Composites Part B 253 (2023) 110536

Contents lists available at ScienceDirect

Composites Part B
journal homepage: www.elsevier.com/locate/composites

Review article

Rapidly predicting the effect of tool geometry on the wrinkling of biaxial

NCFs during composites manufacturing using a deep learning surrogate
model
J.V. Viisainen, F. Yu, A. Codolini ∗, S. Chen, L.T. Harper, M.P.F. Sutcliffe
Department of Engineering, University of Cambridge, CB2 1PZ, Cambridge, UK
Composites Research Group, Faculty of Engineering, University of Nottingham, NG7 2RD, Nottingham, UK

ARTICLE INFO ABSTRACT

Keywords: A deep learning surrogate model is developed to rapidly predict the wrinkling patterns of a biaxial non-crimp
Fabrics/textiles [A] fabric (NCF) layup for any given tool geometry during forming. The underlying dataset of finite element
Defects [B] simulations is used to investigate the effect of tool geometry on wrinkling severity. The trained surrogate
Computational modelling [C]
model is able to make reliable predictions of wrinkling patterns at a very low computational cost, suitable for
Forming [E]
tool design optimisation. Results indicate that certain geometrical features have a greater impact on wrinkling
than others. In particular, forming NCFs over geometries with greater draft angles tends to result in smaller
wrinkles.

1. Introduction
Wrinkling is a critical defect during the manufacturing of tex-
tile composite components which compromises the mechanical perfor-
mance of composites structures [1]. While there has been a multitude
of studies attempting to experimentally characterise [2–4] and simu-
late [5–7] the occurrence and development of fabric wrinkling during
preforming, mitigating these wrinkles remains an underdeveloped area
of research. In contrast to experimental tests and finite element sim-
ulations that are limited by material and computational cost, the use
of deep learning (DL) networks could offer a potential way to use
large empirical datasets to better understand the formation of wrinkles
under different conditions and to optimise preforming conditions for
minimal wrinkling. In particular, image-based DL networks called ‘fully
convolutional networks’ (FCNs) are suited for studying the effect of tool
geometry on fabric forming behaviour, as previously shown by [8],
due to being able to efficiently process image-based data and relate
representations of the tool geometry with the final pre-formed fabric
shape.
The tool geometry is a key factor affecting the presence of wrinkling
defects in the resultant preform [1] but this factor has also been
largely overlooked in the literature in favour of investigating process
conditions that can be modified at the preforming process stage, such as
the blank holder force [2,9,10] and inter-ply friction [7,11]. However,
it has been noted that wrinkling may not be mitigated solely through
changes to the preforming process conditions [12] and thus further un-
derstanding is needed to understand the effect of component geometry
on wrinkling. With better understanding of this and the development
of functional tools that can predict the effect on the manufactured
component, the overall product design cycle time and the cost of
manufacturing composite components could be reduced.
Finite element models have become the most widely used approach
for simulating the preforming process and investigating the formation
of wrinkles via simulation. One type of macroscale fabric model that
is regularly used for modelling fabric forming is a hypoelastic non-
orthogonal constitutive model [13]. When combined with a material
model that decouples in-plane and out-of-plane bending, it has been
shown to be capable of modelling fabric wrinkling while providing
a suitable balance between accuracy and computational cost [14,15].
Such models have been validated against experimental results, includ-
ing for multi-ply forming, and have proven to be able to capture both
the severity and location of wrinkles during press forming [16] and
double diaphragm forming [15,17].
While preforming models have become reasonably mature and the
accuracy of these models is now sufficient to predict wrinkling defects,
the development of methods for virtual process optimisation of pre-
forming with respect to minimising defects and increasing formability
remains largely unexplored [18].
Existing simulation tools for predicting the formability of a fabric
layup are not suitable for design optimisation as they are either too

∗ Corresponding author at: Department of Engineering, University of Cambridge, CB2 1PZ, Cambridge, UK.
E-mail address: ac2386@cam.ac.uk (A. Codolini).

https://doi.org/10.1016/j.compositesb.2023.110536
Received 5 August 2022; Received in revised form 8 December 2022; Accepted 18 January 2023
Available online 21 January 2023
1359-8368/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
J.V. Viisainen et al. Composites Part B 253 (2023) 110536

Fig. 1. (a) Outline of the code-based geometry generator used to generate the initial set of male (globally convex) geometries, consisting of three n-sided polygons fitted within
ellipses that are lofted together. The generator contains 12 independent variables (number of polygon sides (𝑛), major and minor axes of each ellipse (𝑎, 𝑏, 𝑐, 𝑑, 𝑒, 𝑓 ), the relative
height between them (ℎ𝑡 ) and their relative rotation (𝜃1 , 𝜃2 ,) and lateral shift (𝛥𝑥, 𝛥𝑦)) that are selected randomly for each generated geometry based on a set of pre-defined ranges
and subject to the conditions shown. (b) Examples of generated geometries from the initial set of 10,000 geometries. (c) The four investigated geometry characteristics (Mean
Curvature, Gauss Curvature, Angularity and Conicity) shown relative to one of the generated tool geometries. The Mean and Gauss curvatures (𝐻𝑟 and 𝐾𝑟 ) are calculated based
on the method outlined in [23] using the surface coordinates (𝑋 = (𝑥, 𝑦, 𝑧)) of the geometry while the Angularity is calculated from the mean interior angle (𝜃𝑖𝑛𝑡 ) of the polygons
(viewed from above) and the Conicity is the mean overall draft angle (𝜃𝑑𝑜 ) between the top and bottom polygons (viewed from the side).

simplistic to accurately capture the occurrence of wrinkles (e.g. kine-
matic or membrane models) or too computationally costly to produce
the rapid results necessary for optimisation purposes (e.g. shell mod-
els). DL surrogate models offer an attractive solution that could, in
theory, provide the low computational cost of kinematic or simplistic
mechanistic models while maintaining the higher accuracy of advanced
models that are expensive to run. Thus such a model could feasibly be
used for the iterative optimisation of a preforming process for minimal
wrinkling defects.
In particular, a FCN model could be used for this purpose. FCNs, first
introduced by Long et al. [19] in 2015, are an extension of the more
common convolutional neural networks (CNNs) and have images at
both the input and the output of the network. There is typically a one-
to-one connection between the input image pixels and the output image
pixels [20], allowing the model to make predictions of the output based
on a given input. Within the network layers, the input images are first
broken down into high-level feature representations before these are
used to generate a pixel-wise prediction of the expected output. These
networks have been particularly used for image segmentation [21]: to
divide images into key segments (for example to divide a portrait image
into the subject and the background along the subject’s silhouette).
When used without a classifier, the pixel-wise predictions of FCNs have
also been used as image-based surrogate models that learn the phe-
nomenological relationships between pairs of (input, output) images
derived from a physically relevant model [8,22].
DL surrogate models have the potential for improving the fabric
preforming process, for which process optimisation has typically been
limited to the use of simple models [24,25] that do not sufficiently
capture the physics of fabric deformation, in particular wrinkling for-
mation. Zimmerling et al. [8] showed that a surrogate FCN could be
trained to predict the shear angle distributions of a woven fabric formed
over a given tool geometry. This work was novel because it incorpo-
rated predictions of full shear strain fields rather than single scalar
values [26,27], making it a much more powerful and useful tool for
engineers compared to using scalar-based optimisation. While useful,
the accuracy of this model was still limited by the use of a kinematic
forming model for training [8]. For greater accuracy, physics-based FE
forming simulations should be used instead [28,29]. In addition, the use
of the shear angle as the output measure limits the model’s suitability
for wrinkling prediction of NCFs. This is because the shear locking
angle cannot predict the non-shear wrinkling in regions of restricted
shear [4] and cannot indicate wrinkling severity, which is crucial for
complex geometries where wrinkling cannot be fully avoided. Hence
for a more suitable surrogate model, the physically relevant wrinkling
patterns should be simulated directly and the wrinkling severity used
as the measure of performance, as previously attempted in [28,29] for
woven fabrics. In this work, a FCN-based surrogate model is developed
to predict wrinkling patterns for a NCF layup and this model is assessed
based on measures of wrinkling severity.

