You are on page 1of 17

International Journal of Mechanical Sciences 214 (2022) 106891

Contents lists available at ScienceDirect

International Journal of Mechanical Sciences


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

Enhancing the information-richness of sheet metal specimens for inverse


identification of plastic anisotropy through strain fields
Y. Zhang 1, *, S. Gothivarekar 1, M. Conde 2, A. Van de Velde 1, B. Paermentier 1,
A. Andrade-Campos 2, S. Coppieters 1
1
Department of Materials Engineering, KU Leuven, Ghent Technology Campus, Gebroeders de Smetstraat 1, 9000 Gent, Belgium
2
Department of Mechanical Engineering, Centre for Mechanical Technology and Automation, University of Aveiro, 3810-193 Aveiro, Portugal

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

Keywords: Advanced anisotropic constitutive models involve a large number of model parameters requiring multiple
Heterogeneous mechanical tests conventional mechanical experiments for accurate calibration. Inverse identification through a single hetero­
Specimen design geneous experiment combined with full-field deformation fields enables to reduce the experimental effort. User-
Identifiability
dependency in design of such information-rich experiment can be eliminated via shape optimization that is
Material orientation
driven by an indicator expressing the degree of strain field heterogeneity. This approach is validated by assessing
Anisotropic yield criteria
Parameters identification the identification quality of the heterogeneous experiments through the Finite Element Model Updating (FEMU)
technique fed by synthetically generated Digital Image Correlation (DIC) data using a reference material model.
Positive correlation between the adopted heterogeneity indicator and the inverse identification quality is found
for the identification of the Hill 1948 yield criterion. It is shown that identification metrics can be used to
accurately predict the identification quality prior to the FEMU process. Moreover, it is shown that the identi­
fiability of the sought parameters can be significantly improved by considering the orientation of the principal
material axis in the specimen. The reliability and robustness of the proposed identifiability metrics are validated
using a more advanced anisotropic yield function, namely Yld2000-2D. The identification metrics can therefore
be used to enhance the design of information-rich experiments. Finally, it is recommended to treat the material
orientation as a design variable.

1. INTRODUCTION certain strain state, under monotonic loading conditions. To identify the
anisotropic parameters, several standard tests must be performed that
Sheet metals typically exhibit an orthotropic plastic material are usually time and cost consuming. To this end, the second approach
behavior related to the rolling process, causing different plastic behavior considers heterogeneous mechanical testing, where information-rich
along different directions [1]. When forming sheet metal, this plastic experimentation enables to reduce the experimental effort by calibrat­
anisotropy has a significant influence on the final geometry and residual ing the material model parameters from a single heterogeneous me­
stress of the formed part. In addition, it affects the predictive accuracy of chanical test. Such test generates multiple heterogeneous strain fields
metal forming simulations [2]. Therefore, numerous anisotropic plastic that can be captured by Digital Image Correlation (DIC) and used in an
constitutive models exist to precisely reproduce the anisotropic behavior inverse identification strategy. The Virtual Fields Method (VFM) [4] and
[3]. In general, the parameter identification quality for these models Finite Element Model Updating (FEMU) [5] are currently the main
depends on the quantity and quality of experimental test data at hand. methods to inversely calibrate anisotropic yield criteria from full field
Two mechanical approaches exist to acquire this data: quasi- data.
homogeneous and heterogeneous mechanical tests. First and foremost, Strong proof of concept is available showing that heterogeneous
several classical quasi-homogeneous mechanical tests can be performed, mechanical tests can reduce the effort for material parameters identifi­
such as uniaxial tension, simple shear and biaxial tension. These are cation [1, 5-14]. Moreover, the parameters identification quality from
considered as conventional sheet metal material tests. In practice, every complex heterogeneous tests potentially outperforms classical homo­
conventional test requires a specific specimen geometry that exhibits a geneous tests [5]. The reason for this observation comes from the fact

* Corresponding Author: Mr Y. Zhang, KU Leuven: Katholieke Universiteit Leuven, Belgium


E-mail address: yi.zhang@kuleuven.be (Y. Zhang).

https://doi.org/10.1016/j.ijmecsci.2021.106891
Received 31 May 2021; Received in revised form 18 October 2021; Accepted 22 October 2021
Available online 31 October 2021
0020-7403/© 2021 Elsevier Ltd. All rights reserved.
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

that heterogeneous experiments can be tailored towards a specific Table 1


application. Indeed, the stress and strain fields generated in homoge­ Constitutive formulations and Lankford value calculation.
nous mechanical tests generally do not resemble the complex defor­ Swift’s hardening law n
σeq = K(ε0 + εpl
eq ) (1)
mation states that occur in sheet metal forming processes. It is
Plane stress Hill48 yield criterion σeq = Fσ222 + H(σ11 − σ22 )2 + 2Nσ212
Gσ211
well-known that the validity of phenomenological plasticity models is
+
(2)
confined to situations that are comparable to the range of experiments ⎧
H H
Anisotropic parameters based on
used for calibrating these models. In this regard, heterogeneous me­


⎪ r0 = =
Lankford values ⎪

⎪ G 1− H
chanical testing aims at generating stress and strain fields that mimic

⎨ 2N − F − G 2N − F + H − 1
r45 = = (3)
those occurring during metal forming processes. In other words, a het­ ⎪


2(G + F) 2(1 − H + F)

erogeneous experiment can increase the parameter reliability of a ⎩ r90 = H



F
plasticity model, thereby promoting its performance in subsequent
forming simulations. Hence, the design of heterogeneous mechanical
testing could reduce the experimental effort by extracting multiple robustness of the viscoelastic properties extraction. Similarity, Renner
model parameters from a single experiment. et al. [24] used identifiabiltiy index to measure the information richness
In essence, a reduction of experimental effort is achieved when of the indentation curve and concluded that the index could quantita­
maximizing the identifiability of the sought parameters. Therefore, the tively assess parameters identification stability using an inverse method.
shape of the complex specimen is designed in such a way that all rele­ Thereby, the identifiability analysis is a potential useful methodology to
vant anisotropic mechanical effects are “activated”. While shape opti­ assess whether the design specimen has sufficient information for
mization is covered by a number of papers [15–19], most of them unique and reliable gradient-based inverse identification.
consider the design of a complex specimen geometry to obtain a het­ The aim of this paper is to analyze the importance of material
erogeneous strain field. Currently, the design of heterogeneous me­ orientation and parameter identifiability in the design of heterogeneous
chanical tests for the identification of material parameters can be experiments for the inverse identification of plastic anisotropy for sheet
divided into three distinct strategies [15]. The first strategy is based on metal. The starting point of the paper is an IT-based design of a notched
empirical experience and knowledge to design uniaxial tensile speci­ tensile specimen. The effectiveness of this design is investigated through
mens with varying cross-sections, perforations, internal and external FEMU for inversely identifying the anisotropic parameters of the Hill48
notches and other complicated shapes. These specimens are then used yield criterion and a more advanced anisotropic yield function, namely
for uniaxial tests [1, 3-10, 12-14], and non-standard biaxial tensile Yld2000-2D [27]. The assessment is based on synthetic DIC data
specimens [11]. However, this intuitive approach is mainly based on generated with the ground truth material behavior (also referred to as
engineering judgement, where the optimal identification quality is reference material behavior) that is adopted in a FE model of the
driven by trial and error methods as well as user-specific experience. For experiment. This implies that experimental errors are excluded whilst
example, Denys et al. [10] designed a double-perforated tensile spec­ the metrological aspects of the DIC code are included, yielding an honest
imen to identify the anisotropic parameters of the 3D Hill48 yield cri­ assessment of the identification quality. In addition to the FEMU anal­
terion through FEMU and full-field displacement data. ysis, a quantitative identifiability analysis of the IT-based design is
The second design strategy relies on a quantitative evaluation of a introduced in this paper. The practical identifiability analysis enables to
heterogeneity indicator, enabling to rank and design mechanical tests measure individual parameters sensitivity, determine identifiable
through shape optimization considering the level of equivalent plastic parameter subsets and predict the overall identification quality.
strain and strain heterogeneity obtained via a numerical model [16–18]. The paper is structured as follows. First, the IT-based design meth­
For example, Aquino et al. [16] used a heterogeneity indicator along odology and the adopted material model are presented. Second, the
with numerical shape optimization to obtain the so-called butterfly FEMU methodology including the concept of virtual experimentation
specimen. Finally, the third category exploits Topology Optimization and identifiability analysis are introduced. Third, the results of the IT-
(TO) to design heterogeneous specimens [19–20]. For example, Cha­ based optimization and the parameter identification via FEMU are dis­
moin et al. [19] proposed the optimization of the specimen geometry via cussed. Finally, the sensitivity towards the material orientation is
TO to maximize strain-states and richness of strain fields, as well as mapped and the most reliable identifiability metric is scrutinized.
maximize the sensitivity of the measured displacement field with respect
to the sought parameters. 2. NUMERICAL PROCEDURE
In the context of identifying plastic anisotropy of sheet metal, the
orientation of the principal material axis in the specimen plays a central 2.1. Material model
role. However, this was not considered in the currently available design
approaches for heterogeneous experiments. Cooreman et al. [5], The material investigated in this paper is a cold-rolled, low carbon
Lecompte et al. [6] and Wang et al. [14] argued that the material steel sheet named SPCE, with a nominal thickness of 1.2 mm. Due to the
orientation is of fundamental importance when it comes to identifying cold-rolling process, the sheet exhibits significant anisotropic plastic
the plastic anisotropy of steel sheet. Hence, these studies conducted a material behavior, as shown by Coppieters et al. [28]. In the case of
series of perforated tensile tests under a tensile direction of 45o degrees sheet metal, three orthotropic axes can be drawn that are mutually
with respect to the rolling direction. As such, in view of calibrating orthogonal. Practically, ‘RD’ is the rolling direction is the reference to
plastic anisotropy, it was found that it is relevant to include the material which other orientation are defined: ‘TD’ is the transverse direction with
orientation in the design of heterogeneous experiments. 90◦ orientation from the RD, and ‘ND’ is the normal direction (i.e.
With regard to the design driven by maximizing the strain hetero­ through the thickness of the sheet). In addition, a 45◦ (45D) orientation
geneity, it can be stated that this approach inherently assumes a positive can be defined as the intermediate orientation between RD and TD.
correlation between strain heterogeneity and identifiability. It is clear Typically, the variation of the plastic behavior with respect to the ma­
that maximizing strain heterogeneity increases the probability for terial directions is quantified by the Lankford ratios, determined from
arriving at a good identification quality. However, it cannot guarantee tensile tests along RD, TD and 45D.
sufficient identifiability of all constitutive model parameters. When In the present study, the material is assumed to be elastically
dealing with a large set of unknown model parameters, the concept of isotropic and plastically orthotropic. A Young’s modulus E of 219 GPa
identifiability [21–26] could potentially enhance the design methodol­ and a Poisson coefficient v of 0.3 are assumed. The plane stress Hill48
ogy. Barick et al. [21] applied the identifiability index proposed by yield criterion is used to model the plastic anisotropy. The stress-strain
Richard et al. [22] to design nanoindentation experiments to ensure the curve obtained from a tensile test in the RD is used as a reference

