Professional Documents
Culture Documents
Artificial Neural Network and Finite Element Model
Artificial Neural Network and Finite Element Model
net/publication/225474334
CITATIONS READS
62 876
4 authors, including:
Ashok Saxena
University of Arkansas
230 PUBLICATIONS 4,485 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Rami Haj-Ali on 21 May 2014.
This study first presents two-dimensional (2-D) axisymmetric and three-dimensional (3-D) finite
element (FE) models of nanoindentation tests. Calculated load-displacement curves from the FE
models are compared with the load-displacement curves from nanoindentation measurements on
annealed copper. Numerical parametric studies are also performed to examine the effect of indenter
geometry and the material’s stress-strain behavior on the load-displacement response. The 2-D and
3-D FE load-displacement curves compare well with the measured results on annealed copper. The
second aspect of this study introduces a new modeling approach for indentation tests using artificial
neural networks (ANN). In this approach, ANN models are generated to approximate the FE load-
displacement curves for a wide range of material and geometric parameters. The ability of the ANN
models to predict the indentation response is examined against other FE results not used as part of
the training data. These models are shown to accurately predict the load-displacement behavior of a
nonlinear homogeneous material as well as one with a hard film, such as an oxide film, on a relatively
soft substrate. It is shown that the monotonic indentation load-displacement response during loading
contains ample information for the ANN model to extract material flow properties of the indented
material.
I. INTRODUCTION used the finite element (FE) method to solve for an indenta-
tion problem with an axisymmetric cone. The cone shape
THE strength, fatigue, and fracture properties of multi- is constrained to have the same volume as that of a pyramid-
phase metallic alloys strongly depend on the in-situ elastic shaped indenter for a given depth. Their FE hardness results
and plastic flow properties of the individual constituents of compared well with experimental data. They also presented
the microstructure. However, the in-situ mechanical proper- an elastoplastic analysis of axisymmetric conical indentation
ties of the individual constituents are difficult to determine. and showed that the shape of the plastic zone strongly
The objective of this study is to extract stress-strain behavior depends on the indenter angle, Young’s modulus, and yield
of a local region of a material using the load-displacement stress.[8] The FE analyses seem to predict an upper bound
behavior obtained during a nanoindentation test. of the experimental results, while the spherical cavity model
The general indentation problem has old roots in elasticity seems to predict a lower bound behavior. Wang and Ban-
and mechanics of solids. Prandtl[1] developed a slip-line field gret[9] used three-dimensional (3-D) FE models to investigate
solution for a flat punch on a semi-infinite medium. Hill et the indentation response in a coated system. Sun et al.[10]
al.[2] also considered a wedge indentation under plane strain developed an axisymmetric FE model to simulate the inden-
conditions. Ishlinky[3] obtained a solution for spherical tation of a hard coating on a softer substrate using a conical
indentation. Shield[4] used an axisymmetric indentation indenter. The axisymmetric indenter is constrained to have
model by a flat cylindrical punch. Tabor[5] studied the plastic the same projected area of the Berkovich indenter, as a
deformed zone in elastic-plastic material as a result of spheri- function of the indentation depth. Lichinchi et al.,[11] Fivel
cal cavity expansion with an internal hydrostatic pressure, et al.,[12] and Knapp and Follstaedt[13] used 3-D and axisym-
and Johnson[6] extended this study to include the effect of metric FE models for different indentation problems in lay-
tip geometry. However, it is difficult to analytically describe ered and homogeneous materials. A complex deformation
the complex stress-strain field due to inelastic response and field is shown in their sixfold symmetric 3-D FE model for
include detailed geometric parameters that may exist at dif- verifying their axisymmetric model.[11]
ferent scales of the material. The focus of the FE and analytical studies thus far has been
Different numerical techniques have been developed and to predict the load-displacement curves given the plastic flow
can be used in indentation problems. Bhattacharya and Nix[7] properties. In this study, an artificial neural network (ANN)
approach has been used to solve the inverse problem of
extracting the plastic flow properties from measured load-
displacement behavior.
