Machine Learning with Applications 19 (2025) 100609

Contents lists available at ScienceDirect

Machine Learning with Applications

journal homepage: www.elsevier.com/locate/mlwa

Application of convolutional neural networks and ensemble methods in the

fiber volume content analysis of natural fiber composites
Florian Rothenhäusler a ,1 , Rodrigo Queiroz Albuquerque a ,1 , Marcel Sticher a ,
Christopher Kuenneth b , Holger Ruckdaeschel a ,∗
a
Department of Polymer Engineering, University of Bayreuth, Universitätsstrasse 30, Bayreuth 95447, Bavaria, Germany
b
Faculty of Engineering Science, University of Bayreuth, Universitätsstrasse 30, Bayreuth 95447, Bavaria, Germany

ARTICLE INFO ABSTRACT

Dataset link: https://github.com/Polymer-Engi The incorporation of natural fibers into fiber-reinforced polymer composites (FRPC) has the potential to
neering-University-Bayreuth/FiberVolumeCont bolster their sustainability. A critical attribute of FRPC is the fiber volume content (FVC), a parameter that
ent_ConvNet profoundly influences their thermo-mechanical characteristics. However, the determination of FVC in natural
Keywords: fiber composites (NFC) through manual analysis of light microscopy images is a labor-intensive process. In
Flax fibers this work, it is demonstrated that the pixels from light microscopy images of NFC can be utilized to predict
Sustainability FVC using machine learning (ML) models. In this proof-of-concept investigation, it is shown that convolutional
Machine learning neural network-based models predict FVC with an accuracy required in polymer engineering applications, with
Image detection a mean average error of 2.72 % and an 𝑅2 coefficient of 0.85. Finally, it is shown that much simpler ML models,
Natural fiber composites non-specialized in image recognition, besides being much easier and more efficient to optimize and train, can
CNN also deliver good accuracies required for FVC characterization, which not only contributes to the sustainability,
but also facilitates the access of such models by researchers in regions with little computational resources. This
study marks a substantial advancement in the area of automated characterization of NFC, and democratization
of knowledge, offering a promising avenue for the enhancement of sustainable materials.

1. Introduction The mechanical properties of FRPC, influenced by factors such as

fiber length (Fu & Lauke, 1996; Gupta et al., 1989), fiber orienta-
Fiber-reinforced polymer composites (FRPC) find extensive use in tion (Chin et al., 1988; Fu & Lauke, 1998), and fiber volume content
automotive, wind energy generation, and aerospace applications, pri- (FVC) (Affdl & Kardos, 1976), are crucial in diverse engineering con-
marily due to their high weight-specific Young’s modulus and strength texts. The FVC, denoting the ratio of fiber volume to the total volume in
(Ashby, 1992; Ehrenstein, 2006), which are attributed to the reinforc- a FRPC (Schürmann, 2007), is particularly integral in determining spe-
ing fibers (Flemming et al., 1995; Schürmann, 2007). Traditionally, cific mechanical properties (Voigt, 1889) and is expressed by Eq. (1):
synthetic fibers such as glass and carbon have been utilized as rein- 𝑉f iber
𝐹𝑉 𝐶 = (1)
forcements for FRPC. The production of these fibers involves substantial 𝑉composit e
energy consumption, leading to relevant CO2 emissions (Das, 2011).
In contrast, plant-based fibers require less energy for production and For composites reinforced with glass or carbon fibers, FVC analy-
processing (Bos, 2004). Additionally, natural fibers are biodegradable, sis is typically conducted through techniques like thermo-gravimetric
presenting an environmentally friendly alternative to glass and carbon analysis (TGA) (DIN EN ISO 1172:1998-12, 1998; Grund et al., 2019),
fibers. Plant-based fibers, like flax and hemp, also have lower den- density measurements (Abdalla et al., 2008), computed tomography
sity and cost compared to synthetic fibers (Bledzki & Gassan, 1997). (CT) (Hessman et al., 2019), or light microscopy (LM) (Jock, 1986).
However, plant-based fibers exhibit lower strength and modulus than
However, natural fiber composites (NFC) pose challenges due to their
synthetic fibers (Flemming et al., 1995). Despite this, they can be a
organic nature, low thermal stability, and characteristic hollow
sustainable alternative in applications where the highest strength and
structure (Kumar et al., 2022; Mwaikambo & Ansell, 2001). Analysis
modulus are not essential.

∗ Corresponding author.
E-mail addresses: [email protected] (F. Rothenhäusler), [email protected] (R.Q. Albuquerque),
[email protected] (M. Sticher), [email protected] (C. Kuenneth), [email protected] (H. Ruckdaeschel).
1
Contributed equally to this work.

https://doi.org/10.1016/j.mlwa.2024.100609
Received 23 September 2024; Received in revised form 15 November 2024; Accepted 26 November 2024
Available online 2 December 2024
2666-8270/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

