Composite Structures 212 (2019) 199–206

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Prediction of mechanical properties of carbon fiber based on cross-scale FEM T

and machine learning
⁎
Qi Zhenchaoa, , Zhang Nanxib, Liu Yonga, Chen Wenlianga
a
College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China
b
Research Institute for Interdisciplinary Sciences, Shanghai University of Finance and Economics, Shanghai 200433, China

A R T I C LE I N FO A B S T R A C T

Keywords: Carbon fiber is the most common reinforcing phase in composite materials. However, it is difficult to obtain the
Mechanical properties performance parameters of the monofilament. In this study, the relationship between the property variables of
CFRP the carbon fiber monofilament and the macroscopic parameters of the composites is established using a re-
Carbon fiber gression tree, a type of decision tree model, in machine learning. First, in order to obtain the data for machine
Machine learning
learning, representative volume element (RVE) models of single-layer and multi-layer carbon fiber reinforced
Feature selection
plastic (CFRP) are established by a cross-scale finite element method (FEM), and periodic boundary conditions
are loaded. Then, a correlation model between the carbon fiber properties and CFRP and matrix properties is
established. The non-GUI mode is called by Software Isight to generate the sample data. Second, in order to
avoid overfitting, the L1 norm method is used for feature selection before model training. Finally, the four elastic
properties of the carbon fiber are analyzed by a regression tree model. After a series of parameter adjustments
and model selection, the model with a better generalization performance was obtained. The validity of the
models was verified by the validating sample set.

1. Introduction engineers [2]. There are several engineering solutions to this problem,
such as testing the carbon fiber bundle instead of the carbon fiber
Carbon fiber reinforced plastic (CFRP) is a multi-phase solid mate- monofilament or increasing the sample size to exclude random factors.
rial consisting of two or more materials with different physical and However, it is difficult to achieve a satisfactory balance between the
chemical properties, and the use of this material in aerospace systems accuracy and cost.
has increased in recent years [1]. The continuous symmetry of the The matrix phase and reinforcing phase of the composite materials
composites is a matrix, and the dispersed phase is the reinforcing phase. are arranged in order at the micro level to form macroscopic materials;
All phase materials reinforce each other, and the overall performance thus, the mechanical properties have a one to one correlation. The
achieved is superior that of a single material. theory of composite laminates is well established, and the macroscopic
As one of the most commonly used reinforcing phases of aerospace properties of the composite materials can be deduced according to the
composites, the carbon fiber mechanical properties play an important properties and arrangement of the fiber and matrix [3,4]. The theories
role in the evaluation of the composite comprehensive properties, si- of the laminates are based on mathematical formula that ignore the
mulation of component deformation, and prediction of damage beha- stress in the depth of the material and have special boundary condi-
vior. The engineering community utilizes the mechanical property test tions, creating some problems. Under the actual force or the complex
of the carbon fiber, and national associations or enterprises have de- surfaces, the finite element method can directly reflect the structure
veloped relevant testing standards, such as JIS R7608-2007 (Carbon characteristics of the composite materials and construct the relationship
fiber – Determination of tensile properties) in Japan, ASTM D4018- between the micromechanics and macroscopic mechanics. Tang [5]
2011 (Standard Test Methods for Properties of Continuous Filament proposed an analysis method based on a numerical simulation and
Carbon and Graphite Fiber Tows) in America, and GB/T26749 in China. mathematical modeling after the weaving model was established.
However, because of the carbon fiber's tiny size, high artificial effect, Huang [6] proposed a versatile and easy-to-use micro composite model,
and high discrete value of the performance, the testing is expensive. In which can simulate the elastic, elastoplastic, and stress limits of two-
addition, the results are not reliable, which is a problem faced by dimensional braided composites under arbitrary loads. Donadon et al.

⁎
Corresponding author.
E-mail address: [email protected] (Z. Qi).

https://doi.org/10.1016/j.compstruct.2019.01.042
Received 2 October 2018; Received in revised form 19 December 2018; Accepted 4 January 2019
Available online 06 January 2019
0263-8223/ © 2019 Elsevier Ltd. All rights reserved.
Z. Qi et al. Composite Structures 212 (2019) 199–206

Fig. 1. RVE model and boundary conditions.