ANASTASIA MULIANA, Graduate Research Assistant, and RAMI M. The ANN is a collection of computational cells (neurons)
HAJ-ALI, Assistant Professor, are with the School of Civil and Environmen-
tal Engineering, Georgia Institute of Technology, Atlanta, GA 30332-0355. that communicate through different connections with
Contact e-mail: rami.haj-ali@ce.gatech.edu REJANAH STEWARD, for- adjusted weights. A forward-pass-type network consists of
merly Graduate Research Assistant, School of Materials Science and Engi- a number of layers with several cells. Each cell is usually
neering, Georgia Institute of Technology, is with the Department of Material connected with all cells in both the previous and the next
Science Engineering, University of Tennessee-Knoxville, TN 37996.
ASHOK SAXENA, Professor and Chair, is with the School of Materials
layers. The first and last layers include the input and output
Science and Engineering, Georgia Institute of Technology. cells. The ANN learning and training is achieved by
Manuscript submitted July 26, 2001. adjusting the connection strengths (weights) between the
In this investigation, the Nanoindenter-II manufactured The indentation process is simulated both during loading
by MTS (Oak Ridge, TN), equipped with a Berkovich tip and unloading. During the loading process, the indenter is
is used to apply loads in the millinewton (mN) range to pushed into the specimen under load or displacement control,
produce indentations of the nanometer scale depth on a and the changing depth is monitored. For a given displace-
copper sample and generate a load-displacement curve. ment, the corresponding load is calculated from the reaction
Information obtained is analyzed usually by applying the force of the indenter (displacement control). Incremental
冦
E⑀, ⑀ Ɐ ⑀0 shaped indenters are used in the 3-D models. They are mod-
⫽ 1 [1]
k⑀p p冢 冣,
n ⑀ Ɑ ⑀0 eled as rigid surfaces. The cone-shaped indenter has the
same dimensions as the indenter in the axisymmetric model.
⑀p ⫽ ⑀ ⫺ ⑀0 Loading is applied by moving the indenter in the negative
z-axis direction in order to create the penetration in the
where E is Young’s modulus, n is hardening power, and (0, block. A second unloading process is applied by releasing
0) is the yield point. the displacement and moving the indenter back to the ini-
The properties used for annealed copper material are tial position.
Young’s modulus, E ⫽ 107.6 GPa; Poisson’s ratio, v ⫽
0.245; hardening power, n ⫽ 1.85; k ⫽ 0.32 GPa; and yield
stress, 0 ⫽ 34.5 MPa.
Contact with sliding is used to model the interaction C. Analysis of Experimental and FE Results
between the rigid indenter and the block surface. The rigid
indenter and the specimen surfaces are defined within the Figure 4 shows a comparison between the measured load-
FE procedure using a “master-slave” contact scheme. displacement response during several nanoindentation tests
and the 2-D and 3-D FE predictions for annealed copper.
As mentioned before, the 3-D FE analysis uses the Berkovich
A. Axisymmetric Models indenter, and the 2-D FE model uses an equivalent cone
The 2-D FE models consist of four-node axisymmetric indenter. The FE results, solid and dashed lines in Figure 4,
elements with reduced integration (CAX4R). A typical axi- compare favorably with the repeated experimental data. The
symmetric model and its geometry are depicted in Figure FE-simulated load-displacement behavior includes an inden-
2. The region of contact part is 1000 nm along the width tation up to 250 nm, followed by the unloading step. It is
of the specimen surface (9000 nm). A fine mesh is used noted that the 2-D and 3-D curves begin to deviate from
under the contact area and near the tip of the indenter. The each other and from the test results after an indentation depth
mesh is progressively coarser further away from the tip. All of 200 nm, and the 2-D FE response is consistently softer
nodes at the bottom of the model are prevented from moving than the 3-D model. This may be attributed to differences
in the x and y directions. Nodes along the axisymmetric axis in mesh refinement near the tip zone or other factors relating
are constrained in the x direction. to the difference between the indenters. Recall that the only
A mesh convergence study was first performed in order constraint that is satisfied between the two models is only
to determine an adequate number of nodes and elements. the common projected contact area. These differences are
This is done by running different models with increasing explored in more depth in the following discussion.
