ISSN 1648-7788 / eISSN 1822-4180

2024

AVIATION
Volume 28
Issue 2
Pages 64–71
https://doi.org/10.3846/aviation.2024.21472

MACHINE LEARNING METHODS AS APPLIED TO MODELLING THERMAL

CONDUCTIVITY OF EPOXY-BASED COMPOSITES WITH DIFFERENT FILLERS
FOR AIRCRAFT

Oleh YASNIY , Mykola MYTNYK , Pavlo MARUSCHAK , Andriy MYKYTYSHYN ,

Iryna DIDYCH
Ternopil Ivan Puluj National Technical University, Ruska str. 56, 46001 Ternopil, Ukraine

Article History: Abstract. The thermal conductivity coefficient of epoxy composites for aircraft, which are reinforced with
■ received 19 November 2023 glass fiber and filled with aerosil, γ-aminopropylaerosil, aluminum oxide, chromium oxide, respectively, was
■ accepted 24 January 2024 simulated. To this end, various machine learning methods were used, in particular, neural networks and boost-
ed trees. The results obtained were found to be in good agreement with the experimental data. In particular,
the correlation coefficient in the test sample was 0.99%. The prediction error of neural networks in the test
samples was 0.5; 0.3; 0.2%, while that of boosted trees was 1.5; 0.9%.

Keywords: epoxy-based composites, fillers, modelling, aircraft, aerospace applications, machine learning.

Corresponding author. E-mail: maruschak.tu.edu@gmail.com

1. Introduction the authors (Bezerra et al., 2007) in predicting the stress-

strain diagrams that describe the shear of such composites
In the aircraft design, the fuselage, wings, tail unit, engine as carbon fiber/epoxy resin and glass fiber/epoxy resin.
nacelle, and interior parts can be made from composite While the bending properties of 3D-printed carbon/ep-
materials. It is urgent to develop new materials, innova- oxy composites with different processing parameters were
tive solutions for technology and production equipment, predicted by the authors (Monticeli et al., 2022) using an
allowing for high strength and aerodynamic characteristics artificial neural network and statistical methods. In Stephen
with economic efficiency in aircraft production conditions. et al. (2022), the impact characteristics of fiber-reinforced
An increasingly growing automation of technology polymer composites are modeled using the finite element
along with numerous data sets account for the need in method and an artificial neural network. In particular, a
machine learning methods, which, instead of being pro- multi-scale approach based on the stochastic integrated
grammed, are capable of analyzing data and learning from machine training was used in Liu et al. (2022) in order to
them (Konovalenko et al., 2021). In the presence of ex- predict the thermal conductivity of polymer composites
perimental data, machine learning algorithms are known reinforced with carbon nanotubes. Therefore, different ma-
to solve the problems of fracture mechanics with great chine training methods appear relevant when it comes to
accuracy. In particular, they are capable of predicting the predicting the thermal conductivity coefficient of polymer
life of a fatigue crack that occurred in an aluminum alloy composites.
using a neural network (Mohanty et al., 2009) and fatigue In general, polymer composite materials are often
fracture diagrams of aluminum alloy D16T subjected to used in the production of load-bearing structures, such as
the regular loading conditions with the load cycle asym- equipment for gas mains, radio engineering products, and
metry R = 0; 0.2; 0.4; 0.6 (Yasnii et al., 2018). The authors aircraft structures because of their superior physical-me-
(Zhang & Wei, 2022) predicted the propagation of a fa- chanical and thermo-physical properties (Stukhlyak et al.,
tigue crack under conditions of variable amplitude loading 2015; Dobrotvor et al., 2021). The issues related to the
by means of Lagrange interpolation based on an artifi- preferable filler type, the degree of filling and dispersion,
cial neural network. In addition, the stress-strain diagrams and the polymerization conditions of the filled composi-
of aluminum alloy AMg6 (Yasniy et al., 2020; 2022a) and tion are considered while creating composite materials. In
AL-6061 (Yasniy et al., 2022b) were simulated using ma- particular, composites with improved technological and
chine learning methods. Neural networks have also aided service characteristics ensure a strong and stable bond

Copyright © 2024 The Author(s). Published by Vilnius Gediminas Technical University

