Optics and Lasers in Engineering 146 (2021) 106689

Contents lists available at ScienceDirect

Optics and Lasers in Engineering

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

A new methodology for thermoelastic model identification in composite

materials using digital image correlation
Francisco de Sá Rodrigues a, Ricardo Marques b, Isa Emami Tabrizi c,d,e, Adnan Kefal c,d,e,
Hafiz Qasim Ali c,d,e, Mehmet Yildiz c,d,e,∗, Afzal Suleman b,f
a
Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais, Lisbon 1049-001, Portugal
b
Department of Mechanical Engineering, University of Victoria, Victoria, BC, Canada
c
Integrated Manufacturing Technologies Research and Application Center, Sabanci University, Orhanli-Tuzla, Istanbul 34956, Turkey
d
Composite Technologies Center of Excellence, Sabanci University-Kordsa, Istanbul Technology Development Zone, Sanayi Mah. Teknopark Blvd. No: 1/1B, Pendik,
Istanbul 34906, Turkey
e
Faculty of Engineering and Natural Sciences, Materials Science and Nano Engineering, Sabanci University, Orhanli-Tuzla, Istanbul 34956, Turkey
f
IDMEC, Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais, No. 1, Lisboa 1049-001, Portugal

a r t i c l e i n f o a b s t r a c t

Keywords: Local strain measurement techniques, such as strain gauges or extensometers, have been broadly utilized as input
Thermoelastic stress analysis data source for thermoelastic effect model identification in fiber-reinforced polymers, despite their well-known
Digital image correlation high heterogeneity. This experimental setup strongly limits the possibility of assigning thermoelastic models in
Carbon fiber-reinforced polymers
a local-based manner for posterior Thermoelastic Stress Analysis. This issue has been addressed herein through
Structural health monitoring
proposing a novel method for spatial identification of thermoelastic models in composite structures using full-field
Smoothing element analysis
experimental measurements. The proposed concept is validated by conducting tests on laminated tensile coupons
with various stacking sequences. To this end, the displacements and thermal data collected from Digital Image
Correlation and infrared camera, respectively, are interpolated to a mutual background mesh using a Smoothing
Element Analysis. It is shown that this procedure results in a continuous strain field and a reconstructed thermal
map for the laminate domain at various loading stages. The resulting smoothed strain map is used as the input for
three different thermoelastic models, followed by a comparison of the calculated analytic temperature variations
for each model against the infrared camera’s measurements. The relative error associated to this model assignment
process is underlined, and it is revealed a significant effect of material properties variability in the accuracy
of the method. It is shown that the proposed methodology can circumvent inaccuracies of the conventional
Thermoelastic Stress Analysis method, while providing a viable and computationally efficient method for the
selection of appropriate thermoelastic models at the local level.

1. Introduction tinual harsh operating conditions. In order to assess the condition of

aerospace structures, the current approach involves scheduled mainte-
Carbon and glass fiber-reinforced polymers (CFRP and GFRP, respec- nance resorting to Non-Destructive Inspections (NDIs) and preventive
tively) have been gaining increasing usage in the aerospace industry replacements if needed. However, the scheduled maintenance has been
due to their higher strength-to-weight ratios, as well as improved cor- reported to implicate considerable cumulative costs. To address this is-
rosion and fatigue resistances in comparison to metals and alloys. Fur- sue, several Structural Health Monitoring (SHM) techniques have been
thermore, due to their ability of being processed into complex shapes, proposed for the in-operation assessment of the stress/strain states of
fiber-reinforced composite materials lend themselves excellently to the the material through the employment of complex systems. These would
manufacture of aircraft structures with aerodynamically-optimized chal- acquire and process data continuously whereby the damage location can
lenging geometries [1]. To this end, they have become the material of be identified and its severity quantified.
choice in aircraft structures subjected to high loading and frequency Among the existing SHM techniques, optical methods such as Ther-
levels [2]. moelastic Stress Analysis [3] (TSA) and Digital Image Correlation
Aircraft structures suffer a decrease in their operational lifetime due [4] (DIC) have stood out for their non-contact and non-destructive char-
to the initiation and evolution of damage mechanisms caused by con- acteristics in providing the full-field deformation of solid materials un-

∗
Corresponding author at: Integrated Manufacturing Technologies Research and Application Center, Sabanci University, Orhanli-Tuzla, Istanbul 34956, Turkey.
E-mail addresses: mehmet.yildiz@sabanciuniv.edu (M. Yildiz), suleman@uvic.ca (A. Suleman).

https://doi.org/10.1016/j.optlaseng.2021.106689
Received 1 December 2020; Received in revised form 1 April 2021; Accepted 10 May 2021
Available online 2 June 2021
0143-8166/© 2021 Elsevier Ltd. All rights reserved.
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

der loading. TSA is an experimental surface analysis technique based suiting TSA model to quantify the strain distribution of a loaded struc-
on the thermoelastic effect, which refers to the temperature variations ture.
detected in elastic materials due to volume changes. The increasing Unlike metals, laminated composite materials contain several micro-
research interest in TSA has accompanied the technological advance constituents with distinct thermal-mechanical properties. These con-
and commercial availability of infrared cameras witnessed over the stituents can be brought together in a wide range of possibilities de-
last 30 years. Recent efforts in the development of experimental se- pending on stacking sequence and manufacturing techniques employed,
tups composed of in-house developed signal processing systems and mi- thereby leading to macro-structures which can possess different struc-
crobolometers for thermal data acquisition aim at offering more cost- tural and thermal attributes. In this regard, determining the thermal
effective and widely-accessible solutions for TSA [5,6]. The typical out- source responsible for the IR camera’s output becomes a prerequisite
put of TSA includes temperature change (Δ𝑇 ), absolute temperature at for ameliorating the results of TSA on composite materials. The conven-
the mean load level (𝑇0 ), and the phase angle (𝜙) which corresponds to tional methodology for selecting the thermoelastic model is based on
the phase shift between the detected thermal and applied load signals. measuring the local strain change (Δ𝜖𝑥 and Δ𝜖𝑦 ) through the usage of
Under small reversible deformations and adiabatic conditions, the linear biaxial strain gauges or extensometers mounted on the central region
relationship between the small detected Δ𝑇 and mechanical deforma- of the specimen under test. These strain measurements are regarded as
tion of the material can be analytically quantified using a thermoelastic representative of the average deformation of the material subjected to
effect model (hereinafter referred to as thermoelastic model). Thomson loading condition. For each available thermoelastic model, the temper-
[7] first derived the thermoelastic model valid for isotropic materials, ature ratio Δ𝑇 ∕𝑇0 is calculated using the acquired strain as input, here-
inafter, referred to as theoretical. Under the same loading condition, Δ𝑇
and 𝑇0 are recorded by the IR camera and their ratio is defined as ex-
𝛼 ∑
3
Δ𝑇
=− Δ𝜎 (1) perimental. The model whose theoretical predictions yields the smallest
𝑇0 𝜌𝐶𝑝 𝑘=1 𝑘
relative error with respect to the experimental one is selected for strain
quantification in TSA. To the authors’ perspective, this process simply
where Δ𝜎𝑘 is the kth principal component of stress, 𝜌 is the density, 𝛼 is relies on the average strain measurements thus neglecting the pixel-wise
the coefficient of thermal expansion (CTE), 𝑇0 is the absolute tempera- and full-field natures of TSA. In heterogeneous materials such as FRPs,
ture and 𝐶𝑝 is the specific heat capacity. The ratio 𝛼∕𝜌𝐶𝑝 is commonly this assumption may lead to under/overestimation of the mechanical
referred to as thermoelastic constant. deformation in critical regions, which might result in safety issues for
In the current literature, the scope of TSA in fiber-reinforced poly- the aerospace industry.
mers (FRPs) includes damage assessment [8–10] and investigation of Comparison of experimental data acquired from cameras with dif-
the thermoelastic signal’s source [11–14]. The term thermoelastic signal’s ferent resolutions requires interpolation of the respective outputs into
source refers to the element of the meso or macro structure of the com- a common mesh. The Smoothing Element Analysis (SEA) has presented
posite (i.e., surface ply, outer epoxy layer or laminate) which contributes accurate results for stress predictions by smoothing and interpolating
to the measured surface temperature. Additionally, TSA has recently discontinuous stress fields into 𝐶 1 -continuous stress fields with 𝐶 0 -
been also used for predicting the remaining useful life (RUL) of quasi- continuous stress [20] by minimization of a penalized-discrete-least-
isotropic GFRP laminates with a novel methodology based on the strain squares function by comparing discrete finite element stresses with con-
energy density [15]. Emery et al. [16] evaluated the strain field distribu- tinuous recovered stresses. Moreover, the interpolated continuous form
tion in the vicinity of three types of damage mechanisms: delamination, of stress gradients allows for assessment of equilibrium satisfaction in
fiber breakage and matrix cracking for different stacking sequences of the domain. This technique was further improved by considering a four-
unidirectional pre-impregnated E-glass epoxy laminates. Pitarresi and node element [21] instead of a five-node macro-element composed of
Galietti [17] investigated the capability of two different thermoelastic four three-node triangular elements. This newly considered element pro-
models, that is, the Bulk and Resin-Rich Layer (RRL) models for rep- vided equivalent results to the original three-node element, whilst pro-
resenting the CFRP’s thermoelastic response. Their study initially indi- viding considerable decreases in the computational cost.
cated that these two models (which need predetermined material prop- The main objective of this work is to propose a novel experimental
erties) were not able to accurately reproduce experimentally obtained method for identifying the most suitable models for thermoelastic mea-
thermoelastic signals. After accounting for a possible variability of the surements in composite structures by replacing the strain input data
fiber’s CTE between positive and negative values, the Bulk model stood from either strain gauges or extensometers with full-field strain data
out as representative of the CFRP’s thermoelastic signal regardless of the calculated from DIC displacements using SEA. This new methodology
stacking sequence. The argument proposed by the authors to justify the addresses the limitations of the conventional thermoelastic model iden-
poor agreement between the RRL model and the experimental results tification approach in capturing spatial variations of material distribu-
lies in the strong mismatch between the resin’s and fiber’s CTEs. tion by exploring the full-field capabilities of both TSA and DIC, hence
TSA and DIC have been jointly used by different authors to compen- contributing to minimize errors in strain quantification. Here, full-field
sate for the unstable motion caused by high frequency dynamic load- 2D displacement changes between two load levels obtained from DIC are
ing during thermoelastic measurements. Silva et al. [18] performed DIC smoothed by employing SEA and utilized to calculate continuous strains
and TSA on 2024 aluminum alloy and Nylon 6 samples using a single on a background mesh. Analogously, amplitude and mean temperature
Charged Coupled Device (CCD) camera. Both datasets were resultant values obtained from the IR camera between the same load levels are
from the same IR images and missing data due to unpainted regions (sec- also smoothed and reconstructed on the same background mesh. The
tions without speckle pattern) were generated through an interpolation whole strain data set is consecutively used as an input for the different
from neighboring dataset. Wang et al. [19] stressed that, even though thermoelastic model equations to calculate analytic temperature vari-
the work by Silva et al. [18] presented several benefits, the technolog- ations. Afterwards, these were rigorously compared to the data set of
ical state-of-the-art of the involved IR detectors provided limited spa- the experimental temperature variations at the local level of the com-
tial resolutions, hence decreasing the accuracy of the displacement field posite to identify the best thermoelastic model as a function of spatial
calculated from DIC. To correct these weaknesses, Wang et al. [19] em- coordinates.
ployed two separate camera systems with respective position markers to This work starts by introducing a theoretical background for both
address the face sheet debonding problem in a double cantilever beam DIC and TSA, followed by a description of the manufacturing and cut-
(DCB). To the authors knowledge, no attempt has been made to em- ting processes involved in the preparation of epoxy and composite sam-
ploy both TSA and DIC in combination for identifying the thermoelastic ples for testing. The next section firstly provides a general overview of
signal’s source, which is conspicuously critical for determining the best the comparison process between TSA and DIC data, followed by a thor-

