You are on page 1of 10

RESEARCH ARTICLE

www.aem-journal.com

Classifying Tensile Loading History of Continuous Carbon


Fiber Composites Using X-Ray Scattering and Machine
Learning
Alexander Sexton,* Marc Kanters, Hanna Demchenko, Barbara Pacáková,
Jon Otto Fossum, Luigi Balzano, and Matti Knaapila*

1. Introduction
The tensile loading history of continuous carbon fiber composites is classified
using machine learning (ML) and crystallographic data from the polymer matrix. Carbon fiber-reinforced polymers (CFRPs),
Composites with polyamide-4,10 matrix and unidirectional 10° and 45°, and including continuous carbon fiber compo-
sites with macroscopically aligned fibers
0°/90° cross-ply layups are subjected to single-cycle uniaxial tensile loads cor-
within thermoplastic polymer matrices, are
responding to 25–90% of their nominal maximum strain, and mapped by X-ray used in a myriad of aerospace, automotive
diffraction with approximately 1000 data points from each layup. The unit cell and maritime applications, protective
alterations are used as a feature set for optimizing three ML algorithms; linear armors, wind turbines, pressure vessels,
discriminant analysis, support vector machines (SVM), and gradient-boosted sports items, and generally whenever high
decision trees (GBDT), with the objective of predicting five discrete loading durability and high strength-to-weight ratio
are required.[1–5] The strength of these mate-
magnitudes of the respective layups. It is demonstrated that SVMs and GBDTs
rials stems from the transfer of load from the
can be trained to achieve a classification accuracy of >90% on unseen test data, polymer matrix to the reinforcing fibers, and
both in cases where the feature set consists of data points from individual layups thus the mechanical properties of the com-
only, but also when data from the three layups are aggregated. The performance posite depend on the fiber and matrix
of the models is also shown to be similar to a binary problem, in which the strength, their directionality, as well as on
composites are categorized according to a threshold load. the interfacial bonding between the two.
The load-bearing capabilities are anisotropic
with respect to the orientation of the rein-
forcing fibers. Lamina with fibers in several
orientations are therefore commonly stacked
in the interest of producing a material capable of bearing loads
along more than one axis. Composites may be damaged during
the materials processing or during the operational lifetime, e.g.,
A. Sexton, H. Demchenko, B. Pacáková, J. O. Fossum, M. Knaapila upon overloading, fatigue, or accidental impact. Different defect
Department of Physics
Norwegian University of Science and Technology
types reach from nanoscale to macroscopic scale and include
Høgskoleringen 5, 7491 Trondheim, Norway matrix cracking, fiber failure, delamination of the plies and
E-mail: alexander.h.sexton@ntnu.no; matti.knaapila@ntnu.no matrix–fiber debonding,[6] intermediately affecting the stiffness
A. Sexton, L. Balzano,[+] M. Knaapila and strength,[7–9] and ultimately leading to macroscopic failure.
DPI Nondestructive health monitoring techniques are ubiquitous
P.O. Box 902, 5600 AX Eindhoven, The Netherlands for processing control and for determining whether or not the
M. Kanters,[++] L. Balzano, structural integrity of the material is intact, as well as for lifetime
DSM Materials Science Center
prediction.[10] Due to the diversity of failure modes in CFRPs,
Urmonderbaan 22, 6167RD Geleen, The Netherlands
several techniques are often required for ascertaining damage.[11]
The ORCID identification number(s) for the author(s) of this article Existing methods include ultrasonic scans,[12] acoustic emmi-
can be found under https://doi.org/10.1002/adem.202301415.
sions,[13] vibrothermography,[14] surface reflectometry,[15] infra-
[+]
Present address: Envalior, Urmonderbaan 22, 6167 RD Geleen, The red thermography,[16] terahertz imaging,[17] microwaves,[18]
Netherlands X-ray tomography,[19] and X-ray backscattering techniques,[20]
[++]
Present address: Avient Protective Materials, Urmonderbaan 22, to mention a few. Regardless of the origin or defect type,
6167 RD Geleen, The Netherlands macroscopic and eventually detrimental defects originate from
© 2023 The Authors. Advanced Engineering Materials published by Wiley- the nanoscale level, whereby molecular chains, crystals, and
VCH GmbH. This is an open access article under the terms of the Creative amorphous regions manifest local frustrations and dislocations,
Commons Attribution License, which permits use, distribution and and reach their maximum load-bearing capability. Not only
reproduction in any medium, provided the original work is properly cited. macroscopic, but also nanoscale inspections are therefore
DOI: 10.1002/adem.202301415 required for determining and predicting early-stage defects.