2
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Table 2 σeq as a function of the equivalent plastic strain εpl


eq . The identified model
Swift’s hardening law fitted in a strain range from εpl
= 0.002 up to the
eq parameters, i.e., deformation resistance K, the initial strain ε0 and the
maximum uniform strain. The reported r-values are measured at εeng= 0.10 hardening exponent n can be found in Table 2. The anisotropic yield
[28]. criterion is expressed with respect to the principal material axis denoted
Hardening parameters K(MPa) ε0 n 1 (RD) and 2 (TD) with σij the components of the Cauchy stress tensor.
564 0.0059 0.275 Furthermore, the associated flow rule is adopted assuming equivalency
r-values r0 r45 r90 between the plastic potential and yield function.
1.85 1.93 2.82 F, G, H and N are the anisotropic parameters of the Hill48 yield
Anisotropic parameters F H N criterion. Given that the reference datum for work hardening is cali­
0.230 0.649 1.412
brated using tensile test data along RD, the sum of H and G is equal to 1.
The r-values are used to calibrate the anisotropic parameters and their
datum for work hardening. The hardening behavior is approximated by values can be found in Table 2. The parameter set, shown in Table 2, is
Swift’s hardening law fitted to the available pre-necking data in the RD. considered the ground truth, i.e., the reference material behavior that
The associated constitutive equations are summarized in Table 1. can be used to assess the identification quality (accuracy) via FEMU in
Swift’s hardening law describes the evolution of the equivalent stress Section 4.1.3.

Fig. 1. FE model information (a) specimen geometry, (b) design variables for shape optimization and (c) material orientations, mesh information and bound­
ary conditions.

Fig. 2. Python script structure for automatic numerical model generation, data post-processing and optimization process communicating with Abaqus.

3
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 3. The detailed flowchart of FEMU procedure.

2.2. Specimen design and FE model in a python script that communicates with Abaqus CAE. Here, autono­
mous model generation, meshing, submitting the simulation, post-
The optimization problem, tackled in this paper, is relatively simple processing and shape optimization, are all executed using one python
and based on the concept of a notched tensile specimen to generate a script. The optimization process consists of three main building blocks.
heterogeneous strain field. The resulting heterogeneity obviously de­ The first one determines the critical displacement boundary with the aid
pends on the shape of the notch. Fig. 1(a) shows the basic dimensions of of a necking criterion. Once the optimization is launched, two models
specimen that are kept constant, where the red shaded area represents are created, namely model-0 and model-1. Model-0 and Model-1 are
the area of interest (AOI). Fig. 1(b) illustrates an enlargement of the identical (geometry, material definition, etc.) except for the applied
notch that needs to be designed. In order to optimize the shape of this displacement. This process is repeated for every new iteration.
notch, it needs to be parameterized. In this work, four coordinates are Model-0 consistently imposes an excessive vertical displacement
used, as shown in Fig. 1(b). The coordinates of A, B, C and D are (U2) to determine the maximum allowable elongation of the specimen,
controlled by the corresponding length, and connected via cubic splines. according to the necking criterion [29]. This leads to large displace­
An example of the FE model is shown in Fig.1(c) including local seed ments and deformations requiring to account for nonlinear geometric
size, element formulation (S4R), boundary conditions and material effects in the FE model. Obviously, this critical displacement Ucrit de­
orientation. FEA is performed using ABAQUS v2017 with default im­ pends on the specimen geometry and must therefore be determined at
plicit solver settings. every new iteration. Additionally, the magnitude of consecutive
For the initial material orientation, the tensile direction was assumed displacement increment intervals is around 10− 2 mm. Thus, the critical
to be aligned with the TD as shown in Fig. 1(c). The length vector L = displacement is determined in model-0 and then implemented in
[LA, LB, LC, LD] is the design variable subjected to optimization. The goal model-1, where the simulation must terminate at the onset of necking.
of the design process is to maximize the strain heterogeneity in the AOI After model-1 is completed, the second building block starts by calcu­
region. lating the IT value in the AOI. To update the notch radii vector L = [LA,
LB, LC, LD], a cost function is computed that reads as:
2.3. Indicator definition and shape optimization process 1( Target )2
C(L) = I − ITNum (L) , (5)
2 T
The shape optimization process aims at maximizing the strain het­
erogeneity indicator IT, as defined in [16]: with L the notch radii vector of the tensile specimen. The cost function
p p expresses the discrepancy between the target IT value and the numeri­
[Std(ε2 /ε1 )] (ε2 /ε1 )R [Std(ε )] ε Av p
IT = wr1 + wr2 + wr3 Max
+ wr4 + wr5 ε (4) cally computed IT value, where ITNum (L) is associated with the current
wa1 wa2 wa3 wa4 wa5
design. Although the theoretical maximum value of IT is 1 [18], the
It can be inferred from Eq. (4) that IT is a scalar computed using five target IT value is arbitrarily chosen to be IT
Target
= 0.7. From experience,
strain measures, namely the range of the strain state (ε2 /ε1 )R between this slightly lower target value helps to keep the parameters update
the minor (ε2) and major (ε1) principal strains, standard deviation of stable. The Gauss-Newton parallel with Levenberg-Marquardt algorithm
strain state Std(ε2 /ε1 ) and the plastic strain Std(εp ), the mean value of [30–31] is used to minimize the cost function C(L). In the third building
maximum equivalent plastic strain εpMax at the most relevant strain states block, an updated specimen geometry is obtained that is re-entered in
(tension, pure shear, uniaxial tension, plane strain tension and uniaxial the first building block for the next iteration. Once the parameter change
compression values), and the average deformation Avεp . Moreover, wr is sufficiently small (convergence criterion value of 0.05%), the opti­
and wa are the relative and absolute weighting factors, respectively. In mized shape is considered to be found.
this paper, equal importance (50-50%) weighting factors for strain
range/heterogeneity and strain level were used. More information on
the calculation of the individual terms of IT and the values of the
weighting factors can found in Souto et al. [17].
Fig. 2 schematically shows the optimization procedure implemented