methods such as TGA may be hindered by the low thermal stability of Epoxy, based in Stade, Germany. The Heloxy modifier BD was procured
natural fibers, especially when additives with low thermal stability are from Westlake Epoxy GmbH, headquartered in Duisburg, Germany.
present in the matrix. Similarly, the hollow structure of natural fibers Diethyl toluene diamine (DETDA), characterized by an active hydrogen
complicates FVC determination through density measurements, and the equivalent weight of 44.57 g mol−1 , was provided by Schill+Seilacher
organic nature makes precise CT detection challenging. Consequently, GmbH, situated in Böblingen, Germany, and was supplied in the form
LM analysis of polished cross-sections of unidirectional NFC is often of Polyvertec® H81. Jeffamine® D-230 was obtained from Huntsman
preferred. Corporation, located in Salt Lake City, US.
Manual analysis of LM images via image processing software proves For the natural reinforcement component, non-crimp flax fabric
to be laborious and time-intensive. In contrast, artificial neural net- FUSE 200 FLX was employed, sourced from SachsenLeinen GmbH
works (ANN), such as convolutional neural networks (CNN), offer a in Markkleeberg, Germany. FUSE 200 FLX comprises unidirectional,
faster and precise alternative once properly trained (Andrade & Ed- non-twisted yarns bound together by a thermoplastic fiber binder,
uardo, 2011; Chauhan et al., 2018). Previous applications of neural constituting approximately 5% of the total fabric weight.
networks have demonstrated substantial reductions in workload, exem-
plified in defect detection in metal screws and the inline assessment 2.2. Resin formulation & laminate curing
of part porosity in thermoplastic polymers produced via powder-based
additive manufacturing (Schlicht et al., 2022; Song et al., 2018). DGEBF and DETDA were weighed in so that the stoichiometric ratio
Different machine learning (ML) models have been used to model between the active hydrogen atoms in DETDA to the epoxy groups in
mechanical properties of polymer composites, as highlighted in the DGEBF is equal to one. The blending of DGEBF and DETDA involved
recent review by Sorour et al. (2024). Machello et al. (2024) trained de- the utilization of a centrifuge speed mixer from Hauschild Engineering,
cision tree-based models using features like the exposure temperature, located in Hamm, Germany, operating at 3000 min−1 for a duration of
𝑇g and sample thickness to predict the tensile strength retention of a 120 s.
FRPC material, achieving 𝑅2 of 0.88–0.96 for the test set. Osa-Uwagboe Flax fibers, cut to lengths of 10 cm, were manually impregnated
et al. (2024) screened several ML models that used quasi-static inden- with the resin. After aligning the fibers uni-directionally and placing
tation data as features to predict damage properties in FRPC materials, them within a vacuum bag (maintained at approximately 1 mbar), the
achieving 𝑅2 larger than 0.98 for the test set. Shokrollahi et al. (2023) composite underwent curing in a Memmert ULE 400 convection oven
have employed deep CNNs to predict local stress fields within 2D from Memmert GmbH + Co. KG, based in Schwabach, Germany. The
slices of composite microstructure images, revealing that this ML-based
curing process consisted of one hour at 120 ◦ C followed by three hours
technique can be an alternative to costly force electron microscopy
at 160 ◦ C. Subsequently, the composite was gradually cooled to room
measurements.
temperature over a 4-h duration.
The importance of determining FVC via ML models and image
The resulting composite was then sectioned into pieces measuring
recognition lies in their much faster, cheaper and more sustainable
15 mm by 15 mm using a Mutronic DIADISC5200 diamond plate saw
nature as compared to traditional destructive methods. To the best of
from MUTRONIC Präzisionsgerätebau GmbH & Co. KG, located in
our knowledge, no paper has been found describing the use of ML
Rieden am Forggensee, Germany. These specimens were embedded in
models to determine FVC of natural fiber composites from LM images.
a matrix composed of DGEBA, Heloxy modifier BD, and Jeffamine® D-
However, one paper has been found describing the use of CNN models
230, maintaining a stoichiometric ratio between the functional groups.
to determine FVC of carbon fiber composites (Blarr et al., 2024) from
The embedding matrix underwent a curing process lasting 12 h at 40 ◦ C
3D CT images. This model exhibited high variance and was trained
in a Memmert ULE 400 convection oven from Memmert GmbH + Co.
using a very small experimental data set (14 samples) with a small
KG.
range of FVC (22%–29%), and the trained CNN showed a very low
𝑅2 score of −0.11, which is apparently related to the training error.
2.3. Grinding & polishing
The current investigation reveals that it is possible to train an accurate
CNN model with a FVC range five times larger and still get high 𝑅2
coefficients and small errors for the prediction of this property on both The preparation of specimens involved grinding and polishing pro-
the training and test sets. cedures using a Struers RotoPol-21 equipped with a Struers RotoForce-
This proof-of-concept study aims at demonstrating that pixels from 4, both provided by Struers GmbH in Ottensoos, Germany. For the
LM images of NFCs can be used to train CNN models to predict FVC grinding process, silicon carbide papers from Struers with varying grain
with accuracies good enough for being used in usual polymer engi- sizes ranging from 68 μm to 10 μm were employed. Each grinding step,
neering applications. It is also shown that comparatively simpler ML lasting 10 min, applied a force of 10 N to each specimen. The specimens
models like the eXtreme Gradient Boosting (XGBoost), which are much were rotated in the opposite direction to the silicon carbide papers,
easier and faster to optimize and train, but which are not specialized which rotated at 300 rpm. Subsequently, polishing was carried out
in image recognition problems, can also predict FVC and deliver an using a diamond dispersion featuring 6 μm grain sizes, with the spec-
accuracy equally acceptable for FVC characterizations found in polymer imens rotating at 150 rpm under a force of 5 N. The final step involved
engineering problems. Both models are trained using LM images of an cleaning the specimens in an ultra-sonic bath utilizing de-ionized water.
NFC produced through the hand-layup process involving flax fibers and
an epoxy resin matrix. Afterwards, the performance of the ML models 2.4. Light microscopy
is improved by employing ensemble techniques.
The analysis of the composite’s cross-section was conducted using
2. Methods a Leica DM 6000M microscope from Leica Microsystems GmbH in
Wetzlar, Germany, equipped with a lens providing 50 times magni-
2.1. Materials fication. The resulting images had dimensions of 200 μm by 150 μm
(width × height). Gray scale images were captured in bright field mode,
The diglycidyl ether of bisphenol F (DGEBF) utilized in this study, utilizing a brightness setting of 85%, exposure time of 16.8 ms and a
exhibiting an epoxide equivalent weight (EEW) of 169 g mol−1 , was gamma value of 1.3. The images were randomly acquired throughout
obtained in the form of Epilox F17-00 from LEUNA-Harze GmbH, the entire cross-section of the NFC and saved as TIFF files. Exemplary
located in Leuna, Germany. The diglycidyl ether of bisphenol A (EEW = images are presented in Fig. 3(a). The distribution of FVC across all 290
187 g mol−1 ) was sourced from Blue Cube Assets GmbH & Co. KG, Olin recorded images is illustrated in Fig. 3(b).