2
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

ough description of the experimental implementation of the proposed 2.2. Thermoelastic Stress Analysis (TSA) of composite materials
methodology. The later encompasses test methods, equipment and test-
ing conditions for characterization of thermal and mechanical properties Eq. (1) has been further extended to orthotropic material cases where
of the tested materials, DIC and TSA. The relevant mathematical back- Δ𝑇 ∕𝑇0 is a function of the inner product between CTE and stress change,
ground of SEA is next presented. Finally, the results of the thermoelastic Δ𝜎, represented in a vector multiplication format as
signal’s source identification process are discussed in two sections con-
Δ𝑇 1 1
sidering constant and variable CTE values. =− [𝛼]𝑇 [Δ𝜎]1,2 = − [𝛼]𝑇 [Δ𝜎]𝑥,𝑦 . (4)
𝑇0 𝜌𝐶𝑝 1,2 𝜌𝐶𝑝 𝑥,𝑦

2. Theoretical background Here, the subscripts (𝑥, 𝑦) and (1,2) respectively refer to the global (i.e.
parallel and perpendicular to the applied loading direction) and local
2.1. Digital Image Correlation (DIC) (i.e. parallel and perpendicular to the fiber’s direction) coordinate sys-
tems. However, there’s no readily available direct stress measurement
DIC is a contactless SHM technique which allows for the evaluation technique. Instead, stresses at the lamina/resin-rich layer levels must be
of full-field displacements through the acquisition, storage and analysis computed using geometrical and material parameters as input. Given
of a set of images from a loaded object using CCD sensors in digital cam- that lamina stresses in either local or global coordinate systems of a
eras. Motion estimation is done by monitoring and identifying changes symmetric multidirectional laminate subjected to in-plane loading con-
in the surface pattern through the comparison of the acquired images ditions vary as a function of the fiber orientation, the stress calculation
with a reference stage (typically an unloaded configuration). This pro- process in inner layers of the composite can yield additional inaccura-
cess relies on the assumption of object continuity, meaning that the dif- cies. Alternatively, under the same laminate and loading configurations,
ferences detected in the sensor’s plane also occur in the object. To enable every lamina deforms under the same in-plane strain field represented
the tracking of individual material points, the surface must be covered in the global coordinate system. Due to the numerous available surface
with a random distribution of gray dots on a white background to in- strain sensing technologies, it is deemed more convenient to rewrite
crease visual contrast, referred to as speckle pattern. DIC can be classi- the thermoelastic equations as a function of strains rather than stresses.
fied as two- or three-dimensional depending on the number of cameras Hence, under a linear elastic behavior assumption, the orthotropic ther-
employed for motion tracking (one and two, respectively). Recalling the moelastic formulation in Eq. (4) can be rewritten in the local coordinate
thermoelastic model for isotropic materials of Eq. (1), TSA can correlate system as
material temperature changes with the first principal stress invariant. Δ𝑇 1
Given a composite laminate subjected to in-plane loading conditions =− [𝛼]𝑇 [𝑄]1,2 [Δ𝜖]1,2 (5)
𝑇0 𝜌𝐶𝑝 1,2
whose thickness is much smaller than its in-plane dimensions, plane-
stress conditions can be assumed (i.e. 𝜎3 = 𝜎𝑧 = 0). Moreover, as stated where [𝑄]1,2 is the lamina’s reduced stiffness matrix. Eq. (5) can be also
in Section 2.2, the in-plane principal stress directions are independent formulated in the global coordinate system as
of 𝜀3 . Therefore, 2D-DIC is sufficient for fulfilling the current analysis Δ𝑇 1
=− [𝛼]𝑇 [𝑄]1,2 [𝑇 ]𝜖 [Δ𝜖]𝑥,𝑦 (6)
requirements. In-plane deformations are monitored with sensors whose 𝑇0 𝜌𝐶𝑝 1,2
axes are perpendicular to the sample’s surface, provided the material is
under plane stress conditions and no sudden change of lightning occurs where [𝑇 ]𝜖 is the transformation matrix for the strain components.
which might lead to inaccurate subset matching [4]. Image correlation For the case of balanced and symmetric laminates, the open form of
is performed by quantifying the intensity signal of the gray distribution Eq. (6) can be written as
according to a color gradient ranging from black to white, and limited Δ𝑇
= − 𝜌𝐶1 [(𝛼1 𝑄11 + 𝛼2 𝑄12 )(Δ𝜖𝑥 cos2 (𝜃) + Δ𝜖𝑦 sin2 (𝜃))+
𝑇0 𝑝 (7)
by the number of bits available in the sensor. Subset matching is per-
(𝛼1 𝑄12 + 𝛼2 𝑄22 )(Δ𝜖𝑥 sin2 (𝜃) + Δ𝜖𝑦 cos2 (𝜃))]
formed by evaluating a given region of interest (ROI) of the acquired
images, prior and after deformation, through a correlation function. The where it is assumed that shear stresses, 𝜏𝑥𝑦 = 0 for uniaxial loading con-
correlation function 𝐶(𝑥, 𝑦) used in this study is the Normalized Cross ditions, i.e. 𝜎𝑥 ≠ 0, and 𝜃 is the fiber angle. This equation is named as
Correlation [22] formulated as the Bulk model where the laminate’s thermal response is generated at
𝑁 ∑
∑ 𝑀
𝐼 (𝑥𝑖 , 𝑦𝑗 )𝐼 ′ (𝑥′𝑖 , 𝑦′𝑖 )) the orthotropic surface ply.
𝐶(𝑥, 𝑦) = 1 − 1
(2) In the course of manufacturing a composite laminate, an outer thin
𝑖=1 𝑗=1 𝐼 2 (𝑥𝑖 , 𝑦𝑗 )𝐼 ′2 (𝑥′𝑖 , 𝑦′𝑖 ) 2 layer of resin material forms on the surface of the laminate. If the com-
posite with a sufficiently thick resin layer is subjected to high load-
where (𝑥𝑖 , 𝑦𝑗 ) and (𝑥′𝑖 , 𝑦′𝑖 ) are the locations of the subset prior and after
ing frequencies, the resin layer can deform in similitude with the or-
deformation, respectively, and their correspondent continuously inter-
thotropic substrate while leading to heat generation that can be mea-
polated discrete intensity pattern values are given by I and I’.
sured by IR camera. Due to the high frequency of the applied load (re-
An ideal subset matching is given by the minimization of the function
quired for ensuring quasi-adiabatic conditions) as well as the low ther-
given in Eq. (2). Displacements are estimated afterwards through differ-
mal conductivity of the resin layer, the heat generated within the inner
ent order shape functions. In this work, a second order shape function
layers (substrate) can not diffuse into the most outer resin layer [23].
is used,
Under the aforementioned conditions, it is prudent to assume that the
1 1 1
𝑢(𝑥, 𝑦) = 𝑢 + 𝑢𝑥 Δ𝑥 + 𝑢𝑦 Δ𝑦 + 𝑢 Δ𝑥2 + 𝑢𝑦𝑦 Δ𝑦2 + 𝑢𝑥𝑦 Δ𝑦Δ𝑥 (3a) epoxy acts as a strain witness of the entire laminate, i.e. Δ𝜖𝑥𝑐 = Δ𝜖𝑥𝑟 and
2 𝑥𝑥 2 2 Δ𝜖𝑦𝑐 = Δ𝜖𝑦𝑟 = −𝜈𝑟 Δ𝜖𝑥𝑟 , where the subscripts c and r stand for composite
and resin, respectively. These assumptions were first described mathe-
1 1 1 matically by Pitarresi et al.[24] in the Resin-Rich layer (RRL) model.
𝑣(𝑥, 𝑦) = 𝑣 + 𝑣𝑥 Δ𝑥 + 𝑣𝑦 Δ𝑦 + 𝑣 Δ𝑥2 + 𝑣𝑦𝑦 Δ𝑦2 + 𝑣𝑥𝑦 Δ𝑦Δ𝑥 (3b)
2 𝑥𝑥 2 2 Using the stress-based thermoelastic formulation given in Eq. (1), one
where 𝑢(𝑥, 𝑦) and 𝑣(𝑥, 𝑦) are the longitudinal and transverse displace- can write
ments, respectively, Δ𝑥 and Δ𝑦 refer to the distance between the subset
Δ𝑇 𝛼 [ 𝐸𝑟 ]
center and a given point in the subset. First partial derivatives are rep- =− 𝑟 (Δ𝜖𝑥𝑐 + Δ𝜖𝑦𝑐 ) (8)
𝑇0 𝜌𝑟 𝐶𝑝𝑟 1 − 𝜈𝑟
resented with subscripts 𝑥 and 𝑦, while second partial derivatives can be
identified with the subscripts 𝑥𝑥, 𝑦𝑦 and 𝑥𝑦. Accurate displacement cal- Pitarresi et al. [24] investigated this hypothesis on GFRPs by conduct-
culations rely on computing 𝑢, 𝑢𝑥 , 𝑢𝑥𝑥 , 𝑢𝑦 , 𝑢𝑦𝑦 , 𝑢𝑥𝑦 , 𝑣, 𝑣𝑥 , 𝑣𝑥𝑥 , 𝑣, 𝑣𝑦 , 𝑣𝑦𝑦 , 𝑣𝑥𝑦 ing TSA using SPATE and Deltatherm systems on load-controlled unidi-
through the Newton-Raphson method. rectional plain weave E-glass/epoxy samples. They indicated that the