4
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

3. METHODOLOGY FOR MATERIAL PARAMETERS where θ is the vector of unknown material parameters (F, H, N) of the
IDENTIFICATION AND IDENTIFIABILITY ANALYSIS anisotropic yield function, and n the number of data points in the DIC
measurement at certain load step (only one load step is considered in
3.1. Virtual experimentation and FEMU method this paper). The RMSm (root mean square) value can be obtained by Eq.
(7), which was used to normalize each strain component by eliminating
The task of the FEMU procedure is discussed here by means of a the magnitude difference in Eq. (6). The subscripts “exp” and “num”
detailed flowchart in Fig. 3. FEMid is a home-made FEMU module used indicate experimental data and numerical response, respectively. Within
for the minimization of the discrepancy between the numerically the iterative cycle of FEMU, the update of the anisotropic parameters is
computed and experimentally acquired strain fields by tuning the performed by searching for a minimum of the objective function using
anisotropic material parameters. In order to have a consistent strain the Levenberg-Marquardt algorithm. FEMU is launched with an initial
computation, FEMid is initially fed with displacement fields. The FE set of arbitrarily chosen anisotropic parameters. Then the iterative
model delivers the nodal displacement field at the surface of the spec­ process starts and the optimization method generates an updated set of
imen, while the DIC code delivers the displacement field after subset- anisotropic parameters. The process is repeated until the value of the
based correlation. The strain window method is then applied to both objective function reaches a value below the threshold or the parameter
displacement fields to arrive at the strains in each DIC data point. The change is sufficiently small (0.5 %). The derivatives of the objective
latter approach mitigates to some extent the dependency of the DIC user function are calculated numerically via a forward finite difference
settings. The leveraged approach recently applied by Gothivarekar et al. scheme with a perturbation value of δ=0.005.
[32] could further improve the consistency here.
Instead of using actual experimental data, virtual or synthetic data is 3.2. Sensitivity analysis
used in this paper. The lower panel of Fig. 3 illustrates this process. The
Finite Element (FE) model with ground truth anisotropic parameters In this paper, the identifiability analysis [26] is applied to the Hill48
(see Table 2) is used to create the virtual experiment. The FEDEF-module yield criterion and the Yld2000-2D yield function. For the sake of clarity,
[33] in MatchID [34] uses the initial mesh and displacement field(s) we first introduce the methodology for the Hill48 criterion. In general, a
from the FE model to numerically deform a speckle pattern associated sensitivity analysis is defined as a method to assess the influence of an
with a specific load step. The synthetic (reference and deformed) images optimization parameter variation (geometry, material, loading, etc.)
are then post-processed in the DIC code, yielding the experimental with respect to a numerical response (strain, stress, force, etc.) [25]. The
displacement field. The approach enables to mimic the complete mea­ targeted FEMU approach (Section 3.1) minimizes the discrepancy be­
surement chain including camera calibration, image noise, speckle tween the experimentally measured and numerically predicted strain
pattern quality and DIC settings. In this study, however, we aim at field in the AOI of the mechanical test. Output sensitivity functions are
finding the intrinsic relations between the IT-based design and identifi­ obtained by linearization of the parameters vector θ0:
cation quality through FEMU. Therefore, we only consider the essential ⃒
step to arrive at the displacement field, namely the DIC settings (cor­ ∂εnum (θ)⃒⃒
εnum (θ0 + δθ0 ) ≈ εnum (θ0 ) + δθ0 . (8)
relation algorithm, subset and step size). Future work will embark on the ∂θ ⃒θ=θ0
T

assessment of the full measurement chain. In the FEMU procedure,


For the identifiability analysis, the required sensitivity matrix V is
FEMid uses the following cost function [11]:
constructed by computing the partial derivative of the strain field with
[( exp
∑ n
εXX,i − εnum
XX,i (θ)
)2 ( exp
εYY,i − εnum
YY,i (θ)
)2 respect to the vector of unknown anisotropic material parameters θ:
minimise C(θ) = + ⃒
θ∈R n
RMSXX RMSYY ∂εnum (θ)⃒⃒
(9)
i=1
)2 ] V= ,
( exp num
εXY,i − εXY,i (θ) ∂θT ⃒θ=θ0
+ , (6)
RMSXY
with εnum(θ) the generated strain field expressed as:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ [( ) ( num ) ( num )
1∑n ( exp )2 εnum (θ) = εnumXX (θ) 1 , εYY (θ) 1 , εXY (θ) 1 , ...,
RMSm = ε , m = XX, YY, XY, (7) ( num ) ( num ) ( num ) ]T
n i=1 m,i εXX (θ) n , εYY (θ) n , εXY (θ) n . (10)

Table 3
Methodology of identifiability analysis.
Identifiability analysis Definitions Description
Sensitivity Analysis
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Absolute sensitivity ∂εnum (θ) 1∑n
V = {vij } = | (12) i = 1, 2, ..., n j = 1, 2, 3. SCm = (εnum )2 , m = XX, YY, XY1.
∂θT θ=θ0 n i=1 m,i
Non-dimensional sensitivity Δθj ‖Sj‖ is the Euclidian norm of the sensitivity.
S = {sij}, sij = vij (13)
SCm
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Sensitivity measure, δmsqr
j msqr 1 ∑3n 2
δj = s (14)
3n i=1 ij
Normalized sensitivity ̃ = {̃sij }, ̃sij = sij (15)
S
‖ sj ‖
Identifiability Metrics

Collinearity index, γK 1 λK = eigen(H


̃ FEMU ) or eigen(̃T
γK = √̅̅̅̅̅̅̅̅̅̅̅̅̅ (16) ̃ ̃K ).
SK WS
min ̃
(λK )
I –index, IK maxλK λK = eigen(HFEMU ) o r eigen(STK WSK ).
IK = log10 (17)
minλK
Determinant measure,ρK 1/(2k) ∏k 1/(2k)
2
ρK = det(STK WSK ) =( j=1 λj ) (18)

Note: STK SK is the same as STK ⋅SK . Bold notation represent vector or matrix for variable.
1. εnum num num
XX,i , εYY,i and εXY,i represents three strain components at point i.
2. K is a parameter (sub)set of size k from full parameters.

5
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

The vector θ = [θ1,θ2,..., θm]T (m=3) is the parameter vector of the F(θ). In the context of nonlinear least squares, F(θ) approximately
Hill48 yield criterion, namely θ = [F, H, N]T. V is thus a (3n × 3) matrix equals to HFEMU [35]. Therefore, the HFEMU can be written as F(θ) =
evaluated at the reference anisotropic parameters (see table 2) vector θ0 STK WSK , where W=diag(w1, w2, …, wn) is a diagonal matrix related to
= [0.230, 0.649, 1.412]T: measurement noise. Since no measurement noise (e.g. image noise) was
( num ) ( num ) ( num ) introduced in the creation of the synthetic DIC data, W is the identity
∂εXX (θ) ∂εXX (θ) ∂εXX (θ)
⎡ ⎤ matrix I. Thus, the F(θ) can be regarded as a summary of the output
∂F ∂H ∂N
1 1 1
sensitivity functions and the measurement accuracy (if measurement
⎢ ⎥
⎢ ( )( ) ⎥
)( noise is considered), thereby summarizing the information concerning
⎢ ∂εnum ∂εnum ⎥ ∂εnum


YY (θ) YY (θ) ⎥

YY (θ) the model parameters gained from an experiment. If F(θ) is nonsingular,


∂F 1 1 ∂H 1 ⎥

∂N the approximate covariance matrix COV(θ) [26,36]is given as:
⎢ ⎥
⎢ ( num ) ( num ) ( num ) ⎥
⎢ C(θ) − 1
⎢ ∂εXY (θ) ∂εXY (θ) ∂εXY (θ) ⎥ ⎥ COV(θ) = F (θ), (19)
⎢ ⎥ 3n − k
⎢ ∂F ∂ H ∂N ⎥
⎢ 1 1 1 ⎥
⎢ ⎥
⎢ ⎥ where n is total number of measurements points and k is the parameters
⎢ ⎥
V=⎢ ⋮ ⋮ ⋮ ⎥, (11) number. F− 1(θ) could be rewritten as:
⎢ ⎥
⎢ ⎥
⎢ ( num ) ( num ) ( num ) ⎥ 1 1

⎢ ∂εXX (θ) ∂εXX (θ) ∂εXX (θ) ⎥ ⎥ F− 1 (θ) = adj(HFEMU ) = (∏ ) adj(HFEMU ), (20)
⎢ ⎥ det(HFEMU ) K
⎢ ∂F ∂H ∂N j=1 λj
n n n ⎥
⎢ ⎥
⎢ ⎥
⎢ ( num ) ( num ) ( num ) ⎥
⎢ ∂εYY (θ) ∂εYY (θ) ∂εYY (θ) ⎥ where adj() represents adjugate matrix of HFEMU, det() being the
⎢ ⎥



⎥ determinant function and λj being the eigenvalues of HFEMU, indicating
∂F ∂H ∂N