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
 Aviation, 2024, 28(2), 64–71 65

between the surfaces of the filler and the polymer. The ary. In the other case, it leads to a decrease in l due to
thermal expansion of polymers is characterized by cer- a higher defectiveness of the boundary layer. In addition,
tain features in the range of glass transition temperature. thermal resistance also increases in response to the elevat-
These features are known to depend on the temperature ed temperature caused by failure of physical nodes. At the
variation rate. Therefore, investigations into the thermo- same time, the thermal conductivity of the filler accounts
physical properties that affect the service characteristics for an increase in the filler content with an increase in the
remain topical. thermal conductivity of the matrix.
This research aims at predicting the thermal conduc- In general, the thermophysical properties of epoxy
tivity coefficient of the epoxy composites reinforced with composite materials depend both on the nature and
glass fiber and filled with aerosil, γ-aminopropylaerosil, amount of the filler and the specifics of the particle sur-
aluminum oxide, chromium oxide, respectively, during their face. In addition, these characteristics can be improved by
heating. For this purpose, various machine learning meth- introducing the fillers that are chemically active in relation
ods are used, in particular, neural networks and boosted to the matrix. These fillers form a strong bond between the
decision trees. The results obtained are compared. binder macromolecules and the filler surface, which occurs
as a result of the chemical and chemisorption interaction.

2. Materials and methods

2.2. Methods of machine learning
2.1. Thermophysical research of filled epoxy The most common model of a neural network is a multi-
composites layer perceptron, in which neurons are arranged in layers
Epoxy composites with enhanced service properties can be from input to output (Haykin, 2006). The structure of a
obtained subject to optimizing their production technol- three-layer neural network is shown in Figure 1. In particu-
ogy and the process of filling the matrix with the fillers of lar, the concentration of the mass fraction C in the filler
different chemical nature, composition and size, which are and the temperature T serve as input parameters, while
characterized by the improved mechanical and thermo- the thermal conductivity coefficient λ is the output pa-
physical properties. A small amount of fillers introduced rameter. Accordingly, wij and wjk are the synaptic weights
into the epoxy matrix, the former being chemically active between different layers of the network, and (H1, H2, ..., Hn)
in relation to the matrix (1 wt % of aerosil and 1 wt % of are the thresholds of the hidden layer.
g-aminopropylaerosil per 100 wt %s), entails a significant The input signal passes to the neurons of the hidden
increase in l. This is associated with a strong bond formed layer, while the output signals of the hidden layer of the
by the chemical and chemisorption interaction between neural network are the inputs of the next one. Neurons of
the binder macromolecules and the aerosil surface. At the each layer are the output signals of only the previous one.
same time, a large amount of aerosil introduced into the To attain the best neural network architecture, a num-
epoxy matrix causes a decreased thermal conductivity of ber of numerical experiments were conducted using the
the material, which is due to an increased length of the “input-output” pairs and different activation functions,
separation boundary and a weakened interaction of OH such as tangential, logarithmic ones and others, thus vary-
groups between the filler and the binder. ing the number of neurons in the hidden layer. In the pro-
A low concentration of Al2O3 (up to 30 wt %) introduced cess of learning, the network uses error backpropagation
into the binder causes little change in the thermal con- algorithms, that is, it adjusts its weights to minimize the
ductivity coefficient of the epoxy composite, as compared error between the predicted and experimental data.
to the thermal conductivity of the unfilled binder. Epoxy The network operation algorithm is based on Haykin
composites with the Al2O3 content in the matrix exceeding (2006) the following Equations:
30 wt % are characterized by an increased thermal conduc-
tivity compared to the unfilled binder. This is because the NET jl
= ∑w mjl × x mjl ; (1)
thermal conductivity of the filler itself affects the thermal m

conductivity of the material as a whole. In such composites,

there is no bond between the matrix and the filler.
The mechanisms characteristic of both chemically ac-
tive (g-aminopropylaerosil and aerosil) and inactive (alumi-
num oxide) fillers are also inherent in the epoxy composite
filled with chromium oxide. Therefore, for this filler, the
thermal conductivity coefficient is determined by the com-
peting contribution made by the mechanisms that account
for its change as described above. In one case, physical
nodes formed between the matrix and the active centers
on the surface of the filler particles lead to an increase
in l. This normally occurs at low temperatures due to the Figure 1. A three-layer neural network architecture with one
additional hydrogen bonds formed at the phase bound- hidden layer and one output layer
 66 O. Yasniy et al. Machine learning methods as applied to modelling thermal conductivity of epoxy-based composites with different...