3
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Fig. 1. Source of thermoelastic signal for each

mathematical model.

results of RRL model are in better agreement with experimentally-

obtained TSA results. Effectively, the isotropic nature of epoxy resin en-
ables the construction of mathematically simple strain witness or RRL
model which has been proven to be extremely practical in TSA of or-
thotropic materials.
Another proposed model in the literature assumes heat diffusion to
occur between laminae, thus leading to a condition of thermal equilib-
rium attained at the macro rather than meso scale. The TSA equations
are now formulated by regarding the multi-layered structure as a sin-
gle lamina where adiabatic conditions still prevail, and the thermoe-
lastic response is now quantified based on the mechanical and thermal
properties of the laminate. This model denoted as Homogeneous can be
formulated using Eq. (7) as a starting point,
Δ𝑇 1 [ ]
=− 𝛼 (𝐴 Δ𝜖 + 𝐴12 Δ𝜖𝑦 ) + 𝛼𝑦𝐿 (𝐴12 Δ𝜖𝑥 + 𝐴22 Δ𝜖𝑦 ) (9)
𝑇0 𝜌𝐶𝑝 ℎ 𝑥𝐿 11 𝑥

where [𝐴] is the extensional stiffness matrix (with units of [N/m]), 𝛼𝑥

and 𝛼𝑦 are the laminate’s coefficients of thermal expansion measured
with respect to the loading reference axes, and ℎ is the laminate’s thick-
ness. Eqs. (7) and (9) are equivalent for stacking sequences consisting
of equally-oriented layers.
The control volume considered in the derivation of each thermoelas-
tic model is schematically represented in Fig. 1.

3. Materials

3.1. Manufacturing of composite laminates

Unidirectional carbon prepreg composed of carbon fiber T700 and

epoxy with the commercial code OM10 were purchased from Kordsa
Company with the code ”KOM10 T700 12KT UD300 37 600 KOMP”.
The fiber area weight of the prepregs is 300 g/m2 and the resin content
is 37% weight ratio. Layers of carbon fiber prepreg with dimensions
of 300 × 300 mm were stacked by hand lay-up on a steel mold whose
surface was cleaned beforehand with ethanol and covered with sealing
and releasing agents. To ensure that the resin’s curing cycle was prop-
erly followed, a thermocouple was attached to the lower surface of the
laminate and connected to a thermometer. Epoxy was cured in an ”MSE
Teknoloji LTD” hot-press using the following procedure: (i) the press
was initially heated up to a temperature of 80 ◦ C, (ii) the steel mold Fig. 2. DIC-based thermoelastic signal’s source identification process.
(containing the stacked prepregs) was placed inside the press, (iii) a

4
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Table 1 Full-field displacement and thermal data are stored in a pixel-wise

Mechanical and thermal properties of epoxy manner. However, depending on each camera’s resolution, the same ma-
resin and CFRP lamina. terial geometry can be represented by two distinct sets of pixels. There-
Property CFRP lamina OM 10 Epoxy fore, data comparison at each material point is only made possible if
one interpolates both final datasets into a common background mesh
E1 [GPa] 128.382 -
E2 [GPa] 8.284 -
by using a SEA algorithm. This process is detailed in Section 5.
G12 [GPa] 3.58 - The TSA-derived temperature ratio after interpolation, (Δ𝑇 ∕𝑇0 )TSA
𝜈 12 0.3 - will be hereinafter referred to as experimental. On the other hand, the
𝜈 21 0.019 - three different thermoelastic models introduced in Eqs. (7), (8) and
𝜌 [kg/m3 ] 1516.3 -
(9) provided theoretical temperature ratios, (Δ𝑇 ∕𝑇0 )k (where 𝑘 stands for
Cp [J/kg.K] 763.632 -
𝛼1 [K-1 ] 0.29 × 10-6 - Bulk, RRL or Homogeneous), using the full-field strains by subsequent
𝛼2 [K-1 ] 27.4 × 10-6 - interpolation/calculation of DIC’s full-field displacement. For each loca-
𝜈𝑓 (%) 55.95 - tion in the background mesh domain, both theoretical and experimen-
𝜈𝑚 (%) 41.87 -
tal datasets were known. Hence, the percent error of the thermoelastic
𝜈𝑣 (%) 2.177 -
Er [GPa] - 3.583
model 𝑘’s prediction relative to the experimentally-determined temper-
𝜈𝑟 - 0.39 ature ratio, 𝛿k , could be computed as
Cpr [J/kg.K] - 841
|(Δ𝑇 ∕𝑇0 )k − (Δ𝑇 ∕𝑇0 )TSA |
𝜌𝑟 [kg/m3 ] - 1142 𝛿k (%) = ( × 100%. (10)
𝛼𝑟 [K-1 ] - 58 × 10-6 𝑚𝑎𝑥 (Δ𝑇 ∕𝑇0 ))TSA
(
where 𝑚𝑎𝑥 (Δ𝑇 ∕𝑇0 ))TSA corresponds to the maximum value for the tem-
perature ratio obtained throughout the coupon domain.
first pressure level of 3 bar was applied for 5 mins at 80 ◦ C, (iv) temper- The thermoelastic model yielding the lowest 𝛿k was further con-
ature was gradually increased up to 120 ◦ C while keeping pressure at sidered to best characterize the local thermoelastic behavior. After re-
3 bar, (v) pressure was increased to 7.5 bar, (vi) pressure (7.5 bar) and peating this model assignment process for every pixel in the back-
temperature (120 ◦ C) conditions were maintained steady for 60 mins, ground mesh domain, the thermoelastic signal’s identification process
and (vii) the laminate was cooled down to ambient temperature at a was deemed completed. The aforementioned steps comprising this anal-
gradual step of 10 ◦ C/min so that no residual stresses could be present ysis are summarized in a flowchart illustrated in Fig. 2
in the composite material. Four different stacking sequences were man-
ufactured using this process: Unidirectional [0]8 , Cross-Ply [0/902 /0]s , 4.2. Mechanical and thermal characterization of CFRP laminate
Cross-Ply [90/02 /90]s and Quasi-Isotropic [02 /-452 /452 /902 ]s . A more
simplified notation for each stacking sequence is introduced in Table 2. For the determination of the longitudinal and transverse modulus,
five UD and five specimens with fibers aligned with the transverse di-
3.2. Manufacturing of epoxy specimens rection ([90]6 ) were cut from two different plates in accordance with
ASTM D3039/D3039M standards (250 x 15 x 2 mm and 175 x 25 x
Kordsa’s prepreg is composed of a high thoughness, opaque resin for 1.5 mm, respectively). 50 x 15 x 2 mm GFRP tabs with a [±45]2𝑠 stack-
manufacturing of structural composites. In order to determine the me- ing sequence were bonded to the tensile specimens using epoxy Araldite
chanical properties of the neat matrix, with the comercial code KOM10, 2011 and left for curing at room temperature for 24 h. An Instron 5983
samples were prepared through casting in a metal mold. Epoxy resin, servo-hydraulic machine with a load cell capacity of ±50 kN was used
hardener and accelerator were provided by the manufacturer at appro- for conducting the tensile tests on the specimens at a displacement rate
priate ratios to totally obtain 500 g of unmixed material. Mixing of the of 2 mm/min. The value of the Poisson’s ratio (𝜈12 ) was determined us-
constituents was performed using THINKY planetary vacuum mixer for ing a biaxial strain gauge attached to the samples’ surface during the
10 min at 200 rpm to assure an homogeneous and degassed mixture. The tensile tests. Since the ultimate tensile strain of the [90]6 specimen was
surface of aluminum molds, specifically designed to obtain finished sam- lower than the minimum detectable strain by the biaxial extensometer,
ples according to ASTM D790 and ISO 604, was cleaned using sandpaper the result for 𝜈21 was theoretically calculated using the symmetry re-
and acetone before being preheated to 80◦ C for 30 min. The prepared lation between the lamina’s reduced stiffness entries, 𝑄12 and 𝑄21 , i.e.
mixture was transferred into the mold under vacuum conditions and 𝐸1 𝜈21 = 𝐸2 𝜈12 .
placed inside the oven to undergo the same curing profile as the CFRP For determining the density of the manufactured laminates, five sam-
prepreg material. After removal from the molds, neat samples were cut ples with the dimensions of 25 x 25 x 1.5 mm were weighted using a
according to standards and tested accordingly. Shimadzu AUX220 scale. The density was determined following a hy-
drostatic weighing technique, which is based on the load equilibrium
4. Experimental implementation before and after the immersion of a sample in water [15]. The val-
ues for 𝐶𝑝 were determined using a Differential Scanning Calorimeter
4.1. Methodology (Mettler Toledo DSC 3+ model) while following guidelines provided
in ASTM E1269-11 standard. Five small pieces of the laminate were
Under the linear elastic behavior assumption, the undamaged mate- cut and subjected to a temperature increase of 20 ◦ C/min and the val-
rial deformation is rate independent, i.e. the displacement field attained ues obtained for the specimen’s heat flux absorption were then com-
at any load level will be the same regardless of the loading speed. While pared with the ones from sapphire at the same temperature. The CTE of
TSA is limited to dynamic testing at high frequencies, this material be- unidirectional composite materials along the principle directions were
havior allows for DIC to be utilized under quasi-static tensile test condi- obtained through a TMA (Thermomechanical Analysis) using a Mettler
tion without compromising the validity of the proposed methodology if Toledo TMA/SDTA 1 STARe system in accordance with the ASTM E831-
the test is bounded by the same load levels as TSA’s. Among the advan- 19. Twenty nine samples with dimensions of 5 x 5 x 2 mm were cut and
tages, one can mention the significant reduction in stored data, as well then exposed to a slow temperature increase of 0.5 ◦ C/min to allow tem-
as improvements in image stability and resolution, the latter resulting perature/strain homogenization throughout the sample. In this device,
in higher image quality and number of data acquisition points. Conse- a probe was leaned to the sample’s surface for displacement measure-
quently, DIC and TSA on samples under load control were performed ment. From the statistical representation of data using an histogram, the
asynchronously. sample’s most representative CTE along the fiber direction (𝛼1 ) was de-