n n n ⎥
⎦ that COV(θ)is connected with λj. The variance or uncertainty of θj on the
( num ) ( num ) ( num ) evaluated parameter value could be quantified by var(θj ) =
∂εXY (θ) ∂εXY (θ) ∂εXY (θ) diag(cov(θ))jj (diag(cov(θ)) represents the diagonal elements). The
∂F ∂H ∂N
n n n above equations indicate that the F(θ)matrix relates to the uncertainty of
( ) the identified parameters, providing insight into the evaluation of an
∂εnum (θ)
where ∂θj is the partial derivative of the three strain components optimal specimen shape and ideal material orientation resulting in a
n maximum parameter identifiability.
at element n for a certain load step with respect to material parameter j
From a geometrical perspective, given that F(θ) is approximately
in the AOI region. Finite difference approximation was used to calculate
equal to the Hessian matrix HFEMU of the C(θ), a steeper ‘valley’-cur­
the sensitivities adopting the same parameter perturbation value as used
vature (Hessian matrix) of the C(θ) in the vicinity of local minimum is
in the FEMU process (δ=0.005).
desirable for the parameters identification. Specifically, a steeper ‘val­
ley’-curvature of the C(θ) has great potential to build a well-posed
3.3. Practical identifiability gradient-based inverse problem for obtaining nearly unique local or
global optimal identified parameters during the FEMU process, resulting
In general, a good identifiability has to fulfil two conditions. First, in a better identification quality. This can then be quantified by the
the model output εnum(θ) has to be sufficiently sensitive to individual eigenvalue of F(θ) or the Hessian matrix [37]. Thus, several scalar
changes of each parameter. Second, changes in the model output due to metrics based on the eigenvalue of H ̃ FEMU or HFEMU are defined to
variations in single parameter cannot be compensated by appropriate quantify the parameters identifiability:
changes in other parameters in K (K is a parameter (sub)set of size k from
full parameters). The adopted identifiability metrics and corresponding
• Metric a: Collinearity index [26] - γK, smallest eigenvalue of H
̃ FEMU ,
formulas are summarized in Table 3. The first condition is addressed by
Eq. (16) in Table 3;
the sensitivity measure δmsqr
j (Eq. (14) in Table 3) which is calculated for • Metric b: Identifiability index [21–25] - IK(I-index), logarithm be­
each parameter θj separately. The sensitivity measure δmsqr
j is used to tween largest and smallest eigenvalue of HFEMU, Eq. (17) in Table 3;
quantitatively assess the individual parameter importance (parameter • Metric c: Determinant measurement [38–39] - ρK, determinant of
sensitivity strength). The second condition is quantified by the collin­ HFEMU, Eq. (18) in Table 3.
earity index γK (Eq. (16) in Table 3), serving as a tool for the analysis of
parameter interdependencies. Moreover, IK and ρK (Eq. (17) and Eq. (18) First, a γK - design aims at maximizing the smallest eigenvalue of
in Table 3) combine the information provided by the above two condi­ ̃ FEMU , which is the common metric for measuring the collinearity
H
tions [24, 38]. The calculation of IK and ρK based on the index. The calculation of γK based on normalized matrix S
̃K avoids biases
non-dimensional pseudo-Hessian matrix -HFEMU of cost function C(θ) caused by differences in the absolute values of the individual sensitivity
containing the information-richness (e.g. strain fields) for a specific vectors [40]. According to Burn et al. [26], a typically collinearity index
specimen design. of γK is 20, indicating that a variation of parameters θj on the model
The calculation of the sensitivity is a prerequisite for the computa­ output (εnum(θ)) can be compensated up to a value of 1/γK (5%) by
tion of sensitivity measure (δmsqr
j ) and the identifiability metrics (γK , appropriate changes in another parameter in K.
IK and ρK ).For improving the conditioning of the absolute sensitivity Second, the so-called identifiability index- IK design, which is
matrix V (Eq. (11)) or Eq. (12) during the FEMU process, Δθj (j=F, H or calculated based on HFEMU, attempts to minimize the ratio of largest to
N) and SCm are applied to obtain non-dimensional sensitivity S (Eq. (13) the smallest eigenvalue. Gujarati [41] suggested a threshold value to
in Table 3). Moreover, the selection of Δθj and SCm for identifiability discriminate between poor (IK>3), moderate (2 < IK < 3) and good
metrics are identical as those adopted in the FEMU process thereby identifiability (IK < 2).
eliminating the user-dependency for Δθj and SCm. Hence, this approach Finally, another metric is the so-called determinant measure ρK.
guarantees the consistency of HFEMU for the identifiabiltiy metrics According to Brun et al. [38] and Martinez-Lopez et al. [42], ρK com­
calculation and the inverse identification via FEMU technique. bines the sensitivity provided by δmsqr
j (Eq. (14)) and γK in a useful way.
The sensitivity of εnum(θ) to changes in parameters θ could be Hence, the calculation of ρK based on the non-dimensional sensitivity
quantified by Fisher information matrix

6
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Table 4
IT value for different initial geometry parameters set and local optimal geometry parameters.
Mechanical Tests Length vector L - [LA, LB, LC, LD] Prior to necking Equal εpl
eq

LA(mm) LB(mm) LC(mm) LD(mm) IT-n IT-eq

Test A Initial L 1 0.00 0.00 0.00 0.00 0.25 0.01 2


Test B Initial L 2.80 2.80 2.80 2.80 0.23 -
Local Optimal L 2.12 3.54 3.49 2.10 0.26 0.14
Test C Initial L 3.00 3.00 6.00 6.00 0.26 -
Local Optimal L 2.87 4.81 7.23 7.23 0.34 0.15

1. Initial guess of L.
2. IT value at εpl
eq =0.03.

Fig. 4. Geometry of mechanical tests (a) Test A, (b) Test B and (c) Test C.

Fig. 5. (a) Stress states for Test B and Test C at TD material orientation using Hill48 criterion. (b) Strain states for Test B. (c) Strain states for Test C.

7
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 6. Virtual experimentation: (a) FE mesh projection, (b) synthetic image deformation and (c) synthetic strain field after DIC for Test C.

matrix SK instead of ̃
SK . Moreover, recalling to Eq. (19) and Eq. (20), occurs and is denoted as IT-n. It can be inferred that all IT-n values are
parameters covariance matrix COV(θ) could be reformulated as a relatively large. This is mainly driven by the large plastic equivalent
function of ρK (COV(θ) = C(θ) 1
adj(HFEMU )), which is possibly the strain probed prior to local necking. The reason that Test A, in essence a
3n− k ρK
uniaxial tensile test, corresponds with an IT value of 0.25 is due to the
reason why ρK is recommended as useful metric [42]. However, ρK is
substantially dependent on the choice of Δθj, and therefore it is generally onset of diffuse necking. When limiting Test A at a strain level of εpl
eq =

not possible to give a threshold or critical ρK -value [38]. Consequently, 0.03, the corresponding IT value is around 0.01. Given that less contri­
ρK is a relative measure for the quantified comparison of parameter bution for strain heterogeneity provided by plastic equivalent strain and
identifiability for different parameter subsets. In conclusion, the smaller have an honest comparison Test B and Test C, the IT value was recalcu­
γK, IKand the larger ρK, the better the parameter identifiability. Hence, lated for a load step that generated an equal equivalent plastic strain
the identifiability metrics can also be used to assess the effectiveness of a (εpl
eq ≈ 0.5) in for both tests, denoted as IT-eq. From this analysis, it can be
heterogeneous specimen to inversely identify a set of unknown material inferred that Test C is slightly more heterogeneous than Test B.
parameters.
Although the above three metrics are investigated in this paper, it 4.1.2. Heterogeneity analysis for different mechanical tests
needs to be noted that alternative metrics or criteria can be used to In order to better understand the heterogeneity of the above­
assess identifiability, e.g. (1) the correlation among parameters mentioned mechanical tests, the generated stress and strain states are
expressed by off-diagonal terms of the correlation matrix [26, 37], (2) evaluated in the current section. Here, only material points in the plastic
the confidence interval for individual parameters [39] and (3) the size of regime are considered. Fig. 5(a) shows the stress state distribution of
parameter cloud [43]. A well-posed problem requires that the correla­ Test B and Test C along with the reference Hill48 criterion. Fig. 5(b) and
tion between parameters remains below a certain threshold, the confi­ (c) illustrate the strain states for Test B and Test C, respectively. To
dence interval of parameters and the size of the parameter cloud should enable a comparison between both tests, the stress components are
remain small. For detailed information with regard to these alternative normalized by the equivalent stress. Furthermore, the local stresses in
identifiability metrics, the reader is referred to [26, 37, 39, 43]. Fig. 5(a), σ11 and σ22, are aligned with the RD and TD, respectively.
Here, each point (green color and red color represent Test B and Test C,
4. Results respectively) in the graph represents a stress state in the AOI.
It can be inferred that a large amount of the plastic deformation
4.1. Determining the optimal mechanical test shape and heterogeneity comes with a shear stress component σ12 (all points that are not on the
analysis projected contour of the yield surface). Compared to the stress state
distribution (Fig. 5(a)) of Test B, Test C covers a slightly larger area in
4.1.1. Shape optimization stress space, indicating that Test C is a good candidate for inverse
Depending on the initial length vector L = [LA, LB, LC, LD], the shape identification through FEMU.
optimization, described in Section 2.3, yields different heterogeneous Fig. 5(b) and (c) display the corresponding strain states in the AOI
specimens corresponding to several local maxima of IT. The optimized region of the two mechanical tests along TD. The strain state(ε2 /ε1 ) is
vector lengths, derived from two different initial guesses (Test B and Test defined as the ratio between of the minor ε2 and the major ε1 principal
C), and their corresponding IT value are shown in Table 4. strains. In addition, straight lines are added to the graphs that represent
In Fig. 4, Test A represents a tensile test specimen that has the same a certain strain state. It can be observed that the amount of scatter is a
width and height of the AOI as the specimens used in Test B and Test C. It measure for the diversity of the strain states. For both tests, the graphs
must be noted that the outcome of an IT-calculation depends on the load suggest that most of the strain states are located between pure shear and
step at which the strain field is extracted. Indeed, the larger the plastic uniaxial tension. Here again, Test C exhibits a slightly higher diversity in
deformation, the higher resulting IT value. This obviously depends on the data, associated with a higher strain heterogeneity. This test can thus
the chosen weighting factors and adopted necking criterion. In this re­ be advocated as the best candidate for inverse identification. The latter
gard, two values of IT are reported in Table 4. The column ‘Prior to can be confirmed with a quantitative calculation of the standard devi­
necking’ reports the IT value found before the onset of local necking ation of the strain state and stain state range, respectively written as:

8
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Table 5 Subsequently, this comparison serves as a measure for the identifi­


Initial guesses set of anisotropic coefficients F, H and N. cation quality and can be expressed by the average relative error - A.R.E:
Parameters F H N [ ⃒ ⃒]
1∑q 1∑k ⃒⃒θj − θj ⃒⃒
ref id

