Composite Structures 273 (2021) 114290

Contents lists available at ScienceDirect

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

Machine learning assisted characterisation and simulation of compressive

damage in composite laminates
Johannes Reiner a,⇑, Reza Vaziri b, Navid Zobeiry c
a
School of Engineering, Faculty of Science Engineering and Built Environment, Deakin University, Geelong, Australia
b
Composites Research Network, Departments of Civil Engineering and Materials Engineering, The University of British Columbia, Vancouver, Canada
c
Materials Science & Engineering Department, University of Washington, Seattle, USA

A R T I C L E I N F O A B S T R A C T

Keywords: A data‐rich framework is presented to consistently characterise the macroscopic strain‐softening response of
Damage characterisation laminated composites subjected to compressive loading. First, a highly efficient continuum damage finite ele-
Finite element analysis ment (FE) model is used to simulate compact compression tests of quasi‐isotropic IM7/8552 carbon fibre‐
Continuum damage mechanics reinforced polymers in order to generate a large virtual dataset for training of machine learning (ML) models.
Machine learning
Then, two ML methods, one based on theory‐guided neural network architecture to solve the inverse FE prob-
lem, and one based on recurrent neural networks with Long Short‐Term Memory (LSTM) architecture to solve
the forward FE problem, are trained and predictive capabilities are compared. It is found that theory‐guided ML
for the inverse FE problem yields high loss values and is not applicable to compressive damage characterisation
whereas a minimum number of 5,000 FE simulations are needed to train accurate LSTM models for the forward
problem. Numerical calibration using the trained LSTM model is validated successfully against experimental
data obtained from a wide range of compressive tests including size effect studies in open‐hole compression
tests and axial crush tests of composite tubes. The proposed strategy demonstrates the effectiveness and chal-
lenges of ML to reduce experimental efforts for damage characterisation in composites subjected to compres-
sive loads.

1. Introduction failure modes, the investigation of compressive failure mechanisms is

further complicated by high sensitivity to fibre misalignment and
The understanding and simulation of failure mechanisms and dam- imperfections which can lead to unstable failure processes [3] as well
age evolution in fibre‐reinforced polymers (FRPs) remains an ongoing as by limited visibility of failure modes. Moreover, failure mechanisms
challenge. Thereby, research on the initiation, progression and interac- include frictional effects and contact forces between newly formed
tion of various failure modes has been largely focussed on composites surfaces.
subjected to tensile loadings. However, it is equally important to inves- Considering the complexity of failure mechanisms, difficulties asso-
tigate failure mechanisms due to compressive loads. In many cases, the ciated with compression testing, and process‐induced variabilities,
relatively low compressive strength of composite materials limits the results obtained from compression tests on composites can vary signif-
design and the exploitation of its full potential and capabilities. Exam- icantly. In a round‐robin test program on in‐plane compressive proper-
ples of typical compression‐dominated load cases in aerospace or auto- ties of uni‐directional and cross‐ply glass‐ and carbon‐fibre‐reinforced
motive applications are crashworthiness, impact or bearing behaviour composites, Schneider [4] reports differences among four different
of bolted joints. laboratories of up to 15% and 10% in compressive strength and mod-
In contrast to composites under tensile loadings, damage behaviour ulus, respectively.
under compression is governed by kink‐band formation induced by Compared to strength tests, the number of reports on progressive
matrix shear failure [1]. Other typical failure modes observed in exper- fracture tests in compression is even further limited. Pinho et al. [5]
imental fracture tests are fibre splitting, matrix cracking and delamina- studied cross‐ply T300/913 Carbon Fibre Reinforced Polymer (CFRP)
tion [1,2]. In addition to challenges on testing and inspection of tensile laminates in Compact Compression (CC) tests and calculated an initial

⇑ Corresponding author.
E-mail address: [email protected] (J. Reiner).

https://doi.org/10.1016/j.compstruct.2021.114290
Received 11 March 2021; Accepted 15 June 2021
Available online 21 June 2021
0263-8223/© 2021 Elsevier Ltd. All rights reserved.
J. Reiner et al. Composite Structures 273 (2021) 114290

