PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

A methodology is presented to fit material parameters in finite element (FE) models using full displacement field data. Four bulge inflation tests were performed on a PVC-coated polyester material. Digital image correlation (DIC) was used to capture the full displacement field of the material. An inverse analysis was set up to find material parameters in a FE model which replicated the full displacement field of the experimental test. Two different objective functions were considered to quantify the discrepancy between the FE model and test data. One function considered equal weight among displacement components, while the other function weighted the discrepancies to balance different displacement components. The resulting parameters, for isotropic and orthotropic models, were heavily dependent upon which objective function was chosen. Additionally, cross validation was performed to select between the two material models. Similar to the material parameters, the cross validation results preferred different material models depending upon which objective function was being used.
Content may be subject to copyright.
Noname manuscript No.
(will be inserted by the editor)
Eect of Weighting Full-eld Residuals when Fitting Material
Model Parameters
Charles F. Jekel ·Martin P. Venter ·Gerhard
Venter ·Raphael T. Haftka
Received: date / Accepted: date
Abstract A methodology is presented to t material parameters in nite element
(FE) models using full displacement eld data. Four bulge ination tests were per-
formed on a PVC-coated polyester material. Digital image correlation (DIC) was
used to capture the full displacement eld of the material. An inverse analysis was
set up to nd material parameters in a FE model which replicated the full displace-
ment eld of the experimental test. Two dierent objective functions were consid-
ered to quantify the discrepancy between the FE model and test data. One function
considered equal weight among displacement components, while the other function
weighted the discrepancies to balance dierent displacement components. The result-
ing parameters, for isotropic and orthotropic models, were heavily dependent upon
which objective function was chosen. Additionally, cross validation was performed
to select between the two material models. Similar to the material parameters, the
cross validation results preferred dierent material models depending upon which
objective function was being used.
Keywords Inverse analysis ·Objective functions ·material parameter identication
1 Introduction
The nite element (FE) method has become an important design tool for membrane
structures, and selecting material parameters to represent complex material behavior
is a dicult task. This is especially true for the most complex and non-linear ma-
terials including coated woven fabrics. Inverse analyses, or iterative schemes for FE
model updating (FEMU) have been used to nd material parameters from complex
load cases. In this paper, two dierent objective functions were investigated when
F. Author
rst address
Tel.: +123-45-678910
Fax: +123-45-678910
E-mail: fauthor@example.com
S. Author
second address
2 Charles F. Jekel et al.
estimating material parameters from bulge ination tests on PVC-coated polyester.
Objective function refers to how the discrepancy between the FE model and test
data is quantied, and material parameters are typically selected to minimize this
function. The objective function has an important eect on the material parameters
selected using this type of inverse analysis (Zhan et al., 2011; Jekel et al., 2019; Tam,
2020).
PVC-coated polyester is a coated textile. It’s most commonly modeled as a con-
tinuous, homogeneous, orthotropic material (Shaw et al., 2010), which is largely
dependent on the warp-to-ll stiness ratio (Dinh et al., 2017). The typical weave
has the warp yarns pulled taut, while the ll yarns are woven in-between the warp
yarns. The ll yarns run orthogonal to the warp yarns. Various non-linear models
have been used in an attempt to better describe the behavior of the material. Galliot
and Luchsinger (2009) proposed a non-linear material model based on the load ratio
between the material warp and ll directions. Ambroziak and Kłosowski (2014) used
a piecewise linear orthotropic model, and Jekel et al. (2017) used polynomials to
describe the non-linear elastic behavior of the material.
Biaxial tests are commonly used to characterize material parameters for struc-
tural membrane materials (Stranghöner et al., 2016). Several studies (Rachik et al.,
2001; Charalambides et al., 2002; Machado et al., 2012; Tonge et al., 2013) have used
bulge (or bubble ination) tests to induce an equal biaxial load on the material. These
bulge tests typically involve inducing a pressure on one side of a circularly clamped
membrane material. The equal biaxial load occurs at the apex of the deected mem-
brane for an isotropic material. The measured pressure and displacements are then
used to infer material parameters, by comparing the results to a FE simulation.
A variety of applications (Becker et al., 2012; Murienne and Nguyen, 2016; Mejía
and Lantsoght, 2016; Chisari et al., 2017) have used Digital image correlation (DIC)
as non-invasive deformation measuring technique. The technique uses the correlation
between consecutive images, from multiple cameras, to calculate a full 3D displace-
ment eld. For bulge tests, DIC is an ideal tool to obtain a full 3D displacement eld
while not interfering with the ination or deection of the membrane material.
Material parameters in various models have been identied using FEMU (Lovato
et al., 1993; Cailletaud and Pilvin, 1994; Drass and Schneider, 2016). The process can
be generalized by using optimization to nd parameters in a FE model that minimize
the discrepancy between the model and experimental data (Asaadi et al., 2017). In
a forward problem parameters can be directly inferred from an experimental test.
While an inverse problem iteratively updates the FE model to nd parameters that
lead to responses that resemble the experimental response (Lee et al., 2019). This
process is often referred to as model calibration.
Model calibration can be performed using deterministic or probabilistic meth-
ods (Lee et al., 2019). A deterministic model calibration does not include the mea-
surement uncertainty, while probabilistic method can account for both model and
experimental uncertainty. Jung et al. (2016) stated that models calibrated determin-
isticly do not account for material uncertainty, and thus may result in unreliable
solutions. The uncertainty of the experiments needs to be quantied in order to cal-
ibrate material models probalistically. Unfortunately, DIC displacement eld data is
typically treated as a black box where the uncertainties are not well understood. The
bulge ination tests is particularly challenging having out-of-plane displacements on
the curved material surface. While it may be more reliable to calibrate the mate-
rial model parameters using a probabilistic method, this work uses a deterministic
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 3
method because the uncertainty due to DIC errors of the displacement measurements
is currently unknown. However, the experimental measurements are made available,
so that a probabilistic calibration can be performed in the future.
There are many dierent choices of possible objective functions to calibrate ma-
terial models. Zhan et al. (2011) proposed the EARTH method to calibrate model
parameters into phase, magnitude, and slope errors from time history data. A sin-
gle DIC bulge ination test may generate over a hundred thousand unique time
histories. In order to apply EARTH, each of these time histories would need to be
broken into their respective errors and treated as separate objective functions in a
multi-objective optimization. This process to compute the objective functions may
be computationally expensive with lots of full-eld data.
The previous studies (Rachik et al., 2001; Charalambides et al., 2002) used only
the height to determine material parameters from bulge ination tests. Machado
et al. (2012) used curvatures to determine a stress tensor from bulge ination tests
to infer material parameters for an elastic material. While these previous methods
work well for a planar isotropic material, they are dicult to extend to a directional
dependent woven textile like PVC-coated polyester. Jekel et al. (2016) showed that
it was possible to use the displacement eld of a bulge test to select parameters for
a non-linear orthotropic material model using simulated experimental data. Jekel
(2016) then used this process on polynomial displacement elds tted to experimental
data at selected ination pressures. These polynomial surfaces introduce an extra
layer of error in the material parameters. This work will nd material parameters by
tting the full experimental displacement eld directly. A new approach is describe
which allows for any data point in the experimental displacement eld be compared
to a FE model, for any initial surface location and ination pressure. While (Jekel
et al., 2016) considered the discrepancy in the x,yand zdisplacements elds to be
equal weight, the objective function was dominated by the z displacements which
were an order of magnitudes larger than the xand ycomponents. Therefore, this
work explores the eect of weighting down the zdiscrepancy.
