JOM

https://doi.org/10.1007/s11837-019-03666-7
 2019 The Minerals, Metals & Materials Society

MODELING AND SIMULATION OF COMPOSITE MATERIALS

Artificial Neural Network Approach to Determine Elastic

Modulus of Carbon Fiber-Reinforced Laminates
1,2
XIANBO XU and NIKHIL GUPTA1

1.—Composite Materials and Mechanics Laboratory, Mechanical and Aerospace Engineering

Department, New York University Tandon School of Engineering, 6 Metrotech Center, Brooklyn,
NY 11201, USA. 2.—e-mail: [email protected]

Recognized as a prominent material for engineering applications, carbon fiber-

reinforced laminated composites require significant effort to characterize due
to their anisotropic structure and viscoelastic nature. Dynamic mechanical
analysis has been used to accelerate the testing process by transforming the
measured viscoelastic properties to elastic modulus. To expand the transfor-
mation to anisotropic materials, artificial neural network approach is used to
build the master relationship of the storage modulus in three in-plane direc-
tions. Using rotation transformation, the stiffness tensor can be calculated to
extrapolate the frequency domain viscoelastic properties in any orientation
with respect to the fiber direction. The viscoelastic properties are transformed
to time domain relaxation function using the linear relationship of viscoelas-
ticity. Stress response with a certain strain history is predicted and the elastic
modulus is extracted. Compared to the experimental flexural test results, the
artificial neural network-based method achieved an error of less than 7.3%.
The results show that the transformation can predict the anisotropic material
behavior at a wide range of temperatures and strain rates.

viscoelasticity of polymers even within the quasi-

INTRODUCTION
static strain rate range also presents a challenge
The ability to tailor the properties of composite and requires developing a rigorous testing regime.7
materials is their biggest asset. However, the Viscoelasticity is widely displayed by a large
directional properties of fiber-reinforced composites variety of materials, including polymers,8 compos-
require investing significant effort in their charac- ites9 and biomaterials.10 Viscoelastic properties are
terization.1 Carbon fiber-reinforced polymer (CFRP) commonly characterized by tensile or flexural tests
composites with laminated structures are now used and dynamic mechanical analysis (DMA).7,11,12
in a wide range of structural applications, including Tensile and flexural tests can provide Young’s
aircraft, automobiles,2 and sports equipment. Due modulus, which can be used in mechanical design.
to their anisotropic properties, laminated compos- However, a large number of specimens need to be
ites need to be characterized in several different tested if the mechanical properties need to be
orientations. The complexity in their characteriza- characterized with respect to strain rates.3,13,14
tion increases if the properties are also required The campaign will grow exponentially when testing
over a range of temperatures and strain rates.3,4 anisotropic materials. In contrast, the storage (E¢)
The elastic modulus is the most important design and loss (E¢¢) moduli obtained from DMA describe
parameter, which is often the focus of the charac- the in-phase and out-of-phase relationships between
terization efforts.5 However, conducting a large stress and strain when a sinusoidal load is
number of tensile or bending tests with respect to applied.15 However, applications of E¢ and E¢¢ are
fiber direction, temperature, and strain rate to limited because they are obtained in the frequency
determine the modulus can be a prohibitively domain and do not provide a direct correlation with
expensive or time consuming.6 The strong Young’s modulus.
 Xu and Gupta