OUT jl F (NET jl − q jl ) ;
= (2) 3. Results and discussion
x mj (l +1) = OUTml , (3) 3.1. Composition of the filled epoxy polymers
reinforced with basalt and glass fiber
where i is the number of entry into the layer; j is the neu-
ron number in the layer; l is the layer number; xijl is the i-th The introduction of fillers into epoxy composite materials
input signal of the j-th neuron in layer l; wijl is the weight in known to increase the heat resistance of the compos-
coefficient of the i-th input of the j-th neuron of layer l; ites compared to that of the polymer binder. Accordingly,
NETjl is the NET signal of the j-th neuron of layer l; Outjl the heat resistance of epoxy composites depends both on
is the output signal; F is the nonlinear activation function; the filler’s nature and the specifics of the particle surface.
qjl is the threshold level of a particular neuron. In particular, the presence of active (g-aminopropylaerosil
The choice of the neural network architecture, learning and aerosil) and inert (Al2O3 and Cr2O3) fillers allows in-
algorithm, error function, and activation function of the creasing the thermal stability limits of epoxy composites.
hidden and output layers is important when building a Therefore, the introduction of fillers that are chemically
network. In addition, the stopping parameter of the neural active in relation to the matrix provides for a more signifi-
network learning was the number of epochs, which in this cant effect, which manifests itself in increasing the tem-
research was equal to 1000. perature of thermal fracture of the filled polymer materials
The algorithm of boosted trees is a powerful method compared to Al2O3 and Cr2O3. In the formation of epoxy
of the intelligent data analysis (Alpayndin, 2010). It is com- CMs, the ED-20 epoxy dian oligomer (GOST 10587-84) was
monly used to predict data and in situations when the chosen as the main binding component characterized by
results of one decision affect those of the subsequent one, high adhesive and cohesive strength, low shrinkage and
that is, in order to make sequential decisions. The model high manufacturability during application.
structure is tree-like (Figure 2). In general, a boosted tree Polyethylene polyamine PEPA hardener (TU 6-05-241-
consists of internal nodes and leaves that are added dur- 202-78) was used for crosslinking epoxy compositions, mak-
ing learning depending on the problem complexity. In par- ing it possible for the materials to harden at room tempera-
ticular, ovals are the decision nodes, while rectangles are tures. PEPA is known (Dobrotvor et al., 2021) to be a low
the leaf nodes. The univariate decision node splits along molecular weight substance with the following chemical for-
one axis, and successive splits are orthogonal to each mula: [-CH2-CH2-NH-]n. CMs were cross-linked by introduc-
other. After the first split, {x | x1 < w10} is pure and is not ing a hardener into the composition at a stoichiometric ratio
split further. of components by content (wt %) – ED-20: PEPA – 100: 10.
The major advantage of the boosted trees is versatility The scheme of the epoxy composite is given in Table 1.
and ease of data interpretation. Table 1. Composition of the filled epoxy polymers
Using the Mean Absolute Percent Error (MAPE) for- reinforced with basalt and glass fiber
mula, the prediction error is calculated as follows:
Filler Concentration, wt % Reinforcing fibrous filler
yi − yi Aerosil 2 glass fiber
1 n ( true ) ( prediction )
MAPE
= 100% ⋅
n ∑ i =1
yi
, (4) basalt fiber
(true ) 6 glass fiber
basalt fiber
where yi(prediction) is the predicted element of the sample;
12 glass fiber
yi(true) is the real value of the element of the sample; n is
basalt fiber
the size of the study sample.
g-aminopropy- 2 glass fiber
laerosil basalt fiber
6 glass fiber
basalt fiber
12 glass fiber
basalt fiber
Cr2O3 30 glass fiber
basalt fiber
50 glass fiber
basalt fiber
100 glass fiber
basalt fiber
Al2O3 30 glass fiber
basalt fiber
50 glass fiber
Figure 2. Example of a data set and the corresponding basalt fiber
decision tree (Alpayndin, 2010) 100 glass fiber
basalt fiber
 Aviation, 2024, 28(2), 64–71 67

