach was
used: the result of the FEA was used to build an approximation where the subscript 0 denotes the value at the approximation point.
function in terms of the in- and out-of-plane stiffness matrices, For the second approximation function, the damping shape is a
which was consecutively used to build a second approximation regularization matrix, that is added to the Gauss–Newton part of
in terms of the change in fibre angles. To ensure each step is the Hessian to form the ‘complete’ Hessian that is used in the opti-
an improvement step and to let the approximation be conserva- misation. The regularization matrix has the following form:
tive, a damping function was added to the approximations. This 2 3
1 1
approach has been proven to work well: a limited number of 6 1 2 7 0 1
6 1 7 1 ... 1
FEAs was needed to find the optimal fibre angle distribution 16 7 B
while the first, computationally cheaper, approximation was fre- Hc ¼ 2 6 .. .. .
7 þ a@ .. . . . .. C
.
A ð19Þ
nl 6
6
. . 7
7
quently evaluated. 4 2 1 5 1  1
Initial results indicate the effectiveness of the proposed method. 1
1 1
When the manufacturing constraints are active, the use of local
constraints resulted in improved optima compared to the use of where nl denotes the number of layers in the symmetric part of the
global, per layer, constraints. When both global and local con- laminate and a given by
straints are used, it has been shown the maximum local steering   2  ðnl  1Þ
can be reduced without affecting the performance.
a¼ ð20Þ
n3l
The current algorithm does not allow thickness change: the
number of layers is pre-specified by the user. The capability to where  is a small value to penalise the average change in fibre
change the laminate thickness will be addressed in future work. angles, typically 102 is used.
604 D.M.J. Peeters et al. / Composite Structures 125 (2015) 596–604
To find the initial damping factor in the first approximation, the
method proposed by IJsselmuiden was used [25]:
2 X wi  2  2 
qð1Þ ¼  jj/m : A1 jj þ jjwm : Ajj þ jj/b : D1 jj þ jjwb : Djj
n
2
ð21Þ
where wi is defined according to
Ai
wi ¼ P ð22Þ
i Ai
where Ai is the area represented by node i.
The minimal damping for the second approximation function
can be found solving
 
 Thermoplastic composite parts based on online spun commingled hybrid
qmin ¼  kmax ð23Þ
1
where kmax denotes the maximal eigenvalue of the Hessian.
After each iteration, the damping factor is updated: if the point
found is conservative, the damping can be decreased to increase
the step size, if the point is not conservative, the damping needs
to be increased to decrease the step size. Ideally, the exact and
approximate function, denoted f and ^f , are exactly the same at
the new optimum. Hence, the damping is updated using:
f ðx Þ^f ðx Þ
qnew ¼ qold  e dðx Þ ð24Þ

where x denotes the new optimum. This formula is straightfor-
ward, but some bounds need to be imposed: if the point is not con-
servative the ratio must not be larger than 2 to avoid the increase in
damping factor being too large, and minimal 1.05 to avoid too many
iterations that are just not conservative. When the point is conser-
vative a minimal ratio of 0.5 is set. This update is used for both the
first and second level approximation.
References
[1] Gürdal Z, Olmedo R. In-plane response of laminates with spatially varying fiber
orientations – variable stiffness concept. AIAA J 1993;31(4):751–8.
[2] Hyer M, Lee H. The use of curvilinear fiber format to improve buckling
resistance of composite plates with central circular holes. Compos Struct
1991;18(3):239–61.
[3] De Leon D, de Souza C, Fonseca J, da Silva R. Aeroelastic tailoring using fiber
orientation and topology optimization. Struct Multidiscip Optim
2012;46(5):663–77.
[4] Kennedy G, Martins J. A laminate parametrization technique for discrete ply-
angle problems with manufacturing constraints. Struct Multidiscip Optim
2013;48(2):379–93.
[5] Lopes C, Gürdal Z, Camanho P. Tailoring for strength of composite steered-fibre
panels with cutouts. Compos Part A: Appl Sci Manuf 2010;41(12):1760–7.
[6] Gürdal Z, Tatting B, Wu C. Variable stiffness composite panels: effects of
stiffness variation on the in-plane and buckling response. Compos Part A: Appl
Sci Manuf 2008;39(5):911–22.
[7] Hyer MW, Charette RF. Use of curvilinear fiber format in composite structure
design. AIAA J 1991;29(6):1011–5.
[8] van Campen JM, Kassapoglou C, Gürdal Z. Generating realistic laminate fiber
angle distributions for optimal variable stiffness laminates. Compos Part B: Eng
2012;43(2):354–60.
[9] Stodieck O, Cooper J, Weaver P, Kealy P. Improved aeroelastic tailoring using
tow-steered composites. Compos Struct 2013;106:703–15. cited By 7.
[10] Ungwattanapanit T, Baier H. Postbuckling analysis and optimization of
stiffened fuselage panels utilizing variable-stiffness laminates. In: Congress
of the international council of the aeronautical sciences, ICAS [29., 2014, St.
Petersburg]; 2014.
[11] Liu W, Butler R. Buckling optimization for composite panels with elastic
tailoring. In: 49th AIAA/ASME/ASCE/AHS/ASC structures. Structural dynamics
and materials conference; 2008.
[12] Coburn B, Wu Z, Weaver P. Buckling analysis of stiffened variable angle tow
panels. Compos Struct 2014;111(1):259–70.