This work also investigates using cross validation to select the best material
model when performing FEMU. The idea is that cross validation could be used
to determine whether the isotropic or orthotropic material model generalizes the
PVC-coated polyester better in the complex load case. This is especially important
for PVC-coated polyester, which is not a homogeneous (or continuous) material,
but modeled as one in the FE method. Cross validation was performed using the
two dierent objective functions described to select material parameters from the
experimental data.
2 Methods
Bulge ination tests were performed using DIC on PVC-coated polyester.1A FE
model was created to replicate the boundary conditions of the bulge ination tests.
Two dierent objective functions are described. Each objective function represents
a dierent method to quantify the dierence between the experimental data and
1An online repository is available at https://github.com/cjekel/inv_bubble_opt which
includes the source code to perform the inverse analysis, test data, and procedures to reproduce
this work.
4 Charles F. Jekel et al.
DIC
200 mm
Pressure Sensor Line Pressure
Test
Fixture
Inflated
Material
Fig. 1 Bulge ination test overview.
FE model. The process to perform the inverse analysis is briey described. Cross
validation is also discussed as a method for selecting material models in this scheme.
2.1 Experimental tests
The bulge ination tests involve clamping a sample of membrane material into a
circular clamp. Pressure is then induced on one side of the material. The material
deections were recorded using DIC. The DIC system used was the StrainMaster
with DaVis (LaVision GmbH, 2014), which was capable of syncing the ination
pressure with the recorded images. A visual representation of a bulge ination test
is provided in Figure 1.
Four bulge ination tests were performed at Stellenbosch University in South
Africa. Details of the process and test xture are found in section 5.1 of (Jekel, 2016).
The diameter of the circular bulge test was 200 mm. The PVC-coated polyester tested
was Mehler Texnologies VALMEX®7318 (the same material tested in(Jekel et al.,
2017)). Four samples of material were cut from the same roll, with each sample being
a 250 mm square. Spray paint was added to the surface of each specimen, with a
random pattern, to increase the surface contrast of the material for DIC processing.
Each test was inated from zero to three bar, by manually opening a compressed air
valve to the bulge ination test xture. This resulted in each test being inated at
a unique load rate as seen in Figure 2. The internal pressure was recorded with a
Festo SPTE-P10R-S4-V-2.5K pressure transmitter.
Full eld bulge ination tests generated a large amount of data as shown in Ta-
ble 1. The xy-plane was oriented with the surface of the material prior to ination,
with the xdirection occurring in the warp material direction, while the ydirection
occurred in the ll material direction. The material was inated in the zdirection.
The number of data points ranged from test to test. The variation was largely de-
pendent on how nely the DIC data was processed. Here the largest test generated
nearly two million data points, while the smallest test generated only two hundred
thousand data points. Each data point represents a unique combination of ination
pressure pand initial x,ylocation. There are three deformations recorded for each
data point, represented as separate
x,
y,and
zvalues.
The DIC techniques were not perfect, as there are missing data points in some of
the test. This happens when correlations is lost between images, and becomes more
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 5
02468
Time, seconds
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Pressure, bar
Test 1
Test 2
Test 3
Test 4
Fig. 2 Pressure time curves from each bulge ination test.
Table 1 Number of unique (x,y,p)data points from each test and ination pressures.
Test # of data points
1 1,836,961
2 729,718
3 1,201,509
4 289,312
evident with larger deections. Additionally, the severity of missing data varies from
test to test. Figure 3 shows each x,ydata point at the last ination pressure of each
test, where each colored pixel is a DIC data point. The spacing of the x,ydata points
vary from test to test as a result of dierent DIC processing While Tests 1 and 2
have the highest density of data points, Test 1 doesn’t have any missing data points
while test 2 has several small holes on the surface of the material. The worst of the
missing data occurs with test 3, which has large holes on one quarter of the test.
Test 4 only has two small holes on the surface of the data, and also has the least
number of x,ydata points on the surface of the material.
2.2 Finite Element model
An implicit non-linear FE model was constructed in ABAQUS which resembled the
physical boundary conditions of the bulge ination test. Sheplak and Dugundji (1998)
described dierential equations to solve for the displacement eld of bulge ination
tests for isotropic and orthotropic material models, but (Sheplak and Dugundji, 1998)
was not used as the geometry is simple enough to quickly construct in any non-linear
FE code. The FE model uses an implicit solver, with 201 load steps between zero and
three bar. Each load step uses adaptive time stepping. The adaptive time stepping
allows the model to solve a single load-case in one step, or cut back to smaller
increments if needed. Over 900 linear Q4 membrane elements are used to represent
the surface of the material. The linear elements prevent nonphysical out-of-plane
force imbalances. The displacement eld at the nodes of the FE solver are exported
to be used in further calculations when the model is run with a given material model.
The displacement eld of the FE model needs to be computed several times
at various pressures which correspond with the pressures of the experimental data.
Either load steps can be added to the FE model at the exact experimental pressures,
or the displacement eld of the FE model can be interpolated to match the exact
6 Charles F. Jekel et al.
(a) Test 1 (b) Test 2
(c) Test 3 (d) Test 4
Fig. 3 Plots of the x,ydata points of each bulge ination test. The darker color is related to
having a higher density of data points.
experimental pressures. Interpolation was chosen since it didn’t involve manually
editing the input deck for each test, and resulted in a xed number of exported
displacement elds. Specically linear interpolation was used to evaluate the FE
model’s displacement at the nodes for the exact pressures of the experimental test.
Linear interpolation is use to solve the displacement eld
(x,y,p)by interpolating
the displacement eld at the two nearest pressures as
(x,y,p)
(x,y,p1)
pp1
=
(x,y,p2)
(x,y,p1)
p2p1
(1)
where p2and p1represent the nearest pressures of the FE models’ load steps.
Overall, the linear interpolation scheme proved very accurate when interpolating
between the FE model load steps. The linear interpolation accuracy was compared
to FE models with load steps halfway in-between the previous 201 load steps (rep-
resenting the worse possible interpolation case). The linear interpolation error was
negligible, with the interpolation error following on the order of single precision
(108mm) numerical noise. This level of single precision was the same level of pre-
cision used by the FE model.
Linear isotropic and orthotropic material models were investigated. An isotropic
model with one unknown parameter (stiness modulus E), and an isotropic model
with two unknown parameters (stiness modulus Eand the shear modulus G) were
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 7
considered. The orthotropic model was simplied as a three parameter mode, with
parameters for the stiness moduli (E1&E2), and the shear modulus (G12 ). The one
parameter isotropic model and the orthotropic model use a Poisson’s ratio of 0.24
as measured in (Jekel et al., 2017). Using a constant Poisson’s ratio simplies issues
with gradient magnitudes, as it is anticipated that gradients of Poisson’s ratio could
be orders of magnitude dierent than stiness moduli.
The displacement eld of the FE model for an orthotropic material model at
2.0 bar is shown in Figures 4 through 6. The maximum
zvalue is about ten times
larger than the
xor
yvalues. Radial Basis Functions (RBF) are used to interpolate
the displacements from the initial (x,y) node locations at each outputted pressure.