compressive laminate fracture energy of 80 kJ=m2 . However, the tion tools by learning discrepancies between high and low fidelity
authors state that it was not possible to extract meaningful propaga- numerical models [15,16], analyzing high dimensional data for dis-
tion values of fracture energy due to severe presence of delamination covering underlying physical laws [17], analyzing complex micro-
[5]. Zobeiry et al. [6] were able to determine a fracture energy of 85 scopy results [18], and accelerating discovery of novel materials
kJ=m2 in CC tests of quasi‐isotropic IM7/8552 CFRP laminates. [19]. For advanced composites, various successful ML models have
Due to the limited amount of physical compression tests and the been developed in recent years. For example, surrogate Neural Net-
previously mentioned challenges around failure mechanisms, the finite works were recently developed to speed‐up process simulation of com-
element analysis of compressive damage in composites is challenging posites [20,21]. Accurate ML models have been used to predict gaps
and relatively rare. Irrespective of tensile or compressive loading, and shimming geometries during assembly of aircraft components
intra‐laminar (matrix and fibre) and inter‐laminar (delamination) pro- [22]. ML has been used for process inspection including detection of
gressive damage can be simulated effectively by Continuum Damage defects during automated fibre placement processing of composites
Mechanics (CDM) and the Cohesive Zone Method (CZM) [7,8]. [23]. A more active area of research for ML applications in composites
Due to the complex evolution of damage in compression and the is damage detection, characterisation, and simulation [24–26].
critical micromechanical failure mode of fibre kinking, many research- Recently, the authors developed a novel theory‐guided machine
ers have focused on the simulation at the microscale to provide a learning framework for damage characterisation of composites using
physics‐based analysis. Based on the pioneering work on microbuck- the Over‐height Compact Tension (OCT) tests [27]. By integrating
ling by Rosen [9] in the 1960s, latest FE analyses on the microscale knowledge of failure, several theory‐guided Neural Networks (NNs)
consider matrix cracking through CDM and fibre–matrix debonding in series were trained to successfully calibrate parameters of a bi‐
via CZM [3,10,11]. linear strain‐softening curve based on OCT test results. In this research,
These models have high physical accuracy, however these are lim- simulations were initially conducted based on random combinations of
ited to small‐scale studies due to their large computational cost and damage properties representing a wide range of strain‐softening
model complexity. curves. Results from these simulations combined with theory‐based
To simulate composites at the coupon level, meso‐scale analyses feature transformations were used to train a multi‐level NN architec-
consider each ply of the laminate as the representative volume. Pinho ture to calibrate damage properties one at a time. Trained NN models
et al. [12] developed a ply‐based material model considering shear were used to predict damage properties using load–displacement
non‐linearity in the matrix and initial fibre‐misalignment for simula- curves from OCT tests as inputs (i.e. inverse FE modeling). These prop-
tion of kinking. This material model is available as MAT261 in the erties were successfully validated by conducting destructive sectioning
commercial FE software LS‐DYNA. Costa et al. [1] demonstrated tests on failed specimens.
how mesoscopic material models can be applied to progressive crush This paper presents a data‐rich framework to consistently charac-
tests of composites. Shipsha et al. [13] simulated bolt bearing failure terise the macroscopic strain‐softening response of laminated compos-
of non‐crimp fabrics by applying CDM to represent intralaminar dam- ites subjected to compressive loading. Section 2 presents the
age and CZM for delamination. underlying experiments and efficient FE continuum damage model
In aiming at simulating large composite structures, sub‐laminates that are used to create large data sets for the training of neural net-
can be considered for constructing the inelastic response of composite works. The formulation of these neural networks is outlined in Sec-
laminates with the overall goal of capturing the macroscopic beha- tion 3 alongside a strategy to combine the limited physical data with
viour of the laminate. McGregor et al. [14] demonstrated how such over 10,000 FE simulation results. The optimal machine learning
computationally efficient framework can be applied for simulating informed strain‐softening responses are validated against experimental
axial crushing of large braided composite tubes. data obtained from a wide range of compressive tests in Section 4
In addition to the different FE modelling techniques for the simula- before concluding in Section 5.
tion of compressive damage in composites, additional complexities
arise on effective calibration of numerical models based on experimen- 2. Testing & simulation of compressive damage
tal results to best describe the complex nature of damage evolution
under compression. Here, we are proposing an automated damage The progressive fracture tests and FE simulations are conducted on
characterisation method assisted by machine learning to calibrate quasi‐isotropic ½90=45=0=  454s laminates made from Hexcel HexPly
the compressive strain‐softening response at the macroscale. IM7/8552 CFRP with nominal ply thickness of 0.125 mm. IM7/8552
Current calibration processes of damage models for composites rely CFRP has been widely applied to different compressive tests at the cou-
on time consuming data reduction efforts, as well as trial‐and‐error pon [2,28] and structural level [29]. Section 4 will assess these test
numerical simulations. Depending on the choice of material model, results to validate the FE simulations using the characterised strain‐
several material properties (e.g. strengths and fracture energies) and softening response obtained from machine learning as outlined in
model parameters must be calibrated. From a mathematical point of Section 3.
view, the calibration process is equivalent to a multidimensional opti-
misation problem, where several FE inputs are optimised to minimise 2.1. Compact compression test
prediction error(s) compared to experiments. However, given the high
dimensionality of the problem, inherent uncertainties and variabilities The compressive fracture tests depicted in Fig. 1 applied to quasi‐
in experimental results, and noise/errors in numerical simulations, the isotropic dispersed ½90=45=0=  454s laminates produces stable and
optimisation task is generally quite complex. Success of trial‐and‐error self‐similar crack growth where delamination is confined to a narrow
approaches heavily depends on knowledge and expertise of engineers. fracture process zone ahead of the crack. Zobeiry et al. [6] conducted
In addition to being time‐consuming, this may result in incorrect cali- quasi‐static fracture tests with a loading rate of 0.25 mm/min.
bration and low accuracy numerical models. The rise of Machine Thereby, the Pin Opening Displacement (POD) and the corresponding
Learning (ML) methods and advanced analytics offers an opportunity cross‐head loads were recorded. Fig. 2 shows two of these load‐POD
to streamline this calibration process to reduce time and efforts, while curves that will be used as input data for ML models.
increasing the accuracy of calibrated models and reducing risk of
human errors. 2.2. Laminate-based finite element simulation
ML methods are currently used in various engineering applications.
Most notably, ML models are used for increasing accuracy of simula- In order to generate a sufficiently large dataset of virtual results for
training ML models with acceptable accuracies, an efficient laminate‐

2
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 1. Geometry, dimensions and finite element representation of a test sample for compact compression tests on quasi-isotropic ½90=45=0=  454s IM7/8552
laminates.

