materials

Article
A New Method to Predict Damage to Composite Structures
Using Convolutional Neural Networks
Laurent Mezeix 1 , Ainhoa Soldevila Rivas 2 , Antonin Relandeau 2 and Christophe Bouvet 3, *

1 Faculty of Engineering, Burapha University, 169 Long-Hard Bangsaen Road, Chonburi 20131, Thailand;
[email protected]
2 INSA Toulouse, 135 Avenue de Rangueil, CEDEX 4, 31077 Toulouse, France;
[email protected] (A.S.R.); [email protected] (A.R.)
3 INSA/ISAE-SUPAERO/IMT Mines Albi/UPS, Institut Clément Ader (CNRS UMR 5312),
Université de Toulouse, 10 av. E. Belin, CEDEX 4, 31055 Toulouse, France
* Correspondence: [email protected]

Abstract: To reduce the cost of developing composite aeronautical structures, manufacturers and
university researchers are increasingly using “virtual testing” methods. Then, finite element methods
(FEMs) are intensively used to calculate mechanical behavior and to predict the damage to fiber-
reinforced polymer (FRP) composites under impact loading, which is a crucial design aspect for
aeronautical composite structures. But these FEMs require a lot of knowledge and a significant
number of IT resources to run. Therefore, artificial intelligence could be an interesting way of sizing
composites in terms of impact damage tolerance. In this research, the authors propose a methodology
and deep learning-based approach to predict impact damage to composites. The data are both
collected from the literature and created using an impact simulation performed using an FEM. The
data augmentation method is also proposed to increase the data number from 149 to 2725. Firstly,
a CNN model is built and optimized, and secondly, an aggregation of two CNN architectures is
proposed. The results show that the use of an aggregation of two CNNs provides better performance
than a single CNN. Finally, the aggregated CNN model prediction demonstrates the potential for
CNN models to accelerate composite design by showing a 0.15 mm precision for all the length
Citation: Mezeix, L.; Rivas, A.S.;
measurements, an average delaminated surface error of 56 mm2 , and an error rate of 7% for the
Relandeau, A.; Bouvet, C. A New
prediction of the presence of delamination.
Method to Predict Damage to
Composite Structures Using
Keywords: Convolutional Neural Network (CNN); carbon fiber-reinforced polymer; composite;
Convolutional Neural Networks.
Materials 2023, 16, 7213. https://
impact; impact damage
doi.org/10.3390/ma16227213

Academic Editors: Roberto Citarella,

Michele Guida and Alessandro
1. Introduction
De Luca
Carbon fiber-reinforced polymers (CFRPs) are widely adopted by numerous industries
Received: 12 October 2023 due to their high modulus and strength, and low density, which allow for weight reduc-
Revised: 6 November 2023 tion. Moreover, CFRPs show excellent fatigue resistance, creep, and corrosion resistance
Accepted: 7 November 2023 properties [1]. Despite their advantages, CFRPs’ vulnerability to impact is a significant
Published: 17 November 2023
concern [2–4]. Maintenance tool drops or debris impact during service can cause damage
within the composite structure’s laminate, while leaving only subtle surface indentations [5].
Internal damage typically includes delamination, matrix cracking, and fiber rupture [6,7].
Copyright: © 2023 by the authors.
Low-velocity impact can considerably reduce a CFRP’s residual strength, especially its
Licensee MDPI, Basel, Switzerland. compressive strength [8–11]. Even though aramid fibers present good impact performance,
This article is an open access article the weak interfacial adhesion between aramid fibers and the resin matrix is considered a
distributed under the terms and significant limitation to its utilization [12].
conditions of the Creative Commons The issue of impact damage dictates a damage tolerance approach in the field of
Attribution (CC BY) license (https:// composite structure design to ensure that the structure is able to withstand in-service loads,
creativecommons.org/licenses/by/ even if the damage is undetectable. This design philosophy of impact damage tolerance
4.0/). has led to standards for composite coupons under low-velocity/low-energy impact [13]

Materials 2023, 16, 7213. https://doi.org/10.3390/ma16227213 https://www.mdpi.com/journal/materials

 Materials
Materials 2023,2023, 16, x FOR PEER REVIEW
16, 7213 2 of 220of 18

andhasCompression
led to standards forImpact
After composite
[5]. coupons
Figure 1under low‐velocity/low‐energy
illustrates the impact geometry impact [13] by
defined
and Compression After Impact [5]. Figure 1 illustrates the impact geometry defined by
the standard ASTM D7136 and various types of impact damage corresponding to different
the standard ASTM D7136 and various types of impact damage corresponding to different
levels of impact energy [2]. In the initial stage, damage manifests as small matrix cracks as
levels of impact energy [2]. In the initial stage, damage manifests as small matrix cracks
the impact causes minimal denting. As the impact energy increases, delamination occurs,
as the impact causes minimal denting. As the impact energy increases, delamination oc‐
leading to larger dents. In the second stage, all three types of damage—matrix cracks,
curs, leading to larger dents. In the second stage, all three types of damage—matrix cracks,
delamination, and fiber breakage—can occur, making visual inspection easier due to the
delamination, and fiber breakage—can occur, making visual inspection easier due to the
larger
largerdent
dentsize
size(dent
(dentdepth
depth and
and diameter). Thepresence
diameter). The presenceofoffiber
fiber breakage
breakage in in
thisthis stage
stage
aids in achieving more accurate inspection results. However, this can negatively
aids in achieving more accurate inspection results. However, this can negatively impact impact
thethe
material’s
material’sresidual
residual strength afterthe
strength after theimpact,
impact, contributing
contributing to complexity
to the the complexity
of the of
in‐the
interaction between various damage types during an impact event. Finally,
teraction between various damage types during an impact event. Finally, in the third in the third
stage, the damage becomes visible to the naked eye as perforation
stage, the damage becomes visible to the naked eye as perforation occurs. occurs.

Figure (a)(a)Schematic
1. 1.
Figure Schematicofofimpact
impact test
test setup, and(b)
setup, and (b)different
differentimpact
impactdamage
damage stages
stages with
with respect
respect to to
thethe impact
impact energyand
energy andpermanent
permanentindentation
indentation size.
size.

The complex phenomena of composite damage, specifically those developing during

impact loading, depend on several parameters, such as the matrix and fiber materials, the
stacking sequence, the weaving pattern, the ply thickness, etc., and thus, make the use
of models difficult [14,15] for the design of composite structures in the aeronautical field
Materials 2023, 16, 7213 3 of 18

in terms of impact damage tolerance. In particular, it is necessary, at the same time, to