5
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Table 2
Ultimate tensile properties and load limits imposed during DIC and TSA.

Tensile Properties Load Limits [kN]

Configuration Laminate Code 𝑃max [kN] 𝜖max [%] 𝑃min 𝑃max

[0]8 UD 65.49 1.372 2 12.5

[0/90/90/0]s CP0 51.69 1.274 2 7.75
[90/0/0/90]s CP90 47.03 1.517 1 4.49
[02 /-452 /452 /902 ]s QI 77.39 1.696 2 12.5

termined from the average of data points inserted in the bin with the
highest relative frequency. Finally, the volume fractions of matrix, fiber
and void were evaluated for five specimens following ASTM D3171-15
recommendations. One can infer the high quality of the manufactured
laminates from the very small value of the void content (𝑤𝑣𝑜𝑖𝑑 ≈ 2.18%).
A summary of the results obtained in these tests is presented in Table 1.
Unlike metals, in composite materials, an endurance limit, i.e. a
stress level below which an infinite life can be assumed, is a rather dif-
ficult parameter to quantify. However, it is reasonable to assume the
absence of damage being imparted to the structure during TSA if the
applied load levels lie below the yield point where the material behav-
ior is perfectly linear elastic. Values ranging between 10 − 20% of the
ultimate tensile load can be considered as reference for maximum ap-
plied load during a TSA test [25]. In order to determine the ultimate
tensile properties for each stacking sequence, samples with in-plane di-
mensions of 25 × 250 mm were prepared and uniaxially loaded in an
Instron 8801 servo-hydraulic machine while displacement values were
tracked using an Instron AV2 Non-Contacting Video Extensometer until
failure. Table 2 summarizes the measured strain and load at break, 𝜖max
Fig. 3. Frequency test for a UD0 CFRP sample.
and 𝑃max , respectively.

4.3. TSA setup demands the spraying of a black coated surface due to their high re-
flectivity characteristics [12]. On the other hand, the high emissivity of
The setup for TSA involves three different equipments. A FLIR the epoxy layer formed during the manufacture of laminated composite
X6580sc infrared camera with an InSb (Indium Antimonide) sensor and materials enables a more straightforward testing routine. However, the
a 25 mm lens was used for thermal data acquisition. This device is ca- in-service application of the current methodology to large-sized aircraft
pable of an infrared image resolution of 640 × 512 pixels, a maximum structures would become impractical if the DIC’s speckle pattern were
frame rate of 355 Hz (at full resolution) and a thermal sensitivity lower to be removed ahead of a TSA test. Therefore, the thermoelastic signal’s
than 25 mK. The mechanical load was imparted to the material using source identification tests involving both DIC and TSA were performed
an Instron 8801 servo-hydraulic machine with a load cell capacity of in speckle-patterned surfaces, whereas no modifications to the surface
±250 kN. The acquired camera’s signal and load cell’s lock-in signal finishing were carried out for test routines solely involving TSA. The
were processed in real-time using the TSA module available on the Ede- emissivity values of the speckle pattern was determined to be 0.97, in
vis DisplayImg 6 software. Prior to starting any TSA test, the emissivity accordance with the methodology detailed in the MSc thesis of the first
properties of the material under deformation, apparent reflected tem- author [27].
perature, time duration for data acquisition, camera’s frame rate and The presence of an extra material coating can induce thermal sig-
integration time were required as input. nal amplitude variations that must be accounted for. Previous TSA on
During the dynamic tests on the samples, the frame rate of IR camera black-coated metals reported Δ𝑇 ∕𝑇0 amplifications of 20% and phase re-
was set to 355 Hz and the camera was located at a distance of around ductions [28]. Wang et al. [19] used speckle pattern on metals for com-
70 cm with respect to the sample due to physical limitations imposed by bined DIC and TSA as a compensation technique for complex motion. An
the servo-hydraulic machine’s cables and grip orientation. All dynamic almost negligible influence of the speckle pattern in the thermoelastic
tests were conducted under load control, at a frequency of 20 Hz, with response of the material was observed by the authors. Clarification on
a maximum load lower or equal to 15% of 𝑃max as shown in Table 2. these contradictory observations was sought by conducting two series of
To avoid compression loads that might originate from the incapability five TSA tests each on a UD sample with and without a speckle pattern
of the controller to stabilize the applied load at high frequencies, the coating under the load limits presented in Table 2 and at a frequency of
minimum load was kept between 1–2 KN. A summary of the imposed 20 Hz. For each series of tests, an average value of the mean full-field
load limits is also presented in Table 2. Δ𝑇 ∕𝑇0 was extracted, and their ratio evidenced a Δ𝑇 ∕𝑇0 increase of 3%
After each measurement is finalized, the Edevis software outputs a due to the application of a speckle pattern. Hence, initial Δ𝑇 ∕𝑇0 values
single image containing the pixel-wise 𝑇0 , 𝜙 and temperature amplitude were divided by a factor of 1.03 before interpolation.
(Δ𝑇 ∕2). These data are the results of an average over a user-defined Attaining quasi-adiabatic conditions during TSA is dependent upon
time for data acquisition of 240 s in order to increase the signal-to-noise selecting a loading frequency capable of generating a temperature varia-
ratio [26]. A rectangular region of interest (ROI) for data extraction tion with a wavelength small enough to minimize the heat transfer rate
was defined within the area limited by the specimen’s width and gauge between neighboring plies. Under this condition, the temperature de-
length. tected by the thermal camera is expected to be generated in the most
Unlike DIC, there’s no requirement for applying an extra layer of sur- outer surface of the laminate. The numerical models developed by Wong
face coating if high emissivity materials are considered. TSA of metals [23] for understanding heat transfer phenomena in composite materials

6
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Fig. 4. Input and outputs from the Smoothing

Element Analysis for CP0; (a)-Point-wise longi-
tudinal displacement; (b)-Continuous longitu-
dinal displacement; (c)-Continuous longitudi-
nal strain.

Fig. 5. Input and outputs from the Smooth-

ing Element Analysis for CP0; (a)-Point-wise
transverse displacement; (b)- Continuous trans-
verse displacement; (c)- Continuous transverse
strain.