2
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

2.5. Image analysis

The analysis of LM images was conducted using a routine pro-

grammed in ImageJ image processing software (ImageJ 2.14.0/1.54f).
However, automated processing was hampered by inhomogeneous il-
lumination of the images due to the light microscopes’ illumination
setup. The light cone renders the center of the image darker than the
border areas resulting in problems in setting the threshold between gray
scale values of matrix and fibers. Therefore, an artificial background
was created beforehand by overlaying several areas of images where no
fibers are present to obtain an image with only matrix. On this image a
Gaussian blur filter was applied using ImageJ with a sigma (radius) of
100 to average over local gray value fluctuations. The image processing
routine performs the following steps:

1. Directory: creating a directory where the raw images are stored,

one where the processed images are going to be stored and
inquiring an artificial background. The filenames are stored in
a list and a for loop iterates in the range of the length of the list,
i.e. over all files.
2. Background correction: subtracting background from raw im-
ages and saves the result in 5_ramps LUT (pseudo-color im-
age look-up tables) (ImageJ, 2023) to check the illumination
homogeneity.
3. Enhance contrast: increasing contrast by normalization and
equalization of pixel gray value histogram with saturated pixels
set to 0.1%.
4. Thresholding: setting a threshold by localizing the local mini-
mum between the fiber and the matrix peak in the gray scale
histogram using the ‘minimum’ algorithm. Results in a binary
image with white features and black background (matrix), which
is saved.
5. Denoising: The despeckle tool is applied iteratively to eliminate
artifacts.
6. Eroding and Dilating: Erodes (removes pixels) and dilates (adds
pixels) all features (white) in the binary image using the same
number of iterations for both operations. This removes features
that measure only few pixels in at least one dimension, e.g. grind
marks and dust, while retaining the shape and size of larger
structures, i.e. the fibers.
7. Measure: The area fraction (% of white pixels) of the binary
image is measured and appends it together with the filename
to a list.
Fig. 1. Exemplary image during various stages of the analysis.
8. Overlay: The original image and the binary image are overlaid
and flattened. The resulting image is converted to 8-bit for-
mat and saved. The image processing efficacy was assessed by
visually evaluating the mapping accuracy of the overlay images. where 𝑤𝑖 ’s are weight coefficients which are kept constant each time
the patch slides over the image, and 𝑏 is an added bias (a constant). The
This standardized procedure was consistently applied to all images result of this operation is sent through an activation function (here, the
analyzed (see Fig. 1). In this context, FVC was quantified through image sigmoid function, 𝜎), so that a final number (or output) is generated.
analysis, where Eq. (2) expresses the ratio of pixels corresponding to In Eq. (3), 𝜎(𝑥) is given by
fibers (white) to the total number of pixels in the image: 1
𝜎(𝑥) = (4)
𝑛 1 + 𝑒−𝑥
𝐹 𝑉 𝐶 = f iber (2)
𝑛t ot al
The unmodified 16-bit .tiff images, along with the calculated FVC, were In Eq. (4), a number between zero and one is generated, which is
associated with the probability of detecting a specific pattern. Each
provided as input to the CNN.
time the patch slides to a different location, it processes nine different
neurons and a different output is generated, which is stored in the so-
2.6. Convolutional neural network called feature map. Each convolutional layer can have many feature
maps corresponding to the detection of different (local or global) pat-
Background. CNNs are an special type of neural networks specialized terns. Local patterns are detected in the first convolutional layers, while
in image recognition. It basically works as follows. A small patch, typ- global patterns are detected after more convolutional layers have been
ically of size 3 by 3 neurons, slides over the original image, operating applied. The last convolutional layer is in general flattened, i.e., all
on the current nine pixel intensities (𝐼𝑖 ) that fall inside this patch, neurons are depicted as a very long 1D vector instead of many (2D)
according to planes of neurons. Following some fully-connected layers of neurons,
each with different weight coefficients and activation functions, a final
𝜎(𝑤1 𝐼1 + 𝑤2 𝐼2 + ⋯ + 𝑤9 𝐼9 + 𝑏) (3)

3
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

Fig. 2. Optimized architecture of the CNN model. The number of neurons in each layer is shown in parenthesis (except for the dropout layer), where (a,b,c) = (height, width,
number of filters). A batch of 24 samples and the MSE loss function were used. Total number of trainable parameters = 432,301.