simulate the damage developing during impact, the permanent indentation caused by the
impact (which determines whether the damage can be detected during a visual inspection),
and finally, the residual strength (in order to evaluate the loss of mechanical characteristics
due to impact). Additionally, a lot of complex numerical models have been developed in
the literature [4,16–18], but this requires a lot of knowledge and additional experiments in
order to better study the proposed composite structures, and finally, there is a significant
need for IT resources to run the models (mainly finite element models). Therefore, artificial
intelligence could be an interesting way of sizing composites in impact damage tolerance
studies. The initial phase involves simulating and predicting both the impact damage and
the permanent indentation. These aspects are the focus of this article.
Machine learning, a subset of artificial intelligence, focuses on identifying patterns and
correlations within large and diverse datasets. This approach involves a stochastic process
and encompasses a wide range of algorithms, each striving to establish relationships in
the data by performing various learning tasks. Among these algorithms, Artificial Neural
Networks (ANN) are notable as universal approximators and are commonly employed
for classification and regression tasks [19–22]. Another successful approach in image
processing involves Convolutional Neural Networks (CNNs). Subsequently, CNNs have
dominated the popular ImageNet challenge, achieving outstanding results across multiple
evaluation metrics [23].
Among the various non-destructive testing (NDT) methods available for aerospace
composite structures, visual inspection stands out as a widely used approach due to its
rapid assessment of surface damage. Nonetheless, this method heavily relies on human-
related factors, making its effectiveness susceptible to human judgment and subjectivity.
Additionally, it may not be capable of detecting barely visible impact damage (BVID),
which can be challenging to identify with the naked eye. Artificial intelligence (AI)-based
techniques for detecting impact damage in composite materials can be broadly classified
into three primary types: image-based, vibration-based, and acoustic-based methods.
Image-based approaches employ computer vision algorithms to analyze images of the
composite panel’s surface before and after an impact event, enabling the detection of surface
topography changes, such as cracks and fiber breakage. Vibration-based methods focus on
measuring and analyzing the composite structure’s vibrational response to identify changes
in mechanical properties, including stiffness and damping, resulting from impact damage.
Acoustic-based techniques utilize acoustic sensors to detect changes in acoustic emission
signals generated by impact damage [22,24–26]. Numerous studies have investigated
the application of AI-based methods for detecting impact-induced damage in polymer
composite materials [27–33]. Beyond impact damage detection, AI-based techniques have
wider applications in damage classification, damage quantification, and predicting the
remaining useful life of composite materials [34]. These AI-driven approaches have the
potential to advance impact damage assessment in composite materials, enhancing their
reliability and durability. Tabatabaeian et al. successfully applied CNN models to detect
BVID from both impacted and non-impacted surface images of composites [35]. In this
study dataset, images were collected from impact tests carried out with energy from 3
J to 128 J, in a 32-ply CFRP composite [45/0/90/−45]4s configuration. Both C-scan and
visual inspection of the upper and lower surfaces were performed to create a dataset, and
different CNN models were investigated. The predictions achieved an accuracy higher
than 88.46% on the back face and a value between 51.25% and 97.05% for the impacted
surface. In their study, Wei et al. [32] employed infrared thermography data from curved
carbon fiber-reinforced polymer (CFRP) composites that had been impacted. They utilized
these data to train two distinct deep learning models. These models successfully detected
impact damage and accurately predicted the location of the damage, achieving an F1-score
of 92.74% for mid-wave infrared data and 87.39% for long-wave infrared data. In their
research, Hasebe et al. [30] utilized three machine learning models to analyze a dataset
derived from low-velocity impact tests conducted on composites. Special attention was
Materials 2023, 16, 7213 4 of 18

dedicated to three key factors: stacking sequence, impactor shape, and impact energy. The
outcomes of their study revealed that characteristics such as local volume, dent surface
gradient, and pure dent depth could effectively serve as indicators for characterizing
internal damage in CFRP laminates.
The utilization of a machine learning-based approach holds immense potential in
expediting the design process for optimal composite materials, resulting in significant
time and resource savings [36]. The prediction of composite properties takes advantage
of CNN models. The thermal conductivity properties of particle-filled 3D composites
were predicted thanks to 2D CNN models using 3000 multiple cross-section images as the
input [37]. The results showed that the use of five layers instead of three layers improved
the accuracy of the CNN model. The elastic properties of composite materials, E11 , E22
and G12 , were predicted using a CNN model [38]. In order to create the dataset, the values
corresponding to these properties were generated using an FEM. To address the significant
computational resource challenge associated with generating training data, an innovative
data augmentation scheme was introduced that enabled an increase in the dataset size from
9000 to 4.6 million samples. The results showed that the test error decreased from 2.4% to
0.4%. CNN models have been successfully used to predict the properties of composites
beyond the elastic regime, i.e., crack propagation [39]. An FEM was used to obtain training
data consisting of 26,000 configurations. Crack propagation under the quasi-static frac-
ture of elastic solids was simulated using a hybrid formulation, and the elastic modulus,
strength, and toughness were obtained from stress–strain curves. The results showed
that CNN exhibited better performance than traditional models, i.e., linear regression and
random forest. Stress–strain curves hold significance as they depict a material’s mechanical
characteristics, outlining vital traits like the elastic modulus, strength, and toughness. The
computational intensity escalates when generating these curves through numerical tech-
niques like the finite element method (FEM), particularly when encompassing the complete
failure trajectory of a material. The amalgamation of Principal Component Analysis (PCA)
and Convolutional Neural Networks (CNN) has been employed to forecast the complete
stress–strain curve of composite failures that extend beyond the elastic limit [40]. Using an
FEM, a dataset containing 100,000 distinct composite microstructures and their correspond-
ing stress–strain curves were created. This dataset was then utilized for both training and
evaluating the performance of the model. The results showed a mean absolute error of less
than 10%, demonstrating the robustness of the model. A comprehensive examination of
an ANN in the modeling of composite materials was performed [41]. A large number of
potentials used was identified: metamaterials [42–44], the mechanical behavior of yarns
in textile composites [45], and the shape/size optimization of composite structures [46].
Finally, ANN models hold the potential to address a wide array of challenges, includ-
ing unveiling unfamiliar physical principles and expediting computer simulations for
composite materials.
In this work, CNN models to predict low-velocity/low-energy impact damage to FRP
composites are explored. Initially, the data collection and preparation and the finite element
procedure employed to acquire the training data for the CNN models are discussed. Then,
two different CNN models are developed and optimized. Finally, the two CNN models’
results and validation are detailed and their performances are compared and discussed.

2. Methods
The approach to predicting impact damage to FRP composites used in this study can
be summarized as follows (Figure 2):
Data: Data collection and creation, followed by data augmentation and dataset con-
struction.
Training: Building, training, and validation of CNN models.
Prediction: Use of optimized CNN model to predict the impact damage to composites.
 Materials 2023, 16, x FOR PEER REVIEW 5 of 20

Data: Data collection and creation, followed by data augmentation and dataset con‐
struction.
Materials 2023, 16, 7213 Training: Building, training, and validation of CNN models. 5 of 18
Prediction: Use of optimized CNN model to predict the impact damage to compo‐
sites.

Figure 2. Overall methodology to predict

Figure 2. Overall impact
methodology damage
to predict impactto composites.
damage to composites.

2.1. Data 2.1. Data

The dataset was constructed by gathering information (data) identified as represent‐
The dataset was ingconstructed
key factors thatby gathering
influenced information
the impact results of (data) identified
FRP [5,6,9,10]. The aimaswas
representing
to derive
key factors that influenced the impact
valuable insights into theresults of FRP
mechanical [5,6,9,10].
behavior The aim
and performance wasvarious
of the to derive
FRP ma‐ valu-
terials
able insights into the subjected to behavior
mechanical impact. The and
data were organized into
performance of3 the
types:
various FRP materials
Sample parameters: Crucial details concerning the characteristics of the samples were
subjected to impact.captured,
The data were organized into 3 types:
such as dimensions, stacking configurations, and material properties.
Sample parameters: Crucial
Impact details concerning
test parameters: the related
Vital information characteristics
to the impactoftest
theitself
samples
was com‐ were
piled, particularly
captured, such as dimensions, the impact
stacking energy and impact
configurations, and window size. properties.
material This energy parameter
played a crucial role in assessing the material response under dynamic loading conditions.
Impact test parameters: Vital information related to the impact test itself was compiled,
Impact test results: Impact tests results, encompassing pertinent metrics such as per‐
particularly the impact energy
manent indentationandand impact
maximumwindow size. were
displacement, Thiscollated.
energy parameter played a
crucial role in assessing the material response under dynamic loading
To provide a comprehensive overview of the parameters conditions.
used in the dataset, a list is
provided in Table 1.
Impact test results: Impact tests results, encompassing pertinent metrics such as
permanent indentation and maximum displacement, were collated.
To provide a comprehensive overview of the parameters used in the dataset, a list is
provided in Table 1.

Table 1. List of the dataset parameters.

Samples Parameters Impact Test Parameters Impact Test Results

In-plane Young’s modulus (GPa) Impact window (mm2 ) Permanent indentation (mm)
Stacking (orientation of each ply) Energy of impact (J) Maximum displacement (mm)
Symmetrical test plate?
Laminate thickness (mm) Maximum force (kN)
(1 for yes, 0 for no)
Is the test dynamic or static?
Material reference (T700...)
(1 for dynamic, 0 for static) Delaminated area (mm2 )
Is there perforation?
Type of carbon (pre-impregnated or dry)
(1 for yes, 0 for no)
Thermal protection Is there delamination?
(1 for yes, 0 for no) (1 for yes, 0 for no)
Resin type (epoxy. . .)
Resin proportion
(integer between 0 and 1)
Resin property GIIC (N/mm)
Number of plies (integer)
Sandwich core type
(honeycomb, foam. . .)
Sandwich core thickness (mm)
Fiber weaving
(woven or unidirectional)
Materials 2023, 16, 7213 6 of 18

2.2. Dataset Construction

The dataset was built through spreadsheets where each influencing factor (Table 1)
was arranged as a distinct feature in columns, while rows were employed to represent
individual data source entries. To fully capture detailed information, ply characteristics
were divided into four parameters: fiber orientation, material, thickness, and weaving type.
As the largest composite lay-up consisted of 28 plies, a total of 112 columns were required
to describe ply characteristics. Finally, the dataset consisted of 142 columns in order to
cover all input information.