number of elements until the difference between a two con- In depth examination of the mesh deformation shows that
secutive models was negligible. The 2-D mesh that is used the number of elements in the 2-D model are sufficient to
as a result of the convergence study includes a block with capture a smooth deformed shape compared to the 3-D
815 nodes and 761 elements. The characteristic length for model. The latter is stiffer due to smaller number and larger
the elements located directly under the indenter is on the aspect ratio of the elements used. Further, comparing the
order of 1 to 10 nm. effective Mises stress contours for the 2-D and 3-D models,
A conical-shaped indenter is used in the 2-D models and Figures 5 and 6, respectively, show comparable levels of
is defined as a rigid surface with a conical angle of 70.5 stress away from the tip region. The maximum stress magni-
deg. The tip of the cone is trimmed by adding a sphere with tudes, as expected, are in the areas under the tip and have
a radius R ⫽ 20 nm. The tip radius used in the 2-D simula- values of 1.95 and 1.09 GPa for the 2-D and 3-D models,
tions may not exactly match the tip radius of the actual repectively, at indentation level of 250 nm. The stresses
Berkovich indenter. This can lead to initial mismatch appear more spatially distributed in the depth direction than
between the FE model and test results at indentation depths the 2-D case. The situation is reversed when comparing the
close to the radius magnitude. Loading is achieved by mov- contours for the effective plastic strain. The magnitude for
ing the indenter in the negative y direction in order to pene- the plastic strain is larger for the 3-D case and is localized
trate the block. A second unloading step is performed, and in a small area under the Berkovich tip. Thus, sufficient
the indenter is released back to its initial position. differences exist between the 3-D and 2-D simulations to
rationalize the small differences between the resultant load-
displacement behavior. Nevertheless, the 2-D simulations
B. Three-Dimensional Models
seem to provide an accurate overall representation of the
The 3-D indentation model and geometry are depicted in measured trends. Thus, the parametric studies described in
Figure 3. The symmetry of the model can be realized by Section III–have been performed using the 2-D model.
D. Parametric Studies For example, for hard materials, smaller indenter angles may
Three parametric studies are presented using the 2-D axi- be more suitable to reach larger depths for a given maximum
symmetric model. The first study examines the indentation force capacity.
response as a function of different indenter angles (␣ ⫽ 30, The second parametric study involved changing the strain-
40, 50, 60, and 70.5 deg). The cone angle, ␣ ⫽ 70.5 deg, hardening exponent of the indented material. The power
is the equivalent angle that maintains the same projected hardening in the Ramberg–Osgood uniaxial stress-strain
area of a cone equal to that of a Berkovich pyramid-shaped relationship of Eq. [1] is used with different values (n ⫽
indenter. The annealed copper material is modeled using 1.25, 1.5, 1.75, 2, 2.5, 5, and 10), while the indenter angle is
incremental plasticity with isotropic hardening material. The held constant at ␣ ⫽ 70.5 deg. Two predetermined unloading
Ramberg–Osgood constants and the elastic constants used levels are selected (1200 and 7200 nN). The effect of the
in the analysis are described in Section II. Four intermediate material hardening is pronounced even at very low levels
loading levels (1200, 3600, 7200, and 12,000 nN), which of deformation, as shown in Figure 8. Small changes in n
are fractions of the maximum applied load, are used to are significant enough to be resolved by the resultant load-
illustrate the different responses when loads are decreased displacement diagrams. It is then clear that the nanoindenta-
to fractions of these intermediate values. The results are tion technique can be used to determine the stress-strain
summarized in Figure 7. The force-displacement response behavior of metallic materials. However, a methodology is
is dependent on the indenter angle especially at large dis- needed to solve the inverse problem of determining the
placements. Sharper indenter tips create more deformation, stress-strain behavior from measured load-displacement
and their indentation curves are relatively shallower for the characteristics. This will be addressed in greater detail in
same applied load. At small displacements, the effect of the Section IV on ANN.
