Meccanica (2007) 42:9–17

DOI 10.1007/s11012-006-9019-5

O R I G I NA L PA P E R

Identification of Material Damage Model Parameters:

an Inverse Approach Using Digital Image Processing
Giovanni B. Broggiato · Francesca Campana ·
Luca Cortese

Received: 6 May 2006 / Accepted: 9 June 2006 / Published online: 28 September 2006
© Springer Science+Business Media B.V. 2006

Abstract A reliable prediction of ductile failure the measuring of specimen profile during load-
in metals is still a wide-open matter of research. ing. The advantage of this analysis is that more
Several models are available in the literature, rang- data are available for calibration thus allowing a
ing from empirical criteria, porosity-based models greater level of confidence and accuracy in model
and continuum damage mechanics (CDM). One parameter evaluation.
major issue is the accurate identification of
parameters which describe material behavior. For Keywords Damage mechanics · Image analysis ·
some damage models, parameter identification is Inverse methods · Parameter identification ·
more or less straightforward, being possible to per- Machine design
form experiments for their evaluation. For the oth-
ers, direct calibration from laboratory tests is not
possible, so that the approach of inverse methods
is required for a proper identification. In material 1 Introduction
model calibration, the inverse approach consists
in a non-linear iterative fitting of a parameter- Nowadays investigation on damage models for plas-
dependent load–displacement curve (coming from tic failure prediction of ductile metals is becoming
a FEM simulation) on the experimental specimen more and more relevant. Plasticity theory is quite
response. The test is usually a tensile test on a consolidated and capable to predict with great accu-
round-notched cylindrical bar. The present paper racy material response in terms of stress–strain
shows a novel inverse procedure aimed to estimate fields. But this theory can neither give information
the material parameters of the Gurson–Tvergaard– about damage evolution in the material nor about
Needleman (GTN) porosity-based plastic damage its failure, so additional damage models must be
model by means of experimental data collected developed for that task. It is thus evident that a
using image analysis. The use of digital image pro- tool, which could predict damage evolution and
cessing allows to substitute the load–displacement failure of materials during plastic deformation, is
curve with other global quantities resulting from of utmost importance in all fields involving cold
forming processes of all kinds. Not only failure is
G. B. Broggiato (B) · F. Campana · L. Cortese important, but also the knowledge of the percent-
Department of Mechanics and Aeronautics University
of Rome “La Sapienza”
age of damage reached during plastic deformation
via Eudossiana, 18, 00184 Rome, Italy with respect to final fracture. This is needed to pro-
e-mail: giovanni.broggiato@uniroma1.it duce safe components at in-service load conditions.
10 Meccanica (2007) 42:9–17