layer called output layer is used to predict the final property (FVC). using the optuna library (Akiba et al., 2019), which minimized the
Because each feature map has been generated using the same few average MAE evaluated via 5-fold cross validation using only the
weight coefficients, CNNs are not fully-connected neural networks and training set. The following hyperparameters were optimized: total
have therefore much less parameters than fully-connected NNs, which number of convolutional layers (1–5), number of feature maps in each
renders the former more computationally efficient towards complex convolutional layer (2–128), activation function (‘‘sigmoid’’, ‘‘relu’’,
tasks involving image recognition. ‘‘selu’’ and ‘‘tanh’’), batch size (4–128), size of the dropout layer (0–
Dataset and preprocessing. All microscopy images were used in gray scale 50%), patch kernel sizes (3 × 3 or 5 × 5) and the loss function used to
and had resolution of 256 × 192 pixels (width × height), besides being train the model (‘‘mae’’ or ‘‘mse’’). The ADAM optimizer was adopted
normalized to have their pixel intensities in the range of [0,1]. The here. Early stopping and a dropout layer were used in the model to
dataset, composed of 290 images, was split into training set (80% of try to minimize overfitting. The final optimized architecture of the
data or 232 images) and test set (58 images) using the train_test_split CNN model is shown in Fig. 2. After the hyperparameter optimization,
function of the sklearn package and the default random seed number a fresh CNN model was trained using the whole training set and
(of 42) shown in the documentation. The training set was augmented predictions were performed on the test set.
to 2552 images (10 augmented images per original image), where Comparison with other models. There are numerous state-of-the-art CNN-
each original image was either rotated (5–15 degrees), horizontally or
based models, such as those from the ResNet (He et al., 2016), VG-
vertically flipped, or horizontally or vertically translated (up to 60% of
GNet (Simonyan & Zisserman, 2015), and Inception Networks (Szegedy
the total width). During the rotations, the pixels leaving the area of the
et al., 2016) families, that have the potential to outperform the vanilla
image were re-inserted to preserve FVC. The features used to train the
CNN model explored in this study. However, the primary aim of this
ML models were the pixel intensities, while the manually determined
proof-of-concept investigation is not to exhaustively maximize accuracy
FVC (see Section 2.5) was used as the target property. Standardization
by employing these complex models through transfer learning. Given
was used to preprocess FVC.
that the accuracy achieved — especially in the ensemble model pre-
Metrics. The CNN model was evaluated via the mean absolute error sented in the results — already meets the requirements for polymer
(MAE, see Eq. (5)), mean absolute percentage error (MAPE, see Eq. (6)), engineering, we opted to compare the vanilla CNN model with simpler
mean squared error (MSE, see Eq. (7)), and the 𝑅2 score (see Eq. (8)), ML models that are not specialized in image recognition tasks. These
which are given by: simpler models are significantly faster, easier to optimize, and much
1∑
𝑛
more efficient to train, making them a viable alternative for predicting
𝑀 𝐴𝐸 = ∣ 𝑦 − 𝑦̂𝑖 ∣ (5)
𝑛 𝑖=1 𝑖 FVC in regions with limited computational resources. Additionally, the
efficiency in training these simpler models contributes to higher sus-
1 ∑ || 𝑦𝑖 − 𝑦̂𝑖 ||
𝑛
𝑀 𝐴𝑃 𝐸 = × 100 (6) tainability. The XGBoost (eXtreme Gradient Boosting Chen & Guestrin,
𝑛 𝑖=1 || 𝑦𝑖 ||
2016) model was then chosen for this model comparison, as it was
1∑
𝑛
𝑀𝑆𝐸 = (𝑦 − 𝑦̂𝑖 )2 (7) the only simple ML model showing a good prediction on the test
𝑛 𝑖=1 𝑖 set. XGBoost is a highly popular and powerful ML algorithm utilized
∑𝑛
(𝑦𝑖 − 𝑦̂𝑖 )2 for supervised learning tasks, particularly in predictive modeling and
̂ = 1 − ∑𝑖=1
𝑅2 (𝑦, 𝑦) 𝑛 (8) classification problems. It belongs to the ensemble learning family,
̄2
𝑖=1 (𝑦𝑖 − 𝑦)
which involves combining the predictions of multiple individual models
where 𝑛 is the number of samples, 𝑦𝑖 and 𝑦̂𝑖 are the true and predicted (here, decision trees), to create a stronger and more accurate final
target property for sample 𝑖, respectively, and 𝑦̄ is the average target
model. XGBoost employs gradient boosting to construct a sequence
property.
of decision trees, where each tree corrects the errors made by its
Model optimization. The CNN model was created using tensorflow predecessor using gradient descent optimization. This algorithm has
(Abadi et al., 2015) and Python 3. The hyperparameters were optimized L2 (or Ridge) regularization by default for help minimizing overfitting,

4
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

and parallel and distributed computing for speed and efficiency. In the
present work, the number of estimators (or trees), the learning rate,
the maximum depth of each tree, and the L1 and L2 regularization
parameters were optimized via a grid search using the training set and a
5-fold cross-validation approach. As XGBoost is not specialized in image
recognition, the pixels of the 2D images were first flattened to a long
1D vector of 49 152 components and every fifth component (or pixel) of
this vector was used for training and testing the model, as this is known
to increase the model efficiency (Albuquerque et al., 2024). After the
hyperparameter optimization, a fresh XGBoost model was trained using
the whole training set and predictions were performed on the test set.
Ensemble model. The predictions from the CNN and XGBoost models
were used to train an ensemble model, as this approach has been
successfully used to improve ML models in the past (Albuquerque
et al., 2022; Meier et al., 2022). The linear combination of the FVC
predictions from the CNN model (≡ 𝐹 𝑉 𝐶 𝐶 𝑁 𝑁 ) and XGBoost model
(≡ 𝐹 𝑉 𝐶 𝑋 𝐺𝐵 𝑜𝑜𝑠𝑡 ) extracted from Fig. 4ac, is given by:
𝐹 𝑉 𝐶 𝑒𝑛𝑠 = 𝑤1 ⋅ 𝐹 𝑉 𝐶 𝐶 𝑁 𝑁 + 𝑤2 ⋅ 𝐹 𝑉 𝐶 𝑋 𝐺𝐵 𝑜𝑜𝑠𝑡 (9)