indenter angle is not pronounced because the behavior is The third parametric study includes models with different
more dependent on the spherical indenter tip geometry, oxide coating thicknesses. The goal of this investigation is
which is the same for all indenter angles. to simulate the effect of the often unknown oxidation layer
The variation in load-displacement diagram with different thickness that exists in the specimens used for nanoindenta-
indenter angles shows that it is a variable that can be chosen tion testing. The coating material is assumed as the stiffer
based on the material and microstructural characteristics. material, with Young’s modulus being 20 times larger than
However, it is then necessary to adjust the hardness and the annealed copper (E ⫽ 2151 GPa). The coating is assumed
stress-strain behavior interpreted from the measurements. as elastic and the substrate as elastoplastic materials. The
conical indenter equivalent to the Berkovich is used in a Fig. 5—Mises stress contours for 2-D (axisymmetric) FE model (indenta-
tion ⫽ 250 nm).
2-D axisymmetric model. The indentation results are shown
in Figure 9 for different coating thicknesses. As expected, a
consistent softening response exists with decreasing coating
thickness. With additional studies, such as the preceding, it IV. ANN INDENTATION MODELS
may be possible to correct the load-displacement diagrams The topology of the neural network used in this study has
for the substrate if the oxide thickness is known. It also two hidden layers. These are constrained to have the same
shows that it is necessary to conduct the nanoindentation number of neurons (nodes). This number usually determines
tests as soon as possible to prevent the results from being the capacity of the neural network. The neural-network train-
influenced by the oxide coatings, especially in certain materi- ing algorithm includes different criteria that allow ANN size
als, such as copper. growth by adding nodes in the hidden layers. Training is
Fig. 10—ANN architecture and its hidden layers’ evolution during training
of indentation response of films, with different moduli and thicknesses, on
a copper substrate.
n
⑀ ⑀0 ⫽ 0.00032
trained with a large set of input and output vector pairs that
are generated from the different FE models. The connection
冢 冣 冢冣
0
⫽
⑀0 0 ⫽ 0.0345 GPa
[3]
weights are modified during the training until the ANN The first ANN model has an input vector with two compo-
yields a force response that satisfactorily agrees with the nents, ⌬ and n, and the output is the indentation force. The
correct FE simulation results. The term back propagation FE simulation is trained with monotonic response that is
refers to the training algorithm, where the observed differ- generated using displacement control. Figure 11 describes
ence in the force output, ANN vs FE, is used to modify the the FE and ANN indentation response for a wide range of
connection weights. hardening parameters. The case with hardening parameter,
The selection of an appropriate set of training cases and n ⫽ 1.8, is not used as part of the ANN training. Instead,
procedure is very important. As in any approximation, there the ANN is trained with the other cases. The prediction of
is not a clear method to generate a comprehensive a priori ANN is compared with the FE results and is shown in thick
estimate of the required set of training cases before knowing dashed and solid lines, respectively. The trained ANN has
the outcome and convergence of the training process. Several clearly captured the indentation behavior for both data cases
training strategies with different data sets have to be initially (training and verification). It is interesting to note that con-
applied. The training data must contain the knowledge that vergence difficulties are experienced for the elastic case
describes the indentation behavior within a finite domain of (n ⫽ 1), where the analysis is terminated due excessive
input and output. This is needed to allow the trained neural deformation at about an indentation level of 70 nm. However,
network to generalize the indentation response that is implic- the ANN response beyond this level of indentation is clearly
itly included in the training data. The trained neural network reasonable, and the ANN was not degraded due to the lack
is then examined, and its response is compared, often visu- of data in this range. This can only add to the capability of
ally, with the full range of trained and untrained (verifica- ANN in generalizing the data.
tion) data. The second case study is similar to the first; however, the
The term “ANN prediction” is reserved for ANN response yield stress is not specified and is part of the ANN input.