Many works can be found in the literature, microcavities which are responsible for damage
dealing with various aspects of damage [1]. It is growth and subsequent failure. They state that fail-
well known that one of the most influent aspects ure occurs when void volume fraction in the mate-
on damage accumulation is the triaxiality level of rial reaches a critical value. Among them, the most
stress [2]. well known is the Gurson model [4, 6] and its
Triaxiality can be defined as modification due to Tvedgaard and Needleman
σH (GTN model) [4, 7]. It should also be cited the
T= (1)
σEq thermodynamically consistent Rousselier model
where σH is the hydrostatic stress and σEq the von [8] which can be placed in between CDM and
Mises equivalent stress; triaxiality is influent both porosity-based models.
on the level of damage and on its creation mecha- The GTN model is the one adopted for the
nisms. Relying on these considerations, many mod- investigation presented in the paper. Although the
els have been devised and developed to describe proposed calibration approach is general and
how a material reaches failure as a consequence of appropriate to any other damage model, the choice
plastic deformation and triaxial stress state. They of GTN model is due to the fact that it is so far one
can be divided into three main groups: empiri- of the few that is implemented in commercial fi-
cal criteria, continuum damage mechanics (CDM) nite element codes. In the next chapter, a detailed
models and porosity-based models. description of its theoretical background will be
Empirical criteria are a sort of damage accu- provided.
mulation laws where damage D itself grows with It should be pointed out that, in modeling a
equivalent plastic strain εp weighted on a proper material mechanical behavior and its damage
function f (T) of the triaxiality state of stress: mechanisms, the main experimental troubles and
 implementation concerns are associated to the diffi-
D = f (T)dεp (2) cult duty of a correct and reliable evaluation of
model parameters. The GTN model requires at
They differ one another for the form that the least six parameters for its calibration, along with
weighting function assumes. It is worth to remem- the stress–strain curve of the ideal undamaged
ber the Rice–Tracey criterion [2] and its modifica- material. Moreover, GTN parameters (as in many
tion proposed by Hancock and Mackenzie [3]. other models) cannot be directly measured with
CDM models derive damage evolution laws from any kind of standard laboratory test; thus they must
a thermodynamically consistent framework where be assessed indirectly through an inverse method.
damage is observed at a mesoscale level. Damage The inverse approach consists in a minimization
takes into account the progressive degradation of of an error function which is the difference between
the material properties and the loss of stiffness due appropriate global quantities, coming from
to the irreversible process of void nucleation and experimental tests, and the same quantities
growth. Damage variable is a scalar quantity for an obtained as a parameter-dependent response, from
isotropic material and it is based on the effects of a finite element model simulating the experimen-
damage on its elastic behavior: tal setup. The optimization algorithm chosen is the
Eeff zero-order simplex method. Though slow, it is ro-
D=1− (3)
E0 bust and does not require the evaluation of target
where Eeff is the damaged material elasticity function derivatives; in fact, faster gradient meth-
modulus and E0 is its initial value. They are likely ods may lead to numerical problems because of
the most refined and consistent models though response discontinuities at final specimen break.
sometimes hard to be implemented and calibrated. Damage model must be implemented into the
Among them, the Lemaitre model [4], which is a FEM code (in this work MSC Marc 2005 has
coupled elasto-plastic and isotropic model and the been used) and, obviously, data used for fitting
Bonora model [5] should be mentioned. must be sensible to parameter variations. Testing
Porosity-based models start from a micro- usually consists of tensile tests on round-notched
structural investigation on microdefects and cylindrical bars; usually, global quantities are
Meccanica (2007) 42:9–17 11

load–displacement curves acquired through load Then, these data can be used as global quantities
cell and clip gage. Round-notched bars are partic- for fitting replacing load–displacement curves. The
ularly effective because different notch radii are main advantages are that more and more data are
indicative of different levels of triaxiality and for available for fitting purposes and that these data
the reason that they can be studied by axisymmet- refer to specimen cross-sections with different lev-
ric numerical models. els of damage, deformation and triaxiality. Let us
The necessity of defining the undamaged stress– remember that all models described above are lo-
strain curve up to high plastic strains, as requested cal ones, thus every point of the specimen contains
by GTN model, causes two problems. First, information about damage. It will be shown that by
experimental data can only give evidence of mate- adopting this method is it possible to achieve the
rial behavior at its actual damaged status. Second, same level of accuracy (in terms of damage pre-
in the post-necking regime, where experimental diction and transferability) of the multi-test tradi-
(damaged material) data are unavailable, the tional calibration while performing a single test on
undamaged material hardening curve cannot be a round-notched bar specimen.
separately estimated from damage parameters.
These problems have been faced describing the
post-necking regime of the constitutive law as a 2 GTN Model And Traditional Calibration
linear weighted combination of tangent extrapola-
tion and power law extrapolation of the pre-neck- GTN model assumes that damage is locally due to
ing regime (as shown by Ling in Ref. [9]). The linear nucleation, growth and subsequent coalescence of
combination weight, w, enters then in calibration voids inside the material; these three mechanisms
procedure as further damage parameter. cause a loss of resistance that progressively leads
A further difficulty of GTN model is that its to failure. This reduction of resistance is described
parameters are coupled, especially those that de- by the introduction in the yield surface, described
scribe constitutive behavior with the damage ones. by the plastic potential, of a softening term, which
When calibration is performed on a single geome- is a function of the created and accumulated per-
try round-notched test (single-fit calibration), this centage of voids (void volume fraction, f = VV0 ):
problem leads to poor consistency of results (the
fitting parameter set is not unique: more the one σEq
2