[13] Nagendra S, Kodiyalam S, Davis JE, Parthasarathy VN. Optimization of tow
fiber paths for composite design. In: 36th AIAA/American society of
mechanical engineers/American society of civil engineers/American
helicopter society/society for composites structures. Structural dynamics,
and materials conference; 1995.
[14] van den Brink WM, Vankan W, Maas R. Buckling optimized variable stiffness
laminates for a composite fuselage window section. In: 28th International
congress of the aeronautical sciences; 2012.
[15] Wu Z, Weaver PM, Raju G, Kim BC. Buckling analysis and optimisation of
variable angle tow composite plates. Thin-Walled Struct 2012;60(0):163–72.
[16] Gürdal Z, Tatting B, Wu K. Tow-placement technology and fabrication issues
for laminated composite structures. In: 46th AIAA/ASME/ASCE/AHS/ASC
structures. Structural dynamics and materials conference. American Institute
of Aeronautics and Astronautics; 2005.
[17] Crothers P, Drechsler K, Feltin D, Herszberg I, Kruckenberg T. Tailored fibre
placement to minimise stress concentrations. Compos Part A: Appl Sci Manuf
1997;28(7):619–25.
[18] Richter E, Uhlig K, Spickenheuer A, Bittrich L, Maäder E, Heinrich G.
yarns with continuous curvilinear fibre patterns. In: 16th European conference
on composite materials; 2014.
[19] Tosh M, Kelly D. On the design, manufacture and testing of trajectorial fibre
steering for carbon fibre composite laminates. Compos Part A: Appl Sci Manuf
2000;31(10):1047–60.
[20] Kim B, Potter K, Weaver P. Multi-tow shearing mechanism for high-speed
manufacturing of variable angle tow composites, venice, IT; 2012.
[21] Kim B, Potter K, Weaver P. Continuous tow shearing for manufacturing
variable angle tow composites. Compos Part A: Appl Sci Manuf
2012;43(8):1347–56.
[22] Groh R, Weaver P. Mass optimization of variable angle tow, variable thickness
panels with static failure and buckling constraints. In: 56th AIAA/ASME/ASCE/
AHS/ASC structures. Structural dynamics and materials conference; 2015.
[23] Liu W, Butler R. Buckling optimization of variable-angle-tow panels using the
infinite-strip method. AIAA J 2013;51(6):1442–9.
[24] Ghiasi H, Fayazbakhsh K, Pasini D, Lessard L. Optimum stacking sequence
design of composite materials part ii: variable stiffness design. Compos Struct
2010;93(1):1–13.
[25] IJsselmuiden ST. Optimal design of variable stiffness composite structures
using lamination parameters [Ph.D. thesis]. Delft University of Technology;
2011.
[26] Peeters D, van Baalen D, Abdalla M. Combining topology and lamination
parameter optimisation. Struct Multidiscip Optim 2015:1–16.
[27] Autio M. Determining the real lay-up of a laminate corresponding to optimal
lamination parameters by genetic search. Struct Multidiscip Optim
2000;20(4):301–10.
[28] Setoodeh S, Blom A, Abdalla M, Gürdal Z. Generating curvilinear fiber paths
from lamination parameters distribution. In: 47th AIAA/ASME/ASCE/AHS/ASC
structures. Structural dynamics, and materials conference. American Institute
of Aeronautics and Astronautics; 2006.
[29] van Campen J, Gürdal Z. Retrieving variable stiffness laminates from
lamination parameters distribution. In: 50th AIAA/ASME/ASCE/AHS/ASC
structures. Structural dynamics, and materials conference. American
Institute of Aeronautics and Astronautics; 2009.
[30] Blom AW, Abdalla MM, Gürdal Z. Optimization of course locations in fiber-
placed panels for general fiber angle distributions. Compos Sci Technol
2010;70(4):564–70.
[31] Evans D. Fiber placement. In: Peters S, editor. Handbook of
composites. US: Springer; 1998. p. 476–87.
[32] Lukaszewicz DH-J, Ward C, Potter KD. The engineering aspects of automated
prepreg layup: history, present and future. Compos Part B: Eng
2012;43(3):997–1009.
[33] de Wit A, van Keulen F. Numerical comparison of multi-level optimization
techniques. In: 48th AIAA/ASME/ASCE/AHS/ASC structures. Structural
dynamics, and materials conference. American Institute of Aeronautics and
Astronautics; 2007.
[34] Gürdal Z, Haftka R. Elem Struct Optim 1992.
[35] Kumar V, Lee S-J, German M. Finite element design sensitivity analysis and its
integration with numerical optimization techniques for structural design.
Comput Struct 1989;32(3–4):883–97.
[36] Fleury C. Conlin: an efficient dual optimizer based on convex approximation
concepts. Struct Optim 1989;1(2):81–9.
[37] Zillober C. A combined convex approximation-interior point approach for large
scale nonlinear programming. Optim Eng 2001;2(1):51–73.
[38] Svanberg K. A class of globally convergent optimization methods based on
conservative convex separable approximations. SIAM J Optim 2002;2:555–73.
[39] Ijsselmuiden ST, Abdalla MM, Gürdal Z. Optimization of variable-stiffness
panels for maximum buckling load using lamination parameters. AIAA J
2010;48(1):134–43.