simplify the generation of input data, it is assumed that damage initi-

ation coincides with the onset of yielding with the strain at damage
initiation being denoted by ɛ i . This is similar to the manner in which
the MAT81 material model in LS‐DYNA was used in previous studies
involving damage simulation of quasi‐isotropic composite laminates
[6,27,30,31,35]. The computation of the described model takes
around five minutes on 12 CPUs which is considerably faster than con-
ducting a large number of physical CC tests. This efficiency proves to
be essential for generating sufficiently large dataset required for the
machine learning algorithm outlined in the following section.
Fig. 3(a) shows a typical curve for damage evolution as a function
of the effective plastic strain. While tensile damage in quasi‐isotropic
laminates can be modelled by monotonically decreasing bi‐linear dam-
age curves [27], a stress plateau is required before ultimate failure in
compressive damage modelling (i.e. tri‐linear softening response) to
represent kink‐band broadening and slip surfaces where load can be
Fig. 2. Macroscopic force vs Pin Opening Displacement (POD) curves from
transferred between two sides of cracks via friction [6]. Hence, the
compact compression tests on quasi-isotropic ½90=45=0=  454s IM7/8552
damage curve in Fig. 3(a) can be described by the angle θ, intermedi-
laminates [6].
ate effective plastic strain of ɛ 1  ɛ i as well as the ultimate effective
based model is used to speed‐up simulations. Fig. 1 shows the finite plastic strain of ɛs  ɛ i . Here ɛ i and ɛs denote the damage initiation
element discretisation of the CC test sample where the mesh consists and saturation strain, respectively. Fig. 3(b) illustrates a typical
of 4965 nodes and 4848 single shell elements through the thickness. stress–strain curve resulting from the elastic constants and the previ-
The element size within the expected damage zone around the notch ously described damage evolution. Stress increases linearly until
is 1 mm x 1 mm. Note that such discretisation has produced good cor- reaching σ peak (which coincides with both the peak stress and yield
relations between simulations and experiments in transverse impact stress). The tri‐linear post‐peak response consists of an initial decrease
tests [30] and notched tension and compression tests [31–34]. A pre- of stress governed by the angle θ until reaching the strain value ɛ1 at
scribed displacement is imposed to the rigid loading pins (green) in stress σ 1 . The stress plateau at σ 1 until saturation strain ɛs represents
opposite vertical directions. Note that the mesh in Fig. 1 is suitable kink band broadening. The area under the resulting stress–strain curve
to model the quasi‐isotropic dispersed laminate where delamination is the fracture energy density g f which is related to the fracture energy
is assumed to be negligible. If interface failure becomes more critical, Gf by Bazant’s crack band scaling law [36] such that Gf ¼ g f l with the
multiple layers of shells (or continuum elements) can be stacked and characteristic element length l . Overall, we consider the following
connected via cohesive elements/surfaces. However, such higher fide- five input parameters to be calibrated by the proposed machine learn-
lity models increase the computation time significantly and hence are ing algorithm: E; ɛi ; θ; ɛ1 and ɛs . Note that Poisson’s ratio is chosen to
not suited to generate a large amount of data for machine learning be constant ν ¼ 0:32.
purposes.
The quasi‐isotropic composite laminate is modelled by the isotropic 3. Machine learning for damage characterisation
coupled damage‐plasticity material model MAT81 in the commercial
FE software LS‐DYNA. As input material properties, the model requires 3.1. Data generation
elastic laminate properties such as Young’s modulus E and Poisson’s
ratio ν as well as yield stress σ peak and the evolution of a damage vari- By randomly varying material input properties, 15,000 FE simula-
able d as a function of the effective plastic strain. The evolution of tions were conducted using MAT81 in LS‐DYNA. Force‐POD data from
damage can be conveniently incorporated by local stress–strain curves these simulations were recorded as outputs. The following ranges of
of general shapes that include strain‐softening features. In order to parameters were considered to create random inputs using a random

3
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 3. Illustration of laminate response of quasi-isotropic composite laminates by means of material card MAT81 in LS-DYNA to describe (a) damage evolution vs
plastic strain and (b) the resulting laminate stress–strain behaviour.

number generation with a uniform distribution in Python (numpy.ran-

dom.rand):

• 45 GPa ⩽ E ⩽ 60 GPa
• 0.7% ⩽ ɛi ⩽ 2.0%
• 2.0% ⩽ ɛ1 ⩽ 20.0%
• 65 ⩽ θ ⩽ 85
• 6.0% ⩽ ɛs ⩽ 50.0%.

Based on these values, other parameters for the strain‐softening

curve shown in Fig. 3 were calculated. The following ranges were
obtained based on above parameters:

• 50 kJ=m2 ⩽ Gf ⩽ 150 kJ=m2

• 315 MPa ⩽ σ peak ⩽ 1200 MPa
• 25 MPa ⩽ σ 1 ⩽1150 MPa.
Fig. 4. Filtered FE simulation results for noise reduction before application of
machine learning.
3.2. Filtering & data reduction

Following our previous study [27], force‐POD results from FE sim-

3.3. Training and validation
ulations were used to measure several key parameters from each sim-
ulation (i.e. features for ML training). However, before measuring
Two ML‐based methods were investigated for damage characterisa-
these values, given the high level of noise in numerical data, FE data
tion. Initially, the ML method described in our previous study [27] was
was initially filtered to create smooth output curves. For this, the
investigated to train feed‐forward Neural Networks (NN) for inverse
Savitzky‐Golay filter [37] in Python (scipy.signal, savgol filter) was
modeling and damage characterisation (ML method I). In the second
used to fit second order polynomials to sub‐sets of every 7 adjacent
ML approach (ML method II), a Recurrent Neural Network (RNN)
points in the force data, while preserving the maximum load as well
was trained for fast forward simulation and damage characterisation.
as POD data. An example of three filtered curves superposed on initial
For both of these methods, an in‐house developed ML code at the
noisy FE results are shown in Fig. 4.
University of Washington, CompML (Composites Machine Learning),
Upon filtering numerical data and following our previous study on
was utilised [38,39]. CompML is a python code based on Tensorflow
damage characterisation of composites under tension [27], the follow-
and Keras libraries with several modules for data generation, training,
ing three categories of parameters were recoded from each filtered FE
prediction, and visualisation of data.
curve:
3.3.1. ML method I
• Maximum force F max and POD at maximum force d100 .
Following our previous study for tensile characterisation [27], a
• Force values F n at 10% intervals of PODs compared to d100 (from
theory‐based feature transformation was initially used to increase cor-
70% to 190% with respect of d100 ):
relation between original inputs and outputs [17]. Specifically, in pre-
F n ðd ¼ dn Þ where n ∈ ½70; 190: ð1Þ vious study it was observed that the fracture energy of the laminate Gf
is highly correlated to the following combination of outputs from
• Dissipated energies U n determined by evaluating the area under
Over‐height Compact Tension (OCT) simulations [27]:
the force‐POD curves and an assumed linear unloading path at
10% intervals of PODs with respect of d100 : U 170
Gf /
0:75 : ð3Þ
F max
U n ðd ¼ dn Þ where n ∈ ½70; 190: ð2Þ F 170

Here U 170 refers to the dissipated energy under the force‐POD curve
Finally, a dataset with 15,000 datapoints was created based on at a POD equal to 170% of POD at the peak force (1.7  d100 ), and F 170
parameters of strain‐softening curves as inputs, and above measured is the force at this specific POD. However, such high correlation was
features from force‐POD curves as outputs. not observed in this study for Compact Compression (CC) tests. A com-