Initial guess 1 0.5 0.5 1.5 A.R.E = 100 ∗ ⃒ ⃒ , k = 3, q = 3, (23)


q p=1 k j=1 ⃒ θref ⃒
Initial guess 2 0.3 0.4 1 j

Initial guess 3 0.4 0.7 1.3


with θr e f (θ = [F, H, N]T, number of parameters k = 3, number of initial
guesses q=3) the vector of reference anisotropic parameters and θid the
vector of inversely identified anisotropic parameters.
Table 6
Comparison of reference parameters values and identified parameters values for Similarity, individual parameter identification quality could be
Test A, Test B and Test C, respectively. expressed via the individual relative error -I.R.E:
⃒ ⃒
Parameters F H N A.R.E ⃒θref − θid ⃒
⃒ j j ⃒
I.R.E = 100 ∗ ⃒ ⃒. (24)
Reference values 0.2302 0.6492 1.4120 ⃒ θref ⃒
j
Identified values Test A 0.2274 0.6176 1.3419 (5.0%) 1 3.68%
Test B 0.2210 0.6342 1.3679 3.13% The reported A.R.E values in Table 6 represent the average A.R.E
Test C 0.2192 0.6340 1.3936 2.49%
obtained with the different initial guesses stated in Table 5. It can be
1. The data in the parenthesis represents I.R.E (individual identified parameters inferred from Table 6 that Test A (A.R.E = 3.68%), yields the worst
relative error). identification quality, yet unexpectedly good for a uniaxial tensile test.
For a tensile test along the principal material direction, the shear strain
( ) ( ) ( )
ε2 ε2 ε2 should be negligibly small, hence leading to a poor identification quality
= − , (21) for the anisotropic parameter N. Table 6 shows that for Test A, the
ε1 R ε1 max ε1 min
parameter N is indeed poorly identified, compared to the other tests (I.R.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑ [ ]2 E of N = 5.0%). In general, the successful identification should satisfy at
( ) √ n (ε /ε ) − μ
ε2 √ i 2 1 i ε2 /ε1 least two conditions. First, the model output εnum(θ) has to be suffi­
Std = . (22)
ε1 n− 1 ciently sensitive to individual changes for each parameter and the
( ) ( ) collinearity of the sensitivity matrix ̃
S should be sufficiently low. The
Compared to Test B ( εε21 =1.14, Std εε21 =0.12), it is clear that the second condition is confirmed through analyzing the identifiability
R
( ) ( ) metric γK for Test A (γK=4.96, more details can be found in [44]),
standard deviations of Test C ( εε21 =1.43, Std εε21 =0.13) are higher, indicating good correlation between γK and the identification quality.
R From FEA, it can be stated that, the shear strain present in the AOI
implying a stronger diversity and strain heterogeneity.
originates for the sharp shoulders of the tensile specimen. Nevertheless,
the magnitude of the shear strains can be expected to be small. As such, a
4.1.3. Validation via FEMU
crucial condition for a successful identification is related to the accuracy
The main purpose in this section is to validate whether the IT -based
(strain resolution) of the DIC experiment. Indeed, in the case of Test A,
designs (Test B and Test C) exhibit a positive correlation with the iden­
the DIC measurement should be accurate enough to capture the small
tification quality through FEMU. As shown in Table 4 (column “Equal
shear strains. To this end, experimentally measured and numerically
εpl
eq ”) in Section 4.1.1, the IT-eq value of Test A (IT-eq = 0.01) is the computed strain fields are compared in Fig. 7. The strain field obtained
smallest, while that of Test C is the largest (IT-eq = 0.15), yet very close to via the virtual experiment is shown in the first column in Fig. 7. The
Test B (IT-eq= 0.14). The validation of the IT-based design is pursued by numerical “ground truth” strain field is shown in the second column. The
virtual experimentation and the targeted identification method, i.e. third column shows the discrepancy between the experimental and
FEMU, as introduced in Section 3.1. The virtual experiment is generated numerical strain field, represented by Δ and defined as:
with the reference material behavior, shown in Table 2. ⃒( )exp ( )num ⃒
⃒ ⃒
Fig. 6 schematically shows the process of generating virtual DIC data Δ = ⃒ εlnEA
ij − εlnEA
ij ⃒, (25)
for Test C. First, the FE mesh is aligned with an arbitrary experimental
speckle pattern (Fig. 6(a)). The node displacements, extracted at cor­ with(εlnEA
exp
) the experimental strain field and (εlnEA )
num
the numerical
ij ij
responding strain field for calculating the IT-eq value, are then used to
strain field. In this case, Δ represents the intrinsic error of the DIC al­
numerically deform the reference image (Fig. 6(b)). Finally, the gener­
gorithm, as experimental noise was not imposed.
ated images are correlated using MatchID-2D to arrive at the synthetic
It can be seen from Fig. 7 that the experimental and numerical strain
strain field shown in Fig. 6(c). In this way, all metrological aspects of the
fields show a qualitative and quantitative accordance. It can also be
DIC algorithm are included in the optimization [33]. Given the large
inferred that, in general, the AOI is not perfectly subjected to uniaxial
strain gradients and near-edge strain localizations in the IT-based design,
tension. Although the numerical shear strain εlnEA is very small
the approach enables to scrutinize the importance of DIC settings in the − 4
XY

identification process. However, in this study, default DIC settings were (maximum numerical strain εlnEA XY =1.8 × 10 ), some heterogeneity can
adopted, yielding a Virtual Strain Gauge (VSG) of 45 px with an image be observed.
resolution of 0.039 mm/px. It is important to note this feasibility study is More importantly, it can be seen that 2D-DIC enables to reconstruct
confined to 2D DIC which is justified by the planar test specimen and this shear strain heterogeneity (first column, last row). The fact that Test
in-plane loading conditions. A exhibits a sufficient identifiability with respect to the sought aniso­
From literature [8, 18] it is well-known that gradient-based optimi­ tropic parameters and that the strain field is accurately reconstructed are
zation depends on the initial guess. To mitigate this dependency in the the main reasons why a relatively low A.R.E and unexpectedly high
FEMU process, three initial guesses were arbitrarily selected for the identification quality were found. This shows the mutual importance of
material parameters F, H and N, as shown in Table 5. The plasticity identifiability and metrological aspects of the DIC algorithm. Never­
model, used to create the synthetic data, was calibrated in previous theless, it must be noted that the virtual experiment did not account for
research [28] and the reference parameters can be found in Table 2. noise inevitably present in an actual experiment.
These values allow for a direct comparison with the identified aniso­ The strain resolution for the shear strain in Test A has been measured
tropic parameters for Test A, B and C shown in Table 6. by printing the virtual experiment on a blank paper and taking still

9
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 7. Comparison of strain fields for experimentally measured (first column), the numerically computed strain fields (second column) and the discrepancy between
the plots (third column) for Test A.

Fig. 8. Material orientation optimization and validation procedures.

10
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Table 7
The potential identified accuracy rank comparison of IT and three identifiability metrics for different material orientation tests.
Material Orientation IT IT − Rank γK1 γK − Rank IK2 IK − Rank ρK ρK − Rank
RD 0.145 7 1.53 2 2.36 6 32.21 7
R15 0.150 3 1.89 4 2.37 7 38.15 5
R30 0.156 1 1.77 3 1.67 2 54.02 3
R45 0.152 2 1.27 1 1.06 1 66.18 1
R60 0.148 4 2.45 5 1.83 4 59.89 2
R75 0.146 6 4.45 7 2.04 5 40.50 4
TD 0.147 5 3.84 6 1.71 3 33.67 6

1. A typical threshold is 20.


2. A typical threshold is 3.

images with a typical 2D-DIC set up (5MPx camera and a 12.5 mm lens).
Table 8
Here, a strain resolution of εlnEA
XY = 2.66 × 10
− 4
was found. Conse­
Initial guesses set of anisotropic coefficients H, F and N.
quently, an actual experiment the parameter N would be non-
Parameters F H N
identifiable due to the noise floor. Ultimately, a positive correlation
between the IT-value and the identification quality can be inferred from Initial guess 1 0.5 0.5 1.5
the A.R.E of Test B (A.R.E=3.13% and IT-eq=0.14) and Test C (A.R.E= Initial guesses 2 0.3 0.4 1
Initial guesses 3 0.4 0.7 1.3
2.49% and IT-eq=0.15). Initial guesses 4 0.1 0.55 1.6

4.2. Determining the optimal material orientation via identifiability


analysis Table 9
Comparison of reference parameters, identified parameters values and corre­
4.2.1. Optimal material orientation sponding accuracy rank for different material orientation tests.
From previous analyses, it is clear that a positive correlation exists
Parameters F H N A.R.E A.R.E
between the IT value and the identification quality (A.R.E). However, it Rank 1
worth noting that a lower IT value does not necessarily lead to an inferior
Reference 0.2302 0.6492 1.4120 - -
identification quality. Indeed, recall that Test A in Section 4.1.3 has very values
low equivalent IT-eq value, while an acceptable identification quality was Different RD 0.2421 0.6526 1.4015 2.14% 7
found, provided that the strain field can be accurately reconstructed. Material (5.51%) 2