2.2.1. Literature Data

The presented methodology involved a meticulous examination of the literature in
order to collect impact tests results conducted on different materials. Content was based on
over 11 different studies, with over 133 different impact tests conducted on the different
materials [47–56]:
Fiber materials: carbon, Kevlar, and graphite fibers.
Resin materials: Epoxy, PEK, PPS, and PPS44.
Carbon fiber type: T700, T300, G30-500, ASA-8552, and IM7-8552.
Type of ply: prepreg or ‘dry’.
Core material: foam and honeycomb.

2.2.2. Abaqus Model

In order to enhance precision, additional data were needed; however, the available
literature does not supply an adequate amount of data. To augment the dataset’s size, an
ABAQUS model was employed. The final aim of the model was to be able to simulate the
damage, and especially to obtain the delamination surface, maximum force, deflection, and
indentation of laminates under low-velocity/low-energy impact. Therefore, the explicit
model consisted of ply-by-ply laminate and the indentor (Figure 3). Normal
Materials 2023, 16, x FOR PEER REVIEW 7 ofbehavior
20
was used as a contact property in ABAQUS between the laminate and the indentor, while
contact friction was neglected. As the indentor used was composed of a hardened steel, it
was considered
steel, rigid [57].
it was considered rigidTherefore, in the numerical
[57]. Therefore, model,model,
in the numerical the indentor was represented
the indentor was
by an analytical
represented by anrigid shell body.
analytical Duebody.
rigid shell to theDue
out-of-plane shear stress,
to the out‐of‐plane solid
shear elements
stress, solid were
required;were
elements therefore, plies
required; were modeled
therefore, withmodeled
plies were C3D8 solid
withelements.
C3D8 solidIn elements.
this study,In3 this
laminates
were investigated:
study, 3 laminates were 8-, 12-, and 16-ply8‐,laminates
investigated: of the same
12‐, and 16‐ply size of
laminates the×same
(150 size2(150
100 mm ). To×reduce
themm
100 computing time,
2). To reduce theonly a quarter
computing of only
time, the specimen was
a quarter of thesimulated
specimen was(75 × 50 mm2 ) and
simulated
symmetry
(75 × 50 mmconditions were imposed.
2) and symmetry conditionsAwere
window of 125A×window
imposed. 2
75 mm ofwas 125utilized
× 75 mm on the lower
2 was

utilized
surfaceon of the
thelower surface
laminate of the
to fix the out-of-plane
laminate to fixdisplacement.
the out‐of‐plane displacement.

Geometry
Figure3.3.Geometry
Figure of of
thethe numerical
numerical model.
model.

InInthis
thisfirst
first analysis,
analysis, anan FEA
FEA model
model waswas
usedused to determine
to determine the impact
the impact energy
energy required
required
to reach the critical force avoiding delamination. Indeed, even though a lot of
to reach the critical force avoiding delamination. Indeed, even though a lot of complex complex
and relevant FEMs exist in the literature in order to simulate the damage that develops in
composite structures during impact loading [1–7], the objective is to evaluate the ability
of the AI to predict impact damage and its detectability. The critical force necessary to
induce the beginning of delamination at the mid‐thickness of the laminate under a mode
II fracture is obtained using [6]:
Materials 2023, 16, 7213 7 of 18

and relevant FEMs exist in the literature in order to simulate the damage that develops in
composite structures during impact loading [1–7], the objective is to evaluate the ability
of the AI to predict impact damage and its detectability. The critical force necessary to
induce the beginning of delamination at the mid-thickness of the laminate under a mode II
fracture is obtained using [6]:
8π 2 Eh3
Pc2 = G I Ic (1)
9(1 − µ2 )
where E and v are the equivalent in-plane modulus and Poisson ratio for the laminate,
and h is the laminate thickness. GIIc is the fracture toughness in mode II of the composite
laminate and is taken to be equal to 1.5 N/mm [58]. This value depends on the composite
material, but in this first approach, this value was kept constant in order to highlight the
effect of the stacking sequence on the impact damage. As no damage was simulated, a
simple elastic model was used in order to easily and quickly generate additional data to
feed the AI, and linear elastic properties of unidirectional T700 carbon/epoxy ply were
chosen (Table 2).

Table 2. Mechanical properties of unidirectional T700 carbon/epoxy [59].

Properties
Tensile modulus (GPa) 135
Tensile strength (MPa) 2550
Compression strength (MPa) 1470
Flexural modulus (GPa) 120
Flexural strength (MPa) 1670

A large number of different stacking configurations were tested using Abaqus. The
stacking of 8, 12, and 16 plies was investigated, and the stacking rules followed the
aerospace design principles used in industry [60]. A total of 43 laminate stackings were
investigated (Table 3). For each configuration, the FEA model enabled us to obtain the
impact energy given the maximum force matching the one obtained using Equation (1).

Table 3. Laminate stacking investigated through explicit FEA.

Number of Plies Stacking Energy of Impact (J)

[45/−45/45/−45]s 2.3
[45/−45/45/−45]s 4
[0/45/−45/90]s 4
[0/45/−45/90]s 9
[0/45/−45/90]s 11.5
8
[0/90/0/90]s 11.5
[0/90/0/90]s 16.8
[45/90/−45/90]s 3.6
[45/0/−45/0]s 13.6
[45/0/−45/0]s 18.48
[45/−45/−45/45/0/0]s 8.992
[45/−45/90/90/90/90]s 1.688
[45/−45/−45/45/0/0]s 10.8
[45/−45/−45/45/90/90]s 2.558
12
[45/−45/0/0/0/0]s 15.9
[45/−45/−90/−45/45/0]s 6.244
[0/0/45/−45/90/90]s 10.23
[0/45/−45/−45/45/90]s 6.244
[0/0/0/45/−45/90]s 15.985
[90/90/90/45/−45/0]s 5.755
[0/0/45/−45/90/90]s 12.948
[0/45/−45/−45/45/90]s 8.402
[0/0/0/45/−45/90]s 18.473
 Materials 2023, 16, 7213 8 of 18

Table 3. Cont.

Number of Plies Stacking Energy of Impact (J)