where 𝑤1 and 𝑤2 are the coefficients of the linear combination and

𝐹 𝑉 𝐶 𝑒𝑛𝑠 is the final ensemble prediction. These coefficients are op-
timized by minimizing the error between 𝐹 𝑉 𝐶 𝑒𝑛𝑠 and the true FVC
(≡ 𝐹 𝑉 𝐶 𝑡𝑟𝑢𝑒 ) for the training set. The loss function defined by Eq. (10)
being minimized, 𝐿, is defined as:
1 ∑|
𝑁
|
𝐿= |𝐹 𝑉 𝐶 𝑡𝑟𝑢𝑒(𝑖) − 𝐹 𝑉 𝐶 𝑒𝑛𝑠(𝑖) | (10)
𝑁 𝑖=1 | |

where the index i refers to each of the N samples of the training set.

3. Results and discussion

Fig. 3 shows representative LM images, exhibiting various FVCs.

The flax fibers have a polygonal shape and are about 10 to 20 μm
in diameter. Through the irregular hand-layup of flax fibers and the
randomized locations in which images were taken, there is a broad
range of local FVCs in the composite. This is ideal for training a CNN
because the training set contains a wide variety of different input
values. Another consequence of the composite preparation and image
generation process is that the FVC shows an almost bell-shaped prob-
ability distribution function (see blue line in Fig. 3(b)). The minimum
and maximum FVCs are about 5.9% and 54.8%, respectively. The mean Fig. 3. (a) Representative LM images of the studied samples are depicted, showcasing
and median FVCs are 28.3% and 28.6%, respectively, with a standard various FVCs. The fibers, represented in dark gray, are oriented perpendicular to
the plane of observation, while the epoxy matrix is represented in light gray. (b)
deviation of 8.8%. The distribution of FVC covers a wide range of The distribution of FVC values is presented in a bar graph, with the kernel density
values that are relevant for composites for load-bearing applications. estimation illustrated as a blue line. These visualizations are derived from the analysis
The mean FVC is comparable to the FVC of other composites produced of 290 LM images employed in this study. (For interpretation of the references to
via hand-layup. The data set is representative of FVC that might be color in this figure legend, the reader is referred to the web version of this article.)
achieved in industrial applications and is suitable for the training of
ML models (Gandhi et al., 2020; Mallick, 2007).
The CNN has 420,801 trainable parameters, and uses five 2D convo- decreasing the model complexity, the XGBoost model still exhibited
lutional layers and three fully connected layers before the output layer some overfitting, as the training error shown in Fig. 4(c) is somehow
(the optimized hyperparameters are shown in Fig. 2). The number of smaller than the test error (Fig. 4(d)). Interestingly, the prediction
convolutional layers was also tuned by hyperparameter optimization performance on the test set is still acceptable, considering that XGBoost
(see Section 2). Fig. 4(a) shows the training error of the optimized is not specialized in image recognition, that only every fifth pixel
CNN model, where the model was first trained with the training set was used, and that many images were rotated or translated (via data
and then used to predict FVC on the same training set. A small MAE augmentation), which makes more difficult to use flattened pixels for
error was obtained, as well as a good 𝑅2 score. The same trained model predicting FVC. The simplest model generally used as reference (least
still performed well on the test set, showing little overfitting, as the squares model) was not able to predict FVC.
test error (2.89) was only a bit worse than the training error (1.76). The full performance evaluation for both ML models for the training
Fig. 4(b) shows that the CNN model generalizes well and can reliably and test sets are compared in Table 1. The comparison of the MSE
predict FVC using new LM images with a relatively small error. Note and MAPE errors for the training and test sets of both investigated ML
that the test set contains only original (non-augmented) images, which models also confirms the overfitting trend discussed above. Both the
were never seen before by the trained CNN model. XGBoost and CNN models have similar degree of overfitting, but the
The best hyperparameters found for the XGBoost model were: absolute errors are considerably smaller for the latter.
n_estimators = 100, learning_rate = 0.2, max_depth = 3, reg_alpha = The optimized ensemble model had 𝑤1 = 0.671 and 𝑤2 = 0.362
0.5, reg_lambda = 0.01. Even though regularization and a small tree as final parameters (see Eq. (9)) and showed a considerably better
depth were used, which are also supposed to decrease overfitting by performance on the test set, as revealed in Fig. 5. For instance, the 𝑅2

5
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

Fig. 4. Model evaluation using the training (2552 original + augmented images) and test (58 original images) sets employing the optimized CNN and XGBoost models for the
prediction of FVC. Every 5th image from the training set is plotted for clarity.