4
J. Reiner et al. Composite Structures 273 (2021) 114290

parison is shown in Fig. 5 where 10,000 FE results for OCT simulations Table 1
are compared with 15,000 FE results for CC simulations. This shows a Hyper parameters for training of NN models.
high degree of correlation in OCT (99%) and a much lower correlation ML hyper parameter for training Value
in CC simulations (≈ 75%). A module in CompML was used to investi-
Training datapoints 10,500
gate other combinations of output parameters to increase this correla-
Validation datapoints 4,500
tion [17]. However, after an exhaustive search of more than 200,000 Activation function ReLU
power transformations of outputs, no correlations greater than 76% Hidden layers 5
was achieved. This suggests that unlike OCT simulations with bi‐ Nodes per layer 32
linear strain‐softening curves, the correlation of fracture energy Gf to Optimiser Adam, learning rate = 0.0005
Regulariser L2, weight = 0.001
features of force‐POD curves in CC simulations using tri‐linear soften- Training epochs 1,000
ing curves as model input is complex. This has been previously inves- Batch size 200
tigated in a parametric study to measure the sensitivity of force‐POD hline
curves to model input parameters of a tri‐linear strain‐softening curve
[6,40]. It has been shown that, in addition to fracture energy, peak
stress and plateau stress of the tri‐linear strain‐softening curve highly
affect simulation results (force‐POD curve). This confirms that a simple
correlation for fracture energy based on features of force‐POD curves
may not be feasible for the more complex strain‐softening curve used
in this study.
Instead of feature transformation, original inputs and outputs from
FE simulations were used to train feed‐forward Neural Networks (NN)
for inverse modeling and damage characterisation following proce-
dures in our previous study [27]. A module in CompML was used to
train several NNs with different inputs and outputs. However, after
many attempts, training loss could not be reduced within an accept-
able tolerance in any of these trainings. For example, for predicting
fracture energy and damage initiation strain, two NNs were trained
each with 7 inputs (E; F max ; d100 ; U 170 ; F 80 ,F 110 ; F 170 ) and hyper param-
eters listed under Table 1. The network to predict fracture energy is
shown in Fig. 6. Following loss values were obtained after these
trainings:

• Fracture energy Gf : root mean squared error of 11 kJ=m2 and max- Fig. 6. Example of a Feedforward neural network for inverse modelling and
damage characterisation based on outputs of FE simulations.
imum error of 40 kJ=m2 .
• Damage initiation strain ɛ i : root mean squared error of 0.3% and
maximum error of 0.7%. 3.3.2. ML method II
Assuming that the inverse problem is ill‐posed with several plausi-
Given the ranges of parameters previously defined in Section 3.1, ble solutions, a second ML approach was developed. In this approach,
these high loss values were not satisfactory for inverse modeling and using a module in CompML, a Recurrent Neural Network with Long
accurate damage characterisation following previously established Short‐Term Memory (LSTM) architecture was trained to predict the
method [27]. This is due to the fact that the inverse problem is ill‐ entire force‐POD curve as a function of FE input parameters. The LSTM
posed with several acceptable solutions for a given FE output. To mit- architecture is shown in Fig. 7. Different LSTM models with different
igate this issue, a second ML approach was investigated in this study. dataset sizes were investigated. Convergence of error as a function
of dataset size is shown in Fig. 8. This shows that a minimum of
5,000 FE simulations are needed to train an accurate LSTM model.
Hyper parameters for training of the LSTM model with 15,000 FE sim-
ulations are listed under Table 2. Comparison of five FE simulation
results and LSTM predictions are shown in Fig. 9. This shows a high
degree of accuracy in LSTM, with the benefit of increased speed over
FE for fast forward modeling.
Using the trained LSTM model, a large number of simulations can
be conducted in a short amount of time. In a gird search, FE inputs
were varied between pre‐defined ranges in fine steps and 65,000 sim-
ulations were conducted using LSTM in less than 30 min on a typical
computer workstation. Results from these LSTM predictions were com-
pared with two CC experimental data shown in Fig. 2 to identify the
combination of strain‐softening parameters resulting in the lowest
error value defined below:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
∑N ½F LSTM ðdÞ  F EXP ðdÞ2
Error ¼ with d ∈ ½0; dmax ; ð4Þ
N

Fig. 5. Correlation of laminate fracture energy Gf and a combination of FE where F LSTM ðdÞ and F EXP ðdÞ are force values obtained from LSTM simu-
outputs obtained in this study for Compact Compression (CC) tests compared lations and experiments at given PODs d, respectively.
to previously determiend values in Over-height Compact tension (OCT) tests This analysis showed several combinations of input parameters
[27]. with low error values compared to experimental results. For example,

5
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 7. Architecture of the LSTM model to predict load as a function of

parameters of the strain-softening curve and pin opening displacement (POD) Fig. 9. Comparison of FE simulation results and LSTM predictions for five
d. different cases.

damage initiation and fracture energy result in low error values less
than 0.5 kN. This also confirms previous observation that the inverse
problem is ill‐posed. By expanding the above search and conducting
65,000 LSTM simulations, several combinations of input parameters
were identified that result in low error values. Some of the identified
strain‐softening curves are shown in Fig. 11(a), while corresponding
force‐POD curves are compared with experimental data in Fig. 11
(b). This shows that while input curves are quite different, resulting
force‐POD curves are relatively similar. After comparing all results,
five strain‐softening curves with lowest error values were selected to
be further investigated. These cases are listed under Table 3 with asso-
ciated parameters and error values. The white crosses in Fig. 10 indi-
cate the damage initiation strains ɛi and fracture energies Gf of these
five strain‐softening curves in relation to the prediction error.