7
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

revealed neglegible convection at the free surfaces as well as at the fiber- 4.4. DIC setup
matrix interface. However, the heat transfer between plies cannot be
overlooked. The same through-thickness heat transfer phenomenon was Two-dimensional DIC was performed using a GOM Adjustable 12M
also detected in TSA of double overlap aluminum joint specimens, and system with a resolution of 4096 × 3072 pixels. The calibration of the
numerically validated using FEA by Dunn [41]. Nonetheless, frequency camera, image acquisition and further post processing were conducted
calibration tests performed by Bakis et al. [29] in previous TSA of CFRP through the ARAMIS professional software. The calibration for the 2D
materials reported a more unstable thermoelastic signal response trend DIC system yielded a calibration deviation of 0.035 pixels and a scale
in the range of testing frequencies below 20 Hz for the UD configura- deviation of 0.002 mm. DIC was carried out while the specimen was
tion comparatively to cross-ply and angle-ply laminates. Based on the quasi-statically loaded in an Instron 8853 servo-hydraulic machine at a
aforementioned findings, it was decided to subject a UD sample for load rate of 3kN/min up to the maximum load presented in Table 2, and the
control tests between the load limits presented in Table 2 and loading data was recorded at a frame rate of 0.5 Hz. Two 75Ω analog coaxial
frequencies of 5, 8, 10, 12, 15, 18 and 20 Hz. cables were connected from the servo-hydraulic machine to the 2D-DIC
The average Δ𝑇 ∕𝑇0 for each frequency alongside the exponential in order to trigger the start and end of data acquisition at the exact
curve fitting to the experimental data are presented in Fig. 3. As can be load limit points. During data processing using the ARAMIS software,
seen, the trend line levels off at higher frequencies thereby confirming the dimensions of ROI (also referred to as surface component) for each
the need for high loading frequencies to reach temperature measure- specimen were manually redefined using the width of each specimen as
ments solely resultant from the thermoelastic phenomenon at each pixel the reference distance.
location. For frequency values higher than 20 Hz, the servo-hydraulic
machine can induce significant mechanical vibrations to the surround-
ing and the tripod of the IR camera. Consequently, the maximum and 5. Smoothing Element Analysis (SEA)
minimum temperature readings from the camera for a specific pixel
could also be including the thermoelastic effect from adjacent pixels Prior to implementing the model comparison, the experimental set
on the specimen, hence contaminating the results. In literature, it was of measurements is required to be treated numerically since different
also stated that the employment of frequencies higher than 30 Hz gen- meshes were attributed to each data set. For this purpose the SEA
erate mechanical and electrical resonance effects which considerably method is implemented in an in-house code, which is subsequently used
affect the measurements [30]. Therefore, in order to ensure a balance to obtain a common full-field distribution for the aforementioned vari-
between high signal-to-noise ratio and reliable thermoelastic response, ables over the structure’s domain. Utilization of SEA will not only mini-
a value of 20 Hz was selected as it fell into the region where the curve is mize the errors in the reconstruction of the data set, but will also repre-
nearly flat. The aforementioned results obtained from experiments ap- sent the variables as continuous function of the spatial coordinates. This
pear to be in line with the trends described in [23,29,30]. According analysis tool has proven successfull in previous studies [39,40] given
to these authors, adiabatic conditions in CFRPs require the frequency that the SEA simulation yields in C1 -continuous functions of the experi-
value to be much higher than the ones generally accepted for metals or mental data fields; therefore, the first order derivatives of the displace-
GFRPs (ranging between 5 and 10 Hz). ments (strains) will be C0 -continuous.

Fig. 6. Distribution map of thermoelastic mod-

els yielding the lowest pixel-wise 𝛿.

8
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Let us consider that the term 𝜀ℎ𝑖 be a common symbol for the discrete of the triangular element, q𝑒 . Subsequently, a final set of equations can
set of experimental data (DIC displacements/IR temperature). Analo- be obtained by minimizing Eq. (11) with respect to q𝑒 as:
gously, the continuous form of the experimental data can be repre-
sented as 𝜀ℎ𝑖 → 𝜀(x) ≡ 𝜀, i.e., resultant of the smoothing process. For 𝜕𝐹 (𝐪𝑒 )
= 0 ⇒ 𝐀𝑒 𝐪𝑒 = 𝐇𝑒 (12)
a single smoothing element, the penalized-discrete-least-squares error 𝜕𝐪𝑒
functional can be defined as [20]:
∑ ( ) where the right-hand side vector He is a function of the discrete ex-
𝐹 = 1𝑛 𝑛𝑖=1 (𝜀(x𝑖 ) − 𝜀ℎ𝑖 )2 + 𝛼 ∫𝑆 (𝜀,1 − 𝑘1 )2 + (𝜀,2 − 𝑘2 )2 d𝑆 perimental data and the left-hand-side matrix A𝑒 contains the squared
( ) (11)
+ 𝛽𝑆 ∫𝑆 (𝑘1,1 )2 + (𝑘2,2 )2 + 12 (𝑘1,2 + 𝑘2,1 )2 d𝑆 shape functions and their derivatives. The explicit form of these matrix
and vectors can be found in [20].
where 𝑛 is the total number of discrete data-points available in the Given that the coupon domain is discretized with several smoothing
smoothing element, the symbol 𝜀(x𝑖 ) represents the analytical counter- elements, a global matrix-vector equation should be constructed by the
part of 𝜀ℎ𝑖 at the point x𝑖 , the operator (𝑔),𝑎 ≡ 𝜕𝜕𝑥(𝑔 ) henceforth stands assembly of each single element’s contribution through the connectivity
𝛼
for the partial derivative with respect to in-plane coordinates, and the S of elements (analogous to the classical finite element method). Then,
term is the area of the smoothing element. The 𝑘1 ≡ 𝑘1 (x) and 𝑘2 ≡ 𝑘2 (x) the global equation system can be solved to obtain all unknown DOF
represents the continuous analytical terms for the first-order deriva- of the smoothing discretization. Such solution leads to the continuous
tives of the discrete 𝜀ℎ𝑖 experimental displacement/temperature data form of displacements as well as the longitudinal and transverse strains,
with respect to the in-plane coordinates of 𝑥1 and 𝑥2 , respectively. C1 - which are computed through the derivatives of the displacements along
𝜕𝑢 𝜕𝑢
continuity of the 𝜀(x) distribution with C0 -continuous derivatives, 𝑘𝑖 (x), the longitudinal and transverse directions, 𝜕𝑥𝑥 and 𝜕𝑦𝑦 , respectively. The
over the SEA domain can be achieved by controlling the dimensionless SEA results for the CP0 sample are presented in Figs. 4 and 5 for both the
parameters 𝛼 and 𝛽. More details of each term in Eq. (11) and their roles longitudinal and transverse displacements, in the given order. In these
for SEA can be found in [20,21]. figures, (a) represents the point-wise displacements extracted from DIC
The smoothed strain data 𝜀(x) and its respective derivatives 𝑘𝑖 (x) can (in the camera’s mesh); and (b) and (c) represent the continuous dis-
be approximated through finite-element interpolation functions and as- placements and strains, respectively (in the background mesh). These
sociated degrees of freedom. To this end, we utilized the three-node tri- continuous strain fields were used as input to the analytical thermoe-
angular smoothing element introduced in [20]. Such an approximation lastic model Eqs. (7)–(9) to generate the pixel-wise theoretical Δ𝑇 ∕𝑇0
results in Eq. (11) to be written in terms of the nodal degrees of freedom values.

Fig. 7. Distribution map of pixel-wise Δ𝑇 ∕𝑇0

values for the thermoelastic models calculated
using strain output from SEA, and Δ𝑇 ∕𝑇0 val-
ues measured with the IR camera for the UD
configuration.

9
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Fig. 8. Distribution map of pixel-wise Δ𝑇 ∕𝑇0

values for the thermoelastic models calculated
using strain output from SEA, and Δ𝑇 ∕𝑇0 val-
ues measured with the IR camera for the CP0
configuration.

Fig. 9. Distribution map of pixel-wise Δ𝑇 ∕𝑇0

values for the thermoelastic models calculated
using strain output from SEA, and Δ𝑇 ∕𝑇0 val-
ues measured with the IR camera for the CP90
configuration.

10
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Table 3
Average phase angle (𝜙), and results from thermoelastic model identification process: area covered by
each thermoelastic model, average relative error (𝛿), average mixed model temperature ratio prediction
((Δ𝑇 ∕𝑇0 )𝑀𝑖𝑥 ) and average experimentally-measured temperature ratio ((Δ𝑇 ∕𝑇0 )𝑇 𝑆𝐴 ).

Laminate Code UD CP0 CP90 QI

◦
𝜙( ) 5.90 -11.18 -10.77 -16.71
Model Bulk Bulk Bulk RRL Homogeneous
Covered Area (%) 100 100 100 97.77 2.23
̄
𝛿(%) 17.18 25.49 15.06 18.01
m Mix TSA Mix TSA Mix TSA Mix TSA
(Δ𝑇 ∕𝑇0 )𝑚 (×10−6 ) 66.69 84.38 128.83 198.49 242.49 210.87 354.01 300.49

6. Results and discussion ratios are presented for each stacking sequence in Figs. 7–10, comple-
mented by the average temperature ratio computed over the spatial do-
m
Fig. 6 presents a map with the distribution of thermoelastic models main that best fits model k, (Δ𝑇 ∕𝑇0 )k (where 𝑘 stands for Bulk, RRL
yielding the lowest error in each pixel k calculated as per Eq. (10) for or Homogeneous and 𝑚 defines whether the source of input data was
each of the four different stacking sequences mentioned in Table 2. The either TSA or DIC) presented in Table 4.
resulting percentage of pixels assigned to each thermoelastic model is There’s a clear predominance of a single thermoelastic model in each
shown in Table 3. It is important to recall the mathematical equivalence stacking sequence. These results were not initially foreseen as one would
between the Bulk and Homogeneous model equations Eqs. (6) and ((9), expect a more substantial mixed model contribution for the overall ther-
respectively) for the UD case. Hence, the thermoelastic model identifi- moelastic response considering the inherent composite microstructural
cation process for UD solely accounted for the Bulk and RRL models. heterogeneity. However, the full-field distribution of strain in the CP0
Furthermore, high Δ𝑇 ∕𝑇0 values were consistently found at the tab ter- case shown in Figs. 4 and 5 suggests negligible spatial changes in the
mination near the moving grip. These observations can be attributed to strain field. A similar strain pattern can be inferred for the other con-
motion effects that captured the moving grip’s surface temperature, as figurations based on the range of Δ𝑇 ∕𝑇0 values calculated using the
well as high stress concentrations generated by the tab end geometry. three thermoelastic models presented in Figs. 7–10 whose magnitude
Despite the thermoelastic nature of the temperature variations resultant never exceeds 9 × 10−5 . Similarly, a fairly constant full-field distribution
from the latter phenomenon, it does not reflect either material or geo- of Δ𝑇 ∕𝑇0 measured by the IR camera can be reported from observation
metrical effects of the sample in the measured strain. Therefore, these of Figs. 7–10. In addition, the narrow ranges of deformation-induced
regions were disregarded from the analysis and not presented in Fig. 6. and IR camera-based Δ𝑇 ∕𝑇0 solely overlap for a single model per lami-
In addition, the full-field experimental- and strain-based temperature nate configuration, hence yielding the result of Fig. 6. This homogenized

Fig. 10. Distribution map of pixel-wise Δ𝑇 ∕𝑇0

values for the thermoelastic models calculated
using strain output from SEA, and Δ𝑇 ∕𝑇0 val-
ues measured with the IR camera for the QI
configuration.