Fig. 5. Model evaluation using the training (2552 original + augmented images) and test (58 original images) sets employing the ensemble model (Eq. (9) with 𝑤1 = 0.671 and
𝑤2 = 0.362) for the prediction of FVC. Every 5th image from the training set is plotted for clarity.

score increased from 0.81 (CNN model) to 0.85 (ensemble model) for both the training and test sets, from where the superior performance of
the test set. The larger value of 𝑤1 means that predictions from the this model can be seen.
CNN model are considerably more important than that of the XGBoost An error analysis has been performed for the ensemble model
to calculate the final ensemble prediction. This happens because the predictions carried out on the training and test sets since this model
ensemble model (optimized with the training set) had much smaller showed the highest accuracy. Fig. 6(a) shows histograms (vertical bars)
MAE error for the CNN model compared to the XGBoost model. The full of the predicted FVC for both the training and test sets using the
metrics calculated for the ensemble model are also shown in Table 1 for ensemble model where each distribution has also its own kernel density

6
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

AutoML libraries such as aX (Balandat et al., 2023) is still another

way to do that. While this is certainly expected to improve even more
the accuracy, it brings no new insight to the present proof-of-concept
investigation and is out of the scope of this work. This is part of a
future investigation (vide infra). Having compared the CNN discussed
in this work with a much simpler ML model like XGBoost brings, on
the other side, much more insight, as it provides alternatives to easily
and efficiently train and optimize a ML model, therefore enhancing
sustainability, as well as helping democratize knowledge since regions
with low-budget computational resources could also have access to it.
To advance the discussion on the performance of ML models, it is
crucial to explore both the potential and limitations of determining
FVC in NFC using LM and CNN. Accurate analysis of FVC using LM-
based methods hinges on two key assumptions: first, the fibers must
be unidirectionally oriented. Second, the cross-sectional plane must
be perpendicular to the fiber axes. Additionally, the quality of sam-
ple preparation is paramount, requiring polished surfaces free from
scratches. Achieving this can be challenging due to the softness of
natural fibers relative to glass or carbon fibers. Furthermore, suffi-
cient contrast between the natural fibers and the polymeric matrix
is necessary to enable effective binarization of the LM images via
auto-thresholding.
Comparing ML models to existing techniques for determining the
FVC in NFC is essential for objectively evaluating the models’ practical
usability. However, direct comparison between different experimental
methods presents challenges. For instance, LM assesses the FVC of a
microscopic section (200 μm by 150 μm) of the NFC, whereas TGA and
density measurements require macroscopic samples for reliable FVC
determination. It is important to note that, unlike LM, NFC samples
analyzed by TGA or density measurements do not require unidirec-
Fig. 6. (a) Histograms of ensemble predictions on the training and test sets (bars), tional fiber orientation. However, applying these common methods
each with its own Kernel Density Estimate (KDE, lines). (b) Histograms of the errors to NFC remains complex due to the low thermal stability and low
for the predictions shown in (a), each with its own KDE (lines). The inset shows the
error scattering in each case.
atomic density of natural fibers. The primary aim of this study is to
demonstrate how image analysis of NFC via LM can be optimized with
CNNs. Here, it is shown that automated image processing software,
Table 1
Model performance metrics for the investigated ML models trained with the training such as ImageJ, significantly reduces the time needed for FVC analysis
set with predictions done on the training (tr) and test (te) sets. when using macros compared to manual methods. Additionally, once
Model MAE 𝑅2 MAPE/% MSE CNNs and ensemble models are properly trained, they further expedite
XGBoost 𝑡𝑟
2.54 0.87 9.69 10.04 FVC assessment in LM images while maintaining robust predictive
XGBoost 𝑡𝑒 3.23 0.78 11.00 17.78 performance.
CNN 𝑡𝑟 1.76 0.93 7.00 5.10
CNN 𝑡𝑒 2.89 0.81 9.63 15.14
Ensemble 𝑡𝑟 1.56 0.95 6.32 3.85
4. Conclusion
Ensemble 𝑡𝑒 2.72 0.85 9.54 12.22
It has been proved that the pixels from light microscopy images of
NFC could be reliably used to predict FVC using different ML models.
The good accuracy achieved for both models and specially for their
estimate (continuous lines). Fig. 6(a) shows that the distributions of
ensemble already fulfills the minimum accuracy requirements found in
training and test predictions nicely overlap and also have a similar
polymer engineering. The exhaustive maximization of the accuracy by
shape to a Gaussian distribution with two peaks, which arises from the
using more complex deep learning models and transfer learning is out
use of two ML models in the ensemble model. Fig. 6(b) shows that the
of the scope of this proof-of-concept investigation.
errors of the predictions are rather unstructured (see the inset), being
The effectiveness of CNNs and XGBoost models in predicting the
approximately described by a Gaussian distribution with zero mean. As
FVC of NFC from LM images was shown. The distribution of FVCs,
expected, the Gaussian is broader for the test errors, as these predictions
were performed on unseen data and are therefore supposed to exhibit captured through a comprehensive set of LM images, exhibited a wide
larger errors. range, with values pertinent to industrial applications of load-bearing
The final accuracy found for the ensemble model (Table 1, last composites.
line) is already good enough for FVC characterizations required in The CNN model, equipped with 432,301 trainable parameters and
polymer engineering, where acceptable MAE errors lie in the range a series of convolutional and fully connected layers, showed good
of about 3.00–5.00. There are several ways of further improving the performance on both the training and test datasets. It achieved low
accuracy of the model. Re-parameterizing the CNN using Bayesian errors and a high 𝑅2 value on the training set, with little overfitting
optimization would be one a first way. Adding other ML models to the observed when evaluated against the test set. This suggests that the
ensemble approach would be a second way. Another way is to employ CNN model has a strong ability to generalize from the training data
transfer learning to use one (or a combination of) the state-of-the-art to unseen data, making it a reliable tool for predicting FVC in natural
deep learning models such as, for instance, those from the ResNet (He fiber composites based on microscopy images.
et al., 2016), VGGNet (Simonyan & Zisserman, 2015), and Inception The XGBoost model displayed a similar degree of overfitting. It is
Networks (Szegedy et al., 2016) families, and fine tune these models. noteworthy that the XGBoost model still delivered a good prediction