minimum can be found); this can be easily revealed FGurson = − 1 + 2 q1 fv∗

σs2
by their bad transferability to differently shaped  
3 σH  2
specimens. To overcome this obstacle, in an at- × cosh q2 − q1 fv∗ (4)
tempt of decreasing coupling effects on parameter 2 σs
 
identification, a multi-test fit calibration has been σs = σy + R represents the current yield stress
performed on specimens with different levels of described by the isotropic hardening curve R, σH
triaxiality. Doing so, the model portability and reli- stands for the hydrostatic stress and fv∗ is the effec-
ability can be highly improved. tive void volume fraction. Concerning the harden-
The same goal may also be reached exploiting ing function, as already mentioned, an ideal
the measure of displacement fields without using undamaged stress–strain relation must be provided.
force–displacement curves as shown in Ref. [10]. Up to the onset of necking, data can be taken
In this work, a further improvement is provided by from a traditional tensile test, while the post-
means of digital image processing techniques. Con- necking regime is described as a linear weighted
currently with previously described tests, combination of tangent extrapolation and power
specimen profile is acquired through a CCD cam- law extrapolation of the pre-necking regime.
era (silhouette analysis) and then post-processed The effective void volume fraction, fv∗ , is defined
to provide synchronous profile and load time as:
histories for each test. From these data it is pos- 
fv fv < fc
sible to determine mean axial-stress versus radial- ∗
fv = q−1 −f (5)
strain curves for various specimen cross-sections. fc + f1F −fc (fv − fc ) fv ≥ fc
c
12 Meccanica (2007) 42:9–17

where fc is the void volume fraction at which resulting best-fit parameters. In Table 1, a summary
coalescence occurs together with a sudden drop of calibrated model parameters are reported.
of resistance in the material. This kind of inverse methodology, exploiting
Void volume fraction evolution is given by sum- force–displacement global quantities, presents
ming a void growth and a void nucleation rate [4]: some limits:

ḟv = ḟg + ḟn – The FEM model is built on the nominal

p
ḟg = (1 − fv ) ε̇kk specimen geometry; shape irregularities in the
 (6) test specimen are not taken into account. These
p 2
fN εv −εN p
ḟn = √ exp − 12 SN ε̇v defects could give rise to appreciable differ-
SN 2π
ences in the response that leads to an incorrect
p
where εv stands for the equivalent plastic strain estimation of the parameters.
defined by energetic equivalence. – The fitting procedure is based on a limited set
GTN model states that material failure occurs of data (i.e. load and displacement samples at
when the effective void volume fraction, fv , various simulation and test time steps) that
becomes greater then the critical value fF . averages the local effects of damage evolution
From experimental tests, q1 and q2 can be held on global quantities. Since the GTN is a local
constant [4] with values: q1 = 1.5, q2 = 1.0, while damage criterion every point of the continuum
f0 , fc , fF , εN , SN and fN are effective material “contains” specific information of its damaged
parameters. From an earlier analysis, the initial status, so it can be stated that more and more
void volume fraction, f0 , of this material, has been data are available and exploitable for optimi-
stated to have no significant deviation from the zation purposes.
typical literature value: f0 = 0.002. Consequently, – Finally, some studies [1, 11] have demonstrated
being small its influence on material response, it that a particular care must be put in the simula-
has been chosen to exclude it from the set of fitting tion set-up, considering that mesh size,
parameters and to keep it constant and equal to non-linear analysis options (such as plasticity
that value. Finally, the investigated material (the options), time stepping choices and so on, can
same used for optical calibration; its classification, influence the response as soon as specimen gets
mechanical properties and specimen geometries highly damaged.
will be detailed later on) does not show a clear coa-
lescence phase, hence fc value can be set close to fF ; For all these reasons, and for considerations re-
this guarantees that as soon as material approaches lated to strain localization and “ill-posedness” of
coalescence it reaches failure immediately. In con- the problem when approaching high damage lev-
clusion, only fF , εN , SN , fN must be determined els, the application of inverse techniques to dam-
along with the weighting parameter w of the post- age problems is sometimes criticized [11]. In the
necking regime of hardening curve. present paper, the authors would like to show that
Hereafter, the results obtained using the tradi- the use of digital image processing methods al-
tional inverse approach based on multi-test fitting low to improve the robustness and effectiveness
method are presented. Experimental data have of inverse approach in GTN parameters identifica-
been collected running tensile tests on three differ- tion.
ent types of specimens: smooth cylindrical bars and
round-notched bars with notch radii of 10 mm and
2 mm and minimum diameters of 5 mm and 6 mm, 3 Optical Method
respectively. The error function is now evaluated
extending its computation to the data coming from Similarly to the technique proposed in Ref. [12],
the three tests together. Figure 1 shows experi- an easy way to collect data to be used for damage
mental load–displacement curves for the described model calibration consists in grabbing the profile
three classes of tests compared with the of a round-notched specimen through the so called
corresponding numerical ones obtained using the ‘silhouette analysis’.
Meccanica (2007) 42:9–17 13