for cases that were not used in the pretraining stages. This To this end, the ANN input vector has three variables: dis-
is used in order to examine the ANN’s ability to associate placement, hardening parameter, and yield stress (⌬, n, and
and generalize a true physical response that have not been 0). The output is the indentation measured force. Indenta-
previously “seen.” A good prediction for these cases is the tion simulations are performed using all combinations of
ultimate verification test for the ANN models. These tests hardening and yield stress parameters:
have to be applied for (input and output) response within
the domain of training. It should be expected that ANN will n ⫽ 1.7, 1.8, 1.9, 2, 2.1, 2.2, (2.25), 2.3, and 2.4
produce poor results for data that are outside the training 0 ⫽ 0.03, 0.032, (0.0325), 0.033, 0.034, (0.0345), 0.036,
domain.
0.037, and 0.04 GPa [4]
E ⫽ 107.562 GPa
ANN Case Studies
where the values in parenthesis are not used in the training
In order to demonstrate the proposed ANN modeling of cases and are reserved for verification.
indentation response, three different indentation problems Figures 12 and 13 show the indentation load-displacement
are defined, and ANN models are generated for each of response from the FE and ANN models after training. The
these cases. The FE models are used to generate indentation ANN training is successful and can be verified visually by
data, both for training and verification purposes. The first examining how well the dashed curves are close to the FE
case is for indentation in a homogeneous medium that has solid lines. It is interesting to note that, despite the load-
different hardening parameter. The Ramberg–Osgood uniax- deflection variation not being large for the response cases
ial stress-strain relation is used in the form of in Figure 12, the trained ANN response is still very close
Fig. 14—ANN prediction compared to FE indentation response for different Fig. 16—ANN and FE indentation response for different film-substrate
yield stress and hardening values (that were not used in the training). Young’s modulus ratio and fixed coating thickness.
ACKNOWLEDGMENTS
We acknowledge the assistance of Ms. Laura Reister, Oak
Ridge National Laboratory, with nanoindentation testing.
This work was partially supported by the NSF, through the
Fig. 17—ANN prediction compared to FE indentation response for different Civil and Mechanical Systems (CMS) Division, and under
film-substrate Young’s modulus ratio and fixed coating thickness. Grant No. 9876080.
REFERENCES
response of the trained ANN compared with the FE curves
1. L. Prandtl: Goettinger Nachr, Math. Phys. Kl., 1920, pp. 74-85.
for the training cases with tf ⫽ 40 nm and different stiffness 2. R. Hill, E.H. Lee, and S.J. Tupper: Proc. R. Soc. London, 1947, vol.
ratios, Ef /Es . The quality of the ANN for predicting the A188, pp. 273-90.
indentation response of a coated soft substrate is shown in 3. A. Ishlinsky: J. Appl. Math. Mech., 1944, vol. 8, p. 201.
Figure 17 for cases that are not part of the training process, 4. R.T. Shield: Proc. R. Soc. London, 1955, vol. A233, pp. 267-87.
5. D. Tabor: Rev. Phys. Technol., 1970, vol. 1, pp. 145-79.
tf ⫽ 28 nm and Ef /Es ⫽ 1.5, 2.5, 6, and 8. The ANN models 6. K.L. Johnson: J. Mech. Phys. Solids, 1970, vol. 18, pp. 115-26.
for all cases except for Ef /Es ⫽ 8 show excellent predictive 7. A.K. Bhattacharya and W.D. Nix: Int. J. Solids Struct., 1998, vol. 24
capability. This can be explained by examining the different (12), pp. 1287-98.
stiffness ratios used in the training cases in Eq. [5]. Unlike 8. A.K. Bhattacharya and W.D. Nix: Int. J. Solids Struct., 1991, vol. 27
the stiffness ratios used for verification, values in parenthe- (8), pp. 1047-58.
9. H.F. Wang and H. Bangert: Mater. Sci. Eng., 1993, vol. A16, p. 343.
sis, the last one is within a wider interval, between 7 and 10. Y. Sun, T. Bell, and S. Zheng: Thin Solid Films, 1995, vol. 258, pp.