7
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

accuracy on the test set, which was acceptably high for FVC charac- References
terizations required in polymer engineering. This outcome underscores
the versatility of XGBoost in handling diverse datasets and learning Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S.,
Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G.,
tasks, even those for which it is not inherently specialized, like image Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., .... Zheng, X. (2015).
recognition. Given that this model is easier, faster, and more efficient TensorFlow: Large-scale machine learning on heterogeneous systems. Software
to optimize and train compared to CNN-based models, it presents an available from tensorflow.org.
interesting alternative for predicting FVC. Not only is it more sustain- Abdalla, F., Megat Ahmad, M. M. H., Sapuan, S., & Sahari, B. (2008). Determination
of volume fraction values of filament wound glass and carbon fiber reinforced
able, but it also helps to democratize knowledge in regions with limited composites. ARPN Journal of Engineering and Applied Sciences, 3, 7–11.
computational resources. Affdl, J. H., & Kardos, J. (1976). The Halpin-Tsai equations: a review. Polymer
Ensembling the predictions from the CNN and XGBoost models Engineering and Science, 16(5), 344–352.
Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019). Optuna: A next-
enabled one to achieve even higher predictive performances for both
generation hyperparameter optimization framework. In Proceedings of the 25th ACM
the training and test sets. SIGKDD international conference on knowledge discovery and data mining.
The findings advocate for the integration of ML models into the Albuquerque, R. Q., Brütting, C., Standau, T., & Ruckdäschel, H. (2022). A machine
analysis of NFC. The demonstrated capability of these models to ac- learning investigation of low-density polylactide batch foams. e-Polymers, 22,
318–331.
curately predict FVC from LM images can significantly streamline the
Albuquerque, R. Q., Sarhadi, A., Demleitner, M., Ruckdäschel, H., & Eder, M. A. (2024).
material characterization process, providing a rapid and efficient means Fatigue damage reconstruction in glass/epoxy composites via thermal analysis and
of assessing composite quality and performance. Future work may focus machine learning: A theoretical study. Composite Structures, 331, Article 117855.
on maximizing the accuracy using complex deep learning models and Andrade, R. M. d., & Eduardo, A. C. (2011). Methodology for automatic process of the
fired ceramic tile’s internal defect using IR images and artificial neural network.
transfer learning, exploring the incorporation of additional image fea-
Journal of the Brazilian Society of Mechanical Sciences and Engineering, 33, 67–73.
tures, and extending the methodology to other types of composites and Ashby, M. (1992). Materials selection in mechanical design (1st ed.). Oxford:
material characterizations. This research paves the way for advanced Butterworth-Heinemann.
material design and quality control through the application of ML Balandat, M., Bakshy, E., Marthur, L., Scott, S., Zhang, A., Karrer, B., Blei, D., &
Polley, E. (2023). Ax: A python library for adaptive experimentation. Version 0.2.5.
technologies. Blarr, J., Kunze, P., Kresin, N., Liebig, W. V., Inal, K., & Weidenmann, K. A. (2024).
Novel thresholding method and convolutional neural network for fiber volume
content determination from 3D 𝜇CT images. NDT & E International, 144, Article
CRediT authorship contribution statement
103067.
Bledzki, A., & Gassan, J. (1997). Natural fibre reinforced plastics. In Handbook of
Florian Rothenhäusler: Conceptualization, Investigation, Method- engineering polymeric materials. CRC Press.
Bos, H. L. (2004). The potential of flax fibres as reinforcement for composite materials
ology, Writing – original draft. Rodrigo Queiroz Albuquerque: Con-
(Ph.D. thesis), University of Eindhoven, Netherlands.
ceptualization, Machine learning modeling, Methodology, Writing – Chauhan, R., Ghanshala, K. K., & Joshi, R. (2018). Convolutional neural network (CNN)
original draft. Marcel Sticher: Investigation. Christopher Kuenneth: for image detection and recognition. In 2018 first international conference on secure
Writing – review & editing. Holger Ruckdaeschel: Writing – review & cyber computing and communication (pp. 278–282).
Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. CoRR
editing.
abs/1603.02754.
Chin, W.-K., Liu, H.-T., & Lee, Y.-D. (1988). Effects of fiber length and orientation
Declaration of competing interest distribution on the elastic modulus of short fiber reinforced thermoplastics. Polymer
Composites, 9(1), 27–35.
Das, S. (2011). Life cycle assessment of carbon fiber-reinforced polymer composites.
The authors declare that they have no known competing finan- The International Journal of Life Cycle Assessment, 16, 268–282.
cial interests or personal relationships that could have appeared to DIN EN ISO 1172:1998-12 (1998). Textilglasverstärkte kunststoffe - prepregs, formmassen
und laminate - bestimmung des textilglas- und mineralfüllstoffgehalts. Beuth Verlag.
influence the work reported in this paper. Ehrenstein, G. W. (2006). Faserverbund-kunststoffe. Munich: Carl Hanser Verlag GmbH
& Co. KG.
Flemming, M., Ziegmann, G., & Roth, S. (1995). Faserverbundbauweisen - fasern und
Acknowledgments
matrices. Berlin: Springer.
Fu, S.-Y., & Lauke, B. (1996). Effects of fiber length and fiber orientation distributions
Parts of the research documented in this manuscript have been on the tensile strength of short-fiber-reinforced polymers. Composites Science and
Technology, 56(10), 1179–1190.
funded by the German Federal Ministry for Economic Affairs and Cli-
Fu, S.-Y., & Lauke, B. (1998). The elastic modulus of misaligned short-fiber-reinforced
mate Action (BMWK) within the research project ‘‘EcoPrepregs - Grund- polymers. Composites Science and Technology, 58(3–4), 389–400.
lagenforschung zur Klärung der Struktur-Eigenschaftsbeziehungen von Gandhi, U., Goris, S., Osswald, T. A., & Song, Y.-Y. (2020). Discontinuous fiber-reinforced
Epoxidharzen und Fasern aus nachwachsenden Rohstoffen zur composites: fundamentals and applications. Carl Hanser Verlag GmbH Co KG.
Grund, D., Orlishausen, M., & Taha, I. (2019). Determination of fiber volume fraction of
Anwendung in der Sekundärstruktur von Flugzeugen’’ (grant #
carbon fiber-reinforced polymer using thermogravimetric methods. Polymer Testing,
20E1907A, Germany). Parts of this work were funded by the 75.
‘‘Bayerisches Staatsministerium für Wissenschaft und Kunst’’ (grant Gupta, V., Mittal, R., Sharma, P., Mennig, G., & Wolters, J. (1989). Some studies
# F.2-M7426.10.2. 1/4/16, Germany). The authors want to thank on glass fiber-reinforced polypropylene. Part II: Mechanical properties and their
dependence on fiber length, interfacial adhesion, and fiber dispersion. Polymer
Lukas Endner and Adwaith Sudhakaran for their support during the
Composites, 10(1), 16–27.
experiments. We would like to thank all colleagues of the work group He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image
‘‘Resins & Composites’’ at the Department of Polymer Engineering for recognition. In Proceedings of the IEEE conference on computer vision and pattern
their support. recognition (pp. 770–778).
Hessman, P. A., Riedel, T., Welschinger, F., Hornberger, K., & Böhlke, T. (2019).
Microstructural analysis of short glass fiber reinforced thermoplastics based on x-
Data availability ray micro-computed tomography. Composites Science and Technology, 183, Article
107752.
ImageJ (2023). Getting started with simple visualization options in ImageJ. (Accessed
The trained CNN and XGBoost models, the dataset and a demonstra- 06 March 2024).
tor code (Python, Jupyter Notebook) are here: ‘‘https://github.com/P Jock, C. P. (1986). Quantitative optical microscopy fiber volume methods for
composites. Journal of Reinforced Plastics and Composites, 5(2), 110–119.
olymer-Engineering-University-Bayreuth/FiberVolumeContent_ConvNe
Kumar, S., Prasad, L., Bijlwan, P. P., & Yadav, A. (2022). Thermogravimetric analysis
t’’, which is freely accessible in the public GitHub repository. of lignocellulosic leaf-based fiber-reinforced thermosets polymer composites: an
overview. Biomass Conversion and Biorefinery.