3.2. Thermal conductivity coefficients of filled To model the dependence of the thermal conductivity
epoxy polymers coefficient on the mass fraction concentration of the filler
and temperature, the thermal conductivity coefficient λ
To model the dependence of the thermal conductivity co-
served as an input parameter, while the mass fraction con-
efficient on the mass fraction concentration of the filler
centration of the filler and temperature were considered
and temperature, the mass fraction concentration of the
as output parameters.
filler and temperature served as input parameters, while
Machine learning methods were used to construct the
the thermal conductivity coefficient λ was considered as
dependences of the experimental data relating to the ther-
an output parameter.
mal conductivity coefficient on the predicted data (Fig-
The thermal conductivity coefficient of epoxy poly-
mers reinforced with glass fiber and filled with aerosil, ures 3–6). The prediction results were found to be in good
γ-aminopropylaerosil, aluminum oxide, chromium oxide, agreement with the experimental ones.
respectively, was predicted on the basis of the experimen- It is noteworthy that points are located close to the
tal data obtained in Mykytyshyn (2002) by various meth- bisector of the first coordinate angle in Figures 3–6, in-
ods of machine learning, in particular, neural networks and dicating a linear relationship between the predicted
boosted decision trees. and experimental data. The NN method gives an er-
In the process of learning, the data set was divided ror of 0.5% for epoxy composites filled with aerosil and
into two unequal parts – the training sample and the test γ-aminopropylaerosil, 0.3% and 0.2% for epoxy composites
sample. The sample contained 4,000 elements for each filled with aluminum oxide and chromium oxide, respec-
epoxy polymer reinforced with fiberglass and filled with tively. At the same time, the error of the boosted deci-
aerosil and γ-aminopropylaerosil, respectively, and 28,000 sion tree method is 1.5% for epoxy composites filled with
elements for the polymer filled with aluminum oxide and aerosil, γ-aminopropylaerosil, and aluminum oxide, respec-
chromium oxide, respectively, of which 80% were ran- tively, and 0.9% for epoxy composite filled with chromium
domly selected for the learning sample, and 20% were left oxide. Figures 7–10 show the dependence of the predicted
to assess the prediction quality. In particular, to achieve thermal conductivity coefficient on the mass fraction con-
this number of elements, the dataset was expanded using centration of the filler and temperature of the test sample
spline interpolation. obtained by various methods of machine learning.

(a) (b)

Figure 3. Predicted and experimental dependences of the thermal conductivity coefficient for an epoxy composite filled with
aerosil obtained by the method of neural networks (a) and boosted decision trees (b)

(a) (b)

Figure 4. Predicted and experimental dependences of the thermal conductivity coefficient for an epoxy composite filled with
γ-aminopropylaerosil obtained by the method of neural networks (a) and boosted decision trees (b)
 68 O. Yasniy et al. Machine learning methods as applied to modelling thermal conductivity of epoxy-based composites with different...

(a) (b)

Figure 5. Predicted and experimental dependences of the thermal conductivity coefficient for an epoxy composite filled with
aluminum oxide obtained by the method of neural networks (a) and boosted decision trees (b)

(a) (b)

Figure 6. Predicted and experimental dependences of the thermal conductivity coefficient for an epoxy composite filled with
chromium oxide obtained by the method of neural networks (a) and boosted decision trees (b)

(a) (b)

Figure 7. Temperature dependence of changes in the thermal conductivity coefficient for the composite filled with aerosil
obtained by the method of neural networks (a) and boosted decision trees (b)

(a) (b)

Figure 8. Temperature dependence of changes in the thermal conductivity coefficient for the composite filled with
γ-aminopropylaerosil obtained by the method of neural networks (a) and boosted decision trees (b)
 Aviation, 2024, 28(2), 64–71 69

(a) (b)

Figure 9. Temperature dependence of changes in the thermal conductivity coefficient for the composite filled with aluminum
oxide obtained by the method of neural networks (a) and boosted decision trees (b)

(a) (b)

Figure 10. Temperature dependence of changes in the thermal conductivity coefficient for the epoxy composite filled with
chromium oxide obtained by the method of neural networks (a) and boosted decision trees (b)

The main parameters of machine learning algorithms, They were found to be concentrated around zero and have
in particular, neural networks and boosted decision trees, a normal distribution.
are given in Table 2–3. The statistical graph used for the Figures 11–14 show frequency for epoxy compos-
data analysis presents as a histogram of residual values, ites reinforced by glass fiber and filled with Aerosil,
which shows the frequency of each interval of values ver- γ-aminopropylaerosil, aluminum oxide, chromium oxide,
sus residual values. In particular, the residuals show the respectively, using different machine learning methods,
difference between the experimental and predicted values. in particular, neural networks and boosted decision trees.