4. Validation

Fig. 8. Convergence analysis of training error of the LSTM model as a function It is important to further evaluate the chosen strain‐softening
of training size. curves shown in Table 3 and apply them to other load cases and
geometries. First, Section 4.1 will directly compare the optimal stress–-
strain curves listed in Table 3 to experimental measurements obtained
Table 2
from digital image correlation (DIC) analysis on the CC tests shown in
Hyper parameters for training the LSTM model.
Fig. 1. The second and third validation cases in Sections 4.2 and 4.3
ML hyper parameter for training Value consider other compressive load cases than the previously presented
Training datapoints 5,175,000
Validation datapoints 1,295,000
Activation function tanh
Hidden layers 5
Nodes per layer 32
Optimiser Adam, learning rate = 0.001
Training epochs 2,000
Batch size 100,000

consider the following case where only two inputs (damage initiation
strain ɛ i and fracture energy Gf ) are varied:

• E ¼ 45 GPa
• 0.8% ⩽ ɛi ⩽ 2.0%
• σ 1 ¼ 200 MPa
• θ ¼ 82
• 60 kJ=m2 ⩽ Gf ⩽ 120 kJ=m2

By discretising the ranges of the two variables ɛi and Gf into 20

Fig. 10. Error response surface obtained by varying two input parameters ɛi
steps, the LSTM model was used to predict force‐POD curves and cal- and Gf and comparing LSTM predictions with experimental data. White
culate error values. These were used to construct an error response sur- crosses indicate parameters of the five selected strain-softening used for
face as shown in Fig. 10. This shows that many combinations of further investigation shown in Table 3.

6
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 11. (a) Several strain-softening curves identified by LSTM model to result in low error values, and (b) comparison of force-POD curves from experimental data
with LSTM predictions based on strain-softening curves in (a).

Table 3
Strain-softening curves identified by LSTM model with lowest error values compared with experimental results.

Case ID E (MPa) ɛ i (–) θ ð Þ σ 1 (MPa) Gf (kJ=m2 ) Error (kN)

3343 45000 0.011 70 125 100 0.23

10277 45000 0.015 82 200 90 0.26
11845 45000 0.016 83 100 90 0.27
4838 45000 0.012 69 200 95 0.28
1928 45000 0.010 72 100 105 0.29

CC test. Thereby, the FE models are consistent with those described in of central double 0∘ ply block). Xu et al. [41] found that these different
Section 2 on simulations of CC tests using one reduced integrated shell layup sequences show similar failure responses.
element through the thickness and an approximate element size of The size of the FE mesh as shown in Fig. 13 is 1 mm × 1 mm
1 mm × 1 mm. around the expected damage zone which is consistent with previously
outlined discretisation in CC simulations. A prescribed displacement
4.1. Compact compression and stress–strain data was imposed on one edge along the gauge width of the OHC samples
while the nodes at the opposite edge were fully constrained.
In addition to the macroscopic force‐POD data shown in Fig. 2, The simulation results using the five strain‐softening curves from
Zobeiry et al. [6] estimated the stress–strain response within the frac- Table 3 are shown in Table 5 and Fig. 14(a). It can be seen that most
ture process zone of the quasi‐isotropic laminate. The analysis of data of the predicted strength values are within 10% of the experimentally
obtained from DIC on the surface of the outer 90° ply revealed the grey measured data. Most of the simulations yield higher strength values
curves shown in Fig. 12(a). The optimised strain‐softening responses compared to experiments. Moreover, LSTM strain‐softening curves
from LSTM shown in Table 3 correlate generally well with the exper- with a lower error lead to lower errors in the OHC analysis. However,
imentally measured data. However, the results further illustrate the ill‐ it is difficult to identify clear trends between OHC predictions and fea-
posed problem outlined in previous section where many solutions with tures of the stress–strain curves in Fig. 12(a) which underlines the dif-
different features such as softening shape, peak stress or maximum ficulty of finding a unique macroscopic material response to describe
strain values yield low errors. Since the measured damage height of compressive damage in composites. Fig. 14(b) shows that simulation
around 6 mm is implicitly incorporated into the DIC analysis, the opti- errors do not correlate with the peak stresses σ peak or the post‐peak
mal stress–strain curves obtained from previous LSTM analysis are angles θ from the strain‐softening curves that are used as input data.
scaled according to Bazant’s crack band model [36].
Fig. 12(b) shows the force‐POD result obtained from one of the five 4.3. Tube crushing
selected strain‐softening curves. The simulation results agree well with
the experimental data as well as with the LSTM results shown in Crashworthiness is mainly driven by the evolution of compressive
Fig. 11(b). damage modes. Therefore, experimental tube crush tests on quasi‐
isotropic ½02 =  452 =902 s IM7/8552 laminates performed at the Oak
4.2. Size-effects in open-hole compression tests Ridge National Laboratory (ORNL) [29] were considered to provide
more application‐driven validation examples for the characterised
Lee and Soutis [2] studied a wide range of Open‐Hole Compression strain‐softening responses shown in Fig. 12(a). Axial crushing of com-
(OHC) tests on ½45=90=  45=04s IM7/8552 CFRP laminates. From posites mainly initiates two failure modes: fragmentation and/or
these data, three different geometries were selected for analysis as splaying. The former can be attributed to fibre damage whereas the
listed in Table 4. Fig. 13 shows the corresponding geometry of the latter results from delamination. Since the presented macroscopic FE
OHC specimens with hole diameter D, gauge length L and gauge width model is only applicable to cases where delamination is negligible,
W. the two test cases labelled No.46B and 47B from the ORNL database
Note that the stacking sequence is different from the were chosen due to the reported dominant failure mode of fragmenta-
½90=45=0=  454s laminate used in previous sections (e.g. presence tion [42].

7
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 12. (a) Comparison of the approximate strain-softening curves across the damage zone obtained in compact compression tests [6] and those curves
determined by the LSTM model; and (b) comparison of force vs POD data between experiments and FE simulation using one of the five selected strain-softening
curves shown in Table 3.