2
J.V. Viisainen et al.
The aims of this study are to develop a DL surrogate model to
rapidly predict the wrinkling severity for a given convex tool geometry
during double diaphragm forming for a biaxial NCF, and to investigate
the relationship between geometrical characteristics of the tool geome-
try and the severity of wrinkling. The study is structured as follows: the
method section outlines the material and forming process investigated
as well as the steps taken to develop the DL surrogate model. The results
section shows the effect of tool geometry on wrinkling and describes
the performance of the surrogate model. The discussion then expands
on the results and their implications for the mitigation of wrinkling,
before the conclusion outlines the key findings of the study.
The study finds that the conicity of a male tool geometry is neg-
atively correlated with the mean wrinkle amplitude of the resultant
non-crimp fabric ply formed over the geometry, given a linear co-
efficient of correlation of 0.53 between them. Furthermore, it finds
that the trained deep learning model is able to make predictions of
wrinkling patterns for a given tool geometry in only 0.7 s, compared
to 1.33 h for the equivalent finite element model. These predictions are
deemed sufficiently accurate to conduct tool design optimisation as it
is found that mean wrinkle amplitudes of previously unseen geometries
are predicted with a coefficient of determination of 0.71 and a mean
image accuracy (𝑀𝑆 − 𝑆𝑆𝐼𝑀) of 0.93 out of 1. These key findings are
further elaborated on below.
2. Method
2.1. Material, forming process and FE model
This numerical investigation focuses on the double diaphragm form-
ing (DDF) of a two-layer biaxial carbon fibre NCF layup under various
tool geometries. The DL surrogate model is trained based on macroscale
finite element simulations that capture the forming and wrinkling
behaviour of this material.
The fabric material is a carbon fibre biaxial ±45◦ NCF with a pillar
stitch (FCIM-359) manufactured by Hexcel, as outlined in Table 1. Two
plies of this NCF with dimensions of 200 mm × 200 mm and a layup of
[0◦ ∕90◦ , 0◦ ∕90◦ ] are used (Table 1). The bottom ply closest to the tool
is referred to as ‘NCF1’ while the top ply of the layup is referred to
as ‘NCF2’. The pillar stitching on the plies is orientated such that it is
at 45◦ to the global X direction (Table 1). A two-layer NCF layup is
chosen instead of a single NCF in order to consider a multi-ply forming
scenario where interply frictional effects could impact on wrinkling
arrangement [11]. However, in this layup arrangement, this effect is
likely to be less significant compared to when the plies are laid up
at contrasting angles [30]. The simulated DDF process is outlined in
Table 1 and is based on a laboratory scale DDF rig with dimensions
of 1.8 m × 1.5 m [31], developed at the University of Nottingham.
All other parameters related to the forming process beyond the tool
geometry are kept constant throughout the investigation.
A laminate shell forming model for the DDF process, originally
detailed by Yu et al. [14], that decouples the shell and membrane
bending behaviour is used to model the wrinkling behaviour of the NCF
layup over each of the different tool geometries. Reduced quadrilateral
shell elements (S4R) of element edge size of 3 mm are used for the
fabric. The model has been previously validated against experimental
results obtained using the DDF rig referred to above [17]. Additionally,
it has been shown to be able to capture the two distinct wrinkling
mechanisms in positive and negative shear for this ±45◦ NCF [16].
The bending stiffness of the NCF is modelled with a constant
bending stiffness per unit width along the fibre direction (𝐵𝑓 𝑖 ) of
0.00228 Nm. This is 40% lower than the mean experimentally derived
value of 0.0035 Nm for this NCF using the standard cantilever test
BS EN ISO 9073-7 [16]. This lower bending stiffness was chosen to
make the wrinkling behaviour more pronounced and therefore aid the
deep learning model in identifying the differences in wrinkling patterns
for different tool geometries. Direct comparisons of simulation results
across a selection of geometries showed that the lower bending stiffness
leads to a 10% increase in wrinkling severity while maintaining similar
wrinkle patterns.
3
Composites Part B 253 (2023) 110536
2.2. Deep learning model
The method for the development of the DL surrogate model for rapid
wrinkle prediction consists of four stages: tool geometry generation,
forming data generation, data pre-processing, and model training and
evaluation.
2.2.1. Tool geometry generation
Geometry data generation consists of generating a large set of
differing tool geometries, characterising them, filtering out all those
that are unsuitable for forming and finally filleting the edges of the
suitable geometries.
A code-based CAD geometry generator, implemented in ‘CADQuery’
[32], is developed (Fig. 1a) and used to generate an initial set of 10,000
male tool geometries, with examples shown in (Fig. 1b). The generator
is designed to generate a wide variety of contrasting geometries based
on varying 12 independent variables related to three sets of lofted
polygons fitted within ellipses (Fig. 1a). The 12 variables consist of
the number of polygon sides (𝑛), major and minor axes of each ellipse
(𝑎, 𝑏, 𝑐, 𝑑, 𝑒, 𝑓 ), the relative height between them (ℎ𝑡 ) and their
relative rotation (𝜃1 , 𝜃2 ,) and lateral shift (𝛥𝑥, 𝛥𝑦) (Table 2). While
curvature is not explicitly defined as an input variable, lofting between
the vertices of three polygon vertices introduces curvature along the
𝑧 direction by fitting a curve through them. Additionally, geometries
with a large 𝑛 approximate curvature in the 𝑥–𝑦 directions, resulting
in shapes of highly contrasting mean and Gauss curvatures. Alongside
this, the twist (𝜃1 , 𝜃2 ,) and shift (𝛥𝑥, 𝛥𝑦) variables introduce relative
twist and translation respectively.
The overall height of each geometry is kept constant at 50 mm to
allow more direct comparisons between geometrical features in terms
of the resulting wrinkling behaviour. Male (convex) tooling is used
specifically because female (concave) tooling is typically avoided for
DDF as they tend to be very challenging to form with vacuum-only
forming forces due to the higher frictional forces involved.
During the semi-randomised generation process, combinations of
variables are randomly selected from predefined variable ranges (Ta-
ble 2) that were carefully chosen to produce the widest possible dis-
tribution of geometry characteristics. In order to limit the number
of geometries that are unfeasible for preforming such as those with
undercuts, certain conditions are imposed on the parameter values
(Fig. 1a). Despite these conditions, due to the introduction of twisting
and translation during generation, it is not possible to fully eliminate
all geometries with undercuts. Therefore filtering is performed using
an automated algorithm to ensure all shapes are fully formable. Be-
sides undercuts, geometries with sharp vertices are also filtered out to
avoid potential fabric penetration during finite element simulation that
would make any wrinkling patterns unrepresentative of the underlying
geometry and because during a DDF process such geometries might not
be formable due to them piercing through the diaphragm.
Out of the 10,000 geometries, 1802 suitable geometries are taken
forward to the next stage of wrinkle data generation. All generated tool
geometries are characterised based on four geometrical characteristics:
Conicity, Angularity, Mean Curvature and Gauss Curvature. As shown
in Fig. 1c, the curvatures are calculated based on surface coordinates
according to [23]. The Angularity is defined as the mean interior angle
and the Conicity is measured by the mean overall draft angle. The
convex and unimodal (single point of contact) nature of the generated
geometries allows us to make observations on the effect of these
geometrical characteristics on wrinkling severity.
2.2.2. Forming data generation
Using the 1802 suitable geometries generated, the corresponding
wrinkle data for each tool geometry is created by simulating the wrin-
J.V. Viisainen et al. Composites Part B 253 (2023) 110536