11
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Fig. 11. Distribution map of pixel-wise 𝛿 after

model identification process.

Table 4
Experimental- and strain-based Δ𝑇 ∕𝑇0 averaged in regions of best fit to each thermoelastic model and respective 𝛿 𝑘 .

Laminate Code UD CP0 CP90 QI

k Bulk Bulk Bulk RRL Homogeneous

m DIC TSA DIC TSA DIC TSA DIC TSA DIC TSA
𝑚
(Δ𝑇 ∕𝑇0 )𝑘 66.69 84.38 128.83 198.49 359.73 299.42 359.73 325.45 306.85 325.45
̄𝛿𝑘 (%) 17.18 25.49 15.06 18.01 14.80

strain and temperature patterns can be the combined result of the high tensile load-bearing element and yielding a fiber-dominated thermo-
quality prepreg utilized for manufacturing and well-controlled temper- mechanical response. If the thermal energy associated with the fibers
ature/pressure conditions attained while curing the epoxy evidenced is able to propagate through the external epoxy layer and reach the
by the even distribution of fiber and matrix statistically represented in surface of the specimen, it can be detected by the thermal camera and
Table 1 by 𝑣𝑓 and 𝑣𝑚 . Nonetheless, the model superposition in the QI quantified by the Bulk model which takes into account the homogenized
configuration confirms the existing limitations of implementing a global material properties of the laminate’s most outer ply. The success of this
identification of the thermoelastic signal source using localized strain energy propagation process is strongly dependent on the applied loading
input based on either strain gauges or extensometers. frequency and thickness of the external epoxy layer. Our results allow
The predominance of the Bulk model in most of the tested configura- us to conclude that the external epoxy layer is not sufficiently thick so as
tions is consistent with previous findings of TSA on CFRPs [17]. Both UD to prevent heat diffusion from the first ply towards the surface to occur.
and CP0 configurations contain surface plies with fibers oriented along At the surface level of the CP90, the epoxy accumulated in the outer
the externally applied loading direction, hence becoming the primary surface as well as within the most-outer ply of the laminate is the main

12
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

constituent of the composite responsible for carrying the externally- dictions from the RRL and Bulk models is greater for the CFRP case. On
applied uniaxial load. Under quasi-adiabatic conditions, a more pro- the other hand, E-glass fibers coupled with epoxy matrix yield homoge-
nounced matrix-derived thermal response is expected to be produced. nized properties with a thermo-mechanical behavior more analogous to
Based on the previous knowledge of TSA on GFRPs with a transversely- that of the epoxy. It is therefore expected that a matrix-derived thermal
oriented surface ply [31], one would expect the thermoelastic response response originated at the first ply level can be more likely detected by
of the CP90 configuration to be more accurately replicated by the RRL the RRL model in GFRPs.
model given in Eq. (8) as it is a function of the resin’s mechanical and On the other hand, the almost absolute contribution of the RRL
thermal properties. However, the current work shows a surprising con- model to the thermoelastic response in the QI case can be justified by
formity between the results of TSA and the Bulk model, with the full the possible existence of more localized regions with a thick outer epoxy
sample area being associated with the Δ𝑇 ∕𝑇0 from this model. Upon layer acting as a strain witness.
observing a less pronounced influence of the RRL model on the overall The Homogeneous model has proven capabilities in stress quantifica-
thermoelastic response in CFRPs, Pitarresi and Galietti [17] suggested tion of geometrically complex composite structures using the thermoe-
the aforementioned strain witness behavior to be ineffective due to in- lastic signal from TSA in circumstances where loading frequencies high
sufficient epoxy thickness deposited in the material’s surface. Yet, we enough to prevent inter-laminar heat diffusion are not attained [33]. The
believe that this argument is not sufficient to justify the incapability influence of this model is further enhanced in stacking sequences with
of the RRL model to capture the thermoelastic response generated by highly contrasting fiber orientations between neighboring plies. There-
the matrix at the first ply level. Additionally, one should allude to the fore, the CP0, CP90 and QI configurations would theoretically be more
significant differences in thermal and mechanical properties between prone to achieving a thermal equilibrium at the laminate level quanti-
the homogenized CFRP lamina and the epoxy. Although the thermal re- fied by a more pronounced influence of the Homogeneous model in the
sponse is originated in the matrix, its thermo-mechanical properties are distribution of Fig. 6. However, the tenuous agreement attained in this
strongly influenced by the embedded carbon fibers and in turn can no study between the experimental results and the Homogeneous model
longer be quantified by the RRL model. To better elucidate this state- predictions for all stacking sequences demonstrates the effectiveness of
ment, Eqs. (7) and (8) were simplified by replacing the properties with the selected loading frequency of 20 Hz in guaranteeing the satisfaction
their respective values from Table 1 and represented as a function of Δ𝜀𝑥 of quasi-adiabatic conditions at the lamina level.
and Δ𝜀𝑦 , i.e. (Δ𝑇 ∕𝑇0 )RRL = 𝑓 (Δ𝜀𝑥 , Δ𝜀𝑦 ) and (Δ𝑇 ∕𝑇0 )Bulk = 𝑓 (Δ𝜀𝑥 , Δ𝜀𝑦 ), Another reliable indicator of non-adiabatic phenomena in TSA is 𝜙.
where (Δ𝑇 ∕𝑇0 )Bulk and (Δ𝑇 ∕𝑇0 )RRL are the temperature ratios for the Besides heat transfer between laminae, a deviation in 𝜙 from in-phase
Bulk and RRL models, respectively. The difference between both equa- conditions can be also associated to frictional heating in cracked re-
tions, i.e. Δ = (Δ𝑇 ∕𝑇0 )RRL − (Δ𝑇 ∕𝑇0 )Bulk , is our expression of interest gions and delays in the propagation of thermal waves from the sub-
for analysis. For the sake of comparison, the same calculations were strate to the laminate’s surface due to the presence of external coatings.
performed for the case of GFRPs using properties reported in Emery Table 4 presents the phase angle averaged over each sample’s domain,
et al. [32]. The expression for Δ using data from the CFRP case yields 𝜙. It is of utmost importance to point out that edevis software displays
Δ = 0.146Δ𝜀𝑥 + 0.233Δ𝜀𝑦 whereas Δ for the GFRP case is represented 𝜙 in a range comprehended between [-1800 , 1800 ]. In addition, a load
by Δ = 0.138Δ𝜀𝑥 + 0.146Δ𝜀𝑦 . When the CFRP and GFRP are subjected to signal in-phase with the temperature variation yields a default phase
the same deformation, we can infer that the discrepancy between pre- angle of 1800 , in agreement with the symmetric relationship between

Fig. 12. Distribution map of thermoelastic

models yielding the lowest pixel-wise 𝛿 after
model identification process with variable 𝛼1 .

13
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

both signals represented by the negative sign in Eqs. (7)–(9). However, 1.65 × 10−4 was obtained in the QI configuration. However, particular
for a set of 𝜙 values located in the vicinity of both upper and lower attention must be given to all stacking sequences with a 00 fiber-oriented
range bounds, a straightforward arithmetic average would yield erro- surface ply where 𝛿 was the highest. Conformity of experimental results
neous results. Therefore, for the sake of simplicity, all 𝜙 results were in these stacking sequences to Bulk model’s predictions is strongly af-
shifted by a factor of 1800 , and only arithmetic averaging was consid- fected by the variability inherent to the carbon fibers’ CTE as reported in
ered in Table 4. Assuming initially undamaged conditions, together with [17]. Therefore, assigning a unique value for 𝛼1 can explain the higher
the previously discussed negligible contribution from the Homogeneous 𝛿 and relative error for locations of best fit to the Bulk model, 𝛿 Bulk . The
model to the mixed thermoelastic model distribution for all configura- next section will cover the effect of locally readjusting 𝛼1 for minimizing
tions allows us to infer a possible influence of the external layer of paint 𝛿.
in the slight average phase offset detected. A second contribution is pre- For completeness, it is essential to state that there might be other
sumably a consequence of the selected loading frequency greater than possible sources of uncertainties contributing to the relative errors listed
15 Hz which has already proved to introduce an artificial phase gradient in Tables 3 and 4 such as the vibrational noise induced by the high
in TSA [34]. Nevertheless, the almost in-phase nature of the measured frequency loading which might affect the TSA measurements, and the
temperature with respect to the applied load underlines the validity of deterministic nature of thermal and mechanical properties.
the experimental results obtained from TSA.
For a given stacking sequence, maps of error calculated using 6.1. Effect of Fiber’s coefficient of thermal expansion
Eq. (10) for each location of the background mesh are shown in
Fig. 11 and complemented by the respective average over each sam- Most engineering materials have a positive 𝛼, hence leading to a
ple’s domain, 𝛿, and area of best fit to each thermoelastic model, 𝛿𝑘 , on dilatation due to a temperature increase. For carbon fibers, this value
Tables 3 and 4, respectively. There’s a generalized good agreement be- might be negative at room temperature, i.e 𝛼𝑓 = -0.38 × 10-6 K-1 [35],
tween (Δ𝑇 ∕𝑇0 )DIC
k
and (Δ𝑇 ∕𝑇0 )TSA
k
considering the order of magnitude where the subscript 𝑓 corresponds to fiber. On the other hand, under the
involved in thermoelastic measurements. For the sake of argument, the same environmental conditions, epoxy resin is characterized by a posi-
largest difference of Δ𝑇 ∕𝑇0 between model predictions and TSA equal to tive 𝛼𝑟 value. The CTE mismatch between the constituents yields a com-

Fig. 13. Distribution map of pixel-wise 𝛿 after

model identification process with variable 𝛼1 .