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (1 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

Wide-angle X-ray scattering (WAXS) has been used exten- altogether. Additionally, we construct a scenario in which the
sively to probe crystalline order in carbon fiber composites specimens are successfully classified based on a threshold of
and matrix polymers, including crystalline phases of nylon loading history.
sheets,[21] the interphase between carbon and polyamide,[22]
phase transitions of polyimide-confined nylon fibers,[23] and
the local tensile deformations in polyamides,[24] for example. 2. Experimental Section
The studies conclude that early-state defects lead only to minor 2.1. Materials
effects on crystalline information, and any assessment requires
considerable statistical evidence. Furthermore, there are few rig- Figure 1a,b illustrates the employed samples and sample layups
orous physical models that connect microscale crystallographic (DSM Engineering Materials). All samples were continuous car-
information to the macroscale defects, and eventually to the bon fiber-reinforced thermoplastic composites with a
composites’ lifetime, and thus a more phenomenological or polyamide-4,10, and 60% fiber content by weight. Panels were
databased approach must be taken. hot pressed from eight layers of tapes in two configurations,
Machine learning (ML) methods have an advantage in that they UD and 0°/90°, providing panels with a thickness of approx.
can find correlations in large amounts of data, in principle con- 2 mm. From these panels, samples of 200  10 mm (see
necting any input–output pair, given that the information is Figure 1b) were cut using waterjet cutting to obtain UD samples
there.[25] In the context of health monitoring of composites and
with fibers oriented under 10°, 45°, and orthotropic (0°/90°) ori-
polymers, Pawar et al.[26] proposed a system for damage detection entation with respect to the loading direction. To allow proper
in helicopter rotor blades. By using simulated loads of the helicop-
gripping and avoid failure in the grips, tabs were cut from the
ter vibratory hub as input features for a support vector machine
same material and subsequently glued on the ends prior to
(SVM) classifier,[27] the rotor blades could be categorized into three
testing.
classes relating to their structural integrity. Qiao et al.[28] used
The selected three fiber orientations were chosen to introduce
acoustic emission measurements and ML to monitor damage
a variety of damage mechanisms. As an example where multiple
modes in carbon/epoxy composites. Specifically, they used a
layers with varying orientations interact under mechanical load-
combination of SVMs and k-means clustering method,[29]
ing, the 0°/90° x-ply was chosen. Since the deformation behavior
together with real-time readings of specimens under the
of a 0° UD will be dictated by the fibers, and that of a 90° UD by
three-point bending test, and were able to distinguish between
very local deformation in between fibers and at fiber interfaces,
various damage events in the composites. Kurita et al.[30] used
10° and 45° UD were chosen instead since they show a more pro-
electron microscopy together with k-means method to predict
nounced contribution of the matrix to the total deformation,
Young’s modulus in composites with varying fiber porosity
expressed by the distinct nonlinearity in the stress/strain curve
and orientation.
(see, e.g., Figure 2a, vide infra) before triggering failure.
When X-rays are concerned, Obdrup et al.[31] used WAXS
and linear discriminant analysis (LDA)[32] to classify two struc-
tural conditions of polyethylene ropes. In their study, they dem- 2.2. Mechanical Loading
onstrated that a 100% classification accuracy could be obtained
both with preprocessed 1D WAXS data and derived physical Figure 1b shows the loading direction. Uniaxial tensile tests were
parameters as input features for the classifier. Elsewhere, performed using a Zwick Z1484 equipped with a 200 kN load cell
Liu et al.[33] proposed a method for characterizing the extent using a gripping length of 100 mm. The test protocol was accord-
of delamination in carbon fiber composites using lamb wave ing to ISO527, providing a nominal strain rate of 1% min1 for
propagation and ML. The authors used X-ray absorption images strains up to 0.25% to determine the modulus and a strain rate of
to quantify the delamination area, which they correlated with the 5% min1 for the remainder of the test. Strains were measured
acoustic emissions through a comparative ensemble of ML using an optical extensometer (VideoXtens) providing a resolu-
methods. tion of 0.5, which with a gauge length of approx. 70 μm results in
In this study, we demonstrate how WAXS and ML can be com- a resolution in strain below 0.0001%. To obtain an indication of
bined to monitor the health of continuous carbon fiber compo- the strength of the samples, a single sample was loaded up to
sites. Particularly, the focus is on minute changes in the crystal failure. Next, to vary the mechanical histories to the test samples,
structure of the bulk polymer matrix, at the onset of damage to pristine samples were loaded for one cycle to a maximum of 25,
the composites. To generate samples with variations in damage 50, 75, and 90% of their strength. Subsequently, the samples
magnitude, material samples were subjected to single-cycle uni- were unloaded and reloaded to a maximum of 10% to evaluate
axial tensile loads, at magnitudes progressively outside of the lin- their stiffness after the preloading, and samples were unloaded
ear regime. Three layup configurations were investigated, 10° and removed for the scattering experiments.
and 45° unidirectional (UD) and 0°/90° x-ply. We find that subtle
alterations in matrix unit cell and crystallinity can be used as a 2.3. WAXS
feature set for training ML algorithms for predicting the relative
magnitude of exposed loads to different material samples with a Figure 1c,d illustrates the conducted WAXS experiment and
high accuracy. Three commonly used classification algorithms polyamide-4,10 unit cell. The experiments were conducted at
are considered and the performance of models trained on data the Swiss-Norwegian beamline BM01 at the European
from the separate composite layup configurations are compared, Synchrotron Radiation Facility (ESRF), Grenoble, France. The
but also when trained using data from the three layups energy of the beam was E = 17.8 keV and the beam size

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (2 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

(a) (b)

(c) (d)

Figure 1. a) Layups of the composites investigated. The 10° and 45° are UD and the 0°/90° is of a 4S stacking sequence. b) Dimensions of the samples
with tabs. The samples are loaded uniaxially, in the direction indicated by the red arrow. c) Geometry of the WAXS experiments. The samples are mounted
on a motorized stage, with the X-ray beam s in transmission mode perpendicular to the sample surface (x-, z-plane). Between each scan, the sample is
translated along r to probe the full length of the sample. Three parallel scan lines along r are performed for varying z. d) Illustration of the crystalline
polymer sheets within the matrix. The polymer chains are oriented along the c direction, and the intersheet and interchain directions correspond to the a
and b axes, respectively.