Table 1
Details related to the material, layup and forming process.
Fabric
Type Biaxial NCF
Commercial name Hexcel ‘FCIM359’
Thickness/mm (± 0.01 mm) 0.4
Areal weight (GSM)/ g m−2 441 ± 22
Tow orientation/◦ −45/45
Fibre material Carbon fibre
Stitch material Texturised polyester
Stitch type Pillar

Front/back of fabric (same for both)

Layup
Layup/◦ [0/90,0/90]
Sample size/mm × mm 200 × 200
Sample shape Square
Stitch orientation (relative to the global X direction)/◦ 45

Representation of the layup relative to the tool

Forming process
Forming process type Double diaphragm forming (DDF)
Vacuum pressure/kPa 100
Diaphragm material Stretchlon HT-350 (thermoplastic)
Diaphragm thickness/mm 0.08
Base dimensions/m ×m 1.8 × 1.5

Representation of the forming process

Table 2
Outline of the variable input parameters used for the geometry generator and the corresponding parameter ranges. The ranges are manually
selected for each parameter while the combinations of parameter values for each geometry are randomly assigned from these ranges, subject
to certain conditions. The referenced ellipses and polygons are defined in Fig. 1a.
Symbol/unit Description Parameter range
𝑛 Sides (number of sides for each polygon) [3, 4, 5, 6, 8, 10, 30, 50]
ℎ𝑡 /mm Top height (height between top and middle polygon) [1,2… 30]
𝑎/mm Bottom ellipse major axis [50,51… 99,100]
𝑏/mm Bottom ellipse minor axis [50,51… 99,100]
𝑐/mm Middle ellipse major axis [25,26… 99,100]
𝑑/mm Middle ellipse minor axis [25,26… 99,100]
𝑒/mm Top ellipse major axis [0.25… 100]
𝑓 /mm Top ellipse minor axis [0.25… 100]
𝑔 Top ellipse/middle ellipse length ratio [1/100, 1/10,1/4,1/2,3/4,7/8,1]
𝜃1 ∕◦ Bottom polygon rotation anticlockwise relative to x axis [−90,−45,0,45,90]
𝜃2 ∕◦ Middle polygon rotation anticlockwise relative to x axis [−90,−45,−30,−15,0,15, 30, 45, 90]
𝛥𝑥&𝛥𝑦/mm Middle polygon shift in x and y, relative to centre of bottom polygon [−18.75… 18.75]

kling behaviour of a two-ply layup over each geometry during DDF. For
simulation, the generated tool geometries are meshed in Abaqus using
a combination of quadrilateral rigid elements (𝑅3𝐷4) and triangular
elements (𝑅3𝐷3) based on the ‘Free’ meshing algorithm and a target
mesh size of 1 mm. All the FE simulations are run using a high-end
CPU: the Intel i9-10980XE @3.0 GHz with 18 total cores and 64 GB of
RAM, which is able to run up to 16 simulations in parallel. The average
runtime for the completed simulations was found to be 11.5 h (hours)
and all the simulations were completed in less than two months (1307
h or 54 d (days)).

4
J.V. Viisainen et al.
Fig. 2. Processing of finite element data into grayscale images that are used to train the deep learning model: (a) tool height data and (b) wrinkle amplitude data.
Fig. 3. The ten geometries selected to evaluate the performance of the deep learning surrogate model.
2.2.3. Data pre-processing
The DL model utilised in this work requires data in the form
of grayscale images, where the input images correspond to the tool
geometries and the output images correspond to the wrinkle patterns at
the end of forming. As a result, these sets of grayscale images (Fig. 2)
are extracted from the FE data. For the output images, this is done
by first calculating the wrinkle amplitudes at each position on the
deformed fabric and then mapping the wrinkling results back onto the
initially flat, undeformed fabric, such that there is no data loss when
these surfaces are converted into 2D grayscale images. The wrinkle
amplitude surfaces are calculated based on the method in [4], where
the wrinkle amplitude (𝑎𝑤 ) is defined to be the magnitude of the out-
of-plane displacement of the deformed fabric surface in the direction
of the surface normal of a wrinkle-free ‘reference surface’. To achieve
a similar result for the input images, the tool geometry coordinates are
projected (with no scaling) onto a contour map of the same dimensions
as the undeformed fabric and then turned into a grayscale image, with
the local tool height represented by the local intensity of the image.
The tool height and wrinkle amplitude images are used to train the DL
surrogate model.
To assess the severity of wrinkling patterns presented in the grayscale
images, three wrinkling metrics are used: the maximum absolute wrin-
kle amplitude (|𝑎𝑤 |𝑚𝑎𝑥 ), the wrinkle area (𝐴𝑤 ) and the mean absolute
wrinkle amplitude (|𝑎𝑤 |). Based on the linear relationship between the
local intensities of the image (𝐼(𝑖)) and the wrinkle amplitude at any
given point (𝑎𝑤 (𝑖)), the absolute wrinkle amplitude map (|𝑎𝑤 (𝑖)|) from
each image can be obtained as follows:
2𝑎𝑤𝑚𝑎𝑥,𝑠𝑒𝑡
|𝑎𝑤 (𝑖)|[mm] = | 𝐼(𝑖) − 𝑎𝑤𝑚𝑎𝑥,𝑠𝑒𝑡 |
𝐼𝑚𝑎𝑥
Composites Part B 253 (2023) 110536
where 𝑖 is the 𝑖th pixel of the image, 𝑎𝑤𝑚𝑎𝑥,𝑠𝑒𝑡 = 38.8 mm (maximum
amplitude across all geometries) and 𝐼𝑚𝑎𝑥 = 255. Note that the max-
imum amplitude recorded is larger than would typically be expected
of wrinkles due to an artefact of the wrinkle amplitude calculation
method [4]: for a limited set of geometries with sharp vertical edges
(low Conicity), it determines the local reference wrinkle surface vector
to be near parallel to the z axis resulting in large ‘wrinkles’ in these local
regions. The typical wrinkle amplitudes are significantly lower than
this as shown by the mean values recorded in e.g. Fig. 6b. However to
capture the whole range of amplitudes for all geometries, it is necessary
to use the largest recorded value as the scaling factor.
From this, the metrics can be calculated as follows:
• Max wrinkle amplitude:
|𝑎𝑤 |𝑚𝑎𝑥 [mm] = 𝑚𝑎𝑥(|𝑎𝑤 (𝑖)|) (2)
• Wrinkle area:
𝑁|𝑎𝑤 (𝑖)|>1 mm
𝐴𝑤 [%] = (3)
𝑁
• Mean wrinkle amplitude:
|𝑎𝑤 |[mm] = |𝑎𝑤 (𝑖)|>1 mm (4)
where 𝑁|𝑎𝑤 (𝑖)|>1 mm is the number of pixels with an absolute wrinkle
amplitude greater than 1 mm, 𝑁 is the total number of pixels in the
image and |𝑎𝑤 (𝑖)|>1 mm corresponds to all image pixels with an absolute
wrinkle amplitude greater than 1 mm. The threshold limit of 1 mm
(1) was chosen to be consistent with a previous study where this threshold
5
J.V. Viisainen et al. Composites Part B 253 (2023) 110536