[45/−45/0/−45/45/45/−45/90]s 2.248
[45/−45/0/0/−45/45/90/90]s 2.25
[45/0/−45/−45/0/45/92/90]s 2.247
[−45/45/0/0/0/45/−45/90]s 2.247
[45/0/0/−45/−45/90/90/45]s 2.247
[−45/0/0/45/90/90/90/45]s 2.249
[45/0/0/0/0/−45/90/90]s 7.84
[45/0/−45/90/−45/0/45/90]s 7.83
[45/−45/0/−45/45/45/−45/90]s 3.237
Materials 2023, 16, x FOR PEER REVIEW [−45/0/0/45/90/90/90/45]s 6.243
16
[45/−45/0/−45/45/45/−45/90]s 3.996
[45/−45/0/0/−45/45/90/90]s 7.83
[45/0/−45/−45/0/45/90/90]s 7.831
[−45/45/0/0/0/45/−45/90]s 6.242
2.2.3. Data Augmentation [45/0/0/ − 45/ − 45/90/90/45]s 6.243
[45/−45/0/−45/45/45/−45/90]s 4.4
The performance of[−a45/45/0/0/0/45/
CNN model relies on various
−45/90]s 7.83factors, with the d
[45/0/0/−45/−45/90/90/45]s 7.28
crucial one [39]. Boosting[45/0/0/0/0/
the size of the dataset can greatly
−45/90/90]s 8.989 improve the p
[45/0/0/0/0/
racy of the machine learning model −45/90/90]s 11.546
[61]. However, generating an adequat
ing dataset from an FEM can be computationally demanding and time‐con
2.2.3. Data Augmentation
sitating high computing resources. To address this challenge, a data augm
The performance of a CNN model relies on various factors, with the dataset being
egy is presented
a crucial to expand
one [39]. Boosting theofdataset
the size size
the dataset cansubstantially,
greatly improve thus reducing the
the prediction
resources
accuracy of theneeded
machine during
learningtraining.
model [61].The data augmentation
However, process
generating an adequately is divide
large
training dataset from an FEM can be computationally demanding and time-consuming,
egories:
necessitating high computing resources. To address this challenge, a data augmentation
Symmetry:
strategy is presented toExperimental
expand the datasetsamples with a square
size substantially, or circle
thus reducing impact wind
the computa-
tional resources needed during training. The data augmentation
metry; therefore, rotation of 90° was applied as the properties remain process is divided into the sa
two categories:
inal.Symmetry:
This procedure increases the size of the base dataset of 24 lines.
Experimental samples with a square or circle impact window show sym-
metry;Layer
therefore,translation:
rotation of 90The
◦ wasmaximum
applied as thenumber of remain
properties compositethe samelayers
as theis 28, b
original. This procedure increases the size of the base dataset of
investigated composites have only 12 or 16 layers. To ensure the model u 24 lines.
Layer translation: The maximum number of composite layers is 28, but many of the
usability of all layers, not just the initial ones, data entries of fewer than
investigated composites have only 12 or 16 layers. To ensure the model understands the
shifted
usability oftoallthe subsequent
layers, layer,ones,
not just the initial creating newof data,
data entries as depicted
fewer than 28 layers werein Figur
shifted to the
method, thesubsequent
final total layer, creatingof
number new data,reached
lines as depicted in Figure
a value of 4.2725.
Using this
method, the final total number of lines reached a value of 2725.

Figure Translation
Figure 4. 4. of the layers.
Translation of the layers.

2.3. Input and Output Definitions

In order to use the CNN models, data inputs and outputs were determ
Materials 2023, 16, 7213 9 of 18

Figure 4. Translation of the layers.

2.3.Input
2.3. Inputand
andOutput
Output Definitions
Definitions
InInorder
orderto
touse
use the
the CNN
CNN models,
models,data
datainputs
inputsand
andoutputs
outputswere
weredetermined andand
determined were
were
divided in the following way (Figure
divided in the following way (Figure 5): 5):

Figure5.5.Inputs
Figure Inputsand
andoutputs
outputs of
of CNN
CNN models.
models.

Alltextual
All textualinputs,
inputs, such
such as
as fiber
fiberweaving
weavingpattern,
pattern,material type,
material and
type, state
and (prepreg
state (prepreg or or
dry), underwent encoding using one‐hot encoding. A similar approach was
dry), underwent encoding using one-hot encoding. A similar approach was employed foremployed for
non‐continuousvariables,
non-continuous variables,such
suchasasfiber
fiberorientation.
orientation. Since
Since only
only certain
certain values
values (e.g.,
(e.g., 00 and
and 90)
90) were valid orientations, one‐hot encoding was applied to represent these
were valid orientations, one-hot encoding was applied to represent these admissible admissiblevalues.
For numerical features, normalization was performed by dividing each feature by the
maximum value within our dataset. This normalization procedure constrained the value
range of each feature between 0 and 1. These preprocessing steps collectively ensured
that each data point could be represented as a set of floating-point numbers, spanning
an interval from 0 to 1. Subsequently, the data were structured according to a 2D grid
to maximize the pattern detection of the CNN, where columns correspond to distinct
layers, and rows denote the composite layer properties (Table 4). Scalar values were
incorporated as rows within this matrix, repeating the same value 28 times to occupy the
entire row uniformly. The dataset was configured as a float32 tensor, possessing dimensions
of 2725 samples × 42 properties × 28 layers. The normalization process was extended to
the outputs, ensuring the predictions aligned with the same value range.

Table 4. Description of 2D grid for 8 plies [0/45/−45/90]s.

Composite Layer
1 2 3 4 5 6 7 8
0◦ orientation 1 0 0 0 0 0 0 1
+45◦ orientation 0 1 0 0 0 0 1 0
−45◦ orientation 0 0 1 0 0 1 0 0
90◦ orientation 0 0 0 1 1 0 0 0
Unidirectional or woven 1 1 1 0 0 1 1 1
Carbon fibers 1 1 1 1 1 1 1 1
Glass fibers 0 0 0 0 0 0 0 0
Kevlar fibers 0 0 0 0 0 0 0 0
Graphite fibers 0 0 0 0 0 0 0 0
Prepreg or ‘dry’ 1 1 1 1 1 1 1 1
Core material 0 0 0 0 0 0 0 0

In terms of dataset processing, comprehensive and reproducible shuffling of the

tensors was executed. This meticulous shuffling guaranteed homogeneity during the CNN
training process, while simultaneously preserving the association of inputs with their
corresponding expected outcomes. Subsequently, the tensors were sliced based on a split
 Materials 2023, 16, 7213 10 of 18

percentage. By default, 90% of the data were allocated for training, and the remaining 10%
were reserved for validation purposes.

2.4. CNN Model’s Construction

The networks were trained under a cloud computing environment using Python
v3.10.12 and Tensorflow v2.12.0.

2.4.1. Description of the First Supervised Network

A series of tests were run in order to select the best configurations for the net-
work among more than 1700 different architectures. Specifically, 1, 2, 4, and 6 convo-
lutional layers were evaluated, employing ascending or descending combinations of
4, 8, 16, 32, 64, 128, and 256 nodes. Independent kernel sizes ranging from 3 × 3 to 15 × 15
were also considered.
Optimization of the model parameters was achieved using the Adam optimizer,
utilizing Keras’ default implementation. The loss function chosen was the mean absolute
error, with its default Keras implementation chosen for its simplicity and for the low amount
Materials 2023, 16, x FOR PEER REVIEW 11 of 20
of outliers in the dataset. Among the hyperparameters influencing model training, the
batch size holds significance. Integral to the optimization algorithm, it dictates the quantity
of training samples processed before the internal model parameters receive updates. After
triala and
and error, the
preventive defaultbased
strategy batch onsizeearly
of 32, as offered
stopping was byimplemented.
the Keras library, yielded optimal
Specifically, training
outcomes. ifTothe
terminated forestall overfitting,
validation the maximum
loss remained stagnantepoch count
for 10 was capped
consecutive at 200,
epochs. Theand
ulti‐
a preventive strategy based on early stopping was implemented.
mate weights chosen were those associated with the most favorable overall validation Specifically, training
terminated if the validation loss remained stagnant for 10 consecutive epochs. The ultimate
loss.
weights
The chosen were necessitates
best model those associated with the mostlayers,
4 convolutional favorable overall validation
comprising 256, 32, loss.
8, and 16
The best model necessitates 4 convolutional layers, comprising 256, 32, 8, and 16 nodes,
nodes, followed by a fully connected layer housing 80 nodes (Figure 6). To mitigate over‐
followed by a fully connected layer housing 80 nodes (Figure 6). To mitigate overfitting
fitting risks, 2 max pooling operations were employed with a pooling rate of 0.5. Addi‐
risks, 2 max pooling operations were employed with a pooling rate of 0.5. Additionally,
tionally, a rectified linear unit (ReLU) activation function was applied to each convolu‐
a rectified linear unit (ReLU) activation function was applied to each convolutional and
tional and fully connected layer. The convolutional layer kernels adhere to a descending
fully connected layer. The convolutional layer kernels adhere to a descending logic, with
logic, with dimensions
dimensions of 15 × 15of for15the
× 15 for the
initial initial
layer, andlayer, and subsequent
subsequent sizes of 11sizes
× 11,of711× ×7,11,
and7×
7,5and 5 × 5. All convolutional layers incorporate L2 regularization, employing
× 5. All convolutional layers incorporate L2 regularization, employing a regularization a regular‐
ization
parameterparameter ofexcept
of 0.001, 0.001, for
except for the
the first first layer.
layer.

Flowchart
Figure6.6.Flow
Figure chartof
ofthe
the first
first CNN
CNN architecture.
architecture.