The RBFs are exact at the node locations, and result in a smooth displacement eld
from the linear four node FE elements. The InterpolateSimpleRBF object construct
these RBFs to the full displacement eld of the FE analysis. The RBFs were inspired
by the SciPy rbf function (Virtanen et al., 2020).
100 75 50 25 0 25 50 75 100
x
(mm)
100
75
50
25
0
25
50
75
100
y
(mm)
0
4
8
12
16
20
24
28
32
z
(mm)
Fig. 4 Displacement
zof FE model at 2.0 bar with orthotropic properties E1=0.8GPa,
E2=0.15 GPa, G12 =0.025 GPa, and
ν
12 =0.24.
100 75 50 25 0 25 50 75 100
x
(mm)
100
75
50
25
0
25
50
75
100
y
(mm)
2.4
1.8
1.2
0.6
0.0
0.6
1.2
1.8
2.4
x
(mm)
Fig. 5 Displacement
xof FE model at 2.0 bar with orthotropic properties E1=0.8GPa,
E2=0.15 GPa, G12 =0.025 GPa, and
ν
12 =0.24. Note the symmetry about x=0.
The radial basis functions are expressed as
y= (2)
8 Charles F. Jekel et al.
100 75 50 25 0 25 50 75 100
x
(mm)
100
75
50
25
0
25
50
75
100
y
(mm)
3.2
2.4
1.6
0.8
0.0
0.8
1.6
2.4
3.2
y
(mm)
Fig. 6 Displacement
yof FE model at 2.0 bar with orthotropic properties E1=0.8GPa,
E2=0.15 GPa, G12 =0.025 GPa, and
ν
12 =0.24. Note the symmetry about y=0.
where
A=
||x1x1||2||x2x1||2··· ||xnx1||2
||x1x2||2||x2x2||2··· ||xnx2||2
.
.
..
.
.....
.
.
||x1xn||2||x2xn||2··· ||xnxn||2
(3)
which follows a simple linear kernel (Broomhead and Lowe, 1988). The RBF param-
eters λare solved for a general input of x, output of y, with ndata points. Then ˆn
predictions are generated for new ˆxlocations as
ˆy=
ˆ
(4)
where
ˆ
A=
||x1ˆx1||2||x2ˆx1||2· · · ||xnˆx1||2
||x1ˆx2||2||x2ˆx2||2· · · ||xnˆx2||2
.
.
..
.
.....
.
.
||x1ˆxˆn||2||x2ˆxˆn||2··· ||xnˆxˆn||2
.(5)
For this problem, xrefers to the in-plane (x,y) location of the FE model nodes, ˆx
refers to the in-plane (x,y) location of the DIC data, ydenotes the displacement from
the FE model, and ˆyrepresents the experimental displacements. This type of RBF
is ideal for interpolating surfaces of FE models since it is exact at node locations,
and results in a smooth interpolation in-between the FE node locations.
2.3 Objective functions
The objective function which quanties the discrepancy between the physical bulge
ination tests and the FE model dierence is fundamental for the optimization in the
inverse analysis. The average absolute deviation between the xdisplacement of the
FE model and ination test is denoted r
x(j,β)for the jtest and βset of material
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 9
parameters. These average absolute deviations are expressed as
r
x(j,β) = 1
nj
nj
i=1
|
x(xi,yi,pi)t
x(xi,yi,pi,β)f|(6)
r
y(j,β) = 1
nj
nj
i=1
|
y(xi,yi,pi)t
y(xi,yi,pi,β)f|(7)
r
z(j,β) = 1
nj
nj
i=1
|
z(xi,yi,pi)t
z(xi,yi,pi,β)f|(8)
where njis the total number of data points in the jtest. A simple discrepancy is
then formulated as the average L1distance in mm as
e(β) = 1
nt
nt
j
r
x(j,β) + r
y(j,β) + r
z(j,β)(9)
where ntis the total number of tests. The subscripts tis for the physical ination
test data, while the subscript fis from the FE model.
The formulation of econsiders the discrepancy in the x,y, and zdirections to
have an equal weight. This formulation could be problematic if the discrepancy in one
displacement components dominates the others. In the bulge ination test data, the
zcomponent was roughly ten times larger than the in-plane (xor y) components.
This creates the potential for the r
zdiscrepancies to be larger than the other two
directions, because the
zvalues have the potential to be at least ten times larger
than
xor
y.
A second objective function is proposed as ewto deal with the imbalance between
the maximum in-plane and out-of-plane displacements. The function is just a slight
modication of e, and is expressed as
ew(β) = 1
nt
nt
j
r
x(j,β) + r
y(j,β) + wr
z(j,β)(10)
where wis a weighting component2. While there can be many ways to select w, a
simple scheme was chosen as
wz=1
nt
nt
j
max
x(j)+max
y(j)
2 max
z(j)(11)
which represents the ratio of the average xand ydisplacement to the maximum z
displacement. This resulted in w=0.1for the bulge ination tests in consideration.
Two dierent objective functions are presented to quantify the dierence between
the full displacement eld of the FE model and DIC tests. One objective function is
an average L1norm between the x,y,zdisplacements, while the other function is a
relative L1which considers the weighted dierence between x,yand zdisplacements.
A zero for both functions would indicate that the FE model’s displacement eld
exactly matches the experimental data.
The optimization requires the objective function eor ewto be computed multiple
times. There are many steps required to automate this using software, and the process
2The software available online allows a weight to be specied for each directional component
of the displacement eld.
10 Charles F. Jekel et al.
is described in Table 2. Several Python functions were created to interface with the
the ABAQUS solver and post processor for this application. The function essentially
returns eor ewfrom inputted material parameters.
Table 2 Process to compute the objective function for given material parameters.
Step Description
1 Write the material model parameters to the ABAQUS input le
2 Run ABAQUS solver on the input le
3 Run ABAQUS post processor to export displacement eld of FE model
4 Load the FE displacement eld into memory
5 Compute the discrepancy between FE model and DIC data:
i) Linearly interpolate the FE model to match the pressures of the bulge test data
ii) Construct and evaluate RBFs to the FE model displacement eld
iii) Compute r
x,r
y, and r
zfor each set of test data
6 Compute the nal objective function of eor ew
2.4 Optimization
The inverse analysis is the process of nding the material parameters of the FE model
to match the bulge ination test data. The optimization problem can be stated as
minimize: e(β)(12)
subject to:
β
l
β
k
β
u,k=1,2,···,np.(13)
where βis the vector of material parameters which are restricted to some reasonable
lower and upper bounds. The isotropic parameters are expressed as β= (E)or β=
(E,G), and the simplied orthotropic parameters are expressed as β= (E1,E2,G12).
Note that ewis substituted for ewhen minimizing the weighted objective function.
The rst optimization strategy used a global optimizer with an allocated number
of function evaluations. When the global optimizer exhausted the specied budget,
a local optimizer was used for nal convergence to a local optimum. The global
optimization strategy used was Ecient Global Optimization (EGO), which utilized
the expected improvement from a Gaussian process to minimize the function (Jones
et al., 1998; The GPyOpt authors, 2016). A variant of the BFGS (Broyden, 1970;
Fletcher, 1970; Goldfarb, 1970; Shanno, 1970) gradient based optimization was used
as the local optimizer (Virtanen et al., 2020; Byrd et al., 1995). Initially 50 EGO
function evaluations (calculations of e) were performed before switching to the BFGS
implementation. A budget of 200 function evaluations for the BFGS appeared to be
sucient at nding a local optimum.