70000
Multi-test Calibration
60000 Experimental
RNB2
Multi-test Best Fit
50000

Load (N)
40000

30000
Smooth
20000
RNB10
10000

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Displacement (mm)

Fig. 1 Traditional procedure: force–displacement curve fitting for smooth specimen and round-notched bar (RNB) specimens
with 10 mm and 2 mm notch radius respectively

Table 1 Traditional multi-test best-fit parameters

Parameter f0 fC fF εN sN fN w

Value 0.002 0.245 0.245 0.078 0.152 0.044 0.324

This approach requires an experimental setup tion of the two axes of symmetry of notch image.
where the workpiece is interposed between a This analysis requires the following steps:
digital camera and a white screen. To avoid per-
(1) A four quadrant window is overlapped on the
spective errors, the use of a telecentric lens is advis-
picture; its position is defined by the central
able, but a long focal-length standard lens gives
point A and the angle α; the content of each
good results as well. Special cares should be put in
quadrant is then mapped on four subimages
lighting adjustment, in fact specimen surface, be-
using a bilinear warping function; see Fig. 2a.
ing highly reflecting, may not appear completely
(2) A minimization algorithm is used to identify
dark on the grabbed pictures; also, a background
the exact position of A and the value of angle
screen too lighted may generate unwelcome bright
α that allow a perfect overlapping of the four
haloes. Surely, any auto gain control (AGC) sys-
subimages if mirrored along the axes B–F and
tems present on the imaging device must be
D–H (Fig. 2b).
switched off; thus, lens stop and shutter speed
(3) Since the notch profile has to be extracted and
should be trimmed manually, trying to spread the
mapped into the FEM model coordinate sys-
image histogram on the entire gray scale. About
tem, the four subimage contents is summed
focusing, you only need to get focus not on the
to give a single image of one fourth of the
specimen front face but on its silhouette profile
specimen notch (Fig. 2c).
(terminator line).
(4) After that, gradient and threshold operations
In addition, it is of primary importance to gather
allow to highlight and extract pixels that fall
load data from the material testing machine syn-
on profile edge (Fig. 2d).
chronously with image grabbing. Even if the test
(5) The pixels belonging to the specimen cylin-
can be performed quite slowly (typically ramp rate
drical sections are used to define a gauge size
is set equal to 0.02 mm/s), close to final failure, dam-
and compute the image scale (Fig. 2e).
age evolution is fairly fast because it mostly depend
(6) Finally, all the profiles identified on each frame
on plastic instability of neck section; therefore, a
are scaled and regularized to build a sequence
time mismatch between load samples and images
of contours such as those shown in Fig. 2f.
leads to significant errors in damage modeling.
Afterward an image sequence is saved, the first A first advance, which comes from the images
action performed on each frame is the identifica- processed as described above, is the opportunity of
14 Meccanica (2007) 42:9–17