14
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Fig. 14. Distribution map of 𝛼1 yielding mini-

mum pixel-wise 𝛿.

posite material whose 𝛼1 is highly sensitive to changes in fiber/matrix employing an admissible value of 𝛼1 . Since the RRL and Homogeneous
volume fractions. A theoretical relationship between 𝛼1 and each con- model equations are independent of 𝛼1 , temperature ratios using these
stituent volume fraction can be obtained for a stress-free composite el- models were calculated only once. The relative error 𝛿𝑘 served as a met-
ement subjected to a temperature change [36], ric to quantify the accuracy of (Δ𝑇 ∕𝑇0 )k with respect to (Δ𝑇 ∕𝑇0 )TSA .
( ) When first considering the lower bound of the admissible range for 𝛼1 ,
𝛼𝑓 𝐸𝑓 − 𝛼𝑟 𝐸𝑟 𝜈𝑓 + 𝛼𝑟 𝐸𝑟
𝛼1 = ( ) , (13) the model k yielding the minimum relative error was assigned to charac-
𝐸𝑓 − 𝐸𝑟 𝜈𝑓 + 𝐸𝑟
terize the thermoelastic response in that location, and the optimal value
where 𝐸𝑓 is the fiber Young’s modulus. Yales et al. [37] detected a for 𝛼1 was automatically replaced by the same lower bound. For every
change in the sign of 𝛼1 for a fiber volume fraction increase from 50% following increment of 10−8 in 𝛼1 , the calculation of 𝛿𝑘 was repeated,
to 60%. The variable nature of CTE in the composite laminate was also and the minimum value of which was compared against the one stored
pointed out in a study on TSA by Zhang et al. [38]. Upon comparing the up to the previous iteration. Assigned models were replaced if the new
thermoelastic response generated over and between carbon fiber bun- minimum 𝛿𝑘 was associated to a different k model than the one previ-
dles, they showed that the presence of alternating fiber- and resin-rich ously assigned. Moreover, the optimal value for 𝛼1 would be updated if
regions led to rapid changes between positive to negative Stress Pattern a new minimum 𝛿Bulk was attained.
Analysis by Thermal Emission (SPATE) signal values. Figs. 12 and 13 supported by Tables 5 and 6 stress once again the
Assuming that the experimental errors associated with TSA, DIC and effectiveness of the Bulk model to replicate the thermoelastic behavior
SEA interpolation are negligible, and the standard deviation of all other in CFRPs. The reduction in 𝛿 attained for all stacking sequences rein-
thermal-mechanical variables to be negligible compared to the CTE’s forces the authors hypothesis of the impact in the model identification
(if normally distributed), a strategy for minimizing 𝛿 is being proposed process of a variable 𝛼1 . Fig. 14 presents the distribution of optimal 𝛼1
by allowing for pixel-wise variations in 𝛼1 . The range of values mea- that minimizes 𝛿Bulk . Variations of 𝛼1 with respect to the statistically-
[ ]
sured for 𝛼1 ∈ −0.99 × 10−6 , 0.70 × 10−6 K-1 according to the guidelines representative value of 𝛼1 in Table 1 are not significant for the UD and
provided in Section 4.2 were considered. In this analysis, it is assumed CP0 configurations, whereas 𝛼1 values in the QI configuration approach
that any point in the 𝛼1 domain shares an equal probability of success. 𝛼𝑟 . With support of Eq. (13), it can be inferred an extended influence
At each point of the background mesh, (Δ𝑇 ∕𝑇0 )Bulk was updated after

15
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

Table 5
Area covered by each thermoelastic model, average relative error (𝛿), average temperature ratio us-
ing mixed model predictions, ((Δ𝑇 ∕𝑇0 )𝑀𝑖𝑥 ), and average experimentaly-measured temperature ratio,
((Δ𝑇 ∕𝑇0 )TSA ), after thermoelastic model identification process with variable 𝛼1 .

Laminate Code UD CP0 CP90 QI

Model Bulk Bulk Bulk RRL Bulk

Covered Area (%) 100 100 100 0.73 99.27
𝛿̄𝑘 0.8406 0.2652 8.4624 0.2078
m Mix TSA Mix TSA Mix TSA Mix TSA
(Δ𝑇 ∕𝑇0 )m (×10−6 ) 87.79 84.38 202.90 198.49 230.34 210.87 299.49 300.49

Table 6
Experimental- and strain-based Δ𝑇 ∕𝑇0 averaged in regions of best fit to each thermoelastic model and respective 𝛿 for
the model identification process with variable 𝛼1 .

Laminate Code UD CP0 CP90 QI

k Bulk Bulk Bulk RRL Bulk

m DIC TSA DIC TSA DIC TSA DIC TSA DIC TSA
(Δ𝑇 ∕𝑇0 )𝑚𝑘 87.79 84.38 202.90 198.49 230.34 210.87 311.97 319.25 299.40 300.36
𝛿̄𝑘 0.8406 0.2652 8.4624 0.1121 0.2085

of epoxy in the material distribution of the previously mentioned con- Declaration of Competing Interest
figurations through an increase in 𝜈𝑚 . On the other hand, the CP(90/0)
configuration contains a marked presence of negative-valued 𝛼1 , hence The authors declare that they have no known competing financial
indicating the existence of regions with a higher 𝜈𝑓 ∕𝜈𝑚 ratio. Neverthe- interests or personal relationships that could have appeared to influence
less, the considered limits of 𝛼1 still reveal insufficient to further reduce the work reported in this paper.
𝛿 from the value shown in Table 3.
In summary, this conceptual approach to characterize the full-field
CRediT authorship contribution statement
𝛼1 contributed to a substantial error minimization in the comparison of
data collected from TSA and DIC, therefore enhancing the reliability of
Francisco de Sá Rodrigues: Conceptualization, Methodology, Soft-
the thermoelastic model identification process proposed in this study. It
ware, Validation, Formal analysis, Investigation, Writing - original
also provided a further insight into the actual fiber/matrix volume frac-
draft, Writing - review & editing. Ricardo Marques: Conceptualiza-
tions of each stacking sequence, and how a deterministic approach for
tion, Methodology, Validation, Writing - review & editing. Isa Emami
thermoelastic model identification does not fully account for the com-
Tabrizi: Formal analysis, Investigation, Writing - review & editing.
posite material heterogeneity. However, the solution to this problem
Adnan Kefal: Software, Formal analysis, Writing - review & editing.
does not solely arise from the 𝛼1 variability, but would also consist of
Hafiz Qasim Ali: Investigation, Writing - review & editing. Mehmet
quantifying the probability distribution of all thermal-mechanical vari-
Yildiz: Writing - review & editing, Supervision, Resources, Funding ac-
ables included in the various thermoelastic models. Transitioning the
quisition. Afzal Suleman: Writing - review & editing, Supervision, Fund-
model identification approach from deterministic to probabilistic will
ing acquisition.
be the subject of a future work to be conducted by the researchers in-
volved in the current work.
Acknowledgments
7. Conclusions
The authors of this paper would like to thank the financial and in-
An alternative methodology to identify the thermoelastic signal’s frastructural support provided by Sabanci University - Integrated Man-
source using the full-field data of TSA has been proposed. A scattered ufacturing Technologies Research and Application center, Sabanci Uni-
model distribution was obtained for the QI configuration with an over- versity - Kordsa Composite Technologies Center of Excellence. The au-
all good agreement between results from TSA and DIC, thus highlighting thors also acknowledge Fundação para a Ciência e a Tecnologia (FCT),
the weaknesses of a traditional single model-based approach to charac- through IDMEC, under LAETA, project UIDB/50022/2020. Afzal Sule-
terize the material’s thermoelastic response. Moreover, the Bulk model’s man also acknowledges the NSERC Canada Research Chair funding pro-
predominance corroborates the literature findings in TSA of CFRPs. Fur- gram. Moreover, the contributions of Raja Khan and Pouya Yousefi are
ther improvements in model fitting error were achieved by exploring the greatly appreciated.
well-known vulnerability of the composite material’s longitudinal CTE
to fiber/matrix volume fractions. Focus must be also directed to switch- References
ing from deterministic to statistical/probabilistic data input, finding the
minimum loading frequency for a good compromise between adiabatic [1] Irving PE, Soutis C. Polymer Composites in the Aerospace Industry. Woodhead Pub-
conditions and reduced vibrational noise and finding equipment ade- lishing; 2019.
[2] Kassapoglou C. Design and Analysis of Composite Structures With Application to
quate to the geometry, load values and other physical constraints im-
Aerospace Structures. Wiley; 2013.
posed by the lab environment. This methodology could be further vali- [3] Greene RJ. Thermoelastic stress analysis. Springer Handbook of Experimental Me-
dated in GFRP materials as well as samples with thicker laminae, various chanics. Springer US; 2008.
[4] Sutton MA, Orteu J-J, Schreier H. Image Correlation for Shape, Motion and De-
matrix/fiber volume fraction ratios or damaged prior to TSA.
formation Measurements: Basic Concepts,Theory and Applications. 1st ed. Springer
DIC and TSA proved, once again, to be complementary SHM tech- Publishing Company, Incorporated; 2009. doi:10.1007/978-0-387-78747-3.
niques with straightforward applicability in stress/strain quantification [5] Pitarresi G, Cappello R, Catalanotti G. Quantitative thermoelastic stress analysis by
and damage monitoring in composite structures. Together it can pro- means of low-cost setups. Opt. Lasers Eng. 2020;134:106158.
[6] Thatcher J, Crump D, Devivier C, Bailey P, Dulieu-Barton J. Low cost infrared
vide TSA with enhanced full-field capabilities while minimizing errors thermography for automated crack monitoring in fatigue testing. Opt. Lasers Eng.
in future strain measurements. 2020;126:105914.