After several runs of the EGO to BFGS strategy, it became evident that the
optimizer was failing at nding a global optimum. A multi-start optimization strat-
egy was adopted to better deal with the presence of multiple local minima. The
multi-start process ran ve BFGS optimizations from dierent starting points in
the design space (Schutte et al., 2006). Each of the ve BFGS runs were limited
to either 200 function evaluations, or satisfying the convergence criteria. Converge
considered either relative changes in the objective function, absolute changes in the
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 11
objective function, or gradient magnitude. This multi-start optimization was able to
consistently nd better optimums than the EGO to BFGS strategy, but at the cost
of additional objective functions.
There is the possibility that some combination of material parameters may cause
the FE analysis to not converge. This is problematic when the optimization algo-
rithm requires a discrepancy for a particular set of parameters that lead to failed
convergence. To deal with this problem, the maximum objective value from the run-
time history was passed to the optimization algorithm when the FE analysis failed to
converge. Additionally, a discrepancy value of 30 mm was passed if the rst function
evaluation in a given run failed to converge. This strategy works well with EGO,
however it creates a non-dierential objective function which can be problematic for
gradient based optimization algorithms. There are also problems with the use of a
L1based objective function, where gradients can vanish when the residual of a single
points goes to zero.
While there were a number of potential issues using gradient based optimization
with this application, in practice the gradient optimization was able to successfully
minimize the objective function. Looking at the optimization history, the FE analysis
would only fail to converge during the line search stage of the gradient based opti-
mization algorithm, and not the nite dierences which approximate the gradients.
This is less problematic because the gradients were accurate. Lastly, there were no
observed issues with the optimization caused by the L1objective function.
2.5 Cross validation
Cross validation is a model selection or validation tool used in various regression
problems to assess the quality of models (Queipo et al., 2005). Cross validation
provides for a nearly unbiased estimate of the modeling error, and can be used to
diagnose overtting or bias errors. Cross validation can be used to compare the
performance of one material model to another in the context of tting material
models with an inverse analysis (FEMU). This may be important in practice when
the ideal material model is unknown. In this case, cross validation will be used to
quantitatively compare how the linear isotropic and orthotropic material models
represent the behavior of PVC-coated polyester from these bulge ination tests.
The processes proposed is similar to leave-one-out cross validation, and is de-
scribed in Table 3. This cross validation score was computed for both the linear
isotropic and orthotropic material models. The model with the lower cross validation
score is assumed to be a better generalized representation of the material behavior.
Table 3 Process to compute leave-one-test-out cross validation error.
Step Description
1 Perform optimization without test j
2 Calculate the discrepancy eor ewon the left-out test j
3 Repeat 1 & 2 for all tests
4 Cross validation score is the average discrepancy eor ewfrom the left-out tests
12 Charles F. Jekel et al.
3 Results
Inverse analyses were performed on the bulge ination tests to nd material pa-
rameters for the linear isotropic and linear orthotropic models. The results show
parameters when each test were t separately (such that nt=1). Lastly, the cross
validation errors were computed by tting parameters to all combinations of the
three tests (nt=3). The focus of the results is to demonstrate the eect of the two
objective functions.
3.1 Linear isotropic material model
Resulting parameters for the single parameter (E) linear isotropic material model are
found in Table 4, and results for the two parameter model (Eand G) are shown in
Table 5. The one parameter model used a Poisson’s ratio of 0.24, while optimization of
the two parameter resulted in a Poisson’s ratio near 0.5. Eectively the shear modulus
in the two parameter model approached the lower limit3, and the FE model is unable
to run when the Poisson ratio exceed 0.5. Poisson’s ratio appears to signicantly
eect the stiness modulus E, in which the one parameter model resulted in a larger
stiness modulus. For both models there was little dierence between parameters
from minimizing eor ewwith the rst test. The second and third test produced
larger stiness moduli when minimizing ew, and the degree of the increase was larger
with the one parameter model. In general, it appeared that minimizing ewresulted in
stier material parameters. Parameters of the two parameter isotropic model appears
less eected by the weighted objective function than the single objective function.
Table 4 One parameter isotropic material results from each inverse analysis. Note
ν
was xed
to 0.24.
Minimizing eMinimizing ew
E(GPa) E(GPa)
Test 1 0.279 0.283
Test 2 0.222 0.292
Test 3 0.242 0.282
Test 4 0.218 0.253
Table 5 Resulting isotropic material parameters from each inverse analysis. Note
ν
is calculated
from Eand G.
Minimizing eMinimizing ew
E(GPa) G(GPa)
ν
E(GPa) G(GPa)
ν
Test 1 0.279 0.113 0.24 0.283 0.114 0.24
Test 2 0.160 0.054 0.48 0.170 0.057 0.48
Test 3 0.162 0.054 0.50 0.170 0.057 0.49
Test 4 0.154 0.052 0.48 0.155 0.052 0.49
The objective values from the various ts are shown in Table 6. For both objective
functions, the fourth test resulted in the smallest objective values (or the best t).
3The lower limit of Gis E/3, and Poisson’s ratio is expressed as
ν
=E/(2G)1.
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 13
The largest objective values (or the worst t test) depends upon which objective
function was minimized, but appears to be either the second or third test. In most
cases, the two parameter model resulted in lower objective function values than the
one parameter model.
Table 6 Resulting objective values when tting the linear isotropic models to each bulge in-
ation test.
One parameter (E) Two parameter (E,G)
e(mm) ewe(mm) ew
Test 1 1.554 0.497 1.554 0.497
Test 2 1.881 0.795 1.699 0.725
Test 3 1.870 0.588 1.716 0.549
Test 4 1.195 0.457 1.054 0.403
3.2 Linear orthotropic material model
The simplied linear orthotropic material parameters resulting from inverse analysis
on the individual tests are shown in Table 5. The parameters of the orthotropic model
are very dierent depending upon which objective function was minimized. There are
sizable changes to both E1and E2depending upon whether eor ewwas minimized.
The most interesting changes occur in the second and third test, where the choice of
objective function reversed the stiness directions. Minimizing eresulted in E2>E1,
but minimizing ewresulted in E1>E2. Overall, minimizing ewresulted in parameters
that were more consistent from test to test which is expected since the test material
was nominally identical. While the stiness moduli were very dierent depending on
eor ew, the shear modulus was nearly the same in all conditions.
Table 7 Resulting orthotropic material parameters from minimizing tests independently with
each inverse analysis. Note that
ν
12 was xed to 0.24.
Minimizing e(GPa) Minimizing ew(GPa)
E1E2G12 E1E2G12
Test 1 0.343 0.248 0.005 0.303 0.229 0.005
Test 2 0.212 0.241 0.004 0.306 0.230 0.005
Test 3 0.217 0.257 0.004 0.306 0.229 0.005
Test 4 0.239 0.215 0.004 0.280 0.215 0.005
The objective values resulting from each parameter set with the linear orthotropic
model are shown in Table 8. The table shows the resulting values of ewwhen e
was minimized, and vice versa. The denition of the worse t depends upon which
objective function was minimized. In all cases, the value of ewwas worse when e
was minimized than when ewwas minimized. A similar statement can be made
for minimizing ew. If we consider test two, the objective values when minimizing
echanged up to 15% when minimizing ewwhile E1changed over 40% when the
objective function was changed.