Fig. 2 Image analysis procedure used to map notch profile into the same coordinate system of FEM model

building the FEM model on the basis of the actual

notch shape using the profile of the undeformed
specimen as explained in the next chapter.
Also, Bridgman’s formulas [13] may be applied
to evaluate the damaged true-stress true-strain
relation in the post-necking regime; Bridgman’s
relations require neck radius and curvature which axial stress
both can be obtained from silhouette analysis. σA
Though not useful for GTN model, where stress–
strain characterization must be related to the ideal
undamaged material, it is of great importance to radial strain
describe material behavior at high plastic strains εR
whenever a damage model is not used.
In addition, each profile collects a large amount
of data functional for parameter identification, the
proposed technique to exploit this information is
based on the stress–strain curve obtained comput-
ing the mean radial strain εR on a set of neck sec- Fig. 3 Definition of the stress–strain curves obtained on
tions and the mean axial stress σA evaluated by a set of neck sections used in optical damage parameter
calibration
dividing the applied load by their associated cross-
section areas (see Fig. 3). the corresponding parameter-dependent response,
Both the radial strain and the section area fol- coming from FEM simulation, can be fitted on
low directly from the measured profiles; the ap- these experimental data allowing parameter iden-
plied load comes from the load sample assigned tification. Although, even a perfect fitting among
to each saved picture. Similarly to the standard averaged quantities does not prove that the ac-
method, by means of a minimization algorithm, tual and the computed stress and strain fields are
Meccanica (2007) 42:9–17 15

Fig. 4 FEM specimen nominal shape correction obtained from actual undeformed profile

2500
Multi-Section Calibration

2000
Mean Axial Stress (MPa)

1500

1000

Experimental
500 First Guess
Best Fit

0
0 0.05 0.1 0.15 0.2 0.25
Mean Radial Strain (m/m)

Fig. 5 Mean axial stress vs. mean radial strain curves: Experimental, FEM first guess and FEM best-fit results

identical, the use of several cross-section along quired specimen silhouette: a dedicated algorithm
specimen neck significantly reduces the risk to ob- modifies profile nodes of the numerical model to
tain erroneous results. match actual specimen shape and also moves the
inner nodes accordingly, in order to avoid exces-
sive distortion of outer elements. This procedure
4 Application guarantees FEM model to reproduce correspond-
ing experimental test with greater accuracy.
Main results achieved by inverse method through Then, best-fit parameters have been identified
digital image processing are now summarized. A for round-notched bar with 10 mm radius by the
first advantage is that the knowledge of specimen optical inverse method described in the previous
real undeformed profile can be exploited to change chapter. The investigated material is a high strength
its initial shape of the corresponding FEM model quenched and normalized carbon steel with yield
to correct possible geometric defects due to a not stress above 1200 MPa and maximum nominal
precise machining. In Fig. 4, it is shown how fi- stress at 1600 MPa. Figure 5 shows mean axial stress
nite element grid can be modified to match ac- versus mean radial strain curves evaluated at 5
16 Meccanica (2007) 42:9–17

Table 2 Optical best-fit parameters

Parameter f0 fC fF εN sN fN w

Value 0.002 0.153 0.153 0.096 0.141 0.031 0.312

70000
Optical Calibration
60000 Experimental
RNB2
Optical Best Fit
50000
Load (N)

40000

30000
Smooth
20000
RNB10
10000

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Displacement (mm)

Fig. 6 Optical procedure: force–displacement curve fitting for smooth specimen and round-notched bar (RNB) specimens
with 10 mm and 2 mm notch radius respectively