16
F. de Sá Rodrigues, R. Marques, I.E. Tabrizi et al. Optics and Lasers in Engineering 146 (2021) 106689

[7] Thomson W. Ii. on the thermoelastic, thermomagnetic, and pyroelectric proper- [24] Pitarresi G, Found M, Patterson E. An investigation of the influence of macroscopic
ties of matter. Lond. Edinbur. Dublin Philosoph. Mag. J. Sci. 1878;5(28):4–27. heterogeneity on the thermoelastic response of fibre reinforced plastics. Compos.
doi:10.1080/14786447808639378. Sci. Technol. 2005;65(2):269–80. doi:10.1016/j.compscitech.2004.07.008.
[8] Muc A, Barski M, Chwał M, Romanowicz P, Stawiarski A. Fatigue damage growth [25] Salerno A, Costa A, Fantoni G. Calibration of the thermoelastic constants
monitoring for composite structures with holes. Compos. Struct. 2018;189:117–26. for quantitative thermoelastic stress analysis on composites. Rev. Sci. Instrum.
doi:10.1016/j.compstruct.2018.01.065. 2009;80(3):034904.
[9] Krstulovic-Opara L, Klarin B, Neves P, Domazet Z. Thermal imaging and thermoe- [26] Rajic N, Street N. A performance comparison between cooled and uncooled in-
lastic stress analysis of impact damage of composite materials. Eng. Fail. Anal. frared detectors for thermoelastic stress analysis. Quant. InfraRed Thermogr. J.
2011;18(2):713–19. doi:10.1016/j.engfailanal.2010.11.010. 2014;11(2):207–21. doi:10.1080/17686733.2014.962835.
[10] Diaz F, Patterson E, Tomlinson R, Yates J. Measuring stress intensity factor dur- [27] de Sá Rodrigues F. Local-based calibration of the thermoelastic signal using ther-
ing fatigue crack growth using thermoelasticity. Fatigue Fract. Eng. Mater Struct. moelastic sress analysis and digital image correlation. Instituto Superior Técnico,
2004;27:571–83. doi:10.1111/j.1460-2695.2004.00782.x. Universidade de Lisboa; 2019. Master’s thesis.
[11] Quinn S, Dulieu-Barton J. Identification of the sources of non-adiabatic behaviour [28] Welch C, Zickel M. Thermal coating characterization using thermoelasticity. In: Re-
for practical thermoelastic stress analysis. J. Strain Anal. Eng. Des. 2002;37:59–71. view of Progress in Quantitative Nondestructive Evaluation: Volumes 12A and 12B.
doi:10.1243/0309324021514835. Springer US; 1993. p. 1923–30.
[12] Mackenzie AK. Effects of surface coatings on infra-red measurements of thermoe- [29] Bakis CE, Reifsnider KL. The adiabatic thermoelastic effect in laminated fiber com-
lastic responses. In: Stanley P, editor. Stress and Vibration: Recent Developments posites. J. Compos. Mater. 1991;25(7):809–30. doi:10.1177/002199839102500702.
in Industrial Measurement and Analysis, 1084. International Society for Optics and [30] Potter RT, Greaves LJ. The application of thermoelastic stress analysis techniques
Photonics: SPIE; 1989. p. 59–71. doi:10.1117/12.952906. to fibre composites. In: Optomechanical Systems Engineering, 0817. SPIE; 1987.
[13] Dulieu-Barton JM, Emery TR, Quinn S, Cunningham PR. A temperature correction p. 134–46. doi:10.1117/12.967419.
methodology for quantitative thermoelastic stress analysis and damage assessment. [31] Pitarresi G, Conti A, Galietti U. Investigation on the influence of the surface resin
Meas. Sci. Technol. 2006;17(6):1627–37. doi:10.1088/0957-0233/17/6/047. rich layer on the thermoelastic signal from different composite laminate lay-ups. In:
[14] Dulieu-Smith J. Alternative calibration techniques for quantitative thermoelastic Applied Mechanics and Materials, 3. Trans Tech Publ; 2005. p. 167–72.
stress analysis. Strain 1995;31(1):9–16. doi:10.1111/j.1475-1305.1995.tb00949.x. [32] Emery T, Dulieu-Barton J, Earl J, Cunningham P. A generalised approach to the
[15] Marques R, Unel M, Yildiz M, Suleman A. Remaining useful life prediction of calibration of orthotropic materials for thermoelastic stress analysis. Compos. Sci.
laminated composite materials using thermoelastic stress analysis. Compos. Struct. Technol. 2008;68(3–4):743–52.
2019;210:381–90. doi:10.1016/j.compstruct.2018.10.047. [33] Crump D, Dulieu-Barton J. Assessment of non-adiabatic behaviour in thermoelastic
[16] Emery TR, Dulieu-Barton JM. Thermoelastic stress analysis of damage mechanisms stress analysis of composite sandwich panels. Exp. Mech. 2012;52(7):829–42.
in composite materials. Compos. Part A Appl. Sci. Manufact. 2010;41(12):1729–42. [34] Pendleton R, Tuttle M. Manual on Experimental Methods for Mechanical Testing of
doi:10.1016/j.compositesa.2009.08.015. Composites. Springer Science & Business Media; 2012.
[17] Pitarresi G, Galietti U. A quantitative analysis of the thermoelas- [35] . T700s data sheetTech. Rep.. TorayCA; 2018.
tic effect in CFRP composite materials. Strain 2010;46(5):446–59. [36] Hyer MW, White SR. Stress Analysis of Fiber-Reinforced Composite Materials.
doi:10.1111/j.1475-1305.2009.00660.x. DEStech Publications, Inc; 2009.
[18] Silva ML, Ravichandran G. Combined thermoelastic stress analysis and digital image [37] Yates B, Overy MJ, Sargent JP, McCalla BA, Kingston-Lee DM, Phillips LN,
correlation with a single infrared camera. J. Strain Anal. Eng. Des. 2011;46(8):783– et al. The thermal expansion of carbon fibre-reinforced plastics. J. Mater. Sci.
93. doi:10.1177/0309324711418286. 1978;13(2):433–40.
[19] Wang W, Fruehmann R, Dulieu-Barton J. Application of digital image correlation to [38] Zhang D, Enke NF, Sandor BI. Thermographic stress analysis of composite materials.
address complex motions in thermoelastic stress analysis. Strain 2015;51(5):405–18. Exp. Mech. 1990;30(1):68–73. doi:10.1007/BF02322705.
doi:10.1111/str.12151. [39] Kefal A, Tessler A, Oterkus E. An enhanced inverse finite element
[20] Tessler A, Riggs HR, Freese CE, Cook GM. An improved variational method for finite method for displacement and stress monitoring of multilayered compos-
element stress recovery and a posteriori error estimation. Comput. Methods Appl. ite and sandwich structures.. Composite Structures 2017;179:514–40. doi:
Mech. Eng. 1998;155(1):15–30. 10.1016/j.compstruct.2017.07.078.
[21] Tessler A, Riggs HR, Dambach M. A novel four-node quadrilateral smoothing el- [40] Kefal A, Tabrizi I E, Yildiz M, Tessler A. A smoothed iFEM approach for effi-
ement for stress enhancement and error estimation. Int. J. Numer. Methods Eng. cient shape-sensing applications: Numerical and experimental validation on com-
1999;44(10):1527–41. posite structures. Mechanical Systems and Signal Processing 2021;152:107486. doi:
[22] Pan B, Quian K, Xie H, Asundi A. Two-dimensional digital image correlation 10.1016/j.ymssp.2020.107486.
for in-plane displacement and strain measurement: a review. Meas. Sci. Technol. [41] Dunn SA. On the effects of through-thickness thermal conduction on stress mea-
2009;20(6). doi:10.1088/0957-0233/20/6/062001. surement by thermoelastic techniques. Experimental Mechanics 1993;33:32–6. doi:
[23] Wong A. A non-adiabatic thermoelastic theory for composite laminates. J. Phys. 10.1007/BF02322547.
Chem. Solids 1991;52(3):483–94. doi:10.1016/0022-3697(91)90180-8.

17