Comparisons of the dierences between the two parameter isotropic material
model and the linear orthotropic material model are shown in Figures 7 through
14 Charles F. Jekel et al.
Table 8 Objective values when minimizing eor ewfor the linear orthotropic model to each
bulge ination test.
Minimizing e(GPa) Minimizing ew(GPa)
e(mm) ewe(mm) ew
Test 1 1.550 0.512 1.773 0.490
Test 2 1.660 0.765 1.917 0.710
Test 3 1.702 0.570 1.759 0.527
Test 4 1.033 0.406 1.101 0.380
10. The chosen displacement locations occur at the approximate maximums for the
linear orthotropic FE model, as previously shown in Figures 4 through 6. The
x
displacements occur at [x=56,y=0], the
ydisplacements occur at [x=0,y=63],
and the
zdisplacements occur at [x=0,y=0]on the surface of the material. The
displacements were plotted with the ination pressure. The results of the one param-
eter isotropic material model have been omitted, because they were nearly identical
to the two parameter isotropic results with the same objective function.
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Pressure, bar
0
1
2
3
4
x
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(a)
xdisplacement
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Pressure, bar
0
1
2
3
4
5
6
y
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(b)
ydisplacement
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Pressure, bar
5
10
15
20
25
30
35
40
z
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(c)
zdisplacement
Fig. 7 Resulting displacements from the two parameter isotropic and the linear orthotropic
material models compared with test 1. The
xresults shown in A) occur at [x=56,y=0]. The
yresults shown in B) occur at [x=0,y=63]. The
zresults shown in C) occur at [x=0,y=0].
The linear orthotropic FE model appears to match the selected test data better
than the isotropic model in all of the cases presented in Figures 7 through 10. Al-
though it is unclear whether the choice of eor ewas the objective function resulted
in a better ts. It appears that the linear orthotropic model with ewas the objective
function matched the
xand
ydisplacements better, while the eobjective function
matched the
zdisplacements better. There are exceptions to both of these cases,
where the reverse is seen, depending on which test is considered.
The dierences between eand ewfor the two parameter isotropic model were
very subtle. However, there is a more noticeable dierence between eand ewfor
the linear orthotropic material model. This is most evident in the maximum
z
displacement pressure curves for all tests. The non-linearity of the
zdisplacements
is more prevalent than the
xor
ydisplacements.
3.3 Cross validation material model comparison
Inverse analyses were performed to t isotropic and orthotropic material models
to the bulge ination tests. Additional inverse analyses were performed such that
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 15
(a)
xdisplacement
0.0 0.5 1.0 1.5 2.0
Pressure, bar
0.5
1.0
1.5
2.0
2.5
3.0
3.5
y
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(b)
ydisplacement
0.0 0.5 1.0 1.5 2.0
Pressure, bar
10
15
20
25
30
35
40
z
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(c)
zdisplacement
Fig. 8 Resulting displacements from the two parameter isotropic and the linear orthotropic
material models compared with test 2. The
xresults shown in A) occur at [x=56,y=0]. The
yresults shown in B) occur at [x=0,y=63]. The
zresults shown in C) occur at [x=0,y=0].
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Pressure, bar
0
1
2
3
4
x
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(a)
xdisplacement
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Pressure, bar
0
1
2
3
4
5
y
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(b)
ydisplacement
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Pressure, bar
5
10
15
20
25
30
35
40
45
z
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(c)
zdisplacement
Fig. 9 Resulting displacements from the two parameter isotropic and the linear orthotropic
material models compared with test 3. The
xresults shown in A) occur at [x=56,y=0]. The
yresults shown in B) occur at [x=0,y=63]. The
zresults shown in C) occur at [x=0,y=0].
(a)
xdisplacement
0.0 0.5 1.0 1.5 2.0 2.5
Pressure, bar
0
1
2
3
4
5
y
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(b)
ydisplacement
0.0 0.5 1.0 1.5 2.0 2.5
Pressure, bar
5
10
15
20
25
30
35
40
z
Displacement, mm
Two Iso. Obj:
e
Two Iso. Obj:
ew
Lin. Ortho. Obj:
e
Lin. Ortho. Obj:
ew
Test data
(c)
zdisplacement
Fig. 10 Resulting displacements from the two parameter isotropic and the linear orthotropic
material models compared with test 4. The
xresults shown in A) occur at [x=56,y=0]. The
yresults shown in B) occur at [x=0,y=63]. The
zresults shown in C) occur at [x=0,y=0].
a leave-one-test-out cross validation score was computed for each material model.
The resulting discrepancy values are presented in Table 9. The orthotropic material
model had the lowest cross validation error when ewwas minimized, but the two
parameter isotropic material model had the lowest cross validation error when e
was minimized. The dierent objective functions appear to prefer dierent material
models. This implies that the material model which generalizes the bulge ination
test better depends on the denition of the discrepancy between the test data and
FE model.
The two parameter isotropic material model parameters from the full inverse
analysis and the cross validation runs are presented in Table 10. When ewas min-
16 Charles F. Jekel et al.
Table 9 Resulting discrepancy from the inverse analysis and leave-one-test-out cross validation.
Cross validation error Left out test values
Model e(mm) ewe[1, 2, 3, 4] ew[1, 2, 3, 4]
isotropic (E) 1.995 0.590 [2.47, 2.10, 1.87, 1.54] [0.62, 0.52, 0.59, 0.63]
isotropic (Eand G) 1.833 0.554 [2.37, 1.85, 1.72, 1.40] [0.57, 0.50, 0.56, 0.60]
orthotropic 1.856 0.533 [2.47, 1.88, 1.73, 1.35] [0.49, 0.72, 0.53, 0.40]
imized, the stiness modulus varied from 0.155 to 0.170 GPa. However, when ew
was minimized the stiness modulus varied from 0.170 to 0.193 GPa. and the shear
modulus varied from 0.057 to 0.070 GPa. Overall, ewresulted in more consistent and
stier moduli.
Table 10 Resulting two parameter isotropic material parameters from each inverse analysis.
Minimizing e(GPa) Minimizing ew(GPa)
E G E G
Leaving test 1 out 0.155 0.052 0.170 0.057
Leaving test 2 out 0.167 0.056 0.170 0.057
Leaving test 3 out 0.164 0.055 0.170 0.056
Leaving test 4 out 0.170 0.057 0.193 0.066
The orthotropic material parameters from the cross validation study are shown
in Table 11. There is a signicant dierence between the parameter variance depend-
ing upon which objective function was used. For instance, E1varied from 0.224 to
0.280 GPa when ewas minimized. However, when ewwas minimized E1varied from
0.303 to 0.305 GPa. A similar trend occurs for E2and G12, where minimizing ew
resulted in more consistent parameters.
Table 11 Resulting orthotropic material parameters from each inverse analysis.
Minimizing e(GPa) Minimizing ew(GPa)
E1E2G12 E1E2G12
Leaving test 1 out 0.224 0.235 0.003 0.303 0.229 0.005
Leaving test 2 out 0.297 0.231 0.004 0.304 0.229 0.005
Leaving test 3 out 0.284 0.224 0.004 0.304 0.229 0.005
Leaving test 4 out 0.280 0.244 0.003 0.305 0.229 0.005
4 Discussion
The most striking result was the eect of the two objective functions on selecting
parameters for the linear orthotropic material model. When ewas minimized the
linear orthotropic parameters varied signicantly from test to test, with some tests
resulting in E2>E1and others E1>E2. This was not the case when ewwas mini-
mized, in which the parameters from test to test were fairly consistent with E1>E2.