Figure 2. a) Uniaxial strain as a function of stress for all samples of the three layups. The loading of the different samples to various strains is represented
by colored curves, where the sample loaded until failure is shown in black. Following the initial loading, each sample is stretched to 10% of the maximum
load (not shown) to assess differences in stiffness. b) Measured stiffness of the samples before (circles) and after (triangles) loading.

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (3 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

Table 1. Number of data points collected by X-rays of each class for the intensity components from all crystalline peaks Ik, divided by
different datasets. the intensity of the amorphous phase Ia:
Z
Loading [% σmax]
P qmax
Layups I k dq
q
0 25 50 75 90 Total C ¼ kZ q min (2)
max
10° 44 230 233 229 235 968 I a dq
qmin
45° 24 227 233 343 467 1289
0°/90° 24 231 231 231 349 1061 The index k iterates over the seven crystalline peaks that are
10° þ 45° þ 0°/90° 92 688 697 803 1051 3318 present in the 1d scattering profile (Figure 3c). In addition, we
10° þ 45° 68 457 466 572 702 2257 also considered the full width at half-maximum and the relative
(inverse space) distances of the reflections from the polyamide.
These were taken directly from the results of the curve-fitting
procedure.
0.5  0.5 mm2. The samples were mounted horizontally on a
motorized stage, with the thickness of the sample normal to
the incident beam. Three parallel scan lines across the sample
3. Computational Section
lengths were performed (see Figure 1c) with a step size of
0.6 mm and an exposure time of 1 s per scan, totaling approx. Three supervised learning algorithms were chosen for predicting
200 point-like X-ray illuminations per sample (see Table 1). the mechanical testing history of the materials; LDA, SVMs, and
The scattered X-rays were recorded by a Pilatus 2M detector with gradient-boosted decision trees (GBDTs).
a pixel size of 172  172 μm2, placed 0.64 m downstream of the LDA [32] is a classification and dimensionality reduction tech-
sample, accommodating a q-range of approx. 3–30 nm1. nique which uses decision boundaries in the N-dimensional fea-
The isotropic regions of the diffraction signal were reduced to ture space to separate features of data belonging to different
the intensity I as a function of the scattering vector q. First, this classes. The class densities are modeled as a multivariate
required identifying the isotropic regions, which might vary due to Gaussian (where the training data is used to estimate the distri-
the out-of-plane waviness of the fibers.[34] The scattering images bution), and by assuming that they have a common covariance
were thus integrated radially, and a Gaussian was fitted to the matrix, the decision boundaries can be drawn where their poste-
resulting radial profile, giving an azimuthal maximum intensity rior probabilities are equal. Having closed-form solutions and no
χ 0 for each of the scattering images. Subsequently, the images hyperparameters to tune, LDA is both fast to compute and simple
were masked and integrated azimuthally, excluding angles to use.
χ 0  35∘ . This was done to exclude as much of the anisotropic SVMs[27] is a method which also uses hyperplanes to separate
signal as possible (i.e., matrix–fiber interphase), not directly relat- classes of data. The strength of this algorithm lies in its ability to
ing to the bulk matrix. Furthermore, the resulting 1d WAXS pro- achieve this in an optimal way when the classes cannot be sepa-
files were deconvoluted into their respective components by rated by a linear boundary, or when they overlap. The objective of
performing a curve fit to the experimental data from all the scans SVMs is to maximize the margin between the decision boundary
of the samples. The Bragg reflections and the amorphous halo of and two neighboring classes, while under the constraint that
the polymer were each represented by a Pseudo–Voigt function, some points are allowed to be on the wrong side of the margin.
and the background was fitted with a Kohlrausch–Williams–Watts As opposed to LDA, this method has a higher versatility in that
exponential[35] and a linear function. Prior to the fitting procedure, more complicated kernels (decision boundary shapes) can be
the scattering intensity was normalized to sum to one, to account applied, but the training complexity increases rapidly with large
for differences in beam intensity. The curve fitting was done using datasets.[37]
the lmfit library[36] in Python. GBDTs[38] is an ensemble learning method which uses deci-
From the fits to the WAXS profile, in total 21 parameters were sion trees as the base estimator. As opposed to averaging ensem-
extracted. Of these, 20 were associated with the polyamide-4,10 ble methods, gradient-boosted trees are trained sequentially,
unit cell, and one represented the degree of crystallinity in the where subsequent estimators “learn from the mistakes” of their
material. These parameters were subsequently employed in predecessors. This is achieved by more heavily weighing training
the supervised learning (vide infra). First, the relative intensities examples that have been misclassified by earlier estimators, and
of the peaks were calculated as: thus the training instances that are difficult to predict have a
Z q higher influence on estimators later in the chain. Predictions
max
I i dq are then made by combining a weighted majority vote from
qmin all of them.
I ij ¼ Z q , for peaks i 6¼ j (1)
max Table 1 lists the dataset specifications. Five datasets were
I j dq
qmin constructed from parameters of the reduced WAXS curve and
their associated mechanical testing histories. Three datasets con-
with i, j representing the [001], [002], [100], and [010/110] sisted of X-ray scans (illuminations) stemming from samples of
reflections (see e.g., Figure 3c), and the respective fiber orientations. The remaining two were a
[qmin, qmax] = [3,30] nm1, the q-range of the experiment. combination of data collected from all three layups, and the
Similarly, the crystallinity was estimated by the sum of the two UD, respectively. The fiber orientation was not included