Moreover, the positive correlation cannot guarantee maximum identi­ Orientations


R15 0.2314 0.6508 1.4098 0.61% 3
fiability of the sought parameters. In this section, it is scrutinized
R30 0.2291 0.6482 1.4198 0.84% 5
whether an IT-based design (Test C in the remainder of this work) can be R45 0.2306 0.6527 1.4184 0.39% 1
further improved by considering the orientation of the principal mate­ R60 0.2244 0.6472 1.4003 1.22% 6
rial axis. The goal is to enhance the identification quality of the sought R75 0.2293 0.6484 1.4020 0.40% 2
anisotropic parameters by changing the material orientation with TD 0.2279 0.6471 1.4033 0.64% 4

respect to the tensile direction. In addition, three identifiability metrics 1. Accuracy rank: best-1, worst-7.
(Section 3.3) are adopted to assess the influence of the orientation on the 2. Individual identified relative error (I.R.E).
identifiability.
To this end, the orientation is varied with steps of 15o as shown in the quality through FEMU.
left panel of Fig. 8 yielding seven material tests. Both sensitivity and
identifiability analysis are based on solution variables that are expressed 4.2.2. Parameters identification for different material orientations
in a local co-rotational coordinate system. In view of an honest assess­ The aim of this section is to validate the findings concerning the
ment, the identifiability of each test is evaluated at the same amount of optimal material orientation reported in the previous section. In a more
plastic work. As shown in the right panel of Fig. 8, the strain hetero­ general sense, however, it is also important to determine the most reli­
geneity indicator IT and the three identifiabiltiy metrics (γK, IK and ρK) able identifiability metric for assessing the identification quality. The
are derived for each material orientation test. impact of different material orientations is then assessed through the
The results are shown in Table 7. Small values of γK and IK, and large FEMU procedure. In order to focus on the intrinsic performance of the
values for IT and ρK correspond to a potential better identification identifiability metrics, it is opted here to exclude any experimental bias
quality (low A.R.E). Table 7 also shows the rank of the heterogeneity- in the analysis. To this end, the virtual experiment consists of numerical
indicator and three identifiability metrics, whereby rank-1 corre­ strain fields generated by the FE model using the reference material
sponds to the best material orientation for Test C. It can be inferred from behavior. To mitigate the influence of the initial guess for the sought
Table 7 that the variation of the IT value is small, indicating that the anisotropic material parameters, one additional initial guess set was
influence of the material orientation on the strain heterogeneity is arbitrarily chosen and added in Table 8.
moderate. The IT rank shows that the material orientation R30 is most The average identified parameters resulting from the four initial
favorable, yet very close to R45. The metric γK and IK both suggest (γK guesses are reported in Table 9. The relative error (A.R.E) and the
less than 20 and IKless than 3) that the full parameter set (F, H and N) can associated accuracy rank are also reported in Table 9. First, it is clear
be accurately identified for all considered material orientations. Looking that all material orientations yield a small A.R.E, which is not surprising
at the rank of γK and IK, it can be concluded that R45 is the most as FEMid is fed with numerical strain fields. Moreover, the identifiability
adequate material orientation. Moreover, the metric ρK also indicates metric γK and IKas presented in Table 7 indicate that the full parameters
that R45 is the most suitable material orientation and RD is the worst set can be accurately identified for all considered material orientations.
one. From this analysis, it can be concluded that all identifiability The latter is confirmed by the small A.R.E for all material orientation
metrics point towards a material orientation of R45, which is 45o with tests shown in Table 9. It can be seen from Table 9 that R45 is optimal
respect to the rolling direction (RD). The effectiveness of this orientation with an A.R.E of 0.39% (rank-1), while RD shows the worst accuracy
will be validated in the next section by assessing the identification

11
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 9. Importance rank of Hill48 yield criterion parameters for Test RD.

Fig. 10. Strain states (a) for Test C RD, Test C TD and Test C R45 in a single graph, (b) for Test C RD, (c) for Test C TD and (d) for Test C R45.

with an A.R.E of 2.14% (rank-7). It can be stated that all intermediate adequate material orientation associated with Test C R45, the strain and
material orientations, except for R60, yield a similar A.R.E., and, as a stress states are analyzed for Test C RD, Test CTD and Test CR45. Fig. 10
consequence, a general relation between identification quality and (a) presents the principal strain states for the three material orientations.
material orientation cannot be derived. The individual plots (Fig. 10(b), (c), (d)) show that the strain state dis­
Nevertheless, it is clear that testing in the RD is not beneficial with tribution in the principal strain space is virtually identical for the three
regard to the identifiability. A further analysis reveals that this finds its material orientations. This indeed explains why IT is virtually identical
roots in the identification quality of the parameter F (I.R.E error around for Test C RD, Test C R45 and Test C TD (see Table 7).
5.15%). Intuitively, when conducting the experiment in the RD direc­ Fig. 11 shows the stress distribution where the stress components σ11
tion, the sensitivity with regard to F is less because this parameter only and σ22 are normalized by the equivalent stress. Again, it must be noted
depends on σ22 (i.e. the transverse stress component). Consequently, the that σ11 and σ22 are aligned with the RD and the TD, respectively. For the
anisotropic parameter F is indeed less activated when the tensile di­ material test along the RD and the TD shown in Fig. 11(a) and Fig. 11(b),
rection is aligned with the RD. The latter is confirmed by looking at the respectively, the stress distributions are identical, yet almost mirrored
importance ranking for the Test C RD shown in Fig. 9. The importance according the dotted diagonal. Compared to stress distribution of Test C
ranking is calculated based on Eq. (14) in Table 3 and represents the RD and Test C TD, the stress states generated by Test C R45 (Fig. 11(c)) is
parameter sensitivity strength during FEMU. much more dispersed, covering a larger area. Hence, the probability of
In order to improve the understanding with regard to the most activating the anisotropic parameters increases significantly. This

12
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 11. Stress states for (a) Test C RD, (b) Test C TD and (c) Test C R45 using Hill48 criterion.

Table 11
Table 10
Reference parameters, initial guess, identified parameters and identification
Comparison of identified parameters accuracy rank, IT and identifiability metrics
quality for Test C TD and Test C R45.
rank for seven material orientations tests.
Parameters Reference Initial Test C TD Test C R45
Material Accuracy IT − γK − IK − ρK −
lists values guess
Orientation Rank Rank Rank Rank Rank
1
θ Ref θ Identified Identified
RD 7 7 2 6 7
parameters parameters
R15 3 3 4 7 5
α1 0.9394 1.0 1.1565 0.9650
R30 5 1 3 2 3
(23.11%) 2 (2.72%)
R45 1 2 1 1 1
α2 1.1841 1.0 1.2065 1.2271
R60 6 4 5 4 2
(1.92%) (3.63%)
R75 2 6 7 5 4
α3 0.8872 1.0 1.3291 0.9176
TD 4 5 6 3 6
(49.81%) (3.4297%)
Sum of least square for rank 38 60 40 32
α4 0.8765 1.0 0.9416 0.8998
order difference w.r.t Accuracy
(7.43%) (2.66%)
Rank
α5 0.9333 1.0 0.2491 0.9575
(73.31%) (2.59%)
α6 0.802 1.0 -0.2315 0.8083
potentially explains why R45 exhibits the best identification quality (128.9%) (0.78%)
when applying FEMU. α7 1.0462 1.0 1.0917 1.0742
(4.35%) (2.68%)
1.0227 1.0 1.2484 1.0538
4.2.3. Optimal identifiability criterion α8
(22.07%) (3.04%)
Finally, this section aims to determine the most reliable identifi­ M 5.90 6.0 3.7085 5.8895
ability metric for predicting the identification quality. Table 10 com­ (37.14%) (0.1776%)
pares the accuracy rank based on the A.R.E (table 9), the ranking of A.R.E 38.37% 2.41%
indicator IT and the identifiability metrics for the seven material Metric a: γk 83.70 23.97
Metric b: IK 5.90 3.96
orientation tests.
Metric c: ρK 9.58 36.02
The last row in Table 10 is the sum of least squares, expressing the Parameter subset (Full parameters) {α1 , α2 , α3 , α4 , α5 , {α1 , α2 , α3 , α4 , α5 ,
difference between the accuracy rank (second column) and the different α6 , α7 , α8 , M} α6 , α7 , α8 , M}
metrics. It can be inferred from Table 10 that ρK yields the smallest
1. Reference anisotropic parameters for Yld2000-2D.
difference with the accuracy rank and is therefore deemed the most
2. The data in the parenthesis represents I.R.E.
suited criterion to assess identification quality. This metric also enables
the correct prediction of the best and worst material orientation.
Although it is not possible to assign an absolute threshold to ρK, it can be γK for predicting the accuracy rank. Most importantly, the three iden­
used to evaluate the influence of the material orientation on the pa­ tifiability metrics both successfully predict that R45 is the optimal ma­
rameters identifiability. The IT rank also shows a good match with the terial orientation ( as shown in Table 10).
accuracy rank. However, IT cannot provide a threshold value for good It is important to understand that the sensitivity of anisotropic pa­
identifiability. Moreover, a larger IT does not guarantee a good param­ rameters depends on the selected orientation of the principal material
eters identifiability. Compared to γK, it can be stated that IK is better than axis with respect to the loading direction. To find the most adequate
material orientation, the determinant measure ρK is a more suited

13
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 12. (a), (c) are comparisons of the reference Yld2000-2D yield loci (red dash line) with identified yield locus, stress states distribution and corresponding density
contour plot for Test C TD and Test C R45.respectivity; (b), (d) are reference yield loci with identified yield loci for Test C TD and Test C R45 at different levels of σ12
/σ0 , respectively.