The dierence between the two objective functions was that econsidered the dis-
crepancies in the x,yand zdirections to be of equal weight, while ewconsidered the
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 17
zdiscrepancies to be one tenth the weight. This weighting factor corresponds to an
imbalance between the test maximum zdisplacement being about ten times larger
than the maximum xor ydisplacements. It’s important to clarify that both eand
ewwould approach zero if it was possible for a perfect t, however resulting material
parameters occur with the inability to perfectly t the data. In these circumstances
it appears that one may need to careful consider how to eventuate the discrepancy
of the displacement eld, especially under similar imbalanced displacement data.
Tests two and three are of particular interest with the linear orthotropic model,
because these tests resulted in E2>E1when ewas minimized. It could be reasonable
to assume that the test directions were incorrect by a factor of 90, and that this
mistake in the xand ydirections resulted in the set of parameters. Additional inverse
analyses were performed on these tests, where the test data was rotated by factors
of 45and 90. What is interesting is that when ewas minimized the resulting
parameters resulted in E2>E1regardless of the rotation, and there was little change
to the parameters. However, when ewwas minimized both E2and E1would change
signicantly based on the rotation. This hints that giving more weight to the xand
ydisplacements, like in ew, would be better if an inverse analyses was required to
identify orthotropic parameters without knowing the primary and secondary material
directions.
The dierent objective functions had an interesting eect on using leave-one-
test-out cross validation for material model selection. When ewas minimized, the
cross validation error favored the two parameter isotropic material model. Though
when ewwas minimized, the cross validation error favored the orthotropic material
model by a much larger margin. This reemphasizes the importance of selecting an
appropriate objective function, as the choice of objective function not only eects
the resulting parameters, but also eects the perceived generalization error of the
model.
5 Conclusion
An inverse analysis was described to nd material parameters by matching the full
displacement eld from bulge ination tests. Optimization was used to nd material
parameters in a FE model that best matched the experimental displacement eld
from tests on PVC-coated polyester. Material parameters were determined for a lin-
ear isotropic and simplied linear orthotropic material models. Two dierent objec-
tive functions were considered to describe the discrepancy between the experimental
data and numerical model. The rst objective function considered equal weight be-
tween the displacement components, while the other function gave more weight to
the xand ydisplacements. The weighting scheme was chosen to compensate for the
fact that the majority of the deections within the bulge ination test occur out-of-
plane. Resulting material parameters for the linear orthotropic material model were
very dierent depending on which objective function was minimized. Thus the choice
of objective function being considered is very important when performing such an
optimization on the full eld data.
Cross validation was performed to determine whether the isotropic or orthotropic
material model was a better representation of the material behavior. There was little
dierence in the cross validation error according to the equally weighted objective
function, however the weighted objective function heavily favored the orthotropic
18 Charles F. Jekel et al.
material model. This indicates that the choice of objective function was not only
important in material parameter selection, but can also impact material model se-
lection.
Acknowledgements Thanks to Sudharshan Udhayakumar for help constructing the FE model
and comparing the FE model to analytical solutions. Thanks to Andrés Bernardo for helping
set up script to process the full displacement eld data.
Replication of results
The FE models require a commercial code to run. Experimental data, Python scripts
to perform the optimization, and FE model input decks are available online with
instructions at https://github.com/cjekel/inv_bubble_opt
Declarations
Funding
Charles F. Jekel has received the following funding for his PhD research which has
supported this work: University of Florida Graduate Preeminence Award, U.S. De-
partment of Veterans Aairs Educational Assistance, and Stellenbosch University
Merrit Bursary.
Conicts of interest
The authors declare that they have no conict of interest.
Availability of data and material
The experimental data is available online with instructions on how to download at
https://github.com/cjekel/inv_bubble_opt
Code availability
The scientic Python ecosystem was used to produce these results. Specic scripts
used to perform the optimizations are available online at https://github.com/
cjekel/inv_bubble_opt
References
Ambroziak A, Kłosowski P (2014) Mechanical properties for preliminary design of
structures made from PVC coated fabric. Construction and Building Materials
50:74–81, DOI 10.1016/j.conbuildmat.2013.08.060
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 19
Asaadi E, Wilke DN, Heyns PS, Kok S (2017) The use of direct inverse maps to
solve material identication problems: pitfalls and solutions. Structural and Mul-
tidisciplinary Optimization 55(2):613–632, DOI 10.1007/s00158-016-1515-1, URL
https://doi.org/10.1007/s00158-016-1515-1
Becker TH, Mostafavi M, Tait RB, Marrow TJ (2012) An approach to calculate
the J-integral by digital image correlation displacement eld measurement. Fa-
tigue & Fracture of Engineering Materials & Structures 35(10):971–984, DOI 10.
1111/j.1460-2695.2012.01685.x, URL https://onlinelibrary.wiley.com/doi/
abs/10.1111/j.1460-2695.2012.01685.x
Broomhead DS, Lowe D (1988) Radial basis functions, multi-variable functional
interpolation and adaptive networks. Tech. rep., Royal Signals and Radar Estab-
lishment Malvern (United Kingdom)
Broyden CG (1970) The Convergence of a Class of Double-rank Minimization Algo-
rithms 1. General Considerations. IMA Journal of Applied Mathematics 6(1):76–
90, DOI 10.1093/imamat/6.1.76, URL https://dx.doi.org/10.1093/imamat/
6.1.76
Byrd RH, Lu P, Nocedal J, Zhu C (1995) A limited memory algorithm for bound
constrained optimization. SIAM Journal on Scientic Computing 16(5):1190–1208
Cailletaud G, Pilvin P (1994) Identication and inverse problems related to material
behaviour. Inverse problems in engineering mechanics 1:79–86
Charalambides M, Wanigasooriya L, Williams G, Chakrabarti S (2002) Biaxial defor-
mation of dough using the bubble ination technique. I. Experimental. Rheologica
Acta 41(6):532–540, DOI 10.1007/s00397-002-0242- 2
Chisari C, Macorini L, Amadio C, Izzuddin BA (2017) Optimal sensor placement
for structural parameter identication. Structural and Multidisciplinary Optimiza-
tion 55(2):647–662, DOI 10.1007/s00158-016-1531-1, URL https://doi.org/10.
1007/s00158-016-1531-1
Dinh TD, Rezaei A, Daelemans L, Mollaert M, Hemelrijck DV, Paepegem
WV (2017) A hybrid micro-meso-scale unit cell model for homogenization of
the nonlinear orthotropic material behavior of coated fabrics used in ten-
sioned membrane structures. Composite Structures 162:271–279, DOI https:
//doi.org/10.1016/j.compstruct.2016.12.027, URL http://www.sciencedirect.
com/science/article/pii/S0263822316313204
Drass M, Schneider J (2016) On the mechanical behavior of transparent structural
silicone adhesive–TSSA. In: SEMC 2016–Sixth International Conference on Struc-
tural Engineering, Mechanics and Computation, pp 14–16
Fletcher R (1970) A new approach to variable metric algorithms. The Computer
Journal 13(3):317–322, DOI 10.1093/comjnl/13.3.317, URL https://dx.doi.
org/10.1093/comjnl/13.3.317
Galliot C, Luchsinger R (2009) A simple model describing the non-linear biaxial
tensile behaviour of PVC-coated polyester fabrics for use in nite element analysis.