Table 2. Parameters of neural networks

Algorithm of Function of hidden Function of output

Filler Name of network Error function
learning activation activation
Aerosil MLP 2-8-1 BFGS SOS Logarithmic Logarithmic
γ-aminopro-pylaerosil MLP 2-8-1 BFGS SOS Logarithmic Tangential
Al2O3; Cr2O3 MLP 2-7-1 BFGS SOS Tangential Tangential

Table 3. Parameters of reinforced trees

Filler Number of trees

Aerosil, γ-aminopropylaerosil, Al2O3; Cr2O3 1000
 70 O. Yasniy et al. Machine learning methods as applied to modelling thermal conductivity of epoxy-based composites with different...

(a) (b)

Figure 11. Frequency in the test set for the epoxy composite filled with aerosil obtained by the method of neural networks (a)
and boosted decision trees (b)

(a) (b)

Figure 12. Frequency in the test set for the epoxy composite filled with γ-aminopropylaerosil obtained by the method of
neural networks (a) and boosted decision trees (b)

(a) (b)

Figure 13. Frequency in the test set for the epoxy composite filled with aluminum oxide obtained by the method of neural
networks (a) and boosted decision trees (b)

(a) (b)

Figure 14. Frequency in the test set for the epoxy composite filled with chromium oxide obtained by the method of neural
networks (a) and boosted decision trees (b)
 Aviation, 2024, 28(2), 64–71 71