metric. Indeed, the γK and IKare effective measures to verify whether the tests based on Test C, denoted as Test C TD and Test C R45, respectively. It
mechanical test can be used for to properly identify unknown aniso­ must be emphasized that, except for the material orientation, all other
tropic parameters. The metric ρK can be used as comparable tool to find conditions (FE model, etc.) and procedures (virtual experiment and
the optimal material orientation. As a consequence, in the context of this parameters identification, etc.) are kept identical to Test C reported in
work on identifiability of anisotropic parameters, it is deemed appro­ Section 4.2. The full parameter set of Yld2000-2D yield criterion is
priate to combine γK or IK and ρK as joint diagnostic tool to evaluate the considered as unknown in the FEMU process.
parameters identifiability. Table 11 shows the reference anisotropic parameters of the Yld2000-
2D yield function that were calibrated by standard experiments (more
details can be found in Coppieters et al. [28]). In addition, Table 11
4.3. Validation for an advanced anisotropic yield function: Yld2000-2D shows the initial guess values, the inversely identified anisotropic pa­
rameters, the average identification quality (A.R.E), the individual
4.3.1. Parameters identifiabiltiy and identification identification quality (I.R.E) and the three identifiability metrics for Test
The above research findings are based on a relatively simple aniso­ C RD and Test C R45, respectively.
tropic yield criterion. Indeed, the plane stress Hill48 yield criterion in­ As shown in Table 11, γk for Test C TD is significantly larger (83.7)
volves only three anisotropic parameters. This raises the question than the threshold value (20). Similarity, IK(5.90) for Test C TD is also
whether the findings hold for a more complex anisotropic yield function over the threshold value (3). This indicates a poor identification quality,
having multiple anisotropic parameters.With the purpose of validating which is indeed reflected by its large A.R.E (38.67%). Thus, this dem­
the specimen design methodology (importance of material orientation onstrates that Test C TD is not sufficiently rich to uniquely identify the
and parameter identifiability) for a more complex material model, the full parameters set of the Yld2000-2D yield function. Conversely, the full
identifiability analysis is applied in this section to the Yld2000-2D yield parameters set can be remarkably be well identified for Test C R45 (A.R.
function (see Appendix A for more details). Yld2000-2D yield criterion E -2.41%). In addition, γk (23.97) and IK(3.96) for Test C R45 are close to
involves nine parameters (αi(i = 1, ..., 8) and exponent M). Here, two the threshold values γk (20) and IK(3). This suggests that the threshold
material orientations (TD and R45) are considered for the mechanical

14
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

Fig. 13. Evolution of (a) r − valueand (b) normalized yield stress according to tension axis angle from rolling direction for identified parameters and reference
parameters for Test C TD and Test C R45.

values for γk and IKmight need to be slightly tailored for a particular process could not retrieve the reference yield locus due to a lack of in­
constitutive model and the targeted identification quality. The latter will formation. The inversely identified anisotropic parameters found
be investigated in a forthcoming paper. Moreover, γk and IKfor Test C through Test C TD are used in Fig. 13(a) and Fig. 13(b) to predict the r-
R45 are notably lower than observed for Test C TD. In Section 4.2.3, we values and normalized yield stresses along different orientations with
concluded that a higherρKcorresponds with a better parameters identi­ respect to the RD. It can be inferred that both predictions are very poor,
fiability. The latter conclusion is also valid for the Yld2000-2D yield except for the r-values calculated in the vicinity of TD. Indeed, in this
function since ρK (36.02) for Test C R45 is almost four times larger than region, the direction of the plastic strain rate is correctly predicted (see
ρK (9.58) for Test C TD. The results shown in Table 11 confirm that the A. Fig. 12(a)). Given that the strain hardening is assumed to be known in
R.E shows good correlation with γk, IKand ρK for the Yld2000-2D yield the FEMU process, it is not surprising that the yield stress in the RD is
function. As a consequence, this validates the proposed identifiability correctly predicted (see Fig. 13(b)).
analysis via the simultaneous use of three identifiability metrics, namely It can be inferred from Fig.12(c) that Test C R45 enables to have more
γk, IKand ρK. Finally, the results in Table 11 clearly show the importance dispersed stress states than Test C TD. This leads to a much better
of the material orientation for improving the parameter identifiability. identification of the yield locus as shown in Fig. 12(c)-(d). In addition,
the parameters found through Test C R45 enable to accurately predict
4.3.2. Yield stress and plastic anisotropy prediction using the identified the r-values and yield stresses shown in Fig. 13(a)-(b).
parameters
The previous section was based on a direct comparison between the 5. CONCLUSIONS
identified and the reference parameters. Such a direct comparison is
somewhat questionable since small parameter deviations can have a This paper scrutinizes the effectiveness of an IT-based design (i.e. a
strong influence on the predictive accuracy of the model. Therefore, it notched tensile specimen) to inversely identify the plastic anisotropy of
advised to use the identified parameters to plot the normalized locus and sheet metal using heterogeneous strain fields. To this end, virtual DIC
compare it with the reference yield locus. Moreover, the identified data was generated using reference material behavior and fed to a FEMU
model parameters can be used to predict yield stress and the r − values (Finite Element Model Updating) process to assess the identification
as a function of the angle from the RD. The latter predictions can then quality of three IT-based designs. For a relatively simple model such as
again be compared with the reference material behaviour. This the plane stress Hill48 yield criterion, a positive correlation between the
approach gives a more profound understanding of the identification strain heterogeneity indicator-IT and the inverse identification quality is
quality. Fig. 12(a)-(b) and Fig. 12(c)-(d) shows the comparison between observed. Despite this positive correlation, however, IT-based designs
the identified yield loci and the reference yield locus (at different values cannot guarantee maximum and sufficient identifiability of the sought
of σ12 /σ0 ) for Test C TD and Test C R45, respectively. In addition, the anisotropic parameters. In this regard, it is shown that a set of identifi­
stress states (labelled reference data) generated in the AOI for Test TD and ability metrics enables to accurately predict the identification quality
Test R45 are shown. The density of the stress states is indicated with a found through FEMU. The following results are validated for an
color legend normalized to 1. For Test C TD, see Fig. 12(a), it can be advanced anisotropic yield function (i.e., Yld2000-2D):
inferred that the stress states are clustered around uniaxial tension in the
TD, yet are moderately dispersed in the first quadrant of stress space. • It is recommended to combine the γK or IK and ρK as joint diagnostic
The identified yield locus is shown for σ 12 /σ 0 = 0 in Fig. 12(a) and for tool to rank and assess the effectiveness of heterogeneous specimen.
different values of σ12 /σ 0 in Fig. 12(b). It can be inferred that the With regard to the optimization of the material orientation, the
identified locus is only accurate in the densely populated region of the determinant measure ρK is found to be the most reliable identifi­
probed stress states. Clearly, significant deviations occur elsewhere. This ability metric.
means that the inversely identified anisotropic parameters enable to • The identifiability of the IT-based design can be significantly
reproduce the material behavior in Test C TD, however, the FEMU improved by considering the orientation of the principal material

15
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

axis within the specimen. Therefore, the material orientation should Declaration of Competing Interest
be considered as a design variable for increasing identifiability of
anisotropic parameters. The authors have no conflict of interest to declare.

The above findings can be used to enhance the design strategy for Acknowledgements
heterogeneous specimen. The presented identifiability analysis fully
relies on numerical strain fields generated with a FE model. In other Yi Zhang acknowledges the financial support from the program of
words, it is assumed that the actual strain fields can be measured with China Scholarships Council (No.201806460097). Yi Zhang and S. Cop­
sufficient accuracy. Future work will embark on integrating metrolog­ pieters acknowledge MatchID for the use of the MatchID software. S.
ical aspects and measurement noise within the framework of Coppieters and A. Andrade-Campos gratefully acknowledge the support
identifiability. from the Research Fund for Coal and Steel under grant agreement No
888153 (EU-RFCS 2019 project No. 888153 | vForm-xSteels).

Appendix A. : Yld2000-2D

The Yld2000-2D [27] yield function is defined as:


[ ]
1 ( ⃒⃒ ′ ′ ⃒M ⃒ ⃒M ⃒ ⃒M ) 1/M
σ Yld
eq = X1 − X2 ⃒ + ⃒2X2′′ + X1′′ ⃒ + ⃒2X1′′ + X2′′ ⃒ , (A.1)
2

where exponent “M” is a material coefficient and based on the metal micro-structure(M= 8 for FCC and M= 6 for BCC) and Xi , Xi′′ are the principal

values of two stress tensors X , X , which are obtained by employing linear transformations on Cauchy stress deviators in plane stress, σ =
′ ′′

[σxx σ yy σxy ]T
{ ′ ′
X =Lσ
, (A.2)
X = L′′ σ
′′

where Rolling Direction (RD) and the Transverse Direction (TD) are aligned with x and y axis, respectively, The two tensors L and L are function of
′ ′′

eight independent parameters αi, which are given as:


⎛ ⎞
2α − α1 0
1⎝ 1
(A.3)

L = − α2 2α2 0 ⎠ ,
3
0 0 3a7
⎛ ⎞
( − 2α3 + 2α4 + 8α5 − 2α6 ) (α3 − 4α4 − 4α5 + 4α6 ) 0
1⎝
′′
L = (4α3 − 4α4 − 4α5 + α6 ) ( − 2α3 + 8α4 + 2α5 − 2α6 ) 0 ⎠ . (A.4)
9
0 0 9a8