Table 4 where structured meshes demonstrated potential for unstable crushing

Dimensions (mm) of open-hole compression (OHC) specimens [2]. Fig. 13 and unrealistic buckling. A laminate thickness of 2.16 mm is assigned
illustrates the specimen geometry. to the shell elements. The crush initiators were modelled through pro-
Hole diameter D Gauge width W Gauge length L gressive reduction of the laminate thickness as illustrated in Fig. 15. A
prescribed displacement is applied to the rigid impact plate. Contact
Small 3.2 15.9 63.5
between the impact plate and the composite tube is modelled by a con-
Medium 6.4 31.8 127.0
Large 12.7 63.5 254.0 stant friction coefficient of 0.2 [45]. The translational degrees of free-
dom at the bottom of the tube are fully fixed.
The crush responses is quantified by evaluating the contact forces P
at the tube and the displacement d of the impact plate. Fig. 16(a)
shows the crush force vs displacment responses from the five simula-
tions compared with two experimentally measured crush forces. It
can be seen that simulations yield very similar results.
For further comparison, the specific energy absorption (SEA) is cal-
culated as
PCrush
SEA ¼ ð5Þ
ρA
where PCrush is the mean crush force, ρ is the density of the material and
A the cross‐sectional area of the tube specimen. PCrush and SEA are deter-
mined for the crush displacement range between 15 mm and 60 mm
which is in the range of stable steady‐state crushing.
Table 6 compares the maximum crush forces Pmax and SEA values
obtained from the five simulations with different strain‐softening input
as listed in Table 3. In comparison with experimental data listed in
Table 6, all simulations yield reasonable results with lower peak forces
Fig. 13. Geometry, dimensions and finite element representation of the open- and slightly higher SEA values. Moreover, the SEA values are consis-
hole compression (OHC) sample consisting of hole diameter D, gauge length L tently predicted within the range of 72–75 kJ/kg which is around
and gauge width W.
10% higher compared to experimental results. The peak forces vary
among the simulations.
It should be noted that peak force values highly depend on the con-
The tube consists of a square cross‐section with rounded corners
figuration of the crush trigger. The effect of the assumed crush initia-
and bevel‐shaped crush initiators to initiate stable crushing. To be con-
tors on the predicted peak force was not further investigated here.
sistent with previous FE models, the tube was meshed with an approx-
However, simulation results in Table 6 can still be compared with each
imate element size of 1 mm × 1 mm. An unstructured mesh was
chosen based on previous sensitivity analyses on crushing [43,44]

Table 5
Comparison between measured data [2] and predicted compressive strength by simulation [% error to experiments] in MPa in 4 mm thick ½45=90=  45=04s IM7/
8552 laminates over the range of OHC samples listed in Table 4.

OHC sample Experiments Simulation ID

[2] #3343 #10277 #11845 #4838 #1928

Small 351 357 [+1.8%] 365 [+4.0%] 375 [+6.9%] 387 [+10.4%] 324 [−7.5%]
Medium 300 310 [+3.4%] 317 [+5.7%] 316 [+5.4%] 336 [+12.0%] 281 [−6.0%]
Large 285 272 [−4.3%] 281 [−1.3%] 289 [+1.4%] 295 [+3.6%] 248 [−12.9%]

8
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 14. Results from OHC FE simulations showing (a) errors to experimentally reported strength values using the five selected strain-softening curves and (b) lack
of correlation between prediction errors and features of the strain-softening curves.

Fig. 15. Illustration of modelling tube crush test of quasi-isotropic ½02 =  452 =902 s IM7/8552 laminates.

other to identify correlations to the strain‐softening curves as model 5. Conclusion

input.
Fig. 16(b) shows the correlation between the predicted peak crush This paper presents a data‐rich framework to characterise the
forces Pmax and features of the strain‐softening curves depicted in strain‐softening response of quasi‐isotropic IM7/8552 composite lam-
Fig. 12(a). In contrast to previous study on OHC tests, clear trends inates subjected to compressive loads. Only two experimentally deter-
can be observed where peak crush forces Pmax increase with the peak mined load–displacement curves from compact compression (CC) tests
stress σ peak of the strain‐softening curve. In addition, higher fracture are considered in combination with a large virtual dataset obtained
energy input data Gf leads to lower peak crush forces Pmax . from highly efficient Finite Element (FE) simulations.

9
J. Reiner et al. Composite Structures 273 (2021) 114290

Fig. 16. Analysis of crush behaviour by (a) comparing crush force vs displacement response of FE simulations using the five selected strain-sofetning curves with
experimental measurements [29] and (b) correlation of simulated peak crush force Pmax with features of these strain-softening curves.

Table 6
Analysis of peak crush force Pmax and Specific Energy Absorption (SEA) in axial tube crush tests.

Simulation ID Experiments [29]

#3343 #10277 #11845 #4838 #1928 #46B #47B

Peak force Pmax (kN) 54.0 61.8 61.3 57.5 52.4 64.2 70.4
SEA (kJ/kg) 74.5 73.0 72.5 74.4 73.6 68.3 66.8