[7] presented an analytical model for the prediction of the elastic be- tree model has been applied in numerous fields involving data analysis,
havior of plain-weave fabric composites based on classical laminate such as transit service quality analysis [15], listing status prediction of
theory, and the theoretical predictions were compared with the ex- companies [16], software fault prediction [17], soil unit prediction
perimental results and predictions using alternative models available in [18], etc.
the literature. Chun et al. [8] examined the effects of a z-pin re- In this study, the decision tree model is applied to establish the
inforcement on the elastic properties of composite materials as well as relationship between the micro and macro properties of composite
the changes in the elastic properties based on the diameter of the z-pin. materials to achieve accurate and rapid prediction of the micro prop-
Kamiński et al. [9] proposed computational techniques based on the erties. First, ABAQUS software is used to established the finite element
finite element method (FEM) solution of the cell problem to discover model, and sets of data for model training and testing are obtained.
the sensitivities and uncertainties in the homogenized tensor that re- Then, in order to improve the generalization performance of the model,
sulted from variations of the material parameters in the CFRP. Ka- the feature selection of the original data is carried out before the model
marudin et al. [10] established a three-dimensional finite element training, and dimensionality reduction of the data is realized. Finally,
model of unidirectional fiber reinforced composites using a periodic the decision tree model is established according to the object to be
boundary condition method to predict the elastic mechanical behavior predicted, and the optimal prediction model is obtained by adjusting
of a unit cell of such composites. Yang et al. [11] numerically simulated the parameters.
the tensile behavior of a three dimensional (3D) orthogonal woven
composite based on the unit cell model. The influence of the crack 2. FEM modeling and obtaining the data set
damage and the stress on the fiber, resin, and fiber/resin interface were
analyzed. Sun et al. [12] developed a multi-scale computational ana- A sufficient amount of data should be collected for model training
lysis based on representative volume element modeling and molecular and checking. Without considering the cost factors, experimental
dynamics simulations to investigate the microscopic failure mechan- methods cannot meet the requirements of data diversity. Therefore, a
isms of unidirectional CFRP composites. These representative studies finite element model is established to generate the property data using a
predicted the properties of composite materials using a finite element representative volume element (RVE) under periodic boundary condi-
method, and verified by performing experiments, which demonstrated tions. Isight software is used to organize the input data and call
the validity of the finite element method in building the meso-macro Software Abaqus using the noGUI mode. A detailed description of the
performance relationship of the composite material. FEM modeling is presented in a previous article [19].
In practice, macro performance parameters are easy to obtain. If the
microscopic parameters can be obtained by measuring macro para- 2.1. FEM modeling
meters, the cost would be reduced, and the efficiency would be im-
proved. Because of the complex nonlinear relationship between the CFRP is composed of multiple layers of unidirectional materials.
macro and micro performance parameters, the theoretical method First, an RVE model of the unidirectional CFRP (UD-CFRP) should be
should solve complex transcendental equations, which is not a realistic established to predict the mechanical properties.
expectation. In order to realize modeling and solving of the materials, the com-
Therefore, machine learning is used as a reference. Machine posite materials should be idealized at the structural level. The com-
learning is a modeling method for creating the relationship between posite material is assumed to be intact, and the fiber is combined with
input and output quantities from a given data set through algorithmic the matrix. The voids and initial micro cracks produced during fabri-
iteration. Given only historical data, the algorithm can generate the cation are neglected, in which the fiber is transversely isotropic, and the
model automatically according to the iteration rules of the selected matrix is isotropic. Because of the randomness of the fiber arrangement
model. Once the model is built and tested, the model can predict the in the actual structure, the fiber arrangement is simplified to a uniform
output value of the newly generated input information. A regression distribution or staggered distribution in the modeling process. Because
tree model is a machine learning method, which has low modeling and the staggered arrangement is closer to the random distribution of the
calculation costs and a high prediction accuracy [13,14]. The regression fiber, the rectangular staggered RVE is used to predict the anisotropic