10. This can be solved by adding more training cases; how- 198-204.
ever, the overall performance of the ANN is clearly 11. M. Lichinchi, C. Lenardi, J. Haupt, and R. Vitali: Thin Solid Films,
acceptable. 1998, vol. 333, pp. 278-86.
12. M.C. Fivel, C.F. Robertson, G.R. Canova, and L. Boulanger: Acta
Mater., 1998, vol. 46 (17), pp. 6783-94.
V. SUMMARY AND CONCLUSIONS 13. J.A. Knapp and D.M. Follstaedt: Surface Coating Technol., 1998, vols.
103–104, pp. 268-75.
Finite element modeling is an effective tool to simulate 14. J. Hertz, A. Krogh, and R.G. Palmer: Introduction to the Theory of
the indentation response at the nanoscale level. This study Neural Computations, Addison-Wesley, Redwood City, CA, 1991.
15. H. Ritter, T. Martinez, and K. Schulten: Neural Computation and Self-
presents 2-D and 3-D numerical FE models. Numerical para- Organizing Maps: An Introduction, Addison-Wesley, Redwood City,
metric studies are performed for other indentation problems CA, 1991.
to examine the effect of material and geometric parameters 16. J. Ghaboussi, D.A. Pecknold, M.-F. Zhang, and R.M. Haj-Ali: Int. J.
on the overall load-displacement response. The 2-D and 3- Num. Methods Eng., 1998, vol. 42, pp. 105-26.
D FE load-displacement curves compare well with available 17. R.M. Haj-Ali, D.A. Pecknold, and J. Ghaboussi: Modeling and Simula-
tion Based Engineering, Proc. Int. Conf. on Computational Engi-
test results on annealed copper. While the 2-D models are neering Science, ICES98, Atlanta, GA, Oct., S.N. Atluri and P.E.
accurate and computationally affordable, the use of 3-D O’Donoghue, eds., Tech Science Press, pp. 551-57.
models is also needed in cases where multiphase materials 18. R.M. Haj-Ali, K.E. Kurtis, and A.R. Sthapit: ACI Mater. J., 2001,
are investigated. vol. 98 (1), pp. 36-43.
19. J. Ghaboussi, J.H. Garrett, and X. Wu: Proc., NUMETA-90, Int. Conf.
The ANN models are proposed in the second part of this on Numerical Methods in Engineering: Theory and Applications,
study to model the indentation response as a function of Swansea, England, pp. 701-17.
different material and geometric parameters. The ability of 20. J. Ghaboussi, J.H. Garrett, and X. Wu: J. Eng. Mech. Div., ASCE,
the ANN models to predict the indentation response is exam- 1991, vol. 117 (1), pp. 132-53.
ined against other FE results that are not used as part of the 21. G. Ellis, C. Yao, R. Zhao, and D. Penumadu: J. Geotechn. Eng., ASCE,
1995, vol. 102, pp. 429-35.
training data. The proposed ANN models show excellent 22. R. Pidaparti and M. Palakal: AIAA J., 1993, vol. 31, pp. 1533-35.
predictive capability for indentation of a nonlinear homoge- 23. J.-W. Oh, I.-W. Lee, I.-T. Kim, and G.-W. Lee: ACI Mater. J., 1999,
neous material as well as for a hard film on a soft substrate. vol. 96 (1), pp. 61-67.
This is demonstrated by generating ANN models with non- 24. R.M. Haj-Ali, D.A. Pecknold, J. Ghaboussi, and G.Z. Voyiadjis: J.
Eng. Mech. Div., ASCE, 2001, vol. 127 (7), pp. 730-38.
linear material parameters in the form of Ramberg–Osgood 25. W.C. Oliver and G.M. Pharr: J. Mater. Res., 1992, vol. 7, p. 1564.
stress-strain relations as well as geometry of the indented 26. Hibbitt, Karlsson, & Sorensen, Inc., ABAQUS, Version 5.8, User’s
material system. Manual, Pawtucket, RI, 1998.