It is found that a previously developed theory‐guided machine Acknowledgments

learning (ML) approach for inverse FE modelling and damage charac-
terisation under tension is not suitable for the characterisation of com- We would like to acknowledge many fruitful discussions with col-
pressive damage. It was shown that this is due to the ill‐posedness of leagues at the Composites Research Network (CRN) in Canada.
the inverse FE problem. Therefore, a recurrent neural network with
Long Short‐Term Memory (LSTM) (i.e. forward FE modeling) was References
trained to predict the entire load–displacement curve as a function
of FE input parameters. Results show that a minimum number of [1] Costa S, Fagerström M, Olsson R. Development and validation of a finite
deformation fibre kinking model for crushing of composites. Compos Sci
5,000 FE simulations are needed to train accurate LSTM models. Technol 2020;197:. https://doi.org/10.1016/j.compscitech.2020.108236108236.
The LSTM analysis reveals that multiple sets of FE input parameters [2] Lee J, Soutis C. Measuring the notched compressive strength of composite
with vastly different features in the strain‐softening response can be laminates: Specimen size effects. Compos Sci Technol 2008;68(12):2359–66,
deformation and Fracture of Composites: Analytical, Numerical and Experimental
considered to accurately predict the laminate behaviour in progressive Techniques, with regular papers. doi: 10.1016/j.compscitech.2007.09.003..
fracture tests. [3] Sun Q, Zhou G, Guo H, Meng Z, Chen Z, Liu H, Kang H, Su X. Failure mechanisms
Despite the lack of a unique solution, the wide range of validation of cross-ply carbon fiber reinforced polymer laminates under longitudinal
compression with experimental and computational analyses. Compos Part B Eng
cases including simulations of open‐hole compression and axial tube
2019;167:147–60. https://doi.org/10.1016/j.compositesb.2018.12.041. url:
crush tests demonstrate that strength and energy absorption can be http://www.sciencedirect.com/science/article/pii/S1359836818330294.
predicted within an acceptable error of 10%. [4] Schneider K. Determination of compressive properties of fibre composites in the in-
Given the previously completed damage characterisation in ten- plane direction according to iso 14126. part 1: A round robin test. Appl Compos
Mater 2007;14:1–15. https://doi.org/10.1007/s10443-006-9027-6.
sion, and the results on compressive damage shown in the presented [5] Pinho S, Robinson P, Iannucci L. Fracture toughness of the tensile and compressive
study, future work will focus on simulating impact and compression fibre failure modes in laminated composites. Compos Sci Technol 2006;66
after impact behaviour of quasi‐isotropic IM7/8552 laminates in com- (13):2069–79. https://doi.org/10.1016/j.compscitech.2005.12.023. URL: http://
www.sciencedirect.com/science/article/pii/S026635380600011X.
bination with machine learning informed material characterisation. [6] Zobeiry N, Vaziri R, Poursartip A. Characterization of strain-softening behavior
In addition, given that damage characterisation and calibration of and failure mechanisms of composites under tension and compression. Compos
strain‐softening curves based on one CC test geometry alone does Part A Appl Sci Manuf 2015;68:29–41. https://doi.org/10.1016/
j.compositesa.2014.09.009.
not yield a unique solution, the application of multiple geometries [7] Reiner J, Vaziri R. Structural analysis of composites with finite element codes: An
using the developed ML framework for damage characterisation will overview of commonly used computational methods. In: Beaumont P, Zweben C,
be investigated. editors. Comprehensive Composite Materials II, vol. 8. Oxford: Academic Press;
2017. p. 61–84. https://doi.org/10.1016/B978-0-12-803581-8.10050-5.
[8] Reiner J. A computational investigation of failure modes in hybrid titanium
Data availability composite laminates . The University of Queensland, School of Mechanical and
Mining Engineering. https://doi.org/10.14264/uql.2016.524.
[9] Rosen B. Mechanics of composite strengthening 1965.
The raw/processed data required to reproduce these findings can- [10] Arteiro A, Catalanotti G, Melro A, Linde P, Camanho P. Micro-mechanical analysis
not be shared at this time as the data also forms part of an ongoing of the effect of ply thickness on the transverse compressive strength of polymer
study. composites. Compos Part A Appl Sci Manuf 2015;79:127–37.
[11] Sebaey T, Catalanotti G, Lopes C, O’Dowd N. Computational micromechanics of
the effect of fibre misalignment on the longitudinal compression and shear
Declaration of Competing Interest properties of ud fibre-reinforced plastics. Compos Struct 2020;248:. https://doi.
org/10.1016/j.compstruct.2020.112487. URL: http://
www.sciencedirect.com/science/article/pii/S0263822320311685112487.
The authors declare that they have no known competing financial [12] Pinho S, Iannucci L, Robinson P. Physically based failure models and criteria for
interests or personal relationships that could have appeared to influ- laminated fibre-reinforced composites with emphasis on fibre kinking. part ii: Fe
ence the work reported in this paper.