200
Z. Qi et al. Composite Structures 212 (2019) 199–206

Fig. 2. RVE model and boundary conditions: (a) Fiber orientation of the [0/90°] CFRP and (b) RVE of the [0/90°] CFRP.

properties of the composites, as shown in Fig. 1(a). The dimensions in 3. Feature selection based on L1 norm
the RVE model are determined from the basic geometric parameters of
the composite materials and fibers. Given a data set, each of the attributes are referred to as features,
The elastic properties of the carbon fiber and matrix are defined in such as the elastic modulus of the fiber, Poisson's ratio of the matrix,
the software program. Carbon fiber is treated as a transverse isotropic etc. Feature selection is the process of selecting the feature subset with
material, and the elastic constants are applied to characterize the me- the greatest correlation to the output data from the feature set of the
chanical properties, including the Young’s modulus, fE1, fE2 = fE3, sample [20]. Feature selection is an important step in data processing.
shear modulus, fG12 = fG13, fG23, Poisson’s ratio, fν12 = fν13, and fν23. Its purpose is to select the feature subset that has the greatest relevance
The elastic constants of the matrix are defined as an isotropic material, to the current learning task from all feature sets, simplify the model,
including the Young’s modulus mE and Poisson’s ratio mν. decrease the computational cost of the model, and reduce the risk of
A normal stress and shear stress of 1 MPa are applied in three di- over-fitting.
rections, and the corresponding stress nephogram under each load is
obtained. According to the applied load and the strain response, the 3.1. The process of feature selection
predicted values of the elastic modulus and shear modulus of UD-CFRP
in all directions can be calculated, including the Young’s modulus, Feature selection is generally divided into two steps. The first step is
E1 = E2, E3, shear modulus, G12, G13 = G23, Poisson’s ratio, ν12, and a subset search, and the second step is a subset evaluation. An em-
ν13 = ν23. bedded selection is a selection method that combines feature selection
For multidirectional CFRP (MD-CFRP), one single layer could be with the learning process. As one method of embedded selection, the L1
considered a UD-CFRP. With [0/90°] CFRP, the RVE could be estab- regularization algorithm can easily obtain a sparse solution to achieve
lished considering the different fiber orientation, as shown in Fig. 2. feature selection. This method considers the data set and learning tasks
The RVE is combined by two parts with different mechanical properties, and avoids a large computation; therefore, it is applied for feature se-
which are obtained from the calculation of the UD-CFRP. lection in this study.
Similar to the UD-CFRP, a normal stress and shear stress of 1 MPa The purpose of this study is to predict the fiber performance para-
are applied in three directions, and the corresponding stress nephogram meters; thus, the features listed in Table 1 should be recognized. The
under each load is obtained. The values of the elastic modulus and elastic parameters of the CFRP and matrix are defined as the input data,
shear modulus of the UD-CFRP in all directions can be obtained. labeled F0–F5. The fiber elastic parameters are the output data labeled
F6–F9, as listed in Table 2.
The program of feature selection is written in Python according to
2.2. Obtaining the data set the flow chart shown in Fig. 3.

The effectiveness of machine learning is dependent on the large

3.2. Feature selection of the output data
amounts of data sets, which contain a training sample and checking
sample. Since one sample can be obtained for each FEM model, it is
Feature selection of the fiber longitudinal elastic modulus fE1 is
time consuming to obtain all samples individually. Isight software is
carried out, and the weight ratio is shown in Fig. 4. Feature 0, the
used to organize the input data that is generated using the Monte Carlo
elastic modulus of CFRP E1 (E2 = E1), has a significant effect on fE1,
method. The FEM modeling step is recorded to generate the PY file with
while the other feature effects are nearly negligible. This is consistent
parameterized codes. The PY file could be called by Isight and run in
with results of the previous studies. The elastic properties of the fibers
non-GUI mode, which will improve the computational efficiency.
determine the properties of the CFRP in the fiber orientation and vice
Five hundred samples were prepared for model training and
versa.
checking. Several samples are listed as examples in Table 1. Since the
Feature selection of the fiber transverse elastic modulus fE2
Poisson’s ratios could be calculated through the Young’s modulus and
(fE3 = fE2) is carried out, and the weight ratio is shown in Fig. 5. This
shear modulus, they were deleted from the sample and were not treated
situation is different from that in Fig. 4. Features 1 and 4, the elastic
as valid features.
modulus of CFRP E3 and elastic modulus of matrix mE, have a sig-
nificant effect on fE2, while the other feature effects are individually
smaller; however, the effects cannot be ignored.