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (4 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

Figure 3. a) Example of a diffraction pattern from the composites (10° layup). The Bragg reflections of the bulk polymer are annotated in red and
represented by the isotropic rings in the image. b) Intensity as a function of azimuthal angle from radially integrating the intensity in (a). The measured
anisotropy is perpendicular to the fiber axis and is due to the preferred orientation of the polymer crystals in the fiber–matrix. The maximum of the streak
χ 0 is found by fitting a Gaussian to the profile, and angles χ 0  35∘ are not included in the azimuthal integration. The inset depicts the azimuthal angle in
the (qr,qz)-plane. c) Chosen example of the azimuthally integrated diffraction pattern, with a representative curve-fit to the profile. The Bragg reflections of
the polymer unit cell are labeled corresponding to that of (a). The background signal is represented by an exponential and a linear function, and the Bragg
peaks of the crystalline and amorphous polymer phases are represented by the Pseudo-Voigt function. d) Box plots of the measured difference in position
of the 100 and 010/110 reflections in all samples.

as a feature in the mixed datasets. The five datasets were sepa- 1XN

rately used for optimizing, training, and testing instances of the Accuracy ½% ¼ Iðyi ¼ ŷi Þ  100 (3)
N i
three classification algorithms. In total, there were 21 input
features per sample scan, as described in Section 2.3, and five
Here I is the indicator function, N is the number of test instan-
possible outputs, corresponding to the relative amounts of
ces, ŷ is the predicted label, and y is the true label. Bootstrap val-
mechanical loading associated with the samples, described in
Section 2.2. idation of the algorithms with optimal parameters was
The chosen pipeline for training and independent testing of performed for 10 000 rounds (resampling of the training data),
the classifiers was as follows. The data were first split into 70% before retraining the models on the full training dataset, and ulti-
for training and validation and 30% for testing. Subsequently, mately evaluated on the external test data.
mean value μ and standard deviation σ were calculated for each We used the scikit-learn[39] implementations of LDA and
feature (in the training set). These were used for standardizing SVM, and the XGBoost library[40] for GBDT. The radial basis
the features to have a zero mean and a unit variance. An obser- function[41] was chosen as the kernel in the SVM. In addition,
vation xj belonging to feature i was thus transformed according to two other hyperparameters were important for the SVM’s
zj = (xjμi)/σi. Following this, the classification algorithms were performance, which related to the complexity of the decision
optimized by performing an exhaustive grid search of a surface and the influence of individual training samples,
combination of hyperparameters, using stratified fivefold respectively. These were optimized by the grid search routine.
cross-validation on the training data. The optimal set of param- The GBDTs were used with a Softmax objective function.
eters was chosen based on the algorithms’ performances in During optimization, we also tune the depth of each decision
terms of the classification accuracy. tree, the learning rate, and the number of training rounds.

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (5 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

The values of hyperparameters used are listed in Table S2, reflections associated with the α-phase, the profile shows three
Supporting Information. LDA was used with a singular value additional reflections from which the middle peak gives an evi-
decomposition solver, having no hyperparameters to tune. dence for the pseudo-hexagonal phase.[21] We assume that the
effect of mechanical load is easiest to be discerned from the
majority phase and focus on the α peaks alongside the amor-
4. Results and Discussion phous scattering. Particular attention is placed on the relative
peak positions with increasing tensile loading.
4.1. Mechanical Loading
Figure 3d shows a box plot of the normalized (inverse space)
distance between the [100] and [010/110] reflections as a function
Figure 2 shows the mechanical characterization of employed
of the mechanical loading, q̂ ¼ ðq010=110  q100 Þ=½q010=110 
materials. Figure 2a shows the tensile stress as a function of
strain for the three fiber orientations, where each test represents q100 0%σmax . The median value is represented by the orange line,
an individual sample. The curves of the samples with 10° and and the top and bottom represent the 25th and 75th percentiles,
0°/90° orientation display discontinuities, which are indications respectively. Here, the whiskers are determined by multiplying
of damage events such as fiber failure. All curves overlap to a the interquartile range by 1.5, outliers represented by the circles.
good approximation, demonstrating reasonably good reproduc- The variability of the measured values along the scanlines within
ibility. However, when comparing the initial modulus of each samples is relatively low, indicating that the effects from the
experiment (see Figure 2b), the modulus of the pristine samples mechanical loading manifest homogeneously across the sample
displays large deviations for all layups, even up to 20% variation length.
between the minimum and maximum. This is not uncommon Jones et al.[21] showed that in a well-crystallized
for thermoplastic composites and finds its origin in its intrinsic polyamide-4,10 the Bragg peak positions are such that
scatter, i.e., its sensitivity to fiber waviness, (local) fiber orienta- q010/110  q100 = 2.8 nm1, a value substantially higher than in
tion, and (local) fiber volume fraction.[34] our experimental data (approx. 2.55 nm1). Lower inverse space
The statistical relevance of the magnitude of stiffness can be distances between these peaks are commonly associated with lat-
improved by adding more measurements. However, it will be tice defects in polyamides.[43] Therefore, we infer that, due to the
difficult to use classical approaches linking secant modulus to relatively fast cooling rate during processing, in combination
mechanical loading,[7,8] since the variability of the local stiffness with topological constraints (entanglements), the polymer chains
would exceed the anticipated decrease in stiffness caused by end up forming crystals where lattice defects are more frequent
damage. This is also reflected in the stiffness measured during than expected.
the secondary loading, displaying no clear trend of stiffness Figure 3d also shows that the value of q010/110  q100 tends
reduction after preloading. to increase. This indicates that molecules use mechanical energy
Although one might argue that the lack of stiffness reduction to rearrange themselves and let lattice defects migrate out of the
implies no damage is induced by the preload, it is important to crystalline structures. This crystal perfectioning mechanism is
realize that the stiffness of the composite is dictated by the car- pronounced for the samples with 10° fiber orientation and
bon fibers. However, since the matrix material is viscoelastic- much less/not present for samples with 45° and 0°/90° fiber ori-
viscoplastic, there are various stress-induced deformation and entations. This is consistent with the observation that samples
relaxation mechanisms on the molecular level which can intro- with 10° fiber orientation exhibit the largest macroscopic
duce (local) changes that do not affect the macroscopic stiffness. strain-at-break (see Figure 2a) and, therefore experience a larger
As a result, a decrease in stiffness is actually not always present[9] molecular mobility gain.
and it is often difficult to distinguish between compliance
changes caused by creep and actual damage.[6] 4.3. History Classification