10
J. Reiner et al. Composite Structures 273 (2021) 114290

implementation. Compos Part A Appl Sci Manuf 2006;37(5):766–77. doi: [29] Crashworthiness of Carbon Fiber Composites, Oak ridge national laboratory. URL:
10.1016/j.compositesa.2005.06.008.. http://energy.ornl.gov/CFCrush/rate_tests/rate_tests.cgi..
[13] Shipsha A, Burman M. Failure mechanisms in ncf composite bolted joints: [30] Reiner J, Zobeiry N, Vaziri R. A stacked sublaminate-based damage-plasticity
Experiments and fe model. Compos Part B Eng 2020;192:. https://doi.org/ model for simulating progressive damage in composite laminates under impact
10.1016/j.compositesb.2020.107950. URL: http:// loading. Thin-Wall Struct 2020;156:. https://doi.org/10.1016/j.tws.2020.107009.
www.sciencedirect.com/science/article/pii/S135983681935783X107950. URL: http://www.sciencedirect.com/science/article/pii/
[14] McGregor C, Zobeiry N, Vaziri R, Poursartip A, Xiao X. Calibration and validation S0263823120308843107009.
of a continuum damage mechanics model in aid of axial crush simulation of [31] Reiner J. A practical approach for the non-local simulation of progressive damage
braided composite tubes. Compos Part A Appl Sci Manuf 2017;95:208–19. https:// in quasi-isotropic fibre-reinforced composite laminates. Compos Struct
doi.org/10.1016/j.compositesa.2017.01.012. 2021:113761. doi:https://doi.org/10.1016/j.compstruct.2021.113761. URL:
[15] Wang J-X, Wu J-L, Xiao H. Physics-informed machine learning approach for https://www.sciencedirect.com/science/article/pii/S0263822321002221..
reconstructing reynolds stress modeling discrepancies based on dns data. Phys Rev [32] Reiner J, Xu X, Zobeiry N, Vaziri R, Hallett SR, Wisnom MR. Virtual
Fluids 2(3). doi:10.1103/physrevfluids.2.034603. URL: https://doi.org/10.1103/ characterization of nonlocal continuum damage model parameters using a high
PhysRevFluids.2.034603.. fidelity finite element model. Compos Struct 2021;256:. https://doi.org/10.1016/
[16] Wang J, Wu J, Xiao H. A physics-informed machine learning approach of j.compstruct.2020.113073. URL: https://www.sciencedirect.com/science/article/
improving RANS predicted reynolds stresses. arXiv:https://arc.aiaa.org/doi/pdf/ pii/S0263822320329998113073.
10.2514/6.2017-1712, URL: https://arc.aiaa.org/doi/abs/10.2514/6.2017-1712.. [33] Nagaraj M, Reiner J, Vaziri R, Carrera E, Petrolo M. Progressive damage analysis of
[17] Zobeiry N, Humfeld KD. An iterative scientific machine learning approach for composite structures using higher-order layer-wise elements. Compos Part B Eng
discovery of theories underlying physical phenomena. CoRR abs/1909.13718. url: 2020;190:. https://doi.org/10.1016/j.compositesb.2020.107921107921.
http://arxiv.org/abs/1909.13718. [34] Nagaraj M, Reiner J, Vaziri R, Carrera E, Petrolo M. Compressive damage
[18] Belianinov A, He Q, Kravchenko M, Jesse S, Borisevich A, Kalinin SV. modeling of fiber-reinforced composite laminates using 2d higher-order layer-wise
Identification of phases, symmetries and defects through local crystallography. models. Compos Part B Eng 2021;215:. https://doi.org/10.1016/
Nat Commun 2015;6:7801. https://doi.org/10.1038/ncomms8801. j.compositesb.2021.108753. URL: https://www.sciencedirect.com/science/
[19] Aspuru-Guzik A, Persson K. Materials acceleration platform: accelerating advanced article/pii/S1359836821001451108753.
energy materials discovery by integrating high-throughput methods and artificial [35] Forghani A, Vaziri R. Computational modeling of damage development in
intelligence; 2018. URL: http://nrs.harvard.edu/urn-3:HUL.InstRepos:35164974. composite laminates subjected to transverse dynamic loading. J Appl Mech 76
[20] Zobeiry N, Stewart A, Poursartip A. Applications of machine learning for process (5):051304. doi:10.1115/1.3129705..
modeling of composites. In: SAMPE virtual conference, Seattle, WA; 2020.. [36] Bažant ZP, Oh BH. Crack band theory for fracture of concrete. Matér Constr
[21] Zobeiry N, Humfeld KD. A physics-informed machine learning approach for 1983;16(3):155–77. https://doi.org/10.1007/BF02486267.
solving heat transfer equation in advanced manufacturing and engineering [37] Chen J, Jönsson P, Tamura M, Gu Z, Matsushita B, Eklundh L. A simple method for
applications; 2020. arXiv:2010.02011.. reconstructing a high-quality ndvi time-series data set based on the savitzky–golay
[22] Manohar K, Hogan T, Buttrick J, Banerjee AG, Kutz JN, Brunton SL. Predicting filter. Remote Sens Environ 2004;91(3):332–44. https://doi.org/10.1016/j.
shim gaps in aircraft assembly with machine learning and sparse sensing. J Manuf rse.2004.03.014. URL: http://www.sciencedirect.com/science/article/pii/
Syst 2018;48:87–95, special Issue on Smart Manufacturing. doi: 10.1016/j. S003442570400080X.
jmsy.2018.01.011. URL: https://www.sciencedirect.com/science/article/pii/ [38] Kim M, Zobeiry N. Machine learning for reduced-order modeling of composites
S0278612518300116.. processing. In: SAMPE Virtual Conference, Long Beach, CA; 2021..
[23] Sacco C, Baz Radwan A, Anderson A, Harik R, Gregory E. Machine learning in [39] Humfeld K, Zobeiry N. Machine learning-based process simulation approach for
composites manufacturing: A case study of automated fiber placement inspection. real-time optimization and active control of composites autoclave processing. In:
Compos Struct 2020;250:. https://doi.org/10.1016/j.compstruct.2020.112514. SAMPE Virtual Conference, Long Beach, CA; 2021..
URL: https://www.sciencedirect.com/science/article/pii/ [40] Zobeiry N. Extracting the strain-softening response of composites using full-field
S0263822320313659112514. displacement measurement. Ph.D. thesis, University of British Columbia, Canada.
[24] Sause MG, Schmitt S, Kalafat S. Failure load prediction for fiber-reinforced 2010..
composites based on acoustic emission. Compos Sci Technol 2018;164:24–33. [41] Xu X, Wisnom MR, Hallett SR, Zobeiry N, Leslie S, Poursartip A, Vaziri R. Stacking
https://doi.org/10.1016/j.compscitech.2018.04.033. URL: https:// sequence effects in over-height compact tension tests of quasi-isotropic laminates.
www.sciencedirect.com/science/article/pii/S0266353817331627. In: 19th International Conference on Composite Materials; 2013. URL: http://
[25] Patel DK, Parthasarathy T, Przybyla C. Predicting the effects of microstructure on www.iccm-central.org/Proceedings/ICCM19proceedings/..
matrix crack initiation in fiber reinforced ceramic matrix composites via machine [42] Cherniaev A, Butcher C, Montesano J. Predicting the axial crush response of cfrp
learning. Compos Struct 2020;236:. https://doi.org/10.1016/ tubes using three damage-based constitutive models. Thin-Wall Struct
j.compstruct.2019.111702111702. 2018;129:349–64. https://doi.org/10.1016/j.tws.2018.05.003. URL: http://
[26] Califano A, Chandarana N, Grassia L, D’Amore A, Soutis C. Damage detection in www.sciencedirect.com/science/article/pii/S0263823118301964.
composites by artificial neural networks trained by using in situ distributed strains. [43] David M, Johnson AF. Effect of strain rate on the failure mechanisms and energy
Appl Compos Mater 2020;27:657–71. https://doi.org/10.1007/s10443-020- absorption in polymer composite elements under axial loading. Compos Struct
09829-z. 2015;122:430–9. https://doi.org/10.1016/j.compstruct.2014.11.010. URL:
[27] Zobeiry N, Reiner J, Vaziri R. Theory-guided machine learning for damage http://www.sciencedirect.com/science/article/pii/S0263822314005790.
characterization of composites. Compos Struct 2020;246:. https://doi.org/ [44] Waimer M, Siemann M, Feser T. Simulation of cfrp components subjected to
10.1016/j.compstruct.2020.112407. URL: http:// dynamic crash loads. Int J Impact Eng 2017;101:115–31. https://doi.org/
www.sciencedirect.com/science/article/pii/S0263822320301756112407. 10.1016/j.ijimpeng.2016.11.011. URL: http://www.sciencedirect.com/science/
[28] Xu X, Paul A, Sun X, Wisnom MR. An experimental study of scaling effects in article/pii/S0734743X16301154.
notched quasi-isotropic carbon/epoxy laminates under compressive loads. Compos [45] Schueler D, Toso-Pentecôte N, Voggenreiter H. Simulation of high velocity impact
Part A Appl Sci Manuf 2020;137:. https://doi.org/10.1016/ on composite structures - model implementation and validation. Appl Compos
j.compositesa.2020.106029. URL: http://www.sciencedirect.com/science/ Mater 2016;23(4):857–78. https://doi.org/10.1007/s10443-016-9489-0.
article/pii/S1359835X20302682106029.

11