Theoretical models are now available to trans- F~ðE0 ; x; T Þ ¼ F ðE0 ; x; T Þ þ aXðlÞ ð1Þ
form the storage modulus into the elastic modu-
where E¢ is the storage modulus at a certain
lus.16–18 These models require the DMA results to
temperature T and frequency x, l represents the
satisfy the time–temperature superposition princi-
network weights, X is the regulation term, and a is
ple to develop a master curve. The shape of the
the corresponding regularization factor. The ratio of
master curve is modeled by a mathematical function
with respect to frequency, which is then trans- the predicted storage modulus (E~0 ) to the experi-
formed to the time domain. This procedure allows mental E¢ measurement is used to define the
the obtaining of the elastic modulus over a large prediction error, and the L2 regularization, or so-
range of strain rates and eliminates the need to called ridge regression, is used as the regularization
conduct a large number of tensile tests at various term as in:34
strain rates.19 This transformation procedure has N h i2
1X
been verified on isotropic materials, such as neat F ðE0 ; x; T Þ ¼ E~0 ðx; T Þ=E0 ðx; T Þ  1 ð2:aÞ
polymers,16,18,20 syntactic foams17 and nanocompos- N i¼1
ites.21 However, these methods have not been
extended to anisotropic material systems, which 1X n
require setting up a generalized Hooke’s law-based XðlÞ ¼ l2 ð2:bÞ
n j¼1 j
material model. The artificial neural network
(ANN) approach has been used recently in this field
to model the material behavior, and has promise to where N is the size of the dataset and n is the
be applied to materials with complex behaviors.22–24 number of parameters. Temperature and the loga-
The ANN approach has a strong physical founda- rithmic form of frequency are used as the input
tion25 and is good at handling complex thermal neurons. As three is the minimal number to imple-
behaviors.26,27 It is also the paradigm of parallel ment the transformation and predict the perfor-
computing and robustly enables massive computa- mance in other orientations for CFRP, the ANN has
tion tasks,28,29 which is a major advantage in three output neurons to describe the properties in
modeling complex material responses.30,31 three orientations (h = 0, 45 and 90). The defini-
In the current work, a unidirectional CFRP tion of the orientations are shown in Fig. 1a and the
laminate is used as the anisotropic material flow chart of the process is shown in Fig. 1b.
system to implement the ANN approach and use Following previous work,22 the ANN has only one
DMA data to extract the elastic modulus with hidden layer with two neurons and the regulation
respect to strain rate in different orientations factor is set to be 104. The number of neurons and
with respect to the fiber direction. A feed-forward regulation factors are not considered as variables
neural network is applied to build the master for optimization in this work.
relationship of E¢ in different in-plane orienta- The datasets obtained from the DMA tests of the
tions. The viscoelastic properties can be used to laminates in three orientations (h = 0, 45 and 90)
calculate the frequency domain stiffness tensor in are each randomly divided into training, validation
the testing direction using rotation transforma- and test sets in the ratio of 65:15:20, respectively.
tion.32 Then, using rotation transformation again, As a metaheuristic optimization method, particle
the viscoelastic properties in any orientation can swarm optimization (PSO)35 has proved to be effi-
be predicted. The frequency domain viscoelastic cient and robust in solving many nonlinear opti-
properties are transformed to a time domain mization problems.36–38 Hence, PSO is used to train
relaxation function using the linear relationship the ANN in this work. Then, using the linear
of viscoelasticity and then the time domain elastic viscoelastic theory, the E¢ (x, T) in each orientation
modulus is predicted. Since the materials are is transformed to the time domain relaxation func-
tested with the aim of finding a correlation tion E(t, T). The stress response is found by
between the test methods, the quality of the integrating E(t, T) with the strain history, and the
material is not a study parameter and the trans- elastic modulus in the time domain is extracted.
formation should yield accurate results for any
material regardless of composition and quality. The Materials and Experiment Procedure
The unidirectional CFRP laminates of 1.6 mm
MATERIALS AND METHODS thickness were procured from Allred and Associ-
ANN Methodology ates, Elbridge, New York, NY, USA. Laminate
parameters such as the number of layers and fiber
The ANN framework is used to establish the content and defects are not characterized because
master relationship of the storage modulus with they are not of relevance to the present work, which
respect to frequency and temperature. A feed-for- is focused on comparing the mechanical property
ward neural network is used in this work and the results obtained from two different characterization
general loss function is defined as:33 methods.
Artificial Neural Network Approach to Determine Elastic Modulus of Carbon Fiber-Reinforced
Laminates