8
F. Rothenhäusler et al. Machine Learning with Applications 19 (2025) 100609

Machello, C., Baghaei, K. A., Bazli, M., Hadigheh, A., Rajabipour, A., Arashpour, M., Schürmann, H. (2007). Konstruieren mit faser-kunststoff-verbunden. Berlin: Springer.
Rad, H. M., & Hassanli, R. (2024). Tree-based machine learning approach to mod- Shokrollahi, Y., Nikahd, M. M., Gholami, K., & Azamirad, G. (2023). Deep learning
elling tensile strength retention of Fibre Reinforced Polymer composites exposed techniques for predicting stress fields in composite materials: A superior alternative
to elevated temperatures. Composites Part B (Engineering), 270, Article 111132. to finite element analysis. Journal of Composites Science, 7(8).
Mallick, P. K. (2007). Fiber-reinforced composites: materials, manufacturing, and design. Simonyan, K., & Zisserman, A. (2015). Very deep convolutional networks for large-scale
CRC Press. image recognition. arXiv preprint arXiv:1409.1556.
Meier, S., Albuquerque, R. Q., Demleitner, M., & Ruckdäschel, H. (2022). Modeling Song, L., Li, X., Yang, Y., Zhu, X., Guo, Q., & Yang, H. (2018). Detection of micro-defects
glass transition temperatures of epoxy systems: a machine learning study. Journal on metal screw surfaces based on deep convolutional neural networks. Sensors,
of Materials Science, 57(29), 13991–14002. 18(11), 3709.
Mwaikambo, L. Y., & Ansell, M. P. (2001). The determination of porosity and cellulose Sorour, S. S., Saleh, C. A., & Shazly, M. (2024). A review on machine learning im-
content of plant fibers by density methods. Journal of Materials Science Letters, plementation for predicting and optimizing the mechanical behaviour of laminated
20(23), 2095–2096. fiber-reinforced polymer composites. Heliyon.
Osa-Uwagboe, N., Udu, A. G., Ghalati, M. K., Silberschmidt, V. V., Aremu, A., Dong, H., Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the
& Demirci, E. (2024). A machine learning-enabled prediction of damage properties inception architecture for computer vision. In Proceedings of the IEEE conference on
for fiber-reinforced polymer composites under out-of-plane loading. Engineering computer vision and pattern recognition (pp. 2818–2826).
Structures, 308, Article 117970. Voigt, W. (1889). Ueber die Beziehung zwischen den beiden Elasticitätsconstanten
Schlicht, S., Jaksch, A., & Drummer, D. (2022). Inline quality control through optical isotroper Körper. Annalen der Physik, 274(12), 573–587.
deep learning-based porosity determination for powder bed fusion of polymers.
Polymers, 14(5).