Fig. 4. (a) Outline of the fully convolutional network (FCN) used for surrogate model. (b) The effect of increasing the training set size on the maximum mean image accuracy
(MS-SSIM) obtained for the test set. The test set size is kept constant at 186 throughout.

was used for wrinkle area [4]. Choosing a different threshold could Table 3
The chosen hyperparameter values for the surrogate
have an impact on the calculated wrinkle area values and the extent
model.
of wrinkling observed [3], but other thresholds were not considered
Hyperparameter Value/description
in this study as it was deemed that these potential variations could
Loss function 𝐿𝑀𝑆−𝑆𝑆𝐼𝑀−𝐿1
be mitigated through the use of the two other wrinkling metrics in
Alpha (𝛼) 0.4
combination. Learning rate (𝑙𝑟) 2 × 10−4
Batch size (𝑏) 16
2.2.4. Model training and evaluation Weight decay (𝑤𝑑 ) 0
The trained DL surrogate model is obtained by splitting the image Epochs (𝑒𝑝) Early stopping
data into subsets, selecting a suitable neural network architecture and
training the model. The predictive capability of the surrogate model
is then evaluated by comparing predicted results against the ‘ground results for this model than either of the loss functions that it makes
truth’ (or expected results) from the finite element model. up. The model is trained using the popular ‘Adam’ algorithm [39] and
The total dataset is divided into two subsets: the training set and to avoid overfitting, an ‘early stopping’ algorithm is utilised to stop
the test set. Approximately 90% (1616) of the total dataset (1802)
training once a maximum has been reached. The hyperparameters for
is used for the training set, with 10% of the data reserved for the
the model, including the value of 𝛼, are optimised based on achieving
test set (186). The test set is used to optimise the performance of the
the maximum image accuracy (𝑀𝑆 − 𝑆𝑆𝐼𝑀) values on the test set
model for maximum accuracy, and evaluate the distributions in model
(Table 3).
predictions.
The necessary size of the training set to achieve maximum model
Additionally, an evaluation set of ten geometries is created to
performance is evaluated in Fig. 4b, showing that the model converges
evaluate the final performance of the model for contrasting geometries
towards a maximum 𝑀𝑆 − 𝑆𝑆𝐼𝑀 value on the test set when the
(Fig. 3). Rather than being generated like the test set geometries, these
evaluation geometries are selected based on actual forming geometries training set is around 1000 images (keeping the test set constant at
used in previous preforming experiments and thus allow the assessment 186). This suggests that the training set could in fact be reduced while
of the model’s ‘real world’ performance. still achieving the same performance. In spite of this, to make use of
The FCN architecture utilised for the DL model is shown in Fig. 4a. all available data, the training set size is kept at the maximum of 1616
The architecture is primarily inspired by the ‘DeconvNet’ [33] and in the results that follow.
the ‘SegNet’ [34], which are FCN versions of the original ‘VGG16’ The wrinkle predictions of the optimised model are evaluated both
CNN [35]. It is also similar to the FCN used by Zimmerling et al. on the test set and the evaluation set, with the test set used to better
[8] that was shown to be fit for purpose for use as a preforming understand the model’s prediction accuracy while the evaluation set is
surrogate model. To assess the accuracy of the surrogate model, the used to assess the model’s capability to generalise to new geometries.
multi-scale structural similarity metric (𝑀𝑆 − 𝑆𝑆𝐼𝑀) [36] is used and To evaluate how well the model wrinkle predictions (𝑝) compare with
it is implemented as per the ‘PIQ’ metric library [37]. the expected (𝑒) ground truth from the FE simulation, the three wrin-
The loss function used for training the model, 𝐿𝑀𝑆−𝑆𝑆𝐼𝑀−𝐿1 , is a kling metrics are applied to both sets of images. The relative wrinkle
combination of the mean absolute error loss function (𝐿1 ) and a loss errors, based on each of the wrinkling metrics, are calculated as follows:
function based on 𝑀𝑆 − 𝑆𝑆𝐼𝑀 metric:
• Max wrinkle error:
𝐿𝑀𝑆−𝑆𝑆𝐼𝑀−𝐿1 = 𝛼𝐿𝑀𝑆−𝑆𝑆𝐼𝑀 + (1 − 𝛼)𝐿𝐿1 (5) ||𝑎𝑤 |𝑚𝑎𝑥𝑝 − |𝑎𝑤 |𝑚𝑎𝑥𝑒 |
𝛥|𝑎𝑤 |𝑚𝑎𝑥
[%] = (7)
where |𝑎𝑤 |𝑚𝑎𝑥 |𝑎𝑤 |𝑚𝑎𝑥𝑒

𝐿𝑀𝑆−𝑆𝑆𝐼𝑀 = 1 − (𝑀𝑆 − 𝑆𝑆𝐼𝑀) (6) • Wrinkle area error:

and 𝛼 = [0, 1]. This loss function is chosen based on its successful use 𝛥𝐴𝑤 |𝐴𝑤𝑝 − 𝐴𝑤𝑒 |
[%] = (8)
in a previous work [38] and because it was found to achieve better 𝐴𝑤 𝐴𝑤𝑒

6
J.V. Viisainen et al. Composites Part B 253 (2023) 110536

Fig. 5. The linear correlation between the eight tool geometry characteristics and the mean wrinkle amplitude, descending order of the squared coefficients of correlation (𝑟2 ):
(a) Conicity, (b) Gauss Curvature, (c) Mean curvature and (d) Angularity. For each correlation, the 95% prediction limits and the 95% confidence limits are also calculated.