modulus with respect to temperature and frequency

is presented in Fig. 2. It can be observed that E¢
decreases with temperature and increases with
frequency. The E¢ at 100 Hz is higher than that at
1 Hz for each temperature in each orientation,
which indicates the strain rate sensitivity of the
materials. It is also observed in this figure that the
storage modulus in 0 is at least 2 times higher than
that in 45 and 9 times higher than that in 90. The
obvious directionality is due the anisotropic nature
of the material.
The experimental flexural properties are obtained
by flexural tests under three-point bending condi-
tions in h = 0, 45 and 90 orientations at room
temperature (25C). An Instron 4467 universal test
system was used for the tests at different deflection
rates from 1 9 103–1 mm Æ s1 (nominal initial
strain rates of 3 9 106–3 9 103 s1). The geome-
try of the specimens conforms to the ASTM D790
standard.39 A representative set of stress–strain
curves for a unidirectional CFRP laminate in 0, 45
and 90 orientation (deformation rate of 1 mm/s) is
shown in Fig. 3. Further, flexural tests on the
laminate were also conducted in h = 30 orientation.
The data obtained in this orientation have been
used for validation of the results obtained from the
ANN-based model which was developed using the
data in the 0, 45 and 90 orientations. At least
three specimens were repeated for each strain rate
and orientation, and data for all three specimens
are included in the plots.

Transformation of Storage Modulus to Time

Domain
For anisotropic elastic materials, the stress and
strain are related to the stiffness tensor in both the
time and frequency domains and the relationship
can be written using Einstein summation notation
as:
Fig. 1. (a) Definition of 0, 45 and 90 orientations with respect to
the fiber direction in the laminate and (b) the modeling framework rij ¼ Cijkl ekl ð3Þ
using artificial neural network approach. where r is the stress tensor, e is the strain tensor,
and C is the stiffness tensor. The stiffness tensor is
dependent on h. The CFRP laminate is commonly
regarded as a thin-shell structure and its stiffness
The laminates were machined into specimens of properties can be found by calculating only the in-
dimensions 60.0 9 12.7 9 1.6 mm3 (length 9 plane component of the stiffness tensor. The elas-
width 9 height) in 0, 45 and 90 orientations for ticity term for axial properties under a rotation
DMA testing. TA Instruments (New Castle, DE, transformation can be found using:40
USA) Q800 DMA has been used to conduct the  
experiments under strain control mode in dual Ch1111 ¼ C01111 cos4 h þ 2 C01122 þ 2C01212 cos2 h sin2 h
cantilever configuration with a maximum displace- þ C02222 sin4 h
ment of 25 lm using a span length of 35 mm. The ð4Þ
tests were conducted in the temperature range of
30–200C with a step of 5C and a soaking time of C1111, C1122, C1212, and C2222 correspond to C11,
5 min. Isothermal frequency sweeps were conducted C12, C66, and C22, respectively, when using a 2-index
at 20 discrete frequencies logarithmically spaced notation. Although the stiffness properties in cer-
between 1 and 100 Hz. The experiment was termi- tain orientations are related to four components of
nated when the force magnitude dropped below the stiffness tensor, they are linearly dependent on
104 N. The response surface of the storage three terms in Eq. 4. Rather than calculating all the
 Xu and Gupta

Fig. 2. (a) The trend of storage modulus with respect to temperature at 1 Hz frequency for three orientations of laminates shown in Fig. 1a.
Response surface of storage modulus measured with respect to temperature and frequency in (b) 0, (c) 45 and (d) 90 orientation.

four components of the stiffness tensor, the three 0.9897 and 0.9824 for 90, as shown in Fig. 4. The
terms can be calculated using the test data of three frequency domain viscoelastic properties in the
orientations and used to predict the axial stiffness other orientations can be predicted by rotation
in any orientation. The DMA results in three transformation.
orientations are used to train the ANN with the With the master relationship at a certain tem-
PSO algorithm, and the training errors of the tests perature, the integral relationships of viscoelastic-
set in all three orientations are found to be below ity can transform each term in the stiffness tensor
5.4%. The Pearson’s correlation coefficient R for the to a time domain relaxation modulus by numerically
training set and the test set are found to be 0.9897 integrating Eq. 5 from 0 to infinity:41
and 0.9824 for 0, 0.9897 and 0.9824 for 45 and
Artificial Neural Network Approach to Determine Elastic Modulus of Carbon Fiber-Reinforced
Laminates

Zt
dekl ðsÞ
rij ðt; T Þ ¼ E  e_ ¼ Cijkl ðt  s; T Þ ds ð6Þ
ds
0
where s and e_ represent time variable and strain
rate, respectively. For the standard three-point
bending test, the strain rate is assumed to be
constant at small deformations, so that Eq. 6 can
be approximated as:41
Zt
rij ðt; T Þ ¼ Cijkl ðs; T Þ_ekl ds ð7Þ
0