4.2. WAXS Figure 4 presents various aspects of the classification algorithms’


performance. In Figure 4a, the classification accuracies on
Figure 3 summarizes the results of the X-ray experiments. unseen test data are shown for the three classifiers, respectively,
Figure 3a shows a chosen example of the diffraction pattern from for each of the five datasets introduced in Section 3 (Table 1).
one sample with 10° fiber orientation. The labeled isotropic When trained on the datasets consisting of the UD 10° layup (rep-
reflections are contributions from the polyamide-4,10 matrix resented by the first set of columns in Figure 4a), LDA, SVM, and
reported in ref. [21]. There is also an anisotropic reflection appar- GBDT all achieve accuracy scores above 96%, showing good gen-
ent at χ  80∘ , perpendicular to the fiber axis. At the higher scat- eralization from the training data to unseen data samples. The
tering angles, this is attributed to the more oriented crystalline accuracies of the predictions are slightly worse on the individual
domain in the matrix–fiber interphase,[22] and the tail of the sig- UD 45° and 0°/90° datasets (column sets 2 and 3 in Figure 4a),
nal stemming from elongated pore structures within the fibers at with the lowest score at 87.5% test accuracy using LDA, and the
lower angles.[42] The radially averaged profile of the latter is highest using SVMs with 91.3%.
shown in Figure 3b, for scattering angles up to q = 4, used for The datasets consisting of samples from different fiber
determining the local fiber orientation of the samples. orientations, have features belonging to the same class which
Figure 3c shows a representative example of the fit to the are more scattered, i.e., do not necessarily follow similar trends
reduced scattering profile. The α-phase dominates the curve with (conversely to those of the individual layups), and thus the learn-
the distinct [100] and [010/110] reflections.[21] Apart from the ing task is conceivably more complex (see e.g., Figure 3d). Here,

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (6 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

Figure 4. Performance of the classification algorithms trained on the different datasets in terms of classification accuracy on the external test data.
Subfigure a) shows the performance in the multiclass case, and b) shows the corresponding scores of the binary classification setting. c) Accuracy
scores per layup of the classification algorithms trained on the full dataset (all layups), with multiclass classification as the objective. The percentage
change in accuracy compared to the algorithms trained only on data from the respective layups is also shown. d) Confusion matrix of the GBDT pre-
dictions on unseen samples when trained on the full dataset (all layups). The rows and columns represent the number of instances of actual and predicted
loading, respectively.

the SVMs and GBDTs are able to describe the variance in the data instances follow the diagonal as correct predictions. The super-
to a reasonably high degree, whilst the LDA model struggles in and subdiagonals represent the slight miss-classifications, where
comparison, as shown in Figure 4a (column sets 4 and 5). This is samples are predicted to have only a slightly better or worse con-
not all that surprising, since one of the assumptions of LDA is dition, in terms of the loading history. Only a few instances fall
that the covariance of the features across classes is identical,[32] into this category. More severely, there are 17 miss-classifications
which they in this case clearly are not. Figure 4c further demon- in the cell corresponding to the samples loaded to 75% of the
strates the performances using this dataset, by showing the maximum load, which are here classified as having only a
amount of correctly predicted data samples within each of the 25% σmax loading history. This corresponds to approx. 7% of
three fiber orientations. For comparison, the change in accuracy the total number of samples with 75% σmax as the true label.
compared to the algorithms trained on data from the respective The cells of “extreme” miss-classifications (0–90 pairs) have
layups only is also shown. The SVMs and GBDTs show only no entries. The figure also highlights that the 0% σmax class is
slight performance decrease compared to their counterparts underrepresented, with only 19 (68%) samples correctly
trained on single fiber orientation data only, while that of classified.
LDA drops more drastically, more than 30% on the 10° data. Table 2 shows the training scores of the algorithms, together
Figure 4d shows the confusion matrix of the GBDTs predic- with the mean and standard deviations from the bootstrap resam-
tions on unseen samples of the multilayup dataset. Most of the pling routine. Although the GBDTs show the best performance

Table 2. Performance of the classification algorithms in terms of classification accuracy on the training dataset and external test dataset (multiclass). The
corresponding mean scores and standard deviations from the Bootstrap resampling routine are also listed.