• Mean wrinkle error: particular geometries influence the shearing behaviour of the NCFs dur-
ing the DDF process, with wrinkles for this fabric forming due to lateral
𝛥|𝑎𝑤 | ||𝑎𝑤 |𝑝 − |𝑎𝑤 |𝑒 |
[%] = (9) tow compression in positive shear or due to lateral fabric compression
|𝑎𝑤 | |𝑎𝑤 |𝑒 in negative shear once the shear locking angle has been reached [4].
where 𝑝 corresponds to the predicted image and 𝑒 corresponds to the Based on how the tool geometry enables or restricts shear along its
expected image. These error metrics are used to evaluate the predictive surface, it can influence not only the severity of these wrinkles but also
capabilities of the DL model. where are they likely to occur, resulting in a variety of wrinkle patterns.
In addition to calculating the relative prediction error based on each Fig. 5 shows that certain characteristics are significantly correlated with
wrinkle metric, the coefficient of determination (𝑅2 ) is calculated to the resulting wrinkling severity while some characteristics have little to
compare how well the predicted values match up to expected values for no correlation with wrinkling severity.
each wrinkle metric. Based on fitting a line 𝑦𝑝 = 𝑦𝑒 over all predictions Based on Fig. 5a, the mean wrinkle amplitude (|𝑎𝑤 |) is most strongly
from the test set, 𝑅2 for metric 𝑦 can be found as follows: correlated (𝑟2 = 0.53) with the Conicity of the geometry. This is a
∑ negative correlation and thus geometries that are more cone-like (with
𝑆𝑆𝑟𝑒𝑠 (𝑦𝑒 − 𝑦𝑝 ) a higher mean overall draft angle 𝜃𝑑𝑜 ) tend to result in lower mean
𝑅2 = 1 − =1− ∑ (10)
𝑆𝑆𝑡𝑜𝑡 (𝑦𝑒 − 𝑦𝑝 ) wrinkle amplitudes. Based on the line of best fit and 95% prediction
intervals, |𝑎𝑤 | can drop from 4.9 ± 1 mm for 𝜃𝑑𝑜 = 0◦ to 2.9 ± 1 mm for
where 𝑦𝑝 is the mean of 𝑦𝑝 and 𝑦 = |𝑎𝑤 |𝑚𝑎𝑥 , 𝐴𝑤 or |𝑎𝑤 |. To analyse 𝜃𝑑𝑜 = 45◦ , representing a 40% drop in mean severity.
the effect that particular geometrical characteristics of the tool have In contrast to the Conicity, the Gauss Curvature (Fig. 5b, 𝑟2 = 0.16)
on the wrinkling behaviour of the preform, correlation analysis is is found to be weakly correlated with wrinkling severity. This suggests
performed to evaluate the linear coefficient of correlation (𝑟2 ) between that geometries with negative Gauss Curvature (inwardly curving)
each characteristic and the mean wrinkle amplitude metric. After fitting tend to wrinkle more than those with positive 𝐾𝑟 (outwardly curving)
a line of best fit 𝑦̃ = 𝑚𝑥 + 𝑏 between a characteristic 𝑥, and a wrinkling but the confidence in this correlation is dampened by the large 95%
metric 𝑦, 𝑟2 is obtained as follows: confidence intervals for 𝐾𝑟 < 1.5 × 10−3 mm−2 , as most geometries are
∑
𝑆𝑆𝑟𝑒𝑠 (𝑦 − 𝑦)
̃ concentrated close to 𝐾𝑟 = 0 mm−2 . The two remaining characteristics,
𝑟2 = 1 − =1− ∑ (11) Mean Curvature (Fig. 5c, 𝑟2 = 0.04) and Angularity (Fig. 5d, 𝑟2 = 0.00),
𝑆𝑆𝑡𝑜𝑡 (𝑦 − 𝑦)
show little to no correlation with overall wrinkling severity.
where 𝑦 is the mean of 𝑦.
3.2. Surrogate model performance
3. Results
3.2.1. Performance on test set
3.1. Effect of geometry on NCF wrinkling The performance of the surrogate model based on the test set is
shown in Fig. 6. Fig. 6a shows that the predicted images have a high
The wrinkling severity of the NCF plies depends significantly on level of accuracy given that 75% of the predictions have a 𝑀𝑆 −𝑆𝑆𝐼𝑀
the tool geometry being formed, and certain characteristics of these greater than 0.86 (where 𝑀𝑆 − 𝑆𝑆𝐼𝑀 = 1 corresponds to a per-
geometries have a greater influence of the wrinkling behaviour than fect match), and predictions above this value are qualitatively shown
others. The differences observed in wrinkling severity are due to how to be of ‘good’ accuracy relative to the ground truth. Furthermore,

7
J.V. Viisainen et al.
Fig. 6. (a) The probability density distribution of the image accuracy across the 186 test set geometries with the 5th, 25th, 50th (median), 75th and 95th percentiles shown,
including inset examples of poor (left), good (middle) and excellent (right) wrinkle predictions. (b) The predicted mean wrinkle amplitude vs. the expected mean wrinkle amplitude
for all the geometries in the test set, with the coefficient of determination (𝑅2 ) of the data with respect to the line 𝑦𝑝 = 𝑦𝑒 reported. The 95% prediction limits of the data are
additionally shown.
Fig. 6b shows that the mean wrinkle amplitude has a coefficient of
determination (𝑅2 ) of 0.71 compared to the expected values from the
FE model and the line 𝑦𝑝 = 𝑦𝑒 falling within the 95% prediction limits.
While not shown, it was found that the model was less accurate in its
predictions of the max wrinkle amplitude (𝑅2 = 0.53) and the wrinkle
area (𝑅2 = 0.53), which are both somewhat under-predicted by the
model, as shown by their 𝑅2 values being lower compared to the mean
wrinkle amplitude. However, the surrogate model predictions are still
sufficient as first estimates of wrinkling severity. This is because the
relative differences between tool geometries can be established based
on these predictions and therefore the tool geometries resulting in the
lowest wrinkling severity can be identified.
3.2.2. Performance on evaluation set
The model predictions for the previously unseen evaluation set,
shown in Fig. 7, highlight that the surrogate model can successfully
predict the wrinkling patterns for certain geometries but for geometries
that are noticeably different to those in the training set, it is less
accurate.
Fig. 7a shows the normalised wrinkling severity of the ground truth
wrinkle patterns for each evaluation geometry based on three wrinkling
severity metrics, showing that the cylinder, corner, box, triangular
prism and elongated hemisphere have the highest wrinkling severity
out of the ten geometries.
The corresponding error in the wrinkle predictions of the surrogate
model for each evaluation geometry is shown in Fig. 7b, with the
geometries arranged in order of increasing relative prediction error
(the mean of the three relative error metrics) from left to right. Thus
it shows that the cylinder is predicted with the lowest error and the
pyramid is predicted with the largest error. There is a significant range
in the measured errors across the geometries, for example the wrinkle
𝛥𝐴
area error ( 𝐴 𝑤 ), which ranges from near 0% for the cylinder to over
𝑤
120% for the pyramid. Also, the trained model is shown to be better at
predicting the mean wrinkle amplitude rather than the wrinkle area or
𝛥|𝑎𝑤 |
the max wrinkle amplitude, given that the mean wrinkle error (
|𝑎𝑤 |
is generally smaller across the 10 geometries than the other two error
metrics.
Composites Part B 253 (2023) 110536
The qualitative predictive accuracy of the DL model for the eval-
uation geometries is shown in Fig. 7c. The first row (‘Input’) shows
the input tool geometries, the second row (’Exp. Output’) shows the
corresponding ground truth wrinkle patterns from the FE simulations,
the third row shows the predicted wrinkle patterns for each evaluation
geometry from DL model (‘Pred. Output’) and the last row (‘Diff’) shows
the absolute difference in intensity between the FE and DL wrinkle
patterns. The wrinkle prediction accuracy based on the 𝑀𝑆 − 𝑆𝑆𝐼𝑀
metric is also shown above each geometry. This figure allows the
predictions to be qualitatively evaluated and shows that the first four
wrinkle patterns from the left (cylinder, elongated hemisphere, corner
and box) are visually well predicted while more noticeable errors in
the location and/or numbers of wrinkles are present for the other
six geometries. For example, the predicted pattern for the hemisphere
has only one noticeable wrinkle whereas four were expected, although
that one wrinkle is of the expected size and amplitude. The wrinkling
patterns of the last five geometries (double dome 1, tetrahedron, double
dome 2, triangular prism and pyramid) are also less well predicted,
each in their own particular way. The predictions for these last 5
geometries were expected to be poorer, as they represent geometries
that fall outside of the types of geometries the surrogate model was
trained on. Thus they are ‘extrapolations’ whereas the first five ge-
ometries are ‘interpolations’. It is notable that the hemisphere should
be better predicted by the surrogate model given that it represents an
interpolation of the geometries in the dataset.
Comparing the relative wrinkle error in Fig. 7b and the 𝑀𝑆−𝑆𝑆𝐼𝑀
values given in Fig. 7c highlights that while the 𝑀𝑆 − 𝑆𝑆𝐼𝑀 gives
a good first approximation of model prediction accuracy, it does not
measure wrinkling directly and so can be limited in certain instances
where wrinkle patterns are smaller and less severe such as the trian-
gular prism and the pyramid. For these cases, 𝑀𝑆 − 𝑆𝑆𝐼𝑀 values
suggest a higher level of prediction accuracy than in reality because the
non-wrinkled background becomes dominant in the metric calculation
(Fig. 7c). As a result, it is more reliable to use the specific wrinkling
metrics to make conclusions about model performance, rather than an
) image comparison metric such as 𝑀𝑆−𝑆𝑆𝐼𝑀, as the wrinkling metrics
are shown to correspond more closely with the qualitative comparisons
(Fig. 7b).
8
J.V. Viisainen et al. Composites Part B 253 (2023) 110536