201
Z. Qi et al. Composite Structures 212 (2019) 199–206

Table 1
Several samples of the data set (GPa).
Fiber Matrix CFRP

No fE1 fE2/fE3 fG12/fG13 fG23 mE mν E1/E2 E3 G12 G23/G13

1 155183.9 5676.46 16719.06 13699.34 1694.076 0.174561 48.28472 3.498697 2.659311 2.120019
2 240807.7 18162.21 18859.44 7793.484 3640.422 0.366412 77.82365 10.53344 4.465141 3.439052
3 168299.4 12334.73 24852.19 3024.314 1444.104 0.192185 52.66667 4.13852 2.398585 1.585691
4 193,953 13790.01 11116.56 8518.753 3390.328 0.383944 62.88932 9.420868 3.641958 3.103078
5 132242.7 17159.39 23656.16 11831.93 2807.759 0.216146 43.75523 7.199494 4.169607 3.234731
6 233,628 7148.917 12312.97 6362.479 4861.649 0.281865 73.84071 6.914798 5.0241 3.997205
7 186,861 18007.66 16231.13 7607.108 4208.89 0.334704 61.97462 10.66212 4.866547 3.804595

Table 2
Number of the features.
F0 F1 F2 F3 F4 F5 F6 F7 F8 F9
E1/E2 E3 G12 G23/G13 mE mν fE1 fE2/fE3 fG12/fG13 fG23
CFRP Matrix Fiber
Input data Output data

Fig. 5. Weight ratio of fE2.

Fig. 3. Flow chart of feature selection.

Fig. 6. Weight ratio of fG12.

Fig. 4. Weight ratio of fE1.

Feature selection of the fiber shear modulus fG12 (fG13 = fG12) is

carried out, and the weight ratio is shown in Fig. 6. The situation is
similar to that in Fig. 5. Features 2 and 4, the shear modulus of CFRP
G12 and elastic modulus of matrix mE, have a significant effect on fG12,
while the other feature effects are individually smaller; however, the
effects cannot be ignored.
Fig. 7. Weight ratio of fG23.
Feature selection of the fiber shear modulus fG23 is carried out, and
the weight ratio is shown in Fig. 7. The four features have a significant
effect on fG23; features 2, 3, 4, and 5; the shear modulus of CFRP G12 The correlation of the features could be obtained through the cor-
and G23; and the elastic modulus and Poisson’s ratio of the matrix mE relation coefficient calculation. A global reflection of the correlation of
and mν. all the features is shown in Fig. 8

202
Z. Qi et al. Composite Structures 212 (2019) 199–206

Fig. 8. Correlation of the features.

4. Property prediction based on decision tree regression 4.1. Prediction of the longitudinal elastic modulus fE1

The decision tree model can realize data classification and regres- According to the above decision tree training and prediction pro-
sion, such as decision tree classification and decision tree regression. In cess, the prediction results of the tree depths of 2, 3, 4, and 5 are shown
this study, the decision tree regression algorithm classification and re- in Fig. 9. The acronym RMSE represents the root mean square error
gression tree (CART) is used, which was derived from the algorithm between the predicted value and actual value. The term AV represents
proposed by Breiman et al. [21]. The learning process includes the the average of the actual value. Thus, RMSE/AV is a ratio that shows
following steps: the overall deviation degree of the predicted sample. The rate of the
error is the ratio of the predicted error sample to the total number of
(1) Divide all the data into a training set and test set. samples, and the smaller the rate, the better the prediction accuracy.
(2) Generate a regression tree model to fit the training set data. However, a model with a small error rate may also have a risk of over-
(3) Predict the output value of the test set data with the learning model. fitting; therefore, a variety of factors should be selected.
(4) Compare the predicted value of the model with the true value of the For the prediction results of fE1, the accuracy of the model was
test set. The predicted root mean square error and error rate are satisfactory. When the tree depth increased to 3, the error rate and
obtained. RMSE/AV decreased. As the tree depth increased further, the reduction
(5) Determine the two curves of the predicted value and the real value of the error rate and RSEM were not significant; therefore, the decision
of the model using the serial number of the test data as the hor- tree model with a tree depth of 3 was selected for the fE1 prediction.
izontal axis and the output value as the vertical axis, respectively.