2.4.2. Description of the Second Supervised Network (TwIN_Z6_Net)

2.4.2. Description of the Second Supervised Network (TwIN_Z6_Net)
The input data were structured according to a 2D grid (Table 4). However, it can
The inputthat
be observed data were of
2 types structured according
data were to ascalar
used: single 2D grid (Table 4). However,
representing it can
data such as corebe
observed
presence, material composition, or impact energy values, and a matrix representing pres‐
that 2 types of data were used: single scalar representing data such as core the
ence, material composition, or impact energy values, and a matrix representing the com‐
posite stacking. However, in the context of stacking, deeming a scalar as a layer property
appeared illogical. Therefore, in order to differentiate the types of inputs, a different ap‐
proach was proposed. The inputs were separated into two parts: a matrix representing
Materials 2023, 16, 7213 11 of 18

composite stacking. However, in the context of stacking, deeming a scalar as a layer

property appeared illogical. Therefore, in order to differentiate the types of inputs, a
different approach was proposed. The inputs were separated into two parts: a matrix
representing only the composite stacking, and a vector containing scalar parameters. These
two parts were then processed through different paths in a Convolutional Neural Network
(CNN). Afterward, the outputs from these different paths were combined and run through
a simple Artificial Neural Network (Figure 7). This approach allowed us to generate the
original outputs while addressing the challenge of incorporating diverse types of data. The
initial branch (branch 1) is an extension of the previous model, where this pathway was
streamlined by eliminating the final convolutional layer. This adjustment aimed to simplify
the model and reduce training duration. Meanwhile, the second branch encompasses a basic
structure consisting of a two-layer Artificial Neural Network (ANN) featuring 64 nodes in
each layer, activated via the Parametric Rectified Linear Unit (PReLU) activation function.
The fusion of these two branches occurs through concatenation, forming the input for two
32-node fully connected layers, also utilizing PReLU activation. PReLU was chosen as an
Materials enhancement over
2023, 16, x FOR PEER standard ReLU activation. This choice was driven by the aim to retain12
REVIEW of 20
information embedded within the negative activation of neurons, all while preserving the
nonlinear characteristics.

Figure 7. Flow chart of the second

Figure CNN
7. Flow chart of architecture (TwIN_Z6_Net).
the second CNN architecture (TwIN_Z6_Net).

3. Results and Discussion

3. Results and Discussion
3.1. First Supervised3.1.
Network
First Supervised Network
Initially, two models basedtwo
Initially, on models
the proposed
based onCNN architecture
the proposed CNNwere trained
architecture to conduct
were trained to con‐
duct aof
a comparative analysis comparative
outcomesanalysis
with andof outcomes
withoutwith
the and without the normalization
normalization of output
of output data.
The training processdata. The training90%
encompassed process encompassed
of the 90% of
dataset, with the dataset,
identical with identical
samples employed samples
for em‐
ployed for training both models. The remaining subset was dedicated to validation pur‐
training both models. The remaining subset was dedicated to validation purposes. Loss
poses. Loss history was obtained, as shown in Figure 8, divided by the maximum loss
value to allow for a comparison.
Materials 2023, 16, 7213 12 of 18

Materials 2023, 16, x FOR PEER REVIEW 13 of 2

history was obtained, as shown in Figure 8, divided by the maximum loss value to allow
for a comparison.

Figure 8. Training Figure

and validation normalized
8. Training lossnormalized
and validation comparison.
loss comparison.

Upon completing the completing

Upon training process, the average
the training process,error across error
the average all outputs using
across all the using th
outputs
validation data was computed. The analysis revealed that the normalized model exhibited
validation data was computed. The analysis revealed that the normalized model exhibited
superior overallsuperior
performance
overall(Table 5). Notably,
performance (Tablethe marginalthe
5). Notably, reduction
marginalinreduction
precision in for
precision fo
the delaminated thesurface output, while present, was deemed insignificant in relation to
delaminated surface output, while present, was deemed insignificant in relation the to th
actual value of the result.
actual Indeed,
value of thethe concept
result. of scaling
Indeed, or normalizing
the concept of scaling oroutputs by dividing
normalizing outputs by divid
them by the maximum
ing themvalue
by theofmaximum
the dataset is often
value of the applied
dataset isbased
often on common
applied basedsense or
on common sens
specific needs [62].
or specific needs [62].

Table
Table 5. Comparison 5. Comparison
of the of the different
different CNNs’ average CNNs’
errors. average errors.

Average ErrorAverageAverage
Error ofError
Average Error of Error
Average Average Error of Average Error of Error
Average Average Error of
of Maximum Maximum Average Error of
Force Maximum
of Maximum ofDis‐ Permanent
Permanent Inden‐ Delamination
Delamination Index In‐ Delaminated
of Delaminated Sur
Force (kN) Displacement (mm) Indentation (mm) Surface (mm2 )
(kN) placement (mm) tation (mm) dex 2 face (mm )
Standard output 0.30 0.35 0.48 0.30 62.17
Normalized outputs Standard output
0.23 0.30 0.25 0.35 0.28 0.48 0.05 0.30 76.56 62.17
Improvement betweenNormalized outputs 0.23 0.25 0.28 0.05 76.56
standard and 0.07 0.1 0.2 0.25 −14.39
normalized output Improvement be‐
TwIN_Z6_Net tween standard 0.16and 0.07 0.15 0.1 0.13 0.2 0.02 0.25 56.36 −14.39
normalized output
TwIN_Z6_Net 0.16 Network
3.2. Second Supervised 0.15 0.13 0.02 56.36
Due to the previous result, output normalization was applied to the TwIN_Z6_Net
3.2. Second Supervised Network
model. Figure 9 illustrates the progression of the loss value throughout the training of the
proposed architecture.Due toutilization
The the previousof result,
the Adamoutput normalization
optimizer was rapid
facilitates applied to the TwIN_Z6_Ne
convergence
model. Figure 9 illustrates the progression of the loss value throughout
of the model towards the minimum loss value. Notably, the early-stopping mechanism the training of th
proposed architecture. The utilization of the Adam optimizer facilitates rapid
was invoked around epoch 180 to avoid overfitting.
 Materials 2023, 16, x FOR PEER REVIEW 14 of 20

Materials 2023, 16, 7213 13 of 18

convergence of the model towards the minimum loss value. Notably, the early‐stopping
mechanism was invoked around epoch 180 to avoid overfitting.

Figure Training
Figure 9.9.Training lossloss
andand validation
validation loss
loss per per epoch.
epoch.

Forthe
For the evaluation
evaluation phase,phase, the error
the mean meanoferror of each
each output output
was was(Table
calculated calculated
5). A (Table 5). A
clear improvement for all outputs can be observed. The highlight of our model is
clear improvement for all outputs can be observed. The highlight of our model is reachingreaching
0.15 mm
0.15 mmprecision for for
precision all the
alllength measurements.
the length The errorThe
measurements. for each individual
error for eachsample
individual sample
within the validation dataset was assessed, and we subsequently calculated both the mean
within the validation dataset was assessed, and we subsequently calculated both the mean
and standard deviation for each output (Figure 10):
and Output
standard deviation
maximum for each 0.0116
force—mean: outputkN,(Figure
std dev:10):
0.2451
Output maximum force—mean:
Output maximum displacement—mean: −0.0132 0.0116 kN,
mm, stdstddev:
dev: 0.2451
0.3706
Output
Output maximum
permanent displacement—mean:
indentation—mean: −0.0132
−0.0170 mm, mm,
std dev: std dev: 0.3706
0.3458
Materials 2023, 16, x FOR PEER REVIEWOutput delamination surface—mean: −14.4754 mm 2, std dev: 164.2299
Output permanent indentation—mean: −0.0170 mm, std dev: 0.3458 15 of 20

Output delamination surface—mean: −14.4754 mm2 , std dev: 164.2299

Figure
Figure 10. 10. Distributionofofthe
Distribution theerror
error across
across all
allthe
themodel
modeloutputs.
outputs.