Fig. 7. (a) The normalised wrinkling severity for the ten evaluation geometries based on the wrinkling data from the finite element simulation, (b) the relative wrinkle prediction
error (based on the max wrinkle error, the wrinkle area error and the mean wrinkle error) for the ten evaluation geometries, arranged from left to right descending in prediction
error (based on the mean of the three relative wrinkle error metrics), and (c) the wrinkle pattern predictions for all evaluation geometries from the surrogate model (‘Pred. Output’)
compared to the ground truth from the finite element model (‘Exp. Output’), with the tool geometry images (‘Input’) and the inverted pixel-wise difference of the wrinkle patterns
(‘Diff’) also shown.

4. Discussion terms of wrinkling defects could potentially be achieved by introducing

4.1. Mitigating wrinkling through component design
Based on the correlation analysis (Fig. 5), certain practical guide-
lines for designing components or tooling to achieve minimal wrinkling
in the final component can be suggested. Firstly, more tapered, conical
tool geometries tend to wrinkle less during double diaphragm forming
and thus introducing a larger overall draft angle into the tool geometry
is likely to be beneficial from a manufacturing point of view. For prede-
fined designs with limited scope for changes, a similar improvement in
larger fillets or chamfers onto component edges, such that transi-
tions along the geometry topology are more gradual. Furthermore,
introducing positive Gauss curvature i.e. outwardly curving surfaces
is suggested to reduce wrinkling. Finally, the study suggests that ge-
ometries having angular sides has no effect on wrinkling severity. Thus
the angularity of a tool geometry is suggested not to be critical from
a wrinkling mitigation perspective during manufacturing. While these
observations are limited to convex, unimodal geometries, they can be
extended to more complex industrially applicable geometries where
geometrical features are located far enough apart, such that they can
be considered independently from a forming perspective [40].