Using Eq. 7, the stress–strain curve at various

temperatures and strain rates can be predicted and
the time domain elastic response of the materials
can be extracted. The elastic modulus is defined as
0.10% secant modulus. The predicted modulus
Fig. 3. A representative set of flexural stress–strain curves for values with respect to temperature and strain rate
unidirectional CFRP laminate tested in 0, 45 and 90 orientations at are shown in Fig. 6. The elastic modulus decreases
1 mm/s deflection rate. The tests were stopped after acquiring
sufficient data for calculating modulus and were not continued to with temperature and increases with strain rate. It
fracture. can also be observed in this figure that the elastic
modulus in 0 is also at least 2 times higher than
that in 45 and at least 9 times higher than that in
90, which conforms to the anisotropic nature of the
material. To verify the transformation from storage
modulus to elastic modulus, the elastic modulus
values predicted from the DMA experiment data are
extrapolated to room temperature and then com-
pared to those obtained from flexural tests in the 0,
45 and 90 orientations. The comparisons of the
results is shown in Fig. 7. The average errors
between the predicted and experimentally mea-
sured modulus are found to be below 7.3% in any of
the three orientations. The observed variation of the
experimental elastic modulus data in 0 orientation
testing is attributed to possibilities such as a slight
misalignment of the fibers with respect to the
loading direction, as a result of cutting the speci-
mens from a large plate, and gripping procedures in
the test machine.
The accuracy of model prediction is also tested by
Fig. 4. The Pearson’s correlation coefficient R of the experimental predicting the elastic modulus in the 30 orientation
and predicted storage modulus values for training and test set in 0, using the DMA experiment results in the 0, 45 and
45 and 90 orientations.
90 orientations. The predictions are compared to
the results from flexural tests in Fig. 8. The average
Z1 0
error is found to be 3.4% for this case. The high
2 Cijkl ðx; T Þ accuracy proves that the transformation can predict
Cijkl ðt; T Þ ¼ sinðxtÞdx ð5Þ the modulus in any in-plane orientation for a wide
p x
0 range of temperatures and strain rates using a
where t is time, x is the angular frequency, and T is minimum of three specimens. Since the elastic
the temperature. The relaxation function C1111 (t, T) in modulus is a widely used material property in
the 0, 45 and 90 orientations are shown in Fig. 5. mechanical design, the present scheme will signif-
For isothermal processes, Cijkl (t, T) can be used to icantly reduce the time and effort needed to char-
determine the stress response with a certain strain acterize an anisotropic material. In the present
history by:41 work, the fiber volume fraction and ply stacking
 Xu and Gupta

Fig. 5. The relaxation function C1111 (t, T) in (a) 0, (b) 45 and (c) 90 orientations.

sequence are not taken as variables in order to keep scheme without any change in the basic framework.
the experimental testing campaign manageable for The number of neurons in the hidden layer will
measuring the elastic modulus for validation of the have to be adjusted based on the number of
predictions. However, these, and many other parameters in a study.
parameters, can be taken as variables in the ANN
Artificial Neural Network Approach to Determine Elastic Modulus of Carbon Fiber-Reinforced
Laminates

Fig. 6. The predicted elastic modulus of CFRP laminate with respect to temperature and strain rate in (a) 0, (b) 45 and (c) 90 orientations.

CONCLUSION converted to a time domain relaxation function

using a linear relationship of viscoelasticity. The
An ANN-based transformation has been built to
viscoelastic response is predicted using the relax-
predict the elastic modulus of carbon fiber-rein- ation function, and the average error is found to be
forced polymer laminate over a wide range of
below 7.3% in the 0, 45 and 90 orientations.
temperatures and strain rates. The DMA results
Further validation of the method was conducted to
in the 0, 45 and 90 orientations are used as input predict the modulus values in the 30 orientation
to model the directionality. The ANN is trained
based on the learning of the ANN. The error in
using a particle swarm optimization algorithm and
prediction of the elastic modulus in the 30
 Xu and Gupta