The treatment for the delamination index varied due to its binary nature, denoting
the presence or absence of delamination on a sample. A slight post‐processing step was
introduced to enhance the interpretability of the floating‐point output, enabling an easier
assessment of its proximity to 0 or 1. This involved computing the absolute error value
and applying a threshold filter, resulting in a Boolean outcome that reflects prediction
Materials 2023, 16, 7213 14 of 18

The treatment for the delamination index varied due to its binary nature, denoting
the presence or absence of delamination on a sample. A slight post-processing step was
introduced to enhance the interpretability of the floating-point output, enabling an easier
assessment of its proximity to 0 or 1. This involved computing the absolute error value and
applying a threshold filter, resulting in a Boolean outcome that reflects prediction accuracy. In
this work, it was considered that there is no delamination for an index value lower than 0.2. It
was observed that 11 samples without delamination were predicted to be delaminated, and
Materials 2023, 16, x FOR PEER REVIEW
conversely, 9 samples with actual delamination were predicted to have none (Figure 16 of 2011). In

total, the global error reached 20 out of 273 samples, yielding an error rate of 7.3%.

Figure 11. Distribution of errors for delamination index.

Figure 11. Distribution of errors for delamination index.
3.3. Perspective
3.3. Perspective
The most intriguing aspect of designing a composite aeronautical structure for impact
The most intriguing aspect of designing a composite aeronautical structure for im‐
damage tolerance is not the impact itself, but rather, the loss of residual strength resulting
pact damage tolerance is not the impact itself, but rather, the loss of residual strength re‐
from the impact, along with the detectability of the impact. Residual compressive strength
sulting from the impact, along with the detectability of the impact. Residual compressive
after impact is a critical design factor in assessing the damage tolerance of fiber-reinforced
strength after impact is a critical design factor in assessing the damage tolerance of fiber‐
polymers (FRP) for structural applications in aircraft, as it is classically the most affected
reinforced polymers (FRP) for structural applications in aircraft, as it is classically the most
mechanical characteristic by impact damage (due to the buckling of the delaminated plies)
affected mechanical characteristic by impact damage (due to the buckling of the delami‐
(Figure 12a).(Figure
nated plies) The Compression After Impact
12a). The Compression (CAI)
After test is(CAI)
Impact a standardized method for char-
test is a standardized
method for characterizing the residual compressive strength for laminates of FRP,the
acterizing the residual compressive strength for laminates of FRP, employing em‐ASTM
ploying the ASTM D7136 [13] and ASTM D7137 [63] standards (Figure 12b). The full pro‐ two
D7136 [13] and ASTM D7137 [63] standards (Figure 12b). The full procedure follows
stages,follows
cedure where twosamples arewhere
stages, first subjected
samples are to first
low-velocity
subjectedimpact via drop-weight
to low‐velocity impact via impact
testing and impact
drop‐weight are, then, subjected
testing and are, tothen,
in-plane compressive
subjected to in‐plane loading using loading
compressive CAI apparatus.
using To
design
CAI and optimize
apparatus. To design a composite
and optimize structure for impact
a composite damage
structure tolerance,
for impact damage it is essential
toler‐
to simultaneously
ance, it is essential tosimulate the damage
simultaneously simulatedevelopment
the damageduring impact,
development the permanent
during impact, in-
dentation
the permanent leftindentation
by the impact, left byand
theultimately,
impact, andthe residualthe
ultimately, strength.
residualWhile
strength.theWhile
numerical
models
the presented
numerical modelsinpresented
the literature
in the [4,16] are[4,16]
literature highly arevaluable, their effective
highly valuable, utilization
their effective
demands demands
utilization extensiveextensive
knowledge, additional
knowledge, experiments
additional for a more
experiments for ain-depth examination
more in‐depth ex‐ of
amination
the compositeof thestructures
compositeunder structures
study,under
and astudy, and a considerable
considerable number of IT number
resourcesof ITtore‐
run the
sources
models.toApplying
run the models.
the sameApplying the sameto
methodology methodology to an experimental/FEA
an experimental/FEA dataset, proposingda‐ a
taset, proposing a CNN model for Compression After Impact
CNN model for Compression After Impact (CAI) represents an intriguing approach for(CAI) represents an intri‐
guing
sizingapproach
composites for sizing composites
in terms of impact in terms
damageof impact damage
tolerance. Thistolerance.
methodThis method the
circumvents
circumvents the need for FEA iteration, thereby reducing design calculation
need for FEA iteration, thereby reducing design calculation time. Designers would only be time. Design‐
ers would to
required only be required
input parameters to input
suchparameters
as composite such as composite
stacking stacking
definition and definition
impact energy,and and
impact energy, and the CNN model would instantly
the CNN model would instantly provide the desired response. provide the desired response.
 Materials2023,
Materials 2023,16,
16,7213
x FOR PEER REVIEW 17 of 2015 of 18

(a) (b)
Figure
Figure12.
12.(a)
(a)Compression
CompressionAfter
AfterImpact test
Impact set‐up
test and
set-up (b)(b)
and schematic damage
schematic thatthat
damage developed dur‐during
developed
ing Compression After Impact test.
Compression After Impact test.

4.
4. Conclusions
Conclusions
Low‐velocity/‐energy
Low-velocity/-energy impact
impact onona composite
a composite structure
structure hashas
thethe
characteristic
characteristic of caus‐
of causing
ing extensive internal damage within the layers of the material, despite
extensive internal damage within the layers of the material, despite only causing a slightly only causing a
slightly perceptible indentation on the surface. This type of damage
perceptible indentation on the surface. This type of damage results in a decrease in residualresults in a decrease
in residualpost-impact,
strength strength post‐impact,
particularly particularly
in terms in ofterms of its compressive
its compressive strength. strength.
WithinWithin
the field of
the field of aeronautics, this reduction in strength compels designers
aeronautics, this reduction in strength compels designers to consider damage tolerance to consider damage and
tolerance
restrict the and restrict the
utilization of utilization
the material’sof the material’s
full full capabilities.
capabilities. As a result,As a result, accu‐
accurately predicting
rately predicting the effects of impact damage has become
the effects of impact damage has become crucial for enhancing the remaining crucial for enhancing the strength
re‐
maining strength of composites. In this work, CNN models are proposed to predict impact
of composites. In this work, CNN models are proposed to predict impact damage in
damage in FRP composites. On one hand, data are collected from the literature, while on
FRP composites. On one hand, data are collected from the literature, while on the other
the other hand, the finite element method (FEM) is employed to simulate impact, with an
hand, the finite element method (FEM) is employed to simulate impact, with an initial
initial literature‐to‐FEM data ratio of 32%. A large amount of data information is gathered,
literature-to-FEM data ratio of 32%. A large amount of data information is gathered, and
and a data augmentation method is proposed to increase the data number from 149 to
a data augmentation method is proposed to increase the data number from 149 to 2725.
2725. The data are divided into two parts: input and output data. The first one consists of
The data arestacking
a composite divideddefinition
into twoand parts: input
of the andtest
impact output data. The
parameters. Thefirst oneone
second consists
con‐ of a
composite stacking definition and of the impact test parameters.
cerns the damage to the FRP composite, including maximum force, maximum displace‐ The second one concerns
the damage to the FRP composite, including maximum force,
ment, indentation, delamination area, and delamination index. Then, two different CNN maximum displacement,
indentation,
models delamination
are investigated andarea, and delamination
optimized. The first one is index.
basedThen, two different
on a traditional CNN models
architecture
are investigated
and and optimized.
the second consists The first
of an aggregation ofone
twoisCNNs.
basedThe on aproposed
traditional architecture
method enables us and the
second
to predictconsists of an aggregation
the impact damage to an of two
FRP CNNs.
compositeThe forproposed method enables
given stacking us to predict
configurations.
the impacttodamage
However, obtain the to best
an FRP composite
results, for givenofstacking
the aggregation two CNNs configurations.
should be consideredHowever, to
obtain
as the best
it shows betterresults, the aggregation
performance of two CNNs
than the traditional should be considered
architecture—the as it shows
first to process a ma‐better
performance
trix representing than the traditional
composite stackingarchitecture—the
and the second to first to process
process a vector acontaining
matrix representing
scalar
composite stacking
parameters and the second
(sample surface, to process
impact energy, a vector
etc.). Moreover, containing scalar parameters
normalization of the output (sample
data improves
surface, impact theenergy,
CNN model’s performance.
etc.). Moreover, The best model
normalization of reaches
the output0.15 data
mm precision
improves the
for
CNN all the lengthperformance.
model’s measurementsThe andbest
an average delaminated
model reaches 0.15 mmsurface error offor
precision 56.36
all mm ,
the 2length
and a 7% error rate is obtained on the delamination index.
measurements and an average delaminated surface error of 56.36 mm , and a 7% error rate 2
This work
is obtained on thecould facilitate theindex.
delamination process in the area of damage tolerance design by
providing rapid damage prediction
This work could facilitate the process for CFRP composite
in the area solutions. Furthermore,
of damage toleranceit has the by
design
potential to decrease the time and expenses associated with investigating
providing rapid damage prediction for CFRP composite solutions. Furthermore, it has the and formulating
novel
potentialFRPtocomposites.
decrease the In time
orderandto improve
expensesthe model, work
associated with is in progress and
investigating in order to
formulating
novel FRP composites. In order to improve the model, work is in progress in order to
increase the dataset size. Moreover, deeper analysis of the model regarding the influence of
each parameter is in progress.
Materials 2023, 16, 7213 16 of 18