The principal values X and X′′ are:


√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅)
1( ′ ( )2
(A.5)
′ ′
X xx − X yy + 4X ′xy 2 ,
′ ′
Xi = Xxx + Xyy ±
2
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅)
1 ( ′′ ( )2
Xi′′ = Xxx + Xyy′′
± X ′′xx − X ′′yy + 4X ′′xy 2 . (A.6)
2

References [7] Zhang H, Coppieters S, Jimenez-Pena C, et al. Inverse identification of the post-
necking work hardening behavior of thick HSS through full-field strain
measurements during diffuse necking. Mechanics of Materials [J] 2019;129:
[1] Kim J H, Barlat F, Pierron F, et al. Determination of Anisotropic Plastic Constitutive
361–74.
Parameters Using the Virtual Fields Method. Experimental Mechanics [J] 2014;54
[8] Cooreman S, Lecompte D, Sol H, et al. Identification of mechanical material
(7):1189–204.
behavior through inverse modeling and DIC. Experimental Mechanics [J] 2008;48
[2] Ha J, Coppieters S, Korkolis YP. On the expansion of a circular hole in an
(4):421–33.
orthotropic elastoplastic thin sheet. International Journal of Mechanical Sciences
[9] Pottier T, Toussaint F, Vacher P. Contribution of heterogeneous strain field
[J] 2020;182:105706.
measurements and boundary conditions modelling in inverse identification of
[3] Banabic D, Carleer B, Comsa D S, et al. Sheet metal forming processes: Constitutive
material parameters. European Journal of Mechanics A-Solids [J] 2011;30(3):
modelling and numerical simulation [M]. 2010.
373–82.
[4] Lattanzi A, Barlat F, Pierron F, et al. Inverse identification strategies for the
[10] Denys K, Coppieters S, Seefeldt M, et al. Multi-DIC setup for the identification of a
characterization of transformation-based anisotropic plasticity models with the
3D anisotropic yield surface of thick high strength steel using a double perforated
non-linear VFM. International Journal of Mechanical Sciences [J] 2020;173:
specimen. Mechanics of Materials [J] 2016:96–108.
105422.
[11] Coppieters S, Hakoyama T, Debruyne D, et al. Inverse Yield Locus Identification
[5] Cooreman S. Identification of the plastic material behavior through full-field
using a biaxial tension apparatus with link mechanism and displacement fields.
displacement measurements and inverse methods. Belgium: Free University of
Journal of Physics Conference [J] 2018;1063:012039.
Brussels; 2008. PhD thesis.
[12] Güner A, Soyarslan C, Brosius A, et al. Characterization of anisotropy of sheet
[6] Lecompte D, Cooreman S, Coppieters S, et al. Parameter identification for
metals employing inhomogeneous strain fields for Yld2000-2D yield function.
anisotropic plasticity model using digital image correlation: Comparison between
International Journal of Solids and Structures [J] 2012;49(25):3517–27.
uni-axial and bi-axial tensile testing. European Journal of Computational
Mechanics [J] 2009;18(3-4):393–418.

16
Y. Zhang et al. International Journal of Mechanical Sciences 214 (2022) 106891

[13] Prates P A, Oliveira M C, Fernandes J V. A new strategy for the simultaneous [28] Coppieters S, Hakoyama T, Eyckens P, et al. On the synergy between physical and
identification of constitutive laws parameters of metal sheets using a single test. virtual sheet metal testing: calibration of anisotropic yield functions using a
Computational Materials Science [J] 2014;85:102–20. microstructure-based plasticity model. International Journal of Material Forming
[14] Wang Yueqi, et al. Anisotropic yield surface identification of sheet metal through [J] 2018:741–59.
stereo finite element model updating. The Journal of Strain Analysis for [29] Hill R. On discontinuous plastic states, with special reference to localized necking
Engineering Design [J] 2016;51(8):598–611. in thin sheets. Journal of the Mechanics and Physics of Solid [J] 1952;1:19–30.
[15] Pierron F, Grédiac M. A review of test design for identification of constitutive [30] Levenberg K. A method for the solution of certain non-linear problems by least
parameters from full-field measurements. Strain [J] 2020. squares. Quarterly of Applied Mathematics [J] 1944;2(2):164–8.
[16] Aquino J, Andrade-Campos A, Martins JM, et al. Design of heterogeneous [31] Marquardt DW. An Algorithm for Least-Squares Estimation of Nonlinear
mechanical tests: Numerical methodology and experimental validation. Strain [J], Parameters. Journal of the Society for Industrial and Applied Mathematics [J]
2019;55(1):e12313. 1963;11(2):431–41.
[17] Souto N, Thuillier S, Andrade-Campos A. Design of an indicator to characterize and [32] Gothivarekar S, Coppieters S, Van de Velde, et al. Advanced FE model validation of
classify mechanical tests for sheet metals. International Journal of Mechanical cold-forming process using DIC: Air bending of high strength steel. International
Sciences [J] 2015:252–71. 101-102. Journal of Material Forming [J] 2020;13(3):409–21.
[18] Souto N. Computational design of a technological mechanical test for material [33] Lava P, Cooreman S, Coppieters S, et al. Assessment of Measuring Errors in DIC
characterization by inverse analysis, 2015, PhD thesis, University of Aveiro. Using Deformation Fields Generated by Plastic FEA. Optics and Lasers in
[19] Chamoin L, Jailin C, Diaz M, et al. Coupling between topology optimization and Engineering [J] 2009;47(7):747–53.
digital image correlation for the design of specimen dedicated to selected material [34] MatchID. 2021. Available online: https://www.matchid.eu/Software.html
parameters identification. International Journal of Solids and Structures [J] 2020: (accessed on 4 February 2021).
193–4. [35] Machta BB, Chachra R, Transtrum MK, et al. Parameter space compression
[20] Barroqueiro B, Andrade-Campos A, Dias-de-Oliveira João, et al. Design of underlies emergent theories and predictive models. Science [J] 2013;342(6158):
mechanical heterogeneous specimens using topology optimization. International 604–7.
Journal of Mechanical Sciences [J] 2020:105764. [36] Sin G, Meyer AS, Gernaey KV. Assessing reliability of cellulose hydrolysis models to
[21] Barick MC, Gaillard Y, Lejeune A, et al. On the uniqueness of intrinsic viscoelastic support biofuel process design-identifiability and uncertainty analysis. Computers
properties of materials extracted from nanoindentation using FEMU. International & Chemical Engineering [J] 2010;34(9):1385–92.
Journal of Solids and Structures [J] 2020;202(1):929–46. [37] Hartmann S, Gilbert RR. Material parameter identification using finite elements
[22] Richard F, Villars M, Thibaud S. Viscoelastic modeling and quantitative with time-adaptive higher-order time integration and experimental full-field strain
experimental characterization of normal and osteoarthritic human articular information. Computational Mechanics [J] 2021:1–18.
cartilage using indentation. Journal of the mechanical behavior of biomedical [38] Brun R, Kühni M, Siegrist H, et al. Practical identifiability of ASM2d parameters-
materials [J] 2013;24:41–52. systematic selection and tuning of parameter subsets. Water research [J] 2002;36
[23] Pac MJ, Giljean S, Rousselot C, et al. Microstructural and elasto-plastic material (16):4113–27.
parameters identification by inverse finite elements method of Ti sputtered thin [39] Babarenda Gamage, T. Constitutive parameter identifiability and the design of
films from Berkovich nano-indentation experiments. Thin Solid Films [J] 2014; experiments for applications in breast biomechanics. 2016. PhD thesis, University
569:81–92. of Auckland.
[24] Renner E, Bourceret A, Gaillard Y, et al. Identifiability of single crystal plasticity [40] Gábor A, Villaverde Alejandro F, et al. Parameter identifiability analysis and
parameters from residual topographies in Berkovich nanoindentation on FCC visualization in large-scale kinetic models of biosystems. BMC systems biology [J]
nickel. Journal of the Mechanics and Physics of Solids [J] 2020;138:103916. 2017;11(1):1–16.
[25] Hapsari G, Richard F, Hmida RB, et al. Instrumented Incremental Sheet Testing for [41] Gujarati D. Basic Econometrics. Fourth edition, McGraw-Hill, Irwin. 2004.
material behavior extraction under very large strain: Information richness of [42] Martinez-Lopez B, Peyron S, Gontard N, et al. Practical identifiability analysis for
continuous force measurement. Materials & Design [J] 2018;140:317–31. the characterization of mass transport properties in migration tests. Industrial &
[26] Brun R, Reichert P, Künsch HR. Practical identifiability analysis of large Engineering Chemistry Research [J] 2015;54(17):4725–36.
environmental simulation models. Water Resources Research [J] 2001;37(4): [43] Shutov AV, Kaygorodtseva AA. Parameter identification in elasto-plasticity:
1015–30. distance between parameters and impact of measurement errors. Journal of
[27] Barlat F, Brem JC, Yoon. Plane stress yield function for aluminum alloy Applied Mathematics and Mechanics [J] 2019;8(99). 2019, 201800340.
sheets—part 1: theory. International Journal of Plasticity [J] 2003;19(9): [44] Zhang Y, Coppieters S, Gothivarekar S, et al. Independent Validation of Generic
1297–319. Specimen Design for Inverse Identification of Plastic Anisotropy. International
Conference on Material Forming, Esaform 2021.

17

You might also like