3. L. Peroni, M. Scapin, C. Fichera, D. Lehmhus, J. Weise, J.

Baumeister, and M. Avalle, Compos. B 66, 430 (2014).
https://doi.org/10.1016/j.compositesb.2014.06.001.
4. F. Chen, Y. Luo, N.G. Tsoutsos, M. Maniatakos, K. Shahin,
and N. Gupta, Adv. Eng. Mater. (2019). https://doi.org/10.
1002/adem.201800495.
5. A.K. Singh, B. Saltonstall, B. Patil, N. Hoffmann, M. Dod-
damani, and N. Gupta, JOM 70, 310 (2018). https://doi.org/
10.1007/s11837-017-2731-x.
6. F. Chen, G. Mac, and N. Gupta, Mater. Des. 128, 182 (2017).
https://doi.org/10.1016/j.matdes.2017.04.078.
7. A.K. Singh, B. Patil, N. Hoffmann, B. Saltonstall, M. Dod-
damani, and N. Gupta, JOM 70, 303 (2018). https://doi.org/
10.1007/s11837-017-2734-7.
8. D. Ivaneyko, V. Toshchevikov, and M. Saphiannikova,
Polymer 147, 95 (2018). https://doi.org/10.1016/j.polymer.20
18.04.057.
9. Z. Jin, K.P. Pramoda, G. Xu, and S.H. Goh, Chem. Phys.
Lett. 337, 43 (2001). https://doi.org/10.1016/S0009-2614(01)
00186-5.
10. L.J. Dooling, M.E. Buck, W.-B. Zhang, and D.A. Tirrell, Adv.
Mater. 28, 4651 (2016). https://doi.org/10.1002/adma.20150
6216.
Fig. 7. Comparison of elastic modulus of unidirectional CFRP 11. I. Ivaneiko, V. Toshchevikov, K.W. Stöckelhuber, M.
laminate predicted from ANN approach with experimentally Saphiannikova, and G. Heinrich, Polymer 127, 129 (2017).
measured values at room temperature (25C) in 0, 45 and 90 https://doi.org/10.1016/j.polymer.2017.08.051.
orientations. 12. K.P. Sharma, R. Harniman, T. Farrugia, W.H. Briscoe, A.W.
Perriman, and S. Mann, Adv. Mater. 28, 1597 (2016). http
s://doi.org/10.1002/adma.201504740.
13. P.A. Romero, S.F. Zheng, and A.M. Cuitiño, J. Mech. Phys.
Solids 56, 1916 (2008). https://doi.org/10.1016/j.jmps.2007.
11.007.
14. D.D. Luong, D. Pinisetty, and N. Gupta, Compos. B 44, 403
(2013). https://doi.org/10.1016/j.compositesb.2012.04.060.
15. J.D. Ferry, Dependence of Viscoelastic Behavior on Tem-
perature and Pressure (New York: Wiley, 1980), p. 225.
16. S.E. Zeltmann, B.R. Bharath Kumar, M. Doddamani, and N.
Gupta, Polymer 101, 1 (2016). https://doi.org/10.1016/j.poly
mer.2016.08.053.
17. S.E. Zeltmann, K.A. Prakash, M. Doddamani, and N. Gupta,
Compos. B 120, 27 (2017). https://doi.org/10.1016/j.compo
sitesb.2017.03.062.
18. C. Koomson, S.E. Zeltmann, and N. Gupta, Adv. Compos.
Hybrid. Mater. 1, 341 (2018). https://doi.org/10.1007/s42114-
018-0026-5.
19. X. Xu and N. Gupta, Polymer 157, 166 (2018). https://doi.org/
10.1016/j.polymer.2018.10.036.
20. X. Xu and N. Gupta, Materialia 4, 221 (2018). https://doi.org/
10.1016/j.mtla.2018.09.034.
21. X. Xu, C. Koomson, M. Doddamani, R.K. Behera, and N.
Gupta, Compos. B 159, 346 (2019). https://doi.org/10.1016/j.c
ompositesb.2018.10.015.
Fig. 8. Comparison of elastic modulus of unidirectional CFRP 22. X. Xu and N. Gupta, Adv. Theory Simul. (2019). https://doi.
laminate predicted from ANN and measured from flexural tests at org/10.1002/adts.201800131.
room temperature (25C) in 30 orientation. 23. E. Vatankhah, D. Semnani, M.P. Prabhakaran, M. Tadayon,
S. Razavi, and S. Ramakrishna, Acta Biomater. 10, 709
(2014). https://doi.org/10.1016/j.actbio.2013.09.015.