LDA SVM GBDT


Layups Train Test Bootstrap Train Test Bootstrap Train Test Bootstrap
10° 97.5 96.2 96.6 (1.1) 98.1 96.2 97.2 (1.0) 100.0 97.3 97.3 (1.0)
45° 90.5 89.5 89.1 (1.5) 91.9 91.3 90.3 (1.4) 98.0 89.7 89.7 (1.5)
0°/90° 87.0 87.5 85.9 (1.9) 95.3 90.0 90.4 (1.7) 100.0 90.3 89.4 (1.7)
10°, 45°, 0°/90° 74.6 72.4 73.2 (1.3) 97.1 88.2 90.2 (0.9) 100.0 90.0 91.0 (0.9)
10°, 45° 81.7 79.3 80.2 (1.5) 96.9 92.1 91.7 (1.0) 98.6 91.3 92.3 (1.0)

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (7 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

on the test data in most cases, they have training scores close to Supporting Information) of the algorithms.[44] These are qualita-
100%, and test scores that are considerably lower. This indicates tively similar to those presented in Figure 4b, and are shown in
overfitting, and that more regularization could be used to further Figure S2, Supporting Information.
improve the generalization to unseen data samples. This is also
the case for the SVMs, most notably on the dataset of all layups,
4.5. Performance on Smaller Datasets
where the accuracy score drops from 97.1% on the training data
to 88.2% on the external test data. When comparing the test
Collecting X-ray data from material samples can often be both
scores of the classifiers from all experiments to the mean values
expensive and time-consuming. ML algorithms may also require
from the bootstrap resampling, we note that they are mostly
large amounts of data to converge and be viable as a modeling
within the estimated standard deviations. This indicates that
technique. It is therefore interesting, from a practical standpoint,
the test scores are quite representative for the general perfor-
to have an idea of how much data, i.e., X-ray scans, is required to
mance of the algorithms. The distributions of the test scores
successfully train a classifier with a sufficient performance. We
from the Bootstrap resampling are shown in Figure S1,
therefore investigate how the classification algorithms perform
Supporting Information.
when the relatively large amount of data available here is dimin-
ished. This numerical experiment is performed using subsets of
4.4. Binary Classification data from the dataset of all sample layups for training the clas-
sifiers, with classifying the five loading histories as the objective.
The results in Section 4.3 demonstrate that the samples with Figure 5 shows the performance of the three algorithms as a
five different loading histories can be distinguished with high function of the number of data points used for training and vali-
accuracy, using features from the X-ray experiments as data for dation. Data points were drawn randomly from the full dataset,
training classification algorithms. In engineering applications, and training the models followed the same procedure as in
it may be sufficient to have a more granular view of whether a Section 3, with the exception that the full amount of prescribed
certain part needs replacement or is a candidate for further data was used for training and validation (no data was set aside
inspection, i.e., above or below some threshold of exposed load. for external testing). Each point along the curve in Figure 5 is the
Here, this scenario was simulated by redefining the class labels mean value of the validation score from 10 000 bootstrapping
of the X-ray features of the different samples. One class con- rounds, and the shaded areas are the standard deviation.
sisted of the unloaded materials together with those loaded (In the resampling procedure, we allow drawing data samples
to 25% of the maximum while the second class consisted of from the full dataset in a stratified manner, but the number of
those loaded to 50% of the maximum and above. This threshold samples is restricted to the prescribed amount, indicated by the
could in principle have been arbitrary, but here it was chosen x-axis of Figure 5). There is a rapid increase in the accuracy of
based on the failure events of the materials loaded to 50% all three algorithms up to approximately 75% classification accu-
and above, visible in the stress–strain curves in Figure 2. racy, or 10% of the full dataset size. Following this, the SVMs and
Optimizing the algorithms followed the same procedure as GBDTs behave similarly and have slight increase in performance
described in Section 3. with increasing amounts of training data. The curves of these two,
Figure 4b shows the classification accuracies on the unseen do not stabilize, and their trends are still positive when granted the
test data. The overall performance of the GBDTs and SVMs is
similar when compared to the multiclass case (see Figure 4a),
though LDA displays varying results in comparison. On the
combined datasets (all layups þ UD layups), the performance
of LDA increases drastically but worsens when predicting the
45° and 0°/90° samples. Intuitively, the learning task becomes
less complex when going from five to two classes, so one would
expect an increase in performance given that the features of
one class (in this case 0% and 25% σmax) are more similar
to each other than to the second class (50, 75, and 90%
σmax). In contrast, the dataset in the binary case consists of data
points from more than just one material sample, meaning that
the algorithms also need to account for variability of the fea-
tures across samples, which again makes the learning objec-
tive more complex. Considering the performance across the
five discussed datasets, the results in Figure 4a,b suggest that
these “mechanisms” are (for better or worse) affecting the per-
formance of the simpler LDA model, while the GBDTs and
SVMs seem versatile enough to adapt and perform well
regardless. Figure 5. Median validation accuracy as a function of dataset size for the
Since this binary classification problem is unbalanced with three classification algorithms. At each point, bootstrap validation is
respect to the number of instances belonging to each class performed for 10 000 rounds. The standard deviations are indicated by
(see Table 1), we also report the F1-scores (Equation S(2), the shaded areas.