9
J.V. Viisainen et al.
4.2. Assessment of surrogate model
The DL surrogate model is assessed to perform well when inter-
polating to new, similar geometries and its usefulness is aided by the
potential to reduce the number of geometries required for training but,
its capacity to extrapolate to new, contrasting geometries is found to be
limited (Fig. 7). With regards to the extrapolation geometries, the two
double domes consist of two peaks that make simultaneous contact with
the fabric while all the generated geometries only have one initial point
of contact with the fabric, thus significantly affecting the shear distri-
bution and wrinkling patterns in ways that the surrogate model cannot
currently learn about, particularly as these geometries contain certain
concave features. Secondly, the tetrahedron and pyramid both consist
of sharp peaks and these types of geometries had been eliminated from
the training set at the filtering stage due to concerns about fabric
penetration during simulation creating unrepresentative results. This
limited extrapolation capability highlights the importance of selecting
the initial dataset such that it is fit for the purpose of the application.
This finding regarding limited extrapolation capabilities is in agreement
with those of Zimmerling et al. [8] for the prediction of shear angle
patterns, who found their FCN surrogate model was only effective on
convex shapes but not on concave shapes, which were outside of the
training set.
The usefulness of the surrogate model derives from how fast it is
to obtain predictions from the trained model. The computational cost
of using the trained surrogate model to get wrinkle predictions for one
tool geometry is found to be 6000 times lower than the alternative of
running the equivalent finite element model of comparable accuracy.
The average computational cost for one image prediction using the sur-
rogate model is less than 1 s (0.7 s or 0.000215 h) using a conventional
CPU while the cost of running the FE model with a high-performance
CPU and parallelised over 16 cores is 1.33 h. This makes the model
useful in a number of applications (see Section 4.3). Furthermore, the
more complex the phenomenon that is being simulated, the higher the
benefit of using a trained surrogate model as the prediction cost, in
contrast to running the parent model, does not increase with level of
complexity.
However, a limitation of the model is that its development has
significant computational costs due to, primarily, the need to run thou-
sands of simulations to generate the wrinkling dataset for training. The
generation of the wrinkling data using finite element simulations for
1802 geometries is calculated to be approximately 99.3% of the total
development cost of 1316 h (or 55 d). However, this total development
time could be reduced significantly given that it was found that the
performance of the model converges with a training dataset that is
about half the size of the 1616 training set images simulated for this
study (see Fig. 4b).
In contrast, the training of the surrogate model is only 0.7% of
the total time at 9 h, while the cost of prediction is negligible in
comparison. As the time for training the model (9 h) is of the same
order of magnitude as running a FE model for wrinkle prediction
(1.33 h), it makes training new networks efficient once a suitable
training dataset has been obtained.
4.3. Potential applications of surrogate model
Due to the low computational cost of the DL surrogate model com-
pared to its high-fidelity parent model, the DL model holds potential
for component geometry optimisation. As an example, by incorporating
the DL surrogate model within an optimisation scheme consisting of
a parametrised CAD model of an initial component design, all the
potential parameter combinations can be iteratively evaluated until an
optimal geometry for minimal wrinkling is obtained. Such an approach
was recently employed by Hao et al. [41] for the design optimisation of
curvilinearly stiffened panels, showing that it achieved much-improved
component performance at a lower computational cost compared to
10
Composites Part B 253 (2023) 110536
other surrogate-based optimisation methods. The surrogate model can
also identify relative differences in wrinkling between potential geome-
tries and thus allows for a time-efficient exploration of the potential
design space [40].
Furthermore, in contrast to FE models that require significant levels
of technical expertise for use for fabric forming, pre-trained DL sur-
rogate models could become particularly useful within the composites
manufacturing industry due to their ease of use and rapid predictions.
Thus they could replace the default methods of kinematic modelling,
and trial and error testing. While this work focused on predictions
based on changes in tool geometry, similar surrogate models can be
trained based on changes in different forming process parameters [27].
The various factors affecting formability and wrinkling could also be
incorporated within the same surrogate model by expanding dataset
and model inputs, as has been done within metal forming [22], giv-
ing engineers a simple tool to obtain rapid forming results for any
given forming case without having to resort to expensive experimental
testing.
Surrogate modelling could also be useful for the simulation of
large scale components of multiple metres long where the defects
occur within local regions of that component at a much smaller scale
(e.g. aerospace components). Based on a global-to-local modelling strat-
egy [17] and assuming that the regions of wrinkling are far enough
from each other to be independent [40], high accuracy local FE models
could be replaced by a trained surrogate models that obtain predictions
significantly faster.
4.4. Limitations
While advancing the understanding around the effect of tool geom-
etry and showing the possibilities of DL surrogate models for predicting
wrinkling, this study has certain limitations that need to be highlighted.
The wrinkling severity observed in this study is exaggerated by the
use of a lower bending stiffness than measured experimentally and also
by characterising the bending stiffness as linear. Thus for more accurate
results, a non-linear bending model should be implemented [14,16],
although this would add significant computational cost.
The high initial computational cost for generating the simulated
data used to train the surrogate model is a limitation, as it makes
such a surrogate model time-consuming to implement and cumbersome
to make changes to. This illustrates why previous forming surrogate
models (e.g. [8]) have relied upon vastly computationally cheaper par-
ent models, sacrificing accuracy for the sake of accessibility. However,
using transfer learning [42] previously trained models can be retrained
for a new task based on smaller sets of new data and significantly
reduce computational costs.
5. Conclusions
The wrinkling severity during composites forming is influenced by
the geometrical characteristics of the preform. The greater the tapering
of a convex tool geometry is, the less severe the resultant wrinkling is
likely to be. In addition, outwardly curving geometries tend to form less
severe wrinkles than zero curvature or inwardly curving geometries.
However, the angularity of a tool geometry’s vertices tends to have
limited impact on wrinkling severity.
A deep learning surrogate model can be successfully developed
for the prediction of wrinkling severity over a given tool geometry.
Furthermore, the predictions of the surrogate model are found to be
sufficiently accurate for geometries similar to those it was trained on
but its ability to extrapolate to very different geometries is limited.
However, it is shown to be beneficial due to the reduced computational
cost of prediction. Once the training has been completed, the compu-
tational cost of obtaining a wrinkling prediction using the trained deep
learning surrogate model is approximately 6000 times lower than the
equivalent finite element model.
J.V. Viisainen et al.
Declaration of competing interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Data availability
I have shared the link to my data at the Attached File step.
Acknowledgements
This work was funded by the Engineering and Physical Sciences
Research Council (EPSRC), via an ‘‘EPSRC Doctoral Training Partner-
ship (DTP)’’ studentship and grant number EP/P006701/1, as part of
the EPSRC Future Composites Manufacturing Research Hub. For the
purpose of open access, the authors have applied a Creative Commons
Attribution (CC BY) licence to any Author Accepted Manuscript version
arising from this submission.
References
[1] Potter KD, Khan B, Wisnom MR, Bell T, Stevens J. Variability, fibre waviness
and misalignment in the determination of the properties of composite materials
and structures. Composites A 2008;39(9):1343–54. http://dx.doi.org/10.1016/j.
compositesa.2008.04.016.
[2] Arnold SE, Sutcliffe MPF, Oram WLA. Experimental measurement of wrinkle
formation during draping of non-crimp fabric. Composites A 2016;82:159–69.
http://dx.doi.org/10.1016/j.compositesa.2015.12.011.