The model depth represents the complexity of the model; thus, se- 4.2. Prediction of the transverse elastic modulus fE2/fE3
lecting the optimal model selects the optimal tree depth. The depth of
the tree could be adjusted. When the tree is deeper, the model more Similar to the prediction process of fE1, the prediction results of the
precisely describes the details of the training set. However, a tree that is tree depths of 5, 6, 7, and 8 are shown in Fig. 10. When the tree depth
too deep will cause over-fitting. Therefore, in order to obtain an ap- increased to 7, the error rate and RMSE/AV decreased. As the tree
propriate model, experiments are carried out on model trees with depth increased further, the error rate and RSEM increased. This
multiple depths. probability was from over-fitting. Therefore, the decision tree model
Because the training data set is small, and the probability of over- with a tree depth of 7 was selected for the fE2/fE3 prediction.
fitting is high, the ideal model can be obtained by setting a small tree
depth and adjusting the random seed number.

203
Z. Qi et al. Composite Structures 212 (2019) 199–206

Fig. 9. Prediction result of fE1.

4.3. Prediction of the shear modulus fG12/fG13 the non-GUI pattern through Software Isight. The regression tree model
is used for feature selection, model training, and sample verifying.
The prediction results of the tree depths of 6 and 7 are shown in Using [0/90°] CFRP as an example, the properties of the carbon fiber
Fig. 11. When the tree depth increased to 7, the error rate and RMSE/ were analyzed, and the conclusions are as follows.
AV decreased. As the tree depth increased further, the reduction of the
error rate and RSEM were not significant. Thus, the decision tree model 1) Longitudinal elastic modulus fE1. The elastic modulus of CFRP E1
with a tree depth of 7 was selected for the fG12/fG13 prediction. has a significant effect on fE1, while the other feature effects are
nearly negligible. The decision tree model with a tree depth of 3 is
4.4. Prediction of the shear modulus fG23 suitable for the fE1 prediction.
2) Transverse elastic modulus fE2/fE3. The elastic modulus of CFRP E3
The prediction results of the tree depths of 6 and 7 are shown in and the elastic modulus of matrix mE have a significant effect on fE2,
Fig. 12. When the tree depth increased to 7, the error rate and RMSE/ while the other feature effects are individually smaller. The decision
AV decreased. As the tree depth increased further, the reduction of the tree model with a tree depth of 7 is suitable for the fE2/fE3 predic-
error rate and RSEM were not significant. Therefore, the decision tree tion.
model with a tree depth of 7 was selected for the fG23 prediction. 3) Shear modulus fG12/fG13. The shear modulus of CFRP G12 and the
elastic modulus of matrix mE have a significant effect on fG12, while
5. Conclusion the other feature effects are individually smaller. Therefore, the
decision tree model with a tree depth of 7 is suitable for the fG12/
In this study, a new approach for predicting carbon fiber mechanical fG13 prediction.
properties is proposed based on machine learning. The mechanical 4) Shear modulus fG23. The shear modulus of CFRP G12 and G23 and
properties of the CFRP are determined by the mechanical properties of the elastic modulus and Poisson’s ratio of matrix mE and mν have a
the fiber and matrix under a specified structure. Therefore, there is a significant effect on fG23. The decision tree model with a tree depth
correlation between CFRP and fiber/matrix; however, the relationship of 7 is suitable for the fG23 prediction.
is complicated. This study establishes this relationship through a cross-
scale finite element model, and then generates data samples by calling

204
Z. Qi et al. Composite Structures 212 (2019) 199–206

Fig. 10. Prediction result of fE2/fE3.