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (8 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

full dataset, at which point their performance is comparable to Supporting Information


what was found in Figure 4. The LDA classifier’s performance
Supporting Information is available from the Wiley Online Library or from
looks to stabilize at 75% accuracy and does not improve when
the author.
using larger amounts of data for the training.

5. Conclusion Acknowledgements
This research forms part of the research program of DPI, project No. 854.
WAXS and ML can be applied for discriminating This research has received funding from the European Union’s Horizon
between the mechanical loading history of carbon fiber compos- 2020 research and innovation program under the Marie Skłodowska-Curie
ite materials. This was demonstrated for three layup configura- grant agreement no. 956248. B.P. and J.O.F. acknowledge support from
tions (i.e., 10° UD, 45° UD, and 0°/90° x-ply) uniaxially stretched Norwegian Research Council Nano2021 grant no. 250619. The authors
to various loads. This procedure was based on employing char- also thank Marcel Teeuwen of DSM for conducting the mechanical experi-
ments, the staff at beamline BM01 of the ESRF for aiding the scattering
acterizations of the polyamide matrix unit cell as a feature set for
experiments, as well as Enno Klop of Teijin Aramid and Gerwin Ingenbleek
optimizing three supervised classification algorithms, LDA, of Shell for fruitful discussions.
SVMs, and GBDTs, with the learning objective of classifying
the material samples into five distinct classes, corresponding
to the amount of exposed mechanical loading. X-ray measure- Conflict of Interest
ments indicate that the samples with 10° fiber orientation
undergo distinct crystal enhancement when exposed to the The authors declare no conflict of interest.
mechanical loading. The loading history of these samples was
thus easier to predict by the models, compared to those with
45° and 0°/90° fiber orientation. Data Availability Statement
The classifiers were subsequently generalized by expanding
The data that support the findings of this study are available from the
the dataset to include features stemming from the three layups corresponding author upon reasonable request.
altogether. The GBDTs and SVMs performed well also in this
case, having accuracy scores of approx. 90%. A binary classifica-
tion task was also investigated by segregating materials into two
classes of mechanical loading, based on whether there were visi- Keywords
ble failure events in the stress–strain curves from the mechanical
artificial intelligence, carbon fiber composites, structural health
testing. The results showed that the accuracy of SVMs and monitoring, wide-angle X-ray scattering
GBDTs was similar to that of the multiclass case, while LDA dis-
played a varying performance depending on the layup. Received: September 5, 2023
Considering the demonstrated versatility of GBDTs and Revised: November 6, 2023
SVMs, the results suggest that these two algorithms seem the Published online: November 28, 2023
best candidates out of the three for correlating WAXS data to
exposed loads in discussed composites. This takes into account
[1] D. D. L. Chung, Mater. Sci. Eng.: R: Rep. 2017, 113, 1.
that composites with different fiber orientations might display
[2] S.-S. Yao, F.-L. Jin, K. Y. Rhee, S.-J. Hui, D. Park, Composites, Part B
dissimilar trends in the evolution of the unit cell morphology,
2018, 142, 241.
with respect to the magnitude of mechanical loading. ML algo-
[3] M. A. Karatas, H. Gokkaya, Def. Technol. 2018, 14, 318.
rithms with nonlinear decision boundaries might be required to [4] K. Niendorf, R. B., Adv. Eng. Mater. 2022, 23, 2001002.
explain the variance in the data. [5] H. Zheng, W. Zhang, B. Li, J. Zhu, C. Wang, G. Song, G. Wu, X. Yang,
This study shows that correlations between the mechanical Y. Huang, L. Ma, Composites, Part B 2022, 233, 109639.
loading history and nanoscopic information from composites [6] A. Malpot, F. Touchard, S. Bergamo, Procedia Eng. 2015, 133, 136.
without visible macroscopic cracking or delamination can be [7] A. Gagel, D. Lange, K. Schulte, Composites, Part A 2006, 37, 222.
found. The study does not consider the importance of the individ- [8] K. Reifsnider, R. Jamison, Int. J. Fatigue 1982, 4, 187.
ual features used for building the models. For instance, it may be [9] K. B. Pettersson, J. M. Neumeister, E. K. Gamstedt, H. Öberg,
sufficient to use information from a subset of the Bragg reflections Compos. Struct. 2006, 76, 151.
for making these predictions. This may be an important aspect [10] B. Wang, S. Zhong, T.-L. Lee, K. S. Fancey, J. Mi, Adv. Mech. Eng.
from an application stand-point, where peak intensities or reach- 2020, 12, 4.
[11] G. Franz, M. H. Hassan, Structural Integrity and Monitoring for
able scattering angles can be limited. Additionally, the basis for
Composite Materials, Springer, Singapore 2023, p. 126.
these results was limited to using characterizations of the polyam-
[12] M. Kersemans, A. Martens, K. Van Den Abeele, J. Degrieck,
ide unit cell in the bulk matrix only—There are other features pres-
F. Zastavnik, L. Pyl, H. Sol, W. Van Paepegem, J. Nondestr. Eval.
ent in the wide angles of the scattering pattern that might augment 2014, 33, 522.
the feature space, i.e., the fiber–matrix interphase, as known in the [13] J. P. McCrory, S. K. Al-Jumaili, D. Crivelli, M. R. Pearson, M. J. Eaton,
prior art.[22] Scattering fingerprints from the lower scattering C. A. Featherston, M. Guagliano, K. M. Holford, R. Pullin, Composites,
angles might also augment the feature space for making these pre- Part B 2015, 68, 424.
dictions, for example, larger microscopic defects and pores in the [14] J. Segers, S. Hedayatrasa, E. Verboven, G. Poelman, W. Van
carbon fibers, as discussed elsewhere.[42] Paepegem, M. Kersemans, NDT Int. 2019, 102, 218.