[3] Shen H, Wang P, Legrand X, Liu L. Characterisation and optimisation of wrinkling
during the forming of tufted three-dimensional composite preforms. Composites
A 2019;127:105651. http://dx.doi.org/10.1016/j.compositesa.2019.105651.
[4] Viisainen JV, Hosseini A, Sutcliffe MPF. Experimental investigation, using 3D
digital image correlation, into the effect of component geometry on the wrinkling
behaviour and the wrinkling mechanisms of a biaxial NCF during preforming.
Composites A 2021;142:106248. http://dx.doi.org/10.1016/j.compositesa.2020.
106248.
[5] Harrison P, Alvarez MF, Anderson D. Towards comprehensive characterisation
and modelling of the forming and wrinkling mechanics of engineering fabrics.
Int J Solids Struct 2017;1–17. http://dx.doi.org/10.1016/j.ijsolstr.2016.11.008.
[6] Boisse P, Colmars J, Hamila N, Naouar N, Steer Q. Bending and wrinkling of
composite fiber preforms and prepregs. A review and new developments in the
draping simulations. Composites B 2018;141:234–49. http://dx.doi.org/10.1016/
j.compositesb.2017.12.061.
[7] Boisse P, Huang J, Guzman-Maldonado E. Analysis and modeling of wrinkling
in composite forming. J Compos Sci 2021;5(3). http://dx.doi.org/10.3390/
jcs5030081.
[8] Zimmerling C, Trippe D, Fengler B, Kärger L. An approach for rapid prediction
of textile draping results for variable composite component geometries using
deep neural networks. AIP Conf Proc 2019;2113(1):020007. http://dx.doi.org/
10.1063/1.5112512.
[9] Lee JS, Hong SJ, Yu WR, Kang TJ. The effect of blank holder force on the stamp
forming behavior of non-crimp fabric with a chain stitch. Compos Sci Technol
2007;67(3):357–66. http://dx.doi.org/10.1016/j.compscitech.2006.09.009.
[10] Yu WR, Harrison P, Long AC. Finite element forming simulation of NCF
considering natural variability of fiber direction. In: Proceedings of the 8th
international ESAFORM conference on materials forming. Cluj-Napoca; 2005.
[11] Guzman-Maldonado E, Wang P, Hamila N, Boisse P. Experimental and numerical
analysis of wrinkling during forming of multi-layered textile composites. Compos
Struct 2019;208:213–23. http://dx.doi.org/10.1016/j.compstruct.2018.10.018.
[12] Gereke T, Nezami FN, Hübner M, Döbrich O, Cherif C. Active control of textile
forming processes. In: 20th international conference on composite materials.
Copenhagen; 2015.
[13] Yu WR, Harrison P, Long A. Finite element forming simulation for non-
crimp fabrics using a non-orthogonal constitutive equation. Composites A
2005;36(8):1079–93. http://dx.doi.org/10.1016/j.compositesa.2005.01.007.
[14] Yu F, Chen S, Viisainen JV, Sutcliffe MPF, Harper LT, Warrior NA. A macroscale
finite element approach for simulating the bending behaviour of biaxial fabrics.
Compos Sci Technol 2020;191. http://dx.doi.org/10.1016/j.compscitech.2020.
108078.
[15] Thompson AJ, Belnoue JP, Hallett SR. Modelling defect formation in textiles
during the double diaphragm forming process. Composites B 2020;202(Au-
gust):108357. http://dx.doi.org/10.1016/j.compositesb.2020.108357.
[16] Yu F, Chen S, Harper LT, Warrior NA. Simulating the effect of fabric bending
stiffness on the wrinkling behaviour of biaxial fabrics during preforming.
Composites A 2021;143. http://dx.doi.org/10.1016/j.compositesa.2021.106308.
11
Composites Part B 253 (2023) 110536
[17] Yu F, Chen S, Harper L, Warrior N. Double diaphragm forming simulation using a
global-to-local modelling strategy for detailed defect detection in large structures.
Composites A 2021;147:106457. http://dx.doi.org/10.1016/j.compositesa.2021.
106457.
[18] Zimmerling C, Pfrommer J, Liu J, Beyerer J, Henning F, Kärger L. Application
and evaluation of meta-model assisted optimisation strategies for gripper assisted
fabric draping in composite manufacturing. In: 18th European conference on
composite materials. Athens; 2018.
[19] Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic
segmentation. In: Proceedings of the IEEE computer society conference on
computer vision and pattern recognition. Boston: IEEE; 2015, p. 431–40. http:
//dx.doi.org/10.1109/CVPR.2015.7298965.
[20] Zhang A, Lipton ZC, Li M, Smolar AJ. Fully convolutional networks. 2021, URL
https://d2l.ai/chapter_computer-vision/fcn.html.
[21] Seo H, Badiei Khuzani M, Vasudevan V, Huang C, Ren H, Xiao R, et al.
Machine learning techniques for biomedical image segmentation: An overview
of technical aspects and introduction to state-of-art applications. Med Phys
2020;47(5):e148–67. http://dx.doi.org/10.1002/mp.13649, arXiv:1911.02521.
[22] Zhou H, Xu Q, Nie Z, Li N. A study on using image-based machine learning
methods to develop surrogate models of stamp forming simulations. J Manuf Sci
Eng 2021;144(2). http://dx.doi.org/10.1115/1.4051604, 021012.
[23] Kurita T, Boulanger P. Computation of surface curvature from range images
using geometrically intrinsic weights. In: IAPR workshop on machine vision
applications. 1992, p. 389–92.
[24] Chen S, Harper L, Endruweit A, Warrior N. Formability optimisation of fabric
preforms by controlling material draw-in through in-plane constraints. Compos-
ites A 2015;76:10–9. http://dx.doi.org/10.1016/j.compositesa.2015.05.006, URL
http://linkinghub.elsevier.com/retrieve/pii/S1359835X15001608.
[25] Chen S, McGregor OP, Harper LT, Endruweit A, Warrior NA. Optimisation of
local in-plane constraining forces in double diaphragm forming. Compos Struct
2018;201:570–81. http://dx.doi.org/10.1016/j.compstruct.2018.06.062.
[26] Zimmerling C, Dörr D, Henning F, Kärger L. A meta-model based approach for
rapid formability estimation of continuous fibre reinforced components. AIP Conf
Proc 2018;1960(1):020042. http://dx.doi.org/10.1063/1.5034843.
[27] Pfrommer J, Zimmerling C, Liu J, Kärger L, Henning F, Beyerer J. Optimisation
of manufacturing process parameters using deep neural networks as surrogate
models. Proc CIRP 2018;72:426–31. http://dx.doi.org/10.1016/j.procir.2018.03.
046.
[28] Zimmerling C, Poppe C, Kärger L. Estimating optimum process parameters in
textile draping of variable part geometries - a reinforcement learning approach.
Procedia Manuf 2020;47:847–54. http://dx.doi.org/10.1016/j.promfg.2020.04.
263.
[29] Zimmerling C, Poppe C, Stein O, Kärger L. Optimisation of manufacturing process
parameters for variable component geometries using reinforcement learning.
Mater Des 2022;214:110423. http://dx.doi.org/10.1016/j.matdes.2022.110423,
URL https://www.sciencedirect.com/science/article/pii/S0264127522000442.
[30] Viisainen JV, Sutcliffe MPF. Characterising the variability in wrinkling during
the preforming. Composites A 2021;149:106536. http://dx.doi.org/10.1016/j.
compositesa.2021.106536.
[31] Chen S, McGregor OP, Endruweit A, Elsmore MT, De Focatiis DS, Harper LT,
et al. Double diaphragm forming simulation for complex composite structures.
Composites A 2017;95:346–58. http://dx.doi.org/10.1016/j.compositesa.2017.
01.017.
[32] Cowden D. CadQuery. 2020, URL https://github.com/CadQuery/cadquery.
[33] Noh H, Hong S, Han B. Learning deconvolution network for semantic segmen-
tation. In: 2015 IEEE international conference on computer vision. 2015, p.
1520–8. http://dx.doi.org/10.1109/ICCV.2015.178.
[34] Badrinarayanan V, Kendall A, Cipolla R. SegNet: a deep convolutional encoder-
decoder architecture for image segmentation. IEEE Trans Pattern Anal Mach
Intell 2017;39(12):2481–95. http://dx.doi.org/10.1109/TPAMI.2016.2644615,
arXiv:1511.00561.
[35] Simonyan K, Vedaldi A, Zisserman A. Deep inside convolutional networks:
visualising image classification models and saliency maps. In: 2nd international
conference on learning representations, ICLR 2014 - workshop track proceedings.
Banff; 2014, arXiv:1312.6034.
[36] Wang Z, Simoncelli E, Bovik A. Multiscale structural similarity for image quality
assessment. In: The 37th asilomar conference on signals, systems and computers.
Vol. 2, 2003, p. 1398–402. http://dx.doi.org/10.1109/ACSSC.2003.1292216.
[37] Kastryulin S, Zakirov D, Prokopenko D. PyTorch image quality: metrics and mea-
sure for image quality assessment. 2019, URL https://github.com/photosynthesis-
team/piq.
[38] Zhao H, Gallo O, Frosio I, Kautz J. Loss functions for neural networks for image
processing. 2015, arXiv preprint arXiv:1511.08861, URL http://arxiv.org/abs/
1511.08861.
J.V. Viisainen et al.
[39] Kingma DP, Ba JL. Adam: A method for stochastic optimization. In: 3rd
international conference on learning representations, ICLR 2015 - conference
track proceedings. 2015, arXiv:1412.6980.
[40] Zimmerling C, Dörr D, Henning F, Kärger L. A machine learning assisted
approach for textile formability assessment and design improvement of com-
posite components. Composites A 2019;124:105459. http://dx.doi.org/10.1016/
j.compositesa.2019.05.027.
12
Composites Part B 253 (2023) 110536
[41] Hao P, Liu D, Zhang K, Yuan Y, Wang B, Li G, et al. Intelligent layout design
of curvilinearly stiffened panels via deep learning-based method. Mater Des
2021;197:109180. http://dx.doi.org/10.1016/j.matdes.2020.109180.
[42] Yosinski J, Clune J, Bengio Y, Lipson H. How transferable are features in
deep neural networks? Adv Neural Inf Process Syst 2014;4(January):3320–8,
arXiv:1411.1792.