Composite Structures 90(4):438–447, DOI 10.1016/j.compstruct.2009.04.016
Goldfarb D (1970) A family of variable-metric methods derived by varia-
tional means. Math Comp 24 (1970), 23-26 DOI https://doi.org/10.1090/
S0025-5718-1970-0258249-6
Jekel CF (2016) Obtaining non-linear orthotropic material models for PVC-coated
polyester via inverse bubble ination. Master’s thesis, Stellenbosch University,
URL http://hdl.handle.net/10019.1/98627
20 Charles F. Jekel et al.
Jekel CF, Venter G, Venter MP (2016) Obtaining a hyperelastic non-linear or-
thotropic material model via inverse bubble ination analysis. Structural and
Multidisciplinary Optimization pp 1–9, DOI 10.1007/s00158-016-1456-8, URL
http://dx.doi.org/10.1007/s00158-016-1456-8
Jekel CF, Venter G, Venter MP (2017) Modeling PVC-coated polyester as a hypoelas-
tic non-linear orthotropic material. Composite Structures 161:51–64, DOI http://
dx.doi.org/10.1016/j.compstruct.2016.11.019, URL http://www.sciencedirect.
com/science/article/pii/S0263822316320839
Jekel CF, Venter G, Venter MP, Stander N, Haftka RT (2019) Similarity mea-
sures for identifying material parameters from hysteresis loops using inverse anal-
ysis. International Journal of Material Forming 12(3):355–378, DOI 10.1007/
s12289-018-1421-8, URL https://doi.org/10.1007/s12289-018-1421- 8
Jones DR, Schonlau M, Welch WJ (1998) Ecient Global Optimization of Expen-
sive Black-Box Functions. Journal of Global Optimization 13(4):455–492, DOI
10.1023/A:1008306431147, URL https://doi.org/10.1023/A:1008306431147
Jung BC, Yoon H, Oh H, Lee G, Yoo M, Youn BD, Huh YC (2016) Hierarchi-
cal model calibration for designing piezoelectric energy harvester in the pres-
ence of variability in material properties and geometry. Structural and Multi-
disciplinary Optimization 53(1):161–173, DOI 10.1007/s00158-015-1310-4, URL
https://doi.org/10.1007/s00158-015-1310-4
LaVision GmbH (2014) Product-Manual DaVis 8.2 Software, 8th edn. Göttingen,
Germany
Lee G, Kim W, Oh H, Youn BD, Kim NH (2019) Review of statistical model
calibration and validation—from the perspective of uncertainty structures.
Structural and Multidisciplinary Optimization 60(4):1619–1644, DOI 10.1007/
s00158-019-02270-2, URL https://doi.org/10.1007/s00158-019-02270- 2
Lovato G, Moret F, Le Gallo P, Cailletaud G, Pilvin P (1993) Determination of
brazed joint constitutive law by inverse method. Le Journal de Physique IV
3(C7):C7—-1135
Machado G, Favier D, Chagnon G (2012) Membrane curvatures and stress-strain
full elds of axisymmetric bulge tests from 3D-DIC measurements. Theory and
validation on virtual and experimental results. Experimental mechanics 52(7):865–
880
Mejía CA, Lantsoght EOL (2016) Strain and deection analysis in plain concrete
beams and reinforced concrete beams by applying Digital Image Correlation. In:
SEMC 2016–Sixth International Conference on Structural Engineering, Mechanics
and Computation, CRC Press
Murienne BJ, Nguyen TD (2016) A comparison of 2D and 3D digital image
correlation for a membrane under ination. Optics and Lasers in Engineer-
ing 77:92–99, DOI https://doi.org/10.1016/j.optlaseng.2015.07.013, URL http:
//www.sciencedirect.com/science/article/pii/S0143816615001888
Queipo NV, Haftka RT, Shyy W, Goel T, Vaidyanathan R, Kevin Tucker
P (2005) Surrogate-based analysis and optimization. Progress in Aerospace
Sciences 41(1):1–28, DOI 10.1016/j.paerosci.2005.02.001, URL http://www.
sciencedirect.com/science/article/pii/S0376042105000102
Rachik M, Schmidtt F, Reuge N, Le Maoult Y, Abbeé F (2001) Elastomer bi-
axial characterization using bubble ination technique. II: Numerical investiga-
tion of some constitutive models. Polymer Engineering & Science 41(3):532–
541, DOI 10.1002/pen.10750, URL https://onlinelibrary.wiley.com/doi/
Eect of Weighting Full-eld Residuals when Fitting Material Model Parameters 21
abs/10.1002/pen.10750
Schutte JF, Haftka RT, Fregly BJ (2006) Improved global convergence proba-
bility using multiple independent optimizations. International Journal for Nu-
merical Methods in Engineering 71(6):678–702, DOI 10.1002/nme.1960, URL
https://doi.org/10.1002/nme.1960
Shanno DF (1970) Conditioning of quasi-Newton methods for function min-
imization. Math Comp 24 (1970), 647-656 DOI https://doi.org/10.1090/
S0025-5718-1970-0274029-X
Shaw A, Sriramula S, Gosling PD, Chryssanthopoulos MK (2010) A critical reli-
ability evaluation of bre reinforced composite materials based on probabilistic
micro and macro-mechanical analysis. Composites Part B: Engineering 41(6):446–
453, DOI https://doi.org/10.1016/j.compositesb.2010.05.005, URL http://www.
sciencedirect.com/science/article/pii/S1359836810000818
Sheplak M, Dugundji J (1998) Large Deections of Clamped Circular Plates Un-
der Initial Tension and Transitions to Membrane Behavior. Journal of Applied
Mechanics 65(1):107–115, URL http://dx.doi.org/10.1115/1.2789012
Stranghöner N, Uhlemann J, Bilginoglu F, Bletzinger KU, Bögner-Balz H, Corne E,
Gibson N, Gosling P, Houtman R, Llorens J, Others (2016) Prospect for European
Guidance for the Structural Design of Tensile Membrane Structures. Science and
Policy Report (SaP-Report) Draft Version To be published by the Joint Research
Centre (JRC) of the European Commission, publication expected
Tam JH (2020) Identication of elastic properties utilizing non-destructive vi-
brational evaluation methods with emphasis on denition of objective func-
tions: a review. Structural and Multidisciplinary Optimization DOI 10.1007/
s00158-019-02433-1, URL https://doi.org/10.1007/s00158-019-02433- 1
The GPyOpt authors (2016) GPyOpt: A bayesian optimization framework in python.
http://github.com/SheffieldML/GPyOpt
Tonge TK, Atlan LS, Voo LM, Nguyen TD (2013) Full-eld bulge test for pla-
nar anisotropic tissues: Part I Experimental methods applied to human
skin tissue. Acta Biomaterialia 9(4):5913–5925, DOI https://doi.org/10.1016/
j.actbio.2012.11.035, URL http://www.sciencedirect.com/science/article/
pii/S1742706112005983
Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D,
Burovski E, Peterson P, Weckesser W, Bright J, van der Walt SJ, Brett M, Wilson
J, Jarrod Millman K, Mayorov N, Nelson ARJ, Jones E, Kern R, Larson E, Carey
C, Polat I, Feng Y, Moore EW, Vand erPlas J, Laxalde D, Perktold J, Cimrman
R, Henriksen I, Quintero EA, Harris CR, Archibald AM, Ribeiro AH, Pedregosa
F, van Mulbregt P, Contributors S (2020) SciPy 1.0: Fundamental Algorithms for
Scientic Computing in Python. Nature Methods DOI https://doi.org/10.1038/
s41592-019-0686-2
Zhan Z, Fu Y, Yang RJ, Peng Y (2011) An automatic model calibration method
for occupant restraint systems. Structural and Multidisciplinary Optimiza-
tion 44(6):815–822, DOI 10.1007/s00158-011-0671-6, URL https://doi.org/10.