Author Contributions: A.S.R.: data curation, formal analysis, validation, writing—original draft.
A.R.: coding, data curation, formal analysis, validation, writing—original draft. C.B.: writing—review.
L.M.: conceptualization, validation, writing—review and editing. All authors have read and agreed
to the published version of the manuscript.
Funding: The authors received no financial support for the research, authorship, and/or publication
of this article.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The raw/processed data required to reproduce these findings are
available upon request.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Flore, D.; Wegener, K.; Mayer, H.; Karr, U.; Oetting, C.C. Investigation of the high and very high cycle fatigue behaviour of continuous
fibre reinforced plastics by conventional and ultrasonic fatigue testing. Compos. Sci. Technol. 2017, 141, 130–136. [CrossRef]
2. Bouvet, C.; Rivallant, S. 2—Damage tolerance of composite structures under low-velocity impact. In Dynamic Deformation, Damage and
Fracture in Composite Materials and Structures; Silberschmidt, V.V., Ed.; Woodhead Publishing: Sawston, UK, 2016; pp. 7–33. [CrossRef]
3. Rozylo, P.; Debski, H.; Kubiak, T. A model of low-velocity impact damage of composite plates subjected to Compression-After-
Impact (CAI) testing. Compos. Struct. 2017, 181, 158–170. [CrossRef]
4. Dubary, N.; Bouvet, C.; Rivallant, S.; Ratsifandrihana, L. Damage tolerance of an impacted composite laminate. Compos. Struct.
2018, 206, 261–271. [CrossRef]
5. Morteau, E.; Fualdesairbus, C. Damage Tolerance Philosophy. In Proceedings of the FAA Workshop for Composite Damage
Tolerance and Maintenances, Chicago, IL, USA, 19–21 July 2006.
6. Abrate, S. Impact on Composite Structures; Cambridge University Press: New York, NY, USA, 2005.
7. Hongkarnjanakul, N.; Bouvet, C.; Rivallant, S. Validation of low velocity impact modelling on different stacking sequences of
CFRP laminates and influence of fibre failure. Compos. Struct. 2013, 106, 549–559. [CrossRef]
8. Cantwell, W.J.; Morton, J. The impact resistance of composite materials—A review. Composites 1991, 22, 347–362. [CrossRef]
9. Tropis, A.; Thomas, M.; Bounie, J.L.; Lafon, P. Certification of the Composite Outer Wing of the ATR72. Proc. Inst. Mech. Eng. Part
G J. Aerosp. Eng. 1995, 209, 327–339. [CrossRef]
10. Rouchon, J. Certification of Large Aircraft Composite Structures, Recent Progress and New Trends in Compliance Philosophy. In
ICAS Congress Proceedings; American Institute of Aeronautics and Astronautics: Reston, VA, USA, 1990.
11. Caprino, G.; Lopresto, V. The significance of indentation in the inspection of CFRP panels damaged by low-velocity impact.
Compos. Sci. Technol. Compos. Sci. Technol. 2000, 60, 1003–1012. [CrossRef]
12. Cheng, Z.; Zhang, L.; Jiang, C.; Dai, Y.; Meng, C.; Luo, L.; Liu, X. Aramid fiber with excellent interfacial properties suitable for
resin composite in a wide polarity range. Chem. Eng. J. 2018, 347, 483–492. [CrossRef]
13. ASTMD7136/D7136M; Standard Test Method for Measuring the Damage Resistance of a Fiber-Reinforced Polymer Matrix
Composite to a Drop-Weight Impact Event. ASTM International: West Conshohocken, PA, USA, 2015.
14. Huang, K.; de Boer, A.; Akkerman, R. Analytical Modeling of Impact Resistance and Damage Tolerance of Laminated Composite
Plates. AIAA J. 2008, 46, 2760–2772. [CrossRef]
15. Olsson, R. Analytical prediction of large mass impact damage in composite laminates. Compos. Part Appl. Sci. Manuf. 2001, 32,
1207–1215. [CrossRef]
16. Kutlu, Z.; Chang, F. Modeling Compression Failure of laminated Composites Containing Multiple Through-the-Width Delamina-
tions. J. Compos. Mater. 1992, 26, 350–387. [CrossRef]
17. Wisnom, M.R. Modelling discrete failures in composites with interface elements. Compos. Part Appl. Sci. Manuf. 2010, 41, 795–805.
[CrossRef]
18. Lopes, C.S.; Sádaba, S.; González, C.; Llorca, J.; Camanho, P. Physically-Sound Simulation of Low-Velocity Impact on Fibre
Reinforced Laminates. Int. J. Impact Eng. 2015, 92, 3–17. [CrossRef]
19. Hornik, K.; Stinchcombe, M.; White, H. Multilayer feedforward networks are universal approximators. Neural Netw. 1989, 2,
359–366. [CrossRef]
20. Manry, M.T. Neural networks: Algorithms, applications, and programming techniques: By James A. Freeman and David M.
Skapura, Addison-Wesley Publishing, Reading, MA, ISBN 0-201-51376-5. Neural Netw. 1994, 7, 209–212. [CrossRef]
21. Schmidhuber, J. Deep learning in neural networks: An overview. Neural Netw. 2015, 61, 85–117. [CrossRef]
22. Azimi, M.; Eslamlou, A.D.; Pekcan, G. Data-Driven Structural Health Monitoring and Damage Detection through Deep Learning:
State-of-the-Art Review. Sensors 2020, 20, 2778. [CrossRef]
23. Mishkin, D.; Sergievskiy, N.; Matas, J. Systematic evaluation of convolution neural network advances on the Imagenet.
Comput. Vis. Image Underst. 2017, 161, 11–19. [CrossRef]
Materials 2023, 16, 7213 17 of 18