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (9 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH
15272648, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adem.202301415 by Library (Periodicals) Vikram S, Wiley Online Library on [03/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.aem-journal.com

[15] S. W. F. Spronk, M. Kersemans, J. C. A. De Baerdemaeker, [30] H. Kurita, M. Suganuma, Y. Wang, F. Narita, Adv. Eng. Mater. 2022,
F. A. Gilabert, R. D. B. Sevenois, D. Garoz, C. Kassapoglou, 24, 2101072.
W. Van Paepagem, Polym. Test. 2018, 65, 231. [31] A. S. Obdrup, D. F. Wieland, M. K. Huss-Hansen, M. M. Arras,
[16] C. Meola, S. Boccardi, G. Carlomagno, N. Boffa, E. Monaco, F. Ricci, M. Knaapila, Polym. Test. 2022, 109, 107524.
Compos. Struct. 2015, 134, 845. [32] T. Hastie, R. Tibshirani, J. H. Friedman, J. H. Friedman, The Elements
[17] J. Dong, P. Pomarède, L. Chehami, A. Locquet, F. Meraghni, of statistical Learning: Data Mining, Inference, and Prediction, Vol. 2,
N. F. Declercq, D. Citrin, NDT & E Int. 2018, 99, 72. Springer 2009, pp. 106–111, Ch. 4.3.
[18] Z. Li, Z. Meng, C. Soutis, A. Haigh, P. Wang, A. Gibson, Adv. Eng. [33] H. Liu, S. Liu, Z. Liu, N. Mrad, A. S. Milani, IEEE Trans. Ind. Electron.
Mater. 2022, 24, 2100494. 2020, 68, 2532.
[19] F. Sket, A. Enfedaque, C. Alton, C. González, J. M. Molina-Aldareguia, [34] A. Gillet, O. Mantaux, G. Cazaurang, Composites, Part A 2015, 75, 89.
J. Llorca, Compos. Sci. Technol. 2014, 90, 129. [35] G. Williams, D. C. Watts, Trans. Faraday Soc. 1970, 66, 80.
[20] J. E. Engel, G. Georgeson, M. Safai, (The Boeing Co.), US. 10,317,349, [36] M. Newville, T. Stensitzki, D. B. Allen, M. Rawlik, A. Ingargiola,
2019. A. Nelson, Astrophysics Source Code Library 2016, ascl:1606.014.
[21] N. Jones, E. Atkins, M. Hill, S. Cooper, L. Franco, Polymer 1997, 38, 2689. [37] C.-C. Chang, C.-J. Lin, ACM Trans. Intell. Syst. Technol. 2011, 2, 1.
[22] D. Kobayashi, Y.-T. Hsieh, A. Takahara, Polymer 2016, 89, 154. [38] J. H. Friedman, Comput. Stat. Data Anal. 2002, 38, 367.
[23] Y. Liu, L. Cui, F. Guan, Y. Gao, N. E. Hedin, L. Zhu, H. Fong, [39] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
Macromolecules 2007, 40, 6283. O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg,
[24] C. Millot, R. Seguela, O. Lame, L.-A. Fillot, C. Rochas, P. Sotta, J. Vanderplas, A. Passos, D. Cournapeau, J. Mach. Learn. Res.
Macromolecules 2017, 50, 1541. 2011, 12, 2825.
[25] K. Hornik, M. Stinchcombe, H. White, Neural Networks 1989, [40] T. Chen, C. Guestrin, in Proc. of the 22nd ACM SIGKDD Int. Conf. on
2, 359. Knowledge Discovery and Data Mining, ACM, New York 2016, 785.
[26] P. M. Pawar, S. N. Jung, J. Intell. Mater. Syst. Struct. 2008, 19, 1217. [41] B. Schölkopf, K. Tsuda, J.-P. Vert, Kernel Methods in Computational
[27] T. Hastie, R. Tibshirani, J. H. Friedman, J. H. Friedman, The Elements Biology, MIT Press, Cambridge, MA 2004.
of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed., [42] H. Peterlik, P. Fratzl, K. Kromp, Carbon 1994, 32, 939.
Springer, New York, NY, USA 2009, pp. 417–419, Ch. 12.2. [43] D. Cavallo, L. Gardella, G. C. Alfonso, G. Portale, L. Balzano,
[28] S. Qiao, M. Huang, Y.-J. Liang, S.-Z. Zhang, W. Zhou, Polym. Compos. R. Androsch, Colloid Polym. Sci. 2011, 289, 1073.
2023, 44, 2427. [44] N. Chinchor, in MUC4 ’92: Proc. of the 4th Conf. on Message
[29] S. Lloyd, IEEE Trans. Inf. Theory 1982, 28, 129. Understanding, ACL, Virginia 1992, 22.

Adv. Eng. Mater. 2024, 26, 2301415 2301415 (10 of 10) © 2023 The Authors. Advanced Engineering Materials published by Wiley-VCH GmbH

You might also like