24. L.F. Cupertino, O.P.V. Neto, M.A.C. Pacheco, M.B.R. Vel-
lasco, and J.R.M. d’Almeida, Modeling the young modulus of
orientation of the laminate was found to be below nanocomposites: a neural network approach. in The 2011
International Joint Conference on Neural Networks, San
3.4%, which shows the effectiveness of the ANN Jose, California, USA (2011). https://doi.org/10.1109/ijcnn.2
approach for such materials. The high accuracy 011.6033415.
indicates that the transformation can be used to 25. Z. Sun, E. Ambrosi, A. Bricalli, and D. Ielmini, Adv. Mater.
model and predict the elastic properties of aniso- 30, 1802554 (2018). https://doi.org/10.1002/adma.20180255
tropic materials under a wide range of temperatures 4.
26. P. Jarosz, J. Kusiak, S. Małecki, P. Morkisz, P. Oprocha, W.
and strain rates. Pietrucha, and Ł. Sztangret, JOM 68, 1535 (2016). https://d
oi.org/10.1007/s11837-016-1916-z.
REFERENCES 27. G. Zhao, J. Yang, J. Chen, G. Zhu, Z. Jiang, X. Liu, G. Niu,
1. M. MInus and S. Kumar, JOM 57, 52 (2005). https://doi.org/ Z.L. Wang, and B. Zhang, Adv. Mater. Technol. 4, 1800167
10.1007/s11837-005-0217-8. (2019). https://doi.org/10.1002/admt.201800167.
2. R.A. Sullivan, JOM 58, 77 (2006). https://doi.org/10.1007/s1 28. O. Sporns, Adv. Mater. 5, 488 (1993). https://doi.org/10.100
1837-006-0233-3. 2/adma.19930050624.
Artificial Neural Network Approach to Determine Elastic Modulus of Carbon Fiber-Reinforced
Laminates
29. H. Li, Z. Hu, W. Hu, and L. Hua, JOM 10, 10 (2019). 37. I.C. Trelea, Inf. Process. Lett 85, 317 (2003). https://doi.org/
https://doi.org/10.1007/s11837-018-03326-2. 10.1016/S0020-0190(02)00447-7.
30. D.J. Armaghani, E. Tonnizam Mohamad, E. Momeni, M. 38. P.J. Angeline, Evolutionary Optimization Versus Particle
Monjezi, and M. Sundaram Narayanasamy, Arab. J. Geosci. Swarm Optimization: Philosophy and Performance Differ-
9, 48 (2015). https://doi.org/10.1007/s12517-015-2057-3. ences (Berlin: Springer, 1998).
31. RS Anupama Upadhyay, OALib J. 1, 1 (2014). https://doi. 39. Z. Jia, T. Li, F.-P. Chiang, and L. Wang, Compos. Sci.
org/10.4236/oalib.1100903. Technol. 154, 53 (2018). https://doi.org/10.1016/j.compscitec
32. R.B. Heslehurst, Compos. Struct. 35, 369 (1996). https://doi. h.2017.11.015.
org/10.1016/S0263-8223(96)00042-6. 40. R.T.H. Zafer Gürdal and Prabhat Hajela, Design and Opti-
33. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning mization of Laminated Composite Materials (New York:
(Cambridge, MA: MIT Press, 2016). Wiley, 1999).
34. O.A. Mohamed, S.H. Masood, and J.L. Bhowmik, Measure- 41. H. Markovitz, J. Colloid Interface Sci. 98, 292 (1984). http
ment 107, 128 (2017). https://doi.org/10.1016/j.measuremen s://doi.org/10.1016/0021-9797(84)90514-9.
t.2017.05.019.
35. J. Kennedy and R. Eberhart, Particle swarm optimization. Publisher’s Note Springer Nature remains neutral with re-
in Proceedings of ICNN’95 - International Conference on
gard to jurisdictional claims in published maps and institutional
Neural Networks, Perth, WA, Australia, 1995. https://doi.
org/10.1109/icnn.1995.488968. affiliations.
36. M. Clerc and J. Kennedy, IEEE Trans. Evolut. Comput. 6,
58 (2002). https://doi.org/10.1109/4235.985692.