24. Nelon, C.; Myers, O.; Hall, A. The intersection of damage evaluation of fiber-reinforced composite materials with machine
learning: A review. J. Compos. Mater. 2022, 56, 002199832110370. [CrossRef]
25. Gu, J.; Wang, Z.; Kuen, J.; Ma, L.; Shahroudy, A.; Shuai, B.; Liu, T.; Wang, X.; Wang, G.; Cai, J.; et al. Recent advances in
convolutional neural networks. Pattern Recognit. 2018, 77, 354–377. [CrossRef]
26. Sony, S.; Dunphy, K.; Sadhu, A.; Capretz, M. A systematic review of convolutional neural network-based structural condition
assessment techniques. Eng. Struct. 2021, 226, 111347. [CrossRef]
27. Deng, K.; Liu, H.; Yang, L.; Addepalli, S.; Zhao, Y. Classification of barely visible impact damage in composite laminates using
deep learning and pulsed thermographic inspection. Neural Comput. Applic. 2023, 35, 11207–11221. [CrossRef]
28. Hasebe, S.; Higuchi, R.; Yokozeki, T.; Takeda, S. Multi-task learning application for predicting impact damage-related information
using surface profiles of CFRP laminates. Compos. Sci. Technol. 2023, 231, 109820. [CrossRef]
29. Alhammad, M.S.; Avdelidis, N.; Ibarra Castanedo, C.; Torbali, M.E. Automated Impact Damage Detection Technique for
Composites Based on Thermographic Image Processing and Machine Learning Classification. Sensors 2022, 22, 9031. [CrossRef]
30. Hasebe, S.; Higuchi, R.; Yokozeki, T.; Takeda, S. Internal low-velocity impact damage prediction in CFRP laminates using surface
profiles and machine learning. Compos. Part B Eng. 2022, 237, 109844. [CrossRef]
31. Zargar, S.; Yuan, F.-G. Impact diagnosis in stiffened structural panels using a deep learning approach. Struct. Health Monit. 2020,
20, 147592172092504. [CrossRef]
32. Wei, Z.; Fernandes, H.; Herrmann, H. A Deep Learning Method for the Impact Damage Segmentation of Curve-Shaped CFRP
Specimens Inspected by Infrared Thermography. Sensors 2021, 21, 395. [CrossRef]
33. Jung, K.-C.; Chang, S.-H. Advanced Deep Learning Model-Based Impact Characterization Method for Composite Laminates.
Compos. Sci. Technol. 2021, 220, 108713. [CrossRef]
34. Ahmed, O.; Wang, X.; Tran, M.-V.; Ismadi, M.-Z. Advancements in fiber-reinforced polymer composite materials damage detection
methods: Towards achieving energy-efficient SHM systems. Compos. Part B Eng. 2021, 223, 109136. [CrossRef]
35. Tabatabaeian, A.; Jerkovic, B.; Harrison, P.; Marchiori, E.; Fotouhi, M. Barely visible impact damage detection in composite
structures using deep learning networks with varying complexities. Compos. Part B Eng. 2023, 264, 110907. [CrossRef]
36. Pilania, G.; Wang, C.; Jiang, X.; Rajasekaran, S.; Ramprasad, R. Accelerating materials property predictions using machine
learning. Sci. Rep. 2013, 3, 2810. [CrossRef] [PubMed]
37. Rong, Q.; Wei, H.; Bao, H. Deep learning methods based on cross-section images for predicting effective thermal conductivity of
composites. arXiv 2019, arXiv:1904.06104.
38. Ashank; Chakravarty, S.; Garg, P.; Kumar, A.; Agrawal, M.; Agnihotri, P. Deep neural networks based predictive-generative
framework for designing composite materials. Model. Simul. Mater. Sci. Eng. 2022, 30, 075003. [CrossRef]
39. Yang, C.; Kim, Y.; Ryu, S.; Gu, G.X. Using convolutional neural networks to predict composite properties beyond the elastic limit.
MRS Commun. 2019, 9, 609–617. [CrossRef]
40. Yang, C.; Kim, Y.; Ryu, S.; Gu, G.X. Prediction of composite microstructure stress-strain curves using convolutional neural
networks. Mater. Des. 2020, 189, 108509. [CrossRef]
41. Liu, X.; Tian, S.; Tao, F.; Yu, W. A review of artificial neural networks in the constitutive modeling of composite materials. Compos.
Part B Eng. 2021, 224, 109152. [CrossRef]
42. Kollmann, H.T.; Abueidda, D.W.; Koric, S.; Guleryuz, E.; Sobh, N.A. Deep learning for topology optimization of 2D metamaterials.
Mater. Des. 2020, 196, 109098. [CrossRef]
43. Xue, T.; Beatson, A.; Chiaramonte, M.; Roeder, G.; Ash, J.T.; Menguc, Y.; Adriaenssens, S.; Adams, R.P.; Mao, S. A data-driven
computational scheme for the nonlinear mechanical properties of cellular mechanical metamaterials under large deformation.
Soft Matter 2020, 16, 7524–7534. [CrossRef]
44. Jiao, P.; Alavi, A.H. Artificial intelligence-enabled smart mechanical metamaterials: Advent and future trends. Int. Mater. Rev.
2021, 66, 365–393. [CrossRef]
45. Liu, X.; Gasco, F.; Goodsell, J.; Yu, W. Initial failure strength prediction of woven composites using a new yarn failure criterion
constructed by deep learning. Compos. Struct. 2019, 230, 111505. [CrossRef]
46. Abueidda, D.; Koric, S.; Sobh, N. Topology optimization of 2D structures with nonlinearities using deep learning. Comput. Struct.
2020, 237, 106283. [CrossRef]
47. Qiu, A.; Fu, K.; Lin, W.; Zhao, C.; Tang, Y. Modelling low-speed drop-weight impact on composite laminates. Mater. Des. 2014, 60,
520–531. [CrossRef]
48. Vieille, B.; Casado, V.M.; Bouvet, C. Influence of matrix toughness and ductility on the compression-after-impact behavior of
woven-ply thermoplastic- and thermosetting-composites: A comparative study. Compos. Struct. 2014, 110, 207–218. [CrossRef]
49. Bandaru, A.K.; Chavan, V.V.; Ahmad, S.; Alagirusamy, R.; Bhatnagar, N. Low velocity impact response of 2D and 3D
Kevlar/polypropylene composites. Int. J. Impact Eng. 2016, 93, 136–143. [CrossRef]
50. Usta, F.; Türkmen, H.S.; Scarpa, F. Low-velocity impact resistance of composite sandwich panels with various types of auxetic
and non-auxetic core structures. Thin-Walled Struct. 2021, 163, 107738. [CrossRef]
51. Bhuiyan, M.d.A.; Hosur, M.V.; Jeelani, S. Low-velocity impact response of sandwich composites with nanophased foam core and
biaxial (±45◦ ) braided face sheets. Compos. Part B Eng. 2009, 40, 561–571. [CrossRef]
52. Sun, X.C.; Hallett, S.R. Barely visible impact damage in scaled composite laminates: Experiments and numerical simulations.
Int. J. Impact Eng. 2017, 109, 178–195. [CrossRef]
Materials 2023, 16, 7213 18 of 18

53. Petit, S. Contribution à L’étude de L’influence D’une Protection Thermique sur la Tolérance aux Dommages des Structures
Composites des Lanceurs. Ph.D. Thesis, University of ENSAE, Palaiseau, France, 2005.
54. Hongkarnjanakul, N. Modélisation Numérique Pour la Tolérance aux Dommages D’impact sur Stratifié Composite: De L’impact
à la Résistance Résiduelle en Compression. Doctoral Dissertation, ISAE, Toulouse, France, 2013.
55. ISAE-SUPAERO. Fourcroyd Influence de L’armure et des Conditions de Fabrication sur le Comportement a L’impact d’un Stratifie Composite
Tisse Carbone/pps; ISAE-SUPAERO: Toulouse, France, 2013.
56. Broll, B. Experimental Studies on the Damage Mechanisms of CFRP-Structures. Ph.D. Thesis, ISAE-SUPAERO/Lehrstuhl fur
Leichtbau, Toulouse, France, 2008.
57. Warren, T.L.; Tabbara, M.R. Simulations of the penetration of 6061-T6511 aluminum targets by spherical-nosed VAR 4340 steel
projectiles. Int. J. Solids Struct. 2000, 37, 4419–4435. [CrossRef]
58. Abdulhamid, H.; Bouvet, C.; Michel, L.; Aboissiere, J.; Minot, C. Numerical simulation of impact and compression after impact of
asymmetrically tapered laminated CFRP. Int. J. Impact Eng. 2016, 95, 154–164. [CrossRef]
59. TORAYCA T700S DataSheet, n.d. Available online: https://www.rockwestcomposites.com/downloads/T700SDataSheet.pdf
(accessed on 15 May 2015).
60. Bieling, U. AIRBUS Reference Structure Design Principles the basis for A350XWB design. In 8th CFK-Valley Stade Convention;
YUMPU: Ottenbecker, Germany, 2014.
61. Abueidda, D.W.; Almasri, M.; Ammourah, R.; Ravaioli, U.; Jasiuk, I.M.; Sobh, N.A. Prediction and optimization of mechanical
properties of composites using convolutional neural networks. Compos. Struct. 2019, 227, 111264. [CrossRef]
62. Bishop, C.M.; Nasrabadi, N.M. Pattern Recognition and Machine Learning; Springer: New York, NY, USA, 2006.
63. ASTM D7137/D7137M-17; Standard Test Method for Compressive Residual Strength Properties of Damaged Polymer Matrix
Composite Plates. ASTM International: West Conshohocken, PA, USA, 2011.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.