1007/s00158-011-0671-6
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
SciPy is an open-source scientific computing library for the Python programming language. Since its initial release in 2001, SciPy has become a de facto standard for leveraging scientific algorithms in Python, with over 600 unique code contributors, thousands of dependent packages, over 100,000 dependent repositories and millions of downloads per year. In this work, we provide an overview of the capabilities and development practices of SciPy 1.0 and highlight some recent technical developments. This Perspective describes the development and capabilities of SciPy 1.0, an open source scientific computing library for the Python programming language.
Article
Full-text available
Defining a suitable objective function is important as it affects the solution search process and the solution quality. In vibrational material identification, a number of objective functions have been employed since the last decades. It is the intent of the present study to provide a comprehensive review of the previous and current trends of non-destructive vibrational material identification with emphasis on the definition of objective function. Prior to defining an objective function, several aspects should be given utmost attention, including selection of responses, design variables and goals. Along with the critical review of past research, potential areas that could be worth for future investigation are as well explicated.
Article
Full-text available
Computer-aided engineering (CAE) is now an essential instrument that aids in engineering decision-making. Statistical model calibration and validation has recently drawn great attention in the engineering community for its applications in practical CAE models. The objective of this paper is to review the state-of-the-art and trends in statistical model calibration and validation, based on the available extensive literature, from the perspective of uncertainty structures. After a brief discussion about uncertainties, this paper examines three problem categories—the forward problem, the inverse problem, and the validation problem—in the context of techniques and applications for statistical model calibration and validation.
Article
Full-text available
Sum-of-square based error formulations may be difficult to implement on an inverse analysis consisting of multiple tension-compression hysteresis loops. Five alternative measures of similarity between curves are investigated as useful tools to help identify parameters from hysteresis loops with inverse analyses. A new algorithm is presented to calculate the area between curves. Four additional methods are presented from literature, which include the Partial Curve Mapping value, discrete Fréchet distance, Dynamic Time Warping, and Curve Length approach. These similarity measures are compared by solving a non-linear regression problem resembling a single load-unload cycle. The measures are then used to solve more complicated inverse analysis, where material parameters are identified for a kinematic hardening transversely anisotropic material model. The inverse analysis finds material parameters such that a non-linear FE model reproduces the behavior from five experimental hysteresis loops. Each method was shown to find useful parameters for these problems, and should be considered a viable alternative when sum-of-square based methods may be difficult to implement. It is important to consider multiple similarity measures in cases when it is impossible to obtain a perfect match.
Article
Full-text available
The identification of model material parameters is often required when assessing existing structures, in damage analysis and structural health monitoring. A typical procedure considers a set of experimental data for a given problem and the use of a numerical or analytical model for the problem description, with the aim of finding the material characteristics which give a model response as close as possible to the experimental outcomes. Since experimental results are usually affected by errors and limited in number, it is important to specify sensor position(s) to obtain the most informative data. This work proposes a novel method for optimal sensor placement based on the definition of the representativeness of the data with respect to the global displacement field. The method employs an optimisation procedure based on Genetic Algorithms and allows for the assessment of any sensor layout independently from the actual inverse problem solution. Two numerical applications are presented, which show that the representativeness of the data is connected to the error in the inverse analysis solution. These also confirm that the proposed approach, where different practical constraints can be added to the optimisation procedure, can be effective in decreasing the instability of the parameter identification process.
Article
Full-text available
Material parameter identification is a technique that is used to calibrate material models, often a precursor to perform an industrial analysis. Conventional material parameter identification methods estimate the material parameters for a material model by solving an optimisation problem. An alternative but lesser-known approach, called a direct inverse map, directly maps the measured response to the parameters of a material model. In this study we investigate the potential pitfalls of the well-known stochastic noise and lesser-known model errors when constructing direct inverse maps. We show how to address these problems, explaining in particular the importance of projecting the measured response onto the domain of the simulated responses before mapping it to the material parameters. This paper concludes by proposing partial least squares regression as an elegant and computationally efficient approach to address stochastic and systematic (model) errors. This paper also gives insight into the nature of the inverse problem under consideration.
Article
Full-text available
An inverse bubble inflation test is proposed utilizing full displacement field matching to obtain non-linear material models suitable for the Finite Element (FE) method. In this paper a known non-linear orthotropic material model is assumed as the solution for the inverse method to illustrate the process. A bubble inflation FE analysis is performed with the known material model to determine the load and displacement field from the assumed material. Polynomial surfaces are fit to the nodal displacement values of the FE model, such that the entire displacement field is stored as three unique polynomial surfaces. An error formulation was established to quantify the quality of fit between different bubble inflation displacement fields. Gradient based optimization is used to obtain the assumed material model by matching the full displacement field. The inverse bubble inflation test successful produces a non-linear orthotropic model that is analogous to the assumed non-linear orthotropic material, and thus demonstrates that the inverse bubble inflation analysis would be able to characterize other non-linear orthotropic materials.
Article
Abstract Coated fabric materials is a kind of textile composites. However, at micro scale the coating material is not fully impregnated inside the fabric yarn. Thus, the fabric fibers are free to slip relatively to each other during deformation. This article proposes a novel three-dimensional hybrid micro-meso unit cell model to predict the nonlinear mechanical behavior of coated fabrics using the digital element method. In the model, while the coating material is modeled with 3D solid elements, the fabric yarns are modeled as assemblies of truss elements. Thus, the material constitutive law used in this model is far simpler than the one in which the yarn is modeled with solid elements and the interaction among the fibers can be considered. Different load cases are examined, including uniaxial and biaxial stress states. In these analyses the periodic boundary conditions are applied to the unit cell, and contact friction among the fibers is also considered. The results obtained from the analyses have a good correlation with the experimental data, which proves the validity of the model. Moreover, local mechanical deformation mechanisms of the coated fabric can be observed from these analyses, which provide a better understanding of the mechanical behavior of this kind of material.
Article
Piezoelectric energy harvesting which scavenges electric power from ambient vibration energy has received significant attention as an ultimate solution to realize self-powered wireless sensors. For designing a piezoelectric energy harvester, it is of great importance to develop a high-fidelity electromechanical model which predicts the output power under various vibration conditions. To the best of our knowledge, however, there has been no systematic approach to account for variability in the material properties and geometry of a piezoelectric energy harvester. This paper thus presents (1) the hierarchical model calibration to improve the predictive capability of the electromechanical model and (2) the design of energy harvesting (EH) skin to maximize the output power to reliably operate self-powered wireless sensors. In this study, the hierarchical model calibration infers statistical information of unknown model variables (compliance, piezoelectric strain coefficient, and relative permittivity). The calibrated electromechanical model is then used to design EH skin based on the piezoelectric material segmentation to avoid voltage cancellation. The output power predicted by the calibrated electromechanical model is statistically compared with the measured one. Finally, it is concluded from the feasibility demonstration that EH skin can sufficiently generate the output power to realize self-powered wireless sensors without batteries.