4. Conclusions Dobrotvor, I. G., Stukhlyak, P. D., Mykytyshyn, A. G., & Stukhlyak, D. P.

(2021). Influence of thickness and dispersed impurities on residual
Once predicted, thermal conductivity can provide im- stresses in epoxy composite coatings. Strength of Materials, 53,
portant information about the thermal properties of a 283–290. https://doi.org/10.1007/s11223-021-00287-x
material. It also aids a lot in understanding thermal per- Haykin, S. (2006). Neural networks – a comprehensive foundation.
Prentice Hall.
formance of the material, making it easier to investigate
Konovalenko, I., Maruschak, P., Brevus, V., & Prentkovskis, O.
and develop new composites for various applications. In (2021). Recognition of scratches and abrasions on metal sur-
particular, knowing thermal conductivity is crucial in the faces using a classifier based on a convolutional neural net-
design of structures and materials, especially composites work. Metals, 11(4), Article 549.
with different fillers. Materials with improved thermal per- https://doi.org/10.3390/met11040549
formance can be created for such industries as civil engi- Liu, B., Vu-Bac, N., Zhuang, X., Fu, X., & Rabczuk, T. (2022). Sto-
neering, aviation, electronics, and so on. chastic integrated machine learning based multiscale approach
Machine learning methods, such as neural networks for the prediction of the thermal conductivity in carbon nano-
tube reinforced polymeric composites. Composites Science and
and boosted trees, were used to simulate the thermal
Technology, 224, Article 109425.
conductivity coefficient of epoxy polymers reinforced by
https://doi.org/10.1016/j.compscitech.2022.109425
glass fiber and filled with aerosil, γ-aminopropyl aerosil, Mohanty, J. R., Verma, B. B., Parhi, D. R. K., & Ray, P. K. (2009). Ap-
aluminum oxide, chromium oxide, respectively. Neural plication of artificial neural network for predicting fatigue crack
networks had a prediction error of 0.5%, 0.3%, and 0.2% propagation life of aluminum alloys. Archives of Computational
when evaluating different test samples. At the same time, Materials Science and Surface Engineering, 1(3), 133–138.
boosted trees had a slightly higher prediction error, which Monticeli, F. M., Neves, R. M., Ornaghi, H. L., & Almeida, J. H. S.
was 1.5% and 0.9%, respectively. Despite these differences, (2022). Prediction of bending properties for 3D-printed carbon
fibre/epoxy composites with several processing parameters
both machine learning methods proved to be efficient at
using ANN and statistical methods. Polymers, 14(17), Article
predicting thermal conductivity coefficients. This makes
3668. https://doi.org/10.3390/polym14173668
them very promising in materials science. Very low predic- Mykytyshyn, A. G. (2002). Development of technology and study
tion errors further emphasize the accuracy and reliability of parameters of formation of products from epoxy-filled com-
of these models in the analysis of the complex relation- posites, [Dissertation, Lviv Polytechnic National University]. Lviv
ships between the composite materials studied. (in Ukrainian).
Stephen, C., Thekkuden, D. T., Mourad, A. H. I., Shivamurthy, B.,
Selvam, R., & Rohit Behara, S. (2022). Prediction of impact per-
Author contributions formance of fiber reinforced polymer composites using finite
element analysis and artificial neural network. Journal of the
Conceptualization, O. Y. and I. D.; methodology, O. Y. and Brazilian Society of Mechanical Sciences and Engineering, 44,
I. D.; validation, P. M. and A. M.; formal analysis, M. M.; Article 408. https://doi.org/10.1007/s40430-022-03711-8
investigation, I. D., O. Y., P. M., A. M., M. M.; data cura- Stukhlyak, P. D., Buketov, A.V., Panin, S. V., Maruschak, P. O., Mo-
tion, I. D.; writing – original draft preparation, I. D., O. Y., roz, K. M., Poltaranin, M. A., Vukherer, T., Kornienko, L. A., &
P. M., A. M., M. M.; writing – review and editing I. D.; visu- Lyukshin, B. A. (2015). Structural fracture scales in shock-loaded
alization, O. Y.; supervision, M. M. and A. M.; project ad- epoxy composites. Physical Mesomechanics, 18, 58–74.
https://doi.org/10.1134/S1029959915010075
ministration, P. M. All authors have read and agreed to the
Yasnii, О. P., Pastukh, O. А., Pyndus, Y. І., Lutsyk, N. S., & Didych I. S.
published version of the manuscript.
(2018). Prediction of the diagrams of fatigue fracture of D16T
aluminum alloy by the methods of machine learning. Materials
Disclosure statement Science, 54, 333–338.
https://doi.org/10.1007/s11003-018-0189-9
The data are available from the corresponding author Yasniy, O., Didych, I., Fedak, S., & Lapusta, Y. (2020). Modeling
upon reasonable requests. of AMg6 aluminum alloy jump-like deformation properties by
machine learning methods. Procedia Structural Integrity, 28,
1392–1398. https://doi.org/10.1016/j.prostr.2020.10.110
References Yasniy, O., Fedak, S., & Lapusta, Y. (2022a). Prediction of jump-like
creep using preliminary plastic strain. Procedia Structural Integ-
Alpayndin, E. (2010). Introduction to machine learning. The Knowl- rity, 36, 166–170. https://doi.org/10.1016/j.prostr.2022.01.019
edge Engineering Review, 25(3), 353–353. Yasniy, O., Pasternak, I., & Sobashek, L. (2022b). Modelling of
https://doi.org/10.101497/S0269888910000056 AL-6061 aluminum alloy deformation diagrams by machine
Bezerra, E. M., Ancelotti, A. C., Pardini, L. C., Rocco, J. A. F. F., learning methods. Procedia Structural Integrity, 42, 1344–1349.
Iha, K., & Ribeiro, C. H. C. (2007). Artificial neural networks ap- https://doi.org/10.1016/j.prostr.2022.12.171
plied to epoxy composites reinforced with carbon and E-glass Zhang, L., & Wei, X. (2022) Prediction of fatigue crack growth
fibers: Analysis of the shear mechanical properties. Materials under variable amplitude loading by artificial neural network-
Science and Engineering A, 464(1–2), 177–185. based Lagrange interpolation. Mechanics of Materials, 171, Ar-
https://doi.org/10.1016/j.msea.2007.01.131 ticle 104309. https://doi.org/10.1016/j.mechmat.2022.104309