specimen cross-sections: the first one refers to min- Some differences between the two obtained sets
imum diameter section and the others are taken of parameters raise suspicions about non-unique-
along the axis of symmetry at a distance of 0.3 mm ness of the solution; nevertheless the accuracy in
from each other. Section spacing has been chosen failure prediction in both cases assures the correct-
to obtain consistent data for calibration from spec- ness of the calibration procedure. Also, attention
imen volume where different levels of triaxiality, must be put in the failure parameter values, fF ;
and consequently of damage, are present. Exper- in both cases they are higher than the ones usu-
imental response is compared with correspond- ally adopted. This can be justified by the absence,
ing numerical simulation obtained from first guess for the investigated material, of a clear coales-
parameters and finally from best-fit parameters. cence phase that accelerates resistance loss. Con-
It can be noticed how numerical response closely sequently damage proceeds more slowly, leading
matches experimental one for this multi-section to higher values for failure parameter.
analysis in Table 2 the best-fit parameter set is re-
ported.
Finally, a comparison between the new method 5 Conclusion
and the traditional one is put forth in terms of
force-displacement curves for the three types of GTN damage model has been calibrated success-
tests: the results in Fig. 6 show that the optical fully by an inverse approach exploiting digital
method calibrated on single test gives the same image processing techniques. This method is shown
level of agreement of the traditional method tuned to provide the same level of accuracy of a tradi-
through a multi-test fit showed in Fig. 1. For the tional multi-test fitting. To guarantee result
optical method, the error (in terms of elongation at transferability, the standard method needs to be
failure) is below 1% for round-notched specimen calibrated on specimens of different geometry, thus
used for calibration (notch radius 10 mm, minimum requiring more experimental effort. The major
diameter 5 mm) and is contained under 5% when advantage of the optical method is to provide por-
a cross check is performed on the other geometries tability requiring only a single test on a round-
available. notched bar specimen.
Meccanica (2007) 42:9–17 17

This is a consequence of the greater quantities 3. Hancock JW, Mackenzie AC (1976) On the mecha-
of damage-dependent information available from nism of ductile failure in high-strength steel subjected
to multi-axial stresses. J Mech Phys Solids 24:147–169
multi-section analysis with respect to the global 4. Lemaitre J, Desmorat R (2005) Engineering Damage
force–displacement response used for standard Mechanics. Springer
procedure. As a disadvantage, a more complex 5. Bonora N (1997) On the effect of triaxiality state of
experimental equipment and setup is necessary, as stress on ductility using nonlinear CDM model. Int J
Fract 88:359–371
well as a greater post-processing effort on raw data 6. Zhang ZL, Thaulow C, Odegard J (2000) A complete
collected from digital camera. Gurson model approach for metal ductile fracture. Eng
Finally, as a further validation, it is desirable the Fract Mech 67:155–168
application of the calibrated model on different 7. Springmann M, Kuna M (2005) Identification of mate-
rial parameters of the Gurson–Tvergaard–Needleman
and more complex geometries, in particular on model by combined experimental and numerical tech-
critical components of actual applications. niques. Comput Mat Sci 33:501–509
8. Rousselier G (2001) The Rousselier model for porous
Acknowledgements The authors wish to thank the metal plasticity and ductile fracture. In: Lemaitre J (ed)
Mechanical and Automotive Product Division of Tenaris- Handbook of Materials Behavior Models. Academic
Dalmine Company leaded by Dr. Mario Rossi for material Press, pp 436–445
and specimen supplies. They are also grateful to 9. Ling Y (1996) Uniaxial true stress–strain after necking.
Dr. Tommaso Coppola of Centro Sviluppo Materiali S.p.A. AMP J Technol 5:37–48
for his appreciable technical support. 10. Springmann M, Kuna M (2003) Identification of mate-
rial parameters of the Rousselier model by non-linear
optimization. Comput Mat Sci 26:202–209
11. Bonora N (1999) Identification and measurement
References of ductile damage parameters. J Strain Anal 34:
463–478
1. Berdin C, Besson J, Bugat S, Desmorat R, Feyel F, 12. Vavrik D, Korous J, Zemankova J, Jaros P (2001) Opti-
Forest S, Lorentz E, Maire E, Pardoen T, Pineau A, cal determination of constitutive equations parameters
Tanguy B (2004) In: Jacques Besson (ed) Local ap- of nonlinear materials. Proceeding of the 39th Inter-
proach to fracture. Les Presses de l’Ecole des Mines, national Conference on Experimental Stress Analysis
Paris (EAN2001), Tabor 2001.
2. Rice JR, Tracey DM (1969) On the ductile enlarge- 13. McClintock FA, Argon AS (1966) Mechanical behav-
ment of voids in triaxial stress fields. J Mech Phys Solids ior of materials. Addison-Wesley Publishing Company,
17:201–217 pp 322–324