Fig. 11. Prediction result of fG12/fG13.

205
Z. Qi et al. Composite Structures 212 (2019) 199–206

Fig. 12. Prediction result of fG23.

Acknowledgements for woven z-pinned composites. Compos B 2014;64:59–71.

[9] Kamiński M, Pawlak A. Sensitivity and uncertainty in homogenization of the CFRP
composites via the Response Function Method. Compos Struct 2014;118:342–50.
The work reported herein is sponsored by the National Natural [10] Kamarudin KA, Ismail AE. Prediction of elastic properties for unidirectional carbon
Science Foundation of China (51605221, 51875283) and the composites: periodic boundary condition approach. Appl Mech Mater
Aeronautical Science Foundation of China (2017ZE52). The authors 2014;773–774:262–6.
[11] Yang X, Gao X, Ma Y. numerical simulation of tensile behavior of 3D orthogonal
would like to acknowledge the editors and the anonymous referees for woven composites. Fibers Polym 2018;19:641–7.
their insightful comments. [12] Sun Q, Meng Z, Zhou G, Lin SP, Kang H, Keten S, et al. Multi-scale computational
analysis of unidirectional carbon fiber reinforced polymer composites under various
loading conditions. Compos Struct 2018.
References [13] Osborn PG, Osmond DWJ, Thorpe BJ. Compos Mater 1980.
[14] Kotsiantis SB. Decision trees: a recent overview. Artif Intell Rev 2013;39:261–83.
[1] Hu J, Zhang K, Yang Q, et al. An experimental study on mechanical response of [15] Oña JD, Oña RD, López G. Transit service quality analysis using cluster analysis and
single-lap bolted CFRP composite interference-fit joints. Compos Struct decision trees: a step forward to personalized marketing in public transportation.
2018;196:76–88. Transportation 2016;43:1–23.
[2] Hu W, Ton-That MT, Perrin-Sarazin F, Denault J. An improved method for single [16] Zhou L, Si YW, Fujita H. Predicting the listing statuses of Chinese-listed companies
fiber tensile test of natural fibers. Polym Eng Sci 2010;50:819–25. using decision trees combined with an improved filter feature selection method.
[3] Tsai SW, Wu EM. A general theory of strength for anisotropic materials. J Compos Knowl-Based Syst 2017;128.
Mater 1971;5:58–80. [17] Rathore SS, Kumar S. A decision tree logic based recommendation system to select
[4] Bahei-El-Din YA. Multiscale. Mech Compos Mater Struct B 2017. software fault prediction techniques. Springer-Verlag New York, Inc.; 2017.
[5] Tang ZX, Postle R. Mechanics of three-dimensional braided structures for composite [18] Pinheiro HSK, Owens PR, Anjos LHC, Júnior WC, Chagas CS. Tree-based techniques
materials – part I: fabric structure and fibre volume fraction. Compos Struct to predict soil units. Soil Res 2017;55.
2000;49:451–9. [19] Qi Z, Liu Y, Chen W. An approach to predict the mechanical properties of CFRP
[6] Huang ZM, Ramakrishna S. Towards automatic designing of 2D biaxial woven and based on cross-scale simulation. Compos Struct 2018. https://doi.org/10.1016/j.
braided fabric reinforced composites. J Compos Mater 2002;36:1541–79. compstruct.2018.11.056.
[7] Donadon MV, Falzon BG, Iannucci L, Hodgkinson JM. A 3-D micromechanical [20] Machine Flach P. Learning: the art and science of algorithms that make sense of
model for predicting the elastic behaviour of woven laminates. Compos Sci Technol data. Cambridge University Press; 2012.
2007;67:2467–77. [21] Breiman L, Friedman J, Olshen RA, et al. Classification and regression trees[J].
[8] Chun HJ, Son J, Kang KT, Byun JH, Um MK, Lee SK. Prediction of elastic properties Encyclopedia Ecol 2015;57(1):582–8.

206